diff --git a/.gitattributes b/.gitattributes index 013573e7a0eaf8da05441a778d335af8bd8881d7..eef34e96fa83d94e78ff9540f96f619963e128e2 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2588,3 +2588,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/V2PE_[[:space:]]Improving[[:space:]]Multimodal[[:space:]]Long-Context[[:space:]]Capability[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models[[:space:]]with[[:space:]]Variable[[:space:]]Visual[[:space:]]Position[[:space:]]Encoding/75b04339-69b5-48e5-a71f-b17b99c6243a_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/V2XPnP_[[:space:]]Vehicle-to-Everything[[:space:]]Spatio-Temporal[[:space:]]Fusion[[:space:]]for[[:space:]]Multi-Agent[[:space:]]Perception[[:space:]]and[[:space:]]Prediction/36b3719d-faf2-4104-a9f7-9253a3958218_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/V2XScenes_[[:space:]]A[[:space:]]Multiple[[:space:]]Challenging[[:space:]]Traffic[[:space:]]Conditions[[:space:]]Dataset[[:space:]]for[[:space:]]Large-Range[[:space:]]Vehicle-Infrastructure[[:space:]]Collaborative[[:space:]]Perception/e74c8f66-2f14-41e8-a64e-91a69d371436_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VA-MoE_[[:space:]]Variables-Adaptive[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]for[[:space:]]Incremental[[:space:]]Weather[[:space:]]Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VACE_[[:space:]]All-in-One[[:space:]]Video[[:space:]]Creation[[:space:]]and[[:space:]]Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VAFlow_[[:space:]]Video-to-Audio[[:space:]]Generation[[:space:]]with[[:space:]]Cross-Modality[[:space:]]Flow[[:space:]]Matching/7812feed-641c-4918-a582-f934ca29c2a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VAGUE_[[:space:]]Visual[[:space:]]Contexts[[:space:]]Clarify[[:space:]]Ambiguous[[:space:]]Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VALLR_[[:space:]]Visual[[:space:]]ASR[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Lip[[:space:]]Reading/88644d63-e953-49ce-b2d3-f0418686524b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VCA_[[:space:]]Video[[:space:]]Curious[[:space:]]Agent[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Understanding/a73c253b-5afa-42a6-b79a-31559f118997_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VEGGIE_[[:space:]]Instructional[[:space:]]Editing[[:space:]]and[[:space:]]Reasoning[[:space:]]Video[[:space:]]Concepts[[:space:]]with[[:space:]]Grounded[[:space:]]Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VFlowOpt_[[:space:]]A[[:space:]]Token[[:space:]]Pruning[[:space:]]Framework[[:space:]]for[[:space:]]LMMs[[:space:]]with[[:space:]]Visual[[:space:]]Information[[:space:]]Flow-Guided[[:space:]]Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VGGSounder_[[:space:]]Audio-Visual[[:space:]]Evaluations[[:space:]]for[[:space:]]Foundation[[:space:]]Models/a6042e74-fb7f-4488-a331-f3b644cd7772_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VGMamba_[[:space:]]Attribute-to-Location[[:space:]]Clue[[:space:]]Reasoning[[:space:]]for[[:space:]]Quantity-Agnostic[[:space:]]3D[[:space:]]Visual[[:space:]]Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VIGFace_[[:space:]]Virtual[[:space:]]Identity[[:space:]]Generation[[:space:]]for[[:space:]]Privacy-Free[[:space:]]Face[[:space:]]Recognition[[:space:]]Dataset/570e97bc-508f-4feb-946d-7b35db443845_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VIPerson_[[:space:]]Flexibly[[:space:]]Generating[[:space:]]Virtual[[:space:]]Identity[[:space:]]for[[:space:]]Person[[:space:]]Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VISION-XL_[[:space:]]High[[:space:]]Definition[[:space:]]Video[[:space:]]Inverse[[:space:]]Problem[[:space:]]Solver[[:space:]]using[[:space:]]Latent[[:space:]]Image[[:space:]]Diffusion[[:space:]]Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VISO_[[:space:]]Accelerating[[:space:]]In-orbit[[:space:]]Object[[:space:]]Detection[[:space:]]with[[:space:]]Language-Guided[[:space:]]Mask[[:space:]]Learning[[:space:]]and[[:space:]]Sparse[[:space:]]Inference/bca87654-a43f-4442-9368-eee092911b00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VITAL_[[:space:]]More[[:space:]]Understandable[[:space:]]Feature[[:space:]]Visualization[[:space:]]through[[:space:]]Distribution[[:space:]]Alignment[[:space:]]and[[:space:]]Relevant[[:space:]]Information[[:space:]]Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VLABench_[[:space:]]A[[:space:]]Large-Scale[[:space:]]Benchmark[[:space:]]for[[:space:]]Language-Conditioned[[:space:]]Robotics[[:space:]]Manipulation[[:space:]]with[[:space:]]Long-Horizon[[:space:]]Reasoning[[:space:]]Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VLDrive_[[:space:]]Vision-Augmented[[:space:]]Lightweight[[:space:]]MLLMs[[:space:]]for[[:space:]]Efficient[[:space:]]Language-grounded[[:space:]]Autonomous[[:space:]]Driving/03e7ad20-412e-4720-b543-ad870d80aec2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VLIPP_[[:space:]]Towards[[:space:]]Physically[[:space:]]Plausible[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Vision[[:space:]]and[[:space:]]Language[[:space:]]Informed[[:space:]]Physical[[:space:]]Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VLM4D_[[:space:]]Towards[[:space:]]Spatiotemporal[[:space:]]Awareness[[:space:]]in[[:space:]]Vision[[:space:]]Language[[:space:]]Models/7ad15710-76ed-462f-ba27-85159b1f2146_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VLR-Driver_[[:space:]]Large[[:space:]]Vision-Language-Reasoning[[:space:]]Models[[:space:]]for[[:space:]]Embodied[[:space:]]Autonomous[[:space:]]Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VLRMBench_[[:space:]]A[[:space:]]Comprehensive[[:space:]]and[[:space:]]Challenging[[:space:]]Benchmark[[:space:]]for[[:space:]]Vision-Language[[:space:]]Reward[[:space:]]Models/8e730ad2-3996-4a2e-9875-e28304c3e917_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VMBench_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Perception-Aligned[[:space:]]Video[[:space:]]Motion[[:space:]]Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VMem_[[:space:]]Consistent[[:space:]]Interactive[[:space:]]Video[[:space:]]Scene[[:space:]]Generation[[:space:]]with[[:space:]]Surfel-Indexed[[:space:]]View[[:space:]]Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VOVTrack_[[:space:]]Exploring[[:space:]]the[[:space:]]Potentiality[[:space:]]in[[:space:]]Raw[[:space:]]Videos[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Multi-Object[[:space:]]Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VOccl3D_[[:space:]]A[[:space:]]Video[[:space:]]Benchmark[[:space:]]Dataset[[:space:]]for[[:space:]]3D[[:space:]]Human[[:space:]]Pose[[:space:]]and[[:space:]]Shape[[:space:]]Estimation[[:space:]]under[[:space:]]real[[:space:]]Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VPO_[[:space:]]Aligning[[:space:]]Text-to-Video[[:space:]]Generation[[:space:]]Models[[:space:]]with[[:space:]]Prompt[[:space:]]Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VPR-Cloak_[[:space:]]A[[:space:]]First[[:space:]]Look[[:space:]]at[[:space:]]Privacy[[:space:]]Cloak[[:space:]]Against[[:space:]]Visual[[:space:]]Place[[:space:]]Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VQ-SGen_[[:space:]]A[[:space:]]Vector[[:space:]]Quantized[[:space:]]Stroke[[:space:]]Representation[[:space:]]for[[:space:]]Creative[[:space:]]Sketch[[:space:]]Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VQ-VLA_[[:space:]]Improving[[:space:]]Vision-Language-Action[[:space:]]Models[[:space:]]via[[:space:]]Scaling[[:space:]]Vector-Quantized[[:space:]]Action[[:space:]]Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VRBench_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Multi-Step[[:space:]]Reasoning[[:space:]]in[[:space:]]Long[[:space:]]Narrative[[:space:]]Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VRM_[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]via[[:space:]]Virtual[[:space:]]Relation[[:space:]]Matching/4752d372-1ed6-459b-bee6-6b202a028bae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VSC_[[:space:]]Visual[[:space:]]Search[[:space:]]Compositional[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Model/9f76eb31-5853-4235-b96e-6306079810f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VSP_[[:space:]]Diagnosing[[:space:]]the[[:space:]]Dual[[:space:]]Challenges[[:space:]]of[[:space:]]Perception[[:space:]]and[[:space:]]Reasoning[[:space:]]in[[:space:]]Spatial[[:space:]]Planning[[:space:]]Tasks[[:space:]]for[[:space:]]MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VSRM_[[:space:]]A[[:space:]]Robust[[:space:]]Mamba-Based[[:space:]]Framework[[:space:]]for[[:space:]]Video[[:space:]]Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VSSD_[[:space:]]Vision[[:space:]]Mamba[[:space:]]with[[:space:]]Non-Causal[[:space:]]State[[:space:]]Space[[:space:]]Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VTimeCoT_[[:space:]]Thinking[[:space:]]by[[:space:]]Drawing[[:space:]]for[[:space:]]Video[[:space:]]Temporal[[:space:]]Grounding[[:space:]]and[[:space:]]Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Vamba_[[:space:]]Understanding[[:space:]]Hour-Long[[:space:]]Videos[[:space:]]with[[:space:]]Hybrid[[:space:]]Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Variance-Based[[:space:]]Pruning[[:space:]]for[[:space:]]Accelerating[[:space:]]and[[:space:]]Compressing[[:space:]]Trained[[:space:]]Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Vector[[:space:]]Contrastive[[:space:]]Learning[[:space:]]For[[:space:]]Pixel-Wise[[:space:]]Pretraining[[:space:]]In[[:space:]]Medical[[:space:]]Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VehicleMAE_[[:space:]]View-asymmetry[[:space:]]Mutual[[:space:]]Learning[[:space:]]for[[:space:]]Vehicle[[:space:]]Re-identification[[:space:]]Pre-training[[:space:]]via[[:space:]]Masked[[:space:]]AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Verbalized[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Interpretable[[:space:]]Few-Shot[[:space:]]Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Versatile[[:space:]]Transition[[:space:]]Generation[[:space:]]with[[:space:]]Image-to-Video[[:space:]]Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VertexRegen_[[:space:]]Mesh[[:space:]]Generation[[:space:]]with[[:space:]]Continuous[[:space:]]Level[[:space:]]of[[:space:]]Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViCTr_[[:space:]]Vital[[:space:]]Consistency[[:space:]]Transfer[[:space:]]for[[:space:]]Pathology[[:space:]]Aware[[:space:]]Image[[:space:]]Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViLLa_[[:space:]]Video[[:space:]]Reasoning[[:space:]]Segmentation[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Model/9374d1a1-c3e0-498b-83e7-5023233fa5ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViLU_[[:space:]]Learning[[:space:]]Vision-Language[[:space:]]Uncertainties[[:space:]]for[[:space:]]Failure[[:space:]]Prediction/9693ec7c-e63e-4549-b95a-8e863d6bee58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViM-VQ_[[:space:]]Efficient[[:space:]]Post-Training[[:space:]]Vector[[:space:]]Quantization[[:space:]]for[[:space:]]Visual[[:space:]]Mamba/bb5b37a9-8d33-488a-b6b3-ef3b094ba666_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViSpeak_[[:space:]]Visual[[:space:]]Instruction[[:space:]]Feedback[[:space:]]in[[:space:]]Streaming[[:space:]]Videos/4c2298b6-a1cd-4d0c-bdd9-742253ff8150_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViT-EnsembleAttack_[[:space:]]Augmenting[[:space:]]Ensemble[[:space:]]Models[[:space:]]for[[:space:]]Stronger[[:space:]]Adversarial[[:space:]]Transferability[[:space:]]in[[:space:]]Vision[[:space:]]Transformers/9e6fb5c3-a13e-4a97-9aed-2462169214a0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViT-Linearizer_[[:space:]]Distilling[[:space:]]Quadratic[[:space:]]Knowledge[[:space:]]into[[:space:]]Linear-Time[[:space:]]Vision[[:space:]]Models/4d67ff7b-60ce-453d-9e8b-8fe983e07bca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ViT-Split_[[:space:]]Unleashing[[:space:]]the[[:space:]]Power[[:space:]]of[[:space:]]Vision[[:space:]]Foundation[[:space:]]Models[[:space:]]via[[:space:]]Efficient[[:space:]]Splitting[[:space:]]Heads/62f8e79c-d973-4405-a368-6eb7d50f7246_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Vid-Group_[[:space:]]Temporal[[:space:]]Video[[:space:]]Grounding[[:space:]]Pretraining[[:space:]]from[[:space:]]Unlabeled[[:space:]]Videos[[:space:]]in[[:space:]]the[[:space:]]Wild/b7f7fdbe-e75b-4823-a9b7-19405b49fb65_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Video[[:space:]]Color[[:space:]]Grading[[:space:]]via[[:space:]]Look-Up[[:space:]]Table[[:space:]]Generation/b0b9c231-2082-4b4c-ad77-337ee0cc9045_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Video[[:space:]]Individual[[:space:]]Counting[[:space:]]for[[:space:]]Moving[[:space:]]Drones/1cf4d87f-9c8a-44cd-a5ee-135b6f1daa18_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Video[[:space:]]Motion[[:space:]]Graphs/78363dc2-8dfc-4b9b-80e6-1323ef2ac37e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Video-T1_[[:space:]]Test-time[[:space:]]Scaling[[:space:]]for[[:space:]]Video[[:space:]]Generation/1019b1f8-19d0-4e70-9f39-8bbdb082c721_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Video2BEV_[[:space:]]Transforming[[:space:]]Drone[[:space:]]Videos[[:space:]]to[[:space:]]BEVs[[:space:]]for[[:space:]]Video-based[[:space:]]Geo-localization/1c533f62-65f2-4dff-a54f-3aa8e0de6fdc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VideoAds[[:space:]]for[[:space:]]Fast-Paced[[:space:]]Video[[:space:]]Understanding/8a639327-709f-4460-aabc-c77b069d813b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VideoAuteur_[[:space:]]Towards[[:space:]]Long[[:space:]]Narrative[[:space:]]Video[[:space:]]Generation/31b80179-4ba4-49ae-a3df-b51fb25c8cd6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VideoLLaMB_[[:space:]]Long[[:space:]]Streaming[[:space:]]Video[[:space:]]Understanding[[:space:]]with[[:space:]]Recurrent[[:space:]]Memory[[:space:]]Bridges/ef7f527a-2c19-4223-8b21-ac615c5af575_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VideoMiner_[[:space:]]Iteratively[[:space:]]Grounding[[:space:]]Key[[:space:]]Frames[[:space:]]of[[:space:]]Hour-Long[[:space:]]Videos[[:space:]]via[[:space:]]Tree-based[[:space:]]Group[[:space:]]Relative[[:space:]]Policy[[:space:]]Optimization/d7b4639d-d9a0-4b6c-986f-11dd15c86ebf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VideoOrion_[[:space:]]Tokenizing[[:space:]]Object[[:space:]]Dynamics[[:space:]]in[[:space:]]Videos/872b582e-2d26-46e1-93e2-773aad7bd594_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VideoRFSplat_[[:space:]]Direct[[:space:]]Scene-Level[[:space:]]Text-to-3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Generation[[:space:]]with[[:space:]]Flexible[[:space:]]Pose[[:space:]]and[[:space:]]Multi-View[[:space:]]Joint[[:space:]]Modeling/0f70facc-7c84-4b86-b0dc-d3d1bdbeaad1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/VideoSetDiff_[[:space:]]Identifying[[:space:]]and[[:space:]]Reasoning[[:space:]]Similarities[[:space:]]and[[:space:]]Differences[[:space:]]in[[:space:]]Similar[[:space:]]Videos/3d626d6f-d212-46b3-bc48-4c5f49732e8e_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_content_list.json b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..52f5e468dccd982fa1cf50dd64ad24d4f5ce4460 --- /dev/null +++ b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_content_list.json @@ -0,0 +1,1781 @@ +[ + { + "type": "text", + "text": "VA-MoE: Variables-Adaptive Mixture of Experts for", + "text_level": 1, + "bbox": [ + 233, + 130, + 764, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Incremental Weather Forecasting", + "text_level": 1, + "bbox": [ + 308, + 162, + 651, + 184 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/408a686bb8235a3b61e408da6aca9f7a13881c231183d5f5fbc8364f783e3092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 151, + 687, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Chen $^{1\\dagger}$ Han Tao $^{1\\dagger}$ Guo Song $^{1*}$ Jie Zhang $^{1}$ Yonghan Dong $^{2}$ Yunlong Yu $^{3}$ Lei Bai $^{4}$ $^{1}$ Hong Kong University of Science and Technology (HKUST) $^{2}$ Huawei Technologies Ltd. \n $^{3}$ Zhejiang University $^{4}$ Shanghai AI Laboratory", + "bbox": [ + 137, + 210, + 893, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Equal contribution *Corresponding author {hchener, thanad}@connect.ust.hk", + "bbox": [ + 181, + 268, + 841, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{songguo, csejzhang}@ust.hk dongyonghan@huawei.com yuyunlong@zju.edu.cn bailei@pjlab.org.cn", + "bbox": [ + 101, + 285, + 918, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 335, + 326, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents Variables-Adaptive Mixture of Experts (VA-MoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real-time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VA-MoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct sub-patterns of atmospheric variables (e.g., temperature, humidity, wind speed). Moreover, the proposed method employs a variable-adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on ERA5 dataset demonstrate that VA-MoE performs comparable against state-of-the-art models in both short-term (e.g., 1-3 days) and long-term (e.g., 5 days) forecasting tasks, with only about $25\\%$ of trainable parameters and $50\\%$ of the initial training data. Code: https://github.com/chenhao-zju/VAMoE", + "bbox": [ + 89, + 369, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weather forecasting remains one of the most critical scientific challenges with profound socio-economic impacts. While traditional Numerical Weather Prediction (NWP) [2] relies on solving complex partial different equations to simulate atmospheric dynamics, recent advances in data-driven artificial intelligence [3, 4, 9, 22] have revolutionized earth system modeling. Deep learning models, in particular, have demonstrated remarkable potential in capturing diverse spatiotemporal patterns across weather and climate phenomena.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a597efb3a93059332d6e7ca7100ecc4ea802f35eb2a605ebaae8713017f4c997.jpg", + "image_caption": [ + "Figure 1. Illustration of two different training paradigms of weather forecasting when new variables arrive: (c) Full-Train the whole model (d) Incremental-Train only the module corresponding new variables. (b) Performances of these two paradigms." + ], + "image_footnote": [], + "bbox": [ + 517, + 333, + 903, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ena, giving rise to the emerging field of AI for Weather.", + "bbox": [ + 511, + 566, + 880, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, a fundamental limitation of existing AI4Weather approaches lies in their assumption that all variables are available synchronously during training and inference. In reality, meteorological variables are often heterogeneous in terms of data sources, collection frequencies, and spatial-temporal distributions. For instance, the upper-air variables (e.g., temperature profiles) are sparse and sampled via radiosondes/satellites, while the surface variable (e.g., precipitation, wind) are dense but updated in near-real-time. This asynchrony poses significant challenges: when introducing new variables (e.g., satellite-derived aerosol data), existing models must be entirely retrained from scratch, incurring prohibitive computational costs. For example, Pangu [3] required 64 days and 192 V100 GPUs for a full retraining cycle.", + "bbox": [ + 509, + 583, + 906, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this challenge, we propose Incremental Weather Forecasting (IWF), a novel paradigm designed to dynamically expand meteorological forecast variables in response to increasing observational data. By enabling parameter-efficient variable scaling, IWF revolutionizes operational meteorology by offering a streamlined alternative", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7915", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to conventional models. This innovation tackles a critical limitation in the field: the need for scalable modeling solutions as datasets grow gradually, ensuring adaptability without compromising computational efficiency.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A pivotal limitation of existing incremental approaches is catastrophic forgetting, i.e., when introducing new variables, pre-trained parameters drift towards new distributions, causing significant performance degradation on original variables. To overcome this, we present the Variables-Adaptive Mixture of Experts (VA-MoE) model, the first expert-based system explicitly designed for hierarchical atmospheric variables.", + "bbox": [ + 89, + 154, + 480, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture of Experts (MoE) models inherently enable scalable representation learning for multi-variable atmospheric datasets by dynamically expanding expert modules as input dimensions grow. While traditional MoE architectures excel in representational capacity, their parallelized design which distributes uniform weights across experts, often leads to homogeneous outputs, limiting specialized knowledge acquisition. To address this limitation, we introduce VA-MoE, an auxiliary-loss-free framework that decouples expert specialization from computational efficiency. VA-MoE employs a phased training strategy: Initial experts are pretrained on base variables, then incrementally expanded to accommodate new variables. Crucially, the pretrained experts are frozen during expansion phases to prevent catastrophic forgetting, ensuring robust retention of prior expertise while adapting to novel inputs. To foster expert diversity without auxiliary losses, VA-MoE incorporates variable index embeddings, i.e., positional metadata encodings that guide experts to develop domain-specific specialization. These embeddings dynamically activate contextually relevant computational pathways for each variable, optimizing resource allocation at inference time. By integrating variable index embeddings that encode spatial-temporal metadata, VA-MoE directs experts to specialize in specific atmospheric domains. This context-aware routing reduces inference latency compared to static MoE models while maintaining expert diversity.", + "bbox": [ + 91, + 276, + 483, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides, we propose a gradient-scaled variable-adaptive loss function that dynamically aligns variable optimization rates with their inherent spatiotemporal characteristics. By quantifying the distinct temporal evolution patterns of atmospheric variables, our method allocates differential loss weights, i.e., larger gradient magnitudes prioritize fast-transient fields like temperature to capture immediate atmospheric dynamics, while gradual weight adjustments stabilize slow-changing fields such as geopotential height to preserve long-term system behavior.", + "bbox": [ + 89, + 686, + 480, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, our paradigm employs a two-phase training paradigm: an initial phase optimizing base atmospheric variables, followed by an incremental phase integrating new variables while freezing previously trained ex", + "bbox": [ + 89, + 839, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perts. Experiments with the same 40-year dataset reveal that our paradigm Fig. 1(a)+(d) achieves superior z500 predictions performance compared to full retraining Fig. 1(a)+(c).", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, the contributions of this work include:", + "bbox": [ + 532, + 137, + 883, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- This work initiates systematic research on incremental learning paradigms for weather forecasting. We propose the quantitative benchmark to evaluate the trade-offs between model scalability and generalization in incremental weather modeling.", + "- We present Variables-Adaptive Mixture of Experts (VA-MoE), the first framework tailored for incremental atmospheric modeling. VA-MoE achieves expert specialization through contextual variable activation driven by variable index embeddings, enabling dynamic assignment of experts to variables during both training and inference.", + "- Extensive experiments on the ERA5 dataset demonstrate that VA-MoE achieves comparable performance for surface variables, while delivering superior accuracy in upper-air variables against the existing competitors under $50\\%$ reduced dataset size and $25\\%$ fewer parameters." + ], + "bbox": [ + 511, + 152, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 409, + 650, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Data-Driven Weather Forecasting", + "text_level": 1, + "bbox": [ + 513, + 435, + 808, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent decades, traditional NWP [2] methods have dominated the weather forecasting field due to the robust predictions and rigorous mathematical validation. However, NWP methods [26, 33] require training from scratch for new prediction, resulting in slow computation and high costs.", + "bbox": [ + 511, + 457, + 903, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in deep learning have catalyzed a paradigm shift toward data-driven models for medium-range weather forecasting. Pioneering works like FourCastNet [22] leverage Fourier Neural Operators (FNOs) [24] to learn spatiotemporal weather patterns, while PanguWeather [3] employs a 3D vision transformer [13, 25] to model atmospheric dynamics. Subsequent innovations, including Neural Operators [5, 24, 38] and specialized architectures [9, 10, 15, 23, 31, 43, 44, 46], have achieved accuracy rivaling NWP at a fraction of computation.", + "bbox": [ + 511, + 534, + 905, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While existing data-driven models excel at fixed-variable forecasting after costly initial training, their inflexibility poses a critical barrier: retraining from scratch is required to incorporate new variables, rendering cross-domain adaptation impractical. In contrast, our work introduces incremental weather forecasting, a novel paradigm evolving spatiotemporal patterns in real-time data. By decoupling variable-specific expertise from shared dynamics, our approach supports dynamic expansion to new atmospheric parameters with minimal retraining overhead.", + "bbox": [ + 511, + 685, + 905, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Mixture of Experts for Incremental Learning", + "text_level": 1, + "bbox": [ + 511, + 847, + 897, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advances in deep learning[6, 17] have spurred interest in parameter-efficient transfer strategies, where models adapt", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7916", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f6c306f6dc259a2ccf6fca6af91be6c6aae1b11d49c857416c88197d46d07c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 92, + 459, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/124c88fa5df353a7a3ccd95e9e7a9a285c004e52ab5edbcdfdd0876865d9dbfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 93, + 880, + 180 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d77f605a7ffeb8dd32d6de34d598d180acb7152790dd370ed404951bb38af79d.jpg", + "image_caption": [ + "Figure 2. Illustration of the incremental weather forecasting paradigm. (a), (b) are the initial and incremental stages, respectively. (c) presents the detailed structure in both initial and incremental stages. During the initial stage, the model is trained with upper-air variables(UV). In the incremental stage, the fire components are trained while the snow components remain frozen. The index embedding is trained by the Index Encoder and divided into six parts based on the upper-air variables (UV: Z, Q, U, V, T) and surface variables (SV)." + ], + "image_footnote": [], + "bbox": [ + 116, + 181, + 880, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to new domains by fine-tuning only a subset of parameters. A key paradigm, incremental learning, addresses the stability-plasticity dilemma: retaining knowledge of prior tasks while integrating new concepts. However, real-world data often involves domain shifts, such as variations in variable distributions or task objectives, leading to two dominant research threads. Class Incremental Learning (CIL) [16, 21, 35, 47], focuses on expanding classification tasks with new classes while Task Incremental Learning (TIL) [29, 34, 40, 42] adapts models to distinct task sequences.", + "bbox": [ + 88, + 492, + 480, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One of the most popular structures for addressing incremental learning tasks is the Mixture of Experts (MoE) model [19], which learns representations of new concepts through different experts. Due to its sparse architecture, many works have adopted the MoE structure to reduce inference costs and enhance model capacity [37, 41]. An early work introducing MoE to incremental learning is Expert Gate [1], which trains multiple backbones as different experts and assigns new domains to the relevant expert. Lifelong-MoE [11] leverages pretrained experts and gates to retain the representational knowledge of the training domain. In addition to these approaches, MoE is used as an adapter in MoE-Adapter [45], where adapters are attached to the main structure. These works have demonstrated the promising performance of MoE in visual and natural language incremental learning.", + "bbox": [ + 89, + 645, + 482, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to MoE designed for fixed-variable systems,", + "bbox": [ + 109, + 886, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "our propose incremental learning of variable distributions by assigning distinct pressure-level groups to specialized experts. Unlike prior incremental learning literature that primarily focus on task or dataset-specific adaptation, our incremental paradigm targets variable-centric expansion, enabling models to adapt to evolving observational data without retraining from scratch.", + "bbox": [ + 511, + 492, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 613, + 648, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work focuses on a novel paradigm of incremental weather forecasting with VA-MoE. In this section, we first formally define the problem in Sec. 3.1. Then, we present the structure of VA-MoE, consisting with the Channel-Adaptive Expert (CAE) and the Shared Expert in Sec. 3.2. The model during incremental stage is introduced in Sec. 3.3. Finally, loss function is presented in Sec. 3.4.", + "bbox": [ + 511, + 638, + 906, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 756, + 624, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the weather forecasting task, the AI-based model $\\Phi$ aims to predict the future weather variables $\\mathbf{X}^{t + 1}$ with the historical weather variables $\\mathbf{X}^t$ , i.e., $\\mathbf{X}^{t + 1} = \\Phi (\\mathbf{X}^t)$ .", + "bbox": [ + 511, + 777, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the incremental weather forecasting task, the weather variables $\\mathbf{X}^t$ are divided into two sets: (i) the initial training variables at $t$ time $\\mathbf{X}_h^t \\in \\mathbb{R}^{H \\times W \\times N}$ , and, (ii) the incremental variables at $t$ time $\\mathbf{X}_{sv}^t \\in \\mathbb{R}^{H \\times W \\times M}$ . Unlike traditional weather forecasting tasks, where models are trained", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7917", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ec231ab293262e1e38372cb270935bd9ebecfb3ad7ef96b9c46580860695194a.jpg", + "image_caption": [ + "Figure 3. Illustration of the VA-MoE. In the left subgraph, the input features comprise both upper-air and surface variables. During the initial training stage, five distinct CAE modules process the upper-air variables. In the incremental stage, surface variables are handled by a dedicated module, $\\mathrm{CAESV}$ , while the original five CAEs remain frozen to preserve learned representations. The right subgraph details the CAE module. The Index Embedding is selected through a multi-layer process and applied to weight the Input Embedding via multiplication. Initially discrete, the index embedding is transformed into a continuous representation after passing through the GateEmbed layer, enabling it to dynamically reflect the relevance of specific channels to the input variables." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with all variables available simultaneously, the incremental weather forecasting task sequentially incorporates dynamic variables as they become available over time. The incremental weather forecasting process is structured into a sequence of learning stages: an initial training stage followed by subsequent training stages.", + "bbox": [ + 89, + 470, + 483, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the initial training stage, the available variables of 40 years are used to train a weather model, i.e., $\\mathrm{X}_h^{t + 1} = \\Phi_h(\\mathrm{X}_h^t)$ , as shown in Fig. 2(a). In the incremental training stage, the half variables used at the initial stage and the surface variables at the incremental stage are incorporated together, to train the newly added modules for the surface variables while keeping the model parameters from the initial stage frozen, i.e., $\\mathrm{X}^{t + 1} = \\Phi (\\mathrm{X}^t)$ , where $\\Phi$ consists of both the fixed parameters of the pre-trained model $\\Phi_h$ and the trainable parameters of $\\Phi_{sv}$ , $\\mathrm{X}^t$ consists of all the available variables at time $t$ , $\\mathrm{X}^{t + 1}$ consists of the predicted variables at time $t + 1$ , as shown in Fig. 2(b). In this work, the upper-air variables are utilized during the initial training stage while the surface variables become available and are incorporated during the incremental training stage. The upper-air variables include five different types, Z, Q, U, V, and T, each defined across 13 different levels.", + "bbox": [ + 89, + 563, + 483, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. VA-MoE", + "text_level": 1, + "bbox": [ + 89, + 830, + 194, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce index embedding within transformer blocks, a novel mechanism to dynamically guide experts in learning hierarchical relationships between meteorological vari", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ables. As new variables are incrementally integrated, corresponding experts are added to the transformer architecture and optimized via index-based affinity assignments. Building on this framework, we adapt the Mixture-of-Experts (MoE) paradigm which enables task-specific specialization and flexible integration of new experts—into weather forecasting transformers.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The structure of the proposed VA-MoE is illustrated in Fig. 2(c). The input features, extracted by the encoder, first pass through a normalization layer and a self-attention layer. The output of the self-attention layer is then combined with the residual connection. Then, another normalization layer is applied, followed by a VA-MoE module, which is also equipped with a residual connection. This process can be formulated as:", + "bbox": [ + 511, + 196, + 906, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {X} _ {m i d} = \\mathbf {X} _ {i n} + \\operatorname {S A} (\\text {L a y e r N o r m} (\\mathbf {X} _ {\\text {i n}})), \\\\ \\mathbf {X} _ {i n} = \\mathbf {X} _ {i n - 1} + \\operatorname {V A M E} (\\mathbf {I} _ {i n - 1}, \\mathbf {N} _ {i n - 1} (\\mathbf {X} _ {i n})) \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 328, + 903, + 353 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\text {o u t}} = \\mathbf {X} _ {\\text {m i d}} + \\operatorname {V A - M o E} (\\operatorname {L a y e r N o r m} (\\mathbf {X} _ {\\text {m i d}})),\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 349, + 854, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_{in}$ and $\\mathbf{X}_{out}$ are the input and the output of each Transformer Block, respectively.", + "bbox": [ + 511, + 375, + 906, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Variables and Index Embedding", + "text_level": 1, + "bbox": [ + 511, + 416, + 784, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, during the training stage, the input feature $\\mathbf{X}_h^t\\in \\mathbb{R}^{H\\times W\\times N}$ consists of five different types of variables: $\\mathbf{X}_Z^t,\\mathbf{X}_Q^t,\\mathbf{X}_U^t,\\mathbf{X}_V^t,\\mathbf{X}_T^t\\in \\mathbb{R}^{H\\times W\\times (N / 5)}$ , where, $H\\times W$ denotes the spatial resolution. These five types of variables are concatenated and processed by a shared expert and five distinct Channel-Adaptive Experts (CAEs).", + "bbox": [ + 511, + 435, + 906, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to these variables, the framework introduces an extra one-hot index embedding $\\mathbf{I}_h\\in \\mathbb{R}^{5\\times N}$ to guide experts in learning variable affinity. The number $N$ corresponds to the number of variables, and 5 corresponds to the number of variable types. Similar to the input weather features, the index embedding is composed of five separate embeddings: $\\mathbf{I}_Z,\\mathbf{I}_Q,\\mathbf{I}_U,\\mathbf{I}_V,\\mathbf{I}_T\\in \\mathbb{R}^{1\\times 1\\times N}$ . This initial index embedding is then encoded into a latent space using a Linear layer, expressed as $\\mathbf{I}_h = \\mathrm{Linear}(\\mathbf{I}_h)$ .", + "bbox": [ + 511, + 527, + 905, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Channel-Adaptive Expert (CAE)", + "text_level": 1, + "bbox": [ + 511, + 671, + 790, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key design of VA-MoE is the channel-adaptive expert, as shown in Fig. 3. For each variable $Z$ , the corresponding expert module processes both the feature embeddings of the input variables $\\mathbf{X}_h^t$ and the index embedding $\\mathbf{I}_Z$ , which is formulated as:", + "bbox": [ + 511, + 691, + 905, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 777, + 903, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{CAE}_Z$ is the abbreviation of channel-adaptive expert module for variable $Z$ .", + "bbox": [ + 511, + 809, + 903, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the CAE module for variable $Z$ , the index embedding $\\mathbf{I}_Z$ is encoded through a linear layer for feature projection and then multiplied channel-wise with the input embedding $\\mathbf{X}_h^t$ . The resulting fused feature is processed", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7918", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by the GateEmbed layer, which produces a gate embedding to enhance the fusion of the index and variable embeddings. This enhanced feature is then normalized using a SoftMax layer and filtered through a TopK operation, which selects the top-K high-rank channels. This process yields the GateIndex $\\mathrm{GI}_Z\\in \\mathbb{R}^{H\\times W\\times K}$ and GateWeight $\\mathrm{GW}_Z\\in \\mathbb{R}^{H\\times W\\times K}$ . The process is formulated as:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {Z} ^ {t o p k}, \\mathbf {W} _ {Z} ^ {t o p k} = \\mathrm {T O P} _ {k} (\\operatorname {S o f t M a x} (\\operatorname {M L P} _ {Z} (\\mathbf {X} _ {h} ^ {t} \\odot \\mathbf {I} _ {Z}))), (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 208, + 483, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_Z^{topk}$ and $\\mathbf{W}_Z^{topk}$ denote the index and weight of the Top-k features. Once the GateIndex and GateWeight are computed, the input embedding is selectively processed based on the GateIndex, while the variable embedding is weighted channel-wise according to the GateWeight. The GateEmbed layer uses variable embeddings to guide the index embedding, enabling the creation of a variable-specific confidence matrix. Within the CAE module, this embedding enhances the confidence matrix by incorporating both semantic and positional information, rather than relying solely on positional information.", + "bbox": [ + 89, + 239, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The selected features for variable $Z$ are obtained with:", + "bbox": [ + 109, + 409, + 470, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {Z} ^ {t, s e l e c t e d} = \\mathbf {W} _ {Z} ^ {t o p k} * \\operatorname {S e l e c t} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z} ^ {t o p k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 435, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Select}_Z$ denotes selecting Top-K embedding from the input feature $\\mathbf{X}_h^t$ within the $\\mathrm{CAE}_Z$ module.", + "bbox": [ + 89, + 467, + 482, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The selected and weighted embedding captures the key features associated with the specific type of variables. In the $\\mathrm{CAE}_Z$ module, these features serve as the training set to model the distribution of the variable $Z$ with $\\mathrm{Expert}_Z$ . The expert module adopts a parallel structure, splitting the channels into multiple parts and enhancing each part with a sparse MoE. This process is formulated as:", + "bbox": [ + 89, + 498, + 483, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {E x p e r t} _ {Z} (\\mathbf {X} _ {Z} ^ {t, s e l e c t e d}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 614, + 482, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Expert}_Z$ is implemented with a multiple layer.", + "bbox": [ + 89, + 647, + 452, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Shared Expert.", + "text_level": 1, + "bbox": [ + 89, + 672, + 243, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The shared expert processes the overall features of all variables and operates in parallel with the variable-specific modules. Each variable is handled by its corresponding CAE module, and their outputs are summarized to produce the fused features $\\mathbf{X}_h^{t,fused}$ for all variables. Since the channel count of the selected variable embeddings is lower than that of the original embeddings, an additional up-channel linear layer, $\\mathrm{Linear}_{\\mathrm{up}}$ , is introduced. The fused features and the output from the shared expert are then combined via pixel-wise addition. This process is expressed as:", + "bbox": [ + 89, + 691, + 483, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {h} ^ {t, f u s e d} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right) + \\dots + \\operatorname {C A E} _ {T} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {T}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 853, + 482, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {X} ^ {\\prime}\\right) _ {h} ^ {t} = \\operatorname {E x p e r t} _ {s h a r e d} \\left(\\mathbf {X} _ {h} ^ {t}\\right) + \\operatorname {L i n e a r} _ {u p} \\left(\\mathbf {X} _ {h} ^ {t, f u s e d}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 882, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Incremental Learning Stage", + "text_level": 1, + "bbox": [ + 513, + 90, + 767, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When new variables are incrementally introduced to the model, specialized experts are dynamically integrated into the transformer blocks to process these variables. Concurrently, both the encoder and decoder architectures undergo synchronized updates to maintain compatibility with the expanded variable set.", + "bbox": [ + 511, + 112, + 906, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Encoder and Decoder Modules.", + "text_level": 1, + "bbox": [ + 511, + 210, + 779, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When $K$ new variables are introduced incrementally, the encoder's input dimension expands to $H \\times W \\times (N + M)$ , necessitating an adjustment to the convolutional layer's kernel from $(3, 3, N, C)$ to $(3, 3, N + M, C)$ . To accommodate this, the layer's parameters are partitioned into two components: The $(3, 3, N, C)$ segment retains pretrained weights from earlier stages, ensuring continuity. The $(3, 3, M, C)$ segment, corresponding to the new variables, is initialized randomly to learn emergent patterns.", + "bbox": [ + 511, + 229, + 906, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The same strategy is applied to the position embedding and the Decoder module.", + "bbox": [ + 511, + 367, + 905, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. VA-MoE Module.", + "text_level": 1, + "bbox": [ + 513, + 405, + 681, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the incremental learning phase, $M$ new experts are dynamically integrated into the MoE layer of each transformer block to accommodate $M$ newly introduced variables. In this phase, only the new experts and the shared expert are actively trained, while all preexisting experts remain frozen to preserve previously learned knowledge.", + "bbox": [ + 511, + 424, + 905, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Index Embedding Module.", + "text_level": 1, + "bbox": [ + 513, + 523, + 745, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When $M$ new variables become available during the incremental stage, the index embedding changes from $\\mathbf{I}_h\\in$ $\\mathbb{R}^{(N\\times l)\\times N}$ to $\\mathbf{I}_h\\in \\mathbb{R}^{(N\\times l + M\\times r)\\times (N + M)}$ , where each $M$ new variables has $r$ levels. Additionally, the index encoder needs to be retrained.", + "bbox": [ + 511, + 542, + 906, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Objective Function", + "text_level": 1, + "bbox": [ + 513, + 628, + 696, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the model, we introduce a novel dynamic prediction loss combined with a reconstruction loss.", + "bbox": [ + 511, + 650, + 905, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Dynamic Prediction Loss.", + "text_level": 1, + "bbox": [ + 511, + 689, + 736, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prediction loss aims to minimize the difference between the model's predicted results and the actual future weather variables. Unlike existing weather forecasting methods that treat all the variables equally, we argue that different variables, such as temperature and surface pressure, follow distinct distributions. Thus, we introduce a novel dynamic prediction loss that dynamically assigns a weight to each channel of the input data. This loss is formulated as:", + "bbox": [ + 511, + 708, + 905, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO b j _ {p r e d} = \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) / e ^ {\\mathbf {w}} + \\mathbf {w}, (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 839, + 906, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes Hadamard product, $e$ denotes the base of the natural logarithm, $\\mathbf{X}^{t + 1}\\in \\mathbb{R}^{H\\times W\\times C}$ , and $\\mathbf{w}\\in$", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7919", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/041e988ddd89d40a80ac5822062df3d6dd06335024de3d68c9678d208845afb6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NameDescriptionLevels
Upper-Air
ZGeopotential13
QSpecific humidity13
Ux-direction wind13
Vy-direction wind13
TTemperature13
Surface-Incremental Learning
u10x-direction wind at 10m heightSingle
v10y-direction wind at 10m heightSingle
t2mTemperature at 2m heightSingle
mslMean sea-level pressureSingle
spSurface pressureSingle
", + "bbox": [ + 125, + 88, + 455, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{1\\times 1\\times C}$ is a learnable vector. With this loss function, the model could dynamically prioritize variables based on their distributional characteristics.", + "bbox": [ + 89, + 373, + 483, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.2. Reconstruction Loss.", + "text_level": 1, + "bbox": [ + 89, + 428, + 282, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We introduce an additional reconstruction loss that connects the encoder and decoder directly, formulated as:", + "bbox": [ + 89, + 446, + 483, + 477 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nO b j _ {r e c o n} = \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 487, + 482, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}^t = \\mathrm{Dec}(\\mathrm{Enc}(\\mathbf{X}^t))$ represents the reconstructed variables. With the loss function, the encoder and decoder modules specialize in encoding and decoding features, while the intermediate transformer blocks focus on learning the data distribution. This ensures that the decoding process is entirely handled by decoder, allowing the intermediate blocks to focus on their task and not participate in decoding, thereby optimizing network efficiency.", + "bbox": [ + 89, + 516, + 483, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, the model is trained with the final objective function, a combination of the dynamic prediction loss and reconstruction loss. The final objective is formulated as:", + "bbox": [ + 89, + 638, + 483, + 684 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nO b j _ {f i n a l} = O b j _ {p r e d} + \\lambda O b j _ {r e c o n}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 696, + 482, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ denotes a hyper-parameter.", + "bbox": [ + 89, + 724, + 328, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 753, + 223, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. In this work, we conduct experiments on a popular weather dataset, i.e., ERA5[18], provided by the ECMWF. ERA5 dataset is a reanalysis atmospheric dataset, consisting of the atmospheric variables from 1979 to the present day with a $0.25^{\\circ}$ spatial resolution with $721 \\times 1440$ . We train the model on 40-year dataset in the initial stage and continuously train the model on 20-year dataset in the incremental stage, which contains the weather variables from", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c32d76ef3271d634424358f221973a4c9a674afb759538b61aacedf3e343fc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 705, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0d1aa500d574a7e95efd8c0c9b28ef1e85a3a448b254b80980622a66a0bc2a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 90, + 903, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/052ad5c9ca6a4a26d4335087addf28eac50c9b1211888b4204be97c3469dbdba.jpg", + "image_caption": [ + "Figure 4. Comparative analysis of RMSE $\\downarrow$ across 10 data-driven models for four variables, including Z500 and T850 (upper-air variables) in the initial stage, as well as T2M and U10 (incremental surface variables) in the incremental stage." + ], + "image_footnote": [], + "bbox": [ + 517, + 219, + 705, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bd3379cdf299482f4eddfea62f542bb052f9854ee8afec711d077e1cd24bf117.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 219, + 903, + 344 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5c27927caadf954f2ab0a6c58768eeb365f2cd7f4cb19f217f1e1d4f1c9d5cd4.jpg", + "table_caption": [ + "Table 1. A summary of atmospheric variables. The 13 levels are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 hPa. 'Single' denotes the variables under earth's surface." + ], + "table_footnote": [], + "table_body": "
——ClimaX——EWMoE——WeatherGFT——Stormer
——FourCastNet——ClimODE——GraphCast——Ours(IL)
——Panguo-Weather——Keisler
", + "bbox": [ + 540, + 345, + 887, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1979 to 2020 and 2000 to 2020 year, respectively. And we test the model on one-year dataset of weather variables in 2021. In this work, the model processes 5 upper-air variables and 5 surface variables as Tab. 1, where the upper-air variables are exploited in the training stage, and the surface variables are exploited in the incremental stage.", + "bbox": [ + 511, + 455, + 905, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The main structure of this work follows the novel backbones [7, 8, 12]. We apply the AdamW optimizer with 0.0002 and 0.00005 learning rates to initial and incremental stages, respectively. In two stages, we train 100 epochs and set batch size to 16. Our model is trained with PyTorch using 16 A100 GPUs.", + "bbox": [ + 511, + 547, + 905, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 650, + 651, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 presents a quantitative comparison of our structure with existing state-of-the-art approaches, including ClimaX [27], Pangu-Weather [3], ClimODE [36], WeatherGFT [39], FourCastNet [22], EWMoE [14], Keisler [20], GraphCast [23], and Stormer [28], for the 5-day weather prediction task. All the competitors are trained with both upper-air and surface variables on a $0.25^{\\circ}$ weather dataset with a resolution of $721 \\times 1440$ .", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the surface variables, including T2M and U10, our VA-MoE demonstrates performance similar to the leading methods Stormer [28] and Graphcast [23]. Compared to other competitors, our structure shows significant advantages in both short-term and long-term predictions. For the upper-air variable Z500, our structure delivers one of the best performances among all 10 methods, outperforming", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7920", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e1e47cf80a602d88639422c583130b60dbcfc1b6d6cbe0dc96828b9436f47f10.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset(years)Iteration\\( (×10^4) \\)T2M(K)↓U10(m/s)↓V10(m/s)↓MSL(Pa)↓SP(Pa)↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
Plain Training
ViT* [13]1979-2020400.721.351.860.661.983.010.682.023.1140.2208.5393.963.3222.1397.0
IFS [32]1979-2020401.091.381.740.961.872.780.991.932.87------
Pangu-Weather [3]1979-2020400.821.091.530.771.632.540.791.682.65------
FourCastNet [22]1979-2020400.821.021.770.822.083.340.842.113.41------
ClimaX [27]1979-2020401.111.471.831.042.022.79---------
Graphcast [23]1979-2020400.510.941.370.381.512.37---23.4135.2278.2---
Fengwu [9]1979-2020400.581.031.410.421.532.32---23.2137.1276.9---
FuXi [10]1979-2020400.550.991.410.421.502.360.431.542.4427.2136.7282.9---
VA-MoE1979-2020400.571.031.420.431.412.250.441.462.3427.5131.1275.957.1168.9302.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (79-20)
VA-MoE (IL)1979-2020200.581.051.450.481.472.330.471.542.4127.9137.3281.659.3173.4312.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)2000-2020100.731.171.570.541.582.490.551.632.5730.0148.8304.760.6171.4314.8
", + "bbox": [ + 96, + 88, + 906, + 345 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c48125c0977a9f970ee7a1f1510428c67d7c7c44508d3eb463f98acbe402368.jpg", + "table_caption": [ + "Table 2. Prediction performances on Incremental Training with 5 surface variables, i.e., T2M, U10, V10, MSL, and SP. * denotes running by ourselves. The best results are marked in bold. All experiments are in ${0.25}^{ \\circ }$ with ${721} \\times {1440}$ resolutions." + ], + "table_footnote": [], + "table_body": "
Para. (M)Z500(m2/s2) ↓Q500(×e-3, g/kg) ↓U500(m/s) ↓V500(m/s) ↓T500(K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
IFS [32]-28.31154.08333.960.310.610.751.433.235.121.403.585.640.360.981.70
Pangu-Weather [3]-24.88167.90391.260.250.550.690.963.134.730.913.525.150.270.941.56
Graphcast [23]-15.23125.42275.35---0.772.864.490.742.924.670.230.871.48
VA-MoE66519.28134.63295.520.170.490.620.842.994.710.843.044.890.250.761.36
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)13718.23133.14292.630.170.490.610.843.014.740.843.514.930.250.761.37
", + "bbox": [ + 96, + 397, + 906, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Prediction performances on Initial Training with 5 upper-air variables. All experiments are in ${0.25}^{ \\circ }$ with ${721} \\times {1440}$ resolutions.", + "bbox": [ + 91, + 542, + 903, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pangu-Weather [3] and Stormer [28], and significantly surpassing ClimaX [27] and FourCastNet [22], while slightly inferior to GraphCast [23]. For T850, our structure is slightly inferior to EWMoE [14], GraphCast [23], and Stormer [28] in the short-term predictions within 48 hours. After that period, our structure achieves the best performance in the long-term predictions.", + "bbox": [ + 88, + 583, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results of Surface Variables Prediction. This study evaluates model performance on five surface variables: T2M, U10, V10, MSL, and SP. While competing structures are trained with both upper-air and surface variables, VA-MoE (IL) employs a two-stage training approach: initial training with upper-air variables followed by incremental training with all five surface variables.", + "bbox": [ + 88, + 688, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of the results from Tab. 2 demonstrates that VA-MoE outperforms competing models across multiple variables and settings, particularly in U10 and V10. Although VA-MoE shows marginally weaker performance in T2M compared to GraphCast [23] and comparable results to FengWu [9] and FuXi [10], it achieves superior long-term predictions for U10, V10, MSL, and SP. These indicate", + "bbox": [ + 88, + 795, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that while VA-MoE exhibits slight limitations in short-term forecasting, it excels in long-term prediction tasks.", + "bbox": [ + 511, + 583, + 903, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When implementing the incremental training strategy, VA-MoE (IL) trained on a 40-year dataset with only half the training iterations achieves performance comparable to standard VA-MoE. Even with a reduced 20-year dataset, VA-MoE(IL) maintains acceptable performance using just a quarter of the iterations required by the standard approach. These results underscore the efficacy of incremental learning for atmospheric datasets and suggest its potential to replace conventional training methods in operational settings.", + "bbox": [ + 509, + 614, + 905, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results of Upper-Air Variables Prediction. This study evaluates model performance on upper-air variables, comparing three training approaches: (1) VA-MoE, trained exclusively on upper-air variables; (2) VA-MoE (IL), which incorporates incremental training on surface variables after initial upper-air training; and (3) competing models trained simultaneously on both upper-air and surface variables.", + "bbox": [ + 509, + 750, + 905, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 3, VA-MoE achieves performance comparable to GraphCast [23] and outperforms IFS [32] and Pangu-Weather [3] across both short- and long-term", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7921", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a89841f541b62779568827b555ddd995ada9597a86b689891af324f299c2b25e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Para. (M)Z500 (m2/s2) ↓Q500 (×e-3, g/kg) ↓U500 (m/s) ↓V500 (m/s) ↓T500 (K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
ViT*[13]30733.38209.4517.810.220.611.061.243.666.521.223.767.410.421.182.40
ViT+MoE (light)*[30]60937.92207.11405.730.220.600.781.303.845.871.273.896.110.461.232.02
ViT+MoE* [30]111328.31169.61356.020.230.560.721.213.465.441.233.545.690.351.071.83
VA-MoE66520.59139.02302.130.180.490.620.913.024.760.913.084.970.270.921.59
VA-MoE (IL)13720.29138.52301.410.180.500.630.913.044.790.913.105.030.270.931.60
", + "bbox": [ + 98, + 89, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Architectural impact on 5 upper-air variables under ${500}\\mathrm{{hPa}}$ . All experiments are in ${1.5}^{ \\circ }$ with ${128} \\times {256}$ resolutions.", + "bbox": [ + 127, + 205, + 866, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a3f0e25dc8da0ded62ed82c5f1ef56cb67792b1eb0e64e3699a3a53e9edaebc.jpg", + "image_caption": [ + "Figure 5. 6-hour global weather prediction of upper-air and surface variables forecast visualization generated by VA-MoE framework." + ], + "image_footnote": [], + "bbox": [ + 114, + 223, + 875, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "prediction. Notably, VA-MoE (IL) demonstrates robust predictive capabilities for upper-air variables even after incremental training on surface variables, despite using only a 20-year dataset - half the size of the initial training data. Furthermore, VA-MoE (IL) exhibits marginal performance improvements over the baseline in long-term predictions of Z500, confirming the absence of catastrophic forgetting during incremental training. These results highlight the efficacy of incremental training in VA-MoE(IL) for maintaining and enhancing predictive accuracy on original variables.", + "bbox": [ + 89, + 489, + 482, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 650, + 243, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impacts of Architecture. As shown in Tab. 4, we compare VA-MoE with other architectures, including ViT and ViT+MoE. VA-MoE significantly outperforms both ViT and ViT+MoE architectures across 6-hour, 72-hour, and 120-hour prediction horizons, despite ViT+MoE containing nearly twice the number of parameters. This performance gap highlights VA-MoE's suitability for weather forecasting tasks as its channel-adaptive expert design ensures parameter efficiency while maintaining accuracy, particularly in incremental learning. Considering the computational complexity, all ablation studies are in $1.5^{\\circ}$ with $128 \\times 256$ .", + "bbox": [ + 89, + 672, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Visualization", + "text_level": 1, + "bbox": [ + 89, + 848, + 227, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the work, we visualize some predicted results and parameters. In Fig. 5, we visualize the 6-hour predicted results of", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Z500, T850, U10, and T2M. There are max $0.08\\%$ , $0.22\\%$ , $1.59\\%$ , and $0.47\\%$ absolute errors across 4 variables, including 2 upper-air variables and 2 surface variables.", + "bbox": [ + 511, + 489, + 905, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 549, + 633, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed incremental weather forecasting, a novel task that addresses the challenge of dynamically expanding weather models to incorporate new variables without retraining from scratch. Our solution, Variables-Adaptive Mixture-of-Experts, enables parameter-efficient adaptation by selectively fine-tuning experts assigned to new variables through index embedding guidance, while preserving pretrained knowledge in existing parameters. Experimental results demonstrate that our approach achieves forecast accuracy comparable to fully retrained models while drastically reducing computational costs.", + "bbox": [ + 511, + 574, + 906, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 753, + 700, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the Hong Kong RGC General Research Fund (152169/22E, 152228/23E, 162161/24E), Research Impact Fund (No. R5060-19, No. R5011-23), Collaborative Research Fund (No. C1042-23GF), NSFC/RGC Collaborative Research Scheme (No. 62461160332 & CRS_HKUST602/24), Areas of Excellence Scheme (AoE/E-601/22-R), the InnoHK (HKGAI), and Key R&D Program of Zhejiang Province 2025C01075.", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7922", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3", + "[2] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567): 47-55, 2015. 1, 2", + "[3] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 1, 2, 6, 7", + "[4] Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan Weyn, Haiyu Dong, Anna Vaughan, et al. Aurora: A foundation model of the atmosphere. arXiv:2405.13063, 2024. 1", + "[5] Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: learning stable dynamics on the sphere. In Proceedings of International Conference on Machine Learning, 2023. 2", + "[6] Hao Chen, Yonghan Dong, Zhe-Ming Lu, Yunlong Yu, and Jungong Han. Self-prompting perceptual edge learning for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4528-4541, 2023. 2", + "[7] Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel matching network for cross-domain few-shot segmentation. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 978-987, 2024. 6", + "[8] Hao Chen, Yunlong Yu, Yonghan Dong, Zheming Lu, Yingming Li, and Zhongfei Zhang. Multi-content interaction network for few-shot segmentation. ACM Transactions on Multimedia Computing, Communications and Applications, 20 (6):1-20, 2024. 6", + "[9] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv:2304.02948, 2023. 1, 2, 7", + "[10] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 2, 7", + "[11] Wuyang Chen, Yanqi Zhou, Nan Du, Yanping Huang, James Laudon, Zhifeng Chen, and Claire Cui. Lifelong language pretraining with distribution-specialized experts. In International Conference on Machine Learning, pages 5383-5395. PMLR, 2023. 3", + "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems, pages 16344-16359, 2022. 6", + "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Trans-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formers for image recognition at scale. arXiv:2010.11929, 2020.2,7,8", + "[14] Lihao Gan, Xin Man, Chenghong Zhang, and Jie Shao. Ew-moe: An effective model for global weather forecasting with mixture-of-experts. arXiv:2405.06004, 2024. 6, 7", + "[15] Yuan Gao, Hao Wu, Ruiqi Shu, Huanshuo Dong, Fan Xu, Rui Chen, Yibo Yan, Qingsong Wen, Xuming Hu, Kun Wang, et al. Oneforecast: A universal framework for global and regional weather forecasting. arXiv:2502.00338, 2025. 2", + "[16] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 3", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2", + "[18] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquin Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999-2049, 2020. 6", + "[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3", + "[20] Ryan Keisler. Forecasting global weather with graph neural networks. arXiv:2202.07575, 2022. 6", + "[21] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. 3", + "[22] Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. 1, 2, 6, 7", + "[23] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2, 6, 7", + "[24] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv:2010.08895, 2020. 2", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7923", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Franco Molteni, Roberto Buizza, Tim N Palmer, and Thomas Petroliagis. The ecmwf ensemble prediction system: Methodology and validation. Quarterly journal of the royal meteorological society, 122(529):73-119, 1996. 2", + "[27] Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. Climax: A foundation model for weather and climate. In International Conference on Machine Learning, 2023. 6, 7", + "[28] Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. In Advances in Neural Information Processing Systems, pages 68740-68771, 2025. 6, 7", + "[29] Youqi Pan, Wugen Zhou, Yingdian Cao, and Hongbin Zha. Adaptive vio: Deep visual-inertial odometry with online continual learning. arXiv:2405.16754, 2024. 3", + "[30] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. In European Conference on Computer Vision, 2024. 8", + "[31] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Probabilistic weather forecasting with machine learning. Nature, 637(8044):84-90, 2025. 2", + "[32] D Richardson. The thorpex interactive grand global ensemble (tigge). In Geophysical Research Abstracts, page 02815, 2024. 7", + "[33] Harold Ritchie, Clive Temperton, Adrian Simmons, Mariano Hortal, Terry Davies, David Dent, and Mats Hamrud. Implementation of the semi-lagrangian method in a high-resolution version of the ecmwf forecast model. Monthly Weather Review, 123(2):489-514, 1995. 2", + "[34] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23933-23942, 2024. 3", + "[35] Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li. Semantically-shifted incremental adapter-tuning is a continual vitransformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23252-23262, 2024. 3", + "[36] Yogesh Verma, Markus Heinonen, and Vikas Garg. ClimODE: Climate forecasting with physics-informed neural ODEs. In International Conference on Learning Representations, 2024. 6", + "[37] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14205-14215, 2024. 3", + "[38] Wei Xiong, Muyuan Ma, Xiaomeng Huang, Ziyang Zhang, Pei Sun, and Yang Tian. Koopmanlab: machine learning for" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "solving complex physics equations. APL Machine Learning, 1(3), 2023. 2", + "[39] Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, and Lei Bai. Generalizing weather forecast to fine-grained temporal scales via physics-ai hybrid modeling. In Advances in Neural Information Processing Systems, 2024. 6", + "[40] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23670-23680, 2024. 3", + "[41] Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. Multi-task dense prediction via mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27927-27937, 2024. 3", + "[42] Fei Ye and Adrian G Bors. Online task-free continual generative and discriminative learning via dynamic cluster memory. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 26202-26212, 2024. 3", + "[43] Donggeun Yoon, Minseok Seo, Doyi Kim, Yeji Choi, and Donghyeon Cho. Probabilistic weather forecasting with deterministic guidance-based diffusion model. In European Conference on Computer Vision, pages 108-124, 2024. 2", + "[44] Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 27758-27767, 2024. 2", + "[45] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 3", + "[46] Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. Skilful nowcasting of extreme precipitation with nowcastnet. Nature, 619(7970):526-532, 2023. 2", + "[47] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, Cheng-Lin Liu, and Zhaoxiang Zhang. Rcl: Reliable continual learning for unified failure detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12140-12150, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 752 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_model.json b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c57688f97f41f61b679688356175e9801e0688b --- /dev/null +++ b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.766, + 0.153 + ], + "angle": 0, + "content": "VA-MoE: Variables-Adaptive Mixture of Experts for" + }, + { + "type": "title", + "bbox": [ + 0.31, + 0.163, + 0.653, + 0.185 + ], + "angle": 0, + "content": "Incremental Weather Forecasting" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.152, + 0.688, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.211, + 0.895, + 0.266 + ], + "angle": 0, + "content": "Hao Chen\\(^{1\\dagger}\\) Han Tao\\(^{1\\dagger}\\) Guo Song\\(^{1*}\\) Jie Zhang\\(^{1}\\) Yonghan Dong\\(^{2}\\) Yunlong Yu\\(^{3}\\) Lei Bai\\(^{4}\\) \n\\(^{1}\\)Hong Kong University of Science and Technology (HKUST) \\(^{2}\\)Huawei Technologies Ltd. \n\\(^{3}\\)Zhejiang University \\(^{4}\\)Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.269, + 0.843, + 0.284 + ], + "angle": 0, + "content": "Equal contribution *Corresponding author {hchener, thanad}@connect.ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.286, + 0.919, + 0.302 + ], + "angle": 0, + "content": "{songguo, csejzhang}@ust.hk dongyonghan@huawei.com yuyunlong@zju.edu.cn bailei@pjlab.org.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.336, + 0.327, + 0.352 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.485, + 0.705 + ], + "angle": 0, + "content": "This paper presents Variables-Adaptive Mixture of Experts (VA-MoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real-time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VA-MoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct sub-patterns of atmospheric variables (e.g., temperature, humidity, wind speed). Moreover, the proposed method employs a variable-adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on ERA5 dataset demonstrate that VA-MoE performs comparable against state-of-the-art models in both short-term (e.g., 1-3 days) and long-term (e.g., 5 days) forecasting tasks, with only about \\(25\\%\\) of trainable parameters and \\(50\\%\\) of the initial training data. Code: https://github.com/chenhao-zju/VAMoE" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Weather forecasting remains one of the most critical scientific challenges with profound socio-economic impacts. While traditional Numerical Weather Prediction (NWP) [2] relies on solving complex partial different equations to simulate atmospheric dynamics, recent advances in data-driven artificial intelligence [3, 4, 9, 22] have revolutionized earth system modeling. Deep learning models, in particular, have demonstrated remarkable potential in capturing diverse spatiotemporal patterns across weather and climate phenomena." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.334, + 0.904, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.56 + ], + "angle": 0, + "content": "Figure 1. Illustration of two different training paradigms of weather forecasting when new variables arrive: (c) Full-Train the whole model (d) Incremental-Train only the module corresponding new variables. (b) Performances of these two paradigms." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.568, + 0.881, + 0.584 + ], + "angle": 0, + "content": "ena, giving rise to the emerging field of AI for Weather." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.584, + 0.907, + 0.81 + ], + "angle": 0, + "content": "However, a fundamental limitation of existing AI4Weather approaches lies in their assumption that all variables are available synchronously during training and inference. In reality, meteorological variables are often heterogeneous in terms of data sources, collection frequencies, and spatial-temporal distributions. For instance, the upper-air variables (e.g., temperature profiles) are sparse and sampled via radiosondes/satellites, while the surface variable (e.g., precipitation, wind) are dense but updated in near-real-time. This asynchrony poses significant challenges: when introducing new variables (e.g., satellite-derived aerosol data), existing models must be entirely retrained from scratch, incurring prohibitive computational costs. For example, Pangu [3] required 64 days and 192 V100 GPUs for a full retraining cycle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To address this challenge, we propose Incremental Weather Forecasting (IWF), a novel paradigm designed to dynamically expand meteorological forecast variables in response to increasing observational data. By enabling parameter-efficient variable scaling, IWF revolutionizes operational meteorology by offering a streamlined alternative" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7915" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "to conventional models. This innovation tackles a critical limitation in the field: the need for scalable modeling solutions as datasets grow gradually, ensuring adaptability without compromising computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.482, + 0.274 + ], + "angle": 0, + "content": "A pivotal limitation of existing incremental approaches is catastrophic forgetting, i.e., when introducing new variables, pre-trained parameters drift towards new distributions, causing significant performance degradation on original variables. To overcome this, we present the Variables-Adaptive Mixture of Experts (VA-MoE) model, the first expert-based system explicitly designed for hierarchical atmospheric variables." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.685 + ], + "angle": 0, + "content": "Mixture of Experts (MoE) models inherently enable scalable representation learning for multi-variable atmospheric datasets by dynamically expanding expert modules as input dimensions grow. While traditional MoE architectures excel in representational capacity, their parallelized design which distributes uniform weights across experts, often leads to homogeneous outputs, limiting specialized knowledge acquisition. To address this limitation, we introduce VA-MoE, an auxiliary-loss-free framework that decouples expert specialization from computational efficiency. VA-MoE employs a phased training strategy: Initial experts are pretrained on base variables, then incrementally expanded to accommodate new variables. Crucially, the pretrained experts are frozen during expansion phases to prevent catastrophic forgetting, ensuring robust retention of prior expertise while adapting to novel inputs. To foster expert diversity without auxiliary losses, VA-MoE incorporates variable index embeddings, i.e., positional metadata encodings that guide experts to develop domain-specific specialization. These embeddings dynamically activate contextually relevant computational pathways for each variable, optimizing resource allocation at inference time. By integrating variable index embeddings that encode spatial-temporal metadata, VA-MoE directs experts to specialize in specific atmospheric domains. This context-aware routing reduces inference latency compared to static MoE models while maintaining expert diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.482, + 0.838 + ], + "angle": 0, + "content": "Besides, we propose a gradient-scaled variable-adaptive loss function that dynamically aligns variable optimization rates with their inherent spatiotemporal characteristics. By quantifying the distinct temporal evolution patterns of atmospheric variables, our method allocates differential loss weights, i.e., larger gradient magnitudes prioritize fast-transient fields like temperature to capture immediate atmospheric dynamics, while gradual weight adjustments stabilize slow-changing fields such as geopotential height to preserve long-term system behavior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.482, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 1, our paradigm employs a two-phase training paradigm: an initial phase optimizing base atmospheric variables, followed by an incremental phase integrating new variables while freezing previously trained ex" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "perts. Experiments with the same 40-year dataset reveal that our paradigm Fig. 1(a)+(d) achieves superior z500 predictions performance compared to full retraining Fig. 1(a)+(c)." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.138, + 0.884, + 0.152 + ], + "angle": 0, + "content": "In conclusion, the contributions of this work include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- This work initiates systematic research on incremental learning paradigms for weather forecasting. We propose the quantitative benchmark to evaluate the trade-offs between model scalability and generalization in incremental weather modeling." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.319 + ], + "angle": 0, + "content": "- We present Variables-Adaptive Mixture of Experts (VA-MoE), the first framework tailored for incremental atmospheric modeling. VA-MoE achieves expert specialization through contextual variable activation driven by variable index embeddings, enabling dynamic assignment of experts to variables during both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- Extensive experiments on the ERA5 dataset demonstrate that VA-MoE achieves comparable performance for surface variables, while delivering superior accuracy in upper-air variables against the existing competitors under \\(50\\%\\) reduced dataset size and \\(25\\%\\) fewer parameters." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.41, + 0.651, + 0.426 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.436, + 0.81, + 0.452 + ], + "angle": 0, + "content": "2.1. Data-Driven Weather Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.905, + 0.534 + ], + "angle": 0, + "content": "In recent decades, traditional NWP [2] methods have dominated the weather forecasting field due to the robust predictions and rigorous mathematical validation. However, NWP methods [26, 33] require training from scratch for new prediction, resulting in slow computation and high costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.685 + ], + "angle": 0, + "content": "Recent advances in deep learning have catalyzed a paradigm shift toward data-driven models for medium-range weather forecasting. Pioneering works like FourCastNet [22] leverage Fourier Neural Operators (FNOs) [24] to learn spatiotemporal weather patterns, while PanguWeather [3] employs a 3D vision transformer [13, 25] to model atmospheric dynamics. Subsequent innovations, including Neural Operators [5, 24, 38] and specialized architectures [9, 10, 15, 23, 31, 43, 44, 46], have achieved accuracy rivaling NWP at a fraction of computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.838 + ], + "angle": 0, + "content": "While existing data-driven models excel at fixed-variable forecasting after costly initial training, their inflexibility poses a critical barrier: retraining from scratch is required to incorporate new variables, rendering cross-domain adaptation impractical. In contrast, our work introduces incremental weather forecasting, a novel paradigm evolving spatiotemporal patterns in real-time data. By decoupling variable-specific expertise from shared dynamics, our approach supports dynamic expansion to new atmospheric parameters with minimal retraining overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.898, + 0.865 + ], + "angle": 0, + "content": "2.2. Mixture of Experts for Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Advances in deep learning[6, 17] have spurred interest in parameter-efficient transfer strategies, where models adapt" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7916" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.093, + 0.46, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.094, + 0.882, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.182, + 0.882, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.418, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Figure 2. Illustration of the incremental weather forecasting paradigm. (a), (b) are the initial and incremental stages, respectively. (c) presents the detailed structure in both initial and incremental stages. During the initial stage, the model is trained with upper-air variables(UV). In the incremental stage, the fire components are trained while the snow components remain frozen. The index embedding is trained by the Index Encoder and divided into six parts based on the upper-air variables (UV: Z, Q, U, V, T) and surface variables (SV)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.482, + 0.644 + ], + "angle": 0, + "content": "to new domains by fine-tuning only a subset of parameters. A key paradigm, incremental learning, addresses the stability-plasticity dilemma: retaining knowledge of prior tasks while integrating new concepts. However, real-world data often involves domain shifts, such as variations in variable distributions or task objectives, leading to two dominant research threads. Class Incremental Learning (CIL) [16, 21, 35, 47], focuses on expanding classification tasks with new classes while Task Incremental Learning (TIL) [29, 34, 40, 42] adapts models to distinct task sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.886 + ], + "angle": 0, + "content": "One of the most popular structures for addressing incremental learning tasks is the Mixture of Experts (MoE) model [19], which learns representations of new concepts through different experts. Due to its sparse architecture, many works have adopted the MoE structure to reduce inference costs and enhance model capacity [37, 41]. An early work introducing MoE to incremental learning is Expert Gate [1], which trains multiple backbones as different experts and assigns new domains to the relevant expert. Lifelong-MoE [11] leverages pretrained experts and gates to retain the representational knowledge of the training domain. In addition to these approaches, MoE is used as an adapter in MoE-Adapter [45], where adapters are attached to the main structure. These works have demonstrated the promising performance of MoE in visual and natural language incremental learning." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In contrast to MoE designed for fixed-variable systems," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.6 + ], + "angle": 0, + "content": "our propose incremental learning of variable distributions by assigning distinct pressure-level groups to specialized experts. Unlike prior incremental learning literature that primarily focus on task or dataset-specific adaptation, our incremental paradigm targets variable-centric expansion, enabling models to adapt to evolving observational data without retraining from scratch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.614, + 0.649, + 0.632 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Our work focuses on a novel paradigm of incremental weather forecasting with VA-MoE. In this section, we first formally define the problem in Sec. 3.1. Then, we present the structure of VA-MoE, consisting with the Channel-Adaptive Expert (CAE) and the Shared Expert in Sec. 3.2. The model during incremental stage is introduced in Sec. 3.3. Finally, loss function is presented in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.625, + 0.772 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.906, + 0.825 + ], + "angle": 0, + "content": "For the weather forecasting task, the AI-based model \\(\\Phi\\) aims to predict the future weather variables \\(\\mathbf{X}^{t + 1}\\) with the historical weather variables \\(\\mathbf{X}^t\\), i.e., \\(\\mathbf{X}^{t + 1} = \\Phi (\\mathbf{X}^t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the incremental weather forecasting task, the weather variables \\( \\mathbf{X}^t \\) are divided into two sets: (i) the initial training variables at \\( t \\) time \\( \\mathbf{X}_h^t \\in \\mathbb{R}^{H \\times W \\times N} \\), and, (ii) the incremental variables at \\( t \\) time \\( \\mathbf{X}_{sv}^t \\in \\mathbb{R}^{H \\times W \\times M} \\). Unlike traditional weather forecasting tasks, where models are trained" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7917" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.465 + ], + "angle": 0, + "content": "Figure 3. Illustration of the VA-MoE. In the left subgraph, the input features comprise both upper-air and surface variables. During the initial training stage, five distinct CAE modules process the upper-air variables. In the incremental stage, surface variables are handled by a dedicated module, \\(\\mathrm{CAESV}\\), while the original five CAEs remain frozen to preserve learned representations. The right subgraph details the CAE module. The Index Embedding is selected through a multi-layer process and applied to weight the Input Embedding via multiplication. Initially discrete, the index embedding is transformed into a continuous representation after passing through the GateEmbed layer, enabling it to dynamically reflect the relevance of specific channels to the input variables." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.563 + ], + "angle": 0, + "content": "with all variables available simultaneously, the incremental weather forecasting task sequentially incorporates dynamic variables as they become available over time. The incremental weather forecasting process is structured into a sequence of learning stages: an initial training stage followed by subsequent training stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.82 + ], + "angle": 0, + "content": "In the initial training stage, the available variables of 40 years are used to train a weather model, i.e., \\(\\mathrm{X}_h^{t + 1} = \\Phi_h(\\mathrm{X}_h^t)\\), as shown in Fig. 2(a). In the incremental training stage, the half variables used at the initial stage and the surface variables at the incremental stage are incorporated together, to train the newly added modules for the surface variables while keeping the model parameters from the initial stage frozen, i.e., \\(\\mathrm{X}^{t + 1} = \\Phi (\\mathrm{X}^t)\\), where \\(\\Phi\\) consists of both the fixed parameters of the pre-trained model \\(\\Phi_h\\) and the trainable parameters of \\(\\Phi_{sv}\\), \\(\\mathrm{X}^t\\) consists of all the available variables at time \\(t\\), \\(\\mathrm{X}^{t + 1}\\) consists of the predicted variables at time \\(t + 1\\), as shown in Fig. 2(b). In this work, the upper-air variables are utilized during the initial training stage while the surface variables become available and are incorporated during the incremental training stage. The upper-air variables include five different types, Z, Q, U, V, and T, each defined across 13 different levels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.196, + 0.847 + ], + "angle": 0, + "content": "3.2. VA-MoE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We introduce index embedding within transformer blocks, a novel mechanism to dynamically guide experts in learning hierarchical relationships between meteorological vari" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "ables. As new variables are incrementally integrated, corresponding experts are added to the transformer architecture and optimized via index-based affinity assignments. Building on this framework, we adapt the Mixture-of-Experts (MoE) paradigm which enables task-specific specialization and flexible integration of new experts—into weather forecasting transformers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.318 + ], + "angle": 0, + "content": "The structure of the proposed VA-MoE is illustrated in Fig. 2(c). The input features, extracted by the encoder, first pass through a normalization layer and a self-attention layer. The output of the self-attention layer is then combined with the residual connection. Then, another normalization layer is applied, followed by a VA-MoE module, which is also equipped with a residual connection. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.329, + 0.905, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {X} _ {m i d} = \\mathbf {X} _ {i n} + \\operatorname {S A} (\\text {L a y e r N o r m} (\\mathbf {X} _ {\\text {i n}})), \\\\ \\mathbf {X} _ {i n} = \\mathbf {X} _ {i n - 1} + \\operatorname {V A M E} (\\mathbf {I} _ {i n - 1}, \\mathbf {N} _ {i n - 1} (\\mathbf {X} _ {i n})) \\end{array} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.351, + 0.855, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\text {o u t}} = \\mathbf {X} _ {\\text {m i d}} + \\operatorname {V A - M o E} (\\operatorname {L a y e r N o r m} (\\mathbf {X} _ {\\text {m i d}})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.407 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_{in}\\) and \\(\\mathbf{X}_{out}\\) are the input and the output of each Transformer Block, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.417, + 0.785, + 0.432 + ], + "angle": 0, + "content": "3.2.1. Variables and Index Embedding" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.436, + 0.907, + 0.526 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, during the training stage, the input feature \\(\\mathbf{X}_h^t\\in \\mathbb{R}^{H\\times W\\times N}\\) consists of five different types of variables: \\(\\mathbf{X}_Z^t,\\mathbf{X}_Q^t,\\mathbf{X}_U^t,\\mathbf{X}_V^t,\\mathbf{X}_T^t\\in \\mathbb{R}^{H\\times W\\times (N / 5)}\\), where, \\(H\\times W\\) denotes the spatial resolution. These five types of variables are concatenated and processed by a shared expert and five distinct Channel-Adaptive Experts (CAEs)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.664 + ], + "angle": 0, + "content": "In addition to these variables, the framework introduces an extra one-hot index embedding \\(\\mathbf{I}_h\\in \\mathbb{R}^{5\\times N}\\) to guide experts in learning variable affinity. The number \\(N\\) corresponds to the number of variables, and 5 corresponds to the number of variable types. Similar to the input weather features, the index embedding is composed of five separate embeddings: \\(\\mathbf{I}_Z,\\mathbf{I}_Q,\\mathbf{I}_U,\\mathbf{I}_V,\\mathbf{I}_T\\in \\mathbb{R}^{1\\times 1\\times N}\\). This initial index embedding is then encoded into a latent space using a Linear layer, expressed as \\(\\mathbf{I}_h = \\mathrm{Linear}(\\mathbf{I}_h)\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.672, + 0.791, + 0.687 + ], + "angle": 0, + "content": "3.2.2. Channel-Adaptive Expert (CAE)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.766 + ], + "angle": 0, + "content": "The key design of VA-MoE is the channel-adaptive expert, as shown in Fig. 3. For each variable \\( Z \\), the corresponding expert module processes both the feature embeddings of the input variables \\( \\mathbf{X}_h^t \\) and the index embedding \\( \\mathbf{I}_Z \\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.778, + 0.905, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.839 + ], + "angle": 0, + "content": "where \\(\\mathrm{CAE}_Z\\) is the abbreviation of channel-adaptive expert module for variable \\(Z\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In the CAE module for variable \\( Z \\), the index embedding \\( \\mathbf{I}_Z \\) is encoded through a linear layer for feature projection and then multiplied channel-wise with the input embedding \\( \\mathbf{X}_h^t \\). The resulting fused feature is processed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7918" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.198 + ], + "angle": 0, + "content": "by the GateEmbed layer, which produces a gate embedding to enhance the fusion of the index and variable embeddings. This enhanced feature is then normalized using a SoftMax layer and filtered through a TopK operation, which selects the top-K high-rank channels. This process yields the GateIndex \\(\\mathrm{GI}_Z\\in \\mathbb{R}^{H\\times W\\times K}\\) and GateWeight \\(\\mathrm{GW}_Z\\in \\mathbb{R}^{H\\times W\\times K}\\). The process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.209, + 0.484, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {Z} ^ {t o p k}, \\mathbf {W} _ {Z} ^ {t o p k} = \\mathrm {T O P} _ {k} (\\operatorname {S o f t M a x} (\\operatorname {M L P} _ {Z} (\\mathbf {X} _ {h} ^ {t} \\odot \\mathbf {I} _ {Z}))), (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.484, + 0.408 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_Z^{topk}\\) and \\(\\mathbf{W}_Z^{topk}\\) denote the index and weight of the Top-k features. Once the GateIndex and GateWeight are computed, the input embedding is selectively processed based on the GateIndex, while the variable embedding is weighted channel-wise according to the GateWeight. The GateEmbed layer uses variable embeddings to guide the index embedding, enabling the creation of a variable-specific confidence matrix. Within the CAE module, this embedding enhances the confidence matrix by incorporating both semantic and positional information, rather than relying solely on positional information." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.41, + 0.472, + 0.424 + ], + "angle": 0, + "content": "The selected features for variable \\( Z \\) are obtained with:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.436, + 0.483, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {Z} ^ {t, s e l e c t e d} = \\mathbf {W} _ {Z} ^ {t o p k} * \\operatorname {S e l e c t} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z} ^ {t o p k}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.483, + 0.498 + ], + "angle": 0, + "content": "where \\(\\mathrm{Select}_Z\\) denotes selecting Top-K embedding from the input feature \\(\\mathbf{X}_h^t\\) within the \\(\\mathrm{CAE}_Z\\) module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.604 + ], + "angle": 0, + "content": "The selected and weighted embedding captures the key features associated with the specific type of variables. In the \\(\\mathrm{CAE}_Z\\) module, these features serve as the training set to model the distribution of the variable \\(Z\\) with \\(\\mathrm{Expert}_Z\\). The expert module adopts a parallel structure, splitting the channels into multiple parts and enhancing each part with a sparse MoE. This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.615, + 0.483, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {E x p e r t} _ {Z} (\\mathbf {X} _ {Z} ^ {t, s e l e c t e d}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.648, + 0.454, + 0.664 + ], + "angle": 0, + "content": "where \\(\\mathbf{Expert}_Z\\) is implemented with a multiple layer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.245, + 0.688 + ], + "angle": 0, + "content": "3.2.3. Shared Expert." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.844 + ], + "angle": 0, + "content": "The shared expert processes the overall features of all variables and operates in parallel with the variable-specific modules. Each variable is handled by its corresponding CAE module, and their outputs are summarized to produce the fused features \\(\\mathbf{X}_h^{t,fused}\\) for all variables. Since the channel count of the selected variable embeddings is lower than that of the original embeddings, an additional up-channel linear layer, \\(\\mathrm{Linear}_{\\mathrm{up}}\\), is introduced. The fused features and the output from the shared expert are then combined via pixel-wise addition. This process is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.854, + 0.483, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {h} ^ {t, f u s e d} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right) + \\dots + \\operatorname {C A E} _ {T} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {T}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.883, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {X} ^ {\\prime}\\right) _ {h} ^ {t} = \\operatorname {E x p e r t} _ {s h a r e d} \\left(\\mathbf {X} _ {h} ^ {t}\\right) + \\operatorname {L i n e a r} _ {u p} \\left(\\mathbf {X} _ {h} ^ {t, f u s e d}\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.768, + 0.108 + ], + "angle": 0, + "content": "3.3. Incremental Learning Stage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.204 + ], + "angle": 0, + "content": "When new variables are incrementally introduced to the model, specialized experts are dynamically integrated into the transformer blocks to process these variables. Concurrently, both the encoder and decoder architectures undergo synchronized updates to maintain compatibility with the expanded variable set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.212, + 0.78, + 0.226 + ], + "angle": 0, + "content": "3.3.1. Encoder and Decoder Modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.907, + 0.367 + ], + "angle": 0, + "content": "When \\( K \\) new variables are introduced incrementally, the encoder's input dimension expands to \\( H \\times W \\times (N + M) \\), necessitating an adjustment to the convolutional layer's kernel from \\( (3, 3, N, C) \\) to \\( (3, 3, N + M, C) \\). To accommodate this, the layer's parameters are partitioned into two components: The \\( (3, 3, N, C) \\) segment retains pretrained weights from earlier stages, ensuring continuity. The \\( (3, 3, M, C) \\) segment, corresponding to the new variables, is initialized randomly to learn emergent patterns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.906, + 0.397 + ], + "angle": 0, + "content": "The same strategy is applied to the position embedding and the Decoder module." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.406, + 0.682, + 0.42 + ], + "angle": 0, + "content": "3.3.2. VA-MoE Module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.516 + ], + "angle": 0, + "content": "During the incremental learning phase, \\(M\\) new experts are dynamically integrated into the MoE layer of each transformer block to accommodate \\(M\\) newly introduced variables. In this phase, only the new experts and the shared expert are actively trained, while all preexisting experts remain frozen to preserve previously learned knowledge." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.524, + 0.746, + 0.539 + ], + "angle": 0, + "content": "3.3.3. Index Embedding Module." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.543, + 0.907, + 0.618 + ], + "angle": 0, + "content": "When \\(M\\) new variables become available during the incremental stage, the index embedding changes from \\(\\mathbf{I}_h\\in\\) \\(\\mathbb{R}^{(N\\times l)\\times N}\\) to \\(\\mathbf{I}_h\\in \\mathbb{R}^{(N\\times l + M\\times r)\\times (N + M)}\\), where each \\(M\\) new variables has \\(r\\) levels. Additionally, the index encoder needs to be retrained." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.629, + 0.697, + 0.644 + ], + "angle": 0, + "content": "3.4. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.651, + 0.906, + 0.681 + ], + "angle": 0, + "content": "To train the model, we introduce a novel dynamic prediction loss combined with a reconstruction loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.69, + 0.738, + 0.705 + ], + "angle": 0, + "content": "3.4.1. Dynamic Prediction Loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.906, + 0.83 + ], + "angle": 0, + "content": "The prediction loss aims to minimize the difference between the model's predicted results and the actual future weather variables. Unlike existing weather forecasting methods that treat all the variables equally, we argue that different variables, such as temperature and surface pressure, follow distinct distributions. Thus, we introduce a novel dynamic prediction loss that dynamically assigns a weight to each channel of the input data. This loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.84, + 0.907, + 0.861 + ], + "angle": 0, + "content": "\\[\nO b j _ {p r e d} = \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) / e ^ {\\mathbf {w}} + \\mathbf {w}, (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes Hadamard product, \\(e\\) denotes the base of the natural logarithm, \\(\\mathbf{X}^{t + 1}\\in \\mathbb{R}^{H\\times W\\times C}\\), and \\(\\mathbf{w}\\in\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7919" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.457, + 0.314 + ], + "angle": 0, + "content": "
NameDescriptionLevels
Upper-Air
ZGeopotential13
QSpecific humidity13
Ux-direction wind13
Vy-direction wind13
TTemperature13
Surface-Incremental Learning
u10x-direction wind at 10m heightSingle
v10y-direction wind at 10m heightSingle
t2mTemperature at 2m heightSingle
mslMean sea-level pressureSingle
spSurface pressureSingle
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.365 + ], + "angle": 0, + "content": "Table 1. A summary of atmospheric variables. The 13 levels are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 hPa. 'Single' denotes the variables under earth's surface." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.419 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{1\\times 1\\times C}\\) is a learnable vector. With this loss function, the model could dynamically prioritize variables based on their distributional characteristics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.429, + 0.283, + 0.442 + ], + "angle": 0, + "content": "3.4.2. Reconstruction Loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.478 + ], + "angle": 0, + "content": "We introduce an additional reconstruction loss that connects the encoder and decoder directly, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.488, + 0.483, + 0.507 + ], + "angle": 0, + "content": "\\[\nO b j _ {r e c o n} = \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.639 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}^t = \\mathrm{Dec}(\\mathrm{Enc}(\\mathbf{X}^t))\\) represents the reconstructed variables. With the loss function, the encoder and decoder modules specialize in encoding and decoding features, while the intermediate transformer blocks focus on learning the data distribution. This ensures that the decoding process is entirely handled by decoder, allowing the intermediate blocks to focus on their task and not participate in decoding, thereby optimizing network efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.685 + ], + "angle": 0, + "content": "To this end, the model is trained with the final objective function, a combination of the dynamic prediction loss and reconstruction loss. The final objective is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.697, + 0.483, + 0.715 + ], + "angle": 0, + "content": "\\[\nO b j _ {f i n a l} = O b j _ {p r e d} + \\lambda O b j _ {r e c o n}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.726, + 0.33, + 0.74 + ], + "angle": 0, + "content": "where \\(\\lambda\\) denotes a hyper-parameter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.224, + 0.772 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset. In this work, we conduct experiments on a popular weather dataset, i.e., ERA5[18], provided by the ECMWF. ERA5 dataset is a reanalysis atmospheric dataset, consisting of the atmospheric variables from 1979 to the present day with a \\(0.25^{\\circ}\\) spatial resolution with \\(721 \\times 1440\\). We train the model on 40-year dataset in the initial stage and continuously train the model on 20-year dataset in the incremental stage, which contains the weather variables from" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.707, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.092, + 0.904, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.22, + 0.707, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.22, + 0.904, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.346, + 0.888, + 0.375 + ], + "angle": 0, + "content": "
——ClimaX——EWMoE——WeatherGFT——Stormer
——FourCastNet——ClimODE——GraphCast——Ours(IL)
——Panguo-Weather——Keisler
" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Comparative analysis of RMSE \\(\\downarrow\\) across 10 data-driven models for four variables, including Z500 and T850 (upper-air variables) in the initial stage, as well as T2M and U10 (incremental surface variables) in the incremental stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.547 + ], + "angle": 0, + "content": "1979 to 2020 and 2000 to 2020 year, respectively. And we test the model on one-year dataset of weather variables in 2021. In this work, the model processes 5 upper-air variables and 5 surface variables as Tab. 1, where the upper-air variables are exploited in the training stage, and the surface variables are exploited in the incremental stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.639 + ], + "angle": 0, + "content": "Implementation Details. The main structure of this work follows the novel backbones [7, 8, 12]. We apply the AdamW optimizer with 0.0002 and 0.00005 learning rates to initial and incremental stages, respectively. In two stages, we train 100 epochs and set batch size to 16. Our model is trained with PyTorch using 16 A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.652, + 0.665 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Fig. 4 presents a quantitative comparison of our structure with existing state-of-the-art approaches, including ClimaX [27], Pangu-Weather [3], ClimODE [36], WeatherGFT [39], FourCastNet [22], EWMoE [14], Keisler [20], GraphCast [23], and Stormer [28], for the 5-day weather prediction task. All the competitors are trained with both upper-air and surface variables on a \\(0.25^{\\circ}\\) weather dataset with a resolution of \\(721 \\times 1440\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For the surface variables, including T2M and U10, our VA-MoE demonstrates performance similar to the leading methods Stormer [28] and Graphcast [23]. Compared to other competitors, our structure shows significant advantages in both short-term and long-term predictions. For the upper-air variable Z500, our structure delivers one of the best performances among all 10 methods, outperforming" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7920" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.907, + 0.346 + ], + "angle": 0, + "content": "
Dataset(years)Iteration\\( (×10^4) \\)T2M(K)↓U10(m/s)↓V10(m/s)↓MSL(Pa)↓SP(Pa)↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
Plain Training
ViT* [13]1979-2020400.721.351.860.661.983.010.682.023.1140.2208.5393.963.3222.1397.0
IFS [32]1979-2020401.091.381.740.961.872.780.991.932.87------
Pangu-Weather [3]1979-2020400.821.091.530.771.632.540.791.682.65------
FourCastNet [22]1979-2020400.821.021.770.822.083.340.842.113.41------
ClimaX [27]1979-2020401.111.471.831.042.022.79---------
Graphcast [23]1979-2020400.510.941.370.381.512.37---23.4135.2278.2---
Fengwu [9]1979-2020400.581.031.410.421.532.32---23.2137.1276.9---
FuXi [10]1979-2020400.550.991.410.421.502.360.431.542.4427.2136.7282.9---
VA-MoE1979-2020400.571.031.420.431.412.250.441.462.3427.5131.1275.957.1168.9302.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (79-20)
VA-MoE (IL)1979-2020200.581.051.450.481.472.330.471.542.4127.9137.3281.659.3173.4312.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)2000-2020100.731.171.570.541.582.490.551.632.5730.0148.8304.760.6171.4314.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.356, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Table 2. Prediction performances on Incremental Training with 5 surface variables, i.e., T2M, U10, V10, MSL, and SP. * denotes running by ourselves. The best results are marked in bold. All experiments are in \\( {0.25}^{ \\circ } \\) with \\( {721} \\times {1440} \\) resolutions." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.398, + 0.907, + 0.533 + ], + "angle": 0, + "content": "
Para. (M)Z500(m2/s2) ↓Q500(×e-3, g/kg) ↓U500(m/s) ↓V500(m/s) ↓T500(K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
IFS [32]-28.31154.08333.960.310.610.751.433.235.121.403.585.640.360.981.70
Pangu-Weather [3]-24.88167.90391.260.250.550.690.963.134.730.913.525.150.270.941.56
Graphcast [23]-15.23125.42275.35---0.772.864.490.742.924.670.230.871.48
VA-MoE66519.28134.63295.520.170.490.620.842.994.710.843.044.890.250.761.36
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)13718.23133.14292.630.170.490.610.843.014.740.843.514.930.250.761.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.543, + 0.904, + 0.558 + ], + "angle": 0, + "content": "Table 3. Prediction performances on Initial Training with 5 upper-air variables. All experiments are in \\( {0.25}^{ \\circ } \\) with \\( {721} \\times {1440} \\) resolutions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Pangu-Weather [3] and Stormer [28], and significantly surpassing ClimaX [27] and FourCastNet [22], while slightly inferior to GraphCast [23]. For T850, our structure is slightly inferior to EWMoE [14], GraphCast [23], and Stormer [28] in the short-term predictions within 48 hours. After that period, our structure achieves the best performance in the long-term predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Results of Surface Variables Prediction. This study evaluates model performance on five surface variables: T2M, U10, V10, MSL, and SP. While competing structures are trained with both upper-air and surface variables, VA-MoE (IL) employs a two-stage training approach: initial training with upper-air variables followed by incremental training with all five surface variables." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis of the results from Tab. 2 demonstrates that VA-MoE outperforms competing models across multiple variables and settings, particularly in U10 and V10. Although VA-MoE shows marginally weaker performance in T2M compared to GraphCast [23] and comparable results to FengWu [9] and FuXi [10], it achieves superior long-term predictions for U10, V10, MSL, and SP. These indicate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.614 + ], + "angle": 0, + "content": "that while VA-MoE exhibits slight limitations in short-term forecasting, it excels in long-term prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.615, + 0.906, + 0.75 + ], + "angle": 0, + "content": "When implementing the incremental training strategy, VA-MoE (IL) trained on a 40-year dataset with only half the training iterations achieves performance comparable to standard VA-MoE. Even with a reduced 20-year dataset, VA-MoE(IL) maintains acceptable performance using just a quarter of the iterations required by the standard approach. These results underscore the efficacy of incremental learning for atmospheric datasets and suggest its potential to replace conventional training methods in operational settings." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Results of Upper-Air Variables Prediction. This study evaluates model performance on upper-air variables, comparing three training approaches: (1) VA-MoE, trained exclusively on upper-air variables; (2) VA-MoE (IL), which incorporates incremental training on surface variables after initial upper-air training; and (3) competing models trained simultaneously on both upper-air and surface variables." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 3, VA-MoE achieves performance comparable to GraphCast [23] and outperforms IFS [32] and Pangu-Weather [3] across both short- and long-term" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7921" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.09, + 0.904, + 0.198 + ], + "angle": 0, + "content": "
Para. (M)Z500 (m2/s2) ↓Q500 (×e-3, g/kg) ↓U500 (m/s) ↓V500 (m/s) ↓T500 (K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
ViT*[13]30733.38209.4517.810.220.611.061.243.666.521.223.767.410.421.182.40
ViT+MoE (light)*[30]60937.92207.11405.730.220.600.781.303.845.871.273.896.110.461.232.02
ViT+MoE* [30]111328.31169.61356.020.230.560.721.213.465.441.233.545.690.351.071.83
VA-MoE66520.59139.02302.130.180.490.620.913.024.760.913.084.970.270.921.59
VA-MoE (IL)13720.29138.52301.410.180.500.630.913.044.790.913.105.030.270.931.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.207, + 0.867, + 0.221 + ], + "angle": 0, + "content": "Table 4. Architectural impact on 5 upper-air variables under \\( {500}\\mathrm{{hPa}} \\) . All experiments are in \\( {1.5}^{ \\circ } \\) with \\( {128} \\times {256} \\) resolutions." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.224, + 0.877, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.464, + 0.892, + 0.478 + ], + "angle": 0, + "content": "Figure 5. 6-hour global weather prediction of upper-air and surface variables forecast visualization generated by VA-MoE framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.643 + ], + "angle": 0, + "content": "prediction. Notably, VA-MoE (IL) demonstrates robust predictive capabilities for upper-air variables even after incremental training on surface variables, despite using only a 20-year dataset - half the size of the initial training data. Furthermore, VA-MoE (IL) exhibits marginal performance improvements over the baseline in long-term predictions of Z500, confirming the absence of catastrophic forgetting during incremental training. These results highlight the efficacy of incremental training in VA-MoE(IL) for maintaining and enhancing predictive accuracy on original variables." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.245, + 0.667 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Impacts of Architecture. As shown in Tab. 4, we compare VA-MoE with other architectures, including ViT and ViT+MoE. VA-MoE significantly outperforms both ViT and ViT+MoE architectures across 6-hour, 72-hour, and 120-hour prediction horizons, despite ViT+MoE containing nearly twice the number of parameters. This performance gap highlights VA-MoE's suitability for weather forecasting tasks as its channel-adaptive expert design ensures parameter efficiency while maintaining accuracy, particularly in incremental learning. Considering the computational complexity, all ablation studies are in \\(1.5^{\\circ}\\) with \\(128 \\times 256\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.228, + 0.863 + ], + "angle": 0, + "content": "4.3. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In the work, we visualize some predicted results and parameters. In Fig. 5, we visualize the 6-hour predicted results of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.536 + ], + "angle": 0, + "content": "Z500, T850, U10, and T2M. There are max \\(0.08\\%\\), \\(0.22\\%\\), \\(1.59\\%\\), and \\(0.47\\%\\) absolute errors across 4 variables, including 2 upper-air variables and 2 surface variables." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.634, + 0.565 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.742 + ], + "angle": 0, + "content": "In this work, we proposed incremental weather forecasting, a novel task that addresses the challenge of dynamically expanding weather models to incorporate new variables without retraining from scratch. Our solution, Variables-Adaptive Mixture-of-Experts, enables parameter-efficient adaptation by selectively fine-tuning experts assigned to new variables through index embedding guidance, while preserving pretrained knowledge in existing parameters. Experimental results demonstrate that our approach achieves forecast accuracy comparable to fully retrained models while drastically reducing computational costs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.702, + 0.772 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This research was supported by the Hong Kong RGC General Research Fund (152169/22E, 152228/23E, 162161/24E), Research Impact Fund (No. R5060-19, No. R5011-23), Collaborative Research Fund (No. C1042-23GF), NSFC/RGC Collaborative Research Scheme (No. 62461160332 & CRS_HKUST602/24), Areas of Excellence Scheme (AoE/E-601/22-R), the InnoHK (HKGAI), and Key R&D Program of Zhejiang Province 2025C01075." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7922" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567): 47-55, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[3] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[4] Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan Weyn, Haiyu Dong, Anna Vaughan, et al. Aurora: A foundation model of the atmosphere. arXiv:2405.13063, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[5] Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: learning stable dynamics on the sphere. In Proceedings of International Conference on Machine Learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[6] Hao Chen, Yonghan Dong, Zhe-Ming Lu, Yunlong Yu, and Jungong Han. Self-prompting perceptual edge learning for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4528-4541, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[7] Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel matching network for cross-domain few-shot segmentation. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 978-987, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[8] Hao Chen, Yunlong Yu, Yonghan Dong, Zheming Lu, Yingming Li, and Zhongfei Zhang. Multi-content interaction network for few-shot segmentation. ACM Transactions on Multimedia Computing, Communications and Applications, 20 (6):1-20, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.684 + ], + "angle": 0, + "content": "[9] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv:2304.02948, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.484, + 0.756 + ], + "angle": 0, + "content": "[10] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.484, + 0.827 + ], + "angle": 0, + "content": "[11] Wuyang Chen, Yanqi Zhou, Nan Du, Yanping Huang, James Laudon, Zhifeng Chen, and Claire Cui. Lifelong language pretraining with distribution-specialized experts. In International Conference on Machine Learning, pages 5383-5395. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.484, + 0.884 + ], + "angle": 0, + "content": "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems, pages 16344-16359, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Trans-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "formers for image recognition at scale. arXiv:2010.11929, 2020.2,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[14] Lihao Gan, Xin Man, Chenghong Zhang, and Jie Shao. Ew-moe: An effective model for global weather forecasting with mixture-of-experts. arXiv:2405.06004, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[15] Yuan Gao, Hao Wu, Ruiqi Shu, Huanshuo Dong, Fan Xu, Rui Chen, Yibo Yan, Qingsong Wen, Xuming Hu, Kun Wang, et al. Oneforecast: A universal framework for global and regional weather forecasting. arXiv:2502.00338, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[16] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[18] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquin Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999-2049, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[20] Ryan Keisler. Forecasting global weather with graph neural networks. arXiv:2202.07575, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.59 + ], + "angle": 0, + "content": "[21] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.592, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[22] Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[23] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[24] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv:2010.08895, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7923" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[26] Franco Molteni, Roberto Buizza, Tim N Palmer, and Thomas Petroliagis. The ecmwf ensemble prediction system: Methodology and validation. Quarterly journal of the royal meteorological society, 122(529):73-119, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[27] Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. Climax: A foundation model for weather and climate. In International Conference on Machine Learning, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[28] Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. In Advances in Neural Information Processing Systems, pages 68740-68771, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[29] Youqi Pan, Wugen Zhou, Yingdian Cao, and Hongbin Zha. Adaptive vio: Deep visual-inertial odometry with online continual learning. arXiv:2405.16754, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[30] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. In European Conference on Computer Vision, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[31] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Probabilistic weather forecasting with machine learning. Nature, 637(8044):84-90, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[32] D Richardson. The thorpex interactive grand global ensemble (tigge). In Geophysical Research Abstracts, page 02815, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[33] Harold Ritchie, Clive Temperton, Adrian Simmons, Mariano Hortal, Terry Davies, David Dent, and Mats Hamrud. Implementation of the semi-lagrangian method in a high-resolution version of the ecmwf forecast model. Monthly Weather Review, 123(2):489-514, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[34] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23933-23942, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[35] Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li. Semantically-shifted incremental adapter-tuning is a continual vitransformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23252-23262, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[36] Yogesh Verma, Markus Heinonen, and Vikas Garg. ClimODE: Climate forecasting with physics-informed neural ODEs. In International Conference on Learning Representations, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14205-14215, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Wei Xiong, Muyuan Ma, Xiaomeng Huang, Ziyang Zhang, Pei Sun, and Yang Tian. Koopmanlab: machine learning for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "solving complex physics equations. APL Machine Learning, 1(3), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[39] Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, and Lei Bai. Generalizing weather forecast to fine-grained temporal scales via physics-ai hybrid modeling. In Advances in Neural Information Processing Systems, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[40] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23670-23680, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[41] Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. Multi-task dense prediction via mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27927-27937, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[42] Fei Ye and Adrian G Bors. Online task-free continual generative and discriminative learning via dynamic cluster memory. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 26202-26212, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.417, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[43] Donggeun Yoon, Minseok Seo, Doyi Kim, Yeji Choi, and Donghyeon Cho. Probabilistic weather forecasting with deterministic guidance-based diffusion model. In European Conference on Computer Vision, pages 108-124, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.555 + ], + "angle": 0, + "content": "[44] Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 27758-27767, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.558, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[45] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[46] Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. Skilful nowcasting of extreme precipitation with nowcastnet. Nature, 619(7970):526-532, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.753 + ], + "angle": 0, + "content": "[47] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, Cheng-Lin Liu, and Zhaoxiang Zhang. Rcl: Reliable continual learning for unified failure detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12140-12150, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7924" + } + ] +] \ No newline at end of file diff --git a/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_origin.pdf b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..095e7bd3ee8e54a5cf8a30d494522278ce0fd2f5 --- /dev/null +++ b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/122157a9-e33d-4cdd-b0f9-d8b983f2a104_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7eea66fdf5dc129b9d33f8e0ce279a1716764baf57971bd6accca28cff3537d +size 2144797 diff --git a/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/full.md b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..45bb05f8f2cdefb704348d81243be54eb3c9d16a --- /dev/null +++ b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/full.md @@ -0,0 +1,350 @@ +# VA-MoE: Variables-Adaptive Mixture of Experts for + +# Incremental Weather Forecasting + +![](images/408a686bb8235a3b61e408da6aca9f7a13881c231183d5f5fbc8364f783e3092.jpg) + +Hao Chen $^{1\dagger}$ Han Tao $^{1\dagger}$ Guo Song $^{1*}$ Jie Zhang $^{1}$ Yonghan Dong $^{2}$ Yunlong Yu $^{3}$ Lei Bai $^{4}$ $^{1}$ Hong Kong University of Science and Technology (HKUST) $^{2}$ Huawei Technologies Ltd. + $^{3}$ Zhejiang University $^{4}$ Shanghai AI Laboratory + +Equal contribution *Corresponding author {hchener, thanad}@connect.ust.hk + +{songguo, csejzhang}@ust.hk dongyonghan@huawei.com yuyunlong@zju.edu.cn bailei@pjlab.org.cn + +# Abstract + +This paper presents Variables-Adaptive Mixture of Experts (VA-MoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real-time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VA-MoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct sub-patterns of atmospheric variables (e.g., temperature, humidity, wind speed). Moreover, the proposed method employs a variable-adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on ERA5 dataset demonstrate that VA-MoE performs comparable against state-of-the-art models in both short-term (e.g., 1-3 days) and long-term (e.g., 5 days) forecasting tasks, with only about $25\%$ of trainable parameters and $50\%$ of the initial training data. Code: https://github.com/chenhao-zju/VAMoE + +# 1. Introduction + +Weather forecasting remains one of the most critical scientific challenges with profound socio-economic impacts. While traditional Numerical Weather Prediction (NWP) [2] relies on solving complex partial different equations to simulate atmospheric dynamics, recent advances in data-driven artificial intelligence [3, 4, 9, 22] have revolutionized earth system modeling. Deep learning models, in particular, have demonstrated remarkable potential in capturing diverse spatiotemporal patterns across weather and climate phenomena. + +![](images/a597efb3a93059332d6e7ca7100ecc4ea802f35eb2a605ebaae8713017f4c997.jpg) +Figure 1. Illustration of two different training paradigms of weather forecasting when new variables arrive: (c) Full-Train the whole model (d) Incremental-Train only the module corresponding new variables. (b) Performances of these two paradigms. + +ena, giving rise to the emerging field of AI for Weather. + +However, a fundamental limitation of existing AI4Weather approaches lies in their assumption that all variables are available synchronously during training and inference. In reality, meteorological variables are often heterogeneous in terms of data sources, collection frequencies, and spatial-temporal distributions. For instance, the upper-air variables (e.g., temperature profiles) are sparse and sampled via radiosondes/satellites, while the surface variable (e.g., precipitation, wind) are dense but updated in near-real-time. This asynchrony poses significant challenges: when introducing new variables (e.g., satellite-derived aerosol data), existing models must be entirely retrained from scratch, incurring prohibitive computational costs. For example, Pangu [3] required 64 days and 192 V100 GPUs for a full retraining cycle. + +To address this challenge, we propose Incremental Weather Forecasting (IWF), a novel paradigm designed to dynamically expand meteorological forecast variables in response to increasing observational data. By enabling parameter-efficient variable scaling, IWF revolutionizes operational meteorology by offering a streamlined alternative + +to conventional models. This innovation tackles a critical limitation in the field: the need for scalable modeling solutions as datasets grow gradually, ensuring adaptability without compromising computational efficiency. + +A pivotal limitation of existing incremental approaches is catastrophic forgetting, i.e., when introducing new variables, pre-trained parameters drift towards new distributions, causing significant performance degradation on original variables. To overcome this, we present the Variables-Adaptive Mixture of Experts (VA-MoE) model, the first expert-based system explicitly designed for hierarchical atmospheric variables. + +Mixture of Experts (MoE) models inherently enable scalable representation learning for multi-variable atmospheric datasets by dynamically expanding expert modules as input dimensions grow. While traditional MoE architectures excel in representational capacity, their parallelized design which distributes uniform weights across experts, often leads to homogeneous outputs, limiting specialized knowledge acquisition. To address this limitation, we introduce VA-MoE, an auxiliary-loss-free framework that decouples expert specialization from computational efficiency. VA-MoE employs a phased training strategy: Initial experts are pretrained on base variables, then incrementally expanded to accommodate new variables. Crucially, the pretrained experts are frozen during expansion phases to prevent catastrophic forgetting, ensuring robust retention of prior expertise while adapting to novel inputs. To foster expert diversity without auxiliary losses, VA-MoE incorporates variable index embeddings, i.e., positional metadata encodings that guide experts to develop domain-specific specialization. These embeddings dynamically activate contextually relevant computational pathways for each variable, optimizing resource allocation at inference time. By integrating variable index embeddings that encode spatial-temporal metadata, VA-MoE directs experts to specialize in specific atmospheric domains. This context-aware routing reduces inference latency compared to static MoE models while maintaining expert diversity. + +Besides, we propose a gradient-scaled variable-adaptive loss function that dynamically aligns variable optimization rates with their inherent spatiotemporal characteristics. By quantifying the distinct temporal evolution patterns of atmospheric variables, our method allocates differential loss weights, i.e., larger gradient magnitudes prioritize fast-transient fields like temperature to capture immediate atmospheric dynamics, while gradual weight adjustments stabilize slow-changing fields such as geopotential height to preserve long-term system behavior. + +As shown in Fig. 1, our paradigm employs a two-phase training paradigm: an initial phase optimizing base atmospheric variables, followed by an incremental phase integrating new variables while freezing previously trained ex + +perts. Experiments with the same 40-year dataset reveal that our paradigm Fig. 1(a)+(d) achieves superior z500 predictions performance compared to full retraining Fig. 1(a)+(c). + +In conclusion, the contributions of this work include: + +- This work initiates systematic research on incremental learning paradigms for weather forecasting. We propose the quantitative benchmark to evaluate the trade-offs between model scalability and generalization in incremental weather modeling. +- We present Variables-Adaptive Mixture of Experts (VA-MoE), the first framework tailored for incremental atmospheric modeling. VA-MoE achieves expert specialization through contextual variable activation driven by variable index embeddings, enabling dynamic assignment of experts to variables during both training and inference. +- Extensive experiments on the ERA5 dataset demonstrate that VA-MoE achieves comparable performance for surface variables, while delivering superior accuracy in upper-air variables against the existing competitors under $50\%$ reduced dataset size and $25\%$ fewer parameters. + +# 2. Related work + +# 2.1. Data-Driven Weather Forecasting + +In recent decades, traditional NWP [2] methods have dominated the weather forecasting field due to the robust predictions and rigorous mathematical validation. However, NWP methods [26, 33] require training from scratch for new prediction, resulting in slow computation and high costs. + +Recent advances in deep learning have catalyzed a paradigm shift toward data-driven models for medium-range weather forecasting. Pioneering works like FourCastNet [22] leverage Fourier Neural Operators (FNOs) [24] to learn spatiotemporal weather patterns, while PanguWeather [3] employs a 3D vision transformer [13, 25] to model atmospheric dynamics. Subsequent innovations, including Neural Operators [5, 24, 38] and specialized architectures [9, 10, 15, 23, 31, 43, 44, 46], have achieved accuracy rivaling NWP at a fraction of computation. + +While existing data-driven models excel at fixed-variable forecasting after costly initial training, their inflexibility poses a critical barrier: retraining from scratch is required to incorporate new variables, rendering cross-domain adaptation impractical. In contrast, our work introduces incremental weather forecasting, a novel paradigm evolving spatiotemporal patterns in real-time data. By decoupling variable-specific expertise from shared dynamics, our approach supports dynamic expansion to new atmospheric parameters with minimal retraining overhead. + +# 2.2. Mixture of Experts for Incremental Learning + +Advances in deep learning[6, 17] have spurred interest in parameter-efficient transfer strategies, where models adapt + +![](images/0f6c306f6dc259a2ccf6fca6af91be6c6aae1b11d49c857416c88197d46d07c4.jpg) + +![](images/124c88fa5df353a7a3ccd95e9e7a9a285c004e52ab5edbcdfdd0876865d9dbfd.jpg) + +![](images/d77f605a7ffeb8dd32d6de34d598d180acb7152790dd370ed404951bb38af79d.jpg) +Figure 2. Illustration of the incremental weather forecasting paradigm. (a), (b) are the initial and incremental stages, respectively. (c) presents the detailed structure in both initial and incremental stages. During the initial stage, the model is trained with upper-air variables(UV). In the incremental stage, the fire components are trained while the snow components remain frozen. The index embedding is trained by the Index Encoder and divided into six parts based on the upper-air variables (UV: Z, Q, U, V, T) and surface variables (SV). + +to new domains by fine-tuning only a subset of parameters. A key paradigm, incremental learning, addresses the stability-plasticity dilemma: retaining knowledge of prior tasks while integrating new concepts. However, real-world data often involves domain shifts, such as variations in variable distributions or task objectives, leading to two dominant research threads. Class Incremental Learning (CIL) [16, 21, 35, 47], focuses on expanding classification tasks with new classes while Task Incremental Learning (TIL) [29, 34, 40, 42] adapts models to distinct task sequences. + +One of the most popular structures for addressing incremental learning tasks is the Mixture of Experts (MoE) model [19], which learns representations of new concepts through different experts. Due to its sparse architecture, many works have adopted the MoE structure to reduce inference costs and enhance model capacity [37, 41]. An early work introducing MoE to incremental learning is Expert Gate [1], which trains multiple backbones as different experts and assigns new domains to the relevant expert. Lifelong-MoE [11] leverages pretrained experts and gates to retain the representational knowledge of the training domain. In addition to these approaches, MoE is used as an adapter in MoE-Adapter [45], where adapters are attached to the main structure. These works have demonstrated the promising performance of MoE in visual and natural language incremental learning. + +In contrast to MoE designed for fixed-variable systems, + +our propose incremental learning of variable distributions by assigning distinct pressure-level groups to specialized experts. Unlike prior incremental learning literature that primarily focus on task or dataset-specific adaptation, our incremental paradigm targets variable-centric expansion, enabling models to adapt to evolving observational data without retraining from scratch. + +# 3. Methodology + +Our work focuses on a novel paradigm of incremental weather forecasting with VA-MoE. In this section, we first formally define the problem in Sec. 3.1. Then, we present the structure of VA-MoE, consisting with the Channel-Adaptive Expert (CAE) and the Shared Expert in Sec. 3.2. The model during incremental stage is introduced in Sec. 3.3. Finally, loss function is presented in Sec. 3.4. + +# 3.1. Overview + +For the weather forecasting task, the AI-based model $\Phi$ aims to predict the future weather variables $\mathbf{X}^{t + 1}$ with the historical weather variables $\mathbf{X}^t$ , i.e., $\mathbf{X}^{t + 1} = \Phi (\mathbf{X}^t)$ . + +In the incremental weather forecasting task, the weather variables $\mathbf{X}^t$ are divided into two sets: (i) the initial training variables at $t$ time $\mathbf{X}_h^t \in \mathbb{R}^{H \times W \times N}$ , and, (ii) the incremental variables at $t$ time $\mathbf{X}_{sv}^t \in \mathbb{R}^{H \times W \times M}$ . Unlike traditional weather forecasting tasks, where models are trained + +![](images/ec231ab293262e1e38372cb270935bd9ebecfb3ad7ef96b9c46580860695194a.jpg) +Figure 3. Illustration of the VA-MoE. In the left subgraph, the input features comprise both upper-air and surface variables. During the initial training stage, five distinct CAE modules process the upper-air variables. In the incremental stage, surface variables are handled by a dedicated module, $\mathrm{CAESV}$ , while the original five CAEs remain frozen to preserve learned representations. The right subgraph details the CAE module. The Index Embedding is selected through a multi-layer process and applied to weight the Input Embedding via multiplication. Initially discrete, the index embedding is transformed into a continuous representation after passing through the GateEmbed layer, enabling it to dynamically reflect the relevance of specific channels to the input variables. + +with all variables available simultaneously, the incremental weather forecasting task sequentially incorporates dynamic variables as they become available over time. The incremental weather forecasting process is structured into a sequence of learning stages: an initial training stage followed by subsequent training stages. + +In the initial training stage, the available variables of 40 years are used to train a weather model, i.e., $\mathrm{X}_h^{t + 1} = \Phi_h(\mathrm{X}_h^t)$ , as shown in Fig. 2(a). In the incremental training stage, the half variables used at the initial stage and the surface variables at the incremental stage are incorporated together, to train the newly added modules for the surface variables while keeping the model parameters from the initial stage frozen, i.e., $\mathrm{X}^{t + 1} = \Phi (\mathrm{X}^t)$ , where $\Phi$ consists of both the fixed parameters of the pre-trained model $\Phi_h$ and the trainable parameters of $\Phi_{sv}$ , $\mathrm{X}^t$ consists of all the available variables at time $t$ , $\mathrm{X}^{t + 1}$ consists of the predicted variables at time $t + 1$ , as shown in Fig. 2(b). In this work, the upper-air variables are utilized during the initial training stage while the surface variables become available and are incorporated during the incremental training stage. The upper-air variables include five different types, Z, Q, U, V, and T, each defined across 13 different levels. + +# 3.2. VA-MoE + +We introduce index embedding within transformer blocks, a novel mechanism to dynamically guide experts in learning hierarchical relationships between meteorological vari + +ables. As new variables are incrementally integrated, corresponding experts are added to the transformer architecture and optimized via index-based affinity assignments. Building on this framework, we adapt the Mixture-of-Experts (MoE) paradigm which enables task-specific specialization and flexible integration of new experts—into weather forecasting transformers. + +The structure of the proposed VA-MoE is illustrated in Fig. 2(c). The input features, extracted by the encoder, first pass through a normalization layer and a self-attention layer. The output of the self-attention layer is then combined with the residual connection. Then, another normalization layer is applied, followed by a VA-MoE module, which is also equipped with a residual connection. This process can be formulated as: + +$$ +\begin{array}{l} \mathbf {X} _ {m i d} = \mathbf {X} _ {i n} + \operatorname {S A} (\text {L a y e r N o r m} (\mathbf {X} _ {\text {i n}})), \\ \mathbf {X} _ {i n} = \mathbf {X} _ {i n - 1} + \operatorname {V A M E} (\mathbf {I} _ {i n - 1}, \mathbf {N} _ {i n - 1} (\mathbf {X} _ {i n})) \end{array} \tag {1} +$$ + +$$ +\mathbf {X} _ {\text {o u t}} = \mathbf {X} _ {\text {m i d}} + \operatorname {V A - M o E} (\operatorname {L a y e r N o r m} (\mathbf {X} _ {\text {m i d}})), +$$ + +where $\mathbf{X}_{in}$ and $\mathbf{X}_{out}$ are the input and the output of each Transformer Block, respectively. + +# 3.2.1. Variables and Index Embedding + +As illustrated in Fig. 3, during the training stage, the input feature $\mathbf{X}_h^t\in \mathbb{R}^{H\times W\times N}$ consists of five different types of variables: $\mathbf{X}_Z^t,\mathbf{X}_Q^t,\mathbf{X}_U^t,\mathbf{X}_V^t,\mathbf{X}_T^t\in \mathbb{R}^{H\times W\times (N / 5)}$ , where, $H\times W$ denotes the spatial resolution. These five types of variables are concatenated and processed by a shared expert and five distinct Channel-Adaptive Experts (CAEs). + +In addition to these variables, the framework introduces an extra one-hot index embedding $\mathbf{I}_h\in \mathbb{R}^{5\times N}$ to guide experts in learning variable affinity. The number $N$ corresponds to the number of variables, and 5 corresponds to the number of variable types. Similar to the input weather features, the index embedding is composed of five separate embeddings: $\mathbf{I}_Z,\mathbf{I}_Q,\mathbf{I}_U,\mathbf{I}_V,\mathbf{I}_T\in \mathbb{R}^{1\times 1\times N}$ . This initial index embedding is then encoded into a latent space using a Linear layer, expressed as $\mathbf{I}_h = \mathrm{Linear}(\mathbf{I}_h)$ . + +# 3.2.2. Channel-Adaptive Expert (CAE) + +The key design of VA-MoE is the channel-adaptive expert, as shown in Fig. 3. For each variable $Z$ , the corresponding expert module processes both the feature embeddings of the input variables $\mathbf{X}_h^t$ and the index embedding $\mathbf{I}_Z$ , which is formulated as: + +$$ +\mathbf {X} _ {Z} ^ {t, C A E} = \operatorname {C A E} _ {Z} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {Z}\right), \tag {2} +$$ + +where $\mathrm{CAE}_Z$ is the abbreviation of channel-adaptive expert module for variable $Z$ . + +In the CAE module for variable $Z$ , the index embedding $\mathbf{I}_Z$ is encoded through a linear layer for feature projection and then multiplied channel-wise with the input embedding $\mathbf{X}_h^t$ . The resulting fused feature is processed + +by the GateEmbed layer, which produces a gate embedding to enhance the fusion of the index and variable embeddings. This enhanced feature is then normalized using a SoftMax layer and filtered through a TopK operation, which selects the top-K high-rank channels. This process yields the GateIndex $\mathrm{GI}_Z\in \mathbb{R}^{H\times W\times K}$ and GateWeight $\mathrm{GW}_Z\in \mathbb{R}^{H\times W\times K}$ . The process is formulated as: + +$$ +\mathbf {I} _ {Z} ^ {t o p k}, \mathbf {W} _ {Z} ^ {t o p k} = \mathrm {T O P} _ {k} (\operatorname {S o f t M a x} (\operatorname {M L P} _ {Z} (\mathbf {X} _ {h} ^ {t} \odot \mathbf {I} _ {Z}))), (3) +$$ + +where $\mathbf{I}_Z^{topk}$ and $\mathbf{W}_Z^{topk}$ denote the index and weight of the Top-k features. Once the GateIndex and GateWeight are computed, the input embedding is selectively processed based on the GateIndex, while the variable embedding is weighted channel-wise according to the GateWeight. The GateEmbed layer uses variable embeddings to guide the index embedding, enabling the creation of a variable-specific confidence matrix. Within the CAE module, this embedding enhances the confidence matrix by incorporating both semantic and positional information, rather than relying solely on positional information. + +The selected features for variable $Z$ are obtained with: + +$$ +\mathbf {X} _ {Z} ^ {t, s e l e c t e d} = \mathbf {W} _ {Z} ^ {t o p k} * \operatorname {S e l e c t} _ {Z} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {Z} ^ {t o p k}\right), \tag {4} +$$ + +where $\mathrm{Select}_Z$ denotes selecting Top-K embedding from the input feature $\mathbf{X}_h^t$ within the $\mathrm{CAE}_Z$ module. + +The selected and weighted embedding captures the key features associated with the specific type of variables. In the $\mathrm{CAE}_Z$ module, these features serve as the training set to model the distribution of the variable $Z$ with $\mathrm{Expert}_Z$ . The expert module adopts a parallel structure, splitting the channels into multiple parts and enhancing each part with a sparse MoE. This process is formulated as: + +$$ +\mathbf {X} _ {Z} ^ {t, C A E} = \operatorname {E x p e r t} _ {Z} (\mathbf {X} _ {Z} ^ {t, s e l e c t e d}), \tag {5} +$$ + +where $\mathbf{Expert}_Z$ is implemented with a multiple layer. + +# 3.2.3. Shared Expert. + +The shared expert processes the overall features of all variables and operates in parallel with the variable-specific modules. Each variable is handled by its corresponding CAE module, and their outputs are summarized to produce the fused features $\mathbf{X}_h^{t,fused}$ for all variables. Since the channel count of the selected variable embeddings is lower than that of the original embeddings, an additional up-channel linear layer, $\mathrm{Linear}_{\mathrm{up}}$ , is introduced. The fused features and the output from the shared expert are then combined via pixel-wise addition. This process is expressed as: + +$$ +\mathbf {X} _ {h} ^ {t, f u s e d} = \operatorname {C A E} _ {Z} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {Z}\right) + \dots + \operatorname {C A E} _ {T} \left(\mathbf {X} _ {h} ^ {t}, \mathbf {I} _ {T}\right), \tag {6} +$$ + +$$ +\left(\mathbf {X} ^ {\prime}\right) _ {h} ^ {t} = \operatorname {E x p e r t} _ {s h a r e d} \left(\mathbf {X} _ {h} ^ {t}\right) + \operatorname {L i n e a r} _ {u p} \left(\mathbf {X} _ {h} ^ {t, f u s e d}\right). \tag {7} +$$ + +# 3.3. Incremental Learning Stage + +When new variables are incrementally introduced to the model, specialized experts are dynamically integrated into the transformer blocks to process these variables. Concurrently, both the encoder and decoder architectures undergo synchronized updates to maintain compatibility with the expanded variable set. + +# 3.3.1. Encoder and Decoder Modules. + +When $K$ new variables are introduced incrementally, the encoder's input dimension expands to $H \times W \times (N + M)$ , necessitating an adjustment to the convolutional layer's kernel from $(3, 3, N, C)$ to $(3, 3, N + M, C)$ . To accommodate this, the layer's parameters are partitioned into two components: The $(3, 3, N, C)$ segment retains pretrained weights from earlier stages, ensuring continuity. The $(3, 3, M, C)$ segment, corresponding to the new variables, is initialized randomly to learn emergent patterns. + +The same strategy is applied to the position embedding and the Decoder module. + +# 3.3.2. VA-MoE Module. + +During the incremental learning phase, $M$ new experts are dynamically integrated into the MoE layer of each transformer block to accommodate $M$ newly introduced variables. In this phase, only the new experts and the shared expert are actively trained, while all preexisting experts remain frozen to preserve previously learned knowledge. + +# 3.3.3. Index Embedding Module. + +When $M$ new variables become available during the incremental stage, the index embedding changes from $\mathbf{I}_h\in$ $\mathbb{R}^{(N\times l)\times N}$ to $\mathbf{I}_h\in \mathbb{R}^{(N\times l + M\times r)\times (N + M)}$ , where each $M$ new variables has $r$ levels. Additionally, the index encoder needs to be retrained. + +# 3.4. Objective Function + +To train the model, we introduce a novel dynamic prediction loss combined with a reconstruction loss. + +# 3.4.1. Dynamic Prediction Loss. + +The prediction loss aims to minimize the difference between the model's predicted results and the actual future weather variables. Unlike existing weather forecasting methods that treat all the variables equally, we argue that different variables, such as temperature and surface pressure, follow distinct distributions. Thus, we introduce a novel dynamic prediction loss that dynamically assigns a weight to each channel of the input data. This loss is formulated as: + +$$ +O b j _ {p r e d} = \left(\hat {\mathbf {X}} ^ {t + 1} - \mathbf {X} ^ {t + 1}\right) \odot \left(\hat {\mathbf {X}} ^ {t + 1} - \mathbf {X} ^ {t + 1}\right) / e ^ {\mathbf {w}} + \mathbf {w}, (8) +$$ + +where $\odot$ denotes Hadamard product, $e$ denotes the base of the natural logarithm, $\mathbf{X}^{t + 1}\in \mathbb{R}^{H\times W\times C}$ , and $\mathbf{w}\in$ + +
NameDescriptionLevels
Upper-Air
ZGeopotential13
QSpecific humidity13
Ux-direction wind13
Vy-direction wind13
TTemperature13
Surface-Incremental Learning
u10x-direction wind at 10m heightSingle
v10y-direction wind at 10m heightSingle
t2mTemperature at 2m heightSingle
mslMean sea-level pressureSingle
spSurface pressureSingle
+ +$\mathbb{R}^{1\times 1\times C}$ is a learnable vector. With this loss function, the model could dynamically prioritize variables based on their distributional characteristics. + +# 3.4.2. Reconstruction Loss. + +We introduce an additional reconstruction loss that connects the encoder and decoder directly, formulated as: + +$$ +O b j _ {r e c o n} = \left(\hat {\mathbf {X}} ^ {t} - \mathbf {X} ^ {t}\right) \odot \left(\hat {\mathbf {X}} ^ {t} - \mathbf {X} ^ {t}\right), \tag {9} +$$ + +where $\hat{\mathbf{X}}^t = \mathrm{Dec}(\mathrm{Enc}(\mathbf{X}^t))$ represents the reconstructed variables. With the loss function, the encoder and decoder modules specialize in encoding and decoding features, while the intermediate transformer blocks focus on learning the data distribution. This ensures that the decoding process is entirely handled by decoder, allowing the intermediate blocks to focus on their task and not participate in decoding, thereby optimizing network efficiency. + +To this end, the model is trained with the final objective function, a combination of the dynamic prediction loss and reconstruction loss. The final objective is formulated as: + +$$ +O b j _ {f i n a l} = O b j _ {p r e d} + \lambda O b j _ {r e c o n}, \tag {10} +$$ + +where $\lambda$ denotes a hyper-parameter. + +# 4. Experiments + +Dataset. In this work, we conduct experiments on a popular weather dataset, i.e., ERA5[18], provided by the ECMWF. ERA5 dataset is a reanalysis atmospheric dataset, consisting of the atmospheric variables from 1979 to the present day with a $0.25^{\circ}$ spatial resolution with $721 \times 1440$ . We train the model on 40-year dataset in the initial stage and continuously train the model on 20-year dataset in the incremental stage, which contains the weather variables from + +![](images/5c32d76ef3271d634424358f221973a4c9a674afb759538b61aacedf3e343fc8.jpg) + +![](images/a0d1aa500d574a7e95efd8c0c9b28ef1e85a3a448b254b80980622a66a0bc2a2.jpg) + +![](images/052ad5c9ca6a4a26d4335087addf28eac50c9b1211888b4204be97c3469dbdba.jpg) +Figure 4. Comparative analysis of RMSE $\downarrow$ across 10 data-driven models for four variables, including Z500 and T850 (upper-air variables) in the initial stage, as well as T2M and U10 (incremental surface variables) in the incremental stage. + +![](images/bd3379cdf299482f4eddfea62f542bb052f9854ee8afec711d077e1cd24bf117.jpg) + +Table 1. A summary of atmospheric variables. The 13 levels are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 hPa. 'Single' denotes the variables under earth's surface. + +
——ClimaX——EWMoE——WeatherGFT——Stormer
——FourCastNet——ClimODE——GraphCast——Ours(IL)
——Panguo-Weather——Keisler
+ +1979 to 2020 and 2000 to 2020 year, respectively. And we test the model on one-year dataset of weather variables in 2021. In this work, the model processes 5 upper-air variables and 5 surface variables as Tab. 1, where the upper-air variables are exploited in the training stage, and the surface variables are exploited in the incremental stage. + +Implementation Details. The main structure of this work follows the novel backbones [7, 8, 12]. We apply the AdamW optimizer with 0.0002 and 0.00005 learning rates to initial and incremental stages, respectively. In two stages, we train 100 epochs and set batch size to 16. Our model is trained with PyTorch using 16 A100 GPUs. + +# 4.1. Main Results + +Fig. 4 presents a quantitative comparison of our structure with existing state-of-the-art approaches, including ClimaX [27], Pangu-Weather [3], ClimODE [36], WeatherGFT [39], FourCastNet [22], EWMoE [14], Keisler [20], GraphCast [23], and Stormer [28], for the 5-day weather prediction task. All the competitors are trained with both upper-air and surface variables on a $0.25^{\circ}$ weather dataset with a resolution of $721 \times 1440$ . + +For the surface variables, including T2M and U10, our VA-MoE demonstrates performance similar to the leading methods Stormer [28] and Graphcast [23]. Compared to other competitors, our structure shows significant advantages in both short-term and long-term predictions. For the upper-air variable Z500, our structure delivers one of the best performances among all 10 methods, outperforming + +
Dataset(years)Iteration\( (×10^4) \)T2M(K)↓U10(m/s)↓V10(m/s)↓MSL(Pa)↓SP(Pa)↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
Plain Training
ViT* [13]1979-2020400.721.351.860.661.983.010.682.023.1140.2208.5393.963.3222.1397.0
IFS [32]1979-2020401.091.381.740.961.872.780.991.932.87------
Pangu-Weather [3]1979-2020400.821.091.530.771.632.540.791.682.65------
FourCastNet [22]1979-2020400.821.021.770.822.083.340.842.113.41------
ClimaX [27]1979-2020401.111.471.831.042.022.79---------
Graphcast [23]1979-2020400.510.941.370.381.512.37---23.4135.2278.2---
Fengwu [9]1979-2020400.581.031.410.421.532.32---23.2137.1276.9---
FuXi [10]1979-2020400.550.991.410.421.502.360.431.542.4427.2136.7282.9---
VA-MoE1979-2020400.571.031.420.431.412.250.441.462.3427.5131.1275.957.1168.9302.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (79-20)
VA-MoE (IL)1979-2020200.581.051.450.481.472.330.471.542.4127.9137.3281.659.3173.4312.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)2000-2020100.731.171.570.541.582.490.551.632.5730.0148.8304.760.6171.4314.8
+ +Table 2. Prediction performances on Incremental Training with 5 surface variables, i.e., T2M, U10, V10, MSL, and SP. * denotes running by ourselves. The best results are marked in bold. All experiments are in ${0.25}^{ \circ }$ with ${721} \times {1440}$ resolutions. + +
Para. (M)Z500(m2/s2) ↓Q500(×e-3, g/kg) ↓U500(m/s) ↓V500(m/s) ↓T500(K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
IFS [32]-28.31154.08333.960.310.610.751.433.235.121.403.585.640.360.981.70
Pangu-Weather [3]-24.88167.90391.260.250.550.690.963.134.730.913.525.150.270.941.56
Graphcast [23]-15.23125.42275.35---0.772.864.490.742.924.670.230.871.48
VA-MoE66519.28134.63295.520.170.490.620.842.994.710.843.044.890.250.761.36
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)13718.23133.14292.630.170.490.610.843.014.740.843.514.930.250.761.37
+ +Table 3. Prediction performances on Initial Training with 5 upper-air variables. All experiments are in ${0.25}^{ \circ }$ with ${721} \times {1440}$ resolutions. + +Pangu-Weather [3] and Stormer [28], and significantly surpassing ClimaX [27] and FourCastNet [22], while slightly inferior to GraphCast [23]. For T850, our structure is slightly inferior to EWMoE [14], GraphCast [23], and Stormer [28] in the short-term predictions within 48 hours. After that period, our structure achieves the best performance in the long-term predictions. + +Results of Surface Variables Prediction. This study evaluates model performance on five surface variables: T2M, U10, V10, MSL, and SP. While competing structures are trained with both upper-air and surface variables, VA-MoE (IL) employs a two-stage training approach: initial training with upper-air variables followed by incremental training with all five surface variables. + +Analysis of the results from Tab. 2 demonstrates that VA-MoE outperforms competing models across multiple variables and settings, particularly in U10 and V10. Although VA-MoE shows marginally weaker performance in T2M compared to GraphCast [23] and comparable results to FengWu [9] and FuXi [10], it achieves superior long-term predictions for U10, V10, MSL, and SP. These indicate + +that while VA-MoE exhibits slight limitations in short-term forecasting, it excels in long-term prediction tasks. + +When implementing the incremental training strategy, VA-MoE (IL) trained on a 40-year dataset with only half the training iterations achieves performance comparable to standard VA-MoE. Even with a reduced 20-year dataset, VA-MoE(IL) maintains acceptable performance using just a quarter of the iterations required by the standard approach. These results underscore the efficacy of incremental learning for atmospheric datasets and suggest its potential to replace conventional training methods in operational settings. + +Results of Upper-Air Variables Prediction. This study evaluates model performance on upper-air variables, comparing three training approaches: (1) VA-MoE, trained exclusively on upper-air variables; (2) VA-MoE (IL), which incorporates incremental training on surface variables after initial upper-air training; and (3) competing models trained simultaneously on both upper-air and surface variables. + +As shown in Tab. 3, VA-MoE achieves performance comparable to GraphCast [23] and outperforms IFS [32] and Pangu-Weather [3] across both short- and long-term + +
Para. (M)Z500 (m2/s2) ↓Q500 (×e-3, g/kg) ↓U500 (m/s) ↓V500 (m/s) ↓T500 (K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
ViT*[13]30733.38209.4517.810.220.611.061.243.666.521.223.767.410.421.182.40
ViT+MoE (light)*[30]60937.92207.11405.730.220.600.781.303.845.871.273.896.110.461.232.02
ViT+MoE* [30]111328.31169.61356.020.230.560.721.213.465.441.233.545.690.351.071.83
VA-MoE66520.59139.02302.130.180.490.620.913.024.760.913.084.970.270.921.59
VA-MoE (IL)13720.29138.52301.410.180.500.630.913.044.790.913.105.030.270.931.60
+ +Table 4. Architectural impact on 5 upper-air variables under ${500}\mathrm{{hPa}}$ . All experiments are in ${1.5}^{ \circ }$ with ${128} \times {256}$ resolutions. + +![](images/7a3f0e25dc8da0ded62ed82c5f1ef56cb67792b1eb0e64e3699a3a53e9edaebc.jpg) +Figure 5. 6-hour global weather prediction of upper-air and surface variables forecast visualization generated by VA-MoE framework. + +prediction. Notably, VA-MoE (IL) demonstrates robust predictive capabilities for upper-air variables even after incremental training on surface variables, despite using only a 20-year dataset - half the size of the initial training data. Furthermore, VA-MoE (IL) exhibits marginal performance improvements over the baseline in long-term predictions of Z500, confirming the absence of catastrophic forgetting during incremental training. These results highlight the efficacy of incremental training in VA-MoE(IL) for maintaining and enhancing predictive accuracy on original variables. + +# 4.2. Ablation Study + +Impacts of Architecture. As shown in Tab. 4, we compare VA-MoE with other architectures, including ViT and ViT+MoE. VA-MoE significantly outperforms both ViT and ViT+MoE architectures across 6-hour, 72-hour, and 120-hour prediction horizons, despite ViT+MoE containing nearly twice the number of parameters. This performance gap highlights VA-MoE's suitability for weather forecasting tasks as its channel-adaptive expert design ensures parameter efficiency while maintaining accuracy, particularly in incremental learning. Considering the computational complexity, all ablation studies are in $1.5^{\circ}$ with $128 \times 256$ . + +# 4.3. Visualization + +In the work, we visualize some predicted results and parameters. In Fig. 5, we visualize the 6-hour predicted results of + +Z500, T850, U10, and T2M. There are max $0.08\%$ , $0.22\%$ , $1.59\%$ , and $0.47\%$ absolute errors across 4 variables, including 2 upper-air variables and 2 surface variables. + +# 5. Conclusion + +In this work, we proposed incremental weather forecasting, a novel task that addresses the challenge of dynamically expanding weather models to incorporate new variables without retraining from scratch. Our solution, Variables-Adaptive Mixture-of-Experts, enables parameter-efficient adaptation by selectively fine-tuning experts assigned to new variables through index embedding guidance, while preserving pretrained knowledge in existing parameters. Experimental results demonstrate that our approach achieves forecast accuracy comparable to fully retrained models while drastically reducing computational costs. + +# 6. Acknowledgements + +This research was supported by the Hong Kong RGC General Research Fund (152169/22E, 152228/23E, 162161/24E), Research Impact Fund (No. R5060-19, No. R5011-23), Collaborative Research Fund (No. C1042-23GF), NSFC/RGC Collaborative Research Scheme (No. 62461160332 & CRS_HKUST602/24), Areas of Excellence Scheme (AoE/E-601/22-R), the InnoHK (HKGAI), and Key R&D Program of Zhejiang Province 2025C01075. + +# References + +[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3 +[2] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567): 47-55, 2015. 1, 2 +[3] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 1, 2, 6, 7 +[4] Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan Weyn, Haiyu Dong, Anna Vaughan, et al. Aurora: A foundation model of the atmosphere. arXiv:2405.13063, 2024. 1 +[5] Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: learning stable dynamics on the sphere. In Proceedings of International Conference on Machine Learning, 2023. 2 +[6] Hao Chen, Yonghan Dong, Zhe-Ming Lu, Yunlong Yu, and Jungong Han. Self-prompting perceptual edge learning for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4528-4541, 2023. 2 +[7] Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel matching network for cross-domain few-shot segmentation. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 978-987, 2024. 6 +[8] Hao Chen, Yunlong Yu, Yonghan Dong, Zheming Lu, Yingming Li, and Zhongfei Zhang. Multi-content interaction network for few-shot segmentation. ACM Transactions on Multimedia Computing, Communications and Applications, 20 (6):1-20, 2024. 6 +[9] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv:2304.02948, 2023. 1, 2, 7 +[10] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 2, 7 +[11] Wuyang Chen, Yanqi Zhou, Nan Du, Yanping Huang, James Laudon, Zhifeng Chen, and Claire Cui. Lifelong language pretraining with distribution-specialized experts. In International Conference on Machine Learning, pages 5383-5395. PMLR, 2023. 3 +[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems, pages 16344-16359, 2022. 6 +[13] Alexey Dosovitskiy. An image is worth 16x16 words: Trans- + +formers for image recognition at scale. arXiv:2010.11929, 2020.2,7,8 +[14] Lihao Gan, Xin Man, Chenghong Zhang, and Jie Shao. Ew-moe: An effective model for global weather forecasting with mixture-of-experts. arXiv:2405.06004, 2024. 6, 7 +[15] Yuan Gao, Hao Wu, Ruiqi Shu, Huanshuo Dong, Fan Xu, Rui Chen, Yibo Yan, Qingsong Wen, Xuming Hu, Kun Wang, et al. Oneforecast: A universal framework for global and regional weather forecasting. arXiv:2502.00338, 2025. 2 +[16] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 3 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2 +[18] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquin Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999-2049, 2020. 6 +[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3 +[20] Ryan Keisler. Forecasting global weather with graph neural networks. arXiv:2202.07575, 2022. 6 +[21] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. 3 +[22] Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. 1, 2, 6, 7 +[23] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2, 6, 7 +[24] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv:2010.08895, 2020. 2 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 2 + +[26] Franco Molteni, Roberto Buizza, Tim N Palmer, and Thomas Petroliagis. The ecmwf ensemble prediction system: Methodology and validation. Quarterly journal of the royal meteorological society, 122(529):73-119, 1996. 2 +[27] Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. Climax: A foundation model for weather and climate. In International Conference on Machine Learning, 2023. 6, 7 +[28] Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. In Advances in Neural Information Processing Systems, pages 68740-68771, 2025. 6, 7 +[29] Youqi Pan, Wugen Zhou, Yingdian Cao, and Hongbin Zha. Adaptive vio: Deep visual-inertial odometry with online continual learning. arXiv:2405.16754, 2024. 3 +[30] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. In European Conference on Computer Vision, 2024. 8 +[31] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Probabilistic weather forecasting with machine learning. Nature, 637(8044):84-90, 2025. 2 +[32] D Richardson. The thorpex interactive grand global ensemble (tigge). In Geophysical Research Abstracts, page 02815, 2024. 7 +[33] Harold Ritchie, Clive Temperton, Adrian Simmons, Mariano Hortal, Terry Davies, David Dent, and Mats Hamrud. Implementation of the semi-lagrangian method in a high-resolution version of the ecmwf forecast model. Monthly Weather Review, 123(2):489-514, 1995. 2 +[34] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23933-23942, 2024. 3 +[35] Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li. Semantically-shifted incremental adapter-tuning is a continual vitransformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23252-23262, 2024. 3 +[36] Yogesh Verma, Markus Heinonen, and Vikas Garg. ClimODE: Climate forecasting with physics-informed neural ODEs. In International Conference on Learning Representations, 2024. 6 +[37] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14205-14215, 2024. 3 +[38] Wei Xiong, Muyuan Ma, Xiaomeng Huang, Ziyang Zhang, Pei Sun, and Yang Tian. Koopmanlab: machine learning for + +solving complex physics equations. APL Machine Learning, 1(3), 2023. 2 +[39] Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, and Lei Bai. Generalizing weather forecast to fine-grained temporal scales via physics-ai hybrid modeling. In Advances in Neural Information Processing Systems, 2024. 6 +[40] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23670-23680, 2024. 3 +[41] Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. Multi-task dense prediction via mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27927-27937, 2024. 3 +[42] Fei Ye and Adrian G Bors. Online task-free continual generative and discriminative learning via dynamic cluster memory. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 26202-26212, 2024. 3 +[43] Donggeun Yoon, Minseok Seo, Doyi Kim, Yeji Choi, and Donghyeon Cho. Probabilistic weather forecasting with deterministic guidance-based diffusion model. In European Conference on Computer Vision, pages 108-124, 2024. 2 +[44] Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 27758-27767, 2024. 2 +[45] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 3 +[46] Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. Skilful nowcasting of extreme precipitation with nowcastnet. Nature, 619(7970):526-532, 2023. 2 +[47] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, Cheng-Lin Liu, and Zhaoxiang Zhang. Rcl: Reliable continual learning for unified failure detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12140-12150, 2024. 3 \ No newline at end of file diff --git a/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/images.zip b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..653e88df3feefe1076dd30bc29dabb1235370012 --- /dev/null +++ b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bce7286370f0dfd322f20959f8c31c8d2d26e21b49e7e7bbc481daaefe53f0aa +size 818512 diff --git a/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/layout.json b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2bb4efcd7b15f81526fb9f7f22b01ee442e4d95 --- /dev/null +++ b/2025/VA-MoE_ Variables-Adaptive Mixture of Experts for Incremental Weather Forecasting/layout.json @@ -0,0 +1,9100 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 468, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 468, + 121 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 468, + 121 + ], + "type": "text", + "content": "VA-MoE: Variables-Adaptive Mixture of Experts for" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 129, + 399, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 129, + 399, + 146 + ], + "spans": [ + { + "bbox": [ + 189, + 129, + 399, + 146 + ], + "type": "text", + "content": "Incremental Weather Forecasting" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 400, + 120, + 421, + 142 + ], + "blocks": [ + { + "bbox": [ + 400, + 120, + 421, + 142 + ], + "lines": [ + { + "bbox": [ + 400, + 120, + 421, + 142 + ], + "spans": [ + { + "bbox": [ + 400, + 120, + 421, + 142 + ], + "type": "image", + "image_path": "408a686bb8235a3b61e408da6aca9f7a13881c231183d5f5fbc8364f783e3092.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": "Hao Chen" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": " Han Tao" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": " Guo Song" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": " Jie Zhang" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": " Yonghan Dong" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": " Yunlong Yu" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": " Lei Bai" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology (HKUST) " + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": "Huawei Technologies Ltd. \n" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 84, + 167, + 547, + 210 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 213, + 515, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 213, + 515, + 224 + ], + "spans": [ + { + "bbox": [ + 111, + 213, + 515, + 224 + ], + "type": "text", + "content": "Equal contribution *Corresponding author {hchener, thanad}@connect.ust.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 226, + 562, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 226, + 562, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 226, + 562, + 239 + ], + "type": "text", + "content": "{songguo, csejzhang}@ust.hk dongyonghan@huawei.com yuyunlong@zju.edu.cn bailei@pjlab.org.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 266, + 200, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 266, + 200, + 278 + ], + "spans": [ + { + "bbox": [ + 151, + 266, + 200, + 278 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 293, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 558 + ], + "type": "text", + "content": "This paper presents Variables-Adaptive Mixture of Experts (VA-MoE), a novel framework for incremental weather forecasting that dynamically adapts to evolving spatiotemporal patterns in real-time data. Traditional weather prediction models often struggle with exorbitant computational expenditure and the need to continuously update forecasts as new observations arrive. VA-MoE addresses these challenges by leveraging a hybrid architecture of experts, where each expert specializes in capturing distinct sub-patterns of atmospheric variables (e.g., temperature, humidity, wind speed). Moreover, the proposed method employs a variable-adaptive gating mechanism to dynamically select and combine relevant experts based on the input context, enabling efficient knowledge distillation and parameter sharing. This design significantly reduces computational overhead while maintaining high forecast accuracy. Experiments on ERA5 dataset demonstrate that VA-MoE performs comparable against state-of-the-art models in both short-term (e.g., 1-3 days) and long-term (e.g., 5 days) forecasting tasks, with only about " + }, + { + "bbox": [ + 55, + 293, + 296, + 558 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 293, + 296, + 558 + ], + "type": "text", + "content": " of trainable parameters and " + }, + { + "bbox": [ + 55, + 293, + 296, + 558 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 293, + 296, + 558 + ], + "type": "text", + "content": " of the initial training data. Code: https://github.com/chenhao-zju/VAMoE" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Weather forecasting remains one of the most critical scientific challenges with profound socio-economic impacts. While traditional Numerical Weather Prediction (NWP) [2] relies on solving complex partial different equations to simulate atmospheric dynamics, recent advances in data-driven artificial intelligence [3, 4, 9, 22] have revolutionized earth system modeling. Deep learning models, in particular, have demonstrated remarkable potential in capturing diverse spatiotemporal patterns across weather and climate phenomena." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 264, + 553, + 387 + ], + "blocks": [ + { + "bbox": [ + 317, + 264, + 553, + 387 + ], + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 387 + ], + "type": "image", + "image_path": "a597efb3a93059332d6e7ca7100ecc4ea802f35eb2a605ebaae8713017f4c997.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 398, + 555, + 443 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 443 + ], + "type": "text", + "content": "Figure 1. Illustration of two different training paradigms of weather forecasting when new variables arrive: (c) Full-Train the whole model (d) Incremental-Train only the module corresponding new variables. (b) Performances of these two paradigms." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 539, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 539, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 539, + 462 + ], + "type": "text", + "content": "ena, giving rise to the emerging field of AI for Weather." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 462, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 555, + 641 + ], + "type": "text", + "content": "However, a fundamental limitation of existing AI4Weather approaches lies in their assumption that all variables are available synchronously during training and inference. In reality, meteorological variables are often heterogeneous in terms of data sources, collection frequencies, and spatial-temporal distributions. For instance, the upper-air variables (e.g., temperature profiles) are sparse and sampled via radiosondes/satellites, while the surface variable (e.g., precipitation, wind) are dense but updated in near-real-time. This asynchrony poses significant challenges: when introducing new variables (e.g., satellite-derived aerosol data), existing models must be entirely retrained from scratch, incurring prohibitive computational costs. For example, Pangu [3] required 64 days and 192 V100 GPUs for a full retraining cycle." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "To address this challenge, we propose Incremental Weather Forecasting (IWF), a novel paradigm designed to dynamically expand meteorological forecast variables in response to increasing observational data. By enabling parameter-efficient variable scaling, IWF revolutionizes operational meteorology by offering a streamlined alternative" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7915" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "to conventional models. This innovation tackles a critical limitation in the field: the need for scalable modeling solutions as datasets grow gradually, ensuring adaptability without compromising computational efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 294, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 294, + 217 + ], + "type": "text", + "content": "A pivotal limitation of existing incremental approaches is catastrophic forgetting, i.e., when introducing new variables, pre-trained parameters drift towards new distributions, causing significant performance degradation on original variables. To overcome this, we present the Variables-Adaptive Mixture of Experts (VA-MoE) model, the first expert-based system explicitly designed for hierarchical atmospheric variables." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 219, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 542 + ], + "type": "text", + "content": "Mixture of Experts (MoE) models inherently enable scalable representation learning for multi-variable atmospheric datasets by dynamically expanding expert modules as input dimensions grow. While traditional MoE architectures excel in representational capacity, their parallelized design which distributes uniform weights across experts, often leads to homogeneous outputs, limiting specialized knowledge acquisition. To address this limitation, we introduce VA-MoE, an auxiliary-loss-free framework that decouples expert specialization from computational efficiency. VA-MoE employs a phased training strategy: Initial experts are pretrained on base variables, then incrementally expanded to accommodate new variables. Crucially, the pretrained experts are frozen during expansion phases to prevent catastrophic forgetting, ensuring robust retention of prior expertise while adapting to novel inputs. To foster expert diversity without auxiliary losses, VA-MoE incorporates variable index embeddings, i.e., positional metadata encodings that guide experts to develop domain-specific specialization. These embeddings dynamically activate contextually relevant computational pathways for each variable, optimizing resource allocation at inference time. By integrating variable index embeddings that encode spatial-temporal metadata, VA-MoE directs experts to specialize in specific atmospheric domains. This context-aware routing reduces inference latency compared to static MoE models while maintaining expert diversity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 294, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 294, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 294, + 663 + ], + "type": "text", + "content": "Besides, we propose a gradient-scaled variable-adaptive loss function that dynamically aligns variable optimization rates with their inherent spatiotemporal characteristics. By quantifying the distinct temporal evolution patterns of atmospheric variables, our method allocates differential loss weights, i.e., larger gradient magnitudes prioritize fast-transient fields like temperature to capture immediate atmospheric dynamics, while gradual weight adjustments stabilize slow-changing fields such as geopotential height to preserve long-term system behavior." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 714 + ], + "type": "text", + "content": "As shown in Fig. 1, our paradigm employs a two-phase training paradigm: an initial phase optimizing base atmospheric variables, followed by an incremental phase integrating new variables while freezing previously trained ex" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "perts. Experiments with the same 40-year dataset reveal that our paradigm Fig. 1(a)+(d) achieves superior z500 predictions performance compared to full retraining Fig. 1(a)+(c)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 109, + 541, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 109, + 541, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 109, + 541, + 120 + ], + "type": "text", + "content": "In conclusion, the contributions of this work include:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 553, + 312 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 180 + ], + "type": "text", + "content": "- This work initiates systematic research on incremental learning paradigms for weather forecasting. We propose the quantitative benchmark to evaluate the trade-offs between model scalability and generalization in incremental weather modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 181, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 252 + ], + "type": "text", + "content": "- We present Variables-Adaptive Mixture of Experts (VA-MoE), the first framework tailored for incremental atmospheric modeling. VA-MoE achieves expert specialization through contextual variable activation driven by variable index embeddings, enabling dynamic assignment of experts to variables during both training and inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 253, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 312 + ], + "type": "text", + "content": "- Extensive experiments on the ERA5 dataset demonstrate that VA-MoE achieves comparable performance for surface variables, while delivering superior accuracy in upper-air variables against the existing competitors under " + }, + { + "bbox": [ + 313, + 253, + 553, + 312 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 253, + 553, + 312 + ], + "type": "text", + "content": " reduced dataset size and " + }, + { + "bbox": [ + 313, + 253, + 553, + 312 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 253, + 553, + 312 + ], + "type": "text", + "content": " fewer parameters." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 324, + 398, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 398, + 337 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 398, + 337 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 345, + 495, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 345, + 495, + 357 + ], + "spans": [ + { + "bbox": [ + 314, + 345, + 495, + 357 + ], + "type": "text", + "content": "2.1. Data-Driven Weather Forecasting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 362, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 553, + 422 + ], + "type": "text", + "content": "In recent decades, traditional NWP [2] methods have dominated the weather forecasting field due to the robust predictions and rigorous mathematical validation. However, NWP methods [26, 33] require training from scratch for new prediction, resulting in slow computation and high costs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 423, + 554, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 542 + ], + "type": "text", + "content": "Recent advances in deep learning have catalyzed a paradigm shift toward data-driven models for medium-range weather forecasting. Pioneering works like FourCastNet [22] leverage Fourier Neural Operators (FNOs) [24] to learn spatiotemporal weather patterns, while PanguWeather [3] employs a 3D vision transformer [13, 25] to model atmospheric dynamics. Subsequent innovations, including Neural Operators [5, 24, 38] and specialized architectures [9, 10, 15, 23, 31, 43, 44, 46], have achieved accuracy rivaling NWP at a fraction of computation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 543, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 663 + ], + "type": "text", + "content": "While existing data-driven models excel at fixed-variable forecasting after costly initial training, their inflexibility poses a critical barrier: retraining from scratch is required to incorporate new variables, rendering cross-domain adaptation impractical. In contrast, our work introduces incremental weather forecasting, a novel paradigm evolving spatiotemporal patterns in real-time data. By decoupling variable-specific expertise from shared dynamics, our approach supports dynamic expansion to new atmospheric parameters with minimal retraining overhead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 671, + 549, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 549, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 549, + 685 + ], + "type": "text", + "content": "2.2. Mixture of Experts for Incremental Learning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Advances in deep learning[6, 17] have spurred interest in parameter-efficient transfer strategies, where models adapt" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7916" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 73, + 281, + 143 + ], + "blocks": [ + { + "bbox": [ + 71, + 73, + 281, + 143 + ], + "lines": [ + { + "bbox": [ + 71, + 73, + 281, + 143 + ], + "spans": [ + { + "bbox": [ + 71, + 73, + 281, + 143 + ], + "type": "image", + "image_path": "0f6c306f6dc259a2ccf6fca6af91be6c6aae1b11d49c857416c88197d46d07c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 283, + 74, + 539, + 143 + ], + "blocks": [ + { + "bbox": [ + 283, + 74, + 539, + 143 + ], + "lines": [ + { + "bbox": [ + 283, + 74, + 539, + 143 + ], + "spans": [ + { + "bbox": [ + 283, + 74, + 539, + 143 + ], + "type": "image", + "image_path": "124c88fa5df353a7a3ccd95e9e7a9a285c004e52ab5edbcdfdd0876865d9dbfd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 71, + 144, + 539, + 318 + ], + "blocks": [ + { + "bbox": [ + 71, + 144, + 539, + 318 + ], + "lines": [ + { + "bbox": [ + 71, + 144, + 539, + 318 + ], + "spans": [ + { + "bbox": [ + 71, + 144, + 539, + 318 + ], + "type": "image", + "image_path": "d77f605a7ffeb8dd32d6de34d598d180acb7152790dd370ed404951bb38af79d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "type": "text", + "content": "Figure 2. Illustration of the incremental weather forecasting paradigm. (a), (b) are the initial and incremental stages, respectively. (c) presents the detailed structure in both initial and incremental stages. During the initial stage, the model is trained with upper-air variables(UV). In the incremental stage, the fire components are trained while the snow components remain frozen. The index embedding is trained by the Index Encoder and divided into six parts based on the upper-air variables (UV: Z, Q, U, V, T) and surface variables (SV)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 390, + 294, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 294, + 510 + ], + "type": "text", + "content": "to new domains by fine-tuning only a subset of parameters. A key paradigm, incremental learning, addresses the stability-plasticity dilemma: retaining knowledge of prior tasks while integrating new concepts. However, real-world data often involves domain shifts, such as variations in variable distributions or task objectives, leading to two dominant research threads. Class Incremental Learning (CIL) [16, 21, 35, 47], focuses on expanding classification tasks with new classes while Task Incremental Learning (TIL) [29, 34, 40, 42] adapts models to distinct task sequences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 511, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 701 + ], + "type": "text", + "content": "One of the most popular structures for addressing incremental learning tasks is the Mixture of Experts (MoE) model [19], which learns representations of new concepts through different experts. Due to its sparse architecture, many works have adopted the MoE structure to reduce inference costs and enhance model capacity [37, 41]. An early work introducing MoE to incremental learning is Expert Gate [1], which trains multiple backbones as different experts and assigns new domains to the relevant expert. Lifelong-MoE [11] leverages pretrained experts and gates to retain the representational knowledge of the training domain. In addition to these approaches, MoE is used as an adapter in MoE-Adapter [45], where adapters are attached to the main structure. These works have demonstrated the promising performance of MoE in visual and natural language incremental learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "content": "In contrast to MoE designed for fixed-variable systems," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 390, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 475 + ], + "type": "text", + "content": "our propose incremental learning of variable distributions by assigning distinct pressure-level groups to specialized experts. Unlike prior incremental learning literature that primarily focus on task or dataset-specific adaptation, our incremental paradigm targets variable-centric expansion, enabling models to adapt to evolving observational data without retraining from scratch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 486, + 397, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 397, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 397, + 500 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 506, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 592 + ], + "type": "text", + "content": "Our work focuses on a novel paradigm of incremental weather forecasting with VA-MoE. In this section, we first formally define the problem in Sec. 3.1. Then, we present the structure of VA-MoE, consisting with the Channel-Adaptive Expert (CAE) and the Shared Expert in Sec. 3.2. The model during incremental stage is introduced in Sec. 3.3. Finally, loss function is presented in Sec. 3.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 599, + 382, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 382, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 382, + 611 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "text", + "content": "For the weather forecasting task, the AI-based model " + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "text", + "content": " aims to predict the future weather variables " + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t + 1}" + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "text", + "content": " with the historical weather variables " + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t + 1} = \\Phi (\\mathbf{X}^t)" + }, + { + "bbox": [ + 313, + 616, + 554, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "In the incremental weather forecasting task, the weather variables " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " are divided into two sets: (i) the initial training variables at " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " time " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_h^t \\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": ", and, (ii) the incremental variables at " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " time " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{sv}^t \\in \\mathbb{R}^{H \\times W \\times M}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": ". Unlike traditional weather forecasting tasks, where models are trained" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7917" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 224 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 224 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 224 + ], + "type": "image", + "image_path": "ec231ab293262e1e38372cb270935bd9ebecfb3ad7ef96b9c46580860695194a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "type": "text", + "content": "Figure 3. Illustration of the VA-MoE. In the left subgraph, the input features comprise both upper-air and surface variables. During the initial training stage, five distinct CAE modules process the upper-air variables. In the incremental stage, surface variables are handled by a dedicated module, " + }, + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\mathrm{CAESV}" + }, + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "type": "text", + "content": ", while the original five CAEs remain frozen to preserve learned representations. The right subgraph details the CAE module. The Index Embedding is selected through a multi-layer process and applied to weight the Input Embedding via multiplication. Initially discrete, the index embedding is transformed into a continuous representation after passing through the GateEmbed layer, enabling it to dynamically reflect the relevance of specific channels to the input variables." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": "with all variables available simultaneously, the incremental weather forecasting task sequentially incorporates dynamic variables as they become available over time. The incremental weather forecasting process is structured into a sequence of learning stages: an initial training stage followed by subsequent training stages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": "In the initial training stage, the available variables of 40 years are used to train a weather model, i.e., " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\mathrm{X}_h^{t + 1} = \\Phi_h(\\mathrm{X}_h^t)" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": ", as shown in Fig. 2(a). In the incremental training stage, the half variables used at the initial stage and the surface variables at the incremental stage are incorporated together, to train the newly added modules for the surface variables while keeping the model parameters from the initial stage frozen, i.e., " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\mathrm{X}^{t + 1} = \\Phi (\\mathrm{X}^t)" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": " consists of both the fixed parameters of the pre-trained model " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\Phi_h" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": " and the trainable parameters of " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\Phi_{sv}" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\mathrm{X}^t" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": " consists of all the available variables at time " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\mathrm{X}^{t + 1}" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": " consists of the predicted variables at time " + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 446, + 296, + 649 + ], + "type": "text", + "content": ", as shown in Fig. 2(b). In this work, the upper-air variables are utilized during the initial training stage while the surface variables become available and are incorporated during the incremental training stage. The upper-air variables include five different types, Z, Q, U, V, and T, each defined across 13 different levels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 658, + 119, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 119, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 119, + 670 + ], + "type": "text", + "content": "3.2. VA-MoE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We introduce index embedding within transformer blocks, a novel mechanism to dynamically guide experts in learning hierarchical relationships between meteorological vari" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "ables. As new variables are incrementally integrated, corresponding experts are added to the transformer architecture and optimized via index-based affinity assignments. Building on this framework, we adapt the Mixture-of-Experts (MoE) paradigm which enables task-specific specialization and flexible integration of new experts—into weather forecasting transformers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 156, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 251 + ], + "type": "text", + "content": "The structure of the proposed VA-MoE is illustrated in Fig. 2(c). The input features, extracted by the encoder, first pass through a normalization layer and a self-attention layer. The output of the self-attention layer is then combined with the residual connection. Then, another normalization layer is applied, followed by a VA-MoE module, which is also equipped with a residual connection. This process can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 260, + 553, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 260, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 324, + 260, + 553, + 280 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {X} _ {m i d} = \\mathbf {X} _ {i n} + \\operatorname {S A} (\\text {L a y e r N o r m} (\\mathbf {X} _ {\\text {i n}})), \\\\ \\mathbf {X} _ {i n} = \\mathbf {X} _ {i n - 1} + \\operatorname {V A M E} (\\mathbf {I} _ {i n - 1}, \\mathbf {N} _ {i n - 1} (\\mathbf {X} _ {i n})) \\end{array} \\tag {1}", + "image_path": "a8a35b32384539ef5136c17e671fbe86066f4b4c9804482719693a9073215577.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 327, + 277, + 523, + 288 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 277, + 523, + 288 + ], + "spans": [ + { + "bbox": [ + 327, + 277, + 523, + 288 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\text {o u t}} = \\mathbf {X} _ {\\text {m i d}} + \\operatorname {V A - M o E} (\\operatorname {L a y e r N o r m} (\\mathbf {X} _ {\\text {m i d}})),", + "image_path": "3f61904f915e6a31b6845ab57a1b67adf4e1d606c04b479fc2c46f85bf46cf5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 297, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 297, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{in}" + }, + { + "bbox": [ + 313, + 297, + 555, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 297, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{out}" + }, + { + "bbox": [ + 313, + 297, + 555, + 322 + ], + "type": "text", + "content": " are the input and the output of each Transformer Block, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 330, + 480, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 480, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 480, + 342 + ], + "type": "text", + "content": "3.2.1. Variables and Index Embedding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "text", + "content": "As illustrated in Fig. 3, during the training stage, the input feature " + }, + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_h^t\\in \\mathbb{R}^{H\\times W\\times N}" + }, + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "text", + "content": " consists of five different types of variables: " + }, + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_Z^t,\\mathbf{X}_Q^t,\\mathbf{X}_U^t,\\mathbf{X}_V^t,\\mathbf{X}_T^t\\in \\mathbb{R}^{H\\times W\\times (N / 5)}" + }, + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "text", + "content": ", where, " + }, + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 313, + 345, + 555, + 416 + ], + "type": "text", + "content": " denotes the spatial resolution. These five types of variables are concatenated and processed by a shared expert and five distinct Channel-Adaptive Experts (CAEs)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "text", + "content": "In addition to these variables, the framework introduces an extra one-hot index embedding " + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_h\\in \\mathbb{R}^{5\\times N}" + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "text", + "content": " to guide experts in learning variable affinity. The number " + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "text", + "content": " corresponds to the number of variables, and 5 corresponds to the number of variable types. Similar to the input weather features, the index embedding is composed of five separate embeddings: " + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_Z,\\mathbf{I}_Q,\\mathbf{I}_U,\\mathbf{I}_V,\\mathbf{I}_T\\in \\mathbb{R}^{1\\times 1\\times N}" + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "text", + "content": ". This initial index embedding is then encoded into a latent space using a Linear layer, expressed as " + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_h = \\mathrm{Linear}(\\mathbf{I}_h)" + }, + { + "bbox": [ + 313, + 418, + 554, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 532, + 484, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 484, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 484, + 544 + ], + "type": "text", + "content": "3.2.2. Channel-Adaptive Expert (CAE)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "text", + "content": "The key design of VA-MoE is the channel-adaptive expert, as shown in Fig. 3. For each variable " + }, + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "text", + "content": ", the corresponding expert module processes both the feature embeddings of the input variables " + }, + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_h^t" + }, + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "text", + "content": " and the index embedding " + }, + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_Z" + }, + { + "bbox": [ + 313, + 548, + 554, + 606 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 616, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 616, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 376, + 616, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right), \\tag {2}", + "image_path": "ace8f9de15f4cfa0e9083a4971fc2f764327d65785b3d525d42e6e9d9fdb75fa.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 553, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{CAE}_Z" + }, + { + "bbox": [ + 313, + 641, + 553, + 664 + ], + "type": "text", + "content": " is the abbreviation of channel-adaptive expert module for variable " + }, + { + "bbox": [ + 313, + 641, + 553, + 664 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 641, + 553, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "In the CAE module for variable " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": ", the index embedding " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_Z" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " is encoded through a linear layer for feature projection and then multiplied channel-wise with the input embedding " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_h^t" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": ". The resulting fused feature is processed" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7918" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "by the GateEmbed layer, which produces a gate embedding to enhance the fusion of the index and variable embeddings. This enhanced feature is then normalized using a SoftMax layer and filtered through a TopK operation, which selects the top-K high-rank channels. This process yields the GateIndex " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{GI}_Z\\in \\mathbb{R}^{H\\times W\\times K}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and GateWeight " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{GW}_Z\\in \\mathbb{R}^{H\\times W\\times K}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". The process is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 165, + 296, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 165, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 165, + 296, + 182 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {Z} ^ {t o p k}, \\mathbf {W} _ {Z} ^ {t o p k} = \\mathrm {T O P} _ {k} (\\operatorname {S o f t M a x} (\\operatorname {M L P} _ {Z} (\\mathbf {X} _ {h} ^ {t} \\odot \\mathbf {I} _ {Z}))), (3)", + "image_path": "be85a1bc34eedd8b21c0ba216e41c42c5c0e852b7399d8663f17a8379d269166.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_Z^{topk}" + }, + { + "bbox": [ + 55, + 190, + 296, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 190, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Z^{topk}" + }, + { + "bbox": [ + 55, + 190, + 296, + 323 + ], + "type": "text", + "content": " denote the index and weight of the Top-k features. Once the GateIndex and GateWeight are computed, the input embedding is selectively processed based on the GateIndex, while the variable embedding is weighted channel-wise according to the GateWeight. The GateEmbed layer uses variable embeddings to guide the index embedding, enabling the creation of a variable-specific confidence matrix. Within the CAE module, this embedding enhances the confidence matrix by incorporating both semantic and positional information, rather than relying solely on positional information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 324, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 324, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 324, + 288, + 335 + ], + "type": "text", + "content": "The selected features for variable " + }, + { + "bbox": [ + 67, + 324, + 288, + 335 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 67, + 324, + 288, + 335 + ], + "type": "text", + "content": " are obtained with:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 345, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 345, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 86, + 345, + 295, + 361 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {Z} ^ {t, s e l e c t e d} = \\mathbf {W} _ {Z} ^ {t o p k} * \\operatorname {S e l e c t} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z} ^ {t o p k}\\right), \\tag {4}", + "image_path": "58a303b5e91cf34827f681c5b1f066001fa7da63360b4b7d71de4656c8f37eb6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "inline_equation", + "content": "\\mathrm{Select}_Z" + }, + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "text", + "content": " denotes selecting Top-K embedding from the input feature " + }, + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_h^t" + }, + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "text", + "content": " within the " + }, + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "inline_equation", + "content": "\\mathrm{CAE}_Z" + }, + { + "bbox": [ + 55, + 370, + 295, + 394 + ], + "type": "text", + "content": " module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "text", + "content": "The selected and weighted embedding captures the key features associated with the specific type of variables. In the " + }, + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "inline_equation", + "content": "\\mathrm{CAE}_Z" + }, + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "text", + "content": " module, these features serve as the training set to model the distribution of the variable " + }, + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "inline_equation", + "content": "\\mathrm{Expert}_Z" + }, + { + "bbox": [ + 55, + 395, + 296, + 478 + ], + "type": "text", + "content": ". The expert module adopts a parallel structure, splitting the channels into multiple parts and enhancing each part with a sparse MoE. This process is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 487, + 295, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 487, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 107, + 487, + 295, + 504 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {Z} ^ {t, C A E} = \\operatorname {E x p e r t} _ {Z} (\\mathbf {X} _ {Z} ^ {t, s e l e c t e d}), \\tag {5}", + "image_path": "bcf2cc23d39d9f4f8bc9c3ce482925a79a620b33aa7186b3b3e574502368c0e5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 513, + 277, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 277, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 277, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 513, + 277, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{Expert}_Z" + }, + { + "bbox": [ + 55, + 513, + 277, + 525 + ], + "type": "text", + "content": " is implemented with a multiple layer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 149, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 149, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 149, + 544 + ], + "type": "text", + "content": "3.2.3. Shared Expert." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 548, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 668 + ], + "type": "text", + "content": "The shared expert processes the overall features of all variables and operates in parallel with the variable-specific modules. Each variable is handled by its corresponding CAE module, and their outputs are summarized to produce the fused features " + }, + { + "bbox": [ + 55, + 548, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_h^{t,fused}" + }, + { + "bbox": [ + 55, + 548, + 296, + 668 + ], + "type": "text", + "content": " for all variables. Since the channel count of the selected variable embeddings is lower than that of the original embeddings, an additional up-channel linear layer, " + }, + { + "bbox": [ + 55, + 548, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{Linear}_{\\mathrm{up}}" + }, + { + "bbox": [ + 55, + 548, + 296, + 668 + ], + "type": "text", + "content": ", is introduced. The fused features and the output from the shared expert are then combined via pixel-wise addition. This process is expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 676, + 295, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 676, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 62, + 676, + 295, + 692 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {h} ^ {t, f u s e d} = \\operatorname {C A E} _ {Z} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {Z}\\right) + \\dots + \\operatorname {C A E} _ {T} \\left(\\mathbf {X} _ {h} ^ {t}, \\mathbf {I} _ {T}\\right), \\tag {6}", + "image_path": "c809bb864893ff0c565c16ce0ed07cdeea5003f751fd75b79a4d57670ac41bad.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 699, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 699, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 66, + 699, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {X} ^ {\\prime}\\right) _ {h} ^ {t} = \\operatorname {E x p e r t} _ {s h a r e d} \\left(\\mathbf {X} _ {h} ^ {t}\\right) + \\operatorname {L i n e a r} _ {u p} \\left(\\mathbf {X} _ {h} ^ {t, f u s e d}\\right). \\tag {7}", + "image_path": "b032f93bffea71f199f3a81a0f819a10009efc9a5799f6cfc3c8b2fc8f8b2c1a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 470, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 470, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 470, + 85 + ], + "type": "text", + "content": "3.3. Incremental Learning Stage" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 161 + ], + "type": "text", + "content": "When new variables are incrementally introduced to the model, specialized experts are dynamically integrated into the transformer blocks to process these variables. Concurrently, both the encoder and decoder architectures undergo synchronized updates to maintain compatibility with the expanded variable set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 167, + 477, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 477, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 477, + 178 + ], + "type": "text", + "content": "3.3.1. Encoder and Decoder Modules." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "content": " new variables are introduced incrementally, the encoder's input dimension expands to " + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "inline_equation", + "content": "H \\times W \\times (N + M)" + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "content": ", necessitating an adjustment to the convolutional layer's kernel from " + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "inline_equation", + "content": "(3, 3, N, C)" + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "inline_equation", + "content": "(3, 3, N + M, C)" + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "content": ". To accommodate this, the layer's parameters are partitioned into two components: The " + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "inline_equation", + "content": "(3, 3, N, C)" + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "content": " segment retains pretrained weights from earlier stages, ensuring continuity. The " + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "inline_equation", + "content": "(3, 3, M, C)" + }, + { + "bbox": [ + 313, + 182, + 555, + 290 + ], + "type": "text", + "content": " segment, corresponding to the new variables, is initialized randomly to learn emergent patterns." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 291, + 554, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 554, + 314 + ], + "type": "text", + "content": "The same strategy is applied to the position embedding and the Decoder module." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 321, + 417, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 321, + 417, + 332 + ], + "spans": [ + { + "bbox": [ + 314, + 321, + 417, + 332 + ], + "type": "text", + "content": "3.3.2. VA-MoE Module." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "text", + "content": "During the incremental learning phase, " + }, + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "text", + "content": " new experts are dynamically integrated into the MoE layer of each transformer block to accommodate " + }, + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "text", + "content": " newly introduced variables. In this phase, only the new experts and the shared expert are actively trained, while all preexisting experts remain frozen to preserve previously learned knowledge." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 415, + 456, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 456, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 456, + 426 + ], + "type": "text", + "content": "3.3.3. Index Embedding Module." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "content": " new variables become available during the incremental stage, the index embedding changes from " + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_h\\in" + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{(N\\times l)\\times N}" + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_h\\in \\mathbb{R}^{(N\\times l + M\\times r)\\times (N + M)}" + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "content": " new variables has " + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 430, + 555, + 489 + ], + "type": "text", + "content": " levels. Additionally, the index encoder needs to be retrained." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 498, + 426, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 498, + 426, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 498, + 426, + 510 + ], + "type": "text", + "content": "3.4. Objective Function" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 515, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 554, + 539 + ], + "type": "text", + "content": "To train the model, we introduce a novel dynamic prediction loss combined with a reconstruction loss." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 546, + 451, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 451, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 451, + 558 + ], + "type": "text", + "content": "3.4.1. Dynamic Prediction Loss." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 561, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 554, + 657 + ], + "type": "text", + "content": "The prediction loss aims to minimize the difference between the model's predicted results and the actual future weather variables. Unlike existing weather forecasting methods that treat all the variables equally, we argue that different variables, such as temperature and surface pressure, follow distinct distributions. Thus, we introduce a novel dynamic prediction loss that dynamically assigns a weight to each channel of the input data. This loss is formulated as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 665, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 665, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 318, + 665, + 555, + 681 + ], + "type": "interline_equation", + "content": "O b j _ {p r e d} = \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t + 1} - \\mathbf {X} ^ {t + 1}\\right) / e ^ {\\mathbf {w}} + \\mathbf {w}, (8)", + "image_path": "2d90e240e89098f1925b378161a82a3afbb97f99c9015de745901edbf115ae87.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " denotes Hadamard product, " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " denotes the base of the natural logarithm, " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t + 1}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}\\in" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7919" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 279, + 248 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 279, + 248 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 279, + 248 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 279, + 248 + ], + "type": "table", + "html": "
NameDescriptionLevels
Upper-Air
ZGeopotential13
QSpecific humidity13
Ux-direction wind13
Vy-direction wind13
TTemperature13
Surface-Incremental Learning
u10x-direction wind at 10m heightSingle
v10y-direction wind at 10m heightSingle
t2mTemperature at 2m heightSingle
mslMean sea-level pressureSingle
spSurface pressureSingle
", + "image_path": "041e988ddd89d40a80ac5822062df3d6dd06335024de3d68c9678d208845afb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 296, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times 1\\times C}" + }, + { + "bbox": [ + 55, + 296, + 296, + 331 + ], + "type": "text", + "content": " is a learnable vector. With this loss function, the model could dynamically prioritize variables based on their distributional characteristics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 173, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 173, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 173, + 350 + ], + "type": "text", + "content": "3.4.2. Reconstruction Loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 378 + ], + "type": "text", + "content": "We introduce an additional reconstruction loss that connects the encoder and decoder directly, formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 386, + 295, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 386, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 96, + 386, + 295, + 401 + ], + "type": "interline_equation", + "content": "O b j _ {r e c o n} = \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right) \\odot \\left(\\hat {\\mathbf {X}} ^ {t} - \\mathbf {X} ^ {t}\\right), \\tag {9}", + "image_path": "bb83433fe20cccaa34ddc4d36541be7ba859ff5d21fb9bf5e979967bf7db6da6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^t = \\mathrm{Dec}(\\mathrm{Enc}(\\mathbf{X}^t))" + }, + { + "bbox": [ + 55, + 409, + 296, + 506 + ], + "type": "text", + "content": " represents the reconstructed variables. With the loss function, the encoder and decoder modules specialize in encoding and decoding features, while the intermediate transformer blocks focus on learning the data distribution. This ensures that the decoding process is entirely handled by decoder, allowing the intermediate blocks to focus on their task and not participate in decoding, thereby optimizing network efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 542 + ], + "type": "text", + "content": "To this end, the model is trained with the final objective function, a combination of the dynamic prediction loss and reconstruction loss. The final objective is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 552, + 295, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 552, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 102, + 552, + 295, + 566 + ], + "type": "interline_equation", + "content": "O b j _ {f i n a l} = O b j _ {p r e d} + \\lambda O b j _ {r e c o n}, \\tag {10}", + "image_path": "22d35777f8e10615c054df1320cb82a07267a1cbfcc0f2beccd4b2935f3f10fc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 574, + 201, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 201, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 201, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 574, + 201, + 586 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 574, + 201, + 586 + ], + "type": "text", + "content": " denotes a hyper-parameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 597, + 137, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 137, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 137, + 611 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Dataset. In this work, we conduct experiments on a popular weather dataset, i.e., ERA5[18], provided by the ECMWF. ERA5 dataset is a reanalysis atmospheric dataset, consisting of the atmospheric variables from 1979 to the present day with a " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.25^{\\circ}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " spatial resolution with " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "721 \\times 1440" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": ". We train the model on 40-year dataset in the initial stage and continuously train the model on 20-year dataset in the incremental stage, which contains the weather variables from" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 432, + 173 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 432, + 173 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 432, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 432, + 173 + ], + "type": "image", + "image_path": "5c32d76ef3271d634424358f221973a4c9a674afb759538b61aacedf3e343fc8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 437, + 72, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 437, + 72, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 437, + 72, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 437, + 72, + 553, + 173 + ], + "type": "image", + "image_path": "a0d1aa500d574a7e95efd8c0c9b28ef1e85a3a448b254b80980622a66a0bc2a2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 174, + 432, + 273 + ], + "blocks": [ + { + "bbox": [ + 317, + 174, + 432, + 273 + ], + "lines": [ + { + "bbox": [ + 317, + 174, + 432, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 432, + 273 + ], + "type": "image", + "image_path": "052ad5c9ca6a4a26d4335087addf28eac50c9b1211888b4204be97c3469dbdba.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 309, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 354 + ], + "type": "text", + "content": "Figure 4. Comparative analysis of RMSE " + }, + { + "bbox": [ + 313, + 309, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 309, + 555, + 354 + ], + "type": "text", + "content": " across 10 data-driven models for four variables, including Z500 and T850 (upper-air variables) in the initial stage, as well as T2M and U10 (incremental surface variables) in the incremental stage." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 437, + 174, + 553, + 273 + ], + "blocks": [ + { + "bbox": [ + 437, + 174, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 437, + 174, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 437, + 174, + 553, + 273 + ], + "type": "image", + "image_path": "bd3379cdf299482f4eddfea62f542bb052f9854ee8afec711d077e1cd24bf117.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 331, + 274, + 543, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 256, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 289 + ], + "type": "text", + "content": "Table 1. A summary of atmospheric variables. The 13 levels are 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000 hPa. 'Single' denotes the variables under earth's surface." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 274, + 543, + 297 + ], + "lines": [ + { + "bbox": [ + 331, + 274, + 543, + 297 + ], + "spans": [ + { + "bbox": [ + 331, + 274, + 543, + 297 + ], + "type": "table", + "html": "
——ClimaX——EWMoE——WeatherGFT——Stormer
——FourCastNet——ClimODE——GraphCast——Ours(IL)
——Panguo-Weather——Keisler
", + "image_path": "5c27927caadf954f2ab0a6c58768eeb365f2cd7f4cb19f217f1e1d4f1c9d5cd4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 433 + ], + "type": "text", + "content": "1979 to 2020 and 2000 to 2020 year, respectively. And we test the model on one-year dataset of weather variables in 2021. In this work, the model processes 5 upper-air variables and 5 surface variables as Tab. 1, where the upper-air variables are exploited in the training stage, and the surface variables are exploited in the incremental stage." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "type": "text", + "content": "Implementation Details. The main structure of this work follows the novel backbones [7, 8, 12]. We apply the AdamW optimizer with 0.0002 and 0.00005 learning rates to initial and incremental stages, respectively. In two stages, we train 100 epochs and set batch size to 16. Our model is trained with PyTorch using 16 A100 GPUs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 515, + 399, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 399, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 399, + 526 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "Fig. 4 presents a quantitative comparison of our structure with existing state-of-the-art approaches, including ClimaX [27], Pangu-Weather [3], ClimODE [36], WeatherGFT [39], FourCastNet [22], EWMoE [14], Keisler [20], GraphCast [23], and Stormer [28], for the 5-day weather prediction task. All the competitors are trained with both upper-air and surface variables on a " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "0.25^{\\circ}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " weather dataset with a resolution of " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "721 \\times 1440" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "For the surface variables, including T2M and U10, our VA-MoE demonstrates performance similar to the leading methods Stormer [28] and Graphcast [23]. Compared to other competitors, our structure shows significant advantages in both short-term and long-term predictions. For the upper-air variable Z500, our structure delivers one of the best performances among all 10 methods, outperforming" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7920" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 555, + 274 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 274 + ], + "type": "table", + "html": "
Dataset(years)Iteration\\( (×10^4) \\)T2M(K)↓U10(m/s)↓V10(m/s)↓MSL(Pa)↓SP(Pa)↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
Plain Training
ViT* [13]1979-2020400.721.351.860.661.983.010.682.023.1140.2208.5393.963.3222.1397.0
IFS [32]1979-2020401.091.381.740.961.872.780.991.932.87------
Pangu-Weather [3]1979-2020400.821.091.530.771.632.540.791.682.65------
FourCastNet [22]1979-2020400.821.021.770.822.083.340.842.113.41------
ClimaX [27]1979-2020401.111.471.831.042.022.79---------
Graphcast [23]1979-2020400.510.941.370.381.512.37---23.4135.2278.2---
Fengwu [9]1979-2020400.581.031.410.421.532.32---23.2137.1276.9---
FuXi [10]1979-2020400.550.991.410.421.502.360.431.542.4427.2136.7282.9---
VA-MoE1979-2020400.571.031.420.431.412.250.441.462.3427.5131.1275.957.1168.9302.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (79-20)
VA-MoE (IL)1979-2020200.581.051.450.481.472.330.471.542.4127.9137.3281.659.3173.4312.4
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)2000-2020100.731.171.570.541.582.490.551.632.5730.0148.8304.760.6171.4314.8
", + "image_path": "e1e47cf80a602d88639422c583130b60dbcfc1b6d6cbe0dc96828b9436f47f10.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 315, + 555, + 422 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": "Table 2. Prediction performances on Incremental Training with 5 surface variables, i.e., T2M, U10, V10, MSL, and SP. * denotes running by ourselves. The best results are marked in bold. All experiments are in " + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "inline_equation", + "content": "{0.25}^{ \\circ }" + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "inline_equation", + "content": "{721} \\times {1440}" + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": " resolutions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 315, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 59, + 315, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 59, + 315, + 555, + 422 + ], + "type": "table", + "html": "
Para. (M)Z500(m2/s2) ↓Q500(×e-3, g/kg) ↓U500(m/s) ↓V500(m/s) ↓T500(K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
IFS [32]-28.31154.08333.960.310.610.751.433.235.121.403.585.640.360.981.70
Pangu-Weather [3]-24.88167.90391.260.250.550.690.963.134.730.913.525.150.270.941.56
Graphcast [23]-15.23125.42275.35---0.772.864.490.742.924.670.230.871.48
VA-MoE66519.28134.63295.520.170.490.620.842.994.710.843.044.890.250.761.36
Incremental Training from 65 Upper-Air Variables (79-20) to 5 Surface Variables (00-20)
VA-MoE (IL)13718.23133.14292.630.170.490.610.843.014.740.843.514.930.250.761.37
", + "image_path": "6c48125c0977a9f970ee7a1f1510428c67d7c7c44508d3eb463f98acbe402368.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 430, + 553, + 441 + ], + "lines": [ + { + "bbox": [ + 56, + 430, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 553, + 441 + ], + "type": "text", + "content": "Table 3. Prediction performances on Initial Training with 5 upper-air variables. All experiments are in " + }, + { + "bbox": [ + 56, + 430, + 553, + 441 + ], + "type": "inline_equation", + "content": "{0.25}^{ \\circ }" + }, + { + "bbox": [ + 56, + 430, + 553, + 441 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 56, + 430, + 553, + 441 + ], + "type": "inline_equation", + "content": "{721} \\times {1440}" + }, + { + "bbox": [ + 56, + 430, + 553, + 441 + ], + "type": "text", + "content": " resolutions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 462, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 295, + 544 + ], + "type": "text", + "content": "Pangu-Weather [3] and Stormer [28], and significantly surpassing ClimaX [27] and FourCastNet [22], while slightly inferior to GraphCast [23]. For T850, our structure is slightly inferior to EWMoE [14], GraphCast [23], and Stormer [28] in the short-term predictions within 48 hours. After that period, our structure achieves the best performance in the long-term predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 545, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 628 + ], + "type": "text", + "content": "Results of Surface Variables Prediction. This study evaluates model performance on five surface variables: T2M, U10, V10, MSL, and SP. While competing structures are trained with both upper-air and surface variables, VA-MoE (IL) employs a two-stage training approach: initial training with upper-air variables followed by incremental training with all five surface variables." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": "Analysis of the results from Tab. 2 demonstrates that VA-MoE outperforms competing models across multiple variables and settings, particularly in U10 and V10. Although VA-MoE shows marginally weaker performance in T2M compared to GraphCast [23] and comparable results to FengWu [9] and FuXi [10], it achieves superior long-term predictions for U10, V10, MSL, and SP. These indicate" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 462, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 486 + ], + "type": "text", + "content": "that while VA-MoE exhibits slight limitations in short-term forecasting, it excels in long-term prediction tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 487, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 487, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 487, + 554, + 594 + ], + "type": "text", + "content": "When implementing the incremental training strategy, VA-MoE (IL) trained on a 40-year dataset with only half the training iterations achieves performance comparable to standard VA-MoE. Even with a reduced 20-year dataset, VA-MoE(IL) maintains acceptable performance using just a quarter of the iterations required by the standard approach. These results underscore the efficacy of incremental learning for atmospheric datasets and suggest its potential to replace conventional training methods in operational settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 594, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 554, + 677 + ], + "type": "text", + "content": "Results of Upper-Air Variables Prediction. This study evaluates model performance on upper-air variables, comparing three training approaches: (1) VA-MoE, trained exclusively on upper-air variables; (2) VA-MoE (IL), which incorporates incremental training on surface variables after initial upper-air training; and (3) competing models trained simultaneously on both upper-air and surface variables." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "As shown in Tab. 3, VA-MoE achieves performance comparable to GraphCast [23] and outperforms IFS [32] and Pangu-Weather [3] across both short- and long-term" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7921" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 71, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 156 + ], + "type": "table", + "html": "
Para. (M)Z500 (m2/s2) ↓Q500 (×e-3, g/kg) ↓U500 (m/s) ↓V500 (m/s) ↓T500 (K) ↓
6h72h120h6h72h120h6h72h120h6h72h120h6h72h120h
ViT*[13]30733.38209.4517.810.220.611.061.243.666.521.223.767.410.421.182.40
ViT+MoE (light)*[30]60937.92207.11405.730.220.600.781.303.845.871.273.896.110.461.232.02
ViT+MoE* [30]111328.31169.61356.020.230.560.721.213.465.441.233.545.690.351.071.83
VA-MoE66520.59139.02302.130.180.490.620.913.024.760.913.084.970.270.921.59
VA-MoE (IL)13720.29138.52301.410.180.500.630.913.044.790.913.105.030.270.931.60
", + "image_path": "a89841f541b62779568827b555ddd995ada9597a86b689891af324f299c2b25e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "lines": [ + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "type": "text", + "content": "Table 4. Architectural impact on 5 upper-air variables under " + }, + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "type": "inline_equation", + "content": "{500}\\mathrm{{hPa}}" + }, + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "type": "text", + "content": " . All experiments are in " + }, + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "type": "inline_equation", + "content": "{1.5}^{ \\circ }" + }, + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "type": "inline_equation", + "content": "{128} \\times {256}" + }, + { + "bbox": [ + 78, + 163, + 530, + 175 + ], + "type": "text", + "content": " resolutions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 177, + 536, + 361 + ], + "blocks": [ + { + "bbox": [ + 70, + 177, + 536, + 361 + ], + "lines": [ + { + "bbox": [ + 70, + 177, + 536, + 361 + ], + "spans": [ + { + "bbox": [ + 70, + 177, + 536, + 361 + ], + "type": "image", + "image_path": "7a3f0e25dc8da0ded62ed82c5f1ef56cb67792b1eb0e64e3699a3a53e9edaebc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 367, + 545, + 378 + ], + "lines": [ + { + "bbox": [ + 62, + 367, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 545, + 378 + ], + "type": "text", + "content": "Figure 5. 6-hour global weather prediction of upper-air and surface variables forecast visualization generated by VA-MoE framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "text", + "content": "prediction. Notably, VA-MoE (IL) demonstrates robust predictive capabilities for upper-air variables even after incremental training on surface variables, despite using only a 20-year dataset - half the size of the initial training data. Furthermore, VA-MoE (IL) exhibits marginal performance improvements over the baseline in long-term predictions of Z500, confirming the absence of catastrophic forgetting during incremental training. These results highlight the efficacy of incremental training in VA-MoE(IL) for maintaining and enhancing predictive accuracy on original variables." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 149, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 149, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 149, + 528 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": "Impacts of Architecture. As shown in Tab. 4, we compare VA-MoE with other architectures, including ViT and ViT+MoE. VA-MoE significantly outperforms both ViT and ViT+MoE architectures across 6-hour, 72-hour, and 120-hour prediction horizons, despite ViT+MoE containing nearly twice the number of parameters. This performance gap highlights VA-MoE's suitability for weather forecasting tasks as its channel-adaptive expert design ensures parameter efficiency while maintaining accuracy, particularly in incremental learning. Considering the computational complexity, all ablation studies are in " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "1.5^{\\circ}" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "128 \\times 256" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 139, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 139, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 139, + 683 + ], + "type": "text", + "content": "4.3. Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In the work, we visualize some predicted results and parameters. In Fig. 5, we visualize the 6-hour predicted results of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "content": "Z500, T850, U10, and T2M. There are max " + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "inline_equation", + "content": "0.08\\%" + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "inline_equation", + "content": "0.22\\%" + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "inline_equation", + "content": "1.59\\%" + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "inline_equation", + "content": "0.47\\%" + }, + { + "bbox": [ + 313, + 388, + 554, + 424 + ], + "type": "text", + "content": " absolute errors across 4 variables, including 2 upper-air variables and 2 surface variables." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 435, + 388, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 388, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 388, + 447 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 455, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 587 + ], + "type": "text", + "content": "In this work, we proposed incremental weather forecasting, a novel task that addresses the challenge of dynamically expanding weather models to incorporate new variables without retraining from scratch. Our solution, Variables-Adaptive Mixture-of-Experts, enables parameter-efficient adaptation by selectively fine-tuning experts assigned to new variables through index embedding guidance, while preserving pretrained knowledge in existing parameters. Experimental results demonstrate that our approach achieves forecast accuracy comparable to fully retrained models while drastically reducing computational costs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 597, + 429, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 429, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 429, + 611 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "This research was supported by the Hong Kong RGC General Research Fund (152169/22E, 152228/23E, 162161/24E), Research Impact Fund (No. R5060-19, No. R5011-23), Collaborative Research Fund (No. C1042-23GF), NSFC/RGC Collaborative Research Scheme (No. 62461160332 & CRS_HKUST602/24), Areas of Excellence Scheme (AoE/E-601/22-R), the InnoHK (HKGAI), and Key R&D Program of Zhejiang Province 2025C01075." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7922" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "text", + "content": "[2] Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567): 47-55, 2015. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 296, + 214 + ], + "type": "text", + "content": "[3] Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Accurate medium-range global weather forecasting with 3d neural networks. Nature, 619(7970): 533-538, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 296, + 270 + ], + "type": "text", + "content": "[4] Cristian Bodnar, Wessel P Bruinsma, Ana Lucic, Megan Stanley, Johannes Brandstetter, Patrick Garvan, Maik Riechert, Jonathan Weyn, Haiyu Dong, Anna Vaughan, et al. Aurora: A foundation model of the atmosphere. arXiv:2405.13063, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 296, + 327 + ], + "type": "text", + "content": "[5] Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, and Anima Anandkumar. Spherical fourier neural operators: learning stable dynamics on the sphere. In Proceedings of International Conference on Machine Learning, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 373 + ], + "type": "text", + "content": "[6] Hao Chen, Yonghan Dong, Zhe-Ming Lu, Yunlong Yu, and Jungong Han. Self-prompting perceptual edge learning for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4528-4541, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 296, + 428 + ], + "type": "text", + "content": "[7] Hao Chen, Yonghan Dong, Zheming Lu, Yunlong Yu, and Jungong Han. Pixel matching network for cross-domain few-shot segmentation. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 978-987, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 296, + 485 + ], + "type": "text", + "content": "[8] Hao Chen, Yunlong Yu, Yonghan Dong, Zheming Lu, Yingming Li, and Zhongfei Zhang. Multi-content interaction network for few-shot segmentation. ACM Transactions on Multimedia Computing, Communications and Applications, 20 (6):1-20, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 296, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 541 + ], + "type": "text", + "content": "[9] Kang Chen, Tao Han, Junchao Gong, Lei Bai, Fenghua Ling, Jing-Jia Luo, Xi Chen, Leiming Ma, Tianning Zhang, Rui Su, et al. Fengwu: Pushing the skillful global medium-range weather forecast beyond 10 days lead. arXiv:2304.02948, 2023. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 296, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 296, + 598 + ], + "type": "text", + "content": "[10] Lei Chen, Xiaohui Zhong, Feng Zhang, Yuan Cheng, Yinghui Xu, Yuan Qi, and Hao Li. Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast. npj Climate and Atmospheric Science, 6(1):190, 2023. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 600, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 296, + 654 + ], + "type": "text", + "content": "[11] Wuyang Chen, Yanqi Zhou, Nan Du, Yanping Huang, James Laudon, Zhifeng Chen, and Claire Cui. Lifelong language pretraining with distribution-specialized experts. In International Conference on Machine Learning, pages 5383-5395. PMLR, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 296, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 296, + 700 + ], + "type": "text", + "content": "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Processing Systems, pages 16344-16359, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Trans-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "formers for image recognition at scale. arXiv:2010.11929, 2020.2,7,8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[14] Lihao Gan, Xin Man, Chenghong Zhang, and Jie Shao. Ew-moe: An effective model for global weather forecasting with mixture-of-experts. arXiv:2405.06004, 2024. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[15] Yuan Gao, Hao Wu, Ruiqi Shu, Huanshuo Dong, Fan Xu, Rui Chen, Yibo Yan, Qingsong Wen, Xuming Hu, Kun Wang, et al. Oneforecast: A universal framework for global and regional weather forecasting. arXiv:2502.00338, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "text", + "content": "[16] Dipam Goswami, Albin Soutif-Cormerais, Yuyang Liu, Sandesh Kamath, Bart Twardowski, Joost van de Weijer, et al. Resurrecting old classes with new data for exemplar-free continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 28525-28534, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "text", + "content": "[18] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquin Muñoz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al. The era5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730):1999-2049, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[19] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. Neural computation, 3(1):79-87, 1991. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "type": "text", + "content": "[20] Ryan Keisler. Forecasting global weather with graph neural networks. arXiv:2202.07575, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 412, + 553, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 467 + ], + "type": "text", + "content": "[21] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3346-3356, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 468, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 544 + ], + "type": "text", + "content": "[22] Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. In Proceedings of the Platform for Advanced Scientific Computing Conference, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 546, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 600 + ], + "type": "text", + "content": "[23] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "text", + "content": "[24] Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv:2010.08895, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7923" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[26] Franco Molteni, Roberto Buizza, Tim N Palmer, and Thomas Petroliagis. The ecmwf ensemble prediction system: Methodology and validation. Quarterly journal of the royal meteorological society, 122(529):73-119, 1996. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[27] Tung Nguyen, Johannes Brandstetter, Ashish Kapoor, Jayesh K. Gupta, and Aditya Grover. Climax: A foundation model for weather and climate. In International Conference on Machine Learning, 2023. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "type": "text", + "content": "[28] Tung Nguyen, Rohan Shah, Hritik Bansal, Troy Arcomano, Romit Maulik, Rao Kotamarthi, Ian Foster, Sandeep Madireddy, and Aditya Grover. Scaling transformer neural networks for skillful and reliable medium-range weather forecasting. In Advances in Neural Information Processing Systems, pages 68740-68771, 2025. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[29] Youqi Pan, Wugen Zhou, Yingdian Cao, and Hongbin Zha. Adaptive vio: Deep visual-inertial odometry with online continual learning. arXiv:2405.16754, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[30] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. In European Conference on Computer Vision, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[31] Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Probabilistic weather forecasting with machine learning. Nature, 637(8044):84-90, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 408 + ], + "type": "text", + "content": "[32] D Richardson. The thorpex interactive grand global ensemble (tigge). In Geophysical Research Abstracts, page 02815, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[33] Harold Ritchie, Clive Temperton, Adrian Simmons, Mariano Hortal, Terry Davies, David Dent, and Mats Hamrud. Implementation of the semi-lagrangian method in a high-resolution version of the ecmwf forecast model. Monthly Weather Review, 123(2):489-514, 1995. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "type": "text", + "content": "[34] Minhyuk Seo, Hyunseo Koh, Wonje Jeung, Minjae Lee, San Kim, Hankook Lee, Sungjun Cho, Sungik Choi, Hyunwoo Kim, and Jonghyun Choi. Learning equi-angular representations for online continual learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23933-23942, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[35] Yuwen Tan, Qinhao Zhou, Xiang Xiang, Ke Wang, Yuchuan Wu, and Yongbin Li. Semantically-shifted incremental adapter-tuning is a continual vitransformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23252-23262, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[36] Yogesh Verma, Markus Heinonen, and Vikas Garg. ClimODE: Climate forecasting with physics-informed neural ODEs. In International Conference on Learning Representations, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[37] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14205-14215, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Wei Xiong, Muyuan Ma, Xiaomeng Huang, Ziyang Zhang, Pei Sun, and Yang Tian. Koopmanlab: machine learning for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 596 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "solving complex physics equations. APL Machine Learning, 1(3), 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[39] Wanghan Xu, Fenghua Ling, Wenlong Zhang, Tao Han, Hao Chen, Wanli Ouyang, and Lei Bai. Generalizing weather forecast to fine-grained temporal scales via physics-ai hybrid modeling. In Advances in Neural Information Processing Systems, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "text", + "content": "[40] Hongwei Yan, Liyuan Wang, Kaisheng Ma, and Yi Zhong. Orchestrate latent expertise: Advancing online continual learning with multi-level supervision and reverse self-distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 23670-23680, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "text", + "content": "[41] Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. Multi-task dense prediction via mixture of low-rank experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27927-27937, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 327 + ], + "type": "text", + "content": "[42] Fei Ye and Adrian G Bors. Online task-free continual generative and discriminative learning via dynamic cluster memory. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 26202-26212, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 330, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 330, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 330, + 553, + 374 + ], + "type": "text", + "content": "[43] Donggeun Yoon, Minseok Seo, Doyi Kim, Yeji Choi, and Donghyeon Cho. Probabilistic weather forecasting with deterministic guidance-based diffusion model. In European Conference on Computer Vision, pages 108-124, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 439 + ], + "type": "text", + "content": "[44] Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen. Diffcast: A unified framework via residual diffusion for precipitation nowcasting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 27758-27767, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 441, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 496 + ], + "type": "text", + "content": "[45] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23219-23230, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[46] Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. Skilful nowcasting of extreme precipitation with nowcastnet. Nature, 619(7970):526-532, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 543, + 553, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 596 + ], + "type": "text", + "content": "[47] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, Cheng-Lin Liu, and Zhaoxiang Zhang. Rcl: Reliable continual learning for unified failure detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12140-12150, 2024. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7924" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_content_list.json b/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c24f2cc7f057dea11e8ebc5958ba4431b47c6587 --- /dev/null +++ b/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "VACE: All-in-One Video Creation and Editing", + "text_level": 1, + "bbox": [ + 263, + 130, + 736, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeyinzi Jiang*", + "bbox": [ + 153, + 180, + 272, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhen Han*", + "bbox": [ + 285, + 181, + 372, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chaojie Mao*†", + "bbox": [ + 395, + 181, + 514, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingfeng Zhang", + "bbox": [ + 535, + 181, + 661, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yulin Pan", + "bbox": [ + 683, + 181, + 763, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Liu", + "bbox": [ + 784, + 181, + 841, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tongyi Lab, Alibaba Group", + "text_level": 1, + "bbox": [ + 387, + 208, + 609, + 227 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6dc0ac2e79e0ba9bb358ce5358c79ab788f0b963e087806e235c7b43592a9f61.jpg", + "image_caption": [ + "Reference-to-video generation" + ], + "image_footnote": [], + "bbox": [ + 96, + 250, + 486, + 295 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01187571914a5e6d6532ea027270a9ab6354e126bbeba8577fb665a9462afeee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 252, + 602, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85545d4d442ad80243ffb8b78dc40cee7e78444acc8b7f750789189b0c03bbf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 250, + 901, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c9c116f5690eba3d9ee44b3b791916ff76bb2ec58621dbbe86a0ae44980aa0d.jpg", + "image_caption": [ + "Video-to-video editing", + "Pose Transfer (Pose Control)" + ], + "image_footnote": [], + "bbox": [ + 96, + 310, + 486, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d10ef66bf9e895e32d99c420bde8e9d16ef77d13a2666119f69ac5d62c76b41.jpg", + "image_caption": [ + "Motion Transfer (Layout Control)" + ], + "image_footnote": [], + "bbox": [ + 509, + 310, + 901, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58ad5b44490705232a973cf7a586334b054d0bc8c494901a11fe885919a638d8.jpg", + "image_caption": [ + "Structure Transfer (Depth Control)" + ], + "image_footnote": [], + "bbox": [ + 96, + 369, + 486, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80df6d6119b73f2d4005bce8b9c89505b8abb30c091bc04ee8329de809340e56.jpg", + "image_caption": [ + "Colorization (Gray Control)" + ], + "image_footnote": [], + "bbox": [ + 511, + 369, + 901, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95d1d327420388d5737f00901b38c9c4081cca6ef3806a9ed2ec1ed8f43224ce.jpg", + "image_caption": [ + "Masked Video-to-video editing", + "Subject Recreation (Inpainting)" + ], + "image_footnote": [], + "bbox": [ + 96, + 431, + 486, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f9436c54eb21d2c2ced5d62e408773bb87a5fad5ad4a71c1a7bcecc0f3dff35.jpg", + "image_caption": [ + "Subject Removal (Inpainting)" + ], + "image_footnote": [], + "bbox": [ + 509, + 431, + 901, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e856a972006ce8b10c473c7deb34dd2b9911915774764a54da3b6adf70ffa9b.jpg", + "image_caption": [ + "Canvas Extension (Outpainting)" + ], + "image_footnote": [], + "bbox": [ + 96, + 491, + 486, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1285142f3ed6719d624eecacd52d7f643cbad1f1346811431897c717ae26fcc.jpg", + "image_caption": [ + "Temporal Extension (First Clip)" + ], + "image_footnote": [], + "bbox": [ + 524, + 491, + 901, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a8e8ec3fd8352db7fcaa2531b94053d11e401af6e453edc2f6ee0834d1ac198.jpg", + "image_caption": [ + "Task Composition", + "Swap Anything (Subject Reference + Inpainting)", + "Figure 1. Comprehensive capabilities of VACE. We present the outstanding generation results based on Wan2.1 (left) and LTX-Video (right). For each task, the original input image or video (left), the context video (top left corner), and the generated frames are illustrated." + ], + "image_footnote": [], + "bbox": [ + 96, + 551, + 486, + 597 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca5b623143e9929a16ef37e21b1cf7268db1c9a956e3c85c812cdbf5bb037f27.jpg", + "image_caption": [ + "Animate Anything (Frame Reference + Pose Control)" + ], + "image_footnote": [], + "bbox": [ + 511, + 555, + 901, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 666, + 326, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion Transformer has demonstrated powerful capability and scalability in generating high-quality images and videos. Further pursuing the unification of generation and editing tasks has yielded significant progress in the domain of image content creation. However, due to the intrinsic demands for consistency across both temporal and spatial dynamics, achieving a unified approach for video synthesis remains challenging. We introduce VACE, which enables users to perform Video tasks within an All-in-one framework for Creation and Editing. These tasks include reference-to-video generation, video-to-video editing, and masked video-to-video editing. Specifically, we effectively", + "bbox": [ + 88, + 696, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "integrate the requirements of various tasks by organizing video task inputs, such as editing, reference, and masking, into a unified interface referred to as the Video Condition Unit (VCU). Furthermore, by utilizing a Context Adapter structure, we inject different task concepts into the model using formalized representations of temporal and spatial dimensions, allowing it to handle arbitrary video synthesis tasks flexibly. Extensive experiments demonstrate that the unified model of VACE achieves performance on par with task-specific models across various subtasks. Simultaneously, it enables diverse applications through versatile task combinations. Project page: https://alivilab.github.io/VACE-Page/.", + "bbox": [ + 511, + 667, + 908, + 864 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. †Project lead.", + "bbox": [ + 107, + 886, + 297, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17191", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, the domain of visual generation tasks has witnessed remarkable advancements, driven in particular by the rapid evolution of diffusion models [24, 25, 48, 53, 54, 56, 57]. Beyond the early foundational pre-trained models for text-to-image [7, 16, 33] or text-to-video [9, 22, 64] generation in the field, there has been a proliferation of downstream tasks and applications, such as painting [3, 82], editing [4, 42, 68, 70, 75], controllable generation [30, 76], frame reference generation [20, 73], and ID-referenced video synthesis [11, 35, 47, 74]. This array of developments highlights the dynamic and complex nature of the visual generation field. To enhance task flexibility and reduce the overhead associated with deploying multiple models, researchers have begun to focus on constructing unified model architectures [12, 63] (e.g., ACE [23, 41] and Omni-Gen [71]), aiming to integrate different tasks into a single image model, facilitating the creation of various application workflows while maintaining simplicity in usage. In the field of video, due to the collaborative transformations in both temporal and spatial dimensions, leveraging a unified model can present endless possibilities for video creation. However, leveraging diverse input modalities and ensuring spatiotemporal consistency are still challenging for unified video generation and editing.", + "bbox": [ + 89, + 114, + 485, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose VACE, an all-in-one model for video creation and editing that performs tasks including reference-to-video generation, video-to-video editing, masked video-to-video editing, and free composition of these tasks, as illustrated in Fig. 1. The aggregation of various capabilities reduces the costs of service deployment and user interaction, and by combining the capabilities of different tasks within a single model, it addresses challenges faced by existing video generation models, such as controllable generation of long videos, multi-condition and reference-based generation, and continuous video editing, thereby empowering users with greater creativity. To achieve this, we utilize the current mainstream Diffusion Transformers (DiTs) structure as the foundational video framework and pre-trained text-to-video generation models [22, 64], which provide better basic capabilities and scalability for handling long video sequences. Specifically, VACE takes into account the needs of different tasks during its construction and designs a unified interface, dubbed the Video Condition Unit (VCU), which integrates multiple modalities such as images or videos for editing, references, and masks. Furthermore, to differentiate the visual modality information in editing and reference tasks, we introduce the concept decoupling strategy, enabling the model to understand what aspects need to be retained and what should be modified. Meanwhile, by employing a pluggable Context Adapter structure, concepts from different tasks (e.g., the areas or ranges of editing or reference) are injected into the model through col", + "bbox": [ + 89, + 478, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "laborative spatiotemporal representation, enabling it to possess the capability of adaptive processing for unified tasks.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the lack of existing multi-task benchmarks in video synthesis, we construct a dataset of 480 evaluation samples containing 12 different tasks, while evaluating the performance of the VACE unified model by comparing it with existing specialized models. Experimental results demonstrate that our framework exhibits sufficient competitiveness in both quantitative and qualitative analyses. To the best of our knowledge, we are the first all-in-one model based on the video DiT architecture that concurrently supports such a wide range of tasks. Notably, this innovative framework allows for the compositional expansion of base tasks, enabling the construction of scenarios such as long video re-rendering, which provides a versatile and efficient solution for video synthesis, opening new possibilities for user-side video content creation and editing.", + "bbox": [ + 511, + 122, + 906, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 361, + 655, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Generation and Editing. With the rapid development of image [2, 7, 16, 18, 58, 59] and video [22, 32, 73, 77] generation models, they are being used to create high-quality visual content and are widely applied in fields such as advertising, film special effects, game development, and animation production [13, 43-45, 55]. Meanwhile, to meet the diverse needs of visual media production and to enhance efficiency and quality, precise generation and editing methods have emerged. Models are required to perform generative creation based on multimodal inputs, such as depth, structure, pose, scene, and characters. According to the purposes of the input conditions, we can categorize them into two types: editing of the input and concept-guided re-creation. A significant portion of the work, such as ControlNet [76], ControlVideo,Composer [26], VideoComposer [68], and SCEdit [30], focuses on single-condition editing and multi-condition composite editing based on temporal and spatial alignment conditions. Additionally, some works that focus on interactive local editing scenarios, such as DragGAN [46] and MagicBrush [75]. Methods that guide generation based on semantic information from the input, such as Cone [38], Cone2 [39], InstantID [67], and PuLID [21], can achieve conceptual understanding of the input and inject it into the model for creative purposes.", + "bbox": [ + 511, + 387, + 906, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task-unified Visual Generative Model. As the complexity and diversity of user creations increase, relying solely on a single model or a complicated chain of multiple models can no longer provide a convenient and efficient path for implementing creative ideas. In image generation, a unified generation and editing framework has begun to emerge, allowing for more flexible creative approaches. Methods such as UltraEdit [81] and SEED-Data-Edit [19] provide general-purpose editing datasets, while techniques like In", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17192", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2bdea4dc0e5c25dd5529b7d9f6e57e9084415154530503123357e880e21f4386.jpg", + "image_caption": [ + "Figure 2. Task categories covered by VACE. Four basic tasks can be combined to create a vast number of possibilities." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 897, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "structPix2Pix [4], MagicBrush [61], and CosXL [60] offer general instruction-based editing features. Additionally, methods like UniControl [50] and UNIC-Adapter [15] have unified controllable generation. Further advancements have led to the development of ACE [23, 41], OmniGen [71], OmniControl [63], and UniReal [12], which expand the scope of tasks by providing flexible controllable generation, local editing, and reference-guided generation. In the video domain, due to the increased difficulty of generation, approaches often manifest as single-task single-model frameworks, offering capabilities for editing or reference generation, as seen in Video-P2P [37], MagicEdit [34], MotionCtrl [69], Magic Mirror [80], and Phantom [35]. VACE aims to fill the gap for a unified model within the video domain, providing possibilities for complex creative scenarios.", + "bbox": [ + 89, + 330, + 483, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 577, + 181, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VACE is designed as a multimodal-to-video generation model, where text, image, video, and mask are integrated into a unified conditioning input. To cover as many video generation and editing tasks as possible, we conduct in-depth research into existing tasks, then divide them into 4 categories according to their individual requirements of multimodal inputs. Without losing generality, we specifically design a novel multimodal input format for each category under a Video Condition Unit (VCU) paradigm. Finally, we restructure the DiT model for VCU inputs, making it a versatile model for a wide range of video tasks.", + "bbox": [ + 89, + 604, + 483, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multimodal Inputs and Video Tasks.", + "text_level": 1, + "bbox": [ + 89, + 786, + 406, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite existing video tasks varying in complex user inputs and ambitious creative goals, we found that most of their inputs can be fully represented in 4 modalities: text, image, video, and mask. Overall, as illustrated in Fig. 2, we group these video tasks into 5 categories based on their requirements of these four multimodal inputs.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Text-to-Video Generation (T2V) is a basic video creation task, and text is the only input.", + "- Reference-to-Video Generation (R2V) requires additional images as reference inputs, making sure that specified contents, such as subjects of faces, animals, and other objects, or video frames, appear in the generated video.", + "- Video-to-Video Editing (V2V) makes an entire change to a provided video, such as colorization, stylization, controllable generation, etc. We use video control types whose control signals can be represented and stored as RGB videos, including depth, grayscale, pose, scribble, optical flow, and layout; however, the method itself is not limited to these.", + "- Masked Video-to-Video Editing (MV2V) makes changes to an input video only within the provided 3D regions of interest (3D ROI), seamlessly blending in with the other unchanged regions, such as inpainting, outpainting, video extension, etc. We use an extra spatiotemporal mask to represent the 3D ROI.", + "- Task Composition includes all the compositional possibilities of the 4 types of video tasks above." + ], + "bbox": [ + 513, + 330, + 903, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Video Condition Unit", + "text_level": 1, + "bbox": [ + 513, + 657, + 714, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an input paradigm, Video Condition Unit (VCU), to unify diverse input conditions into textual input, frame sequence, and mask sequence. A VCU can be denoted as", + "bbox": [ + 511, + 680, + 906, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV = [ T; F; M ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 741, + 903, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T$ is a text prompt, while $F$ and $M$ are sequences of context video frames $\\{u_1, u_2, \\dots, u_n\\}$ and masks $\\{m_1, m_2, \\dots, m_n\\}$ respectively. Here, $u$ is in RGB space, normalized to $[-1, 1]$ and $m$ is binary, in which \"1\"s and \"0\"s symbolize where to edit or not. $F$ and $M$ are aligned both in spatial size $h \\times w$ and temporal size $n$ . In T2V, no context frame or mask is required. To keep generality, we assign a default value $0_{h \\times w}$ to each $u$ denoting empty input, and set every $m$ to $1_{h \\times w}$ meaning that all these pixels", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17193", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2a2cdb8973973f507ef1873f7f8e7f67b97950d6faa55f40b8c825386eb1ca8e.jpg", + "image_caption": [ + "Figure 3. Overview of VACE Framework. Frames and masks are tokenized through Concept Decoupling, Context Latent Encode, and Context Embedder. To achieve training with VCU as input, we employ two strategies: (a) Fully Fine-tuning and (b) Context Adapter Tuning. The latter converges faster and supports pluggable features." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 486, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/572bd3aa43bbd8d28bf5ab05a3e7a83f3135cec10cac1bc8df624e2eb426a3ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 88, + 901, + 257 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/92a45626850099a38ad40633be9b967221966e8df2a8a8367cad01631bff0ba4.jpg", + "table_caption": [ + "Table 1. The formal representation of frames $(F\\mathrm{s})$ and masks $(M\\mathrm{s})$ under the four basic tasks. Frames and masks are aligned spatially and temporally." + ], + "table_footnote": [], + "table_body": "
TasksFrames (Fs) & Masks (Ms)
T2VF = {0h×w} × n\nM = {1h×w} × n
R2VF = {r1, r2, ..., rl} + {0h×w} × n\nM = {0h×w} × l + {1h×w} × n
V2VF = {u1, u2, ..., un}\nM = {1h×w} × n
MV2VF = {u1, u2, ..., un}\nM = {m1, m2, ..., mn}
", + "bbox": [ + 91, + 376, + 470, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are about to be re-generated. For R2V, additional reference frames $r_i$ are inserted in front of the default frame sequence, while all-zero masks $0_{h \\times w}$ are inserted in front of the mask sequence. These all-zero masks mean that the corresponding frames should be kept unchanged. In V2V, the context frame sequence is the input video frames, and the context mask is a sequence of $1_{h \\times w}$ . For MV2V, both context video and mask are required. The formal mathematical representations are shown in Tab. 1.", + "bbox": [ + 89, + 575, + 482, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VCU can also support task composition. For example, the context frames of reference-inpainting task are $\\{r_1,r_2,\\dots,r_l,u_1,u_2,\\dots,u_n\\}$ and the context masks are $\\{0_{h\\times w}\\} \\times l + \\{m_1,m_2,\\dots,m_n\\}$ . In this case, users can modify $l$ objects in the video and regenerate based on the provided reference images. For another example, users only have a scribble image and want to generate a video beginning with the contents described by this scribble image, which is a scribble-based video extension task. The context frames are $\\{u\\} +\\{0_{h\\times w}\\} \\times (n - 1)$ and the context masks are $\\{1_{h\\times w}\\} \\times n$ . In this way, we can achieve multi-condition and reference control generation for long videos.", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Architecture", + "text_level": 1, + "bbox": [ + 511, + 325, + 648, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We restructure the DiT model for VACE, as shown in Fig. 3, aiming to support multimodal VCU inputs. Since there is an existing pipeline for text tokenization, we only consider the tokenization of context frames and masks. After being tokenized, the context tokens are combined with noisy video tokens and used to fine-tune the DiT model. Different from that, we also propose a Context Adapter Tuning strategy, which allows context tokens to pass through Context Blocks and be added back to the original DiT Blocks.", + "bbox": [ + 511, + 347, + 906, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Context Tokenization", + "text_level": 1, + "bbox": [ + 511, + 489, + 709, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept Decoupling. Two different visual concepts of natural video and control signals, like depth, pose, are encoded in $F$ simultaneously. We believe that explicitly separating these data of different modalities and distributions is essential for model convergence. The concept decoupling is based on masks and yields two frame sequences identical in shape: $F_{c} = F \\times M$ and $F_{k} = F \\times (1 - M)$ , where $F_{c}$ is called the reactive frames contain all the pixels to be changed, while all the pixels to be kept are stored in $F_{k}$ , named inactive frames. Specifically, the reference images and the unchanged part of V2V and MV2V go to $F_{k}$ , while control signals and those pixels about to change, such as gray pixels, are collected to $F_{c}$ .", + "bbox": [ + 511, + 508, + 906, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Context Latent Encoding. A typical DiT processes noisy video latents $X \\in \\mathbb{R}^{n' \\times h' \\times w' \\times d}$ , where $n'$ , $h'$ , and $w'$ are the temporal and spatial shapes of the latent space. Similar to $X$ , $F_{c}$ , $F_{k}$ , and $M$ need to be encoded into a high-dimensional feature space to ensure the property of significant spatiotemporal correlations. Therefore, we reorganize them together with $X$ into a hierarchical and spatiotemporal aligned visual features. $F_{c}$ , $F_{k}$ are processed by video VAE and mapped into the same latent space of $X$ , maintaining their spatiotemporal consistency. To avoid any unexpected mishmash of images and videos, reference images are separately encoded by VAE encoder and concatenated back along the temporal dimension, while the correspond", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17194", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing parts need to be removed during decoding. $M$ is directly reshaped and interpolated. After that, $F_{c}, F_{k}$ , and $M$ are all mapping into latent spaces and are spatiotemporal aligned with $X$ in the shape of $n' \\times h' \\times w'$ .", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Context Embedder. We extend the embedder layer by concatenating $F_{c}$ , $F_{k}$ , and $M$ in the channel dimension and tokenizing them into context tokens, which we refer to as the Context Embedder. The corresponding weights to tokenize $F_{c}$ and $F_{k}$ are directly copied from the original video embedder, and the weights to tokenize $M$ are initialized to zeros.", + "bbox": [ + 89, + 152, + 480, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Fully Fine-Tuning and Context Adapter Tuning", + "text_level": 1, + "bbox": [ + 89, + 263, + 467, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve training with VCU as input, a simple methodology is to fully fine-tune the whole DiT model, as shown in Fig. 3a. Context tokens are added together with noisy tokens $X$ , and all the parameters in DiT and the newly introduced Context Embedder will be updated during training. To avoid fully fine-tuning and achieve faster convergence, as well as to establish a pluggable feature with the foundation model, we also propose another methodology processing the context token in a Res-Tuning [29] manner, as shown in Fig. 3b. Particularly, we select and copy several Transformer Blocks from the original DiT, forming a distributed and cascade-type Context Blocks. The original DiT processes video tokens and text tokens, while the newly added Transformer Blocks process context tokens and text tokens. The output of each Context Block is inserted back into the DiT blocks as an additive signal to assist the main branch in performing generation and editing tasks. In this manner, the parameters of DiT are frozen. Only the Context Embedder and Context Blocks are trainable.", + "bbox": [ + 89, + 282, + 482, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Datasets", + "text_level": 1, + "bbox": [ + 89, + 582, + 187, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data Construction", + "text_level": 1, + "bbox": [ + 89, + 606, + 269, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain an all-in-one model, the diversity and complexity of the required data construction also increase. Existing common text-to-video and image-to-video tasks only require constructing pairs of text and video. However, for the tasks in VACE, the modalities need to be further expanded to include target videos, source videos, local masks, reference, and more. To efficiently and rapidly acquire data for various tasks, it is imperative to maintain video quality while also conducting instance-level analysis and understanding of the video data.", + "bbox": [ + 89, + 628, + 482, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve this, we first analyze the video data itself by performing shot slicing and preliminarily filtering out data based on resolution, aesthetic score, and motion amplitude. Next, we label the first frame of the video using RAM [78] and combine it with Grounding DINO [36] for detection, utilizing the localization results to perform secondary filtering on videos with target areas that are either too small or too large. Furthermore, we employ the propagate operation", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of SAM2 [52] for video segmentation to obtain instance-level information across the video. Leveraging the results of video segmentation, we filter instances in the temporal dimension by calculating the effective frame ratio based on the mask area threshold.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the actual training process, the construction for different tasks also needs to be tailored according to the characteristics of each task: 1) For some controllable video generation tasks, we pre-extract depth [51], scribble [6], pose [5, 72], and optical flow [65] from the filtered videos. For gray and layout tasks, we create data on the fly. 2) For repainting tasks, random instances from the videos can be masked for inpainting, while the inverse of the mask enables the construction of outpainting data. Augmentation of the masks [62] allows for unconditional inpainting. 3) In the case of extension tasks, we extract key frames such as the first frame, last frame, frames from both ends, random frames, and segments from both ends to support a wider variety of extension types. 4) For reference tasks, we can extract several face or object instances from the videos and apply offline or online augmentation operations to create paired data. Notably, we randomly combine all the previously mentioned tasks for training to accommodate a broader range of model application scenarios. Additionally, for all operations involving masks, we perform arbitrary augmentation to satisfy various granular local generation requirements.", + "bbox": [ + 511, + 167, + 906, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. VACE-Benchmark", + "text_level": 1, + "bbox": [ + 511, + 513, + 692, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Significant progress has been made in the field of video generation. However, a scientific and thorough evaluation of the performance of these models remains an urgent issue that needs to be addressed. VBench [27] and VBench++ [28] have established a precise evaluation framework for text-to-video and image-to-video tasks through an extensive assessment suite and dimensional design. Nevertheless, as the video generation ecosystem continues to evolve, more derivative tasks have begun to emerge, such as video reference generation and video editing, for which a comprehensive benchmark is still lacking. To address this gap, we propose VACE-Benchmark to evaluate various downstream tasks related to video in a systematic manner.", + "bbox": [ + 511, + 537, + 906, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Starting from the data sources, we recognize that real videos and generated videos may exhibit different performance characteristics during evaluation. Thus, we collected a total of 240 high-quality videos categorized by their sources, encompassing various data types, including text-to-video, inpainting, outpainting, extension, grayscale, depth, scribble, pose, optical flow, layout, reference face, and reference object tasks, with an average of 20 samples for each task. The input modalities include videos, masks, and references, and we also provide the original videos", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1fa4a54822a61bf1cc9e1b5b49f7de831ea2fd8e0e4c0076d2e8a8520a89fe66.jpg", + "table_caption": [ + "Table 2. Quantitative evaluations on VACE-Benchmark. We compare the automated score metrics of the unified VACE based on LTX-Video and the proprietary model on the dimensions of video quality and video consistency, as well as results of human user studies." + ], + "table_footnote": [], + "table_body": "
TypeMethodVideo Quality & Video ConsistencyUser Study
Aesthetic QualityBackground ConsistencyDynamic DegreeImaging QualityMotion SmoothnessOverall ConsistencySubject ConsistencyTemporal FlickeringNormalized AveragePrompt FollowingTemporal ConsistencyVideo QualityAverage
I2VI2VGenXL [77]55.20%92.87%60.00%63.31%97.43%23.78%89.58%95.67%71.54%2.651.602.342.20
CogVideoX-I2V [73]57.78%94.80%40.00%68.23%98.69%24.38%93.84%97.84%73.66%3.302.283.192.92
LTX-Video [22]56.12%94.57%35.00%62.72%99.27%24.92%92.83%98.41%72.89%2.952.282.282.50
VACE (Ours)57.53%95.32%45.00%68.03%99.08%25.13%93.61%97.80%74.38%3.204.002.543.24
InpaintProPainter [82]44.70%95.64%50.00%61.57%99.01%18.48%92.99%98.47%70.15%2.354.002.993.11
VACE (Ours)51.30%96.30%50.00%60.39%99.12%21.12%94.59%98.21%72.05%2.404.002.603.00
OutpaintFollow-Your-Canvas [8]53.30%95.99%5.00%69.53%98.08%25.90%95.38%97.20%71.54%3.052.001.632.23
M3DDM [17]53.34%95.87%30.00%65.07%99.22%25.43%93.65%98.85%73.16%3.703.882.283.29
VACE (Ours)57.04%96.55%30.00%69.49%99.20%25.36%94.47%98.47%74.25%3.903.923.583.80
DepthControl-A-Video [10]50.62%91.71%70.00%67.76%97.58%24.48%88.10%96.58%72.35%2.702.281.542.17
VideoComposer [68]50.03%94.18%70.00%59.44%96.23%24.95%89.79%94.38%70.74%2.602.442.172.40
ControlVideo [79]63.30%95.02%10.00%65.13%96.49%24.20%92.29%95.42%70.07%2.552.501.822.29
VACE (Ours)56.72%96.12%60.00%66.41%98.84%25.27%94.09%97.27%74.99%3.103.922.663.23
PoseText2Video-Zero [31]57.63%87.67%100.00%70.74%79.65%23.94%84.82%76.57%59.69%2.152.001.882.01
ControlVideo [79]65.37%94.56%25.00%65.28%97.32%25.19%92.76%96.82%72.45%2.151.802.031.99
Follow-Your-Pose [40]48.79%86.80%100.00%67.41%90.12%26.10%80.18%88.02%66.43%2.002.601.582.06
VACE (Ours)60.17%94.92%75.00%64.71%98.63%26.44%94.82%96.60%76.13%2.953.962.633.18
FlowFLATTEN [14]56.23%95.80%70.00%61.65%97.86%26.23%93.94%96.17%74.42%3.502.403.193.03
VACE (Ours)55.76%96.07%75.00%65.37%98.98%25.89%94.63%96.93%75.90%2.903.752.603.08
R2VKling1.6 [1]62.13%96.04%85.00%69.27%99.38%27.82%93.79%97.79%78.81%4.224.103.804.04
Pika2.2 [49]62.48%96.79%65.00%69.87%99.37%26.02%95.93%98.90%77.87%4.003.853.873.91
Vidu2.0 [66]64.30%96.85%35.00%67.03%99.66%26.53%96.73%99.41%76.47%3.903.853.773.84
VACE (Ours)63.25%98.03%30.00%72.29%99.51%25.85%98.54%99.15%76.76%3.473.423.303.40
", + "bbox": [ + 91, + 128, + 913, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to enable further processing by developers based on the specific characteristics of each task. Regarding the data prompts, we supply the original captions of the videos for quantitative assessment, as well as rewritten prompts tailored to the specific tasks to evaluate the models' creativity.", + "bbox": [ + 89, + 554, + 482, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 642, + 223, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 666, + 282, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. VACE is trained based on Diffusion Transformers for text-to-video generation at various scales. It utilizes LTX-Video-2B [22] for faster generation, while Wan-T2V-14B [64] is used specifically for higher-quality outputs, supporting resolutions of up to 720p. The training employs a phased approach. Initially, we focus on foundational tasks such as inpainting and extension, which are considered modal complementary to the pre-trained text-to-video models. This includes the incorporation of masks and the learning of contextual generation in both spatial and temporal dimensions. Next, from a task expansion perspective, we progressively transition from single input reference frames to multiple input reference frames and from single tasks to composite tasks. Fi", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nally, we fine-tune the model's quality using higher-quality data and longer sequences. The input for model training accommodates arbitrary resolutions, dynamic durations, and variable frame rates to support diverse input needs of users.", + "bbox": [ + 511, + 554, + 906, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Our goal is to achieve the unification of video creation and editing tasks, and currently, there is no comparable all-in-one video generation model available, which leads us to focus our evaluation on comparing our general model with proprietary task-specific models. Moreover, due to the numerous tasks involved and the lack of open-sourced methods for many of them, we conduct our comparisons on models that are available either offline or online. Specifically for the tasks, we compare the following: 1) For the I2V task, we examine I2VGenXL [77], CogVideoX-I2V [73], and LTX-Video-I2V [22]; 2) In the repainting task, we compare the ProPainter [82] for removal inpainting, while Follow-Your-Canvas [8] and M3DDM [17] are compared for outpainting; 3) For controllable task, we use Control-A-Video [10], VideoComposer [68], and ControlVideo [79] under depth conditions, and compare Text2Video-Zero [31], ControlVideo [79], and Follow-Your-Pose [40] under pose conditions, as well as FLAT", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17196", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7fa9ea987d1bb71f5051d3e521300faee221cee1c690174509157a93ceb2484.jpg", + "image_caption": [ + "Reference Anything" + ], + "image_footnote": [], + "bbox": [ + 98, + 93, + 243, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6416d4a007a14dd68aefc6bfd0f1fe6cb773c3b035d97636f24ebb9a0c54d90.jpg", + "image_caption": [], + "image_footnote": [ + "The elegant lady carefully selects bags in the boutique, and she shows the charm of a mature woman in a black slim dress with a pearl necklace. Holding a vintage-inspired brown leather handbag, she is carefully observing its craftsmanship and texture. The interior of the store." + ], + "bbox": [ + 254, + 88, + 897, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8434dacb2e4a7709e3f78e2229e213784dba4a91b4b1f46122d498181dd52856.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 183, + 254, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7755152ca82d7f0fbc13b2b7b203bfcb16d5f3adeb3ff41ce7b58ef1ace9a95a.jpg", + "image_caption": [], + "image_footnote": [ + "A young boy rises from his chair and walks briskly to the right side of the frame towards the edge of the sun-drenched frame, as if chasing a new adventure. His eyes were bright, and the corners of his mouth were slightly upturned, revealing curiosity and excitement about the unknown..." + ], + "bbox": [ + 254, + 180, + 898, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a2a7dc2056bb9eefe68f4fd0f438de07109f6e820e51ec5c8399a3d85f38d3d.jpg", + "image_caption": [ + "Move Anything" + ], + "image_footnote": [ + "\"Anime-style, hot-blooded teenager in bright orange long-sleeved pants sportswear, standing on a surfboard, facing the golden sunshine in the rough sea. The teenager's short yellow hair is flying in the wind, his eyes are firm, and he has a confident smile on the corner of his mouth.\"" + ], + "bbox": [ + 98, + 271, + 254, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ddbe8eeb9b05ec19d551d6405830a378b3c62dc317578be5cdea44cf0957355.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 271, + 898, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04fcf82eeef0b0e26c17bc7b5f184eadced2d8d689a0a494b4537ebd63f4d9cb.jpg", + "image_caption": [ + "Animate Anything" + ], + "image_footnote": [ + "The video shows an old movie-style scene in retro tones, with a little penguin and a kitten having a joyous bike race. Little penguins and kittens, both dressed in orange-and-red race suits, ride vintage multi-wheeled bikes on a nostalgic dirt road flanked by spectators..." + ], + "bbox": [ + 96, + 362, + 254, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a20893df72f446a06fe28fb383ae5b5f8fa37e8addd8069ae865846de816e160.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 362, + 897, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5105cdd07923abce73ff9a89c0e5639e4918777310ce414986bec976cd654a25.jpg", + "image_caption": [ + "Swap Anything" + ], + "image_footnote": [], + "bbox": [ + 96, + 462, + 254, + 516 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9751db892a0a52597f8e5f5d005db3142b8067aa9d728a4fd80dd8f895ed60a5.jpg", + "image_caption": [ + "Expand Anything", + "Figure 4. Visualization results of compositional tasks. VACE creatively enables reference-, move-, animate-, swap-, and expand-anything." + ], + "image_footnote": [ + "\"In the style of classical oil painting, the background is a river, and in the center of the picture is a mature and elegant woman, wearing a long skirt and sitting on a chair. She took the open red heart-shaped sunglasses from her arms with both hands and put them on...\"" + ], + "bbox": [ + 254, + 452, + 897, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TEN [14] under optical flow conditions; 4) In reference generation, given the absence of open-sourced models, we compare commercial products Kling1.6 [1], Pika2.2 [49], and Vidu2.0 [66].", + "bbox": [ + 89, + 597, + 482, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. To comprehensively evaluate the performance of various tasks, we employ the VACE-Benchmark for assessment. Specifically, we divide the evaluation into automatic scoring and a user study for manual assessment. For the automatic scoring, we utilize select metrics from VBench [27] to assess video quality and video consistency, including eight indicators: aesthetic quality, background consistency, dynamic degree, imaging quality, motion smoothness, overall consistency, subject consistency, and temporal flickering. For the manual assessment, we utilize the mean opinion score (MOS) as our evaluation metric, focusing on three aspects: prompt following, temporal consistency, and video quality. In practice, we anonymize the generated data and randomly distribute it to different participants for scoring on a scale from 1 to 5.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 595, + 651, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. We compare VACE comprehensive model based on LTX-Video with task proprietary approaches on VACE-Benchmark. For certain tasks, we follow existing methods; for example, although we support generating based on any frame, we conduct comparisons using the first-frame reference approach from current open-source methods to ensure fairness. From Tab. 2, we can see that for the tasks of I2V, inpainting, outpainting, depth, pose, and optical flow, our method demonstrates better performance than other open-source methods across eight indicators of video quality and video consistency, with normalized average metrics showing superior results. Some competing methods can only generate at a resolution of 256, have very short generation durations, and exhibit instability in temporal coherence, resulting in poorer performance on automatic metric calculations. For the R2V task, there is still a certain gap in metrics compared to commercial models for a small-scale model that aims for fast generation,", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17197", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfdbd7dd56c30c3a6af8e7f6a707e50ca6f36261f186b92d725ce4882711d134.jpg", + "image_caption": [ + "(a) Base structure setting." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 285, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/271e1884078ac46369e29ea040268eec90e4a0725fcd2514331506d8f08d1a3c.jpg", + "image_caption": [ + "(b) Hyperparameter settings." + ], + "image_footnote": [], + "bbox": [ + 287, + 95, + 473, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95c6ecc97504c4d853d22a32ff59fd3c44365b48d077815be8a79fc91f3a7089.jpg", + "image_caption": [ + "(c) Context adapter configurations." + ], + "image_footnote": [], + "bbox": [ + 98, + 227, + 284, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14fdebd9ed061f7d0e1c6ae123885833ea7a70db27684d65351cc2fd5fe7e5d8.jpg", + "image_caption": [ + "(d) Concept decouple setting.", + "Figure 5. Ablation Studies of the VACE regarding structures, hyperparameters, and module configurations." + ], + "image_footnote": [], + "bbox": [ + 287, + 227, + 473, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "while being comparable to the metrics of Vidu 2.0. According to the results of human user studies, our method consistently performs better in evaluation metrics across multiple tasks, aligning well with user preferences.", + "bbox": [ + 89, + 422, + 482, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 1, we present the results of the VACE single model across various tasks. It is evident that the model achieves a high level of performance in video quality and temporal consistency. Furthermore, in composition tasks shown in Fig. 4, our model showcases impressive abilities, effectively integrating different modalities and tasks to produce results that cannot be generated by existing single or multiple models, thereby demonstrating its strong potential in the fields of video generation and editing. For example, in the \"Move Anything\" case, by providing a single input image and a movement trajectory, we are able to precisely move the characters in the scene with a specified direction while maintaining coherence and narrative consistency.", + "bbox": [ + 89, + 484, + 482, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 709, + 254, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To better understand the impact of different independent modules on a unified video generation framework, we conducted a series of systematic comparative experiments based on the LTX-Video model to achieve a better model structure and configuration. To accurately assess the different experimental settings, we sample 250 data points for each task as a validation set and calculate the training loss, reflecting the model's training progress through the mean curve changes of different tasks.", + "bbox": [ + 89, + 732, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Base Structure. Text-guided image or video generation models only take noise as inference input. When extended", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to our unified input paradigm, VCU, we can conduct training using fully fine-tuning or by incorporating additional parameter fine-tuning. Specifically, as shown in Fig. 5a, we compare the concatenation of different inputs along the channel dimension and modify the input dimensions of the patchify projection layer to achieve the loading and fully fine-tuning of the pre-trained model. Additionally, we introduce some extra training parameters in the form of RestTuning [29], which serializes VCU in a bypass branch and injects information into the main branch. The results indicate that both methods yielded similar effects; however, since the additional parameter fine-tuning converges faster, we base our subsequent experiments on this approach. As shown in Fig. 5b, we further conduct hyperparameter experiments based on this structure, focusing on aspects such as weighting schemes, timestamp shifting, and p-zero.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Context Adapter. Since the number of context blocks will significantly affect the model size and inference time consumption, we attempt to find an optimal number and distribution of context blocks. We begin with selecting continuous blocks at the input side and make comparisons between the first 1/4 blocks, 1/2 blocks, and all blocks. Inspired by the Res-Tuning [29] method, we also experiment with evenly distributing the injection blocks instead of selecting a continuous block series. As shown in Fig. 5c, we can see that when using the same number of blocks, the distributed arrangement of blocks outperforms the continuous arrangement in shallow blocks. Furthermore, a greater number of blocks generally yields better results, but due to the limited improvement in effectiveness and the constraints of training resources, we adopt a partially distributed arrangement of blocks.", + "bbox": [ + 511, + 332, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Concept Decouple. During training, we introduce a Concept Decouple processing module to further disassemble the visual units, clarifying what content the model needs to learn to modify or retain. As shown in Fig. 5d, using this module results in a more significant reduction in loss.", + "bbox": [ + 511, + 574, + 905, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces VACE, an all-in-one video generation and editing framework. It unifies the diverse and complex multimodal inputs required for various video tasks, bridging the gap between specialized models for each individual task. This enables most video AI creation tasks to be completed with a single inference of a single model. While broadly covering various video tasks, VACE also supports flexible and free combinations of these tasks, greatly expanding the application scenarios of video generation models and meeting a wide range of user creative needs. The VACE framework paves the way for the development of unified visual generative models with multimodal inputs and represents a significant milestone in the field of visual generation.", + "bbox": [ + 511, + 689, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17198", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We would like to express our sincere appreciation for the contributions of many colleagues for their insightful discussions, valuable suggestions, and constructive feedback, including: Yuwei Wang, Haiming Zhao, Chenwei Xie, and Sheng Yao for their data contributions, and Shiwei Zhang, Tao Fang, and Xiang Wang for their discussions and suggestions.", + "bbox": [ + 89, + 90, + 485, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 219, + 187, + 234 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] KLING AI. KLING AI, https://klingai.com/, 2025.6,7,13", + "[2] Runway AI. Stable Diffusion v1.5 Model Card, https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022.2", + "[3] Runway AI. Stable Diffusion Inpainting Model Card, https://huggingface.co/runwayml/stable-diffusion-inpainting, 2022.2", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning To Follow Image Editing Instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 2, 3", + "[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 5", + "[6] Caroline Chan, Frédo Durand, and Phillip Isola. Learning To Generate Line Drawings That Convey Geometry and Semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 5", + "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- $\\alpha$ : Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv preprint arXiv:2310.00426, 2023. 2", + "[8] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-Your-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation. In Assoc. Adv. Artif. Intell., 2025. 6, 13", + "[9] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, Ting-Che Lin, Shilong Zhang, Fu Li, Chuan Li, Xing Wang, Yanghua Peng, Peize Sun, Ping Luo, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Goku: Flow Based Video Generative Foundation Models. arXiv preprint arXiv:2502.04896, 2025. 2", + "[10] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable Text-to-Video Diffusion Models with Motion Prior and Reward Feedback Learning. arXiv preprint arXiv:2305.13840, 2023. 6, 13", + "[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. AnyDoor: Zero-shot Object-level Image Customization. arXiv preprint arXiv:2307.09481, 2023. 2" + ], + "bbox": [ + 93, + 243, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang Zhao. UniReal: Universal Image Generation and Editing via Learning Real-world Dynamics. arXiv preprint arXiv:2412.07774, 2024. 2, 3", + "[13] Alibaba Cloud. Tongyi Wanxiang, https://tongyi. aliyun.com/wanxiang, 2023. 2", + "[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: Optical FLOW-guided ATTENtion for consistent text-to-video editing. In Int. Conf. Learn. Represent., 2024. 6, 7, 13", + "[15] Lunhao Duan, Shanshan Zhao, Wenjun Yan, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Mingming Gong, and Gui-Song Xia. UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation. arXiv preprint arXiv:2412.18928, 2024.3", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Int. Conf. Mach. Learn., 2024. 2", + "[17] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical Masked 3D Diffusion Model for Video Outpainting. In ACM Int. Conf. Multimedia, pages 7890-7900, 2023. 6, 13", + "[18] FLUX. FLUX, https://blackforestlabs.ai/, 2024.2", + "[19] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. SEED-Data-Edit Technical Report: A Hybrid Dataset for Instructional Image Editing. arXiv preprint arXiv:2405.04007, 2024. 2", + "[20] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, Haibin Huang, and Chongyang Ma. 12V-Adapter: A General Image-to-Video Adapter for Diffusion Models. In ACM SIGGRAPH, pages 1-12, 2024. 2", + "[21] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Adv. Neural Inform. Process. Syst., 2024. 2", + "[22] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion. arXiv preprint arXiv:2501.00103, 2025. 2, 6, 13", + "[23] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer. In Int. Conf. Learn. Represent., 2025. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 897 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17199", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In Adv. Neural Inform. Process. Syst., 2021. 2", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Adv. Neural Inform. Process. Syst. Curran Associates, Inc., 2020. 2", + "[26] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and Controllable Image Synthesis with Composable Conditions. In Int. Conf. Mach.Learn., 2023.2", + "[27] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive Benchmark Suite for Video Generative Models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21807-21818, 2024. 5, 7", + "[28] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, YingCong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench++: Comprehensive and Versatile Benchmark Suite for Video Generative Models. arXiv preprint arXiv:2411.13503, 2024. 5", + "[29] Zeyinzi Jiang, Chaojie Mao, Ziyuan Huang, Ao Ma, Yiliang Lv, Yujun Shen, Deli Zhao, and Jingren Zhou. Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone. In Adv. Neural Inform. Process. Syst., 2023. 5, 8", + "[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8995-9004, 2024. 2", + "[31] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In Int. Conf. Comput. Vis., pages 15954–15964, 2023. 6, 13", + "[32] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyuan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models. arXiv preprint arXiv:2412.03603, 2024. 2", + "[33] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 2" + ], + "bbox": [ + 89, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. MagicEdit: High-Fidelity and Temporally Coherent Video Editing. arXiv preprint arXiv:2308.14749, 2023. 3", + "[35] Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv preprint arXiv:2502.11079, 2025. 2, 3", + "[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. arXiv preprint arXiv:2303.05499, 2023. 5", + "[37] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-P2P: Video Editing with Cross-attention Control. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8599–8608, 2024. 3", + "[38] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept Neurons in Diffusion Models for Customized Generation. In Int. Conf. Mach. Learn., 2023. 2", + "[39] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable Image Synthesis with Multiple Subjects. In Adv. Neural Inform. Process. Syst., 2023. 2", + "[40] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Ying Shan, Xiu Li, and Qifeng Chen. Follow Your Pose: Pose-Guided Text-to-Video Generation using Pose-Free Videos. In Assoc. Adv. Artif. Intell., 2024. 6, 13", + "[41] Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. ACE++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling. arXiv preprint arXiv:2501.02487, 2025. 2, 3", + "[42] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2", + "[43] Midjourney. Midjourney, https://www.midjourney.com, 2023.2", + "[44] MiniMax. Hailuo AI Video, https://hailuoai.com/video, 2024.", + "[45] OpenAI. DALL·E 3, https://openai.com/dall-e-3, 2023.2", + "[46] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold. In ACM SIGGRAPH, 2023. 2", + "[47] Yulin Pan, Chaojie Mao, Zeyinzi Jiang, Zhen Han, and Jingfeng Zhang. Locate, Assign, Refine: Taming Customized Image Inpainting with Text-Subject Guidance. arXiv preprint arXiv:2403.19534, 2024. 2", + "[48] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In Int. Conf. Comput. Vis., pages 4195-4305, 2023. 2", + "[49] PiKa. PiKa, https://pika.art/, 2025.6, 7, 13", + "[50] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming" + ], + "bbox": [ + 514, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17200", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild. In Adv. Neural Inform. Process. Syst., 2023. 3", + "[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. IEEE Trans. Pattern Anal. Mach. Intell., pages 1623-1637, 2022. 5", + "[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. In Int. Conf. Learn. Represent., 2025. 5", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 2", + "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. Med. Image Comput. Computer-Assisted Interv., 2015. 2", + "[55] Runway. Gen-3, https://app.runwayml.com/video-tools, 2025.2", + "[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In Int. Conf. Learn. Represent., 2021. 2", + "[57] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2", + "[58] StabilityAI. Stable Diffusion v2-1 Model Card, https: //huggingface.co/stabilityai/stable-diffusion-2-1,2022.2", + "[59] StabilityAI. Stable Diffusion XL Model Card, https: //huggingface.co/stabilityai/stable-diffusion-x1-base-1.0,2022.2", + "[60] StabilityAI. CosXL Model Card, https://huggingface.co/stabilityai/cosxl, 2024.3", + "[61] Ya Sheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation. In Adv. Neural Inform. Process. Syst., 2023. 3", + "[62] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-Robust Large Mask Inpainting With Fourier Convolutions. In IEEE Winter Conf. Appl. Comput. Vis., pages 2149-2159, 2022. 5", + "[63] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. OminiControl: Minimal and Universal Control for Diffusion Transformer. arXiv preprint arXiv:2411.15098, 2024. 2, 3", + "[64] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2, 6, 13" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In Eur. Conf. Comput. Vis., pages 402-419, 2020. 5", + "[66] Vidu. Vidu, https://www.vidu.cn/, 2025. 6, 7, 13", + "[67] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 2", + "[68] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In Adv. Neural Inform. Process. Syst., 2023. 2, 6, 13", + "[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A Unified and Flexible Motion Controller for Video Generation. In ACM SIGGRAPH, pages 1-11, 2024. 3", + "[70] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, Yingya Zhang, and Hongming Shan. DreamVideo-2: Zero-Shot Subject-Driven Video Customization with Precise Motion Control. arXiv preprint arXiv:2410.13830, 2024. 2", + "[71] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. OmniGen: Unified Image Generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3", + "[72] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective Whole-body Pose Estimation with Two-stages Distillation. In Int. Conf. Comput. Vis., pages 4210-4220, 2023. 5", + "[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. In Int. Conf. Learn. Represent., 2025. 2, 6, 13", + "[74] Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-Preserving Text-to-Video Generation by Frequency Decomposition. In IEEE Conf. Comput. Vis. Pattern Recog., 2025. 2", + "[75] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. In Adv. Neural Inform. Process. Syst., 2023. 2", + "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In Int. Conf. Comput. Vis., pages 3836–3847, 2023. 2", + "[77] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv preprint arXiv:2311.04145, 2023. 2, 6, 13", + "[78] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo," + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17201", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize Anything: A Strong Image Tagging Model. arXiv preprint arXiv:2306.03514, 2023. 5", + "[79] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo: Training-free Controllable Text-to-Video Generation. In Int. Conf. Learn. Represent., 2024. 6, 13", + "[80] Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. Magic Mirror: ID-Preserved Video Generation in Video Diffusion Transformers. arXiv preprint arXiv:2411.13503, 2025. 3", + "[81] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based Fine-Grained Image Editing at Scale. arXiv preprint arXiv:2407.05282v1, 2024. 2", + "[82] Shangchen Zhou, Chongyi Li, Kelvin C. K. Chan, and Chen Change Loy. ProPainter: Improving Propagation and Transformer for Video Inpainting. In Int. Conf. Comput. Vis., pages 10477-10486, 2023. 2, 6, 13" + ], + "bbox": [ + 91, + 90, + 480, + 373 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17202", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_model.json b/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0c6c498b9acdf74be00f36ad1677c0e54239fa5 --- /dev/null +++ b/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_model.json @@ -0,0 +1,2765 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.264, + 0.131, + 0.737, + 0.154 + ], + "angle": 0, + "content": "VACE: All-in-One Video Creation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.181, + 0.273, + 0.2 + ], + "angle": 0, + "content": "Zeyinzi Jiang*" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.182, + 0.374, + 0.198 + ], + "angle": 0, + "content": "Zhen Han*" + }, + { + "type": "text", + "bbox": [ + 0.396, + 0.182, + 0.515, + 0.199 + ], + "angle": 0, + "content": "Chaojie Mao*†" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.182, + 0.663, + 0.199 + ], + "angle": 0, + "content": "Jingfeng Zhang" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.182, + 0.764, + 0.198 + ], + "angle": 0, + "content": "Yulin Pan" + }, + { + "type": "text", + "bbox": [ + 0.785, + 0.182, + 0.842, + 0.197 + ], + "angle": 0, + "content": "Yu Liu" + }, + { + "type": "title", + "bbox": [ + 0.388, + 0.209, + 0.61, + 0.228 + ], + "angle": 0, + "content": "Tongyi Lab, Alibaba Group" + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.241, + 0.582, + 0.251 + ], + "angle": 0, + "content": "Reference-to-video generation" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.251, + 0.488, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.253, + 0.603, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.251, + 0.903, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.301, + 0.561, + 0.312 + ], + "angle": 0, + "content": "Video-to-video editing" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.311, + 0.488, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.352, + 0.357, + 0.361 + ], + "angle": 0, + "content": "Pose Transfer (Pose Control)" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.311, + 0.902, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.352, + 0.769, + 0.361 + ], + "angle": 0, + "content": "Motion Transfer (Layout Control)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.37, + 0.488, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.412, + 0.37, + 0.421 + ], + "angle": 0, + "content": "Structure Transfer (Depth Control)" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.37, + 0.902, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.412, + 0.757, + 0.421 + ], + "angle": 0, + "content": "Colorization (Gray Control)" + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.423, + 0.583, + 0.434 + ], + "angle": 0, + "content": "Masked Video-to-video editing" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.433, + 0.488, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.474, + 0.363, + 0.484 + ], + "angle": 0, + "content": "Subject Recreation (Inpainting)" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.433, + 0.902, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.474, + 0.761, + 0.484 + ], + "angle": 0, + "content": "Subject Removal (Inpainting)" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.492, + 0.488, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.535, + 0.364, + 0.544 + ], + "angle": 0, + "content": "Canvas Extension (Outpainting)" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.492, + 0.902, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.535, + 0.754, + 0.544 + ], + "angle": 0, + "content": "Temporal Extension (First Clip)" + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.546, + 0.549, + 0.556 + ], + "angle": 0, + "content": "Task Composition" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.552, + 0.488, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.598, + 0.4, + 0.607 + ], + "angle": 0, + "content": "Swap Anything (Subject Reference + Inpainting)" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.556, + 0.902, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.598, + 0.815, + 0.607 + ], + "angle": 0, + "content": "Animate Anything (Frame Reference + Pose Control)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.619, + 0.907, + 0.649 + ], + "angle": 0, + "content": "Figure 1. Comprehensive capabilities of VACE. We present the outstanding generation results based on Wan2.1 (left) and LTX-Video (right). For each task, the original input image or video (left), the context video (top left corner), and the generated frames are illustrated." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.667, + 0.327, + 0.683 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.698, + 0.486, + 0.88 + ], + "angle": 0, + "content": "Diffusion Transformer has demonstrated powerful capability and scalability in generating high-quality images and videos. Further pursuing the unification of generation and editing tasks has yielded significant progress in the domain of image content creation. However, due to the intrinsic demands for consistency across both temporal and spatial dynamics, achieving a unified approach for video synthesis remains challenging. We introduce VACE, which enables users to perform Video tasks within an All-in-one framework for Creation and Editing. These tasks include reference-to-video generation, video-to-video editing, and masked video-to-video editing. Specifically, we effectively" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.909, + 0.865 + ], + "angle": 0, + "content": "integrate the requirements of various tasks by organizing video task inputs, such as editing, reference, and masking, into a unified interface referred to as the Video Condition Unit (VCU). Furthermore, by utilizing a Context Adapter structure, we inject different task concepts into the model using formalized representations of temporal and spatial dimensions, allowing it to handle arbitrary video synthesis tasks flexibly. Extensive experiments demonstrate that the unified model of VACE achieves performance on par with task-specific models across various subtasks. Simultaneously, it enables diverse applications through versatile task combinations. Project page: https://alivilab.github.io/VACE-Page/." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.299, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution. †Project lead." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17191" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.478 + ], + "angle": 0, + "content": "In recent years, the domain of visual generation tasks has witnessed remarkable advancements, driven in particular by the rapid evolution of diffusion models [24, 25, 48, 53, 54, 56, 57]. Beyond the early foundational pre-trained models for text-to-image [7, 16, 33] or text-to-video [9, 22, 64] generation in the field, there has been a proliferation of downstream tasks and applications, such as painting [3, 82], editing [4, 42, 68, 70, 75], controllable generation [30, 76], frame reference generation [20, 73], and ID-referenced video synthesis [11, 35, 47, 74]. This array of developments highlights the dynamic and complex nature of the visual generation field. To enhance task flexibility and reduce the overhead associated with deploying multiple models, researchers have begun to focus on constructing unified model architectures [12, 63] (e.g., ACE [23, 41] and Omni-Gen [71]), aiming to integrate different tasks into a single image model, facilitating the creation of various application workflows while maintaining simplicity in usage. In the field of video, due to the collaborative transformations in both temporal and spatial dimensions, leveraging a unified model can present endless possibilities for video creation. However, leveraging diverse input modalities and ensuring spatiotemporal consistency are still challenging for unified video generation and editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We propose VACE, an all-in-one model for video creation and editing that performs tasks including reference-to-video generation, video-to-video editing, masked video-to-video editing, and free composition of these tasks, as illustrated in Fig. 1. The aggregation of various capabilities reduces the costs of service deployment and user interaction, and by combining the capabilities of different tasks within a single model, it addresses challenges faced by existing video generation models, such as controllable generation of long videos, multi-condition and reference-based generation, and continuous video editing, thereby empowering users with greater creativity. To achieve this, we utilize the current mainstream Diffusion Transformers (DiTs) structure as the foundational video framework and pre-trained text-to-video generation models [22, 64], which provide better basic capabilities and scalability for handling long video sequences. Specifically, VACE takes into account the needs of different tasks during its construction and designs a unified interface, dubbed the Video Condition Unit (VCU), which integrates multiple modalities such as images or videos for editing, references, and masks. Furthermore, to differentiate the visual modality information in editing and reference tasks, we introduce the concept decoupling strategy, enabling the model to understand what aspects need to be retained and what should be modified. Meanwhile, by employing a pluggable Context Adapter structure, concepts from different tasks (e.g., the areas or ranges of editing or reference) are injected into the model through col" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "laborative spatiotemporal representation, enabling it to possess the capability of adaptive processing for unified tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Due to the lack of existing multi-task benchmarks in video synthesis, we construct a dataset of 480 evaluation samples containing 12 different tasks, while evaluating the performance of the VACE unified model by comparing it with existing specialized models. Experimental results demonstrate that our framework exhibits sufficient competitiveness in both quantitative and qualitative analyses. To the best of our knowledge, we are the first all-in-one model based on the video DiT architecture that concurrently supports such a wide range of tasks. Notably, this innovative framework allows for the compositional expansion of base tasks, enabling the construction of scenarios such as long video re-rendering, which provides a versatile and efficient solution for video synthesis, opening new possibilities for user-side video content creation and editing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.362, + 0.656, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.908, + 0.765 + ], + "angle": 0, + "content": "Visual Generation and Editing. With the rapid development of image [2, 7, 16, 18, 58, 59] and video [22, 32, 73, 77] generation models, they are being used to create high-quality visual content and are widely applied in fields such as advertising, film special effects, game development, and animation production [13, 43-45, 55]. Meanwhile, to meet the diverse needs of visual media production and to enhance efficiency and quality, precise generation and editing methods have emerged. Models are required to perform generative creation based on multimodal inputs, such as depth, structure, pose, scene, and characters. According to the purposes of the input conditions, we can categorize them into two types: editing of the input and concept-guided re-creation. A significant portion of the work, such as ControlNet [76], ControlVideo,Composer [26], VideoComposer [68], and SCEdit [30], focuses on single-condition editing and multi-condition composite editing based on temporal and spatial alignment conditions. Additionally, some works that focus on interactive local editing scenarios, such as DragGAN [46] and MagicBrush [75]. Methods that guide generation based on semantic information from the input, such as Cone [38], Cone2 [39], InstantID [67], and PuLID [21], can achieve conceptual understanding of the input and inject it into the model for creative purposes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Task-unified Visual Generative Model. As the complexity and diversity of user creations increase, relying solely on a single model or a complicated chain of multiple models can no longer provide a convenient and efficient path for implementing creative ideas. In image generation, a unified generation and editing framework has begun to emerge, allowing for more flexible creative approaches. Methods such as UltraEdit [81] and SEED-Data-Edit [19] provide general-purpose editing datasets, while techniques like In" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "17192" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.898, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.301, + 0.851, + 0.316 + ], + "angle": 0, + "content": "Figure 2. Task categories covered by VACE. Four basic tasks can be combined to create a vast number of possibilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.485, + 0.56 + ], + "angle": 0, + "content": "structPix2Pix [4], MagicBrush [61], and CosXL [60] offer general instruction-based editing features. Additionally, methods like UniControl [50] and UNIC-Adapter [15] have unified controllable generation. Further advancements have led to the development of ACE [23, 41], OmniGen [71], OmniControl [63], and UniReal [12], which expand the scope of tasks by providing flexible controllable generation, local editing, and reference-guided generation. In the video domain, due to the increased difficulty of generation, approaches often manifest as single-task single-model frameworks, offering capabilities for editing or reference generation, as seen in Video-P2P [37], MagicEdit [34], MotionCtrl [69], Magic Mirror [80], and Phantom [35]. VACE aims to fill the gap for a unified model within the video domain, providing possibilities for complex creative scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.182, + 0.594 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.772 + ], + "angle": 0, + "content": "VACE is designed as a multimodal-to-video generation model, where text, image, video, and mask are integrated into a unified conditioning input. To cover as many video generation and editing tasks as possible, we conduct in-depth research into existing tasks, then divide them into 4 categories according to their individual requirements of multimodal inputs. Without losing generality, we specifically design a novel multimodal input format for each category under a Video Condition Unit (VCU) paradigm. Finally, we restructure the DiT model for VCU inputs, making it a versatile model for a wide range of video tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.407, + 0.803 + ], + "angle": 0, + "content": "3.1. Multimodal Inputs and Video Tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Despite existing video tasks varying in complex user inputs and ambitious creative goals, we found that most of their inputs can be fully represented in 4 modalities: text, image, video, and mask. Overall, as illustrated in Fig. 2, we group these video tasks into 5 categories based on their requirements of these four multimodal inputs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.332, + 0.905, + 0.361 + ], + "angle": 0, + "content": "- Text-to-Video Generation (T2V) is a basic video creation task, and text is the only input." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.363, + 0.905, + 0.422 + ], + "angle": 0, + "content": "- Reference-to-Video Generation (R2V) requires additional images as reference inputs, making sure that specified contents, such as subjects of faces, animals, and other objects, or video frames, appear in the generated video." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.424, + 0.905, + 0.527 + ], + "angle": 0, + "content": "- Video-to-Video Editing (V2V) makes an entire change to a provided video, such as colorization, stylization, controllable generation, etc. We use video control types whose control signals can be represented and stored as RGB videos, including depth, grayscale, pose, scribble, optical flow, and layout; however, the method itself is not limited to these." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.529, + 0.905, + 0.618 + ], + "angle": 0, + "content": "- Masked Video-to-Video Editing (MV2V) makes changes to an input video only within the provided 3D regions of interest (3D ROI), seamlessly blending in with the other unchanged regions, such as inpainting, outpainting, video extension, etc. We use an extra spatiotemporal mask to represent the 3D ROI." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.62, + 0.904, + 0.65 + ], + "angle": 0, + "content": "- Task Composition includes all the compositional possibilities of the 4 types of video tasks above." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.332, + 0.905, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.659, + 0.715, + 0.674 + ], + "angle": 0, + "content": "3.2. Video Condition Unit" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.74 + ], + "angle": 0, + "content": "We propose an input paradigm, Video Condition Unit (VCU), to unify diverse input conditions into textual input, frame sequence, and mask sequence. A VCU can be denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.742, + 0.905, + 0.759 + ], + "angle": 0, + "content": "\\[\nV = [ T; F; M ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\( T \\) is a text prompt, while \\( F \\) and \\( M \\) are sequences of context video frames \\( \\{u_1, u_2, \\dots, u_n\\} \\) and masks \\( \\{m_1, m_2, \\dots, m_n\\} \\) respectively. Here, \\( u \\) is in RGB space, normalized to \\( [-1, 1] \\) and \\( m \\) is binary, in which \"1\"s and \"0\"s symbolize where to edit or not. \\( F \\) and \\( M \\) are aligned both in spatial size \\( h \\times w \\) and temporal size \\( n \\). In T2V, no context frame or mask is required. To keep generality, we assign a default value \\( 0_{h \\times w} \\) to each \\( u \\) denoting empty input, and set every \\( m \\) to \\( 1_{h \\times w} \\) meaning that all these pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17193" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.487, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.089, + 0.902, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.908, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Overview of VACE Framework. Frames and masks are tokenized through Concept Decoupling, Context Latent Encode, and Context Embedder. To achieve training with VCU as input, we employ two strategies: (a) Fully Fine-tuning and (b) Context Adapter Tuning. The latter converges faster and supports pluggable features." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Table 1. The formal representation of frames \\((F\\mathrm{s})\\) and masks \\((M\\mathrm{s})\\) under the four basic tasks. Frames and masks are aligned spatially and temporally." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.377, + 0.472, + 0.55 + ], + "angle": 0, + "content": "
TasksFrames (Fs) & Masks (Ms)
T2VF = {0h×w} × n\nM = {1h×w} × n
R2VF = {r1, r2, ..., rl} + {0h×w} × n\nM = {0h×w} × l + {1h×w} × n
V2VF = {u1, u2, ..., un}\nM = {1h×w} × n
MV2VF = {u1, u2, ..., un}\nM = {m1, m2, ..., mn}
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.712 + ], + "angle": 0, + "content": "are about to be re-generated. For R2V, additional reference frames \\( r_i \\) are inserted in front of the default frame sequence, while all-zero masks \\( 0_{h \\times w} \\) are inserted in front of the mask sequence. These all-zero masks mean that the corresponding frames should be kept unchanged. In V2V, the context frame sequence is the input video frames, and the context mask is a sequence of \\( 1_{h \\times w} \\). For MV2V, both context video and mask are required. The formal mathematical representations are shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "VCU can also support task composition. For example, the context frames of reference-inpainting task are \\(\\{r_1,r_2,\\dots,r_l,u_1,u_2,\\dots,u_n\\}\\) and the context masks are \\(\\{0_{h\\times w}\\} \\times l + \\{m_1,m_2,\\dots,m_n\\}\\). In this case, users can modify \\(l\\) objects in the video and regenerate based on the provided reference images. For another example, users only have a scribble image and want to generate a video beginning with the contents described by this scribble image, which is a scribble-based video extension task. The context frames are \\(\\{u\\} +\\{0_{h\\times w}\\} \\times (n - 1)\\) and the context masks are \\(\\{1_{h\\times w}\\} \\times n\\). In this way, we can achieve multi-condition and reference control generation for long videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.326, + 0.649, + 0.34 + ], + "angle": 0, + "content": "3.3. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.484 + ], + "angle": 0, + "content": "We restructure the DiT model for VACE, as shown in Fig. 3, aiming to support multimodal VCU inputs. Since there is an existing pipeline for text tokenization, we only consider the tokenization of context frames and masks. After being tokenized, the context tokens are combined with noisy video tokens and used to fine-tune the DiT model. Different from that, we also propose a Context Adapter Tuning strategy, which allows context tokens to pass through Context Blocks and be added back to the original DiT Blocks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.49, + 0.71, + 0.503 + ], + "angle": 0, + "content": "3.3.1. Context Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Concept Decoupling. Two different visual concepts of natural video and control signals, like depth, pose, are encoded in \\( F \\) simultaneously. We believe that explicitly separating these data of different modalities and distributions is essential for model convergence. The concept decoupling is based on masks and yields two frame sequences identical in shape: \\( F_{c} = F \\times M \\) and \\( F_{k} = F \\times (1 - M) \\), where \\( F_{c} \\) is called the reactive frames contain all the pixels to be changed, while all the pixels to be kept are stored in \\( F_{k} \\), named inactive frames. Specifically, the reference images and the unchanged part of V2V and MV2V go to \\( F_{k} \\), while control signals and those pixels about to change, such as gray pixels, are collected to \\( F_{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Context Latent Encoding. A typical DiT processes noisy video latents \\(X \\in \\mathbb{R}^{n' \\times h' \\times w' \\times d}\\), where \\(n'\\), \\(h'\\), and \\(w'\\) are the temporal and spatial shapes of the latent space. Similar to \\(X\\), \\(F_{c}\\), \\(F_{k}\\), and \\(M\\) need to be encoded into a high-dimensional feature space to ensure the property of significant spatiotemporal correlations. Therefore, we reorganize them together with \\(X\\) into a hierarchical and spatiotemporal aligned visual features. \\(F_{c}\\), \\(F_{k}\\) are processed by video VAE and mapped into the same latent space of \\(X\\), maintaining their spatiotemporal consistency. To avoid any unexpected mishmash of images and videos, reference images are separately encoded by VAE encoder and concatenated back along the temporal dimension, while the correspond" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17194" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "ing parts need to be removed during decoding. \\(M\\) is directly reshaped and interpolated. After that, \\(F_{c}, F_{k}\\), and \\(M\\) are all mapping into latent spaces and are spatiotemporal aligned with \\(X\\) in the shape of \\(n' \\times h' \\times w'\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.257 + ], + "angle": 0, + "content": "Context Embedder. We extend the embedder layer by concatenating \\( F_{c} \\), \\( F_{k} \\), and \\( M \\) in the channel dimension and tokenizing them into context tokens, which we refer to as the Context Embedder. The corresponding weights to tokenize \\( F_{c} \\) and \\( F_{k} \\) are directly copied from the original video embedder, and the weights to tokenize \\( M \\) are initialized to zeros." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.468, + 0.28 + ], + "angle": 0, + "content": "3.3.2. Fully Fine-Tuning and Context Adapter Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.569 + ], + "angle": 0, + "content": "To achieve training with VCU as input, a simple methodology is to fully fine-tune the whole DiT model, as shown in Fig. 3a. Context tokens are added together with noisy tokens \\( X \\), and all the parameters in DiT and the newly introduced Context Embedder will be updated during training. To avoid fully fine-tuning and achieve faster convergence, as well as to establish a pluggable feature with the foundation model, we also propose another methodology processing the context token in a Res-Tuning [29] manner, as shown in Fig. 3b. Particularly, we select and copy several Transformer Blocks from the original DiT, forming a distributed and cascade-type Context Blocks. The original DiT processes video tokens and text tokens, while the newly added Transformer Blocks process context tokens and text tokens. The output of each Context Block is inserted back into the DiT blocks as an additive signal to assist the main branch in performing generation and editing tasks. In this manner, the parameters of DiT are frozen. Only the Context Embedder and Context Blocks are trainable." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.188, + 0.598 + ], + "angle": 0, + "content": "4. Datasets" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.27, + 0.622 + ], + "angle": 0, + "content": "4.1. Data Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To obtain an all-in-one model, the diversity and complexity of the required data construction also increase. Existing common text-to-video and image-to-video tasks only require constructing pairs of text and video. However, for the tasks in VACE, the modalities need to be further expanded to include target videos, source videos, local masks, reference, and more. To efficiently and rapidly acquire data for various tasks, it is imperative to maintain video quality while also conducting instance-level analysis and understanding of the video data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To achieve this, we first analyze the video data itself by performing shot slicing and preliminarily filtering out data based on resolution, aesthetic score, and motion amplitude. Next, we label the first frame of the video using RAM [78] and combine it with Grounding DINO [36] for detection, utilizing the localization results to perform secondary filtering on videos with target areas that are either too small or too large. Furthermore, we employ the propagate operation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "of SAM2 [52] for video segmentation to obtain instance-level information across the video. Leveraging the results of video segmentation, we filter instances in the temporal dimension by calculating the effective frame ratio based on the mask area threshold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.907, + 0.501 + ], + "angle": 0, + "content": "In the actual training process, the construction for different tasks also needs to be tailored according to the characteristics of each task: 1) For some controllable video generation tasks, we pre-extract depth [51], scribble [6], pose [5, 72], and optical flow [65] from the filtered videos. For gray and layout tasks, we create data on the fly. 2) For repainting tasks, random instances from the videos can be masked for inpainting, while the inverse of the mask enables the construction of outpainting data. Augmentation of the masks [62] allows for unconditional inpainting. 3) In the case of extension tasks, we extract key frames such as the first frame, last frame, frames from both ends, random frames, and segments from both ends to support a wider variety of extension types. 4) For reference tasks, we can extract several face or object instances from the videos and apply offline or online augmentation operations to create paired data. Notably, we randomly combine all the previously mentioned tasks for training to accommodate a broader range of model application scenarios. Additionally, for all operations involving masks, we perform arbitrary augmentation to satisfy various granular local generation requirements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.514, + 0.694, + 0.529 + ], + "angle": 0, + "content": "4.2. VACE-Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Significant progress has been made in the field of video generation. However, a scientific and thorough evaluation of the performance of these models remains an urgent issue that needs to be addressed. VBench [27] and VBench++ [28] have established a precise evaluation framework for text-to-video and image-to-video tasks through an extensive assessment suite and dimensional design. Nevertheless, as the video generation ecosystem continues to evolve, more derivative tasks have begun to emerge, such as video reference generation and video editing, for which a comprehensive benchmark is still lacking. To address this gap, we propose VACE-Benchmark to evaluate various downstream tasks related to video in a systematic manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Starting from the data sources, we recognize that real videos and generated videos may exhibit different performance characteristics during evaluation. Thus, we collected a total of 240 high-quality videos categorized by their sources, encompassing various data types, including text-to-video, inpainting, outpainting, extension, grayscale, depth, scribble, pose, optical flow, layout, reference face, and reference object tasks, with an average of 20 samples for each task. The input modalities include videos, masks, and references, and we also provide the original videos" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17195" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluations on VACE-Benchmark. We compare the automated score metrics of the unified VACE based on LTX-Video and the proprietary model on the dimensions of video quality and video consistency, as well as results of human user studies." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.13, + 0.915, + 0.539 + ], + "angle": 0, + "content": "
TypeMethodVideo Quality & Video ConsistencyUser Study
Aesthetic QualityBackground ConsistencyDynamic DegreeImaging QualityMotion SmoothnessOverall ConsistencySubject ConsistencyTemporal FlickeringNormalized AveragePrompt FollowingTemporal ConsistencyVideo QualityAverage
I2VI2VGenXL [77]55.20%92.87%60.00%63.31%97.43%23.78%89.58%95.67%71.54%2.651.602.342.20
CogVideoX-I2V [73]57.78%94.80%40.00%68.23%98.69%24.38%93.84%97.84%73.66%3.302.283.192.92
LTX-Video [22]56.12%94.57%35.00%62.72%99.27%24.92%92.83%98.41%72.89%2.952.282.282.50
VACE (Ours)57.53%95.32%45.00%68.03%99.08%25.13%93.61%97.80%74.38%3.204.002.543.24
InpaintProPainter [82]44.70%95.64%50.00%61.57%99.01%18.48%92.99%98.47%70.15%2.354.002.993.11
VACE (Ours)51.30%96.30%50.00%60.39%99.12%21.12%94.59%98.21%72.05%2.404.002.603.00
OutpaintFollow-Your-Canvas [8]53.30%95.99%5.00%69.53%98.08%25.90%95.38%97.20%71.54%3.052.001.632.23
M3DDM [17]53.34%95.87%30.00%65.07%99.22%25.43%93.65%98.85%73.16%3.703.882.283.29
VACE (Ours)57.04%96.55%30.00%69.49%99.20%25.36%94.47%98.47%74.25%3.903.923.583.80
DepthControl-A-Video [10]50.62%91.71%70.00%67.76%97.58%24.48%88.10%96.58%72.35%2.702.281.542.17
VideoComposer [68]50.03%94.18%70.00%59.44%96.23%24.95%89.79%94.38%70.74%2.602.442.172.40
ControlVideo [79]63.30%95.02%10.00%65.13%96.49%24.20%92.29%95.42%70.07%2.552.501.822.29
VACE (Ours)56.72%96.12%60.00%66.41%98.84%25.27%94.09%97.27%74.99%3.103.922.663.23
PoseText2Video-Zero [31]57.63%87.67%100.00%70.74%79.65%23.94%84.82%76.57%59.69%2.152.001.882.01
ControlVideo [79]65.37%94.56%25.00%65.28%97.32%25.19%92.76%96.82%72.45%2.151.802.031.99
Follow-Your-Pose [40]48.79%86.80%100.00%67.41%90.12%26.10%80.18%88.02%66.43%2.002.601.582.06
VACE (Ours)60.17%94.92%75.00%64.71%98.63%26.44%94.82%96.60%76.13%2.953.962.633.18
FlowFLATTEN [14]56.23%95.80%70.00%61.65%97.86%26.23%93.94%96.17%74.42%3.502.403.193.03
VACE (Ours)55.76%96.07%75.00%65.37%98.98%25.89%94.63%96.93%75.90%2.903.752.603.08
R2VKling1.6 [1]62.13%96.04%85.00%69.27%99.38%27.82%93.79%97.79%78.81%4.224.103.804.04
Pika2.2 [49]62.48%96.79%65.00%69.87%99.37%26.02%95.93%98.90%77.87%4.003.853.873.91
Vidu2.0 [66]64.30%96.85%35.00%67.03%99.66%26.53%96.73%99.41%76.47%3.903.853.773.84
VACE (Ours)63.25%98.03%30.00%72.29%99.51%25.85%98.54%99.15%76.76%3.473.423.303.40
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.631 + ], + "angle": 0, + "content": "to enable further processing by developers based on the specific characteristics of each task. Regarding the data prompts, we supply the original captions of the videos for quantitative assessment, as well as rewritten prompts tailored to the specific tasks to evaluate the models' creativity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.643, + 0.224, + 0.66 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.283, + 0.684 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. VACE is trained based on Diffusion Transformers for text-to-video generation at various scales. It utilizes LTX-Video-2B [22] for faster generation, while Wan-T2V-14B [64] is used specifically for higher-quality outputs, supporting resolutions of up to 720p. The training employs a phased approach. Initially, we focus on foundational tasks such as inpainting and extension, which are considered modal complementary to the pre-trained text-to-video models. This includes the incorporation of masks and the learning of contextual generation in both spatial and temporal dimensions. Next, from a task expansion perspective, we progressively transition from single input reference frames to multiple input reference frames and from single tasks to composite tasks. Fi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.616 + ], + "angle": 0, + "content": "nally, we fine-tune the model's quality using higher-quality data and longer sequences. The input for model training accommodates arbitrary resolutions, dynamic durations, and variable frame rates to support diverse input needs of users." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Baselines. Our goal is to achieve the unification of video creation and editing tasks, and currently, there is no comparable all-in-one video generation model available, which leads us to focus our evaluation on comparing our general model with proprietary task-specific models. Moreover, due to the numerous tasks involved and the lack of open-sourced methods for many of them, we conduct our comparisons on models that are available either offline or online. Specifically for the tasks, we compare the following: 1) For the I2V task, we examine I2VGenXL [77], CogVideoX-I2V [73], and LTX-Video-I2V [22]; 2) In the repainting task, we compare the ProPainter [82] for removal inpainting, while Follow-Your-Canvas [8] and M3DDM [17] are compared for outpainting; 3) For controllable task, we use Control-A-Video [10], VideoComposer [68], and ControlVideo [79] under depth conditions, and compare Text2Video-Zero [31], ControlVideo [79], and Follow-Your-Pose [40] under pose conditions, as well as FLAT" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17196" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.094, + 0.244, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.167, + 0.226, + 0.178 + ], + "angle": 0, + "content": "Reference Anything" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.089, + 0.898, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.255, + 0.161, + 0.888, + 0.179 + ], + "angle": 0, + "content": "The elegant lady carefully selects bags in the boutique, and she shows the charm of a mature woman in a black slim dress with a pearl necklace. Holding a vintage-inspired brown leather handbag, she is carefully observing its craftsmanship and texture. The interior of the store." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.184, + 0.255, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.258, + 0.215, + 0.269 + ], + "angle": 0, + "content": "Move Anything" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.181, + 0.899, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.256, + 0.252, + 0.889, + 0.27 + ], + "angle": 0, + "content": "A young boy rises from his chair and walks briskly to the right side of the frame towards the edge of the sun-drenched frame, as if chasing a new adventure. His eyes were bright, and the corners of his mouth were slightly upturned, revealing curiosity and excitement about the unknown..." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.272, + 0.255, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.349, + 0.223, + 0.359 + ], + "angle": 0, + "content": "Animate Anything" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.272, + 0.899, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.256, + 0.343, + 0.873, + 0.361 + ], + "angle": 0, + "content": "\"Anime-style, hot-blooded teenager in bright orange long-sleeved pants sportswear, standing on a surfboard, facing the golden sunshine in the rough sea. The teenager's short yellow hair is flying in the wind, his eyes are firm, and he has a confident smile on the corner of his mouth.\"" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.363, + 0.255, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.44, + 0.214, + 0.45 + ], + "angle": 0, + "content": "Swap Anything" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.363, + 0.898, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.256, + 0.434, + 0.886, + 0.452 + ], + "angle": 0, + "content": "The video shows an old movie-style scene in retro tones, with a little penguin and a kitten having a joyous bike race. Little penguins and kittens, both dressed in orange-and-red race suits, ride vintage multi-wheeled bikes on a nostalgic dirt road flanked by spectators..." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.463, + 0.256, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.527, + 0.22, + 0.538 + ], + "angle": 0, + "content": "Expand Anything" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.453, + 0.898, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.256, + 0.525, + 0.896, + 0.544 + ], + "angle": 0, + "content": "\"In the style of classical oil painting, the background is a river, and in the center of the picture is a mature and elegant woman, wearing a long skirt and sitting on a chair. She took the open red heart-shaped sunglasses from her arms with both hands and put them on...\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.557, + 0.905, + 0.572 + ], + "angle": 0, + "content": "Figure 4. Visualization results of compositional tasks. VACE creatively enables reference-, move-, animate-, swap-, and expand-anything." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.658 + ], + "angle": 0, + "content": "TEN [14] under optical flow conditions; 4) In reference generation, given the absence of open-sourced models, we compare commercial products Kling1.6 [1], Pika2.2 [49], and Vidu2.0 [66]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation. To comprehensively evaluate the performance of various tasks, we employ the VACE-Benchmark for assessment. Specifically, we divide the evaluation into automatic scoring and a user study for manual assessment. For the automatic scoring, we utilize select metrics from VBench [27] to assess video quality and video consistency, including eight indicators: aesthetic quality, background consistency, dynamic degree, imaging quality, motion smoothness, overall consistency, subject consistency, and temporal flickering. For the manual assessment, we utilize the mean opinion score (MOS) as our evaluation metric, focusing on three aspects: prompt following, temporal consistency, and video quality. In practice, we anonymize the generated data and randomly distribute it to different participants for scoring on a scale from 1 to 5." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.597, + 0.652, + 0.612 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantitative Evaluation. We compare VACE comprehensive model based on LTX-Video with task proprietary approaches on VACE-Benchmark. For certain tasks, we follow existing methods; for example, although we support generating based on any frame, we conduct comparisons using the first-frame reference approach from current open-source methods to ensure fairness. From Tab. 2, we can see that for the tasks of I2V, inpainting, outpainting, depth, pose, and optical flow, our method demonstrates better performance than other open-source methods across eight indicators of video quality and video consistency, with normalized average metrics showing superior results. Some competing methods can only generate at a resolution of 256, have very short generation durations, and exhibit instability in temporal coherence, resulting in poorer performance on automatic metric calculations. For the R2V task, there is still a certain gap in metrics compared to commercial models for a small-scale model that aims for fast generation," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17197" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.287, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.215, + 0.262, + 0.226 + ], + "angle": 0, + "content": "(a) Base structure setting." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.097, + 0.475, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.215, + 0.458, + 0.226 + ], + "angle": 0, + "content": "(b) Hyperparameter settings." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.228, + 0.285, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.346, + 0.285, + 0.358 + ], + "angle": 0, + "content": "(c) Context adapter configurations." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.228, + 0.474, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.305, + 0.346, + 0.46, + 0.358 + ], + "angle": 0, + "content": "(d) Concept decouple setting." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Figure 5. Ablation Studies of the VACE regarding structures, hyperparameters, and module configurations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.484 + ], + "angle": 0, + "content": "while being comparable to the metrics of Vidu 2.0. According to the results of human user studies, our method consistently performs better in evaluation metrics across multiple tasks, aligning well with user preferences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 1, we present the results of the VACE single model across various tasks. It is evident that the model achieves a high level of performance in video quality and temporal consistency. Furthermore, in composition tasks shown in Fig. 4, our model showcases impressive abilities, effectively integrating different modalities and tasks to produce results that cannot be generated by existing single or multiple models, thereby demonstrating its strong potential in the fields of video generation and editing. For example, in the \"Move Anything\" case, by providing a single input image and a movement trajectory, we are able to precisely move the characters in the scene with a specified direction while maintaining coherence and narrative consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.255, + 0.725 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.869 + ], + "angle": 0, + "content": "To better understand the impact of different independent modules on a unified video generation framework, we conducted a series of systematic comparative experiments based on the LTX-Video model to achieve a better model structure and configuration. To accurately assess the different experimental settings, we sample 250 data points for each task as a validation set and calculate the training loss, reflecting the model's training progress through the mean curve changes of different tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Base Structure. Text-guided image or video generation models only take noise as inference input. When extended" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "to our unified input paradigm, VCU, we can conduct training using fully fine-tuning or by incorporating additional parameter fine-tuning. Specifically, as shown in Fig. 5a, we compare the concatenation of different inputs along the channel dimension and modify the input dimensions of the patchify projection layer to achieve the loading and fully fine-tuning of the pre-trained model. Additionally, we introduce some extra training parameters in the form of RestTuning [29], which serializes VCU in a bypass branch and injects information into the main branch. The results indicate that both methods yielded similar effects; however, since the additional parameter fine-tuning converges faster, we base our subsequent experiments on this approach. As shown in Fig. 5b, we further conduct hyperparameter experiments based on this structure, focusing on aspects such as weighting schemes, timestamp shifting, and p-zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Context Adapter. Since the number of context blocks will significantly affect the model size and inference time consumption, we attempt to find an optimal number and distribution of context blocks. We begin with selecting continuous blocks at the input side and make comparisons between the first 1/4 blocks, 1/2 blocks, and all blocks. Inspired by the Res-Tuning [29] method, we also experiment with evenly distributing the injection blocks instead of selecting a continuous block series. As shown in Fig. 5c, we can see that when using the same number of blocks, the distributed arrangement of blocks outperforms the continuous arrangement in shallow blocks. Furthermore, a greater number of blocks generally yields better results, but due to the limited improvement in effectiveness and the constraints of training resources, we adopt a partially distributed arrangement of blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.651 + ], + "angle": 0, + "content": "Concept Decouple. During training, we introduce a Concept Decouple processing module to further disassemble the visual units, clarifying what content the model needs to learn to modify or retain. As shown in Fig. 5d, using this module results in a more significant reduction in loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This paper introduces VACE, an all-in-one video generation and editing framework. It unifies the diverse and complex multimodal inputs required for various video tasks, bridging the gap between specialized models for each individual task. This enables most video AI creation tasks to be completed with a single inference of a single model. While broadly covering various video tasks, VACE also supports flexible and free combinations of these tasks, greatly expanding the application scenarios of video generation models and meeting a wide range of user creative needs. The VACE framework paves the way for the development of unified visual generative models with multimodal inputs and represents a significant milestone in the field of visual generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.192 + ], + "angle": 0, + "content": "Acknowledgments. We would like to express our sincere appreciation for the contributions of many colleagues for their insightful discussions, valuable suggestions, and constructive feedback, including: Yuwei Wang, Haiming Zhao, Chenwei Xie, and Sheng Yao for their data contributions, and Shiwei Zhang, Tao Fang, and Xiang Wang for their discussions and suggestions." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.22, + 0.188, + 0.235 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.482, + 0.271 + ], + "angle": 0, + "content": "[1] KLING AI. KLING AI, https://klingai.com/, 2025.6,7,13" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[2] Runway AI. Stable Diffusion v1.5 Model Card, https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[3] Runway AI. Stable Diffusion Inpainting Model Card, https://huggingface.co/runwayml/stable-diffusion-inpainting, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning To Follow Image Editing Instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[6] Caroline Chan, Frédo Durand, and Phillip Isola. Learning To Generate Line Drawings That Convey Geometry and Semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-\\(\\alpha\\): Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[8] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-Your-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation. In Assoc. Adv. Artif. Intell., 2025. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[9] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, Ting-Che Lin, Shilong Zhang, Fu Li, Chuan Li, Xing Wang, Yanghua Peng, Peize Sun, Ping Luo, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Goku: Flow Based Video Generative Foundation Models. arXiv preprint arXiv:2502.04896, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable Text-to-Video Diffusion Models with Motion Prior and Reward Feedback Learning. arXiv preprint arXiv:2305.13840, 2023. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. AnyDoor: Zero-shot Object-level Image Customization. arXiv preprint arXiv:2307.09481, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[12] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang Zhao. UniReal: Universal Image Generation and Editing via Learning Real-world Dynamics. arXiv preprint arXiv:2412.07774, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.903, + 0.204 + ], + "angle": 0, + "content": "[13] Alibaba Cloud. Tongyi Wanxiang, https://tongyi. aliyun.com/wanxiang, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: Optical FLOW-guided ATTENtion for consistent text-to-video editing. In Int. Conf. Learn. Represent., 2024. 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[15] Lunhao Duan, Shanshan Zhao, Wenjun Yan, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Mingming Gong, and Gui-Song Xia. UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation. arXiv preprint arXiv:2412.18928, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.458 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Int. Conf. Mach. Learn., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[17] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical Masked 3D Diffusion Model for Video Outpainting. In ACM Int. Conf. Multimedia, pages 7890-7900, 2023. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.904, + 0.558 + ], + "angle": 0, + "content": "[18] FLUX. FLUX, https://blackforestlabs.ai/, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[19] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. SEED-Data-Edit Technical Report: A Hybrid Dataset for Instructional Image Editing. arXiv preprint arXiv:2405.04007, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[20] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, Haibin Huang, and Chongyang Ma. 12V-Adapter: A General Image-to-Video Adapter for Diffusion Models. In ACM SIGGRAPH, pages 1-12, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[21] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Adv. Neural Inform. Process. Syst., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[22] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion. arXiv preprint arXiv:2501.00103, 2025. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.898 + ], + "angle": 0, + "content": "[23] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer. In Int. Conf. Learn. Represent., 2025. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17199" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In Adv. Neural Inform. Process. Syst., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Adv. Neural Inform. Process. Syst. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.163, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[26] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and Controllable Image Synthesis with Composable Conditions. In Int. Conf. Mach.Learn., 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.219, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[27] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive Benchmark Suite for Video Generative Models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21807-21818, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.316, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[28] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, YingCong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench++: Comprehensive and Versatile Benchmark Suite for Video Generative Models. arXiv preprint arXiv:2411.13503, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.414, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[29] Zeyinzi Jiang, Chaojie Mao, Ziyuan Huang, Ao Ma, Yiliang Lv, Yujun Shen, Deli Zhao, and Jingren Zhou. Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone. In Adv. Neural Inform. Process. Syst., 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.484, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8995-9004, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.553, + 0.483, + 0.623 + ], + "angle": 0, + "content": "[31] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In Int. Conf. Comput. Vis., pages 15954–15964, 2023. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.623, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[32] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyuan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models. arXiv preprint arXiv:2412.03603, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[33] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[34] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. MagicEdit: High-Fidelity and Temporally Coherent Video Editing. arXiv preprint arXiv:2308.14749, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[35] Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv preprint arXiv:2502.11079, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.206, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. arXiv preprint arXiv:2303.05499, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.276, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[37] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-P2P: Video Editing with Cross-attention Control. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8599–8608, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.332, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[38] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept Neurons in Diffusion Models for Customized Generation. In Int. Conf. Mach. Learn., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.389, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[39] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable Image Synthesis with Multiple Subjects. In Adv. Neural Inform. Process. Syst., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.446, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[40] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Ying Shan, Xiu Li, and Qifeng Chen. Follow Your Pose: Pose-Guided Text-to-Video Generation using Pose-Free Videos. In Assoc. Adv. Artif. Intell., 2024. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.502, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[41] Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. ACE++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling. arXiv preprint arXiv:2501.02487, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.559, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[42] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.616, + 0.904, + 0.642 + ], + "angle": 0, + "content": "[43] Midjourney. Midjourney, https://www.midjourney.com, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.643, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[44] MiniMax. Hailuo AI Video, https://hailuoai.com/video, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.672, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[45] OpenAI. DALL·E 3, https://openai.com/dall-e-3, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.702, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[46] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold. In ACM SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.759, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[47] Yulin Pan, Chaojie Mao, Zeyinzi Jiang, Zhen Han, and Jingfeng Zhang. Locate, Assign, Refine: Taming Customized Image Inpainting with Text-Subject Guidance. arXiv preprint arXiv:2403.19534, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.814, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[48] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In Int. Conf. Comput. Vis., pages 4195-4305, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.858, + 0.866, + 0.872 + ], + "angle": 0, + "content": "[49] PiKa. PiKa, https://pika.art/, 2025.6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17200" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild. In Adv. Neural Inform. Process. Syst., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. IEEE Trans. Pattern Anal. Mach. Intell., pages 1623-1637, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. In Int. Conf. Learn. Represent., 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. Med. Image Comput. Computer-Assisted Interv., 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[55] Runway. Gen-3, https://app.runwayml.com/video-tools, 2025.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In Int. Conf. Learn. Represent., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[57] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[58] StabilityAI. Stable Diffusion v2-1 Model Card, https: //huggingface.co/stabilityai/stable-diffusion-2-1,2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[59] StabilityAI. Stable Diffusion XL Model Card, https: //huggingface.co/stabilityai/stable-diffusion-x1-base-1.0,2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[60] StabilityAI. CosXL Model Card, https://huggingface.co/stabilityai/cosxl, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[61] Ya Sheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation. In Adv. Neural Inform. Process. Syst., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[62] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-Robust Large Mask Inpainting With Fourier Convolutions. In IEEE Winter Conf. Appl. Comput. Vis., pages 2149-2159, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[63] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. OminiControl: Minimal and Universal Control for Diffusion Transformer. arXiv preprint arXiv:2411.15098, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[64] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2, 6, 13" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[65] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In Eur. Conf. Comput. Vis., pages 402-419, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.89, + 0.149 + ], + "angle": 0, + "content": "[66] Vidu. Vidu, https://www.vidu.cn/, 2025. 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[67] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[68] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In Adv. Neural Inform. Process. Syst., 2023. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A Unified and Flexible Motion Controller for Video Generation. In ACM SIGGRAPH, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[70] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, Yingya Zhang, and Hongming Shan. DreamVideo-2: Zero-Shot Subject-Driven Video Customization with Precise Motion Control. arXiv preprint arXiv:2410.13830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[71] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. OmniGen: Unified Image Generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[72] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective Whole-body Pose Estimation with Two-stages Distillation. In Int. Conf. Comput. Vis., pages 4210-4220, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. In Int. Conf. Learn. Represent., 2025. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[74] Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-Preserving Text-to-Video Generation by Frequency Decomposition. In IEEE Conf. Comput. Vis. Pattern Recog., 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[75] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. In Adv. Neural Inform. Process. Syst., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In Int. Conf. Comput. Vis., pages 3836–3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[77] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv preprint arXiv:2311.04145, 2023. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[78] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17201" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize Anything: A Strong Image Tagging Model. arXiv preprint arXiv:2306.03514, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[79] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo: Training-free Controllable Text-to-Video Generation. In Int. Conf. Learn. Represent., 2024. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[80] Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. Magic Mirror: ID-Preserved Video Generation in Video Diffusion Transformers. arXiv preprint arXiv:2411.13503, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[81] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based Fine-Grained Image Editing at Scale. arXiv preprint arXiv:2407.05282v1, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[82] Shangchen Zhou, Chongyi Li, Kelvin C. K. Chan, and Chen Change Loy. ProPainter: Improving Propagation and Transformer for Video Inpainting. In Int. Conf. Comput. Vis., pages 10477-10486, 2023. 2, 6, 13" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17202" + } + ] +] \ No newline at end of file diff --git a/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_origin.pdf b/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dbde8f9e7868f54c0169b1015bc6c05d1aba7601 --- /dev/null +++ b/2025/VACE_ All-in-One Video Creation and Editing/bb0c9cf5-4f56-4fe0-bdd0-bf9454fc4819_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b659ea208919e8bc322093dde0fe027f5c2956ee57f56e559fe1aad0dedca67 +size 5291197 diff --git a/2025/VACE_ All-in-One Video Creation and Editing/full.md b/2025/VACE_ All-in-One Video Creation and Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..20c9c48208de08e62975596e471c8e297ee55293 --- /dev/null +++ b/2025/VACE_ All-in-One Video Creation and Editing/full.md @@ -0,0 +1,343 @@ +# VACE: All-in-One Video Creation and Editing + +Zeyinzi Jiang* + +Zhen Han* + +Chaojie Mao*† + +Jingfeng Zhang + +Yulin Pan + +Yu Liu + +# Tongyi Lab, Alibaba Group + +![](images/6dc0ac2e79e0ba9bb358ce5358c79ab788f0b963e087806e235c7b43592a9f61.jpg) +Reference-to-video generation + +![](images/01187571914a5e6d6532ea027270a9ab6354e126bbeba8577fb665a9462afeee.jpg) + +![](images/85545d4d442ad80243ffb8b78dc40cee7e78444acc8b7f750789189b0c03bbf4.jpg) + +![](images/6c9c116f5690eba3d9ee44b3b791916ff76bb2ec58621dbbe86a0ae44980aa0d.jpg) +Video-to-video editing +Pose Transfer (Pose Control) + +![](images/1d10ef66bf9e895e32d99c420bde8e9d16ef77d13a2666119f69ac5d62c76b41.jpg) +Motion Transfer (Layout Control) + +![](images/58ad5b44490705232a973cf7a586334b054d0bc8c494901a11fe885919a638d8.jpg) +Structure Transfer (Depth Control) + +![](images/80df6d6119b73f2d4005bce8b9c89505b8abb30c091bc04ee8329de809340e56.jpg) +Colorization (Gray Control) + +![](images/95d1d327420388d5737f00901b38c9c4081cca6ef3806a9ed2ec1ed8f43224ce.jpg) +Masked Video-to-video editing +Subject Recreation (Inpainting) + +![](images/0f9436c54eb21d2c2ced5d62e408773bb87a5fad5ad4a71c1a7bcecc0f3dff35.jpg) +Subject Removal (Inpainting) + +![](images/6e856a972006ce8b10c473c7deb34dd2b9911915774764a54da3b6adf70ffa9b.jpg) +Canvas Extension (Outpainting) + +![](images/f1285142f3ed6719d624eecacd52d7f643cbad1f1346811431897c717ae26fcc.jpg) +Temporal Extension (First Clip) + +![](images/5a8e8ec3fd8352db7fcaa2531b94053d11e401af6e453edc2f6ee0834d1ac198.jpg) +Task Composition +Swap Anything (Subject Reference + Inpainting) +Figure 1. Comprehensive capabilities of VACE. We present the outstanding generation results based on Wan2.1 (left) and LTX-Video (right). For each task, the original input image or video (left), the context video (top left corner), and the generated frames are illustrated. + +![](images/ca5b623143e9929a16ef37e21b1cf7268db1c9a956e3c85c812cdbf5bb037f27.jpg) +Animate Anything (Frame Reference + Pose Control) + +# Abstract + +Diffusion Transformer has demonstrated powerful capability and scalability in generating high-quality images and videos. Further pursuing the unification of generation and editing tasks has yielded significant progress in the domain of image content creation. However, due to the intrinsic demands for consistency across both temporal and spatial dynamics, achieving a unified approach for video synthesis remains challenging. We introduce VACE, which enables users to perform Video tasks within an All-in-one framework for Creation and Editing. These tasks include reference-to-video generation, video-to-video editing, and masked video-to-video editing. Specifically, we effectively + +integrate the requirements of various tasks by organizing video task inputs, such as editing, reference, and masking, into a unified interface referred to as the Video Condition Unit (VCU). Furthermore, by utilizing a Context Adapter structure, we inject different task concepts into the model using formalized representations of temporal and spatial dimensions, allowing it to handle arbitrary video synthesis tasks flexibly. Extensive experiments demonstrate that the unified model of VACE achieves performance on par with task-specific models across various subtasks. Simultaneously, it enables diverse applications through versatile task combinations. Project page: https://alivilab.github.io/VACE-Page/. + +# 1. Introduction + +In recent years, the domain of visual generation tasks has witnessed remarkable advancements, driven in particular by the rapid evolution of diffusion models [24, 25, 48, 53, 54, 56, 57]. Beyond the early foundational pre-trained models for text-to-image [7, 16, 33] or text-to-video [9, 22, 64] generation in the field, there has been a proliferation of downstream tasks and applications, such as painting [3, 82], editing [4, 42, 68, 70, 75], controllable generation [30, 76], frame reference generation [20, 73], and ID-referenced video synthesis [11, 35, 47, 74]. This array of developments highlights the dynamic and complex nature of the visual generation field. To enhance task flexibility and reduce the overhead associated with deploying multiple models, researchers have begun to focus on constructing unified model architectures [12, 63] (e.g., ACE [23, 41] and Omni-Gen [71]), aiming to integrate different tasks into a single image model, facilitating the creation of various application workflows while maintaining simplicity in usage. In the field of video, due to the collaborative transformations in both temporal and spatial dimensions, leveraging a unified model can present endless possibilities for video creation. However, leveraging diverse input modalities and ensuring spatiotemporal consistency are still challenging for unified video generation and editing. + +We propose VACE, an all-in-one model for video creation and editing that performs tasks including reference-to-video generation, video-to-video editing, masked video-to-video editing, and free composition of these tasks, as illustrated in Fig. 1. The aggregation of various capabilities reduces the costs of service deployment and user interaction, and by combining the capabilities of different tasks within a single model, it addresses challenges faced by existing video generation models, such as controllable generation of long videos, multi-condition and reference-based generation, and continuous video editing, thereby empowering users with greater creativity. To achieve this, we utilize the current mainstream Diffusion Transformers (DiTs) structure as the foundational video framework and pre-trained text-to-video generation models [22, 64], which provide better basic capabilities and scalability for handling long video sequences. Specifically, VACE takes into account the needs of different tasks during its construction and designs a unified interface, dubbed the Video Condition Unit (VCU), which integrates multiple modalities such as images or videos for editing, references, and masks. Furthermore, to differentiate the visual modality information in editing and reference tasks, we introduce the concept decoupling strategy, enabling the model to understand what aspects need to be retained and what should be modified. Meanwhile, by employing a pluggable Context Adapter structure, concepts from different tasks (e.g., the areas or ranges of editing or reference) are injected into the model through col + +laborative spatiotemporal representation, enabling it to possess the capability of adaptive processing for unified tasks. + +Due to the lack of existing multi-task benchmarks in video synthesis, we construct a dataset of 480 evaluation samples containing 12 different tasks, while evaluating the performance of the VACE unified model by comparing it with existing specialized models. Experimental results demonstrate that our framework exhibits sufficient competitiveness in both quantitative and qualitative analyses. To the best of our knowledge, we are the first all-in-one model based on the video DiT architecture that concurrently supports such a wide range of tasks. Notably, this innovative framework allows for the compositional expansion of base tasks, enabling the construction of scenarios such as long video re-rendering, which provides a versatile and efficient solution for video synthesis, opening new possibilities for user-side video content creation and editing. + +# 2. Related Work + +Visual Generation and Editing. With the rapid development of image [2, 7, 16, 18, 58, 59] and video [22, 32, 73, 77] generation models, they are being used to create high-quality visual content and are widely applied in fields such as advertising, film special effects, game development, and animation production [13, 43-45, 55]. Meanwhile, to meet the diverse needs of visual media production and to enhance efficiency and quality, precise generation and editing methods have emerged. Models are required to perform generative creation based on multimodal inputs, such as depth, structure, pose, scene, and characters. According to the purposes of the input conditions, we can categorize them into two types: editing of the input and concept-guided re-creation. A significant portion of the work, such as ControlNet [76], ControlVideo,Composer [26], VideoComposer [68], and SCEdit [30], focuses on single-condition editing and multi-condition composite editing based on temporal and spatial alignment conditions. Additionally, some works that focus on interactive local editing scenarios, such as DragGAN [46] and MagicBrush [75]. Methods that guide generation based on semantic information from the input, such as Cone [38], Cone2 [39], InstantID [67], and PuLID [21], can achieve conceptual understanding of the input and inject it into the model for creative purposes. + +Task-unified Visual Generative Model. As the complexity and diversity of user creations increase, relying solely on a single model or a complicated chain of multiple models can no longer provide a convenient and efficient path for implementing creative ideas. In image generation, a unified generation and editing framework has begun to emerge, allowing for more flexible creative approaches. Methods such as UltraEdit [81] and SEED-Data-Edit [19] provide general-purpose editing datasets, while techniques like In + +![](images/2bdea4dc0e5c25dd5529b7d9f6e57e9084415154530503123357e880e21f4386.jpg) +Figure 2. Task categories covered by VACE. Four basic tasks can be combined to create a vast number of possibilities. + +structPix2Pix [4], MagicBrush [61], and CosXL [60] offer general instruction-based editing features. Additionally, methods like UniControl [50] and UNIC-Adapter [15] have unified controllable generation. Further advancements have led to the development of ACE [23, 41], OmniGen [71], OmniControl [63], and UniReal [12], which expand the scope of tasks by providing flexible controllable generation, local editing, and reference-guided generation. In the video domain, due to the increased difficulty of generation, approaches often manifest as single-task single-model frameworks, offering capabilities for editing or reference generation, as seen in Video-P2P [37], MagicEdit [34], MotionCtrl [69], Magic Mirror [80], and Phantom [35]. VACE aims to fill the gap for a unified model within the video domain, providing possibilities for complex creative scenarios. + +# 3. Method + +VACE is designed as a multimodal-to-video generation model, where text, image, video, and mask are integrated into a unified conditioning input. To cover as many video generation and editing tasks as possible, we conduct in-depth research into existing tasks, then divide them into 4 categories according to their individual requirements of multimodal inputs. Without losing generality, we specifically design a novel multimodal input format for each category under a Video Condition Unit (VCU) paradigm. Finally, we restructure the DiT model for VCU inputs, making it a versatile model for a wide range of video tasks. + +# 3.1. Multimodal Inputs and Video Tasks. + +Despite existing video tasks varying in complex user inputs and ambitious creative goals, we found that most of their inputs can be fully represented in 4 modalities: text, image, video, and mask. Overall, as illustrated in Fig. 2, we group these video tasks into 5 categories based on their requirements of these four multimodal inputs. + +- Text-to-Video Generation (T2V) is a basic video creation task, and text is the only input. +- Reference-to-Video Generation (R2V) requires additional images as reference inputs, making sure that specified contents, such as subjects of faces, animals, and other objects, or video frames, appear in the generated video. +- Video-to-Video Editing (V2V) makes an entire change to a provided video, such as colorization, stylization, controllable generation, etc. We use video control types whose control signals can be represented and stored as RGB videos, including depth, grayscale, pose, scribble, optical flow, and layout; however, the method itself is not limited to these. +- Masked Video-to-Video Editing (MV2V) makes changes to an input video only within the provided 3D regions of interest (3D ROI), seamlessly blending in with the other unchanged regions, such as inpainting, outpainting, video extension, etc. We use an extra spatiotemporal mask to represent the 3D ROI. +- Task Composition includes all the compositional possibilities of the 4 types of video tasks above. + +# 3.2. Video Condition Unit + +We propose an input paradigm, Video Condition Unit (VCU), to unify diverse input conditions into textual input, frame sequence, and mask sequence. A VCU can be denoted as + +$$ +V = [ T; F; M ], \tag {1} +$$ + +where $T$ is a text prompt, while $F$ and $M$ are sequences of context video frames $\{u_1, u_2, \dots, u_n\}$ and masks $\{m_1, m_2, \dots, m_n\}$ respectively. Here, $u$ is in RGB space, normalized to $[-1, 1]$ and $m$ is binary, in which "1"s and "0"s symbolize where to edit or not. $F$ and $M$ are aligned both in spatial size $h \times w$ and temporal size $n$ . In T2V, no context frame or mask is required. To keep generality, we assign a default value $0_{h \times w}$ to each $u$ denoting empty input, and set every $m$ to $1_{h \times w}$ meaning that all these pixels + +![](images/2a2cdb8973973f507ef1873f7f8e7f67b97950d6faa55f40b8c825386eb1ca8e.jpg) +Figure 3. Overview of VACE Framework. Frames and masks are tokenized through Concept Decoupling, Context Latent Encode, and Context Embedder. To achieve training with VCU as input, we employ two strategies: (a) Fully Fine-tuning and (b) Context Adapter Tuning. The latter converges faster and supports pluggable features. + +![](images/572bd3aa43bbd8d28bf5ab05a3e7a83f3135cec10cac1bc8df624e2eb426a3ba.jpg) + +Table 1. The formal representation of frames $(F\mathrm{s})$ and masks $(M\mathrm{s})$ under the four basic tasks. Frames and masks are aligned spatially and temporally. + +
TasksFrames (Fs) & Masks (Ms)
T2VF = {0h×w} × n +M = {1h×w} × n
R2VF = {r1, r2, ..., rl} + {0h×w} × n +M = {0h×w} × l + {1h×w} × n
V2VF = {u1, u2, ..., un} +M = {1h×w} × n
MV2VF = {u1, u2, ..., un} +M = {m1, m2, ..., mn}
+ +are about to be re-generated. For R2V, additional reference frames $r_i$ are inserted in front of the default frame sequence, while all-zero masks $0_{h \times w}$ are inserted in front of the mask sequence. These all-zero masks mean that the corresponding frames should be kept unchanged. In V2V, the context frame sequence is the input video frames, and the context mask is a sequence of $1_{h \times w}$ . For MV2V, both context video and mask are required. The formal mathematical representations are shown in Tab. 1. + +VCU can also support task composition. For example, the context frames of reference-inpainting task are $\{r_1,r_2,\dots,r_l,u_1,u_2,\dots,u_n\}$ and the context masks are $\{0_{h\times w}\} \times l + \{m_1,m_2,\dots,m_n\}$ . In this case, users can modify $l$ objects in the video and regenerate based on the provided reference images. For another example, users only have a scribble image and want to generate a video beginning with the contents described by this scribble image, which is a scribble-based video extension task. The context frames are $\{u\} +\{0_{h\times w}\} \times (n - 1)$ and the context masks are $\{1_{h\times w}\} \times n$ . In this way, we can achieve multi-condition and reference control generation for long videos. + +# 3.3. Architecture + +We restructure the DiT model for VACE, as shown in Fig. 3, aiming to support multimodal VCU inputs. Since there is an existing pipeline for text tokenization, we only consider the tokenization of context frames and masks. After being tokenized, the context tokens are combined with noisy video tokens and used to fine-tune the DiT model. Different from that, we also propose a Context Adapter Tuning strategy, which allows context tokens to pass through Context Blocks and be added back to the original DiT Blocks. + +# 3.3.1. Context Tokenization + +Concept Decoupling. Two different visual concepts of natural video and control signals, like depth, pose, are encoded in $F$ simultaneously. We believe that explicitly separating these data of different modalities and distributions is essential for model convergence. The concept decoupling is based on masks and yields two frame sequences identical in shape: $F_{c} = F \times M$ and $F_{k} = F \times (1 - M)$ , where $F_{c}$ is called the reactive frames contain all the pixels to be changed, while all the pixels to be kept are stored in $F_{k}$ , named inactive frames. Specifically, the reference images and the unchanged part of V2V and MV2V go to $F_{k}$ , while control signals and those pixels about to change, such as gray pixels, are collected to $F_{c}$ . + +Context Latent Encoding. A typical DiT processes noisy video latents $X \in \mathbb{R}^{n' \times h' \times w' \times d}$ , where $n'$ , $h'$ , and $w'$ are the temporal and spatial shapes of the latent space. Similar to $X$ , $F_{c}$ , $F_{k}$ , and $M$ need to be encoded into a high-dimensional feature space to ensure the property of significant spatiotemporal correlations. Therefore, we reorganize them together with $X$ into a hierarchical and spatiotemporal aligned visual features. $F_{c}$ , $F_{k}$ are processed by video VAE and mapped into the same latent space of $X$ , maintaining their spatiotemporal consistency. To avoid any unexpected mishmash of images and videos, reference images are separately encoded by VAE encoder and concatenated back along the temporal dimension, while the correspond + +ing parts need to be removed during decoding. $M$ is directly reshaped and interpolated. After that, $F_{c}, F_{k}$ , and $M$ are all mapping into latent spaces and are spatiotemporal aligned with $X$ in the shape of $n' \times h' \times w'$ . + +Context Embedder. We extend the embedder layer by concatenating $F_{c}$ , $F_{k}$ , and $M$ in the channel dimension and tokenizing them into context tokens, which we refer to as the Context Embedder. The corresponding weights to tokenize $F_{c}$ and $F_{k}$ are directly copied from the original video embedder, and the weights to tokenize $M$ are initialized to zeros. + +# 3.3.2. Fully Fine-Tuning and Context Adapter Tuning + +To achieve training with VCU as input, a simple methodology is to fully fine-tune the whole DiT model, as shown in Fig. 3a. Context tokens are added together with noisy tokens $X$ , and all the parameters in DiT and the newly introduced Context Embedder will be updated during training. To avoid fully fine-tuning and achieve faster convergence, as well as to establish a pluggable feature with the foundation model, we also propose another methodology processing the context token in a Res-Tuning [29] manner, as shown in Fig. 3b. Particularly, we select and copy several Transformer Blocks from the original DiT, forming a distributed and cascade-type Context Blocks. The original DiT processes video tokens and text tokens, while the newly added Transformer Blocks process context tokens and text tokens. The output of each Context Block is inserted back into the DiT blocks as an additive signal to assist the main branch in performing generation and editing tasks. In this manner, the parameters of DiT are frozen. Only the Context Embedder and Context Blocks are trainable. + +# 4. Datasets + +# 4.1. Data Construction + +To obtain an all-in-one model, the diversity and complexity of the required data construction also increase. Existing common text-to-video and image-to-video tasks only require constructing pairs of text and video. However, for the tasks in VACE, the modalities need to be further expanded to include target videos, source videos, local masks, reference, and more. To efficiently and rapidly acquire data for various tasks, it is imperative to maintain video quality while also conducting instance-level analysis and understanding of the video data. + +To achieve this, we first analyze the video data itself by performing shot slicing and preliminarily filtering out data based on resolution, aesthetic score, and motion amplitude. Next, we label the first frame of the video using RAM [78] and combine it with Grounding DINO [36] for detection, utilizing the localization results to perform secondary filtering on videos with target areas that are either too small or too large. Furthermore, we employ the propagate operation + +of SAM2 [52] for video segmentation to obtain instance-level information across the video. Leveraging the results of video segmentation, we filter instances in the temporal dimension by calculating the effective frame ratio based on the mask area threshold. + +In the actual training process, the construction for different tasks also needs to be tailored according to the characteristics of each task: 1) For some controllable video generation tasks, we pre-extract depth [51], scribble [6], pose [5, 72], and optical flow [65] from the filtered videos. For gray and layout tasks, we create data on the fly. 2) For repainting tasks, random instances from the videos can be masked for inpainting, while the inverse of the mask enables the construction of outpainting data. Augmentation of the masks [62] allows for unconditional inpainting. 3) In the case of extension tasks, we extract key frames such as the first frame, last frame, frames from both ends, random frames, and segments from both ends to support a wider variety of extension types. 4) For reference tasks, we can extract several face or object instances from the videos and apply offline or online augmentation operations to create paired data. Notably, we randomly combine all the previously mentioned tasks for training to accommodate a broader range of model application scenarios. Additionally, for all operations involving masks, we perform arbitrary augmentation to satisfy various granular local generation requirements. + +# 4.2. VACE-Benchmark + +Significant progress has been made in the field of video generation. However, a scientific and thorough evaluation of the performance of these models remains an urgent issue that needs to be addressed. VBench [27] and VBench++ [28] have established a precise evaluation framework for text-to-video and image-to-video tasks through an extensive assessment suite and dimensional design. Nevertheless, as the video generation ecosystem continues to evolve, more derivative tasks have begun to emerge, such as video reference generation and video editing, for which a comprehensive benchmark is still lacking. To address this gap, we propose VACE-Benchmark to evaluate various downstream tasks related to video in a systematic manner. + +Starting from the data sources, we recognize that real videos and generated videos may exhibit different performance characteristics during evaluation. Thus, we collected a total of 240 high-quality videos categorized by their sources, encompassing various data types, including text-to-video, inpainting, outpainting, extension, grayscale, depth, scribble, pose, optical flow, layout, reference face, and reference object tasks, with an average of 20 samples for each task. The input modalities include videos, masks, and references, and we also provide the original videos + +Table 2. Quantitative evaluations on VACE-Benchmark. We compare the automated score metrics of the unified VACE based on LTX-Video and the proprietary model on the dimensions of video quality and video consistency, as well as results of human user studies. + +
TypeMethodVideo Quality & Video ConsistencyUser Study
Aesthetic QualityBackground ConsistencyDynamic DegreeImaging QualityMotion SmoothnessOverall ConsistencySubject ConsistencyTemporal FlickeringNormalized AveragePrompt FollowingTemporal ConsistencyVideo QualityAverage
I2VI2VGenXL [77]55.20%92.87%60.00%63.31%97.43%23.78%89.58%95.67%71.54%2.651.602.342.20
CogVideoX-I2V [73]57.78%94.80%40.00%68.23%98.69%24.38%93.84%97.84%73.66%3.302.283.192.92
LTX-Video [22]56.12%94.57%35.00%62.72%99.27%24.92%92.83%98.41%72.89%2.952.282.282.50
VACE (Ours)57.53%95.32%45.00%68.03%99.08%25.13%93.61%97.80%74.38%3.204.002.543.24
InpaintProPainter [82]44.70%95.64%50.00%61.57%99.01%18.48%92.99%98.47%70.15%2.354.002.993.11
VACE (Ours)51.30%96.30%50.00%60.39%99.12%21.12%94.59%98.21%72.05%2.404.002.603.00
OutpaintFollow-Your-Canvas [8]53.30%95.99%5.00%69.53%98.08%25.90%95.38%97.20%71.54%3.052.001.632.23
M3DDM [17]53.34%95.87%30.00%65.07%99.22%25.43%93.65%98.85%73.16%3.703.882.283.29
VACE (Ours)57.04%96.55%30.00%69.49%99.20%25.36%94.47%98.47%74.25%3.903.923.583.80
DepthControl-A-Video [10]50.62%91.71%70.00%67.76%97.58%24.48%88.10%96.58%72.35%2.702.281.542.17
VideoComposer [68]50.03%94.18%70.00%59.44%96.23%24.95%89.79%94.38%70.74%2.602.442.172.40
ControlVideo [79]63.30%95.02%10.00%65.13%96.49%24.20%92.29%95.42%70.07%2.552.501.822.29
VACE (Ours)56.72%96.12%60.00%66.41%98.84%25.27%94.09%97.27%74.99%3.103.922.663.23
PoseText2Video-Zero [31]57.63%87.67%100.00%70.74%79.65%23.94%84.82%76.57%59.69%2.152.001.882.01
ControlVideo [79]65.37%94.56%25.00%65.28%97.32%25.19%92.76%96.82%72.45%2.151.802.031.99
Follow-Your-Pose [40]48.79%86.80%100.00%67.41%90.12%26.10%80.18%88.02%66.43%2.002.601.582.06
VACE (Ours)60.17%94.92%75.00%64.71%98.63%26.44%94.82%96.60%76.13%2.953.962.633.18
FlowFLATTEN [14]56.23%95.80%70.00%61.65%97.86%26.23%93.94%96.17%74.42%3.502.403.193.03
VACE (Ours)55.76%96.07%75.00%65.37%98.98%25.89%94.63%96.93%75.90%2.903.752.603.08
R2VKling1.6 [1]62.13%96.04%85.00%69.27%99.38%27.82%93.79%97.79%78.81%4.224.103.804.04
Pika2.2 [49]62.48%96.79%65.00%69.87%99.37%26.02%95.93%98.90%77.87%4.003.853.873.91
Vidu2.0 [66]64.30%96.85%35.00%67.03%99.66%26.53%96.73%99.41%76.47%3.903.853.773.84
VACE (Ours)63.25%98.03%30.00%72.29%99.51%25.85%98.54%99.15%76.76%3.473.423.303.40
+ +to enable further processing by developers based on the specific characteristics of each task. Regarding the data prompts, we supply the original captions of the videos for quantitative assessment, as well as rewritten prompts tailored to the specific tasks to evaluate the models' creativity. + +# 5. Experiments + +# 5.1. Experimental Setup + +Implementation Details. VACE is trained based on Diffusion Transformers for text-to-video generation at various scales. It utilizes LTX-Video-2B [22] for faster generation, while Wan-T2V-14B [64] is used specifically for higher-quality outputs, supporting resolutions of up to 720p. The training employs a phased approach. Initially, we focus on foundational tasks such as inpainting and extension, which are considered modal complementary to the pre-trained text-to-video models. This includes the incorporation of masks and the learning of contextual generation in both spatial and temporal dimensions. Next, from a task expansion perspective, we progressively transition from single input reference frames to multiple input reference frames and from single tasks to composite tasks. Fi + +nally, we fine-tune the model's quality using higher-quality data and longer sequences. The input for model training accommodates arbitrary resolutions, dynamic durations, and variable frame rates to support diverse input needs of users. + +Baselines. Our goal is to achieve the unification of video creation and editing tasks, and currently, there is no comparable all-in-one video generation model available, which leads us to focus our evaluation on comparing our general model with proprietary task-specific models. Moreover, due to the numerous tasks involved and the lack of open-sourced methods for many of them, we conduct our comparisons on models that are available either offline or online. Specifically for the tasks, we compare the following: 1) For the I2V task, we examine I2VGenXL [77], CogVideoX-I2V [73], and LTX-Video-I2V [22]; 2) In the repainting task, we compare the ProPainter [82] for removal inpainting, while Follow-Your-Canvas [8] and M3DDM [17] are compared for outpainting; 3) For controllable task, we use Control-A-Video [10], VideoComposer [68], and ControlVideo [79] under depth conditions, and compare Text2Video-Zero [31], ControlVideo [79], and Follow-Your-Pose [40] under pose conditions, as well as FLAT + +![](images/c7fa9ea987d1bb71f5051d3e521300faee221cee1c690174509157a93ceb2484.jpg) +Reference Anything + +![](images/f6416d4a007a14dd68aefc6bfd0f1fe6cb773c3b035d97636f24ebb9a0c54d90.jpg) +The elegant lady carefully selects bags in the boutique, and she shows the charm of a mature woman in a black slim dress with a pearl necklace. Holding a vintage-inspired brown leather handbag, she is carefully observing its craftsmanship and texture. The interior of the store. + +![](images/8434dacb2e4a7709e3f78e2229e213784dba4a91b4b1f46122d498181dd52856.jpg) + +![](images/7755152ca82d7f0fbc13b2b7b203bfcb16d5f3adeb3ff41ce7b58ef1ace9a95a.jpg) +A young boy rises from his chair and walks briskly to the right side of the frame towards the edge of the sun-drenched frame, as if chasing a new adventure. His eyes were bright, and the corners of his mouth were slightly upturned, revealing curiosity and excitement about the unknown... + +Move Anything +![](images/4a2a7dc2056bb9eefe68f4fd0f438de07109f6e820e51ec5c8399a3d85f38d3d.jpg) +"Anime-style, hot-blooded teenager in bright orange long-sleeved pants sportswear, standing on a surfboard, facing the golden sunshine in the rough sea. The teenager's short yellow hair is flying in the wind, his eyes are firm, and he has a confident smile on the corner of his mouth." + +![](images/2ddbe8eeb9b05ec19d551d6405830a378b3c62dc317578be5cdea44cf0957355.jpg) + +Animate Anything +![](images/04fcf82eeef0b0e26c17bc7b5f184eadced2d8d689a0a494b4537ebd63f4d9cb.jpg) +The video shows an old movie-style scene in retro tones, with a little penguin and a kitten having a joyous bike race. Little penguins and kittens, both dressed in orange-and-red race suits, ride vintage multi-wheeled bikes on a nostalgic dirt road flanked by spectators... + +![](images/a20893df72f446a06fe28fb383ae5b5f8fa37e8addd8069ae865846de816e160.jpg) + +![](images/5105cdd07923abce73ff9a89c0e5639e4918777310ce414986bec976cd654a25.jpg) +Swap Anything + +Expand Anything +Figure 4. Visualization results of compositional tasks. VACE creatively enables reference-, move-, animate-, swap-, and expand-anything. +![](images/9751db892a0a52597f8e5f5d005db3142b8067aa9d728a4fd80dd8f895ed60a5.jpg) +"In the style of classical oil painting, the background is a river, and in the center of the picture is a mature and elegant woman, wearing a long skirt and sitting on a chair. She took the open red heart-shaped sunglasses from her arms with both hands and put them on..." + +TEN [14] under optical flow conditions; 4) In reference generation, given the absence of open-sourced models, we compare commercial products Kling1.6 [1], Pika2.2 [49], and Vidu2.0 [66]. + +Evaluation. To comprehensively evaluate the performance of various tasks, we employ the VACE-Benchmark for assessment. Specifically, we divide the evaluation into automatic scoring and a user study for manual assessment. For the automatic scoring, we utilize select metrics from VBench [27] to assess video quality and video consistency, including eight indicators: aesthetic quality, background consistency, dynamic degree, imaging quality, motion smoothness, overall consistency, subject consistency, and temporal flickering. For the manual assessment, we utilize the mean opinion score (MOS) as our evaluation metric, focusing on three aspects: prompt following, temporal consistency, and video quality. In practice, we anonymize the generated data and randomly distribute it to different participants for scoring on a scale from 1 to 5. + +# 5.2. Main Results + +Quantitative Evaluation. We compare VACE comprehensive model based on LTX-Video with task proprietary approaches on VACE-Benchmark. For certain tasks, we follow existing methods; for example, although we support generating based on any frame, we conduct comparisons using the first-frame reference approach from current open-source methods to ensure fairness. From Tab. 2, we can see that for the tasks of I2V, inpainting, outpainting, depth, pose, and optical flow, our method demonstrates better performance than other open-source methods across eight indicators of video quality and video consistency, with normalized average metrics showing superior results. Some competing methods can only generate at a resolution of 256, have very short generation durations, and exhibit instability in temporal coherence, resulting in poorer performance on automatic metric calculations. For the R2V task, there is still a certain gap in metrics compared to commercial models for a small-scale model that aims for fast generation, + +![](images/cfdbd7dd56c30c3a6af8e7f6a707e50ca6f36261f186b92d725ce4882711d134.jpg) +(a) Base structure setting. + +![](images/271e1884078ac46369e29ea040268eec90e4a0725fcd2514331506d8f08d1a3c.jpg) +(b) Hyperparameter settings. + +![](images/95c6ecc97504c4d853d22a32ff59fd3c44365b48d077815be8a79fc91f3a7089.jpg) +(c) Context adapter configurations. + +![](images/14fdebd9ed061f7d0e1c6ae123885833ea7a70db27684d65351cc2fd5fe7e5d8.jpg) +(d) Concept decouple setting. +Figure 5. Ablation Studies of the VACE regarding structures, hyperparameters, and module configurations. + +while being comparable to the metrics of Vidu 2.0. According to the results of human user studies, our method consistently performs better in evaluation metrics across multiple tasks, aligning well with user preferences. + +Qualitative Results. In Fig. 1, we present the results of the VACE single model across various tasks. It is evident that the model achieves a high level of performance in video quality and temporal consistency. Furthermore, in composition tasks shown in Fig. 4, our model showcases impressive abilities, effectively integrating different modalities and tasks to produce results that cannot be generated by existing single or multiple models, thereby demonstrating its strong potential in the fields of video generation and editing. For example, in the "Move Anything" case, by providing a single input image and a movement trajectory, we are able to precisely move the characters in the scene with a specified direction while maintaining coherence and narrative consistency. + +# 5.3. Ablation Studies + +To better understand the impact of different independent modules on a unified video generation framework, we conducted a series of systematic comparative experiments based on the LTX-Video model to achieve a better model structure and configuration. To accurately assess the different experimental settings, we sample 250 data points for each task as a validation set and calculate the training loss, reflecting the model's training progress through the mean curve changes of different tasks. + +Base Structure. Text-guided image or video generation models only take noise as inference input. When extended + +to our unified input paradigm, VCU, we can conduct training using fully fine-tuning or by incorporating additional parameter fine-tuning. Specifically, as shown in Fig. 5a, we compare the concatenation of different inputs along the channel dimension and modify the input dimensions of the patchify projection layer to achieve the loading and fully fine-tuning of the pre-trained model. Additionally, we introduce some extra training parameters in the form of RestTuning [29], which serializes VCU in a bypass branch and injects information into the main branch. The results indicate that both methods yielded similar effects; however, since the additional parameter fine-tuning converges faster, we base our subsequent experiments on this approach. As shown in Fig. 5b, we further conduct hyperparameter experiments based on this structure, focusing on aspects such as weighting schemes, timestamp shifting, and p-zero. + +Context Adapter. Since the number of context blocks will significantly affect the model size and inference time consumption, we attempt to find an optimal number and distribution of context blocks. We begin with selecting continuous blocks at the input side and make comparisons between the first 1/4 blocks, 1/2 blocks, and all blocks. Inspired by the Res-Tuning [29] method, we also experiment with evenly distributing the injection blocks instead of selecting a continuous block series. As shown in Fig. 5c, we can see that when using the same number of blocks, the distributed arrangement of blocks outperforms the continuous arrangement in shallow blocks. Furthermore, a greater number of blocks generally yields better results, but due to the limited improvement in effectiveness and the constraints of training resources, we adopt a partially distributed arrangement of blocks. + +Concept Decouple. During training, we introduce a Concept Decouple processing module to further disassemble the visual units, clarifying what content the model needs to learn to modify or retain. As shown in Fig. 5d, using this module results in a more significant reduction in loss. + +# 6. Conclusion + +This paper introduces VACE, an all-in-one video generation and editing framework. It unifies the diverse and complex multimodal inputs required for various video tasks, bridging the gap between specialized models for each individual task. This enables most video AI creation tasks to be completed with a single inference of a single model. While broadly covering various video tasks, VACE also supports flexible and free combinations of these tasks, greatly expanding the application scenarios of video generation models and meeting a wide range of user creative needs. The VACE framework paves the way for the development of unified visual generative models with multimodal inputs and represents a significant milestone in the field of visual generation. + +Acknowledgments. We would like to express our sincere appreciation for the contributions of many colleagues for their insightful discussions, valuable suggestions, and constructive feedback, including: Yuwei Wang, Haiming Zhao, Chenwei Xie, and Sheng Yao for their data contributions, and Shiwei Zhang, Tao Fang, and Xiang Wang for their discussions and suggestions. + +# References + +[1] KLING AI. KLING AI, https://klingai.com/, 2025.6,7,13 +[2] Runway AI. Stable Diffusion v1.5 Model Card, https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022.2 +[3] Runway AI. Stable Diffusion Inpainting Model Card, https://huggingface.co/runwayml/stable-diffusion-inpainting, 2022.2 +[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning To Follow Image Editing Instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 2, 3 +[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 5 +[6] Caroline Chan, Frédo Durand, and Phillip Isola. Learning To Generate Line Drawings That Convey Geometry and Semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 5 +[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt- $\alpha$ : Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv preprint arXiv:2310.00426, 2023. 2 +[8] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-Your-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation. In Assoc. Adv. Artif. Intell., 2025. 6, 13 +[9] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, Ting-Che Lin, Shilong Zhang, Fu Li, Chuan Li, Xing Wang, Yanghua Peng, Peize Sun, Ping Luo, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Goku: Flow Based Video Generative Foundation Models. arXiv preprint arXiv:2502.04896, 2025. 2 +[10] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable Text-to-Video Diffusion Models with Motion Prior and Reward Feedback Learning. arXiv preprint arXiv:2305.13840, 2023. 6, 13 +[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. AnyDoor: Zero-shot Object-level Image Customization. arXiv preprint arXiv:2307.09481, 2023. 2 + +[12] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang Zhao. UniReal: Universal Image Generation and Editing via Learning Real-world Dynamics. arXiv preprint arXiv:2412.07774, 2024. 2, 3 +[13] Alibaba Cloud. Tongyi Wanxiang, https://tongyi. aliyun.com/wanxiang, 2023. 2 +[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: Optical FLOW-guided ATTENtion for consistent text-to-video editing. In Int. Conf. Learn. Represent., 2024. 6, 7, 13 +[15] Lunhao Duan, Shanshan Zhao, Wenjun Yan, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Mingming Gong, and Gui-Song Xia. UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation. arXiv preprint arXiv:2412.18928, 2024.3 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Int. Conf. Mach. Learn., 2024. 2 +[17] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical Masked 3D Diffusion Model for Video Outpainting. In ACM Int. Conf. Multimedia, pages 7890-7900, 2023. 6, 13 +[18] FLUX. FLUX, https://blackforestlabs.ai/, 2024.2 +[19] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. SEED-Data-Edit Technical Report: A Hybrid Dataset for Instructional Image Editing. arXiv preprint arXiv:2405.04007, 2024. 2 +[20] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, Haibin Huang, and Chongyang Ma. 12V-Adapter: A General Image-to-Video Adapter for Diffusion Models. In ACM SIGGRAPH, pages 1-12, 2024. 2 +[21] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Adv. Neural Inform. Process. Syst., 2024. 2 +[22] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion. arXiv preprint arXiv:2501.00103, 2025. 2, 6, 13 +[23] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer. In Int. Conf. Learn. Represent., 2025. 2, 3 + +[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In Adv. Neural Inform. Process. Syst., 2021. 2 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Adv. Neural Inform. Process. Syst. Curran Associates, Inc., 2020. 2 +[26] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and Controllable Image Synthesis with Composable Conditions. In Int. Conf. Mach.Learn., 2023.2 +[27] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive Benchmark Suite for Video Generative Models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21807-21818, 2024. 5, 7 +[28] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, YingCong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench++: Comprehensive and Versatile Benchmark Suite for Video Generative Models. arXiv preprint arXiv:2411.13503, 2024. 5 +[29] Zeyinzi Jiang, Chaojie Mao, Ziyuan Huang, Ao Ma, Yiliang Lv, Yujun Shen, Deli Zhao, and Jingren Zhou. Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone. In Adv. Neural Inform. Process. Syst., 2023. 5, 8 +[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8995-9004, 2024. 2 +[31] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In Int. Conf. Comput. Vis., pages 15954–15964, 2023. 6, 13 +[32] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyuan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models. arXiv preprint arXiv:2412.03603, 2024. 2 +[33] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 2 + +[34] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. MagicEdit: High-Fidelity and Temporally Coherent Video Editing. arXiv preprint arXiv:2308.14749, 2023. 3 +[35] Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv preprint arXiv:2502.11079, 2025. 2, 3 +[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. arXiv preprint arXiv:2303.05499, 2023. 5 +[37] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-P2P: Video Editing with Cross-attention Control. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8599–8608, 2024. 3 +[38] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept Neurons in Diffusion Models for Customized Generation. In Int. Conf. Mach. Learn., 2023. 2 +[39] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable Image Synthesis with Multiple Subjects. In Adv. Neural Inform. Process. Syst., 2023. 2 +[40] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Ying Shan, Xiu Li, and Qifeng Chen. Follow Your Pose: Pose-Guided Text-to-Video Generation using Pose-Free Videos. In Assoc. Adv. Artif. Intell., 2024. 6, 13 +[41] Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. ACE++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling. arXiv preprint arXiv:2501.02487, 2025. 2, 3 +[42] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2 +[43] Midjourney. Midjourney, https://www.midjourney.com, 2023.2 +[44] MiniMax. Hailuo AI Video, https://hailuoai.com/video, 2024. +[45] OpenAI. DALL·E 3, https://openai.com/dall-e-3, 2023.2 +[46] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold. In ACM SIGGRAPH, 2023. 2 +[47] Yulin Pan, Chaojie Mao, Zeyinzi Jiang, Zhen Han, and Jingfeng Zhang. Locate, Assign, Refine: Taming Customized Image Inpainting with Text-Subject Guidance. arXiv preprint arXiv:2403.19534, 2024. 2 +[48] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In Int. Conf. Comput. Vis., pages 4195-4305, 2023. 2 +[49] PiKa. PiKa, https://pika.art/, 2025.6, 7, 13 +[50] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming + +Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild. In Adv. Neural Inform. Process. Syst., 2023. 3 +[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. IEEE Trans. Pattern Anal. Mach. Intell., pages 1623-1637, 2022. 5 +[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. In Int. Conf. Learn. Represent., 2025. 5 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 2 +[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. Med. Image Comput. Computer-Assisted Interv., 2015. 2 +[55] Runway. Gen-3, https://app.runwayml.com/video-tools, 2025.2 +[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In Int. Conf. Learn. Represent., 2021. 2 +[57] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2 +[58] StabilityAI. Stable Diffusion v2-1 Model Card, https: //huggingface.co/stabilityai/stable-diffusion-2-1,2022.2 +[59] StabilityAI. Stable Diffusion XL Model Card, https: //huggingface.co/stabilityai/stable-diffusion-x1-base-1.0,2022.2 +[60] StabilityAI. CosXL Model Card, https://huggingface.co/stabilityai/cosxl, 2024.3 +[61] Ya Sheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation. In Adv. Neural Inform. Process. Syst., 2023. 3 +[62] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-Robust Large Mask Inpainting With Fourier Convolutions. In IEEE Winter Conf. Appl. Comput. Vis., pages 2149-2159, 2022. 5 +[63] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. OminiControl: Minimal and Universal Control for Diffusion Transformer. arXiv preprint arXiv:2411.15098, 2024. 2, 3 +[64] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2, 6, 13 + +[65] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In Eur. Conf. Comput. Vis., pages 402-419, 2020. 5 +[66] Vidu. Vidu, https://www.vidu.cn/, 2025. 6, 7, 13 +[67] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 2 +[68] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In Adv. Neural Inform. Process. Syst., 2023. 2, 6, 13 +[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A Unified and Flexible Motion Controller for Video Generation. In ACM SIGGRAPH, pages 1-11, 2024. 3 +[70] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, Yingya Zhang, and Hongming Shan. DreamVideo-2: Zero-Shot Subject-Driven Video Customization with Precise Motion Control. arXiv preprint arXiv:2410.13830, 2024. 2 +[71] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. OmniGen: Unified Image Generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3 +[72] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective Whole-body Pose Estimation with Two-stages Distillation. In Int. Conf. Comput. Vis., pages 4210-4220, 2023. 5 +[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. In Int. Conf. Learn. Represent., 2025. 2, 6, 13 +[74] Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-Preserving Text-to-Video Generation by Frequency Decomposition. In IEEE Conf. Comput. Vis. Pattern Recog., 2025. 2 +[75] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. In Adv. Neural Inform. Process. Syst., 2023. 2 +[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In Int. Conf. Comput. Vis., pages 3836–3847, 2023. 2 +[77] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv preprint arXiv:2311.04145, 2023. 2, 6, 13 +[78] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, + +Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize Anything: A Strong Image Tagging Model. arXiv preprint arXiv:2306.03514, 2023. 5 +[79] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo: Training-free Controllable Text-to-Video Generation. In Int. Conf. Learn. Represent., 2024. 6, 13 +[80] Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. Magic Mirror: ID-Preserved Video Generation in Video Diffusion Transformers. arXiv preprint arXiv:2411.13503, 2025. 3 +[81] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based Fine-Grained Image Editing at Scale. arXiv preprint arXiv:2407.05282v1, 2024. 2 +[82] Shangchen Zhou, Chongyi Li, Kelvin C. K. Chan, and Chen Change Loy. ProPainter: Improving Propagation and Transformer for Video Inpainting. In Int. Conf. Comput. Vis., pages 10477-10486, 2023. 2, 6, 13 \ No newline at end of file diff --git a/2025/VACE_ All-in-One Video Creation and Editing/images.zip b/2025/VACE_ All-in-One Video Creation and Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..28c76d7bbef8090a8380af00788ac14cc6c30f0a --- /dev/null +++ b/2025/VACE_ All-in-One Video Creation and Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d0088dcd57ebce66515463fd6ba5126521846337b34eb2a20fe627ffa904654 +size 824666 diff --git a/2025/VACE_ All-in-One Video Creation and Editing/layout.json b/2025/VACE_ All-in-One Video Creation and Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8dd27621dad46f7eead2416d87619889daca9453 --- /dev/null +++ b/2025/VACE_ All-in-One Video Creation and Editing/layout.json @@ -0,0 +1,9750 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 103, + 451, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 103, + 451, + 121 + ], + "spans": [ + { + "bbox": [ + 161, + 103, + 451, + 121 + ], + "type": "text", + "content": "VACE: All-in-One Video Creation and Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 143, + 167, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 143, + 167, + 158 + ], + "spans": [ + { + "bbox": [ + 94, + 143, + 167, + 158 + ], + "type": "text", + "content": "Zeyinzi Jiang*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 144, + 228, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 144, + 228, + 156 + ], + "spans": [ + { + "bbox": [ + 175, + 144, + 228, + 156 + ], + "type": "text", + "content": "Zhen Han*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 144, + 315, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 144, + 315, + 157 + ], + "spans": [ + { + "bbox": [ + 242, + 144, + 315, + 157 + ], + "type": "text", + "content": "Chaojie Mao*†" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 144, + 405, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 144, + 405, + 157 + ], + "spans": [ + { + "bbox": [ + 328, + 144, + 405, + 157 + ], + "type": "text", + "content": "Jingfeng Zhang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 418, + 144, + 467, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 144, + 467, + 156 + ], + "spans": [ + { + "bbox": [ + 418, + 144, + 467, + 156 + ], + "type": "text", + "content": "Yulin Pan" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 480, + 144, + 515, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 480, + 144, + 515, + 156 + ], + "spans": [ + { + "bbox": [ + 480, + 144, + 515, + 156 + ], + "type": "text", + "content": "Yu Liu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 237, + 165, + 373, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 165, + 373, + 180 + ], + "spans": [ + { + "bbox": [ + 237, + 165, + 373, + 180 + ], + "type": "text", + "content": "Tongyi Lab, Alibaba Group" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 198, + 298, + 234 + ], + "blocks": [ + { + "bbox": [ + 254, + 190, + 356, + 198 + ], + "lines": [ + { + "bbox": [ + 254, + 190, + 356, + 198 + ], + "spans": [ + { + "bbox": [ + 254, + 190, + 356, + 198 + ], + "type": "text", + "content": "Reference-to-video generation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 198, + 298, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 198, + 298, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 198, + 298, + 234 + ], + "type": "image", + "image_path": "6dc0ac2e79e0ba9bb358ce5358c79ab788f0b963e087806e235c7b43592a9f61.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 332, + 200, + 369, + 232 + ], + "blocks": [ + { + "bbox": [ + 332, + 200, + 369, + 232 + ], + "lines": [ + { + "bbox": [ + 332, + 200, + 369, + 232 + ], + "spans": [ + { + "bbox": [ + 332, + 200, + 369, + 232 + ], + "type": "image", + "image_path": "01187571914a5e6d6532ea027270a9ab6354e126bbeba8577fb665a9462afeee.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 378, + 198, + 552, + 232 + ], + "blocks": [ + { + "bbox": [ + 378, + 198, + 552, + 232 + ], + "lines": [ + { + "bbox": [ + 378, + 198, + 552, + 232 + ], + "spans": [ + { + "bbox": [ + 378, + 198, + 552, + 232 + ], + "type": "image", + "image_path": "85545d4d442ad80243ffb8b78dc40cee7e78444acc8b7f750789189b0c03bbf4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 59, + 246, + 298, + 277 + ], + "blocks": [ + { + "bbox": [ + 268, + 238, + 343, + 247 + ], + "lines": [ + { + "bbox": [ + 268, + 238, + 343, + 247 + ], + "spans": [ + { + "bbox": [ + 268, + 238, + 343, + 247 + ], + "type": "text", + "content": "Video-to-video editing" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 246, + 298, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 246, + 298, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 246, + 298, + 277 + ], + "type": "image", + "image_path": "6c9c116f5690eba3d9ee44b3b791916ff76bb2ec58621dbbe86a0ae44980aa0d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 278, + 218, + 285 + ], + "lines": [ + { + "bbox": [ + 146, + 278, + 218, + 285 + ], + "spans": [ + { + "bbox": [ + 146, + 278, + 218, + 285 + ], + "type": "text", + "content": "Pose Transfer (Pose Control)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 312, + 246, + 552, + 277 + ], + "blocks": [ + { + "bbox": [ + 312, + 246, + 552, + 277 + ], + "lines": [ + { + "bbox": [ + 312, + 246, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 246, + 552, + 277 + ], + "type": "image", + "image_path": "1d10ef66bf9e895e32d99c420bde8e9d16ef77d13a2666119f69ac5d62c76b41.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 278, + 470, + 285 + ], + "lines": [ + { + "bbox": [ + 386, + 278, + 470, + 285 + ], + "spans": [ + { + "bbox": [ + 386, + 278, + 470, + 285 + ], + "type": "text", + "content": "Motion Transfer (Layout Control)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 59, + 293, + 298, + 325 + ], + "blocks": [ + { + "bbox": [ + 59, + 293, + 298, + 325 + ], + "lines": [ + { + "bbox": [ + 59, + 293, + 298, + 325 + ], + "spans": [ + { + "bbox": [ + 59, + 293, + 298, + 325 + ], + "type": "image", + "image_path": "58ad5b44490705232a973cf7a586334b054d0bc8c494901a11fe885919a638d8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 326, + 226, + 333 + ], + "lines": [ + { + "bbox": [ + 138, + 326, + 226, + 333 + ], + "spans": [ + { + "bbox": [ + 138, + 326, + 226, + 333 + ], + "type": "text", + "content": "Structure Transfer (Depth Control)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 313, + 293, + 552, + 325 + ], + "blocks": [ + { + "bbox": [ + 313, + 293, + 552, + 325 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 552, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 552, + 325 + ], + "type": "image", + "image_path": "80df6d6119b73f2d4005bce8b9c89505b8abb30c091bc04ee8329de809340e56.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 326, + 463, + 333 + ], + "lines": [ + { + "bbox": [ + 394, + 326, + 463, + 333 + ], + "spans": [ + { + "bbox": [ + 394, + 326, + 463, + 333 + ], + "type": "text", + "content": "Colorization (Gray Control)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 59, + 342, + 298, + 375 + ], + "blocks": [ + { + "bbox": [ + 254, + 335, + 356, + 343 + ], + "lines": [ + { + "bbox": [ + 254, + 335, + 356, + 343 + ], + "spans": [ + { + "bbox": [ + 254, + 335, + 356, + 343 + ], + "type": "text", + "content": "Masked Video-to-video editing" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 342, + 298, + 375 + ], + "lines": [ + { + "bbox": [ + 59, + 342, + 298, + 375 + ], + "spans": [ + { + "bbox": [ + 59, + 342, + 298, + 375 + ], + "type": "image", + "image_path": "95d1d327420388d5737f00901b38c9c4081cca6ef3806a9ed2ec1ed8f43224ce.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 375, + 222, + 383 + ], + "lines": [ + { + "bbox": [ + 143, + 375, + 222, + 383 + ], + "spans": [ + { + "bbox": [ + 143, + 375, + 222, + 383 + ], + "type": "text", + "content": "Subject Recreation (Inpainting)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 312, + 342, + 552, + 375 + ], + "blocks": [ + { + "bbox": [ + 312, + 342, + 552, + 375 + ], + "lines": [ + { + "bbox": [ + 312, + 342, + 552, + 375 + ], + "spans": [ + { + "bbox": [ + 312, + 342, + 552, + 375 + ], + "type": "image", + "image_path": "0f9436c54eb21d2c2ced5d62e408773bb87a5fad5ad4a71c1a7bcecc0f3dff35.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 375, + 465, + 383 + ], + "lines": [ + { + "bbox": [ + 391, + 375, + 465, + 383 + ], + "spans": [ + { + "bbox": [ + 391, + 375, + 465, + 383 + ], + "type": "text", + "content": "Subject Removal (Inpainting)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 59, + 389, + 298, + 422 + ], + "blocks": [ + { + "bbox": [ + 59, + 389, + 298, + 422 + ], + "lines": [ + { + "bbox": [ + 59, + 389, + 298, + 422 + ], + "spans": [ + { + "bbox": [ + 59, + 389, + 298, + 422 + ], + "type": "image", + "image_path": "6e856a972006ce8b10c473c7deb34dd2b9911915774764a54da3b6adf70ffa9b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 423, + 222, + 430 + ], + "lines": [ + { + "bbox": [ + 143, + 423, + 222, + 430 + ], + "spans": [ + { + "bbox": [ + 143, + 423, + 222, + 430 + ], + "type": "text", + "content": "Canvas Extension (Outpainting)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 321, + 389, + 552, + 422 + ], + "blocks": [ + { + "bbox": [ + 321, + 389, + 552, + 422 + ], + "lines": [ + { + "bbox": [ + 321, + 389, + 552, + 422 + ], + "spans": [ + { + "bbox": [ + 321, + 389, + 552, + 422 + ], + "type": "image", + "image_path": "f1285142f3ed6719d624eecacd52d7f643cbad1f1346811431897c717ae26fcc.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 423, + 461, + 430 + ], + "lines": [ + { + "bbox": [ + 382, + 423, + 461, + 430 + ], + "spans": [ + { + "bbox": [ + 382, + 423, + 461, + 430 + ], + "type": "text", + "content": "Temporal Extension (First Clip)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 59, + 437, + 298, + 473 + ], + "blocks": [ + { + "bbox": [ + 274, + 432, + 335, + 440 + ], + "lines": [ + { + "bbox": [ + 274, + 432, + 335, + 440 + ], + "spans": [ + { + "bbox": [ + 274, + 432, + 335, + 440 + ], + "type": "text", + "content": "Task Composition" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 437, + 298, + 473 + ], + "lines": [ + { + "bbox": [ + 59, + 437, + 298, + 473 + ], + "spans": [ + { + "bbox": [ + 59, + 437, + 298, + 473 + ], + "type": "image", + "image_path": "5a8e8ec3fd8352db7fcaa2531b94053d11e401af6e453edc2f6ee0834d1ac198.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 473, + 244, + 480 + ], + "lines": [ + { + "bbox": [ + 124, + 473, + 244, + 480 + ], + "spans": [ + { + "bbox": [ + 124, + 473, + 244, + 480 + ], + "type": "text", + "content": "Swap Anything (Subject Reference + Inpainting)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 490, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 490, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 555, + 514 + ], + "type": "text", + "content": "Figure 1. Comprehensive capabilities of VACE. We present the outstanding generation results based on Wan2.1 (left) and LTX-Video (right). For each task, the original input image or video (left), the context video (top left corner), and the generated frames are illustrated." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 313, + 440, + 552, + 473 + ], + "blocks": [ + { + "bbox": [ + 313, + 440, + 552, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 552, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 552, + 473 + ], + "type": "image", + "image_path": "ca5b623143e9929a16ef37e21b1cf7268db1c9a956e3c85c812cdbf5bb037f27.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 473, + 498, + 480 + ], + "lines": [ + { + "bbox": [ + 369, + 473, + 498, + 480 + ], + "spans": [ + { + "bbox": [ + 369, + 473, + 498, + 480 + ], + "type": "text", + "content": "Animate Anything (Frame Reference + Pose Control)" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 152, + 528, + 200, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 528, + 200, + 540 + ], + "spans": [ + { + "bbox": [ + 152, + 528, + 200, + 540 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 54, + 552, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 552, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 552, + 297, + 696 + ], + "type": "text", + "content": "Diffusion Transformer has demonstrated powerful capability and scalability in generating high-quality images and videos. Further pursuing the unification of generation and editing tasks has yielded significant progress in the domain of image content creation. However, due to the intrinsic demands for consistency across both temporal and spatial dynamics, achieving a unified approach for video synthesis remains challenging. We introduce VACE, which enables users to perform Video tasks within an All-in-one framework for Creation and Editing. These tasks include reference-to-video generation, video-to-video editing, and masked video-to-video editing. Specifically, we effectively" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "text", + "content": "integrate the requirements of various tasks by organizing video task inputs, such as editing, reference, and masking, into a unified interface referred to as the Video Condition Unit (VCU). Furthermore, by utilizing a Context Adapter structure, we inject different task concepts into the model using formalized representations of temporal and spatial dimensions, allowing it to handle arbitrary video synthesis tasks flexibly. Extensive experiments demonstrate that the unified model of VACE achieves performance on par with task-specific models across various subtasks. Simultaneously, it enables diverse applications through versatile task combinations. Project page: https://alivilab.github.io/VACE-Page/." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 182, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 182, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 182, + 712 + ], + "type": "text", + "content": "*Equal Contribution. †Project lead." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17191" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 378 + ], + "type": "text", + "content": "In recent years, the domain of visual generation tasks has witnessed remarkable advancements, driven in particular by the rapid evolution of diffusion models [24, 25, 48, 53, 54, 56, 57]. Beyond the early foundational pre-trained models for text-to-image [7, 16, 33] or text-to-video [9, 22, 64] generation in the field, there has been a proliferation of downstream tasks and applications, such as painting [3, 82], editing [4, 42, 68, 70, 75], controllable generation [30, 76], frame reference generation [20, 73], and ID-referenced video synthesis [11, 35, 47, 74]. This array of developments highlights the dynamic and complex nature of the visual generation field. To enhance task flexibility and reduce the overhead associated with deploying multiple models, researchers have begun to focus on constructing unified model architectures [12, 63] (e.g., ACE [23, 41] and Omni-Gen [71]), aiming to integrate different tasks into a single image model, facilitating the creation of various application workflows while maintaining simplicity in usage. In the field of video, due to the collaborative transformations in both temporal and spatial dimensions, leveraging a unified model can present endless possibilities for video creation. However, leveraging diverse input modalities and ensuring spatiotemporal consistency are still challenging for unified video generation and editing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 379, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 714 + ], + "type": "text", + "content": "We propose VACE, an all-in-one model for video creation and editing that performs tasks including reference-to-video generation, video-to-video editing, masked video-to-video editing, and free composition of these tasks, as illustrated in Fig. 1. The aggregation of various capabilities reduces the costs of service deployment and user interaction, and by combining the capabilities of different tasks within a single model, it addresses challenges faced by existing video generation models, such as controllable generation of long videos, multi-condition and reference-based generation, and continuous video editing, thereby empowering users with greater creativity. To achieve this, we utilize the current mainstream Diffusion Transformers (DiTs) structure as the foundational video framework and pre-trained text-to-video generation models [22, 64], which provide better basic capabilities and scalability for handling long video sequences. Specifically, VACE takes into account the needs of different tasks during its construction and designs a unified interface, dubbed the Video Condition Unit (VCU), which integrates multiple modalities such as images or videos for editing, references, and masks. Furthermore, to differentiate the visual modality information in editing and reference tasks, we introduce the concept decoupling strategy, enabling the model to understand what aspects need to be retained and what should be modified. Meanwhile, by employing a pluggable Context Adapter structure, concepts from different tasks (e.g., the areas or ranges of editing or reference) are injected into the model through col" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "laborative spatiotemporal representation, enabling it to possess the capability of adaptive processing for unified tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 97, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 555, + 276 + ], + "type": "text", + "content": "Due to the lack of existing multi-task benchmarks in video synthesis, we construct a dataset of 480 evaluation samples containing 12 different tasks, while evaluating the performance of the VACE unified model by comparing it with existing specialized models. Experimental results demonstrate that our framework exhibits sufficient competitiveness in both quantitative and qualitative analyses. To the best of our knowledge, we are the first all-in-one model based on the video DiT architecture that concurrently supports such a wide range of tasks. Notably, this innovative framework allows for the compositional expansion of base tasks, enabling the construction of scenarios such as long video re-rendering, which provides a versatile and efficient solution for video synthesis, opening new possibilities for user-side video content creation and editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 286, + 401, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 401, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 401, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 307, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 605 + ], + "type": "text", + "content": "Visual Generation and Editing. With the rapid development of image [2, 7, 16, 18, 58, 59] and video [22, 32, 73, 77] generation models, they are being used to create high-quality visual content and are widely applied in fields such as advertising, film special effects, game development, and animation production [13, 43-45, 55]. Meanwhile, to meet the diverse needs of visual media production and to enhance efficiency and quality, precise generation and editing methods have emerged. Models are required to perform generative creation based on multimodal inputs, such as depth, structure, pose, scene, and characters. According to the purposes of the input conditions, we can categorize them into two types: editing of the input and concept-guided re-creation. A significant portion of the work, such as ControlNet [76], ControlVideo,Composer [26], VideoComposer [68], and SCEdit [30], focuses on single-condition editing and multi-condition composite editing based on temporal and spatial alignment conditions. Additionally, some works that focus on interactive local editing scenarios, such as DragGAN [46] and MagicBrush [75]. Methods that guide generation based on semantic information from the input, such as Cone [38], Cone2 [39], InstantID [67], and PuLID [21], can achieve conceptual understanding of the input and inject it into the model for creative purposes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Task-unified Visual Generative Model. As the complexity and diversity of user creations increase, relying solely on a single model or a complicated chain of multiple models can no longer provide a convenient and efficient path for implementing creative ideas. In image generation, a unified generation and editing framework has begun to emerge, allowing for more flexible creative approaches. Methods such as UltraEdit [81] and SEED-Data-Edit [19] provide general-purpose editing datasets, while techniques like In" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17192" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 549, + 230 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 549, + 230 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 549, + 230 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 549, + 230 + ], + "type": "image", + "image_path": "2bdea4dc0e5c25dd5529b7d9f6e57e9084415154530503123357e880e21f4386.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 238, + 520, + 250 + ], + "lines": [ + { + "bbox": [ + 88, + 238, + 520, + 250 + ], + "spans": [ + { + "bbox": [ + 88, + 238, + 520, + 250 + ], + "type": "text", + "content": "Figure 2. Task categories covered by VACE. Four basic tasks can be combined to create a vast number of possibilities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 443 + ], + "type": "text", + "content": "structPix2Pix [4], MagicBrush [61], and CosXL [60] offer general instruction-based editing features. Additionally, methods like UniControl [50] and UNIC-Adapter [15] have unified controllable generation. Further advancements have led to the development of ACE [23, 41], OmniGen [71], OmniControl [63], and UniReal [12], which expand the scope of tasks by providing flexible controllable generation, local editing, and reference-guided generation. In the video domain, due to the increased difficulty of generation, approaches often manifest as single-task single-model frameworks, offering capabilities for editing or reference generation, as seen in Video-P2P [37], MagicEdit [34], MotionCtrl [69], Magic Mirror [80], and Phantom [35]. VACE aims to fill the gap for a unified model within the video domain, providing possibilities for complex creative scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 457, + 111, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 111, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 111, + 470 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": "VACE is designed as a multimodal-to-video generation model, where text, image, video, and mask are integrated into a unified conditioning input. To cover as many video generation and editing tasks as possible, we conduct in-depth research into existing tasks, then divide them into 4 categories according to their individual requirements of multimodal inputs. Without losing generality, we specifically design a novel multimodal input format for each category under a Video Condition Unit (VCU) paradigm. Finally, we restructure the DiT model for VCU inputs, making it a versatile model for a wide range of video tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 623, + 249, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 249, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 249, + 635 + ], + "type": "text", + "content": "3.1. Multimodal Inputs and Video Tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Despite existing video tasks varying in complex user inputs and ambitious creative goals, we found that most of their inputs can be fully represented in 4 modalities: text, image, video, and mask. Overall, as illustrated in Fig. 2, we group these video tasks into 5 categories based on their requirements of these four multimodal inputs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 262, + 553, + 514 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 262, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 553, + 285 + ], + "type": "text", + "content": "- Text-to-Video Generation (T2V) is a basic video creation task, and text is the only input." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 287, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 287, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 314, + 287, + 553, + 334 + ], + "type": "text", + "content": "- Reference-to-Video Generation (R2V) requires additional images as reference inputs, making sure that specified contents, such as subjects of faces, animals, and other objects, or video frames, appear in the generated video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 335, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 335, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 553, + 417 + ], + "type": "text", + "content": "- Video-to-Video Editing (V2V) makes an entire change to a provided video, such as colorization, stylization, controllable generation, etc. We use video control types whose control signals can be represented and stored as RGB videos, including depth, grayscale, pose, scribble, optical flow, and layout; however, the method itself is not limited to these." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 418, + 553, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 418, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 553, + 489 + ], + "type": "text", + "content": "- Masked Video-to-Video Editing (MV2V) makes changes to an input video only within the provided 3D regions of interest (3D ROI), seamlessly blending in with the other unchanged regions, such as inpainting, outpainting, video extension, etc. We use an extra spatiotemporal mask to represent the 3D ROI." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 491, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 553, + 514 + ], + "type": "text", + "content": "- Task Composition includes all the compositional possibilities of the 4 types of video tasks above." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 521, + 437, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 521, + 437, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 521, + 437, + 533 + ], + "type": "text", + "content": "3.2. Video Condition Unit" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": "We propose an input paradigm, Video Condition Unit (VCU), to unify diverse input conditions into textual input, frame sequence, and mask sequence. A VCU can be denoted as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 400, + 587, + 553, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 587, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 400, + 587, + 553, + 601 + ], + "type": "interline_equation", + "content": "V = [ T; F; M ], \\tag {1}", + "image_path": "2e2ca69e819b76863c83d555f2c9db00228551a0a38a80225ca24ea18b616050.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " is a text prompt, while " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " are sequences of context video frames " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{u_1, u_2, \\dots, u_n\\}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " and masks " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{m_1, m_2, \\dots, m_n\\}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " respectively. Here, " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " is in RGB space, normalized to " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " is binary, in which \"1\"s and \"0\"s symbolize where to edit or not. " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " are aligned both in spatial size " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " and temporal size " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ". In T2V, no context frame or mask is required. To keep generality, we assign a default value " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "0_{h \\times w}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " to each " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " denoting empty input, and set every " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "1_{h \\times w}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " meaning that all these pixels" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17193" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 298, + 203 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 298, + 203 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 298, + 203 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 298, + 203 + ], + "type": "image", + "image_path": "2a2cdb8973973f507ef1873f7f8e7f67b97950d6faa55f40b8c825386eb1ca8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "text", + "content": "Figure 3. Overview of VACE Framework. Frames and masks are tokenized through Concept Decoupling, Context Latent Encode, and Context Embedder. To achieve training with VCU as input, we employ two strategies: (a) Fully Fine-tuning and (b) Context Adapter Tuning. The latter converges faster and supports pluggable features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 300, + 70, + 552, + 204 + ], + "blocks": [ + { + "bbox": [ + 300, + 70, + 552, + 204 + ], + "lines": [ + { + "bbox": [ + 300, + 70, + 552, + 204 + ], + "spans": [ + { + "bbox": [ + 300, + 70, + 552, + 204 + ], + "type": "image", + "image_path": "572bd3aa43bbd8d28bf5ab05a3e7a83f3135cec10cac1bc8df624e2eb426a3ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 298, + 288, + 435 + ], + "blocks": [ + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "type": "text", + "content": "Table 1. The formal representation of frames " + }, + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "type": "inline_equation", + "content": "(F\\mathrm{s})" + }, + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "type": "text", + "content": " and masks " + }, + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "type": "inline_equation", + "content": "(M\\mathrm{s})" + }, + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "type": "text", + "content": " under the four basic tasks. Frames and masks are aligned spatially and temporally." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 298, + 288, + 435 + ], + "lines": [ + { + "bbox": [ + 56, + 298, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 288, + 435 + ], + "type": "table", + "html": "
TasksFrames (Fs) & Masks (Ms)
T2VF = {0h×w} × n\nM = {1h×w} × n
R2VF = {r1, r2, ..., rl} + {0h×w} × n\nM = {0h×w} × l + {1h×w} × n
V2VF = {u1, u2, ..., un}\nM = {1h×w} × n
MV2VF = {u1, u2, ..., un}\nM = {m1, m2, ..., mn}
", + "image_path": "92a45626850099a38ad40633be9b967221966e8df2a8a8367cad01631bff0ba4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "text", + "content": "are about to be re-generated. For R2V, additional reference frames " + }, + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "text", + "content": " are inserted in front of the default frame sequence, while all-zero masks " + }, + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "inline_equation", + "content": "0_{h \\times w}" + }, + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "text", + "content": " are inserted in front of the mask sequence. These all-zero masks mean that the corresponding frames should be kept unchanged. In V2V, the context frame sequence is the input video frames, and the context mask is a sequence of " + }, + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "inline_equation", + "content": "1_{h \\times w}" + }, + { + "bbox": [ + 55, + 456, + 295, + 563 + ], + "type": "text", + "content": ". For MV2V, both context video and mask are required. The formal mathematical representations are shown in Tab. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "VCU can also support task composition. For example, the context frames of reference-inpainting task are " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{r_1,r_2,\\dots,r_l,u_1,u_2,\\dots,u_n\\}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and the context masks are " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{0_{h\\times w}\\} \\times l + \\{m_1,m_2,\\dots,m_n\\}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ". In this case, users can modify " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " objects in the video and regenerate based on the provided reference images. For another example, users only have a scribble image and want to generate a video beginning with the contents described by this scribble image, which is a scribble-based video extension task. The context frames are " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{u\\} +\\{0_{h\\times w}\\} \\times (n - 1)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and the context masks are " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{1_{h\\times w}\\} \\times n" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ". In this way, we can achieve multi-condition and reference control generation for long videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 397, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 397, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 397, + 269 + ], + "type": "text", + "content": "3.3. Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 275, + 555, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 383 + ], + "type": "text", + "content": "We restructure the DiT model for VACE, as shown in Fig. 3, aiming to support multimodal VCU inputs. Since there is an existing pipeline for text tokenization, we only consider the tokenization of context frames and masks. After being tokenized, the context tokens are combined with noisy video tokens and used to fine-tune the DiT model. Different from that, we also propose a Context Adapter Tuning strategy, which allows context tokens to pass through Context Blocks and be added back to the original DiT Blocks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 388, + 434, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 434, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 434, + 398 + ], + "type": "text", + "content": "3.3.1. Context Tokenization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": "Concept Decoupling. Two different visual concepts of natural video and control signals, like depth, pose, are encoded in " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": " simultaneously. We believe that explicitly separating these data of different modalities and distributions is essential for model convergence. The concept decoupling is based on masks and yields two frame sequences identical in shape: " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "F_{c} = F \\times M" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "F_{k} = F \\times (1 - M)" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": " is called the reactive frames contain all the pixels to be changed, while all the pixels to be kept are stored in " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": ", named inactive frames. Specifically, the reference images and the unchanged part of V2V and MV2V go to " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": ", while control signals and those pixels about to change, such as gray pixels, are collected to " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Context Latent Encoding. A typical DiT processes noisy video latents " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{n' \\times h' \\times w' \\times d}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "n'" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "h'" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "w'" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " are the temporal and spatial shapes of the latent space. Similar to " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " need to be encoded into a high-dimensional feature space to ensure the property of significant spatiotemporal correlations. Therefore, we reorganize them together with " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " into a hierarchical and spatiotemporal aligned visual features. " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " are processed by video VAE and mapped into the same latent space of " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ", maintaining their spatiotemporal consistency. To avoid any unexpected mishmash of images and videos, reference images are separately encoded by VAE encoder and concatenated back along the temporal dimension, while the correspond" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17194" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "ing parts need to be removed during decoding. " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " is directly reshaped and interpolated. After that, " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "F_{c}, F_{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " are all mapping into latent spaces and are spatiotemporal aligned with " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " in the shape of " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "n' \\times h' \\times w'" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "content": "Context Embedder. We extend the embedder layer by concatenating " + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "content": " in the channel dimension and tokenizing them into context tokens, which we refer to as the Context Embedder. The corresponding weights to tokenize " + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "content": " are directly copied from the original video embedder, and the weights to tokenize " + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 121, + 294, + 203 + ], + "type": "text", + "content": " are initialized to zeros." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 286, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 286, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 286, + 221 + ], + "type": "text", + "content": "3.3.2. Fully Fine-Tuning and Context Adapter Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 224, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 450 + ], + "type": "text", + "content": "To achieve training with VCU as input, a simple methodology is to fully fine-tune the whole DiT model, as shown in Fig. 3a. Context tokens are added together with noisy tokens " + }, + { + "bbox": [ + 55, + 224, + 295, + 450 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 224, + 295, + 450 + ], + "type": "text", + "content": ", and all the parameters in DiT and the newly introduced Context Embedder will be updated during training. To avoid fully fine-tuning and achieve faster convergence, as well as to establish a pluggable feature with the foundation model, we also propose another methodology processing the context token in a Res-Tuning [29] manner, as shown in Fig. 3b. Particularly, we select and copy several Transformer Blocks from the original DiT, forming a distributed and cascade-type Context Blocks. The original DiT processes video tokens and text tokens, while the newly added Transformer Blocks process context tokens and text tokens. The output of each Context Block is inserted back into the DiT blocks as an additive signal to assist the main branch in performing generation and editing tasks. In this manner, the parameters of DiT are frozen. Only the Context Embedder and Context Blocks are trainable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 115, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 115, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 115, + 473 + ], + "type": "text", + "content": "4. Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 165, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 165, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 165, + 492 + ], + "type": "text", + "content": "4.1. Data Construction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 616 + ], + "type": "text", + "content": "To obtain an all-in-one model, the diversity and complexity of the required data construction also increase. Existing common text-to-video and image-to-video tasks only require constructing pairs of text and video. However, for the tasks in VACE, the modalities need to be further expanded to include target videos, source videos, local masks, reference, and more. To efficiently and rapidly acquire data for various tasks, it is imperative to maintain video quality while also conducting instance-level analysis and understanding of the video data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "To achieve this, we first analyze the video data itself by performing shot slicing and preliminarily filtering out data based on resolution, aesthetic score, and motion amplitude. Next, we label the first frame of the video using RAM [78] and combine it with Grounding DINO [36] for detection, utilizing the localization results to perform secondary filtering on videos with target areas that are either too small or too large. Furthermore, we employ the propagate operation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "of SAM2 [52] for video segmentation to obtain instance-level information across the video. Leveraging the results of video segmentation, we filter instances in the temporal dimension by calculating the effective frame ratio based on the mask area threshold." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 396 + ], + "type": "text", + "content": "In the actual training process, the construction for different tasks also needs to be tailored according to the characteristics of each task: 1) For some controllable video generation tasks, we pre-extract depth [51], scribble [6], pose [5, 72], and optical flow [65] from the filtered videos. For gray and layout tasks, we create data on the fly. 2) For repainting tasks, random instances from the videos can be masked for inpainting, while the inverse of the mask enables the construction of outpainting data. Augmentation of the masks [62] allows for unconditional inpainting. 3) In the case of extension tasks, we extract key frames such as the first frame, last frame, frames from both ends, random frames, and segments from both ends to support a wider variety of extension types. 4) For reference tasks, we can extract several face or object instances from the videos and apply offline or online augmentation operations to create paired data. Notably, we randomly combine all the previously mentioned tasks for training to accommodate a broader range of model application scenarios. Additionally, for all operations involving masks, we perform arbitrary augmentation to satisfy various granular local generation requirements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 407, + 424, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 424, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 424, + 418 + ], + "type": "text", + "content": "4.2. VACE-Benchmark" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 592 + ], + "type": "text", + "content": "Significant progress has been made in the field of video generation. However, a scientific and thorough evaluation of the performance of these models remains an urgent issue that needs to be addressed. VBench [27] and VBench++ [28] have established a precise evaluation framework for text-to-video and image-to-video tasks through an extensive assessment suite and dimensional design. Nevertheless, as the video generation ecosystem continues to evolve, more derivative tasks have begun to emerge, such as video reference generation and video editing, for which a comprehensive benchmark is still lacking. To address this gap, we propose VACE-Benchmark to evaluate various downstream tasks related to video in a systematic manner." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Starting from the data sources, we recognize that real videos and generated videos may exhibit different performance characteristics during evaluation. Thus, we collected a total of 240 high-quality videos categorized by their sources, encompassing various data types, including text-to-video, inpainting, outpainting, extension, grayscale, depth, scribble, pose, optical flow, layout, reference face, and reference object tasks, with an average of 20 samples for each task. The input modalities include videos, masks, and references, and we also provide the original videos" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17195" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 559, + 426 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative evaluations on VACE-Benchmark. We compare the automated score metrics of the unified VACE based on LTX-Video and the proprietary model on the dimensions of video quality and video consistency, as well as results of human user studies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 559, + 426 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 559, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 559, + 426 + ], + "type": "table", + "html": "
TypeMethodVideo Quality & Video ConsistencyUser Study
Aesthetic QualityBackground ConsistencyDynamic DegreeImaging QualityMotion SmoothnessOverall ConsistencySubject ConsistencyTemporal FlickeringNormalized AveragePrompt FollowingTemporal ConsistencyVideo QualityAverage
I2VI2VGenXL [77]55.20%92.87%60.00%63.31%97.43%23.78%89.58%95.67%71.54%2.651.602.342.20
CogVideoX-I2V [73]57.78%94.80%40.00%68.23%98.69%24.38%93.84%97.84%73.66%3.302.283.192.92
LTX-Video [22]56.12%94.57%35.00%62.72%99.27%24.92%92.83%98.41%72.89%2.952.282.282.50
VACE (Ours)57.53%95.32%45.00%68.03%99.08%25.13%93.61%97.80%74.38%3.204.002.543.24
InpaintProPainter [82]44.70%95.64%50.00%61.57%99.01%18.48%92.99%98.47%70.15%2.354.002.993.11
VACE (Ours)51.30%96.30%50.00%60.39%99.12%21.12%94.59%98.21%72.05%2.404.002.603.00
OutpaintFollow-Your-Canvas [8]53.30%95.99%5.00%69.53%98.08%25.90%95.38%97.20%71.54%3.052.001.632.23
M3DDM [17]53.34%95.87%30.00%65.07%99.22%25.43%93.65%98.85%73.16%3.703.882.283.29
VACE (Ours)57.04%96.55%30.00%69.49%99.20%25.36%94.47%98.47%74.25%3.903.923.583.80
DepthControl-A-Video [10]50.62%91.71%70.00%67.76%97.58%24.48%88.10%96.58%72.35%2.702.281.542.17
VideoComposer [68]50.03%94.18%70.00%59.44%96.23%24.95%89.79%94.38%70.74%2.602.442.172.40
ControlVideo [79]63.30%95.02%10.00%65.13%96.49%24.20%92.29%95.42%70.07%2.552.501.822.29
VACE (Ours)56.72%96.12%60.00%66.41%98.84%25.27%94.09%97.27%74.99%3.103.922.663.23
PoseText2Video-Zero [31]57.63%87.67%100.00%70.74%79.65%23.94%84.82%76.57%59.69%2.152.001.882.01
ControlVideo [79]65.37%94.56%25.00%65.28%97.32%25.19%92.76%96.82%72.45%2.151.802.031.99
Follow-Your-Pose [40]48.79%86.80%100.00%67.41%90.12%26.10%80.18%88.02%66.43%2.002.601.582.06
VACE (Ours)60.17%94.92%75.00%64.71%98.63%26.44%94.82%96.60%76.13%2.953.962.633.18
FlowFLATTEN [14]56.23%95.80%70.00%61.65%97.86%26.23%93.94%96.17%74.42%3.502.403.193.03
VACE (Ours)55.76%96.07%75.00%65.37%98.98%25.89%94.63%96.93%75.90%2.903.752.603.08
R2VKling1.6 [1]62.13%96.04%85.00%69.27%99.38%27.82%93.79%97.79%78.81%4.224.103.804.04
Pika2.2 [49]62.48%96.79%65.00%69.87%99.37%26.02%95.93%98.90%77.87%4.003.853.873.91
Vidu2.0 [66]64.30%96.85%35.00%67.03%99.66%26.53%96.73%99.41%76.47%3.903.853.773.84
VACE (Ours)63.25%98.03%30.00%72.29%99.51%25.85%98.54%99.15%76.76%3.473.423.303.40
", + "image_path": "1fa4a54822a61bf1cc9e1b5b49f7de831ea2fd8e0e4c0076d2e8a8520a89fe66.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 439, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 499 + ], + "type": "text", + "content": "to enable further processing by developers based on the specific characteristics of each task. Regarding the data prompts, we supply the original captions of the videos for quantitative assessment, as well as rewritten prompts tailored to the specific tasks to evaluate the models' creativity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 528, + 173, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 173, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 173, + 541 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Implementation Details. VACE is trained based on Diffusion Transformers for text-to-video generation at various scales. It utilizes LTX-Video-2B [22] for faster generation, while Wan-T2V-14B [64] is used specifically for higher-quality outputs, supporting resolutions of up to 720p. The training employs a phased approach. Initially, we focus on foundational tasks such as inpainting and extension, which are considered modal complementary to the pre-trained text-to-video models. This includes the incorporation of masks and the learning of contextual generation in both spatial and temporal dimensions. Next, from a task expansion perspective, we progressively transition from single input reference frames to multiple input reference frames and from single tasks to composite tasks. Fi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 439, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 487 + ], + "type": "text", + "content": "nally, we fine-tune the model's quality using higher-quality data and longer sequences. The input for model training accommodates arbitrary resolutions, dynamic durations, and variable frame rates to support diverse input needs of users." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "Baselines. Our goal is to achieve the unification of video creation and editing tasks, and currently, there is no comparable all-in-one video generation model available, which leads us to focus our evaluation on comparing our general model with proprietary task-specific models. Moreover, due to the numerous tasks involved and the lack of open-sourced methods for many of them, we conduct our comparisons on models that are available either offline or online. Specifically for the tasks, we compare the following: 1) For the I2V task, we examine I2VGenXL [77], CogVideoX-I2V [73], and LTX-Video-I2V [22]; 2) In the repainting task, we compare the ProPainter [82] for removal inpainting, while Follow-Your-Canvas [8] and M3DDM [17] are compared for outpainting; 3) For controllable task, we use Control-A-Video [10], VideoComposer [68], and ControlVideo [79] under depth conditions, and compare Text2Video-Zero [31], ControlVideo [79], and Follow-Your-Pose [40] under pose conditions, as well as FLAT" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17196" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 74, + 149, + 118 + ], + "blocks": [ + { + "bbox": [ + 60, + 74, + 149, + 118 + ], + "lines": [ + { + "bbox": [ + 60, + 74, + 149, + 118 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 149, + 118 + ], + "type": "image", + "image_path": "c7fa9ea987d1bb71f5051d3e521300faee221cee1c690174509157a93ceb2484.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 132, + 138, + 140 + ], + "lines": [ + { + "bbox": [ + 70, + 132, + 138, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 132, + 138, + 140 + ], + "type": "text", + "content": "Reference Anything" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 70, + 549, + 127 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 549, + 127 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 549, + 127 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 549, + 127 + ], + "type": "image", + "image_path": "f6416d4a007a14dd68aefc6bfd0f1fe6cb773c3b035d97636f24ebb9a0c54d90.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 127, + 543, + 141 + ], + "lines": [ + { + "bbox": [ + 156, + 127, + 543, + 141 + ], + "spans": [ + { + "bbox": [ + 156, + 127, + 543, + 141 + ], + "type": "text", + "content": "The elegant lady carefully selects bags in the boutique, and she shows the charm of a mature woman in a black slim dress with a pearl necklace. Holding a vintage-inspired brown leather handbag, she is carefully observing its craftsmanship and texture. The interior of the store." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 145, + 156, + 196 + ], + "blocks": [ + { + "bbox": [ + 59, + 145, + 156, + 196 + ], + "lines": [ + { + "bbox": [ + 59, + 145, + 156, + 196 + ], + "spans": [ + { + "bbox": [ + 59, + 145, + 156, + 196 + ], + "type": "image", + "image_path": "8434dacb2e4a7709e3f78e2229e213784dba4a91b4b1f46122d498181dd52856.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 143, + 550, + 199 + ], + "blocks": [ + { + "bbox": [ + 156, + 143, + 550, + 199 + ], + "lines": [ + { + "bbox": [ + 156, + 143, + 550, + 199 + ], + "spans": [ + { + "bbox": [ + 156, + 143, + 550, + 199 + ], + "type": "image", + "image_path": "7755152ca82d7f0fbc13b2b7b203bfcb16d5f3adeb3ff41ce7b58ef1ace9a95a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 199, + 544, + 213 + ], + "lines": [ + { + "bbox": [ + 156, + 199, + 544, + 213 + ], + "spans": [ + { + "bbox": [ + 156, + 199, + 544, + 213 + ], + "type": "text", + "content": "A young boy rises from his chair and walks briskly to the right side of the frame towards the edge of the sun-drenched frame, as if chasing a new adventure. His eyes were bright, and the corners of his mouth were slightly upturned, revealing curiosity and excitement about the unknown..." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 215, + 156, + 271 + ], + "blocks": [ + { + "bbox": [ + 79, + 204, + 131, + 213 + ], + "lines": [ + { + "bbox": [ + 79, + 204, + 131, + 213 + ], + "spans": [ + { + "bbox": [ + 79, + 204, + 131, + 213 + ], + "type": "text", + "content": "Move Anything" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 215, + 156, + 271 + ], + "lines": [ + { + "bbox": [ + 60, + 215, + 156, + 271 + ], + "spans": [ + { + "bbox": [ + 60, + 215, + 156, + 271 + ], + "type": "image", + "image_path": "4a2a7dc2056bb9eefe68f4fd0f438de07109f6e820e51ec5c8399a3d85f38d3d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 271, + 534, + 285 + ], + "lines": [ + { + "bbox": [ + 156, + 271, + 534, + 285 + ], + "spans": [ + { + "bbox": [ + 156, + 271, + 534, + 285 + ], + "type": "text", + "content": "\"Anime-style, hot-blooded teenager in bright orange long-sleeved pants sportswear, standing on a surfboard, facing the golden sunshine in the rough sea. The teenager's short yellow hair is flying in the wind, his eyes are firm, and he has a confident smile on the corner of his mouth.\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 156, + 215, + 550, + 271 + ], + "blocks": [ + { + "bbox": [ + 156, + 215, + 550, + 271 + ], + "lines": [ + { + "bbox": [ + 156, + 215, + 550, + 271 + ], + "spans": [ + { + "bbox": [ + 156, + 215, + 550, + 271 + ], + "type": "image", + "image_path": "2ddbe8eeb9b05ec19d551d6405830a378b3c62dc317578be5cdea44cf0957355.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 59, + 287, + 156, + 343 + ], + "blocks": [ + { + "bbox": [ + 74, + 276, + 136, + 284 + ], + "lines": [ + { + "bbox": [ + 74, + 276, + 136, + 284 + ], + "spans": [ + { + "bbox": [ + 74, + 276, + 136, + 284 + ], + "type": "text", + "content": "Animate Anything" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 287, + 156, + 343 + ], + "lines": [ + { + "bbox": [ + 59, + 287, + 156, + 343 + ], + "spans": [ + { + "bbox": [ + 59, + 287, + 156, + 343 + ], + "type": "image", + "image_path": "04fcf82eeef0b0e26c17bc7b5f184eadced2d8d689a0a494b4537ebd63f4d9cb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 343, + 542, + 357 + ], + "lines": [ + { + "bbox": [ + 156, + 343, + 542, + 357 + ], + "spans": [ + { + "bbox": [ + 156, + 343, + 542, + 357 + ], + "type": "text", + "content": "The video shows an old movie-style scene in retro tones, with a little penguin and a kitten having a joyous bike race. Little penguins and kittens, both dressed in orange-and-red race suits, ride vintage multi-wheeled bikes on a nostalgic dirt road flanked by spectators..." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 156, + 287, + 549, + 343 + ], + "blocks": [ + { + "bbox": [ + 156, + 287, + 549, + 343 + ], + "lines": [ + { + "bbox": [ + 156, + 287, + 549, + 343 + ], + "spans": [ + { + "bbox": [ + 156, + 287, + 549, + 343 + ], + "type": "image", + "image_path": "a20893df72f446a06fe28fb383ae5b5f8fa37e8addd8069ae865846de816e160.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 59, + 366, + 156, + 409 + ], + "blocks": [ + { + "bbox": [ + 78, + 348, + 130, + 356 + ], + "lines": [ + { + "bbox": [ + 78, + 348, + 130, + 356 + ], + "spans": [ + { + "bbox": [ + 78, + 348, + 130, + 356 + ], + "type": "text", + "content": "Swap Anything" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 366, + 156, + 409 + ], + "lines": [ + { + "bbox": [ + 59, + 366, + 156, + 409 + ], + "spans": [ + { + "bbox": [ + 59, + 366, + 156, + 409 + ], + "type": "image", + "image_path": "5105cdd07923abce73ff9a89c0e5639e4918777310ce414986bec976cd654a25.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 156, + 358, + 549, + 415 + ], + "blocks": [ + { + "bbox": [ + 74, + 417, + 134, + 426 + ], + "lines": [ + { + "bbox": [ + 74, + 417, + 134, + 426 + ], + "spans": [ + { + "bbox": [ + 74, + 417, + 134, + 426 + ], + "type": "text", + "content": "Expand Anything" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 358, + 549, + 415 + ], + "lines": [ + { + "bbox": [ + 156, + 358, + 549, + 415 + ], + "spans": [ + { + "bbox": [ + 156, + 358, + 549, + 415 + ], + "type": "image", + "image_path": "9751db892a0a52597f8e5f5d005db3142b8067aa9d728a4fd80dd8f895ed60a5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 415, + 548, + 430 + ], + "lines": [ + { + "bbox": [ + 156, + 415, + 548, + 430 + ], + "spans": [ + { + "bbox": [ + 156, + 415, + 548, + 430 + ], + "type": "text", + "content": "\"In the style of classical oil painting, the background is a river, and in the center of the picture is a mature and elegant woman, wearing a long skirt and sitting on a chair. She took the open red heart-shaped sunglasses from her arms with both hands and put them on...\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 441, + 553, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 553, + 453 + ], + "type": "text", + "content": "Figure 4. Visualization results of compositional tasks. VACE creatively enables reference-, move-, animate-, swap-, and expand-anything." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 473, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 521 + ], + "type": "text", + "content": "TEN [14] under optical flow conditions; 4) In reference generation, given the absence of open-sourced models, we compare commercial products Kling1.6 [1], Pika2.2 [49], and Vidu2.0 [66]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Evaluation. To comprehensively evaluate the performance of various tasks, we employ the VACE-Benchmark for assessment. Specifically, we divide the evaluation into automatic scoring and a user study for manual assessment. For the automatic scoring, we utilize select metrics from VBench [27] to assess video quality and video consistency, including eight indicators: aesthetic quality, background consistency, dynamic degree, imaging quality, motion smoothness, overall consistency, subject consistency, and temporal flickering. For the manual assessment, we utilize the mean opinion score (MOS) as our evaluation metric, focusing on three aspects: prompt following, temporal consistency, and video quality. In practice, we anonymize the generated data and randomly distribute it to different participants for scoring on a scale from 1 to 5." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 472, + 399, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 472, + 399, + 484 + ], + "spans": [ + { + "bbox": [ + 314, + 472, + 399, + 484 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Quantitative Evaluation. We compare VACE comprehensive model based on LTX-Video with task proprietary approaches on VACE-Benchmark. For certain tasks, we follow existing methods; for example, although we support generating based on any frame, we conduct comparisons using the first-frame reference approach from current open-source methods to ensure fairness. From Tab. 2, we can see that for the tasks of I2V, inpainting, outpainting, depth, pose, and optical flow, our method demonstrates better performance than other open-source methods across eight indicators of video quality and video consistency, with normalized average metrics showing superior results. Some competing methods can only generate at a resolution of 256, have very short generation durations, and exhibit instability in temporal coherence, resulting in poorer performance on automatic metric calculations. For the R2V task, there is still a certain gap in metrics compared to commercial models for a small-scale model that aims for fast generation," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17197" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 175, + 167 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 175, + 167 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 175, + 167 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 175, + 167 + ], + "type": "image", + "image_path": "cfdbd7dd56c30c3a6af8e7f6a707e50ca6f36261f186b92d725ce4882711d134.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 170, + 160, + 178 + ], + "lines": [ + { + "bbox": [ + 77, + 170, + 160, + 178 + ], + "spans": [ + { + "bbox": [ + 77, + 170, + 160, + 178 + ], + "type": "text", + "content": "(a) Base structure setting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 76, + 290, + 167 + ], + "blocks": [ + { + "bbox": [ + 176, + 76, + 290, + 167 + ], + "lines": [ + { + "bbox": [ + 176, + 76, + 290, + 167 + ], + "spans": [ + { + "bbox": [ + 176, + 76, + 290, + 167 + ], + "type": "image", + "image_path": "271e1884078ac46369e29ea040268eec90e4a0725fcd2514331506d8f08d1a3c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 170, + 280, + 178 + ], + "lines": [ + { + "bbox": [ + 187, + 170, + 280, + 178 + ], + "spans": [ + { + "bbox": [ + 187, + 170, + 280, + 178 + ], + "type": "text", + "content": "(b) Hyperparameter settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 180, + 174, + 271 + ], + "blocks": [ + { + "bbox": [ + 60, + 180, + 174, + 271 + ], + "lines": [ + { + "bbox": [ + 60, + 180, + 174, + 271 + ], + "spans": [ + { + "bbox": [ + 60, + 180, + 174, + 271 + ], + "type": "image", + "image_path": "95c6ecc97504c4d853d22a32ff59fd3c44365b48d077815be8a79fc91f3a7089.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 274, + 174, + 283 + ], + "lines": [ + { + "bbox": [ + 62, + 274, + 174, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 174, + 283 + ], + "type": "text", + "content": "(c) Context adapter configurations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 180, + 290, + 271 + ], + "blocks": [ + { + "bbox": [ + 176, + 180, + 290, + 271 + ], + "lines": [ + { + "bbox": [ + 176, + 180, + 290, + 271 + ], + "spans": [ + { + "bbox": [ + 176, + 180, + 290, + 271 + ], + "type": "image", + "image_path": "14fdebd9ed061f7d0e1c6ae123885833ea7a70db27684d65351cc2fd5fe7e5d8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 274, + 281, + 283 + ], + "lines": [ + { + "bbox": [ + 186, + 274, + 281, + 283 + ], + "spans": [ + { + "bbox": [ + 186, + 274, + 281, + 283 + ], + "type": "text", + "content": "(d) Concept decouple setting." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 289, + 295, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 312 + ], + "type": "text", + "content": "Figure 5. Ablation Studies of the VACE regarding structures, hyperparameters, and module configurations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 335, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 383 + ], + "type": "text", + "content": "while being comparable to the metrics of Vidu 2.0. According to the results of human user studies, our method consistently performs better in evaluation metrics across multiple tasks, aligning well with user preferences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 552 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 1, we present the results of the VACE single model across various tasks. It is evident that the model achieves a high level of performance in video quality and temporal consistency. Furthermore, in composition tasks shown in Fig. 4, our model showcases impressive abilities, effectively integrating different modalities and tasks to produce results that cannot be generated by existing single or multiple models, thereby demonstrating its strong potential in the fields of video generation and editing. For example, in the \"Move Anything\" case, by providing a single input image and a movement trajectory, we are able to precisely move the characters in the scene with a specified direction while maintaining coherence and narrative consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 562, + 156, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 156, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 156, + 574 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 580, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 688 + ], + "type": "text", + "content": "To better understand the impact of different independent modules on a unified video generation framework, we conducted a series of systematic comparative experiments based on the LTX-Video model to achieve a better model structure and configuration. To accurately assess the different experimental settings, we sample 250 data points for each task as a validation set and calculate the training loss, reflecting the model's training progress through the mean curve changes of different tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Base Structure. Text-guided image or video generation models only take noise as inference input. When extended" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "to our unified input paradigm, VCU, we can conduct training using fully fine-tuning or by incorporating additional parameter fine-tuning. Specifically, as shown in Fig. 5a, we compare the concatenation of different inputs along the channel dimension and modify the input dimensions of the patchify projection layer to achieve the loading and fully fine-tuning of the pre-trained model. Additionally, we introduce some extra training parameters in the form of RestTuning [29], which serializes VCU in a bypass branch and injects information into the main branch. The results indicate that both methods yielded similar effects; however, since the additional parameter fine-tuning converges faster, we base our subsequent experiments on this approach. As shown in Fig. 5b, we further conduct hyperparameter experiments based on this structure, focusing on aspects such as weighting schemes, timestamp shifting, and p-zero." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 263, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 453 + ], + "type": "text", + "content": "Context Adapter. Since the number of context blocks will significantly affect the model size and inference time consumption, we attempt to find an optimal number and distribution of context blocks. We begin with selecting continuous blocks at the input side and make comparisons between the first 1/4 blocks, 1/2 blocks, and all blocks. Inspired by the Res-Tuning [29] method, we also experiment with evenly distributing the injection blocks instead of selecting a continuous block series. As shown in Fig. 5c, we can see that when using the same number of blocks, the distributed arrangement of blocks outperforms the continuous arrangement in shallow blocks. Furthermore, a greater number of blocks generally yields better results, but due to the limited improvement in effectiveness and the constraints of training resources, we adopt a partially distributed arrangement of blocks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 455, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 515 + ], + "type": "text", + "content": "Concept Decouple. During training, we introduce a Concept Decouple processing module to further disassemble the visual units, clarifying what content the model needs to learn to modify or retain. As shown in Fig. 5d, using this module results in a more significant reduction in loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 546, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 712 + ], + "type": "text", + "content": "This paper introduces VACE, an all-in-one video generation and editing framework. It unifies the diverse and complex multimodal inputs required for various video tasks, bridging the gap between specialized models for each individual task. This enables most video AI creation tasks to be completed with a single inference of a single model. While broadly covering various video tasks, VACE also supports flexible and free combinations of these tasks, greatly expanding the application scenarios of video generation models and meeting a wide range of user creative needs. The VACE framework paves the way for the development of unified visual generative models with multimodal inputs and represents a significant milestone in the field of visual generation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17198" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 152 + ], + "type": "text", + "content": "Acknowledgments. We would like to express our sincere appreciation for the contributions of many colleagues for their insightful discussions, valuable suggestions, and constructive feedback, including: Yuwei Wang, Haiming Zhao, Chenwei Xie, and Sheng Yao for their data contributions, and Shiwei Zhang, Tao Fang, and Xiang Wang for their discussions and suggestions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 193, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 193, + 294, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 294, + 214 + ], + "type": "text", + "content": "[1] KLING AI. KLING AI, https://klingai.com/, 2025.6,7,13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 217, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 249 + ], + "type": "text", + "content": "[2] Runway AI. Stable Diffusion v1.5 Model Card, https://huggingface.co/runwayml/stable-diffusion-v1-5, 2022.2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 283 + ], + "type": "text", + "content": "[3] Runway AI. Stable Diffusion Inpainting Model Card, https://huggingface.co/runwayml/stable-diffusion-inpainting, 2022.2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 286, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 295, + 329 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning To Follow Image Editing Instructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 18392-18402, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 374 + ], + "type": "text", + "content": "[5] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "type": "text", + "content": "[6] Caroline Chan, Frédo Durand, and Phillip Isola. Learning To Generate Line Drawings That Convey Geometry and Semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 475 + ], + "type": "text", + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-" + }, + { + "bbox": [ + 62, + 422, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 62, + 422, + 295, + 475 + ], + "type": "text", + "content": ": Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 533 + ], + "type": "text", + "content": "[8] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-Your-Canvas: Higher-Resolution Video Outpainting with Extensive Content Generation. In Assoc. Adv. Artif. Intell., 2025. 6, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 610 + ], + "type": "text", + "content": "[9] Shoufa Chen, Chongjian Ge, Yuqi Zhang, Yida Zhang, Fengda Zhu, Hao Yang, Hongxiang Hao, Hui Wu, Zhichao Lai, Yifei Hu, Ting-Che Lin, Shilong Zhang, Fu Li, Chuan Li, Xing Wang, Yanghua Peng, Peize Sun, Ping Luo, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Goku: Flow Based Video Generative Foundation Models. arXiv preprint arXiv:2502.04896, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[10] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Ji-ashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-A-Video: Controllable Text-to-Video Diffusion Models with Motion Prior and Reward Feedback Learning. arXiv preprint arXiv:2305.13840, 2023. 6, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. AnyDoor: Zero-shot Object-level Image Customization. arXiv preprint arXiv:2307.09481, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[12] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang Zhao. UniReal: Universal Image Generation and Editing via Learning Real-world Dynamics. arXiv preprint arXiv:2412.07774, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 552, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 552, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 552, + 161 + ], + "type": "text", + "content": "[13] Alibaba Cloud. Tongyi Wanxiang, https://tongyi. aliyun.com/wanxiang, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "text", + "content": "[14] Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: Optical FLOW-guided ATTENtion for consistent text-to-video editing. In Int. Conf. Learn. Represent., 2024. 6, 7, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 283 + ], + "type": "text", + "content": "[15] Lunhao Duan, Shanshan Zhao, Wenjun Yan, Yinglun Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, Mingming Gong, and Gui-Song Xia. UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation. arXiv preprint arXiv:2412.18928, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 288, + 553, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 362 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In Int. Conf. Mach. Learn., 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[17] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical Masked 3D Diffusion Model for Video Outpainting. In ACM Int. Conf. Multimedia, pages 7890-7900, 2023. 6, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 441 + ], + "type": "text", + "content": "[18] FLUX. FLUX, https://blackforestlabs.ai/, 2024.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[19] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. SEED-Data-Edit Technical Report: A Hybrid Dataset for Instructional Image Editing. arXiv preprint arXiv:2405.04007, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "text", + "content": "[20] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Pengfei Wan, Di Zhang, Yufan Liu, Weiming Hu, Zhengjun Zha, Haibin Huang, and Chongyang Ma. 12V-Adapter: A General Image-to-Video Adapter for Diffusion Models. In ACM SIGGRAPH, pages 1-12, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 589 + ], + "type": "text", + "content": "[21] Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, and Qian He. PuLID: Pure and Lightning ID Customization via Contrastive Alignment. In Adv. Neural Inform. Process. Syst., 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "type": "text", + "content": "[22] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion. arXiv preprint arXiv:2501.00103, 2025. 2, 6, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 711 + ], + "type": "text", + "content": "[23] Zhen Han, Zeyinzi Jiang, Yulin Pan, Jingfeng Zhang, Chaojie Mao, Chenwei Xie, Yu Liu, and Jingren Zhou. ACE: All-round Creator and Editor Following Instructions via Diffusion Transformer. In Int. Conf. Learn. Represent., 2025. 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17199" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In Adv. Neural Inform. Process. Syst., 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 295, + 128 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In Adv. Neural Inform. Process. Syst. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 172 + ], + "type": "text", + "content": "[26] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and Controllable Image Synthesis with Composable Conditions. In Int. Conf. Mach.Learn., 2023.2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 249 + ], + "type": "text", + "content": "[27] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive Benchmark Suite for Video Generative Models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 21807-21818, 2024. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 327 + ], + "type": "text", + "content": "[28] Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, Yaohui Wang, Xinyuan Chen, YingCong Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench++: Comprehensive and Versatile Benchmark Suite for Video Generative Models. arXiv preprint arXiv:2411.13503, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "text", + "content": "[29] Zeyinzi Jiang, Chaojie Mao, Ziyuan Huang, Ao Ma, Yiliang Lv, Yujun Shen, Deli Zhao, and Jingren Zhou. Res-Tuning: A Flexible and Efficient Tuning Paradigm via Unbinding Tuner from Backbone. In Adv. Neural Inform. Process. Syst., 2023. 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 437 + ], + "type": "text", + "content": "[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. SCEdit: Efficient and Controllable Image Diffusion Generation via Skip Connection Editing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8995-9004, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 295, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 493 + ], + "type": "text", + "content": "[31] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators. In Int. Conf. Comput. Vis., pages 15954–15964, 2023. 6, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 646 + ], + "type": "text", + "content": "[32] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyuan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models. arXiv preprint arXiv:2412.03603, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "text", + "content": "[33] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[34] Jun Hao Liew, Hanshu Yan, Jianfeng Zhang, Zhongcong Xu, and Jiashi Feng. MagicEdit: High-Fidelity and Temporally Coherent Video Editing. arXiv preprint arXiv:2308.14749, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "text", + "content": "[35] Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv preprint arXiv:2502.11079, 2025. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 553, + 217 + ], + "type": "text", + "content": "[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. arXiv preprint arXiv:2303.05499, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 218, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 218, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 218, + 553, + 262 + ], + "type": "text", + "content": "[37] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-P2P: Video Editing with Cross-attention Control. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8599–8608, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 262, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 262, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 315, + 262, + 553, + 307 + ], + "type": "text", + "content": "[38] Zhiheng Liu, Ruili Feng, Kai Zhu, Yifei Zhang, Kecheng Zheng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones: Concept Neurons in Diffusion Models for Customized Generation. In Int. Conf. Mach. Learn., 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 553, + 352 + ], + "type": "text", + "content": "[39] Zhiheng Liu, Yifei Zhang, Yujun Shen, Kecheng Zheng, Kai Zhu, Ruili Feng, Yu Liu, Deli Zhao, Jingren Zhou, and Yang Cao. Cones 2: Customizable Image Synthesis with Multiple Subjects. In Adv. Neural Inform. Process. Syst., 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 353, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 353, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 553, + 397 + ], + "type": "text", + "content": "[40] Yue Ma, Yingqing He, Xiaodong Cun, Xintao Wang, Siran Chen, Ying Shan, Xiu Li, and Qifeng Chen. Follow Your Pose: Pose-Guided Text-to-Video Generation using Pose-Free Videos. In Assoc. Adv. Artif. Intell., 2024. 6, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 397, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 397, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 315, + 397, + 553, + 442 + ], + "type": "text", + "content": "[41] Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. ACE++: Instruction-Based Image Creation and Editing via Context-Aware Content Filling. arXiv preprint arXiv:2501.02487, 2025. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 442, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 442, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 315, + 442, + 553, + 487 + ], + "type": "text", + "content": "[42] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 487, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 487, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 315, + 487, + 553, + 508 + ], + "type": "text", + "content": "[43] Midjourney. Midjourney, https://www.midjourney.com, 2023.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 509, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 509, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 315, + 509, + 553, + 531 + ], + "type": "text", + "content": "[44] MiniMax. Hailuo AI Video, https://hailuoai.com/video, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 532, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 532, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 532, + 553, + 555 + ], + "type": "text", + "content": "[45] OpenAI. DALL·E 3, https://openai.com/dall-e-3, 2023.2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 555, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 553, + 600 + ], + "type": "text", + "content": "[46] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold. In ACM SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 553, + 644 + ], + "type": "text", + "content": "[47] Yulin Pan, Chaojie Mao, Zeyinzi Jiang, Zhen Han, and Jingfeng Zhang. Locate, Assign, Refine: Taming Customized Image Inpainting with Text-Subject Guidance. arXiv preprint arXiv:2403.19534, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 644, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 644, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 644, + 553, + 677 + ], + "type": "text", + "content": "[48] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers. In Int. Conf. Comput. Vis., pages 4195-4305, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 679, + 529, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 679, + 529, + 690 + ], + "spans": [ + { + "bbox": [ + 315, + 679, + 529, + 690 + ], + "type": "text", + "content": "[49] PiKa. PiKa, https://pika.art/, 2025.6, 7, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "text", + "content": "[50] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17200" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "Xiong, Silvio Savarese, Stefano Ermon, Yun Fu, and Ran Xu. UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild. In Adv. Neural Inform. Process. Syst., 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "text", + "content": "[51] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer. IEEE Trans. Pattern Anal. Mach. Intell., pages 1623-1637, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 249 + ], + "type": "text", + "content": "[52] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. In Int. Conf. Learn. Represent., 2025. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "type": "text", + "content": "[54] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. Med. Image Comput. Computer-Assisted Interv., 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 358 + ], + "type": "text", + "content": "[55] Runway. Gen-3, https://app.runwayml.com/video-tools, 2025.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "text", + "content": "[56] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising Diffusion Implicit Models. In Int. Conf. Learn. Represent., 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[57] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In Int. Conf. Learn. Represent., 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[58] StabilityAI. Stable Diffusion v2-1 Model Card, https: //huggingface.co/stabilityai/stable-diffusion-2-1,2022.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "type": "text", + "content": "[59] StabilityAI. Stable Diffusion XL Model Card, https: //huggingface.co/stabilityai/stable-diffusion-x1-base-1.0,2022.2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 525 + ], + "type": "text", + "content": "[60] StabilityAI. CosXL Model Card, https://huggingface.co/stabilityai/cosxl, 2024.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "text", + "content": "[61] Ya Sheng Sun, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. ImageBrush: Learning Visual In-Context Instructions for Exemplar-Based Image Manipulation. In Adv. Neural Inform. Process. Syst., 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 647 + ], + "type": "text", + "content": "[62] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-Robust Large Mask Inpainting With Fourier Convolutions. In IEEE Winter Conf. Appl. Comput. Vis., pages 2149-2159, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[63] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. OminiControl: Minimal and Universal Control for Diffusion Transformer. arXiv preprint arXiv:2411.15098, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[64] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2, 6, 13" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[65] Zachary Teed and Jia Deng. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In Eur. Conf. Comput. Vis., pages 402-419, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 544, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 544, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 544, + 118 + ], + "type": "text", + "content": "[66] Vidu. Vidu, https://www.vidu.cn/, 2025. 6, 7, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[67] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv preprint arXiv:2401.07519, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 164, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 219 + ], + "type": "text", + "content": "[68] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. VideoComposer: Compositional Video Synthesis with Motion Controllability. In Adv. Neural Inform. Process. Syst., 2023. 2, 6, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. MotionCtrl: A Unified and Flexible Motion Controller for Video Generation. In ACM SIGGRAPH, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 265, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 331 + ], + "type": "text", + "content": "[70] Yujie Wei, Shiwei Zhang, Hangjie Yuan, Xiang Wang, Haonan Qiu, Rui Zhao, Yutong Feng, Feng Liu, Zhizhong Huang, Jiaxin Ye, Yingya Zhang, and Hongming Shan. DreamVideo-2: Zero-Shot Subject-Driven Video Customization with Precise Motion Control. arXiv preprint arXiv:2410.13830, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[71] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Chaofan Li, Shuting Wang, Tiejun Huang, and Zheng Liu. OmniGen: Unified Image Generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "text", + "content": "[72] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective Whole-body Pose Estimation with Two-stages Distillation. In Int. Conf. Comput. Vis., pages 4210-4220, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 422, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 498 + ], + "type": "text", + "content": "[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. In Int. Conf. Learn. Represent., 2025. 2, 6, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "type": "text", + "content": "[74] Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yu-jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-Preserving Text-to-Video Generation by Frequency Decomposition. In IEEE Conf. Comput. Vis. Pattern Recog., 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "text", + "content": "[75] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. In Adv. Neural Inform. Process. Syst., 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In Int. Conf. Comput. Vis., pages 3836–3847, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[77] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv preprint arXiv:2311.04145, 2023. 2, 6, 13" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[78] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17201" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 296 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize Anything: A Strong Image Tagging Model. arXiv preprint arXiv:2306.03514, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[79] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. ControlVideo: Training-free Controllable Text-to-Video Generation. In Int. Conf. Learn. Represent., 2024. 6, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[80] Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. Magic Mirror: ID-Preserved Video Generation in Video Diffusion Transformers. arXiv preprint arXiv:2411.13503, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "text", + "content": "[81] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based Fine-Grained Image Editing at Scale. arXiv preprint arXiv:2407.05282v1, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "text", + "content": "[82] Shangchen Zhou, Chongyi Li, Kelvin C. K. Chan, and Chen Change Loy. ProPainter: Improving Propagation and Transformer for Video Inpainting. In Int. Conf. Comput. Vis., pages 10477-10486, 2023. 2, 6, 13" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17202" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_content_list.json b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6e034fef1947ce0b4022309e24cc6923f8498e4c --- /dev/null +++ b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_content_list.json @@ -0,0 +1,1493 @@ +[ + { + "type": "text", + "text": "VAFlow: Video-to-Audio Generation with Cross-Modality Flow Matching", + "text_level": 1, + "bbox": [ + 127, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xihua Wang $^{1}$ Xin Cheng $^{1}$ Yuyue Wang $^{1}$ Ruihua Song $^{1*}$ Yunfeng Wang $^{2}$ $^{1}$ Gaoling School of Artificial Intelligence, Renmin University of China \nxihuaw@ruc.edu.cn, songruihua.bloon@outlook.com", + "bbox": [ + 112, + 178, + 887, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even textaugmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching.", + "bbox": [ + 86, + 301, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 681, + 220, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-to-audio (V2A) generation plays a vital role in multimedia content generation, including Foley for films and AI-generated silent videos. In recent years, this task has attracted increasing attention within the generative community. Early V2A approaches [15, 24, 26, 29, 34] adopted language modeling strategies by discretizing audio into tokens, and used parallel mask-prediction models (BERT-like) or token-by-token autoregressive models (GPT-like) borrowed from NLP community as shown in Figure 1. However, these methods, reliant on discrete tokens, incur inherent information losses by the discretization process.", + "bbox": [ + 89, + 707, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00b726f5ccb3264f0217ff01f4838780de47120879ff2588490bb01d049831e2.jpg", + "image_caption": [ + "Figure 1. Four different V2A paradigms, along with state-of-the-art performance under each. The first row shows discrete token-based approaches (mask-prediction and autoregressive based generation), which are limited by information loss during tokenization. The second row presents methods generating continuous data. Previous diffusion and flow methods (bottom-left) model the noise-to-audio mapping conditioning on video. We find this approach suboptimal and demonstrate that directly modeling the video-to-audio transformation (bottom-right) yields better results." + ], + "image_footnote": [], + "bbox": [ + 516, + 266, + 906, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Diffusion and Flow Matching (FM) models in image [4, 10, 33] and video [3, 31] generation have spurred similar approaches for V2A [9, 25, 41, 43, 45]. Generally, these methods model V2A as a conditional generation task, transforming Gaussian noise conditioned on video into audio, as shown in Figure 1 (bottom-left). The generation process typically involves two steps: (1) sampling a random noise from Gaussian prior and (2) using this noise and video condition as input to iteratively denoise the noise into audio. This approach avoids discrete audio tokens, directly fitting the continuous distribution of audio data, and currently holds state-of-the-art results.", + "bbox": [ + 511, + 641, + 908, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing diffusion/FM-based V2A methods focus solely on the second step, i.e., denoising Gaussian noise into audio, assuming the first step is trivial. Even state-of-the-art methods [40, 43] begin with Gaussian noise, but we find that this approach is not optimal. Unlike text-to", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11777", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image or text-to-video tasks, where textual conditions are abstract and highly defined by human, the video conditions in V2A are more complex and require both spatial and temporal understanding. The generation goal also demands a high degree of alignment between the generated audio and the video conditions, not just semantically, but also temporally. There are two inherent difficulties: (1) different conditions on the same noise require the model to generate different audio outputs, demanding advanced condition processing [8, 41, 46], and (2) noise variations under the same condition lead to unstable generation results, prompting research into methods for noise control [44], such as finding a \"golden noise\" for consistent results [47].", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, we propose a new perspective by shifting focus to the first step of the generation process. We question: If there are better priors than Gaussian noise? Given that many diffusion and FM models dedicate significant effort to integrating video conditions into the denoising process for better generation performance, we further question: If V2A can work better when directly transforming the video distribution to audio rather than from noise? Fortunately, both answers are affirmative.", + "bbox": [ + 89, + 290, + 480, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce VAFlow, a novel framework for V2A generation, which directly denoises from the video space to the audio space. To implement this video-to-audio paradigm, we adopt the FM framework, as it theoretically supports any source distribution. There are three technical challenges for this implementation: (1) aligning source and target distributions: video and audio feature spaces typically have different time and spatial resolutions, yet FM requires the source and target distributions to be aligned (the same shape). (2) uncertainty of the feature type: identifying the optimal video representations/feature types as the source distribution. (3) conditioning mechanism: exploring additional video information integration ways beyond encoding it in the source distribution. In VAFlow, we first employ an alignment VAE to adjust video features to match the audio latent space. It then explores three different video feature types (semantic CLIP [32], spatiotemporal CAVP [25], reconstruction-focused VidTok [38]) to systematically investigate source video information. Experiments show that using video features both as a source distribution and as conditioning information yields the best results for VAFlow.", + "bbox": [ + 89, + 429, + 482, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, the recent success of large language models (LLMs) [13, 28] has highlighted the effectiveness of scaling laws, and various generative works consequently explore similar scaling properties in each domain [30, 42]. However, existing V2A models have not demonstrated clear scaling trends, where increasing model size does not consistently achieve improvement [8, 43]. In contrast, our experiments show that VAFlow exhibits clear scaling property, with model size increases leading to consistent improvements in generation quality, further confirming its potential.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Without elaborate training tricks or external data (e.g., text-audio pairs), VAFlow achieves state-of-the-art results on standard benchmarks (VGGSound), outperforming even text-augmented models in generation fidelity, diversity, and cross-modality alignment with input video. Playable audi-los generated with VAFlow can be accessed at the demo page: https://vaflow.github.io/demo.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 212, + 653, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Flow Matching and Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 238, + 834, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [17, 27, 35] reconstruct data distributions by iteratively denoising Gaussian-sampled latents, governmented by various designed schedulers achieving state-of-the-art performance in multimodal generation tasks. Unlike diffusions, flow-based models [1, 20, 22, 23] establish deterministic mappings between source (e.g., Gaussian) and target distributions through invertible neural transformations, with recent flow matching optimizing vector fields through optimal transport theory, enabling direct learning of straight transport paths between source and target distributions, thereby reducing training instability and inference steps. While both paradigms typically rely on Gaussian priors as their source distribution, existing video-to-audio systems on both paradigms model noise-to-audio generation conditioned on video features. In contrast, our approach directly aligns video latent spaces with target audio distributions via linear transformations under a flow matching framework, bypassing iterative noise modeling, preserving efficiency, and ensuring precise frame-to-frame alignment.", + "bbox": [ + 511, + 260, + 903, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video-to-Audio Generation", + "text_level": 1, + "bbox": [ + 511, + 559, + 758, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Framework. Recent advances in generative models have driven significant progress in video-to-audio (V2A) synthesis. Current V2A methods fall into three main paradigms: (1) Autoregressive (AR) models [15, 26, 34] discretize audio into token (codec) sequences and utilize video features as prefixes to model unified video-audio sequences. While straightforward, their token-by-token generation process results in quadratic inference time scaling, limiting efficiency for long audio sequences. (2) Mask-based generation paradigms [24, 29, 37] also employ discrete tokens but decode audio in parallel by predicting masked tokens. Video features serve as the condition for mask-to-audio prediction. However, both AR and mask-based methods suffer from an inherent upper bound of fidelity due to lossy compression from audio discretization. (3) Flow/diffusion-based methods [7-9, 25, 40, 41, 43, 45, 46] leverage continuous representations and currently achieve state-of-the-art V2A performance [8, 40, 43]. These techniques iteratively transform noise into target audio distributions while conditioning on video features to steer the denoising trajectory.", + "bbox": [ + 511, + 582, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Conditional Modeling. Unlike typical conditioned", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11778", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generation tasks like text-to-image [33], V2A involves fully aligned sequential correspondences between video and audio. Existing methods address this through various strategies: some AR-based methods investigate various attention mechanisms and masking schemes for temporally sequential alignment [26]; some mask-based approaches integrate hybrid ways to fuse video features into the mask-prediction process to synchronize audio generation [29]; and some flow/diffusion-based methods augment their pipelines with auxiliary modules (e.g., loudness or audio layout predictors [41, 46]) to enforce temporal consistency. In contrast, our approach builds on flow-based techniques by discarding the conventional noise-to-audio paradigm and complex video-conditioning designs. Instead, we propose a more intuitive method that directly transforms the video temporal space to the audio temporal space via a flow transport.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method:", + "text_level": 1, + "bbox": [ + 89, + 345, + 187, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Flow Matching", + "text_level": 1, + "bbox": [ + 89, + 369, + 349, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $x_0 \\sim p(x_0)$ and $x_1 \\sim p(x_1)$ denote samples from source and target distributions, respectively, with $c$ representing an optional conditioning signal. Flow Matching (FM) [20, 23] establishes an inter-distribution transport process governed by the ordinary differential equation (ODE):", + "bbox": [ + 89, + 392, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd x (t) = v _ {\\theta} \\left(x _ {t}, t, c\\right) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 479, + 482, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\in [0,1]$ denotes the continuous time step, and $v_{\\theta}$ represents a neural network-parameterized velocity field. The condition $c$ is incorporated for conditional flow implementations. Optimal transport flows enforce velocity alignment with the direction $(x_{1} - x_{0})$ of the linear path pointing from $x_{0}$ to $x_{1}$ through regression:", + "bbox": [ + 89, + 506, + 483, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, x _ {0}, x _ {1}, c} \\| (x _ {1} - x _ {0}) - v _ {\\theta} (x _ {t}, t, c) \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 607, + 482, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{t} = (1 - t)x_{0} + tx_{1}$ constitutes the linear interpolation between $x_{0}$ and $x_{1}$ . The learned velocity field enables target generation by solving the ODE in Equation 1.", + "bbox": [ + 89, + 636, + 482, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous video-to-audio flow approaches [8, 43] (and related diffusion methods [25, 41]) use a Gaussian prior $p(x_0) = \\mathcal{N}(0,1)$ for the source distribution, with audio as $p(x_{1})$ and video as conditioning input $c$ . During training, random noise $x_0$ is paired with ground-truth audio $x_{1}$ , prompting $v_{\\theta}$ to model the paths from noise to audio under visual conditions.", + "bbox": [ + 89, + 681, + 482, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work challenges the Gaussian prior by proposing a direct video-to-audio transport. Theoretically, FM permits arbitrary source distributions, motivating our formulation:", + "bbox": [ + 89, + 787, + 483, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd x (t) = v _ {\\theta} \\left(x _ {t}, t\\right) d t, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 843, + 482, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{t} = (1 - t)x_{v} + tx_{a}$ linearly interpolates between paired video and audio samples $x_{v},x_{a}\\sim p(v,a)$", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. VAFlow Architecture", + "text_level": 1, + "bbox": [ + 513, + 90, + 714, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Overview. As illustrated in Figure 2, the input video features $x_{v}$ are extracted via a visual encoder, while the target audio waveform can be obtained from latent representations $x_{a}$ through WaveVAE. VAFlow attempts to establish FM between the video feature space $x_{v}$ and audio latent space $x_{a}$ , enabling direct audio generation from video features via ODE solving instead of Gaussian noise initialization. However, since FM requires consistent dimensional shapes between source and target distributions (see Equation 3), the inherent spatiotemporal resolution mismatch between $x_{v}$ and $x_{a}$ necessitates shape alignment. To address this, we employ a 1D VAE (namely alignment VAE) that compresses interpolated $x_{v}$ into aligned features $x_{v}^{\\mathrm{AE}}$ matching $x_{a}$ 's dimensions.", + "bbox": [ + 511, + 114, + 906, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A diffusion transformer (DiT) subsequently works as a velocity field estimator to capture the transport dynamics between $x_{v}^{\\mathrm{AE}}$ and $x_{a}$ . This enables continuous trajectory sampling from $x_{v}^{\\mathrm{AE}}$ to $x_{a}$ through ODE integration. The sampled latent $\\hat{x}_{a}$ is finally decoded by WaveVAE to produce the output audio. The following contents detail: the visual/audio representations $(x_{v}, x_{v}^{\\mathrm{AE}}, x_{a})$ , and the velocity field estimator in VAFlow.", + "bbox": [ + 511, + 330, + 905, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video and Audio Representations. For an input video-audio pair $(v, a)$ , the audio waveform $a$ is compressed by a pretrained 1D WaveVAE's [11] encoder into $x_{a} \\in \\mathbb{R}^{T_{a} \\times D_{a}}$ . This latent $x_{a}$ serves as the target for flow generation and is later decoded to reconstruct the waveform.", + "bbox": [ + 511, + 453, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For video representation, a pretrained visual encoder encodes $v$ into $x_v \\in \\mathbb{R}^{T_v \\times D_v}$ (e.g., CLIP [32], CAVP [25], or VidTok [38] as encoder). Given the discrepancies (e.g., $x_a \\in \\mathbb{R}^{215 \\times 64}$ versus $x_v \\in \\mathbb{R}^{100 \\times 768}$ for a 10-second clip), we propose an alignment VAE $\\epsilon_{\\phi}$ to further compress and align video latents with audio latents' space while preserving essential visual information. Specifically, we utilize a 1D convolutional encoder-decoder as $\\epsilon_{\\phi}$ to interpolate and project $x_v$ into a variational space $x_v^{\\mathrm{AE}} \\in \\mathbb{R}^{T_a \\times D_a}$ , matching the shape of $x_a$ . The encoder of $\\epsilon_{\\phi}$ predicts the parameters $\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}} = E_{\\epsilon_{\\phi}}(x_v)$ for a Gaussian distribution, from which $x_v^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}})$ is sampled. The decoder reconstructs $x_v$ as $\\hat{x}_v = D_{\\epsilon_{\\phi}}(x_v^{\\mathrm{AE}})$ . This process ensures dimensional alignment with audio latents while retaining critical spatiotemporal features of the original video representation.", + "bbox": [ + 511, + 532, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Velocity Estimator. Built upon the diffusion transformer (DiT) architecture [11], the velocity estimator processes two key inputs: time embeddings of $t$ and the interpolated state $x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a}$ , following the Equation 3. The time embedding is concatenated with the audio sequence to form an input tensor $\\in \\mathbb{R}^{(T_a + 1)\\times D_a}$ , which is processed through stacked transformer blocks. The first $T_{a}$ dimensions of the output sequence constitute the pre", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11779", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ff1ff64fb37921e45705859648878aee89608f3208a30432c279d825228be95d.jpg", + "image_caption": [ + "Figure 2. VAFlow Architecture. The figure illustrates VAFow in the inference manner. A visual encoder initially extracts temporal features $x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}}$ from variable-frame-rate video inputs. Then, a 1D alignment VAE (Variational Autoencoder) $\\epsilon_{\\phi}$ compresses interpolated $x_{v}$ into audio-aligned latent features $x_{v}^{\\mathrm{AE}}$ with matched dimensions $(T_{v}^{\\mathrm{AE}} \\times D_{v}^{\\mathrm{AE}} = T_{a} \\times D_{a})$ . Next, a diffusion transformer $(v_{\\theta})$ predicts the velocity field for the transport $(x_{a} - x_{v}^{\\mathrm{AE}})$ , and by solving the corresponding ODE (see Equation 3), an estimated $\\hat{x}_{a}$ is obtained. Finally, the WaveVAE decoder converts $\\hat{x}_{a}$ into the waveform representing the predicted audio. Note that during diffusion transformer prediction, $x_{v}$ can optionally be incorporated via cross-attention, detail discussion is conducted in the experiment Section 4.5." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 898, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dicted velocity $v_{\\theta}(x_t, t)$ , with 1D Rotary Position Embeddings (RoPE [36]) to preserve positional awareness in attention layers. Notably, while the interpolated state $x_t$ implicitly encodes visual information through $x_v^{\\mathrm{AE}}$ , this signal diminishes as $t \\to 1$ . To address this limitation, we retain cross-attention layers in the transformer blocks that persistently integrate the original video features $x_v$ . Further studies (Section 4.5) demonstrate that this architectural choice is critical to enhance visual perception throughout the flow trajectory and improve audio generation quality.", + "bbox": [ + 88, + 349, + 480, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. VAFlow Training and Inference", + "text_level": 1, + "bbox": [ + 89, + 510, + 372, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct three-stage training: (1) pretrain the alignment VAE, (2) train the VAFlow estimator with frozen alignment VAE, and (3) jointly fine-tune both alignment VAE and estimator. We detail training and inference as follows:", + "bbox": [ + 89, + 532, + 482, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alignment VAE Training. Given encoded visual features $x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}}$ , the alignment VAE encoder predicts distribution parameters $(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}})$ , while the decoder reconstructs $\\hat{x}_{v}$ from a sampled $x_{v}^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}})$ . The model optimizes three objectives to preserve visual semantics while regularizing $x_{v}^{\\mathrm{AE}}$ within $\\mathbb{R}^{T_{a} \\times D_{a}}$ Gaussian space:", + "bbox": [ + 89, + 593, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- KL Regularization: enforces Gaussian latent priors via:", + "bbox": [ + 89, + 688, + 470, + 702 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(\\mathcal {N} \\left(\\mu_ {x _ {0} ^ {\\mathrm {A E}}}, \\sigma_ {x _ {0} ^ {\\mathrm {A E}}}\\right) \\| \\mathcal {N} (0, I)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 714, + 482, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Reconstruction Loss: ensures feature-level consistency:", + "bbox": [ + 89, + 742, + 470, + 757 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\left\\| x _ {v} - \\hat {x} _ {v} \\right\\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 768, + 482, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Contrastive Alignment: strengthens feature preservation:", + "bbox": [ + 89, + 797, + 478, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {s \\in \\mathcal {S}} \\log \\frac {\\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x _ {v}\\right) / \\tau\\right)}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 821, + 482, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S = \\{\\mathrm{frame},\\mathrm{clip},\\mathrm{global}\\}$ defines temporal granularities and $\\tau = 0.7$ serves as a temperature parameter [6].", + "bbox": [ + 102, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For scale $s$ , $f_{s}(\\cdot)$ extracts features via three ways: frame-level: single-frame sampling + mean pooling; clip-level: $K = 8$ -adjacent-frame sampling + pooling; global-level: full temporal mean pooling at $T_{v}$ dimension of feature $x$ . Negative samples $x'$ are drawn from training batch $\\mathcal{B}$ .", + "bbox": [ + 524, + 349, + 906, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The composite objective integrates these terms:", + "bbox": [ + 513, + 428, + 826, + 443 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V A E}} = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {\\mathrm {r e c}} \\mathcal {L} _ {\\mathrm {r e c}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 453, + 903, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\lambda_{\\mathrm{KL}} = 8\\times 10^{-4}$ $\\lambda_{\\mathrm{rec}} = 1.0$ , and $\\lambda_{\\mathrm{con}} = 1.0$", + "bbox": [ + 511, + 478, + 839, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Velocity Estimator Training Given encoded features $x_{v}$ , $x_{v}^{\\mathrm{AE}}$ , and $x_{a}$ , we train the velocity estimator $v_{\\theta}$ to predict the linear transport direction $x_{a} - x_{v}^{\\mathrm{AE}}$ through:", + "bbox": [ + 511, + 494, + 905, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} \\left\\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 547, + 903, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a}$ . For the conditioned version, visual features $x_{v}$ are incorporated via cross-attention (with a $10\\%$ probability of zeroing out $x_{v}$ during training). The vanilla version omits this conditioning, simplifying the input of estimator in Equation 8 to $v_{\\theta}(x_t,t)$ .", + "bbox": [ + 511, + 578, + 905, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint Turning After separate training of the alignment VAE and the velocity estimator, a joint optimization stage is performed. In this stage, the VAE encoder and $v_{\\theta}$ are co-trained (discarding the VAE decoder), while the WaveVAE and visual feature extractor remain frozen. For a pair $(v, a)$ , the visual extractor and WaveVAE produce $x_{v}$ and $x_{a}$ . The VAE encoder compresses $x_{v}$ to obtain $x_{v}^{\\mathrm{AE}} \\sim q_{\\phi}(x_{v}^{\\mathrm{AE}}|x_{v})$ which together with $x_{a}$ and a random $t \\sim \\mathcal{U}(0,1)$ forms $x_{t}$ . The joint objective integrates KL regularization in Equation 4 with velocity prediction in Equation 8:", + "bbox": [ + 511, + 655, + 906, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta , \\phi_ {\\text {E n c}}) = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\mathbb {E} \\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\| _ {2} ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 813, + 906, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference Upon training, VAFlow synthesizes audio $a$ from an input video $v$ through the following pipeline: The visual extractor first encodes $v$ into a feature $x_{v}$ , which is", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11780", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "compressed by the alignment VAE encoder into distribution parameters $(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}})$ . A latent sample $x_{v}^{\\mathrm{AE}}\\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}}^2)$ is drawn as the initial state $(x_0)$ for solving the ODE in Equation 1 with learned estimator. The predicted $\\hat{x}_a(x_1)$ is then decoded into waveform via the frozen WaveVAE. During inference, the variational sampling of $x_{v}^{\\mathrm{AE}}$ introduces stochasticity - even with fixed ODE integration paths, distinct samples yield diverse outputs.", + "bbox": [ + 89, + 90, + 480, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For conditional velocity estimators, classifier-free guidance (CFG) [14] enhances synthesis quality by interpolating conditioned and unconditioned velocity predictions:", + "bbox": [ + 89, + 215, + 480, + 260 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\theta} ^ {\\mathrm {C F G}} \\left(x _ {t}, t, x _ {v}\\right) = (1 + \\gamma) v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) - \\gamma v _ {\\theta} \\left(x _ {t}, t, \\emptyset\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 268, + 480, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ controls guidance strength and $\\varnothing$ denotes null (zero) conditioning.", + "bbox": [ + 89, + 311, + 480, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 356, + 222, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 380, + 279, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We employ two datasets: VGGSound [5] and AudioSet-V2A [25]. For VGGSound, we adhere to the original split, which comprises 186k training samples and 15k testing samples. AudioSet-V2A contains 390k videos that are exclusively used for training. During training, the alignment VAE is trained on a combined set of VGGSound training and AudioSet-V2A to yield more generalized visual feature compression. The subsequent velocity estimator training and joint tuning are conducted solely on the VGGSound training set. At inference, we evaluate on the VGGSound test set as in prior work.", + "bbox": [ + 89, + 402, + 480, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For the estimator's DiT architecture, we adapted the DiT implementation from StableAudio-Open-1.0 [11] (SDA1.0) by eliminating duration encoding, applying global time encoding solely to time steps, and restricting the cross-attention layer's key/value inputs to video features. For the visual extractor, we use three variants: CLIP [32] (a widely adopted visual semantic extractor in V2A tasks), CAVP [25] (which captures both visual semantic and temporal features), and VidTok [38] (a visual tokenizer for reconstruction). We train three alignment VAEs with comparable parameter scales—one for each visual extractor. In the velocity estimator training stage, the alignment VAE remains frozen; in the joint training stage, both the estimator and the alignment VAE are trained. In all stages, the audio WaveVAE and the visual extractor are kept frozen. Notably, we do not perform any further rectified iterations on the estimator; all VAFlow variants are trained for a single rectified iteration (Rectified-1, RF-1). During sampling, we employ ODE solvers of various orders—first-order Euler, second-order Midpoint, and fifth-order dopri5—with differing sampling steps and classifier-free guidance scales.", + "bbox": [ + 89, + 569, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We employ both objective and subjective evaluations across three dimensions: generation quality, audiovisual relevance, and audiovisual synchronization. For objective evaluation, generation quality is measured using FD score [21], KL divergence [15], and IS score [21]. In previous work, each metric may be computed with different feature extraction models (e.g., Melception, PasST $^1$ ). Audiovisual relevance is quantified by computing the cosine similarity between the embeddings of the input video and the generated audio using the ImageBind model [12]. For synchronization, we use the synchronization accuracy (Acc.) [25] reflects the alignment between the input video and generated audio.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Our method is comprehensively compared against all major generative paradigms, including autoregressive (AR), masked prediction (Mask), diffusion-based, and flow-based approaches, as detailed in Table 1. All baseline models were evaluated using either officially released code and models or directly downloaded audio outputs by their authors, with consistent experimental settings maintained across all evaluations on the same hardware platform. Implementation specifics include: (1) Im2Wav generates audio in 4s clips, truncated to 10s after three clips. (2) Diff-Foley here employs its double-guidance variant (demonstrating superior performance); (3) MMAudio-L here is the largest variant of MMAudio; and (4) Frieren here employs its final no-reflow configuration (achieving optimal Fréchet distance). For comprehensive evaluation, we also include text-augmented V2A models (denoted in gray rows) that leverage external textual data beyond benchmark audiovisual corpora during training or inference.", + "bbox": [ + 511, + 287, + 903, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Diffusion vs. Flow vs. VAFlow Models", + "text_level": 1, + "bbox": [ + 511, + 571, + 841, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We examine framework-specific performance by training same VAFlow-CLIP-142M models under three paradigms, maintaining identical configurations: CLIP visual features, architecture parameters, initialization weights, and other hyperparameters. The frameworks include: Diffusion: Standard DDPM [27, 33] with cross-attention video conditioning. Flow: Standard continuous-time flow (Equation 2) using Gaussian prior. VAFlow: Our framework employing aligned VAE-compressed video features as source distribution. All frameworks use their respective optimal inference settings (DDPM: DPMSolver; Flows: dopri5). Joint tuning in VAFlow begins at 100K steps.", + "bbox": [ + 511, + 594, + 903, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frechet Distance (FD) results in Figure 3 reveals: 1) Flow paradigms converge faster with lower FD than diffusion, demonstrating inherent optimization benefits. 2) VAFlow attains lower final FD than standard flow despite delayed early-stage convergence (<100K steps). This per", + "bbox": [ + 511, + 776, + 903, + 851 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1We adopt PasST [19] model over PANN [18] for its improved robustness following the practice introduced in AudioLDM [21] evaluation repository: https://github.com/haoheliu/audioldm_eval.", + "bbox": [ + 511, + 862, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11781", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ddbe06928bea183e718f5bdb4312d42104ea62742c41fc44089c0837a7187e51.jpg", + "image_caption": [ + "Figure 3. Comparison of different training frameworks. We examine the same model under different training paradigms by controlling training data, steps, and other hyperparameters. Results demonstrate that VAFlow converges to superior optima, validating the effectiveness of its direct video-to-audio transformation design and highlighting this paradigm advantages in V2A generation." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 436, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formance attributes to VAFlow's explicit modeling of deterministic video-audio transport paths, avoiding the ambiguous noise-to-audio mappings in Gaussian-prior approaches. VAFlow learns optimal transport trajectories between video-audio distribution, bypassing error-prone intermediate transformations in Gaussian-noise-to-audio paths.", + "bbox": [ + 89, + 363, + 483, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. VAFlow Scaling Analysis", + "text_level": 1, + "bbox": [ + 89, + 465, + 318, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We investigate VAFlow's scaling capabilities by developing four model variants (142M, 264M, 527M, 1.05B parameters) through width/depth scaling of DiT blocks while maintaining fixed training protocols (data, steps, etc.). Evaluation metrics (FD, IS) versus parameter counts are visualized in Figure 4 (bubble sizes denote model scales).", + "bbox": [ + 89, + 488, + 482, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are two key findings here: 1) Parameter Efficiency: VAFlow consistently outperforms flow-based baselines (Frieren, MMAudio) at comparable sizes (e.g., 142M vs. 159M, 1.05B vs. 1.03B) across both metrics. 2) Scaling Capability: Previous flow-based studies have struggled to show consistent scalability [8, 43], where larger models fail to achieve better metrics consistently. To the best of our knowledge, VAFlow is the first flow-based V2A framework to demonstrate stable scaling properties. As the parameter size increases, the models achieve better FD and IS scores. Interestingly, larger models tend to improve IS scores more significantly than FD scores, suggesting that as the VAFlow model scales, the quality of generated audio gradually saturates, and the model becomes more capable of deriving diverse audio from different video-sampled latents.", + "bbox": [ + 89, + 579, + 482, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Benchmark Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 277, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluations are presented in Table 1, where VAFlow demonstrates state-of-the-art performance across key audio generation metrics: $\\mathrm{FD}_{\\text {pas. }}$ , $\\mathrm{FD}_{\\text {mel. }}$ , $\\mathrm{IS}_{\\text {pas. }}$ , $\\mathrm{KL}_{\\text {pas. }}$ , and $\\mathrm{KL}_{\\text {mel. }}$ . For audio-visual synchronization,", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f09fb5f06231f54fb4c79ca7f5346440e949c46f53aa2c0415e37e982158f883.jpg", + "image_caption": [ + "Figure 4. Performance of flow-based baselines and VAFlow variants in scale. VAFlow models (142M, 1.05B) consistently outperform comparable-scale baselines (Frieren-159M, MMAudio-1.03B), with stable scaling property indicated by progressive IS gains (diversity) and saturating FD improvements (quality)." + ], + "image_footnote": [], + "bbox": [ + 562, + 90, + 859, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VAFlow achieves $96.3\\%$ accuracy (Acc.), comparable to Frieren $(97.0\\%)$ . In semantic relevance, it attains an IB-VA score of 28.6, second only to TiVA's 30.3. Notably, VAFlow maintains a simple and straightforward model design without relying on complex video conditioning modules [8, 41, 46], carefully designed noise schedules [9], or additional guidance modules [25, 45]. Its performance across various metrics, achieved with a standard Flow Matching training approach and basic ODE solvers, underscores the simplicity and effectiveness of the direct video-to-audio mapping design. When compared to text-enhanced baselines (gray rows), VAFlow maintains advantages in $\\mathrm{FD}_{\\text{pas.}}$ , $\\mathrm{FD}_{\\text{mel.}}$ , $\\mathrm{IS}_{\\text{pas.}}$ , and synchronization Acc., though text-augmented models exhibit superior audio semantic metrics (KL distances and IB-VA). This aligns with expectations since text conditioning explicitly constrains semantic space. We note that Seeing&Hearing's elevated IB-VA score (36.7) stems from classifier guidance using the same ImageBind model, which however compromises generation quality and synchronization performance.", + "bbox": [ + 511, + 369, + 906, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further conduct human evaluation with sampled 50 videos from the test set to compare our method with the flow-based baseline, Frieren. The generated results from both models were randomized, and 10 experts were invited to score the models based on three criteria: sound quality (Quality), audio-visual semantic consistency (Semantic) and audio-visual synchronization (Sync.). The scoring was done by selecting a winner or declaring a tie for each comparison. The results, summarized in Table 2, show that our model outperforms the baseline in all three dimensions.", + "bbox": [ + 511, + 672, + 908, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablative Analysis", + "text_level": 1, + "bbox": [ + 511, + 832, + 684, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Features. Three visual feature types were experimented with: semantic features (CLIP embeddings extracted per frame at 10 fps), temporally-enhanced semantic", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11782", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b5c13a95ac5535f0a7bfdf235fa260cbc27c52e8be3881895d050442f03d78b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ParadigmMethodVis.Feat.QualitySync. Acc. ↑Semantic IB-VA ↑
\\(FD_{pas.}\\)↓\\(FD_{mel.}\\)↓\\(IS_{mel.}\\)↑\\(KL_{pas.}\\)↓\\(KL_{mel.}\\)↓
ARSpecVQGAN [15]RGB,Opti.342.418.1520.53.106.8455.914.1
Im2Wav [34]CLIP252.213.0335.82.265.4178.919.6
MaskVATT [24]eva-CLIP+154.17.0764.31.343.4684.825.0
VAB [37]eva-CLIP193.112.2134.42.315.0982.425.9
DiffusionMultiFoley [7]CAVP241.613.0067.71.593.3485.227.0
Seeing&Hearing [45]ImageBind+251.512.630.12.656.3754.236.7
Diff-Foley [25]CAVP512.111.4550.42.916.1888.020.7
LoVA [9]CLIP149.16.9058.72.105.1287.026.3
FoleyCrafter [46]CLIP134.27.4056.62.295.6583.527.8
TiVA [41]CLIP111.94.3167.02.025.2187.530.3
V2A-Mapper [40]CLIP94.94.2164.92.466.0678.622.2
FlowMMAudio-L [8]CLIP+159.29.3845.41.953.8477.533.9
Frieren [43]CAVP259.96.7052.42.925.6497.023.0
VAFlow (ours)CLIP87.73.8673.61.914.8188.128.6
VAFlow (ours)CAVP91.84.7364.82.415.9696.325.1
VAFlow (ours)VidTok128.56.1748.13.648.9449.912.9
", + "bbox": [ + 94, + 88, + 908, + 380 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb0cc149a1b8270bb5b2fdc91aff8972184a97bd26578f2298e92afecff4b222.jpg", + "table_caption": [ + "Table 1. Quantitative video-to-audio results on the VGGSound test set. “Vis. Feat.” indicates the employed visual features, with a “+” denoting the inclusion of auxiliary visual information (e.g., visual captions via LLama-2 [39] or Qwen-VL [2] in VATT and Seeing&Hearing, or SyncFormer [16] features in MMAudio). All results are obtained from official code or released audio outputs. The MultiFoley results here is an 8k subset filtered by ImageBind from the authors. Rows in gray denote methods that incorporate extra text data (e.g., video captions); for fairness, these results are provided for reference only and are not directly compared with standard video-to-audio models. The best score is bolded, and the second-best is underlined. Our three VAFlow variants, differing in visual features, achieve definitive gains in audio quality (FD, IS, KL) while delivering near state-of-the-art performance for synchronization (Acc.) and correlation (IB-VA)." + ], + "table_footnote": [], + "table_body": "
ComparisonQualityWin rate(%) Sync.Semantic
Ours vs. Frieren66.00(±4.18)60.44(±5.72)60.89(±4.48)
", + "bbox": [ + 96, + 516, + 477, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Human evaluation results, with comparison between our model and the flow-based baseline Frieren across three criteria: sound quality, audio-visual synchronization, and audio-visual semantic consistency. Ours outperforms the baseline in all criteria.", + "bbox": [ + 89, + 582, + 482, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "features (CAVP features via video-audio contrastive learning at 4 fps), reconstruction features (VidTok tokenized latents at 15 fps). Utilizing the same model size and training process for these features, we developed three variants: VAFlow-CLIP, VAFlow-CAVP, and VAFlow-VidTok, with results presented in the last three rows of Table 1. The results indicate that the widely-used CLIP features overall perform the best, yielding superior audio quality and semantic relevance. The temporally enhanced CAVP features, on the other hand, demonstrate the best video-audio synchronization but slightly lag in other metrics compared to VAFlow-CLIP. Models based on reconstruction features exhibit the poorest performance across all evaluated aspects. These outcomes suggest two key insights: 1) The type of video features directly influences VAFlow's performance,", + "bbox": [ + 88, + 672, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with different feature types leading to models with distinct performance preferences (either better semantic performance or synchronization); 2) Modality alignment matters more than information preservation. Although reconstruction features retain the most comprehensive video information among the three, their lack of modality alignment hampers the V2A model's ability to learn effective cross-modal generative capabilities.", + "bbox": [ + 511, + 518, + 906, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Flow Matching Source Distributions. We evaluate VAFlow under two source distributions: video latent priors from alignment VAE and standard Gaussian noise. For video prior, we further test variants with/without joint training of the alignment VAE encoder (Method 3.3). All models share same training budgets (300K steps for non-joint training variants; 170K+130K for joint training).", + "bbox": [ + 511, + 641, + 908, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results in Table 3 demonstrate that video latent priors trained solely under alignment VAE objective in Equation 7 fail to surpass Gaussian noise baselines. Only when coupled with joint training—adapting the prior to flow-matching tasks—do video latent priors outperform Gaussian prior. This highlights the necessity of co-optimizing prior distributions for flow-based generation rather than relying on standalone pretraining on reconstruction task.", + "bbox": [ + 511, + 748, + 908, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditioning Mechanisms and Guidance. Prior work highlights the critical role of condition and guidance mech", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11783", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/72a285c0140b286cc05b89f5a361a88df831d6a8cdedaf0be2c027104301c080.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
V-PriorTuned\\(FD_{pas.}\\)↓\\(IS_{mel.}\\)↑Acc.↑IB-VA ↑
X-102.667.687.927.9
✓X111.857.482.325.4
✓✓87.773.688.128.6
", + "bbox": [ + 107, + 88, + 468, + 157 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b3747903e3af065be5805bc844baf66fb78cff9f80dc36e605eea64dee29096e.jpg", + "table_caption": [ + "Table 3. Ablation study on flow-matching source distributions. 'V-Prior' denotes sampling starts from video latent priors (vs. Gaussian priors). 'Tuned' indicates joint optimization of the alignment VAE encoder. Results show that jointly tuned video priors outperform Gaussian priors as source distributions for V2A flow models." + ], + "table_footnote": [], + "table_body": "
Cond.CACFGFDpas. ↓ISmel. ↑Acc. ↑IB-VA ↑
XX-135.941.431.78.5
X✓X160.536.518.95.4
✓✓X133.842.378.021.0
✓✓✓87.773.688.128.6
", + "bbox": [ + 96, + 268, + 478, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on conditioning mechanisms and guidance. 'Cond.' denotes explicit video conditioning during inference; 'CA' indicates the presence of cross-attention layers of different variants; 'CFG' refers to classifier-free guidance. Results demonstrate that optimal performance requires both explicit conditioning and CFG when solving VAFlow ODE.", + "bbox": [ + 89, + 361, + 482, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "anisms (e.g., classifier-free guidance, CFG) in enhancing model performance. We address two questions:", + "bbox": [ + 89, + 489, + 482, + 521 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Can VAFlow generate high-quality audio using only video priors, omitting explicit video conditioning during sampling? We experimented with two variants without explicit condition: No cross-attention (CA): Remove CA layers in DiT blocks and retrain the model (Table 4, Row 1). Zero-input CA: Retain CA layers but zero out video inputs during inference (Table 4, Row 2). 'No CA' variant exhibit comparable audio quality metrics (FD, IS) to the video-condition-visible variant (Table 4, Row 3: explicit video conditioning without CFG), yet suffer significant degradation in semantic relevance (IB-VA) and synchronization (Acc.). This indicates that while video priors partially preserve visual information, explicit video conditioning during sampling is essential for temporal and semantic alignment. We think that as the latent trajectory approaches the target audio distribution $(t\\to 1)$ , initial visual cues in the prior degrade, necessitating direct video information.", + "(2) Does explicit video conditioning during sampling still require CFG for optimal performance? Adding CFG to the video-condition-visible variant (Row 4 in Table 4) further improves all metrics, demonstrating VAFlow's compatibility with established guidance techniques. This underscores the dual necessity of explicit video conditioning and CFG for optimal V2A synthesis with VAFlow framework." + ], + "bbox": [ + 89, + 529, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/541834df9a135b3f673f2135fc672b6335d888f01e8ee58e170b788fb154322c.jpg", + "image_caption": [ + "A case video:", + "At a seaside with rolling waves, a group of people are swimming and playing. A woman walks toward the camera, points to the right, and says something.", + "Figure 5. Mel-spectrograms of audio generated by VAFlow and the baseline for a complex seaside scene, including background waves and varying voices, demonstrating VAFlow's superior content understanding and synchronization with the video." + ], + "image_footnote": [], + "bbox": [ + 516, + 118, + 903, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Case Studies", + "text_level": 1, + "bbox": [ + 513, + 544, + 648, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 illustrates the performance of VAFlow and the baselines on the complex scene video. The video depicts a seaside scene with various sounds include background waves and different voices associated with shifting characters (noisy laughter, woman speaking). We present the ground truth and the mel-spectrograms of the audio generated by different methods. Our results show that VAFlow not only accurately understands the video content and generates all the necessary sounds but also maintains synchronization with the visual timing.", + "bbox": [ + 511, + 566, + 906, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 737, + 782, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce VAFlow, a novel framework for V2A task, which directly denoises from the video space to the audio space via flow matching. demonstrating superior performance compared to previous methods. Future work will explore VAFlow's potential in more diverse audio domains, such as speech and music. Expanding the range of tasks and datasets will enable VAFlow's scaling properties to shine even further, laying a more general-purpose the foundation model for content generation community.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11784", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (No. 62276268), Public Computing Cloud, Renmin University of China and ZHI-TECH GROUP.", + "bbox": [ + 89, + 114, + 483, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 2", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 7", + "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. 1", + "[4] Li-Wen Chang, Wenlei Bao, Qi Hou, Chengquan Jiang, Ningxin Zheng, Yinmin Zhong, Xuanrun Zhang, Zuquan Song, Chengji Yao, Ziheng Jiang, et al. Flux: fast software-based communication overlap on gpus through kernel fusion. arXiv preprint arXiv:2406.06858, 2024. 1", + "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, pages 721–725, 2020. 5", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 4", + "[7] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. arXiv preprint arXiv:2411.17698, 2024. 2, 7", + "[8] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. arXiv preprint arXiv:2412.15322, 2024. 2, 3, 6, 7", + "[9] Xin Cheng, Xihua Wang, Yihan Wu, Yuyue Wang, and Ruihua Song. Lova: Long-form video-to-audio generation. arXiv preprint arXiv:2409.15157, 2024. 1, 2, 6, 7", + "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 1", + "[11] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv preprint arXiv:2407.14358, 2024. 3, 5", + "[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, pages 15180-15190, 2023. 5" + ], + "bbox": [ + 93, + 213, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2", + "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[15] Vladimir iashin and Esa Rahtu. Taming visually guided sound generation. In BMVC, 2021. 1, 2, 5, 7", + "[16] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, pages 5325-5329, 2024. 7", + "[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577, 2022. 2", + "[18] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5", + "[19] Khaled Koutini, Jan Schlüter, Hamid Eghbal-Zadeh, and Gerhard Widmer. Efficient training of audio transformers with patchout. arXiv preprint arXiv:2110.05069, 2021. 5", + "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3", + "[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In International Conference on Machine Learning, pages 21450-21474, 2023. 5", + "[22] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In CVPR, pages 2755-2765, 2025. 2", + "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3", + "[24] Xiulong Liu, Kun Su, and Eli Shlizerman. Tell what you hear from what you see - video to audio generation through text. In Advances in Neural Information Processing Systems, pages 101337-101366, 2024. 1, 2, 7", + "[25] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In Advances in Neural Information Processing Systems, pages 48855-48876, 2023. 1, 2, 3, 5, 6, 7", + "[26] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing, pages 1-6, 2024. 1, 2, 3", + "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171, 2021. 2, 5" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11785", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text. arXiv preprint arXiv:2410.18514, 2024. 2", + "[29] Santiago Pascual, Chunghsin Yeh, Ioannis Tsiamas, and Joan Serra. Masked generative video-to-audio transformers with enhanced synchronicity. In ECCV, pages 247-264, 2024. 1, 2, 3", + "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2", + "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 5", + "[34] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP, pages 1-5, 2023. 1, 2, 7", + "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2", + "[36] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4", + "[37] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In International Conference on Machine Learning, pages 46804–46822, 2024. 2, 7" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 2, 3, 5", + "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7", + "[40] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, pages 15492-15501, 2024. 1, 2, 7", + "[41] Xihua Wang, Yuyue Wang, Yihan Wu, Ruihua Song, Xu Tan, Zehua Chen, Hongteng Xu, and Guodong Sui. Tiva: Time-aligned video-to-audio generation. In ACM MM, pages 573-582, 2024. 1, 2, 3, 6, 7", + "[42] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR, pages 6572-6582, 2024. 2", + "[43] Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation network with rectified flow matching. In Advances in Neural Information Processing Systems, pages 128118-128138, 2024. 1, 2, 3, 6, 7", + "[44] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In ECCV, pages 378-394, 2024. 2", + "[45] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, pages 7151-7161, 2024. 1, 2, 6, 7", + "[46] Yiming Zhang, Yicheng Gu, Yanhong Zeng, Zhening Xing, Yuancheng Wang, Zhizheng Wu, and Kai Chen. Foley-crafter: Bring silent videos to life with lifelike and synchronized sounds. arXiv preprint arXiv:2407.01494, 2024. 2, 3, 6, 7", + "[47]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.2" + ], + "bbox": [ + 516, + 90, + 903, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11786", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_model.json b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b496c1302517c1b066e60bbf5b36fa9841d4dbf4 --- /dev/null +++ b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "VAFlow: Video-to-Audio Generation with Cross-Modality Flow Matching" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.179, + 0.888, + 0.234 + ], + "angle": 0, + "content": "Xihua Wang\\(^{1}\\) Xin Cheng\\(^{1}\\) Yuyue Wang\\(^{1}\\) Ruihua Song\\(^{1*}\\) Yunfeng Wang\\(^{2}\\) \n\\(^{1}\\)Gaoling School of Artificial Intelligence, Renmin University of China \nxihuaw@ruc.edu.cn, songruihua.bloon@outlook.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.303, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even textaugmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.682, + 0.222, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Video-to-audio (V2A) generation plays a vital role in multimedia content generation, including Foley for films and AI-generated silent videos. In recent years, this task has attracted increasing attention within the generative community. Early V2A approaches [15, 24, 26, 29, 34] adopted language modeling strategies by discretizing audio into tokens, and used parallel mask-prediction models (BERT-like) or token-by-token autoregressive models (GPT-like) borrowed from NLP community as shown in Figure 1. However, these methods, reliant on discrete tokens, incur inherent information losses by the discretization process." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.267, + 0.907, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.612 + ], + "angle": 0, + "content": "Figure 1. Four different V2A paradigms, along with state-of-the-art performance under each. The first row shows discrete token-based approaches (mask-prediction and autoregressive based generation), which are limited by information loss during tokenization. The second row presents methods generating continuous data. Previous diffusion and flow methods (bottom-left) model the noise-to-audio mapping conditioning on video. We find this approach suboptimal and demonstrate that directly modeling the video-to-audio transformation (bottom-right) yields better results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.824 + ], + "angle": 0, + "content": "Recent advancements in Diffusion and Flow Matching (FM) models in image [4, 10, 33] and video [3, 31] generation have spurred similar approaches for V2A [9, 25, 41, 43, 45]. Generally, these methods model V2A as a conditional generation task, transforming Gaussian noise conditioned on video into audio, as shown in Figure 1 (bottom-left). The generation process typically involves two steps: (1) sampling a random noise from Gaussian prior and (2) using this noise and video condition as input to iteratively denoise the noise into audio. This approach avoids discrete audio tokens, directly fitting the continuous distribution of audio data, and currently holds state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, existing diffusion/FM-based V2A methods focus solely on the second step, i.e., denoising Gaussian noise into audio, assuming the first step is trivial. Even state-of-the-art methods [40, 43] begin with Gaussian noise, but we find that this approach is not optimal. Unlike text-to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11777" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "image or text-to-video tasks, where textual conditions are abstract and highly defined by human, the video conditions in V2A are more complex and require both spatial and temporal understanding. The generation goal also demands a high degree of alignment between the generated audio and the video conditions, not just semantically, but also temporally. There are two inherent difficulties: (1) different conditions on the same noise require the model to generate different audio outputs, demanding advanced condition processing [8, 41, 46], and (2) noise variations under the same condition lead to unstable generation results, prompting research into methods for noise control [44], such as finding a \"golden noise\" for consistent results [47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.482, + 0.426 + ], + "angle": 0, + "content": "In contrast, we propose a new perspective by shifting focus to the first step of the generation process. We question: If there are better priors than Gaussian noise? Given that many diffusion and FM models dedicate significant effort to integrating video conditions into the denoising process for better generation performance, we further question: If V2A can work better when directly transforming the video distribution to audio rather than from noise? Fortunately, both answers are affirmative." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.747 + ], + "angle": 0, + "content": "We introduce VAFlow, a novel framework for V2A generation, which directly denoises from the video space to the audio space. To implement this video-to-audio paradigm, we adopt the FM framework, as it theoretically supports any source distribution. There are three technical challenges for this implementation: (1) aligning source and target distributions: video and audio feature spaces typically have different time and spatial resolutions, yet FM requires the source and target distributions to be aligned (the same shape). (2) uncertainty of the feature type: identifying the optimal video representations/feature types as the source distribution. (3) conditioning mechanism: exploring additional video information integration ways beyond encoding it in the source distribution. In VAFlow, we first employ an alignment VAE to adjust video features to match the audio latent space. It then explores three different video feature types (semantic CLIP [32], spatiotemporal CAVP [25], reconstruction-focused VidTok [38]) to systematically investigate source video information. Experiments show that using video features both as a source distribution and as conditioning information yields the best results for VAFlow." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Moreover, the recent success of large language models (LLMs) [13, 28] has highlighted the effectiveness of scaling laws, and various generative works consequently explore similar scaling properties in each domain [30, 42]. However, existing V2A models have not demonstrated clear scaling trends, where increasing model size does not consistently achieve improvement [8, 43]. In contrast, our experiments show that VAFlow exhibits clear scaling property, with model size increases leading to consistent improvements in generation quality, further confirming its potential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Without elaborate training tricks or external data (e.g., text-audio pairs), VAFlow achieves state-of-the-art results on standard benchmarks (VGGSound), outperforming even text-augmented models in generation fidelity, diversity, and cross-modality alignment with input video. Playable audi-los generated with VAFlow can be accessed at the demo page: https://vaflow.github.io/demo." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.213, + 0.655, + 0.229 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.239, + 0.836, + 0.255 + ], + "angle": 0, + "content": "2.1. Flow Matching and Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.905, + 0.549 + ], + "angle": 0, + "content": "Diffusion models [17, 27, 35] reconstruct data distributions by iteratively denoising Gaussian-sampled latents, governmented by various designed schedulers achieving state-of-the-art performance in multimodal generation tasks. Unlike diffusions, flow-based models [1, 20, 22, 23] establish deterministic mappings between source (e.g., Gaussian) and target distributions through invertible neural transformations, with recent flow matching optimizing vector fields through optimal transport theory, enabling direct learning of straight transport paths between source and target distributions, thereby reducing training instability and inference steps. While both paradigms typically rely on Gaussian priors as their source distribution, existing video-to-audio systems on both paradigms model noise-to-audio generation conditioned on video features. In contrast, our approach directly aligns video latent spaces with target audio distributions via linear transformations under a flow matching framework, bypassing iterative noise modeling, preserving efficiency, and ensuring precise frame-to-frame alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.759, + 0.575 + ], + "angle": 0, + "content": "2.2. Video-to-Audio Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Framework. Recent advances in generative models have driven significant progress in video-to-audio (V2A) synthesis. Current V2A methods fall into three main paradigms: (1) Autoregressive (AR) models [15, 26, 34] discretize audio into token (codec) sequences and utilize video features as prefixes to model unified video-audio sequences. While straightforward, their token-by-token generation process results in quadratic inference time scaling, limiting efficiency for long audio sequences. (2) Mask-based generation paradigms [24, 29, 37] also employ discrete tokens but decode audio in parallel by predicting masked tokens. Video features serve as the condition for mask-to-audio prediction. However, both AR and mask-based methods suffer from an inherent upper bound of fidelity due to lossy compression from audio discretization. (3) Flow/diffusion-based methods [7-9, 25, 40, 41, 43, 45, 46] leverage continuous representations and currently achieve state-of-the-art V2A performance [8, 40, 43]. These techniques iteratively transform noise into target audio distributions while conditioning on video features to steer the denoising trajectory." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Video Conditional Modeling. Unlike typical conditioned" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11778" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.334 + ], + "angle": 0, + "content": "generation tasks like text-to-image [33], V2A involves fully aligned sequential correspondences between video and audio. Existing methods address this through various strategies: some AR-based methods investigate various attention mechanisms and masking schemes for temporally sequential alignment [26]; some mask-based approaches integrate hybrid ways to fuse video features into the mask-prediction process to synchronize audio generation [29]; and some flow/diffusion-based methods augment their pipelines with auxiliary modules (e.g., loudness or audio layout predictors [41, 46]) to enforce temporal consistency. In contrast, our approach builds on flow-based techniques by discarding the conventional noise-to-audio paradigm and complex video-conditioning designs. Instead, we propose a more intuitive method that directly transforms the video temporal space to the audio temporal space via a flow transport." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.188, + 0.363 + ], + "angle": 0, + "content": "3. Method:" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.371, + 0.35, + 0.387 + ], + "angle": 0, + "content": "3.1. Preliminary: Flow Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.471 + ], + "angle": 0, + "content": "Let \\( x_0 \\sim p(x_0) \\) and \\( x_1 \\sim p(x_1) \\) denote samples from source and target distributions, respectively, with \\( c \\) representing an optional conditioning signal. Flow Matching (FM) [20, 23] establishes an inter-distribution transport process governed by the ordinary differential equation (ODE):" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.48, + 0.483, + 0.498 + ], + "angle": 0, + "content": "\\[\nd x (t) = v _ {\\theta} \\left(x _ {t}, t, c\\right) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.6 + ], + "angle": 0, + "content": "where \\( t \\in [0,1] \\) denotes the continuous time step, and \\( v_{\\theta} \\) represents a neural network-parameterized velocity field. The condition \\( c \\) is incorporated for conditional flow implementations. Optimal transport flows enforce velocity alignment with the direction \\( (x_{1} - x_{0}) \\) of the linear path pointing from \\( x_{0} \\) to \\( x_{1} \\) through regression:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.608, + 0.483, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, x _ {0}, x _ {1}, c} \\| (x _ {1} - x _ {0}) - v _ {\\theta} (x _ {t}, t, c) \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.681 + ], + "angle": 0, + "content": "where \\( x_{t} = (1 - t)x_{0} + tx_{1} \\) constitutes the linear interpolation between \\( x_{0} \\) and \\( x_{1} \\). The learned velocity field enables target generation by solving the ODE in Equation 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Previous video-to-audio flow approaches [8, 43] (and related diffusion methods [25, 41]) use a Gaussian prior \\( p(x_0) = \\mathcal{N}(0,1) \\) for the source distribution, with audio as \\( p(x_{1}) \\) and video as conditioning input \\( c \\). During training, random noise \\( x_0 \\) is paired with ground-truth audio \\( x_{1} \\), prompting \\( v_{\\theta} \\) to model the paths from noise to audio under visual conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.833 + ], + "angle": 0, + "content": "This work challenges the Gaussian prior by proposing a direct video-to-audio transport. Theoretically, FM permits arbitrary source distributions, motivating our formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.844, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nd x (t) = v _ {\\theta} \\left(x _ {t}, t\\right) d t, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\( x_{t} = (1 - t)x_{v} + tx_{a} \\) linearly interpolates between paired video and audio samples \\( x_{v},x_{a}\\sim p(v,a) \\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.715, + 0.106 + ], + "angle": 0, + "content": "3.2. VAFlow Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Model Overview. As illustrated in Figure 2, the input video features \\( x_{v} \\) are extracted via a visual encoder, while the target audio waveform can be obtained from latent representations \\( x_{a} \\) through WaveVAE. VAFlow attempts to establish FM between the video feature space \\( x_{v} \\) and audio latent space \\( x_{a} \\), enabling direct audio generation from video features via ODE solving instead of Gaussian noise initialization. However, since FM requires consistent dimensional shapes between source and target distributions (see Equation 3), the inherent spatiotemporal resolution mismatch between \\( x_{v} \\) and \\( x_{a} \\) necessitates shape alignment. To address this, we employ a 1D VAE (namely alignment VAE) that compresses interpolated \\( x_{v} \\) into aligned features \\( x_{v}^{\\mathrm{AE}} \\) matching \\( x_{a} \\)'s dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.906, + 0.451 + ], + "angle": 0, + "content": "A diffusion transformer (DiT) subsequently works as a velocity field estimator to capture the transport dynamics between \\( x_{v}^{\\mathrm{AE}} \\) and \\( x_{a} \\). This enables continuous trajectory sampling from \\( x_{v}^{\\mathrm{AE}} \\) to \\( x_{a} \\) through ODE integration. The sampled latent \\( \\hat{x}_{a} \\) is finally decoded by WaveVAE to produce the output audio. The following contents detail: the visual/audio representations \\( (x_{v}, x_{v}^{\\mathrm{AE}}, x_{a}) \\), and the velocity field estimator in VAFlow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.529 + ], + "angle": 0, + "content": "Video and Audio Representations. For an input video-audio pair \\((v, a)\\), the audio waveform \\(a\\) is compressed by a pretrained 1D WaveVAE's [11] encoder into \\(x_{a} \\in \\mathbb{R}^{T_{a} \\times D_{a}}\\). This latent \\(x_{a}\\) serves as the target for flow generation and is later decoded to reconstruct the waveform." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.778 + ], + "angle": 0, + "content": "For video representation, a pretrained visual encoder encodes \\( v \\) into \\( x_v \\in \\mathbb{R}^{T_v \\times D_v} \\) (e.g., CLIP [32], CAVP [25], or VidTok [38] as encoder). Given the discrepancies (e.g., \\( x_a \\in \\mathbb{R}^{215 \\times 64} \\) versus \\( x_v \\in \\mathbb{R}^{100 \\times 768} \\) for a 10-second clip), we propose an alignment VAE \\( \\epsilon_{\\phi} \\) to further compress and align video latents with audio latents' space while preserving essential visual information. Specifically, we utilize a 1D convolutional encoder-decoder as \\( \\epsilon_{\\phi} \\) to interpolate and project \\( x_v \\) into a variational space \\( x_v^{\\mathrm{AE}} \\in \\mathbb{R}^{T_a \\times D_a} \\), matching the shape of \\( x_a \\). The encoder of \\( \\epsilon_{\\phi} \\) predicts the parameters \\( \\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}} = E_{\\epsilon_{\\phi}}(x_v) \\) for a Gaussian distribution, from which \\( x_v^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}}) \\) is sampled. The decoder reconstructs \\( x_v \\) as \\( \\hat{x}_v = D_{\\epsilon_{\\phi}}(x_v^{\\mathrm{AE}}) \\). This process ensures dimensional alignment with audio latents while retaining critical spatiotemporal features of the original video representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Velocity Estimator. Built upon the diffusion transformer (DiT) architecture [11], the velocity estimator processes two key inputs: time embeddings of \\( t \\) and the interpolated state \\( x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a} \\), following the Equation 3. The time embedding is concatenated with the audio sequence to form an input tensor \\( \\in \\mathbb{R}^{(T_a + 1)\\times D_a} \\), which is processed through stacked transformer blocks. The first \\( T_{a} \\) dimensions of the output sequence constitute the pre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.9, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Figure 2. VAFlow Architecture. The figure illustrates VAFow in the inference manner. A visual encoder initially extracts temporal features \\( x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}} \\) from variable-frame-rate video inputs. Then, a 1D alignment VAE (Variational Autoencoder) \\( \\epsilon_{\\phi} \\) compresses interpolated \\( x_{v} \\) into audio-aligned latent features \\( x_{v}^{\\mathrm{AE}} \\) with matched dimensions \\( (T_{v}^{\\mathrm{AE}} \\times D_{v}^{\\mathrm{AE}} = T_{a} \\times D_{a}) \\). Next, a diffusion transformer \\( (v_{\\theta}) \\) predicts the velocity field for the transport \\( (x_{a} - x_{v}^{\\mathrm{AE}}) \\), and by solving the corresponding ODE (see Equation 3), an estimated \\( \\hat{x}_{a} \\) is obtained. Finally, the WaveVAE decoder converts \\( \\hat{x}_{a} \\) into the waveform representing the predicted audio. Note that during diffusion transformer prediction, \\( x_{v} \\) can optionally be incorporated via cross-attention, detail discussion is conducted in the experiment Section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.482, + 0.502 + ], + "angle": 0, + "content": "dicted velocity \\( v_{\\theta}(x_t, t) \\), with 1D Rotary Position Embeddings (RoPE [36]) to preserve positional awareness in attention layers. Notably, while the interpolated state \\( x_t \\) implicitly encodes visual information through \\( x_v^{\\mathrm{AE}} \\), this signal diminishes as \\( t \\to 1 \\). To address this limitation, we retain cross-attention layers in the transformer blocks that persistently integrate the original video features \\( x_v \\). Further studies (Section 4.5) demonstrate that this architectural choice is critical to enhance visual perception throughout the flow trajectory and improve audio generation quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.373, + 0.527 + ], + "angle": 0, + "content": "3.3. VAFlow Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.594 + ], + "angle": 0, + "content": "We conduct three-stage training: (1) pretrain the alignment VAE, (2) train the VAFlow estimator with frozen alignment VAE, and (3) jointly fine-tune both alignment VAE and estimator. We detail training and inference as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Alignment VAE Training. Given encoded visual features \\( x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}} \\), the alignment VAE encoder predicts distribution parameters \\( (\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}}) \\), while the decoder reconstructs \\( \\hat{x}_{v} \\) from a sampled \\( x_{v}^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}}) \\). The model optimizes three objectives to preserve visual semantics while regularizing \\( x_{v}^{\\mathrm{AE}} \\) within \\( \\mathbb{R}^{T_{a} \\times D_{a}} \\) Gaussian space:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.689, + 0.472, + 0.703 + ], + "angle": 0, + "content": "- KL Regularization: enforces Gaussian latent priors via:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.715, + 0.483, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(\\mathcal {N} \\left(\\mu_ {x _ {0} ^ {\\mathrm {A E}}}, \\sigma_ {x _ {0} ^ {\\mathrm {A E}}}\\right) \\| \\mathcal {N} (0, I)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.743, + 0.472, + 0.758 + ], + "angle": 0, + "content": "- Reconstruction Loss: ensures feature-level consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.769, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\left\\| x _ {v} - \\hat {x} _ {v} \\right\\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.798, + 0.479, + 0.813 + ], + "angle": 0, + "content": "- Contrastive Alignment: strengthens feature preservation:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.822, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {s \\in \\mathcal {S}} \\log \\frac {\\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x _ {v}\\right) / \\tau\\right)}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(S = \\{\\mathrm{frame},\\mathrm{clip},\\mathrm{global}\\}\\) defines temporal granularities and \\(\\tau = 0.7\\) serves as a temperature parameter [6]." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.351, + 0.907, + 0.427 + ], + "angle": 0, + "content": "For scale \\( s \\), \\( f_{s}(\\cdot) \\) extracts features via three ways: frame-level: single-frame sampling + mean pooling; clip-level: \\( K = 8 \\)-adjacent-frame sampling + pooling; global-level: full temporal mean pooling at \\( T_{v} \\) dimension of feature \\( x \\). Negative samples \\( x' \\) are drawn from training batch \\( \\mathcal{B} \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.429, + 0.828, + 0.444 + ], + "angle": 0, + "content": "The composite objective integrates these terms:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.454, + 0.905, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V A E}} = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {\\mathrm {r e c}} \\mathcal {L} _ {\\mathrm {r e c}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.84, + 0.494 + ], + "angle": 0, + "content": "with \\(\\lambda_{\\mathrm{KL}} = 8\\times 10^{-4}\\) \\(\\lambda_{\\mathrm{rec}} = 1.0\\) , and \\(\\lambda_{\\mathrm{con}} = 1.0\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.495, + 0.906, + 0.54 + ], + "angle": 0, + "content": "Velocity Estimator Training Given encoded features \\( x_{v} \\), \\( x_{v}^{\\mathrm{AE}} \\), and \\( x_{a} \\), we train the velocity estimator \\( v_{\\theta} \\) to predict the linear transport direction \\( x_{a} - x_{v}^{\\mathrm{AE}} \\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.549, + 0.905, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} \\left\\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.906, + 0.655 + ], + "angle": 0, + "content": "where \\( x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a} \\). For the conditioned version, visual features \\( x_{v} \\) are incorporated via cross-attention (with a \\( 10\\% \\) probability of zeroing out \\( x_{v} \\) during training). The vanilla version omits this conditioning, simplifying the input of estimator in Equation 8 to \\( v_{\\theta}(x_t,t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.806 + ], + "angle": 0, + "content": "Joint Turning After separate training of the alignment VAE and the velocity estimator, a joint optimization stage is performed. In this stage, the VAE encoder and \\( v_{\\theta} \\) are co-trained (discarding the VAE decoder), while the WaveVAE and visual feature extractor remain frozen. For a pair \\( (v, a) \\), the visual extractor and WaveVAE produce \\( x_{v} \\) and \\( x_{a} \\). The VAE encoder compresses \\( x_{v} \\) to obtain \\( x_{v}^{\\mathrm{AE}} \\sim q_{\\phi}(x_{v}^{\\mathrm{AE}}|x_{v}) \\) which together with \\( x_{a} \\) and a random \\( t \\sim \\mathcal{U}(0,1) \\) forms \\( x_{t} \\). The joint objective integrates KL regularization in Equation 4 with velocity prediction in Equation 8:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.814, + 0.907, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta , \\phi_ {\\text {E n c}}) = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\mathbb {E} \\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\| _ {2} ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Inference Upon training, VAFlow synthesizes audio \\(a\\) from an input video \\(v\\) through the following pipeline: The visual extractor first encodes \\(v\\) into a feature \\(x_{v}\\), which is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.215 + ], + "angle": 0, + "content": "compressed by the alignment VAE encoder into distribution parameters \\((\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}})\\). A latent sample \\(x_{v}^{\\mathrm{AE}}\\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}}^2)\\) is drawn as the initial state \\((x_0)\\) for solving the ODE in Equation 1 with learned estimator. The predicted \\(\\hat{x}_a(x_1)\\) is then decoded into waveform via the frozen WaveVAE. During inference, the variational sampling of \\(x_{v}^{\\mathrm{AE}}\\) introduces stochasticity - even with fixed ODE integration paths, distinct samples yield diverse outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.482, + 0.261 + ], + "angle": 0, + "content": "For conditional velocity estimators, classifier-free guidance (CFG) [14] enhances synthesis quality by interpolating conditioned and unconditioned velocity predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.269, + 0.482, + 0.302 + ], + "angle": 0, + "content": "\\[\nv _ {\\theta} ^ {\\mathrm {C F G}} \\left(x _ {t}, t, x _ {v}\\right) = (1 + \\gamma) v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) - \\gamma v _ {\\theta} \\left(x _ {t}, t, \\emptyset\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.482, + 0.344 + ], + "angle": 0, + "content": "where \\(\\gamma\\) controls guidance strength and \\(\\varnothing\\) denotes null (zero) conditioning." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.357, + 0.223, + 0.373 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.381, + 0.281, + 0.397 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.482, + 0.569 + ], + "angle": 0, + "content": "Datasets. We employ two datasets: VGGSound [5] and AudioSet-V2A [25]. For VGGSound, we adhere to the original split, which comprises 186k training samples and 15k testing samples. AudioSet-V2A contains 390k videos that are exclusively used for training. During training, the alignment VAE is trained on a combined set of VGGSound training and AudioSet-V2A to yield more generalized visual feature compression. The subsequent velocity estimator training and joint tuning are conducted solely on the VGGSound training set. At inference, we evaluate on the VGGSound test set as in prior work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For the estimator's DiT architecture, we adapted the DiT implementation from StableAudio-Open-1.0 [11] (SDA1.0) by eliminating duration encoding, applying global time encoding solely to time steps, and restricting the cross-attention layer's key/value inputs to video features. For the visual extractor, we use three variants: CLIP [32] (a widely adopted visual semantic extractor in V2A tasks), CAVP [25] (which captures both visual semantic and temporal features), and VidTok [38] (a visual tokenizer for reconstruction). We train three alignment VAEs with comparable parameter scales—one for each visual extractor. In the velocity estimator training stage, the alignment VAE remains frozen; in the joint training stage, both the estimator and the alignment VAE are trained. In all stages, the audio WaveVAE and the visual extractor are kept frozen. Notably, we do not perform any further rectified iterations on the estimator; all VAFlow variants are trained for a single rectified iteration (Rectified-1, RF-1). During sampling, we employ ODE solvers of various orders—first-order Euler, second-order Midpoint, and fifth-order dopri5—with differing sampling steps and classifier-free guidance scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Metrics. We employ both objective and subjective evaluations across three dimensions: generation quality, audiovisual relevance, and audiovisual synchronization. For objective evaluation, generation quality is measured using FD score [21], KL divergence [15], and IS score [21]. In previous work, each metric may be computed with different feature extraction models (e.g., Melception, PasST\\(^1\\)). Audiovisual relevance is quantified by computing the cosine similarity between the embeddings of the input video and the generated audio using the ImageBind model [12]. For synchronization, we use the synchronization accuracy (Acc.) [25] reflects the alignment between the input video and generated audio." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.56 + ], + "angle": 0, + "content": "Baselines. Our method is comprehensively compared against all major generative paradigms, including autoregressive (AR), masked prediction (Mask), diffusion-based, and flow-based approaches, as detailed in Table 1. All baseline models were evaluated using either officially released code and models or directly downloaded audio outputs by their authors, with consistent experimental settings maintained across all evaluations on the same hardware platform. Implementation specifics include: (1) Im2Wav generates audio in 4s clips, truncated to 10s after three clips. (2) Diff-Foley here employs its double-guidance variant (demonstrating superior performance); (3) MMAudio-L here is the largest variant of MMAudio; and (4) Frieren here employs its final no-reflow configuration (achieving optimal Fréchet distance). For comprehensive evaluation, we also include text-augmented V2A models (denoted in gray rows) that leverage external textual data beyond benchmark audiovisual corpora during training or inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.843, + 0.587 + ], + "angle": 0, + "content": "4.2. Diffusion vs. Flow vs. VAFlow Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.905, + 0.775 + ], + "angle": 0, + "content": "We examine framework-specific performance by training same VAFlow-CLIP-142M models under three paradigms, maintaining identical configurations: CLIP visual features, architecture parameters, initialization weights, and other hyperparameters. The frameworks include: Diffusion: Standard DDPM [27, 33] with cross-attention video conditioning. Flow: Standard continuous-time flow (Equation 2) using Gaussian prior. VAFlow: Our framework employing aligned VAE-compressed video features as source distribution. All frameworks use their respective optimal inference settings (DDPM: DPMSolver; Flows: dopri5). Joint tuning in VAFlow begins at 100K steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.905, + 0.852 + ], + "angle": 0, + "content": "Frechet Distance (FD) results in Figure 3 reveals: 1) Flow paradigms converge faster with lower FD than diffusion, demonstrating inherent optimization benefits. 2) VAFlow attains lower final FD than standard flow despite delayed early-stage convergence (<100K steps). This per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.863, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1We adopt PasST [19] model over PANN [18] for its improved robustness following the practice introduced in AudioLDM [21] evaluation repository: https://github.com/haoheliu/audioldm_eval." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11781" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.091, + 0.437, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.485, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Comparison of different training frameworks. We examine the same model under different training paradigms by controlling training data, steps, and other hyperparameters. Results demonstrate that VAFlow converges to superior optima, validating the effectiveness of its direct video-to-audio transformation design and highlighting this paradigm advantages in V2A generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.456 + ], + "angle": 0, + "content": "formance attributes to VAFlow's explicit modeling of deterministic video-audio transport paths, avoiding the ambiguous noise-to-audio mappings in Gaussian-prior approaches. VAFlow learns optimal transport trajectories between video-audio distribution, bypassing error-prone intermediate transformations in Gaussian-noise-to-audio paths." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.466, + 0.32, + 0.483 + ], + "angle": 0, + "content": "4.3. VAFlow Scaling Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.579 + ], + "angle": 0, + "content": "We investigate VAFlow's scaling capabilities by developing four model variants (142M, 264M, 527M, 1.05B parameters) through width/depth scaling of DiT blocks while maintaining fixed training protocols (data, steps, etc.). Evaluation metrics (FD, IS) versus parameter counts are visualized in Figure 4 (bubble sizes denote model scales)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.807 + ], + "angle": 0, + "content": "There are two key findings here: 1) Parameter Efficiency: VAFlow consistently outperforms flow-based baselines (Frieren, MMAudio) at comparable sizes (e.g., 142M vs. 159M, 1.05B vs. 1.03B) across both metrics. 2) Scaling Capability: Previous flow-based studies have struggled to show consistent scalability [8, 43], where larger models fail to achieve better metrics consistently. To the best of our knowledge, VAFlow is the first flow-based V2A framework to demonstrate stable scaling properties. As the parameter size increases, the models achieve better FD and IS scores. Interestingly, larger models tend to improve IS scores more significantly than FD scores, suggesting that as the VAFlow model scales, the quality of generated audio gradually saturates, and the model becomes more capable of deriving diverse audio from different video-sampled latents." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.279, + 0.833 + ], + "angle": 0, + "content": "4.4. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Quantitative evaluations are presented in Table 1, where VAFlow demonstrates state-of-the-art performance across key audio generation metrics: \\(\\mathrm{FD}_{\\text {pas. }}\\), \\(\\mathrm{FD}_{\\text {mel. }}\\), \\(\\mathrm{IS}_{\\text {pas. }}\\), \\(\\mathrm{KL}_{\\text {pas. }}\\), and \\(\\mathrm{KL}_{\\text {mel. }}\\). For audio-visual synchronization," + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.091, + 0.86, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.274, + 0.908, + 0.346 + ], + "angle": 0, + "content": "Figure 4. Performance of flow-based baselines and VAFlow variants in scale. VAFlow models (142M, 1.05B) consistently outperform comparable-scale baselines (Frieren-159M, MMAudio-1.03B), with stable scaling property indicated by progressive IS gains (diversity) and saturating FD improvements (quality)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.673 + ], + "angle": 0, + "content": "VAFlow achieves \\(96.3\\%\\) accuracy (Acc.), comparable to Frieren \\((97.0\\%)\\). In semantic relevance, it attains an IB-VA score of 28.6, second only to TiVA's 30.3. Notably, VAFlow maintains a simple and straightforward model design without relying on complex video conditioning modules [8, 41, 46], carefully designed noise schedules [9], or additional guidance modules [25, 45]. Its performance across various metrics, achieved with a standard Flow Matching training approach and basic ODE solvers, underscores the simplicity and effectiveness of the direct video-to-audio mapping design. When compared to text-enhanced baselines (gray rows), VAFlow maintains advantages in \\(\\mathrm{FD}_{\\text{pas.}}\\), \\(\\mathrm{FD}_{\\text{mel.}}\\), \\(\\mathrm{IS}_{\\text{pas.}}\\), and synchronization Acc., though text-augmented models exhibit superior audio semantic metrics (KL distances and IB-VA). This aligns with expectations since text conditioning explicitly constrains semantic space. We note that Seeing&Hearing's elevated IB-VA score (36.7) stems from classifier guidance using the same ImageBind model, which however compromises generation quality and synchronization performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.825 + ], + "angle": 0, + "content": "We further conduct human evaluation with sampled 50 videos from the test set to compare our method with the flow-based baseline, Frieren. The generated results from both models were randomized, and 10 experts were invited to score the models based on three criteria: sound quality (Quality), audio-visual semantic consistency (Semantic) and audio-visual synchronization (Sync.). The scoring was done by selecting a winner or declaring a tie for each comparison. The results, summarized in Table 2, show that our model outperforms the baseline in all three dimensions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.686, + 0.85 + ], + "angle": 0, + "content": "4.5. Ablative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Visual Features. Three visual feature types were experimented with: semantic features (CLIP embeddings extracted per frame at 10 fps), temporally-enhanced semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11782" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.91, + 0.381 + ], + "angle": 0, + "content": "
ParadigmMethodVis.Feat.QualitySync. Acc. ↑Semantic IB-VA ↑
\\(FD_{pas.}\\)↓\\(FD_{mel.}\\)↓\\(IS_{mel.}\\)↑\\(KL_{pas.}\\)↓\\(KL_{mel.}\\)↓
ARSpecVQGAN [15]RGB,Opti.342.418.1520.53.106.8455.914.1
Im2Wav [34]CLIP252.213.0335.82.265.4178.919.6
MaskVATT [24]eva-CLIP+154.17.0764.31.343.4684.825.0
VAB [37]eva-CLIP193.112.2134.42.315.0982.425.9
DiffusionMultiFoley [7]CAVP241.613.0067.71.593.3485.227.0
Seeing&Hearing [45]ImageBind+251.512.630.12.656.3754.236.7
Diff-Foley [25]CAVP512.111.4550.42.916.1888.020.7
LoVA [9]CLIP149.16.9058.72.105.1287.026.3
FoleyCrafter [46]CLIP134.27.4056.62.295.6583.527.8
TiVA [41]CLIP111.94.3167.02.025.2187.530.3
V2A-Mapper [40]CLIP94.94.2164.92.466.0678.622.2
FlowMMAudio-L [8]CLIP+159.29.3845.41.953.8477.533.9
Frieren [43]CAVP259.96.7052.42.925.6497.023.0
VAFlow (ours)CLIP87.73.8673.61.914.8188.128.6
VAFlow (ours)CAVP91.84.7364.82.415.9696.325.1
VAFlow (ours)VidTok128.56.1748.13.648.9449.912.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.391, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Table 1. Quantitative video-to-audio results on the VGGSound test set. “Vis. Feat.” indicates the employed visual features, with a “+” denoting the inclusion of auxiliary visual information (e.g., visual captions via LLama-2 [39] or Qwen-VL [2] in VATT and Seeing&Hearing, or SyncFormer [16] features in MMAudio). All results are obtained from official code or released audio outputs. The MultiFoley results here is an 8k subset filtered by ImageBind from the authors. Rows in gray denote methods that incorporate extra text data (e.g., video captions); for fairness, these results are provided for reference only and are not directly compared with standard video-to-audio models. The best score is bolded, and the second-best is underlined. Our three VAFlow variants, differing in visual features, achieve definitive gains in audio quality (FD, IS, KL) while delivering near state-of-the-art performance for synchronization (Acc.) and correlation (IB-VA)." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.517, + 0.478, + 0.572 + ], + "angle": 0, + "content": "
ComparisonQualityWin rate(%) Sync.Semantic
Ours vs. Frieren66.00(±4.18)60.44(±5.72)60.89(±4.48)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.639 + ], + "angle": 0, + "content": "Table 2. Human evaluation results, with comparison between our model and the flow-based baseline Frieren across three criteria: sound quality, audio-visual synchronization, and audio-visual semantic consistency. Ours outperforms the baseline in all criteria." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "features (CAVP features via video-audio contrastive learning at 4 fps), reconstruction features (VidTok tokenized latents at 15 fps). Utilizing the same model size and training process for these features, we developed three variants: VAFlow-CLIP, VAFlow-CAVP, and VAFlow-VidTok, with results presented in the last three rows of Table 1. The results indicate that the widely-used CLIP features overall perform the best, yielding superior audio quality and semantic relevance. The temporally enhanced CAVP features, on the other hand, demonstrate the best video-audio synchronization but slightly lag in other metrics compared to VAFlow-CLIP. Models based on reconstruction features exhibit the poorest performance across all evaluated aspects. These outcomes suggest two key insights: 1) The type of video features directly influences VAFlow's performance," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.641 + ], + "angle": 0, + "content": "with different feature types leading to models with distinct performance preferences (either better semantic performance or synchronization); 2) Modality alignment matters more than information preservation. Although reconstruction features retain the most comprehensive video information among the three, their lack of modality alignment hampers the V2A model's ability to learn effective cross-modal generative capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Flow Matching Source Distributions. We evaluate VAFlow under two source distributions: video latent priors from alignment VAE and standard Gaussian noise. For video prior, we further test variants with/without joint training of the alignment VAE encoder (Method 3.3). All models share same training budgets (300K steps for non-joint training variants; 170K+130K for joint training)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Results in Table 3 demonstrate that video latent priors trained solely under alignment VAE objective in Equation 7 fail to surpass Gaussian noise baselines. Only when coupled with joint training—adapting the prior to flow-matching tasks—do video latent priors outperform Gaussian prior. This highlights the necessity of co-optimizing prior distributions for flow-based generation rather than relying on standalone pretraining on reconstruction task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Conditioning Mechanisms and Guidance. Prior work highlights the critical role of condition and guidance mech" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11783" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.47, + 0.158 + ], + "angle": 0, + "content": "
V-PriorTuned\\(FD_{pas.}\\)↓\\(IS_{mel.}\\)↑Acc.↑IB-VA ↑
X-102.667.687.927.9
✓X111.857.482.325.4
✓✓87.773.688.128.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.168, + 0.483, + 0.239 + ], + "angle": 0, + "content": "Table 3. Ablation study on flow-matching source distributions. 'V-Prior' denotes sampling starts from video latent priors (vs. Gaussian priors). 'Tuned' indicates joint optimization of the alignment VAE encoder. Results show that jointly tuned video priors outperform Gaussian priors as source distributions for V2A flow models." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.269, + 0.48, + 0.351 + ], + "angle": 0, + "content": "
Cond.CACFGFDpas. ↓ISmel. ↑Acc. ↑IB-VA ↑
XX-135.941.431.78.5
X✓X160.536.518.95.4
✓✓X133.842.378.021.0
✓✓✓87.773.688.128.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Table 4. Ablation study on conditioning mechanisms and guidance. 'Cond.' denotes explicit video conditioning during inference; 'CA' indicates the presence of cross-attention layers of different variants; 'CFG' refers to classifier-free guidance. Results demonstrate that optimal performance requires both explicit conditioning and CFG when solving VAFlow ODE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.522 + ], + "angle": 0, + "content": "anisms (e.g., classifier-free guidance, CFG) in enhancing model performance. We address two questions:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.786 + ], + "angle": 0, + "content": "(1) Can VAFlow generate high-quality audio using only video priors, omitting explicit video conditioning during sampling? We experimented with two variants without explicit condition: No cross-attention (CA): Remove CA layers in DiT blocks and retrain the model (Table 4, Row 1). Zero-input CA: Retain CA layers but zero out video inputs during inference (Table 4, Row 2). 'No CA' variant exhibit comparable audio quality metrics (FD, IS) to the video-condition-visible variant (Table 4, Row 3: explicit video conditioning without CFG), yet suffer significant degradation in semantic relevance (IB-VA) and synchronization (Acc.). This indicates that while video priors partially preserve visual information, explicit video conditioning during sampling is essential for temporal and semantic alignment. We think that as the latent trajectory approaches the target audio distribution \\((t\\to 1)\\), initial visual cues in the prior degrade, necessitating direct video information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "(2) Does explicit video conditioning during sampling still require CFG for optimal performance? Adding CFG to the video-condition-visible variant (Row 4 in Table 4) further improves all metrics, demonstrating VAFlow's compatibility with established guidance techniques. This underscores the dual necessity of explicit video conditioning and CFG for optimal V2A synthesis with VAFlow framework." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.089, + 0.596, + 0.099 + ], + "angle": 0, + "content": "A case video:" + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.099, + 0.888, + 0.118 + ], + "angle": 0, + "content": "At a seaside with rolling waves, a group of people are swimming and playing. A woman walks toward the camera, points to the right, and says something." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.119, + 0.905, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.459, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 5. Mel-spectrograms of audio generated by VAFlow and the baseline for a complex seaside scene, including background waves and varying voices, demonstrating VAFlow's superior content understanding and synchronization with the video." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.545, + 0.649, + 0.558 + ], + "angle": 0, + "content": "4.6. Case Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Figure 5 illustrates the performance of VAFlow and the baselines on the complex scene video. The video depicts a seaside scene with various sounds include background waves and different voices associated with shifting characters (noisy laughter, woman speaking). We present the ground truth and the mel-spectrograms of the audio generated by different methods. Our results show that VAFlow not only accurately understands the video content and generates all the necessary sounds but also maintains synchronization with the visual timing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.738, + 0.783, + 0.753 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce VAFlow, a novel framework for V2A task, which directly denoises from the video space to the audio space via flow matching. demonstrating superior performance compared to previous methods. Future work will explore VAFlow's potential in more diverse audio domains, such as speech and music. Expanding the range of tasks and datasets will enable VAFlow's scaling properties to shine even further, laying a more general-purpose the foundation model for content generation community." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11784" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.176 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (No. 62276268), Public Computing Cloud, Renmin University of China and ZHI-TECH GROUP." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.256, + 0.484, + 0.326 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.327, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.396, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[4] Li-Wen Chang, Wenlei Bao, Qi Hou, Chengquan Jiang, Ningxin Zheng, Yinmin Zhong, Xuanrun Zhang, Zuquan Song, Chengji Yao, Ziheng Jiang, et al. Flux: fast software-based communication overlap on gpus through kernel fusion. arXiv preprint arXiv:2406.06858, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, pages 721–725, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[7] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. arXiv preprint arXiv:2411.17698, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[8] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. arXiv preprint arXiv:2412.15322, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[9] Xin Cheng, Xihua Wang, Yihan Wu, Yuyue Wang, and Ruihua Song. Lova: Long-form video-to-audio generation. arXiv preprint arXiv:2409.15157, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv preprint arXiv:2407.14358, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, pages 15180-15190, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[15] Vladimir iashin and Esa Rahtu. Taming visually guided sound generation. In BMVC, 2021. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[16] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, pages 5325-5329, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[18] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.434 + ], + "angle": 0, + "content": "[19] Khaled Koutini, Jan Schlüter, Hamid Eghbal-Zadeh, and Gerhard Widmer. Efficient training of audio transformers with patchout. arXiv preprint arXiv:2110.05069, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In International Conference on Machine Learning, pages 21450-21474, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[22] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In CVPR, pages 2755-2765, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[24] Xiulong Liu, Kun Su, and Eli Shlizerman. Tell what you hear from what you see - video to audio generation through text. In Advances in Neural Information Processing Systems, pages 101337-101366, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[25] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In Advances in Neural Information Processing Systems, pages 48855-48876, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[26] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing, pages 1-6, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11785" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text. arXiv preprint arXiv:2410.18514, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[29] Santiago Pascual, Chunghsin Yeh, Ioannis Tsiamas, and Joan Serra. Masked generative video-to-audio transformers with enhanced synchronicity. In ECCV, pages 247-264, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.208, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.482, + 0.568 + ], + "angle": 0, + "content": "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[34] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP, pages 1-5, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[36] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In International Conference on Machine Learning, pages 46804–46822, 2024. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[38] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[40] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, pages 15492-15501, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[41] Xihua Wang, Yuyue Wang, Yihan Wu, Ruihua Song, Xu Tan, Zehua Chen, Hongteng Xu, and Guodong Sui. Tiva: Time-aligned video-to-audio generation. In ACM MM, pages 573-582, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[42] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR, pages 6572-6582, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.47 + ], + "angle": 0, + "content": "[43] Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation network with rectified flow matching. In Advances in Neural Information Processing Systems, pages 128118-128138, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[44] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In ECCV, pages 378-394, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[45] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, pages 7151-7161, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[46] Yiming Zhang, Yicheng Gu, Yanhong Zeng, Zhening Xing, Yuancheng Wang, Zhizheng Wu, and Kai Chen. Foley-crafter: Bring silent videos to life with lifelike and synchronized sounds. arXiv preprint arXiv:2407.01494, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[47]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11786" + } + ] +] \ No newline at end of file diff --git a/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_origin.pdf b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8acfc90e9f8f554bfba008af0919945b5d0272bd --- /dev/null +++ b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/7812feed-641c-4918-a582-f934ca29c2a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68daea7515223e0114cf193f42366c9da7d466916cb1d8394e2a99954eb6c279 +size 3628600 diff --git a/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/full.md b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c8c708baca3b911296f9fbc2a1c38ec73e0823a3 --- /dev/null +++ b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/full.md @@ -0,0 +1,295 @@ +# VAFlow: Video-to-Audio Generation with Cross-Modality Flow Matching + +Xihua Wang $^{1}$ Xin Cheng $^{1}$ Yuyue Wang $^{1}$ Ruihua Song $^{1*}$ Yunfeng Wang $^{2}$ $^{1}$ Gaoling School of Artificial Intelligence, Renmin University of China +xihuaw@ruc.edu.cn, songruihua.bloon@outlook.com + +# Abstract + +Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even textaugmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching. + +# 1. Introduction + +Video-to-audio (V2A) generation plays a vital role in multimedia content generation, including Foley for films and AI-generated silent videos. In recent years, this task has attracted increasing attention within the generative community. Early V2A approaches [15, 24, 26, 29, 34] adopted language modeling strategies by discretizing audio into tokens, and used parallel mask-prediction models (BERT-like) or token-by-token autoregressive models (GPT-like) borrowed from NLP community as shown in Figure 1. However, these methods, reliant on discrete tokens, incur inherent information losses by the discretization process. + +![](images/00b726f5ccb3264f0217ff01f4838780de47120879ff2588490bb01d049831e2.jpg) +Figure 1. Four different V2A paradigms, along with state-of-the-art performance under each. The first row shows discrete token-based approaches (mask-prediction and autoregressive based generation), which are limited by information loss during tokenization. The second row presents methods generating continuous data. Previous diffusion and flow methods (bottom-left) model the noise-to-audio mapping conditioning on video. We find this approach suboptimal and demonstrate that directly modeling the video-to-audio transformation (bottom-right) yields better results. + +Recent advancements in Diffusion and Flow Matching (FM) models in image [4, 10, 33] and video [3, 31] generation have spurred similar approaches for V2A [9, 25, 41, 43, 45]. Generally, these methods model V2A as a conditional generation task, transforming Gaussian noise conditioned on video into audio, as shown in Figure 1 (bottom-left). The generation process typically involves two steps: (1) sampling a random noise from Gaussian prior and (2) using this noise and video condition as input to iteratively denoise the noise into audio. This approach avoids discrete audio tokens, directly fitting the continuous distribution of audio data, and currently holds state-of-the-art results. + +However, existing diffusion/FM-based V2A methods focus solely on the second step, i.e., denoising Gaussian noise into audio, assuming the first step is trivial. Even state-of-the-art methods [40, 43] begin with Gaussian noise, but we find that this approach is not optimal. Unlike text-to + +image or text-to-video tasks, where textual conditions are abstract and highly defined by human, the video conditions in V2A are more complex and require both spatial and temporal understanding. The generation goal also demands a high degree of alignment between the generated audio and the video conditions, not just semantically, but also temporally. There are two inherent difficulties: (1) different conditions on the same noise require the model to generate different audio outputs, demanding advanced condition processing [8, 41, 46], and (2) noise variations under the same condition lead to unstable generation results, prompting research into methods for noise control [44], such as finding a "golden noise" for consistent results [47]. + +In contrast, we propose a new perspective by shifting focus to the first step of the generation process. We question: If there are better priors than Gaussian noise? Given that many diffusion and FM models dedicate significant effort to integrating video conditions into the denoising process for better generation performance, we further question: If V2A can work better when directly transforming the video distribution to audio rather than from noise? Fortunately, both answers are affirmative. + +We introduce VAFlow, a novel framework for V2A generation, which directly denoises from the video space to the audio space. To implement this video-to-audio paradigm, we adopt the FM framework, as it theoretically supports any source distribution. There are three technical challenges for this implementation: (1) aligning source and target distributions: video and audio feature spaces typically have different time and spatial resolutions, yet FM requires the source and target distributions to be aligned (the same shape). (2) uncertainty of the feature type: identifying the optimal video representations/feature types as the source distribution. (3) conditioning mechanism: exploring additional video information integration ways beyond encoding it in the source distribution. In VAFlow, we first employ an alignment VAE to adjust video features to match the audio latent space. It then explores three different video feature types (semantic CLIP [32], spatiotemporal CAVP [25], reconstruction-focused VidTok [38]) to systematically investigate source video information. Experiments show that using video features both as a source distribution and as conditioning information yields the best results for VAFlow. + +Moreover, the recent success of large language models (LLMs) [13, 28] has highlighted the effectiveness of scaling laws, and various generative works consequently explore similar scaling properties in each domain [30, 42]. However, existing V2A models have not demonstrated clear scaling trends, where increasing model size does not consistently achieve improvement [8, 43]. In contrast, our experiments show that VAFlow exhibits clear scaling property, with model size increases leading to consistent improvements in generation quality, further confirming its potential. + +Without elaborate training tricks or external data (e.g., text-audio pairs), VAFlow achieves state-of-the-art results on standard benchmarks (VGGSound), outperforming even text-augmented models in generation fidelity, diversity, and cross-modality alignment with input video. Playable audi-los generated with VAFlow can be accessed at the demo page: https://vaflow.github.io/demo. + +# 2. Related Work + +# 2.1. Flow Matching and Diffusion Models + +Diffusion models [17, 27, 35] reconstruct data distributions by iteratively denoising Gaussian-sampled latents, governmented by various designed schedulers achieving state-of-the-art performance in multimodal generation tasks. Unlike diffusions, flow-based models [1, 20, 22, 23] establish deterministic mappings between source (e.g., Gaussian) and target distributions through invertible neural transformations, with recent flow matching optimizing vector fields through optimal transport theory, enabling direct learning of straight transport paths between source and target distributions, thereby reducing training instability and inference steps. While both paradigms typically rely on Gaussian priors as their source distribution, existing video-to-audio systems on both paradigms model noise-to-audio generation conditioned on video features. In contrast, our approach directly aligns video latent spaces with target audio distributions via linear transformations under a flow matching framework, bypassing iterative noise modeling, preserving efficiency, and ensuring precise frame-to-frame alignment. + +# 2.2. Video-to-Audio Generation + +Framework. Recent advances in generative models have driven significant progress in video-to-audio (V2A) synthesis. Current V2A methods fall into three main paradigms: (1) Autoregressive (AR) models [15, 26, 34] discretize audio into token (codec) sequences and utilize video features as prefixes to model unified video-audio sequences. While straightforward, their token-by-token generation process results in quadratic inference time scaling, limiting efficiency for long audio sequences. (2) Mask-based generation paradigms [24, 29, 37] also employ discrete tokens but decode audio in parallel by predicting masked tokens. Video features serve as the condition for mask-to-audio prediction. However, both AR and mask-based methods suffer from an inherent upper bound of fidelity due to lossy compression from audio discretization. (3) Flow/diffusion-based methods [7-9, 25, 40, 41, 43, 45, 46] leverage continuous representations and currently achieve state-of-the-art V2A performance [8, 40, 43]. These techniques iteratively transform noise into target audio distributions while conditioning on video features to steer the denoising trajectory. + +Video Conditional Modeling. Unlike typical conditioned + +generation tasks like text-to-image [33], V2A involves fully aligned sequential correspondences between video and audio. Existing methods address this through various strategies: some AR-based methods investigate various attention mechanisms and masking schemes for temporally sequential alignment [26]; some mask-based approaches integrate hybrid ways to fuse video features into the mask-prediction process to synchronize audio generation [29]; and some flow/diffusion-based methods augment their pipelines with auxiliary modules (e.g., loudness or audio layout predictors [41, 46]) to enforce temporal consistency. In contrast, our approach builds on flow-based techniques by discarding the conventional noise-to-audio paradigm and complex video-conditioning designs. Instead, we propose a more intuitive method that directly transforms the video temporal space to the audio temporal space via a flow transport. + +# 3. Method: + +# 3.1. Preliminary: Flow Matching + +Let $x_0 \sim p(x_0)$ and $x_1 \sim p(x_1)$ denote samples from source and target distributions, respectively, with $c$ representing an optional conditioning signal. Flow Matching (FM) [20, 23] establishes an inter-distribution transport process governed by the ordinary differential equation (ODE): + +$$ +d x (t) = v _ {\theta} \left(x _ {t}, t, c\right) d t, \tag {1} +$$ + +where $t \in [0,1]$ denotes the continuous time step, and $v_{\theta}$ represents a neural network-parameterized velocity field. The condition $c$ is incorporated for conditional flow implementations. Optimal transport flows enforce velocity alignment with the direction $(x_{1} - x_{0})$ of the linear path pointing from $x_{0}$ to $x_{1}$ through regression: + +$$ +\mathcal {L} (\theta) = \mathbb {E} _ {t, x _ {0}, x _ {1}, c} \| (x _ {1} - x _ {0}) - v _ {\theta} (x _ {t}, t, c) \| ^ {2}, \tag {2} +$$ + +where $x_{t} = (1 - t)x_{0} + tx_{1}$ constitutes the linear interpolation between $x_{0}$ and $x_{1}$ . The learned velocity field enables target generation by solving the ODE in Equation 1. + +Previous video-to-audio flow approaches [8, 43] (and related diffusion methods [25, 41]) use a Gaussian prior $p(x_0) = \mathcal{N}(0,1)$ for the source distribution, with audio as $p(x_{1})$ and video as conditioning input $c$ . During training, random noise $x_0$ is paired with ground-truth audio $x_{1}$ , prompting $v_{\theta}$ to model the paths from noise to audio under visual conditions. + +This work challenges the Gaussian prior by proposing a direct video-to-audio transport. Theoretically, FM permits arbitrary source distributions, motivating our formulation: + +$$ +d x (t) = v _ {\theta} \left(x _ {t}, t\right) d t, \tag {3} +$$ + +where $x_{t} = (1 - t)x_{v} + tx_{a}$ linearly interpolates between paired video and audio samples $x_{v},x_{a}\sim p(v,a)$ + +# 3.2. VAFlow Architecture + +Model Overview. As illustrated in Figure 2, the input video features $x_{v}$ are extracted via a visual encoder, while the target audio waveform can be obtained from latent representations $x_{a}$ through WaveVAE. VAFlow attempts to establish FM between the video feature space $x_{v}$ and audio latent space $x_{a}$ , enabling direct audio generation from video features via ODE solving instead of Gaussian noise initialization. However, since FM requires consistent dimensional shapes between source and target distributions (see Equation 3), the inherent spatiotemporal resolution mismatch between $x_{v}$ and $x_{a}$ necessitates shape alignment. To address this, we employ a 1D VAE (namely alignment VAE) that compresses interpolated $x_{v}$ into aligned features $x_{v}^{\mathrm{AE}}$ matching $x_{a}$ 's dimensions. + +A diffusion transformer (DiT) subsequently works as a velocity field estimator to capture the transport dynamics between $x_{v}^{\mathrm{AE}}$ and $x_{a}$ . This enables continuous trajectory sampling from $x_{v}^{\mathrm{AE}}$ to $x_{a}$ through ODE integration. The sampled latent $\hat{x}_{a}$ is finally decoded by WaveVAE to produce the output audio. The following contents detail: the visual/audio representations $(x_{v}, x_{v}^{\mathrm{AE}}, x_{a})$ , and the velocity field estimator in VAFlow. + +Video and Audio Representations. For an input video-audio pair $(v, a)$ , the audio waveform $a$ is compressed by a pretrained 1D WaveVAE's [11] encoder into $x_{a} \in \mathbb{R}^{T_{a} \times D_{a}}$ . This latent $x_{a}$ serves as the target for flow generation and is later decoded to reconstruct the waveform. + +For video representation, a pretrained visual encoder encodes $v$ into $x_v \in \mathbb{R}^{T_v \times D_v}$ (e.g., CLIP [32], CAVP [25], or VidTok [38] as encoder). Given the discrepancies (e.g., $x_a \in \mathbb{R}^{215 \times 64}$ versus $x_v \in \mathbb{R}^{100 \times 768}$ for a 10-second clip), we propose an alignment VAE $\epsilon_{\phi}$ to further compress and align video latents with audio latents' space while preserving essential visual information. Specifically, we utilize a 1D convolutional encoder-decoder as $\epsilon_{\phi}$ to interpolate and project $x_v$ into a variational space $x_v^{\mathrm{AE}} \in \mathbb{R}^{T_a \times D_a}$ , matching the shape of $x_a$ . The encoder of $\epsilon_{\phi}$ predicts the parameters $\mu_{x_v^{\mathrm{AE}}}, \sigma_{x_v^{\mathrm{AE}}} = E_{\epsilon_{\phi}}(x_v)$ for a Gaussian distribution, from which $x_v^{\mathrm{AE}} \sim \mathcal{N}(\mu_{x_v^{\mathrm{AE}}}, \sigma_{x_v^{\mathrm{AE}}})$ is sampled. The decoder reconstructs $x_v$ as $\hat{x}_v = D_{\epsilon_{\phi}}(x_v^{\mathrm{AE}})$ . This process ensures dimensional alignment with audio latents while retaining critical spatiotemporal features of the original video representation. + +Velocity Estimator. Built upon the diffusion transformer (DiT) architecture [11], the velocity estimator processes two key inputs: time embeddings of $t$ and the interpolated state $x_{t} = (1 - t)x_{v}^{\mathrm{AE}} + tx_{a}$ , following the Equation 3. The time embedding is concatenated with the audio sequence to form an input tensor $\in \mathbb{R}^{(T_a + 1)\times D_a}$ , which is processed through stacked transformer blocks. The first $T_{a}$ dimensions of the output sequence constitute the pre + +![](images/ff1ff64fb37921e45705859648878aee89608f3208a30432c279d825228be95d.jpg) +Figure 2. VAFlow Architecture. The figure illustrates VAFow in the inference manner. A visual encoder initially extracts temporal features $x_{v} \in \mathbb{R}^{T_{v} \times D_{v}}$ from variable-frame-rate video inputs. Then, a 1D alignment VAE (Variational Autoencoder) $\epsilon_{\phi}$ compresses interpolated $x_{v}$ into audio-aligned latent features $x_{v}^{\mathrm{AE}}$ with matched dimensions $(T_{v}^{\mathrm{AE}} \times D_{v}^{\mathrm{AE}} = T_{a} \times D_{a})$ . Next, a diffusion transformer $(v_{\theta})$ predicts the velocity field for the transport $(x_{a} - x_{v}^{\mathrm{AE}})$ , and by solving the corresponding ODE (see Equation 3), an estimated $\hat{x}_{a}$ is obtained. Finally, the WaveVAE decoder converts $\hat{x}_{a}$ into the waveform representing the predicted audio. Note that during diffusion transformer prediction, $x_{v}$ can optionally be incorporated via cross-attention, detail discussion is conducted in the experiment Section 4.5. + +dicted velocity $v_{\theta}(x_t, t)$ , with 1D Rotary Position Embeddings (RoPE [36]) to preserve positional awareness in attention layers. Notably, while the interpolated state $x_t$ implicitly encodes visual information through $x_v^{\mathrm{AE}}$ , this signal diminishes as $t \to 1$ . To address this limitation, we retain cross-attention layers in the transformer blocks that persistently integrate the original video features $x_v$ . Further studies (Section 4.5) demonstrate that this architectural choice is critical to enhance visual perception throughout the flow trajectory and improve audio generation quality. + +# 3.3. VAFlow Training and Inference + +We conduct three-stage training: (1) pretrain the alignment VAE, (2) train the VAFlow estimator with frozen alignment VAE, and (3) jointly fine-tune both alignment VAE and estimator. We detail training and inference as follows: + +Alignment VAE Training. Given encoded visual features $x_{v} \in \mathbb{R}^{T_{v} \times D_{v}}$ , the alignment VAE encoder predicts distribution parameters $(\mu_{x_{v}^{\mathrm{AE}}}, \sigma_{x_{v}^{\mathrm{AE}}})$ , while the decoder reconstructs $\hat{x}_{v}$ from a sampled $x_{v}^{\mathrm{AE}} \sim \mathcal{N}(\mu_{x_{v}^{\mathrm{AE}}}, \sigma_{x_{v}^{\mathrm{AE}}})$ . The model optimizes three objectives to preserve visual semantics while regularizing $x_{v}^{\mathrm{AE}}$ within $\mathbb{R}^{T_{a} \times D_{a}}$ Gaussian space: + +- KL Regularization: enforces Gaussian latent priors via: + +$$ +\mathcal {L} _ {\mathrm {K L}} = D _ {\mathrm {K L}} \left(\mathcal {N} \left(\mu_ {x _ {0} ^ {\mathrm {A E}}}, \sigma_ {x _ {0} ^ {\mathrm {A E}}}\right) \| \mathcal {N} (0, I)\right). \tag {4} +$$ + +- Reconstruction Loss: ensures feature-level consistency: + +$$ +\mathcal {L} _ {\text {r e c}} = \left\| x _ {v} - \hat {x} _ {v} \right\| _ {2} ^ {2}. \tag {5} +$$ + +- Contrastive Alignment: strengthens feature preservation: + +$$ +\mathcal {L} _ {\text {c o n}} = - \sum_ {s \in \mathcal {S}} \log \frac {\exp \left(f _ {s} \left(\hat {x} _ {v}\right) \cdot f _ {s} \left(x _ {v}\right) / \tau\right)}{\sum_ {x ^ {\prime} \in \mathcal {B}} \exp \left(f _ {s} \left(\hat {x} _ {v}\right) \cdot f _ {s} \left(x ^ {\prime}\right) / \tau\right)}, \tag {6} +$$ + +where $S = \{\mathrm{frame},\mathrm{clip},\mathrm{global}\}$ defines temporal granularities and $\tau = 0.7$ serves as a temperature parameter [6]. + +For scale $s$ , $f_{s}(\cdot)$ extracts features via three ways: frame-level: single-frame sampling + mean pooling; clip-level: $K = 8$ -adjacent-frame sampling + pooling; global-level: full temporal mean pooling at $T_{v}$ dimension of feature $x$ . Negative samples $x'$ are drawn from training batch $\mathcal{B}$ . + +The composite objective integrates these terms: + +$$ +\mathcal {L} _ {\mathrm {V A E}} = \lambda_ {\mathrm {K L}} \mathcal {L} _ {\mathrm {K L}} + \lambda_ {\mathrm {r e c}} \mathcal {L} _ {\mathrm {r e c}} + \lambda_ {\mathrm {c o n}} \mathcal {L} _ {\mathrm {c o n}}, \tag {7} +$$ + +with $\lambda_{\mathrm{KL}} = 8\times 10^{-4}$ $\lambda_{\mathrm{rec}} = 1.0$ , and $\lambda_{\mathrm{con}} = 1.0$ + +Velocity Estimator Training Given encoded features $x_{v}$ , $x_{v}^{\mathrm{AE}}$ , and $x_{a}$ , we train the velocity estimator $v_{\theta}$ to predict the linear transport direction $x_{a} - x_{v}^{\mathrm{AE}}$ through: + +$$ +\mathcal {L} (\theta) = \mathbb {E} _ {t \sim \mathcal {U} (0, 1)} \left\| \left(x _ {a} - x _ {v} ^ {\mathrm {A E}}\right) - v _ {\theta} \left(x _ {t}, t, x _ {v}\right) \right\| _ {2} ^ {2}, \tag {8} +$$ + +where $x_{t} = (1 - t)x_{v}^{\mathrm{AE}} + tx_{a}$ . For the conditioned version, visual features $x_{v}$ are incorporated via cross-attention (with a $10\%$ probability of zeroing out $x_{v}$ during training). The vanilla version omits this conditioning, simplifying the input of estimator in Equation 8 to $v_{\theta}(x_t,t)$ . + +Joint Turning After separate training of the alignment VAE and the velocity estimator, a joint optimization stage is performed. In this stage, the VAE encoder and $v_{\theta}$ are co-trained (discarding the VAE decoder), while the WaveVAE and visual feature extractor remain frozen. For a pair $(v, a)$ , the visual extractor and WaveVAE produce $x_{v}$ and $x_{a}$ . The VAE encoder compresses $x_{v}$ to obtain $x_{v}^{\mathrm{AE}} \sim q_{\phi}(x_{v}^{\mathrm{AE}}|x_{v})$ which together with $x_{a}$ and a random $t \sim \mathcal{U}(0,1)$ forms $x_{t}$ . The joint objective integrates KL regularization in Equation 4 with velocity prediction in Equation 8: + +$$ +\mathcal {L} (\theta , \phi_ {\text {E n c}}) = \lambda_ {\mathrm {K L}} \mathcal {L} _ {\mathrm {K L}} + \mathbb {E} \| \left(x _ {a} - x _ {v} ^ {\mathrm {A E}}\right) - v _ {\theta} \left(x _ {t}, t, x _ {v}\right) \| _ {2} ^ {2}. \tag {9} +$$ + +Inference Upon training, VAFlow synthesizes audio $a$ from an input video $v$ through the following pipeline: The visual extractor first encodes $v$ into a feature $x_{v}$ , which is + +compressed by the alignment VAE encoder into distribution parameters $(\mu_{x_v^{\mathrm{AE}}},\sigma_{x_v^{\mathrm{AE}}})$ . A latent sample $x_{v}^{\mathrm{AE}}\sim \mathcal{N}(\mu_{x_v^{\mathrm{AE}}},\sigma_{x_v^{\mathrm{AE}}}^2)$ is drawn as the initial state $(x_0)$ for solving the ODE in Equation 1 with learned estimator. The predicted $\hat{x}_a(x_1)$ is then decoded into waveform via the frozen WaveVAE. During inference, the variational sampling of $x_{v}^{\mathrm{AE}}$ introduces stochasticity - even with fixed ODE integration paths, distinct samples yield diverse outputs. + +For conditional velocity estimators, classifier-free guidance (CFG) [14] enhances synthesis quality by interpolating conditioned and unconditioned velocity predictions: + +$$ +v _ {\theta} ^ {\mathrm {C F G}} \left(x _ {t}, t, x _ {v}\right) = (1 + \gamma) v _ {\theta} \left(x _ {t}, t, x _ {v}\right) - \gamma v _ {\theta} \left(x _ {t}, t, \emptyset\right), \tag {10} +$$ + +where $\gamma$ controls guidance strength and $\varnothing$ denotes null (zero) conditioning. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We employ two datasets: VGGSound [5] and AudioSet-V2A [25]. For VGGSound, we adhere to the original split, which comprises 186k training samples and 15k testing samples. AudioSet-V2A contains 390k videos that are exclusively used for training. During training, the alignment VAE is trained on a combined set of VGGSound training and AudioSet-V2A to yield more generalized visual feature compression. The subsequent velocity estimator training and joint tuning are conducted solely on the VGGSound training set. At inference, we evaluate on the VGGSound test set as in prior work. + +Implementation Details. For the estimator's DiT architecture, we adapted the DiT implementation from StableAudio-Open-1.0 [11] (SDA1.0) by eliminating duration encoding, applying global time encoding solely to time steps, and restricting the cross-attention layer's key/value inputs to video features. For the visual extractor, we use three variants: CLIP [32] (a widely adopted visual semantic extractor in V2A tasks), CAVP [25] (which captures both visual semantic and temporal features), and VidTok [38] (a visual tokenizer for reconstruction). We train three alignment VAEs with comparable parameter scales—one for each visual extractor. In the velocity estimator training stage, the alignment VAE remains frozen; in the joint training stage, both the estimator and the alignment VAE are trained. In all stages, the audio WaveVAE and the visual extractor are kept frozen. Notably, we do not perform any further rectified iterations on the estimator; all VAFlow variants are trained for a single rectified iteration (Rectified-1, RF-1). During sampling, we employ ODE solvers of various orders—first-order Euler, second-order Midpoint, and fifth-order dopri5—with differing sampling steps and classifier-free guidance scales. + +Metrics. We employ both objective and subjective evaluations across three dimensions: generation quality, audiovisual relevance, and audiovisual synchronization. For objective evaluation, generation quality is measured using FD score [21], KL divergence [15], and IS score [21]. In previous work, each metric may be computed with different feature extraction models (e.g., Melception, PasST $^1$ ). Audiovisual relevance is quantified by computing the cosine similarity between the embeddings of the input video and the generated audio using the ImageBind model [12]. For synchronization, we use the synchronization accuracy (Acc.) [25] reflects the alignment between the input video and generated audio. + +Baselines. Our method is comprehensively compared against all major generative paradigms, including autoregressive (AR), masked prediction (Mask), diffusion-based, and flow-based approaches, as detailed in Table 1. All baseline models were evaluated using either officially released code and models or directly downloaded audio outputs by their authors, with consistent experimental settings maintained across all evaluations on the same hardware platform. Implementation specifics include: (1) Im2Wav generates audio in 4s clips, truncated to 10s after three clips. (2) Diff-Foley here employs its double-guidance variant (demonstrating superior performance); (3) MMAudio-L here is the largest variant of MMAudio; and (4) Frieren here employs its final no-reflow configuration (achieving optimal Fréchet distance). For comprehensive evaluation, we also include text-augmented V2A models (denoted in gray rows) that leverage external textual data beyond benchmark audiovisual corpora during training or inference. + +# 4.2. Diffusion vs. Flow vs. VAFlow Models + +We examine framework-specific performance by training same VAFlow-CLIP-142M models under three paradigms, maintaining identical configurations: CLIP visual features, architecture parameters, initialization weights, and other hyperparameters. The frameworks include: Diffusion: Standard DDPM [27, 33] with cross-attention video conditioning. Flow: Standard continuous-time flow (Equation 2) using Gaussian prior. VAFlow: Our framework employing aligned VAE-compressed video features as source distribution. All frameworks use their respective optimal inference settings (DDPM: DPMSolver; Flows: dopri5). Joint tuning in VAFlow begins at 100K steps. + +Frechet Distance (FD) results in Figure 3 reveals: 1) Flow paradigms converge faster with lower FD than diffusion, demonstrating inherent optimization benefits. 2) VAFlow attains lower final FD than standard flow despite delayed early-stage convergence (<100K steps). This per + +![](images/ddbe06928bea183e718f5bdb4312d42104ea62742c41fc44089c0837a7187e51.jpg) +Figure 3. Comparison of different training frameworks. We examine the same model under different training paradigms by controlling training data, steps, and other hyperparameters. Results demonstrate that VAFlow converges to superior optima, validating the effectiveness of its direct video-to-audio transformation design and highlighting this paradigm advantages in V2A generation. + +formance attributes to VAFlow's explicit modeling of deterministic video-audio transport paths, avoiding the ambiguous noise-to-audio mappings in Gaussian-prior approaches. VAFlow learns optimal transport trajectories between video-audio distribution, bypassing error-prone intermediate transformations in Gaussian-noise-to-audio paths. + +# 4.3. VAFlow Scaling Analysis + +We investigate VAFlow's scaling capabilities by developing four model variants (142M, 264M, 527M, 1.05B parameters) through width/depth scaling of DiT blocks while maintaining fixed training protocols (data, steps, etc.). Evaluation metrics (FD, IS) versus parameter counts are visualized in Figure 4 (bubble sizes denote model scales). + +There are two key findings here: 1) Parameter Efficiency: VAFlow consistently outperforms flow-based baselines (Frieren, MMAudio) at comparable sizes (e.g., 142M vs. 159M, 1.05B vs. 1.03B) across both metrics. 2) Scaling Capability: Previous flow-based studies have struggled to show consistent scalability [8, 43], where larger models fail to achieve better metrics consistently. To the best of our knowledge, VAFlow is the first flow-based V2A framework to demonstrate stable scaling properties. As the parameter size increases, the models achieve better FD and IS scores. Interestingly, larger models tend to improve IS scores more significantly than FD scores, suggesting that as the VAFlow model scales, the quality of generated audio gradually saturates, and the model becomes more capable of deriving diverse audio from different video-sampled latents. + +# 4.4. Benchmark Results + +Quantitative evaluations are presented in Table 1, where VAFlow demonstrates state-of-the-art performance across key audio generation metrics: $\mathrm{FD}_{\text {pas. }}$ , $\mathrm{FD}_{\text {mel. }}$ , $\mathrm{IS}_{\text {pas. }}$ , $\mathrm{KL}_{\text {pas. }}$ , and $\mathrm{KL}_{\text {mel. }}$ . For audio-visual synchronization, + +![](images/f09fb5f06231f54fb4c79ca7f5346440e949c46f53aa2c0415e37e982158f883.jpg) +Figure 4. Performance of flow-based baselines and VAFlow variants in scale. VAFlow models (142M, 1.05B) consistently outperform comparable-scale baselines (Frieren-159M, MMAudio-1.03B), with stable scaling property indicated by progressive IS gains (diversity) and saturating FD improvements (quality). + +VAFlow achieves $96.3\%$ accuracy (Acc.), comparable to Frieren $(97.0\%)$ . In semantic relevance, it attains an IB-VA score of 28.6, second only to TiVA's 30.3. Notably, VAFlow maintains a simple and straightforward model design without relying on complex video conditioning modules [8, 41, 46], carefully designed noise schedules [9], or additional guidance modules [25, 45]. Its performance across various metrics, achieved with a standard Flow Matching training approach and basic ODE solvers, underscores the simplicity and effectiveness of the direct video-to-audio mapping design. When compared to text-enhanced baselines (gray rows), VAFlow maintains advantages in $\mathrm{FD}_{\text{pas.}}$ , $\mathrm{FD}_{\text{mel.}}$ , $\mathrm{IS}_{\text{pas.}}$ , and synchronization Acc., though text-augmented models exhibit superior audio semantic metrics (KL distances and IB-VA). This aligns with expectations since text conditioning explicitly constrains semantic space. We note that Seeing&Hearing's elevated IB-VA score (36.7) stems from classifier guidance using the same ImageBind model, which however compromises generation quality and synchronization performance. + +We further conduct human evaluation with sampled 50 videos from the test set to compare our method with the flow-based baseline, Frieren. The generated results from both models were randomized, and 10 experts were invited to score the models based on three criteria: sound quality (Quality), audio-visual semantic consistency (Semantic) and audio-visual synchronization (Sync.). The scoring was done by selecting a winner or declaring a tie for each comparison. The results, summarized in Table 2, show that our model outperforms the baseline in all three dimensions. + +# 4.5. Ablative Analysis + +Visual Features. Three visual feature types were experimented with: semantic features (CLIP embeddings extracted per frame at 10 fps), temporally-enhanced semantic + +
ParadigmMethodVis.Feat.QualitySync. Acc. ↑Semantic IB-VA ↑
\(FD_{pas.}\)↓\(FD_{mel.}\)↓\(IS_{mel.}\)↑\(KL_{pas.}\)↓\(KL_{mel.}\)↓
ARSpecVQGAN [15]RGB,Opti.342.418.1520.53.106.8455.914.1
Im2Wav [34]CLIP252.213.0335.82.265.4178.919.6
MaskVATT [24]eva-CLIP+154.17.0764.31.343.4684.825.0
VAB [37]eva-CLIP193.112.2134.42.315.0982.425.9
DiffusionMultiFoley [7]CAVP241.613.0067.71.593.3485.227.0
Seeing&Hearing [45]ImageBind+251.512.630.12.656.3754.236.7
Diff-Foley [25]CAVP512.111.4550.42.916.1888.020.7
LoVA [9]CLIP149.16.9058.72.105.1287.026.3
FoleyCrafter [46]CLIP134.27.4056.62.295.6583.527.8
TiVA [41]CLIP111.94.3167.02.025.2187.530.3
V2A-Mapper [40]CLIP94.94.2164.92.466.0678.622.2
FlowMMAudio-L [8]CLIP+159.29.3845.41.953.8477.533.9
Frieren [43]CAVP259.96.7052.42.925.6497.023.0
VAFlow (ours)CLIP87.73.8673.61.914.8188.128.6
VAFlow (ours)CAVP91.84.7364.82.415.9696.325.1
VAFlow (ours)VidTok128.56.1748.13.648.9449.912.9
+ +Table 1. Quantitative video-to-audio results on the VGGSound test set. “Vis. Feat.” indicates the employed visual features, with a “+” denoting the inclusion of auxiliary visual information (e.g., visual captions via LLama-2 [39] or Qwen-VL [2] in VATT and Seeing&Hearing, or SyncFormer [16] features in MMAudio). All results are obtained from official code or released audio outputs. The MultiFoley results here is an 8k subset filtered by ImageBind from the authors. Rows in gray denote methods that incorporate extra text data (e.g., video captions); for fairness, these results are provided for reference only and are not directly compared with standard video-to-audio models. The best score is bolded, and the second-best is underlined. Our three VAFlow variants, differing in visual features, achieve definitive gains in audio quality (FD, IS, KL) while delivering near state-of-the-art performance for synchronization (Acc.) and correlation (IB-VA). + +
ComparisonQualityWin rate(%) Sync.Semantic
Ours vs. Frieren66.00(±4.18)60.44(±5.72)60.89(±4.48)
+ +Table 2. Human evaluation results, with comparison between our model and the flow-based baseline Frieren across three criteria: sound quality, audio-visual synchronization, and audio-visual semantic consistency. Ours outperforms the baseline in all criteria. + +features (CAVP features via video-audio contrastive learning at 4 fps), reconstruction features (VidTok tokenized latents at 15 fps). Utilizing the same model size and training process for these features, we developed three variants: VAFlow-CLIP, VAFlow-CAVP, and VAFlow-VidTok, with results presented in the last three rows of Table 1. The results indicate that the widely-used CLIP features overall perform the best, yielding superior audio quality and semantic relevance. The temporally enhanced CAVP features, on the other hand, demonstrate the best video-audio synchronization but slightly lag in other metrics compared to VAFlow-CLIP. Models based on reconstruction features exhibit the poorest performance across all evaluated aspects. These outcomes suggest two key insights: 1) The type of video features directly influences VAFlow's performance, + +with different feature types leading to models with distinct performance preferences (either better semantic performance or synchronization); 2) Modality alignment matters more than information preservation. Although reconstruction features retain the most comprehensive video information among the three, their lack of modality alignment hampers the V2A model's ability to learn effective cross-modal generative capabilities. + +Flow Matching Source Distributions. We evaluate VAFlow under two source distributions: video latent priors from alignment VAE and standard Gaussian noise. For video prior, we further test variants with/without joint training of the alignment VAE encoder (Method 3.3). All models share same training budgets (300K steps for non-joint training variants; 170K+130K for joint training). + +Results in Table 3 demonstrate that video latent priors trained solely under alignment VAE objective in Equation 7 fail to surpass Gaussian noise baselines. Only when coupled with joint training—adapting the prior to flow-matching tasks—do video latent priors outperform Gaussian prior. This highlights the necessity of co-optimizing prior distributions for flow-based generation rather than relying on standalone pretraining on reconstruction task. + +Conditioning Mechanisms and Guidance. Prior work highlights the critical role of condition and guidance mech + +
V-PriorTuned\(FD_{pas.}\)↓\(IS_{mel.}\)↑Acc.↑IB-VA ↑
X-102.667.687.927.9
✓X111.857.482.325.4
✓✓87.773.688.128.6
+ +Table 3. Ablation study on flow-matching source distributions. 'V-Prior' denotes sampling starts from video latent priors (vs. Gaussian priors). 'Tuned' indicates joint optimization of the alignment VAE encoder. Results show that jointly tuned video priors outperform Gaussian priors as source distributions for V2A flow models. + +
Cond.CACFGFDpas. ↓ISmel. ↑Acc. ↑IB-VA ↑
XX-135.941.431.78.5
X✓X160.536.518.95.4
✓✓X133.842.378.021.0
✓✓✓87.773.688.128.6
+ +Table 4. Ablation study on conditioning mechanisms and guidance. 'Cond.' denotes explicit video conditioning during inference; 'CA' indicates the presence of cross-attention layers of different variants; 'CFG' refers to classifier-free guidance. Results demonstrate that optimal performance requires both explicit conditioning and CFG when solving VAFlow ODE. + +anisms (e.g., classifier-free guidance, CFG) in enhancing model performance. We address two questions: + +(1) Can VAFlow generate high-quality audio using only video priors, omitting explicit video conditioning during sampling? We experimented with two variants without explicit condition: No cross-attention (CA): Remove CA layers in DiT blocks and retrain the model (Table 4, Row 1). Zero-input CA: Retain CA layers but zero out video inputs during inference (Table 4, Row 2). 'No CA' variant exhibit comparable audio quality metrics (FD, IS) to the video-condition-visible variant (Table 4, Row 3: explicit video conditioning without CFG), yet suffer significant degradation in semantic relevance (IB-VA) and synchronization (Acc.). This indicates that while video priors partially preserve visual information, explicit video conditioning during sampling is essential for temporal and semantic alignment. We think that as the latent trajectory approaches the target audio distribution $(t\to 1)$ , initial visual cues in the prior degrade, necessitating direct video information. +(2) Does explicit video conditioning during sampling still require CFG for optimal performance? Adding CFG to the video-condition-visible variant (Row 4 in Table 4) further improves all metrics, demonstrating VAFlow's compatibility with established guidance techniques. This underscores the dual necessity of explicit video conditioning and CFG for optimal V2A synthesis with VAFlow framework. + +![](images/541834df9a135b3f673f2135fc672b6335d888f01e8ee58e170b788fb154322c.jpg) +A case video: +At a seaside with rolling waves, a group of people are swimming and playing. A woman walks toward the camera, points to the right, and says something. +Figure 5. Mel-spectrograms of audio generated by VAFlow and the baseline for a complex seaside scene, including background waves and varying voices, demonstrating VAFlow's superior content understanding and synchronization with the video. + +# 4.6. Case Studies + +Figure 5 illustrates the performance of VAFlow and the baselines on the complex scene video. The video depicts a seaside scene with various sounds include background waves and different voices associated with shifting characters (noisy laughter, woman speaking). We present the ground truth and the mel-spectrograms of the audio generated by different methods. Our results show that VAFlow not only accurately understands the video content and generates all the necessary sounds but also maintains synchronization with the visual timing. + +# 5. Conclusion and Future Work + +In this paper, we introduce VAFlow, a novel framework for V2A task, which directly denoises from the video space to the audio space via flow matching. demonstrating superior performance compared to previous methods. Future work will explore VAFlow's potential in more diverse audio domains, such as speech and music. Expanding the range of tasks and datasets will enable VAFlow's scaling properties to shine even further, laying a more general-purpose the foundation model for content generation community. + +# Acknowledgement + +This work is supported by the National Natural Science Foundation of China (No. 62276268), Public Computing Cloud, Renmin University of China and ZHI-TECH GROUP. + +# References + +[1] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 2 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 7 +[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. 1 +[4] Li-Wen Chang, Wenlei Bao, Qi Hou, Chengquan Jiang, Ningxin Zheng, Yinmin Zhong, Xuanrun Zhang, Zuquan Song, Chengji Yao, Ziheng Jiang, et al. Flux: fast software-based communication overlap on gpus through kernel fusion. arXiv preprint arXiv:2406.06858, 2024. 1 +[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, pages 721–725, 2020. 5 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 4 +[7] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. arXiv preprint arXiv:2411.17698, 2024. 2, 7 +[8] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. arXiv preprint arXiv:2412.15322, 2024. 2, 3, 6, 7 +[9] Xin Cheng, Xihua Wang, Yihan Wu, Yuyue Wang, and Ruihua Song. Lova: Long-form video-to-audio generation. arXiv preprint arXiv:2409.15157, 2024. 1, 2, 6, 7 +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 1 +[11] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv preprint arXiv:2407.14358, 2024. 3, 5 +[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, pages 15180-15190, 2023. 5 + +[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2 +[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[15] Vladimir iashin and Esa Rahtu. Taming visually guided sound generation. In BMVC, 2021. 1, 2, 5, 7 +[16] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, pages 5325-5329, 2024. 7 +[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577, 2022. 2 +[18] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5 +[19] Khaled Koutini, Jan Schlüter, Hamid Eghbal-Zadeh, and Gerhard Widmer. Efficient training of audio transformers with patchout. arXiv preprint arXiv:2110.05069, 2021. 5 +[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3 +[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In International Conference on Machine Learning, pages 21450-21474, 2023. 5 +[22] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In CVPR, pages 2755-2765, 2025. 2 +[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3 +[24] Xiulong Liu, Kun Su, and Eli Shlizerman. Tell what you hear from what you see - video to audio generation through text. In Advances in Neural Information Processing Systems, pages 101337-101366, 2024. 1, 2, 7 +[25] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In Advances in Neural Information Processing Systems, pages 48855-48876, 2023. 1, 2, 3, 5, 6, 7 +[26] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing, pages 1-6, 2024. 1, 2, 3 +[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171, 2021. 2, 5 + +[28] Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text. arXiv preprint arXiv:2410.18514, 2024. 2 +[29] Santiago Pascual, Chunghsin Yeh, Ioannis Tsiamas, and Joan Serra. Masked generative video-to-audio transformers with enhanced synchronicity. In ECCV, pages 247-264, 2024. 1, 2, 3 +[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2 +[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 5 +[34] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP, pages 1-5, 2023. 1, 2, 7 +[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2 +[36] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4 +[37] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In International Conference on Machine Learning, pages 46804–46822, 2024. 2, 7 + +[38] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 2, 3, 5 +[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7 +[40] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, pages 15492-15501, 2024. 1, 2, 7 +[41] Xihua Wang, Yuyue Wang, Yihan Wu, Ruihua Song, Xu Tan, Zehua Chen, Hongteng Xu, and Guodong Sui. Tiva: Time-aligned video-to-audio generation. In ACM MM, pages 573-582, 2024. 1, 2, 3, 6, 7 +[42] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR, pages 6572-6582, 2024. 2 +[43] Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation network with rectified flow matching. In Advances in Neural Information Processing Systems, pages 128118-128138, 2024. 1, 2, 3, 6, 7 +[44] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In ECCV, pages 378-394, 2024. 2 +[45] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, pages 7151-7161, 2024. 1, 2, 6, 7 +[46] Yiming Zhang, Yicheng Gu, Yanhong Zeng, Zhening Xing, Yuancheng Wang, Zhizheng Wu, and Kai Chen. Foley-crafter: Bring silent videos to life with lifelike and synchronized sounds. arXiv preprint arXiv:2407.01494, 2024. 2, 3, 6, 7 +[47]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.2 \ No newline at end of file diff --git a/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/images.zip b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..27eb08d929b05a603dea7594100e2ce159cf5a8c --- /dev/null +++ b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03bd79028fb9b67a086f910a95625ac525c7b72fdb4cece8ed2e54def3d9f3e5 +size 415117 diff --git a/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/layout.json b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5a1b51b7993414b16fba5362cb67c0136c795397 --- /dev/null +++ b/2025/VAFlow_ Video-to-Audio Generation with Cross-Modality Flow Matching/layout.json @@ -0,0 +1,8869 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "text", + "content": "VAFlow: Video-to-Audio Generation with Cross-Modality Flow Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "spans": [ + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "text", + "content": "Xihua Wang" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "text", + "content": " Xin Cheng" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "text", + "content": " Yuyue Wang" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "text", + "content": " Ruihua Song" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "text", + "content": " Yunfeng Wang" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 141, + 543, + 185 + ], + "type": "text", + "content": "Gaoling School of Artificial Intelligence, Renmin University of China \nxihuaw@ruc.edu.cn, songruihua.bloon@outlook.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "type": "text", + "content": "Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even textaugmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 540, + 135, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 540, + 135, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 540, + 135, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": "Video-to-audio (V2A) generation plays a vital role in multimedia content generation, including Foley for films and AI-generated silent videos. In recent years, this task has attracted increasing attention within the generative community. Early V2A approaches [15, 24, 26, 29, 34] adopted language modeling strategies by discretizing audio into tokens, and used parallel mask-prediction models (BERT-like) or token-by-token autoregressive models (GPT-like) borrowed from NLP community as shown in Figure 1. However, these methods, reliant on discrete tokens, incur inherent information losses by the discretization process." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 211, + 555, + 375 + ], + "blocks": [ + { + "bbox": [ + 316, + 211, + 555, + 375 + ], + "lines": [ + { + "bbox": [ + 316, + 211, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 555, + 375 + ], + "type": "image", + "image_path": "00b726f5ccb3264f0217ff01f4838780de47120879ff2588490bb01d049831e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 384, + 556, + 484 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 484 + ], + "type": "text", + "content": "Figure 1. Four different V2A paradigms, along with state-of-the-art performance under each. The first row shows discrete token-based approaches (mask-prediction and autoregressive based generation), which are limited by information loss during tokenization. The second row presents methods generating continuous data. Previous diffusion and flow methods (bottom-left) model the noise-to-audio mapping conditioning on video. We find this approach suboptimal and demonstrate that directly modeling the video-to-audio transformation (bottom-right) yields better results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 508, + 556, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 652 + ], + "type": "text", + "content": "Recent advancements in Diffusion and Flow Matching (FM) models in image [4, 10, 33] and video [3, 31] generation have spurred similar approaches for V2A [9, 25, 41, 43, 45]. Generally, these methods model V2A as a conditional generation task, transforming Gaussian noise conditioned on video into audio, as shown in Figure 1 (bottom-left). The generation process typically involves two steps: (1) sampling a random noise from Gaussian prior and (2) using this noise and video condition as input to iteratively denoise the noise into audio. This approach avoids discrete audio tokens, directly fitting the continuous distribution of audio data, and currently holds state-of-the-art results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "However, existing diffusion/FM-based V2A methods focus solely on the second step, i.e., denoising Gaussian noise into audio, assuming the first step is trivial. Even state-of-the-art methods [40, 43] begin with Gaussian noise, but we find that this approach is not optimal. Unlike text-to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11777" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "image or text-to-video tasks, where textual conditions are abstract and highly defined by human, the video conditions in V2A are more complex and require both spatial and temporal understanding. The generation goal also demands a high degree of alignment between the generated audio and the video conditions, not just semantically, but also temporally. There are two inherent difficulties: (1) different conditions on the same noise require the model to generate different audio outputs, demanding advanced condition processing [8, 41, 46], and (2) noise variations under the same condition lead to unstable generation results, prompting research into methods for noise control [44], such as finding a \"golden noise\" for consistent results [47]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 294, + 337 + ], + "type": "text", + "content": "In contrast, we propose a new perspective by shifting focus to the first step of the generation process. We question: If there are better priors than Gaussian noise? Given that many diffusion and FM models dedicate significant effort to integrating video conditions into the denoising process for better generation performance, we further question: If V2A can work better when directly transforming the video distribution to audio rather than from noise? Fortunately, both answers are affirmative." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 591 + ], + "type": "text", + "content": "We introduce VAFlow, a novel framework for V2A generation, which directly denoises from the video space to the audio space. To implement this video-to-audio paradigm, we adopt the FM framework, as it theoretically supports any source distribution. There are three technical challenges for this implementation: (1) aligning source and target distributions: video and audio feature spaces typically have different time and spatial resolutions, yet FM requires the source and target distributions to be aligned (the same shape). (2) uncertainty of the feature type: identifying the optimal video representations/feature types as the source distribution. (3) conditioning mechanism: exploring additional video information integration ways beyond encoding it in the source distribution. In VAFlow, we first employ an alignment VAE to adjust video features to match the audio latent space. It then explores three different video feature types (semantic CLIP [32], spatiotemporal CAVP [25], reconstruction-focused VidTok [38]) to systematically investigate source video information. Experiments show that using video features both as a source distribution and as conditioning information yields the best results for VAFlow." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Moreover, the recent success of large language models (LLMs) [13, 28] has highlighted the effectiveness of scaling laws, and various generative works consequently explore similar scaling properties in each domain [30, 42]. However, existing V2A models have not demonstrated clear scaling trends, where increasing model size does not consistently achieve improvement [8, 43]. In contrast, our experiments show that VAFlow exhibits clear scaling property, with model size increases leading to consistent improvements in generation quality, further confirming its potential." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "Without elaborate training tricks or external data (e.g., text-audio pairs), VAFlow achieves state-of-the-art results on standard benchmarks (VGGSound), outperforming even text-augmented models in generation fidelity, diversity, and cross-modality alignment with input video. Playable audi-los generated with VAFlow can be accessed at the demo page: https://vaflow.github.io/demo." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 168, + 400, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 400, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 400, + 181 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 189, + 511, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 511, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 511, + 201 + ], + "type": "text", + "content": "2.1. Flow Matching and Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 206, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 434 + ], + "type": "text", + "content": "Diffusion models [17, 27, 35] reconstruct data distributions by iteratively denoising Gaussian-sampled latents, governmented by various designed schedulers achieving state-of-the-art performance in multimodal generation tasks. Unlike diffusions, flow-based models [1, 20, 22, 23] establish deterministic mappings between source (e.g., Gaussian) and target distributions through invertible neural transformations, with recent flow matching optimizing vector fields through optimal transport theory, enabling direct learning of straight transport paths between source and target distributions, thereby reducing training instability and inference steps. While both paradigms typically rely on Gaussian priors as their source distribution, existing video-to-audio systems on both paradigms model noise-to-audio generation conditioned on video features. In contrast, our approach directly aligns video latent spaces with target audio distributions via linear transformations under a flow matching framework, bypassing iterative noise modeling, preserving efficiency, and ensuring precise frame-to-frame alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 443, + 464, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 464, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 464, + 455 + ], + "type": "text", + "content": "2.2. Video-to-Audio Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 461, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 700 + ], + "type": "text", + "content": "Framework. Recent advances in generative models have driven significant progress in video-to-audio (V2A) synthesis. Current V2A methods fall into three main paradigms: (1) Autoregressive (AR) models [15, 26, 34] discretize audio into token (codec) sequences and utilize video features as prefixes to model unified video-audio sequences. While straightforward, their token-by-token generation process results in quadratic inference time scaling, limiting efficiency for long audio sequences. (2) Mask-based generation paradigms [24, 29, 37] also employ discrete tokens but decode audio in parallel by predicting masked tokens. Video features serve as the condition for mask-to-audio prediction. However, both AR and mask-based methods suffer from an inherent upper bound of fidelity due to lossy compression from audio discretization. (3) Flow/diffusion-based methods [7-9, 25, 40, 41, 43, 45, 46] leverage continuous representations and currently achieve state-of-the-art V2A performance [8, 40, 43]. These techniques iteratively transform noise into target audio distributions while conditioning on video features to steer the denoising trajectory." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Video Conditional Modeling. Unlike typical conditioned" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11778" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "generation tasks like text-to-image [33], V2A involves fully aligned sequential correspondences between video and audio. Existing methods address this through various strategies: some AR-based methods investigate various attention mechanisms and masking schemes for temporally sequential alignment [26]; some mask-based approaches integrate hybrid ways to fuse video features into the mask-prediction process to synchronize audio generation [29]; and some flow/diffusion-based methods augment their pipelines with auxiliary modules (e.g., loudness or audio layout predictors [41, 46]) to enforce temporal consistency. In contrast, our approach builds on flow-based techniques by discarding the conventional noise-to-audio paradigm and complex video-conditioning designs. Instead, we propose a more intuitive method that directly transforms the video temporal space to the audio temporal space via a flow transport." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 274, + 115, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 115, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 115, + 287 + ], + "type": "text", + "content": "3. Method:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 214, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 214, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 214, + 306 + ], + "type": "text", + "content": "3.1. Preliminary: Flow Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "inline_equation", + "content": "x_0 \\sim p(x_0)" + }, + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "inline_equation", + "content": "x_1 \\sim p(x_1)" + }, + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "text", + "content": " denote samples from source and target distributions, respectively, with " + }, + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 311, + 296, + 373 + ], + "type": "text", + "content": " representing an optional conditioning signal. Flow Matching (FM) [20, 23] establishes an inter-distribution transport process governed by the ordinary differential equation (ODE):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 380, + 295, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 380, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 128, + 380, + 295, + 394 + ], + "type": "interline_equation", + "content": "d x (t) = v _ {\\theta} \\left(x _ {t}, t, c\\right) d t, \\tag {1}", + "image_path": "e6878073826b5f115ea9385be6517d07d4032dbb2f5242eb16bdb66e9e73b342.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "content": " denotes the continuous time step, and " + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "content": " represents a neural network-parameterized velocity field. The condition " + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "content": " is incorporated for conditional flow implementations. Optimal transport flows enforce velocity alignment with the direction " + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "inline_equation", + "content": "(x_{1} - x_{0})" + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "content": " of the linear path pointing from " + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 401, + 296, + 475 + ], + "type": "text", + "content": " through regression:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 481, + 295, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 481, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 82, + 481, + 295, + 496 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\mathbb {E} _ {t, x _ {0}, x _ {1}, c} \\| (x _ {1} - x _ {0}) - v _ {\\theta} (x _ {t}, t, c) \\| ^ {2}, \\tag {2}", + "image_path": "304215aa9da1cde6cbe344dbff36a76359c701993aec658a1fff3d5107153c56.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "inline_equation", + "content": "x_{t} = (1 - t)x_{0} + tx_{1}" + }, + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "text", + "content": " constitutes the linear interpolation between " + }, + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 504, + 295, + 539 + ], + "type": "text", + "content": ". The learned velocity field enables target generation by solving the ODE in Equation 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "content": "Previous video-to-audio flow approaches [8, 43] (and related diffusion methods [25, 41]) use a Gaussian prior " + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "inline_equation", + "content": "p(x_0) = \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "content": " for the source distribution, with audio as " + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "inline_equation", + "content": "p(x_{1})" + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "content": " and video as conditioning input " + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "content": ". During training, random noise " + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "content": " is paired with ground-truth audio " + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "content": ", prompting " + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 540, + 295, + 622 + ], + "type": "text", + "content": " to model the paths from noise to audio under visual conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 659 + ], + "type": "text", + "content": "This work challenges the Gaussian prior by proposing a direct video-to-audio transport. Theoretically, FM permits arbitrary source distributions, motivating our formulation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 668, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 668, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 132, + 668, + 295, + 682 + ], + "type": "interline_equation", + "content": "d x (t) = v _ {\\theta} \\left(x _ {t}, t\\right) d t, \\tag {3}", + "image_path": "cddedd00612ba49fab0148c1f27f728c53225b4966d19564cf8b6dc2f98c45ce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{t} = (1 - t)x_{v} + tx_{a}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " linearly interpolates between paired video and audio samples " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{v},x_{a}\\sim p(v,a)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 437, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 437, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 437, + 83 + ], + "type": "text", + "content": "3.2. VAFlow Architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": "Model Overview. As illustrated in Figure 2, the input video features " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": " are extracted via a visual encoder, while the target audio waveform can be obtained from latent representations " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": " through WaveVAE. VAFlow attempts to establish FM between the video feature space " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": " and audio latent space " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": ", enabling direct audio generation from video features via ODE solving instead of Gaussian noise initialization. However, since FM requires consistent dimensional shapes between source and target distributions (see Equation 3), the inherent spatiotemporal resolution mismatch between " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": " necessitates shape alignment. To address this, we employ a 1D VAE (namely alignment VAE) that compresses interpolated " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": " into aligned features " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": " matching " + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 91, + 555, + 259 + ], + "type": "text", + "content": "'s dimensions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "content": "A diffusion transformer (DiT) subsequently works as a velocity field estimator to capture the transport dynamics between " + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "content": ". This enables continuous trajectory sampling from " + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "content": " through ODE integration. The sampled latent " + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "inline_equation", + "content": "\\hat{x}_{a}" + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "content": " is finally decoded by WaveVAE to produce the output audio. The following contents detail: the visual/audio representations " + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "inline_equation", + "content": "(x_{v}, x_{v}^{\\mathrm{AE}}, x_{a})" + }, + { + "bbox": [ + 313, + 262, + 554, + 357 + ], + "type": "text", + "content": ", and the velocity field estimator in VAFlow." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "text", + "content": "Video and Audio Representations. For an input video-audio pair " + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "inline_equation", + "content": "(v, a)" + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "text", + "content": ", the audio waveform " + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "text", + "content": " is compressed by a pretrained 1D WaveVAE's [11] encoder into " + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "inline_equation", + "content": "x_{a} \\in \\mathbb{R}^{T_{a} \\times D_{a}}" + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "text", + "content": ". This latent " + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 359, + 555, + 418 + ], + "type": "text", + "content": " serves as the target for flow generation and is later decoded to reconstruct the waveform." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": "For video representation, a pretrained visual encoder encodes " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_v \\in \\mathbb{R}^{T_v \\times D_v}" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " (e.g., CLIP [32], CAVP [25], or VidTok [38] as encoder). Given the discrepancies (e.g., " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_a \\in \\mathbb{R}^{215 \\times 64}" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " versus " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_v \\in \\mathbb{R}^{100 \\times 768}" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " for a 10-second clip), we propose an alignment VAE " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " to further compress and align video latents with audio latents' space while preserving essential visual information. Specifically, we utilize a 1D convolutional encoder-decoder as " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " to interpolate and project " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_v" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " into a variational space " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_v^{\\mathrm{AE}} \\in \\mathbb{R}^{T_a \\times D_a}" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": ", matching the shape of " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_a" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": ". The encoder of " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " predicts the parameters " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}} = E_{\\epsilon_{\\phi}}(x_v)" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " for a Gaussian distribution, from which " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_v^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}}, \\sigma_{x_v^{\\mathrm{AE}}})" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " is sampled. The decoder reconstructs " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "x_v" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\hat{x}_v = D_{\\epsilon_{\\phi}}(x_v^{\\mathrm{AE}})" + }, + { + "bbox": [ + 313, + 422, + 555, + 616 + ], + "type": "text", + "content": ". This process ensures dimensional alignment with audio latents while retaining critical spatiotemporal features of the original video representation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Velocity Estimator. Built upon the diffusion transformer (DiT) architecture [11], the velocity estimator processes two key inputs: time embeddings of " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " and the interpolated state " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ", following the Equation 3. The time embedding is concatenated with the audio sequence to form an input tensor " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{(T_a + 1)\\times D_a}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ", which is processed through stacked transformer blocks. The first " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_{a}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " dimensions of the output sequence constitute the pre" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11779" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 550, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 550, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 550, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 550, + 178 + ], + "type": "image", + "image_path": "ff1ff64fb37921e45705859648878aee89608f3208a30432c279d825228be95d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. VAFlow Architecture. The figure illustrates VAFow in the inference manner. A visual encoder initially extracts temporal features " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}}" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " from variable-frame-rate video inputs. Then, a 1D alignment VAE (Variational Autoencoder) " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " compresses interpolated " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " into audio-aligned latent features " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " with matched dimensions " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "(T_{v}^{\\mathrm{AE}} \\times D_{v}^{\\mathrm{AE}} = T_{a} \\times D_{a})" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": ". Next, a diffusion transformer " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "(v_{\\theta})" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " predicts the velocity field for the transport " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "(x_{a} - x_{v}^{\\mathrm{AE}})" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": ", and by solving the corresponding ODE (see Equation 3), an estimated " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\hat{x}_{a}" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " is obtained. Finally, the WaveVAE decoder converts " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\hat{x}_{a}" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " into the waveform representing the predicted audio. Note that during diffusion transformer prediction, " + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 55, + 190, + 555, + 257 + ], + "type": "text", + "content": " can optionally be incorporated via cross-attention, detail discussion is conducted in the experiment Section 4.5." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "text", + "content": "dicted velocity " + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_t, t)" + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "text", + "content": ", with 1D Rotary Position Embeddings (RoPE [36]) to preserve positional awareness in attention layers. Notably, while the interpolated state " + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "text", + "content": " implicitly encodes visual information through " + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "inline_equation", + "content": "x_v^{\\mathrm{AE}}" + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "text", + "content": ", this signal diminishes as " + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "inline_equation", + "content": "t \\to 1" + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "text", + "content": ". To address this limitation, we retain cross-attention layers in the transformer blocks that persistently integrate the original video features " + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "inline_equation", + "content": "x_v" + }, + { + "bbox": [ + 54, + 277, + 294, + 397 + ], + "type": "text", + "content": ". Further studies (Section 4.5) demonstrate that this architectural choice is critical to enhance visual perception throughout the flow trajectory and improve audio generation quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 404, + 228, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 228, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 228, + 417 + ], + "type": "text", + "content": "3.3. VAFlow Training and Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 470 + ], + "type": "text", + "content": "We conduct three-stage training: (1) pretrain the alignment VAE, (2) train the VAFlow estimator with frozen alignment VAE, and (3) jointly fine-tune both alignment VAE and estimator. We detail training and inference as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "content": "Alignment VAE Training. Given encoded visual features " + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "inline_equation", + "content": "x_{v} \\in \\mathbb{R}^{T_{v} \\times D_{v}}" + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "content": ", the alignment VAE encoder predicts distribution parameters " + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "inline_equation", + "content": "(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}})" + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "content": ", while the decoder reconstructs " + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\hat{x}_{v}" + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "content": " from a sampled " + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}} \\sim \\mathcal{N}(\\mu_{x_{v}^{\\mathrm{AE}}}, \\sigma_{x_{v}^{\\mathrm{AE}}})" + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "content": ". The model optimizes three objectives to preserve visual semantics while regularizing " + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T_{a} \\times D_{a}}" + }, + { + "bbox": [ + 55, + 470, + 295, + 544 + ], + "type": "text", + "content": " Gaussian space:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 288, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 288, + 556 + ], + "type": "text", + "content": "- KL Regularization: enforces Gaussian latent priors via:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 566, + 295, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 566, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 91, + 566, + 295, + 580 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {K L}} = D _ {\\mathrm {K L}} \\left(\\mathcal {N} \\left(\\mu_ {x _ {0} ^ {\\mathrm {A E}}}, \\sigma_ {x _ {0} ^ {\\mathrm {A E}}}\\right) \\| \\mathcal {N} (0, I)\\right). \\tag {4}", + "image_path": "f14f9e7bb3df9bd43d09feba44d2bd6f84d4ea7680838ee5ccb86b74bf38a6d3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 288, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 288, + 600 + ], + "type": "text", + "content": "- Reconstruction Loss: ensures feature-level consistency:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 609, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 609, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 137, + 609, + 295, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\left\\| x _ {v} - \\hat {x} _ {v} \\right\\| _ {2} ^ {2}. \\tag {5}", + "image_path": "15acbf16995dfcb5874d4441483c99dd90ab066f68d9c081d36b6a9cdada5bd5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 632, + 293, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 293, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 293, + 643 + ], + "type": "text", + "content": "- Contrastive Alignment: strengthens feature preservation:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 651, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 651, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 70, + 651, + 295, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n}} = - \\sum_ {s \\in \\mathcal {S}} \\log \\frac {\\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x _ {v}\\right) / \\tau\\right)}{\\sum_ {x ^ {\\prime} \\in \\mathcal {B}} \\exp \\left(f _ {s} \\left(\\hat {x} _ {v}\\right) \\cdot f _ {s} \\left(x ^ {\\prime}\\right) / \\tau\\right)}, \\tag {6}", + "image_path": "b8dff3acf95149d88232fb483c3d57074962431fbb34e62bd27f8b1fdbcc3182.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 63, + 689, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "S = \\{\\mathrm{frame},\\mathrm{clip},\\mathrm{global}\\}" + }, + { + "bbox": [ + 63, + 689, + 295, + 713 + ], + "type": "text", + "content": " defines temporal granularities and " + }, + { + "bbox": [ + 63, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.7" + }, + { + "bbox": [ + 63, + 689, + 295, + 713 + ], + "type": "text", + "content": " serves as a temperature parameter [6]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": "For scale " + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "f_{s}(\\cdot)" + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": " extracts features via three ways: frame-level: single-frame sampling + mean pooling; clip-level: " + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": "-adjacent-frame sampling + pooling; global-level: full temporal mean pooling at " + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "T_{v}" + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": " dimension of feature " + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": ". Negative samples " + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": " are drawn from training batch " + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 321, + 277, + 555, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 339, + 506, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 339, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 314, + 339, + 506, + 351 + ], + "type": "text", + "content": "The composite objective integrates these terms:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 359, + 553, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 359, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 353, + 359, + 553, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V A E}} = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {\\mathrm {r e c}} \\mathcal {L} _ {\\mathrm {r e c}} + \\lambda_ {\\mathrm {c o n}} \\mathcal {L} _ {\\mathrm {c o n}}, \\tag {7}", + "image_path": "623d06976355c95defdc32e63705c5584c9fcffa8b59e1f0084d01fea5f2d049.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 379, + 514, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 514, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 514, + 391 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 379, + 514, + 391 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{KL}} = 8\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 379, + 514, + 391 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{rec}} = 1.0" + }, + { + "bbox": [ + 313, + 379, + 514, + 391 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 379, + 514, + 391 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{con}} = 1.0" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "text", + "content": "Velocity Estimator Training Given encoded features " + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "text", + "content": ", we train the velocity estimator " + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "text", + "content": " to predict the linear transport direction " + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "inline_equation", + "content": "x_{a} - x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 313, + 392, + 554, + 427 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 434, + 553, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 434, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 325, + 434, + 553, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} \\left\\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\right\\| _ {2} ^ {2}, \\tag {8}", + "image_path": "bc5b7c0f3a13f909dd14ae8f9d7c3e73c47c9708cc3f2136b7d534407bce6def.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "inline_equation", + "content": "x_{t} = (1 - t)x_{v}^{\\mathrm{AE}} + tx_{a}" + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": ". For the conditioned version, visual features " + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": " are incorporated via cross-attention (with a " + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": " probability of zeroing out " + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": " during training). The vanilla version omits this conditioning, simplifying the input of estimator in Equation 8 to " + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": "Joint Turning After separate training of the alignment VAE and the velocity estimator, a joint optimization stage is performed. In this stage, the VAE encoder and " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": " are co-trained (discarding the VAE decoder), while the WaveVAE and visual feature extractor remain frozen. For a pair " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "(v, a)" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": ", the visual extractor and WaveVAE produce " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": ". The VAE encoder compresses " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}} \\sim q_{\\phi}(x_{v}^{\\mathrm{AE}}|x_{v})" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": " which together with " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": " and a random " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": " forms " + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 519, + 555, + 638 + ], + "type": "text", + "content": ". The joint objective integrates KL regularization in Equation 4 with velocity prediction in Equation 8:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 644, + 555, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 644, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 644, + 555, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta , \\phi_ {\\text {E n c}}) = \\lambda_ {\\mathrm {K L}} \\mathcal {L} _ {\\mathrm {K L}} + \\mathbb {E} \\| \\left(x _ {a} - x _ {v} ^ {\\mathrm {A E}}\\right) - v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) \\| _ {2} ^ {2}. \\tag {9}", + "image_path": "280758403897054e3bb2d7c18b999ddf22746ceffdcadb3988d9e62bb1ee9931.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Inference Upon training, VAFlow synthesizes audio " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " from an input video " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " through the following pipeline: The visual extractor first encodes " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " into a feature " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{v}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", which is" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11780" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": "compressed by the alignment VAE encoder into distribution parameters " + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "inline_equation", + "content": "(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": ". A latent sample " + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}\\sim \\mathcal{N}(\\mu_{x_v^{\\mathrm{AE}}},\\sigma_{x_v^{\\mathrm{AE}}}^2)" + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": " is drawn as the initial state " + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "inline_equation", + "content": "(x_0)" + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": " for solving the ODE in Equation 1 with learned estimator. The predicted " + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "inline_equation", + "content": "\\hat{x}_a(x_1)" + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": " is then decoded into waveform via the frozen WaveVAE. During inference, the variational sampling of " + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "inline_equation", + "content": "x_{v}^{\\mathrm{AE}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 170 + ], + "type": "text", + "content": " introduces stochasticity - even with fixed ODE integration paths, distinct samples yield diverse outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 294, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 294, + 206 + ], + "type": "text", + "content": "For conditional velocity estimators, classifier-free guidance (CFG) [14] enhances synthesis quality by interpolating conditioned and unconditioned velocity predictions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 213, + 294, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 213, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 63, + 213, + 294, + 239 + ], + "type": "interline_equation", + "content": "v _ {\\theta} ^ {\\mathrm {C F G}} \\left(x _ {t}, t, x _ {v}\\right) = (1 + \\gamma) v _ {\\theta} \\left(x _ {t}, t, x _ {v}\\right) - \\gamma v _ {\\theta} \\left(x _ {t}, t, \\emptyset\\right), \\tag {10}", + "image_path": "a4867d985b258f515e4c44a33ae6656538e306923a15d020552a9ec42ff7ba35.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 294, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 294, + 272 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 247, + 294, + 272 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 247, + 294, + 272 + ], + "type": "text", + "content": " controls guidance strength and " + }, + { + "bbox": [ + 55, + 247, + 294, + 272 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 55, + 247, + 294, + 272 + ], + "type": "text", + "content": " denotes null (zero) conditioning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 282, + 136, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 136, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 136, + 295 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 301, + 171, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 171, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 171, + 314 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 319, + 294, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 294, + 450 + ], + "type": "text", + "content": "Datasets. We employ two datasets: VGGSound [5] and AudioSet-V2A [25]. For VGGSound, we adhere to the original split, which comprises 186k training samples and 15k testing samples. AudioSet-V2A contains 390k videos that are exclusively used for training. During training, the alignment VAE is trained on a combined set of VGGSound training and AudioSet-V2A to yield more generalized visual feature compression. The subsequent velocity estimator training and joint tuning are conducted solely on the VGGSound training set. At inference, we evaluate on the VGGSound test set as in prior work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 451, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 713 + ], + "type": "text", + "content": "Implementation Details. For the estimator's DiT architecture, we adapted the DiT implementation from StableAudio-Open-1.0 [11] (SDA1.0) by eliminating duration encoding, applying global time encoding solely to time steps, and restricting the cross-attention layer's key/value inputs to video features. For the visual extractor, we use three variants: CLIP [32] (a widely adopted visual semantic extractor in V2A tasks), CAVP [25] (which captures both visual semantic and temporal features), and VidTok [38] (a visual tokenizer for reconstruction). We train three alignment VAEs with comparable parameter scales—one for each visual extractor. In the velocity estimator training stage, the alignment VAE remains frozen; in the joint training stage, both the estimator and the alignment VAE are trained. In all stages, the audio WaveVAE and the visual extractor are kept frozen. Notably, we do not perform any further rectified iterations on the estimator; all VAFlow variants are trained for a single rectified iteration (Rectified-1, RF-1). During sampling, we employ ODE solvers of various orders—first-order Euler, second-order Midpoint, and fifth-order dopri5—with differing sampling steps and classifier-free guidance scales." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "Metrics. We employ both objective and subjective evaluations across three dimensions: generation quality, audiovisual relevance, and audiovisual synchronization. For objective evaluation, generation quality is measured using FD score [21], KL divergence [15], and IS score [21]. In previous work, each metric may be computed with different feature extraction models (e.g., Melception, PasST" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "). Audiovisual relevance is quantified by computing the cosine similarity between the embeddings of the input video and the generated audio using the ImageBind model [12]. For synchronization, we use the synchronization accuracy (Acc.) [25] reflects the alignment between the input video and generated audio." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 443 + ], + "type": "text", + "content": "Baselines. Our method is comprehensively compared against all major generative paradigms, including autoregressive (AR), masked prediction (Mask), diffusion-based, and flow-based approaches, as detailed in Table 1. All baseline models were evaluated using either officially released code and models or directly downloaded audio outputs by their authors, with consistent experimental settings maintained across all evaluations on the same hardware platform. Implementation specifics include: (1) Im2Wav generates audio in 4s clips, truncated to 10s after three clips. (2) Diff-Foley here employs its double-guidance variant (demonstrating superior performance); (3) MMAudio-L here is the largest variant of MMAudio; and (4) Frieren here employs its final no-reflow configuration (achieving optimal Fréchet distance). For comprehensive evaluation, we also include text-augmented V2A models (denoted in gray rows) that leverage external textual data beyond benchmark audiovisual corpora during training or inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 453, + 515, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 515, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 515, + 464 + ], + "type": "text", + "content": "4.2. Diffusion vs. Flow vs. VAFlow Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 471, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 613 + ], + "type": "text", + "content": "We examine framework-specific performance by training same VAFlow-CLIP-142M models under three paradigms, maintaining identical configurations: CLIP visual features, architecture parameters, initialization weights, and other hyperparameters. The frameworks include: Diffusion: Standard DDPM [27, 33] with cross-attention video conditioning. Flow: Standard continuous-time flow (Equation 2) using Gaussian prior. VAFlow: Our framework employing aligned VAE-compressed video features as source distribution. All frameworks use their respective optimal inference settings (DDPM: DPMSolver; Flows: dopri5). Joint tuning in VAFlow begins at 100K steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 615, + 553, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 553, + 674 + ], + "type": "text", + "content": "Frechet Distance (FD) results in Figure 3 reveals: 1) Flow paradigms converge faster with lower FD than diffusion, demonstrating inherent optimization benefits. 2) VAFlow attains lower final FD than standard flow despite delayed early-stage convergence (<100K steps). This per" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "1We adopt PasST [19] model over PANN [18] for its improved robustness following the practice introduced in AudioLDM [21] evaluation repository: https://github.com/haoheliu/audioldm_eval." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11781" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 267, + 190 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 267, + 190 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 267, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 267, + 190 + ], + "type": "image", + "image_path": "ddbe06928bea183e718f5bdb4312d42104ea62742c41fc44089c0837a7187e51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 200, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 296, + 266 + ], + "type": "text", + "content": "Figure 3. Comparison of different training frameworks. We examine the same model under different training paradigms by controlling training data, steps, and other hyperparameters. Results demonstrate that VAFlow converges to superior optima, validating the effectiveness of its direct video-to-audio transformation design and highlighting this paradigm advantages in V2A generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 361 + ], + "type": "text", + "content": "formance attributes to VAFlow's explicit modeling of deterministic video-audio transport paths, avoiding the ambiguous noise-to-audio mappings in Gaussian-prior approaches. VAFlow learns optimal transport trajectories between video-audio distribution, bypassing error-prone intermediate transformations in Gaussian-noise-to-audio paths." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 195, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 195, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 195, + 382 + ], + "type": "text", + "content": "4.3. VAFlow Scaling Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 458 + ], + "type": "text", + "content": "We investigate VAFlow's scaling capabilities by developing four model variants (142M, 264M, 527M, 1.05B parameters) through width/depth scaling of DiT blocks while maintaining fixed training protocols (data, steps, etc.). Evaluation metrics (FD, IS) versus parameter counts are visualized in Figure 4 (bubble sizes denote model scales)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 639 + ], + "type": "text", + "content": "There are two key findings here: 1) Parameter Efficiency: VAFlow consistently outperforms flow-based baselines (Frieren, MMAudio) at comparable sizes (e.g., 142M vs. 159M, 1.05B vs. 1.03B) across both metrics. 2) Scaling Capability: Previous flow-based studies have struggled to show consistent scalability [8, 43], where larger models fail to achieve better metrics consistently. To the best of our knowledge, VAFlow is the first flow-based V2A framework to demonstrate stable scaling properties. As the parameter size increases, the models achieve better FD and IS scores. Interestingly, larger models tend to improve IS scores more significantly than FD scores, suggesting that as the VAFlow model scales, the quality of generated audio gradually saturates, and the model becomes more capable of deriving diverse audio from different video-sampled latents." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 170, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 170, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 170, + 659 + ], + "type": "text", + "content": "4.4. Benchmark Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Quantitative evaluations are presented in Table 1, where VAFlow demonstrates state-of-the-art performance across key audio generation metrics: " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FD}_{\\text {pas. }}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{FD}_{\\text {mel. }}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{IS}_{\\text {pas. }}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}_{\\text {pas. }}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}_{\\text {mel. }}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". For audio-visual synchronization," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 344, + 72, + 526, + 207 + ], + "blocks": [ + { + "bbox": [ + 344, + 72, + 526, + 207 + ], + "lines": [ + { + "bbox": [ + 344, + 72, + 526, + 207 + ], + "spans": [ + { + "bbox": [ + 344, + 72, + 526, + 207 + ], + "type": "image", + "image_path": "f09fb5f06231f54fb4c79ca7f5346440e949c46f53aa2c0415e37e982158f883.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 274 + ], + "type": "text", + "content": "Figure 4. Performance of flow-based baselines and VAFlow variants in scale. VAFlow models (142M, 1.05B) consistently outperform comparable-scale baselines (Frieren-159M, MMAudio-1.03B), with stable scaling property indicated by progressive IS gains (diversity) and saturating FD improvements (quality)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "text", + "content": "VAFlow achieves " + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "inline_equation", + "content": "96.3\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "text", + "content": " accuracy (Acc.), comparable to Frieren " + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "inline_equation", + "content": "(97.0\\%)" + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "text", + "content": ". In semantic relevance, it attains an IB-VA score of 28.6, second only to TiVA's 30.3. Notably, VAFlow maintains a simple and straightforward model design without relying on complex video conditioning modules [8, 41, 46], carefully designed noise schedules [9], or additional guidance modules [25, 45]. Its performance across various metrics, achieved with a standard Flow Matching training approach and basic ODE solvers, underscores the simplicity and effectiveness of the direct video-to-audio mapping design. When compared to text-enhanced baselines (gray rows), VAFlow maintains advantages in " + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{FD}_{\\text{pas.}}" + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{FD}_{\\text{mel.}}" + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{IS}_{\\text{pas.}}" + }, + { + "bbox": [ + 313, + 293, + 555, + 533 + ], + "type": "text", + "content": ", and synchronization Acc., though text-augmented models exhibit superior audio semantic metrics (KL distances and IB-VA). This aligns with expectations since text conditioning explicitly constrains semantic space. We note that Seeing&Hearing's elevated IB-VA score (36.7) stems from classifier guidance using the same ImageBind model, which however compromises generation quality and synchronization performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": "We further conduct human evaluation with sampled 50 videos from the test set to compare our method with the flow-based baseline, Frieren. The generated results from both models were randomized, and 10 experts were invited to score the models based on three criteria: sound quality (Quality), audio-visual semantic consistency (Semantic) and audio-visual synchronization (Sync.). The scoring was done by selecting a winner or declaring a tie for each comparison. The results, summarized in Table 2, show that our model outperforms the baseline in all three dimensions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 659, + 419, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 419, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 419, + 673 + ], + "type": "text", + "content": "4.5. Ablative Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Visual Features. Three visual feature types were experimented with: semantic features (CLIP embeddings extracted per frame at 10 fps), temporally-enhanced semantic" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11782" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 556, + 301 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 556, + 301 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 556, + 301 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 556, + 301 + ], + "type": "table", + "html": "
ParadigmMethodVis.Feat.QualitySync. Acc. ↑Semantic IB-VA ↑
\\(FD_{pas.}\\)↓\\(FD_{mel.}\\)↓\\(IS_{mel.}\\)↑\\(KL_{pas.}\\)↓\\(KL_{mel.}\\)↓
ARSpecVQGAN [15]RGB,Opti.342.418.1520.53.106.8455.914.1
Im2Wav [34]CLIP252.213.0335.82.265.4178.919.6
MaskVATT [24]eva-CLIP+154.17.0764.31.343.4684.825.0
VAB [37]eva-CLIP193.112.2134.42.315.0982.425.9
DiffusionMultiFoley [7]CAVP241.613.0067.71.593.3485.227.0
Seeing&Hearing [45]ImageBind+251.512.630.12.656.3754.236.7
Diff-Foley [25]CAVP512.111.4550.42.916.1888.020.7
LoVA [9]CLIP149.16.9058.72.105.1287.026.3
FoleyCrafter [46]CLIP134.27.4056.62.295.6583.527.8
TiVA [41]CLIP111.94.3167.02.025.2187.530.3
V2A-Mapper [40]CLIP94.94.2164.92.466.0678.622.2
FlowMMAudio-L [8]CLIP+159.29.3845.41.953.8477.533.9
Frieren [43]CAVP259.96.7052.42.925.6497.023.0
VAFlow (ours)CLIP87.73.8673.61.914.8188.128.6
VAFlow (ours)CAVP91.84.7364.82.415.9696.325.1
VAFlow (ours)VidTok128.56.1748.13.648.9449.912.9
", + "image_path": "b5c13a95ac5535f0a7bfdf235fa260cbc27c52e8be3881895d050442f03d78b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 409, + 292, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 309, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 309, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 555, + 390 + ], + "type": "text", + "content": "Table 1. Quantitative video-to-audio results on the VGGSound test set. “Vis. Feat.” indicates the employed visual features, with a “+” denoting the inclusion of auxiliary visual information (e.g., visual captions via LLama-2 [39] or Qwen-VL [2] in VATT and Seeing&Hearing, or SyncFormer [16] features in MMAudio). All results are obtained from official code or released audio outputs. The MultiFoley results here is an 8k subset filtered by ImageBind from the authors. Rows in gray denote methods that incorporate extra text data (e.g., video captions); for fairness, these results are provided for reference only and are not directly compared with standard video-to-audio models. The best score is bolded, and the second-best is underlined. Our three VAFlow variants, differing in visual features, achieve definitive gains in audio quality (FD, IS, KL) while delivering near state-of-the-art performance for synchronization (Acc.) and correlation (IB-VA)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 409, + 292, + 453 + ], + "lines": [ + { + "bbox": [ + 59, + 409, + 292, + 453 + ], + "spans": [ + { + "bbox": [ + 59, + 409, + 292, + 453 + ], + "type": "table", + "html": "
ComparisonQualityWin rate(%) Sync.Semantic
Ours vs. Frieren66.00(±4.18)60.44(±5.72)60.89(±4.48)
", + "image_path": "cb0cc149a1b8270bb5b2fdc91aff8972184a97bd26578f2298e92afecff4b222.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 295, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 506 + ], + "type": "text", + "content": "Table 2. Human evaluation results, with comparison between our model and the flow-based baseline Frieren across three criteria: sound quality, audio-visual synchronization, and audio-visual semantic consistency. Ours outperforms the baseline in all criteria." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 713 + ], + "type": "text", + "content": "features (CAVP features via video-audio contrastive learning at 4 fps), reconstruction features (VidTok tokenized latents at 15 fps). Utilizing the same model size and training process for these features, we developed three variants: VAFlow-CLIP, VAFlow-CAVP, and VAFlow-VidTok, with results presented in the last three rows of Table 1. The results indicate that the widely-used CLIP features overall perform the best, yielding superior audio quality and semantic relevance. The temporally enhanced CAVP features, on the other hand, demonstrate the best video-audio synchronization but slightly lag in other metrics compared to VAFlow-CLIP. Models based on reconstruction features exhibit the poorest performance across all evaluated aspects. These outcomes suggest two key insights: 1) The type of video features directly influences VAFlow's performance," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 411, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 507 + ], + "type": "text", + "content": "with different feature types leading to models with distinct performance preferences (either better semantic performance or synchronization); 2) Modality alignment matters more than information preservation. Although reconstruction features retain the most comprehensive video information among the three, their lack of modality alignment hampers the V2A model's ability to learn effective cross-modal generative capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 508, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 592 + ], + "type": "text", + "content": "Flow Matching Source Distributions. We evaluate VAFlow under two source distributions: video latent priors from alignment VAE and standard Gaussian noise. For video prior, we further test variants with/without joint training of the alignment VAE encoder (Method 3.3). All models share same training budgets (300K steps for non-joint training variants; 170K+130K for joint training)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "type": "text", + "content": "Results in Table 3 demonstrate that video latent priors trained solely under alignment VAE objective in Equation 7 fail to surpass Gaussian noise baselines. Only when coupled with joint training—adapting the prior to flow-matching tasks—do video latent priors outperform Gaussian prior. This highlights the necessity of co-optimizing prior distributions for flow-based generation rather than relying on standalone pretraining on reconstruction task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Conditioning Mechanisms and Guidance. Prior work highlights the critical role of condition and guidance mech" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11783" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 287, + 125 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 287, + 125 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 287, + 125 + ], + "type": "table", + "html": "
V-PriorTuned\\(FD_{pas.}\\)↓\\(IS_{mel.}\\)↑Acc.↑IB-VA ↑
X-102.667.687.927.9
✓X111.857.482.325.4
✓✓87.773.688.128.6
", + "image_path": "72a285c0140b286cc05b89f5a361a88df831d6a8cdedaf0be2c027104301c080.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 213, + 293, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 133, + 295, + 189 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 189 + ], + "type": "text", + "content": "Table 3. Ablation study on flow-matching source distributions. 'V-Prior' denotes sampling starts from video latent priors (vs. Gaussian priors). 'Tuned' indicates joint optimization of the alignment VAE encoder. Results show that jointly tuned video priors outperform Gaussian priors as source distributions for V2A flow models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 213, + 293, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 213, + 293, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 213, + 293, + 277 + ], + "type": "table", + "html": "
Cond.CACFGFDpas. ↓ISmel. ↑Acc. ↑IB-VA ↑
XX-135.941.431.78.5
X✓X160.536.518.95.4
✓✓X133.842.378.021.0
✓✓✓87.773.688.128.6
", + "image_path": "b3747903e3af065be5805bc844baf66fb78cff9f80dc36e605eea64dee29096e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 295, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 353 + ], + "type": "text", + "content": "Table 4. Ablation study on conditioning mechanisms and guidance. 'Cond.' denotes explicit video conditioning during inference; 'CA' indicates the presence of cross-attention layers of different variants; 'CFG' refers to classifier-free guidance. Results demonstrate that optimal performance requires both explicit conditioning and CFG when solving VAFlow ODE." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 388, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 413 + ], + "type": "text", + "content": "anisms (e.g., classifier-free guidance, CFG) in enhancing model performance. We address two questions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 419, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 622 + ], + "type": "text", + "content": "(1) Can VAFlow generate high-quality audio using only video priors, omitting explicit video conditioning during sampling? We experimented with two variants without explicit condition: No cross-attention (CA): Remove CA layers in DiT blocks and retrain the model (Table 4, Row 1). Zero-input CA: Retain CA layers but zero out video inputs during inference (Table 4, Row 2). 'No CA' variant exhibit comparable audio quality metrics (FD, IS) to the video-condition-visible variant (Table 4, Row 3: explicit video conditioning without CFG), yet suffer significant degradation in semantic relevance (IB-VA) and synchronization (Acc.). This indicates that while video priors partially preserve visual information, explicit video conditioning during sampling is essential for temporal and semantic alignment. We think that as the latent trajectory approaches the target audio distribution " + }, + { + "bbox": [ + 55, + 419, + 296, + 622 + ], + "type": "inline_equation", + "content": "(t\\to 1)" + }, + { + "bbox": [ + 55, + 419, + 296, + 622 + ], + "type": "text", + "content": ", initial visual cues in the prior degrade, necessitating direct video information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "(2) Does explicit video conditioning during sampling still require CFG for optimal performance? Adding CFG to the video-condition-visible variant (Row 4 in Table 4) further improves all metrics, demonstrating VAFlow's compatibility with established guidance techniques. This underscores the dual necessity of explicit video conditioning and CFG for optimal V2A synthesis with VAFlow framework." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 94, + 553, + 356 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 364, + 78 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 364, + 78 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 364, + 78 + ], + "type": "text", + "content": "A case video:" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 78, + 543, + 93 + ], + "lines": [ + { + "bbox": [ + 324, + 78, + 543, + 93 + ], + "spans": [ + { + "bbox": [ + 324, + 78, + 543, + 93 + ], + "type": "text", + "content": "At a seaside with rolling waves, a group of people are swimming and playing. A woman walks toward the camera, points to the right, and says something." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 94, + 553, + 356 + ], + "lines": [ + { + "bbox": [ + 316, + 94, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 94, + 553, + 356 + ], + "type": "image", + "image_path": "541834df9a135b3f673f2135fc672b6335d888f01e8ee58e170b788fb154322c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 363, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 407 + ], + "type": "text", + "content": "Figure 5. Mel-spectrograms of audio generated by VAFlow and the baseline for a complex seaside scene, including background waves and varying voices, demonstrating VAFlow's superior content understanding and synchronization with the video." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 431, + 397, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 397, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 397, + 441 + ], + "type": "text", + "content": "4.6. Case Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 570 + ], + "type": "text", + "content": "Figure 5 illustrates the performance of VAFlow and the baselines on the complex scene video. The video depicts a seaside scene with various sounds include background waves and different voices associated with shifting characters (noisy laughter, woman speaking). We present the ground truth and the mel-spectrograms of the audio generated by different methods. Our results show that VAFlow not only accurately understands the video content and generates all the necessary sounds but also maintains synchronization with the visual timing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 584, + 479, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 479, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 479, + 596 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we introduce VAFlow, a novel framework for V2A task, which directly denoises from the video space to the audio space via flow matching. demonstrating superior performance compared to previous methods. Future work will explore VAFlow's potential in more diverse audio domains, such as speech and music. Expanding the range of tasks and datasets will enable VAFlow's scaling properties to shine even further, laying a more general-purpose the foundation model for content generation community." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11784" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (No. 62276268), Public Computing Cloud, Renmin University of China and ZHI-TECH GROUP." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "text", + "content": "[1] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 202, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 202, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 296, + 258 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 295, + 312 + ], + "type": "text", + "content": "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 313, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 313, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 313, + 295, + 369 + ], + "type": "text", + "content": "[4] Li-Wen Chang, Wenlei Bao, Qi Hou, Chengquan Jiang, Ningxin Zheng, Yinmin Zhong, Xuanrun Zhang, Zuquan Song, Chengji Yao, Ziheng Jiang, et al. Flux: fast software-based communication overlap on gpus through kernel fusion. arXiv preprint arXiv:2406.06858, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 402 + ], + "type": "text", + "content": "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, pages 721–725, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, pages 1597-1607, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "text", + "content": "[7] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. arXiv preprint arXiv:2411.17698, 2024. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 536 + ], + "type": "text", + "content": "[8] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. arXiv preprint arXiv:2412.15322, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 568 + ], + "type": "text", + "content": "[9] Xin Cheng, Xihua Wang, Yihan Wu, Yuyue Wang, and Ruihua Song. Lova: Long-form video-to-audio generation. arXiv preprint arXiv:2409.15157, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 634 + ], + "type": "text", + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning, pages 12606-12633, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "text", + "content": "[11] Zach Evans, Julian D Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open. arXiv preprint arXiv:2407.14358, 2024. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, pages 15180-15190, 2023. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[13] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "text", + "content": "[14] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 175 + ], + "type": "text", + "content": "[15] Vladimir iashin and Esa Rahtu. Taming visually guided sound generation. In BMVC, 2021. 1, 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "type": "text", + "content": "[16] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, pages 5325-5329, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "type": "text", + "content": "[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[18] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 343 + ], + "type": "text", + "content": "[19] Khaled Koutini, Jan Schlüter, Hamid Eghbal-Zadeh, and Gerhard Widmer. Efficient training of audio transformers with patchout. arXiv preprint arXiv:2110.05069, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "text", + "content": "[20] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 378, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 432 + ], + "type": "text", + "content": "[21] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. In International Conference on Machine Learning, pages 21450-21474, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "text", + "content": "[22] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A noise-free framework for cross-modality evolution. In CVPR, pages 2755-2765, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "text", + "content": "[23] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 512, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 556 + ], + "type": "text", + "content": "[24] Xiulong Liu, Kun Su, and Eli Shlizerman. Tell what you hear from what you see - video to audio generation through text. In Advances in Neural Information Processing Systems, pages 101337-101366, 2024. 1, 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "text", + "content": "[25] Simian Luo, Chuanhao Yan, Chenxu Hu, and Hang Zhao. Diff-foley: Synchronized video-to-audio synthesis with latent diffusion models. In Advances in Neural Information Processing Systems, pages 48855-48876, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "text", + "content": "[26] Xinhao Mei, Varun Nagaraja, Gael Le Lan, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra. Foleygen: Visually-guided audio generation. In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing, pages 1-6, 2024. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171, 2021. 2, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11785" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[28] Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text. arXiv preprint arXiv:2410.18514, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[29] Santiago Pascual, Chunghsin Yeh, Ioannis Tsiamas, and Joan Serra. Masked generative video-to-audio transformers with enhanced synchronicity. In ECCV, pages 247-264, 2024. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 164, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 164, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 164, + 294, + 186 + ], + "type": "text", + "content": "[30] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 294, + 449 + ], + "type": "text", + "content": "[31] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 453, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 518 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "text", + "content": "[34] Roy Sheffer and Yossi Adi. I hear your true colors: Image guided audio generation. In ICASSP, pages 1-5, 2023. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 600, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 622 + ], + "type": "text", + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 666 + ], + "type": "text", + "content": "[36] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[37] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In International Conference on Machine Learning, pages 46804–46822, 2024. 2, 7" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[38] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "type": "text", + "content": "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 163, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 215 + ], + "type": "text", + "content": "[40] Heng Wang, Jianbo Ma, Santiago Pascual, Richard Cartwright, and Weidong Cai. V2a-mapper: A lightweight solution for vision-to-audio generation by connecting foundation models. In AAAI, pages 15492-15501, 2024. 1, 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "type": "text", + "content": "[41] Xihua Wang, Yuyue Wang, Yihan Wu, Ruihua Song, Xu Tan, Zehua Chen, Hongteng Xu, and Guodong Sui. Tiva: Time-aligned video-to-audio generation. In ACM MM, pages 573-582, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "text", + "content": "[42] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR, pages 6572-6582, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 308, + 553, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 372 + ], + "type": "text", + "content": "[43] Yongqi Wang, Wenxiang Guo, Rongjie Huang, Jiawei Huang, Zehan Wang, Fuming You, Ruiqi Li, and Zhou Zhao. Frieren: Efficient video-to-audio generation network with rectified flow matching. In Advances in Neural Information Processing Systems, pages 128118-128138, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "type": "text", + "content": "[44] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In ECCV, pages 378-394, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[45] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, pages 7151-7161, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 454, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 507 + ], + "type": "text", + "content": "[46] Yiming Zhang, Yicheng Gu, Yanhong Zeng, Zhening Xing, Yuancheng Wang, Zhizheng Wu, and Kai Chen. Foley-crafter: Bring silent videos to life with lifelike and synchronized sounds. arXiv preprint arXiv:2407.01494, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 510, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 552 + ], + "type": "text", + "content": "[47]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11786" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_content_list.json b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ae78dba989180a24e934656668fcf2e5f740351 --- /dev/null +++ b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_content_list.json @@ -0,0 +1,1640 @@ +[ + { + "type": "text", + "text": "VAGUE: Visual Contexts Clarify Ambiguous Expressions", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heejeong $\\mathrm{Nam}^{*1}$ , Jinwoo Ahn $^{*2}$ , Keummin $\\mathrm{Ka}^3$ , Jiwan Chung $^3$ , and Youngjae $\\mathrm{Yu}^{\\dagger 3}$", + "bbox": [ + 169, + 191, + 826, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Brown University", + "bbox": [ + 421, + 223, + 575, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ UC Berkeley", + "bbox": [ + 441, + 241, + 557, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Yonsei University", + "bbox": [ + 423, + 258, + 573, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human communication often relies on visual cues to resolve ambiguity. While humans can intuitively integrate these cues, AI systems often find it challenging to engage in sophisticated multimodal reasoning. We introduce VAGUE, a benchmark evaluating multimodal AI systems' ability to integrate visual context for intent disambiguation. VAGUE consists of 1.6K ambiguous textual expressions, each paired with an image and multiple-choice interpretations, where the correct answer is only apparent with visual context. The dataset spans both staged, complex (Visual Commonsense Reasoning) and natural, personal (Ego4D) scenes, ensuring diversity. Our experiments reveal that existing multimodal AI models struggle to infer the speaker's true intent. While performance consistently improves from the introduction of more visual cues, the overall accuracy remains far below human performance, highlighting a critical gap in multimodal reasoning. Analysis of failure cases demonstrates that current models fail to distinguish true intent from superficial correlations in the visual scene, indicating that they perceive images but do not effectively reason with them. We release our code and data at https://hazel-heejeong-nam.github.io/vague/.", + "bbox": [ + 88, + 354, + 485, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 700, + 220, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human communication is inherently contextual; for example, exclaiming \"Hey, this is a disaster!\" upon seeing a cluttered room conveys frustration or exaggeration rather than referring to an actual catastrophe. Without surrounding cues, textual dialogues can be ambiguous, making it difficult for models to accurately capture intent and nuance.", + "bbox": [ + 89, + 727, + 483, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider the case of visual contextual cues. Consider Fig. 1, which depicts a speaker making a remark in a certain situation. Without specifying contexts introduced", + "bbox": [ + 89, + 818, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2a8dbca01ca7f9cf2de59e57ff233135cba5e760ade85d2b736c88db6c48b0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 329, + 599, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hey person1, looks like you're planning a party for the neighborhood squirrels.", + "bbox": [ + 604, + 330, + 846, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb8ba85a6b859c036abd1d76f868a67d75e2753b51dd3ebe0919d8588c4362f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 368, + 580, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b146a4622b5effe44329b07299bd76f8a88078e788e652b22affa5b4f4e0d2a.jpg", + "image_caption": [ + "Visual Context 1" + ], + "image_footnote": [], + "bbox": [ + 517, + 380, + 705, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/509ea48f33bcd7e74ab1e7eabd010e041181ef09744a26e395c2c99caa9b002f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 368, + 769, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1bdd4bc388a4d7802e8453cf79b21f4526a541f221b7420bec508b25359cb457.jpg", + "image_caption": [ + "Visual Context 2" + ], + "image_footnote": [], + "bbox": [ + 712, + 378, + 900, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87898e9bddf2979562ffd934554fc2f25bb756e79aba17f08f153bf2b21ff6c4.jpg", + "image_caption": [ + "The speaker wants person1 to..." + ], + "image_footnote": [], + "bbox": [ + 602, + 481, + 627, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f39f695adddd5796b67d5a45b0ac7a1636c83143ef23955c8f89b99045f1c29a.jpg", + "image_caption": [ + "close the cereal box", + "hand out the envelop to the squirrels", + "Figure 1. A motivating example demonstrating the importance of visual context in understanding intention. Without a predetermined context, a single expression can convey multiple different intentions. The textual expression and Context 1 are from our dataset, while Context 2 is generated using DALL-E 3 [2] to help understanding." + ], + "image_footnote": [], + "bbox": [ + 699, + 518, + 718, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from visual cues, the speaker's intention can vary, thus remaining ambiguous. This implies that the visual contexts play important roles in communication, raising the question: can AI systems integrate visual cues with ambiguous dialogue to infer the speaker's intent?", + "bbox": [ + 511, + 656, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Visual Contexts ClArify ambiGUous Expressions (VAGUE), a benchmark consisting of 1.6K ambiguous textual expressions, each paired with a single image. VAGUE aims to model diverse and natural human-to-human interactions by setting each image as the speaker's viewpoint, where the speaker implicitly requests a certain action from a person within their field of view. We define the problem addressed through this setup as Multimodal Intention Disambiguation (MID), which involves reasoning about the most plausible request conditioned on visual context. Each sample in VAGUE is annotated with four multiple-choice candi", + "bbox": [ + 509, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally.", + "bbox": [ + 107, + 875, + 297, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1537", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dates, ensuring clarity and preventing multiple valid answers caused by paraphrasing or hierarchical inclusion of meaning. The dataset is meticulously curated to ensure visual dependency; the ground-truth candidate is only preferable when considering the visual context. VAGUE includes visual scenes from both artificial sources (Visual Commonsense Reasoning [46]) and real-world scenarios (Ego4D [41]), capturing a broad spectrum of scene complexity and naturalness. The textual expressions in VAGUE are initially generated by GPT-4o [31] following instructions, then reviewed through extensive human rating and filtering to ensure naturalness and alignment with the corresponding images.", + "bbox": [ + 89, + 90, + 485, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on VAGUE demonstrate that existing multimodal AI models struggle to infer a speaker's true intent in a multimodal setting. First, although models can leverage visual context—as seen by a performance progression from text-only Language Models (LMs) to pipelined Socratic Models (SMs) [47] and ultimately to end-to-end Visual Language Models (VLMs)—their overall accuracy remains significantly lower than that of humans, indicating a failure to capture the true intent. A closer analysis of failure cases reveals that the primary source of error is the models' inability to distinguish the true intent from a superficial understanding of the visual context. In other words, even though these multimodal systems can perceive the image content, they cannot effectively use this information to reason about the speaker's true intent.", + "bbox": [ + 89, + 285, + 485, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, we introduce a benchmark that exposes the limitations of current models in integrating visual cues with intent comprehension and identifies their primary failure mode. We anticipate that VAGUE will serve as a testing ground for the development of future multimodal conversational or embodied agents—systems that combine robust visual perception with nuanced conversational reasoning to effectively respond to user requests in complex scenes.", + "bbox": [ + 89, + 523, + 483, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are threefold:", + "bbox": [ + 109, + 657, + 321, + 672 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VAGUE: a novel benchmark for evaluating multimodal intention disambiguation. Validated through extensive human filtering, VAGUE is designed for robust quantitative assessment by ensuring both the ambiguity of queries and the visual (in)dependency to answer candidates.", + "- Carefully curated 1,677 scene images sourced from VCR [46] and Ego4D [41], capturing a wide range of scene complexity, diversity, and naturalness to ensure VAGUE's generalizability across various contexts.", + "- Experimental results highlighting a critical challenge in multimodal intention disambiguation: while existing models can perceive visual cues, they fail to effectively integrate this information into reasoning to deduce the speaker's true intent." + ], + "bbox": [ + 91, + 686, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Theory of Mind", + "text_level": 1, + "bbox": [ + 513, + 114, + 767, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Theory of Mind (ToM) refers to the ability to infer and reason about the intentions of others based on available information [34], where recent language models still struggle with relevant tasks [7] highlighting the need for dedicated research in this area. Initially, various methods and benchmarks have been proposed in unimodal settings, relying on text-based approaches [11, 36]. However, these methods often fail to capture the richness of real-world interactions, which often require integrating both linguistic and visual cues.", + "bbox": [ + 511, + 138, + 906, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moving beyond text-only contexts, recent work has incorporated visual information. MMToM [17] introduces a benchmark where models must process both visual and textual cues to solve question-answering tasks related to ToM. The BOSS dataset [9] is a multimodal dataset collected in situations where nonverbal communication is required. It is used to evaluate whether human beliefs can be inferred based on nonverbal cues during social interactions. Similarly, Chen et al. (2024) [5] propose a Video ToM model that leverages key video frames and transcripts, demonstrating improved reasoning on the Social-IQ 2.0 dataset [45]. MuMA-ToM [37] further extends this direction by assessing ToM reasoning in multi-agent interactions, evaluating a model's ability to infer human beliefs and goals based on video and text inputs. MToMnet [3] introduces a ToM-based neural network that integrates contextual cues, such as scene videos and object locations, with person-specific cues, to predict human beliefs in specific scenarios.", + "bbox": [ + 511, + 275, + 908, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, progress in multimodal ToM remains constrained not only by the scarcity of high-quality datasets [5] but also by the lack of explicit consideration for the ambiguity and indirectness inherent in human communication.", + "bbox": [ + 511, + 547, + 908, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multimodal Implicature Understanding", + "text_level": 1, + "bbox": [ + 513, + 619, + 856, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicature and the ambiguity that arises from it naturally emerge in everyday human conversation, requiring pragmatic understanding [38]. Early research on implicature understanding has primarily been conducted in text-only settings [27, 30, 40], with some studies specifically focusing on figurative language and metaphor [4, 21, 39]. However, since the ambiguity of standalone text is inherently limited, recent studies have extended to multiple modalities. One example is multimodal sarcasm understanding (MSU). WITS [19] and MOSES [20] are benchmarks for sarcasm explanation, both providing the speaker's emotion and voice tone as cues. DocMSU [8] is a document-level benchmark for sarcasm localization and detection. To improve MSU, EDGE [32], a graph-based approach, achieved strong performance. UR-FUNNY [13] is a benchmark for multimodal humor comprehension, incorporating facial expressions and voice tones as in MSU [19, 20]. Hessel et al. (2023) [14] intro", + "bbox": [ + 511, + 643, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1538", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Q. Select the option that best explains the underlying intention of the utterance based on the given image.", + "bbox": [ + 106, + 99, + 465, + 127 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5e3b0a1b1a27751d011a6dcbe11cb25d073beac26e2f82c50585b0bb63611535.jpg", + "image_caption": [ + "Figure 2. Description of the Multimodal Intention Disambiguation (MID) task in a Multiple-Choice Question format: Given an input image $(I)$ and an indirect expression $(p_i)$ , the goal is to infer the speaker's hidden intent $(T)$ and select the most likely answer." + ], + "image_footnote": [], + "bbox": [ + 106, + 143, + 151, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hey person1, spot the difference, this parking's a bit too special isn't it?", + "bbox": [ + 163, + 146, + 421, + 172 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d314ec8014afa1b1655e660cbf27f3b14d3fc17bd409690d07d9a53e5c01bf22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 183, + 478, + 311 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "a) The speaker wants person1 to admire the unusually decorated motorcycle in the parking lot.", + "b) The speaker wants person1 to enjoy playing a puzzle game and spot the difference.", + "c) The speaker wants person1 to move the sedan because it's in a handicapped parking spot.", + "d) The speaker wants person1 to move the sedan because it's parked in front of a fire hydrant." + ], + "bbox": [ + 98, + 316, + 475, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duced a benchmark derived from a Cartoon Caption Contest, exploring humor identification and explanation. Baluja et al. (2024) [1] demonstrated that models benefit from multimodal cues in humor understanding. Memes also involve implicature, with multimodal datasets such as MemeCap [15] and MultiBully-Ex [16].", + "bbox": [ + 89, + 551, + 483, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the cues used in multimodal implicature understanding remain simple, primarily appearing in images with a single main object or person, overlooking the importance of interactions between multiple objects and people in real-world scenarios. These limitations underscore the need for more complex cues, as addressed in VAGUE.", + "bbox": [ + 89, + 643, + 483, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multimodal Intention Disambiguation", + "text_level": 1, + "bbox": [ + 89, + 750, + 433, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we outline the structure and rationale behind the format of our primary task, which we term Multimodal Intention Disambiguation (MID). Then, we further specify the necessary components that form the basis of the task.", + "bbox": [ + 89, + 775, + 482, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting", + "text_level": 1, + "bbox": [ + 89, + 847, + 253, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each MID problem comprises an input image $I$ , a direct text expression $p_d$ , and an indirect text expression $p_i$ . Here,", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the direct expression $p_d$ clearly shows the underlying intention of the corresponding $p_i$ and serves as an essential intermediate step of generating $p_i$ .", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To clarify our problem, we assume that all reasoning is confined to the depicted scene and that each expression is spoken by a human who intends for the listener to take a particular action based on the situation. The ultimate objective of the task is to interpret the hidden intention $T$ effectively by leveraging the contextual cues within the image.", + "bbox": [ + 511, + 137, + 906, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate how well models capture such intentions, we adopt a multiple-choice (MCQ) format as the primary setup, as shown in Fig. 2. This decision reflects the fact that certain prompts can lead to multiple plausible outcomes, driven by hierarchical relations (e.g., pick up the chips - snack - food) or by the inherent uncertainty of what action best satisfies the speaker's goal (e.g., an indirect prompt complaining about darkness could be addressed by either turning on a light or opening curtains). Exploring all possible valid interpretations is labor-intensive and often infeasible. Consequently, each MID instance is presented as four distinct options, one correct and three intentionally designed to be incorrect for different reasons (see Sec. 4.2.3), challenging models in both linguistic and visual reasoning.", + "bbox": [ + 511, + 229, + 908, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, let $C$ be the set of all multiple-choice options $c_{n}$ . Given an image $I$ and an indirect prompt $p_i$ , the task is to select the most valid interpretation of $p_i$ from the predefined options, conditioned on the visual context in $I$ . We define this task as follows:", + "bbox": [ + 511, + 441, + 905, + 516 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT (I, p _ {i}) := \\operatorname {a r g m a x} _ {c _ {n} \\in C} \\Pr \\left(c _ {n} \\mid I, p _ {i}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 537, + 906, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Direct and Indirect Expressions", + "text_level": 1, + "bbox": [ + 511, + 566, + 794, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By the design of our task, curating effective input prompts $p_d$ and $p_i$ is crucial for ensuring accurate interpretation. What makes a good prompt, though? In this section, we define and explain the criteria that both direct and indirect expressions must satisfy. For more details on good and bad examples for each criterion, please refer to our Appendix A.", + "bbox": [ + 511, + 589, + 905, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Directness: Relevance and Solvability", + "text_level": 1, + "bbox": [ + 511, + 691, + 821, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relevance The direct prompt is an utterance from the speaker that explicitly conveys its intended meaning without ambiguity. However, it is equally important that this intention aligns with the visual context of the scene. For example, a direct prompt $p_d$ such as \"Hey person1, I want you to stop the fireworks\" clearly expresses its intended action. However, if the corresponding image, as shown in Fig. 2, contains no elements related to fireworks, the prompt is misaligned with the scene. Thus, a direct prompt must not only reveal its intention but also maintain relevance to the image. In the context of our task, relevance is determined by whether a human can reasonably establish a connection between the prompt and the depicted scene.", + "bbox": [ + 511, + 710, + 906, + 907 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1539", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Solvability Relevance alone does not guarantee that a prompt is useful. As outlined in Sec. 3.1, the prompt $p_d$ must explicitly request an action that the listener can reasonably perform. This introduces solvability, which requires that the prompt present a clear and actionable problem. A solvable prompt defines a specific issue that can be addressed independently, ensuring that the listener is not left with multiple competing actions to choose from.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Indirectness: Consistency and Ambiguity", + "text_level": 1, + "bbox": [ + 89, + 224, + 424, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consistency Indirect prompts are designed to obscure their true intention, but they must still convey the same underlying intention as their direct counterpart—essentially requesting the same solution. Since indirect prompts are derived from direct prompts, consistency serves as a key criterion. We define a direct prompt $p_d$ and an indirect prompt $p_i$ as consistent if their interpretations could potentially align in intention. The term \"potentially\" is used because indirect prompts, by nature, may have multiple valid interpretations. For instance, in the earlier \"this is a disaster!\" example, it conveys distress but allows for multiple reasonable responses, such as cleaning the room or providing reassurance.", + "bbox": [ + 89, + 244, + 482, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ambiguity If an indirect prompt is entirely consistent with its direct counterpart without introducing any additional complexity, it becomes indistinguishable from a direct prompt. Therefore, an indirect prompt should conceal its underlying intention, which we define as ambiguity. The key principle behind this criterion is that neither the specific action required nor the key entity involved should be explicitly or implicitly mentioned within the prompt. Once these two elements are concealed, further refinements—such as adjusting the tone to be more indirect, sarcastic, or humorous—can enhance the overall nuance and difficulty of interpretation.", + "bbox": [ + 89, + 446, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. VAGUE Benchmark Construction", + "text_level": 1, + "bbox": [ + 89, + 631, + 401, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VAGUE is a novel benchmark that extends single-modal or simple multimodal ambiguity to more realistic domains and evaluates whether concurrent vision-language models can perform human-like reasoning with complex visual contexts. It comprises 1,677 images, with 1,144 sourced from the VCR [46] dataset and 533 from Ego4D [41], covering diverse contextual scenarios as well as real-world human interactions. On average, VAGUE contains seven objects and four people per image. Each image is paired with a direct expression $p_d$ , an indirect expression $p_i$ , and four multiple-choice answers, along with relevant meta-information. All textual components are generated using GPT-4o, then refined through extensive human rating, selection, and filtering, ensuring a carefully curated benchmark dataset for testing and advancing multimodal reasoning. We provide detailed benchmark statistics and a diversity analysis in Appendix B.2.", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Visual Data Curation", + "text_level": 1, + "bbox": [ + 513, + 90, + 714, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Sampling", + "text_level": 1, + "bbox": [ + 513, + 113, + 627, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VCR [46] The VCR dataset consists of 110K movie scenes sourced from the Large Scale Movie Description Challenge [35] and YouTube clips. These images are curated based on an \"interestingness\" criterion [46], ensuring the presence of at least two people, which promotes interactive scenarios. To prevent redundancy in our dataset, we sample 10K images while carefully avoiding neighboring frames, as adjacent frames exhibit minimal variation. This selection process preserves the contextual diversity of the dataset while maintaining its focus on complex, multi-entity interactions.", + "bbox": [ + 511, + 132, + 906, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ego4D [41] While VCR provides a wide range of contextual diversity, it often includes artificially composed settings that may not fully capture real-world interactions. To address this, we integrate frames from the Ego4D dataset, which offers a more naturalistic depiction of human interactions. We specifically leverage the AV (Audio-Visual), which indicate conversational exchanges between individuals, to ensure the presence of people in the selected frames. Similar to VCR, we avoid neighboring frames to maintain diversity and filter out heavily blurry images to enhance data quality. This process results in 888 candidate images from 94 videos, which serve as the basis for further text processing.", + "bbox": [ + 511, + 297, + 908, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. Object Extraction", + "text_level": 1, + "bbox": [ + 513, + 488, + 684, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure the complexity of the visual information, we extract a list of physical objects present in each image using a tagging model, RAM [49]. This step allows us to easily identify scenes with sufficient visual detail. In the case of VCR, many scenes are relatively simple, often containing only a few objects. Therefore, we sort VCR images by the number of detected objects and retained the top 4,000 as candidates for text processing, ensuring that our benchmark primarily consists of rich visual cues in contextually diverse scenes. Please refer to Appendix B.3 for more details.", + "bbox": [ + 511, + 508, + 908, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.3. Person Indicator", + "text_level": 1, + "bbox": [ + 513, + 669, + 676, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our task, we assume that the speaker is outside the scene, viewing the image and talking to a person in it. However, identifying the addressee is not always straightforward, as images mostly contain multiple individuals. While grounding the specific person referenced in the utterance could introduce additional complexity, it is not the primary focus of our evaluation. To clarify the listener, we assign an indicator tag to each person in the image as shown in Fig. 3. For VCR, we use their existing annotations [46], while for Ego4D, where bounding boxes are not fully available, we employ YOLOv11 [18] to detect and annotate humans. While this provides a straightforward method for grounding the target person, it requires models to perform basic Optical Character Recognition (OCR). Therefore, we conduct exper", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1540", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e1c2cc12c6b85ca62f868102b42281b9ec99a2fd7df1e4755d09e2cd23bc692.jpg", + "image_caption": [ + "Figure 3. Overview of the data generation process. Based on human-defined criteria and instructions, GPT-4o [31] generates initial data, which are then rated and filtered by humans to ensure quality. Since generating high-quality indirect expressions from raw images is challenging, the process follows these steps: generating a direct expression and intention from the image, creating an indirect expression using information from the previous step, and producing answer candidates based on all the information gathered so far. In the answer candidates, FS stands for Fake Scene Understanding, SU stands for Superficial Understanding, and NE stands for Nonexistent Entity. FS evaluates global hallucination, where the model misinterprets visual context from an entirely different image (e.g., an outdoor camping scene with a candle), while NE addresses local hallucination, where only a specific object is replaced with a fabricated one (e.g., curtains)." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 898, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iments to evaluate the OCR capabilities of the models used in our study. See Appendix B.4 for details.", + "bbox": [ + 89, + 430, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Multimodal Expression Synthesis", + "text_level": 1, + "bbox": [ + 89, + 470, + 382, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, VAGUE's textual expressions are generated first by the model and undergo an extensive process of human rating and filtering. We use GPT-4o [31] for all text processing. The instructions used for generating direct $(p_d)$ and indirect $(p_i)$ expressions are provided in Appendix B.5, while the instructions for generating answer candidates for multiple-choice questions are detailed in Appendix B.7.", + "bbox": [ + 89, + 494, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Direct Expressions", + "text_level": 1, + "bbox": [ + 89, + 608, + 266, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The direct expression $p_d$ serves as a crucial foundation for crafting the indirect one $p_i$ , as both share the same underlying intention. To ensure that $p_d$ adhere to the principles of relevance and solvability discussed in Sec. 3.2.1, we generate $p_d$ conditioned on the input image $I$ and a task prompt that explicitly defines these criteria. During the generation process, we also instruct the model to output a solution triplet in the format: (subject, action, object). Since direct expressions explicitly state their intentions, extracting each component of the solution triplet is straightforward. To maintain consistency with the visual context, the \"object\" in the triplet is restricted to physical objects we extracted in Sec. 4.1.2.", + "bbox": [ + 89, + 628, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After generating $p_d$ for all candidate images, human raters evaluate each prompt based on relevance and solvability, assigning scores from 1 to 5. Only those prompts that receive a rating of 4 or 5 are retained for further use. The detailed rating criteria for human verification and an example of the rating process are provided in Fig. J13.", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Indirect Expressions", + "text_level": 1, + "bbox": [ + 513, + 430, + 702, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that the indirect expressions $p_i$ maintain both ambiguity and fluency while aligning with the true intent $T$ , we adopt a two-stage process: proposal and selection.", + "bbox": [ + 511, + 449, + 906, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the initial step, the model is prompted to generate three distinct candidate options. Each candidate follows the criteria outlined in Section 3.2.2, but with explicit instructions to incorporate different linguistic strategies: sarcasm, humor, and meme/idiomatic expressions, respectively. This approach ensures diversity in the generated responses. In the second step, human annotators evaluate the three candidates and select the one that best aligns with the intended indirectness. The selected prompt is then rated on a scale from 1 to 5 based on more specific criteria. Only those prompts that receive a score of 3 or higher are retained for use in the dataset. The detailed rating criteria for human verification and an example of rating process are provided in Fig. J14", + "bbox": [ + 511, + 496, + 908, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3. Counterfactual Choices", + "text_level": 1, + "bbox": [ + 511, + 702, + 723, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating high-quality counterfactual choices is crucial. To enable detailed analysis of model weaknesses in multimodal intent disambiguation, we design interpretable counterfactual choices that provide more plausible alternatives.", + "bbox": [ + 511, + 720, + 906, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fake Scene Understanding The first counterfactual choice is an interpretation that could arise when the model largely misinterprets the image. This process is conducted in two steps. In the first step, a fake caption is generated by assuming an imaginary scene that can be aligned with the indirect expression but is inconsistent with the true intent. The caption of fake scene is then combined with the speaker's", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1541", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "indirect statement to derive the most likely interpretation.", + "bbox": [ + 89, + 90, + 468, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Superficial Understanding The subsequent choice corresponds to an interpretation generated when the model fails to deeply reason about the implicit intent of the sentence and instead relies on surface-level meaning. We enforce the model to focus only on the literal wording, without considering any implied or deeper meaning of the indirect sentence. These answer choices are generated alongside the indirect expression $p_i$ . During the indirect selection phase, the corresponding superficially understood choice is selected together, maintaining coherence between them.", + "bbox": [ + 89, + 117, + 485, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nonexistent Entity The last choice arises when the model interprets the text correctly but fails to adequately consider the details of the image, resulting in a plausible yet incorrect choice. This resembles the correct answer in structure, but replaces the key object in the solution with one that does not exist in the image. To prevent the task from becoming too easy by generating highly irrelevant objects, we constrain the substituted object to one that, while absent from the image, is highly expected to be present in the scene and could replace the original entity. To identify such entities, the model is provided with the image as input to choose objects that align with the scene's context. This method ensures that the counterfactual choice leverages the expected coherence between the scene and its potential entities while rigorously testing the model's attention to visual details.", + "bbox": [ + 89, + 280, + 483, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 522, + 222, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models We use the following models in our experiments. The detailed descriptions of each model are in Appendix C.", + "bbox": [ + 89, + 547, + 482, + 578 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Phi3.5-Vision-Instruct (4B) [29]", + "- LLaVA Onevision (7B) [23]", + "Qwen2.5-VL-Instruct (7B, 72B) [43]", + "- InternVL-2.5-MPO (8B, 26B) [6]", + "- Idefics2 (8B) [22]", + "LLaVA NeXT Vicuna (13B) [24]", + "- Ovis2 (16B) [28]", + "GPT-4o [31]", + "Gemini 1.5 Pro [12]", + "- InternVL-3 (38B) [42]" + ], + "bbox": [ + 93, + 579, + 351, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. MLLMs Benefit from Visual Cues", + "text_level": 1, + "bbox": [ + 89, + 741, + 387, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our first objective is to assess how effectively MLLMs leverage visual cues to resolve ambiguity in utterances. To this end, we systematically control the level of detail in the visual cues provided to the models and measure their accuracy in inferring the speaker's true intent. Performance is evaluated in both multiple-choice and free-form settings. For clarity, we primarily report multiple-choice accuracy, deferring free-form results to Appendix I.", + "bbox": [ + 89, + 763, + 483, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider three levels of visual cues:", + "bbox": [ + 109, + 885, + 370, + 898 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Language Models (LMs) receive no visual input, requiring models to rely on superficial textual priors such as common-sense knowledge of sarcasm or humor to determine intent.", + "- Socratic Models (SMs) [47] use text-only LMs but incorporate short image captions (up to two or three sentences) as additional input. This short generic caption may lack sufficient detail, which may be insufficient for accurately inferring intent. Each SM model generated its own image captions and used them in subsequent processing.", + "- Visual Language Models (VLMs) receive the raw image input, enabling a more direct interpretation of visual cues." + ], + "bbox": [ + 513, + 90, + 906, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results The results in Tab. 1 indicate that MLLMs can leverage visual cues, albeit to a limited extent, since SMs and VLMs consistently outperform LMs across all evaluated models. Additionally, more detailed visual input generally improves performance, with VLMs surpassing SMs in most cases, except in the case of proprietary models. This exception is further analyzed in Sec. 5.2. Both the VCR and Ego4D subsets exhibit similar performance trends, demonstrating the generalizability of our findings across both staged and real-world scenarios. Finally, the consistently low performance of LMs further reinforces the validity of our dataset as a multimodal benchmark.", + "bbox": [ + 511, + 284, + 908, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Analysis on Failure Modes", + "text_level": 1, + "bbox": [ + 511, + 476, + 754, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we examine the ways in which models fail to infer the true intent and how these failure patterns vary with the level of visual cues provided. As shown in Fig. 3, our multiple-choice questions include three distinct types of incorrect answer candidates. We assess the model's raw visual understanding using the Fake Scene Understanding (FS) and Nonexistent Entity (NE) candidates, which test whether the model can correctly interpret the scene without being misled by fabricated or nonexistent elements. Conversely, the Superficial Understanding (SU) candidate evaluates the model's reasoning ability, testing whether it can go beyond surface-level perception to infer intent. We provide the full table of failure modes in the MCQ setting in Appendix I", + "bbox": [ + 511, + 498, + 908, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results Figure 4 illustrates how frequently each model selects different types of incorrect answers instead of the correct intent. Among the error types, Superficial Understanding (SU) is the most common. This indicates that while models generally succeed in recognizing basic visual details, they often fail to reason deeply about those visual cues to accurately infer the underlying intent of the speaker. However, proprietary models exhibit fewer SU-related errors, indicating stronger reasoning capabilities.", + "bbox": [ + 511, + 703, + 908, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, stronger visual cues improve accuracy across all models and failure types. This improvement highlights the crucial role of visual conditioning in reducing both vision-based errors (FS and NE) and reasoning-related failures (SU).", + "bbox": [ + 511, + 840, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1542", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/62bf245532a064985b3cff71bbdd8486b026798d8780805b5785c577faafa5ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelVAGUE-VCRVAGUE-Ego4D
LM (L)SM (L+V)VLM (L+V)LM (L)SM (L+V)VLM (L+V)
Phi3.5-Vision-Instruct (4B)26.635.3 (↑ 8.7)46.0 (↑ 19.4)22.531.1 (↑ 8.6)42.4 (↑ 19.9)
LLaVA-Onevision (7B)13.129.4 (↑ 16.3)43.1 (↑ 30.0)11.329.5 (↑ 18.2)43.2 (↑ 31.9)
Qwen2.5-VL-Instruct (7B)11.125.6 (↑ 14.5)46.8 (↑ 35.7)9.828.0 (↑ 18.2)48.4 (↑ 38.6)
InternVL-2.5-MPO (8B)23.048.4 (↑ 25.4)63.9 (↑ 40.9)24.254.0 (↑ 29.8)66.8 (↑ 42.6)
Idefics2 (8B)13.921.1 (↑ 7.2)58.7 (↑ 44.8)14.818.2 (↑ 3.4)58.3 (↑ 43.5)
LLaVA-NeXT-vicuna (13B)24.237.2 (↑ 13)46.4 (↑ 22.2)20.334.1 (↑ 13.8)52.5 (↑ 32.2)
Ovis2 (16B)21.923.8 (↑ 1.9)24.5 (↑ 3.6)20.525.3 (↑ 4.8)25.7 (↑ 5.2)
InternVL-2.5-MPO (26B)21.248.5 (↑ 27.3)63.7 (↑ 42.5)21.855.2 (↑ 33.4)68.7 (↑ 46.9)
InternVL-3 (38B)24.847.2 (↑ 22.4)63.6 (↑ 38.8)18.047.5 (↑ 29.5)59.8 (↑ 41.8)
Qwen2.5-VL-Instruct (72B)29.655.6 (↑ 26.0)74.2 (↑ 44.6)26.859.3 (↑ 32.5)69.8 (↑ 43.0)
GPT-4o46.469.5 (↑ 23.1)65.1 (↑ 18.7)48.267.5 (↑ 19.3)63.6 (↑ 15.3)
Gemini-1.5-Pro43.262.4 (↑ 19.2)60.6 (↑ 17.4)40.360.6 (↑ 20.3)60.6 (↑ 20.3)
", + "bbox": [ + 91, + 88, + 906, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Experiments on the Multimodal Intention Disambiguation (MID) task with varying levels of visual cues. We report the accuracy $(\\%)$ of the Multiple-Choice Question. $\\uparrow$ indicates the performance gain from visual cues, i.e. increment compared to the LM setting.(L) denotes the use of language input only, while $(\\mathrm{L} + \\mathrm{V})$ indicates the incorporation of visual cues. The noticeable increase in accuracy across LM, SM, and VLM demonstrates that the introduction of detailed visual cues is beneficial for the task.", + "bbox": [ + 89, + 347, + 906, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f063bf44d3c5ca75ea7e8831036cb2b6bff129c5ecbd71bdc54b08faf9ac3d72.jpg", + "image_caption": [ + "Figure 4. We present a bar plot to analyze the distribution of incorrect answer choices selected by each model. Each number represents how frequently a given choice was selected from the 1,677 items in the dataset. The counterfactual choice categories are FS (Fake Scene Understanding), SU (Superficial Understanding), and NE (Nonexistent Entity). We use distinct colors to represent LM (Language Models), SM (Socratic Models), and VLM (Visual-Language Models)." + ], + "image_footnote": [], + "bbox": [ + 91, + 417, + 908, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, proprietary models perform better with captioned inputs (SM) than with raw images (VLM). A closer examination of Fig. 4 reveals that this discrepancy arises from vision-based failures (FS and NE) rather than reasoning-centric failure (SU). This indicates that their captioning ability potentially allows them to obtain more sophisticated information while reducing hallucination in the images. The supporting experiments and explanations are provided in Appendix F.", + "bbox": [ + 89, + 672, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison with human", + "text_level": 1, + "bbox": [ + 89, + 803, + 318, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate our benchmark and establish an upper bound for performance, we assess human accuracy, highlighting the gap between existing models and human capability. This evaluation follows the multiple-choice setup within the VLM setting, using a subset of 400 samples. As shown", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Tab. 2, human performance reaches $94\\%$ , demonstrating near-perfect accuracy. Although proprietary models and some large-sized models show a decent performance, a notable performance gap ( $\\sim 20\\%$ ) still exists compared to human evaluators. This result underscores the significant gap between AI models' multimodal reasoning capabilities and human-level understanding when inferring hidden intent, suggesting that visual perception alone, even when accurate, is insufficient without deeper cognitive integration. This performance gap is due to the models' tendency to rely on surface-level text rather than understanding deeper visual-textual implications. Thus, advancing multimodal reasoning likely requires models to integrate higher-order cognitive processes, like commonsense reasoning and pragmatic understanding, into visual interpretation tasks. Refer", + "bbox": [ + 511, + 672, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1543", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/54538f880812ae8c5bba8b5ff31d0c85d4580b2bec406d2a9e499f683d40dfea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAcc (%)FSSUNECorrect
Ovis2 (16B)23.01191711892
LLaVA-Onevision (7B)41.04318310164
Phi3.5-Vision-Instruct (4B)44.36013231177
Qwen2.5-VL-Instruct (7B)47.04715213188
LLaVA-NeXT-icuna (13B)48.04814317192
Idefics2 (8b)57.02812024228
Gemini-1.5-Pro60.3607326241
InternVL-3 (38B)61.5389422246
InternVL-2.5-MPO (8B)61.8429516247
GPT-4o62.3616327249
InternVL-2.5-MPO (26B)63.03610111252
Qwen2.5-VL-Instruct (72B)72.3396012289
Human94.01248374
", + "bbox": [ + 91, + 88, + 482, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1c4c73acf4369df3333c7bbaad8ef0c6b89d6f8a77aa4e034ea3936ac1d2dce3.jpg", + "table_caption": [ + "Table 2. Performance across models and humans on a sampled set of 400 questions. The results show that humans outperform models by a margin of over $20\\%$ ." + ], + "table_footnote": [], + "table_body": "
ModelTypeAcc (%)Incorrect count
FSSUNE
GPT-4oSM68.9165240117
SM+CoT69.5 (↑ 0.6)165241105
VLM64.6226226141
VLM+CoT66.4 (↑ 1.8)16215685
Gemini-1.5-ProSM61.817637483
SM+CoT61.0 (↓ 0.8)19036794
VLM60.6249264145
VLM+CoT64.4 (↑ 3.8)213267117
", + "bbox": [ + 91, + 345, + 478, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Result of Chain-of-Thought (CoT) experiments on proprietary models, in both SM and VLM settings. $\\uparrow$ and $\\downarrow$ indicate an increase and decrease in accuracy when zero-shot CoT is applied.", + "bbox": [ + 89, + 526, + 482, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to Appendix E for details on the selected subset and human evaluation setup.", + "bbox": [ + 89, + 595, + 482, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Chain-of-Thought Experiments", + "text_level": 1, + "bbox": [ + 89, + 638, + 370, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the strong reasoning demands of multimodal intent deduction, we further explore the effectiveness of Chain-of-Thought (CoT) prompting [44] in enhancing the reasoning capabilities of MLLMs. The CoT prompt templates, provided in Fig. J21 and Fig. J22, are designed to explicitly ground the reasoning process, reducing hallucinations. While our main results focus on proprietary models, owing to their superior suitability for zero-shot CoT reasoning, we also present experiments and analyses for open-source models ranging from 4B to 72B in Appendix G.", + "bbox": [ + 89, + 661, + 482, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results As shown in Tab. 3, CoT prompting improves performance for raw image inputs (VLM), while showing no clear trend and remaining at a similar level for image caption inputs (SM). One possible explanation for this discrepancy is that CoT primarily enhances reasoning by improving", + "bbox": [ + 89, + 825, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "grounding and reducing hallucinations. Since image captions inherently contain fewer hallucinations, SMs see some benefit from CoT prompting albeit at the cost of reduced detail. Additionally, the performance improvements observed with CoT prompting are consistent across different types of false answer candidates, suggesting a generalizable effect in enhancing reasoning quality.", + "bbox": [ + 511, + 90, + 906, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 210, + 633, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present VAGUE (Visual Contexts CIArify ambiGUous Expressions), a new benchmark aimed at assessing models' ability to interpret nuanced communication in complex multimodal scenarios. Our results show that models benefit from visual information when inferring the underlying intention of indirect expressions, as evidenced by their improved performance with increasing levels of visual cues. However, a significant disparity persists between machine capabilities and humans. To gain deeper insights into model inaccuracies, we design multiple-choice questions that explicitly address failure points, enabling a systematic and quantitative evaluation of the reasons behind performance. The primary challenge identified is the tendency of multimodal models to inadequately integrate visual cues, relying instead on the literal interpretation of textual information. This shortcoming highlights the need for further research, and we anticipate that VAGUE will open up promising avenues for developing systems capable of deeper multimodal reasoning to enhance AI's ability to engage in human-like interactions.", + "bbox": [ + 511, + 234, + 908, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations", + "text_level": 1, + "bbox": [ + 511, + 536, + 635, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we acknowledge that there are cultural and linguistic limitations. We incorporate sarcastic, humorous, and idiomatic implicatures in generating indirect expressions. However, since the initial data drafts are created by a model (GPT-4o [31]), they may reflect cultural biases present in its training data. To prevent any potential ethical issues, all human annotators are instructed to remove any content considered problematic or discriminatory during the rating and filtering process. Also, all textual expressions in VAGUE are limited to English. Therefore, we encourage future exploration of indirect expressions across diverse languages and cultures. The second limitation pertains to the dependency of certain meta-information on the quality of the parent dataset and the performance of the models utilized during the dataset generation. We observe that bounding box annotations from YOLOv11 [18] are occasionally duplicated, leading to a reported number of people higher than actually present. Likewise, the tagging model RAM [49] sometimes misidentified objects. Therefore, we remove any corrupted instances that could undermine the integrity of the task during the human rating and filtering process.", + "bbox": [ + 511, + 561, + 908, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1544", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ashwin Baluja. Text is not all you need: Multimodal prompting helps llms understand humor. arXiv preprint arXiv:2412.05315, 2024. 3", + "[2] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1", + "[3] Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi, and Andreas Bulling. Explicit modelling of theory of mind for belief prediction in nonverbal social interactions. In Proc. 27th European Conference on Artificial Intelligence (ECAI), pages 1-8, 2024. 2", + "[4] Tuhin Chakrabarty, Arkadiy Saakyan, Debanjan Ghosh, and Smaranda Muresan. FLUTE: Figurative language understanding through textual explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7139-7159, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2", + "[5] Zhawnen Chen, Tianchun Wang, Yizhou Wang, Michal Kosinski, Xiang Zhang, Yun Fu, and Sheng Li. Through the theory of mind's eye: Reading minds with multimodal video large language models. arXiv preprint arXiv:2406.13763, 2024. 2", + "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6, 3", + "[7] Christine Cuskley, Rebecca Woods, and Molly Flaherty. The limitations of large language models for understanding human language and cognition. Open Mind, 8:1058-1083, 2024. 2", + "[8] Hang Du, Guoshun Nan, Sicheng Zhang, Binzhu Xie, Junrui Xu, Hehe Fan, Qimei Cui, Xiaofeng Tao, and Xudong Jiang. Docmsu: A comprehensive benchmark for document-level multimodal sarcasm understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17933-17941, 2024. 2", + "[9] Jiafei Duan, Samson Yu, Nicholas Tan, Li Yi, and Cheston Tan. Boss: A benchmark for human belief prediction in object-context scenarios. arXiv preprint arXiv:2206.10665, 2022. 2", + "[10] Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. CVPR 2024, 2024. 4", + "[11] Kanishk Gandhi, Jan-Philipp Fraenkel, Tobias Gerstenberg, and Noah D. Goodman. Understanding social reasoning in language models with language models. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track, 2023. Dataset and Benchmarks Track. 2", + "[12] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6, 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Md Kamrul Hasan, Wasifur Rahman, AmirAli Bagher Zadeh, Jianyuan Zhong, Md Iftekhar Tanveer, Louis-Philippe Morency, and Mohammed (Ehsan) Hoque. UR-FUNNY: A multimodal language dataset for understanding humor. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2046–2056, Hong Kong, China, 2019. Association for Computational Linguistics. 2", + "[14] Jack Hessel, Ana Marasovic, Jena D. Hwang, Lillian Lee, Jeff Da, Rowan Zellers, Robert Mankoff, and Yejin Choi. Do androids laugh at electric sheep? humor \"understanding\" benchmarks from the new yorker caption contest. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 688-714, Toronto, Canada, 2023. Association for Computational Linguistics. 2", + "[15] EunJeong Hwang and Vered Shwartz. Memecap: A dataset for captioning and interpreting memes. arXiv preprint arXiv:2305.13703, 2023. 3", + "[16] Prince Jha, Krishanu Maity, Raghav Jain, Apoorv Verma, Sriparna Saha, and Pushpak Bhattacharyya. Meme-ingful analysis: Enhanced understanding of cyberbullying in memes through multimodal explanations. arXiv preprint arXiv:2401.09899, 2024.3", + "[17] Chuanyang Jin, Yutong Wu, Jing Cao, Jiannan Xiang, Yen Ling Kuo, Zhiting Hu, Tomer Ullman, Antonio Torralba, Joshua Tenenbaum, and Tianmin Shu. MMToM-QA: Multimodal theory of mind question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16077-16102, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2", + "[18] Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements, 2024. 4, 8", + "[19] Shivani Kumar, Atharva Kulkarni, Md Shad Akhtar, and Tan-moy Chakraborty. When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5956-5968, Dublin, Ireland, 2022. Association for Computational Linguistics. 2", + "[20] Shivani Kumar, Ishani Mondai, Md Shad Akhtar, and Tanmoy Chakraborty. Explaining (sarcastic) utterances to enhance affect understanding in multimodal dialogues. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 2", + "[21] Yash Kumar Lal and Mohaddeseh Bastan. SBU figures it out: Models explain figurative language. In Proceedings of the 3rd Workshop on Figurative Language Processing (FLP), pages 143-149, Abu Dhabi, United Arab Emirates (Hybrid), 2022. Association for Computational Linguistics. 2", + "[22] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language mod" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1545", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "els? Advances in Neural Information Processing Systems, 37: 87874-87907, 2025. 6, 3", + "[23] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 3", + "[24] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 6, 3", + "[25] Kevin Y. Li, Sachin Goyal, Joao D. Semedo, and J. Zico Kolter. Inference optimal vlms need fewer visual tokens and more parameters, 2025. 4", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing. 1, 3", + "[27] Annie Louis, Dan Roth, and Filip Radlinski. \"I'd rather just go to bed\": Understanding indirect answers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7411-7425, Online, 2020. Association for Computational Linguistics. 2", + "[28] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 6, 3", + "[29] Microsoft. Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6, 3", + "[30] Maryam Sadat Mirzaei, Kourosh Meshgi, and Satoshi Sekine What is the real intention behind this question? dataset collection and intention classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13606-13622, Toronto, Canada, 2023. Association for Computational Linguistics. 2", + "[31] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 8, 3", + "[32] Kun Ouyang, Liqiang Jing, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Sentiment-enhanced graph-based sarcasm explanation in dialogue. arXiv:2402.11414, 2024. 2", + "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 3, 4", + "[34] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 2", + "[35] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Chris Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. International Journal of Computer Vision, 2017. 4", + "[36] Melanie Sclar, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia Tsvetkov. Minding language models' (lack of) theory of mind: A plug-and-play multi-character belief" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 13960-13980, Toronto, Canada, 2023. Association for Computational Linguistics. 2", + "[37] Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin, Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. Muma-tom: Multi-modal multi-agent theory of mind, 2024. 2", + "[38] Settaluri Sravanthi, Meet Doshi, Pavan Tankala, Rudra Murthy, Raj Dabre, and Pushpak Bhattacharyya. PUB: A pragmatics understanding benchmark for assessing LLMs' pragmatics capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12075-12097, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2", + "[39] Chuandong Su, Fumiyo Fukumoto, Xiaoxi Huang, Jiyi Li, Rongbo Wang, and Zhiqun Chen. DeepMet: A reading comprehension paradigm for token-level metaphor detection. In Proceedings of the Second Workshop on Figurative Language Processing, pages 30-39, Online, 2020. Association for Computational Linguistics. 2", + "[40] Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase. DIRECT: Direct and indirect responses in conversational text corpus. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1980-1989, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2", + "[41] Ego4d Team. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 2, 4", + "[42] OpenGVLab Team. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025. 6, 3", + "[43] Qwen Team. Qwen2.5-v1 technical report, 2025. 6, 3", + "[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 8", + "[45] Alex Wilf, Leena Mathur, Sheryl Mathew, Claire Ko, Youssof Kebe, Paul Pu Liang, and Louis-Philippe Morency. Social-iq 2.0 challenge: Benchmarking multimodal social understanding. https://github.com/abwilf/Social-IQ-2.0-Challenge, 2023.2", + "[46] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6713-6724, 2018. 2, 4, 1", + "[47] Andy Zeng, Maria Attarian, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, 2023. 2, 6", + "[48] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1546", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[49] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1724-1732, 2024. 4, 8, 1", + "bbox": [ + 91, + 90, + 486, + 176 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1547", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_model.json b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6564ca44920f36c16dabf4dc0e5f7e1aad4744d7 --- /dev/null +++ b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.154 + ], + "angle": 0, + "content": "VAGUE: Visual Contexts Clarify Ambiguous Expressions" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.193, + 0.828, + 0.213 + ], + "angle": 0, + "content": "Heejeong \\(\\mathrm{Nam}^{*1}\\), Jinwoo Ahn\\(^{*2}\\), Keummin \\(\\mathrm{Ka}^3\\), Jiwan Chung\\(^3\\), and Youngjae \\(\\mathrm{Yu}^{\\dagger 3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.224, + 0.576, + 0.242 + ], + "angle": 0, + "content": "Brown University" + }, + { + "type": "text", + "bbox": [ + 0.442, + 0.242, + 0.558, + 0.26 + ], + "angle": 0, + "content": "\\(^{2}\\)UC Berkeley" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.26, + 0.575, + 0.278 + ], + "angle": 0, + "content": "3Yonsei University" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.323, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.486, + 0.673 + ], + "angle": 0, + "content": "Human communication often relies on visual cues to resolve ambiguity. While humans can intuitively integrate these cues, AI systems often find it challenging to engage in sophisticated multimodal reasoning. We introduce VAGUE, a benchmark evaluating multimodal AI systems' ability to integrate visual context for intent disambiguation. VAGUE consists of 1.6K ambiguous textual expressions, each paired with an image and multiple-choice interpretations, where the correct answer is only apparent with visual context. The dataset spans both staged, complex (Visual Commonsense Reasoning) and natural, personal (Ego4D) scenes, ensuring diversity. Our experiments reveal that existing multimodal AI models struggle to infer the speaker's true intent. While performance consistently improves from the introduction of more visual cues, the overall accuracy remains far below human performance, highlighting a critical gap in multimodal reasoning. Analysis of failure cases demonstrates that current models fail to distinguish true intent from superficial correlations in the visual scene, indicating that they perceive images but do not effectively reason with them. We release our code and data at https://hazel-heejeong-nam.github.io/vague/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.702, + 0.222, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.485, + 0.818 + ], + "angle": 0, + "content": "Human communication is inherently contextual; for example, exclaiming \"Hey, this is a disaster!\" upon seeing a cluttered room conveys frustration or exaggeration rather than referring to an actual catastrophe. Without surrounding cues, textual dialogues can be ambiguous, making it difficult for models to accurately capture intent and nuance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.485, + 0.865 + ], + "angle": 0, + "content": "We consider the case of visual contextual cues. Consider Fig. 1, which depicts a speaker making a remark in a certain situation. Without specifying contexts introduced" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.33, + 0.601, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.605, + 0.331, + 0.847, + 0.355 + ], + "angle": 0, + "content": "Hey person1, looks like you're planning a party for the neighborhood squirrels." + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.369, + 0.581, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.369, + 0.667, + 0.378 + ], + "angle": 0, + "content": "Visual Context 1" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.381, + 0.707, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.369, + 0.771, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.369, + 0.859, + 0.378 + ], + "angle": 0, + "content": "Visual Context 2" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.38, + 0.901, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.482, + 0.629, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.486, + 0.809, + 0.497 + ], + "angle": 0, + "content": "The speaker wants person1 to..." + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.521, + 0.672, + 0.531 + ], + "angle": 0, + "content": "close the cereal box" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.52, + 0.719, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.516, + 0.875, + 0.538 + ], + "angle": 0, + "content": "hand out the envelop to the squirrels" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.555, + 0.908, + 0.626 + ], + "angle": 0, + "content": "Figure 1. A motivating example demonstrating the importance of visual context in understanding intention. Without a predetermined context, a single expression can convey multiple different intentions. The textual expression and Context 1 are from our dataset, while Context 2 is generated using DALL-E 3 [2] to help understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.908, + 0.734 + ], + "angle": 0, + "content": "from visual cues, the speaker's intention can vary, thus remaining ambiguous. This implies that the visual contexts play important roles in communication, raising the question: can AI systems integrate visual cues with ambiguous dialogue to infer the speaker's intent?" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.91, + 0.903 + ], + "angle": 0, + "content": "We introduce Visual Contexts ClArify ambiGUous Expressions (VAGUE), a benchmark consisting of 1.6K ambiguous textual expressions, each paired with a single image. VAGUE aims to model diverse and natural human-to-human interactions by setting each image as the speaker's viewpoint, where the speaker implicitly requests a certain action from a person within their field of view. We define the problem addressed through this setup as Multimodal Intention Disambiguation (MID), which involves reasoning about the most plausible request conditioned on visual context. Each sample in VAGUE is annotated with four multiple-choice candi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.298, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.298, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1537" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.274 + ], + "angle": 0, + "content": "dates, ensuring clarity and preventing multiple valid answers caused by paraphrasing or hierarchical inclusion of meaning. The dataset is meticulously curated to ensure visual dependency; the ground-truth candidate is only preferable when considering the visual context. VAGUE includes visual scenes from both artificial sources (Visual Commonsense Reasoning [46]) and real-world scenarios (Ego4D [41]), capturing a broad spectrum of scene complexity and naturalness. The textual expressions in VAGUE are initially generated by GPT-4o [31] following instructions, then reviewed through extensive human rating and filtering to ensure naturalness and alignment with the corresponding images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.486, + 0.513 + ], + "angle": 0, + "content": "Experiments on VAGUE demonstrate that existing multimodal AI models struggle to infer a speaker's true intent in a multimodal setting. First, although models can leverage visual context—as seen by a performance progression from text-only Language Models (LMs) to pipelined Socratic Models (SMs) [47] and ultimately to end-to-end Visual Language Models (VLMs)—their overall accuracy remains significantly lower than that of humans, indicating a failure to capture the true intent. A closer analysis of failure cases reveals that the primary source of error is the models' inability to distinguish the true intent from a superficial understanding of the visual context. In other words, even though these multimodal systems can perceive the image content, they cannot effectively use this information to reason about the speaker's true intent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.485, + 0.648 + ], + "angle": 0, + "content": "In conclusion, we introduce a benchmark that exposes the limitations of current models in integrating visual cues with intent comprehension and identifies their primary failure mode. We anticipate that VAGUE will serve as a testing ground for the development of future multimodal conversational or embodied agents—systems that combine robust visual perception with nuanced conversational reasoning to effectively respond to user requests in complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.659, + 0.322, + 0.674 + ], + "angle": 0, + "content": "Our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.687, + 0.483, + 0.76 + ], + "angle": 0, + "content": "- VAGUE: a novel benchmark for evaluating multimodal intention disambiguation. Validated through extensive human filtering, VAGUE is designed for robust quantitative assessment by ensuring both the ambiguity of queries and the visual (in)dependency to answer candidates." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.762, + 0.483, + 0.821 + ], + "angle": 0, + "content": "- Carefully curated 1,677 scene images sourced from VCR [46] and Ego4D [41], capturing a wide range of scene complexity, diversity, and naturalness to ensure VAGUE's generalizability across various contexts." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.823, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- Experimental results highlighting a critical challenge in multimodal intention disambiguation: while existing models can perceive visual cues, they fail to effectively integrate this information into reasoning to deduce the speaker's true intent." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.687, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.768, + 0.132 + ], + "angle": 0, + "content": "2.1. Multimodal Theory of Mind" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.908, + 0.274 + ], + "angle": 0, + "content": "Theory of Mind (ToM) refers to the ability to infer and reason about the intentions of others based on available information [34], where recent language models still struggle with relevant tasks [7] highlighting the need for dedicated research in this area. Initially, various methods and benchmarks have been proposed in unimodal settings, relying on text-based approaches [11, 36]. However, these methods often fail to capture the richness of real-world interactions, which often require integrating both linguistic and visual cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.909, + 0.548 + ], + "angle": 0, + "content": "Moving beyond text-only contexts, recent work has incorporated visual information. MMToM [17] introduces a benchmark where models must process both visual and textual cues to solve question-answering tasks related to ToM. The BOSS dataset [9] is a multimodal dataset collected in situations where nonverbal communication is required. It is used to evaluate whether human beliefs can be inferred based on nonverbal cues during social interactions. Similarly, Chen et al. (2024) [5] propose a Video ToM model that leverages key video frames and transcripts, demonstrating improved reasoning on the Social-IQ 2.0 dataset [45]. MuMA-ToM [37] further extends this direction by assessing ToM reasoning in multi-agent interactions, evaluating a model's ability to infer human beliefs and goals based on video and text inputs. MToMnet [3] introduces a ToM-based neural network that integrates contextual cues, such as scene videos and object locations, with person-specific cues, to predict human beliefs in specific scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.549, + 0.909, + 0.609 + ], + "angle": 0, + "content": "However, progress in multimodal ToM remains constrained not only by the scarcity of high-quality datasets [5] but also by the lack of explicit consideration for the ambiguity and indirectness inherent in human communication." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.857, + 0.637 + ], + "angle": 0, + "content": "2.2. Multimodal Implicature Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Implicature and the ambiguity that arises from it naturally emerge in everyday human conversation, requiring pragmatic understanding [38]. Early research on implicature understanding has primarily been conducted in text-only settings [27, 30, 40], with some studies specifically focusing on figurative language and metaphor [4, 21, 39]. However, since the ambiguity of standalone text is inherently limited, recent studies have extended to multiple modalities. One example is multimodal sarcasm understanding (MSU). WITS [19] and MOSES [20] are benchmarks for sarcasm explanation, both providing the speaker's emotion and voice tone as cues. DocMSU [8] is a document-level benchmark for sarcasm localization and detection. To improve MSU, EDGE [32], a graph-based approach, achieved strong performance. UR-FUNNY [13] is a benchmark for multimodal humor comprehension, incorporating facial expressions and voice tones as in MSU [19, 20]. Hessel et al. (2023) [14] intro" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1538" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.107, + 0.1, + 0.466, + 0.128 + ], + "angle": 0, + "content": "Q. Select the option that best explains the underlying intention of the utterance based on the given image." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.144, + 0.153, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.147, + 0.422, + 0.174 + ], + "angle": 0, + "content": "Hey person1, spot the difference, this parking's a bit too special isn't it?" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.184, + 0.48, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.318, + 0.46, + 0.347 + ], + "angle": 0, + "content": "a) The speaker wants person1 to admire the unusually decorated motorcycle in the parking lot." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.353, + 0.41, + 0.381 + ], + "angle": 0, + "content": "b) The speaker wants person1 to enjoy playing a puzzle game and spot the difference." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.383, + 0.472, + 0.416 + ], + "angle": 0, + "content": "c) The speaker wants person1 to move the sedan because it's in a handicapped parking spot." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.422, + 0.477, + 0.451 + ], + "angle": 0, + "content": "d) The speaker wants person1 to move the sedan because it's parked in front of a fire hydrant." + }, + { + "type": "list", + "bbox": [ + 0.099, + 0.318, + 0.477, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.468, + 0.483, + 0.524 + ], + "angle": 0, + "content": "Figure 2. Description of the Multimodal Intention Disambiguation (MID) task in a Multiple-Choice Question format: Given an input image \\((I)\\) and an indirect expression \\((p_i)\\), the goal is to infer the speaker's hidden intent \\((T)\\) and select the most likely answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.643 + ], + "angle": 0, + "content": "duced a benchmark derived from a Cartoon Caption Contest, exploring humor identification and explanation. Baluja et al. (2024) [1] demonstrated that models benefit from multimodal cues in humor understanding. Memes also involve implicature, with multimodal datasets such as MemeCap [15] and MultiBully-Ex [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.485, + 0.734 + ], + "angle": 0, + "content": "However, the cues used in multimodal implicature understanding remain simple, primarily appearing in images with a single main object or person, overlooking the importance of interactions between multiple objects and people in real-world scenarios. These limitations underscore the need for more complex cues, as addressed in VAGUE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.434, + 0.767 + ], + "angle": 0, + "content": "3. Multimodal Intention Disambiguation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.837 + ], + "angle": 0, + "content": "In this section, we outline the structure and rationale behind the format of our primary task, which we term Multimodal Intention Disambiguation (MID). Then, we further specify the necessary components that form the basis of the task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.254, + 0.865 + ], + "angle": 0, + "content": "3.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Each MID problem comprises an input image \\(I\\), a direct text expression \\(p_d\\), and an indirect text expression \\(p_i\\). Here," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "the direct expression \\( p_d \\) clearly shows the underlying intention of the corresponding \\( p_i \\) and serves as an essential intermediate step of generating \\( p_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.229 + ], + "angle": 0, + "content": "To clarify our problem, we assume that all reasoning is confined to the depicted scene and that each expression is spoken by a human who intends for the listener to take a particular action based on the situation. The ultimate objective of the task is to interpret the hidden intention \\( T \\) effectively by leveraging the contextual cues within the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.909, + 0.442 + ], + "angle": 0, + "content": "To evaluate how well models capture such intentions, we adopt a multiple-choice (MCQ) format as the primary setup, as shown in Fig. 2. This decision reflects the fact that certain prompts can lead to multiple plausible outcomes, driven by hierarchical relations (e.g., pick up the chips - snack - food) or by the inherent uncertainty of what action best satisfies the speaker's goal (e.g., an indirect prompt complaining about darkness could be addressed by either turning on a light or opening curtains). Exploring all possible valid interpretations is labor-intensive and often infeasible. Consequently, each MID instance is presented as four distinct options, one correct and three intentionally designed to be incorrect for different reasons (see Sec. 4.2.3), challenging models in both linguistic and visual reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.906, + 0.517 + ], + "angle": 0, + "content": "Formally, let \\( C \\) be the set of all multiple-choice options \\( c_{n} \\). Given an image \\( I \\) and an indirect prompt \\( p_i \\), the task is to select the most valid interpretation of \\( p_i \\) from the predefined options, conditioned on the visual context in \\( I \\). We define this task as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.538, + 0.907, + 0.556 + ], + "angle": 0, + "content": "\\[\nT (I, p _ {i}) := \\operatorname {a r g m a x} _ {c _ {n} \\in C} \\Pr \\left(c _ {n} \\mid I, p _ {i}\\right). \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.567, + 0.795, + 0.583 + ], + "angle": 0, + "content": "3.2. Direct and Indirect Expressions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.68 + ], + "angle": 0, + "content": "By the design of our task, curating effective input prompts \\( p_d \\) and \\( p_i \\) is crucial for ensuring accurate interpretation. What makes a good prompt, though? In this section, we define and explain the criteria that both direct and indirect expressions must satisfy. For more details on good and bad examples for each criterion, please refer to our Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.822, + 0.707 + ], + "angle": 0, + "content": "3.2.1. Directness: Relevance and Solvability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.907, + 0.909 + ], + "angle": 0, + "content": "Relevance The direct prompt is an utterance from the speaker that explicitly conveys its intended meaning without ambiguity. However, it is equally important that this intention aligns with the visual context of the scene. For example, a direct prompt \\( p_d \\) such as \"Hey person1, I want you to stop the fireworks\" clearly expresses its intended action. However, if the corresponding image, as shown in Fig. 2, contains no elements related to fireworks, the prompt is misaligned with the scene. Thus, a direct prompt must not only reveal its intention but also maintain relevance to the image. In the context of our task, relevance is determined by whether a human can reasonably establish a connection between the prompt and the depicted scene." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1539" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "Solvability Relevance alone does not guarantee that a prompt is useful. As outlined in Sec. 3.1, the prompt \\( p_d \\) must explicitly request an action that the listener can reasonably perform. This introduces solvability, which requires that the prompt present a clear and actionable problem. A solvable prompt defines a specific issue that can be addressed independently, ensuring that the listener is not left with multiple competing actions to choose from." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.226, + 0.426, + 0.243 + ], + "angle": 0, + "content": "3.2.2. Indirectness: Consistency and Ambiguity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.483, + 0.429 + ], + "angle": 0, + "content": "Consistency Indirect prompts are designed to obscure their true intention, but they must still convey the same underlying intention as their direct counterpart—essentially requesting the same solution. Since indirect prompts are derived from direct prompts, consistency serves as a key criterion. We define a direct prompt \\( p_d \\) and an indirect prompt \\( p_i \\) as consistent if their interpretations could potentially align in intention. The term \"potentially\" is used because indirect prompts, by nature, may have multiple valid interpretations. For instance, in the earlier \"this is a disaster!\" example, it conveys distress but allows for multiple reasonable responses, such as cleaning the room or providing reassurance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Ambiguity If an indirect prompt is entirely consistent with its direct counterpart without introducing any additional complexity, it becomes indistinguishable from a direct prompt. Therefore, an indirect prompt should conceal its underlying intention, which we define as ambiguity. The key principle behind this criterion is that neither the specific action required nor the key entity involved should be explicitly or implicitly mentioned within the prompt. Once these two elements are concealed, further refinements—such as adjusting the tone to be more indirect, sarcastic, or humorous—can enhance the overall nuance and difficulty of interpretation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.632, + 0.402, + 0.648 + ], + "angle": 0, + "content": "4. VAGUE Benchmark Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "VAGUE is a novel benchmark that extends single-modal or simple multimodal ambiguity to more realistic domains and evaluates whether concurrent vision-language models can perform human-like reasoning with complex visual contexts. It comprises 1,677 images, with 1,144 sourced from the VCR [46] dataset and 533 from Ego4D [41], covering diverse contextual scenarios as well as real-world human interactions. On average, VAGUE contains seven objects and four people per image. Each image is paired with a direct expression \\( p_d \\), an indirect expression \\( p_i \\), and four multiple-choice answers, along with relevant meta-information. All textual components are generated using GPT-4o, then refined through extensive human rating, selection, and filtering, ensuring a carefully curated benchmark dataset for testing and advancing multimodal reasoning. We provide detailed benchmark statistics and a diversity analysis in Appendix B.2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.715, + 0.106 + ], + "angle": 0, + "content": "4.1. Visual Data Curation" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.114, + 0.628, + 0.129 + ], + "angle": 0, + "content": "4.1.1. Sampling" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.133, + 0.907, + 0.286 + ], + "angle": 0, + "content": "VCR [46] The VCR dataset consists of 110K movie scenes sourced from the Large Scale Movie Description Challenge [35] and YouTube clips. These images are curated based on an \"interestingness\" criterion [46], ensuring the presence of at least two people, which promotes interactive scenarios. To prevent redundancy in our dataset, we sample 10K images while carefully avoiding neighboring frames, as adjacent frames exhibit minimal variation. This selection process preserves the contextual diversity of the dataset while maintaining its focus on complex, multi-entity interactions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.298, + 0.909, + 0.48 + ], + "angle": 0, + "content": "Ego4D [41] While VCR provides a wide range of contextual diversity, it often includes artificially composed settings that may not fully capture real-world interactions. To address this, we integrate frames from the Ego4D dataset, which offers a more naturalistic depiction of human interactions. We specifically leverage the AV (Audio-Visual), which indicate conversational exchanges between individuals, to ensure the presence of people in the selected frames. Similar to VCR, we avoid neighboring frames to maintain diversity and filter out heavily blurry images to enhance data quality. This process results in 888 candidate images from 94 videos, which serve as the basis for further text processing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.489, + 0.686, + 0.505 + ], + "angle": 0, + "content": "4.1.2. Object Extraction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.509, + 0.909, + 0.66 + ], + "angle": 0, + "content": "To ensure the complexity of the visual information, we extract a list of physical objects present in each image using a tagging model, RAM [49]. This step allows us to easily identify scenes with sufficient visual detail. In the case of VCR, many scenes are relatively simple, often containing only a few objects. Therefore, we sort VCR images by the number of detected objects and retained the top 4,000 as candidates for text processing, ensuring that our benchmark primarily consists of rich visual cues in contextually diverse scenes. Please refer to Appendix B.3 for more details." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.67, + 0.677, + 0.684 + ], + "angle": 0, + "content": "4.1.3. Person Indicator" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In our task, we assume that the speaker is outside the scene, viewing the image and talking to a person in it. However, identifying the addressee is not always straightforward, as images mostly contain multiple individuals. While grounding the specific person referenced in the utterance could introduce additional complexity, it is not the primary focus of our evaluation. To clarify the listener, we assign an indicator tag to each person in the image as shown in Fig. 3. For VCR, we use their existing annotations [46], while for Ego4D, where bounding boxes are not fully available, we employ YOLOv11 [18] to detect and annotate humans. While this provides a straightforward method for grounding the target person, it requires models to perform basic Optical Character Recognition (OCR). Therefore, we conduct exper" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1540" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.093, + 0.9, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.306, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Figure 3. Overview of the data generation process. Based on human-defined criteria and instructions, GPT-4o [31] generates initial data, which are then rated and filtered by humans to ensure quality. Since generating high-quality indirect expressions from raw images is challenging, the process follows these steps: generating a direct expression and intention from the image, creating an indirect expression using information from the previous step, and producing answer candidates based on all the information gathered so far. In the answer candidates, FS stands for Fake Scene Understanding, SU stands for Superficial Understanding, and NE stands for Nonexistent Entity. FS evaluates global hallucination, where the model misinterprets visual context from an entirely different image (e.g., an outdoor camping scene with a candle), while NE addresses local hallucination, where only a specific object is replaced with a fabricated one (e.g., curtains)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.462 + ], + "angle": 0, + "content": "iments to evaluate the OCR capabilities of the models used in our study. See Appendix B.4 for details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.472, + 0.383, + 0.487 + ], + "angle": 0, + "content": "4.2. Multimodal Expression Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.601 + ], + "angle": 0, + "content": "As shown in Fig. 3, VAGUE's textual expressions are generated first by the model and undergo an extensive process of human rating and filtering. We use GPT-4o [31] for all text processing. The instructions used for generating direct \\((p_d)\\) and indirect \\((p_i)\\) expressions are provided in Appendix B.5, while the instructions for generating answer candidates for multiple-choice questions are detailed in Appendix B.7." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.609, + 0.267, + 0.624 + ], + "angle": 0, + "content": "4.2.1. Direct Expressions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.809 + ], + "angle": 0, + "content": "The direct expression \\( p_d \\) serves as a crucial foundation for crafting the indirect one \\( p_i \\), as both share the same underlying intention. To ensure that \\( p_d \\) adhere to the principles of relevance and solvability discussed in Sec. 3.2.1, we generate \\( p_d \\) conditioned on the input image \\( I \\) and a task prompt that explicitly defines these criteria. During the generation process, we also instruct the model to output a solution triplet in the format: (subject, action, object). Since direct expressions explicitly state their intentions, extracting each component of the solution triplet is straightforward. To maintain consistency with the visual context, the \"object\" in the triplet is restricted to physical objects we extracted in Sec. 4.1.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "After generating \\( p_d \\) for all candidate images, human raters evaluate each prompt based on relevance and solvability, assigning scores from 1 to 5. Only those prompts that receive a rating of 4 or 5 are retained for further use. The detailed rating criteria for human verification and an example of the rating process are provided in Fig. J13." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.431, + 0.703, + 0.445 + ], + "angle": 0, + "content": "4.2.2. Indirect Expressions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.496 + ], + "angle": 0, + "content": "To ensure that the indirect expressions \\( p_i \\) maintain both ambiguity and fluency while aligning with the true intent \\( T \\), we adopt a two-stage process: proposal and selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.693 + ], + "angle": 0, + "content": "In the initial step, the model is prompted to generate three distinct candidate options. Each candidate follows the criteria outlined in Section 3.2.2, but with explicit instructions to incorporate different linguistic strategies: sarcasm, humor, and meme/idiomatic expressions, respectively. This approach ensures diversity in the generated responses. In the second step, human annotators evaluate the three candidates and select the one that best aligns with the intended indirectness. The selected prompt is then rated on a scale from 1 to 5 based on more specific criteria. Only those prompts that receive a score of 3 or higher are retained for use in the dataset. The detailed rating criteria for human verification and an example of rating process are provided in Fig. J14" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.703, + 0.724, + 0.716 + ], + "angle": 0, + "content": "4.2.3. Counterfactual Choices" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.722, + 0.907, + 0.782 + ], + "angle": 0, + "content": "Generating high-quality counterfactual choices is crucial. To enable detailed analysis of model weaknesses in multimodal intent disambiguation, we design interpretable counterfactual choices that provide more plausible alternatives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Fake Scene Understanding The first counterfactual choice is an interpretation that could arise when the model largely misinterprets the image. This process is conducted in two steps. In the first step, a fake caption is generated by assuming an imaginary scene that can be aligned with the indirect expression but is inconsistent with the true intent. The caption of fake scene is then combined with the speaker's" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1541" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.47, + 0.108 + ], + "angle": 0, + "content": "indirect statement to derive the most likely interpretation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.118, + 0.486, + 0.27 + ], + "angle": 0, + "content": "Superficial Understanding The subsequent choice corresponds to an interpretation generated when the model fails to deeply reason about the implicit intent of the sentence and instead relies on surface-level meaning. We enforce the model to focus only on the literal wording, without considering any implied or deeper meaning of the indirect sentence. These answer choices are generated alongside the indirect expression \\( p_i \\). During the indirect selection phase, the corresponding superficially understood choice is selected together, maintaining coherence between them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.485, + 0.508 + ], + "angle": 0, + "content": "Nonexistent Entity The last choice arises when the model interprets the text correctly but fails to adequately consider the details of the image, resulting in a plausible yet incorrect choice. This resembles the correct answer in structure, but replaces the key object in the solution with one that does not exist in the image. To prevent the task from becoming too easy by generating highly irrelevant objects, we constrain the substituted object to one that, while absent from the image, is highly expected to be present in the scene and could replace the original entity. To identify such entities, the model is provided with the image as input to choose objects that align with the scene's context. This method ensures that the counterfactual choice leverages the expected coherence between the scene and its potential entities while rigorously testing the model's attention to visual details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.523, + 0.223, + 0.54 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Models We use the following models in our experiments. The detailed descriptions of each model are in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.58, + 0.32, + 0.595 + ], + "angle": 0, + "content": "- Phi3.5-Vision-Instruct (4B) [29]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.596, + 0.295, + 0.61 + ], + "angle": 0, + "content": "- LLaVA Onevision (7B) [23]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.611, + 0.352, + 0.625 + ], + "angle": 0, + "content": "Qwen2.5-VL-Instruct (7B, 72B) [43]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.626, + 0.329, + 0.64 + ], + "angle": 0, + "content": "- InternVL-2.5-MPO (8B, 26B) [6]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.641, + 0.227, + 0.655 + ], + "angle": 0, + "content": "- Idefics2 (8B) [22]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.656, + 0.326, + 0.67 + ], + "angle": 0, + "content": "LLaVA NeXT Vicuna (13B) [24]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.671, + 0.222, + 0.685 + ], + "angle": 0, + "content": "- Ovis2 (16B) [28]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.686, + 0.192, + 0.7 + ], + "angle": 0, + "content": "GPT-4o [31]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.701, + 0.241, + 0.715 + ], + "angle": 0, + "content": "Gemini 1.5 Pro [12]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.716, + 0.257, + 0.73 + ], + "angle": 0, + "content": "- InternVL-3 (38B) [42]" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.58, + 0.352, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.388, + 0.757 + ], + "angle": 0, + "content": "5.1. MLLMs Benefit from Visual Cues" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.485, + 0.885 + ], + "angle": 0, + "content": "Our first objective is to assess how effectively MLLMs leverage visual cues to resolve ambiguity in utterances. To this end, we systematically control the level of detail in the visual cues provided to the models and measure their accuracy in inferring the speaker's true intent. Performance is evaluated in both multiple-choice and free-form settings. For clarity, we primarily report multiple-choice accuracy, deferring free-form results to Appendix I." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.372, + 0.9 + ], + "angle": 0, + "content": "We consider three levels of visual cues:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.908, + 0.151 + ], + "angle": 0, + "content": "- Language Models (LMs) receive no visual input, requiring models to rely on superficial textual priors such as common-sense knowledge of sarcasm or humor to determine intent." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.152, + 0.908, + 0.243 + ], + "angle": 0, + "content": "- Socratic Models (SMs) [47] use text-only LMs but incorporate short image captions (up to two or three sentences) as additional input. This short generic caption may lack sufficient detail, which may be insufficient for accurately inferring intent. Each SM model generated its own image captions and used them in subsequent processing." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.243, + 0.907, + 0.273 + ], + "angle": 0, + "content": "- Visual Language Models (VLMs) receive the raw image input, enabling a more direct interpretation of visual cues." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.908, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.909, + 0.466 + ], + "angle": 0, + "content": "Results The results in Tab. 1 indicate that MLLMs can leverage visual cues, albeit to a limited extent, since SMs and VLMs consistently outperform LMs across all evaluated models. Additionally, more detailed visual input generally improves performance, with VLMs surpassing SMs in most cases, except in the case of proprietary models. This exception is further analyzed in Sec. 5.2. Both the VCR and Ego4D subsets exhibit similar performance trends, demonstrating the generalizability of our findings across both staged and real-world scenarios. Finally, the consistently low performance of LMs further reinforces the validity of our dataset as a multimodal benchmark." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.755, + 0.492 + ], + "angle": 0, + "content": "5.2. Analysis on Failure Modes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.909, + 0.697 + ], + "angle": 0, + "content": "Here, we examine the ways in which models fail to infer the true intent and how these failure patterns vary with the level of visual cues provided. As shown in Fig. 3, our multiple-choice questions include three distinct types of incorrect answer candidates. We assess the model's raw visual understanding using the Fake Scene Understanding (FS) and Nonexistent Entity (NE) candidates, which test whether the model can correctly interpret the scene without being misled by fabricated or nonexistent elements. Conversely, the Superficial Understanding (SU) candidate evaluates the model's reasoning ability, testing whether it can go beyond surface-level perception to infer intent. We provide the full table of failure modes in the MCQ setting in Appendix I" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Results Figure 4 illustrates how frequently each model selects different types of incorrect answers instead of the correct intent. Among the error types, Superficial Understanding (SU) is the most common. This indicates that while models generally succeed in recognizing basic visual details, they often fail to reason deeply about those visual cues to accurately infer the underlying intent of the speaker. However, proprietary models exhibit fewer SU-related errors, indicating stronger reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Moreover, stronger visual cues improve accuracy across all models and failure types. This improvement highlights the crucial role of visual conditioning in reducing both vision-based errors (FS and NE) and reasoning-related failures (SU)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1542" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.338 + ], + "angle": 0, + "content": "
ModelVAGUE-VCRVAGUE-Ego4D
LM (L)SM (L+V)VLM (L+V)LM (L)SM (L+V)VLM (L+V)
Phi3.5-Vision-Instruct (4B)26.635.3 (↑ 8.7)46.0 (↑ 19.4)22.531.1 (↑ 8.6)42.4 (↑ 19.9)
LLaVA-Onevision (7B)13.129.4 (↑ 16.3)43.1 (↑ 30.0)11.329.5 (↑ 18.2)43.2 (↑ 31.9)
Qwen2.5-VL-Instruct (7B)11.125.6 (↑ 14.5)46.8 (↑ 35.7)9.828.0 (↑ 18.2)48.4 (↑ 38.6)
InternVL-2.5-MPO (8B)23.048.4 (↑ 25.4)63.9 (↑ 40.9)24.254.0 (↑ 29.8)66.8 (↑ 42.6)
Idefics2 (8B)13.921.1 (↑ 7.2)58.7 (↑ 44.8)14.818.2 (↑ 3.4)58.3 (↑ 43.5)
LLaVA-NeXT-vicuna (13B)24.237.2 (↑ 13)46.4 (↑ 22.2)20.334.1 (↑ 13.8)52.5 (↑ 32.2)
Ovis2 (16B)21.923.8 (↑ 1.9)24.5 (↑ 3.6)20.525.3 (↑ 4.8)25.7 (↑ 5.2)
InternVL-2.5-MPO (26B)21.248.5 (↑ 27.3)63.7 (↑ 42.5)21.855.2 (↑ 33.4)68.7 (↑ 46.9)
InternVL-3 (38B)24.847.2 (↑ 22.4)63.6 (↑ 38.8)18.047.5 (↑ 29.5)59.8 (↑ 41.8)
Qwen2.5-VL-Instruct (72B)29.655.6 (↑ 26.0)74.2 (↑ 44.6)26.859.3 (↑ 32.5)69.8 (↑ 43.0)
GPT-4o46.469.5 (↑ 23.1)65.1 (↑ 18.7)48.267.5 (↑ 19.3)63.6 (↑ 15.3)
Gemini-1.5-Pro43.262.4 (↑ 19.2)60.6 (↑ 17.4)40.360.6 (↑ 20.3)60.6 (↑ 20.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.348, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Table 1. Experiments on the Multimodal Intention Disambiguation (MID) task with varying levels of visual cues. We report the accuracy \\((\\%)\\) of the Multiple-Choice Question. \\(\\uparrow\\) indicates the performance gain from visual cues, i.e. increment compared to the LM setting.(L) denotes the use of language input only, while \\((\\mathrm{L} + \\mathrm{V})\\) indicates the incorporation of visual cues. The noticeable increase in accuracy across LM, SM, and VLM demonstrates that the introduction of detailed visual cues is beneficial for the task." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.418, + 0.909, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.592, + 0.908, + 0.649 + ], + "angle": 0, + "content": "Figure 4. We present a bar plot to analyze the distribution of incorrect answer choices selected by each model. Each number represents how frequently a given choice was selected from the 1,677 items in the dataset. The counterfactual choice categories are FS (Fake Scene Understanding), SU (Superficial Understanding), and NE (Nonexistent Entity). We use distinct colors to represent LM (Language Models), SM (Socratic Models), and VLM (Visual-Language Models)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Notably, proprietary models perform better with captioned inputs (SM) than with raw images (VLM). A closer examination of Fig. 4 reveals that this discrepancy arises from vision-based failures (FS and NE) rather than reasoning-centric failure (SU). This indicates that their captioning ability potentially allows them to obtain more sophisticated information while reducing hallucination in the images. The supporting experiments and explanations are provided in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.32, + 0.82 + ], + "angle": 0, + "content": "5.3. Comparison with human" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To validate our benchmark and establish an upper bound for performance, we assess human accuracy, highlighting the gap between existing models and human capability. This evaluation follows the multiple-choice setup within the VLM setting, using a subset of 400 samples. As shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.901 + ], + "angle": 0, + "content": "in Tab. 2, human performance reaches \\(94\\%\\), demonstrating near-perfect accuracy. Although proprietary models and some large-sized models show a decent performance, a notable performance gap (\\(\\sim 20\\%\\)) still exists compared to human evaluators. This result underscores the significant gap between AI models' multimodal reasoning capabilities and human-level understanding when inferring hidden intent, suggesting that visual perception alone, even when accurate, is insufficient without deeper cognitive integration. This performance gap is due to the models' tendency to rely on surface-level text rather than understanding deeper visual-textual implications. Thus, advancing multimodal reasoning likely requires models to integrate higher-order cognitive processes, like commonsense reasoning and pragmatic understanding, into visual interpretation tasks. Refer" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1543" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.483, + 0.279 + ], + "angle": 0, + "content": "
ModelAcc (%)FSSUNECorrect
Ovis2 (16B)23.01191711892
LLaVA-Onevision (7B)41.04318310164
Phi3.5-Vision-Instruct (4B)44.36013231177
Qwen2.5-VL-Instruct (7B)47.04715213188
LLaVA-NeXT-icuna (13B)48.04814317192
Idefics2 (8b)57.02812024228
Gemini-1.5-Pro60.3607326241
InternVL-3 (38B)61.5389422246
InternVL-2.5-MPO (8B)61.8429516247
GPT-4o62.3616327249
InternVL-2.5-MPO (26B)63.03610111252
Qwen2.5-VL-Instruct (72B)72.3396012289
Human94.01248374
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.29, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Table 2. Performance across models and humans on a sampled set of 400 questions. The results show that humans outperform models by a margin of over \\(20\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.347, + 0.48, + 0.515 + ], + "angle": 0, + "content": "
ModelTypeAcc (%)Incorrect count
FSSUNE
GPT-4oSM68.9165240117
SM+CoT69.5 (↑ 0.6)165241105
VLM64.6226226141
VLM+CoT66.4 (↑ 1.8)16215685
Gemini-1.5-ProSM61.817637483
SM+CoT61.0 (↓ 0.8)19036794
VLM60.6249264145
VLM+CoT64.4 (↑ 3.8)213267117
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Table 3. Result of Chain-of-Thought (CoT) experiments on proprietary models, in both SM and VLM settings. \\(\\uparrow\\) and \\(\\downarrow\\) indicate an increase and decrease in accuracy when zero-shot CoT is applied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.629 + ], + "angle": 0, + "content": "to Appendix E for details on the selected subset and human evaluation setup." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.639, + 0.371, + 0.656 + ], + "angle": 0, + "content": "5.4. Chain-of-Thought Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Given the strong reasoning demands of multimodal intent deduction, we further explore the effectiveness of Chain-of-Thought (CoT) prompting [44] in enhancing the reasoning capabilities of MLLMs. The CoT prompt templates, provided in Fig. J21 and Fig. J22, are designed to explicitly ground the reasoning process, reducing hallucinations. While our main results focus on proprietary models, owing to their superior suitability for zero-shot CoT reasoning, we also present experiments and analyses for open-source models ranging from 4B to 72B in Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Results As shown in Tab. 3, CoT prompting improves performance for raw image inputs (VLM), while showing no clear trend and remaining at a similar level for image caption inputs (SM). One possible explanation for this discrepancy is that CoT primarily enhances reasoning by improving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.199 + ], + "angle": 0, + "content": "grounding and reducing hallucinations. Since image captions inherently contain fewer hallucinations, SMs see some benefit from CoT prompting albeit at the cost of reduced detail. Additionally, the performance improvements observed with CoT prompting are consistent across different types of false answer candidates, suggesting a generalizable effect in enhancing reasoning quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.211, + 0.634, + 0.227 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.909, + 0.525 + ], + "angle": 0, + "content": "We present VAGUE (Visual Contexts CIArify ambiGUous Expressions), a new benchmark aimed at assessing models' ability to interpret nuanced communication in complex multimodal scenarios. Our results show that models benefit from visual information when inferring the underlying intention of indirect expressions, as evidenced by their improved performance with increasing levels of visual cues. However, a significant disparity persists between machine capabilities and humans. To gain deeper insights into model inaccuracies, we design multiple-choice questions that explicitly address failure points, enabling a systematic and quantitative evaluation of the reasons behind performance. The primary challenge identified is the tendency of multimodal models to inadequately integrate visual cues, relying instead on the literal interpretation of textual information. This shortcoming highlights the need for further research, and we anticipate that VAGUE will open up promising avenues for developing systems capable of deeper multimodal reasoning to enhance AI's ability to engage in human-like interactions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.537, + 0.637, + 0.553 + ], + "angle": 0, + "content": "7. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.909, + 0.88 + ], + "angle": 0, + "content": "First, we acknowledge that there are cultural and linguistic limitations. We incorporate sarcastic, humorous, and idiomatic implicatures in generating indirect expressions. However, since the initial data drafts are created by a model (GPT-4o [31]), they may reflect cultural biases present in its training data. To prevent any potential ethical issues, all human annotators are instructed to remove any content considered problematic or discriminatory during the rating and filtering process. Also, all textual expressions in VAGUE are limited to English. Therefore, we encourage future exploration of indirect expressions across diverse languages and cultures. The second limitation pertains to the dependency of certain meta-information on the quality of the parent dataset and the performance of the models utilized during the dataset generation. We observe that bounding box annotations from YOLOv11 [18] are occasionally duplicated, leading to a reported number of people higher than actually present. Likewise, the tagging model RAM [49] sometimes misidentified objects. Therefore, we remove any corrupted instances that could undermine the integrity of the task during the human rating and filtering process." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1544" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Ashwin Baluja. Text is not all you need: Multimodal prompting helps llms understand humor. arXiv preprint arXiv:2412.05315, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[3] Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi, and Andreas Bulling. Explicit modelling of theory of mind for belief prediction in nonverbal social interactions. In Proc. 27th European Conference on Artificial Intelligence (ECAI), pages 1-8, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[4] Tuhin Chakrabarty, Arkadiy Saakyan, Debanjan Ghosh, and Smaranda Muresan. FLUTE: Figurative language understanding through textual explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7139-7159, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.443 + ], + "angle": 0, + "content": "[5] Zhawnen Chen, Tianchun Wang, Yizhou Wang, Michal Kosinski, Xiang Zhang, Yun Fu, and Sheng Li. Through the theory of mind's eye: Reading minds with multimodal video large language models. arXiv preprint arXiv:2406.13763, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[7] Christine Cuskley, Rebecca Woods, and Molly Flaherty. The limitations of large language models for understanding human language and cognition. Open Mind, 8:1058-1083, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[8] Hang Du, Guoshun Nan, Sicheng Zhang, Binzhu Xie, Junrui Xu, Hehe Fan, Qimei Cui, Xiaofeng Tao, and Xudong Jiang. Docmsu: A comprehensive benchmark for document-level multimodal sarcasm understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17933-17941, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[9] Jiafei Duan, Samson Yu, Nicholas Tan, Li Yi, and Cheston Tan. Boss: A benchmark for human belief prediction in object-context scenarios. arXiv preprint arXiv:2206.10665, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[10] Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. CVPR 2024, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[11] Kanishk Gandhi, Jan-Philipp Fraenkel, Tobias Gerstenberg, and Noah D. Goodman. Understanding social reasoning in language models with language models. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track, 2023. Dataset and Benchmarks Track. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[13] Md Kamrul Hasan, Wasifur Rahman, AmirAli Bagher Zadeh, Jianyuan Zhong, Md Iftekhar Tanveer, Louis-Philippe Morency, and Mohammed (Ehsan) Hoque. UR-FUNNY: A multimodal language dataset for understanding humor. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2046–2056, Hong Kong, China, 2019. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.908, + 0.33 + ], + "angle": 0, + "content": "[14] Jack Hessel, Ana Marasovic, Jena D. Hwang, Lillian Lee, Jeff Da, Rowan Zellers, Robert Mankoff, and Yejin Choi. Do androids laugh at electric sheep? humor \"understanding\" benchmarks from the new yorker caption contest. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 688-714, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.333, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[15] EunJeong Hwang and Vered Shwartz. Memecap: A dataset for captioning and interpreting memes. arXiv preprint arXiv:2305.13703, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.376, + 0.907, + 0.444 + ], + "angle": 0, + "content": "[16] Prince Jha, Krishanu Maity, Raghav Jain, Apoorv Verma, Sriparna Saha, and Pushpak Bhattacharyya. Meme-ingful analysis: Enhanced understanding of cyberbullying in memes through multimodal explanations. arXiv preprint arXiv:2401.09899, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.908, + 0.558 + ], + "angle": 0, + "content": "[17] Chuanyang Jin, Yutong Wu, Jing Cao, Jiannan Xiang, Yen Ling Kuo, Zhiting Hu, Tomer Ullman, Antonio Torralba, Joshua Tenenbaum, and Tianmin Shu. MMToM-QA: Multimodal theory of mind question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16077-16102, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[18] Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.908, + 0.687 + ], + "angle": 0, + "content": "[19] Shivani Kumar, Atharva Kulkarni, Md Shad Akhtar, and Tan-moy Chakraborty. When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5956-5968, Dublin, Ireland, 2022. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.908, + 0.798 + ], + "angle": 0, + "content": "[20] Shivani Kumar, Ishani Mondai, Md Shad Akhtar, and Tanmoy Chakraborty. Explaining (sarcastic) utterances to enhance affect understanding in multimodal dialogues. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.908, + 0.871 + ], + "angle": 0, + "content": "[21] Yash Kumar Lal and Mohaddeseh Bastan. SBU figures it out: Models explain figurative language. In Proceedings of the 3rd Workshop on Figurative Language Processing (FLP), pages 143-149, Abu Dhabi, United Arab Emirates (Hybrid), 2022. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[22] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language mod" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1545" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "els? Advances in Neural Information Processing Systems, 37: 87874-87907, 2025. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[23] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[24] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.482, + 0.279 + ], + "angle": 0, + "content": "[25] Kevin Y. Li, Sachin Goyal, Joao D. Semedo, and J. Zico Kolter. Inference optimal vlms need fewer visual tokens and more parameters, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[27] Annie Louis, Dan Roth, and Filip Radlinski. \"I'd rather just go to bed\": Understanding indirect answers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7411-7425, Online, 2020. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.425, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[28] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[29] Microsoft. Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[30] Maryam Sadat Mirzaei, Kourosh Meshgi, and Satoshi Sekine What is the real intention behind this question? dataset collection and intention classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13606-13622, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[31] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[32] Kun Ouyang, Liqiang Jing, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Sentiment-enhanced graph-based sarcasm explanation in dialogue. arXiv:2402.11414, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[34] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[35] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Chris Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. International Journal of Computer Vision, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Melanie Sclar, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia Tsvetkov. Minding language models' (lack of) theory of mind: A plug-and-play multi-character belief" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 13960-13980, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[37] Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin, Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. Muma-tom: Multi-modal multi-agent theory of mind, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.907, + 0.286 + ], + "angle": 0, + "content": "[38] Settaluri Sravanthi, Meet Doshi, Pavan Tankala, Rudra Murthy, Raj Dabre, and Pushpak Bhattacharyya. PUB: A pragmatics understanding benchmark for assessing LLMs' pragmatics capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12075-12097, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[39] Chuandong Su, Fumiyo Fukumoto, Xiaoxi Huang, Jiyi Li, Rongbo Wang, and Zhiqun Chen. DeepMet: A reading comprehension paradigm for token-level metaphor detection. In Proceedings of the Second Workshop on Figurative Language Processing, pages 30-39, Online, 2020. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[40] Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase. DIRECT: Direct and indirect responses in conversational text corpus. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1980-1989, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.907, + 0.509 + ], + "angle": 0, + "content": "[41] Ego4d Team. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.907, + 0.551 + ], + "angle": 0, + "content": "[42] OpenGVLab Team. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.864, + 0.566 + ], + "angle": 0, + "content": "[43] Qwen Team. Qwen2.5-v1 technical report, 2025. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.908, + 0.635 + ], + "angle": 0, + "content": "[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.908, + 0.704 + ], + "angle": 0, + "content": "[45] Alex Wilf, Leena Mathur, Sheryl Mathew, Claire Ko, Youssof Kebe, Paul Pu Liang, and Louis-Philippe Morency. Social-iq 2.0 challenge: Benchmarking multimodal social understanding. https://github.com/abwilf/Social-IQ-2.0-Challenge, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.908, + 0.761 + ], + "angle": 0, + "content": "[46] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6713-6724, 2018. 2, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[47] Andy Zeng, Maria Attarian, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[48] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1546" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.487, + 0.177 + ], + "angle": 0, + "content": "[49] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1724-1732, 2024. 4, 8, 1" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1547" + } + ] +] \ No newline at end of file diff --git a/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_origin.pdf b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2ec2eb18338b90e9a49306afbed2205281362926 --- /dev/null +++ b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/321d7adc-1768-447e-aad1-8d77c13702f1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:048b39b5ad94deb989147ab1faf39a05a487a5307433e61be817438ea8058139 +size 1392500 diff --git a/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/full.md b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9934df65accb811e54cd147b1a388457a4acf6b1 --- /dev/null +++ b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/full.md @@ -0,0 +1,312 @@ +# VAGUE: Visual Contexts Clarify Ambiguous Expressions + +Heejeong $\mathrm{Nam}^{*1}$ , Jinwoo Ahn $^{*2}$ , Keummin $\mathrm{Ka}^3$ , Jiwan Chung $^3$ , and Youngjae $\mathrm{Yu}^{\dagger 3}$ + +Brown University + +$^{2}$ UC Berkeley + +3Yonsei University + +# Abstract + +Human communication often relies on visual cues to resolve ambiguity. While humans can intuitively integrate these cues, AI systems often find it challenging to engage in sophisticated multimodal reasoning. We introduce VAGUE, a benchmark evaluating multimodal AI systems' ability to integrate visual context for intent disambiguation. VAGUE consists of 1.6K ambiguous textual expressions, each paired with an image and multiple-choice interpretations, where the correct answer is only apparent with visual context. The dataset spans both staged, complex (Visual Commonsense Reasoning) and natural, personal (Ego4D) scenes, ensuring diversity. Our experiments reveal that existing multimodal AI models struggle to infer the speaker's true intent. While performance consistently improves from the introduction of more visual cues, the overall accuracy remains far below human performance, highlighting a critical gap in multimodal reasoning. Analysis of failure cases demonstrates that current models fail to distinguish true intent from superficial correlations in the visual scene, indicating that they perceive images but do not effectively reason with them. We release our code and data at https://hazel-heejeong-nam.github.io/vague/. + +# 1. Introduction + +Human communication is inherently contextual; for example, exclaiming "Hey, this is a disaster!" upon seeing a cluttered room conveys frustration or exaggeration rather than referring to an actual catastrophe. Without surrounding cues, textual dialogues can be ambiguous, making it difficult for models to accurately capture intent and nuance. + +We consider the case of visual contextual cues. Consider Fig. 1, which depicts a speaker making a remark in a certain situation. Without specifying contexts introduced + +![](images/e2a8dbca01ca7f9cf2de59e57ff233135cba5e760ade85d2b736c88db6c48b0f.jpg) + +Hey person1, looks like you're planning a party for the neighborhood squirrels. + +![](images/bb8ba85a6b859c036abd1d76f868a67d75e2753b51dd3ebe0919d8588c4362f9.jpg) + +![](images/9b146a4622b5effe44329b07299bd76f8a88078e788e652b22affa5b4f4e0d2a.jpg) +Visual Context 1 + +![](images/509ea48f33bcd7e74ab1e7eabd010e041181ef09744a26e395c2c99caa9b002f.jpg) + +![](images/1bdd4bc388a4d7802e8453cf79b21f4526a541f221b7420bec508b25359cb457.jpg) +Visual Context 2 + +![](images/87898e9bddf2979562ffd934554fc2f25bb756e79aba17f08f153bf2b21ff6c4.jpg) +The speaker wants person1 to... + +![](images/f39f695adddd5796b67d5a45b0ac7a1636c83143ef23955c8f89b99045f1c29a.jpg) +close the cereal box +hand out the envelop to the squirrels +Figure 1. A motivating example demonstrating the importance of visual context in understanding intention. Without a predetermined context, a single expression can convey multiple different intentions. The textual expression and Context 1 are from our dataset, while Context 2 is generated using DALL-E 3 [2] to help understanding. + +from visual cues, the speaker's intention can vary, thus remaining ambiguous. This implies that the visual contexts play important roles in communication, raising the question: can AI systems integrate visual cues with ambiguous dialogue to infer the speaker's intent? + +We introduce Visual Contexts ClArify ambiGUous Expressions (VAGUE), a benchmark consisting of 1.6K ambiguous textual expressions, each paired with a single image. VAGUE aims to model diverse and natural human-to-human interactions by setting each image as the speaker's viewpoint, where the speaker implicitly requests a certain action from a person within their field of view. We define the problem addressed through this setup as Multimodal Intention Disambiguation (MID), which involves reasoning about the most plausible request conditioned on visual context. Each sample in VAGUE is annotated with four multiple-choice candi + +dates, ensuring clarity and preventing multiple valid answers caused by paraphrasing or hierarchical inclusion of meaning. The dataset is meticulously curated to ensure visual dependency; the ground-truth candidate is only preferable when considering the visual context. VAGUE includes visual scenes from both artificial sources (Visual Commonsense Reasoning [46]) and real-world scenarios (Ego4D [41]), capturing a broad spectrum of scene complexity and naturalness. The textual expressions in VAGUE are initially generated by GPT-4o [31] following instructions, then reviewed through extensive human rating and filtering to ensure naturalness and alignment with the corresponding images. + +Experiments on VAGUE demonstrate that existing multimodal AI models struggle to infer a speaker's true intent in a multimodal setting. First, although models can leverage visual context—as seen by a performance progression from text-only Language Models (LMs) to pipelined Socratic Models (SMs) [47] and ultimately to end-to-end Visual Language Models (VLMs)—their overall accuracy remains significantly lower than that of humans, indicating a failure to capture the true intent. A closer analysis of failure cases reveals that the primary source of error is the models' inability to distinguish the true intent from a superficial understanding of the visual context. In other words, even though these multimodal systems can perceive the image content, they cannot effectively use this information to reason about the speaker's true intent. + +In conclusion, we introduce a benchmark that exposes the limitations of current models in integrating visual cues with intent comprehension and identifies their primary failure mode. We anticipate that VAGUE will serve as a testing ground for the development of future multimodal conversational or embodied agents—systems that combine robust visual perception with nuanced conversational reasoning to effectively respond to user requests in complex scenes. + +Our contributions are threefold: + +- VAGUE: a novel benchmark for evaluating multimodal intention disambiguation. Validated through extensive human filtering, VAGUE is designed for robust quantitative assessment by ensuring both the ambiguity of queries and the visual (in)dependency to answer candidates. +- Carefully curated 1,677 scene images sourced from VCR [46] and Ego4D [41], capturing a wide range of scene complexity, diversity, and naturalness to ensure VAGUE's generalizability across various contexts. +- Experimental results highlighting a critical challenge in multimodal intention disambiguation: while existing models can perceive visual cues, they fail to effectively integrate this information into reasoning to deduce the speaker's true intent. + +# 2. Related Work + +# 2.1. Multimodal Theory of Mind + +Theory of Mind (ToM) refers to the ability to infer and reason about the intentions of others based on available information [34], where recent language models still struggle with relevant tasks [7] highlighting the need for dedicated research in this area. Initially, various methods and benchmarks have been proposed in unimodal settings, relying on text-based approaches [11, 36]. However, these methods often fail to capture the richness of real-world interactions, which often require integrating both linguistic and visual cues. + +Moving beyond text-only contexts, recent work has incorporated visual information. MMToM [17] introduces a benchmark where models must process both visual and textual cues to solve question-answering tasks related to ToM. The BOSS dataset [9] is a multimodal dataset collected in situations where nonverbal communication is required. It is used to evaluate whether human beliefs can be inferred based on nonverbal cues during social interactions. Similarly, Chen et al. (2024) [5] propose a Video ToM model that leverages key video frames and transcripts, demonstrating improved reasoning on the Social-IQ 2.0 dataset [45]. MuMA-ToM [37] further extends this direction by assessing ToM reasoning in multi-agent interactions, evaluating a model's ability to infer human beliefs and goals based on video and text inputs. MToMnet [3] introduces a ToM-based neural network that integrates contextual cues, such as scene videos and object locations, with person-specific cues, to predict human beliefs in specific scenarios. + +However, progress in multimodal ToM remains constrained not only by the scarcity of high-quality datasets [5] but also by the lack of explicit consideration for the ambiguity and indirectness inherent in human communication. + +# 2.2. Multimodal Implicature Understanding + +Implicature and the ambiguity that arises from it naturally emerge in everyday human conversation, requiring pragmatic understanding [38]. Early research on implicature understanding has primarily been conducted in text-only settings [27, 30, 40], with some studies specifically focusing on figurative language and metaphor [4, 21, 39]. However, since the ambiguity of standalone text is inherently limited, recent studies have extended to multiple modalities. One example is multimodal sarcasm understanding (MSU). WITS [19] and MOSES [20] are benchmarks for sarcasm explanation, both providing the speaker's emotion and voice tone as cues. DocMSU [8] is a document-level benchmark for sarcasm localization and detection. To improve MSU, EDGE [32], a graph-based approach, achieved strong performance. UR-FUNNY [13] is a benchmark for multimodal humor comprehension, incorporating facial expressions and voice tones as in MSU [19, 20]. Hessel et al. (2023) [14] intro + +Q. Select the option that best explains the underlying intention of the utterance based on the given image. + +![](images/5e3b0a1b1a27751d011a6dcbe11cb25d073beac26e2f82c50585b0bb63611535.jpg) +Figure 2. Description of the Multimodal Intention Disambiguation (MID) task in a Multiple-Choice Question format: Given an input image $(I)$ and an indirect expression $(p_i)$ , the goal is to infer the speaker's hidden intent $(T)$ and select the most likely answer. + +Hey person1, spot the difference, this parking's a bit too special isn't it? + +![](images/d314ec8014afa1b1655e660cbf27f3b14d3fc17bd409690d07d9a53e5c01bf22.jpg) + +a) The speaker wants person1 to admire the unusually decorated motorcycle in the parking lot. +b) The speaker wants person1 to enjoy playing a puzzle game and spot the difference. +c) The speaker wants person1 to move the sedan because it's in a handicapped parking spot. +d) The speaker wants person1 to move the sedan because it's parked in front of a fire hydrant. + +duced a benchmark derived from a Cartoon Caption Contest, exploring humor identification and explanation. Baluja et al. (2024) [1] demonstrated that models benefit from multimodal cues in humor understanding. Memes also involve implicature, with multimodal datasets such as MemeCap [15] and MultiBully-Ex [16]. + +However, the cues used in multimodal implicature understanding remain simple, primarily appearing in images with a single main object or person, overlooking the importance of interactions between multiple objects and people in real-world scenarios. These limitations underscore the need for more complex cues, as addressed in VAGUE. + +# 3. Multimodal Intention Disambiguation + +In this section, we outline the structure and rationale behind the format of our primary task, which we term Multimodal Intention Disambiguation (MID). Then, we further specify the necessary components that form the basis of the task. + +# 3.1. Problem Setting + +Each MID problem comprises an input image $I$ , a direct text expression $p_d$ , and an indirect text expression $p_i$ . Here, + +the direct expression $p_d$ clearly shows the underlying intention of the corresponding $p_i$ and serves as an essential intermediate step of generating $p_i$ . + +To clarify our problem, we assume that all reasoning is confined to the depicted scene and that each expression is spoken by a human who intends for the listener to take a particular action based on the situation. The ultimate objective of the task is to interpret the hidden intention $T$ effectively by leveraging the contextual cues within the image. + +To evaluate how well models capture such intentions, we adopt a multiple-choice (MCQ) format as the primary setup, as shown in Fig. 2. This decision reflects the fact that certain prompts can lead to multiple plausible outcomes, driven by hierarchical relations (e.g., pick up the chips - snack - food) or by the inherent uncertainty of what action best satisfies the speaker's goal (e.g., an indirect prompt complaining about darkness could be addressed by either turning on a light or opening curtains). Exploring all possible valid interpretations is labor-intensive and often infeasible. Consequently, each MID instance is presented as four distinct options, one correct and three intentionally designed to be incorrect for different reasons (see Sec. 4.2.3), challenging models in both linguistic and visual reasoning. + +Formally, let $C$ be the set of all multiple-choice options $c_{n}$ . Given an image $I$ and an indirect prompt $p_i$ , the task is to select the most valid interpretation of $p_i$ from the predefined options, conditioned on the visual context in $I$ . We define this task as follows: + +$$ +T (I, p _ {i}) := \operatorname {a r g m a x} _ {c _ {n} \in C} \Pr \left(c _ {n} \mid I, p _ {i}\right). \tag {1} +$$ + +# 3.2. Direct and Indirect Expressions + +By the design of our task, curating effective input prompts $p_d$ and $p_i$ is crucial for ensuring accurate interpretation. What makes a good prompt, though? In this section, we define and explain the criteria that both direct and indirect expressions must satisfy. For more details on good and bad examples for each criterion, please refer to our Appendix A. + +# 3.2.1. Directness: Relevance and Solvability + +Relevance The direct prompt is an utterance from the speaker that explicitly conveys its intended meaning without ambiguity. However, it is equally important that this intention aligns with the visual context of the scene. For example, a direct prompt $p_d$ such as "Hey person1, I want you to stop the fireworks" clearly expresses its intended action. However, if the corresponding image, as shown in Fig. 2, contains no elements related to fireworks, the prompt is misaligned with the scene. Thus, a direct prompt must not only reveal its intention but also maintain relevance to the image. In the context of our task, relevance is determined by whether a human can reasonably establish a connection between the prompt and the depicted scene. + +Solvability Relevance alone does not guarantee that a prompt is useful. As outlined in Sec. 3.1, the prompt $p_d$ must explicitly request an action that the listener can reasonably perform. This introduces solvability, which requires that the prompt present a clear and actionable problem. A solvable prompt defines a specific issue that can be addressed independently, ensuring that the listener is not left with multiple competing actions to choose from. + +# 3.2.2. Indirectness: Consistency and Ambiguity + +Consistency Indirect prompts are designed to obscure their true intention, but they must still convey the same underlying intention as their direct counterpart—essentially requesting the same solution. Since indirect prompts are derived from direct prompts, consistency serves as a key criterion. We define a direct prompt $p_d$ and an indirect prompt $p_i$ as consistent if their interpretations could potentially align in intention. The term "potentially" is used because indirect prompts, by nature, may have multiple valid interpretations. For instance, in the earlier "this is a disaster!" example, it conveys distress but allows for multiple reasonable responses, such as cleaning the room or providing reassurance. + +Ambiguity If an indirect prompt is entirely consistent with its direct counterpart without introducing any additional complexity, it becomes indistinguishable from a direct prompt. Therefore, an indirect prompt should conceal its underlying intention, which we define as ambiguity. The key principle behind this criterion is that neither the specific action required nor the key entity involved should be explicitly or implicitly mentioned within the prompt. Once these two elements are concealed, further refinements—such as adjusting the tone to be more indirect, sarcastic, or humorous—can enhance the overall nuance and difficulty of interpretation. + +# 4. VAGUE Benchmark Construction + +VAGUE is a novel benchmark that extends single-modal or simple multimodal ambiguity to more realistic domains and evaluates whether concurrent vision-language models can perform human-like reasoning with complex visual contexts. It comprises 1,677 images, with 1,144 sourced from the VCR [46] dataset and 533 from Ego4D [41], covering diverse contextual scenarios as well as real-world human interactions. On average, VAGUE contains seven objects and four people per image. Each image is paired with a direct expression $p_d$ , an indirect expression $p_i$ , and four multiple-choice answers, along with relevant meta-information. All textual components are generated using GPT-4o, then refined through extensive human rating, selection, and filtering, ensuring a carefully curated benchmark dataset for testing and advancing multimodal reasoning. We provide detailed benchmark statistics and a diversity analysis in Appendix B.2. + +# 4.1. Visual Data Curation + +# 4.1.1. Sampling + +VCR [46] The VCR dataset consists of 110K movie scenes sourced from the Large Scale Movie Description Challenge [35] and YouTube clips. These images are curated based on an "interestingness" criterion [46], ensuring the presence of at least two people, which promotes interactive scenarios. To prevent redundancy in our dataset, we sample 10K images while carefully avoiding neighboring frames, as adjacent frames exhibit minimal variation. This selection process preserves the contextual diversity of the dataset while maintaining its focus on complex, multi-entity interactions. + +Ego4D [41] While VCR provides a wide range of contextual diversity, it often includes artificially composed settings that may not fully capture real-world interactions. To address this, we integrate frames from the Ego4D dataset, which offers a more naturalistic depiction of human interactions. We specifically leverage the AV (Audio-Visual), which indicate conversational exchanges between individuals, to ensure the presence of people in the selected frames. Similar to VCR, we avoid neighboring frames to maintain diversity and filter out heavily blurry images to enhance data quality. This process results in 888 candidate images from 94 videos, which serve as the basis for further text processing. + +# 4.1.2. Object Extraction + +To ensure the complexity of the visual information, we extract a list of physical objects present in each image using a tagging model, RAM [49]. This step allows us to easily identify scenes with sufficient visual detail. In the case of VCR, many scenes are relatively simple, often containing only a few objects. Therefore, we sort VCR images by the number of detected objects and retained the top 4,000 as candidates for text processing, ensuring that our benchmark primarily consists of rich visual cues in contextually diverse scenes. Please refer to Appendix B.3 for more details. + +# 4.1.3. Person Indicator + +In our task, we assume that the speaker is outside the scene, viewing the image and talking to a person in it. However, identifying the addressee is not always straightforward, as images mostly contain multiple individuals. While grounding the specific person referenced in the utterance could introduce additional complexity, it is not the primary focus of our evaluation. To clarify the listener, we assign an indicator tag to each person in the image as shown in Fig. 3. For VCR, we use their existing annotations [46], while for Ego4D, where bounding boxes are not fully available, we employ YOLOv11 [18] to detect and annotate humans. While this provides a straightforward method for grounding the target person, it requires models to perform basic Optical Character Recognition (OCR). Therefore, we conduct exper + +![](images/9e1c2cc12c6b85ca62f868102b42281b9ec99a2fd7df1e4755d09e2cd23bc692.jpg) +Figure 3. Overview of the data generation process. Based on human-defined criteria and instructions, GPT-4o [31] generates initial data, which are then rated and filtered by humans to ensure quality. Since generating high-quality indirect expressions from raw images is challenging, the process follows these steps: generating a direct expression and intention from the image, creating an indirect expression using information from the previous step, and producing answer candidates based on all the information gathered so far. In the answer candidates, FS stands for Fake Scene Understanding, SU stands for Superficial Understanding, and NE stands for Nonexistent Entity. FS evaluates global hallucination, where the model misinterprets visual context from an entirely different image (e.g., an outdoor camping scene with a candle), while NE addresses local hallucination, where only a specific object is replaced with a fabricated one (e.g., curtains). + +iments to evaluate the OCR capabilities of the models used in our study. See Appendix B.4 for details. + +# 4.2. Multimodal Expression Synthesis + +As shown in Fig. 3, VAGUE's textual expressions are generated first by the model and undergo an extensive process of human rating and filtering. We use GPT-4o [31] for all text processing. The instructions used for generating direct $(p_d)$ and indirect $(p_i)$ expressions are provided in Appendix B.5, while the instructions for generating answer candidates for multiple-choice questions are detailed in Appendix B.7. + +# 4.2.1. Direct Expressions + +The direct expression $p_d$ serves as a crucial foundation for crafting the indirect one $p_i$ , as both share the same underlying intention. To ensure that $p_d$ adhere to the principles of relevance and solvability discussed in Sec. 3.2.1, we generate $p_d$ conditioned on the input image $I$ and a task prompt that explicitly defines these criteria. During the generation process, we also instruct the model to output a solution triplet in the format: (subject, action, object). Since direct expressions explicitly state their intentions, extracting each component of the solution triplet is straightforward. To maintain consistency with the visual context, the "object" in the triplet is restricted to physical objects we extracted in Sec. 4.1.2. + +After generating $p_d$ for all candidate images, human raters evaluate each prompt based on relevance and solvability, assigning scores from 1 to 5. Only those prompts that receive a rating of 4 or 5 are retained for further use. The detailed rating criteria for human verification and an example of the rating process are provided in Fig. J13. + +# 4.2.2. Indirect Expressions + +To ensure that the indirect expressions $p_i$ maintain both ambiguity and fluency while aligning with the true intent $T$ , we adopt a two-stage process: proposal and selection. + +In the initial step, the model is prompted to generate three distinct candidate options. Each candidate follows the criteria outlined in Section 3.2.2, but with explicit instructions to incorporate different linguistic strategies: sarcasm, humor, and meme/idiomatic expressions, respectively. This approach ensures diversity in the generated responses. In the second step, human annotators evaluate the three candidates and select the one that best aligns with the intended indirectness. The selected prompt is then rated on a scale from 1 to 5 based on more specific criteria. Only those prompts that receive a score of 3 or higher are retained for use in the dataset. The detailed rating criteria for human verification and an example of rating process are provided in Fig. J14 + +# 4.2.3. Counterfactual Choices + +Generating high-quality counterfactual choices is crucial. To enable detailed analysis of model weaknesses in multimodal intent disambiguation, we design interpretable counterfactual choices that provide more plausible alternatives. + +Fake Scene Understanding The first counterfactual choice is an interpretation that could arise when the model largely misinterprets the image. This process is conducted in two steps. In the first step, a fake caption is generated by assuming an imaginary scene that can be aligned with the indirect expression but is inconsistent with the true intent. The caption of fake scene is then combined with the speaker's + +indirect statement to derive the most likely interpretation. + +Superficial Understanding The subsequent choice corresponds to an interpretation generated when the model fails to deeply reason about the implicit intent of the sentence and instead relies on surface-level meaning. We enforce the model to focus only on the literal wording, without considering any implied or deeper meaning of the indirect sentence. These answer choices are generated alongside the indirect expression $p_i$ . During the indirect selection phase, the corresponding superficially understood choice is selected together, maintaining coherence between them. + +Nonexistent Entity The last choice arises when the model interprets the text correctly but fails to adequately consider the details of the image, resulting in a plausible yet incorrect choice. This resembles the correct answer in structure, but replaces the key object in the solution with one that does not exist in the image. To prevent the task from becoming too easy by generating highly irrelevant objects, we constrain the substituted object to one that, while absent from the image, is highly expected to be present in the scene and could replace the original entity. To identify such entities, the model is provided with the image as input to choose objects that align with the scene's context. This method ensures that the counterfactual choice leverages the expected coherence between the scene and its potential entities while rigorously testing the model's attention to visual details. + +# 5. Experiments + +Models We use the following models in our experiments. The detailed descriptions of each model are in Appendix C. + +- Phi3.5-Vision-Instruct (4B) [29] +- LLaVA Onevision (7B) [23] +Qwen2.5-VL-Instruct (7B, 72B) [43] +- InternVL-2.5-MPO (8B, 26B) [6] +- Idefics2 (8B) [22] +LLaVA NeXT Vicuna (13B) [24] +- Ovis2 (16B) [28] +GPT-4o [31] +Gemini 1.5 Pro [12] +- InternVL-3 (38B) [42] + +# 5.1. MLLMs Benefit from Visual Cues + +Our first objective is to assess how effectively MLLMs leverage visual cues to resolve ambiguity in utterances. To this end, we systematically control the level of detail in the visual cues provided to the models and measure their accuracy in inferring the speaker's true intent. Performance is evaluated in both multiple-choice and free-form settings. For clarity, we primarily report multiple-choice accuracy, deferring free-form results to Appendix I. + +We consider three levels of visual cues: + +- Language Models (LMs) receive no visual input, requiring models to rely on superficial textual priors such as common-sense knowledge of sarcasm or humor to determine intent. +- Socratic Models (SMs) [47] use text-only LMs but incorporate short image captions (up to two or three sentences) as additional input. This short generic caption may lack sufficient detail, which may be insufficient for accurately inferring intent. Each SM model generated its own image captions and used them in subsequent processing. +- Visual Language Models (VLMs) receive the raw image input, enabling a more direct interpretation of visual cues. + +Results The results in Tab. 1 indicate that MLLMs can leverage visual cues, albeit to a limited extent, since SMs and VLMs consistently outperform LMs across all evaluated models. Additionally, more detailed visual input generally improves performance, with VLMs surpassing SMs in most cases, except in the case of proprietary models. This exception is further analyzed in Sec. 5.2. Both the VCR and Ego4D subsets exhibit similar performance trends, demonstrating the generalizability of our findings across both staged and real-world scenarios. Finally, the consistently low performance of LMs further reinforces the validity of our dataset as a multimodal benchmark. + +# 5.2. Analysis on Failure Modes + +Here, we examine the ways in which models fail to infer the true intent and how these failure patterns vary with the level of visual cues provided. As shown in Fig. 3, our multiple-choice questions include three distinct types of incorrect answer candidates. We assess the model's raw visual understanding using the Fake Scene Understanding (FS) and Nonexistent Entity (NE) candidates, which test whether the model can correctly interpret the scene without being misled by fabricated or nonexistent elements. Conversely, the Superficial Understanding (SU) candidate evaluates the model's reasoning ability, testing whether it can go beyond surface-level perception to infer intent. We provide the full table of failure modes in the MCQ setting in Appendix I + +Results Figure 4 illustrates how frequently each model selects different types of incorrect answers instead of the correct intent. Among the error types, Superficial Understanding (SU) is the most common. This indicates that while models generally succeed in recognizing basic visual details, they often fail to reason deeply about those visual cues to accurately infer the underlying intent of the speaker. However, proprietary models exhibit fewer SU-related errors, indicating stronger reasoning capabilities. + +Moreover, stronger visual cues improve accuracy across all models and failure types. This improvement highlights the crucial role of visual conditioning in reducing both vision-based errors (FS and NE) and reasoning-related failures (SU). + +
ModelVAGUE-VCRVAGUE-Ego4D
LM (L)SM (L+V)VLM (L+V)LM (L)SM (L+V)VLM (L+V)
Phi3.5-Vision-Instruct (4B)26.635.3 (↑ 8.7)46.0 (↑ 19.4)22.531.1 (↑ 8.6)42.4 (↑ 19.9)
LLaVA-Onevision (7B)13.129.4 (↑ 16.3)43.1 (↑ 30.0)11.329.5 (↑ 18.2)43.2 (↑ 31.9)
Qwen2.5-VL-Instruct (7B)11.125.6 (↑ 14.5)46.8 (↑ 35.7)9.828.0 (↑ 18.2)48.4 (↑ 38.6)
InternVL-2.5-MPO (8B)23.048.4 (↑ 25.4)63.9 (↑ 40.9)24.254.0 (↑ 29.8)66.8 (↑ 42.6)
Idefics2 (8B)13.921.1 (↑ 7.2)58.7 (↑ 44.8)14.818.2 (↑ 3.4)58.3 (↑ 43.5)
LLaVA-NeXT-vicuna (13B)24.237.2 (↑ 13)46.4 (↑ 22.2)20.334.1 (↑ 13.8)52.5 (↑ 32.2)
Ovis2 (16B)21.923.8 (↑ 1.9)24.5 (↑ 3.6)20.525.3 (↑ 4.8)25.7 (↑ 5.2)
InternVL-2.5-MPO (26B)21.248.5 (↑ 27.3)63.7 (↑ 42.5)21.855.2 (↑ 33.4)68.7 (↑ 46.9)
InternVL-3 (38B)24.847.2 (↑ 22.4)63.6 (↑ 38.8)18.047.5 (↑ 29.5)59.8 (↑ 41.8)
Qwen2.5-VL-Instruct (72B)29.655.6 (↑ 26.0)74.2 (↑ 44.6)26.859.3 (↑ 32.5)69.8 (↑ 43.0)
GPT-4o46.469.5 (↑ 23.1)65.1 (↑ 18.7)48.267.5 (↑ 19.3)63.6 (↑ 15.3)
Gemini-1.5-Pro43.262.4 (↑ 19.2)60.6 (↑ 17.4)40.360.6 (↑ 20.3)60.6 (↑ 20.3)
+ +Table 1. Experiments on the Multimodal Intention Disambiguation (MID) task with varying levels of visual cues. We report the accuracy $(\%)$ of the Multiple-Choice Question. $\uparrow$ indicates the performance gain from visual cues, i.e. increment compared to the LM setting.(L) denotes the use of language input only, while $(\mathrm{L} + \mathrm{V})$ indicates the incorporation of visual cues. The noticeable increase in accuracy across LM, SM, and VLM demonstrates that the introduction of detailed visual cues is beneficial for the task. + +![](images/f063bf44d3c5ca75ea7e8831036cb2b6bff129c5ecbd71bdc54b08faf9ac3d72.jpg) +Figure 4. We present a bar plot to analyze the distribution of incorrect answer choices selected by each model. Each number represents how frequently a given choice was selected from the 1,677 items in the dataset. The counterfactual choice categories are FS (Fake Scene Understanding), SU (Superficial Understanding), and NE (Nonexistent Entity). We use distinct colors to represent LM (Language Models), SM (Socratic Models), and VLM (Visual-Language Models). + +Notably, proprietary models perform better with captioned inputs (SM) than with raw images (VLM). A closer examination of Fig. 4 reveals that this discrepancy arises from vision-based failures (FS and NE) rather than reasoning-centric failure (SU). This indicates that their captioning ability potentially allows them to obtain more sophisticated information while reducing hallucination in the images. The supporting experiments and explanations are provided in Appendix F. + +# 5.3. Comparison with human + +To validate our benchmark and establish an upper bound for performance, we assess human accuracy, highlighting the gap between existing models and human capability. This evaluation follows the multiple-choice setup within the VLM setting, using a subset of 400 samples. As shown + +in Tab. 2, human performance reaches $94\%$ , demonstrating near-perfect accuracy. Although proprietary models and some large-sized models show a decent performance, a notable performance gap ( $\sim 20\%$ ) still exists compared to human evaluators. This result underscores the significant gap between AI models' multimodal reasoning capabilities and human-level understanding when inferring hidden intent, suggesting that visual perception alone, even when accurate, is insufficient without deeper cognitive integration. This performance gap is due to the models' tendency to rely on surface-level text rather than understanding deeper visual-textual implications. Thus, advancing multimodal reasoning likely requires models to integrate higher-order cognitive processes, like commonsense reasoning and pragmatic understanding, into visual interpretation tasks. Refer + +
ModelAcc (%)FSSUNECorrect
Ovis2 (16B)23.01191711892
LLaVA-Onevision (7B)41.04318310164
Phi3.5-Vision-Instruct (4B)44.36013231177
Qwen2.5-VL-Instruct (7B)47.04715213188
LLaVA-NeXT-icuna (13B)48.04814317192
Idefics2 (8b)57.02812024228
Gemini-1.5-Pro60.3607326241
InternVL-3 (38B)61.5389422246
InternVL-2.5-MPO (8B)61.8429516247
GPT-4o62.3616327249
InternVL-2.5-MPO (26B)63.03610111252
Qwen2.5-VL-Instruct (72B)72.3396012289
Human94.01248374
+ +Table 2. Performance across models and humans on a sampled set of 400 questions. The results show that humans outperform models by a margin of over $20\%$ . + +
ModelTypeAcc (%)Incorrect count
FSSUNE
GPT-4oSM68.9165240117
SM+CoT69.5 (↑ 0.6)165241105
VLM64.6226226141
VLM+CoT66.4 (↑ 1.8)16215685
Gemini-1.5-ProSM61.817637483
SM+CoT61.0 (↓ 0.8)19036794
VLM60.6249264145
VLM+CoT64.4 (↑ 3.8)213267117
+ +Table 3. Result of Chain-of-Thought (CoT) experiments on proprietary models, in both SM and VLM settings. $\uparrow$ and $\downarrow$ indicate an increase and decrease in accuracy when zero-shot CoT is applied. + +to Appendix E for details on the selected subset and human evaluation setup. + +# 5.4. Chain-of-Thought Experiments + +Given the strong reasoning demands of multimodal intent deduction, we further explore the effectiveness of Chain-of-Thought (CoT) prompting [44] in enhancing the reasoning capabilities of MLLMs. The CoT prompt templates, provided in Fig. J21 and Fig. J22, are designed to explicitly ground the reasoning process, reducing hallucinations. While our main results focus on proprietary models, owing to their superior suitability for zero-shot CoT reasoning, we also present experiments and analyses for open-source models ranging from 4B to 72B in Appendix G. + +Results As shown in Tab. 3, CoT prompting improves performance for raw image inputs (VLM), while showing no clear trend and remaining at a similar level for image caption inputs (SM). One possible explanation for this discrepancy is that CoT primarily enhances reasoning by improving + +grounding and reducing hallucinations. Since image captions inherently contain fewer hallucinations, SMs see some benefit from CoT prompting albeit at the cost of reduced detail. Additionally, the performance improvements observed with CoT prompting are consistent across different types of false answer candidates, suggesting a generalizable effect in enhancing reasoning quality. + +# 6. Conclusion + +We present VAGUE (Visual Contexts CIArify ambiGUous Expressions), a new benchmark aimed at assessing models' ability to interpret nuanced communication in complex multimodal scenarios. Our results show that models benefit from visual information when inferring the underlying intention of indirect expressions, as evidenced by their improved performance with increasing levels of visual cues. However, a significant disparity persists between machine capabilities and humans. To gain deeper insights into model inaccuracies, we design multiple-choice questions that explicitly address failure points, enabling a systematic and quantitative evaluation of the reasons behind performance. The primary challenge identified is the tendency of multimodal models to inadequately integrate visual cues, relying instead on the literal interpretation of textual information. This shortcoming highlights the need for further research, and we anticipate that VAGUE will open up promising avenues for developing systems capable of deeper multimodal reasoning to enhance AI's ability to engage in human-like interactions. + +# 7. Limitations + +First, we acknowledge that there are cultural and linguistic limitations. We incorporate sarcastic, humorous, and idiomatic implicatures in generating indirect expressions. However, since the initial data drafts are created by a model (GPT-4o [31]), they may reflect cultural biases present in its training data. To prevent any potential ethical issues, all human annotators are instructed to remove any content considered problematic or discriminatory during the rating and filtering process. Also, all textual expressions in VAGUE are limited to English. Therefore, we encourage future exploration of indirect expressions across diverse languages and cultures. The second limitation pertains to the dependency of certain meta-information on the quality of the parent dataset and the performance of the models utilized during the dataset generation. We observe that bounding box annotations from YOLOv11 [18] are occasionally duplicated, leading to a reported number of people higher than actually present. Likewise, the tagging model RAM [49] sometimes misidentified objects. Therefore, we remove any corrupted instances that could undermine the integrity of the task during the human rating and filtering process. + +# References + +[1] Ashwin Baluja. Text is not all you need: Multimodal prompting helps llms understand humor. arXiv preprint arXiv:2412.05315, 2024. 3 +[2] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1 +[3] Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi, and Andreas Bulling. Explicit modelling of theory of mind for belief prediction in nonverbal social interactions. In Proc. 27th European Conference on Artificial Intelligence (ECAI), pages 1-8, 2024. 2 +[4] Tuhin Chakrabarty, Arkadiy Saakyan, Debanjan Ghosh, and Smaranda Muresan. FLUTE: Figurative language understanding through textual explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7139-7159, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2 +[5] Zhawnen Chen, Tianchun Wang, Yizhou Wang, Michal Kosinski, Xiang Zhang, Yun Fu, and Sheng Li. Through the theory of mind's eye: Reading minds with multimodal video large language models. arXiv preprint arXiv:2406.13763, 2024. 2 +[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6, 3 +[7] Christine Cuskley, Rebecca Woods, and Molly Flaherty. The limitations of large language models for understanding human language and cognition. Open Mind, 8:1058-1083, 2024. 2 +[8] Hang Du, Guoshun Nan, Sicheng Zhang, Binzhu Xie, Junrui Xu, Hehe Fan, Qimei Cui, Xiaofeng Tao, and Xudong Jiang. Docmsu: A comprehensive benchmark for document-level multimodal sarcasm understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17933-17941, 2024. 2 +[9] Jiafei Duan, Samson Yu, Nicholas Tan, Li Yi, and Cheston Tan. Boss: A benchmark for human belief prediction in object-context scenarios. arXiv preprint arXiv:2206.10665, 2022. 2 +[10] Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. CVPR 2024, 2024. 4 +[11] Kanishk Gandhi, Jan-Philipp Fraenkel, Tobias Gerstenberg, and Noah D. Goodman. Understanding social reasoning in language models with language models. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track, 2023. Dataset and Benchmarks Track. 2 +[12] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6, 3 + +[13] Md Kamrul Hasan, Wasifur Rahman, AmirAli Bagher Zadeh, Jianyuan Zhong, Md Iftekhar Tanveer, Louis-Philippe Morency, and Mohammed (Ehsan) Hoque. UR-FUNNY: A multimodal language dataset for understanding humor. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2046–2056, Hong Kong, China, 2019. Association for Computational Linguistics. 2 +[14] Jack Hessel, Ana Marasovic, Jena D. Hwang, Lillian Lee, Jeff Da, Rowan Zellers, Robert Mankoff, and Yejin Choi. Do androids laugh at electric sheep? humor "understanding" benchmarks from the new yorker caption contest. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 688-714, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[15] EunJeong Hwang and Vered Shwartz. Memecap: A dataset for captioning and interpreting memes. arXiv preprint arXiv:2305.13703, 2023. 3 +[16] Prince Jha, Krishanu Maity, Raghav Jain, Apoorv Verma, Sriparna Saha, and Pushpak Bhattacharyya. Meme-ingful analysis: Enhanced understanding of cyberbullying in memes through multimodal explanations. arXiv preprint arXiv:2401.09899, 2024.3 +[17] Chuanyang Jin, Yutong Wu, Jing Cao, Jiannan Xiang, Yen Ling Kuo, Zhiting Hu, Tomer Ullman, Antonio Torralba, Joshua Tenenbaum, and Tianmin Shu. MMToM-QA: Multimodal theory of mind question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16077-16102, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2 +[18] Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements, 2024. 4, 8 +[19] Shivani Kumar, Atharva Kulkarni, Md Shad Akhtar, and Tan-moy Chakraborty. When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5956-5968, Dublin, Ireland, 2022. Association for Computational Linguistics. 2 +[20] Shivani Kumar, Ishani Mondai, Md Shad Akhtar, and Tanmoy Chakraborty. Explaining (sarcastic) utterances to enhance affect understanding in multimodal dialogues. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 2 +[21] Yash Kumar Lal and Mohaddeseh Bastan. SBU figures it out: Models explain figurative language. In Proceedings of the 3rd Workshop on Figurative Language Processing (FLP), pages 143-149, Abu Dhabi, United Arab Emirates (Hybrid), 2022. Association for Computational Linguistics. 2 +[22] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language mod + +els? Advances in Neural Information Processing Systems, 37: 87874-87907, 2025. 6, 3 +[23] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 3 +[24] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 6, 3 +[25] Kevin Y. Li, Sachin Goyal, Joao D. Semedo, and J. Zico Kolter. Inference optimal vlms need fewer visual tokens and more parameters, 2025. 4 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing. 1, 3 +[27] Annie Louis, Dan Roth, and Filip Radlinski. "I'd rather just go to bed": Understanding indirect answers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7411-7425, Online, 2020. Association for Computational Linguistics. 2 +[28] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 6, 3 +[29] Microsoft. Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6, 3 +[30] Maryam Sadat Mirzaei, Kourosh Meshgi, and Satoshi Sekine What is the real intention behind this question? dataset collection and intention classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13606-13622, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[31] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 8, 3 +[32] Kun Ouyang, Liqiang Jing, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Sentiment-enhanced graph-based sarcasm explanation in dialogue. arXiv:2402.11414, 2024. 2 +[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 3, 4 +[34] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 2 +[35] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Chris Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. International Journal of Computer Vision, 2017. 4 +[36] Melanie Sclar, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia Tsvetkov. Minding language models' (lack of) theory of mind: A plug-and-play multi-character belief + +tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 13960-13980, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[37] Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin, Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. Muma-tom: Multi-modal multi-agent theory of mind, 2024. 2 +[38] Settaluri Sravanthi, Meet Doshi, Pavan Tankala, Rudra Murthy, Raj Dabre, and Pushpak Bhattacharyya. PUB: A pragmatics understanding benchmark for assessing LLMs' pragmatics capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12075-12097, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2 +[39] Chuandong Su, Fumiyo Fukumoto, Xiaoxi Huang, Jiyi Li, Rongbo Wang, and Zhiqun Chen. DeepMet: A reading comprehension paradigm for token-level metaphor detection. In Proceedings of the Second Workshop on Figurative Language Processing, pages 30-39, Online, 2020. Association for Computational Linguistics. 2 +[40] Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase. DIRECT: Direct and indirect responses in conversational text corpus. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1980-1989, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2 +[41] Ego4d Team. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 2, 4 +[42] OpenGVLab Team. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025. 6, 3 +[43] Qwen Team. Qwen2.5-v1 technical report, 2025. 6, 3 +[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 8 +[45] Alex Wilf, Leena Mathur, Sheryl Mathew, Claire Ko, Youssof Kebe, Paul Pu Liang, and Louis-Philippe Morency. Social-iq 2.0 challenge: Benchmarking multimodal social understanding. https://github.com/abwilf/Social-IQ-2.0-Challenge, 2023.2 +[46] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6713-6724, 2018. 2, 4, 1 +[47] Andy Zeng, Maria Attarian, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, 2023. 2, 6 +[48] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. 4 + +[49] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1724-1732, 2024. 4, 8, 1 \ No newline at end of file diff --git a/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/images.zip b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..92c51f3e728e94d067db8bcbb28a6d71aa09f318 --- /dev/null +++ b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba7a1da700708ff8ed1331c24bb4074d291f99e3a6ce7b95ac436f4c672d79ee +size 451798 diff --git a/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/layout.json b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..88be50d8142f41443c26b536d50a2a70f0311690 --- /dev/null +++ b/2025/VAGUE_ Visual Contexts Clarify Ambiguous Expressions/layout.json @@ -0,0 +1,8169 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "VAGUE: Visual Contexts Clarify Ambiguous Expressions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "text", + "content": "Heejeong " + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{Nam}^{*1}" + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "text", + "content": ", Jinwoo Ahn" + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "inline_equation", + "content": "^{*2}" + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "text", + "content": ", Keummin " + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{Ka}^3" + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "text", + "content": ", Jiwan Chung" + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "text", + "content": ", and Youngjae " + }, + { + "bbox": [ + 104, + 152, + 506, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^{\\dagger 3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 177, + 352, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 177, + 352, + 191 + ], + "spans": [ + { + "bbox": [ + 258, + 177, + 352, + 191 + ], + "type": "text", + "content": "Brown University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 270, + 191, + 341, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 270, + 191, + 341, + 205 + ], + "spans": [ + { + "bbox": [ + 270, + 191, + 341, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 270, + 191, + 341, + 205 + ], + "type": "text", + "content": "UC Berkeley" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 259, + 205, + 351, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 205, + 351, + 220 + ], + "spans": [ + { + "bbox": [ + 259, + 205, + 351, + 220 + ], + "type": "text", + "content": "3Yonsei University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 152, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 281, + 297, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 297, + 533 + ], + "type": "text", + "content": "Human communication often relies on visual cues to resolve ambiguity. While humans can intuitively integrate these cues, AI systems often find it challenging to engage in sophisticated multimodal reasoning. We introduce VAGUE, a benchmark evaluating multimodal AI systems' ability to integrate visual context for intent disambiguation. VAGUE consists of 1.6K ambiguous textual expressions, each paired with an image and multiple-choice interpretations, where the correct answer is only apparent with visual context. The dataset spans both staged, complex (Visual Commonsense Reasoning) and natural, personal (Ego4D) scenes, ensuring diversity. Our experiments reveal that existing multimodal AI models struggle to infer the speaker's true intent. While performance consistently improves from the introduction of more visual cues, the overall accuracy remains far below human performance, highlighting a critical gap in multimodal reasoning. Analysis of failure cases demonstrates that current models fail to distinguish true intent from superficial correlations in the visual scene, indicating that they perceive images but do not effectively reason with them. We release our code and data at https://hazel-heejeong-nam.github.io/vague/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 576, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 647 + ], + "type": "text", + "content": "Human communication is inherently contextual; for example, exclaiming \"Hey, this is a disaster!\" upon seeing a cluttered room conveys frustration or exaggeration rather than referring to an actual catastrophe. Without surrounding cues, textual dialogues can be ambiguous, making it difficult for models to accurately capture intent and nuance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "type": "text", + "content": "We consider the case of visual contextual cues. Consider Fig. 1, which depicts a speaker making a remark in a certain situation. Without specifying contexts introduced" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 348, + 261, + 367, + 281 + ], + "blocks": [ + { + "bbox": [ + 348, + 261, + 367, + 281 + ], + "lines": [ + { + "bbox": [ + 348, + 261, + 367, + 281 + ], + "spans": [ + { + "bbox": [ + 348, + 261, + 367, + 281 + ], + "type": "image", + "image_path": "e2a8dbca01ca7f9cf2de59e57ff233135cba5e760ade85d2b736c88db6c48b0f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 262, + 518, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 262, + 518, + 281 + ], + "spans": [ + { + "bbox": [ + 370, + 262, + 518, + 281 + ], + "type": "text", + "content": "Hey person1, looks like you're planning a party for the neighborhood squirrels." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 344, + 292, + 355, + 299 + ], + "blocks": [ + { + "bbox": [ + 344, + 292, + 355, + 299 + ], + "lines": [ + { + "bbox": [ + 344, + 292, + 355, + 299 + ], + "spans": [ + { + "bbox": [ + 344, + 292, + 355, + 299 + ], + "type": "image", + "image_path": "bb8ba85a6b859c036abd1d76f868a67d75e2753b51dd3ebe0919d8588c4362f9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 301, + 432, + 376 + ], + "blocks": [ + { + "bbox": [ + 356, + 292, + 408, + 299 + ], + "lines": [ + { + "bbox": [ + 356, + 292, + 408, + 299 + ], + "spans": [ + { + "bbox": [ + 356, + 292, + 408, + 299 + ], + "type": "text", + "content": "Visual Context 1" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 301, + 432, + 376 + ], + "lines": [ + { + "bbox": [ + 317, + 301, + 432, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 301, + 432, + 376 + ], + "type": "image", + "image_path": "9b146a4622b5effe44329b07299bd76f8a88078e788e652b22affa5b4f4e0d2a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 460, + 292, + 471, + 299 + ], + "blocks": [ + { + "bbox": [ + 460, + 292, + 471, + 299 + ], + "lines": [ + { + "bbox": [ + 460, + 292, + 471, + 299 + ], + "spans": [ + { + "bbox": [ + 460, + 292, + 471, + 299 + ], + "type": "image", + "image_path": "509ea48f33bcd7e74ab1e7eabd010e041181ef09744a26e395c2c99caa9b002f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 436, + 300, + 551, + 377 + ], + "blocks": [ + { + "bbox": [ + 472, + 292, + 525, + 299 + ], + "lines": [ + { + "bbox": [ + 472, + 292, + 525, + 299 + ], + "spans": [ + { + "bbox": [ + 472, + 292, + 525, + 299 + ], + "type": "text", + "content": "Visual Context 2" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 300, + 551, + 377 + ], + "lines": [ + { + "bbox": [ + 436, + 300, + 551, + 377 + ], + "spans": [ + { + "bbox": [ + 436, + 300, + 551, + 377 + ], + "type": "image", + "image_path": "1bdd4bc388a4d7802e8453cf79b21f4526a541f221b7420bec508b25359cb457.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 369, + 381, + 384, + 398 + ], + "blocks": [ + { + "bbox": [ + 369, + 381, + 384, + 398 + ], + "lines": [ + { + "bbox": [ + 369, + 381, + 384, + 398 + ], + "spans": [ + { + "bbox": [ + 369, + 381, + 384, + 398 + ], + "type": "image", + "image_path": "87898e9bddf2979562ffd934554fc2f25bb756e79aba17f08f153bf2b21ff6c4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 384, + 495, + 393 + ], + "lines": [ + { + "bbox": [ + 388, + 384, + 495, + 393 + ], + "spans": [ + { + "bbox": [ + 388, + 384, + 495, + 393 + ], + "type": "text", + "content": "The speaker wants person1 to..." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 428, + 411, + 440, + 422 + ], + "blocks": [ + { + "bbox": [ + 338, + 412, + 411, + 420 + ], + "lines": [ + { + "bbox": [ + 338, + 412, + 411, + 420 + ], + "spans": [ + { + "bbox": [ + 338, + 412, + 411, + 420 + ], + "type": "text", + "content": "close the cereal box" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 411, + 440, + 422 + ], + "lines": [ + { + "bbox": [ + 428, + 411, + 440, + 422 + ], + "spans": [ + { + "bbox": [ + 428, + 411, + 440, + 422 + ], + "type": "image", + "image_path": "f39f695adddd5796b67d5a45b0ac7a1636c83143ef23955c8f89b99045f1c29a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 408, + 535, + 426 + ], + "lines": [ + { + "bbox": [ + 459, + 408, + 535, + 426 + ], + "spans": [ + { + "bbox": [ + 459, + 408, + 535, + 426 + ], + "type": "text", + "content": "hand out the envelop to the squirrels" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 439, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 495 + ], + "type": "text", + "content": "Figure 1. A motivating example demonstrating the importance of visual context in understanding intention. Without a predetermined context, a single expression can convey multiple different intentions. The textual expression and Context 1 are from our dataset, while Context 2 is generated using DALL-E 3 [2] to help understanding." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 520, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 581 + ], + "type": "text", + "content": "from visual cues, the speaker's intention can vary, thus remaining ambiguous. This implies that the visual contexts play important roles in communication, raising the question: can AI systems integrate visual cues with ambiguous dialogue to infer the speaker's intent?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "content": "We introduce Visual Contexts ClArify ambiGUous Expressions (VAGUE), a benchmark consisting of 1.6K ambiguous textual expressions, each paired with a single image. VAGUE aims to model diverse and natural human-to-human interactions by setting each image as the speaker's viewpoint, where the speaker implicitly requests a certain action from a person within their field of view. We define the problem addressed through this setup as Multimodal Intention Disambiguation (MID), which involves reasoning about the most plausible request conditioned on visual context. Each sample in VAGUE is annotated with four multiple-choice candi" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 182, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 182, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 182, + 703 + ], + "type": "text", + "content": "*These authors contributed equally." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1537" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 217 + ], + "type": "text", + "content": "dates, ensuring clarity and preventing multiple valid answers caused by paraphrasing or hierarchical inclusion of meaning. The dataset is meticulously curated to ensure visual dependency; the ground-truth candidate is only preferable when considering the visual context. VAGUE includes visual scenes from both artificial sources (Visual Commonsense Reasoning [46]) and real-world scenarios (Ego4D [41]), capturing a broad spectrum of scene complexity and naturalness. The textual expressions in VAGUE are initially generated by GPT-4o [31] following instructions, then reviewed through extensive human rating and filtering to ensure naturalness and alignment with the corresponding images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 226, + 297, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 297, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 297, + 406 + ], + "type": "text", + "content": "Experiments on VAGUE demonstrate that existing multimodal AI models struggle to infer a speaker's true intent in a multimodal setting. First, although models can leverage visual context—as seen by a performance progression from text-only Language Models (LMs) to pipelined Socratic Models (SMs) [47] and ultimately to end-to-end Visual Language Models (VLMs)—their overall accuracy remains significantly lower than that of humans, indicating a failure to capture the true intent. A closer analysis of failure cases reveals that the primary source of error is the models' inability to distinguish the true intent from a superficial understanding of the visual context. In other words, even though these multimodal systems can perceive the image content, they cannot effectively use this information to reason about the speaker's true intent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 415, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 513 + ], + "type": "text", + "content": "In conclusion, we introduce a benchmark that exposes the limitations of current models in integrating visual cues with intent comprehension and identifies their primary failure mode. We anticipate that VAGUE will serve as a testing ground for the development of future multimodal conversational or embodied agents—systems that combine robust visual perception with nuanced conversational reasoning to effectively respond to user requests in complex scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 521, + 197, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 521, + 197, + 533 + ], + "spans": [ + { + "bbox": [ + 67, + 521, + 197, + 533 + ], + "type": "text", + "content": "Our contributions are threefold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 544, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 544, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 601 + ], + "type": "text", + "content": "- VAGUE: a novel benchmark for evaluating multimodal intention disambiguation. Validated through extensive human filtering, VAGUE is designed for robust quantitative assessment by ensuring both the ambiguity of queries and the visual (in)dependency to answer candidates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 603, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 650 + ], + "type": "text", + "content": "- Carefully curated 1,677 scene images sourced from VCR [46] and Ego4D [41], capturing a wide range of scene complexity, diversity, and naturalness to ensure VAGUE's generalizability across various contexts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 651, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 295, + 711 + ], + "type": "text", + "content": "- Experimental results highlighting a critical challenge in multimodal intention disambiguation: while existing models can perceive visual cues, they fail to effectively integrate this information into reasoning to deduce the speaker's true intent." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 91, + 470, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 470, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 470, + 104 + ], + "type": "text", + "content": "2.1. Multimodal Theory of Mind" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "text", + "content": "Theory of Mind (ToM) refers to the ability to infer and reason about the intentions of others based on available information [34], where recent language models still struggle with relevant tasks [7] highlighting the need for dedicated research in this area. Initially, various methods and benchmarks have been proposed in unimodal settings, relying on text-based approaches [11, 36]. However, these methods often fail to capture the richness of real-world interactions, which often require integrating both linguistic and visual cues." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 218, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 556, + 434 + ], + "type": "text", + "content": "Moving beyond text-only contexts, recent work has incorporated visual information. MMToM [17] introduces a benchmark where models must process both visual and textual cues to solve question-answering tasks related to ToM. The BOSS dataset [9] is a multimodal dataset collected in situations where nonverbal communication is required. It is used to evaluate whether human beliefs can be inferred based on nonverbal cues during social interactions. Similarly, Chen et al. (2024) [5] propose a Video ToM model that leverages key video frames and transcripts, demonstrating improved reasoning on the Social-IQ 2.0 dataset [45]. MuMA-ToM [37] further extends this direction by assessing ToM reasoning in multi-agent interactions, evaluating a model's ability to infer human beliefs and goals based on video and text inputs. MToMnet [3] introduces a ToM-based neural network that integrates contextual cues, such as scene videos and object locations, with person-specific cues, to predict human beliefs in specific scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 482 + ], + "type": "text", + "content": "However, progress in multimodal ToM remains constrained not only by the scarcity of high-quality datasets [5] but also by the lack of explicit consideration for the ambiguity and indirectness inherent in human communication." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 491, + 524, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 524, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 524, + 504 + ], + "type": "text", + "content": "2.2. Multimodal Implicature Understanding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 557, + 715 + ], + "type": "text", + "content": "Implicature and the ambiguity that arises from it naturally emerge in everyday human conversation, requiring pragmatic understanding [38]. Early research on implicature understanding has primarily been conducted in text-only settings [27, 30, 40], with some studies specifically focusing on figurative language and metaphor [4, 21, 39]. However, since the ambiguity of standalone text is inherently limited, recent studies have extended to multiple modalities. One example is multimodal sarcasm understanding (MSU). WITS [19] and MOSES [20] are benchmarks for sarcasm explanation, both providing the speaker's emotion and voice tone as cues. DocMSU [8] is a document-level benchmark for sarcasm localization and detection. To improve MSU, EDGE [32], a graph-based approach, achieved strong performance. UR-FUNNY [13] is a benchmark for multimodal humor comprehension, incorporating facial expressions and voice tones as in MSU [19, 20]. Hessel et al. (2023) [14] intro" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1538" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 65, + 79, + 285, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 79, + 285, + 101 + ], + "spans": [ + { + "bbox": [ + 65, + 79, + 285, + 101 + ], + "type": "text", + "content": "Q. Select the option that best explains the underlying intention of the utterance based on the given image." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 114, + 93, + 139 + ], + "blocks": [ + { + "bbox": [ + 65, + 114, + 93, + 139 + ], + "lines": [ + { + "bbox": [ + 65, + 114, + 93, + 139 + ], + "spans": [ + { + "bbox": [ + 65, + 114, + 93, + 139 + ], + "type": "image", + "image_path": "5e3b0a1b1a27751d011a6dcbe11cb25d073beac26e2f82c50585b0bb63611535.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "type": "text", + "content": "Figure 2. Description of the Multimodal Intention Disambiguation (MID) task in a Multiple-Choice Question format: Given an input image " + }, + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "type": "text", + "content": " and an indirect expression " + }, + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "type": "inline_equation", + "content": "(p_i)" + }, + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "type": "text", + "content": ", the goal is to infer the speaker's hidden intent " + }, + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 55, + 370, + 295, + 415 + ], + "type": "text", + "content": " and select the most likely answer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 116, + 258, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 116, + 258, + 137 + ], + "spans": [ + { + "bbox": [ + 100, + 116, + 258, + 137 + ], + "type": "text", + "content": "Hey person1, spot the difference, this parking's a bit too special isn't it?" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 145, + 293, + 247 + ], + "blocks": [ + { + "bbox": [ + 60, + 145, + 293, + 247 + ], + "lines": [ + { + "bbox": [ + 60, + 145, + 293, + 247 + ], + "spans": [ + { + "bbox": [ + 60, + 145, + 293, + 247 + ], + "type": "image", + "image_path": "d314ec8014afa1b1655e660cbf27f3b14d3fc17bd409690d07d9a53e5c01bf22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 251, + 291, + 357 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 61, + 251, + 281, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 281, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 281, + 274 + ], + "type": "text", + "content": "a) The speaker wants person1 to admire the unusually decorated motorcycle in the parking lot." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 279, + 250, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 279, + 250, + 301 + ], + "spans": [ + { + "bbox": [ + 60, + 279, + 250, + 301 + ], + "type": "text", + "content": "b) The speaker wants person1 to enjoy playing a puzzle game and spot the difference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 303, + 288, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 303, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 61, + 303, + 288, + 329 + ], + "type": "text", + "content": "c) The speaker wants person1 to move the sedan because it's in a handicapped parking spot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 334, + 291, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 334, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 334, + 291, + 357 + ], + "type": "text", + "content": "d) The speaker wants person1 to move the sedan because it's parked in front of a fire hydrant." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 509 + ], + "type": "text", + "content": "duced a benchmark derived from a Cartoon Caption Contest, exploring humor identification and explanation. Baluja et al. (2024) [1] demonstrated that models benefit from multimodal cues in humor understanding. Memes also involve implicature, with multimodal datasets such as MemeCap [15] and MultiBully-Ex [16]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 510, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 581 + ], + "type": "text", + "content": "However, the cues used in multimodal implicature understanding remain simple, primarily appearing in images with a single main object or person, overlooking the importance of interactions between multiple objects and people in real-world scenarios. These limitations underscore the need for more complex cues, as addressed in VAGUE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 265, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 265, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 265, + 607 + ], + "type": "text", + "content": "3. Multimodal Intention Disambiguation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 614, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 662 + ], + "type": "text", + "content": "In this section, we outline the structure and rationale behind the format of our primary task, which we term Multimodal Intention Disambiguation (MID). Then, we further specify the necessary components that form the basis of the task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 671, + 155, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 155, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 155, + 685 + ], + "type": "text", + "content": "3.1. Problem Setting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Each MID problem comprises an input image " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", a direct text expression " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", and an indirect text expression " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". Here," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "the direct expression " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": " clearly shows the underlying intention of the corresponding " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": " and serves as an essential intermediate step of generating " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 109, + 555, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 181 + ], + "type": "text", + "content": "To clarify our problem, we assume that all reasoning is confined to the depicted scene and that each expression is spoken by a human who intends for the listener to take a particular action based on the situation. The ultimate objective of the task is to interpret the hidden intention " + }, + { + "bbox": [ + 313, + 109, + 555, + 181 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 109, + 555, + 181 + ], + "type": "text", + "content": " effectively by leveraging the contextual cues within the image." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 182, + 556, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 556, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 556, + 350 + ], + "type": "text", + "content": "To evaluate how well models capture such intentions, we adopt a multiple-choice (MCQ) format as the primary setup, as shown in Fig. 2. This decision reflects the fact that certain prompts can lead to multiple plausible outcomes, driven by hierarchical relations (e.g., pick up the chips - snack - food) or by the inherent uncertainty of what action best satisfies the speaker's goal (e.g., an indirect prompt complaining about darkness could be addressed by either turning on a light or opening curtains). Exploring all possible valid interpretations is labor-intensive and often infeasible. Consequently, each MID instance is presented as four distinct options, one correct and three intentionally designed to be incorrect for different reasons (see Sec. 4.2.3), challenging models in both linguistic and visual reasoning." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": "Formally, let " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": " be the set of all multiple-choice options " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "c_{n}" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": ". Given an image " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": " and an indirect prompt " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": ", the task is to select the most valid interpretation of " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": " from the predefined options, conditioned on the visual context in " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": ". We define this task as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 351, + 426, + 555, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 426, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 351, + 426, + 555, + 440 + ], + "type": "interline_equation", + "content": "T (I, p _ {i}) := \\operatorname {a r g m a x} _ {c _ {n} \\in C} \\Pr \\left(c _ {n} \\mid I, p _ {i}\\right). \\tag {1}", + "image_path": "30fad751d875947069a6c9c3b17c621a9a1bbbd693af8454b941a398b08b414a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 449, + 486, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 486, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 486, + 461 + ], + "type": "text", + "content": "3.2. Direct and Indirect Expressions" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "content": "By the design of our task, curating effective input prompts " + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 467, + 554, + 538 + ], + "type": "text", + "content": " is crucial for ensuring accurate interpretation. What makes a good prompt, though? In this section, we define and explain the criteria that both direct and indirect expressions must satisfy. For more details on good and bad examples for each criterion, please refer to our Appendix A." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 548, + 503, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 503, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 503, + 559 + ], + "type": "text", + "content": "3.2.1. Directness: Relevance and Solvability" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 563, + 555, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 719 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 719 + ], + "type": "text", + "content": "Relevance The direct prompt is an utterance from the speaker that explicitly conveys its intended meaning without ambiguity. However, it is equally important that this intention aligns with the visual context of the scene. For example, a direct prompt " + }, + { + "bbox": [ + 313, + 563, + 555, + 719 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 563, + 555, + 719 + ], + "type": "text", + "content": " such as \"Hey person1, I want you to stop the fireworks\" clearly expresses its intended action. However, if the corresponding image, as shown in Fig. 2, contains no elements related to fireworks, the prompt is misaligned with the scene. Thus, a direct prompt must not only reveal its intention but also maintain relevance to the image. In the context of our task, relevance is determined by whether a human can reasonably establish a connection between the prompt and the depicted scene." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1539" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "Solvability Relevance alone does not guarantee that a prompt is useful. As outlined in Sec. 3.1, the prompt " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " must explicitly request an action that the listener can reasonably perform. This introduces solvability, which requires that the prompt present a clear and actionable problem. A solvable prompt defines a specific issue that can be addressed independently, ensuring that the listener is not left with multiple competing actions to choose from." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 260, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 260, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 260, + 192 + ], + "type": "text", + "content": "3.2.2. Indirectness: Consistency and Ambiguity" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 339 + ], + "type": "text", + "content": "Consistency Indirect prompts are designed to obscure their true intention, but they must still convey the same underlying intention as their direct counterpart—essentially requesting the same solution. Since indirect prompts are derived from direct prompts, consistency serves as a key criterion. We define a direct prompt " + }, + { + "bbox": [ + 55, + 194, + 295, + 339 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 194, + 295, + 339 + ], + "type": "text", + "content": " and an indirect prompt " + }, + { + "bbox": [ + 55, + 194, + 295, + 339 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 194, + 295, + 339 + ], + "type": "text", + "content": " as consistent if their interpretations could potentially align in intention. The term \"potentially\" is used because indirect prompts, by nature, may have multiple valid interpretations. For instance, in the earlier \"this is a disaster!\" example, it conveys distress but allows for multiple reasonable responses, such as cleaning the room or providing reassurance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 485 + ], + "type": "text", + "content": "Ambiguity If an indirect prompt is entirely consistent with its direct counterpart without introducing any additional complexity, it becomes indistinguishable from a direct prompt. Therefore, an indirect prompt should conceal its underlying intention, which we define as ambiguity. The key principle behind this criterion is that neither the specific action required nor the key entity involved should be explicitly or implicitly mentioned within the prompt. Once these two elements are concealed, further refinements—such as adjusting the tone to be more indirect, sarcastic, or humorous—can enhance the overall nuance and difficulty of interpretation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 500, + 246, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 246, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 246, + 513 + ], + "type": "text", + "content": "4. VAGUE Benchmark Construction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "VAGUE is a novel benchmark that extends single-modal or simple multimodal ambiguity to more realistic domains and evaluates whether concurrent vision-language models can perform human-like reasoning with complex visual contexts. It comprises 1,677 images, with 1,144 sourced from the VCR [46] dataset and 533 from Ego4D [41], covering diverse contextual scenarios as well as real-world human interactions. On average, VAGUE contains seven objects and four people per image. Each image is paired with a direct expression " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", an indirect expression " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", and four multiple-choice answers, along with relevant meta-information. All textual components are generated using GPT-4o, then refined through extensive human rating, selection, and filtering, ensuring a carefully curated benchmark dataset for testing and advancing multimodal reasoning. We provide detailed benchmark statistics and a diversity analysis in Appendix B.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 437, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 437, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 437, + 83 + ], + "type": "text", + "content": "4.1. Visual Data Curation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 90, + 384, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 90, + 384, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 90, + 384, + 102 + ], + "type": "text", + "content": "4.1.1. Sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 105, + 555, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 555, + 226 + ], + "type": "text", + "content": "VCR [46] The VCR dataset consists of 110K movie scenes sourced from the Large Scale Movie Description Challenge [35] and YouTube clips. These images are curated based on an \"interestingness\" criterion [46], ensuring the presence of at least two people, which promotes interactive scenarios. To prevent redundancy in our dataset, we sample 10K images while carefully avoiding neighboring frames, as adjacent frames exhibit minimal variation. This selection process preserves the contextual diversity of the dataset while maintaining its focus on complex, multi-entity interactions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "content": "Ego4D [41] While VCR provides a wide range of contextual diversity, it often includes artificially composed settings that may not fully capture real-world interactions. To address this, we integrate frames from the Ego4D dataset, which offers a more naturalistic depiction of human interactions. We specifically leverage the AV (Audio-Visual), which indicate conversational exchanges between individuals, to ensure the presence of people in the selected frames. Similar to VCR, we avoid neighboring frames to maintain diversity and filter out heavily blurry images to enhance data quality. This process results in 888 candidate images from 94 videos, which serve as the basis for further text processing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 387, + 419, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 387, + 419, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 387, + 419, + 399 + ], + "type": "text", + "content": "4.1.2. Object Extraction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 556, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 522 + ], + "type": "text", + "content": "To ensure the complexity of the visual information, we extract a list of physical objects present in each image using a tagging model, RAM [49]. This step allows us to easily identify scenes with sufficient visual detail. In the case of VCR, many scenes are relatively simple, often containing only a few objects. Therefore, we sort VCR images by the number of detected objects and retained the top 4,000 as candidates for text processing, ensuring that our benchmark primarily consists of rich visual cues in contextually diverse scenes. Please refer to Appendix B.3 for more details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 530, + 414, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 530, + 414, + 541 + ], + "spans": [ + { + "bbox": [ + 314, + 530, + 414, + 541 + ], + "type": "text", + "content": "4.1.3. Person Indicator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "In our task, we assume that the speaker is outside the scene, viewing the image and talking to a person in it. However, identifying the addressee is not always straightforward, as images mostly contain multiple individuals. While grounding the specific person referenced in the utterance could introduce additional complexity, it is not the primary focus of our evaluation. To clarify the listener, we assign an indicator tag to each person in the image as shown in Fig. 3. For VCR, we use their existing annotations [46], while for Ego4D, where bounding boxes are not fully available, we employ YOLOv11 [18] to detect and annotate humans. While this provides a straightforward method for grounding the target person, it requires models to perform basic Optical Character Recognition (OCR). Therefore, we conduct exper" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1540" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 550, + 232 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 550, + 232 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 550, + 232 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 550, + 232 + ], + "type": "image", + "image_path": "9e1c2cc12c6b85ca62f868102b42281b9ec99a2fd7df1e4755d09e2cd23bc692.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 242, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 555, + 319 + ], + "type": "text", + "content": "Figure 3. Overview of the data generation process. Based on human-defined criteria and instructions, GPT-4o [31] generates initial data, which are then rated and filtered by humans to ensure quality. Since generating high-quality indirect expressions from raw images is challenging, the process follows these steps: generating a direct expression and intention from the image, creating an indirect expression using information from the previous step, and producing answer candidates based on all the information gathered so far. In the answer candidates, FS stands for Fake Scene Understanding, SU stands for Superficial Understanding, and NE stands for Nonexistent Entity. FS evaluates global hallucination, where the model misinterprets visual context from an entirely different image (e.g., an outdoor camping scene with a candle), while NE addresses local hallucination, where only a specific object is replaced with a fabricated one (e.g., curtains)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "text", + "content": "iments to evaluate the OCR capabilities of the models used in our study. See Appendix B.4 for details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 234, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 234, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 234, + 385 + ], + "type": "text", + "content": "4.2. Multimodal Expression Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": "As shown in Fig. 3, VAGUE's textual expressions are generated first by the model and undergo an extensive process of human rating and filtering. We use GPT-4o [31] for all text processing. The instructions used for generating direct " + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "inline_equation", + "content": "(p_d)" + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": " and indirect " + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "inline_equation", + "content": "(p_i)" + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": " expressions are provided in Appendix B.5, while the instructions for generating answer candidates for multiple-choice questions are detailed in Appendix B.7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 482, + 163, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 163, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 163, + 494 + ], + "type": "text", + "content": "4.2.1. Direct Expressions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "text", + "content": "The direct expression " + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "text", + "content": " serves as a crucial foundation for crafting the indirect one " + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "text", + "content": ", as both share the same underlying intention. To ensure that " + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "text", + "content": " adhere to the principles of relevance and solvability discussed in Sec. 3.2.1, we generate " + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "text", + "content": " conditioned on the input image " + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 498, + 296, + 640 + ], + "type": "text", + "content": " and a task prompt that explicitly defines these criteria. During the generation process, we also instruct the model to output a solution triplet in the format: (subject, action, object). Since direct expressions explicitly state their intentions, extracting each component of the solution triplet is straightforward. To maintain consistency with the visual context, the \"object\" in the triplet is restricted to physical objects we extracted in Sec. 4.1.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "After generating " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " for all candidate images, human raters evaluate each prompt based on relevance and solvability, assigning scores from 1 to 5. Only those prompts that receive a rating of 4 or 5 are retained for further use. The detailed rating criteria for human verification and an example of the rating process are provided in Fig. J13." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 341, + 430, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 341, + 430, + 352 + ], + "spans": [ + { + "bbox": [ + 314, + 341, + 430, + 352 + ], + "type": "text", + "content": "4.2.2. Indirect Expressions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "text", + "content": "To ensure that the indirect expressions " + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "text", + "content": " maintain both ambiguity and fluency while aligning with the true intent " + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 356, + 555, + 392 + ], + "type": "text", + "content": ", we adopt a two-stage process: proposal and selection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 393, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 548 + ], + "type": "text", + "content": "In the initial step, the model is prompted to generate three distinct candidate options. Each candidate follows the criteria outlined in Section 3.2.2, but with explicit instructions to incorporate different linguistic strategies: sarcasm, humor, and meme/idiomatic expressions, respectively. This approach ensures diversity in the generated responses. In the second step, human annotators evaluate the three candidates and select the one that best aligns with the intended indirectness. The selected prompt is then rated on a scale from 1 to 5 based on more specific criteria. Only those prompts that receive a score of 3 or higher are retained for use in the dataset. The detailed rating criteria for human verification and an example of rating process are provided in Fig. J14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 443, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 443, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 443, + 567 + ], + "type": "text", + "content": "4.2.3. Counterfactual Choices" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 571, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 555, + 619 + ], + "type": "text", + "content": "Generating high-quality counterfactual choices is crucial. To enable detailed analysis of model weaknesses in multimodal intent disambiguation, we design interpretable counterfactual choices that provide more plausible alternatives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Fake Scene Understanding The first counterfactual choice is an interpretation that could arise when the model largely misinterprets the image. This process is conducted in two steps. In the first step, a fake caption is generated by assuming an imaginary scene that can be aligned with the indirect expression but is inconsistent with the true intent. The caption of fake scene is then combined with the speaker's" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1541" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 287, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 287, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 287, + 85 + ], + "type": "text", + "content": "indirect statement to derive the most likely interpretation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 93, + 297, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 93, + 297, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 93, + 297, + 213 + ], + "type": "text", + "content": "Superficial Understanding The subsequent choice corresponds to an interpretation generated when the model fails to deeply reason about the implicit intent of the sentence and instead relies on surface-level meaning. We enforce the model to focus only on the literal wording, without considering any implied or deeper meaning of the indirect sentence. These answer choices are generated alongside the indirect expression " + }, + { + "bbox": [ + 55, + 93, + 297, + 213 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 93, + 297, + 213 + ], + "type": "text", + "content": ". During the indirect selection phase, the corresponding superficially understood choice is selected together, maintaining coherence between them." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 402 + ], + "type": "text", + "content": "Nonexistent Entity The last choice arises when the model interprets the text correctly but fails to adequately consider the details of the image, resulting in a plausible yet incorrect choice. This resembles the correct answer in structure, but replaces the key object in the solution with one that does not exist in the image. To prevent the task from becoming too easy by generating highly irrelevant objects, we constrain the substituted object to one that, while absent from the image, is highly expected to be present in the scene and could replace the original entity. To identify such entities, the model is provided with the image as input to choose objects that align with the scene's context. This method ensures that the counterfactual choice leverages the expected coherence between the scene and its potential entities while rigorously testing the model's attention to visual details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 414, + 136, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 136, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 136, + 427 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 458 + ], + "type": "text", + "content": "Models We use the following models in our experiments. The detailed descriptions of each model are in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 459, + 215, + 578 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 57, + 459, + 195, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 195, + 471 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 195, + 471 + ], + "type": "text", + "content": "- Phi3.5-Vision-Instruct (4B) [29]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 472, + 180, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 472, + 180, + 483 + ], + "spans": [ + { + "bbox": [ + 57, + 472, + 180, + 483 + ], + "type": "text", + "content": "- LLaVA Onevision (7B) [23]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 483, + 215, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 483, + 215, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 483, + 215, + 495 + ], + "type": "text", + "content": "Qwen2.5-VL-Instruct (7B, 72B) [43]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 495, + 201, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 495, + 201, + 506 + ], + "spans": [ + { + "bbox": [ + 57, + 495, + 201, + 506 + ], + "type": "text", + "content": "- InternVL-2.5-MPO (8B, 26B) [6]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 507, + 138, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 507, + 138, + 518 + ], + "spans": [ + { + "bbox": [ + 57, + 507, + 138, + 518 + ], + "type": "text", + "content": "- Idefics2 (8B) [22]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 519, + 199, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 519, + 199, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 519, + 199, + 530 + ], + "type": "text", + "content": "LLaVA NeXT Vicuna (13B) [24]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 531, + 135, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 135, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 135, + 542 + ], + "type": "text", + "content": "- Ovis2 (16B) [28]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 543, + 117, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 543, + 117, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 543, + 117, + 554 + ], + "type": "text", + "content": "GPT-4o [31]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 555, + 147, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 147, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 147, + 566 + ], + "type": "text", + "content": "Gemini 1.5 Pro [12]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 157, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 157, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 157, + 578 + ], + "type": "text", + "content": "- InternVL-3 (38B) [42]" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 587, + 237, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 237, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 237, + 599 + ], + "type": "text", + "content": "5.1. MLLMs Benefit from Visual Cues" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 605, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 700 + ], + "type": "text", + "content": "Our first objective is to assess how effectively MLLMs leverage visual cues to resolve ambiguity in utterances. To this end, we systematically control the level of detail in the visual cues provided to the models and measure their accuracy in inferring the speaker's true intent. Performance is evaluated in both multiple-choice and free-form settings. For clarity, we primarily report multiple-choice accuracy, deferring free-form results to Appendix I." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 701, + 227, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 227, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 227, + 712 + ], + "type": "text", + "content": "We consider three levels of visual cues:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 72, + 555, + 216 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 314, + 72, + 555, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 555, + 119 + ], + "type": "text", + "content": "- Language Models (LMs) receive no visual input, requiring models to rely on superficial textual priors such as common-sense knowledge of sarcasm or humor to determine intent." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 120, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 120, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 120, + 555, + 192 + ], + "type": "text", + "content": "- Socratic Models (SMs) [47] use text-only LMs but incorporate short image captions (up to two or three sentences) as additional input. This short generic caption may lack sufficient detail, which may be insufficient for accurately inferring intent. Each SM model generated its own image captions and used them in subsequent processing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 192, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 555, + 216 + ], + "type": "text", + "content": "- Visual Language Models (VLMs) receive the raw image input, enabling a more direct interpretation of visual cues." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 225, + 556, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 556, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 556, + 369 + ], + "type": "text", + "content": "Results The results in Tab. 1 indicate that MLLMs can leverage visual cues, albeit to a limited extent, since SMs and VLMs consistently outperform LMs across all evaluated models. Additionally, more detailed visual input generally improves performance, with VLMs surpassing SMs in most cases, except in the case of proprietary models. This exception is further analyzed in Sec. 5.2. Both the VCR and Ego4D subsets exhibit similar performance trends, demonstrating the generalizability of our findings across both staged and real-world scenarios. Finally, the consistently low performance of LMs further reinforces the validity of our dataset as a multimodal benchmark." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 377, + 462, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 462, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 462, + 389 + ], + "type": "text", + "content": "5.2. Analysis on Failure Modes" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 395, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 556, + 552 + ], + "type": "text", + "content": "Here, we examine the ways in which models fail to infer the true intent and how these failure patterns vary with the level of visual cues provided. As shown in Fig. 3, our multiple-choice questions include three distinct types of incorrect answer candidates. We assess the model's raw visual understanding using the Fake Scene Understanding (FS) and Nonexistent Entity (NE) candidates, which test whether the model can correctly interpret the scene without being misled by fabricated or nonexistent elements. Conversely, the Superficial Understanding (SU) candidate evaluates the model's reasoning ability, testing whether it can go beyond surface-level perception to infer intent. We provide the full table of failure modes in the MCQ setting in Appendix I" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "text", + "content": "Results Figure 4 illustrates how frequently each model selects different types of incorrect answers instead of the correct intent. Among the error types, Superficial Understanding (SU) is the most common. This indicates that while models generally succeed in recognizing basic visual details, they often fail to reason deeply about those visual cues to accurately infer the underlying intent of the speaker. However, proprietary models exhibit fewer SU-related errors, indicating stronger reasoning capabilities." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 666, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 713 + ], + "type": "text", + "content": "Moreover, stronger visual cues improve accuracy across all models and failure types. This improvement highlights the crucial role of visual conditioning in reducing both vision-based errors (FS and NE) and reasoning-related failures (SU)." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1542" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 267 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 267 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 267 + ], + "type": "table", + "html": "
ModelVAGUE-VCRVAGUE-Ego4D
LM (L)SM (L+V)VLM (L+V)LM (L)SM (L+V)VLM (L+V)
Phi3.5-Vision-Instruct (4B)26.635.3 (↑ 8.7)46.0 (↑ 19.4)22.531.1 (↑ 8.6)42.4 (↑ 19.9)
LLaVA-Onevision (7B)13.129.4 (↑ 16.3)43.1 (↑ 30.0)11.329.5 (↑ 18.2)43.2 (↑ 31.9)
Qwen2.5-VL-Instruct (7B)11.125.6 (↑ 14.5)46.8 (↑ 35.7)9.828.0 (↑ 18.2)48.4 (↑ 38.6)
InternVL-2.5-MPO (8B)23.048.4 (↑ 25.4)63.9 (↑ 40.9)24.254.0 (↑ 29.8)66.8 (↑ 42.6)
Idefics2 (8B)13.921.1 (↑ 7.2)58.7 (↑ 44.8)14.818.2 (↑ 3.4)58.3 (↑ 43.5)
LLaVA-NeXT-vicuna (13B)24.237.2 (↑ 13)46.4 (↑ 22.2)20.334.1 (↑ 13.8)52.5 (↑ 32.2)
Ovis2 (16B)21.923.8 (↑ 1.9)24.5 (↑ 3.6)20.525.3 (↑ 4.8)25.7 (↑ 5.2)
InternVL-2.5-MPO (26B)21.248.5 (↑ 27.3)63.7 (↑ 42.5)21.855.2 (↑ 33.4)68.7 (↑ 46.9)
InternVL-3 (38B)24.847.2 (↑ 22.4)63.6 (↑ 38.8)18.047.5 (↑ 29.5)59.8 (↑ 41.8)
Qwen2.5-VL-Instruct (72B)29.655.6 (↑ 26.0)74.2 (↑ 44.6)26.859.3 (↑ 32.5)69.8 (↑ 43.0)
GPT-4o46.469.5 (↑ 23.1)65.1 (↑ 18.7)48.267.5 (↑ 19.3)63.6 (↑ 15.3)
Gemini-1.5-Pro43.262.4 (↑ 19.2)60.6 (↑ 17.4)40.360.6 (↑ 20.3)60.6 (↑ 20.3)
", + "image_path": "62bf245532a064985b3cff71bbdd8486b026798d8780805b5785c577faafa5ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "type": "text", + "content": "Table 1. Experiments on the Multimodal Intention Disambiguation (MID) task with varying levels of visual cues. We report the accuracy " + }, + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "type": "text", + "content": " of the Multiple-Choice Question. " + }, + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "type": "text", + "content": " indicates the performance gain from visual cues, i.e. increment compared to the LM setting.(L) denotes the use of language input only, while " + }, + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "type": "inline_equation", + "content": "(\\mathrm{L} + \\mathrm{V})" + }, + { + "bbox": [ + 55, + 275, + 555, + 319 + ], + "type": "text", + "content": " indicates the incorporation of visual cues. The noticeable increase in accuracy across LM, SM, and VLM demonstrates that the introduction of detailed visual cues is beneficial for the task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 331, + 556, + 460 + ], + "blocks": [ + { + "bbox": [ + 56, + 331, + 556, + 460 + ], + "lines": [ + { + "bbox": [ + 56, + 331, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 556, + 460 + ], + "type": "image", + "image_path": "f063bf44d3c5ca75ea7e8831036cb2b6bff129c5ecbd71bdc54b08faf9ac3d72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 468, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 555, + 514 + ], + "type": "text", + "content": "Figure 4. We present a bar plot to analyze the distribution of incorrect answer choices selected by each model. Each number represents how frequently a given choice was selected from the 1,677 items in the dataset. The counterfactual choice categories are FS (Fake Scene Understanding), SU (Superficial Understanding), and NE (Nonexistent Entity). We use distinct colors to represent LM (Language Models), SM (Socratic Models), and VLM (Visual-Language Models)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 629 + ], + "type": "text", + "content": "Notably, proprietary models perform better with captioned inputs (SM) than with raw images (VLM). A closer examination of Fig. 4 reveals that this discrepancy arises from vision-based failures (FS and NE) rather than reasoning-centric failure (SU). This indicates that their captioning ability potentially allows them to obtain more sophisticated information while reducing hallucination in the images. The supporting experiments and explanations are provided in Appendix F." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 636, + 195, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 195, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 195, + 649 + ], + "type": "text", + "content": "5.3. Comparison with human" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "To validate our benchmark and establish an upper bound for performance, we assess human accuracy, highlighting the gap between existing models and human capability. This evaluation follows the multiple-choice setup within the VLM setting, using a subset of 400 samples. As shown" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": "in Tab. 2, human performance reaches " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": ", demonstrating near-perfect accuracy. Although proprietary models and some large-sized models show a decent performance, a notable performance gap (" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": ") still exists compared to human evaluators. This result underscores the significant gap between AI models' multimodal reasoning capabilities and human-level understanding when inferring hidden intent, suggesting that visual perception alone, even when accurate, is insufficient without deeper cognitive integration. This performance gap is due to the models' tendency to rely on surface-level text rather than understanding deeper visual-textual implications. Thus, advancing multimodal reasoning likely requires models to integrate higher-order cognitive processes, like commonsense reasoning and pragmatic understanding, into visual interpretation tasks. Refer" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1543" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 295, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 220 + ], + "type": "table", + "html": "
ModelAcc (%)FSSUNECorrect
Ovis2 (16B)23.01191711892
LLaVA-Onevision (7B)41.04318310164
Phi3.5-Vision-Instruct (4B)44.36013231177
Qwen2.5-VL-Instruct (7B)47.04715213188
LLaVA-NeXT-icuna (13B)48.04814317192
Idefics2 (8b)57.02812024228
Gemini-1.5-Pro60.3607326241
InternVL-3 (38B)61.5389422246
InternVL-2.5-MPO (8B)61.8429516247
GPT-4o62.3616327249
InternVL-2.5-MPO (26B)63.03610111252
Qwen2.5-VL-Instruct (72B)72.3396012289
Human94.01248374
", + "image_path": "54538f880812ae8c5bba8b5ff31d0c85d4580b2bec406d2a9e499f683d40dfea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 274, + 293, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "type": "text", + "content": "Table 2. Performance across models and humans on a sampled set of 400 questions. The results show that humans outperform models by a margin of over " + }, + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 274, + 293, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 293, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 293, + 407 + ], + "type": "table", + "html": "
ModelTypeAcc (%)Incorrect count
FSSUNE
GPT-4oSM68.9165240117
SM+CoT69.5 (↑ 0.6)165241105
VLM64.6226226141
VLM+CoT66.4 (↑ 1.8)16215685
Gemini-1.5-ProSM61.817637483
SM+CoT61.0 (↓ 0.8)19036794
VLM60.6249264145
VLM+CoT64.4 (↑ 3.8)213267117
", + "image_path": "1c4c73acf4369df3333c7bbaad8ef0c6b89d6f8a77aa4e034ea3936ac1d2dce3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 417, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 452 + ], + "type": "text", + "content": "Table 3. Result of Chain-of-Thought (CoT) experiments on proprietary models, in both SM and VLM settings. " + }, + { + "bbox": [ + 55, + 417, + 295, + 452 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 417, + 295, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 417, + 295, + 452 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 417, + 295, + 452 + ], + "type": "text", + "content": " indicate an increase and decrease in accuracy when zero-shot CoT is applied." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 498 + ], + "type": "text", + "content": "to Appendix E for details on the selected subset and human evaluation setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 227, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 227, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 227, + 519 + ], + "type": "text", + "content": "5.4. Chain-of-Thought Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 524, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 295, + 644 + ], + "type": "text", + "content": "Given the strong reasoning demands of multimodal intent deduction, we further explore the effectiveness of Chain-of-Thought (CoT) prompting [44] in enhancing the reasoning capabilities of MLLMs. The CoT prompt templates, provided in Fig. J21 and Fig. J22, are designed to explicitly ground the reasoning process, reducing hallucinations. While our main results focus on proprietary models, owing to their superior suitability for zero-shot CoT reasoning, we also present experiments and analyses for open-source models ranging from 4B to 72B in Appendix G." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "content": "Results As shown in Tab. 3, CoT prompting improves performance for raw image inputs (VLM), while showing no clear trend and remaining at a similar level for image caption inputs (SM). One possible explanation for this discrepancy is that CoT primarily enhances reasoning by improving" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "content": "grounding and reducing hallucinations. Since image captions inherently contain fewer hallucinations, SMs see some benefit from CoT prompting albeit at the cost of reduced detail. Additionally, the performance improvements observed with CoT prompting are consistent across different types of false answer candidates, suggesting a generalizable effect in enhancing reasoning quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 167, + 388, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 388, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 388, + 179 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 556, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 556, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 556, + 415 + ], + "type": "text", + "content": "We present VAGUE (Visual Contexts CIArify ambiGUous Expressions), a new benchmark aimed at assessing models' ability to interpret nuanced communication in complex multimodal scenarios. Our results show that models benefit from visual information when inferring the underlying intention of indirect expressions, as evidenced by their improved performance with increasing levels of visual cues. However, a significant disparity persists between machine capabilities and humans. To gain deeper insights into model inaccuracies, we design multiple-choice questions that explicitly address failure points, enabling a systematic and quantitative evaluation of the reasons behind performance. The primary challenge identified is the tendency of multimodal models to inadequately integrate visual cues, relying instead on the literal interpretation of textual information. This shortcoming highlights the need for further research, and we anticipate that VAGUE will open up promising avenues for developing systems capable of deeper multimodal reasoning to enhance AI's ability to engage in human-like interactions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 425, + 389, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 389, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 389, + 437 + ], + "type": "text", + "content": "7. Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 445, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 696 + ], + "type": "text", + "content": "First, we acknowledge that there are cultural and linguistic limitations. We incorporate sarcastic, humorous, and idiomatic implicatures in generating indirect expressions. However, since the initial data drafts are created by a model (GPT-4o [31]), they may reflect cultural biases present in its training data. To prevent any potential ethical issues, all human annotators are instructed to remove any content considered problematic or discriminatory during the rating and filtering process. Also, all textual expressions in VAGUE are limited to English. Therefore, we encourage future exploration of indirect expressions across diverse languages and cultures. The second limitation pertains to the dependency of certain meta-information on the quality of the parent dataset and the performance of the models utilized during the dataset generation. We observe that bounding box annotations from YOLOv11 [18] are occasionally duplicated, leading to a reported number of people higher than actually present. Likewise, the tagging model RAM [49] sometimes misidentified objects. Therefore, we remove any corrupted instances that could undermine the integrity of the task during the human rating and filtering process." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1544" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Ashwin Baluja. Text is not all you need: Multimodal prompting helps llms understand humor. arXiv preprint arXiv:2412.05315, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "text", + "content": "[2] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 236 + ], + "type": "text", + "content": "[3] Matteo Bortoletto, Constantin Ruhdorfer, Lei Shi, and Andreas Bulling. Explicit modelling of theory of mind for belief prediction in nonverbal social interactions. In Proc. 27th European Conference on Artificial Intelligence (ECAI), pages 1-8, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 296, + 304 + ], + "type": "text", + "content": "[4] Tuhin Chakrabarty, Arkadiy Saakyan, Debanjan Ghosh, and Smaranda Muresan. FLUTE: Figurative language understanding through textual explanations. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7139-7159, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "type": "text", + "content": "[5] Zhawnen Chen, Tianchun Wang, Yizhou Wang, Michal Kosinski, Xiang Zhang, Yun Fu, and Sheng Li. Through the theory of mind's eye: Reading minds with multimodal video large language models. arXiv preprint arXiv:2406.13763, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "type": "text", + "content": "[6] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 441 + ], + "type": "text", + "content": "[7] Christine Cuskley, Rebecca Woods, and Molly Flaherty. The limitations of large language models for understanding human language and cognition. Open Mind, 8:1058-1083, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 443, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 509 + ], + "type": "text", + "content": "[8] Hang Du, Guoshun Nan, Sicheng Zhang, Binzhu Xie, Junrui Xu, Hehe Fan, Qimei Cui, Xiaofeng Tao, and Xudong Jiang. Docmsu: A comprehensive benchmark for document-level multimodal sarcasm understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17933-17941, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 553 + ], + "type": "text", + "content": "[9] Jiafei Duan, Samson Yu, Nicholas Tan, Li Yi, and Cheston Tan. Boss: A benchmark for human belief prediction in object-context scenarios. arXiv preprint arXiv:2206.10665, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 296, + 610 + ], + "type": "text", + "content": "[10] Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. CVPR 2024, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 677 + ], + "type": "text", + "content": "[11] Kanishk Gandhi, Jan-Philipp Fraenkel, Tobias Gerstenberg, and Noah D. Goodman. Understanding social reasoning in language models with language models. In Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track, 2023. Dataset and Benchmarks Track. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "text", + "content": "[12] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 6, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 172 + ], + "type": "text", + "content": "[13] Md Kamrul Hasan, Wasifur Rahman, AmirAli Bagher Zadeh, Jianyuan Zhong, Md Iftekhar Tanveer, Louis-Philippe Morency, and Mohammed (Ehsan) Hoque. UR-FUNNY: A multimodal language dataset for understanding humor. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2046–2056, Hong Kong, China, 2019. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 174, + 555, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 555, + 261 + ], + "type": "text", + "content": "[14] Jack Hessel, Ana Marasovic, Jena D. Hwang, Lillian Lee, Jeff Da, Rowan Zellers, Robert Mankoff, and Yejin Choi. Do androids laugh at electric sheep? humor \"understanding\" benchmarks from the new yorker caption contest. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 688-714, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 263, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 263, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 263, + 555, + 295 + ], + "type": "text", + "content": "[15] EunJeong Hwang and Vered Shwartz. Memecap: A dataset for captioning and interpreting memes. arXiv preprint arXiv:2305.13703, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 297, + 555, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 555, + 351 + ], + "type": "text", + "content": "[16] Prince Jha, Krishanu Maity, Raghav Jain, Apoorv Verma, Sriparna Saha, and Pushpak Bhattacharyya. Meme-ingful analysis: Enhanced understanding of cyberbullying in memes through multimodal explanations. arXiv preprint arXiv:2401.09899, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 354, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 555, + 441 + ], + "type": "text", + "content": "[17] Chuanyang Jin, Yutong Wu, Jing Cao, Jiannan Xiang, Yen Ling Kuo, Zhiting Hu, Tomer Ullman, Antonio Torralba, Joshua Tenenbaum, and Tianmin Shu. MMToM-QA: Multimodal theory of mind question answering. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16077-16102, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 444, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 554, + 464 + ], + "type": "text", + "content": "[18] Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements, 2024. 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 467, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 544 + ], + "type": "text", + "content": "[19] Shivani Kumar, Atharva Kulkarni, Md Shad Akhtar, and Tan-moy Chakraborty. When did you become so smart, oh wise one?! sarcasm explanation in multi-modal multi-party dialogues. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5956-5968, Dublin, Ireland, 2022. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 545, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 555, + 632 + ], + "type": "text", + "content": "[20] Shivani Kumar, Ishani Mondai, Md Shad Akhtar, and Tanmoy Chakraborty. Explaining (sarcastic) utterances to enhance affect understanding in multimodal dialogues. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence. AAAI Press, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 689 + ], + "type": "text", + "content": "[21] Yash Kumar Lal and Mohaddeseh Bastan. SBU figures it out: Models explain figurative language. In Proceedings of the 3rd Workshop on Figurative Language Processing (FLP), pages 143-149, Abu Dhabi, United Arab Emirates (Hybrid), 2022. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[22] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language mod" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1545" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "els? Advances in Neural Information Processing Systems, 37: 87874-87907, 2025. 6, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 296, + 140 + ], + "type": "text", + "content": "[23] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "text", + "content": "[24] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 6, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 294, + 220 + ], + "type": "text", + "content": "[25] Kevin Y. Li, Sachin Goyal, Joao D. Semedo, and J. Zico Kolter. Inference optimal vlms need fewer visual tokens and more parameters, 2025. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 294, + 277 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision – ECCV 2014, pages 740–755, Cham, 2014. Springer International Publishing. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 334 + ], + "type": "text", + "content": "[27] Annie Louis, Dan Roth, and Filip Radlinski. \"I'd rather just go to bed\": Understanding indirect answers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7411-7425, Online, 2020. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 336, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 294, + 380 + ], + "type": "text", + "content": "[28] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 6, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 404 + ], + "type": "text", + "content": "[29] Microsoft. Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 406, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 472 + ], + "type": "text", + "content": "[30] Maryam Sadat Mirzaei, Kourosh Meshgi, and Satoshi Sekine What is the real intention behind this question? dataset collection and intention classification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13606-13622, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 474, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 495 + ], + "type": "text", + "content": "[31] OpenAI. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5, 6, 8, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 498, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 529 + ], + "type": "text", + "content": "[32] Kun Ouyang, Liqiang Jing, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Sentiment-enhanced graph-based sarcasm explanation in dialogue. arXiv:2402.11414, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 597 + ], + "type": "text", + "content": "[33] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "text", + "content": "[34] David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences, 1(4): 515-526, 1978. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "text", + "content": "[35] Anna Rohrbach, Atousa Torabi, Marcus Rohrbach, Niket Tandon, Chris Pal, Hugo Larochelle, Aaron Courville, and Bernt Schiele. Movie description. International Journal of Computer Vision, 2017. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[36] Melanie Sclar, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia Tsvetkov. Minding language models' (lack of) theory of mind: A plug-and-play multi-character belief" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), pages 13960-13980, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 149 + ], + "type": "text", + "content": "[37] Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin, Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. Muma-tom: Multi-modal multi-agent theory of mind, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 151, + 555, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 555, + 226 + ], + "type": "text", + "content": "[38] Settaluri Sravanthi, Meet Doshi, Pavan Tankala, Rudra Murthy, Raj Dabre, and Pushpak Bhattacharyya. PUB: A pragmatics understanding benchmark for assessing LLMs' pragmatics capabilities. In Findings of the Association for Computational Linguistics: ACL 2024, pages 12075-12097, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "text", + "content": "[39] Chuandong Su, Fumiyo Fukumoto, Xiaoxi Huang, Jiyi Li, Rongbo Wang, and Zhiqun Chen. DeepMet: A reading comprehension paradigm for token-level metaphor detection. In Proceedings of the Second Workshop on Figurative Language Processing, pages 30-39, Online, 2020. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 358 + ], + "type": "text", + "content": "[40] Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase. DIRECT: Direct and indirect responses in conversational text corpus. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1980-1989, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "type": "text", + "content": "[41] Ego4d Team. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 555, + 436 + ], + "type": "text", + "content": "[42] OpenGVLab Team. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, 2025. 6, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 437, + 528, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 528, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 528, + 448 + ], + "type": "text", + "content": "[43] Qwen Team. Qwen2.5-v1 technical report, 2025. 6, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "type": "text", + "content": "[44] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "type": "text", + "content": "[45] Alex Wilf, Leena Mathur, Sheryl Mathew, Claire Ko, Youssof Kebe, Paul Pu Liang, and Louis-Philippe Morency. Social-iq 2.0 challenge: Benchmarking multimodal social understanding. https://github.com/abwilf/Social-IQ-2.0-Challenge, 2023.2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "text", + "content": "[46] Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual commonsense reasoning. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6713-6724, 2018. 2, 4, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 604, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 669 + ], + "type": "text", + "content": "[47] Andy Zeng, Maria Attarian, Krzysztof Marcin Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael S Ryoo, Vikas Sindhwani, Johnny Lee, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In The Eleventh International Conference on Learning Representations, 2023. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[48] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2020. 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1546" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 140 + ], + "type": "text", + "content": "[49] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1724-1732, 2024. 4, 8, 1" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1547" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_content_list.json b/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aca89923b2b213f2ceb979aeabf3c8d50075f23f --- /dev/null +++ b/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_content_list.json @@ -0,0 +1,1825 @@ +[ + { + "type": "text", + "text": "VALLR: Visual ASR Language Model for Lip Reading", + "text_level": 1, + "bbox": [ + 222, + 130, + 776, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marshall Thomas Edward Fish Richard Bowden University of Surrey", + "bbox": [ + 287, + 181, + 709, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mt00893@surrey.ac.uk edward.fish@surrey.ac.uk r.bowden@surrey.ac.uk", + "bbox": [ + 181, + 219, + 805, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 256, + 326, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), is a complex task requiring the interpretation of spoken language exclusively from visual cues, primarily lip movements and facial expressions. This task is especially challenging due to the absence of auditory information and the inherent ambiguity when visually distinguishing phonemes that have overlapping visemes, where different phonemes appear identical on the lips. Current methods typically attempt to predict words or characters directly from these visual cues, but this approach frequently encounters high error rates due to coarticulation effects and viseme ambiguity. We propose a novel two-stage, phoneme-centric framework for Visual Automatic Speech Recognition (V-ASR) that addresses these longstanding challenges. First, our model predicts a compact sequence of phonemes from visual inputs using a Video Transformer with a CTC head, thereby reducing the task complexity and achieving robust speaker invariance. This phoneme output then serves as the input to a fine-tuned Large Language Model (LLM), which reconstructs coherent words and sentences by leveraging broader linguistic context. Unlike existing methods that either predict words directly or rely on large-scale multimodal pre-training, our approach explicitly encodes intermediate linguistic structure while remaining highly data efficient. We demonstrate state-of-the-art performance on two challenging datasets, LRS2 and LRS3, where our method achieves significant reductions in Word Error Rate (WER) achieving a SOTA WER of 18.7 on LRS3 despite using $99.4\\%$ less labelled video data than the next best approach. Code is available here: https://gitlab.surrey.ac.uk/", + "bbox": [ + 86, + 297, + 485, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 800, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), involves interpreting spoken language from visual cues such as lip movements and facial expressions. As a natural skill, humans use it to supplement auditory information, and as a technology, it has profound potential for en", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ae34c4388a4f19abe6b22865a145bffff65591cbe287bfbeb007bdb92f19060.jpg", + "image_caption": [ + "Figure 1. Comparison between different models' performances in WER for Visual Automatic Speech Recognition on the LRS3 dataset [3] when compared with the amount of labelled training data. Circle size (green) denotes scale of pre-training data, while fine-tuned models (grey) are not pre-trained. Our model (orange) outperforms all existing approaches with just 30 hours of video training data, no self-supervised pre-training, and without the requirement for additional labeled visual data during fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 521, + 258, + 913, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hancing accessibility, particularly for the Deaf and hard-of-hearing communities, as well as for applications in noisy or privacy-sensitive environments [1]. Despite substantial advancements in ASR, the visual-only interpretation of speech remains an unresolved challenge due to inherent uncertainties in lip movements and their complex temporal dynamics. A central difficulty in V-ASR stems from the inherent ambiguity of visemes, the visual equivalents of phonemes, which often appear nearly identical for different sounds (e.g., 'p' and 'b') [1, 7]. Further complicating this task are coarticulation effects [33], where adjacent sounds blur one another's articulation, making phoneme boundaries visually unclear. Although some recent methods have focussed on robust intermediary representations to deal with these cases, such as leveraging subword-level predictions [41], or quantized latent representations [13], these approaches can still fail to capture broader contextual cues, limiting their ability to resolve visually similar phonemes that might otherwise be", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2846", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85b1f3f5dae7dfb135fece6cf0a7069536516d7e82e8e5024ee5d2691a090d08.jpg", + "image_caption": [ + "Figure 2. An overview of our approach. First we extract facial regions from 16 frames of input video. We apply random pixel masking at $50\\%$ probability, add positional embedding, and encode visual features via a vision transformer encoder (ViT-base [14]). We implement temporal downsampling via 1D convolution and then use a CTC linear head to predict sequences of phonemes. During training, we also fine-tune an LLM to reconstruct sentences from phonemes with a text-only dataset [52]. During inference, the phonemes from the CTC head are processed via the LLM to reconstruct the predicted text. This can be performed end-to-end or in two stages, depending on available resources." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 906, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distinguishable through sentence-level semantics.", + "bbox": [ + 89, + 463, + 416, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One approach to this challenge is to leverage large language models (LLM's) and attempt to map raw lip movements directly to sentences, thus bypassing the issue of viseme ambiguity [58, 59]. However, bridging the gap between high-dimensional, unstructured visual inputs and detailed textual representations is resource intensive and typically requires huge datasets, or specialized architectures. Moreover, these purely end-to-end pipelines often lack an interpretable intermediary step, making it unclear how the model is handling viseme ambiguity at a finer linguistic granularity. This can lead to hallucination effects, which become more critical in lip reading technologies designed for accessibility, or security applications.", + "bbox": [ + 88, + 493, + 482, + 690 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our two-stage, phoneme-centric method addresses both issues. In the first stage, we map short windows of video frames to a discrete, interpretable phoneme sequence. This intermediate representation is significantly easier to learn and more robust against speaker or environmental variations, as phonemes abstract away speaker-specific attributes. Subsequently, in the second stage, we fine-tune an LLM for the task of reconstructing sentences from phonemes. By segmenting the problem, we combine the interpretability and efficiency of phoneme-based prediction with the sophisticated contextual reasoning of modern LLMs. This phoneme $\\rightarrow$ sentence pipeline also aligns with psycholinguistic models of speech perception, where", + "bbox": [ + 88, + 703, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "phoneme-level processing naturally precedes lexical access [31, 32].", + "bbox": [ + 511, + 463, + 906, + 493 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several advantages emerge from this design:", + "bbox": [ + 511, + 501, + 808, + 516 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Compact Target Space: Predicting only 38 phoneme classes avoids learning entire word-level vocabularies, making the model more data-efficient and simpler to train.", + "(ii) LLM-Enhanced Accuracy: A large language model, pre-trained for phoneme $\\rightarrow$ sentence reconstruction, removes the requirement of direct word-level predictions from RGB data, reducing the need for complex multimodal training.", + "(iii) Data Availability: Exploiting easily available phoneme $\\rightarrow$ sentence text pairs eliminates reliance on extensive lip-reading video data for pre-training.", + "(iv) Interpretability & Error Analysis: The two-stage design yields an explicit phoneme-level output, constraining recognition errors to local phoneme segments, which the LLM subsequently repairs at the word level reducing errors at a sentence level." + ], + "bbox": [ + 496, + 523, + 906, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 792, + 653, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we review key categories of approaches and highlight their advantages and limitations in the context of our proposed phoneme-centric, language-aware approach.", + "bbox": [ + 511, + 816, + 906, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lip Reading: Early methods predominantly relied on statistical models such as Hidden Markov Models (HMMs)", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2847", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to aggregate temporal sequences of lip movements alongside hand-crafted feature extractors [1, 9, 35, 36, 40, 62]. With the advent of deep learning, coupled with the availability of large-scale datasets such as Grid [12], LRW [11], LRS2 [47], and LRS3 [3], significant progress has been made in recent years. Initially, Convolutional Neural Networks (CNNs) were applied to the task of word recognition [11], with additional temporal processing via RNN's [30, 54, 55]. With greater compute came the possibility to interpret full sentences with approaches applying Automatic Speech Recognition (ASR) methodologies to the Visual Automatic Speech Recognition (V-ASR) task. These included sequence-to-sequence models [2], CTC approaches [5, 46], and hybrid models [37, 38]. These models improved recognition accuracy by learning hierarchical representations of visual features and modelling temporal dependencies. The application of transformer models [14, 21, 44, 50, 53] in V-ASR brought further performance gains in combination with extensive self-supervised pretraining. Methods such as AV-HuBERT [45] and LiteVSR [24] leveraged self-supervised learning to learn robust representations from large amounts of unlabelled video data.", + "bbox": [ + 89, + 90, + 483, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AV-ASR vs V-ASR: Lip reading approaches can be separated into Visual Automatic Speech Recognition (V-ASR) and Audio-Visual Automatic Speech Recognition (AV-ASR). In the AV-ASR task, labelled audio data is available at training time which can be used to guide the visual encoder in various ways. These can include generating synthetic additional pre-training data [25, 27], or simply fusing audio-visual input features [44-46, 57]. Other methods have used knowledge distillation to inductively transfer knowledge from pre-trained ASR models to visual encoders [4, 28]. However, these methods depend on audio data or pre-trained ASR models, which may not be available or reliable in scenarios where lip reading is most needed, such as noisy environments. V-ASR, on the other hand, is a more challenging task that relies only on visual data in the training stage [1, 56, 60, 62]. Recent methods have also shown the effectiveness of adapting visual inputs to pre-trained ASR models [13, 24, 42] without the need for audio data during training. This method is particularly powerful since the ASR model already has the ability to reconstruct latent representations to words from the pre-training task, however it relies on large amounts of additional data to learn robust visual representations and map them to the audio latent space [42].", + "bbox": [ + 89, + 435, + 483, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lip Reading with LLMs: Recent approaches have explored visual attention mechanisms for sub-word units [15, 39, 41, 48] to capture fine-grained linguistic features in lip reading. However, reconstructing these units into coherent sentences remains challenging for networks lacking robust contextual and semantic reasoning capabilities. Early", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "phoneme-based methods [16, 49] similarly struggled with effective decoding to word-level representations and these methods were outpaced in performance by end-to-end approaches [5] which could leverage extensive data. As such they have not been fully explored in the context of modern architectures and techniques.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For example, recent approaches which have integrated Large Language Models (LLMs) [43, 51] into lip-reading pipelines [58, 59] have started from mapping visual features directly to LLM text embeddings. However, by omitting explicit intermediate representations, these approaches inherit phonetic ambiguities that propagate through the network. This manifests as increased word error rates and reduced robustness, thus requiring substantially more training data to achieve generalization.", + "bbox": [ + 511, + 189, + 906, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work re-examines phonemes as an interpretable discrete representation bridging visual and linguistic domains. Rather than forcing LLMs to perform visual $\\rightarrow$ sentence translation requiring expensive visual-text alignment – we constrain the LLM's role to phoneme $\\rightarrow$ sentence reconstruction. This decomposition offers three key advantages: (1) Phoneme-to-text translation is a well-constrained task requiring only lightweight LLM fine-tuning using abundant textual corpora; (2) Speaker-independent phoneme representations eliminate the need for visual adaptation in the language model; (3) The modular architecture prevents error propagation between visual and linguistic domains while maintaining interpretability.", + "bbox": [ + 511, + 334, + 908, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 544, + 604, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our two-stage phoneme-centric approach to visual-only lip reading. Figure 2 provides an overview of the framework. Our goal is to learn a function", + "bbox": [ + 511, + 569, + 905, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f: \\quad \\mathbf {X} = \\left\\{x _ {1}, x _ {2}, \\dots , x _ {T} \\right\\} \\\\ \\longmapsto \\mathbf {P h} = \\left\\{p h _ {1}, p h _ {2}, \\dots , p h _ {m} \\right\\} \\tag {1} \\\\ \\longmapsto \\mathbf {S} = \\left\\{s _ {1}, s _ {2}, \\dots , s _ {M} \\right\\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 627, + 905, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ is a video sequence of $T$ frames (with height $H$ and width $W$ ), $\\mathbf{Ph}$ denotes the phoneme sequence, and $\\mathbf{S}$ is the reconstructed sentence. Specifically, we decompose the task into:", + "bbox": [ + 511, + 693, + 905, + 755 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Video $\\rightarrow$ Phoneme: Map the sequence of video frames to phonemes using a Vision Transformer and CTC head.", + "2. Phoneme $\\rightarrow$ Sentence: Convert phonemes to a coherent word sequence with a fine-tuned Large Language Model (LLM)." + ], + "bbox": [ + 517, + 763, + 905, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model comprises four primary components:", + "bbox": [ + 511, + 847, + 834, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Visual Feature Extractor: Captures relevant features from each frame via a Vision Transformer.", + "bbox": [ + 511, + 869, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2848", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Adapter Network with Temporal Downsampling: Reduces sequence length to accommodate CTC alignment.", + "- CTC Head: Predicts phoneme sequences without requiring explicit temporal labels.", + "- LLM for Phoneme-to-Sentence Reconstruction: Translates predicted phonemes into final word sequences." + ], + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Pre-Processing Pipeline", + "text_level": 1, + "bbox": [ + 89, + 191, + 307, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each frame $x_{i}$ is first passed through a face-detection model [26] to identify and crop the speaker's face region, centered on the speaker's lips. The detected region is then cropped, resized to $224 \\times 224$ , and normalized for batch processing.", + "bbox": [ + 89, + 214, + 480, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Video Transformer", + "text_level": 1, + "bbox": [ + 89, + 299, + 272, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a Vision Transformer (ViT) [44] to encode the spatio-temporal information in the face region. Let ViT(\\cdot) denote this encoding function. The input $\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ is chunked into patches, embedded, and equipped with positional encodings before passing through multiple transformer blocks. We obtain:", + "bbox": [ + 89, + 321, + 480, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\operatorname {V i T} (\\mathbf {X}) \\in \\mathbb {R} ^ {T \\times D}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 422, + 480, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D$ is the transformer output dimension per frame. During training, we randomly mask portions of the input patches to promote robust feature learning and improve generalization.", + "bbox": [ + 89, + 454, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adapter Network with Temporal Downsampling", + "text_level": 1, + "bbox": [ + 89, + 523, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sequence length $T$ of the ViT output can be prohibitively large for the subsequent CTC module. To reduce the temporal dimension, we apply a sequence of 1D convolutions and pooling layers, as shown in Table ?? Formally,", + "bbox": [ + 89, + 561, + 482, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {\\text {d o w n}} = \\operatorname {A d a p t} (\\mathbf {Z}) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C _ {\\text {a d p a t e r}}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 633, + 480, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T' < T$ and $C_{\\mathrm{adapter}}$ is an intermediate feature dimension aligned to the CTC head input.", + "bbox": [ + 89, + 665, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. CTC Head", + "text_level": 1, + "bbox": [ + 89, + 705, + 210, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An MLP with one hidden layer transforms $\\mathbf{G}_{\\mathrm{down}} \\in \\mathbb{R}^{T' \\times C_{\\mathrm{adapter}}} \\text{ into phoneme logits:}$", + "bbox": [ + 89, + 728, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {\\mathrm {c t c}} = \\operatorname {M L P} \\left(\\mathbf {G} _ {\\text {d o w n}}\\right) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times N _ {\\text {p h o n e m e s}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 770, + 480, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{\\text{phonemes}}$ is the size of our phoneme set (39 English phonemes plus a blank symbol). We then apply a logarithmic softmax along the phoneme dimension to obtain log probabilities $\\log p_{\\text{ctc}}(ph_t \\mid \\mathbf{X})$ at each time step.", + "bbox": [ + 89, + 801, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since lip movements do not align cleanly with discrete phoneme boundaries, we adopt the Connectionist Temporal", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1b981d1a719a18feb087fe2df03f3b6b59d1263151fbcdd9f66245eeb8421ddb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StageModuleIn → Out ChannelsKernelStridePaddingDownsample FactorOutput Length
1Conv1dfeature_size → adapter_dim522×2L/2
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
2Conv1dadapter_dim → adapter_dim321×2L/4
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
3Conv1dadapter_dim → adapter_dim321×2L/8
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
4Conv1dadapter_dim → adapter_dim331×3L/24
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
5AvgPool1dadapter_dim → adapter_dim580×8L/192
6Linearadapter_dim → CTC_dim---×1L/192
ReLU----
", + "bbox": [ + 516, + 88, + 923, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Architecture of the adapter for temporal downsampling, where the input length is $T = 1568$ and final output length is $T = 8$ .", + "bbox": [ + 511, + 239, + 903, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classification (CTC) loss [18]:", + "bbox": [ + 513, + 305, + 718, + 320 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C T C}} = - \\ln \\left(\\sum_ {\\alpha \\in \\mathcal {A} (\\mathbf {P h})} P _ {\\mathrm {c t c}} (\\alpha \\mid \\mathbf {X})\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 325, + 903, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Ph} = (ph_1, \\ldots, ph_m)$ is the ground-truth phoneme sequence, $\\mathcal{A}(\\mathbf{Ph})$ is the set of all valid alignments, and $P_{\\mathrm{ctc}}(\\alpha \\mid \\mathbf{X})$ is the probability of alignment $\\alpha$ . Minimizing $\\mathcal{L}_{\\mathrm{CTC}}$ enables the model to learn frame-to-phoneme mappings without explicit per-frame labels. At inference, we perform beam search over the CTC log probabilities to decode the final phoneme sequence.", + "bbox": [ + 511, + 369, + 906, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Phoneme-to-Sentence Reconstruction Using LLM", + "text_level": 1, + "bbox": [ + 511, + 483, + 903, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The predicted phoneme sequence $\\widehat{\\mathbf{Ph}}$ is derived by decoding the CTC logits from Eq. (4) via beam search. We then feed $\\widehat{\\mathbf{Ph}}$ into a Large Language Model (LLM) that is finetuned to map phonemes to sentences:", + "bbox": [ + 511, + 518, + 905, + 582 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\operatorname {L L M} \\left(\\widehat {\\mathbf {P h}}\\right) = \\left(s _ {1}, s _ {2}, \\dots , s _ {M}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 589, + 903, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LoRA Fine-Tuning. To efficiently adapt the LLM, we employ Low-Rank Adaptation (LoRA) [22], injecting two low-rank matrices $\\mathbf{A}$ and $\\mathbf{B}$ into each linear layer. This greatly reduces the number of trainable parameters compared to full-model fine-tuning. Specifically, for a linear layer $\\mathbf{W}_{\\mathrm{orig}} \\in \\mathbb{R}^{d \\times d}$ :", + "bbox": [ + 511, + 623, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} ^ {\\prime} = \\mathbf {W} _ {\\text {o r i g}} + \\mathbf {A B}, \\quad \\mathbf {A} \\in \\mathbb {R} ^ {d \\times r}, \\mathbf {B} \\in \\mathbb {R} ^ {r \\times d}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 722, + 903, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r \\ll d$ . Only $\\mathbf{A}$ and $\\mathbf{B}$ are updated during training.", + "bbox": [ + 511, + 747, + 893, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining LLM Objective. We fine-tune the LLM on a large phoneme $\\rightarrow$ sentence corpus generated from WikiText [34]. Let $\\mathbf{S} = (s_1, \\ldots, s_M)$ be the ground-truth sentence for a phoneme sequence $\\mathbf{Ph}$ . We employ the cross-entropy loss:", + "bbox": [ + 511, + 779, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {t = 1} ^ {M} \\ln p \\left(s _ {t} \\mid \\mathbf {P h}, s _ {1: t - 1}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 862, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2849", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p(s_{t} \\mid \\mathbf{Ph}, s_{1:t-1})$ is the likelihood of predicting the correct word $s_{t}$ , given the phoneme sequence and previously generated words. During inference, we prompt the LLM with $\\widehat{\\mathbf{Ph}}$ to generate $\\widehat{\\mathbf{S}}$ .", + "bbox": [ + 89, + 90, + 480, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Training Procedure", + "text_level": 1, + "bbox": [ + 89, + 162, + 277, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1: Video $\\rightarrow$ Phoneme. We train the visual extractor, adapter, and CTC head to minimize $\\mathcal{L}_{\\mathrm{CTC}}$ (Eq. 5) on video-phoneme pairs. By learning without strict alignment constraints, the model remains flexible to a wide range of speaking speeds and styles.", + "bbox": [ + 89, + 184, + 482, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Phoneme $\\rightarrow$ Sentence. We independently fine-tune the LLM via cross-entropy loss (Eq. 8) on text-based phoneme $\\rightarrow$ sentence datasets. This leverages easily obtained text data, allowing the LLM to learn phonetic-linguistic mappings separately from visual modeling.", + "bbox": [ + 89, + 273, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. We compose the trained modules to form:", + "bbox": [ + 89, + 361, + 442, + 376 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} \\xrightarrow {f _ {\\mathrm {v i s}}} \\widehat {\\mathbf {P h}} \\xrightarrow {f _ {\\mathrm {L L M}}} \\widehat {\\mathbf {S}},\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 388, + 370, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widehat{\\mathbf{Ph}}$ is decoded from CTC logits, and $\\widehat{\\mathbf{S}}$ is the final sentence. This approach is highly modular and data-efficient, as each stage (visual and linguistic) is learned with its own objective and dataset, yet easily integrated at inference time.", + "bbox": [ + 89, + 419, + 482, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 511, + 223, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section outlines the datasets and pre-processing methods used for training the model architecture.", + "bbox": [ + 89, + 537, + 482, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data", + "text_level": 1, + "bbox": [ + 89, + 575, + 165, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The visual feature extractor, the adapter with downsampling, and CTC head are trained on the LRS2 [47] and LRS3 [3] datasets. The LLM is fine-tuned on the WikiText [34] dataset.", + "bbox": [ + 89, + 598, + 482, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LRS2 [47]: The LRS2 dataset consists of 144,482 video clips of spoken sentences from BBC television, consisting of approximately 224.5 hours of footage and each sentences is up to 100 characters in length. The videos are divided into a pre-training set with 96,318 utterances (195 hours), a training set with 45,839 utterances (28 hours), a validation set with 1,082 utterances (0.6 hours) and a test set with 1,243 utterances (0.5 hours).", + "bbox": [ + 89, + 666, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LRS3 [3]: The LRS3 dataset describes the largest public audio-visual English dataset collected consists of clips from over 5,000 TED and TEDx talks totaling 438.9 hours. It contains 438.9 hours with 151,819 utterances. Specifically, there are 118,516 utterances in the 'pre-train' set (408 hours), 31,982 utterances in the 'train-val' set (30 hours) and 1,321 utterances in the 'test' set (0.9 hours).", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our setting, unlike other approaches, we do not utilise the pre-training set and train our model on only the 28-hour and 30-hour partitions.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phoneme Sentence Pairs. Finally, the dataset used for training the LLM was the WikiText [34] dataset with sentences converted into lists of phonemes using the CMU dictionary [52]. These phonemes were then masked and paired with the original sentences.", + "bbox": [ + 511, + 143, + 903, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 228, + 697, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous works [11, 47] we adopt Word Error Rate [23] (WER) as our evaluation metric for Lip-Reading. WER calculates the percentage of errors in the predicted text compared to the ground truth, accounting for substitutions, insertions, and deletions. Similar to previous studies [3, 11, 41, 42, 47], we report results on all recent AV-ASR and V-ASR approaches.", + "bbox": [ + 511, + 250, + 903, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 369, + 599, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the results of our method and compare them against those of existing SOTA approaches.", + "bbox": [ + 511, + 395, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Public vs Private Data on LRS3 [3]: In Table 1, we compare different V-ASR and AV-ASR models based on their WER on the LRS3 dataset, we split the models into two categories; Fully Supervised models with publicly available data and models trained on large-scale non-publicly available datasets.", + "bbox": [ + 511, + 431, + 905, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Fully Supervised Models rely solely on publicly available labelled datasets with the best WER being $40.6\\%$ , which was achieved by VTP [41]. The non-public dataset models leverage extensive, proprietary datasets, achieving much lower WERs with the best being $12.5\\%$ by LP [10].", + "bbox": [ + 511, + 530, + 903, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LRS3 [3]: In Table 1 we compare our results against other methods based on results from the LRS3 dataset. The results in Table 1 demonstrate a consistent trend: models incorporating extensive self-supervised pre-training followed by fine-tuning achieve lower WERs than those using only supervised approaches. However, our model achieves the lowest WER on the LRS3 dataset at $18.7\\%$ without self-supervised pre-training or additional fine-tuning data.", + "bbox": [ + 511, + 613, + 905, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LRS2 [47]: In Table 2, we compare our method against several other state-of-the-art approaches for visual-only lip reading, focusing on Word Error Rate (WER) on the LRS2 dataset. From Table 2 we can see that increasing the amount of data available helps improve a model's WER. However, we can also see that even without increasing the size of the dataset, our model still outperforms other models that only use publicly available data, achieving the lowest WER at $20.8\\%$ .", + "bbox": [ + 511, + 741, + 905, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore as shown in Tables 2 and 3, we can see that our", + "bbox": [ + 513, + 885, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2850", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4bb0f2d946f065d9dbc0973f67487c4e9b6a98182344b3591187de9a2b301982.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
ASR distillation [4]-59068.8
Conv-Seq2Seq [60]-85560.1
Discriminative AVSR [57]-59057.8
Hyb. + Conformer [29]-59043.3
VTP [41]-69840.6
Trained on large-scale non-publicly available datasets
Deep-AV-SR [2]-1,51958.9
Large-scale AV-SR [46]-3,88655.1
RNN-T [30]-31,00033.6
VTP [41]-2,67630.7
ViT-3D [44]-90,00017.0
LP [10]-1,000,00012.5
Self-supervised pre-training + Supervised fine-tuning on LRS3
LiRA [28]4333071.9
ASR distillation [4]33459059.8
LiteVSR [24]6395945.7
ES3 [61]4333043.5
AV-HuBERT Large [45]1,7593032.5
Lip2Vec [13]1,7593031.2
Sub-Word[41]2,6762,67630.7
SynthVSR [25]3,65243827.9
Whisper [42]1,7593025.5
Auto-AVSR [27]1,75943325.0
LLaMA-AVSR [8]-175624.0
RAVeN [19]175943323.1
USR [20]1,32643321.5
Supervised finetuning only on LRS3
Ours-3018.7
", + "bbox": [ + 94, + 88, + 496, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2309680e9fc8df15b81d4649f6ee4748410ca448eec1e3eceae9d15922ee1af3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 635, + 169, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7754bf5651d54cce94bc54485751dcdadbf2b73bbffaf38ff8144f09528b3941.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 636, + 194, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5241836870d9d791c95f8c9159eb00903de7d664b498f268b9fcd5a3aec71514.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 635, + 225, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/276819e67a971584a49e4d208e6f4455af7997117101357a4747b9be5c95485c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 635, + 253, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/513f852c48a8ad848e828353e103328bee41306c6fde1fbc3a53f93a75960893.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 635, + 281, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/015424b644c8663954efc1f779652ecdfffb7b59c66d495a4885f69ecca87fda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 635, + 308, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0aa3d8102ac7868c0e1bcdc5bebd27111bf0d66867bd2c741b52605324fd8002.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 635, + 334, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cf7ce9a562052ae9dd55dc21fbc0e6a62e0fbb85558cbb0d60521588ee3f667.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 635, + 362, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0a1a3fecf95803ab3592d83a37edcb4301f5f641561753b2b269d90f613aa441.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 635, + 390, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a4c6362e1c7e6d602efec1dab005d55e5828f46a9a4f415b27088b2f964df35e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 635, + 419, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5af5aa88bf55dd63f55f2c27f5fb8caddf21de23c2f3aca92a2e4b6dfa0146fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 635, + 446, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/861cf9126b1057f0a3876f194369d5f406de253a33de9d56ccb82e45cfed29e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 635, + 473, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fa901c1b1ede2859a34c2e0b0be5fd01e9a440a82cb5245f16a3afbdcb64885.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 635, + 500, + 667 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4409cabae598a4e370a820bb58e44f6d466d251279aa3f089505cb0d8a8076d4.jpg", + "table_caption": [ + "Table 2. Comparison of Word Error Rate (WER) across different models for visual-only speech recognition on the LRS3 dataset. The table is divided into three sections: fully supervised models trained on publicly available data, models trained on large-scale non-public datasets, and models that leverage self-supervised pre-training with supervised fine-tuning on LRS3 [3]. Each entry details the amount of labelled and unlabelled data used and the resulting WER. Our approach, tested with various language models, achieves the lowest WER of $22.1\\%$ with the Llama 3.2-3B model, demonstrating the effectiveness of our phoneme-based, LLM-assisted approach." + ], + "table_footnote": [], + "table_body": "
Predicted \nPhonemeW, AA, ZDH, AHK, AY, N, DAH, V
Predicted \nWordWasthekindof
GTWasthekindof
Predicted \nPhonemeAH, B, S, EH, SH, \nAH, NDH, AE, THH, AE, DHH, ER
Predicted \nWordobsessionthathadher
GTobsessionthathadher
", + "bbox": [ + 94, + 676, + 500, + 773 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c720ad6324dd4c6cc3507a3073064c3edc26b7b5fb6b9cf8752d83efaa366110.jpg", + "table_caption": [ + "Figure 3. Example of the model's phonetic and sentence outputs from a sample in the LRS3 dataset [3]. The table illustrates the model's ability to predict a sequence of phonemes from visual input, which are then reconstructed into a coherent sentence by the LLM. In this example all of the phonemes are predicted correctly and the words are recreated correctly." + ], + "table_footnote": [], + "table_body": "
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
Auto-AV-SR [27]-81827.9
Auto-AV-SR [27]-344814.6
Self-supervised pre-training + Supervised finetuning on LRS2 for AVSR
LiRA [28]1,75943338.8
ES3 [61]1,75922326.7
Sub-Word [41]2,6762,67622.6
RAVeN [19]1,75922317.9
USR [20]1,75922315.4
Supervised finetuning only on LRS2
Ours-2820.8
", + "bbox": [ + 517, + 88, + 919, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of Word Error Rate (WER) for various self-supervised pre-training and supervised fine-tuning methods evaluated on the LRS2 [47] dataset. Our method achieves competitive performance with other methods, without the requirement of extensive pre-training, additional labelled data, or additional audio modality as in the best performing approaches. We are the only method to train on only the LRS2 train dataset.", + "bbox": [ + 511, + 270, + 906, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method provides a lower WER with the same amount, or less data, without the requirement of self-supervised pretraining as current approaches [4, 13, 24, 25, 27, 28, 42, 45]. This large increase in performance and generalisation can be explained by the powerful combination of the multi-task objectives and the ability of the LLM to reconstruct sentences accurately even when phonemes are predicted incorrectly. In Fig. 4, we find that the phonetic output of the visual encoder is sometimes incorrect or misses phonemes, but the LLM is able to still reconstruct the correct word (as in the example $z \\rightarrow s$ ). We also show how the model still struggles with tricky homophones such as your and you're. In Fig. 3, we show an additional example of the model's outputs at both phonetic and word levels, demonstrating how phonemes are correctly predicted and combined by the LLM for sentence reconstruction. For further examples please see the appendix.", + "bbox": [ + 511, + 393, + 906, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Ablations", + "text_level": 1, + "bbox": [ + 511, + 659, + 624, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we aim to understand the performance contribution of each component in the network and their performance on visual $\\rightarrow$ phoneme prediction and phoneme $\\rightarrow$ word reconstruction, alongside ablations related to architectural decisions, and the limitations of the approach.", + "bbox": [ + 511, + 681, + 905, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual $\\rightarrow$ Phoneme: The confusion matrix in Fig. 5 shows phoneme prediction performance of the model on the LRS3 [3] dataset. The matrix indicates that some phonemes are misclassified more frequently than others, as seen by the off-diagonal elements, with the worst missed classifications being highlighted with a red bounding box. This bias could be due to the similarity in the visual representation of these sounds, making it challenging for the model to distinguish between them. For instance, /S/ and /Z/ may often be con-", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2851", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e8fc1ec0b0814031a2d09ff8bf89e914bb5512b94eccdc80d0511cedb025fbc.jpg", + "image_caption": [ + "Figure 4. Comparison of the model's phonetic and sentence outputs with ground truth from a sample in the LRS3 [3] dataset. Red shows an incorrect prediction, M shows a missing prediction and green shows a correction from phonemes to words. In this example, even though the model incorrectly predicts certain phonemes, the LLM can correctly recreate the word but struggles to recreate homophones." + ], + "image_footnote": [], + "bbox": [ + 127, + 87, + 500, + 119 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e711fe5e567d49e2d32bb3875ffaa3e363b360ab9840fb1300f0144e21978e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Predicted \nPhonemeY, UH, RJH, AH, Z, TW, AH, NAH, V
Predicted \nWordYourjustoneof
GTYou'rejustoneof
Predicted \nPhonemeDH, AHP, IY, P, <M>, LHH, UWHH, AE, S, N, T
Predicted \nWordthepeoplewhohasn't
GTthepeoplewhohasn't
", + "bbox": [ + 93, + 128, + 500, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fused due to similar lip movements, leading to miss classifications between these phonemes.", + "bbox": [ + 89, + 359, + 482, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7d26b30e521eb6a97cd06f3b94f654e12b2cf2e2cd055439660d384e40abed9.jpg", + "image_caption": [ + "Figure 5. Confusion matrix showing the performance on isolated phonemes of the LRS3 [3] dataset. We observe a very high match rate between the predicted phonemes and the ground truth. In red, we show the most difficult phonemes for our model to identify." + ], + "image_footnote": [], + "bbox": [ + 89, + 400, + 501, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Phoneme $\\rightarrow$ Word: In Table 4 we take the top 5 words in alphabetical order that are misclassified when predicting words directly from phonemes after the LLM pretraining stage. We can observe that these are words that may occur infrequently in the training set and are missclassified due to some phonetic similarities between the words.", + "bbox": [ + 89, + 771, + 482, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different LLMs: In this ablation study, we evaluate three different Language models on our model architecture to", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9d3d4061eb0bfc8f590cb45171de428e906f503bc03b483afafcb52a7cd45f4b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PhonemesTrue WordPredicted Word
/EH/ /R/ /AH/ /N/ /S/ /AH/ /N/AaronsonErasmus
/AA/ /B/ /ER/ /G/AabergBraggartly
/AH/ /S/ /EH/ /R /AH/AcerraEsraa
/AE/ /S/ /AH/ /T/ /EY /T/AcetateAsquette
/AH/ /K/ /AH/ /S/ /T/ /AH/ /M/ /D/AccustomedAccosted
", + "bbox": [ + 517, + 87, + 919, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "understand the impact on word error rate (WER) performance. The first model, GPT 2 [43], serves as our baseline, consisting of the default GPT 2 [43] without fine-tuning on lip-reading datasets. This model achieves a WER of $23.9\\%$ , indicating the core components' foundational capabilities. The second model, Llama 3.2-1B [51] introduces fine-tuning on phoneme word pairs, enhancing the model's contextual understanding of phoneme sequences. With this addition, Llama 3.2-1B [51] has a substantial improvement, reducing the WER to $22.8\\%$ . Finally, Llama 3.2-3B [51] also incorporates fine-tuning for phoneme-to-word reconstruction. However, Llama 3.2-3B [51] is a larger model with more parameters than the Llama 3.2-1B [51] model, further lowering the WER to $18.7\\%$ . This progression highlights the significant contributions of fine-tuning for purpose-specific LLMs.", + "bbox": [ + 511, + 290, + 906, + 532 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb11ee012b8e5c7758268e42e9f87c27a56e49644ba4c76ffcde0a43b1da4199.jpg", + "table_caption": [ + "Table 4. Examples of misclassification's between predicted words and ground truth based on phoneme sequences. Each row presents a sequence of phonemes, the intended (true) word, and the model's predicted word. Differences in predictions illustrate common misclassifications, often due to phonetic similarities between visually indistinct sounds." + ], + "table_footnote": [], + "table_body": "
ModelParameters (B)WER (%)
GPT-2 Small [43]0.1233.9
Llama 3.2-1B [51]1.022.8
Llama 3.2-3B [51]3.018.7
", + "bbox": [ + 527, + 546, + 893, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Comparison of Word Error Rate (WER) for different large language models (LLMs) tested with our model. The table includes each model's parameter count (in billions) and resulting WER on the LRS3 [3] dataset. GPT-2 [43] Small with 0.12B parameters has the highest WER at $23.9\\%$ . In contrast, Llama 3.2-3B [51] achieves the best WER of $18.7\\%$ , highlighting that increased model capacity improves phoneme-to-word reconstruction accuracy.", + "bbox": [ + 511, + 633, + 906, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Freezing the CTC Head: Inspired by works that directly map visual features to pretrained ASR networks [42] we investigate the effect of initialising the CTC head in our model with weights from a pre-trained Wav2Vec2 ASR model [6]. The CTC head plays a crucial role in mapping visual features to the phoneme sequences, and we hypothesized that initialising and freezing its parameters could reduce computational overhead while preserving feature alignment from audio pre-training. To test this hypothesis a baseline model", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2852", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with a non-frozen CTC head and a comparison model with a frozen CTC head were trained and a comparison of the results was made as shown in Table 6. Freezing the CTC head led to a slight increase in WER of $0.4\\%$ , indicating a decrease in performance. This result suggests that while the frozen CTC head still provided a foundation for phoneme alignment, we did not require extensive audio pre-training to obtain superior performance.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Varying Sample Sizes: To understand the impact of additional data on performance, we vary the quantity of data used to train both the Video $\\rightarrow$ Phoneme network and the Phoneme $\\rightarrow$ Sentence LLM and investigate the effect this has on the WER when evaluated on the LRS3 [3] dataset. By varying the percentages of data used in training, we can see the effect on the model in Figure 6. Decreasing the amount of data used to train both models has the expected effect on the WER by increasing it. However, increasing the amount of training data for the Feature Extractor has a negligible effect after 30 hours of data. However, by increasing the amount of training data for the LLM we can also increase the effectiveness of the model and decrease the total WER down further to $17.5\\%$ . The additional data is randomly sampled from both AvSpeech [17] for the visual encoder, and a BookCorpus [63] replica to match the quantity of data in WikiText.", + "bbox": [ + 91, + 220, + 483, + 477 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9039c2c35aa9c82d656a47cc63694fed24bdefa3cff3157abb122551594fff59.jpg", + "image_caption": [ + "Figure 6. Graph showing the effect of varying the sample size of the training datasets for both the Video $\\rightarrow$ Phoneme network (in blue) and for the Phoneme $\\rightarrow$ Sentence LLM (in red). The graph shows that decreasing the amount of data negatively impacts the model for both networks, increasing the amount of training data for the Video $\\rightarrow$ Phoneme network has negligible effect and increasing the amount of training data for the Phoneme $\\rightarrow$ Sentence LLM has a noticeable positive effect." + ], + "image_footnote": [], + "bbox": [ + 114, + 507, + 464, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "End-To-End: For our final study, we investigate the effect of removing the video $\\rightarrow$ phoneme stage and train the", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model to directly predict sentences from the video features, reproducing the method in [41] but with the same ViT encoder [44], LLama LLM [51], and limited data set. As shown in Table 6 we achieve aWER of $25.6\\%$ on LRS3, demonstrating that the more parameterised transformer encoder improves performance over the CNN encoder used in [41]. However, our two stage approach still performs significantly better.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f233a11b44b49368eb2ad65b5c63bc99b6dc35c265d76a66a54f5192fb4b4bff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelWER (%)
Ours (Two Stage)18.7
ViT + LLM (End to End)25.6
Sub-Word (End to End) [41]30.7
", + "bbox": [ + 560, + 227, + 859, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Comparison of Word Error Rate (WER) between using the intermediary Phonemes and skipping the CTC head. Removing the intermediary representation increases the WER from 18.7 to 25.6, showing the importance of the intermediary step and the effectiveness of the phoneme centric fine-tuning.", + "bbox": [ + 511, + 314, + 906, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 410, + 769, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a two-stage, phoneme-centric framework for visual-only lip reading that first predicts phonemes from lip movements and then reconstructs coherent sentences via a fine-tuned LLM. This design sharply reduces word error rates on benchmark datasets (LRS2 [47] and LRS3 [3]), providing improvements of $3.5\\%$ and $6.8\\%$ WER, respectively, over existing state-of-the-art approaches that focus on end-to-end connectionist approaches.", + "bbox": [ + 511, + 435, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite these gains, the approach remains limited by the visual ambiguity of certain phonemes. Lip movements alone cannot always differentiate minimal pairs such as /s/ vs. /z/, and words that share near-identical phoneme sequences (e.g., \"Hello\" vs. \"Hallow\") can still result in misclassification. We plan to address these issues by refining the phoneme-to-word generation process, potentially through deeper linguistic context modeling and more specialized phoneme embeddings. Future work may also explore speaker-adaptive fine-tuning or multi-frame alignment strategies to better handle subtle visual distinctions.", + "bbox": [ + 511, + 582, + 908, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 763, + 681, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the SNSF project 'SMILE II' (CRSII5 193686), the Innosuisse IICT Flagship (PFFS-21-47), EPSRC grant APP24554 (SignGPT-EP/Z535370/1) and through funding from Google.org via the AI for Global Goals scheme. This work reflects only the author's views and the funders are not responsible for any use that may be made of the information it contains.", + "bbox": [ + 511, + 787, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2853", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emre Afacan and Gültekin Demirci. A survey on automatic lip-reading: Challenges and future directions. Multimedia Tools and Applications, 2019. 1, 3", + "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 3, 6", + "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 3, 5, 6, 7, 8", + "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 3, 6", + "[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 3", + "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 2020. 7", + "[7] Helen L Bear, Richard W Harvey, Barry-John Theobald, and Yuxuan Lan. Which phoneme-to-viseme maps best improve visual-only computer lip-reading? In International symposium on visual computing. Springer, 2014. 1", + "[8] Umberto Cappellazzo, Minsu Kim, Honglie Chen, Pingchuan Ma, Stavros Petridis, Daniele Falavigna, Alessio Brutti, and Maja Pantic. Large language models are strong audio-visual speech recognition learners, 2025. 6", + "[9] Luca Cappelletta and Naomi Harte. Viseme definitions comparison for visual-only speech recognition. In 2011 19th European Signal Processing Conference. IEEE, 2011. 3", + "[10] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 5, 6", + "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016. 3, 5", + "[12] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. Grid av speech corpus sample. The Journal of the Accoustical Society of America, 2013. 3", + "[13] Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Haithem Boussaid, Ebtessam Almazrouei, and Merouane Debbah. Lip2vec: Efficient and robust visual speech recognition via latent-to-latent visual to audio representation mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6", + "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[15] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H. Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "speech recognition. CAAI Transactions on Intelligence Technology, 8(1):129-138, 2023. 3", + "[16] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent speech recognition. CAAI Transactions on Intelligence Technology, 2023. 3", + "[17] A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K Wilson, A. Hassidim, W. T. Freeman, and M. Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 8", + "[18] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, 2006. 4", + "[19] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. arXiv preprint arXiv:2212.06246, 2022. 6", + "[20] Alexandros Haliassos, Rodrigo Mira, Honglie Chen, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Unified speech recognition: A single model for auditory, visual, and audio-visual inputs. arXiv preprint arXiv:2411.02256, 2024. 6", + "[21] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 2021. 3", + "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2022. 4", + "[23] F. Jelinek, L. Bahl, and R. Mercer. Design of a linguistic statistical decoder for the recognition of continuous speech. IEEE Transactions on Information Theory, 21(3), 1975. 5", + "[24] Hendrik Laux, Emil Mededovic, Ahmed Hallawa, Lukas Martin, Arne Peine, and Anke Schmeink. Litevsr: Efficient visual speech recognition by learning from speech representations of unlabeled data. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 3, 6", + "[25] Xubo Liu, Egor Lakomkin, Konstantinos Vougioukas, Pingchuan Ma, Honglie Chen, Ruiming Xie, Morrie Doulaty, Niko Moritz, Jachym Kolar, Stavros Petridis, et al. Synthvsr: Scaling up visual speech recognition with synthetic supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 6", + "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4", + "[27] Pingchuan Ma and Alexandros et al. Haliassos. Auto-avsr: Audio-visual speech recognition with automatic labels. In" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2854", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. 3, 6", + "[28] Pingchuan Ma and Rodrigo et al. Mira. Lira: Learning visual speech representations from audio through self-supervision. arXiv, 2021. 3, 6", + "[29] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021. 6", + "[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2019. 3, 6", + "[31] William D Marslen-Wilson. Functional parallelism in spoken word-recognition. Cognition, 25(1-2), 1987. 2", + "[32] William D Marslen-Wilson and Alan Welsh. Processing interactions and lexical access during word recognition in continuous speech. Cognitive psychology, 10(1), 1978. 2", + "[33] Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 1976. 1", + "[34] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 4, 5", + "[35] Ara V Nefian, Luhong Liang, Xiaobo Pi, Liu Xiaoxiang, Crusoe Mao, and Kevin Murphy. A coupled hmm for audiovisual speech recognition. In 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing. IEEE, 2002. 3", + "[36] Eng-Jon Ong and Richard Bowden. Learning temporal signatures for lip reading. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops). IEEE, 2011. 3", + "[37] Stavros Petridis and Maja Pantic. Deep complementary bottleneck features for visual speech recognition. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3", + "[38] Stavros Petridis, Themos Stafylakis, Pingchuan Ma, Georgios Tzimiropoulos, and Maja Pantic. Audio-visual speech recognition with a hybrid ctc/attention architecture. In 2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018. 3", + "[39] Javad Peymanfard, Vahid Saeedi, Mohammad Reza Mohammadi, Hossein Zeinali, and Nasser Mozayani. Leveraging visemes for better visual speech representation and lip reading, 2023. 3", + "[40] Gerasimos Potamianos, Chalapathy Neti, Guillaume Gravier, Ashutosh Garg, and Andrew W Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9), 2003. 3", + "[41] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 5, 6, 8", + "[42] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Speech recognition models are strong lip-readers. In Proc. Interspeech 2024, 2024. 3, 5, 6, 7" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. arXiv, 2018. 3, 7", + "[44] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. arXiv preprint arXiv:2201.10439, 2022. 3, 4, 6, 8", + "[45] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. *ICLR*, 2022. 3, 6", + "[46] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. InterSpeech, 2018. 3, 6", + "[47] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3, 5, 6, 8", + "[48] Themos Stafylakis and Georgios Tzimiropoulos. Zero-shot keyword spotting for visual speech recognition in-the-wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-529, 2018. 3", + "[49] Kwanchiva Thangthai, Helen L Bear, and Richard Harvey. Comparing phonemes and visemes with dnn-based lipreading. arXiv preprint arXiv:1805.02924, 2018. 3", + "[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 3", + "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 7, 8", + "[52] Carnegie Mellon University. The cmu pronouncing dictionary, 1993. http://www.speech.cs.cmu.edu/cgi-bin/cmudict.2,5", + "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 3", + "[54] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3", + "[55] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3", + "[56] Xinshuo Weng and Kris Kitani. Learning spatio-temporal features with two-stream deep 3d cnn for lipreading. arXiv preprint arXiv:1905.02540, 2019. 3", + "[57] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2020. 3, 6", + "[58] Jeong Hun Yeo, Seunghee Han, Minsu Kim, and Yong Man Ro. Where visual speech meets language: Vsp-llm frame" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2855", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "work for efficient and context-aware visual speech processing. arXiv preprint arXiv:2402.15151, 2024. 2, 3", + "[59] Jeong Hun Yeo, Chae Won Kim, Hyunjun Kim, Hyeongseop Rha, Seunghee Han, Wen-Huang Cheng, and Yong Man Ro. Personalized lip reading: Adapting to your unique lip movements with vision and language. arXiv preprint arXiv:2409.00986, 2024. 2, 3", + "[60] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International conference on Computer Vision, 2019. 3, 6", + "[61] Yuanhang Zhang, Shuang Yang, Shiguang Shan, and Xilin Chen. Es3: Evolving self-supervised learning of robust audio-visual speech representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6", + "[62] Ziheng Zhou, Guoying Zhao, Xiaopeng Hong, and Matti Pietikainen. A review of recent advances in visual speech decoding. Image and vision computing, 32(9), 2014. 3", + "[63] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, 2015. 8" + ], + "bbox": [ + 91, + 90, + 482, + 443 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2856", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_model.json b/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b6deef22369d01ece475d61e5e7def932f3e133f --- /dev/null +++ b/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.777, + 0.154 + ], + "angle": 0, + "content": "VALLR: Visual ASR Language Model for Lip Reading" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.182, + 0.71, + 0.218 + ], + "angle": 0, + "content": "Marshall Thomas Edward Fish Richard Bowden University of Surrey" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.22, + 0.807, + 0.234 + ], + "angle": 0, + "content": "mt00893@surrey.ac.uk edward.fish@surrey.ac.uk r.bowden@surrey.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.257, + 0.327, + 0.273 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.299, + 0.486, + 0.766 + ], + "angle": 0, + "content": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), is a complex task requiring the interpretation of spoken language exclusively from visual cues, primarily lip movements and facial expressions. This task is especially challenging due to the absence of auditory information and the inherent ambiguity when visually distinguishing phonemes that have overlapping visemes, where different phonemes appear identical on the lips. Current methods typically attempt to predict words or characters directly from these visual cues, but this approach frequently encounters high error rates due to coarticulation effects and viseme ambiguity. We propose a novel two-stage, phoneme-centric framework for Visual Automatic Speech Recognition (V-ASR) that addresses these longstanding challenges. First, our model predicts a compact sequence of phonemes from visual inputs using a Video Transformer with a CTC head, thereby reducing the task complexity and achieving robust speaker invariance. This phoneme output then serves as the input to a fine-tuned Large Language Model (LLM), which reconstructs coherent words and sentences by leveraging broader linguistic context. Unlike existing methods that either predict words directly or rely on large-scale multimodal pre-training, our approach explicitly encodes intermediate linguistic structure while remaining highly data efficient. We demonstrate state-of-the-art performance on two challenging datasets, LRS2 and LRS3, where our method achieves significant reductions in Word Error Rate (WER) achieving a SOTA WER of 18.7 on LRS3 despite using \\(99.4\\%\\) less labelled video data than the next best approach. Code is available here: https://gitlab.surrey.ac.uk/" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.801, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), involves interpreting spoken language from visual cues such as lip movements and facial expressions. As a natural skill, humans use it to supplement auditory information, and as a technology, it has profound potential for en" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.259, + 0.915, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.476, + 0.908, + 0.59 + ], + "angle": 0, + "content": "Figure 1. Comparison between different models' performances in WER for Visual Automatic Speech Recognition on the LRS3 dataset [3] when compared with the amount of labelled training data. Circle size (green) denotes scale of pre-training data, while fine-tuned models (grey) are not pre-trained. Our model (orange) outperforms all existing approaches with just 30 hours of video training data, no self-supervised pre-training, and without the requirement for additional labeled visual data during fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "hancing accessibility, particularly for the Deaf and hard-of-hearing communities, as well as for applications in noisy or privacy-sensitive environments [1]. Despite substantial advancements in ASR, the visual-only interpretation of speech remains an unresolved challenge due to inherent uncertainties in lip movements and their complex temporal dynamics. A central difficulty in V-ASR stems from the inherent ambiguity of visemes, the visual equivalents of phonemes, which often appear nearly identical for different sounds (e.g., 'p' and 'b') [1, 7]. Further complicating this task are coarticulation effects [33], where adjacent sounds blur one another's articulation, making phoneme boundaries visually unclear. Although some recent methods have focussed on robust intermediary representations to deal with these cases, such as leveraging subword-level predictions [41], or quantized latent representations [13], these approaches can still fail to capture broader contextual cues, limiting their ability to resolve visually similar phonemes that might otherwise be" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2846" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.086, + 0.907, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.353, + 0.907, + 0.437 + ], + "angle": 0, + "content": "Figure 2. An overview of our approach. First we extract facial regions from 16 frames of input video. We apply random pixel masking at \\(50\\%\\) probability, add positional embedding, and encode visual features via a vision transformer encoder (ViT-base [14]). We implement temporal downsampling via 1D convolution and then use a CTC linear head to predict sequences of phonemes. During training, we also fine-tune an LLM to reconstruct sentences from phonemes with a text-only dataset [52]. During inference, the phonemes from the CTC head are processed via the LLM to reconstruct the predicted text. This can be performed end-to-end or in two stages, depending on available resources." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.464, + 0.418, + 0.479 + ], + "angle": 0, + "content": "distinguishable through sentence-level semantics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.494, + 0.483, + 0.691 + ], + "angle": 0, + "content": "One approach to this challenge is to leverage large language models (LLM's) and attempt to map raw lip movements directly to sentences, thus bypassing the issue of viseme ambiguity [58, 59]. However, bridging the gap between high-dimensional, unstructured visual inputs and detailed textual representations is resource intensive and typically requires huge datasets, or specialized architectures. Moreover, these purely end-to-end pipelines often lack an interpretable intermediary step, making it unclear how the model is handling viseme ambiguity at a finer linguistic granularity. This can lead to hallucination effects, which become more critical in lip reading technologies designed for accessibility, or security applications." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Our two-stage, phoneme-centric method addresses both issues. In the first stage, we map short windows of video frames to a discrete, interpretable phoneme sequence. This intermediate representation is significantly easier to learn and more robust against speaker or environmental variations, as phonemes abstract away speaker-specific attributes. Subsequently, in the second stage, we fine-tune an LLM for the task of reconstructing sentences from phonemes. By segmenting the problem, we combine the interpretability and efficiency of phoneme-based prediction with the sophisticated contextual reasoning of modern LLMs. This phoneme \\(\\rightarrow\\) sentence pipeline also aligns with psycholinguistic models of speech perception, where" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.464, + 0.907, + 0.494 + ], + "angle": 0, + "content": "phoneme-level processing naturally precedes lexical access [31, 32]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.502, + 0.81, + 0.517 + ], + "angle": 0, + "content": "Several advantages emerge from this design:" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.524, + 0.907, + 0.583 + ], + "angle": 0, + "content": "(i) Compact Target Space: Predicting only 38 phoneme classes avoids learning entire word-level vocabularies, making the model more data-efficient and simpler to train." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.585, + 0.906, + 0.659 + ], + "angle": 0, + "content": "(ii) LLM-Enhanced Accuracy: A large language model, pre-trained for phoneme \\(\\rightarrow\\) sentence reconstruction, removes the requirement of direct word-level predictions from RGB data, reducing the need for complex multimodal training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.905, + 0.706 + ], + "angle": 0, + "content": "(iii) Data Availability: Exploiting easily available phoneme \\(\\rightarrow\\) sentence text pairs eliminates reliance on extensive lip-reading video data for pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.706, + 0.905, + 0.78 + ], + "angle": 0, + "content": "(iv) Interpretability & Error Analysis: The two-stage design yields an explicit phoneme-level output, constraining recognition errors to local phoneme segments, which the LLM subsequently repairs at the word level reducing errors at a sentence level." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.524, + 0.907, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.793, + 0.655, + 0.808 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.818, + 0.907, + 0.864 + ], + "angle": 0, + "content": "Here, we review key categories of approaches and highlight their advantages and limitations in the context of our proposed phoneme-centric, language-aware approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Lip Reading: Early methods predominantly relied on statistical models such as Hidden Markov Models (HMMs)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2847" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.425 + ], + "angle": 0, + "content": "to aggregate temporal sequences of lip movements alongside hand-crafted feature extractors [1, 9, 35, 36, 40, 62]. With the advent of deep learning, coupled with the availability of large-scale datasets such as Grid [12], LRW [11], LRS2 [47], and LRS3 [3], significant progress has been made in recent years. Initially, Convolutional Neural Networks (CNNs) were applied to the task of word recognition [11], with additional temporal processing via RNN's [30, 54, 55]. With greater compute came the possibility to interpret full sentences with approaches applying Automatic Speech Recognition (ASR) methodologies to the Visual Automatic Speech Recognition (V-ASR) task. These included sequence-to-sequence models [2], CTC approaches [5, 46], and hybrid models [37, 38]. These models improved recognition accuracy by learning hierarchical representations of visual features and modelling temporal dependencies. The application of transformer models [14, 21, 44, 50, 53] in V-ASR brought further performance gains in combination with extensive self-supervised pretraining. Methods such as AV-HuBERT [45] and LiteVSR [24] leveraged self-supervised learning to learn robust representations from large amounts of unlabelled video data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.8 + ], + "angle": 0, + "content": "AV-ASR vs V-ASR: Lip reading approaches can be separated into Visual Automatic Speech Recognition (V-ASR) and Audio-Visual Automatic Speech Recognition (AV-ASR). In the AV-ASR task, labelled audio data is available at training time which can be used to guide the visual encoder in various ways. These can include generating synthetic additional pre-training data [25, 27], or simply fusing audio-visual input features [44-46, 57]. Other methods have used knowledge distillation to inductively transfer knowledge from pre-trained ASR models to visual encoders [4, 28]. However, these methods depend on audio data or pre-trained ASR models, which may not be available or reliable in scenarios where lip reading is most needed, such as noisy environments. V-ASR, on the other hand, is a more challenging task that relies only on visual data in the training stage [1, 56, 60, 62]. Recent methods have also shown the effectiveness of adapting visual inputs to pre-trained ASR models [13, 24, 42] without the need for audio data during training. This method is particularly powerful since the ASR model already has the ability to reconstruct latent representations to words from the pre-training task, however it relies on large amounts of additional data to learn robust visual representations and map them to the audio latent space [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Lip Reading with LLMs: Recent approaches have explored visual attention mechanisms for sub-word units [15, 39, 41, 48] to capture fine-grained linguistic features in lip reading. However, reconstructing these units into coherent sentences remains challenging for networks lacking robust contextual and semantic reasoning capabilities. Early" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "phoneme-based methods [16, 49] similarly struggled with effective decoding to word-level representations and these methods were outpaced in performance by end-to-end approaches [5] which could leverage extensive data. As such they have not been fully explored in the context of modern architectures and techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.907, + 0.327 + ], + "angle": 0, + "content": "For example, recent approaches which have integrated Large Language Models (LLMs) [43, 51] into lip-reading pipelines [58, 59] have started from mapping visual features directly to LLM text embeddings. However, by omitting explicit intermediate representations, these approaches inherit phonetic ambiguities that propagate through the network. This manifests as increased word error rates and reduced robustness, thus requiring substantially more training data to achieve generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.909, + 0.532 + ], + "angle": 0, + "content": "Our work re-examines phonemes as an interpretable discrete representation bridging visual and linguistic domains. Rather than forcing LLMs to perform visual \\(\\rightarrow\\) sentence translation requiring expensive visual-text alignment – we constrain the LLM's role to phoneme \\(\\rightarrow\\) sentence reconstruction. This decomposition offers three key advantages: (1) Phoneme-to-text translation is a well-constrained task requiring only lightweight LLM fine-tuning using abundant textual corpora; (2) Speaker-independent phoneme representations eliminate the need for visual adaptation in the language model; (3) The modular architecture prevents error propagation between visual and linguistic domains while maintaining interpretability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.545, + 0.605, + 0.56 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.906, + 0.616 + ], + "angle": 0, + "content": "In this section, we present our two-stage phoneme-centric approach to visual-only lip reading. Figure 2 provides an overview of the framework. Our goal is to learn a function" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.628, + 0.906, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f: \\quad \\mathbf {X} = \\left\\{x _ {1}, x _ {2}, \\dots , x _ {T} \\right\\} \\\\ \\longmapsto \\mathbf {P h} = \\left\\{p h _ {1}, p h _ {2}, \\dots , p h _ {m} \\right\\} \\tag {1} \\\\ \\longmapsto \\mathbf {S} = \\left\\{s _ {1}, s _ {2}, \\dots , s _ {M} \\right\\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.694, + 0.906, + 0.756 + ], + "angle": 0, + "content": "where \\(\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\) is a video sequence of \\(T\\) frames (with height \\(H\\) and width \\(W\\)), \\(\\mathbf{Ph}\\) denotes the phoneme sequence, and \\(\\mathbf{S}\\) is the reconstructed sentence. Specifically, we decompose the task into:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.764, + 0.906, + 0.794 + ], + "angle": 0, + "content": "1. Video \\(\\rightarrow\\) Phoneme: Map the sequence of video frames to phonemes using a Vision Transformer and CTC head." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.795, + 0.906, + 0.839 + ], + "angle": 0, + "content": "2. Phoneme \\(\\rightarrow\\) Sentence: Convert phonemes to a coherent word sequence with a fine-tuned Large Language Model (LLM)." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.764, + 0.906, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.848, + 0.835, + 0.864 + ], + "angle": 0, + "content": "The model comprises four primary components:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "- Visual Feature Extractor: Captures relevant features from each frame via a Vision Transformer." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2848" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.481, + 0.121 + ], + "angle": 0, + "content": "- Adapter Network with Temporal Downsampling: Reduces sequence length to accommodate CTC alignment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.122, + 0.481, + 0.152 + ], + "angle": 0, + "content": "- CTC Head: Predicts phoneme sequences without requiring explicit temporal labels." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.153, + 0.481, + 0.182 + ], + "angle": 0, + "content": "- LLM for Phoneme-to-Sentence Reconstruction: Translates predicted phonemes into final word sequences." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.481, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.193, + 0.308, + 0.209 + ], + "angle": 0, + "content": "3.1. Pre-Processing Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.482, + 0.29 + ], + "angle": 0, + "content": "Each frame \\( x_{i} \\) is first passed through a face-detection model [26] to identify and crop the speaker's face region, centered on the speaker's lips. The detected region is then cropped, resized to \\( 224 \\times 224 \\), and normalized for batch processing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.3, + 0.274, + 0.315 + ], + "angle": 0, + "content": "3.2. Video Transformer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.482, + 0.412 + ], + "angle": 0, + "content": "We employ a Vision Transformer (ViT) [44] to encode the spatio-temporal information in the face region. Let ViT(\\cdot) denote this encoding function. The input \\(\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\) is chunked into patches, embedded, and equipped with positional encodings before passing through multiple transformer blocks. We obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.424, + 0.482, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\operatorname {V i T} (\\mathbf {X}) \\in \\mathbb {R} ^ {T \\times D}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.514 + ], + "angle": 0, + "content": "where \\(D\\) is the transformer output dimension per frame. During training, we randomly mask portions of the input patches to promote robust feature learning and improve generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.525, + 0.483, + 0.557 + ], + "angle": 0, + "content": "3.3. Adapter Network with Temporal Downsampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.624 + ], + "angle": 0, + "content": "The sequence length \\( T \\) of the ViT output can be prohibitively large for the subsequent CTC module. To reduce the temporal dimension, we apply a sequence of 1D convolutions and pooling layers, as shown in Table ?? Formally," + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.635, + 0.482, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {\\text {d o w n}} = \\operatorname {A d a p t} (\\mathbf {Z}) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C _ {\\text {a d p a t e r}}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.697 + ], + "angle": 0, + "content": "where \\( T' < T \\) and \\( C_{\\mathrm{adapter}} \\) is an intermediate feature dimension aligned to the CTC head input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.212, + 0.721 + ], + "angle": 0, + "content": "3.4. CTC Head" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.759 + ], + "angle": 0, + "content": "An MLP with one hidden layer transforms \\(\\mathbf{G}_{\\mathrm{down}} \\in \\mathbb{R}^{T' \\times C_{\\mathrm{adapter}}} \\text{ into phoneme logits:}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.771, + 0.482, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {\\mathrm {c t c}} = \\operatorname {M L P} \\left(\\mathbf {G} _ {\\text {d o w n}}\\right) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times N _ {\\text {p h o n e m e s}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.483, + 0.864 + ], + "angle": 0, + "content": "where \\( N_{\\text{phonemes}} \\) is the size of our phoneme set (39 English phonemes plus a blank symbol). We then apply a logarithmic softmax along the phoneme dimension to obtain log probabilities \\( \\log p_{\\text{ctc}}(ph_t \\mid \\mathbf{X}) \\) at each time step." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Since lip movements do not align cleanly with discrete phoneme boundaries, we adopt the Connectionist Temporal" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.924, + 0.23 + ], + "angle": 0, + "content": "
StageModuleIn → Out ChannelsKernelStridePaddingDownsample FactorOutput Length
1Conv1dfeature_size → adapter_dim522×2L/2
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
2Conv1dadapter_dim → adapter_dim321×2L/4
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
3Conv1dadapter_dim → adapter_dim321×2L/8
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
4Conv1dadapter_dim → adapter_dim331×3L/24
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
5AvgPool1dadapter_dim → adapter_dim580×8L/192
6Linearadapter_dim → CTC_dim---×1L/192
ReLU----
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.241, + 0.905, + 0.281 + ], + "angle": 0, + "content": "Table 1. Architecture of the adapter for temporal downsampling, where the input length is \\( T = 1568 \\) and final output length is \\( T = 8 \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.306, + 0.72, + 0.321 + ], + "angle": 0, + "content": "Classification (CTC) loss [18]:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.327, + 0.905, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C T C}} = - \\ln \\left(\\sum_ {\\alpha \\in \\mathcal {A} (\\mathbf {P h})} P _ {\\mathrm {c t c}} (\\alpha \\mid \\mathbf {X})\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.907, + 0.477 + ], + "angle": 0, + "content": "where \\(\\mathbf{Ph} = (ph_1, \\ldots, ph_m)\\) is the ground-truth phoneme sequence, \\(\\mathcal{A}(\\mathbf{Ph})\\) is the set of all valid alignments, and \\(P_{\\mathrm{ctc}}(\\alpha \\mid \\mathbf{X})\\) is the probability of alignment \\(\\alpha\\). Minimizing \\(\\mathcal{L}_{\\mathrm{CTC}}\\) enables the model to learn frame-to-phoneme mappings without explicit per-frame labels. At inference, we perform beam search over the CTC log probabilities to decode the final phoneme sequence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.905, + 0.514 + ], + "angle": 0, + "content": "3.5. Phoneme-to-Sentence Reconstruction Using LLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.583 + ], + "angle": 0, + "content": "The predicted phoneme sequence \\(\\widehat{\\mathbf{Ph}}\\) is derived by decoding the CTC logits from Eq. (4) via beam search. We then feed \\(\\widehat{\\mathbf{Ph}}\\) into a Large Language Model (LLM) that is finetuned to map phonemes to sentences:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.59, + 0.905, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\operatorname {L L M} \\left(\\widehat {\\mathbf {P h}}\\right) = \\left(s _ {1}, s _ {2}, \\dots , s _ {M}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.906, + 0.716 + ], + "angle": 0, + "content": "LoRA Fine-Tuning. To efficiently adapt the LLM, we employ Low-Rank Adaptation (LoRA) [22], injecting two low-rank matrices \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) into each linear layer. This greatly reduces the number of trainable parameters compared to full-model fine-tuning. Specifically, for a linear layer \\(\\mathbf{W}_{\\mathrm{orig}} \\in \\mathbb{R}^{d \\times d}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.723, + 0.905, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} ^ {\\prime} = \\mathbf {W} _ {\\text {o r i g}} + \\mathbf {A B}, \\quad \\mathbf {A} \\in \\mathbb {R} ^ {d \\times r}, \\mathbf {B} \\in \\mathbb {R} ^ {r \\times d}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.748, + 0.894, + 0.763 + ], + "angle": 0, + "content": "where \\(r \\ll d\\). Only \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) are updated during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Pretraining LLM Objective. We fine-tune the LLM on a large phoneme \\(\\rightarrow\\) sentence corpus generated from WikiText [34]. Let \\(\\mathbf{S} = (s_1, \\ldots, s_M)\\) be the ground-truth sentence for a phoneme sequence \\(\\mathbf{Ph}\\). We employ the cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.863, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {t = 1} ^ {M} \\ln p \\left(s _ {t} \\mid \\mathbf {P h}, s _ {1: t - 1}\\right), \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2849" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.155 + ], + "angle": 0, + "content": "where \\( p(s_{t} \\mid \\mathbf{Ph}, s_{1:t-1}) \\) is the likelihood of predicting the correct word \\( s_{t} \\), given the phoneme sequence and previously generated words. During inference, we prompt the LLM with \\( \\widehat{\\mathbf{Ph}} \\) to generate \\( \\widehat{\\mathbf{S}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.164, + 0.279, + 0.18 + ], + "angle": 0, + "content": "3.6. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.262 + ], + "angle": 0, + "content": "Stage 1: Video \\(\\rightarrow\\) Phoneme. We train the visual extractor, adapter, and CTC head to minimize \\(\\mathcal{L}_{\\mathrm{CTC}}\\) (Eq. 5) on video-phoneme pairs. By learning without strict alignment constraints, the model remains flexible to a wide range of speaking speeds and styles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.35 + ], + "angle": 0, + "content": "Stage 2: Phoneme \\(\\rightarrow\\) Sentence. We independently fine-tune the LLM via cross-entropy loss (Eq. 8) on text-based phoneme \\(\\rightarrow\\) sentence datasets. This leverages easily obtained text data, allowing the LLM to learn phonetic-linguistic mappings separately from visual modeling." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.362, + 0.444, + 0.377 + ], + "angle": 0, + "content": "Inference. We compose the trained modules to form:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.389, + 0.371, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} \\xrightarrow {f _ {\\mathrm {v i s}}} \\widehat {\\mathbf {P h}} \\xrightarrow {f _ {\\mathrm {L L M}}} \\widehat {\\mathbf {S}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.483, + 0.498 + ], + "angle": 0, + "content": "where \\(\\widehat{\\mathbf{Ph}}\\) is decoded from CTC logits, and \\(\\widehat{\\mathbf{S}}\\) is the final sentence. This approach is highly modular and data-efficient, as each stage (visual and linguistic) is learned with its own objective and dataset, yet easily integrated at inference time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.224, + 0.529 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.568 + ], + "angle": 0, + "content": "This section outlines the datasets and pre-processing methods used for training the model architecture." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.166, + 0.592 + ], + "angle": 0, + "content": "4.1. Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.659 + ], + "angle": 0, + "content": "The visual feature extractor, the adapter with downsampling, and CTC head are trained on the LRS2 [47] and LRS3 [3] datasets. The LLM is fine-tuned on the WikiText [34] dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.789 + ], + "angle": 0, + "content": "LRS2 [47]: The LRS2 dataset consists of 144,482 video clips of spoken sentences from BBC television, consisting of approximately 224.5 hours of footage and each sentences is up to 100 characters in length. The videos are divided into a pre-training set with 96,318 utterances (195 hours), a training set with 45,839 utterances (28 hours), a validation set with 1,082 utterances (0.6 hours) and a test set with 1,243 utterances (0.5 hours)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "LRS3 [3]: The LRS3 dataset describes the largest public audio-visual English dataset collected consists of clips from over 5,000 TED and TEDx talks totaling 438.9 hours. It contains 438.9 hours with 151,819 utterances. Specifically, there are 118,516 utterances in the 'pre-train' set (408 hours), 31,982 utterances in the 'train-val' set (30 hours) and 1,321 utterances in the 'test' set (0.9 hours)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "In our setting, unlike other approaches, we do not utilise the pre-training set and train our model on only the 28-hour and 30-hour partitions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.145, + 0.905, + 0.22 + ], + "angle": 0, + "content": "Phoneme Sentence Pairs. Finally, the dataset used for training the LLM was the WikiText [34] dataset with sentences converted into lists of phonemes using the CMU dictionary [52]. These phonemes were then masked and paired with the original sentences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.229, + 0.698, + 0.244 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.251, + 0.905, + 0.357 + ], + "angle": 0, + "content": "Following previous works [11, 47] we adopt Word Error Rate [23] (WER) as our evaluation metric for Lip-Reading. WER calculates the percentage of errors in the predicted text compared to the ground truth, accounting for substitutions, insertions, and deletions. Similar to previous studies [3, 11, 41, 42, 47], we report results on all recent AV-ASR and V-ASR approaches." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.37, + 0.6, + 0.385 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.396, + 0.905, + 0.427 + ], + "angle": 0, + "content": "In this section, we present the results of our method and compare them against those of existing SOTA approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.906, + 0.523 + ], + "angle": 0, + "content": "Public vs Private Data on LRS3 [3]: In Table 1, we compare different V-ASR and AV-ASR models based on their WER on the LRS3 dataset, we split the models into two categories; Fully Supervised models with publicly available data and models trained on large-scale non-publicly available datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.531, + 0.905, + 0.607 + ], + "angle": 0, + "content": "The Fully Supervised Models rely solely on publicly available labelled datasets with the best WER being \\(40.6\\%\\), which was achieved by VTP [41]. The non-public dataset models leverage extensive, proprietary datasets, achieving much lower WERs with the best being \\(12.5\\%\\) by LP [10]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.906, + 0.736 + ], + "angle": 0, + "content": "LRS3 [3]: In Table 1 we compare our results against other methods based on results from the LRS3 dataset. The results in Table 1 demonstrate a consistent trend: models incorporating extensive self-supervised pre-training followed by fine-tuning achieve lower WERs than those using only supervised approaches. However, our model achieves the lowest WER on the LRS3 dataset at \\(18.7\\%\\) without self-supervised pre-training or additional fine-tuning data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.742, + 0.906, + 0.877 + ], + "angle": 0, + "content": "LRS2 [47]: In Table 2, we compare our method against several other state-of-the-art approaches for visual-only lip reading, focusing on Word Error Rate (WER) on the LRS2 dataset. From Table 2 we can see that increasing the amount of data available helps improve a model's WER. However, we can also see that even without increasing the size of the dataset, our model still outperforms other models that only use publicly available data, achieving the lowest WER at \\(20.8\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Therefore as shown in Tables 2 and 3, we can see that our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2850" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.498, + 0.462 + ], + "angle": 0, + "content": "
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
ASR distillation [4]-59068.8
Conv-Seq2Seq [60]-85560.1
Discriminative AVSR [57]-59057.8
Hyb. + Conformer [29]-59043.3
VTP [41]-69840.6
Trained on large-scale non-publicly available datasets
Deep-AV-SR [2]-1,51958.9
Large-scale AV-SR [46]-3,88655.1
RNN-T [30]-31,00033.6
VTP [41]-2,67630.7
ViT-3D [44]-90,00017.0
LP [10]-1,000,00012.5
Self-supervised pre-training + Supervised fine-tuning on LRS3
LiRA [28]4333071.9
ASR distillation [4]33459059.8
LiteVSR [24]6395945.7
ES3 [61]4333043.5
AV-HuBERT Large [45]1,7593032.5
Lip2Vec [13]1,7593031.2
Sub-Word[41]2,6762,67630.7
SynthVSR [25]3,65243827.9
Whisper [42]1,7593025.5
Auto-AVSR [27]1,75943325.0
LLaMA-AVSR [8]-175624.0
RAVeN [19]175943323.1
USR [20]1,32643321.5
Supervised finetuning only on LRS3
Ours-3018.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Table 2. Comparison of Word Error Rate (WER) across different models for visual-only speech recognition on the LRS3 dataset. The table is divided into three sections: fully supervised models trained on publicly available data, models trained on large-scale non-public datasets, and models that leverage self-supervised pre-training with supervised fine-tuning on LRS3 [3]. Each entry details the amount of labelled and unlabelled data used and the resulting WER. Our approach, tested with various language models, achieves the lowest WER of \\(22.1\\%\\) with the Llama 3.2-3B model, demonstrating the effectiveness of our phoneme-based, LLM-assisted approach." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.636, + 0.171, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.637, + 0.196, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.636, + 0.226, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.636, + 0.254, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.636, + 0.282, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.636, + 0.309, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.636, + 0.336, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.636, + 0.364, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.636, + 0.392, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.636, + 0.42, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.636, + 0.447, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.636, + 0.474, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.636, + 0.501, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.678, + 0.5, + 0.774 + ], + "angle": 0, + "content": "
Predicted \nPhonemeW, AA, ZDH, AHK, AY, N, DAH, V
Predicted \nWordWasthekindof
GTWasthekindof
Predicted \nPhonemeAH, B, S, EH, SH, \nAH, NDH, AE, THH, AE, DHH, ER
Predicted \nWordobsessionthathadher
GTobsessionthathadher
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.785, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Figure 3. Example of the model's phonetic and sentence outputs from a sample in the LRS3 dataset [3]. The table illustrates the model's ability to predict a sequence of phonemes from visual input, which are then reconstructed into a coherent sentence by the LLM. In this example all of the phonemes are predicted correctly and the words are recreated correctly." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.921, + 0.261 + ], + "angle": 0, + "content": "
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
Auto-AV-SR [27]-81827.9
Auto-AV-SR [27]-344814.6
Self-supervised pre-training + Supervised finetuning on LRS2 for AVSR
LiRA [28]1,75943338.8
ES3 [61]1,75922326.7
Sub-Word [41]2,6762,67622.6
RAVeN [19]1,75922317.9
USR [20]1,75922315.4
Supervised finetuning only on LRS2
Ours-2820.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.271, + 0.907, + 0.369 + ], + "angle": 0, + "content": "Table 3. Comparison of Word Error Rate (WER) for various self-supervised pre-training and supervised fine-tuning methods evaluated on the LRS2 [47] dataset. Our method achieves competitive performance with other methods, without the requirement of extensive pre-training, additional labelled data, or additional audio modality as in the best performing approaches. We are the only method to train on only the LRS2 train dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.652 + ], + "angle": 0, + "content": "method provides a lower WER with the same amount, or less data, without the requirement of self-supervised pretraining as current approaches [4, 13, 24, 25, 27, 28, 42, 45]. This large increase in performance and generalisation can be explained by the powerful combination of the multi-task objectives and the ability of the LLM to reconstruct sentences accurately even when phonemes are predicted incorrectly. In Fig. 4, we find that the phonetic output of the visual encoder is sometimes incorrect or misses phonemes, but the LLM is able to still reconstruct the correct word (as in the example \\( z \\rightarrow s \\)). We also show how the model still struggles with tricky homophones such as your and you're. In Fig. 3, we show an additional example of the model's outputs at both phonetic and word levels, demonstrating how phonemes are correctly predicted and combined by the LLM for sentence reconstruction. For further examples please see the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.66, + 0.625, + 0.674 + ], + "angle": 0, + "content": "5.1. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.759 + ], + "angle": 0, + "content": "In this section, we aim to understand the performance contribution of each component in the network and their performance on visual \\(\\rightarrow\\) phoneme prediction and phoneme \\(\\rightarrow\\) word reconstruction, alongside ablations related to architectural decisions, and the limitations of the approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Visual \\(\\rightarrow\\) Phoneme: The confusion matrix in Fig. 5 shows phoneme prediction performance of the model on the LRS3 [3] dataset. The matrix indicates that some phonemes are misclassified more frequently than others, as seen by the off-diagonal elements, with the worst missed classifications being highlighted with a red bounding box. This bias could be due to the similarity in the visual representation of these sounds, making it challenging for the model to distinguish between them. For instance, /S/ and /Z/ may often be con-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.088, + 0.501, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.13, + 0.5, + 0.226 + ], + "angle": 0, + "content": "
Predicted \nPhonemeY, UH, RJH, AH, Z, TW, AH, NAH, V
Predicted \nWordYourjustoneof
GTYou'rejustoneof
Predicted \nPhonemeDH, AHP, IY, P, <M>, LHH, UWHH, AE, S, N, T
Predicted \nWordthepeoplewhohasn't
GTthepeoplewhohasn't
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.237, + 0.484, + 0.336 + ], + "angle": 0, + "content": "Figure 4. Comparison of the model's phonetic and sentence outputs with ground truth from a sample in the LRS3 [3] dataset. Red shows an incorrect prediction, M shows a missing prediction and green shows a correction from phonemes to words. In this example, even though the model incorrectly predicts certain phonemes, the LLM can correctly recreate the word but struggles to recreate homophones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.483, + 0.391 + ], + "angle": 0, + "content": "fused due to similar lip movements, leading to miss classifications between these phonemes." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.401, + 0.502, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Figure 5. Confusion matrix showing the performance on isolated phonemes of the LRS3 [3] dataset. We observe a very high match rate between the predicted phonemes and the ground truth. In red, we show the most difficult phonemes for our model to identify." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Phoneme \\(\\rightarrow\\) Word: In Table 4 we take the top 5 words in alphabetical order that are misclassified when predicting words directly from phonemes after the LLM pretraining stage. We can observe that these are words that may occur infrequently in the training set and are missclassified due to some phonetic similarities between the words." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Different LLMs: In this ablation study, we evaluate three different Language models on our model architecture to" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.088, + 0.92, + 0.169 + ], + "angle": 0, + "content": "
PhonemesTrue WordPredicted Word
/EH/ /R/ /AH/ /N/ /S/ /AH/ /N/AaronsonErasmus
/AA/ /B/ /ER/ /G/AabergBraggartly
/AH/ /S/ /EH/ /R /AH/AcerraEsraa
/AE/ /S/ /AH/ /T/ /EY /T/AcetateAsquette
/AH/ /K/ /AH/ /S/ /T/ /AH/ /M/ /D/AccustomedAccosted
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.179, + 0.907, + 0.262 + ], + "angle": 0, + "content": "Table 4. Examples of misclassification's between predicted words and ground truth based on phoneme sequences. Each row presents a sequence of phonemes, the intended (true) word, and the model's predicted word. Differences in predictions illustrate common misclassifications, often due to phonetic similarities between visually indistinct sounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.907, + 0.534 + ], + "angle": 0, + "content": "understand the impact on word error rate (WER) performance. The first model, GPT 2 [43], serves as our baseline, consisting of the default GPT 2 [43] without fine-tuning on lip-reading datasets. This model achieves a WER of \\(23.9\\%\\), indicating the core components' foundational capabilities. The second model, Llama 3.2-1B [51] introduces fine-tuning on phoneme word pairs, enhancing the model's contextual understanding of phoneme sequences. With this addition, Llama 3.2-1B [51] has a substantial improvement, reducing the WER to \\(22.8\\%\\). Finally, Llama 3.2-3B [51] also incorporates fine-tuning for phoneme-to-word reconstruction. However, Llama 3.2-3B [51] is a larger model with more parameters than the Llama 3.2-1B [51] model, further lowering the WER to \\(18.7\\%\\). This progression highlights the significant contributions of fine-tuning for purpose-specific LLMs." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.547, + 0.895, + 0.625 + ], + "angle": 0, + "content": "
ModelParameters (B)WER (%)
GPT-2 Small [43]0.1233.9
Llama 3.2-1B [51]1.022.8
Llama 3.2-3B [51]3.018.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Table 5. Comparison of Word Error Rate (WER) for different large language models (LLMs) tested with our model. The table includes each model's parameter count (in billions) and resulting WER on the LRS3 [3] dataset. GPT-2 [43] Small with 0.12B parameters has the highest WER at \\(23.9\\%\\). In contrast, Llama 3.2-3B [51] achieves the best WER of \\(18.7\\%\\), highlighting that increased model capacity improves phoneme-to-word reconstruction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Freezing the CTC Head: Inspired by works that directly map visual features to pretrained ASR networks [42] we investigate the effect of initialising the CTC head in our model with weights from a pre-trained Wav2Vec2 ASR model [6]. The CTC head plays a crucial role in mapping visual features to the phoneme sequences, and we hypothesized that initialising and freezing its parameters could reduce computational overhead while preserving feature alignment from audio pre-training. To test this hypothesis a baseline model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2852" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "with a non-frozen CTC head and a comparison model with a frozen CTC head were trained and a comparison of the results was made as shown in Table 6. Freezing the CTC head led to a slight increase in WER of \\(0.4\\%\\), indicating a decrease in performance. This result suggests that while the frozen CTC head still provided a foundation for phoneme alignment, we did not require extensive audio pre-training to obtain superior performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.222, + 0.485, + 0.478 + ], + "angle": 0, + "content": "Varying Sample Sizes: To understand the impact of additional data on performance, we vary the quantity of data used to train both the Video \\(\\rightarrow\\) Phoneme network and the Phoneme \\(\\rightarrow\\) Sentence LLM and investigate the effect this has on the WER when evaluated on the LRS3 [3] dataset. By varying the percentages of data used in training, we can see the effect on the model in Figure 6. Decreasing the amount of data used to train both models has the expected effect on the WER by increasing it. However, increasing the amount of training data for the Feature Extractor has a negligible effect after 30 hours of data. However, by increasing the amount of training data for the LLM we can also increase the effectiveness of the model and decrease the total WER down further to \\(17.5\\%\\). The additional data is randomly sampled from both AvSpeech [17] for the visual encoder, and a BookCorpus [63] replica to match the quantity of data in WikiText." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.508, + 0.465, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Figure 6. Graph showing the effect of varying the sample size of the training datasets for both the Video \\(\\rightarrow\\) Phoneme network (in blue) and for the Phoneme \\(\\rightarrow\\) Sentence LLM (in red). The graph shows that decreasing the amount of data negatively impacts the model for both networks, increasing the amount of training data for the Video \\(\\rightarrow\\) Phoneme network has negligible effect and increasing the amount of training data for the Phoneme \\(\\rightarrow\\) Sentence LLM has a noticeable positive effect." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "End-To-End: For our final study, we investigate the effect of removing the video \\(\\rightarrow\\) phoneme stage and train the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "model to directly predict sentences from the video features, reproducing the method in [41] but with the same ViT encoder [44], LLama LLM [51], and limited data set. As shown in Table 6 we achieve aWER of \\(25.6\\%\\) on LRS3, demonstrating that the more parameterised transformer encoder improves performance over the CNN encoder used in [41]. However, our two stage approach still performs significantly better." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.228, + 0.861, + 0.306 + ], + "angle": 0, + "content": "
ModelWER (%)
Ours (Two Stage)18.7
ViT + LLM (End to End)25.6
Sub-Word (End to End) [41]30.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.315, + 0.907, + 0.386 + ], + "angle": 0, + "content": "Table 6. Comparison of Word Error Rate (WER) between using the intermediary Phonemes and skipping the CTC head. Removing the intermediary representation increases the WER from 18.7 to 25.6, showing the importance of the intermediary step and the effectiveness of the phoneme centric fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.411, + 0.771, + 0.427 + ], + "angle": 0, + "content": "6. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.572 + ], + "angle": 0, + "content": "We have introduced a two-stage, phoneme-centric framework for visual-only lip reading that first predicts phonemes from lip movements and then reconstructs coherent sentences via a fine-tuned LLM. This design sharply reduces word error rates on benchmark datasets (LRS2 [47] and LRS3 [3]), providing improvements of \\(3.5\\%\\) and \\(6.8\\%\\) WER, respectively, over existing state-of-the-art approaches that focus on end-to-end connectionist approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.75 + ], + "angle": 0, + "content": "Despite these gains, the approach remains limited by the visual ambiguity of certain phonemes. Lip movements alone cannot always differentiate minimal pairs such as /s/ vs. /z/, and words that share near-identical phoneme sequences (e.g., \"Hello\" vs. \"Hallow\") can still result in misclassification. We plan to address these issues by refining the phoneme-to-word generation process, potentially through deeper linguistic context modeling and more specialized phoneme embeddings. Future work may also explore speaker-adaptive fine-tuning or multi-frame alignment strategies to better handle subtle visual distinctions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.682, + 0.782 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.907, + 0.888 + ], + "angle": 0, + "content": "This work was supported by the SNSF project 'SMILE II' (CRSII5 193686), the Innosuisse IICT Flagship (PFFS-21-47), EPSRC grant APP24554 (SignGPT-EP/Z535370/1) and through funding from Google.org via the AI for Global Goals scheme. This work reflects only the author's views and the funders are not responsible for any use that may be made of the information it contains." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2853" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Emre Afacan and Gültekin Demirci. A survey on automatic lip-reading: Challenges and future directions. Multimedia Tools and Applications, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.483, + 0.267 + ], + "angle": 0, + "content": "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.482, + 0.337 + ], + "angle": 0, + "content": "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.339, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[7] Helen L Bear, Richard W Harvey, Barry-John Theobald, and Yuxuan Lan. Which phoneme-to-viseme maps best improve visual-only computer lip-reading? In International symposium on visual computing. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[8] Umberto Cappellazzo, Minsu Kim, Honglie Chen, Pingchuan Ma, Stavros Petridis, Daniele Falavigna, Alessio Brutti, and Maja Pantic. Large language models are strong audio-visual speech recognition learners, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[9] Luca Cappelletta and Naomi Harte. Viseme definitions comparison for visual-only speech recognition. In 2011 19th European Signal Processing Conference. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.593, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[10] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[12] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. Grid av speech corpus sample. The Journal of the Accoustical Society of America, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[13] Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Haithem Boussaid, Ebtessam Almazrouei, and Merouane Debbah. Lip2vec: Efficient and robust visual speech recognition via latent-to-latent visual to audio representation mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H. Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "speech recognition. CAAI Transactions on Intelligence Technology, 8(1):129-138, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[16] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent speech recognition. CAAI Transactions on Intelligence Technology, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[17] A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K Wilson, A. Hassidim, W. T. Freeman, and M. Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[18] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.33, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[19] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. arXiv preprint arXiv:2212.06246, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.386, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[20] Alexandros Haliassos, Rodrigo Mira, Honglie Chen, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Unified speech recognition: A single model for auditory, visual, and audio-visual inputs. arXiv preprint arXiv:2411.02256, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[21] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.623 + ], + "angle": 0, + "content": "[23] F. Jelinek, L. Bahl, and R. Mercer. Design of a linguistic statistical decoder for the recognition of continuous speech. IEEE Transactions on Information Theory, 21(3), 1975. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[24] Hendrik Laux, Emil Mededovic, Ahmed Hallawa, Lukas Martin, Arne Peine, and Anke Schmeink. Litevsr: Efficient visual speech recognition by learning from speech representations of unlabeled data. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[25] Xubo Liu, Egor Lakomkin, Konstantinos Vougioukas, Pingchuan Ma, Honglie Chen, Ruiming Xie, Morrie Doulaty, Niko Moritz, Jachym Kolar, Stavros Petridis, et al. Synthvsr: Scaling up visual speech recognition with synthetic supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Pingchuan Ma and Alexandros et al. Haliassos. Auto-avsr: Audio-visual speech recognition with automatic labels. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2854" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[28] Pingchuan Ma and Rodrigo et al. Mira. Lira: Learning visual speech representations from audio through self-supervision. arXiv, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[29] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.292, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[31] William D Marslen-Wilson. Functional parallelism in spoken word-recognition. Cognition, 25(1-2), 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.483, + 0.362 + ], + "angle": 0, + "content": "[32] William D Marslen-Wilson and Alan Welsh. Processing interactions and lexical access during word recognition in continuous speech. Cognitive psychology, 10(1), 1978. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.363, + 0.482, + 0.391 + ], + "angle": 0, + "content": "[33] Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 1976. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.392, + 0.483, + 0.434 + ], + "angle": 0, + "content": "[34] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.436, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[35] Ara V Nefian, Luhong Liang, Xiaobo Pi, Liu Xiaoxiang, Crusoe Mao, and Kevin Murphy. A coupled hmm for audiovisual speech recognition. In 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing. IEEE, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[36] Eng-Jon Ong and Richard Bowden. Learning temporal signatures for lip reading. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops). IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[37] Stavros Petridis and Maja Pantic. Deep complementary bottleneck features for visual speech recognition. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[38] Stavros Petridis, Themos Stafylakis, Pingchuan Ma, Georgios Tzimiropoulos, and Maja Pantic. Audio-visual speech recognition with a hybrid ctc/attention architecture. In 2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[39] Javad Peymanfard, Vahid Saeedi, Mohammad Reza Mohammadi, Hossein Zeinali, and Nasser Mozayani. Leveraging visemes for better visual speech representation and lip reading, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[40] Gerasimos Potamianos, Chalapathy Neti, Guillaume Gravier, Ashutosh Garg, and Andrew W Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9), 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[41] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[42] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Speech recognition models are strong lip-readers. In Proc. Interspeech 2024, 2024. 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[43] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. arXiv, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[44] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. arXiv preprint arXiv:2201.10439, 2022. 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[45] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. *ICLR*, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[46] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. InterSpeech, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[47] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[48] Themos Stafylakis and Georgios Tzimiropoulos. Zero-shot keyword spotting for visual speech recognition in-the-wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-529, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[49] Kwanchiva Thangthai, Helen L Bear, and Richard Harvey. Comparing phonemes and visemes with dnn-based lipreading. arXiv preprint arXiv:1805.02924, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[52] Carnegie Mellon University. The cmu pronouncing dictionary, 1993. http://www.speech.cs.cmu.edu/cgi-bin/cmudict.2,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[54] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[55] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[56] Xinshuo Weng and Kris Kitani. Learning spatio-temporal features with two-stream deep 3d cnn for lipreading. arXiv preprint arXiv:1905.02540, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[57] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[58] Jeong Hun Yeo, Seunghee Han, Minsu Kim, and Yong Man Ro. Where visual speech meets language: Vsp-llm frame" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2855" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "work for efficient and context-aware visual speech processing. arXiv preprint arXiv:2402.15151, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[59] Jeong Hun Yeo, Chae Won Kim, Hyunjun Kim, Hyeongseop Rha, Seunghee Han, Wen-Huang Cheng, and Yong Man Ro. Personalized lip reading: Adapting to your unique lip movements with vision and language. arXiv preprint arXiv:2409.00986, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[60] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International conference on Computer Vision, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[61] Yuanhang Zhang, Shuang Yang, Shiguang Shan, and Xilin Chen. Es3: Evolving self-supervised learning of robust audio-visual speech representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[62] Ziheng Zhou, Guoying Zhao, Xiaopeng Hong, and Matti Pietikainen. A review of recent advances in visual speech decoding. Image and vision computing, 32(9), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[63] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, 2015. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2856" + } + ] +] \ No newline at end of file diff --git a/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_origin.pdf b/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e0fd1ff0ca9c87efd96f724951e952a088b2618 --- /dev/null +++ b/2025/VALLR_ Visual ASR Language Model for Lip Reading/88644d63-e953-49ce-b2d3-f0418686524b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c314d458f2fa8c83c775c6acd30b3e01aeec072736d457fd9c94a2db792ba2b8 +size 893483 diff --git a/2025/VALLR_ Visual ASR Language Model for Lip Reading/full.md b/2025/VALLR_ Visual ASR Language Model for Lip Reading/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6ffcd658f3dc53c28740be2866bb2fd519ad8ce --- /dev/null +++ b/2025/VALLR_ Visual ASR Language Model for Lip Reading/full.md @@ -0,0 +1,374 @@ +# VALLR: Visual ASR Language Model for Lip Reading + +Marshall Thomas Edward Fish Richard Bowden University of Surrey + +mt00893@surrey.ac.uk edward.fish@surrey.ac.uk r.bowden@surrey.ac.uk + +# Abstract + +Lip Reading, or Visual Automatic Speech Recognition (V-ASR), is a complex task requiring the interpretation of spoken language exclusively from visual cues, primarily lip movements and facial expressions. This task is especially challenging due to the absence of auditory information and the inherent ambiguity when visually distinguishing phonemes that have overlapping visemes, where different phonemes appear identical on the lips. Current methods typically attempt to predict words or characters directly from these visual cues, but this approach frequently encounters high error rates due to coarticulation effects and viseme ambiguity. We propose a novel two-stage, phoneme-centric framework for Visual Automatic Speech Recognition (V-ASR) that addresses these longstanding challenges. First, our model predicts a compact sequence of phonemes from visual inputs using a Video Transformer with a CTC head, thereby reducing the task complexity and achieving robust speaker invariance. This phoneme output then serves as the input to a fine-tuned Large Language Model (LLM), which reconstructs coherent words and sentences by leveraging broader linguistic context. Unlike existing methods that either predict words directly or rely on large-scale multimodal pre-training, our approach explicitly encodes intermediate linguistic structure while remaining highly data efficient. We demonstrate state-of-the-art performance on two challenging datasets, LRS2 and LRS3, where our method achieves significant reductions in Word Error Rate (WER) achieving a SOTA WER of 18.7 on LRS3 despite using $99.4\%$ less labelled video data than the next best approach. Code is available here: https://gitlab.surrey.ac.uk/ + +# 1. Introduction + +Lip Reading, or Visual Automatic Speech Recognition (V-ASR), involves interpreting spoken language from visual cues such as lip movements and facial expressions. As a natural skill, humans use it to supplement auditory information, and as a technology, it has profound potential for en + +![](images/9ae34c4388a4f19abe6b22865a145bffff65591cbe287bfbeb007bdb92f19060.jpg) +Figure 1. Comparison between different models' performances in WER for Visual Automatic Speech Recognition on the LRS3 dataset [3] when compared with the amount of labelled training data. Circle size (green) denotes scale of pre-training data, while fine-tuned models (grey) are not pre-trained. Our model (orange) outperforms all existing approaches with just 30 hours of video training data, no self-supervised pre-training, and without the requirement for additional labeled visual data during fine-tuning. + +hancing accessibility, particularly for the Deaf and hard-of-hearing communities, as well as for applications in noisy or privacy-sensitive environments [1]. Despite substantial advancements in ASR, the visual-only interpretation of speech remains an unresolved challenge due to inherent uncertainties in lip movements and their complex temporal dynamics. A central difficulty in V-ASR stems from the inherent ambiguity of visemes, the visual equivalents of phonemes, which often appear nearly identical for different sounds (e.g., 'p' and 'b') [1, 7]. Further complicating this task are coarticulation effects [33], where adjacent sounds blur one another's articulation, making phoneme boundaries visually unclear. Although some recent methods have focussed on robust intermediary representations to deal with these cases, such as leveraging subword-level predictions [41], or quantized latent representations [13], these approaches can still fail to capture broader contextual cues, limiting their ability to resolve visually similar phonemes that might otherwise be + +![](images/85b1f3f5dae7dfb135fece6cf0a7069536516d7e82e8e5024ee5d2691a090d08.jpg) +Figure 2. An overview of our approach. First we extract facial regions from 16 frames of input video. We apply random pixel masking at $50\%$ probability, add positional embedding, and encode visual features via a vision transformer encoder (ViT-base [14]). We implement temporal downsampling via 1D convolution and then use a CTC linear head to predict sequences of phonemes. During training, we also fine-tune an LLM to reconstruct sentences from phonemes with a text-only dataset [52]. During inference, the phonemes from the CTC head are processed via the LLM to reconstruct the predicted text. This can be performed end-to-end or in two stages, depending on available resources. + +distinguishable through sentence-level semantics. + +One approach to this challenge is to leverage large language models (LLM's) and attempt to map raw lip movements directly to sentences, thus bypassing the issue of viseme ambiguity [58, 59]. However, bridging the gap between high-dimensional, unstructured visual inputs and detailed textual representations is resource intensive and typically requires huge datasets, or specialized architectures. Moreover, these purely end-to-end pipelines often lack an interpretable intermediary step, making it unclear how the model is handling viseme ambiguity at a finer linguistic granularity. This can lead to hallucination effects, which become more critical in lip reading technologies designed for accessibility, or security applications. + +Our two-stage, phoneme-centric method addresses both issues. In the first stage, we map short windows of video frames to a discrete, interpretable phoneme sequence. This intermediate representation is significantly easier to learn and more robust against speaker or environmental variations, as phonemes abstract away speaker-specific attributes. Subsequently, in the second stage, we fine-tune an LLM for the task of reconstructing sentences from phonemes. By segmenting the problem, we combine the interpretability and efficiency of phoneme-based prediction with the sophisticated contextual reasoning of modern LLMs. This phoneme $\rightarrow$ sentence pipeline also aligns with psycholinguistic models of speech perception, where + +phoneme-level processing naturally precedes lexical access [31, 32]. + +Several advantages emerge from this design: + +(i) Compact Target Space: Predicting only 38 phoneme classes avoids learning entire word-level vocabularies, making the model more data-efficient and simpler to train. +(ii) LLM-Enhanced Accuracy: A large language model, pre-trained for phoneme $\rightarrow$ sentence reconstruction, removes the requirement of direct word-level predictions from RGB data, reducing the need for complex multimodal training. +(iii) Data Availability: Exploiting easily available phoneme $\rightarrow$ sentence text pairs eliminates reliance on extensive lip-reading video data for pre-training. +(iv) Interpretability & Error Analysis: The two-stage design yields an explicit phoneme-level output, constraining recognition errors to local phoneme segments, which the LLM subsequently repairs at the word level reducing errors at a sentence level. + +# 2. Related Work + +Here, we review key categories of approaches and highlight their advantages and limitations in the context of our proposed phoneme-centric, language-aware approach. + +Lip Reading: Early methods predominantly relied on statistical models such as Hidden Markov Models (HMMs) + +to aggregate temporal sequences of lip movements alongside hand-crafted feature extractors [1, 9, 35, 36, 40, 62]. With the advent of deep learning, coupled with the availability of large-scale datasets such as Grid [12], LRW [11], LRS2 [47], and LRS3 [3], significant progress has been made in recent years. Initially, Convolutional Neural Networks (CNNs) were applied to the task of word recognition [11], with additional temporal processing via RNN's [30, 54, 55]. With greater compute came the possibility to interpret full sentences with approaches applying Automatic Speech Recognition (ASR) methodologies to the Visual Automatic Speech Recognition (V-ASR) task. These included sequence-to-sequence models [2], CTC approaches [5, 46], and hybrid models [37, 38]. These models improved recognition accuracy by learning hierarchical representations of visual features and modelling temporal dependencies. The application of transformer models [14, 21, 44, 50, 53] in V-ASR brought further performance gains in combination with extensive self-supervised pretraining. Methods such as AV-HuBERT [45] and LiteVSR [24] leveraged self-supervised learning to learn robust representations from large amounts of unlabelled video data. + +AV-ASR vs V-ASR: Lip reading approaches can be separated into Visual Automatic Speech Recognition (V-ASR) and Audio-Visual Automatic Speech Recognition (AV-ASR). In the AV-ASR task, labelled audio data is available at training time which can be used to guide the visual encoder in various ways. These can include generating synthetic additional pre-training data [25, 27], or simply fusing audio-visual input features [44-46, 57]. Other methods have used knowledge distillation to inductively transfer knowledge from pre-trained ASR models to visual encoders [4, 28]. However, these methods depend on audio data or pre-trained ASR models, which may not be available or reliable in scenarios where lip reading is most needed, such as noisy environments. V-ASR, on the other hand, is a more challenging task that relies only on visual data in the training stage [1, 56, 60, 62]. Recent methods have also shown the effectiveness of adapting visual inputs to pre-trained ASR models [13, 24, 42] without the need for audio data during training. This method is particularly powerful since the ASR model already has the ability to reconstruct latent representations to words from the pre-training task, however it relies on large amounts of additional data to learn robust visual representations and map them to the audio latent space [42]. + +Lip Reading with LLMs: Recent approaches have explored visual attention mechanisms for sub-word units [15, 39, 41, 48] to capture fine-grained linguistic features in lip reading. However, reconstructing these units into coherent sentences remains challenging for networks lacking robust contextual and semantic reasoning capabilities. Early + +phoneme-based methods [16, 49] similarly struggled with effective decoding to word-level representations and these methods were outpaced in performance by end-to-end approaches [5] which could leverage extensive data. As such they have not been fully explored in the context of modern architectures and techniques. + +For example, recent approaches which have integrated Large Language Models (LLMs) [43, 51] into lip-reading pipelines [58, 59] have started from mapping visual features directly to LLM text embeddings. However, by omitting explicit intermediate representations, these approaches inherit phonetic ambiguities that propagate through the network. This manifests as increased word error rates and reduced robustness, thus requiring substantially more training data to achieve generalization. + +Our work re-examines phonemes as an interpretable discrete representation bridging visual and linguistic domains. Rather than forcing LLMs to perform visual $\rightarrow$ sentence translation requiring expensive visual-text alignment – we constrain the LLM's role to phoneme $\rightarrow$ sentence reconstruction. This decomposition offers three key advantages: (1) Phoneme-to-text translation is a well-constrained task requiring only lightweight LLM fine-tuning using abundant textual corpora; (2) Speaker-independent phoneme representations eliminate the need for visual adaptation in the language model; (3) The modular architecture prevents error propagation between visual and linguistic domains while maintaining interpretability. + +# 3. Method + +In this section, we present our two-stage phoneme-centric approach to visual-only lip reading. Figure 2 provides an overview of the framework. Our goal is to learn a function + +$$ +\begin{array}{l} f: \quad \mathbf {X} = \left\{x _ {1}, x _ {2}, \dots , x _ {T} \right\} \\ \longmapsto \mathbf {P h} = \left\{p h _ {1}, p h _ {2}, \dots , p h _ {m} \right\} \tag {1} \\ \longmapsto \mathbf {S} = \left\{s _ {1}, s _ {2}, \dots , s _ {M} \right\}. \\ \end{array} +$$ + +where $\mathbf{X} \in \mathbb{R}^{T \times H \times W \times 3}$ is a video sequence of $T$ frames (with height $H$ and width $W$ ), $\mathbf{Ph}$ denotes the phoneme sequence, and $\mathbf{S}$ is the reconstructed sentence. Specifically, we decompose the task into: + +1. Video $\rightarrow$ Phoneme: Map the sequence of video frames to phonemes using a Vision Transformer and CTC head. +2. Phoneme $\rightarrow$ Sentence: Convert phonemes to a coherent word sequence with a fine-tuned Large Language Model (LLM). + +The model comprises four primary components: + +- Visual Feature Extractor: Captures relevant features from each frame via a Vision Transformer. + +- Adapter Network with Temporal Downsampling: Reduces sequence length to accommodate CTC alignment. +- CTC Head: Predicts phoneme sequences without requiring explicit temporal labels. +- LLM for Phoneme-to-Sentence Reconstruction: Translates predicted phonemes into final word sequences. + +# 3.1. Pre-Processing Pipeline + +Each frame $x_{i}$ is first passed through a face-detection model [26] to identify and crop the speaker's face region, centered on the speaker's lips. The detected region is then cropped, resized to $224 \times 224$ , and normalized for batch processing. + +# 3.2. Video Transformer + +We employ a Vision Transformer (ViT) [44] to encode the spatio-temporal information in the face region. Let ViT(\cdot) denote this encoding function. The input $\mathbf{X} \in \mathbb{R}^{T \times H \times W \times 3}$ is chunked into patches, embedded, and equipped with positional encodings before passing through multiple transformer blocks. We obtain: + +$$ +\mathbf {Z} = \operatorname {V i T} (\mathbf {X}) \in \mathbb {R} ^ {T \times D}, \tag {2} +$$ + +where $D$ is the transformer output dimension per frame. During training, we randomly mask portions of the input patches to promote robust feature learning and improve generalization. + +# 3.3. Adapter Network with Temporal Downsampling + +The sequence length $T$ of the ViT output can be prohibitively large for the subsequent CTC module. To reduce the temporal dimension, we apply a sequence of 1D convolutions and pooling layers, as shown in Table ?? Formally, + +$$ +\mathbf {G} _ {\text {d o w n}} = \operatorname {A d a p t} (\mathbf {Z}) \in \mathbb {R} ^ {T ^ {\prime} \times C _ {\text {a d p a t e r}}}, \tag {3} +$$ + +where $T' < T$ and $C_{\mathrm{adapter}}$ is an intermediate feature dimension aligned to the CTC head input. + +# 3.4. CTC Head + +An MLP with one hidden layer transforms $\mathbf{G}_{\mathrm{down}} \in \mathbb{R}^{T' \times C_{\mathrm{adapter}}} \text{ into phoneme logits:}$ + +$$ +\mathbf {H} _ {\mathrm {c t c}} = \operatorname {M L P} \left(\mathbf {G} _ {\text {d o w n}}\right) \in \mathbb {R} ^ {T ^ {\prime} \times N _ {\text {p h o n e m e s}}}, \tag {4} +$$ + +where $N_{\text{phonemes}}$ is the size of our phoneme set (39 English phonemes plus a blank symbol). We then apply a logarithmic softmax along the phoneme dimension to obtain log probabilities $\log p_{\text{ctc}}(ph_t \mid \mathbf{X})$ at each time step. + +Since lip movements do not align cleanly with discrete phoneme boundaries, we adopt the Connectionist Temporal + +
StageModuleIn → Out ChannelsKernelStridePaddingDownsample FactorOutput Length
1Conv1dfeature_size → adapter_dim522×2L/2
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
2Conv1dadapter_dim → adapter_dim321×2L/4
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
3Conv1dadapter_dim → adapter_dim321×2L/8
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
4Conv1dadapter_dim → adapter_dim331×3L/24
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
5AvgPool1dadapter_dim → adapter_dim580×8L/192
6Linearadapter_dim → CTC_dim---×1L/192
ReLU----
+ +Table 1. Architecture of the adapter for temporal downsampling, where the input length is $T = 1568$ and final output length is $T = 8$ . + +Classification (CTC) loss [18]: + +$$ +\mathcal {L} _ {\mathrm {C T C}} = - \ln \left(\sum_ {\alpha \in \mathcal {A} (\mathbf {P h})} P _ {\mathrm {c t c}} (\alpha \mid \mathbf {X})\right), \tag {5} +$$ + +where $\mathbf{Ph} = (ph_1, \ldots, ph_m)$ is the ground-truth phoneme sequence, $\mathcal{A}(\mathbf{Ph})$ is the set of all valid alignments, and $P_{\mathrm{ctc}}(\alpha \mid \mathbf{X})$ is the probability of alignment $\alpha$ . Minimizing $\mathcal{L}_{\mathrm{CTC}}$ enables the model to learn frame-to-phoneme mappings without explicit per-frame labels. At inference, we perform beam search over the CTC log probabilities to decode the final phoneme sequence. + +# 3.5. Phoneme-to-Sentence Reconstruction Using LLM + +The predicted phoneme sequence $\widehat{\mathbf{Ph}}$ is derived by decoding the CTC logits from Eq. (4) via beam search. We then feed $\widehat{\mathbf{Ph}}$ into a Large Language Model (LLM) that is finetuned to map phonemes to sentences: + +$$ +\mathbf {S} = \operatorname {L L M} \left(\widehat {\mathbf {P h}}\right) = \left(s _ {1}, s _ {2}, \dots , s _ {M}\right). \tag {6} +$$ + +LoRA Fine-Tuning. To efficiently adapt the LLM, we employ Low-Rank Adaptation (LoRA) [22], injecting two low-rank matrices $\mathbf{A}$ and $\mathbf{B}$ into each linear layer. This greatly reduces the number of trainable parameters compared to full-model fine-tuning. Specifically, for a linear layer $\mathbf{W}_{\mathrm{orig}} \in \mathbb{R}^{d \times d}$ : + +$$ +\mathbf {W} ^ {\prime} = \mathbf {W} _ {\text {o r i g}} + \mathbf {A B}, \quad \mathbf {A} \in \mathbb {R} ^ {d \times r}, \mathbf {B} \in \mathbb {R} ^ {r \times d}, \tag {7} +$$ + +where $r \ll d$ . Only $\mathbf{A}$ and $\mathbf{B}$ are updated during training. + +Pretraining LLM Objective. We fine-tune the LLM on a large phoneme $\rightarrow$ sentence corpus generated from WikiText [34]. Let $\mathbf{S} = (s_1, \ldots, s_M)$ be the ground-truth sentence for a phoneme sequence $\mathbf{Ph}$ . We employ the cross-entropy loss: + +$$ +\mathcal {L} _ {\mathrm {C E}} = - \sum_ {t = 1} ^ {M} \ln p \left(s _ {t} \mid \mathbf {P h}, s _ {1: t - 1}\right), \tag {8} +$$ + +where $p(s_{t} \mid \mathbf{Ph}, s_{1:t-1})$ is the likelihood of predicting the correct word $s_{t}$ , given the phoneme sequence and previously generated words. During inference, we prompt the LLM with $\widehat{\mathbf{Ph}}$ to generate $\widehat{\mathbf{S}}$ . + +# 3.6. Training Procedure + +Stage 1: Video $\rightarrow$ Phoneme. We train the visual extractor, adapter, and CTC head to minimize $\mathcal{L}_{\mathrm{CTC}}$ (Eq. 5) on video-phoneme pairs. By learning without strict alignment constraints, the model remains flexible to a wide range of speaking speeds and styles. + +Stage 2: Phoneme $\rightarrow$ Sentence. We independently fine-tune the LLM via cross-entropy loss (Eq. 8) on text-based phoneme $\rightarrow$ sentence datasets. This leverages easily obtained text data, allowing the LLM to learn phonetic-linguistic mappings separately from visual modeling. + +Inference. We compose the trained modules to form: + +$$ +\mathbf {X} \xrightarrow {f _ {\mathrm {v i s}}} \widehat {\mathbf {P h}} \xrightarrow {f _ {\mathrm {L L M}}} \widehat {\mathbf {S}}, +$$ + +where $\widehat{\mathbf{Ph}}$ is decoded from CTC logits, and $\widehat{\mathbf{S}}$ is the final sentence. This approach is highly modular and data-efficient, as each stage (visual and linguistic) is learned with its own objective and dataset, yet easily integrated at inference time. + +# 4. Experiments + +This section outlines the datasets and pre-processing methods used for training the model architecture. + +# 4.1. Data + +The visual feature extractor, the adapter with downsampling, and CTC head are trained on the LRS2 [47] and LRS3 [3] datasets. The LLM is fine-tuned on the WikiText [34] dataset. + +LRS2 [47]: The LRS2 dataset consists of 144,482 video clips of spoken sentences from BBC television, consisting of approximately 224.5 hours of footage and each sentences is up to 100 characters in length. The videos are divided into a pre-training set with 96,318 utterances (195 hours), a training set with 45,839 utterances (28 hours), a validation set with 1,082 utterances (0.6 hours) and a test set with 1,243 utterances (0.5 hours). + +LRS3 [3]: The LRS3 dataset describes the largest public audio-visual English dataset collected consists of clips from over 5,000 TED and TEDx talks totaling 438.9 hours. It contains 438.9 hours with 151,819 utterances. Specifically, there are 118,516 utterances in the 'pre-train' set (408 hours), 31,982 utterances in the 'train-val' set (30 hours) and 1,321 utterances in the 'test' set (0.9 hours). + +In our setting, unlike other approaches, we do not utilise the pre-training set and train our model on only the 28-hour and 30-hour partitions. + +Phoneme Sentence Pairs. Finally, the dataset used for training the LLM was the WikiText [34] dataset with sentences converted into lists of phonemes using the CMU dictionary [52]. These phonemes were then masked and paired with the original sentences. + +# 4.2. Evaluation Metrics + +Following previous works [11, 47] we adopt Word Error Rate [23] (WER) as our evaluation metric for Lip-Reading. WER calculates the percentage of errors in the predicted text compared to the ground truth, accounting for substitutions, insertions, and deletions. Similar to previous studies [3, 11, 41, 42, 47], we report results on all recent AV-ASR and V-ASR approaches. + +# 5. Results + +In this section, we present the results of our method and compare them against those of existing SOTA approaches. + +Public vs Private Data on LRS3 [3]: In Table 1, we compare different V-ASR and AV-ASR models based on their WER on the LRS3 dataset, we split the models into two categories; Fully Supervised models with publicly available data and models trained on large-scale non-publicly available datasets. + +The Fully Supervised Models rely solely on publicly available labelled datasets with the best WER being $40.6\%$ , which was achieved by VTP [41]. The non-public dataset models leverage extensive, proprietary datasets, achieving much lower WERs with the best being $12.5\%$ by LP [10]. + +LRS3 [3]: In Table 1 we compare our results against other methods based on results from the LRS3 dataset. The results in Table 1 demonstrate a consistent trend: models incorporating extensive self-supervised pre-training followed by fine-tuning achieve lower WERs than those using only supervised approaches. However, our model achieves the lowest WER on the LRS3 dataset at $18.7\%$ without self-supervised pre-training or additional fine-tuning data. + +LRS2 [47]: In Table 2, we compare our method against several other state-of-the-art approaches for visual-only lip reading, focusing on Word Error Rate (WER) on the LRS2 dataset. From Table 2 we can see that increasing the amount of data available helps improve a model's WER. However, we can also see that even without increasing the size of the dataset, our model still outperforms other models that only use publicly available data, achieving the lowest WER at $20.8\%$ . + +Therefore as shown in Tables 2 and 3, we can see that our + +
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
ASR distillation [4]-59068.8
Conv-Seq2Seq [60]-85560.1
Discriminative AVSR [57]-59057.8
Hyb. + Conformer [29]-59043.3
VTP [41]-69840.6
Trained on large-scale non-publicly available datasets
Deep-AV-SR [2]-1,51958.9
Large-scale AV-SR [46]-3,88655.1
RNN-T [30]-31,00033.6
VTP [41]-2,67630.7
ViT-3D [44]-90,00017.0
LP [10]-1,000,00012.5
Self-supervised pre-training + Supervised fine-tuning on LRS3
LiRA [28]4333071.9
ASR distillation [4]33459059.8
LiteVSR [24]6395945.7
ES3 [61]4333043.5
AV-HuBERT Large [45]1,7593032.5
Lip2Vec [13]1,7593031.2
Sub-Word[41]2,6762,67630.7
SynthVSR [25]3,65243827.9
Whisper [42]1,7593025.5
Auto-AVSR [27]1,75943325.0
LLaMA-AVSR [8]-175624.0
RAVeN [19]175943323.1
USR [20]1,32643321.5
Supervised finetuning only on LRS3
Ours-3018.7
+ +![](images/2309680e9fc8df15b81d4649f6ee4748410ca448eec1e3eceae9d15922ee1af3.jpg) + +![](images/7754bf5651d54cce94bc54485751dcdadbf2b73bbffaf38ff8144f09528b3941.jpg) + +![](images/5241836870d9d791c95f8c9159eb00903de7d664b498f268b9fcd5a3aec71514.jpg) + +![](images/276819e67a971584a49e4d208e6f4455af7997117101357a4747b9be5c95485c.jpg) + +![](images/513f852c48a8ad848e828353e103328bee41306c6fde1fbc3a53f93a75960893.jpg) + +![](images/015424b644c8663954efc1f779652ecdfffb7b59c66d495a4885f69ecca87fda.jpg) + +![](images/0aa3d8102ac7868c0e1bcdc5bebd27111bf0d66867bd2c741b52605324fd8002.jpg) + +![](images/4cf7ce9a562052ae9dd55dc21fbc0e6a62e0fbb85558cbb0d60521588ee3f667.jpg) + +![](images/0a1a3fecf95803ab3592d83a37edcb4301f5f641561753b2b269d90f613aa441.jpg) + +![](images/a4c6362e1c7e6d602efec1dab005d55e5828f46a9a4f415b27088b2f964df35e.jpg) + +![](images/5af5aa88bf55dd63f55f2c27f5fb8caddf21de23c2f3aca92a2e4b6dfa0146fa.jpg) + +![](images/861cf9126b1057f0a3876f194369d5f406de253a33de9d56ccb82e45cfed29e5.jpg) + +![](images/9fa901c1b1ede2859a34c2e0b0be5fd01e9a440a82cb5245f16a3afbdcb64885.jpg) + +Table 2. Comparison of Word Error Rate (WER) across different models for visual-only speech recognition on the LRS3 dataset. The table is divided into three sections: fully supervised models trained on publicly available data, models trained on large-scale non-public datasets, and models that leverage self-supervised pre-training with supervised fine-tuning on LRS3 [3]. Each entry details the amount of labelled and unlabelled data used and the resulting WER. Our approach, tested with various language models, achieves the lowest WER of $22.1\%$ with the Llama 3.2-3B model, demonstrating the effectiveness of our phoneme-based, LLM-assisted approach. + +
Predicted +PhonemeW, AA, ZDH, AHK, AY, N, DAH, V
Predicted +WordWasthekindof
GTWasthekindof
Predicted +PhonemeAH, B, S, EH, SH, +AH, NDH, AE, THH, AE, DHH, ER
Predicted +Wordobsessionthathadher
GTobsessionthathadher
+ +Figure 3. Example of the model's phonetic and sentence outputs from a sample in the LRS3 dataset [3]. The table illustrates the model's ability to predict a sequence of phonemes from visual input, which are then reconstructed into a coherent sentence by the LLM. In this example all of the phonemes are predicted correctly and the words are recreated correctly. + +
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
Auto-AV-SR [27]-81827.9
Auto-AV-SR [27]-344814.6
Self-supervised pre-training + Supervised finetuning on LRS2 for AVSR
LiRA [28]1,75943338.8
ES3 [61]1,75922326.7
Sub-Word [41]2,6762,67622.6
RAVeN [19]1,75922317.9
USR [20]1,75922315.4
Supervised finetuning only on LRS2
Ours-2820.8
+ +Table 3. Comparison of Word Error Rate (WER) for various self-supervised pre-training and supervised fine-tuning methods evaluated on the LRS2 [47] dataset. Our method achieves competitive performance with other methods, without the requirement of extensive pre-training, additional labelled data, or additional audio modality as in the best performing approaches. We are the only method to train on only the LRS2 train dataset. + +method provides a lower WER with the same amount, or less data, without the requirement of self-supervised pretraining as current approaches [4, 13, 24, 25, 27, 28, 42, 45]. This large increase in performance and generalisation can be explained by the powerful combination of the multi-task objectives and the ability of the LLM to reconstruct sentences accurately even when phonemes are predicted incorrectly. In Fig. 4, we find that the phonetic output of the visual encoder is sometimes incorrect or misses phonemes, but the LLM is able to still reconstruct the correct word (as in the example $z \rightarrow s$ ). We also show how the model still struggles with tricky homophones such as your and you're. In Fig. 3, we show an additional example of the model's outputs at both phonetic and word levels, demonstrating how phonemes are correctly predicted and combined by the LLM for sentence reconstruction. For further examples please see the appendix. + +# 5.1. Ablations + +In this section, we aim to understand the performance contribution of each component in the network and their performance on visual $\rightarrow$ phoneme prediction and phoneme $\rightarrow$ word reconstruction, alongside ablations related to architectural decisions, and the limitations of the approach. + +Visual $\rightarrow$ Phoneme: The confusion matrix in Fig. 5 shows phoneme prediction performance of the model on the LRS3 [3] dataset. The matrix indicates that some phonemes are misclassified more frequently than others, as seen by the off-diagonal elements, with the worst missed classifications being highlighted with a red bounding box. This bias could be due to the similarity in the visual representation of these sounds, making it challenging for the model to distinguish between them. For instance, /S/ and /Z/ may often be con- + +![](images/4e8fc1ec0b0814031a2d09ff8bf89e914bb5512b94eccdc80d0511cedb025fbc.jpg) +Figure 4. Comparison of the model's phonetic and sentence outputs with ground truth from a sample in the LRS3 [3] dataset. Red shows an incorrect prediction, M shows a missing prediction and green shows a correction from phonemes to words. In this example, even though the model incorrectly predicts certain phonemes, the LLM can correctly recreate the word but struggles to recreate homophones. + +
Predicted +PhonemeY, UH, RJH, AH, Z, TW, AH, NAH, V
Predicted +WordYourjustoneof
GTYou'rejustoneof
Predicted +PhonemeDH, AHP, IY, P, <M>, LHH, UWHH, AE, S, N, T
Predicted +Wordthepeoplewhohasn't
GTthepeoplewhohasn't
+ +fused due to similar lip movements, leading to miss classifications between these phonemes. + +![](images/f7d26b30e521eb6a97cd06f3b94f654e12b2cf2e2cd055439660d384e40abed9.jpg) +Figure 5. Confusion matrix showing the performance on isolated phonemes of the LRS3 [3] dataset. We observe a very high match rate between the predicted phonemes and the ground truth. In red, we show the most difficult phonemes for our model to identify. + +Phoneme $\rightarrow$ Word: In Table 4 we take the top 5 words in alphabetical order that are misclassified when predicting words directly from phonemes after the LLM pretraining stage. We can observe that these are words that may occur infrequently in the training set and are missclassified due to some phonetic similarities between the words. + +Different LLMs: In this ablation study, we evaluate three different Language models on our model architecture to + +
PhonemesTrue WordPredicted Word
/EH/ /R/ /AH/ /N/ /S/ /AH/ /N/AaronsonErasmus
/AA/ /B/ /ER/ /G/AabergBraggartly
/AH/ /S/ /EH/ /R /AH/AcerraEsraa
/AE/ /S/ /AH/ /T/ /EY /T/AcetateAsquette
/AH/ /K/ /AH/ /S/ /T/ /AH/ /M/ /D/AccustomedAccosted
+ +understand the impact on word error rate (WER) performance. The first model, GPT 2 [43], serves as our baseline, consisting of the default GPT 2 [43] without fine-tuning on lip-reading datasets. This model achieves a WER of $23.9\%$ , indicating the core components' foundational capabilities. The second model, Llama 3.2-1B [51] introduces fine-tuning on phoneme word pairs, enhancing the model's contextual understanding of phoneme sequences. With this addition, Llama 3.2-1B [51] has a substantial improvement, reducing the WER to $22.8\%$ . Finally, Llama 3.2-3B [51] also incorporates fine-tuning for phoneme-to-word reconstruction. However, Llama 3.2-3B [51] is a larger model with more parameters than the Llama 3.2-1B [51] model, further lowering the WER to $18.7\%$ . This progression highlights the significant contributions of fine-tuning for purpose-specific LLMs. + +Table 4. Examples of misclassification's between predicted words and ground truth based on phoneme sequences. Each row presents a sequence of phonemes, the intended (true) word, and the model's predicted word. Differences in predictions illustrate common misclassifications, often due to phonetic similarities between visually indistinct sounds. + +
ModelParameters (B)WER (%)
GPT-2 Small [43]0.1233.9
Llama 3.2-1B [51]1.022.8
Llama 3.2-3B [51]3.018.7
+ +Table 5. Comparison of Word Error Rate (WER) for different large language models (LLMs) tested with our model. The table includes each model's parameter count (in billions) and resulting WER on the LRS3 [3] dataset. GPT-2 [43] Small with 0.12B parameters has the highest WER at $23.9\%$ . In contrast, Llama 3.2-3B [51] achieves the best WER of $18.7\%$ , highlighting that increased model capacity improves phoneme-to-word reconstruction accuracy. + +Freezing the CTC Head: Inspired by works that directly map visual features to pretrained ASR networks [42] we investigate the effect of initialising the CTC head in our model with weights from a pre-trained Wav2Vec2 ASR model [6]. The CTC head plays a crucial role in mapping visual features to the phoneme sequences, and we hypothesized that initialising and freezing its parameters could reduce computational overhead while preserving feature alignment from audio pre-training. To test this hypothesis a baseline model + +with a non-frozen CTC head and a comparison model with a frozen CTC head were trained and a comparison of the results was made as shown in Table 6. Freezing the CTC head led to a slight increase in WER of $0.4\%$ , indicating a decrease in performance. This result suggests that while the frozen CTC head still provided a foundation for phoneme alignment, we did not require extensive audio pre-training to obtain superior performance. + +Varying Sample Sizes: To understand the impact of additional data on performance, we vary the quantity of data used to train both the Video $\rightarrow$ Phoneme network and the Phoneme $\rightarrow$ Sentence LLM and investigate the effect this has on the WER when evaluated on the LRS3 [3] dataset. By varying the percentages of data used in training, we can see the effect on the model in Figure 6. Decreasing the amount of data used to train both models has the expected effect on the WER by increasing it. However, increasing the amount of training data for the Feature Extractor has a negligible effect after 30 hours of data. However, by increasing the amount of training data for the LLM we can also increase the effectiveness of the model and decrease the total WER down further to $17.5\%$ . The additional data is randomly sampled from both AvSpeech [17] for the visual encoder, and a BookCorpus [63] replica to match the quantity of data in WikiText. + +![](images/9039c2c35aa9c82d656a47cc63694fed24bdefa3cff3157abb122551594fff59.jpg) +Figure 6. Graph showing the effect of varying the sample size of the training datasets for both the Video $\rightarrow$ Phoneme network (in blue) and for the Phoneme $\rightarrow$ Sentence LLM (in red). The graph shows that decreasing the amount of data negatively impacts the model for both networks, increasing the amount of training data for the Video $\rightarrow$ Phoneme network has negligible effect and increasing the amount of training data for the Phoneme $\rightarrow$ Sentence LLM has a noticeable positive effect. + +End-To-End: For our final study, we investigate the effect of removing the video $\rightarrow$ phoneme stage and train the + +model to directly predict sentences from the video features, reproducing the method in [41] but with the same ViT encoder [44], LLama LLM [51], and limited data set. As shown in Table 6 we achieve aWER of $25.6\%$ on LRS3, demonstrating that the more parameterised transformer encoder improves performance over the CNN encoder used in [41]. However, our two stage approach still performs significantly better. + +
ModelWER (%)
Ours (Two Stage)18.7
ViT + LLM (End to End)25.6
Sub-Word (End to End) [41]30.7
+ +Table 6. Comparison of Word Error Rate (WER) between using the intermediary Phonemes and skipping the CTC head. Removing the intermediary representation increases the WER from 18.7 to 25.6, showing the importance of the intermediary step and the effectiveness of the phoneme centric fine-tuning. + +# 6. Conclusion and Limitations + +We have introduced a two-stage, phoneme-centric framework for visual-only lip reading that first predicts phonemes from lip movements and then reconstructs coherent sentences via a fine-tuned LLM. This design sharply reduces word error rates on benchmark datasets (LRS2 [47] and LRS3 [3]), providing improvements of $3.5\%$ and $6.8\%$ WER, respectively, over existing state-of-the-art approaches that focus on end-to-end connectionist approaches. + +Despite these gains, the approach remains limited by the visual ambiguity of certain phonemes. Lip movements alone cannot always differentiate minimal pairs such as /s/ vs. /z/, and words that share near-identical phoneme sequences (e.g., "Hello" vs. "Hallow") can still result in misclassification. We plan to address these issues by refining the phoneme-to-word generation process, potentially through deeper linguistic context modeling and more specialized phoneme embeddings. Future work may also explore speaker-adaptive fine-tuning or multi-frame alignment strategies to better handle subtle visual distinctions. + +# Acknowledgements + +This work was supported by the SNSF project 'SMILE II' (CRSII5 193686), the Innosuisse IICT Flagship (PFFS-21-47), EPSRC grant APP24554 (SignGPT-EP/Z535370/1) and through funding from Google.org via the AI for Global Goals scheme. This work reflects only the author's views and the funders are not responsible for any use that may be made of the information it contains. + +# References + +[1] Emre Afacan and Gültekin Demirci. A survey on automatic lip-reading: Challenges and future directions. Multimedia Tools and Applications, 2019. 1, 3 +[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 3, 6 +[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 3, 5, 6, 7, 8 +[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 3, 6 +[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 3 +[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 2020. 7 +[7] Helen L Bear, Richard W Harvey, Barry-John Theobald, and Yuxuan Lan. Which phoneme-to-viseme maps best improve visual-only computer lip-reading? In International symposium on visual computing. Springer, 2014. 1 +[8] Umberto Cappellazzo, Minsu Kim, Honglie Chen, Pingchuan Ma, Stavros Petridis, Daniele Falavigna, Alessio Brutti, and Maja Pantic. Large language models are strong audio-visual speech recognition learners, 2025. 6 +[9] Luca Cappelletta and Naomi Harte. Viseme definitions comparison for visual-only speech recognition. In 2011 19th European Signal Processing Conference. IEEE, 2011. 3 +[10] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 5, 6 +[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016. 3, 5 +[12] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. Grid av speech corpus sample. The Journal of the Accoustical Society of America, 2013. 3 +[13] Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Haithem Boussaid, Ebtessam Almazrouei, and Merouane Debbah. Lip2vec: Efficient and robust visual speech recognition via latent-to-latent visual to audio representation mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6 +[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[15] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H. Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent + +speech recognition. CAAI Transactions on Intelligence Technology, 8(1):129-138, 2023. 3 +[16] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent speech recognition. CAAI Transactions on Intelligence Technology, 2023. 3 +[17] A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K Wilson, A. Hassidim, W. T. Freeman, and M. Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 8 +[18] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, 2006. 4 +[19] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. arXiv preprint arXiv:2212.06246, 2022. 6 +[20] Alexandros Haliassos, Rodrigo Mira, Honglie Chen, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Unified speech recognition: A single model for auditory, visual, and audio-visual inputs. arXiv preprint arXiv:2411.02256, 2024. 6 +[21] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 2021. 3 +[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2022. 4 +[23] F. Jelinek, L. Bahl, and R. Mercer. Design of a linguistic statistical decoder for the recognition of continuous speech. IEEE Transactions on Information Theory, 21(3), 1975. 5 +[24] Hendrik Laux, Emil Mededovic, Ahmed Hallawa, Lukas Martin, Arne Peine, and Anke Schmeink. Litevsr: Efficient visual speech recognition by learning from speech representations of unlabeled data. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 3, 6 +[25] Xubo Liu, Egor Lakomkin, Konstantinos Vougioukas, Pingchuan Ma, Honglie Chen, Ruiming Xie, Morrie Doulaty, Niko Moritz, Jachym Kolar, Stavros Petridis, et al. Synthvsr: Scaling up visual speech recognition with synthetic supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 6 +[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4 +[27] Pingchuan Ma and Alexandros et al. Haliassos. Auto-avsr: Audio-visual speech recognition with automatic labels. In + +IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. 3, 6 +[28] Pingchuan Ma and Rodrigo et al. Mira. Lira: Learning visual speech representations from audio through self-supervision. arXiv, 2021. 3, 6 +[29] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021. 6 +[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2019. 3, 6 +[31] William D Marslen-Wilson. Functional parallelism in spoken word-recognition. Cognition, 25(1-2), 1987. 2 +[32] William D Marslen-Wilson and Alan Welsh. Processing interactions and lexical access during word recognition in continuous speech. Cognitive psychology, 10(1), 1978. 2 +[33] Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 1976. 1 +[34] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 4, 5 +[35] Ara V Nefian, Luhong Liang, Xiaobo Pi, Liu Xiaoxiang, Crusoe Mao, and Kevin Murphy. A coupled hmm for audiovisual speech recognition. In 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing. IEEE, 2002. 3 +[36] Eng-Jon Ong and Richard Bowden. Learning temporal signatures for lip reading. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops). IEEE, 2011. 3 +[37] Stavros Petridis and Maja Pantic. Deep complementary bottleneck features for visual speech recognition. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3 +[38] Stavros Petridis, Themos Stafylakis, Pingchuan Ma, Georgios Tzimiropoulos, and Maja Pantic. Audio-visual speech recognition with a hybrid ctc/attention architecture. In 2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018. 3 +[39] Javad Peymanfard, Vahid Saeedi, Mohammad Reza Mohammadi, Hossein Zeinali, and Nasser Mozayani. Leveraging visemes for better visual speech representation and lip reading, 2023. 3 +[40] Gerasimos Potamianos, Chalapathy Neti, Guillaume Gravier, Ashutosh Garg, and Andrew W Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9), 2003. 3 +[41] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 5, 6, 8 +[42] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Speech recognition models are strong lip-readers. In Proc. Interspeech 2024, 2024. 3, 5, 6, 7 + +[43] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. arXiv, 2018. 3, 7 +[44] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. arXiv preprint arXiv:2201.10439, 2022. 3, 4, 6, 8 +[45] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. *ICLR*, 2022. 3, 6 +[46] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. InterSpeech, 2018. 3, 6 +[47] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3, 5, 6, 8 +[48] Themos Stafylakis and Georgios Tzimiropoulos. Zero-shot keyword spotting for visual speech recognition in-the-wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-529, 2018. 3 +[49] Kwanchiva Thangthai, Helen L Bear, and Richard Harvey. Comparing phonemes and visemes with dnn-based lipreading. arXiv preprint arXiv:1805.02924, 2018. 3 +[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 3 +[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 7, 8 +[52] Carnegie Mellon University. The cmu pronouncing dictionary, 1993. http://www.speech.cs.cmu.edu/cgi-bin/cmudict.2,5 +[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 3 +[54] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3 +[55] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3 +[56] Xinshuo Weng and Kris Kitani. Learning spatio-temporal features with two-stream deep 3d cnn for lipreading. arXiv preprint arXiv:1905.02540, 2019. 3 +[57] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2020. 3, 6 +[58] Jeong Hun Yeo, Seunghee Han, Minsu Kim, and Yong Man Ro. Where visual speech meets language: Vsp-llm frame + +work for efficient and context-aware visual speech processing. arXiv preprint arXiv:2402.15151, 2024. 2, 3 +[59] Jeong Hun Yeo, Chae Won Kim, Hyunjun Kim, Hyeongseop Rha, Seunghee Han, Wen-Huang Cheng, and Yong Man Ro. Personalized lip reading: Adapting to your unique lip movements with vision and language. arXiv preprint arXiv:2409.00986, 2024. 2, 3 +[60] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International conference on Computer Vision, 2019. 3, 6 +[61] Yuanhang Zhang, Shuang Yang, Shiguang Shan, and Xilin Chen. Es3: Evolving self-supervised learning of robust audio-visual speech representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6 +[62] Ziheng Zhou, Guoying Zhao, Xiaopeng Hong, and Matti Pietikainen. A review of recent advances in visual speech decoding. Image and vision computing, 32(9), 2014. 3 +[63] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, 2015. 8 \ No newline at end of file diff --git a/2025/VALLR_ Visual ASR Language Model for Lip Reading/images.zip b/2025/VALLR_ Visual ASR Language Model for Lip Reading/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dd2f98291b318790b38eebffaebf06f11a0ac5d0 --- /dev/null +++ b/2025/VALLR_ Visual ASR Language Model for Lip Reading/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0556c17ccf71be7e1012a391f3b4b84247ff5e661599d1da12f4b89a7c77124f +size 540249 diff --git a/2025/VALLR_ Visual ASR Language Model for Lip Reading/layout.json b/2025/VALLR_ Visual ASR Language Model for Lip Reading/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c3c6b98791182694668764e061eb3292a67df1f1 --- /dev/null +++ b/2025/VALLR_ Visual ASR Language Model for Lip Reading/layout.json @@ -0,0 +1,9517 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 475, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 475, + 121 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 475, + 121 + ], + "type": "text", + "content": "VALLR: Visual ASR Language Model for Lip Reading" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 144, + 434, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 144, + 434, + 172 + ], + "spans": [ + { + "bbox": [ + 176, + 144, + 434, + 172 + ], + "type": "text", + "content": "Marshall Thomas Edward Fish Richard Bowden University of Surrey" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 174, + 493, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 174, + 493, + 185 + ], + "spans": [ + { + "bbox": [ + 111, + 174, + 493, + 185 + ], + "type": "text", + "content": "mt00893@surrey.ac.uk edward.fish@surrey.ac.uk r.bowden@surrey.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 203, + 200, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 203, + 200, + 216 + ], + "spans": [ + { + "bbox": [ + 152, + 203, + 200, + 216 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 236, + 297, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 297, + 606 + ], + "type": "text", + "content": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), is a complex task requiring the interpretation of spoken language exclusively from visual cues, primarily lip movements and facial expressions. This task is especially challenging due to the absence of auditory information and the inherent ambiguity when visually distinguishing phonemes that have overlapping visemes, where different phonemes appear identical on the lips. Current methods typically attempt to predict words or characters directly from these visual cues, but this approach frequently encounters high error rates due to coarticulation effects and viseme ambiguity. We propose a novel two-stage, phoneme-centric framework for Visual Automatic Speech Recognition (V-ASR) that addresses these longstanding challenges. First, our model predicts a compact sequence of phonemes from visual inputs using a Video Transformer with a CTC head, thereby reducing the task complexity and achieving robust speaker invariance. This phoneme output then serves as the input to a fine-tuned Large Language Model (LLM), which reconstructs coherent words and sentences by leveraging broader linguistic context. Unlike existing methods that either predict words directly or rely on large-scale multimodal pre-training, our approach explicitly encodes intermediate linguistic structure while remaining highly data efficient. We demonstrate state-of-the-art performance on two challenging datasets, LRS2 and LRS3, where our method achieves significant reductions in Word Error Rate (WER) achieving a SOTA WER of 18.7 on LRS3 despite using " + }, + { + "bbox": [ + 53, + 236, + 297, + 606 + ], + "type": "inline_equation", + "content": "99.4\\%" + }, + { + "bbox": [ + 53, + 236, + 297, + 606 + ], + "type": "text", + "content": " less labelled video data than the next best approach. Code is available here: https://gitlab.surrey.ac.uk/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 634, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Lip Reading, or Visual Automatic Speech Recognition (V-ASR), involves interpreting spoken language from visual cues such as lip movements and facial expressions. As a natural skill, humans use it to supplement auditory information, and as a technology, it has profound potential for en" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 205, + 559, + 365 + ], + "blocks": [ + { + "bbox": [ + 319, + 205, + 559, + 365 + ], + "lines": [ + { + "bbox": [ + 319, + 205, + 559, + 365 + ], + "spans": [ + { + "bbox": [ + 319, + 205, + 559, + 365 + ], + "type": "image", + "image_path": "9ae34c4388a4f19abe6b22865a145bffff65591cbe287bfbeb007bdb92f19060.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 376, + 555, + 467 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 467 + ], + "type": "text", + "content": "Figure 1. Comparison between different models' performances in WER for Visual Automatic Speech Recognition on the LRS3 dataset [3] when compared with the amount of labelled training data. Circle size (green) denotes scale of pre-training data, while fine-tuned models (grey) are not pre-trained. Our model (orange) outperforms all existing approaches with just 30 hours of video training data, no self-supervised pre-training, and without the requirement for additional labeled visual data during fine-tuning." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "hancing accessibility, particularly for the Deaf and hard-of-hearing communities, as well as for applications in noisy or privacy-sensitive environments [1]. Despite substantial advancements in ASR, the visual-only interpretation of speech remains an unresolved challenge due to inherent uncertainties in lip movements and their complex temporal dynamics. A central difficulty in V-ASR stems from the inherent ambiguity of visemes, the visual equivalents of phonemes, which often appear nearly identical for different sounds (e.g., 'p' and 'b') [1, 7]. Further complicating this task are coarticulation effects [33], where adjacent sounds blur one another's articulation, making phoneme boundaries visually unclear. Although some recent methods have focussed on robust intermediary representations to deal with these cases, such as leveraging subword-level predictions [41], or quantized latent representations [13], these approaches can still fail to capture broader contextual cues, limiting their ability to resolve visually similar phonemes that might otherwise be" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2846" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 555, + 264 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 264 + ], + "type": "image", + "image_path": "85b1f3f5dae7dfb135fece6cf0a7069536516d7e82e8e5024ee5d2691a090d08.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 279, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 279, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 555, + 346 + ], + "type": "text", + "content": "Figure 2. An overview of our approach. First we extract facial regions from 16 frames of input video. We apply random pixel masking at " + }, + { + "bbox": [ + 54, + 279, + 555, + 346 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 279, + 555, + 346 + ], + "type": "text", + "content": " probability, add positional embedding, and encode visual features via a vision transformer encoder (ViT-base [14]). We implement temporal downsampling via 1D convolution and then use a CTC linear head to predict sequences of phonemes. During training, we also fine-tune an LLM to reconstruct sentences from phonemes with a text-only dataset [52]. During inference, the phonemes from the CTC head are processed via the LLM to reconstruct the predicted text. This can be performed end-to-end or in two stages, depending on available resources." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 367, + 255, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 255, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 255, + 379 + ], + "type": "text", + "content": "distinguishable through sentence-level semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 391, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 391, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 54, + 391, + 295, + 547 + ], + "type": "text", + "content": "One approach to this challenge is to leverage large language models (LLM's) and attempt to map raw lip movements directly to sentences, thus bypassing the issue of viseme ambiguity [58, 59]. However, bridging the gap between high-dimensional, unstructured visual inputs and detailed textual representations is resource intensive and typically requires huge datasets, or specialized architectures. Moreover, these purely end-to-end pipelines often lack an interpretable intermediary step, making it unclear how the model is handling viseme ambiguity at a finer linguistic granularity. This can lead to hallucination effects, which become more critical in lip reading technologies designed for accessibility, or security applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 557, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 715 + ], + "type": "text", + "content": "Our two-stage, phoneme-centric method addresses both issues. In the first stage, we map short windows of video frames to a discrete, interpretable phoneme sequence. This intermediate representation is significantly easier to learn and more robust against speaker or environmental variations, as phonemes abstract away speaker-specific attributes. Subsequently, in the second stage, we fine-tune an LLM for the task of reconstructing sentences from phonemes. By segmenting the problem, we combine the interpretability and efficiency of phoneme-based prediction with the sophisticated contextual reasoning of modern LLMs. This phoneme " + }, + { + "bbox": [ + 54, + 557, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 557, + 295, + 715 + ], + "type": "text", + "content": " sentence pipeline also aligns with psycholinguistic models of speech perception, where" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 367, + 555, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 391 + ], + "type": "text", + "content": "phoneme-level processing naturally precedes lexical access [31, 32]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 397, + 495, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 495, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 495, + 409 + ], + "type": "text", + "content": "Several advantages emerge from this design:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 415, + 555, + 617 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 309, + 415, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 415, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 309, + 415, + 555, + 461 + ], + "type": "text", + "content": "(i) Compact Target Space: Predicting only 38 phoneme classes avoids learning entire word-level vocabularies, making the model more data-efficient and simpler to train." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 463, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 554, + 521 + ], + "type": "text", + "content": "(ii) LLM-Enhanced Accuracy: A large language model, pre-trained for phoneme " + }, + { + "bbox": [ + 307, + 463, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 307, + 463, + 554, + 521 + ], + "type": "text", + "content": " sentence reconstruction, removes the requirement of direct word-level predictions from RGB data, reducing the need for complex multimodal training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 553, + 559 + ], + "type": "text", + "content": "(iii) Data Availability: Exploiting easily available phoneme " + }, + { + "bbox": [ + 304, + 522, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 522, + 553, + 559 + ], + "type": "text", + "content": " sentence text pairs eliminates reliance on extensive lip-reading video data for pre-training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 559, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 553, + 617 + ], + "type": "text", + "content": "(iv) Interpretability & Error Analysis: The two-stage design yields an explicit phoneme-level output, constraining recognition errors to local phoneme segments, which the LLM subsequently repairs at the word level reducing errors at a sentence level." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 628, + 400, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 400, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 400, + 639 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 647, + 555, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 684 + ], + "type": "text", + "content": "Here, we review key categories of approaches and highlight their advantages and limitations in the context of our proposed phoneme-centric, language-aware approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Lip Reading: Early methods predominantly relied on statistical models such as Hidden Markov Models (HMMs)" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2847" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 336 + ], + "type": "text", + "content": "to aggregate temporal sequences of lip movements alongside hand-crafted feature extractors [1, 9, 35, 36, 40, 62]. With the advent of deep learning, coupled with the availability of large-scale datasets such as Grid [12], LRW [11], LRS2 [47], and LRS3 [3], significant progress has been made in recent years. Initially, Convolutional Neural Networks (CNNs) were applied to the task of word recognition [11], with additional temporal processing via RNN's [30, 54, 55]. With greater compute came the possibility to interpret full sentences with approaches applying Automatic Speech Recognition (ASR) methodologies to the Visual Automatic Speech Recognition (V-ASR) task. These included sequence-to-sequence models [2], CTC approaches [5, 46], and hybrid models [37, 38]. These models improved recognition accuracy by learning hierarchical representations of visual features and modelling temporal dependencies. The application of transformer models [14, 21, 44, 50, 53] in V-ASR brought further performance gains in combination with extensive self-supervised pretraining. Methods such as AV-HuBERT [45] and LiteVSR [24] leveraged self-supervised learning to learn robust representations from large amounts of unlabelled video data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 345, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 633 + ], + "type": "text", + "content": "AV-ASR vs V-ASR: Lip reading approaches can be separated into Visual Automatic Speech Recognition (V-ASR) and Audio-Visual Automatic Speech Recognition (AV-ASR). In the AV-ASR task, labelled audio data is available at training time which can be used to guide the visual encoder in various ways. These can include generating synthetic additional pre-training data [25, 27], or simply fusing audio-visual input features [44-46, 57]. Other methods have used knowledge distillation to inductively transfer knowledge from pre-trained ASR models to visual encoders [4, 28]. However, these methods depend on audio data or pre-trained ASR models, which may not be available or reliable in scenarios where lip reading is most needed, such as noisy environments. V-ASR, on the other hand, is a more challenging task that relies only on visual data in the training stage [1, 56, 60, 62]. Recent methods have also shown the effectiveness of adapting visual inputs to pre-trained ASR models [13, 24, 42] without the need for audio data during training. This method is particularly powerful since the ASR model already has the ability to reconstruct latent representations to words from the pre-training task, however it relies on large amounts of additional data to learn robust visual representations and map them to the audio latent space [42]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Lip Reading with LLMs: Recent approaches have explored visual attention mechanisms for sub-word units [15, 39, 41, 48] to capture fine-grained linguistic features in lip reading. However, reconstructing these units into coherent sentences remains challenging for networks lacking robust contextual and semantic reasoning capabilities. Early" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "phoneme-based methods [16, 49] similarly struggled with effective decoding to word-level representations and these methods were outpaced in performance by end-to-end approaches [5] which could leverage extensive data. As such they have not been fully explored in the context of modern architectures and techniques." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 258 + ], + "type": "text", + "content": "For example, recent approaches which have integrated Large Language Models (LLMs) [43, 51] into lip-reading pipelines [58, 59] have started from mapping visual features directly to LLM text embeddings. However, by omitting explicit intermediate representations, these approaches inherit phonetic ambiguities that propagate through the network. This manifests as increased word error rates and reduced robustness, thus requiring substantially more training data to achieve generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 265, + 556, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 556, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 556, + 421 + ], + "type": "text", + "content": "Our work re-examines phonemes as an interpretable discrete representation bridging visual and linguistic domains. Rather than forcing LLMs to perform visual " + }, + { + "bbox": [ + 313, + 265, + 556, + 421 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 265, + 556, + 421 + ], + "type": "text", + "content": " sentence translation requiring expensive visual-text alignment – we constrain the LLM's role to phoneme " + }, + { + "bbox": [ + 313, + 265, + 556, + 421 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 265, + 556, + 421 + ], + "type": "text", + "content": " sentence reconstruction. This decomposition offers three key advantages: (1) Phoneme-to-text translation is a well-constrained task requiring only lightweight LLM fine-tuning using abundant textual corpora; (2) Speaker-independent phoneme representations eliminate the need for visual adaptation in the language model; (3) The modular architecture prevents error propagation between visual and linguistic domains while maintaining interpretability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 431, + 370, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 370, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 370, + 443 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 451, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 554, + 487 + ], + "type": "text", + "content": "In this section, we present our two-stage phoneme-centric approach to visual-only lip reading. Figure 2 provides an overview of the framework. Our goal is to learn a function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 497, + 554, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 497, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 347, + 497, + 554, + 540 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f: \\quad \\mathbf {X} = \\left\\{x _ {1}, x _ {2}, \\dots , x _ {T} \\right\\} \\\\ \\longmapsto \\mathbf {P h} = \\left\\{p h _ {1}, p h _ {2}, \\dots , p h _ {m} \\right\\} \\tag {1} \\\\ \\longmapsto \\mathbf {S} = \\left\\{s _ {1}, s _ {2}, \\dots , s _ {M} \\right\\}. \\\\ \\end{array}", + "image_path": "1c4e0152e61f3485d387c2cc5bcb8e2f4a1c92ddbb74501827a936af39131e6b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "content": " is a video sequence of " + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "content": " frames (with height " + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{Ph}" + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "content": " denotes the phoneme sequence, and " + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 549, + 554, + 598 + ], + "type": "text", + "content": " is the reconstructed sentence. Specifically, we decompose the task into:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 605, + 554, + 664 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 317, + 605, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 605, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 317, + 605, + 554, + 628 + ], + "type": "text", + "content": "1. Video " + }, + { + "bbox": [ + 317, + 605, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 317, + 605, + 554, + 628 + ], + "type": "text", + "content": " Phoneme: Map the sequence of video frames to phonemes using a Vision Transformer and CTC head." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 629, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 629, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 317, + 629, + 554, + 664 + ], + "type": "text", + "content": "2. Phoneme " + }, + { + "bbox": [ + 317, + 629, + 554, + 664 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 317, + 629, + 554, + 664 + ], + "type": "text", + "content": " Sentence: Convert phonemes to a coherent word sequence with a fine-tuned Large Language Model (LLM)." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 671, + 511, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 511, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 511, + 684 + ], + "type": "text", + "content": "The model comprises four primary components:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "- Visual Feature Extractor: Captures relevant features from each frame via a Vision Transformer." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2848" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "- Adapter Network with Temporal Downsampling: Reduces sequence length to accommodate CTC alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 294, + 120 + ], + "type": "text", + "content": "- CTC Head: Predicts phoneme sequences without requiring explicit temporal labels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 144 + ], + "type": "text", + "content": "- LLM for Phoneme-to-Sentence Reconstruction: Translates predicted phonemes into final word sequences." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 152, + 188, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 188, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 188, + 165 + ], + "type": "text", + "content": "3.1. Pre-Processing Pipeline" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 170, + 294, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 294, + 229 + ], + "type": "text", + "content": "Each frame " + }, + { + "bbox": [ + 55, + 170, + 294, + 229 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 170, + 294, + 229 + ], + "type": "text", + "content": " is first passed through a face-detection model [26] to identify and crop the speaker's face region, centered on the speaker's lips. The detected region is then cropped, resized to " + }, + { + "bbox": [ + 55, + 170, + 294, + 229 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 170, + 294, + 229 + ], + "type": "text", + "content": ", and normalized for batch processing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 237, + 167, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 167, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 167, + 249 + ], + "type": "text", + "content": "3.2. Video Transformer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 255, + 294, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 294, + 326 + ], + "type": "text", + "content": "We employ a Vision Transformer (ViT) [44] to encode the spatio-temporal information in the face region. Let ViT(\\cdot) denote this encoding function. The input " + }, + { + "bbox": [ + 55, + 255, + 294, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 255, + 294, + 326 + ], + "type": "text", + "content": " is chunked into patches, embedded, and equipped with positional encodings before passing through multiple transformer blocks. We obtain:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 335, + 294, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 335, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 119, + 335, + 294, + 350 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\operatorname {V i T} (\\mathbf {X}) \\in \\mathbb {R} ^ {T \\times D}, \\tag {2}", + "image_path": "d98f970f738a90fde074ddb4e45b53b8ea4fcbae49faa5b480ef340f5ba38ad0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "text", + "content": " is the transformer output dimension per frame. During training, we randomly mask portions of the input patches to promote robust feature learning and improve generalization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 415, + 295, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 441 + ], + "type": "text", + "content": "3.3. Adapter Network with Temporal Downsampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 445, + 295, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 494 + ], + "type": "text", + "content": "The sequence length " + }, + { + "bbox": [ + 55, + 445, + 295, + 494 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 445, + 295, + 494 + ], + "type": "text", + "content": " of the ViT output can be prohibitively large for the subsequent CTC module. To reduce the temporal dimension, we apply a sequence of 1D convolutions and pooling layers, as shown in Table ?? Formally," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 97, + 502, + 294, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 502, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 97, + 502, + 294, + 517 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {\\text {d o w n}} = \\operatorname {A d a p t} (\\mathbf {Z}) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times C _ {\\text {a d p a t e r}}}, \\tag {3}", + "image_path": "0a33c65084aaf963147e3c4918ff05ab8d8b6d0d865dba5cdba3a2eed2f292ae.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 527, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 527, + 295, + 552 + ], + "type": "inline_equation", + "content": "T' < T" + }, + { + "bbox": [ + 55, + 527, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 527, + 295, + 552 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{adapter}}" + }, + { + "bbox": [ + 55, + 527, + 295, + 552 + ], + "type": "text", + "content": " is an intermediate feature dimension aligned to the CTC head input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 559, + 129, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 129, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 129, + 571 + ], + "type": "text", + "content": "3.4. CTC Head" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "text", + "content": "An MLP with one hidden layer transforms " + }, + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{down}} \\in \\mathbb{R}^{T' \\times C_{\\mathrm{adapter}}} \\text{ into phoneme logits:}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 88, + 610, + 294, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 610, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 88, + 610, + 294, + 625 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {\\mathrm {c t c}} = \\operatorname {M L P} \\left(\\mathbf {G} _ {\\text {d o w n}}\\right) \\in \\mathbb {R} ^ {T ^ {\\prime} \\times N _ {\\text {p h o n e m e s}}}, \\tag {4}", + "image_path": "e6347d20a541ac7d0e217c412e04fef76cad5fdf8b0dec15c141181dbd205210.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "inline_equation", + "content": "N_{\\text{phonemes}}" + }, + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "text", + "content": " is the size of our phoneme set (39 English phonemes plus a blank symbol). We then apply a logarithmic softmax along the phoneme dimension to obtain log probabilities " + }, + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\log p_{\\text{ctc}}(ph_t \\mid \\mathbf{X})" + }, + { + "bbox": [ + 55, + 635, + 295, + 684 + ], + "type": "text", + "content": " at each time step." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Since lip movements do not align cleanly with discrete phoneme boundaries, we adopt the Connectionist Temporal" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 565, + 182 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 565, + 182 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 565, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 565, + 182 + ], + "type": "table", + "html": "
StageModuleIn → Out ChannelsKernelStridePaddingDownsample FactorOutput Length
1Conv1dfeature_size → adapter_dim522×2L/2
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
2Conv1dadapter_dim → adapter_dim321×2L/4
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
3Conv1dadapter_dim → adapter_dim321×2L/8
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
4Conv1dadapter_dim → adapter_dim331×3L/24
BatchNorm1dadapter_dim → adapter_dim---
ReLU----
5AvgPool1dadapter_dim → adapter_dim580×8L/192
6Linearadapter_dim → CTC_dim---×1L/192
ReLU----
", + "image_path": "1b981d1a719a18feb087fe2df03f3b6b59d1263151fbcdd9f66245eeb8421ddb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 190, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 222 + ], + "type": "text", + "content": "Table 1. Architecture of the adapter for temporal downsampling, where the input length is " + }, + { + "bbox": [ + 313, + 190, + 553, + 222 + ], + "type": "inline_equation", + "content": "T = 1568" + }, + { + "bbox": [ + 313, + 190, + 553, + 222 + ], + "type": "text", + "content": " and final output length is " + }, + { + "bbox": [ + 313, + 190, + 553, + 222 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 313, + 190, + 553, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 242, + 440, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 242, + 440, + 254 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 440, + 254 + ], + "type": "text", + "content": "Classification (CTC) loss [18]:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 353, + 258, + 553, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 258, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 353, + 258, + 553, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C T C}} = - \\ln \\left(\\sum_ {\\alpha \\in \\mathcal {A} (\\mathbf {P h})} P _ {\\mathrm {c t c}} (\\alpha \\mid \\mathbf {X})\\right), \\tag {5}", + "image_path": "e828c004fa14e4bbe1ffccee8205180351bb807c59dc39275620f70abd710a44.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{Ph} = (ph_1, \\ldots, ph_m)" + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "text", + "content": " is the ground-truth phoneme sequence, " + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\mathbf{Ph})" + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "text", + "content": " is the set of all valid alignments, and " + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ctc}}(\\alpha \\mid \\mathbf{X})" + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "text", + "content": " is the probability of alignment " + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "text", + "content": ". Minimizing " + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CTC}}" + }, + { + "bbox": [ + 313, + 293, + 555, + 377 + ], + "type": "text", + "content": " enables the model to learn frame-to-phoneme mappings without explicit per-frame labels. At inference, we perform beam search over the CTC log probabilities to decode the final phoneme sequence." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 383, + 553, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 553, + 407 + ], + "type": "text", + "content": "3.5. Phoneme-to-Sentence Reconstruction Using LLM" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 411, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 461 + ], + "type": "text", + "content": "The predicted phoneme sequence " + }, + { + "bbox": [ + 313, + 411, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Ph}}" + }, + { + "bbox": [ + 313, + 411, + 554, + 461 + ], + "type": "text", + "content": " is derived by decoding the CTC logits from Eq. (4) via beam search. We then feed " + }, + { + "bbox": [ + 313, + 411, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Ph}}" + }, + { + "bbox": [ + 313, + 411, + 554, + 461 + ], + "type": "text", + "content": " into a Large Language Model (LLM) that is finetuned to map phonemes to sentences:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 353, + 467, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 467, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 353, + 467, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\operatorname {L L M} \\left(\\widehat {\\mathbf {P h}}\\right) = \\left(s _ {1}, s _ {2}, \\dots , s _ {M}\\right). \\tag {6}", + "image_path": "2cd34f4a1f29a870095a9b7a9b7e88242a8be188da04af7fa981ffb33613566f.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "text", + "content": "LoRA Fine-Tuning. To efficiently adapt the LLM, we employ Low-Rank Adaptation (LoRA) [22], injecting two low-rank matrices " + }, + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "text", + "content": " into each linear layer. This greatly reduces the number of trainable parameters compared to full-model fine-tuning. Specifically, for a linear layer " + }, + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{orig}} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 494, + 554, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 328, + 572, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 572, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 328, + 572, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {W} ^ {\\prime} = \\mathbf {W} _ {\\text {o r i g}} + \\mathbf {A B}, \\quad \\mathbf {A} \\in \\mathbb {R} ^ {d \\times r}, \\mathbf {B} \\in \\mathbb {R} ^ {r \\times d}, \\tag {7}", + "image_path": "ccebbd6ae82c0ccf4ced8b2f917942d22b673431ca3a83a6965884577103095f.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "inline_equation", + "content": "r \\ll d" + }, + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "text", + "content": ". Only " + }, + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 592, + 547, + 604 + ], + "type": "text", + "content": " are updated during training." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": "Pretraining LLM Objective. We fine-tune the LLM on a large phoneme " + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": " sentence corpus generated from WikiText [34]. Let " + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = (s_1, \\ldots, s_M)" + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": " be the ground-truth sentence for a phoneme sequence " + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{Ph}" + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": ". We employ the cross-entropy loss:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 358, + 683, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 683, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 358, + 683, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {t = 1} ^ {M} \\ln p \\left(s _ {t} \\mid \\mathbf {P h}, s _ {1: t - 1}\\right), \\tag {8}", + "image_path": "848cf371eeba7b5f880bfcf43d1c8149917d637217d67f488bf1a031fb1a6cbe.jpg" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2849" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "p(s_{t} \\mid \\mathbf{Ph}, s_{1:t-1})" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": " is the likelihood of predicting the correct word " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "s_{t}" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": ", given the phoneme sequence and previously generated words. During inference, we prompt the LLM with " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Ph}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{S}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 170, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 170, + 142 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 170, + 142 + ], + "type": "text", + "content": "3.6. Training Procedure" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 295, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 207 + ], + "type": "text", + "content": "Stage 1: Video " + }, + { + "bbox": [ + 55, + 146, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 146, + 295, + 207 + ], + "type": "text", + "content": " Phoneme. We train the visual extractor, adapter, and CTC head to minimize " + }, + { + "bbox": [ + 55, + 146, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CTC}}" + }, + { + "bbox": [ + 55, + 146, + 295, + 207 + ], + "type": "text", + "content": " (Eq. 5) on video-phoneme pairs. By learning without strict alignment constraints, the model remains flexible to a wide range of speaking speeds and styles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 217, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 277 + ], + "type": "text", + "content": "Stage 2: Phoneme " + }, + { + "bbox": [ + 55, + 217, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 217, + 295, + 277 + ], + "type": "text", + "content": " Sentence. We independently fine-tune the LLM via cross-entropy loss (Eq. 8) on text-based phoneme " + }, + { + "bbox": [ + 55, + 217, + 295, + 277 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 217, + 295, + 277 + ], + "type": "text", + "content": " sentence datasets. This leverages easily obtained text data, allowing the LLM to learn phonetic-linguistic mappings separately from visual modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 286, + 271, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 271, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 271, + 298 + ], + "type": "text", + "content": "Inference. We compose the trained modules to form:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 308, + 227, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 308, + 227, + 323 + ], + "spans": [ + { + "bbox": [ + 124, + 308, + 227, + 323 + ], + "type": "interline_equation", + "content": "\\mathbf {X} \\xrightarrow {f _ {\\mathrm {v i s}}} \\widehat {\\mathbf {P h}} \\xrightarrow {f _ {\\mathrm {L L M}}} \\widehat {\\mathbf {S}},", + "image_path": "d97a329befb0d97f356cbbcd302d4d8d1363e4c1882506d7d2ad95ea4b768c5d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 332, + 295, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 295, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 332, + 295, + 394 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Ph}}" + }, + { + "bbox": [ + 55, + 332, + 295, + 394 + ], + "type": "text", + "content": " is decoded from CTC logits, and " + }, + { + "bbox": [ + 55, + 332, + 295, + 394 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{S}}" + }, + { + "bbox": [ + 55, + 332, + 295, + 394 + ], + "type": "text", + "content": " is the final sentence. This approach is highly modular and data-efficient, as each stage (visual and linguistic) is learned with its own objective and dataset, yet easily integrated at inference time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 405, + 137, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 137, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 137, + 418 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 449 + ], + "type": "text", + "content": "This section outlines the datasets and pre-processing methods used for training the model architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 456, + 101, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 101, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 101, + 468 + ], + "type": "text", + "content": "4.1. Data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 521 + ], + "type": "text", + "content": "The visual feature extractor, the adapter with downsampling, and CTC head are trained on the LRS2 [47] and LRS3 [3] datasets. The LLM is fine-tuned on the WikiText [34] dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 528, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 624 + ], + "type": "text", + "content": "LRS2 [47]: The LRS2 dataset consists of 144,482 video clips of spoken sentences from BBC television, consisting of approximately 224.5 hours of footage and each sentences is up to 100 characters in length. The videos are divided into a pre-training set with 96,318 utterances (195 hours), a training set with 45,839 utterances (28 hours), a validation set with 1,082 utterances (0.6 hours) and a test set with 1,243 utterances (0.5 hours)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "LRS3 [3]: The LRS3 dataset describes the largest public audio-visual English dataset collected consists of clips from over 5,000 TED and TEDx talks totaling 438.9 hours. It contains 438.9 hours with 151,819 utterances. Specifically, there are 118,516 utterances in the 'pre-train' set (408 hours), 31,982 utterances in the 'train-val' set (30 hours) and 1,321 utterances in the 'test' set (0.9 hours)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "In our setting, unlike other approaches, we do not utilise the pre-training set and train our model on only the 28-hour and 30-hour partitions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "type": "text", + "content": "Phoneme Sentence Pairs. Finally, the dataset used for training the LLM was the WikiText [34] dataset with sentences converted into lists of phonemes using the CMU dictionary [52]. These phonemes were then masked and paired with the original sentences." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 181, + 427, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 427, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 427, + 193 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 198, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 282 + ], + "type": "text", + "content": "Following previous works [11, 47] we adopt Word Error Rate [23] (WER) as our evaluation metric for Lip-Reading. WER calculates the percentage of errors in the predicted text compared to the ground truth, accounting for substitutions, insertions, and deletions. Similar to previous studies [3, 11, 41, 42, 47], we report results on all recent AV-ASR and V-ASR approaches." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 293, + 367, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 293, + 367, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 293, + 367, + 304 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 313, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 553, + 338 + ], + "type": "text", + "content": "In this section, we present the results of our method and compare them against those of existing SOTA approaches." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 342, + 554, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 414 + ], + "type": "text", + "content": "Public vs Private Data on LRS3 [3]: In Table 1, we compare different V-ASR and AV-ASR models based on their WER on the LRS3 dataset, we split the models into two categories; Fully Supervised models with publicly available data and models trained on large-scale non-publicly available datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 420, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 480 + ], + "type": "text", + "content": "The Fully Supervised Models rely solely on publicly available labelled datasets with the best WER being " + }, + { + "bbox": [ + 313, + 420, + 553, + 480 + ], + "type": "inline_equation", + "content": "40.6\\%" + }, + { + "bbox": [ + 313, + 420, + 553, + 480 + ], + "type": "text", + "content": ", which was achieved by VTP [41]. The non-public dataset models leverage extensive, proprietary datasets, achieving much lower WERs with the best being " + }, + { + "bbox": [ + 313, + 420, + 553, + 480 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 313, + 420, + 553, + 480 + ], + "type": "text", + "content": " by LP [10]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "type": "text", + "content": "LRS3 [3]: In Table 1 we compare our results against other methods based on results from the LRS3 dataset. The results in Table 1 demonstrate a consistent trend: models incorporating extensive self-supervised pre-training followed by fine-tuning achieve lower WERs than those using only supervised approaches. However, our model achieves the lowest WER on the LRS3 dataset at " + }, + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 313, + 486, + 554, + 582 + ], + "type": "text", + "content": " without self-supervised pre-training or additional fine-tuning data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 587, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 554, + 694 + ], + "type": "text", + "content": "LRS2 [47]: In Table 2, we compare our method against several other state-of-the-art approaches for visual-only lip reading, focusing on Word Error Rate (WER) on the LRS2 dataset. From Table 2 we can see that increasing the amount of data available helps improve a model's WER. However, we can also see that even without increasing the size of the dataset, our model still outperforms other models that only use publicly available data, achieving the lowest WER at " + }, + { + "bbox": [ + 313, + 587, + 554, + 694 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 313, + 587, + 554, + 694 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "text", + "content": "Therefore as shown in Tables 2 and 3, we can see that our" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2850" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 304, + 365 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 304, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 304, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 304, + 365 + ], + "type": "table", + "html": "
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
ASR distillation [4]-59068.8
Conv-Seq2Seq [60]-85560.1
Discriminative AVSR [57]-59057.8
Hyb. + Conformer [29]-59043.3
VTP [41]-69840.6
Trained on large-scale non-publicly available datasets
Deep-AV-SR [2]-1,51958.9
Large-scale AV-SR [46]-3,88655.1
RNN-T [30]-31,00033.6
VTP [41]-2,67630.7
ViT-3D [44]-90,00017.0
LP [10]-1,000,00012.5
Self-supervised pre-training + Supervised fine-tuning on LRS3
LiRA [28]4333071.9
ASR distillation [4]33459059.8
LiteVSR [24]6395945.7
ES3 [61]4333043.5
AV-HuBERT Large [45]1,7593032.5
Lip2Vec [13]1,7593031.2
Sub-Word[41]2,6762,67630.7
SynthVSR [25]3,65243827.9
Whisper [42]1,7593025.5
Auto-AVSR [27]1,75943325.0
LLaMA-AVSR [8]-175624.0
RAVeN [19]175943323.1
USR [20]1,32643321.5
Supervised finetuning only on LRS3
Ours-3018.7
", + "image_path": "4bb0f2d946f065d9dbc0973f67487c4e9b6a98182344b3591187de9a2b301982.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 503, + 104, + 529 + ], + "blocks": [ + { + "bbox": [ + 78, + 503, + 104, + 529 + ], + "lines": [ + { + "bbox": [ + 78, + 503, + 104, + 529 + ], + "spans": [ + { + "bbox": [ + 78, + 503, + 104, + 529 + ], + "type": "image", + "image_path": "2309680e9fc8df15b81d4649f6ee4748410ca448eec1e3eceae9d15922ee1af3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 105, + 504, + 119, + 529 + ], + "blocks": [ + { + "bbox": [ + 105, + 504, + 119, + 529 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 119, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 119, + 529 + ], + "type": "image", + "image_path": "7754bf5651d54cce94bc54485751dcdadbf2b73bbffaf38ff8144f09528b3941.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 120, + 503, + 138, + 529 + ], + "blocks": [ + { + "bbox": [ + 120, + 503, + 138, + 529 + ], + "lines": [ + { + "bbox": [ + 120, + 503, + 138, + 529 + ], + "spans": [ + { + "bbox": [ + 120, + 503, + 138, + 529 + ], + "type": "image", + "image_path": "5241836870d9d791c95f8c9159eb00903de7d664b498f268b9fcd5a3aec71514.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 138, + 503, + 155, + 529 + ], + "blocks": [ + { + "bbox": [ + 138, + 503, + 155, + 529 + ], + "lines": [ + { + "bbox": [ + 138, + 503, + 155, + 529 + ], + "spans": [ + { + "bbox": [ + 138, + 503, + 155, + 529 + ], + "type": "image", + "image_path": "276819e67a971584a49e4d208e6f4455af7997117101357a4747b9be5c95485c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 156, + 503, + 172, + 529 + ], + "blocks": [ + { + "bbox": [ + 156, + 503, + 172, + 529 + ], + "lines": [ + { + "bbox": [ + 156, + 503, + 172, + 529 + ], + "spans": [ + { + "bbox": [ + 156, + 503, + 172, + 529 + ], + "type": "image", + "image_path": "513f852c48a8ad848e828353e103328bee41306c6fde1fbc3a53f93a75960893.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 173, + 503, + 189, + 529 + ], + "blocks": [ + { + "bbox": [ + 173, + 503, + 189, + 529 + ], + "lines": [ + { + "bbox": [ + 173, + 503, + 189, + 529 + ], + "spans": [ + { + "bbox": [ + 173, + 503, + 189, + 529 + ], + "type": "image", + "image_path": "015424b644c8663954efc1f779652ecdfffb7b59c66d495a4885f69ecca87fda.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 189, + 503, + 205, + 529 + ], + "blocks": [ + { + "bbox": [ + 189, + 503, + 205, + 529 + ], + "lines": [ + { + "bbox": [ + 189, + 503, + 205, + 529 + ], + "spans": [ + { + "bbox": [ + 189, + 503, + 205, + 529 + ], + "type": "image", + "image_path": "0aa3d8102ac7868c0e1bcdc5bebd27111bf0d66867bd2c741b52605324fd8002.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 503, + 222, + 529 + ], + "blocks": [ + { + "bbox": [ + 206, + 503, + 222, + 529 + ], + "lines": [ + { + "bbox": [ + 206, + 503, + 222, + 529 + ], + "spans": [ + { + "bbox": [ + 206, + 503, + 222, + 529 + ], + "type": "image", + "image_path": "4cf7ce9a562052ae9dd55dc21fbc0e6a62e0fbb85558cbb0d60521588ee3f667.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 503, + 239, + 529 + ], + "blocks": [ + { + "bbox": [ + 223, + 503, + 239, + 529 + ], + "lines": [ + { + "bbox": [ + 223, + 503, + 239, + 529 + ], + "spans": [ + { + "bbox": [ + 223, + 503, + 239, + 529 + ], + "type": "image", + "image_path": "0a1a3fecf95803ab3592d83a37edcb4301f5f641561753b2b269d90f613aa441.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 240, + 503, + 257, + 529 + ], + "blocks": [ + { + "bbox": [ + 240, + 503, + 257, + 529 + ], + "lines": [ + { + "bbox": [ + 240, + 503, + 257, + 529 + ], + "spans": [ + { + "bbox": [ + 240, + 503, + 257, + 529 + ], + "type": "image", + "image_path": "a4c6362e1c7e6d602efec1dab005d55e5828f46a9a4f415b27088b2f964df35e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 257, + 503, + 273, + 529 + ], + "blocks": [ + { + "bbox": [ + 257, + 503, + 273, + 529 + ], + "lines": [ + { + "bbox": [ + 257, + 503, + 273, + 529 + ], + "spans": [ + { + "bbox": [ + 257, + 503, + 273, + 529 + ], + "type": "image", + "image_path": "5af5aa88bf55dd63f55f2c27f5fb8caddf21de23c2f3aca92a2e4b6dfa0146fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 274, + 503, + 290, + 529 + ], + "blocks": [ + { + "bbox": [ + 274, + 503, + 290, + 529 + ], + "lines": [ + { + "bbox": [ + 274, + 503, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 274, + 503, + 290, + 529 + ], + "type": "image", + "image_path": "861cf9126b1057f0a3876f194369d5f406de253a33de9d56ccb82e45cfed29e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 290, + 503, + 306, + 529 + ], + "blocks": [ + { + "bbox": [ + 290, + 503, + 306, + 529 + ], + "lines": [ + { + "bbox": [ + 290, + 503, + 306, + 529 + ], + "spans": [ + { + "bbox": [ + 290, + 503, + 306, + 529 + ], + "type": "image", + "image_path": "9fa901c1b1ede2859a34c2e0b0be5fd01e9a440a82cb5245f16a3afbdcb64885.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 58, + 536, + 306, + 613 + ], + "blocks": [ + { + "bbox": [ + 55, + 373, + 296, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 495 + ], + "type": "text", + "content": "Table 2. Comparison of Word Error Rate (WER) across different models for visual-only speech recognition on the LRS3 dataset. The table is divided into three sections: fully supervised models trained on publicly available data, models trained on large-scale non-public datasets, and models that leverage self-supervised pre-training with supervised fine-tuning on LRS3 [3]. Each entry details the amount of labelled and unlabelled data used and the resulting WER. Our approach, tested with various language models, achieves the lowest WER of " + }, + { + "bbox": [ + 55, + 373, + 296, + 495 + ], + "type": "inline_equation", + "content": "22.1\\%" + }, + { + "bbox": [ + 55, + 373, + 296, + 495 + ], + "type": "text", + "content": " with the Llama 3.2-3B model, demonstrating the effectiveness of our phoneme-based, LLM-assisted approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 536, + 306, + 613 + ], + "lines": [ + { + "bbox": [ + 58, + 536, + 306, + 613 + ], + "spans": [ + { + "bbox": [ + 58, + 536, + 306, + 613 + ], + "type": "table", + "html": "
Predicted \nPhonemeW, AA, ZDH, AHK, AY, N, DAH, V
Predicted \nWordWasthekindof
GTWasthekindof
Predicted \nPhonemeAH, B, S, EH, SH, \nAH, NDH, AE, THH, AE, DHH, ER
Predicted \nWordobsessionthathadher
GTobsessionthathadher
", + "image_path": "4409cabae598a4e370a820bb58e44f6d466d251279aa3f089505cb0d8a8076d4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 563, + 206 + ], + "blocks": [ + { + "bbox": [ + 55, + 621, + 295, + 687 + ], + "lines": [ + { + "bbox": [ + 55, + 621, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 295, + 687 + ], + "type": "text", + "content": "Figure 3. Example of the model's phonetic and sentence outputs from a sample in the LRS3 dataset [3]. The table illustrates the model's ability to predict a sequence of phonemes from visual input, which are then reconstructed into a coherent sentence by the LLM. In this example all of the phonemes are predicted correctly and the words are recreated correctly." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 563, + 206 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 563, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 563, + 206 + ], + "type": "table", + "html": "
MethodUnlabeled (hrs)Labeled (hrs)WER (%)
Fully supervised models with publicly available data
Auto-AV-SR [27]-81827.9
Auto-AV-SR [27]-344814.6
Self-supervised pre-training + Supervised finetuning on LRS2 for AVSR
LiRA [28]1,75943338.8
ES3 [61]1,75922326.7
Sub-Word [41]2,6762,67622.6
RAVeN [19]1,75922317.9
USR [20]1,75922315.4
Supervised finetuning only on LRS2
Ours-2820.8
", + "image_path": "c720ad6324dd4c6cc3507a3073064c3edc26b7b5fb6b9cf8752d83efaa366110.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 214, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 292 + ], + "type": "text", + "content": "Table 3. Comparison of Word Error Rate (WER) for various self-supervised pre-training and supervised fine-tuning methods evaluated on the LRS2 [47] dataset. Our method achieves competitive performance with other methods, without the requirement of extensive pre-training, additional labelled data, or additional audio modality as in the best performing approaches. We are the only method to train on only the LRS2 train dataset." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 312, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 516 + ], + "type": "text", + "content": "method provides a lower WER with the same amount, or less data, without the requirement of self-supervised pretraining as current approaches [4, 13, 24, 25, 27, 28, 42, 45]. This large increase in performance and generalisation can be explained by the powerful combination of the multi-task objectives and the ability of the LLM to reconstruct sentences accurately even when phonemes are predicted incorrectly. In Fig. 4, we find that the phonetic output of the visual encoder is sometimes incorrect or misses phonemes, but the LLM is able to still reconstruct the correct word (as in the example " + }, + { + "bbox": [ + 313, + 312, + 555, + 516 + ], + "type": "inline_equation", + "content": "z \\rightarrow s" + }, + { + "bbox": [ + 313, + 312, + 555, + 516 + ], + "type": "text", + "content": "). We also show how the model still struggles with tricky homophones such as your and you're. In Fig. 3, we show an additional example of the model's outputs at both phonetic and word levels, demonstrating how phonemes are correctly predicted and combined by the LLM for sentence reconstruction. For further examples please see the appendix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 382, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 382, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 382, + 533 + ], + "type": "text", + "content": "5.1. Ablations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "text", + "content": "In this section, we aim to understand the performance contribution of each component in the network and their performance on visual " + }, + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "text", + "content": " phoneme prediction and phoneme " + }, + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 540, + 554, + 601 + ], + "type": "text", + "content": " word reconstruction, alongside ablations related to architectural decisions, and the limitations of the approach." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Visual " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " Phoneme: The confusion matrix in Fig. 5 shows phoneme prediction performance of the model on the LRS3 [3] dataset. The matrix indicates that some phonemes are misclassified more frequently than others, as seen by the off-diagonal elements, with the worst missed classifications being highlighted with a red bounding box. This bias could be due to the similarity in the visual representation of these sounds, making it challenging for the model to distinguish between them. For instance, /S/ and /Z/ may often be con-" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2851" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 69, + 306, + 95 + ], + "blocks": [ + { + "bbox": [ + 78, + 69, + 306, + 95 + ], + "lines": [ + { + "bbox": [ + 78, + 69, + 306, + 95 + ], + "spans": [ + { + "bbox": [ + 78, + 69, + 306, + 95 + ], + "type": "image", + "image_path": "4e8fc1ec0b0814031a2d09ff8bf89e914bb5512b94eccdc80d0511cedb025fbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 187, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 266 + ], + "type": "text", + "content": "Figure 4. Comparison of the model's phonetic and sentence outputs with ground truth from a sample in the LRS3 [3] dataset. Red shows an incorrect prediction, M shows a missing prediction and green shows a correction from phonemes to words. In this example, even though the model incorrectly predicts certain phonemes, the LLM can correctly recreate the word but struggles to recreate homophones." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 102, + 306, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 102, + 306, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 102, + 306, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 102, + 306, + 178 + ], + "type": "table", + "html": "
Predicted \nPhonemeY, UH, RJH, AH, Z, TW, AH, NAH, V
Predicted \nWordYourjustoneof
GTYou'rejustoneof
Predicted \nPhonemeDH, AHP, IY, P, <M>, LHH, UWHH, AE, S, N, T
Predicted \nWordthepeoplewhohasn't
GTthepeoplewhohasn't
", + "image_path": "8e711fe5e567d49e2d32bb3875ffaa3e363b360ab9840fb1300f0144e21978e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 309 + ], + "type": "text", + "content": "fused due to similar lip movements, leading to miss classifications between these phonemes." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 55, + 317, + 307, + 548 + ], + "blocks": [ + { + "bbox": [ + 55, + 317, + 307, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 307, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 307, + 548 + ], + "type": "image", + "image_path": "f7d26b30e521eb6a97cd06f3b94f654e12b2cf2e2cd055439660d384e40abed9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 555, + 295, + 601 + ], + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 601 + ], + "type": "text", + "content": "Figure 5. Confusion matrix showing the performance on isolated phonemes of the LRS3 [3] dataset. We observe a very high match rate between the predicted phonemes and the ground truth. In red, we show the most difficult phonemes for our model to identify." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 684 + ], + "type": "text", + "content": "Phoneme " + }, + { + "bbox": [ + 55, + 611, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 611, + 295, + 684 + ], + "type": "text", + "content": " Word: In Table 4 we take the top 5 words in alphabetical order that are misclassified when predicting words directly from phonemes after the LLM pretraining stage. We can observe that these are words that may occur infrequently in the training set and are missclassified due to some phonetic similarities between the words." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Different LLMs: In this ablation study, we evaluate three different Language models on our model architecture to" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 69, + 563, + 133 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 563, + 133 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 563, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 563, + 133 + ], + "type": "table", + "html": "
PhonemesTrue WordPredicted Word
/EH/ /R/ /AH/ /N/ /S/ /AH/ /N/AaronsonErasmus
/AA/ /B/ /ER/ /G/AabergBraggartly
/AH/ /S/ /EH/ /R /AH/AcerraEsraa
/AE/ /S/ /AH/ /T/ /EY /T/AcetateAsquette
/AH/ /K/ /AH/ /S/ /T/ /AH/ /M/ /D/AccustomedAccosted
", + "image_path": "9d3d4061eb0bfc8f590cb45171de428e906f503bc03b483afafcb52a7cd45f4b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "text", + "content": "understand the impact on word error rate (WER) performance. The first model, GPT 2 [43], serves as our baseline, consisting of the default GPT 2 [43] without fine-tuning on lip-reading datasets. This model achieves a WER of " + }, + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "inline_equation", + "content": "23.9\\%" + }, + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "text", + "content": ", indicating the core components' foundational capabilities. The second model, Llama 3.2-1B [51] introduces fine-tuning on phoneme word pairs, enhancing the model's contextual understanding of phoneme sequences. With this addition, Llama 3.2-1B [51] has a substantial improvement, reducing the WER to " + }, + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "inline_equation", + "content": "22.8\\%" + }, + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "text", + "content": ". Finally, Llama 3.2-3B [51] also incorporates fine-tuning for phoneme-to-word reconstruction. However, Llama 3.2-3B [51] is a larger model with more parameters than the Llama 3.2-1B [51] model, further lowering the WER to " + }, + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 313, + 230, + 555, + 422 + ], + "type": "text", + "content": ". This progression highlights the significant contributions of fine-tuning for purpose-specific LLMs." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 433, + 547, + 495 + ], + "blocks": [ + { + "bbox": [ + 313, + 141, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 141, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 555, + 207 + ], + "type": "text", + "content": "Table 4. Examples of misclassification's between predicted words and ground truth based on phoneme sequences. Each row presents a sequence of phonemes, the intended (true) word, and the model's predicted word. Differences in predictions illustrate common misclassifications, often due to phonetic similarities between visually indistinct sounds." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 433, + 547, + 495 + ], + "lines": [ + { + "bbox": [ + 323, + 433, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 323, + 433, + 547, + 495 + ], + "type": "table", + "html": "
ModelParameters (B)WER (%)
GPT-2 Small [43]0.1233.9
Llama 3.2-1B [51]1.022.8
Llama 3.2-3B [51]3.018.7
", + "image_path": "cb11ee012b8e5c7758268e42e9f87c27a56e49644ba4c76ffcde0a43b1da4199.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 502, + 555, + 591 + ], + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 591 + ], + "type": "text", + "content": "Table 5. Comparison of Word Error Rate (WER) for different large language models (LLMs) tested with our model. The table includes each model's parameter count (in billions) and resulting WER on the LRS3 [3] dataset. GPT-2 [43] Small with 0.12B parameters has the highest WER at " + }, + { + "bbox": [ + 313, + 502, + 555, + 591 + ], + "type": "inline_equation", + "content": "23.9\\%" + }, + { + "bbox": [ + 313, + 502, + 555, + 591 + ], + "type": "text", + "content": ". In contrast, Llama 3.2-3B [51] achieves the best WER of " + }, + { + "bbox": [ + 313, + 502, + 555, + 591 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 313, + 502, + 555, + 591 + ], + "type": "text", + "content": ", highlighting that increased model capacity improves phoneme-to-word reconstruction accuracy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Freezing the CTC Head: Inspired by works that directly map visual features to pretrained ASR networks [42] we investigate the effect of initialising the CTC head in our model with weights from a pre-trained Wav2Vec2 ASR model [6]. The CTC head plays a crucial role in mapping visual features to the phoneme sequences, and we hypothesized that initialising and freezing its parameters could reduce computational overhead while preserving feature alignment from audio pre-training. To test this hypothesis a baseline model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2852" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "with a non-frozen CTC head and a comparison model with a frozen CTC head were trained and a comparison of the results was made as shown in Table 6. Freezing the CTC head led to a slight increase in WER of " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", indicating a decrease in performance. This result suggests that while the frozen CTC head still provided a foundation for phoneme alignment, we did not require extensive audio pre-training to obtain superior performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "text", + "content": "Varying Sample Sizes: To understand the impact of additional data on performance, we vary the quantity of data used to train both the Video " + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "text", + "content": " Phoneme network and the Phoneme " + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "text", + "content": " Sentence LLM and investigate the effect this has on the WER when evaluated on the LRS3 [3] dataset. By varying the percentages of data used in training, we can see the effect on the model in Figure 6. Decreasing the amount of data used to train both models has the expected effect on the WER by increasing it. However, increasing the amount of training data for the Feature Extractor has a negligible effect after 30 hours of data. However, by increasing the amount of training data for the LLM we can also increase the effectiveness of the model and decrease the total WER down further to " + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "inline_equation", + "content": "17.5\\%" + }, + { + "bbox": [ + 56, + 175, + 296, + 378 + ], + "type": "text", + "content": ". The additional data is randomly sampled from both AvSpeech [17] for the visual encoder, and a BookCorpus [63] replica to match the quantity of data in WikiText." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 402, + 284, + 571 + ], + "blocks": [ + { + "bbox": [ + 70, + 402, + 284, + 571 + ], + "lines": [ + { + "bbox": [ + 70, + 402, + 284, + 571 + ], + "spans": [ + { + "bbox": [ + 70, + 402, + 284, + 571 + ], + "type": "image", + "image_path": "9039c2c35aa9c82d656a47cc63694fed24bdefa3cff3157abb122551594fff59.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "text", + "content": "Figure 6. Graph showing the effect of varying the sample size of the training datasets for both the Video " + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "text", + "content": " Phoneme network (in blue) and for the Phoneme " + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "text", + "content": " Sentence LLM (in red). The graph shows that decreasing the amount of data negatively impacts the model for both networks, increasing the amount of training data for the Video " + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "text", + "content": " Phoneme network has negligible effect and increasing the amount of training data for the Phoneme " + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 585, + 295, + 673 + ], + "type": "text", + "content": " Sentence LLM has a noticeable positive effect." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "End-To-End: For our final study, we investigate the effect of removing the video " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " phoneme stage and train the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "model to directly predict sentences from the video features, reproducing the method in [41] but with the same ViT encoder [44], LLama LLM [51], and limited data set. As shown in Table 6 we achieve aWER of " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "25.6\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": " on LRS3, demonstrating that the more parameterised transformer encoder improves performance over the CNN encoder used in [41]. However, our two stage approach still performs significantly better." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 343, + 180, + 526, + 242 + ], + "blocks": [ + { + "bbox": [ + 343, + 180, + 526, + 242 + ], + "lines": [ + { + "bbox": [ + 343, + 180, + 526, + 242 + ], + "spans": [ + { + "bbox": [ + 343, + 180, + 526, + 242 + ], + "type": "table", + "html": "
ModelWER (%)
Ours (Two Stage)18.7
ViT + LLM (End to End)25.6
Sub-Word (End to End) [41]30.7
", + "image_path": "f233a11b44b49368eb2ad65b5c63bc99b6dc35c265d76a66a54f5192fb4b4bff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 249, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 305 + ], + "type": "text", + "content": "Table 6. Comparison of Word Error Rate (WER) between using the intermediary Phonemes and skipping the CTC head. Removing the intermediary representation increases the WER from 18.7 to 25.6, showing the importance of the intermediary step and the effectiveness of the phoneme centric fine-tuning." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 325, + 471, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 471, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 471, + 338 + ], + "type": "text", + "content": "6. Conclusion and Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": "We have introduced a two-stage, phoneme-centric framework for visual-only lip reading that first predicts phonemes from lip movements and then reconstructs coherent sentences via a fine-tuned LLM. This design sharply reduces word error rates on benchmark datasets (LRS2 [47] and LRS3 [3]), providing improvements of " + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 313, + 345, + 555, + 453 + ], + "type": "text", + "content": " WER, respectively, over existing state-of-the-art approaches that focus on end-to-end connectionist approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 594 + ], + "type": "text", + "content": "Despite these gains, the approach remains limited by the visual ambiguity of certain phonemes. Lip movements alone cannot always differentiate minimal pairs such as /s/ vs. /z/, and words that share near-identical phoneme sequences (e.g., \"Hello\" vs. \"Hallow\") can still result in misclassification. We plan to address these issues by refining the phoneme-to-word generation process, potentially through deeper linguistic context modeling and more specialized phoneme embeddings. Future work may also explore speaker-adaptive fine-tuning or multi-frame alignment strategies to better handle subtle visual distinctions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 417, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 417, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 417, + 619 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 703 + ], + "type": "text", + "content": "This work was supported by the SNSF project 'SMILE II' (CRSII5 193686), the Innosuisse IICT Flagship (PFFS-21-47), EPSRC grant APP24554 (SignGPT-EP/Z535370/1) and through funding from Google.org via the AI for Global Goals scheme. This work reflects only the author's views and the funders are not responsible for any use that may be made of the information it contains." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2853" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Emre Afacan and Gültekin Demirci. A survey on automatic lip-reading: Challenges and future directions. Multimedia Tools and Applications, 2019. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 167 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE transactions on pattern analysis and machine intelligence, 2018. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 295, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 295, + 211 + ], + "type": "text", + "content": "[3] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 294, + 266 + ], + "type": "text", + "content": "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2143-2147. IEEE, 2020. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 268, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 294, + 301 + ], + "type": "text", + "content": "[5] Yannis M Assael, Brendan Shillingford, Shimon Whiteson, and Nando De Freitas. Lipnet: End-to-end sentence-level lipreading. arXiv preprint arXiv:1611.01599, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 345 + ], + "type": "text", + "content": "[6] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 294, + 390 + ], + "type": "text", + "content": "[7] Helen L Bear, Richard W Harvey, Barry-John Theobald, and Yuxuan Lan. Which phoneme-to-viseme maps best improve visual-only computer lip-reading? In International symposium on visual computing. Springer, 2014. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 294, + 434 + ], + "type": "text", + "content": "[8] Umberto Cappellazzo, Minsu Kim, Honglie Chen, Pingchuan Ma, Stavros Petridis, Daniele Falavigna, Alessio Brutti, and Maja Pantic. Large language models are strong audio-visual speech recognition learners, 2025. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 435, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 294, + 468 + ], + "type": "text", + "content": "[9] Luca Cappelletta and Naomi Harte. Viseme definitions comparison for visual-only speech recognition. In 2011 19th European Signal Processing Conference. IEEE, 2011. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 469, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 469, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 57, + 469, + 294, + 523 + ], + "type": "text", + "content": "[10] Oscar Chang, Hank Liao, Dmitriy Serdyuk, Ankit Shahy, and Olivier Siohan. Conformer is all you need for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 525, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 546 + ], + "type": "text", + "content": "[11] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 547, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 580 + ], + "type": "text", + "content": "[12] Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao. Grid av speech corpus sample. The Journal of the Accoustical Society of America, 2013. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 581, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 646 + ], + "type": "text", + "content": "[13] Yasser Abdelaziz Dahou Djilali, Sanath Narayan, Haithem Boussaid, Ebtessam Almazrouei, and Merouane Debbah. Lip2vec: Efficient and robust visual speech recognition via latent-to-latent visual to audio representation mapping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 679 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[15] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H. Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "speech recognition. CAAI Transactions on Intelligence Technology, 8(1):129-138, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "text", + "content": "[16] Randa El-Bialy, Daqing Chen, Souheil Fenghour, Walid Hussein, Perry Xiao, Omar H Karam, and Bo Li. Developing phoneme-based lip-reading sentences system for silent speech recognition. CAAI Transactions on Intelligence Technology, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "text", + "content": "[17] A. Ephrat, I. Mosseri, O. Lang, T. Dekel, K Wilson, A. Hassidim, W. T. Freeman, and M. Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 260 + ], + "type": "text", + "content": "[18] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning, 2006. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "type": "text", + "content": "[19] Alexandros Haliassos, Pingchuan Ma, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Jointly learning visual and auditory speech representations from raw data. arXiv preprint arXiv:2212.06246, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 317, + 305, + 553, + 349 + ], + "type": "text", + "content": "[20] Alexandros Haliassos, Rodrigo Mira, Honglie Chen, Zoe Landgraf, Stavros Petridis, and Maja Pantic. Unified speech recognition: A single model for auditory, visual, and audio-visual inputs. arXiv preprint arXiv:2411.02256, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 350, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 414 + ], + "type": "text", + "content": "[21] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 458 + ], + "type": "text", + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 460, + 553, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 493 + ], + "type": "text", + "content": "[23] F. Jelinek, L. Bahl, and R. Mercer. Design of a linguistic statistical decoder for the recognition of continuous speech. IEEE Transactions on Information Theory, 21(3), 1975. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "text", + "content": "[24] Hendrik Laux, Emil Mededovic, Ahmed Hallawa, Lukas Martin, Arne Peine, and Anke Schmeink. Litevsr: Efficient visual speech recognition by learning from speech representations of unlabeled data. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 559, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 634 + ], + "type": "text", + "content": "[25] Xubo Liu, Egor Lakomkin, Konstantinos Vougioukas, Pingchuan Ma, Honglie Chen, Ruiming Xie, Morrie Doulaty, Niko Moritz, Jachym Kolar, Stavros Petridis, et al. Synthvsr: Scaling up visual speech recognition with synthetic supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[26] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[27] Pingchuan Ma and Alexandros et al. Haliassos. Auto-avsr: Audio-visual speech recognition with automatic labels. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2854" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "text", + "content": "IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 296, + 129 + ], + "type": "text", + "content": "[28] Pingchuan Ma and Rodrigo et al. Mira. Lira: Learning visual speech representations from audio through self-supervision. arXiv, 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "text", + "content": "[29] Pingchuan Ma, Stavros Petridis, and Maja Pantic. End-to-end audio-visual speech recognition with conformers. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "text", + "content": "[30] Takaki Makino, Hank Liao, Yannis Assael, Brendan Shillingford, Basilio Garcia, Otavio Braga, and Olivier Siohan. Recurrent neural network transducer for audio-visual speech recognition. In IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2019. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 231, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 231, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 57, + 231, + 294, + 253 + ], + "type": "text", + "content": "[31] William D Marslen-Wilson. Functional parallelism in spoken word-recognition. Cognition, 25(1-2), 1987. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 254, + 295, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 295, + 286 + ], + "type": "text", + "content": "[32] William D Marslen-Wilson and Alan Welsh. Processing interactions and lexical access during word recognition in continuous speech. Cognitive psychology, 10(1), 1978. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 287, + 294, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 287, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 57, + 287, + 294, + 309 + ], + "type": "text", + "content": "[33] Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 1976. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 310, + 295, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 310, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 57, + 310, + 295, + 343 + ], + "type": "text", + "content": "[34] Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. arXiv preprint arXiv:1609.07843, 2016. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 345, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 345, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 345, + 295, + 398 + ], + "type": "text", + "content": "[35] Ara V Nefian, Luhong Liang, Xiaobo Pi, Liu Xiaoxiang, Crusoe Mao, and Kevin Murphy. A coupled hmm for audiovisual speech recognition. In 2002 IEEE International Conference on Acoustics, Speech, and Signal Processing. IEEE, 2002. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 400, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 295, + 443 + ], + "type": "text", + "content": "[36] Eng-Jon Ong and Richard Bowden. Learning temporal signatures for lip reading. In 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops). IEEE, 2011. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "type": "text", + "content": "[37] Stavros Petridis and Maja Pantic. Deep complementary bottleneck features for visual speech recognition. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 544 + ], + "type": "text", + "content": "[38] Stavros Petridis, Themos Stafylakis, Pingchuan Ma, Georgios Tzimiropoulos, and Maja Pantic. Audio-visual speech recognition with a hybrid ctc/attention architecture. In 2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 590 + ], + "type": "text", + "content": "[39] Javad Peymanfard, Vahid Saeedi, Mohammad Reza Mohammadi, Hossein Zeinali, and Nasser Mozayani. Leveraging visemes for better visual speech representation and lip reading, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 295, + 634 + ], + "type": "text", + "content": "[40] Gerasimos Potamianos, Chalapathy Neti, Guillaume Gravier, Ashutosh Garg, and Andrew W Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9), 2003. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 679 + ], + "type": "text", + "content": "[41] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Sub-word level lip reading with visual attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[42] KR Prajwal, Triantafyllos Afouras, and Andrew Zisserman. Speech recognition models are strong lip-readers. In Proc. Interspeech 2024, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[43] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. arXiv, 2018. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[44] Dmitriy Serdyuk, Otavio Braga, and Olivier Siohan. Transformer-based video front-ends for audio-visual speech recognition for single and multi-person video. arXiv preprint arXiv:2201.10439, 2022. 3, 4, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[45] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. *ICLR*, 2022. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "text", + "content": "[46] Brendan Shillingford, Yannis Assael, Matthew W Hoffman, Thomas Paine, Cian Hughes, Utsav Prabhu, Hank Liao, Hasim Sak, Kanishka Rao, Lorrayne Bennett, et al. Large-scale visual speech recognition. InterSpeech, 2018. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[47] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 3, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[48] Themos Stafylakis and Georgios Tzimiropoulos. Zero-shot keyword spotting for visual speech recognition in-the-wild. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-529, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "text", + "content": "[49] Kwanchiva Thangthai, Helen L Bear, and Richard Harvey. Comparing phonemes and visemes with dnn-based lipreading. arXiv preprint arXiv:1805.02924, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "text", + "content": "[50] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "text", + "content": "[51] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "text", + "content": "[52] Carnegie Mellon University. The cmu pronouncing dictionary, 1993. http://www.speech.cs.cmu.edu/cgi-bin/cmudict.2,5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[54] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 612 + ], + "type": "text", + "content": "[55] Michael Wand, Jan Koutnik, and Jürgen Schmidhuber. Lipreading with long short-term memory. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[56] Xinshuo Weng and Kris Kitani. Learning spatio-temporal features with two-stream deep 3d cnn for lipreading. arXiv preprint arXiv:1905.02540, 2019. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[57] Bo Xu, Cheng Lu, Yandong Guo, and Jacob Wang. Discriminative multi-modality speech recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, 2020. 3, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[58] Jeong Hun Yeo, Seunghee Han, Minsu Kim, and Yong Man Ro. Where visual speech meets language: Vsp-llm frame" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2855" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 351 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "work for efficient and context-aware visual speech processing. arXiv preprint arXiv:2402.15151, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[59] Jeong Hun Yeo, Chae Won Kim, Hyunjun Kim, Hyeongseop Rha, Seunghee Han, Wen-Huang Cheng, and Yong Man Ro. Personalized lip reading: Adapting to your unique lip movements with vision and language. arXiv preprint arXiv:2409.00986, 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[60] Xingxuan Zhang, Feng Cheng, and Shilin Wang. Spatiotemporal fusion based convolutional sequence learning for lip reading. In Proceedings of the IEEE/CVF International conference on Computer Vision, 2019. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[61] Yuanhang Zhang, Shuang Yang, Shiguang Shan, and Xilin Chen. Es3: Evolving self-supervised learning of robust audio-visual speech representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[62] Ziheng Zhou, Guoying Zhao, Xiaopeng Hong, and Matti Pietikainen. A review of recent advances in visual speech decoding. Image and vision computing, 32(9), 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 351 + ], + "type": "text", + "content": "[63] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, 2015. 8" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2856" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_content_list.json b/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9b97bd4adc31ffe84444d55b9573778d7f962941 --- /dev/null +++ b/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_content_list.json @@ -0,0 +1,1768 @@ +[ + { + "type": "text", + "text": "VCA: Video Curious Agent for Long Video Understanding", + "text_level": 1, + "bbox": [ + 202, + 130, + 795, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeyuan Yang† \nUniversity of Massachusetts, Amherst \nyangzeyuan@umass.edu", + "bbox": [ + 173, + 179, + 475, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xueyang Yu \nUniversity of Massachusetts, Amherst \nxueyangyu@umass.edu", + "bbox": [ + 173, + 241, + 475, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Delin Chen† \nUniversity of Massachusetts, Amherst delinchen@umass.edu", + "bbox": [ + 521, + 179, + 823, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maohao Shen \nMassachusetts Institute of Technology \nmaohao@mit.edu", + "bbox": [ + 521, + 241, + 823, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuang Gan \nUniversity of Massachusetts, Amherst \nganchuang1990@gmail.com", + "bbox": [ + 346, + 301, + 650, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 388, + 326, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long video understanding poses unique challenges due to their temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as \"VCA\". Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences. Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency.", + "bbox": [ + 88, + 421, + 485, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a growing demand for developing long-form video understanding techniques [48, 69] capable of extracting valuable information from extended visual narratives, such as surveillance footage, documentaries, and instructional videos. Analyzing these videos [53, 71], which can range from minutes to hours, requires models that can process multimodal data and perform reasoning over extremely long sequences [4, 31, 60].", + "bbox": [ + 89, + 724, + 482, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous works commonly utilize two-stream networks to capture spatial and temporal information [18, 19] or in", + "bbox": [ + 89, + 847, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3a34f2280701b001980d124b23ff96bfa7878b2f2423e84306fc895c68a5841.jpg", + "image_caption": [ + "Figure 1. Accuracy v.s. Computation Cost and Number of Observed Frames. Compared to baselines, our agent demonstrates superior performance with higher efficiency (observing fewer frames and incurring comparable or lower computational cost) on EgoSchema, using both GPT-4o and Qwen2-VL-72B." + ], + "image_footnote": [], + "bbox": [ + 535, + 388, + 710, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91ceaf3d095c177613ad02b2b753cad522a8335adc05cf71027f60062516008a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 388, + 888, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "troduce 3D operations [8, 57]. Recent approach [32, 75] leverages the long-sequence reasoning capabilities of Large Language Models (LLMs) [1, 64] for long-form video question answering, often by converting videos into densely and uniformly sampled frames [33, 38, 55]. However, this approach is usually inefficient due to the inherent imbalance and varying information density across video segments [61, 78]. Some segments are rich in critical details necessary for answering questions, while others contain only trivial information [66]. Uniform sampling [49, 60], which might focus on redundant frames that increase the computational burden and may overlook frames with essential information, often fails to efficiently utilize the video's dynamic structure for question answering [17, 62].", + "bbox": [ + 511, + 641, + 906, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this limitation, recent approaches [29, 34, 36, 68] employ complex architectures [63, 85] or specialized preprocessing techniques [44, 65] to improve the efficiency", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Equal Contribution", + "bbox": [ + 107, + 886, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20168", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of long video content processing. These methods [62, 66] often convert videos into text by captioning densely sampled frames or incorporating memory modules [5, 17, 39] to store essential information. However, these approaches still incur significant computational costs. For example, captioning an hour-long video with Qwen2-VL-72B at just 1 fps requires over 3 hours on a single H100. Additionally, these methods [12] identify key frames based on image similarity rather than a holistic understanding of the video content, which can result in missing relevant video features. This necessitates the need for an efficient and scalable solution for long-form video understanding.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce VCA, a curiosity-driven agent inspired by how humans tackle the long video understanding task. Given a question, human is usually intrinsically motivated to identify relevant information in the video without external knowledge. In practise, human tends to build a global understanding of the video, then curiously zoom in to explore specific sections in greater detail. Motivated by this process, we propose a novel video agent framework, dubbed as \"VCA\", to emulate such behavior. VCA treats the video as an environment to explore, actively navigating its contents and identifying segments that merit deeper analysis. VCA incorporates three essential techniques: (1) Tree-search Exploration: Rather than selecting specific frames [62], VCA uses a tree-search structure to adaptively explore the video segments in a structured way. (2) Reward Model: VCA leverages an intrinsic reward scoring mechanism to guide exploration toward the segments most relevant to the query. (3) Memory Management: VCA employs a fixed memory buffer to store key information and discard irrelevant frames, effectively reducing computational costs and avoiding memory overflow. The key difference between VCA and prior works is outlined in Fig. 2. Extensive experiments and superior performance across multiple benchmarks demonstrate that our proposed VCA framework significantly improves an agent's effectiveness and efficiency in reasoning over extended video content. More concretely, our contributions are threefold,", + "bbox": [ + 91, + 313, + 483, + 720 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VCA is a long video understanding agent that can self-exlore the most informative video segments based on its intrinsic reward, without relying on any auxiliary tools.", + "- VCA is easy to implement, training-free, and is flexible to integrate with any vision-language model.", + "- As highlighted in Fig. 1, VCA outperforms existing baselines while observing fewer frames and incurring comparable or lower computational cost, demonstrating both effectiveness and efficiency." + ], + "bbox": [ + 89, + 761, + 483, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/86d97b7ae7cef89e8e72f2b8f1a26808351cdd81b23ff0ab9e2edc6c90c89ae7.jpg", + "image_caption": [ + "Figure 2. Comparison of VCA with prior works. Difference-1: When answering a query, VCA uniformly samples (a few) frames, whereas previous methods (e.g., VideoTree) require captioning each frame for subsequent image retrieval, which can be computationally expensive for long videos. Difference-2: previous methods rely on image similarity to locate key frames, potentially losing important video features. In contrast, VCA uses segment-level exploration to identify the most informative segments." + ], + "image_footnote": [], + "bbox": [ + 540, + 92, + 880, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 382, + 655, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Long-context Multi-modal Agents", + "text_level": 1, + "bbox": [ + 511, + 407, + 810, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large multi-modal models (LMMs) [2, 6, 9, 35, 54] have demonstrated promising performance in addressing complex multimodal tasks [20, 45, 49, 81]. However, these LMMs typically lack self-directed exploration capabilities required for handling more tricky tasks [24, 74]. To bridge this gap, multimodal agents are designed to enhance LMMs with the ability to autonomously make decisions and execute tasks within specific environments [15, 73], such as interpreting visual cues and text prompts to make real-time decisions within websites [28, 76].", + "bbox": [ + 511, + 429, + 906, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key feature of multimodal agents is their long-context capability [84], i.e., the ability to process and utilize information across extended periods to make informed decisions [26]. An example of this is vision-language model (VLM) agents for video understanding tasks. Despite their advancements, even sophisticated VLM agents struggle to extract crucial information and perform effective reasoning over lengthy inputs [62, 79]. Therefore, efficiently extracting essential information and making informed decisions remains a primary challenge. This work aims to address this challenge by enabling agents to self-exlore the environment and self-gather the most relevant information to efficiently solve video-understanding tasks.", + "bbox": [ + 511, + 580, + 908, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Long Video Understanding", + "text_level": 1, + "bbox": [ + 511, + 787, + 759, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early deep learning models for video understanding primarily followed two main approaches [52]. One uses two-stream networks [18, 19, 42, 58] to integrate spatial and motion information, while the second focuses on 3D CNNs [8, 57, 72] for spatial-temporal feature extraction. Vision Transformers [14] then introduce self-attention mechanisms", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20169", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36be31cb5baa1db2c03e215ae3ccb61924a453519badc0eb3c88049bd76108e3.jpg", + "image_caption": [ + "Figure 3. Pipeline of VCA Framework. First, the reward model predicts the relevance scores of each segment. Next, the memory buffer is updated by adding new frames and discarding irrelevant ones. Finally, emulating human reasoning, the agent actively selects segments to explore until the correct answer is found." + ], + "image_footnote": [], + "bbox": [ + 197, + 80, + 803, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to handle temporal modeling more effectively [3, 7, 16]. Methods with self-supervised pre-training [50, 56, 86] exhibit transferability across different tasks and achieve more robust video understanding performance.", + "bbox": [ + 89, + 486, + 482, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal LLM [22, 59, 83] shows strong capabilities in video understanding tasks, but efficiently processing information within a prolonged video environment is the key challenge. Recent approaches often use frame sampling [32, 46, 63, 75, 77] and feature extraction [27, 49, 51, 70] to reduce computational costs. Frame sampling selects a subset of frames, which risks missing essential context in lengthy videos, where transitions and subtle cues may be easily neglected. Additionally, processing entire video sequences as tokens presents significant memory and computational challenges [11, 43, 64, 80].", + "bbox": [ + 89, + 565, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent work has explored auxiliary techniques and modules to improve efficiency, such as pre-extracted content like captions [44, 62], multimodal summaries [40, 65], and combinations of various auxiliary tools [17, 30]. However, these tools introduce additional computational and memory costs. Moreover, some approaches still require traversing the entire video content [17]. Notably, VideoTree [66] requires captioning every frame in the video, making it prohibitively expensive. In contrast, our VCA can solve the task using a single VLM, without relying on any supplementary tools.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 484, + 604, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by human reasoning, we introduce a self-exploration agent that adaptively navigates video environments. Unlike approaches that depend on dense frame sampling or captioning [62], which require VLMs to process numerous frames, our agent analyzes videos iteratively within a limited context window in a process we call self-exploration. This approach allows essential information to be extracted from long videos in a coarse-to-fine manner.", + "bbox": [ + 511, + 513, + 906, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we propose a segment-based tree-search exploration structure. Given a video $v$ and a user query $q$ , the agent $\\pi$ begins by randomly sampling $N$ frames $\\{f_v^i\\}_{i=1}^N$ to obtain an initial overview of the content in each video segment $v_i \\in \\{v_i\\}_{i=1}^{N+1}$ . The agent then decides whether additional exploration is needed to answer the query. If further investigation is necessary, the agent selects a specific segment $s^* \\in \\{v_i\\}_{i=1}^{N+1}$ for deeper exploration until sufficient information is gathered to respond to the query.", + "bbox": [ + 511, + 638, + 908, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, a one-hour movie typically contains approximately 86,400 frames, posing a challenge for current VLM-based agents to efficiently identify the most relevant segment [17, 62, 66]. Moreover, even with a small $N$ , conducting multiple rounds of exploration accumulates a substantial number of frames, leading to high computational costs. To address these issues, we introduce a reward model $R$ to guide the agent's exploration and propose a memory", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "buffer $M$ to reduce storage and inference costs.", + "bbox": [ + 89, + 90, + 403, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first explain how the reward model assists the agent throughout the exploration process (Sec. 3.1). Then we introduce our memory management strategy in Sec. 3.2. Finally, we explain how VCA adaptively explores the video using a segment-based tree structure (Sec. 3.3).", + "bbox": [ + 89, + 107, + 483, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Reward Model", + "text_level": 1, + "bbox": [ + 89, + 195, + 241, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the segment-based tree-search exploration, the agent is required to select the most informative video segment for exploration at each step. However, we observe that current VLM-based agents struggle to independently identify optimal exploration paths (see Sec. 4.3). To this end, inspired by the selective attention mechanism in human cognition [13], we incorporate a reward model $R$ to guide the agent's selection process, i.e., the agent will gain a higher reward if it plans to take the action to explore more relevant segment.", + "bbox": [ + 89, + 218, + 483, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, during the exploration of a current selected segment $s^*$ , the reward model observes a set of frames $\\{f_{s^*}^i\\}_{i=1}^N$ sampled from $s^*$ . These frames $\\{f_{s^*}^i\\}_{i=1}^N$ then form finer sub-segments $\\{s_i^* = [l_{s^*}^i, l_{s^*}^{i+1}]\\}_{i=1}^{N+1}$ , where $l_{s^*}^i$ and $l_{s^*}^{i+1}$ represent the indices of the beginning and end frame of the sub-segment, respectively. The reward model aims to generate a reward score for each sub-segment $\\{s_i^*\\}_{i=1}^{N+1}$ based on its relevance to the user query $q$ . We utilize chain-of-thoughts [67] technique to let the reward model provide a verbal explanation of each segment and then generate a relevant score, i.e., $\\{t_{s_i^*}; r_{s_i^*}\\}_{i=1}^{N+1}$ . Additionally, to maintain consistency across different exploration steps, we incorporate the reward history $H_r$ from previous steps alongside the new segments during evaluation. A simplified prompt template for the reward model is presented here, with detailed prompts in Appendix ??", + "bbox": [ + 89, + 356, + 483, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reward Model Prompt (Simplified)", + "text_level": 1, + "bbox": [ + 117, + 623, + 352, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given:", + "bbox": [ + 116, + 651, + 165, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$H_{r}$ , reward score history in previous steps.", + "bbox": [ + 116, + 665, + 398, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\left\\{f_{s^*}^i\\right\\}_{i = 1}^N$ newly sampled frames from current segment $s^{*}$", + "bbox": [ + 116, + 680, + 455, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Please first explain each sub-segment $\\{s_i^*\\}_{i = 1}^{N + 1}$ and then assign a relevance score of each sub-segment to the user query $q$ .", + "bbox": [ + 114, + 710, + 457, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, rather than training a separate specialized model, we utilize a same VLM-based agent to function as the reward model. While the reward model $R$ and exploration agent $\\pi$ serve distinct purposes, both can be implemented using the same model. In this work, we employ a single VLM-based agent for both roles across different settings, providing a neat and efficient agent framework that is guided by its self-generated intrinsic reward and without", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relying on auxiliary models or tools.", + "bbox": [ + 511, + 90, + 754, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Memory Management", + "text_level": 1, + "bbox": [ + 511, + 116, + 720, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As exploration proceeds over multiple rounds, the accumulation of frames can lead to significant computational costs for long videos, even when $N$ is small. However, as illustrated in Fig. 2, they tend to gradually disregard irrelevant frames, especially from earlier steps, retaining only the most crucial information in their working memory for later analysis. Inspired by this, we introduce a fixed-size memory buffer $M$ that stores only the most relevant frames throughout exploration.", + "bbox": [ + 511, + 138, + 903, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, when the agent $\\pi$ explores a video segment $s^*$ , the associated frames $\\{f_{s^*}^i\\}_{i=1}^N$ are added to the memory buffer. If the total number of frames exceeds the buffer limit, irrelevant frames are removed based on relevance scores $r_{s_i^*}$ defined in Sec. 3.1. We discard the frames with the lowest scores and only retain the most relevant ones. While these less relevant frames are removed, their visual information is extracted and encoded as text descriptions $\\{t_{s_i^*}\\}_{i=1}^N$ within the exploration history $H_r$ , allowing the reward model to utilize. This memory management strategy enables the agent to conduct reasoning within a fixed resource limit, regardless of the exploration trajectory's length, thereby preventing excessive memory usage.", + "bbox": [ + 511, + 275, + 905, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tree-Search Exploration", + "text_level": 1, + "bbox": [ + 511, + 484, + 738, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by humans adaptively focusing on different video segments during exploration rather than targeting specific frames directly, our proposed agent VCA treats the video as an unexplored environment, where each segment represents a node on an exploration tree. Given the entire video as the root node, the agent's goal is to efficiently identify an optimal short path leading to a leaf node (segment) containing the most informative content needed to answer the query.", + "bbox": [ + 511, + 507, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each step, the agent $\\pi$ receives both textual and visual guidance: a set of candidate segments $S$ , which includes a collection of segments with associated reward scores, and a memory buffer $M$ containing multiple frames. If the agent has gathered sufficient information to answer the user's query $q$ , it directly generates a response $a$ . Otherwise, it continues exploring by selecting a specific segment $s^* \\in S$ to expand the tree based on reward scores.", + "bbox": [ + 511, + 628, + 905, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, due to the limited capacity of current foundational models, these reward scores are not always accurate (see further analysis in Sec. 5.3). Selecting segments based on reward scores in a purely greedy manner can potentially lead to a suboptimal solution. Instead, we use the reward scores as guidance for agent $\\pi$ , allowing it to make the final segment selection independently. In this case, the agent is not forced to select the segment with the highest reward, balancing the exploitation and exploration. More importantly, the candidate set $S$ includes both unexplored", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coarse segments from prior steps and fine-grained segments from the current step, enabling the agent to backtrack and select earlier segments if it detects that the current exploration path may be suboptimal. Below are the simplified prompts. The full prompt is provided in Appendix ??", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Exploration Agent Prompt (Simplified)", + "text_level": 1, + "bbox": [ + 116, + 181, + 377, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given:", + "bbox": [ + 116, + 210, + 165, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Candidate segments set: $S$", + "bbox": [ + 116, + 224, + 295, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frames in Memory buffer: $M$", + "bbox": [ + 116, + 241, + 315, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Please choose one segment to explore based on the query $q$ . Output answer if you have gathered sufficient information.", + "bbox": [ + 116, + 256, + 457, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall framework of our proposed method is summarized in Algo. 1. Leveraging a tree-search exploration technique boosted with a reward model and a memory buffer, our self-exploration agent adaptively and efficiently explores the video like a human. Empirical results in Sec. 4.3 further validate the effectiveness of our tree-search exploration over naive frame selection methods.", + "bbox": [ + 89, + 321, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 430, + 223, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 455, + 297, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark. We mainly evaluate VCA on two long-term video question-answering benchmarks. EgoSchema [41] is an egocentric video dataset sourced from Ego4D [21], with a duration of 180 seconds. Here we use the validation set of 500 QA pairs. LVBench [61] covers about 1,500 samples for evaluation of reasoning abilities from high-quality videos over 30 minutes in length. We also conduct experiments on other relevant benchmark datasets in Appendix ??", + "bbox": [ + 89, + 481, + 483, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare with most relevant baselines to our work, including state-of-the-art open-source Video-LLMs [10, 23, 37, 47, 59, 82], the close-source model GPT-4o [25], and agent-based systems [44, 62, 66]. For video models, we report the results obtained with uniform sampling frames, while for the agent-based systems, we follow their official implementation with GPT-4o [25], consistent with VCA's setup to ensure a fair comparison.", + "bbox": [ + 89, + 637, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In this work, we use the same model for the exploration agent and the reward model. Unless otherwise stated, all experiments were performed using the August 2024 version of `gpt-4` as the base model, leveraging the OpenAI LLM API service1 with a temperature of 0.5. We set memory buffer sizes set to 8 and 16 frames for EgoSchema and LVBench, respectively. Further implementation details are provided in Appendix ??", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 VCA: Video Curious Agent", + "Ensure: Answer $a$" + ], + "code_body": "Require: Video environment $v$ , question $q$ , agent $\\pi$ , reward model $R$ , memory buffer $M$ , candidate segments set $S$ , reward score history $H_{r}$ , sampling frame number $N$ .", + "guess_lang": "txt", + "bbox": [ + 514, + 109, + 906, + 155 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize selected segment $s^* \\gets v$", + "2: Initialize candidate segments set $S \\gets \\emptyset$", + "3: Initialize memory buffer $M\\gets \\emptyset$", + "4: Initialize reward score history $H_{r}\\gets \\emptyset$", + "5: continue $\\leftarrow$ true", + "6: while continue do", + "7: # Sampling frames & form segments", + "8: $\\left\\{f_{s^*}^*\\right\\}_{i=1}^{N}, \\left\\{s_i^*\\right\\}_{i=1}^{N+1} \\gets \\text{UniformSample}(s^*)$", + "9: # Scoring each segment", + "10: $\\left\\{t_{s_i^*};r_{s_i^*}\\right\\}_{i = 1}^{N + 1}\\gets R(\\{s_i^*\\}_{i = 1}^{N + 1};q,\\{f_{s_*}^i\\}_{i = 1}^N,H_r)$", + "11: # Update candidate segment set", + "12: $S\\gets S\\cup \\{s_i^*;r_{s_i^*}\\}_{i = 1}^{N + 1}$", + "13: # Update history with thoughts", + "14: $H_{r}\\gets H_{r}\\cup \\left\\{t_{s_{i}^{*}};r_{s_{i}^{*}}\\right\\}_{i = 1}^{N + 1}$", + "15: # Update memory with frames", + "16: $M\\gets \\mathrm{UpdateMemory}(M,\\{f_{s^{*}}^{i}\\}_{i = 1}^{N};S)$", + "17: # Output answer or select segment", + "18: $a,s\\gets \\pi (q,S;M)$", + "19: if $a = \\emptyset$ then", + "20: # Continue exploration", + "21: $s^* \\gets s$", + "22: else", + "23: continue $\\leftarrow$ false", + "24: end if", + "25: end while", + "26: return $a$" + ], + "bbox": [ + 517, + 170, + 885, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 599, + 715, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental results on LVBench and EgoSchema are presented in Tab. 2 and Tab. 1, respectively. By adaptively focusing on critical segments rather than scanning the entire video, our agent not only achieves stronger performance but also incurs less computational cost. With less than $30\\%$ of the frames, VCA achieves substantial improvements over directly feeding uniformly sampled frames into GPT-4o. Specifically, we observe performance gains of $3.2\\%$ on the EgoSchema subset and $6.6\\%$ on LVBench. Significant increases in the Event Understanding and Temporal Grounding domains further underscore our effectiveness in identifying crucial segments within long videos.", + "bbox": [ + 511, + 621, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, we compare our framework with recently proposed SOTA video agent systems, VideoAgent [62] and VideoTree [66]. For a fair comparison, we re-implement both frameworks using GPT-4o for all reasoning components, consistent with our setup. However, captioning all", + "bbox": [ + 511, + 801, + 906, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://platform.openai.com/docs/models", + "bbox": [ + 531, + 886, + 846, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20172", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cdc65838a4e11839a1a1465235c041052cb787d85754c744b501c8f8c7688785.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFramesEREUKIRTGReaSumAvg.
TimeChat* [47]>9621.921.725.922.725.024.122.3
LLaVA-NeXT* [83]3230.131.234.131.435.027.632.2
InternVL2-40B* [10]1637.439.743.431.442.541.439.8
GLM4V-Plus* [23]2039.935.834.837.740.032.838.3
GPT-4o* [25]6435.930.835.528.333.534.534.7
VideoAgent† [62]Avg. 25.528.030.328.029.328.036.429.3
VideoTree† [66]Avg. 103.230.325.126.527.731.925.528.8
VCA (ours)Avg. 20.043.740.737.838.046.227.341.3
", + "bbox": [ + 207, + 88, + 787, + 253 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/58c7945053aceb0fd40e6d71d931795c4e8194cd4f36892dd1f0f3f85477e13b.jpg", + "table_caption": [ + "Table 1. Experimental Results on LVBench. Results of methods marked with * are sourced from LVBench [61]. Methods marked with † indicate implementations where GPT-4o serves as the main component. Our agent's memory buffer size is set to 16 frames. We report the average number of observed frames for agent-based methods, with the best performance highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodFramesSubset
GPT-4o [25]869.0
3270.4
\\(VideoAgent^\\dagger\\) [62]Avg. 8.163.2
Avg. 10.760.8
\\(VideoTree^\\dagger\\) [66]Avg. 69.567.0
\\(LVNet^\\dagger\\) [44]1268.2
VCA (ours)Avg. 7.273.6
", + "bbox": [ + 151, + 323, + 421, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "frames with GPT-4o is prohibitively expensive. Hence, following the baseline implementations, we use open-source VLMs for captioning. A comparison using the same model for all components is provided in Sec. 5.1. As shown in Tab. 1, VCA outperforms all baselines over $5\\%$ on EgoSchema, while requiring processing fewer frames. These results suggest the limitations of dense captioning for video preprocessing tasks, which can incur more computational costs while introducing more redundant information. In contrast, our proposed memory management strategy effectively improves efficiency by only retaining the most crucial information for reasoning. Similar insights can be drawn from results on LVBench in Tab. 2. Additional results are provided in Appendix ??", + "bbox": [ + 88, + 578, + 482, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we observe that VCA explores approximately three times more frames on LVBench than on EgoSchema. Recall that EgoSchema videos have an average duration of 180 seconds, while videos in LVBench extend to at least 30 minutes. This phenomenon suggests that our proposed tree-search exploration technique actively explores the video to locate the most relevant information and adaptively in-", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f6591466ea61c795db00ac6cff0bc93214d577866bef4bd46d75d4ca9daf4aaa.jpg", + "table_caption": [ + "Table 2. Experimental Results on EgoSchema. Methods marked with $\\dagger$ indicate implementations where GPT-4o is the main component. The results of LVNet are from [44]. Our agent's memory buffer size is set to 8 frames. We report the average number of observed frames, with the best performance highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodEgoSchemaLVBench
FramesAcc.FramesAcc.
VCA (ours)7.273.620.041.3
- w/o Reward7.572.622.239.5
- w/o Tree Search10.269.839.936.2
", + "bbox": [ + 540, + 323, + 879, + 412 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/12eb3b064a7106c503cac75584b9acfb634f0a1eb1846cbde7e0a2a197d01119.jpg", + "table_caption": [ + "Table 3. Ablation Study on EgoSchema and LVBench. Both the tree-search exploration and reward model contribute significantly to improvements in performance and efficiency. Detailed ablation study results on LVBench are included in Appendix ??" + ], + "table_footnote": [], + "table_body": "
Buffer Size81632
Acc.38.941.343.2
", + "bbox": [ + 602, + 486, + 818, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Ablation Study of Buffer Size on LVBench. Increasing the buffer size allows the model to observe more frames during reasoning, leading to improved performance.", + "bbox": [ + 511, + 541, + 906, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "creases the exploration budget based on task complexity. For example, VCA exhibits both exploitation and exploration: when the agent determines that the current exploration path is suboptimal, it backtracks to previous segments to explore a new path (see Sec. 5.2). These empirical observations further validate the robustness and scalability of our approach, especially in diverse long-video contexts.", + "bbox": [ + 511, + 597, + 905, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 712, + 666, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct an ablation study to evaluate the components of our agent. Results across both benchmarks are presented in Tab. 3. As illustrated in the table, performance consistently declines when the reward model is removed, highlighting that the agent's inherent reasoning ability alone is insufficient for accurately identifying the most informative segments. This finding emphasizes the effectiveness of incorporating an intrinsic reward to guide the agent's segment selection process. A detailed analysis of our reward model's performance is provided in Sec. 5.3.", + "bbox": [ + 511, + 734, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we observe a significant performance drop", + "bbox": [ + 532, + 886, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20173", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Why does the mice run away?", + "(A) Because the black cat sheriff calls the police", + "(C) Because their boss is injured" + ], + "bbox": [ + 222, + 95, + 455, + 127 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(B) Because they meet more powerful mice", + "(D) Because the owner of the house is back" + ], + "bbox": [ + 460, + 107, + 665, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/494c2e39d33787b489c52c42d9cf3f62f97d802bf3cb5b76c77ffd4cb93b8d38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 132, + 779, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reward Model", + "text_level": 1, + "bbox": [ + 223, + 188, + 310, + 199 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Segment1: Thought: The segment shows a black cat with a gun, possibly indicating a threat. This could suggest that the mice might run away due to the presence of the cat, which aligns with option (A). Score: 0.7", + "Segment4: Thought: The segment features the black cat with a communication device, possibly calling for help or backup. This aligns with option (A), suggesting the mice might run because the cat is calling the police. Score: 0.8" + ], + "bbox": [ + 223, + 200, + 602, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cab27bdb7ff0b0177e4565f738e98f8d8b8a7fec79b098210f2663bdfcdbd937.jpg", + "image_caption": [ + "Agent Choice: (segment: 4)" + ], + "image_footnote": [], + "bbox": [ + 609, + 210, + 696, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db534f7962bd217c6a8212442a281185644644533499078be1979c927fe8315b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 210, + 777, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9709989d9606dbfa36835e193981ef9fd1b3c4452fb7a2442047d274c61b2bbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 272, + 779, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reward Model", + "text_level": 1, + "bbox": [ + 223, + 327, + 308, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segment4.1: Thought: The segment shows a close-up of a document with red text and exclamation marks, indicating urgency or importance. This could imply a critical message or alert being communicated. Score: 0.7", + "bbox": [ + 223, + 339, + 601, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "…", + "bbox": [ + 223, + 369, + 233, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0be7582fe7139255d5979ad1a0a4b2f8416fde86fabbd6434c0c69aee2357f20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 387, + 777, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reward Model", + "text_level": 1, + "bbox": [ + 223, + 443, + 310, + 453 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segment1.2: Thought: The segment features mice characters in mid-air, possibly fleeing or being thrown. This suggests a direct reaction to a threat, supporting the idea that they are running away from the cat. Score: 0.8", + "bbox": [ + 223, + 454, + 584, + 496 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e82b2c43037a928c41960aef4203fd12a1f1ce59e083cac00c6ba6a3ec87c9ad.jpg", + "image_caption": [ + "Figure 4. Exploration Trajectory Example. The agent generally tends to select video segments with higher reward scores for exploration, yet it can adaptively redirect to earlier segments, effectively balancing exploitation and exploration." + ], + "image_footnote": [], + "bbox": [ + 223, + 502, + 777, + 539 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5386d87e480fee48366b58408db92687c9c944cdedb4c56e93e21002e2bd9557.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 542, + 602, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Agent: Answer: A", + "bbox": [ + 640, + 544, + 746, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "when the tree-search exploration technique is ablated, i.e., the agent directly selects frame indices without the segment-based tree structure. Interestingly, this modification also results in a substantial increase in the average number of frames acquired, especially on LVBench. This indicates that the agent struggles to locate the most crucial information and consequently produces a longer but redundant trajectory. These ablation study results validate the effectiveness of the tree-search exploration framework assisted by the reward model. Additionally, this underscores VCA's potential to achieve greater efficiency and improved performance with more advanced VLMs and reward models.", + "bbox": [ + 89, + 611, + 482, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, as shown in Tab. 4, increasing the memory buffer size yields consistently improved performance for VCA, demonstrating that our framework benefits from a broader reasoning window while maintaining effectiveness with fewer frames. Due to limited computation resources, we only tested up to 32 frames; however, the upward trend indicates the potential for gains with additional frames.", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 513, + 609, + 609, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Can VCA generalize to Open-source Models?", + "text_level": 1, + "bbox": [ + 511, + 635, + 897, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While VCA demonstrates strong performance using GPT4o as the base model, we also analyze whether these positive results can be extrapolated to open-source models. We implement the VCA framework with Qwen2-VL-72B-Instruct-GPTQ-Int4 [59] as both the exploration agent and reward model, as well as the captioner for baselines. The results in Tab. 5 reveal that, compared to directly feeding uniformly sampled frames into Qwen2-VL, our framework achieves higher accuracy on the EgoSchema benchmark despite using only about $60\\%$ of the frames. In contrast, VideoAgent performs significantly worse. Furthermore, our approach outperforms VideoTree with far fewer frames. Notably, VideoTree employs clustering to build a caption similarity-based tree structure for frame selection, requiring caption generation for each frame. This process is computationally expensive, demanding over 60 GPU hours", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20174", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/de5d40b1cd7f8593e0c298b92c7fc57eb36751dc7abde60dd39b7f4435f5e892.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFramesSubset
Qwen2-VL870.2
3274.2
VideoAgentAvg. 8.665.2
VideoTreeAvg. 85.471.0
VCA (ours)Avg. 16.975.2
- w/ Image RewardAvg. 20.974.0
", + "bbox": [ + 148, + 88, + 426, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Performance of Using Open-source Model on EgoSchema. Our approach achieves higher accuracy with fewer observed frames, showing consistency improvements over baseline methods. Memory buffer size is set to 16 frames.", + "bbox": [ + 89, + 233, + 482, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on a single H100, as shown in Fig. 1. These results further underscore the robustness of VCA.", + "bbox": [ + 89, + 321, + 482, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. What's the Exploration Behavior of VCA?", + "text_level": 1, + "bbox": [ + 89, + 357, + 452, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we present a real exploration trajectory from LVBench to illustrate our agent's decision-making process. As shown in Fig. 4, given a user query, the reward model initially evaluates each segment's relevance based on uniformly sampled frames, and the agent selects the fourth segment, which has the highest relevance score. However, upon further investigation of segment 4, both the reward model and the agent find it relatively irrelevant to the query. Rather than continuing with it, the agent redirects its exploration to segment 1 in the last round. After selecting segment 1, the agent chooses to investigate the third segment, despite the second segment having the highest reward score. This behavior reflects VCA's ability to adaptively adjust its exploration trajectory based on visual context rather than purely following the reward score, resulting in a more robust system.", + "bbox": [ + 89, + 382, + 482, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also examine the limitations of VCA by analyzing its failure cases, which fall into the following scenarios: The most common failure occurs when the agent overlooks subtle clues, resulting in an inaccurate response to the user query even after prolonged exploration, particularly for queries requiring specific visual details. Other failure cases include instances where the agent is misled by inaccurate reward scores or fails to provide the correct answer despite observing key frames. Detailed cases and further discussion are provided in the Appendix ??", + "bbox": [ + 89, + 611, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, unlike previous methods, VCA employs a segment-level exploration strategy instead of direct image similarity to select key frames. To further investigate this difference, we evaluate the performance of selecting the segment whose start and end images have the highest visual relevance. As shown in Tab. 5, relying solely on image relevance instead of inferring segment semantics results in a performance drop on EgoSchema, confirming the effectiveness of our segment-level exploration.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fca8834cee31104ddaa6b77c3ab66dbfffdf08de055321278fd4a45f2a317be3.jpg", + "image_caption": [ + "Figure 5. Segment Selection Accuracy on LVBench. The y-axis shows segment distance from ground truth; x-axis shows count ratio. Red curves are fitted normal distributions. The reward-guided agent selects segments closer to ground truth than the reward model." + ], + "image_footnote": [], + "bbox": [ + 540, + 89, + 879, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. How Reliable is the Reward Model?", + "text_level": 1, + "bbox": [ + 511, + 421, + 823, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further assess the capability of our reward model, we visualize the matching accuracy between segments with the highest reward scores and ground truth time references from LVBench. The results are shown at the bottom of Fig. 5. We observe that segment distances follow an approximately normal distribution with a mean of 1.28, indicating that the segment with the highest reward score often closely aligns with the ground truth segment. This validates the reward model's effectiveness in providing meaningful guidance.", + "bbox": [ + 511, + 443, + 905, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, as discussed in Sec 3.3, we encourage the agent to make independent decisions when selecting segments rather than blindly following reward scores in a purely greedy manner. The benefit of this design choice is demonstrated at the top of Fig. 5, where we present the matching accuracy of the agent's choices. Compared to the greedy approach, the agent's selections result in a sharper distribution of segment distances, with an average distance gap of $-0.60$ , indicating closer alignment with the ground truth. In summary, our reward model assigns reliable relevance scores to segments, while the agent further refines these choices to achieve even more accurate selections.", + "bbox": [ + 511, + 580, + 905, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 777, + 633, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a curiosity-driven video agent framework for long video understanding tasks with self-exploration capability. VCA achieves state-of-the-art performance and efficiency without relying on auxiliary tools. Appendix ?? shows that improved reward guidance could further boost results, highlighting the potential of synthetic data and specialized reward models for future work.", + "bbox": [ + 511, + 801, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2", + "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 3", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1", + "[5] Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J Henaff. Memory consolidation enables long-context video understanding. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for ai agents, 2023. 2", + "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding?, 2021. 3", + "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset, 2018. 1, 2", + "[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2", + "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5, 6", + "[11] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3", + "[12] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. arXiv preprint arXiv:2311.01233, 2023. 2", + "[13] Robert Desimone, John Duncan, et al. Neural mechanisms of selective visual attention. Annual review of neuroscience, 18(1):193-222, 1995. 4" + ], + "bbox": [ + 93, + 143, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2", + "[15] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, Katsushi Ikeuchi, Hoi Vo, Li Fei-Fei, and Jianfeng Gao. Agent ai: Surveying the horizons of multimodal interaction, 2024. 2", + "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers, 2021. 3", + "[17] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2025. 1, 2, 3", + "[18] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition, 2016. 1, 2", + "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2", + "[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2", + "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5", + "[22] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, Lei Zhao, Zhuoyi Yang, Xiaotao Gu, Xiaohan Zhang, Guanyu Feng, Da Yin, Zihan Wang, Ji Qi, Xixuan Song, Peng Zhang, Debing Liu, Bin Xu, Juanzi Li, Yuxiao Dong, and Jie Tang. CogvIm2: Visual language models for image and video understanding, 2024. 3", + "[23] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 5, 6", + "[24] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 2", + "[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20176", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5, 6", + "[26] Lawrence Jang, Yinheng Li, Charles Ding, Justin Lin, Paul Pu Liang, Dan Zhao, Rogerio Bonatti, and Kazuhito Koishida. Videowebarena: Evaluating long context multimodal agents with video understanding web tasks, 2024. 2", + "[27] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding, 2023. 3", + "[28] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2", + "[29] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In European Conference on Computer Vision, pages 271-288. Springer, 2025. 1", + "[30] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning, 2024. 3", + "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1", + "[32] Yunxin Li, Xinyu Chen, Baotain Hu, and Min Zhang. Llms meet long video: Advancing long video comprehension with an interactive visual adapter in llms, 2024. 1, 3", + "[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1", + "[34] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1", + "[35] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2", + "[36] Ming C Lin and Shan Yang. Vila: Efficient video-language alignment for video question answering. 1", + "[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5", + "[38] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1", + "[39] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846, 2024. 2", + "[40] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 5", + "[42] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification, 2015. 2", + "[43] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14386-14397, 2024. 3", + "[44] Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya, Wonjeong Ryoo, Donghyun Kim, and Michael S Ryoo. Too many frames, not all useful: Efficient strategies for long-form video qa. arXiv preprint arXiv:2406.09396, 2024. 1, 3, 5, 6", + "[45] Ruchit Rawal, Khalid Saifullah, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. In Synthetic Data for Computer Vision Workshop @ CVPR 2024. 2", + "[46] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. arXiv preprint arXiv:2312.02051, 2023. 3", + "[47] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 5, 6", + "[48] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 1", + "[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 2, 3", + "[50] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning, 2019. 3", + "[51] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Advances in neural information processing systems, 35:38032-38045, 2022. 3", + "[52] Yunlong Tang, Jing Bi, Sitting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, Ali Vosoughi, Chao Huang, Zeliang Zhang, Pinxin Liu, Mingqian Feng, Feng Zheng, Jianguo Zhang," + ], + "bbox": [ + 514, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ping Luo, Jiebo Luo, and Chenliang Xu. Video understanding with large language models: A survey, 2024. 2", + "[53] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 1", + "[54] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2", + "[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1", + "[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training, 2022. 3", + "[57] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks, 2015. 1, 2", + "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition, 2016. 2", + "[59] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5, 7", + "[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1", + "[61] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Shiyu Huang, Bin Xu, Yuxiao Dong, Ming Ding, and Jie Tang. Lvbench: An extreme long video understanding benchmark, 2024. 1, 5, 6", + "[62] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 6", + "[63] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 1, 3", + "[64] Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recurrent memory bridges. arXiv preprint arXiv:2409.01071, 2024. 1, 3", + "[65] Ying WANG, Yanlai YANG, and Mengye REN. Lifelong-memory: Leveraging llms for answering queries in long" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "form egocentric videos. arXiv preprint arXiv:2312.05269, 2024. 1, 3", + "[66] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos, 2024. 1, 2, 3, 5, 6", + "[67] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4", + "[68] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision, pages 453-470. Springer, 2025. 1", + "[69] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 1", + "[70] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 3", + "[71] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. CoRR, 2024. 1", + "[72] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification, 2018. 2", + "[73] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. In ICLR 2024 Workshop on Large Language Model (LLM) Agents. 2", + "[74] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 2", + "[75] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Retrieval-based video language model for efficient long video question answering, 2023. 1, 3", + "[76] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2", + "[77] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 2023. 3", + "[78] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[79] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20178", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 3", + "[81] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023. 2", + "[82] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5", + "[83] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3, 6", + "[84] Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang. Longagent: Scaling language models to 128k context through multi-agent collaboration. arXiv preprint arXiv:2402.11550, 2024. 2", + "[85] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In *Forty-first International Conference on Machine Learning*. 1", + "[86] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations, 2020. 3" + ], + "bbox": [ + 91, + 90, + 480, + 529 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20179", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_model.json b/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eccd7f60a9f11a723c0b971d295293fe687ca911 --- /dev/null +++ b/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_model.json @@ -0,0 +1,3051 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.797, + 0.154 + ], + "angle": 0, + "content": "VCA: Video Curious Agent for Long Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.18, + 0.477, + 0.234 + ], + "angle": 0, + "content": "Zeyuan Yang† \nUniversity of Massachusetts, Amherst \nyangzeyuan@umass.edu" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.242, + 0.476, + 0.294 + ], + "angle": 0, + "content": "Xueyang Yu \nUniversity of Massachusetts, Amherst \nxueyangyu@umass.edu" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.18, + 0.825, + 0.234 + ], + "angle": 0, + "content": "Delin Chen† \nUniversity of Massachusetts, Amherst delinchen@umass.edu" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.242, + 0.825, + 0.293 + ], + "angle": 0, + "content": "Maohao Shen \nMassachusetts Institute of Technology \nmaohao@mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.303, + 0.651, + 0.354 + ], + "angle": 0, + "content": "Chuang Gan \nUniversity of Massachusetts, Amherst \nganchuang1990@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.39, + 0.327, + 0.405 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.486, + 0.68 + ], + "angle": 0, + "content": "Long video understanding poses unique challenges due to their temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as \"VCA\". Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences. Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.701, + 0.222, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.847 + ], + "angle": 0, + "content": "There is a growing demand for developing long-form video understanding techniques [48, 69] capable of extracting valuable information from extended visual narratives, such as surveillance footage, documentaries, and instructional videos. Analyzing these videos [53, 71], which can range from minutes to hours, requires models that can process multimodal data and perform reasoning over extremely long sequences [4, 31, 60]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Previous works commonly utilize two-stream networks to capture spatial and temporal information [18, 19] or in" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.389, + 0.712, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.389, + 0.889, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.561, + 0.908, + 0.631 + ], + "angle": 0, + "content": "Figure 1. Accuracy v.s. Computation Cost and Number of Observed Frames. Compared to baselines, our agent demonstrates superior performance with higher efficiency (observing fewer frames and incurring comparable or lower computational cost) on EgoSchema, using both GPT-4o and Qwen2-VL-72B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.853 + ], + "angle": 0, + "content": "troduce 3D operations [8, 57]. Recent approach [32, 75] leverages the long-sequence reasoning capabilities of Large Language Models (LLMs) [1, 64] for long-form video question answering, often by converting videos into densely and uniformly sampled frames [33, 38, 55]. However, this approach is usually inefficient due to the inherent imbalance and varying information density across video segments [61, 78]. Some segments are rich in critical details necessary for answering questions, while others contain only trivial information [66]. Uniform sampling [49, 60], which might focus on redundant frames that increase the computational burden and may overlook frames with essential information, often fails to efficiently utilize the video's dynamic structure for question answering [17, 62]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address this limitation, recent approaches [29, 34, 36, 68] employ complex architectures [63, 85] or specialized preprocessing techniques [44, 65] to improve the efficiency" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.22, + 0.9 + ], + "angle": 0, + "content": "† Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20168" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "of long video content processing. These methods [62, 66] often convert videos into text by captioning densely sampled frames or incorporating memory modules [5, 17, 39] to store essential information. However, these approaches still incur significant computational costs. For example, captioning an hour-long video with Qwen2-VL-72B at just 1 fps requires over 3 hours on a single H100. Additionally, these methods [12] identify key frames based on image similarity rather than a holistic understanding of the video content, which can result in missing relevant video features. This necessitates the need for an efficient and scalable solution for long-form video understanding." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.314, + 0.485, + 0.722 + ], + "angle": 0, + "content": "In this work, we introduce VCA, a curiosity-driven agent inspired by how humans tackle the long video understanding task. Given a question, human is usually intrinsically motivated to identify relevant information in the video without external knowledge. In practise, human tends to build a global understanding of the video, then curiously zoom in to explore specific sections in greater detail. Motivated by this process, we propose a novel video agent framework, dubbed as \"VCA\", to emulate such behavior. VCA treats the video as an environment to explore, actively navigating its contents and identifying segments that merit deeper analysis. VCA incorporates three essential techniques: (1) Tree-search Exploration: Rather than selecting specific frames [62], VCA uses a tree-search structure to adaptively explore the video segments in a structured way. (2) Reward Model: VCA leverages an intrinsic reward scoring mechanism to guide exploration toward the segments most relevant to the query. (3) Memory Management: VCA employs a fixed memory buffer to store key information and discard irrelevant frames, effectively reducing computational costs and avoiding memory overflow. The key difference between VCA and prior works is outlined in Fig. 2. Extensive experiments and superior performance across multiple benchmarks demonstrate that our proposed VCA framework significantly improves an agent's effectiveness and efficiency in reasoning over extended video content. More concretely, our contributions are threefold," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.762, + 0.483, + 0.808 + ], + "angle": 0, + "content": "- VCA is a long video understanding agent that can self-exlore the most informative video segments based on its intrinsic reward, without relying on any auxiliary tools." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.808, + 0.483, + 0.839 + ], + "angle": 0, + "content": "- VCA is easy to implement, training-free, and is flexible to integrate with any vision-language model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.838, + 0.484, + 0.899 + ], + "angle": 0, + "content": "- As highlighted in Fig. 1, VCA outperforms existing baselines while observing fewer frames and incurring comparable or lower computational cost, demonstrating both effectiveness and efficiency." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.762, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.093, + 0.882, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Comparison of VCA with prior works. Difference-1: When answering a query, VCA uniformly samples (a few) frames, whereas previous methods (e.g., VideoTree) require captioning each frame for subsequent image retrieval, which can be computationally expensive for long videos. Difference-2: previous methods rely on image similarity to locate key frames, potentially losing important video features. In contrast, VCA uses segment-level exploration to identify the most informative segments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.383, + 0.656, + 0.398 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.811, + 0.424 + ], + "angle": 0, + "content": "2.1. Long-context Multi-modal Agents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Large multi-modal models (LMMs) [2, 6, 9, 35, 54] have demonstrated promising performance in addressing complex multimodal tasks [20, 45, 49, 81]. However, these LMMs typically lack self-directed exploration capabilities required for handling more tricky tasks [24, 74]. To bridge this gap, multimodal agents are designed to enhance LMMs with the ability to autonomously make decisions and execute tasks within specific environments [15, 73], such as interpreting visual cues and text prompts to make real-time decisions within websites [28, 76]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.779 + ], + "angle": 0, + "content": "A key feature of multimodal agents is their long-context capability [84], i.e., the ability to process and utilize information across extended periods to make informed decisions [26]. An example of this is vision-language model (VLM) agents for video understanding tasks. Despite their advancements, even sophisticated VLM agents struggle to extract crucial information and perform effective reasoning over lengthy inputs [62, 79]. Therefore, efficiently extracting essential information and making informed decisions remains a primary challenge. This work aims to address this challenge by enabling agents to self-exlore the environment and self-gather the most relevant information to efficiently solve video-understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.761, + 0.804 + ], + "angle": 0, + "content": "2.2. Long Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Early deep learning models for video understanding primarily followed two main approaches [52]. One uses two-stream networks [18, 19, 42, 58] to integrate spatial and motion information, while the second focuses on 3D CNNs [8, 57, 72] for spatial-temporal feature extraction. Vision Transformers [14] then introduce self-attention mechanisms" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.198, + 0.082, + 0.805, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.444, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 3. Pipeline of VCA Framework. First, the reward model predicts the relevance scores of each segment. Next, the memory buffer is updated by adding new frames and discarding irrelevant ones. Finally, emulating human reasoning, the agent actively selects segments to explore until the correct answer is found." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.549 + ], + "angle": 0, + "content": "to handle temporal modeling more effectively [3, 7, 16]. Methods with self-supervised pre-training [50, 56, 86] exhibit transferability across different tasks and achieve more robust video understanding performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Multimodal LLM [22, 59, 83] shows strong capabilities in video understanding tasks, but efficiently processing information within a prolonged video environment is the key challenge. Recent approaches often use frame sampling [32, 46, 63, 75, 77] and feature extraction [27, 49, 51, 70] to reduce computational costs. Frame sampling selects a subset of frames, which risks missing essential context in lengthy videos, where transitions and subtle cues may be easily neglected. Additionally, processing entire video sequences as tokens presents significant memory and computational challenges [11, 43, 64, 80]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Recent work has explored auxiliary techniques and modules to improve efficiency, such as pre-extracted content like captions [44, 62], multimodal summaries [40, 65], and combinations of various auxiliary tools [17, 30]. However, these tools introduce additional computational and memory costs. Moreover, some approaches still require traversing the entire video content [17]. Notably, VideoTree [66] requires captioning every frame in the video, making it prohibitively expensive. In contrast, our VCA can solve the task using a single VLM, without relying on any supplementary tools." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.486, + 0.605, + 0.501 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Inspired by human reasoning, we introduce a self-exploration agent that adaptively navigates video environments. Unlike approaches that depend on dense frame sampling or captioning [62], which require VLMs to process numerous frames, our agent analyzes videos iteratively within a limited context window in a process we call self-exploration. This approach allows essential information to be extracted from long videos in a coarse-to-fine manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Specifically, we propose a segment-based tree-search exploration structure. Given a video \\( v \\) and a user query \\( q \\), the agent \\( \\pi \\) begins by randomly sampling \\( N \\) frames \\( \\{f_v^i\\}_{i=1}^N \\) to obtain an initial overview of the content in each video segment \\( v_i \\in \\{v_i\\}_{i=1}^{N+1} \\). The agent then decides whether additional exploration is needed to answer the query. If further investigation is necessary, the agent selects a specific segment \\( s^* \\in \\{v_i\\}_{i=1}^{N+1} \\) for deeper exploration until sufficient information is gathered to respond to the query." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Nevertheless, a one-hour movie typically contains approximately 86,400 frames, posing a challenge for current VLM-based agents to efficiently identify the most relevant segment [17, 62, 66]. Moreover, even with a small \\( N \\), conducting multiple rounds of exploration accumulates a substantial number of frames, leading to high computational costs. To address these issues, we introduce a reward model \\( R \\) to guide the agent's exploration and propose a memory" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20170" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.405, + 0.106 + ], + "angle": 0, + "content": "buffer \\(M\\) to reduce storage and inference costs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.184 + ], + "angle": 0, + "content": "In this section, we first explain how the reward model assists the agent throughout the exploration process (Sec. 3.1). Then we introduce our memory management strategy in Sec. 3.2. Finally, we explain how VCA adaptively explores the video using a segment-based tree structure (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.243, + 0.211 + ], + "angle": 0, + "content": "3.1. Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.484, + 0.356 + ], + "angle": 0, + "content": "With the segment-based tree-search exploration, the agent is required to select the most informative video segment for exploration at each step. However, we observe that current VLM-based agents struggle to independently identify optimal exploration paths (see Sec. 4.3). To this end, inspired by the selective attention mechanism in human cognition [13], we incorporate a reward model \\( R \\) to guide the agent's selection process, i.e., the agent will gain a higher reward if it plans to take the action to explore more relevant segment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Specifically, during the exploration of a current selected segment \\( s^* \\), the reward model observes a set of frames \\( \\{f_{s^*}^i\\}_{i=1}^N \\) sampled from \\( s^* \\). These frames \\( \\{f_{s^*}^i\\}_{i=1}^N \\) then form finer sub-segments \\( \\{s_i^* = [l_{s^*}^i, l_{s^*}^{i+1}]\\}_{i=1}^{N+1} \\), where \\( l_{s^*}^i \\) and \\( l_{s^*}^{i+1} \\) represent the indices of the beginning and end frame of the sub-segment, respectively. The reward model aims to generate a reward score for each sub-segment \\( \\{s_i^*\\}_{i=1}^{N+1} \\) based on its relevance to the user query \\( q \\). We utilize chain-of-thoughts [67] technique to let the reward model provide a verbal explanation of each segment and then generate a relevant score, i.e., \\( \\{t_{s_i^*}; r_{s_i^*}\\}_{i=1}^{N+1} \\). Additionally, to maintain consistency across different exploration steps, we incorporate the reward history \\( H_r \\) from previous steps alongside the new segments during evaluation. A simplified prompt template for the reward model is presented here, with detailed prompts in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.624, + 0.354, + 0.64 + ], + "angle": 0, + "content": "Reward Model Prompt (Simplified)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.652, + 0.166, + 0.665 + ], + "angle": 0, + "content": "Given:" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.666, + 0.4, + 0.682 + ], + "angle": 0, + "content": "\\(H_{r}\\), reward score history in previous steps." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.681, + 0.457, + 0.713 + ], + "angle": 0, + "content": "\\(\\left\\{f_{s^*}^i\\right\\}_{i = 1}^N\\) newly sampled frames from current segment \\(s^{*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.712, + 0.458, + 0.761 + ], + "angle": 0, + "content": "Please first explain each sub-segment \\(\\{s_i^*\\}_{i = 1}^{N + 1}\\) and then assign a relevance score of each sub-segment to the user query \\(q\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In practice, rather than training a separate specialized model, we utilize a same VLM-based agent to function as the reward model. While the reward model \\( R \\) and exploration agent \\( \\pi \\) serve distinct purposes, both can be implemented using the same model. In this work, we employ a single VLM-based agent for both roles across different settings, providing a neat and efficient agent framework that is guided by its self-generated intrinsic reward and without" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.756, + 0.107 + ], + "angle": 0, + "content": "relying on auxiliary models or tools." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.117, + 0.722, + 0.133 + ], + "angle": 0, + "content": "3.2. Memory Management" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.905, + 0.274 + ], + "angle": 0, + "content": "As exploration proceeds over multiple rounds, the accumulation of frames can lead to significant computational costs for long videos, even when \\( N \\) is small. However, as illustrated in Fig. 2, they tend to gradually disregard irrelevant frames, especially from earlier steps, retaining only the most crucial information in their working memory for later analysis. Inspired by this, we introduce a fixed-size memory buffer \\( M \\) that stores only the most relevant frames throughout exploration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.906, + 0.477 + ], + "angle": 0, + "content": "Specifically, when the agent \\(\\pi\\) explores a video segment \\(s^*\\), the associated frames \\(\\{f_{s^*}^i\\}_{i=1}^N\\) are added to the memory buffer. If the total number of frames exceeds the buffer limit, irrelevant frames are removed based on relevance scores \\(r_{s_i^*}\\) defined in Sec. 3.1. We discard the frames with the lowest scores and only retain the most relevant ones. While these less relevant frames are removed, their visual information is extracted and encoded as text descriptions \\(\\{t_{s_i^*}\\}_{i=1}^N\\) within the exploration history \\(H_r\\), allowing the reward model to utilize. This memory management strategy enables the agent to conduct reasoning within a fixed resource limit, regardless of the exploration trajectory's length, thereby preventing excessive memory usage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.74, + 0.503 + ], + "angle": 0, + "content": "3.3. Tree-Search Exploration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Inspired by humans adaptively focusing on different video segments during exploration rather than targeting specific frames directly, our proposed agent VCA treats the video as an unexplored environment, where each segment represents a node on an exploration tree. Given the entire video as the root node, the agent's goal is to efficiently identify an optimal short path leading to a leaf node (segment) containing the most informative content needed to answer the query." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.749 + ], + "angle": 0, + "content": "At each step, the agent \\(\\pi\\) receives both textual and visual guidance: a set of candidate segments \\(S\\), which includes a collection of segments with associated reward scores, and a memory buffer \\(M\\) containing multiple frames. If the agent has gathered sufficient information to answer the user's query \\(q\\), it directly generates a response \\(a\\). Otherwise, it continues exploring by selecting a specific segment \\(s^* \\in S\\) to expand the tree based on reward scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "However, due to the limited capacity of current foundational models, these reward scores are not always accurate (see further analysis in Sec. 5.3). Selecting segments based on reward scores in a purely greedy manner can potentially lead to a suboptimal solution. Instead, we use the reward scores as guidance for agent \\(\\pi\\), allowing it to make the final segment selection independently. In this case, the agent is not forced to select the segment with the highest reward, balancing the exploitation and exploration. More importantly, the candidate set \\(S\\) includes both unexplored" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "coarse segments from prior steps and fine-grained segments from the current step, enabling the agent to backtrack and select earlier segments if it detects that the current exploration path may be suboptimal. Below are the simplified prompts. The full prompt is provided in Appendix ??" + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.183, + 0.378, + 0.2 + ], + "angle": 0, + "content": "Exploration Agent Prompt (Simplified)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.211, + 0.166, + 0.224 + ], + "angle": 0, + "content": "Given:" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.226, + 0.296, + 0.241 + ], + "angle": 0, + "content": "Candidate segments set: \\(S\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.242, + 0.316, + 0.256 + ], + "angle": 0, + "content": "Frames in Memory buffer: \\(M\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.257, + 0.458, + 0.3 + ], + "angle": 0, + "content": "Please choose one segment to explore based on the query \\( q \\). Output answer if you have gathered sufficient information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.428 + ], + "angle": 0, + "content": "The overall framework of our proposed method is summarized in Algo. 1. Leveraging a tree-search exploration technique boosted with a reward model and a memory buffer, our self-exploration agent adaptively and efficiently explores the video like a human. Empirical results in Sec. 4.3 further validate the effectiveness of our tree-search exploration over naive frame selection methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.431, + 0.224, + 0.448 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.457, + 0.298, + 0.474 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Benchmark. We mainly evaluate VCA on two long-term video question-answering benchmarks. EgoSchema [41] is an egocentric video dataset sourced from Ego4D [21], with a duration of 180 seconds. Here we use the validation set of 500 QA pairs. LVBench [61] covers about 1,500 samples for evaluation of reasoning abilities from high-quality videos over 30 minutes in length. We also conduct experiments on other relevant benchmark datasets in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.484, + 0.76 + ], + "angle": 0, + "content": "Baselines. We compare with most relevant baselines to our work, including state-of-the-art open-source Video-LLMs [10, 23, 37, 47, 59, 82], the close-source model GPT-4o [25], and agent-based systems [44, 62, 66]. For video models, we report the results obtained with uniform sampling frames, while for the agent-based systems, we follow their official implementation with GPT-4o [25], consistent with VCA's setup to ensure a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. In this work, we use the same model for the exploration agent and the reward model. Unless otherwise stated, all experiments were performed using the August 2024 version of `gpt-4` as the base model, leveraging the OpenAI LLM API service1 with a temperature of 0.5. We set memory buffer sizes set to 8 and 16 frames for EgoSchema and LVBench, respectively. Further implementation details are provided in Appendix ??" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.79, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 VCA: Video Curious Agent" + }, + { + "type": "code", + "bbox": [ + 0.516, + 0.111, + 0.907, + 0.156 + ], + "angle": 0, + "content": "Require: Video environment \\( v \\), question \\( q \\), agent \\( \\pi \\), reward model \\( R \\), memory buffer \\( M \\), candidate segments set \\( S \\), reward score history \\( H_{r} \\), sampling frame number \\( N \\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.157, + 0.646, + 0.169 + ], + "angle": 0, + "content": "Ensure: Answer \\( a \\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.171, + 0.773, + 0.186 + ], + "angle": 0, + "content": "1: Initialize selected segment \\( s^* \\gets v \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.187, + 0.806, + 0.201 + ], + "angle": 0, + "content": "2: Initialize candidate segments set \\( S \\gets \\emptyset \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.203, + 0.761, + 0.216 + ], + "angle": 0, + "content": "3: Initialize memory buffer \\(M\\gets \\emptyset\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.217, + 0.799, + 0.231 + ], + "angle": 0, + "content": "4: Initialize reward score history \\(H_{r}\\gets \\emptyset\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.233, + 0.679, + 0.245 + ], + "angle": 0, + "content": "5: continue \\(\\leftarrow\\) true" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.247, + 0.687, + 0.26 + ], + "angle": 0, + "content": "6: while continue do" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.263, + 0.885, + 0.276 + ], + "angle": 0, + "content": "7: # Sampling frames & form segments" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.277, + 0.853, + 0.295 + ], + "angle": 0, + "content": "8: \\(\\left\\{f_{s^*}^*\\right\\}_{i=1}^{N}, \\left\\{s_i^*\\right\\}_{i=1}^{N+1} \\gets \\text{UniformSample}(s^*)\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.296, + 0.777, + 0.308 + ], + "angle": 0, + "content": "9: # Scoring each segment" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.309, + 0.887, + 0.327 + ], + "angle": 0, + "content": "10: \\(\\left\\{t_{s_i^*};r_{s_i^*}\\right\\}_{i = 1}^{N + 1}\\gets R(\\{s_i^*\\}_{i = 1}^{N + 1};q,\\{f_{s_*}^i\\}_{i = 1}^N,H_r)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.328, + 0.854, + 0.34 + ], + "angle": 0, + "content": "11: # Update candidate segment set" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.341, + 0.716, + 0.359 + ], + "angle": 0, + "content": "12: \\(S\\gets S\\cup \\{s_i^*;r_{s_i^*}\\}_{i = 1}^{N + 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.36, + 0.855, + 0.372 + ], + "angle": 0, + "content": "13: # Update history with thoughts" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.373, + 0.739, + 0.391 + ], + "angle": 0, + "content": "14: \\(H_{r}\\gets H_{r}\\cup \\left\\{t_{s_{i}^{*}};r_{s_{i}^{*}}\\right\\}_{i = 1}^{N + 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.392, + 0.826, + 0.404 + ], + "angle": 0, + "content": "15: # Update memory with frames" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.406, + 0.828, + 0.424 + ], + "angle": 0, + "content": "16: \\(M\\gets \\mathrm{UpdateMemory}(M,\\{f_{s^{*}}^{i}\\}_{i = 1}^{N};S)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.424, + 0.884, + 0.436 + ], + "angle": 0, + "content": "17: # Output answer or select segment" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.438, + 0.684, + 0.451 + ], + "angle": 0, + "content": "18: \\(a,s\\gets \\pi (q,S;M)\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.453, + 0.65, + 0.465 + ], + "angle": 0, + "content": "19: if \\(a = \\emptyset\\) then" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.468, + 0.794, + 0.482 + ], + "angle": 0, + "content": "20: # Continue exploration" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.484, + 0.627, + 0.494 + ], + "angle": 0, + "content": "21: \\(s^* \\gets s\\)" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.498, + 0.589, + 0.51 + ], + "angle": 0, + "content": "22: else" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.513, + 0.714, + 0.525 + ], + "angle": 0, + "content": "23: continue \\(\\leftarrow\\) false" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.528, + 0.604, + 0.54 + ], + "angle": 0, + "content": "24: end if" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.544, + 0.613, + 0.556 + ], + "angle": 0, + "content": "25: end while" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.559, + 0.609, + 0.571 + ], + "angle": 0, + "content": "26: return \\( a \\)" + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.171, + 0.887, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.716, + 0.616 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.907, + 0.802 + ], + "angle": 0, + "content": "The experimental results on LVBench and EgoSchema are presented in Tab. 2 and Tab. 1, respectively. By adaptively focusing on critical segments rather than scanning the entire video, our agent not only achieves stronger performance but also incurs less computational cost. With less than \\(30\\%\\) of the frames, VCA achieves substantial improvements over directly feeding uniformly sampled frames into GPT-4o. Specifically, we observe performance gains of \\(3.2\\%\\) on the EgoSchema subset and \\(6.6\\%\\) on LVBench. Significant increases in the Event Understanding and Temporal Grounding domains further underscore our effectiveness in identifying crucial segments within long videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Furthermore, we compare our framework with recently proposed SOTA video agent systems, VideoAgent [62] and VideoTree [66]. For a fair comparison, we re-implement both frameworks using GPT-4o for all reasoning components, consistent with our setup. However, captioning all" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.848, + 0.9 + ], + "angle": 0, + "content": "1https://platform.openai.com/docs/models" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20172" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.209, + 0.089, + 0.789, + 0.254 + ], + "angle": 0, + "content": "
MethodFramesEREUKIRTGReaSumAvg.
TimeChat* [47]>9621.921.725.922.725.024.122.3
LLaVA-NeXT* [83]3230.131.234.131.435.027.632.2
InternVL2-40B* [10]1637.439.743.431.442.541.439.8
GLM4V-Plus* [23]2039.935.834.837.740.032.838.3
GPT-4o* [25]6435.930.835.528.333.534.534.7
VideoAgent† [62]Avg. 25.528.030.328.029.328.036.429.3
VideoTree† [66]Avg. 103.230.325.126.527.731.925.528.8
VCA (ours)Avg. 20.043.740.737.838.046.227.341.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.265, + 0.907, + 0.308 + ], + "angle": 0, + "content": "Table 1. Experimental Results on LVBench. Results of methods marked with * are sourced from LVBench [61]. Methods marked with † indicate implementations where GPT-4o serves as the main component. Our agent's memory buffer size is set to 16 frames. We report the average number of observed frames for agent-based methods, with the best performance highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.324, + 0.423, + 0.475 + ], + "angle": 0, + "content": "
MethodFramesSubset
GPT-4o [25]869.0
3270.4
\\(VideoAgent^\\dagger\\) [62]Avg. 8.163.2
Avg. 10.760.8
\\(VideoTree^\\dagger\\) [66]Avg. 69.567.0
\\(LVNet^\\dagger\\) [44]1268.2
VCA (ours)Avg. 7.273.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Table 2. Experimental Results on EgoSchema. Methods marked with \\(\\dagger\\) indicate implementations where GPT-4o is the main component. The results of LVNet are from [44]. Our agent's memory buffer size is set to 8 frames. We report the average number of observed frames, with the best performance highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.579, + 0.483, + 0.791 + ], + "angle": 0, + "content": "frames with GPT-4o is prohibitively expensive. Hence, following the baseline implementations, we use open-source VLMs for captioning. A comparison using the same model for all components is provided in Sec. 5.1. As shown in Tab. 1, VCA outperforms all baselines over \\(5\\%\\) on EgoSchema, while requiring processing fewer frames. These results suggest the limitations of dense captioning for video preprocessing tasks, which can incur more computational costs while introducing more redundant information. In contrast, our proposed memory management strategy effectively improves efficiency by only retaining the most crucial information for reasoning. Similar insights can be drawn from results on LVBench in Tab. 2. Additional results are provided in Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Moreover, we observe that VCA explores approximately three times more frames on LVBench than on EgoSchema. Recall that EgoSchema videos have an average duration of 180 seconds, while videos in LVBench extend to at least 30 minutes. This phenomenon suggests that our proposed tree-search exploration technique actively explores the video to locate the most relevant information and adaptively in-" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.324, + 0.88, + 0.413 + ], + "angle": 0, + "content": "
MethodEgoSchemaLVBench
FramesAcc.FramesAcc.
VCA (ours)7.273.620.041.3
- w/o Reward7.572.622.239.5
- w/o Tree Search10.269.839.936.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.423, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Table 3. Ablation Study on EgoSchema and LVBench. Both the tree-search exploration and reward model contribute significantly to improvements in performance and efficiency. Detailed ablation study results on LVBench are included in Appendix ??" + }, + { + "type": "table", + "bbox": [ + 0.604, + 0.487, + 0.819, + 0.531 + ], + "angle": 0, + "content": "
Buffer Size81632
Acc.38.941.343.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.542, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Table 4. Ablation Study of Buffer Size on LVBench. Increasing the buffer size allows the model to observe more frames during reasoning, leading to improved performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.705 + ], + "angle": 0, + "content": "creases the exploration budget based on task complexity. For example, VCA exhibits both exploitation and exploration: when the agent determines that the current exploration path is suboptimal, it backtracks to previous segments to explore a new path (see Sec. 5.2). These empirical observations further validate the robustness and scalability of our approach, especially in diverse long-video contexts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.667, + 0.73 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.885 + ], + "angle": 0, + "content": "In this section, we conduct an ablation study to evaluate the components of our agent. Results across both benchmarks are presented in Tab. 3. As illustrated in the table, performance consistently declines when the reward model is removed, highlighting that the agent's inherent reasoning ability alone is insufficient for accurately identifying the most informative segments. This finding emphasizes the effectiveness of incorporating an intrinsic reward to guide the agent's segment selection process. A detailed analysis of our reward model's performance is provided in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.887, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Additionally, we observe a significant performance drop" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20173" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.223, + 0.097, + 0.365, + 0.109 + ], + "angle": 0, + "content": "Why does the mice run away?" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.109, + 0.456, + 0.119 + ], + "angle": 0, + "content": "(A) Because the black cat sheriff calls the police" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.119, + 0.38, + 0.128 + ], + "angle": 0, + "content": "(C) Because their boss is injured" + }, + { + "type": "list", + "bbox": [ + 0.223, + 0.097, + 0.456, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.461, + 0.108, + 0.664, + 0.118 + ], + "angle": 0, + "content": "(B) Because they meet more powerful mice" + }, + { + "type": "text", + "bbox": [ + 0.461, + 0.118, + 0.666, + 0.127 + ], + "angle": 0, + "content": "(D) Because the owner of the house is back" + }, + { + "type": "list", + "bbox": [ + 0.461, + 0.108, + 0.666, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.133, + 0.78, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.189, + 0.311, + 0.2 + ], + "angle": 0, + "content": "Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.201, + 0.599, + 0.23 + ], + "angle": 0, + "content": "Segment1: Thought: The segment shows a black cat with a gun, possibly indicating a threat. This could suggest that the mice might run away due to the presence of the cat, which aligns with option (A). Score: 0.7" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.236, + 0.603, + 0.267 + ], + "angle": 0, + "content": "Segment4: Thought: The segment features the black cat with a communication device, possibly calling for help or backup. This aligns with option (A), suggesting the mice might run because the cat is calling the police. Score: 0.8" + }, + { + "type": "list", + "bbox": [ + 0.225, + 0.201, + 0.603, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.188, + 0.714, + 0.21 + ], + "angle": 0, + "content": "Agent Choice: (segment: 4)" + }, + { + "type": "image", + "bbox": [ + 0.611, + 0.212, + 0.697, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.212, + 0.778, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.273, + 0.78, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.224, + 0.328, + 0.309, + 0.339 + ], + "angle": 0, + "content": "Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.34, + 0.602, + 0.37 + ], + "angle": 0, + "content": "Segment4.1: Thought: The segment shows a close-up of a document with red text and exclamation marks, indicating urgency or importance. This could imply a critical message or alert being communicated. Score: 0.7" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.37, + 0.234, + 0.38 + ], + "angle": 0, + "content": "…" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.388, + 0.779, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.444, + 0.311, + 0.454 + ], + "angle": 0, + "content": "Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.455, + 0.585, + 0.497 + ], + "angle": 0, + "content": "Segment1.2: Thought: The segment features mice characters in mid-air, possibly fleeing or being thrown. This suggests a direct reaction to a threat, supporting the idea that they are running away from the cat. Score: 0.8" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.503, + 0.779, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.543, + 0.604, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.642, + 0.545, + 0.748, + 0.558 + ], + "angle": 0, + "content": "Agent: Answer: A" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.57, + 0.905, + 0.598 + ], + "angle": 0, + "content": "Figure 4. Exploration Trajectory Example. The agent generally tends to select video segments with higher reward scores for exploration, yet it can adaptively redirect to earlier segments, effectively balancing exploitation and exploration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.792 + ], + "angle": 0, + "content": "when the tree-search exploration technique is ablated, i.e., the agent directly selects frame indices without the segment-based tree structure. Interestingly, this modification also results in a substantial increase in the average number of frames acquired, especially on LVBench. This indicates that the agent struggles to locate the most crucial information and consequently produces a longer but redundant trajectory. These ablation study results validate the effectiveness of the tree-search exploration framework assisted by the reward model. Additionally, this underscores VCA's potential to achieve greater efficiency and improved performance with more advanced VLMs and reward models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Moreover, as shown in Tab. 4, increasing the memory buffer size yields consistently improved performance for VCA, demonstrating that our framework benefits from a broader reasoning window while maintaining effectiveness with fewer frames. Due to limited computation resources, we only tested up to 32 frames; however, the upward trend indicates the potential for gains with additional frames." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.61, + 0.61, + 0.627 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.898, + 0.652 + ], + "angle": 0, + "content": "5.1. Can VCA generalize to Open-source Models?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "While VCA demonstrates strong performance using GPT4o as the base model, we also analyze whether these positive results can be extrapolated to open-source models. We implement the VCA framework with Qwen2-VL-72B-Instruct-GPTQ-Int4 [59] as both the exploration agent and reward model, as well as the captioner for baselines. The results in Tab. 5 reveal that, compared to directly feeding uniformly sampled frames into Qwen2-VL, our framework achieves higher accuracy on the EgoSchema benchmark despite using only about \\(60\\%\\) of the frames. In contrast, VideoAgent performs significantly worse. Furthermore, our approach outperforms VideoTree with far fewer frames. Notably, VideoTree employs clustering to build a caption similarity-based tree structure for frame selection, requiring caption generation for each frame. This process is computationally expensive, demanding over 60 GPU hours" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20174" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.089, + 0.428, + 0.224 + ], + "angle": 0, + "content": "
MethodFramesSubset
Qwen2-VL870.2
3274.2
VideoAgentAvg. 8.665.2
VideoTreeAvg. 85.471.0
VCA (ours)Avg. 16.975.2
- w/ Image RewardAvg. 20.974.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.234, + 0.483, + 0.292 + ], + "angle": 0, + "content": "Table 5. Performance of Using Open-source Model on EgoSchema. Our approach achieves higher accuracy with fewer observed frames, showing consistency improvements over baseline methods. Memory buffer size is set to 16 frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.353 + ], + "angle": 0, + "content": "on a single H100, as shown in Fig. 1. These results further underscore the robustness of VCA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.453, + 0.374 + ], + "angle": 0, + "content": "5.2. What's the Exploration Behavior of VCA?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.609 + ], + "angle": 0, + "content": "Here we present a real exploration trajectory from LVBench to illustrate our agent's decision-making process. As shown in Fig. 4, given a user query, the reward model initially evaluates each segment's relevance based on uniformly sampled frames, and the agent selects the fourth segment, which has the highest relevance score. However, upon further investigation of segment 4, both the reward model and the agent find it relatively irrelevant to the query. Rather than continuing with it, the agent redirects its exploration to segment 1 in the last round. After selecting segment 1, the agent chooses to investigate the third segment, despite the second segment having the highest reward score. This behavior reflects VCA's ability to adaptively adjust its exploration trajectory based on visual context rather than purely following the reward score, resulting in a more robust system." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.763 + ], + "angle": 0, + "content": "We also examine the limitations of VCA by analyzing its failure cases, which fall into the following scenarios: The most common failure occurs when the agent overlooks subtle clues, resulting in an inaccurate response to the user query even after prolonged exploration, particularly for queries requiring specific visual details. Other failure cases include instances where the agent is misled by inaccurate reward scores or fails to provide the correct answer despite observing key frames. Detailed cases and further discussion are provided in the Appendix ??" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In addition, unlike previous methods, VCA employs a segment-level exploration strategy instead of direct image similarity to select key frames. To further investigate this difference, we evaluate the performance of selecting the segment whose start and end images have the highest visual relevance. As shown in Tab. 5, relying solely on image relevance instead of inferring segment semantics results in a performance drop on EgoSchema, confirming the effectiveness of our segment-level exploration." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.09, + 0.88, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.327, + 0.906, + 0.396 + ], + "angle": 0, + "content": "Figure 5. Segment Selection Accuracy on LVBench. The y-axis shows segment distance from ground truth; x-axis shows count ratio. Red curves are fitted normal distributions. The reward-guided agent selects segments closer to ground truth than the reward model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.825, + 0.437 + ], + "angle": 0, + "content": "5.3. How Reliable is the Reward Model?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.581 + ], + "angle": 0, + "content": "To further assess the capability of our reward model, we visualize the matching accuracy between segments with the highest reward scores and ground truth time references from LVBench. The results are shown at the bottom of Fig. 5. We observe that segment distances follow an approximately normal distribution with a mean of 1.28, indicating that the segment with the highest reward score often closely aligns with the ground truth segment. This validates the reward model's effectiveness in providing meaningful guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Additionally, as discussed in Sec 3.3, we encourage the agent to make independent decisions when selecting segments rather than blindly following reward scores in a purely greedy manner. The benefit of this design choice is demonstrated at the top of Fig. 5, where we present the matching accuracy of the agent's choices. Compared to the greedy approach, the agent's selections result in a sharper distribution of segment distances, with an average distance gap of \\(-0.60\\), indicating closer alignment with the ground truth. In summary, our reward model assigns reliable relevance scores to segments, while the agent further refines these choices to achieve even more accurate selections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.778, + 0.634, + 0.794 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a curiosity-driven video agent framework for long video understanding tasks with self-exploration capability. VCA achieves state-of-the-art performance and efficiency without relying on auxiliary tools. Appendix ?? shows that improved reward guidance could further boost results, highlighting the potential of synthetic data and specialized reward models for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20175" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.482, + 0.212 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.482, + 0.282 + ], + "angle": 0, + "content": "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[5] Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J Henaff. Memory consolidation enables long-context video understanding. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for ai agents, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding?, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[11] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[12] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. arXiv preprint arXiv:2311.01233, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Robert Desimone, John Duncan, et al. Neural mechanisms of selective visual attention. Annual review of neuroscience, 18(1):193-222, 1995. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, Katsushi Ikeuchi, Hoi Vo, Li Fei-Fei, and Jianfeng Gao. Agent ai: Surveying the horizons of multimodal interaction, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[17] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2025. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[18] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[22] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, Lei Zhao, Zhuoyi Yang, Xiaotao Gu, Xiaohan Zhang, Guanyu Feng, Da Yin, Zihan Wang, Ji Qi, Xixuan Song, Peng Zhang, Debing Liu, Bin Xu, Juanzi Li, Yuxiao Dong, and Jie Tang. CogvIm2: Visual language models for image and video understanding, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[23] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20176" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[26] Lawrence Jang, Yinheng Li, Charles Ding, Justin Lin, Paul Pu Liang, Dan Zhao, Rogerio Bonatti, and Kazuhito Koishida. Videowebarena: Evaluating long context multimodal agents with video understanding web tasks, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[28] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[29] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In European Conference on Computer Vision, pages 271-288. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[30] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[32] Yunxin Li, Xinyu Chen, Baotain Hu, and Min Zhang. Llms meet long video: Advancing long video comprehension with an interactive visual adapter in llms, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[34] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[35] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[36] Ming C Lin and Shan Yang. Vila: Efficient video-language alignment for video question answering. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[38] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[39] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[42] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[43] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14386-14397, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[44] Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya, Wonjeong Ryoo, Donghyun Kim, and Michael S Ryoo. Too many frames, not all useful: Efficient strategies for long-form video qa. arXiv preprint arXiv:2406.09396, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[45] Ruchit Rawal, Khalid Saifullah, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. In Synthetic Data for Computer Vision Workshop @ CVPR 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[46] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. arXiv preprint arXiv:2312.02051, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[47] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[48] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[50] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[51] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Advances in neural information processing systems, 35:38032-38045, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Yunlong Tang, Jing Bi, Sitting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, Ali Vosoughi, Chao Huang, Zeliang Zhang, Pinxin Liu, Mingqian Feng, Feng Zheng, Jianguo Zhang," + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20177" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Ping Luo, Jiebo Luo, and Chenliang Xu. Video understanding with large language models: A survey, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[53] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[54] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[57] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[59] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[61] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Shiyu Huang, Bin Xu, Yuxiao Dong, Ming Ding, and Jie Tang. Lvbench: An extreme long video understanding benchmark, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[62] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[63] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[64] Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recurrent memory bridges. arXiv preprint arXiv:2409.01071, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[65] Ying WANG, Yanlai YANG, and Mengye REN. Lifelong-memory: Leveraging llms for answering queries in long" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "form egocentric videos. arXiv preprint arXiv:2312.05269, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[66] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[67] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[68] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision, pages 453-470. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[69] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[70] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[71] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. CoRR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[72] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[73] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. In ICLR 2024 Workshop on Large Language Model (LLM) Agents. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[74] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[75] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Retrieval-based video language model for efficient long video question answering, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[76] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[77] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[78] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[79] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20178" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[81] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[82] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[83] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[84] Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang. Longagent: Scaling language models to 128k context through multi-agent collaboration. arXiv preprint arXiv:2402.11550, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[85] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In *Forty-first International Conference on Machine Learning*. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[86] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20179" + } + ] +] \ No newline at end of file diff --git a/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_origin.pdf b/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..348be723a3636a7c8a6d319269d1bce4399fb008 --- /dev/null +++ b/2025/VCA_ Video Curious Agent for Long Video Understanding/a73c253b-5afa-42a6-b79a-31559f118997_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b2d12f8092f3508d39a82b284d887acba4d240d7e9bea4e4a37aecc06ab16f8 +size 5127347 diff --git a/2025/VCA_ Video Curious Agent for Long Video Understanding/full.md b/2025/VCA_ Video Curious Agent for Long Video Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7b61253f7522f7ef69ae488953c509a69800ee6 --- /dev/null +++ b/2025/VCA_ Video Curious Agent for Long Video Understanding/full.md @@ -0,0 +1,383 @@ +# VCA: Video Curious Agent for Long Video Understanding + +Zeyuan Yang† +University of Massachusetts, Amherst +yangzeyuan@umass.edu + +Xueyang Yu +University of Massachusetts, Amherst +xueyangyu@umass.edu + +Delin Chen† +University of Massachusetts, Amherst delinchen@umass.edu + +Maohao Shen +Massachusetts Institute of Technology +maohao@mit.edu + +Chuang Gan +University of Massachusetts, Amherst +ganchuang1990@gmail.com + +# Abstract + +Long video understanding poses unique challenges due to their temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as "VCA". Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences. Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency. + +# 1. Introduction + +There is a growing demand for developing long-form video understanding techniques [48, 69] capable of extracting valuable information from extended visual narratives, such as surveillance footage, documentaries, and instructional videos. Analyzing these videos [53, 71], which can range from minutes to hours, requires models that can process multimodal data and perform reasoning over extremely long sequences [4, 31, 60]. + +Previous works commonly utilize two-stream networks to capture spatial and temporal information [18, 19] or in + +![](images/b3a34f2280701b001980d124b23ff96bfa7878b2f2423e84306fc895c68a5841.jpg) +Figure 1. Accuracy v.s. Computation Cost and Number of Observed Frames. Compared to baselines, our agent demonstrates superior performance with higher efficiency (observing fewer frames and incurring comparable or lower computational cost) on EgoSchema, using both GPT-4o and Qwen2-VL-72B. + +![](images/91ceaf3d095c177613ad02b2b753cad522a8335adc05cf71027f60062516008a.jpg) + +troduce 3D operations [8, 57]. Recent approach [32, 75] leverages the long-sequence reasoning capabilities of Large Language Models (LLMs) [1, 64] for long-form video question answering, often by converting videos into densely and uniformly sampled frames [33, 38, 55]. However, this approach is usually inefficient due to the inherent imbalance and varying information density across video segments [61, 78]. Some segments are rich in critical details necessary for answering questions, while others contain only trivial information [66]. Uniform sampling [49, 60], which might focus on redundant frames that increase the computational burden and may overlook frames with essential information, often fails to efficiently utilize the video's dynamic structure for question answering [17, 62]. + +To address this limitation, recent approaches [29, 34, 36, 68] employ complex architectures [63, 85] or specialized preprocessing techniques [44, 65] to improve the efficiency + +of long video content processing. These methods [62, 66] often convert videos into text by captioning densely sampled frames or incorporating memory modules [5, 17, 39] to store essential information. However, these approaches still incur significant computational costs. For example, captioning an hour-long video with Qwen2-VL-72B at just 1 fps requires over 3 hours on a single H100. Additionally, these methods [12] identify key frames based on image similarity rather than a holistic understanding of the video content, which can result in missing relevant video features. This necessitates the need for an efficient and scalable solution for long-form video understanding. + +In this work, we introduce VCA, a curiosity-driven agent inspired by how humans tackle the long video understanding task. Given a question, human is usually intrinsically motivated to identify relevant information in the video without external knowledge. In practise, human tends to build a global understanding of the video, then curiously zoom in to explore specific sections in greater detail. Motivated by this process, we propose a novel video agent framework, dubbed as "VCA", to emulate such behavior. VCA treats the video as an environment to explore, actively navigating its contents and identifying segments that merit deeper analysis. VCA incorporates three essential techniques: (1) Tree-search Exploration: Rather than selecting specific frames [62], VCA uses a tree-search structure to adaptively explore the video segments in a structured way. (2) Reward Model: VCA leverages an intrinsic reward scoring mechanism to guide exploration toward the segments most relevant to the query. (3) Memory Management: VCA employs a fixed memory buffer to store key information and discard irrelevant frames, effectively reducing computational costs and avoiding memory overflow. The key difference between VCA and prior works is outlined in Fig. 2. Extensive experiments and superior performance across multiple benchmarks demonstrate that our proposed VCA framework significantly improves an agent's effectiveness and efficiency in reasoning over extended video content. More concretely, our contributions are threefold, + +- VCA is a long video understanding agent that can self-exlore the most informative video segments based on its intrinsic reward, without relying on any auxiliary tools. +- VCA is easy to implement, training-free, and is flexible to integrate with any vision-language model. +- As highlighted in Fig. 1, VCA outperforms existing baselines while observing fewer frames and incurring comparable or lower computational cost, demonstrating both effectiveness and efficiency. + +![](images/86d97b7ae7cef89e8e72f2b8f1a26808351cdd81b23ff0ab9e2edc6c90c89ae7.jpg) +Figure 2. Comparison of VCA with prior works. Difference-1: When answering a query, VCA uniformly samples (a few) frames, whereas previous methods (e.g., VideoTree) require captioning each frame for subsequent image retrieval, which can be computationally expensive for long videos. Difference-2: previous methods rely on image similarity to locate key frames, potentially losing important video features. In contrast, VCA uses segment-level exploration to identify the most informative segments. + +# 2. Related Work + +# 2.1. Long-context Multi-modal Agents + +Large multi-modal models (LMMs) [2, 6, 9, 35, 54] have demonstrated promising performance in addressing complex multimodal tasks [20, 45, 49, 81]. However, these LMMs typically lack self-directed exploration capabilities required for handling more tricky tasks [24, 74]. To bridge this gap, multimodal agents are designed to enhance LMMs with the ability to autonomously make decisions and execute tasks within specific environments [15, 73], such as interpreting visual cues and text prompts to make real-time decisions within websites [28, 76]. + +A key feature of multimodal agents is their long-context capability [84], i.e., the ability to process and utilize information across extended periods to make informed decisions [26]. An example of this is vision-language model (VLM) agents for video understanding tasks. Despite their advancements, even sophisticated VLM agents struggle to extract crucial information and perform effective reasoning over lengthy inputs [62, 79]. Therefore, efficiently extracting essential information and making informed decisions remains a primary challenge. This work aims to address this challenge by enabling agents to self-exlore the environment and self-gather the most relevant information to efficiently solve video-understanding tasks. + +# 2.2. Long Video Understanding + +Early deep learning models for video understanding primarily followed two main approaches [52]. One uses two-stream networks [18, 19, 42, 58] to integrate spatial and motion information, while the second focuses on 3D CNNs [8, 57, 72] for spatial-temporal feature extraction. Vision Transformers [14] then introduce self-attention mechanisms + +![](images/36be31cb5baa1db2c03e215ae3ccb61924a453519badc0eb3c88049bd76108e3.jpg) +Figure 3. Pipeline of VCA Framework. First, the reward model predicts the relevance scores of each segment. Next, the memory buffer is updated by adding new frames and discarding irrelevant ones. Finally, emulating human reasoning, the agent actively selects segments to explore until the correct answer is found. + +to handle temporal modeling more effectively [3, 7, 16]. Methods with self-supervised pre-training [50, 56, 86] exhibit transferability across different tasks and achieve more robust video understanding performance. + +Multimodal LLM [22, 59, 83] shows strong capabilities in video understanding tasks, but efficiently processing information within a prolonged video environment is the key challenge. Recent approaches often use frame sampling [32, 46, 63, 75, 77] and feature extraction [27, 49, 51, 70] to reduce computational costs. Frame sampling selects a subset of frames, which risks missing essential context in lengthy videos, where transitions and subtle cues may be easily neglected. Additionally, processing entire video sequences as tokens presents significant memory and computational challenges [11, 43, 64, 80]. + +Recent work has explored auxiliary techniques and modules to improve efficiency, such as pre-extracted content like captions [44, 62], multimodal summaries [40, 65], and combinations of various auxiliary tools [17, 30]. However, these tools introduce additional computational and memory costs. Moreover, some approaches still require traversing the entire video content [17]. Notably, VideoTree [66] requires captioning every frame in the video, making it prohibitively expensive. In contrast, our VCA can solve the task using a single VLM, without relying on any supplementary tools. + +# 3. Method + +Inspired by human reasoning, we introduce a self-exploration agent that adaptively navigates video environments. Unlike approaches that depend on dense frame sampling or captioning [62], which require VLMs to process numerous frames, our agent analyzes videos iteratively within a limited context window in a process we call self-exploration. This approach allows essential information to be extracted from long videos in a coarse-to-fine manner. + +Specifically, we propose a segment-based tree-search exploration structure. Given a video $v$ and a user query $q$ , the agent $\pi$ begins by randomly sampling $N$ frames $\{f_v^i\}_{i=1}^N$ to obtain an initial overview of the content in each video segment $v_i \in \{v_i\}_{i=1}^{N+1}$ . The agent then decides whether additional exploration is needed to answer the query. If further investigation is necessary, the agent selects a specific segment $s^* \in \{v_i\}_{i=1}^{N+1}$ for deeper exploration until sufficient information is gathered to respond to the query. + +Nevertheless, a one-hour movie typically contains approximately 86,400 frames, posing a challenge for current VLM-based agents to efficiently identify the most relevant segment [17, 62, 66]. Moreover, even with a small $N$ , conducting multiple rounds of exploration accumulates a substantial number of frames, leading to high computational costs. To address these issues, we introduce a reward model $R$ to guide the agent's exploration and propose a memory + +buffer $M$ to reduce storage and inference costs. + +In this section, we first explain how the reward model assists the agent throughout the exploration process (Sec. 3.1). Then we introduce our memory management strategy in Sec. 3.2. Finally, we explain how VCA adaptively explores the video using a segment-based tree structure (Sec. 3.3). + +# 3.1. Reward Model + +With the segment-based tree-search exploration, the agent is required to select the most informative video segment for exploration at each step. However, we observe that current VLM-based agents struggle to independently identify optimal exploration paths (see Sec. 4.3). To this end, inspired by the selective attention mechanism in human cognition [13], we incorporate a reward model $R$ to guide the agent's selection process, i.e., the agent will gain a higher reward if it plans to take the action to explore more relevant segment. + +Specifically, during the exploration of a current selected segment $s^*$ , the reward model observes a set of frames $\{f_{s^*}^i\}_{i=1}^N$ sampled from $s^*$ . These frames $\{f_{s^*}^i\}_{i=1}^N$ then form finer sub-segments $\{s_i^* = [l_{s^*}^i, l_{s^*}^{i+1}]\}_{i=1}^{N+1}$ , where $l_{s^*}^i$ and $l_{s^*}^{i+1}$ represent the indices of the beginning and end frame of the sub-segment, respectively. The reward model aims to generate a reward score for each sub-segment $\{s_i^*\}_{i=1}^{N+1}$ based on its relevance to the user query $q$ . We utilize chain-of-thoughts [67] technique to let the reward model provide a verbal explanation of each segment and then generate a relevant score, i.e., $\{t_{s_i^*}; r_{s_i^*}\}_{i=1}^{N+1}$ . Additionally, to maintain consistency across different exploration steps, we incorporate the reward history $H_r$ from previous steps alongside the new segments during evaluation. A simplified prompt template for the reward model is presented here, with detailed prompts in Appendix ?? + +# Reward Model Prompt (Simplified) + +Given: + +$H_{r}$ , reward score history in previous steps. + +$\left\{f_{s^*}^i\right\}_{i = 1}^N$ newly sampled frames from current segment $s^{*}$ + +Please first explain each sub-segment $\{s_i^*\}_{i = 1}^{N + 1}$ and then assign a relevance score of each sub-segment to the user query $q$ . + +In practice, rather than training a separate specialized model, we utilize a same VLM-based agent to function as the reward model. While the reward model $R$ and exploration agent $\pi$ serve distinct purposes, both can be implemented using the same model. In this work, we employ a single VLM-based agent for both roles across different settings, providing a neat and efficient agent framework that is guided by its self-generated intrinsic reward and without + +relying on auxiliary models or tools. + +# 3.2. Memory Management + +As exploration proceeds over multiple rounds, the accumulation of frames can lead to significant computational costs for long videos, even when $N$ is small. However, as illustrated in Fig. 2, they tend to gradually disregard irrelevant frames, especially from earlier steps, retaining only the most crucial information in their working memory for later analysis. Inspired by this, we introduce a fixed-size memory buffer $M$ that stores only the most relevant frames throughout exploration. + +Specifically, when the agent $\pi$ explores a video segment $s^*$ , the associated frames $\{f_{s^*}^i\}_{i=1}^N$ are added to the memory buffer. If the total number of frames exceeds the buffer limit, irrelevant frames are removed based on relevance scores $r_{s_i^*}$ defined in Sec. 3.1. We discard the frames with the lowest scores and only retain the most relevant ones. While these less relevant frames are removed, their visual information is extracted and encoded as text descriptions $\{t_{s_i^*}\}_{i=1}^N$ within the exploration history $H_r$ , allowing the reward model to utilize. This memory management strategy enables the agent to conduct reasoning within a fixed resource limit, regardless of the exploration trajectory's length, thereby preventing excessive memory usage. + +# 3.3. Tree-Search Exploration + +Inspired by humans adaptively focusing on different video segments during exploration rather than targeting specific frames directly, our proposed agent VCA treats the video as an unexplored environment, where each segment represents a node on an exploration tree. Given the entire video as the root node, the agent's goal is to efficiently identify an optimal short path leading to a leaf node (segment) containing the most informative content needed to answer the query. + +At each step, the agent $\pi$ receives both textual and visual guidance: a set of candidate segments $S$ , which includes a collection of segments with associated reward scores, and a memory buffer $M$ containing multiple frames. If the agent has gathered sufficient information to answer the user's query $q$ , it directly generates a response $a$ . Otherwise, it continues exploring by selecting a specific segment $s^* \in S$ to expand the tree based on reward scores. + +However, due to the limited capacity of current foundational models, these reward scores are not always accurate (see further analysis in Sec. 5.3). Selecting segments based on reward scores in a purely greedy manner can potentially lead to a suboptimal solution. Instead, we use the reward scores as guidance for agent $\pi$ , allowing it to make the final segment selection independently. In this case, the agent is not forced to select the segment with the highest reward, balancing the exploitation and exploration. More importantly, the candidate set $S$ includes both unexplored + +coarse segments from prior steps and fine-grained segments from the current step, enabling the agent to backtrack and select earlier segments if it detects that the current exploration path may be suboptimal. Below are the simplified prompts. The full prompt is provided in Appendix ?? + +# Exploration Agent Prompt (Simplified) + +Given: + +Candidate segments set: $S$ + +Frames in Memory buffer: $M$ + +Please choose one segment to explore based on the query $q$ . Output answer if you have gathered sufficient information. + +The overall framework of our proposed method is summarized in Algo. 1. Leveraging a tree-search exploration technique boosted with a reward model and a memory buffer, our self-exploration agent adaptively and efficiently explores the video like a human. Empirical results in Sec. 4.3 further validate the effectiveness of our tree-search exploration over naive frame selection methods. + +# 4. Experiments + +# 4.1. Experimental Settings + +Benchmark. We mainly evaluate VCA on two long-term video question-answering benchmarks. EgoSchema [41] is an egocentric video dataset sourced from Ego4D [21], with a duration of 180 seconds. Here we use the validation set of 500 QA pairs. LVBench [61] covers about 1,500 samples for evaluation of reasoning abilities from high-quality videos over 30 minutes in length. We also conduct experiments on other relevant benchmark datasets in Appendix ?? + +Baselines. We compare with most relevant baselines to our work, including state-of-the-art open-source Video-LLMs [10, 23, 37, 47, 59, 82], the close-source model GPT-4o [25], and agent-based systems [44, 62, 66]. For video models, we report the results obtained with uniform sampling frames, while for the agent-based systems, we follow their official implementation with GPT-4o [25], consistent with VCA's setup to ensure a fair comparison. + +Implementation Details. In this work, we use the same model for the exploration agent and the reward model. Unless otherwise stated, all experiments were performed using the August 2024 version of `gpt-4` as the base model, leveraging the OpenAI LLM API service1 with a temperature of 0.5. We set memory buffer sizes set to 8 and 16 frames for EgoSchema and LVBench, respectively. Further implementation details are provided in Appendix ?? + +Algorithm 1 VCA: Video Curious Agent +Ensure: Answer $a$ +```txt +Require: Video environment $v$ , question $q$ , agent $\pi$ , reward model $R$ , memory buffer $M$ , candidate segments set $S$ , reward score history $H_{r}$ , sampling frame number $N$ . +``` + +1: Initialize selected segment $s^* \gets v$ +2: Initialize candidate segments set $S \gets \emptyset$ +3: Initialize memory buffer $M\gets \emptyset$ +4: Initialize reward score history $H_{r}\gets \emptyset$ +5: continue $\leftarrow$ true +6: while continue do +7: # Sampling frames & form segments +8: $\left\{f_{s^*}^*\right\}_{i=1}^{N}, \left\{s_i^*\right\}_{i=1}^{N+1} \gets \text{UniformSample}(s^*)$ +9: # Scoring each segment +10: $\left\{t_{s_i^*};r_{s_i^*}\right\}_{i = 1}^{N + 1}\gets R(\{s_i^*\}_{i = 1}^{N + 1};q,\{f_{s_*}^i\}_{i = 1}^N,H_r)$ +11: # Update candidate segment set +12: $S\gets S\cup \{s_i^*;r_{s_i^*}\}_{i = 1}^{N + 1}$ +13: # Update history with thoughts +14: $H_{r}\gets H_{r}\cup \left\{t_{s_{i}^{*}};r_{s_{i}^{*}}\right\}_{i = 1}^{N + 1}$ +15: # Update memory with frames +16: $M\gets \mathrm{UpdateMemory}(M,\{f_{s^{*}}^{i}\}_{i = 1}^{N};S)$ +17: # Output answer or select segment +18: $a,s\gets \pi (q,S;M)$ +19: if $a = \emptyset$ then +20: # Continue exploration +21: $s^* \gets s$ +22: else +23: continue $\leftarrow$ false +24: end if +25: end while +26: return $a$ + +# 4.2. Experimental Results + +The experimental results on LVBench and EgoSchema are presented in Tab. 2 and Tab. 1, respectively. By adaptively focusing on critical segments rather than scanning the entire video, our agent not only achieves stronger performance but also incurs less computational cost. With less than $30\%$ of the frames, VCA achieves substantial improvements over directly feeding uniformly sampled frames into GPT-4o. Specifically, we observe performance gains of $3.2\%$ on the EgoSchema subset and $6.6\%$ on LVBench. Significant increases in the Event Understanding and Temporal Grounding domains further underscore our effectiveness in identifying crucial segments within long videos. + +Furthermore, we compare our framework with recently proposed SOTA video agent systems, VideoAgent [62] and VideoTree [66]. For a fair comparison, we re-implement both frameworks using GPT-4o for all reasoning components, consistent with our setup. However, captioning all + +
MethodFramesEREUKIRTGReaSumAvg.
TimeChat* [47]>9621.921.725.922.725.024.122.3
LLaVA-NeXT* [83]3230.131.234.131.435.027.632.2
InternVL2-40B* [10]1637.439.743.431.442.541.439.8
GLM4V-Plus* [23]2039.935.834.837.740.032.838.3
GPT-4o* [25]6435.930.835.528.333.534.534.7
VideoAgent† [62]Avg. 25.528.030.328.029.328.036.429.3
VideoTree† [66]Avg. 103.230.325.126.527.731.925.528.8
VCA (ours)Avg. 20.043.740.737.838.046.227.341.3
+ +Table 1. Experimental Results on LVBench. Results of methods marked with * are sourced from LVBench [61]. Methods marked with † indicate implementations where GPT-4o serves as the main component. Our agent's memory buffer size is set to 16 frames. We report the average number of observed frames for agent-based methods, with the best performance highlighted in bold. + +
MethodFramesSubset
GPT-4o [25]869.0
3270.4
\(VideoAgent^\dagger\) [62]Avg. 8.163.2
Avg. 10.760.8
\(VideoTree^\dagger\) [66]Avg. 69.567.0
\(LVNet^\dagger\) [44]1268.2
VCA (ours)Avg. 7.273.6
+ +frames with GPT-4o is prohibitively expensive. Hence, following the baseline implementations, we use open-source VLMs for captioning. A comparison using the same model for all components is provided in Sec. 5.1. As shown in Tab. 1, VCA outperforms all baselines over $5\%$ on EgoSchema, while requiring processing fewer frames. These results suggest the limitations of dense captioning for video preprocessing tasks, which can incur more computational costs while introducing more redundant information. In contrast, our proposed memory management strategy effectively improves efficiency by only retaining the most crucial information for reasoning. Similar insights can be drawn from results on LVBench in Tab. 2. Additional results are provided in Appendix ?? + +Moreover, we observe that VCA explores approximately three times more frames on LVBench than on EgoSchema. Recall that EgoSchema videos have an average duration of 180 seconds, while videos in LVBench extend to at least 30 minutes. This phenomenon suggests that our proposed tree-search exploration technique actively explores the video to locate the most relevant information and adaptively in- + +Table 2. Experimental Results on EgoSchema. Methods marked with $\dagger$ indicate implementations where GPT-4o is the main component. The results of LVNet are from [44]. Our agent's memory buffer size is set to 8 frames. We report the average number of observed frames, with the best performance highlighted in bold. + +
MethodEgoSchemaLVBench
FramesAcc.FramesAcc.
VCA (ours)7.273.620.041.3
- w/o Reward7.572.622.239.5
- w/o Tree Search10.269.839.936.2
+ +Table 3. Ablation Study on EgoSchema and LVBench. Both the tree-search exploration and reward model contribute significantly to improvements in performance and efficiency. Detailed ablation study results on LVBench are included in Appendix ?? + +
Buffer Size81632
Acc.38.941.343.2
+ +Table 4. Ablation Study of Buffer Size on LVBench. Increasing the buffer size allows the model to observe more frames during reasoning, leading to improved performance. + +creases the exploration budget based on task complexity. For example, VCA exhibits both exploitation and exploration: when the agent determines that the current exploration path is suboptimal, it backtracks to previous segments to explore a new path (see Sec. 5.2). These empirical observations further validate the robustness and scalability of our approach, especially in diverse long-video contexts. + +# 4.3. Ablation Study + +In this section, we conduct an ablation study to evaluate the components of our agent. Results across both benchmarks are presented in Tab. 3. As illustrated in the table, performance consistently declines when the reward model is removed, highlighting that the agent's inherent reasoning ability alone is insufficient for accurately identifying the most informative segments. This finding emphasizes the effectiveness of incorporating an intrinsic reward to guide the agent's segment selection process. A detailed analysis of our reward model's performance is provided in Sec. 5.3. + +Additionally, we observe a significant performance drop + +Why does the mice run away? +(A) Because the black cat sheriff calls the police +(C) Because their boss is injured + +(B) Because they meet more powerful mice +(D) Because the owner of the house is back + +![](images/494c2e39d33787b489c52c42d9cf3f62f97d802bf3cb5b76c77ffd4cb93b8d38.jpg) + +# Reward Model + +Segment1: Thought: The segment shows a black cat with a gun, possibly indicating a threat. This could suggest that the mice might run away due to the presence of the cat, which aligns with option (A). Score: 0.7 +Segment4: Thought: The segment features the black cat with a communication device, possibly calling for help or backup. This aligns with option (A), suggesting the mice might run because the cat is calling the police. Score: 0.8 + +![](images/cab27bdb7ff0b0177e4565f738e98f8d8b8a7fec79b098210f2663bdfcdbd937.jpg) +Agent Choice: (segment: 4) + +![](images/db534f7962bd217c6a8212442a281185644644533499078be1979c927fe8315b.jpg) + +![](images/9709989d9606dbfa36835e193981ef9fd1b3c4452fb7a2442047d274c61b2bbb.jpg) + +# Reward Model + +Segment4.1: Thought: The segment shows a close-up of a document with red text and exclamation marks, indicating urgency or importance. This could imply a critical message or alert being communicated. Score: 0.7 + +… + +![](images/0be7582fe7139255d5979ad1a0a4b2f8416fde86fabbd6434c0c69aee2357f20.jpg) + +# Reward Model + +Segment1.2: Thought: The segment features mice characters in mid-air, possibly fleeing or being thrown. This suggests a direct reaction to a threat, supporting the idea that they are running away from the cat. Score: 0.8 + +![](images/e82b2c43037a928c41960aef4203fd12a1f1ce59e083cac00c6ba6a3ec87c9ad.jpg) +Figure 4. Exploration Trajectory Example. The agent generally tends to select video segments with higher reward scores for exploration, yet it can adaptively redirect to earlier segments, effectively balancing exploitation and exploration. + +![](images/5386d87e480fee48366b58408db92687c9c944cdedb4c56e93e21002e2bd9557.jpg) + +Agent: Answer: A + +when the tree-search exploration technique is ablated, i.e., the agent directly selects frame indices without the segment-based tree structure. Interestingly, this modification also results in a substantial increase in the average number of frames acquired, especially on LVBench. This indicates that the agent struggles to locate the most crucial information and consequently produces a longer but redundant trajectory. These ablation study results validate the effectiveness of the tree-search exploration framework assisted by the reward model. Additionally, this underscores VCA's potential to achieve greater efficiency and improved performance with more advanced VLMs and reward models. + +Moreover, as shown in Tab. 4, increasing the memory buffer size yields consistently improved performance for VCA, demonstrating that our framework benefits from a broader reasoning window while maintaining effectiveness with fewer frames. Due to limited computation resources, we only tested up to 32 frames; however, the upward trend indicates the potential for gains with additional frames. + +# 5. Analysis + +# 5.1. Can VCA generalize to Open-source Models? + +While VCA demonstrates strong performance using GPT4o as the base model, we also analyze whether these positive results can be extrapolated to open-source models. We implement the VCA framework with Qwen2-VL-72B-Instruct-GPTQ-Int4 [59] as both the exploration agent and reward model, as well as the captioner for baselines. The results in Tab. 5 reveal that, compared to directly feeding uniformly sampled frames into Qwen2-VL, our framework achieves higher accuracy on the EgoSchema benchmark despite using only about $60\%$ of the frames. In contrast, VideoAgent performs significantly worse. Furthermore, our approach outperforms VideoTree with far fewer frames. Notably, VideoTree employs clustering to build a caption similarity-based tree structure for frame selection, requiring caption generation for each frame. This process is computationally expensive, demanding over 60 GPU hours + +
MethodFramesSubset
Qwen2-VL870.2
3274.2
VideoAgentAvg. 8.665.2
VideoTreeAvg. 85.471.0
VCA (ours)Avg. 16.975.2
- w/ Image RewardAvg. 20.974.0
+ +Table 5. Performance of Using Open-source Model on EgoSchema. Our approach achieves higher accuracy with fewer observed frames, showing consistency improvements over baseline methods. Memory buffer size is set to 16 frames. + +on a single H100, as shown in Fig. 1. These results further underscore the robustness of VCA. + +# 5.2. What's the Exploration Behavior of VCA? + +Here we present a real exploration trajectory from LVBench to illustrate our agent's decision-making process. As shown in Fig. 4, given a user query, the reward model initially evaluates each segment's relevance based on uniformly sampled frames, and the agent selects the fourth segment, which has the highest relevance score. However, upon further investigation of segment 4, both the reward model and the agent find it relatively irrelevant to the query. Rather than continuing with it, the agent redirects its exploration to segment 1 in the last round. After selecting segment 1, the agent chooses to investigate the third segment, despite the second segment having the highest reward score. This behavior reflects VCA's ability to adaptively adjust its exploration trajectory based on visual context rather than purely following the reward score, resulting in a more robust system. + +We also examine the limitations of VCA by analyzing its failure cases, which fall into the following scenarios: The most common failure occurs when the agent overlooks subtle clues, resulting in an inaccurate response to the user query even after prolonged exploration, particularly for queries requiring specific visual details. Other failure cases include instances where the agent is misled by inaccurate reward scores or fails to provide the correct answer despite observing key frames. Detailed cases and further discussion are provided in the Appendix ?? + +In addition, unlike previous methods, VCA employs a segment-level exploration strategy instead of direct image similarity to select key frames. To further investigate this difference, we evaluate the performance of selecting the segment whose start and end images have the highest visual relevance. As shown in Tab. 5, relying solely on image relevance instead of inferring segment semantics results in a performance drop on EgoSchema, confirming the effectiveness of our segment-level exploration. + +![](images/fca8834cee31104ddaa6b77c3ab66dbfffdf08de055321278fd4a45f2a317be3.jpg) +Figure 5. Segment Selection Accuracy on LVBench. The y-axis shows segment distance from ground truth; x-axis shows count ratio. Red curves are fitted normal distributions. The reward-guided agent selects segments closer to ground truth than the reward model. + +# 5.3. How Reliable is the Reward Model? + +To further assess the capability of our reward model, we visualize the matching accuracy between segments with the highest reward scores and ground truth time references from LVBench. The results are shown at the bottom of Fig. 5. We observe that segment distances follow an approximately normal distribution with a mean of 1.28, indicating that the segment with the highest reward score often closely aligns with the ground truth segment. This validates the reward model's effectiveness in providing meaningful guidance. + +Additionally, as discussed in Sec 3.3, we encourage the agent to make independent decisions when selecting segments rather than blindly following reward scores in a purely greedy manner. The benefit of this design choice is demonstrated at the top of Fig. 5, where we present the matching accuracy of the agent's choices. Compared to the greedy approach, the agent's selections result in a sharper distribution of segment distances, with an average distance gap of $-0.60$ , indicating closer alignment with the ground truth. In summary, our reward model assigns reliable relevance scores to segments, while the agent further refines these choices to achieve even more accurate selections. + +# 6. Conclusion + +In this work, we propose a curiosity-driven video agent framework for long video understanding tasks with self-exploration capability. VCA achieves state-of-the-art performance and efficiency without relying on auxiliary tools. Appendix ?? shows that improved reward guidance could further boost results, highlighting the potential of synthetic data and specialized reward models for future work. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2 +[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 3 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1 +[5] Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J Henaff. Memory consolidation enables long-context video understanding. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for ai agents, 2023. 2 +[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding?, 2021. 3 +[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset, 2018. 1, 2 +[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2 +[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5, 6 +[11] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3 +[12] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. arXiv preprint arXiv:2311.01233, 2023. 2 +[13] Robert Desimone, John Duncan, et al. Neural mechanisms of selective visual attention. Annual review of neuroscience, 18(1):193-222, 1995. 4 + +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2 +[15] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, Katsushi Ikeuchi, Hoi Vo, Li Fei-Fei, and Jianfeng Gao. Agent ai: Surveying the horizons of multimodal interaction, 2024. 2 +[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers, 2021. 3 +[17] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2025. 1, 2, 3 +[18] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition, 2016. 1, 2 +[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2 +[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2 +[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5 +[22] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, Lei Zhao, Zhuoyi Yang, Xiaotao Gu, Xiaohan Zhang, Guanyu Feng, Da Yin, Zihan Wang, Ji Qi, Xixuan Song, Peng Zhang, Debing Liu, Bin Xu, Juanzi Li, Yuxiao Dong, and Jie Tang. CogvIm2: Visual language models for image and video understanding, 2024. 3 +[23] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 5, 6 +[24] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 2 +[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli + +hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5, 6 +[26] Lawrence Jang, Yinheng Li, Charles Ding, Justin Lin, Paul Pu Liang, Dan Zhao, Rogerio Bonatti, and Kazuhito Koishida. Videowebarena: Evaluating long context multimodal agents with video understanding web tasks, 2024. 2 +[27] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding, 2023. 3 +[28] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2 +[29] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In European Conference on Computer Vision, pages 271-288. Springer, 2025. 1 +[30] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning, 2024. 3 +[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1 +[32] Yunxin Li, Xinyu Chen, Baotain Hu, and Min Zhang. Llms meet long video: Advancing long video comprehension with an interactive visual adapter in llms, 2024. 1, 3 +[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1 +[34] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1 +[35] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2 +[36] Ming C Lin and Shan Yang. Vila: Efficient video-language alignment for video question answering. 1 +[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5 +[38] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1 +[39] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846, 2024. 2 +[40] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding, 2024. 3 + +[41] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 5 +[42] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification, 2015. 2 +[43] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14386-14397, 2024. 3 +[44] Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya, Wonjeong Ryoo, Donghyun Kim, and Michael S Ryoo. Too many frames, not all useful: Efficient strategies for long-form video qa. arXiv preprint arXiv:2406.09396, 2024. 1, 3, 5, 6 +[45] Ruchit Rawal, Khalid Saifullah, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. In Synthetic Data for Computer Vision Workshop @ CVPR 2024. 2 +[46] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. arXiv preprint arXiv:2312.02051, 2023. 3 +[47] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 5, 6 +[48] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 1 +[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 2, 3 +[50] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning, 2019. 3 +[51] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Advances in neural information processing systems, 35:38032-38045, 2022. 3 +[52] Yunlong Tang, Jing Bi, Sitting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, Ali Vosoughi, Chao Huang, Zeliang Zhang, Pinxin Liu, Mingqian Feng, Feng Zheng, Jianguo Zhang, + +Ping Luo, Jiebo Luo, and Chenliang Xu. Video understanding with large language models: A survey, 2024. 2 +[53] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 1 +[54] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2 +[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1 +[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training, 2022. 3 +[57] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks, 2015. 1, 2 +[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition, 2016. 2 +[59] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5, 7 +[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1 +[61] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Shiyu Huang, Bin Xu, Yuxiao Dong, Ming Ding, and Jie Tang. Lvbench: An extreme long video understanding benchmark, 2024. 1, 5, 6 +[62] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 6 +[63] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 1, 3 +[64] Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recurrent memory bridges. arXiv preprint arXiv:2409.01071, 2024. 1, 3 +[65] Ying WANG, Yanlai YANG, and Mengye REN. Lifelong-memory: Leveraging llms for answering queries in long + +form egocentric videos. arXiv preprint arXiv:2312.05269, 2024. 1, 3 +[66] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos, 2024. 1, 2, 3, 5, 6 +[67] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4 +[68] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision, pages 453-470. Springer, 2025. 1 +[69] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 1 +[70] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 3 +[71] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. CoRR, 2024. 1 +[72] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification, 2018. 2 +[73] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. In ICLR 2024 Workshop on Large Language Model (LLM) Agents. 2 +[74] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 2 +[75] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Retrieval-based video language model for efficient long video question answering, 2023. 1, 3 +[76] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2 +[77] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 2023. 3 +[78] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 1 +[79] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2 + +[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 3 +[81] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023. 2 +[82] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5 +[83] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3, 6 +[84] Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang. Longagent: Scaling language models to 128k context through multi-agent collaboration. arXiv preprint arXiv:2402.11550, 2024. 2 +[85] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In *Forty-first International Conference on Machine Learning*. 1 +[86] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations, 2020. 3 \ No newline at end of file diff --git a/2025/VCA_ Video Curious Agent for Long Video Understanding/images.zip b/2025/VCA_ Video Curious Agent for Long Video Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e07566d6f0e1df59df4f7ee30acb687e7ef30018 --- /dev/null +++ b/2025/VCA_ Video Curious Agent for Long Video Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f79d775bae197443e1f123dc4992a2b2bf38d4f1358bd21c18627b16e731cc65 +size 439132 diff --git a/2025/VCA_ Video Curious Agent for Long Video Understanding/layout.json b/2025/VCA_ Video Curious Agent for Long Video Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c11f8318e87a0733e61bcef0b51974ac229b27e4 --- /dev/null +++ b/2025/VCA_ Video Curious Agent for Long Video Understanding/layout.json @@ -0,0 +1,10608 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 487, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 487, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 487, + 121 + ], + "type": "text", + "content": "VCA: Video Curious Agent for Long Video Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 142, + 291, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 142, + 291, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 291, + 185 + ], + "type": "text", + "content": "Zeyuan Yang† \nUniversity of Massachusetts, Amherst \nyangzeyuan@umass.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 191, + 291, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 191, + 291, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 291, + 232 + ], + "type": "text", + "content": "Xueyang Yu \nUniversity of Massachusetts, Amherst \nxueyangyu@umass.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 319, + 142, + 504, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 142, + 504, + 185 + ], + "spans": [ + { + "bbox": [ + 319, + 142, + 504, + 185 + ], + "type": "text", + "content": "Delin Chen† \nUniversity of Massachusetts, Amherst delinchen@umass.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 319, + 191, + 504, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 191, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 319, + 191, + 504, + 232 + ], + "type": "text", + "content": "Maohao Shen \nMassachusetts Institute of Technology \nmaohao@mit.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 212, + 239, + 398, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 239, + 398, + 280 + ], + "spans": [ + { + "bbox": [ + 212, + 239, + 398, + 280 + ], + "type": "text", + "content": "Chuang Gan \nUniversity of Massachusetts, Amherst \nganchuang1990@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 308, + 200, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 308, + 200, + 320 + ], + "spans": [ + { + "bbox": [ + 152, + 308, + 200, + 320 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 334, + 297, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 297, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 297, + 538 + ], + "type": "text", + "content": "Long video understanding poses unique challenges due to their temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as \"VCA\". Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences. Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 574, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 670 + ], + "type": "text", + "content": "There is a growing demand for developing long-form video understanding techniques [48, 69] capable of extracting valuable information from extended visual narratives, such as surveillance footage, documentaries, and instructional videos. Analyzing these videos [53, 71], which can range from minutes to hours, requires models that can process multimodal data and perform reasoning over extremely long sequences [4, 31, 60]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 694 + ], + "type": "text", + "content": "Previous works commonly utilize two-stream networks to capture spatial and temporal information [18, 19] or in" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 308, + 435, + 439 + ], + "blocks": [ + { + "bbox": [ + 328, + 308, + 435, + 439 + ], + "lines": [ + { + "bbox": [ + 328, + 308, + 435, + 439 + ], + "spans": [ + { + "bbox": [ + 328, + 308, + 435, + 439 + ], + "type": "image", + "image_path": "b3a34f2280701b001980d124b23ff96bfa7878b2f2423e84306fc895c68a5841.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 444, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 499 + ], + "type": "text", + "content": "Figure 1. Accuracy v.s. Computation Cost and Number of Observed Frames. Compared to baselines, our agent demonstrates superior performance with higher efficiency (observing fewer frames and incurring comparable or lower computational cost) on EgoSchema, using both GPT-4o and Qwen2-VL-72B." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 436, + 308, + 544, + 439 + ], + "blocks": [ + { + "bbox": [ + 436, + 308, + 544, + 439 + ], + "lines": [ + { + "bbox": [ + 436, + 308, + 544, + 439 + ], + "spans": [ + { + "bbox": [ + 436, + 308, + 544, + 439 + ], + "type": "image", + "image_path": "91ceaf3d095c177613ad02b2b753cad522a8335adc05cf71027f60062516008a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 508, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 675 + ], + "type": "text", + "content": "troduce 3D operations [8, 57]. Recent approach [32, 75] leverages the long-sequence reasoning capabilities of Large Language Models (LLMs) [1, 64] for long-form video question answering, often by converting videos into densely and uniformly sampled frames [33, 38, 55]. However, this approach is usually inefficient due to the inherent imbalance and varying information density across video segments [61, 78]. Some segments are rich in critical details necessary for answering questions, while others contain only trivial information [66]. Uniform sampling [49, 60], which might focus on redundant frames that increase the computational burden and may overlook frames with essential information, often fails to efficiently utilize the video's dynamic structure for question answering [17, 62]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "To address this limitation, recent approaches [29, 34, 36, 68] employ complex architectures [63, 85] or specialized preprocessing techniques [44, 65] to improve the efficiency" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 134, + 712 + ], + "type": "text", + "content": "† Equal Contribution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20168" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "of long video content processing. These methods [62, 66] often convert videos into text by captioning densely sampled frames or incorporating memory modules [5, 17, 39] to store essential information. However, these approaches still incur significant computational costs. For example, captioning an hour-long video with Qwen2-VL-72B at just 1 fps requires over 3 hours on a single H100. Additionally, these methods [12] identify key frames based on image similarity rather than a holistic understanding of the video content, which can result in missing relevant video features. This necessitates the need for an efficient and scalable solution for long-form video understanding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 248, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 248, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 296, + 571 + ], + "type": "text", + "content": "In this work, we introduce VCA, a curiosity-driven agent inspired by how humans tackle the long video understanding task. Given a question, human is usually intrinsically motivated to identify relevant information in the video without external knowledge. In practise, human tends to build a global understanding of the video, then curiously zoom in to explore specific sections in greater detail. Motivated by this process, we propose a novel video agent framework, dubbed as \"VCA\", to emulate such behavior. VCA treats the video as an environment to explore, actively navigating its contents and identifying segments that merit deeper analysis. VCA incorporates three essential techniques: (1) Tree-search Exploration: Rather than selecting specific frames [62], VCA uses a tree-search structure to adaptively explore the video segments in a structured way. (2) Reward Model: VCA leverages an intrinsic reward scoring mechanism to guide exploration toward the segments most relevant to the query. (3) Memory Management: VCA employs a fixed memory buffer to store key information and discard irrelevant frames, effectively reducing computational costs and avoiding memory overflow. The key difference between VCA and prior works is outlined in Fig. 2. Extensive experiments and superior performance across multiple benchmarks demonstrate that our proposed VCA framework significantly improves an agent's effectiveness and efficiency in reasoning over extended video content. More concretely, our contributions are threefold," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 603, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 603, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 639 + ], + "type": "text", + "content": "- VCA is a long video understanding agent that can self-exlore the most informative video segments based on its intrinsic reward, without relying on any auxiliary tools." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 639, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 295, + 664 + ], + "type": "text", + "content": "- VCA is easy to implement, training-free, and is flexible to integrate with any vision-language model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": "- As highlighted in Fig. 1, VCA outperforms existing baselines while observing fewer frames and incurring comparable or lower computational cost, demonstrating both effectiveness and efficiency." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 331, + 73, + 539, + 205 + ], + "blocks": [ + { + "bbox": [ + 331, + 73, + 539, + 205 + ], + "lines": [ + { + "bbox": [ + 331, + 73, + 539, + 205 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 539, + 205 + ], + "type": "image", + "image_path": "86d97b7ae7cef89e8e72f2b8f1a26808351cdd81b23ff0ab9e2edc6c90c89ae7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. Comparison of VCA with prior works. Difference-1: When answering a query, VCA uniformly samples (a few) frames, whereas previous methods (e.g., VideoTree) require captioning each frame for subsequent image retrieval, which can be computationally expensive for long videos. Difference-2: previous methods rely on image similarity to locate key frames, potentially losing important video features. In contrast, VCA uses segment-level exploration to identify the most informative segments." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 303, + 401, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 303, + 401, + 315 + ], + "spans": [ + { + "bbox": [ + 314, + 303, + 401, + 315 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 323, + 496, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 496, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 496, + 335 + ], + "type": "text", + "content": "2.1. Long-context Multi-modal Agents" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "type": "text", + "content": "Large multi-modal models (LMMs) [2, 6, 9, 35, 54] have demonstrated promising performance in addressing complex multimodal tasks [20, 45, 49, 81]. However, these LMMs typically lack self-directed exploration capabilities required for handling more tricky tasks [24, 74]. To bridge this gap, multimodal agents are designed to enhance LMMs with the ability to autonomously make decisions and execute tasks within specific environments [15, 73], such as interpreting visual cues and text prompts to make real-time decisions within websites [28, 76]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "content": "A key feature of multimodal agents is their long-context capability [84], i.e., the ability to process and utilize information across extended periods to make informed decisions [26]. An example of this is vision-language model (VLM) agents for video understanding tasks. Despite their advancements, even sophisticated VLM agents struggle to extract crucial information and perform effective reasoning over lengthy inputs [62, 79]. Therefore, efficiently extracting essential information and making informed decisions remains a primary challenge. This work aims to address this challenge by enabling agents to self-exlore the environment and self-gather the most relevant information to efficiently solve video-understanding tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 465, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 465, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 465, + 636 + ], + "type": "text", + "content": "2.2. Long Video Understanding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Early deep learning models for video understanding primarily followed two main approaches [52]. One uses two-stream networks [18, 19, 42, 58] to integrate spatial and motion information, while the second focuses on 3D CNNs [8, 57, 72] for spatial-temporal feature extraction. Vision Transformers [14] then introduce self-attention mechanisms" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20169" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 121, + 64, + 492, + 345 + ], + "blocks": [ + { + "bbox": [ + 121, + 64, + 492, + 345 + ], + "lines": [ + { + "bbox": [ + 121, + 64, + 492, + 345 + ], + "spans": [ + { + "bbox": [ + 121, + 64, + 492, + 345 + ], + "type": "image", + "image_path": "36be31cb5baa1db2c03e215ae3ccb61924a453519badc0eb3c88049bd76108e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 351, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 54, + 351, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 555, + 384 + ], + "type": "text", + "content": "Figure 3. Pipeline of VCA Framework. First, the reward model predicts the relevance scores of each segment. Next, the memory buffer is updated by adding new frames and discarding irrelevant ones. Finally, emulating human reasoning, the agent actively selects segments to explore until the correct answer is found." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 385, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 434 + ], + "type": "text", + "content": "to handle temporal modeling more effectively [3, 7, 16]. Methods with self-supervised pre-training [50, 56, 86] exhibit transferability across different tasks and achieve more robust video understanding performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 448, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 581 + ], + "type": "text", + "content": "Multimodal LLM [22, 59, 83] shows strong capabilities in video understanding tasks, but efficiently processing information within a prolonged video environment is the key challenge. Recent approaches often use frame sampling [32, 46, 63, 75, 77] and feature extraction [27, 49, 51, 70] to reduce computational costs. Frame sampling selects a subset of frames, which risks missing essential context in lengthy videos, where transitions and subtle cues may be easily neglected. Additionally, processing entire video sequences as tokens presents significant memory and computational challenges [11, 43, 64, 80]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Recent work has explored auxiliary techniques and modules to improve efficiency, such as pre-extracted content like captions [44, 62], multimodal summaries [40, 65], and combinations of various auxiliary tools [17, 30]. However, these tools introduce additional computational and memory costs. Moreover, some approaches still require traversing the entire video content [17]. Notably, VideoTree [66] requires captioning every frame in the video, making it prohibitively expensive. In contrast, our VCA can solve the task using a single VLM, without relying on any supplementary tools." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 384, + 370, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 370, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 370, + 396 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 407, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 503 + ], + "type": "text", + "content": "Inspired by human reasoning, we introduce a self-exploration agent that adaptively navigates video environments. Unlike approaches that depend on dense frame sampling or captioning [62], which require VLMs to process numerous frames, our agent analyzes videos iteratively within a limited context window in a process we call self-exploration. This approach allows essential information to be extracted from long videos in a coarse-to-fine manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": "Specifically, we propose a segment-based tree-search exploration structure. Given a video " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " and a user query " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": ", the agent " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " begins by randomly sampling " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\{f_v^i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " to obtain an initial overview of the content in each video segment " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "v_i \\in \\{v_i\\}_{i=1}^{N+1}" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": ". The agent then decides whether additional exploration is needed to answer the query. If further investigation is necessary, the agent selects a specific segment " + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "inline_equation", + "content": "s^* \\in \\{v_i\\}_{i=1}^{N+1}" + }, + { + "bbox": [ + 313, + 506, + 556, + 616 + ], + "type": "text", + "content": " for deeper exploration until sufficient information is gathered to respond to the query." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Nevertheless, a one-hour movie typically contains approximately 86,400 frames, posing a challenge for current VLM-based agents to efficiently identify the most relevant segment [17, 62, 66]. Moreover, even with a small " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": ", conducting multiple rounds of exploration accumulates a substantial number of frames, leading to high computational costs. To address these issues, we introduce a reward model " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": " to guide the agent's exploration and propose a memory" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20170" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 247, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 247, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 247, + 83 + ], + "type": "text", + "content": "buffer " + }, + { + "bbox": [ + 55, + 72, + 247, + 83 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 72, + 247, + 83 + ], + "type": "text", + "content": " to reduce storage and inference costs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 145 + ], + "type": "text", + "content": "In this section, we first explain how the reward model assists the agent throughout the exploration process (Sec. 3.1). Then we introduce our memory management strategy in Sec. 3.2. Finally, we explain how VCA adaptively explores the video using a segment-based tree structure (Sec. 3.3)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 155, + 148, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 148, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 148, + 167 + ], + "type": "text", + "content": "3.1. Reward Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 173, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 281 + ], + "type": "text", + "content": "With the segment-based tree-search exploration, the agent is required to select the most informative video segment for exploration at each step. However, we observe that current VLM-based agents struggle to independently identify optimal exploration paths (see Sec. 4.3). To this end, inspired by the selective attention mechanism in human cognition [13], we incorporate a reward model " + }, + { + "bbox": [ + 55, + 173, + 296, + 281 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 173, + 296, + 281 + ], + "type": "text", + "content": " to guide the agent's selection process, i.e., the agent will gain a higher reward if it plans to take the action to explore more relevant segment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": "Specifically, during the exploration of a current selected segment " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "s^*" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": ", the reward model observes a set of frames " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\{f_{s^*}^i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "s^*" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": ". These frames " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\{f_{s^*}^i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": " then form finer sub-segments " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\{s_i^* = [l_{s^*}^i, l_{s^*}^{i+1}]\\}_{i=1}^{N+1}" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "l_{s^*}^i" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "l_{s^*}^{i+1}" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": " represent the indices of the beginning and end frame of the sub-segment, respectively. The reward model aims to generate a reward score for each sub-segment " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\{s_i^*\\}_{i=1}^{N+1}" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": " based on its relevance to the user query " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": ". We utilize chain-of-thoughts [67] technique to let the reward model provide a verbal explanation of each segment and then generate a relevant score, i.e., " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\{t_{s_i^*}; r_{s_i^*}\\}_{i=1}^{N+1}" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": ". Additionally, to maintain consistency across different exploration steps, we incorporate the reward history " + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "inline_equation", + "content": "H_r" + }, + { + "bbox": [ + 55, + 282, + 296, + 485 + ], + "type": "text", + "content": " from previous steps alongside the new segments during evaluation. A simplified prompt template for the reward model is presented here, with detailed prompts in Appendix ??" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 494, + 216, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 494, + 216, + 506 + ], + "spans": [ + { + "bbox": [ + 72, + 494, + 216, + 506 + ], + "type": "text", + "content": "Reward Model Prompt (Simplified)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 516, + 101, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 516, + 101, + 526 + ], + "spans": [ + { + "bbox": [ + 71, + 516, + 101, + 526 + ], + "type": "text", + "content": "Given:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 527, + 244, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 527, + 244, + 540 + ], + "spans": [ + { + "bbox": [ + 71, + 527, + 244, + 540 + ], + "type": "inline_equation", + "content": "H_{r}" + }, + { + "bbox": [ + 71, + 527, + 244, + 540 + ], + "type": "text", + "content": ", reward score history in previous steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 539, + 279, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 539, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 71, + 539, + 279, + 564 + ], + "type": "inline_equation", + "content": "\\left\\{f_{s^*}^i\\right\\}_{i = 1}^N" + }, + { + "bbox": [ + 71, + 539, + 279, + 564 + ], + "type": "text", + "content": " newly sampled frames from current segment " + }, + { + "bbox": [ + 71, + 539, + 279, + 564 + ], + "type": "inline_equation", + "content": "s^{*}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 563, + 280, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 563, + 280, + 602 + ], + "spans": [ + { + "bbox": [ + 70, + 563, + 280, + 602 + ], + "type": "text", + "content": "Please first explain each sub-segment " + }, + { + "bbox": [ + 70, + 563, + 280, + 602 + ], + "type": "inline_equation", + "content": "\\{s_i^*\\}_{i = 1}^{N + 1}" + }, + { + "bbox": [ + 70, + 563, + 280, + 602 + ], + "type": "text", + "content": " and then assign a relevance score of each sub-segment to the user query " + }, + { + "bbox": [ + 70, + 563, + 280, + 602 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 70, + 563, + 280, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "In practice, rather than training a separate specialized model, we utilize a same VLM-based agent to function as the reward model. While the reward model " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " and exploration agent " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " serve distinct purposes, both can be implemented using the same model. In this work, we employ a single VLM-based agent for both roles across different settings, providing a neat and efficient agent framework that is guided by its self-generated intrinsic reward and without" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 462, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 462, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 462, + 84 + ], + "type": "text", + "content": "relying on auxiliary models or tools." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 92, + 441, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 92, + 441, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 92, + 441, + 105 + ], + "type": "text", + "content": "3.2. Memory Management" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "content": "As exploration proceeds over multiple rounds, the accumulation of frames can lead to significant computational costs for long videos, even when " + }, + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "content": " is small. However, as illustrated in Fig. 2, they tend to gradually disregard irrelevant frames, especially from earlier steps, retaining only the most crucial information in their working memory for later analysis. Inspired by this, we introduce a fixed-size memory buffer " + }, + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "content": " that stores only the most relevant frames throughout exploration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "content": "Specifically, when the agent " + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "content": " explores a video segment " + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "inline_equation", + "content": "s^*" + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "content": ", the associated frames " + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "inline_equation", + "content": "\\{f_{s^*}^i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "content": " are added to the memory buffer. If the total number of frames exceeds the buffer limit, irrelevant frames are removed based on relevance scores " + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "inline_equation", + "content": "r_{s_i^*}" + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "content": " defined in Sec. 3.1. We discard the frames with the lowest scores and only retain the most relevant ones. While these less relevant frames are removed, their visual information is extracted and encoded as text descriptions " + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "inline_equation", + "content": "\\{t_{s_i^*}\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "content": " within the exploration history " + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "inline_equation", + "content": "H_r" + }, + { + "bbox": [ + 313, + 218, + 554, + 377 + ], + "type": "text", + "content": ", allowing the reward model to utilize. This memory management strategy enables the agent to conduct reasoning within a fixed resource limit, regardless of the exploration trajectory's length, thereby preventing excessive memory usage." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 384, + 452, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 452, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 452, + 398 + ], + "type": "text", + "content": "3.3. Tree-Search Exploration" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 402, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 498 + ], + "type": "text", + "content": "Inspired by humans adaptively focusing on different video segments during exploration rather than targeting specific frames directly, our proposed agent VCA treats the video as an unexplored environment, where each segment represents a node on an exploration tree. Given the entire video as the root node, the agent's goal is to efficiently identify an optimal short path leading to a leaf node (segment) containing the most informative content needed to answer the query." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": "At each step, the agent " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": " receives both textual and visual guidance: a set of candidate segments " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": ", which includes a collection of segments with associated reward scores, and a memory buffer " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": " containing multiple frames. If the agent has gathered sufficient information to answer the user's query " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": ", it directly generates a response " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": ". Otherwise, it continues exploring by selecting a specific segment " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "s^* \\in S" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": " to expand the tree based on reward scores." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "However, due to the limited capacity of current foundational models, these reward scores are not always accurate (see further analysis in Sec. 5.3). Selecting segments based on reward scores in a purely greedy manner can potentially lead to a suboptimal solution. Instead, we use the reward scores as guidance for agent " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", allowing it to make the final segment selection independently. In this case, the agent is not forced to select the segment with the highest reward, balancing the exploitation and exploration. More importantly, the candidate set " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " includes both unexplored" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20171" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "coarse segments from prior steps and fine-grained segments from the current step, enabling the agent to backtrack and select earlier segments if it detects that the current exploration path may be suboptimal. Below are the simplified prompts. The full prompt is provided in Appendix ??" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 144, + 231, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 144, + 231, + 158 + ], + "spans": [ + { + "bbox": [ + 71, + 144, + 231, + 158 + ], + "type": "text", + "content": "Exploration Agent Prompt (Simplified)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 167, + 101, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 167, + 101, + 177 + ], + "spans": [ + { + "bbox": [ + 71, + 167, + 101, + 177 + ], + "type": "text", + "content": "Given:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 178, + 181, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 178, + 181, + 190 + ], + "spans": [ + { + "bbox": [ + 71, + 178, + 181, + 190 + ], + "type": "text", + "content": "Candidate segments set: " + }, + { + "bbox": [ + 71, + 178, + 181, + 190 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 191, + 193, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 191, + 193, + 202 + ], + "spans": [ + { + "bbox": [ + 71, + 191, + 193, + 202 + ], + "type": "text", + "content": "Frames in Memory buffer: " + }, + { + "bbox": [ + 71, + 191, + 193, + 202 + ], + "type": "inline_equation", + "content": "M" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 203, + 280, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 203, + 280, + 237 + ], + "spans": [ + { + "bbox": [ + 71, + 203, + 280, + 237 + ], + "type": "text", + "content": "Please choose one segment to explore based on the query " + }, + { + "bbox": [ + 71, + 203, + 280, + 237 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 71, + 203, + 280, + 237 + ], + "type": "text", + "content": ". Output answer if you have gathered sufficient information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 255, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 338 + ], + "type": "text", + "content": "The overall framework of our proposed method is summarized in Algo. 1. Leveraging a tree-search exploration technique boosted with a reward model and a memory buffer, our self-exploration agent adaptively and efficiently explores the video like a human. Empirical results in Sec. 4.3 further validate the effectiveness of our tree-search exploration over naive frame selection methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 341, + 137, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 137, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 137, + 354 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 361, + 182, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 182, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 182, + 375 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 381, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 489 + ], + "type": "text", + "content": "Benchmark. We mainly evaluate VCA on two long-term video question-answering benchmarks. EgoSchema [41] is an egocentric video dataset sourced from Ego4D [21], with a duration of 180 seconds. Here we use the validation set of 500 QA pairs. LVBench [61] covers about 1,500 samples for evaluation of reasoning abilities from high-quality videos over 30 minutes in length. We also conduct experiments on other relevant benchmark datasets in Appendix ??" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 505, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 296, + 601 + ], + "type": "text", + "content": "Baselines. We compare with most relevant baselines to our work, including state-of-the-art open-source Video-LLMs [10, 23, 37, 47, 59, 82], the close-source model GPT-4o [25], and agent-based systems [44, 62, 66]. For video models, we report the results obtained with uniform sampling frames, while for the agent-based systems, we follow their official implementation with GPT-4o [25], consistent with VCA's setup to ensure a fair comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. In this work, we use the same model for the exploration agent and the reward model. Unless otherwise stated, all experiments were performed using the August 2024 version of `gpt-4` as the base model, leveraging the OpenAI LLM API service1 with a temperature of 0.5. We set memory buffer sizes set to 8 and 16 frames for EgoSchema and LVBench, respectively. Further implementation details are provided in Appendix ??" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 315, + 87, + 555, + 123 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 483, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 483, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 483, + 84 + ], + "type": "text", + "content": "Algorithm 1 VCA: Video Curious Agent" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "lines": [ + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": "Require: Video environment " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", question " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", agent " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", reward model " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", memory buffer " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", candidate segments set " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", reward score history " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "H_{r}" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", sampling frame number " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + }, + { + "bbox": [ + 315, + 124, + 395, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 124, + 395, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 124, + 395, + 133 + ], + "type": "text", + "content": "Ensure: Answer " + }, + { + "bbox": [ + 315, + 124, + 395, + 133 + ], + "type": "inline_equation", + "content": "a" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_caption" + } + ], + "index": 13, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 317, + 135, + 542, + 452 + ], + "type": "list", + "angle": 0, + "index": 41, + "blocks": [ + { + "bbox": [ + 319, + 135, + 473, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 135, + 473, + 147 + ], + "spans": [ + { + "bbox": [ + 319, + 135, + 473, + 147 + ], + "type": "text", + "content": "1: Initialize selected segment " + }, + { + "bbox": [ + 319, + 135, + 473, + 147 + ], + "type": "inline_equation", + "content": "s^* \\gets v" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 148, + 493, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 148, + 493, + 159 + ], + "spans": [ + { + "bbox": [ + 321, + 148, + 493, + 159 + ], + "type": "text", + "content": "2: Initialize candidate segments set " + }, + { + "bbox": [ + 321, + 148, + 493, + 159 + ], + "type": "inline_equation", + "content": "S \\gets \\emptyset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 160, + 465, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 160, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 160, + 465, + 171 + ], + "type": "text", + "content": "3: Initialize memory buffer " + }, + { + "bbox": [ + 321, + 160, + 465, + 171 + ], + "type": "inline_equation", + "content": "M\\gets \\emptyset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 171, + 488, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 171, + 488, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 171, + 488, + 182 + ], + "type": "text", + "content": "4: Initialize reward score history " + }, + { + "bbox": [ + 321, + 171, + 488, + 182 + ], + "type": "inline_equation", + "content": "H_{r}\\gets \\emptyset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 184, + 415, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 184, + 415, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 184, + 415, + 194 + ], + "type": "text", + "content": "5: continue " + }, + { + "bbox": [ + 321, + 184, + 415, + 194 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 321, + 184, + 415, + 194 + ], + "type": "text", + "content": " true" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 195, + 420, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 195, + 420, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 195, + 420, + 205 + ], + "type": "text", + "content": "6: while continue do" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 319, + 208, + 541, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 208, + 541, + 218 + ], + "spans": [ + { + "bbox": [ + 319, + 208, + 541, + 218 + ], + "type": "text", + "content": "7: # Sampling frames & form segments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 219, + 522, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 219, + 522, + 233 + ], + "spans": [ + { + "bbox": [ + 321, + 219, + 522, + 233 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 219, + 522, + 233 + ], + "type": "inline_equation", + "content": "\\left\\{f_{s^*}^*\\right\\}_{i=1}^{N}, \\left\\{s_i^*\\right\\}_{i=1}^{N+1} \\gets \\text{UniformSample}(s^*)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 234, + 475, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 234, + 475, + 243 + ], + "spans": [ + { + "bbox": [ + 321, + 234, + 475, + 243 + ], + "type": "text", + "content": "9: # Scoring each segment" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 244, + 542, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 244, + 542, + 258 + ], + "spans": [ + { + "bbox": [ + 318, + 244, + 542, + 258 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 318, + 244, + 542, + 258 + ], + "type": "inline_equation", + "content": "\\left\\{t_{s_i^*};r_{s_i^*}\\right\\}_{i = 1}^{N + 1}\\gets R(\\{s_i^*\\}_{i = 1}^{N + 1};q,\\{f_{s_*}^i\\}_{i = 1}^N,H_r)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 259, + 522, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 259, + 522, + 269 + ], + "spans": [ + { + "bbox": [ + 318, + 259, + 522, + 269 + ], + "type": "text", + "content": "11: # Update candidate segment set" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 270, + 438, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 270, + 438, + 284 + ], + "spans": [ + { + "bbox": [ + 318, + 270, + 438, + 284 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 270, + 438, + 284 + ], + "type": "inline_equation", + "content": "S\\gets S\\cup \\{s_i^*;r_{s_i^*}\\}_{i = 1}^{N + 1}" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 285, + 523, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 285, + 523, + 294 + ], + "spans": [ + { + "bbox": [ + 318, + 285, + 523, + 294 + ], + "type": "text", + "content": "13: # Update history with thoughts" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 295, + 452, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 295, + 452, + 309 + ], + "spans": [ + { + "bbox": [ + 318, + 295, + 452, + 309 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 318, + 295, + 452, + 309 + ], + "type": "inline_equation", + "content": "H_{r}\\gets H_{r}\\cup \\left\\{t_{s_{i}^{*}};r_{s_{i}^{*}}\\right\\}_{i = 1}^{N + 1}" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 310, + 505, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 310, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 318, + 310, + 505, + 319 + ], + "type": "text", + "content": "15: # Update memory with frames" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 321, + 506, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 318, + 321, + 506, + 335 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 318, + 321, + 506, + 335 + ], + "type": "inline_equation", + "content": "M\\gets \\mathrm{UpdateMemory}(M,\\{f_{s^{*}}^{i}\\}_{i = 1}^{N};S)" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 318, + 335, + 541, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 335, + 541, + 345 + ], + "spans": [ + { + "bbox": [ + 318, + 335, + 541, + 345 + ], + "type": "text", + "content": "17: # Output answer or select segment" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 346, + 418, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 346, + 418, + 357 + ], + "spans": [ + { + "bbox": [ + 318, + 346, + 418, + 357 + ], + "type": "text", + "content": "18: " + }, + { + "bbox": [ + 318, + 346, + 418, + 357 + ], + "type": "inline_equation", + "content": "a,s\\gets \\pi (q,S;M)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 318, + 358, + 397, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 358, + 397, + 368 + ], + "spans": [ + { + "bbox": [ + 318, + 358, + 397, + 368 + ], + "type": "text", + "content": "19: if " + }, + { + "bbox": [ + 318, + 358, + 397, + 368 + ], + "type": "inline_equation", + "content": "a = \\emptyset" + }, + { + "bbox": [ + 318, + 358, + 397, + 368 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 370, + 485, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 370, + 485, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 370, + 485, + 381 + ], + "type": "text", + "content": "20: # Continue exploration" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 318, + 383, + 383, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 383, + 383, + 391 + ], + "spans": [ + { + "bbox": [ + 318, + 383, + 383, + 391 + ], + "type": "text", + "content": "21: " + }, + { + "bbox": [ + 318, + 383, + 383, + 391 + ], + "type": "inline_equation", + "content": "s^* \\gets s" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 317, + 394, + 360, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 360, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 360, + 403 + ], + "type": "text", + "content": "22: else" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 317, + 406, + 436, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 436, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 436, + 415 + ], + "type": "text", + "content": "23: continue " + }, + { + "bbox": [ + 317, + 406, + 436, + 415 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 317, + 406, + 436, + 415 + ], + "type": "text", + "content": " false" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 317, + 418, + 369, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 369, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 369, + 427 + ], + "type": "text", + "content": "24: end if" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 317, + 430, + 375, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 430, + 375, + 440 + ], + "spans": [ + { + "bbox": [ + 317, + 430, + 375, + 440 + ], + "type": "text", + "content": "25: end while" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 317, + 442, + 372, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 442, + 372, + 452 + ], + "spans": [ + { + "bbox": [ + 317, + 442, + 372, + 452 + ], + "type": "text", + "content": "26: return " + }, + { + "bbox": [ + 317, + 442, + 372, + 452 + ], + "type": "inline_equation", + "content": "a" + } + ] + } + ], + "index": 40 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 475, + 438, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 438, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 438, + 487 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "text", + "content": "The experimental results on LVBench and EgoSchema are presented in Tab. 2 and Tab. 1, respectively. By adaptively focusing on critical segments rather than scanning the entire video, our agent not only achieves stronger performance but also incurs less computational cost. With less than " + }, + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "text", + "content": " of the frames, VCA achieves substantial improvements over directly feeding uniformly sampled frames into GPT-4o. Specifically, we observe performance gains of " + }, + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "text", + "content": " on the EgoSchema subset and " + }, + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 313, + 492, + 555, + 635 + ], + "type": "text", + "content": " on LVBench. Significant increases in the Event Understanding and Temporal Grounding domains further underscore our effectiveness in identifying crucial segments within long videos." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 635, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 696 + ], + "type": "text", + "content": "Furthermore, we compare our framework with recently proposed SOTA video agent systems, VideoAgent [62] and VideoTree [66]. For a fair comparison, we re-implement both frameworks using GPT-4o for all reasoning components, consistent with our setup. However, captioning all" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 518, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 518, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 518, + 712 + ], + "type": "text", + "content": "1https://platform.openai.com/docs/models" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20172" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 127, + 70, + 482, + 201 + ], + "blocks": [ + { + "bbox": [ + 127, + 70, + 482, + 201 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 482, + 201 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 482, + 201 + ], + "type": "table", + "html": "
MethodFramesEREUKIRTGReaSumAvg.
TimeChat* [47]>9621.921.725.922.725.024.122.3
LLaVA-NeXT* [83]3230.131.234.131.435.027.632.2
InternVL2-40B* [10]1637.439.743.431.442.541.439.8
GLM4V-Plus* [23]2039.935.834.837.740.032.838.3
GPT-4o* [25]6435.930.835.528.333.534.534.7
VideoAgent† [62]Avg. 25.528.030.328.029.328.036.429.3
VideoTree† [66]Avg. 103.230.325.126.527.731.925.528.8
VCA (ours)Avg. 20.043.740.737.838.046.227.341.3
", + "image_path": "cdc65838a4e11839a1a1465235c041052cb787d85754c744b501c8f8c7688785.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 93, + 256, + 258, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "type": "text", + "content": "Table 1. Experimental Results on LVBench. Results of methods marked with * are sourced from LVBench [61]. Methods marked with † indicate implementations where GPT-4o serves as the main component. Our agent's memory buffer size is set to 16 frames. We report the average number of observed frames for agent-based methods, with the best performance highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 256, + 258, + 376 + ], + "lines": [ + { + "bbox": [ + 93, + 256, + 258, + 376 + ], + "spans": [ + { + "bbox": [ + 93, + 256, + 258, + 376 + ], + "type": "table", + "html": "
MethodFramesSubset
GPT-4o [25]869.0
3270.4
\\(VideoAgent^\\dagger\\) [62]Avg. 8.163.2
Avg. 10.760.8
\\(VideoTree^\\dagger\\) [66]Avg. 69.567.0
\\(LVNet^\\dagger\\) [44]1268.2
VCA (ours)Avg. 7.273.6
", + "image_path": "58c7945053aceb0fd40e6d71d931795c4e8194cd4f36892dd1f0f3f85477e13b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 458, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 458, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 54, + 458, + 295, + 626 + ], + "type": "text", + "content": "frames with GPT-4o is prohibitively expensive. Hence, following the baseline implementations, we use open-source VLMs for captioning. A comparison using the same model for all components is provided in Sec. 5.1. As shown in Tab. 1, VCA outperforms all baselines over " + }, + { + "bbox": [ + 54, + 458, + 295, + 626 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 458, + 295, + 626 + ], + "type": "text", + "content": " on EgoSchema, while requiring processing fewer frames. These results suggest the limitations of dense captioning for video preprocessing tasks, which can incur more computational costs while introducing more redundant information. In contrast, our proposed memory management strategy effectively improves efficiency by only retaining the most crucial information for reasoning. Similar insights can be drawn from results on LVBench in Tab. 2. Additional results are provided in Appendix ??" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Moreover, we observe that VCA explores approximately three times more frames on LVBench than on EgoSchema. Recall that EgoSchema videos have an average duration of 180 seconds, while videos in LVBench extend to at least 30 minutes. This phenomenon suggests that our proposed tree-search exploration technique actively explores the video to locate the most relevant information and adaptively in-" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 331, + 256, + 538, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 384, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 440 + ], + "type": "text", + "content": "Table 2. Experimental Results on EgoSchema. Methods marked with " + }, + { + "bbox": [ + 55, + 384, + 295, + 440 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 384, + 295, + 440 + ], + "type": "text", + "content": " indicate implementations where GPT-4o is the main component. The results of LVNet are from [44]. Our agent's memory buffer size is set to 8 frames. We report the average number of observed frames, with the best performance highlighted in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 256, + 538, + 327 + ], + "lines": [ + { + "bbox": [ + 331, + 256, + 538, + 327 + ], + "spans": [ + { + "bbox": [ + 331, + 256, + 538, + 327 + ], + "type": "table", + "html": "
MethodEgoSchemaLVBench
FramesAcc.FramesAcc.
VCA (ours)7.273.620.041.3
- w/o Reward7.572.622.239.5
- w/o Tree Search10.269.839.936.2
", + "image_path": "f6591466ea61c795db00ac6cff0bc93214d577866bef4bd46d75d4ca9daf4aaa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 369, + 385, + 501, + 420 + ], + "blocks": [ + { + "bbox": [ + 313, + 335, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 380 + ], + "type": "text", + "content": "Table 3. Ablation Study on EgoSchema and LVBench. Both the tree-search exploration and reward model contribute significantly to improvements in performance and efficiency. Detailed ablation study results on LVBench are included in Appendix ??" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 369, + 385, + 501, + 420 + ], + "lines": [ + { + "bbox": [ + 369, + 385, + 501, + 420 + ], + "spans": [ + { + "bbox": [ + 369, + 385, + 501, + 420 + ], + "type": "table", + "html": "
Buffer Size81632
Acc.38.941.343.2
", + "image_path": "12eb3b064a7106c503cac75584b9acfb634f0a1eb1846cbde7e0a2a197d01119.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 429, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 462 + ], + "type": "text", + "content": "Table 4. Ablation Study of Buffer Size on LVBench. Increasing the buffer size allows the model to observe more frames during reasoning, leading to improved performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 473, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 558 + ], + "type": "text", + "content": "creases the exploration budget based on task complexity. For example, VCA exhibits both exploitation and exploration: when the agent determines that the current exploration path is suboptimal, it backtracks to previous segments to explore a new path (see Sec. 5.2). These empirical observations further validate the robustness and scalability of our approach, especially in diverse long-video contexts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 408, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 408, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 408, + 578 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 700 + ], + "type": "text", + "content": "In this section, we conduct an ablation study to evaluate the components of our agent. Results across both benchmarks are presented in Tab. 3. As illustrated in the table, performance consistently declines when the reward model is removed, highlighting that the agent's inherent reasoning ability alone is insufficient for accurately identifying the most informative segments. This finding emphasizes the effectiveness of incorporating an intrinsic reward to guide the agent's segment selection process. A detailed analysis of our reward model's performance is provided in Sec. 5.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 702, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 702, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 702, + 554, + 714 + ], + "type": "text", + "content": "Additionally, we observe a significant performance drop" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20173" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 136, + 76, + 279, + 101 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 136, + 76, + 223, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 76, + 223, + 86 + ], + "spans": [ + { + "bbox": [ + 136, + 76, + 223, + 86 + ], + "type": "text", + "content": "Why does the mice run away?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 138, + 86, + 279, + 94 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 86, + 279, + 94 + ], + "spans": [ + { + "bbox": [ + 138, + 86, + 279, + 94 + ], + "type": "text", + "content": "(A) Because the black cat sheriff calls the police" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 138, + 94, + 232, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 94, + 232, + 101 + ], + "spans": [ + { + "bbox": [ + 138, + 94, + 232, + 101 + ], + "type": "text", + "content": "(C) Because their boss is injured" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 282, + 85, + 407, + 100 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 282, + 85, + 406, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 85, + 406, + 93 + ], + "spans": [ + { + "bbox": [ + 282, + 85, + 406, + 93 + ], + "type": "text", + "content": "(B) Because they meet more powerful mice" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 282, + 93, + 407, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 93, + 407, + 100 + ], + "spans": [ + { + "bbox": [ + 282, + 93, + 407, + 100 + ], + "type": "text", + "content": "(D) Because the owner of the house is back" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 136, + 105, + 477, + 145 + ], + "blocks": [ + { + "bbox": [ + 136, + 105, + 477, + 145 + ], + "lines": [ + { + "bbox": [ + 136, + 105, + 477, + 145 + ], + "spans": [ + { + "bbox": [ + 136, + 105, + 477, + 145 + ], + "type": "image", + "image_path": "494c2e39d33787b489c52c42d9cf3f62f97d802bf3cb5b76c77ffd4cb93b8d38.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 149, + 190, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 149, + 190, + 158 + ], + "spans": [ + { + "bbox": [ + 137, + 149, + 190, + 158 + ], + "type": "text", + "content": "Reward Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 159, + 369, + 211 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 137, + 159, + 366, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 159, + 366, + 182 + ], + "spans": [ + { + "bbox": [ + 137, + 159, + 366, + 182 + ], + "type": "text", + "content": "Segment1: Thought: The segment shows a black cat with a gun, possibly indicating a threat. This could suggest that the mice might run away due to the presence of the cat, which aligns with option (A). Score: 0.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 186, + 369, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 186, + 369, + 211 + ], + "spans": [ + { + "bbox": [ + 137, + 186, + 369, + 211 + ], + "type": "text", + "content": "Segment4: Thought: The segment features the black cat with a communication device, possibly calling for help or backup. This aligns with option (A), suggesting the mice might run because the cat is calling the police. Score: 0.8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 373, + 167, + 426, + 198 + ], + "blocks": [ + { + "bbox": [ + 374, + 148, + 436, + 166 + ], + "lines": [ + { + "bbox": [ + 374, + 148, + 436, + 166 + ], + "spans": [ + { + "bbox": [ + 374, + 148, + 436, + 166 + ], + "type": "text", + "content": "Agent Choice: (segment: 4)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 373, + 167, + 426, + 198 + ], + "lines": [ + { + "bbox": [ + 373, + 167, + 426, + 198 + ], + "spans": [ + { + "bbox": [ + 373, + 167, + 426, + 198 + ], + "type": "image", + "image_path": "cab27bdb7ff0b0177e4565f738e98f8d8b8a7fec79b098210f2663bdfcdbd937.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 427, + 167, + 476, + 198 + ], + "blocks": [ + { + "bbox": [ + 427, + 167, + 476, + 198 + ], + "lines": [ + { + "bbox": [ + 427, + 167, + 476, + 198 + ], + "spans": [ + { + "bbox": [ + 427, + 167, + 476, + 198 + ], + "type": "image", + "image_path": "db534f7962bd217c6a8212442a281185644644533499078be1979c927fe8315b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 137, + 216, + 477, + 256 + ], + "blocks": [ + { + "bbox": [ + 137, + 216, + 477, + 256 + ], + "lines": [ + { + "bbox": [ + 137, + 216, + 477, + 256 + ], + "spans": [ + { + "bbox": [ + 137, + 216, + 477, + 256 + ], + "type": "image", + "image_path": "9709989d9606dbfa36835e193981ef9fd1b3c4452fb7a2442047d274c61b2bbb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 137, + 259, + 189, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 259, + 189, + 268 + ], + "spans": [ + { + "bbox": [ + 137, + 259, + 189, + 268 + ], + "type": "text", + "content": "Reward Model" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 137, + 269, + 368, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 269, + 368, + 293 + ], + "spans": [ + { + "bbox": [ + 137, + 269, + 368, + 293 + ], + "type": "text", + "content": "Segment4.1: Thought: The segment shows a close-up of a document with red text and exclamation marks, indicating urgency or importance. This could imply a critical message or alert being communicated. Score: 0.7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 137, + 293, + 143, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 293, + 143, + 300 + ], + "spans": [ + { + "bbox": [ + 137, + 293, + 143, + 300 + ], + "type": "text", + "content": "…" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 137, + 307, + 476, + 347 + ], + "blocks": [ + { + "bbox": [ + 137, + 307, + 476, + 347 + ], + "lines": [ + { + "bbox": [ + 137, + 307, + 476, + 347 + ], + "spans": [ + { + "bbox": [ + 137, + 307, + 476, + 347 + ], + "type": "image", + "image_path": "0be7582fe7139255d5979ad1a0a4b2f8416fde86fabbd6434c0c69aee2357f20.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 137, + 351, + 190, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 351, + 190, + 359 + ], + "spans": [ + { + "bbox": [ + 137, + 351, + 190, + 359 + ], + "type": "text", + "content": "Reward Model" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 137, + 360, + 358, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 360, + 358, + 393 + ], + "spans": [ + { + "bbox": [ + 137, + 360, + 358, + 393 + ], + "type": "text", + "content": "Segment1.2: Thought: The segment features mice characters in mid-air, possibly fleeing or being thrown. This suggests a direct reaction to a threat, supporting the idea that they are running away from the cat. Score: 0.8" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 137, + 398, + 476, + 427 + ], + "blocks": [ + { + "bbox": [ + 137, + 398, + 476, + 427 + ], + "lines": [ + { + "bbox": [ + 137, + 398, + 476, + 427 + ], + "spans": [ + { + "bbox": [ + 137, + 398, + 476, + 427 + ], + "type": "image", + "image_path": "e82b2c43037a928c41960aef4203fd12a1f1ce59e083cac00c6ba6a3ec87c9ad.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 451, + 553, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 451, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 553, + 473 + ], + "type": "text", + "content": "Figure 4. Exploration Trajectory Example. The agent generally tends to select video segments with higher reward scores for exploration, yet it can adaptively redirect to earlier segments, effectively balancing exploitation and exploration." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 351, + 430, + 369, + 442 + ], + "blocks": [ + { + "bbox": [ + 351, + 430, + 369, + 442 + ], + "lines": [ + { + "bbox": [ + 351, + 430, + 369, + 442 + ], + "spans": [ + { + "bbox": [ + 351, + 430, + 369, + 442 + ], + "type": "image", + "image_path": "5386d87e480fee48366b58408db92687c9c944cdedb4c56e93e21002e2bd9557.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 392, + 431, + 457, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 431, + 457, + 441 + ], + "spans": [ + { + "bbox": [ + 392, + 431, + 457, + 441 + ], + "type": "text", + "content": "Agent: Answer: A" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 484, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 627 + ], + "type": "text", + "content": "when the tree-search exploration technique is ablated, i.e., the agent directly selects frame indices without the segment-based tree structure. Interestingly, this modification also results in a substantial increase in the average number of frames acquired, especially on LVBench. This indicates that the agent struggles to locate the most crucial information and consequently produces a longer but redundant trajectory. These ablation study results validate the effectiveness of the tree-search exploration framework assisted by the reward model. Additionally, this underscores VCA's potential to achieve greater efficiency and improved performance with more advanced VLMs and reward models." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Moreover, as shown in Tab. 4, increasing the memory buffer size yields consistently improved performance for VCA, demonstrating that our framework benefits from a broader reasoning window while maintaining effectiveness with fewer frames. Due to limited computation resources, we only tested up to 32 frames; however, the upward trend indicates the potential for gains with additional frames." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 483, + 373, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 373, + 496 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 373, + 496 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 503, + 549, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 549, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 549, + 516 + ], + "type": "text", + "content": "5.1. Can VCA generalize to Open-source Models?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "While VCA demonstrates strong performance using GPT4o as the base model, we also analyze whether these positive results can be extrapolated to open-source models. We implement the VCA framework with Qwen2-VL-72B-Instruct-GPTQ-Int4 [59] as both the exploration agent and reward model, as well as the captioner for baselines. The results in Tab. 5 reveal that, compared to directly feeding uniformly sampled frames into Qwen2-VL, our framework achieves higher accuracy on the EgoSchema benchmark despite using only about " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " of the frames. In contrast, VideoAgent performs significantly worse. Furthermore, our approach outperforms VideoTree with far fewer frames. Notably, VideoTree employs clustering to build a caption similarity-based tree structure for frame selection, requiring caption generation for each frame. This process is computationally expensive, demanding over 60 GPU hours" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20174" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 70, + 261, + 177 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 261, + 177 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 261, + 177 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 261, + 177 + ], + "type": "table", + "html": "
MethodFramesSubset
Qwen2-VL870.2
3274.2
VideoAgentAvg. 8.665.2
VideoTreeAvg. 85.471.0
VCA (ours)Avg. 16.975.2
- w/ Image RewardAvg. 20.974.0
", + "image_path": "de5d40b1cd7f8593e0c298b92c7fc57eb36751dc7abde60dd39b7f4435f5e892.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 185, + 295, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 295, + 231 + ], + "type": "text", + "content": "Table 5. Performance of Using Open-source Model on EgoSchema. Our approach achieves higher accuracy with fewer observed frames, showing consistency improvements over baseline methods. Memory buffer size is set to 16 frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 255, + 295, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 279 + ], + "type": "text", + "content": "on a single H100, as shown in Fig. 1. These results further underscore the robustness of VCA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 277, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 277, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 277, + 296 + ], + "type": "text", + "content": "5.2. What's the Exploration Behavior of VCA?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 303, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 482 + ], + "type": "text", + "content": "Here we present a real exploration trajectory from LVBench to illustrate our agent's decision-making process. As shown in Fig. 4, given a user query, the reward model initially evaluates each segment's relevance based on uniformly sampled frames, and the agent selects the fourth segment, which has the highest relevance score. However, upon further investigation of segment 4, both the reward model and the agent find it relatively irrelevant to the query. Rather than continuing with it, the agent redirects its exploration to segment 1 in the last round. After selecting segment 1, the agent chooses to investigate the third segment, despite the second segment having the highest reward score. This behavior reflects VCA's ability to adaptively adjust its exploration trajectory based on visual context rather than purely following the reward score, resulting in a more robust system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 604 + ], + "type": "text", + "content": "We also examine the limitations of VCA by analyzing its failure cases, which fall into the following scenarios: The most common failure occurs when the agent overlooks subtle clues, resulting in an inaccurate response to the user query even after prolonged exploration, particularly for queries requiring specific visual details. Other failure cases include instances where the agent is misled by inaccurate reward scores or fails to provide the correct answer despite observing key frames. Detailed cases and further discussion are provided in the Appendix ??" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "In addition, unlike previous methods, VCA employs a segment-level exploration strategy instead of direct image similarity to select key frames. To further investigate this difference, we evaluate the performance of selecting the segment whose start and end images have the highest visual relevance. As shown in Tab. 5, relying solely on image relevance instead of inferring segment semantics results in a performance drop on EgoSchema, confirming the effectiveness of our segment-level exploration." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 71, + 538, + 255 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 538, + 255 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 538, + 255 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 538, + 255 + ], + "type": "image", + "image_path": "fca8834cee31104ddaa6b77c3ab66dbfffdf08de055321278fd4a45f2a317be3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 258, + 554, + 313 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 313 + ], + "type": "text", + "content": "Figure 5. Segment Selection Accuracy on LVBench. The y-axis shows segment distance from ground truth; x-axis shows count ratio. Red curves are fitted normal distributions. The reward-guided agent selects segments closer to ground truth than the reward model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 334, + 504, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 504, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 504, + 346 + ], + "type": "text", + "content": "5.3. How Reliable is the Reward Model?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 351, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 460 + ], + "type": "text", + "content": "To further assess the capability of our reward model, we visualize the matching accuracy between segments with the highest reward scores and ground truth time references from LVBench. The results are shown at the bottom of Fig. 5. We observe that segment distances follow an approximately normal distribution with a mean of 1.28, indicating that the segment with the highest reward score often closely aligns with the ground truth segment. This validates the reward model's effectiveness in providing meaningful guidance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "type": "text", + "content": "Additionally, as discussed in Sec 3.3, we encourage the agent to make independent decisions when selecting segments rather than blindly following reward scores in a purely greedy manner. The benefit of this design choice is demonstrated at the top of Fig. 5, where we present the matching accuracy of the agent's choices. Compared to the greedy approach, the agent's selections result in a sharper distribution of segment distances, with an average distance gap of " + }, + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "type": "inline_equation", + "content": "-0.60" + }, + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "type": "text", + "content": ", indicating closer alignment with the ground truth. In summary, our reward model assigns reliable relevance scores to segments, while the agent further refines these choices to achieve even more accurate selections." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 616, + 388, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 388, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 388, + 628 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 635, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 713 + ], + "type": "text", + "content": "In this work, we propose a curiosity-driven video agent framework for long video understanding tasks with self-exploration capability. VCA achieves state-of-the-art performance and efficiency without relying on auxiliary tools. Appendix ?? shows that improved reward guidance could further boost results, highlighting the potential of synthetic data and specialized reward models for future work." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20175" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 114, + 294, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 294, + 167 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 294, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 294, + 223 + ], + "type": "text", + "content": "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Devendra Chaplot, Jessica Chudnovsky, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amélie Héliou, Paul Jacob, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 225, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 294, + 278 + ], + "type": "text", + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6816-6826, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 280, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 324 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 325, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 294, + 378 + ], + "type": "text", + "content": "[5] Ivana Balazevic, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J Henaff. Memory consolidation enables long-context video understanding. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "text", + "content": "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Fuyu-8b: A multimodal architecture for ai agents, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 445 + ], + "type": "text", + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding?, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 448, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 294, + 479 + ], + "type": "text", + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset, 2018. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 534 + ], + "type": "text", + "content": "[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "text", + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 646 + ], + "type": "text", + "content": "[11] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "text", + "content": "[12] Jiwan Chung and Youngjae Yu. Long story short: a summarize-then-search method for long video question answering. arXiv preprint arXiv:2311.01233, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[13] Robert Desimone, John Duncan, et al. Neural mechanisms of selective visual attention. Annual review of neuroscience, 18(1):193-222, 1995. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[15] Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, Katsushi Ikeuchi, Hoi Vo, Li Fei-Fei, and Jianfeng Gao. Agent ai: Surveying the horizons of multimodal interaction, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[16] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 230, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 283 + ], + "type": "text", + "content": "[17] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2025. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[18] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition, 2016. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 321, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 365 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 366, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 420 + ], + "type": "text", + "content": "[20] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 488 + ], + "type": "text", + "content": "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 567 + ], + "type": "text", + "content": "[22] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, Lei Zhao, Zhuoyi Yang, Xiaotao Gu, Xiaohan Zhang, Guanyu Feng, Da Yin, Zihan Wang, Ji Qi, Xixuan Song, Peng Zhang, Debing Liu, Bin Xu, Juanzi Li, Yuxiao Dong, and Jie Tang. CogvIm2: Visual language models for image and video understanding, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "text", + "content": "[23] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "type": "text", + "content": "[24] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20176" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[26] Lawrence Jang, Yinheng Li, Charles Ding, Justin Lin, Paul Pu Liang, Dan Zhao, Rogerio Bonatti, and Kazuhito Koishida. Videowebarena: Evaluating long context multimodal agents with video understanding web tasks, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[27] Peng Jin, Ryuichi Takanobu, Caiwan Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[28] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[29] Bruno Korbar, Yongqin Xian, Alessio Tonioni, Andrew Zisserman, and Federico Tombari. Text-conditioned resampler for long form video understanding. In European Conference on Computer Vision, pages 271-288. Springer, 2025. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 316 + ], + "type": "text", + "content": "[30] Somnath Kumar, Yash Gadhia, Tanuja Ganu, and Akshay Nambi. Mmctagent: Multi-modal critical thinking agent framework for complex visual reasoning, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "text", + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "text", + "content": "[32] Yunxin Li, Xinyu Chen, Baotain Hu, and Min Zhang. Llms meet long video: Advancing long video comprehension with an interactive visual adapter in llms, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[33] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2025. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[34] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "text", + "content": "[35] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 570 + ], + "type": "text", + "content": "[36] Ming C Lin and Shan Yang. Vila: Efficient video-language alignment for video question answering. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 571, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 603 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[38] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint arXiv:2402.08268, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "text", + "content": "[39] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. arXiv preprint arXiv:2406.12846, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[40] Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[41] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. arXiv preprint arXiv:2308.09126, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[42] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification, 2015. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 240 + ], + "type": "text", + "content": "[43] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14386-14397, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "text", + "content": "[44] Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya, Wonjeong Ryoo, Donghyun Kim, and Michael S Ryoo. Too many frames, not all useful: Efficient strategies for long-form video qa. arXiv preprint arXiv:2406.09396, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[45] Ruchit Rawal, Khalid Saifullah, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. In Synthetic Data for Computer Vision Workshop @ CVPR 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[46] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. arXiv preprint arXiv:2312.02051, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[47] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 521 + ], + "type": "text", + "content": "[48] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "type": "text", + "content": "[49] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Xun Guo, Tian Ye, Yan Lu, Jenq-Neng Hwang, et al. Moviechat: From dense token to sparse memory for long video understanding. arXiv preprint arXiv:2307.16449, 2023. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 612 + ], + "type": "text", + "content": "[50] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "text", + "content": "[51] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. Advances in neural information processing systems, 35:38032-38045, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[52] Yunlong Tang, Jing Bi, Sitting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, Ali Vosoughi, Chao Huang, Zeliang Zhang, Pinxin Liu, Mingqian Feng, Feng Zheng, Jianguo Zhang," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20177" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Ping Luo, Jiebo Luo, and Chenliang Xu. Video understanding with large language models: A survey, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "text", + "content": "[53] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "text", + "content": "[54] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 286 + ], + "type": "text", + "content": "[56] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "text", + "content": "[57] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks, 2015. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "text", + "content": "[58] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 442 + ], + "type": "text", + "content": "[59] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[61] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Shiyu Huang, Bin Xu, Yuxiao Dong, Ming Ding, and Jie Tang. Lvbench: An extreme long video understanding benchmark, 2024. 1, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[62] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[63] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[64] Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recurrent memory bridges. arXiv preprint arXiv:2409.01071, 2024. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[65] Ying WANG, Yanlai YANG, and Mengye REN. Lifelong-memory: Leveraging llms for answering queries in long" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "form egocentric videos. arXiv preprint arXiv:2312.05269, 2024. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[66] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[67] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[68] Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision, pages 453-470. Springer, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 281 + ], + "type": "text", + "content": "[69] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "text", + "content": "[70] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "text", + "content": "[71] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. CoRR, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "text", + "content": "[72] Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, and Kevin Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "text", + "content": "[73] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. In ICLR 2024 Workshop on Large Language Model (LLM) Agents. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "text", + "content": "[74] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[75] Jiaqi Xu, Cuiling Lan, Wenxuan Xie, Xuejin Chen, and Yan Lu. Retrieval-based video language model for efficient long video question answering, 2023. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[76] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[77] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 626, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 553, + 669 + ], + "type": "text", + "content": "[78] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[79] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20178" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 419 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[81] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[82] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 283 + ], + "type": "text", + "content": "[83] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "type": "text", + "content": "[84] Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding, Tao Gui, Qi Zhang, and Xuanjing Huang. Longagent: Scaling language models to 128k context through multi-agent collaboration. arXiv preprint arXiv:2402.11550, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "type": "text", + "content": "[85] Long Zhao, Nitesh Bharadwaj Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. In *Forty-first International Conference on Machine Learning*. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 419 + ], + "type": "text", + "content": "[86] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations, 2020. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20179" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_content_list.json b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e7a8f51ea87ef75a55d98e37e13f2da4bc265cd8 --- /dev/null +++ b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_content_list.json @@ -0,0 +1,1315 @@ +[ + { + "type": "text", + "text": "VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation", + "text_level": 1, + "bbox": [ + 250, + 128, + 748, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shoubin Yu $^{1,3*}$ Difan Liu $^{1*}$ Ziqiao Ma $^{1,2*}$ Yicong Hong $^{1}$ Yang Zhou $^{1}$ Hao Tan $^{1}$ Joyce Chai $^{2}$ Mohit Bansal $^{3}$", + "bbox": [ + 202, + 202, + 794, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Adobe Research $^{2}$ University of Michigan $^{3}$ UNC Chapel Hill", + "bbox": [ + 235, + 255, + 761, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://veggie-gen.github.io/", + "bbox": [ + 351, + 276, + 640, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6de461b48922321aa8d110e2c86339f81fccd8681d7979d3daf8439499923d75.jpg", + "image_caption": [ + "Figure 1. We propose VEGGIE, a unified and versatile video generative model that handles various tasks for both video concept grounding and editing according to user instructions. With VEGGIE, users can locate, add, delete, and change concepts in a given video through diverse instruction formats (direct referring instruction or reasoning-demanding questions). Users can also edit videos with multimodal instruction empowered by MLLM, enabling applications like video editing from a reference image." + ], + "image_footnote": [], + "bbox": [ + 133, + 309, + 857, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 635, + 326, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While recent video diffusion models enable video editing, unifying diverse instructional editing tasks (e.g., add, remove, modify) under a single framework remains a significant challenge. In this paper, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions, a simple end-to-end framework that unifies video concept editing, grounding, and reasoning based on diverse user instructions. Specifically, given a video and text query, VEGGIE first utilizes an MLLM to interpret user intentions in instructions and ground them to the video contexts, generating framespecific grounded task queries for pixel-space responses. A diffusion model then renders these plans and generates edited videos that align with user intent. To support diverse", + "bbox": [ + 88, + 667, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks and complex instructions, we employ a curriculum learning strategy: first aligning the MLLM and video diffusion model with large-scale instructional image editing data, followed by end-to-end fine-tuning on high-quality multitask video data. Additionally, we introduce a novel data synthesis pipeline to generate paired instructional video editing data for model training. It transforms static image data into diverse, high-quality video editing samples by leveraging Image-to-Video models to inject dynamics. VEGGIE shows strong performance in instructional video editing with different editing skills, outperforming the best instructional baseline as a versatile model, while other models struggle with multi-tasking. VEGGIE also excels in video object grounding and reasoning segmentation, where other baselines fail. We further reveal how the multiple tasks help each other and highlight promising applications like zero-shot multimodal instructional and in-context video editing.", + "bbox": [ + 511, + 637, + 908, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 218, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done during internship at Adobe Research.", + "bbox": [ + 109, + 887, + 372, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15147", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the advances in Video Diffusion Models (VIDDMs) [2, 4, 24, 25, 74], video editing methods have emerged as video design tools, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 86] have been developed, using triples of text prompts, source videos, and target videos for training. Due to their limited performance in understanding user intent and multimodal semantics [17], several methods have incorporated multimodal large language models (MLLMs) to handle complex instructions/reasoning [17, 29, 75].", + "bbox": [ + 89, + 114, + 485, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing methods fall short of the goal of a simple, versatile video concept editor, facing three primary challenges. First, most methods are not end-to-end, requiring intermediate layout/mask/human or model caption guidance [35, 53, 63, 75], which adds workload on users and disrupts a seamless editing experience. Second, existing pipelines connecting MLLMs to VidDMs require multiple training objectives beyond simple pixel-space diffusion loss, such as language loss [75] or mask losses [66]. This increases optimization difficulty and often requires additional hyperparameter tuning or annotations. Third, existing video editing models, both instructional and non-instructional, struggle with handling other diverse editing tasks, ranging from addition, and deletion to stylization. For example, LGVI [66] fails in global edits such as stylization and color change, while VidToMe [41] struggles with local edits such as adding or removing objects. These methods also struggle with input videos that contain multiple objects or when user instructions require complex reasoning.", + "bbox": [ + 89, + 296, + 485, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These challenges result from two limitations: First, there is a lack of multitasking fine-tuning on well-curated instructional video editing datasets that span a broad range of skills. Second, models often lack two critical capabilities needed to interpret user intentions and accurately locate concepts: multimodal reasoning to infer the intended modification from the user's instruction; and grounding language to the input video to precisely identify the region or object to be edited. For example, in Figure 1, one can effortlessly locate the girl given \"identify the little girl.\" When asked to \"add a hat to the little girl,\" we intuitively imagine the hat placed on her head from commonsense, even without seeing an actual hat.", + "bbox": [ + 89, + 584, + 485, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions. VEGGIE unifies video concept grounding and editing without relying on additional layout, mask guidance, or intermediate caption [35, 39, 42, 75, 78, 79]. Instead, we formulate the problem as end-to-end grounded generation in pixel space, using only a diffusion loss. Specifically, given a video and a text query, VEGGIE first leverages an MLLM to interpret complex instructions, generating frame-wise con", + "bbox": [ + 89, + 765, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ditions. Unlike prior methods [43, 75] that use discrete text tokens as conditions, which disconnect the pipeline and block gradient propagation, VEGGIE employs continuous, learnable task query embeddings per frame. This enables end-to-end training and effectively captures grounded task representations for diffusion model conditioning. To handle diverse tasks and accurately interpret complex queries, we employ a curriculum learning strategy that begins by aligning MLLMs with diffusion models using massive paired instructional image editing data and then fine-tuning the model end-to-end on high-quality multitask video data to adapt video. Unlike tool-use methods [16, 35, 75], VEGGIE formulates both video grounding and instructional editing in the same video-to-video task formulation, enabling efficient handling through a unified single model. To further support end-to-end training, we introduce a novel automatic instructional video data generation pipeline that lifts high-quality instructional image editing data into the video domain using image-to-video and video evaluation tools.", + "bbox": [ + 511, + 90, + 906, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing video editing benchmarks do not provide wide and uniform coverage of diverse editing skills [59, 68]. To address this gap, we contribute VEG-Bench, an instructional video editing benchmark that spans 8 editing skills: concept addition, removal, object changing, environment background changing, visual feature changing, stylization, object grounding, and reasoning segmentation. Each skill is evaluated using a dedicated suite of metrics. We assess the proposed VEGGIE alongside 6 baselines on VEG-Bench. VEGGIE demonstrates strong performance across diverse editing skills, outperforming the best instructional baseline as a versatile, all-in-one model, while other models struggle with multi-tasking. Additionally, VEGGIE excels in video object grounding and reasoning segmentation tasks, where other baselines fall short. We show further analysis of how multi-task learning enhances our framework and highlight applications such as zero-shot multimodal instructional following (Fig. 2) and few-shot in-context editing (Fig. 3). Our contributions are summarized as follows:", + "bbox": [ + 511, + 381, + 908, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose VEGGIE, an end-to-end model that integrates an MLLM and a VidDM. VEGGIE is a versatile framework that handles diverse instructional requests for editing and grounding various video concepts. Unlike existing work that achieves multitasking via tool use, VEGGIE unifies diverse tasks in a single model, thus simplifying the training with only diffusion loss.", + "- We propose a data synthesis pipeline, scaling high-quality instructional video editing data for future work.", + "- We propose VEG-Bench, an instructional video editing benchmark that spans 8 editing skills with dedicated metrics for each skill.", + "- VEGGIE achieves strong performance across diverse editing skills compared with SoTA methods, and shows potentials for multimodal instruction and in-context following." + ], + "bbox": [ + 514, + 674, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15148", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e0d151b937adc7407e14b52cd875ca0bfb673a4ac4838074cdae58f7cce26412.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 161, + 89, + 243, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4540fb3623e1042b72cdc51d40fae52db70bda03718617114a98d44bd5cf175c.jpg", + "image_caption": [ + "Transfer the style in the reference image.", + "Add the object in the reference image on the women." + ], + "image_footnote": [], + "bbox": [ + 248, + 90, + 493, + 218 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a83a9ab1de823a8e4032e1dedd9635d6aa3ee84abd0b1589dedcb0a1c5522e76.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 500, + 101, + 581, + 142 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/22d931f8d1d6d4cd46b69a5e91dbd7bdc650d603c3c3e56bcf3baecacfd0144e.jpg", + "image_caption": [ + "Figure 2. Multimodal instruction following emerges in VEGGIE, allowing for style transfer or object addition from reference images." + ], + "image_footnote": [], + "bbox": [ + 586, + 90, + 831, + 217 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b5f1369b96b80acff89b8832473d68767bd364283daa8b18fad10e9742759d4e.jpg", + "image_caption": [ + "Figure 3. In-context editing emerges in VEGGIE, allowing for few-shot learning of editing tasks with paired image demonstrations." + ], + "image_footnote": [], + "bbox": [ + 161, + 262, + 831, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 448, + 232, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instructional Video Editing Video Diffusion Models (VidDMs) [2, 4, 6, 7, 15, 26, 49, 67, 71], enable high-quality video generation across a wide range of video concepts. Building on these advances, video editing methods have emerged as tools for video design, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 70, 86] have been developed, using triples of video instructions, source videos, and target videos for training. These methods demonstrate limited performance when complex multimodal reasoning is required, as noted by previous research on instructional image editing [17]. Moreover, they struggle with diverse editing tasks, from addition and deletion to stylization. For example, LGVI [66] is primarily designed for removal tasks, while TokenFlow [20] struggles with local edits such as adding, removing, or changing objects. We address this limitation with pixel-level multitasking fine-tuning on well-curated instructional video editing datasets covering various grounding and editing skills.", + "bbox": [ + 89, + 474, + 483, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Grounding and Segmentation Visual grounding requires models to connect language to its corresponding visual concept in the visual context [40, 47]. This is commonly evaluated via the language-guided semantic localization tasks, ranging from simple referring expressions in RefCOCO series [48, 77] and their generalized variant [44] that takes no-target and multi-target into account. Recently, grounded multimodal large lan", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guage models (MLLMs) are trained for object grounding to bounding boxes [8, 51, 52, 76, 82, 87] and segmentation masks [56, 69, 81, 84] using text-image pairs with fine-grained annotations linking phrases to entities. These models unlock the potential of reasoning segmentation [11, 36], bringing language-informed reasoning into semantic segmentation. Instead of using dedicated object detection or segmentation modules, we achieve video grounding through end-to-end training with only diffusion loss.", + "bbox": [ + 511, + 450, + 906, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method: VEGGIE", + "text_level": 1, + "bbox": [ + 511, + 599, + 725, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce a Video Editor with Grounded Generation from Instructions (VEGGIE), a unified and versatile generative video model. It combines the complex instruction understanding and reasoning capabilities of MLLMs with the generative capacity of VidDMs. The model is trained end-to-end with diffusion loss only. VEGGIE efficiently handles diverse user inputs, including direct instructions, complex questions requiring in-depth reasoning, and multimodal conditioning. It performs various pixel-level manipulations, enabling tasks such as video concept addition, removal, changing, stylization, grounding, and reasoning segmentation based on user instructions. We elaborate on the model design (Sec. 3.1), training and inference process (Sec. 3.2), and data curation (Sec. 3.3).", + "bbox": [ + 511, + 626, + 908, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 511, + 848, + 700, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VEGGIE consists of four main components (see Fig. 4): (1) a multimodal large language model, (2) a set of learnable", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15149", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/419c74734a879ff33d66c796ab67f7d5591f5073932ab3c43c6916f3395226b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 92, + 225, + 109 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4614bdaab383a2d7be34477abb6dd43be30e9fc13b804982c124645ab6ae1578.jpg", + "image_caption": [ + "Figure 4. Overview of our proposed end-to-end VEGGIE framework. Our Multimodal Large Language Model first understands input video frames and diverse user instructions, then it generates frame-wise reasoning queries that maintain per-frame editing conditions for the video diffusion model. The video diffusion model will render the MLLM-generated conditions to the pixel space for diverse tasks, including video editing, video grounding, and video reasoning segmentation with questions. We only apply diffusion loss for the whole pipeline training." + ], + "image_footnote": [], + "bbox": [ + 112, + 118, + 883, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "grounded task queries, (3) an alignment network (single-layer MLP) that projects the MLLM output into the condition space of the diffusion model, and (4) a video diffusion model initialized from an instructional image editing model [83]. Our model first generates latent conditions for target video frames by querying multimodal context using an MLLM, then renders these conditions at the pixel level through a video diffusion model, as detailed below.", + "bbox": [ + 88, + 443, + 485, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MLLM for Generating Grounded Task Guidance. As illustrated in the left of Fig. 4, given a video consisting of a sequence of frames $V = [f_1, \\dots, f_n]$ , where $n$ is the frame number of the given video, a user instruction/question $I$ , our goal is to obtain the response $\\widehat{V} = [\\widehat{f}_1, \\dots, \\widehat{f}_n]$ at pixel space that faithfully reflects user instruction about the given video. The MLLM module processes both the input video $V$ and a user instruction $I$ to generate a sequence of grounded task tokens per frame: $C = [c_1, \\dots, c_n]$ , which are input and output in parallel. These tokens serve as task guidance and implicitly encode the target manipulation, such as object attributes, spatial relationships, or style transfer parameters. The MLLM ensures the model captures both explicit user instructions and implicit reasoning needs.", + "bbox": [ + 88, + 565, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VidDM for Rendering MLLM Guidance at Pixel Space. As illustrated in the right of Table 4, the VidDM takes the original video $V$ and the grounded task tokens $C$ as conditions to synthesize the target video $\\widehat{V}$ . The original video is concatenated with the noise volume, and the task tokens are input to the cross-attention. With grounded task guidance in denoising steps, the generation process ensures that the output faithfully follows user instructions while preserving the", + "bbox": [ + 88, + 779, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "video's structure and motion dynamics. Through iterative denoising, it refines each frame while maintaining temporal consistency, applying pixel modifications coherently for a smooth and visually consistent output video $\\widehat{V}$ .", + "bbox": [ + 511, + 443, + 906, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Curriculum Learning from Images to Videos", + "text_level": 1, + "bbox": [ + 511, + 515, + 893, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training the model directly on video tasks presents two key challenges: (1) misalignment between MLLM and diffusion model representations, making it difficult for the diffusion model to interpret MLLM-generated task queries with limited fine-tuning data, and (2) the diffusion model's lack of multitasking capability, even for image tasks, due to insufficient training on diverse tasks. Our initial experiments also found the model collapsed when the whole pipeline was directly trained with all data. These challenges/observations underscore the need for pre-alignment between MLLM and the diffusion model to enable seamless adaptation from language-space task queries to pixel-space modifications. To this end, we adopt a two-stage curriculum learning strategy for the proposed VEGGIE framework.", + "bbox": [ + 511, + 537, + 906, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1: Aligning Diffusion and Language Spaces. In the first stage, we align the diffusion model with the MLLM using large-scale image-level instructional editing data. The MLLM remains frozen while we update the alignment network, grounded task queries, and diffusion UNet. This process fine-tunes the diffusion model weights to align with the language space, enabling the model to interpret MLLM-generated guidance and translate user instructions into pixel-level edits while preserving the MLLM's strong ability to understand instructions and user intentions.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "VEGGIE: A Unified and Versatile Instructional Video Generative Model", + "bbox": [ + 225, + 92, + 761, + 108 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15150", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c6400e6808829343ebe8bff1b8ea156dad0faa011c49ed265b4bc83db3d484f0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Summary of our data for training. R.: Reasoning, E.: Editing, G.: Grounding. #Img/Vid: the number of images/videos, and #Ins.: the number of instruction-image/video pairs." + ], + "table_body": "
TypeSourceR.E.G.# Img./Vid.# Ins.
VideoROVI [66]✓✓✓4.3K27.4K
VPLM [75]✓4.3K5.5K
GroundMoRe [12]✓✓1.3K5.5K
RVoS [60]✓1.9K6.1K
MeViS [14]✓1.5K17.1K
InstructV2V [9]✓68.3K68.3K
VEG-Edit (Ours)✓4.0K6.2K
Total136.1K
ImageSeed-Data-Edit [18]✓3M3M
LISA [37]✓✓0.2K1.3K
gRefCoCo [44]✓13.6K73.4K
PhraseCut [65]✓310.8K310.8K
EraseDraw [5]✓64.9K42.4K
MagicBrush [83]✓9.3K9.3K
SmartEdit [29]✓0.5K0.9K
Total3438.1K
", + "bbox": [ + 93, + 88, + 478, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Enhancing Temporal Consistency and Dynamics. With the MLLM and diffusion model aligned, fine-tuning diverse instructional video editing data becomes more effective for improved instruction following at pixel-space including temporal consistency, dynamic coherence, and editing faithfulness. In this stage, we fine-tune the framework with the MLLM, including the alignment network, grounded task queries, and all 3 dimensions in diffusion UNet, end-to-end with carefully curated multitasking instructional video editing data. Following prior work [22, 67], we inflated the 2D UNet from Stage 1 with temporal attention layers for video adaptation. For both stages 1 and 2, we optimize the framework with a single diffusion loss, enabling unified learning for improved instructional video editing performance while maintaining simplicity and efficiency.", + "bbox": [ + 88, + 383, + 485, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classifier-Free Guidance during Testing. We employ classifier-free guidance to balance quality and diversity in diffusion-generated samples. Following prior work [3, 17], we apply classifier-free guidance to instructional visual editing considering two conditions: the grounded task tokens and the original video. To obtain unconditional guidance, we set null values $(\\varnothing)$ for both task tokens and input video. In this case, our score estimate is:", + "bbox": [ + 88, + 614, + 486, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {e _ {\\theta}} (z _ {t}, c _ {T}, c _ {V}) = e _ {\\theta} (z _ {t}, \\varnothing , \\varnothing) \\\\ + g _ {T} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, c _ {T}\\right) - e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right)\\right) \\\\ + g _ {V} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right) - e _ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 750, + 472, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ represents the model parameters, $C_T$ and $C_V$ denote the task tokens and video conditions, $\\varnothing$ is the null value, $z_t$ is the noised latent at timestamp $t$ , and $g_T$ and $g_V$ are the task guidance and video guidance scales, respectively. More training details are included later in Appendix.", + "bbox": [ + 88, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Data Curation Pipeline", + "text_level": 1, + "bbox": [ + 513, + 90, + 728, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing video editing models, both instructional and non-instructional, struggle with diverse editing skills due to the lack of high-quality multitasking fine-tuning data. In this section, we introduce our data curation strategy to support VEGGIE in achieving versatile video editing skills. As listed in Tab. 1, we collect 3.4M image and 133.9K video data from diverse sources to support our VEGGIE curriculum learning as discussed in Sec. 3.2. We create our training dataset from two sources: (1) collecting existing image and video data and converting it into an instructional editing format, and (2) synthesizing new instructional video editing samples using existing datasets and generative models.", + "bbox": [ + 509, + 112, + 906, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collecting Diverse Multitask Image and Video Data. We bring together instructional editing data from both image (Seed-Data-Edit [18], MagicBrush [83], EraseDraw [5]) and video (InstructV2V [9], VPLM [75]) sources. These datasets provide pairs of original and edited visual contents with user instructions. The tasks include adding, removing, and changing objects, stylizing, and performing global/local edits. Beyond editing datasets, we incorporate segmentation data at both the image level (gRefCoCo [44] and Phrase-Cut [65]) and the video level (RVoS and MeViS). These segmentation tasks are reformulated as color-filling challenges, which guide the model in learning referring grounding (i.e., understanding which object or region to edit) and strengthen its conceptual learning. To further unlock complex instruction understanding via MLLM, we include data that requires more advanced reasoning and implicit referencing. Specifically, we include: reasoning segmentation (LISA [37]), reasoning editing (SmartEdit [29]), interactive video inpainting (LGVI [66]), and motion-grounded video reasoning (GroundMoRe [12]). These tasks help VEGGIE learn implicit references and reasoning.", + "bbox": [ + 509, + 295, + 908, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthesizing Instructional Video Editing Data via Image-to-Video Animation. Recent methods [3, 54] generate synthetic instructional video-editing data by first creating text instructions with LLM, then getting edited videos via T2V models and prompt-to-prompt editing [23]. While these methods adapt image-based editing pipelines [9] to videos, the generated data suffer from temporal-consistency issues. To address this gap, we propose a novel image-to-video animation strategy that leverages the abundance of high-quality image-level instructional editing datasets [64, 83], which provide well-annotated instructions, paired edited images, and well-organized editing skill categories. As illustrated in Fig. 5, given an original image $I$ , an edited image $\\bar{I}$ , and an instruction from an instructional image editing dataset [64], our approach involves three key steps. First, we use an offline MLLM [61, 72] to generate an image caption and an animation prompt that describes plausible motion within the image. Next, an image-to-video (I2V) model animates the image into a video $V$ . Finally,", + "bbox": [ + 509, + 613, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/519aa3cbb726c3f46600125ec3400403ec4db3182b9494a57667d8cbec97125a.jpg", + "image_caption": [ + "Figure 5. Our data generation pipeline for synthetic instructional video editing data. It injects dynamics into well-constructed instructional image editing datasets via the Image-to-Video (I2V) Model, and generates paired video data for instruction editing." + ], + "image_footnote": [], + "bbox": [ + 119, + 87, + 450, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we generate the corresponding edited video $\\bar{V}$ using a first-frame-conditioned video editing model [35], leveraging $\\bar{I}$ as a strong prior to ensure consistent edits across frames. Finally, to ensure data quality, we evaluate each original-edited video pair with automatic video quality evaluation metrics [30], which assess the generated videos from diverse dimensions, e.g., motion smoothness, image quality, and background consistency. This pipeline transforms carefully curated image-based datasets into instructional video-editing resources while preserving the precision of the original edits. As a result, our data method expands the availability of high-quality synthetic video-editing data, supporting a wider range of editing tasks in our end-to-end unified framework. More details on data generation, prompting, examples, and pre/post-processing are in the Appendix.", + "bbox": [ + 88, + 430, + 483, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 672, + 223, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first introduce the VEG-Bench Benchmark and then demonstrate the superiority of VEGGIE across diverse video instructional editing skills. More experiments, visualization, and implementation details are in the Appendix.", + "bbox": [ + 89, + 699, + 485, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. VEG-Bench and Metrics", + "text_level": 1, + "bbox": [ + 89, + 772, + 313, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As no existing benchmark is designed for fine-grained instructional video editing skills, we manually collect and annotate VEG-Bench, containing 132 video-instruction pairs that balanced cover 8 different video generative skills (15-20 for each). Beyond standard metrics, including text-to-video alignment (CLIP-Text [55]), video smoothness (CLIP-F [55]), and image quality (MUSIQ [32]), we also", + "bbox": [ + 89, + 794, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "first introduce MLLM-as-a-Judge to give a holistic evaluation score according to the given original video, edited video, and user instruction. It is achieved by prompting GPT4o [21] to evaluate whether the requested semantic change has been fulfilled, using a scale from 1 to 10. For addition and removal, we also introduce an object detector (GroundingDiNo [45]) to detect if the object is added/removed faithfully. For grounding and reasoning segmentation, we following video grounding tasks [12, 14, 33, 58] and adopt the Jaccard index $(\\mathcal{I})$ [31], F-measure $(\\mathcal{F})$ [13], and their mean $(\\mathcal{J} \\& \\mathcal{F})$ . We also compute SSIM between the generated video and the original video masked with GT masks. More evaluation/metrics details are included in Appendix.", + "bbox": [ + 511, + 90, + 908, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 300, + 718, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instructional Video Editing over Diverse Skills. As shown in Tab. 2, we evaluate 7 different models on VEG-Bench across 8 distinct editing skills. Overall, VEGGIE demonstrates the best performance among instructional video editing models. Compared to VEGGIE, non-instructional models often struggle with concept removal and addition. This limitation arises because these models rely on attention control or additional conditions (e.g., depth maps) that impose strong priors, constraining the model and making object addition or removal challenging. We also observe that InsV2V achieves high scores in quality and smoothness metrics, but underperforms in alignment and MLLM judgment, which demand faithful semantic changes. Qualitative examples in Fig. 6 illustrate that InsV2V often makes minimal changes to the input video, resulting in high video quality but unfaithful outputs. In contrast, VEGGIE strikes a better balance, delivering both high-quality visuals and accurate semantic alignment with the intended edits.", + "bbox": [ + 511, + 323, + 908, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Can Multi-Task Help Each Other? To test the previous hypothesis, we train our model on the VPLM [75] dataset, which includes paired grounding and removal tasks (approximately 5.5K samples for each task). We focus on these tasks as representative examples due to their straightforward evaluation against ground truth. As shown in Table 3, multitask training yields a lower FVD score and a higher SSIM score, demonstrating that learning to locate and remove a video concept can mutually reinforce performance. We show an example in Fig. 7. However, this conclusion only holds with a balanced data combination. We also observe that an excessive amount of grounding data can introduce more artifacts and negatively impact visual editing skills.", + "bbox": [ + 511, + 597, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Emergent Zero-shot Multimodal Instruction Following. We also highlight the emergent behavior of VEGGIE on multi-modal instruction following, even without dedicated training data for this specific editing instruction. Notably, VEGGIE demonstrates the ability to perform zero-shot multimodal instructional video editing. As illustrated in Fig. 2, VEGGIE can transfer styles or add objects from a reference", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15152", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/448541813648206c696b8060be6836952b32cd5e03312d900e585ca11af2f430.jpg", + "image_caption": [ + "[Addition] Please add a ball in the given video frames." + ], + "image_footnote": [], + "bbox": [ + 114, + 99, + 367, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67afd19caa3fbcb6b392c5ad751d2b0f2e3161441709a94940152b6d86723f7e.jpg", + "image_caption": [ + "[Removal] Please remove the man in black in given video frames." + ], + "image_footnote": [], + "bbox": [ + 370, + 99, + 625, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/373ec45a7f7d566ba4d3a78d8e06b09d7bad2b84179f4cdf62f91a27bbbf530a.jpg", + "image_caption": [ + "[Swap] Replace golden building with a white mountain." + ], + "image_footnote": [], + "bbox": [ + 625, + 99, + 880, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41198c72d6faf22421eb58cefcc01dfd03a7dbb2d35994a01bd56dc226535d45.jpg", + "image_caption": [ + "[Environment] Make it on the beach." + ], + "image_footnote": [], + "bbox": [ + 114, + 349, + 367, + 589 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2aaed0e07fec7c8843e49de92130946b5b89bbd966dcb8e209bd72feffe8789e.jpg", + "image_caption": [ + "[Color] Make the swan white." + ], + "image_footnote": [], + "bbox": [ + 370, + 349, + 625, + 589 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a238051c85810f38dffc43dcef81eb1d2021aded79c1a77a3508a3bafc592ab6.jpg", + "image_caption": [ + "[Texture] Make the rhinocero furry." + ], + "image_footnote": [], + "bbox": [ + 625, + 349, + 880, + 589 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a41b7f9d6e96f7f28878ef89d70835e2d12c0c4a44228a64db4c7f7f34d6c0d5.jpg", + "image_caption": [ + "[Style] Make it chinese ink style.", + "Figure 6. Qualitative comparison of editing results across 8 different abilities (splitting visual features into color and texture). We provide zoom-in details for a more detailed comparison. Best viewed in color. More in Appendix." + ], + "image_footnote": [], + "bbox": [ + 114, + 606, + 367, + 845 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fecf891dfe80f439d7ac4732ec66bb78527fab1d6f6d18a6235978f3c780c18.jpg", + "image_caption": [ + "[Grounding] Could you locate the knife" + ], + "image_footnote": [], + "bbox": [ + 370, + 599, + 625, + 845 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7c113e68f1b140f6b2fd89134382a5e2e6aa73949f6da185815fa51ef01fdde.jpg", + "image_caption": [ + "[Reasoning] What can be used for heating food?" + ], + "image_footnote": [], + "bbox": [ + 625, + 599, + 880, + 845 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15153", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24831cb68b0c27cf147e918fb58dc3a60803aa169edf93235ef0f987dc167bca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods\nGeneratorNon-Instructional Editing ModelInstructional Editing Model
VidToMe [41] (SD1.5)TokenFlow [20] (SD2.1)Flatten [10] (SD2.1)InstructDiff [19] (SD1.5)LGVI [66] (SD1.5)InsV2V [9] (SD1.5)VEGGIE (Ours) (SD1.5)
Concept Addition
MLLM-Judge (↑)5.005.806.627.262.735.697.44
Alignment (↑)27.8029.3028.2228.1025.0628.2729.27
Smoothness (↑)96.7997.2695.7493.6696.0996.9494.93
Quality (↑)62.8165.6255.4554.0841.5956.2461.31
Detection (↑)47.9849.5349.7455.3614.4248.0157.96
Concept Removal
MLLM-Judge (↑)2.603.734.466.126.592.785.07
Alignment (↑)75.0175.9978.4075.5175.6774.4175.63
Smoothness (↑)96.1396.4795.8291.8397.0396.9995.04
Quality (↑)66.3271.5250.7755.0842.3158.7950.99
Detection (↑)34.3155.1670.9164.8178.4025.6470.22
Object Changing
MLLM-Judge (↑)5.006.537.377.002.066.606.63
Alignment (↑)25.6928.7627.0627.3622.1726.6027.77
Smoothness (↑)96.2397.2196.1392.0795.6696.7495.44
Quality (↑)64.0669.9759.3755.0138.2060.9058.15
Environment & Background Changing
MLLM-Judge (↑)5.817.357.376.052.376.607.18
Alignment (↑)28.1730.0030.0428.0321.9428.2729.15
Smoothness (↑)95.7696.9695.9089.8595.6696.0394.58
Quality (↑)61.9567.0654.5853.0638.9754.9454.25
Visual Feature Changing (Color & Texture)
MLLM-Judge (↑)5.866.856.606.432.147.537.33
Alignment (↑)27.9929.2529.4627.5423.1828.8828.69
Smoothness (↑)95.9397.1095.8391.7194.7596.6694.52
Quality (↑)65.8069.3153.3258.2936.2759.3657.91
Stylization
MLLM-Judge (↑)7.237.628.317.413.718.078.26
Alignment (↑)29.8430.2529.0027.7422.8029.1429.38
Smoothness (↑)96.3197.2396.7188.9795.6296.5095.69
Quality (↑)64.0568.2253.1854.1535.7662.5957.00
Object Grounding
SSIM (↑)40.4750.4647.2137.9866.8449.6570.90
Jaccard Index J (↑)13.8519.2925.6219.881.5213.8937.74
F-measure F (↑)15.5016.8617.6012.813.0717.3721.83
Reasoning Segmentation
SSIM (↑)---32.3944.4759.8668.41
Jaccard Index J (↑)---14.0210.1216.8922.53
F-measure F (↑)---8.079.0610.4515.97
Avg. Ranking2.611.411.963.003.212.001.78
", + "bbox": [ + 148, + 88, + 841, + 512 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b625bcc9addcaa022aa6854cf74a6df0f31f087476d9e9b0190a15cb855a15d.jpg", + "table_caption": [ + "Table 2. Comparison of video editing task with instructional / non-instructional models on VEG-Bench. -: the task is not capable of non-instructional models. We gray out numbers of non-instructional models that are in different categories." + ], + "table_footnote": [], + "table_body": "
SettingsRemoval (FVD ↓)Grounding (SSIM ↑)
Grd.-only-52.34
Rmv.-only1098.52-
Mixed987.8055.21
", + "bbox": [ + 117, + 556, + 464, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. An ablation study on whether multi-task learning provides transferable benefits that enhance performance across tasks. We focus on removal and grounding tasks as representative examples.", + "bbox": [ + 89, + 632, + 482, + 674 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16dd7320589fd8681a64e1346642e6554bace0aed5cea65f14131a25d7187071.jpg", + "image_caption": [ + "Figure 7. Comparison between single- and multi-skill models with different data training. We find tasks can help each other." + ], + "image_footnote": [], + "bbox": [ + 114, + 686, + 460, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "image into the input video based on instructions.", + "bbox": [ + 511, + 559, + 834, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Emergent Few-shot In-Context Editing. As shown in Fig. 3, VEGGIE can effectively utilize a few example image pairs to transfer the intended editing changes seamlessly to the input video. We observe that VEGGIE exhibits in-context learning for image editing without the need for language instructions. Instead, it uses image pairs as examples to infer and apply the desired editing intention directly.", + "bbox": [ + 511, + 574, + 906, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present VEGGIE, a unified end-to-end model for instructional video editing that handles diverse pixel-level tasks. VEGGIE leverages MLLM for robust instruction understanding and employs a video diffusion model to execute pixel-level edits. Our framework uses a single diffusion loss for end-to-end optimization across varied tasks/skills. We also introduce a novel synthetic data generation pipeline and VEG-Bench, a benchmark that assesses a broad range of editing skills. Our VEGGIE outperforms previous methods as a versatile, all-in-one solution. We hope our model, data, and benchmark to advance research on instructional generative video models.", + "bbox": [ + 511, + 719, + 908, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15154", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 2", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 5", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI, https://openai.com/research/video-generation-models-as-world-simulators, 2024.2.3", + "[5] Alper Canberk, Maksym Bondarenko, Ege Ozguroglu, Ruoshi Liu, and Carl Vondrick. Erasedraw: Learning to insert objects by erasing them from images. arXiv preprint arXiv:2409.00522, 2024. 5", + "[6] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3", + "[7] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 3", + "[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3", + "[9] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In The Twelfth International Conference on Learning Representations, 2024. 5, 8, 1, 2", + "[10] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENtion for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 8, 1", + "[11] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Mohit Bansal, and Chen Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. arXiv preprint arXiv:2411.09921, 2024. 3", + "[12] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Bansal Mohit, and Chen. Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. 2024. 5, 6, 2" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 6", + "[14] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 5, 6", + "[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In CVPR, pages 7346-7356, 2023. 3", + "[16] Hao Fei, Shengqiong Wu, Hanwang Zhang, Tat-Seng Chua, and Shuicheng Yan. Vitron: A unified pixel-level vision llm for understanding, generating, segmenting, editing, 2024. 2", + "[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 3, 5", + "[18] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 5", + "[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 8, 1, 2", + "[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2, 3, 8, 1", + "[21] gpt 4o. https://openai.com/index/hello-gpt-4o/.2024.6", + "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 5, 1", + "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.5", + "[24] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 2022. 2", + "[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Jiahao Hu, Tianxiong Zhong, Xuebo Wang, Boyuan Jiang, Xingye Tian, Fei Yang, Pengfei Wan, and Di Zhang. Vivid-10m: A dataset and baseline for versatile and interactive video local editing. arXiv preprint arXiv:2411.15260, 2024. 2", + "[29] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 5", + "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 2", + "[31] Paul Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37-50, 1912. 6", + "[32] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6", + "[33] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 6", + "[34] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2", + "[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 6", + "[36] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3", + "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 5", + "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1", + "[39] Jialu Li, Shoubin Yu, Han Lin, Jaemin Cho, Jaehong Yoon, and Mohit Bansal. Training-free guidance in text-to-video generation via multimodal planning and structured noise initialization. arXiv preprint arXiv:2504.08641, 2025. 2", + "[40] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3", + "[41] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 2, 8, 1", + "[42] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2", + "[43] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2", + "[44] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. 3, 5", + "[45] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6", + "[46] Shaoteng Liu, Tianyu Wang, Jui-Hsien Wang, Qing Liu, Zhifei Zhang, Joon-Young Lee, Yijun Li, Bei Yu, Zhe Lin, Soo Ye Kim, et al. Generative video propagation. arXiv preprint arXiv:2412.19761, 2024. 2", + "[47] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 3", + "[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3", + "[49] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 9117-9125, 2023. 3", + "[50] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, Mubarak Shah, and Ajmal Mian. Towards temporally consistent referring video object segmentation. https://arxiv.org/abs/2403.19407, 2024. 2", + "[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[52] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. Detgpt: Detect what you need via reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14172-14189, 2023. 3", + "[53] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15156", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + " attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3", + "[54] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 3, 5", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6", + "[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[58] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6", + "[59] Shangkun Sun, Xiaoyu Liang, Songlin Fan, Wenxu Gao, and Wei Gao. Ve-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2", + "[60] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 5", + "[61] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 2", + "[62] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[63] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal IIm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 2", + "[64] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image edit" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 5, 2", + "[65] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5", + "[66] Jianzong Wu, Xiangtai Li, Chenyang Si, Shangchen Zhou, Jingkang Yang, Jiangning Zhang, Yining Li, Kai Chen, Yunhai Tong, Ziwei Liu, et al. Towards language-driven video inpainting via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12501-12511, 2024. 2, 3, 5, 8, 1", + "[67] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 5", + "[68] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2", + "[69] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[70] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3", + "[71] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3", + "[72] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5", + "[73] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1", + "[74] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Jaehong Yoon, Shoubin Yu, and Mohit Bansal. Raccoon: Remove, add, and change video content with auto-generated narratives. arXiv preprint arXiv:2405.18406, 2024. 2, 5, 6", + "[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations, 2023. 3", + "[77] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3", + "[78] Shoubin Yu, Jacob Zhiyuan Fang, Jian Zheng, Gunnar Sigurdsson, Vicente Ordonez, Robinson Piramuthu, and Mohit Bansal. Zero-shot controllable image-to-video animation via motion decomposition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3332-3341, 2024. 2", + "[79] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2, 3", + "[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 1", + "[81] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949, 2023. 3", + "[82] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferret-v2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 3", + "[83] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 1", + "[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[85] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. arXiv preprint arXiv:2312.03816, 2023. 2, 3", + "[86] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 2, 3", + "[87] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 3" + ], + "bbox": [ + 91, + 92, + 483, + 897 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15158", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_model.json b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0da94a368e5c85fc273a0833099a4c9c46d1e86b --- /dev/null +++ b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_model.json @@ -0,0 +1,2413 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.13, + 0.749, + 0.177 + ], + "angle": 0, + "content": "VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.203, + 0.795, + 0.244 + ], + "angle": 0, + "content": "Shoubin Yu\\(^{1,3*}\\) Difan Liu\\(^{1*}\\) Ziqiao Ma\\(^{1,2*}\\) Yicong Hong\\(^{1}\\) Yang Zhou\\(^{1}\\) Hao Tan\\(^{1}\\) Joyce Chai\\(^{2}\\) Mohit Bansal\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.256, + 0.762, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Adobe Research \\(^{2}\\)University of Michigan \\(^{3}\\)UNC Chapel Hill" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.277, + 0.642, + 0.293 + ], + "angle": 0, + "content": "https://veggie-gen.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.31, + 0.859, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.565, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 1. We propose VEGGIE, a unified and versatile video generative model that handles various tasks for both video concept grounding and editing according to user instructions. With VEGGIE, users can locate, add, delete, and change concepts in a given video through diverse instruction formats (direct referring instruction or reasoning-demanding questions). Users can also edit videos with multimodal instruction empowered by MLLM, enabling applications like video editing from a reference image." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.636, + 0.327, + 0.652 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.668, + 0.486, + 0.865 + ], + "angle": 0, + "content": "While recent video diffusion models enable video editing, unifying diverse instructional editing tasks (e.g., add, remove, modify) under a single framework remains a significant challenge. In this paper, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions, a simple end-to-end framework that unifies video concept editing, grounding, and reasoning based on diverse user instructions. Specifically, given a video and text query, VEGGIE first utilizes an MLLM to interpret user intentions in instructions and ground them to the video contexts, generating framespecific grounded task queries for pixel-space responses. A diffusion model then renders these plans and generates edited videos that align with user intent. To support diverse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.909, + 0.895 + ], + "angle": 0, + "content": "tasks and complex instructions, we employ a curriculum learning strategy: first aligning the MLLM and video diffusion model with large-scale instructional image editing data, followed by end-to-end fine-tuning on high-quality multitask video data. Additionally, we introduce a novel data synthesis pipeline to generate paired instructional video editing data for model training. It transforms static image data into diverse, high-quality video editing samples by leveraging Image-to-Video models to inject dynamics. VEGGIE shows strong performance in instructional video editing with different editing skills, outperforming the best instructional baseline as a versatile model, while other models struggle with multi-tasking. VEGGIE also excels in video object grounding and reasoning segmentation, where other baselines fail. We further reveal how the multiple tasks help each other and highlight promising applications like zero-shot multimodal instructional and in-context video editing." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.22, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.374, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done during internship at Adobe Research." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.374, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15147" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.107 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.297 + ], + "angle": 0, + "content": "Building on the advances in Video Diffusion Models (VIDDMs) [2, 4, 24, 25, 74], video editing methods have emerged as video design tools, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 86] have been developed, using triples of text prompts, source videos, and target videos for training. Due to their limited performance in understanding user intent and multimodal semantics [17], several methods have incorporated multimodal large language models (MLLMs) to handle complex instructions/reasoning [17, 29, 75]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.486, + 0.584 + ], + "angle": 0, + "content": "However, existing methods fall short of the goal of a simple, versatile video concept editor, facing three primary challenges. First, most methods are not end-to-end, requiring intermediate layout/mask/human or model caption guidance [35, 53, 63, 75], which adds workload on users and disrupts a seamless editing experience. Second, existing pipelines connecting MLLMs to VidDMs require multiple training objectives beyond simple pixel-space diffusion loss, such as language loss [75] or mask losses [66]. This increases optimization difficulty and often requires additional hyperparameter tuning or annotations. Third, existing video editing models, both instructional and non-instructional, struggle with handling other diverse editing tasks, ranging from addition, and deletion to stylization. For example, LGVI [66] fails in global edits such as stylization and color change, while VidToMe [41] struggles with local edits such as adding or removing objects. These methods also struggle with input videos that contain multiple objects or when user instructions require complex reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.486, + 0.765 + ], + "angle": 0, + "content": "These challenges result from two limitations: First, there is a lack of multitasking fine-tuning on well-curated instructional video editing datasets that span a broad range of skills. Second, models often lack two critical capabilities needed to interpret user intentions and accurately locate concepts: multimodal reasoning to infer the intended modification from the user's instruction; and grounding language to the input video to precisely identify the region or object to be edited. For example, in Figure 1, one can effortlessly locate the girl given \"identify the little girl.\" When asked to \"add a hat to the little girl,\" we intuitively imagine the hat placed on her head from commonsense, even without seeing an actual hat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions. VEGGIE unifies video concept grounding and editing without relying on additional layout, mask guidance, or intermediate caption [35, 39, 42, 75, 78, 79]. Instead, we formulate the problem as end-to-end grounded generation in pixel space, using only a diffusion loss. Specifically, given a video and a text query, VEGGIE first leverages an MLLM to interpret complex instructions, generating frame-wise con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.379 + ], + "angle": 0, + "content": "ditions. Unlike prior methods [43, 75] that use discrete text tokens as conditions, which disconnect the pipeline and block gradient propagation, VEGGIE employs continuous, learnable task query embeddings per frame. This enables end-to-end training and effectively captures grounded task representations for diffusion model conditioning. To handle diverse tasks and accurately interpret complex queries, we employ a curriculum learning strategy that begins by aligning MLLMs with diffusion models using massive paired instructional image editing data and then fine-tuning the model end-to-end on high-quality multitask video data to adapt video. Unlike tool-use methods [16, 35, 75], VEGGIE formulates both video grounding and instructional editing in the same video-to-video task formulation, enabling efficient handling through a unified single model. To further support end-to-end training, we introduce a novel automatic instructional video data generation pipeline that lifts high-quality instructional image editing data into the video domain using image-to-video and video evaluation tools." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Existing video editing benchmarks do not provide wide and uniform coverage of diverse editing skills [59, 68]. To address this gap, we contribute VEG-Bench, an instructional video editing benchmark that spans 8 editing skills: concept addition, removal, object changing, environment background changing, visual feature changing, stylization, object grounding, and reasoning segmentation. Each skill is evaluated using a dedicated suite of metrics. We assess the proposed VEGGIE alongside 6 baselines on VEG-Bench. VEGGIE demonstrates strong performance across diverse editing skills, outperforming the best instructional baseline as a versatile, all-in-one model, while other models struggle with multi-tasking. Additionally, VEGGIE excels in video object grounding and reasoning segmentation tasks, where other baselines fall short. We show further analysis of how multi-task learning enhances our framework and highlight applications such as zero-shot multimodal instructional following (Fig. 2) and few-shot in-context editing (Fig. 3). Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.675, + 0.907, + 0.78 + ], + "angle": 0, + "content": "- We propose VEGGIE, an end-to-end model that integrates an MLLM and a VidDM. VEGGIE is a versatile framework that handles diverse instructional requests for editing and grounding various video concepts. Unlike existing work that achieves multitasking via tool use, VEGGIE unifies diverse tasks in a single model, thus simplifying the training with only diffusion loss." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.781, + 0.905, + 0.81 + ], + "angle": 0, + "content": "- We propose a data synthesis pipeline, scaling high-quality instructional video editing data for future work." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.811, + 0.907, + 0.855 + ], + "angle": 0, + "content": "- We propose VEG-Bench, an instructional video editing benchmark that spans 8 editing skills with dedicated metrics for each skill." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.857, + 0.909, + 0.901 + ], + "angle": 0, + "content": "- VEGGIE achieves strong performance across diverse editing skills compared with SoTA methods, and shows potentials for multimodal instruction and in-context following." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.09, + 0.245, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.145, + 0.239, + 0.154 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.162, + 0.242, + 0.21 + ], + "angle": 0, + "content": "Transfer the style in the reference image." + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.091, + 0.495, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.102, + 0.583, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.143, + 0.581, + 0.153 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.091, + 0.832, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.155, + 0.578, + 0.216 + ], + "angle": 0, + "content": "Add the object in the reference image on the women." + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.231, + 0.892, + 0.246 + ], + "angle": 0, + "content": "Figure 2. Multimodal instruction following emerges in VEGGIE, allowing for style transfer or object addition from reference images." + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.263, + 0.832, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.41, + 0.885, + 0.424 + ], + "angle": 0, + "content": "Figure 3. In-context editing emerges in VEGGIE, allowing for few-shot learning of editing tasks with paired image demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.449, + 0.233, + 0.464 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Instructional Video Editing Video Diffusion Models (VidDMs) [2, 4, 6, 7, 15, 26, 49, 67, 71], enable high-quality video generation across a wide range of video concepts. Building on these advances, video editing methods have emerged as tools for video design, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 70, 86] have been developed, using triples of video instructions, source videos, and target videos for training. These methods demonstrate limited performance when complex multimodal reasoning is required, as noted by previous research on instructional image editing [17]. Moreover, they struggle with diverse editing tasks, from addition and deletion to stylization. For example, LGVI [66] is primarily designed for removal tasks, while TokenFlow [20] struggles with local edits such as adding, removing, or changing objects. We address this limitation with pixel-level multitasking fine-tuning on well-curated instructional video editing datasets covering various grounding and editing skills." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Video Grounding and Segmentation Visual grounding requires models to connect language to its corresponding visual concept in the visual context [40, 47]. This is commonly evaluated via the language-guided semantic localization tasks, ranging from simple referring expressions in RefCOCO series [48, 77] and their generalized variant [44] that takes no-target and multi-target into account. Recently, grounded multimodal large lan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.908, + 0.587 + ], + "angle": 0, + "content": "guage models (MLLMs) are trained for object grounding to bounding boxes [8, 51, 52, 76, 82, 87] and segmentation masks [56, 69, 81, 84] using text-image pairs with fine-grained annotations linking phrases to entities. These models unlock the potential of reasoning segmentation [11, 36], bringing language-informed reasoning into semantic segmentation. Instead of using dedicated object detection or segmentation modules, we achieve video grounding through end-to-end training with only diffusion loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.727, + 0.616 + ], + "angle": 0, + "content": "3. Our Method: VEGGIE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.838 + ], + "angle": 0, + "content": "In this paper, we introduce a Video Editor with Grounded Generation from Instructions (VEGGIE), a unified and versatile generative video model. It combines the complex instruction understanding and reasoning capabilities of MLLMs with the generative capacity of VidDMs. The model is trained end-to-end with diffusion loss only. VEGGIE efficiently handles diverse user inputs, including direct instructions, complex questions requiring in-depth reasoning, and multimodal conditioning. It performs various pixel-level manipulations, enabling tasks such as video concept addition, removal, changing, stylization, grounding, and reasoning segmentation based on user instructions. We elaborate on the model design (Sec. 3.1), training and inference process (Sec. 3.2), and data curation (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.702, + 0.863 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "VEGGIE consists of four main components (see Fig. 4): (1) a multimodal large language model, (2) a set of learnable" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15149" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.202, + 0.093, + 0.226, + 0.11 + ], + "angle": 0, + "content": null + }, + { + "type": "header", + "bbox": [ + 0.226, + 0.093, + 0.763, + 0.109 + ], + "angle": 0, + "content": "VEGGIE: A Unified and Versatile Instructional Video Generative Model" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.119, + 0.885, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.362, + 0.908, + 0.418 + ], + "angle": 0, + "content": "Figure 4. Overview of our proposed end-to-end VEGGIE framework. Our Multimodal Large Language Model first understands input video frames and diverse user instructions, then it generates frame-wise reasoning queries that maintain per-frame editing conditions for the video diffusion model. The video diffusion model will render the MLLM-generated conditions to the pixel space for diverse tasks, including video editing, video grounding, and video reasoning segmentation with questions. We only apply diffusion loss for the whole pipeline training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.486, + 0.564 + ], + "angle": 0, + "content": "grounded task queries, (3) an alignment network (single-layer MLP) that projects the MLLM output into the condition space of the diffusion model, and (4) a video diffusion model initialized from an instructional image editing model [83]. Our model first generates latent conditions for target video frames by querying multimodal context using an MLLM, then renders these conditions at the pixel level through a video diffusion model, as detailed below." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.483, + 0.779 + ], + "angle": 0, + "content": "MLLM for Generating Grounded Task Guidance. As illustrated in the left of Fig. 4, given a video consisting of a sequence of frames \\( V = [f_1, \\dots, f_n] \\), where \\( n \\) is the frame number of the given video, a user instruction/question \\( I \\), our goal is to obtain the response \\( \\widehat{V} = [\\widehat{f}_1, \\dots, \\widehat{f}_n] \\) at pixel space that faithfully reflects user instruction about the given video. The MLLM module processes both the input video \\( V \\) and a user instruction \\( I \\) to generate a sequence of grounded task tokens per frame: \\( C = [c_1, \\dots, c_n] \\), which are input and output in parallel. These tokens serve as task guidance and implicitly encode the target manipulation, such as object attributes, spatial relationships, or style transfer parameters. The MLLM ensures the model captures both explicit user instructions and implicit reasoning needs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "VidDM for Rendering MLLM Guidance at Pixel Space. As illustrated in the right of Table 4, the VidDM takes the original video \\( V \\) and the grounded task tokens \\( C \\) as conditions to synthesize the target video \\( \\widehat{V} \\). The original video is concatenated with the noise volume, and the task tokens are input to the cross-attention. With grounded task guidance in denoising steps, the generation process ensures that the output faithfully follows user instructions while preserving the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.505 + ], + "angle": 0, + "content": "video's structure and motion dynamics. Through iterative denoising, it refines each frame while maintaining temporal consistency, applying pixel modifications coherently for a smooth and visually consistent output video \\(\\widehat{V}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.895, + 0.532 + ], + "angle": 0, + "content": "3.2. Curriculum Learning from Images to Videos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.908, + 0.749 + ], + "angle": 0, + "content": "Training the model directly on video tasks presents two key challenges: (1) misalignment between MLLM and diffusion model representations, making it difficult for the diffusion model to interpret MLLM-generated task queries with limited fine-tuning data, and (2) the diffusion model's lack of multitasking capability, even for image tasks, due to insufficient training on diverse tasks. Our initial experiments also found the model collapsed when the whole pipeline was directly trained with all data. These challenges/observations underscore the need for pre-alignment between MLLM and the diffusion model to enable seamless adaptation from language-space task queries to pixel-space modifications. To this end, we adopt a two-stage curriculum learning strategy for the proposed VEGGIE framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Stage 1: Aligning Diffusion and Language Spaces. In the first stage, we align the diffusion model with the MLLM using large-scale image-level instructional editing data. The MLLM remains frozen while we update the alignment network, grounded task queries, and diffusion UNet. This process fine-tunes the diffusion model weights to align with the language space, enabling the model to interpret MLLM-generated guidance and translate user instructions into pixel-level edits while preserving the MLLM's strong ability to understand instructions and user intentions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15150" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.48, + 0.312 + ], + "angle": 0, + "content": "
TypeSourceR.E.G.# Img./Vid.# Ins.
VideoROVI [66]✓✓✓4.3K27.4K
VPLM [75]✓4.3K5.5K
GroundMoRe [12]✓✓1.3K5.5K
RVoS [60]✓1.9K6.1K
MeViS [14]✓1.5K17.1K
InstructV2V [9]✓68.3K68.3K
VEG-Edit (Ours)✓4.0K6.2K
Total136.1K
ImageSeed-Data-Edit [18]✓3M3M
LISA [37]✓✓0.2K1.3K
gRefCoCo [44]✓13.6K73.4K
PhraseCut [65]✓310.8K310.8K
EraseDraw [5]✓64.9K42.4K
MagicBrush [83]✓9.3K9.3K
SmartEdit [29]✓0.5K0.9K
Total3438.1K
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.361 + ], + "angle": 0, + "content": "Table 1. Summary of our data for training. R.: Reasoning, E.: Editing, G.: Grounding. #Img/Vid: the number of images/videos, and #Ins.: the number of instruction-image/video pairs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.486, + 0.611 + ], + "angle": 0, + "content": "Stage 2: Enhancing Temporal Consistency and Dynamics. With the MLLM and diffusion model aligned, fine-tuning diverse instructional video editing data becomes more effective for improved instruction following at pixel-space including temporal consistency, dynamic coherence, and editing faithfulness. In this stage, we fine-tune the framework with the MLLM, including the alignment network, grounded task queries, and all 3 dimensions in diffusion UNet, end-to-end with carefully curated multitasking instructional video editing data. Following prior work [22, 67], we inflated the 2D UNet from Stage 1 with temporal attention layers for video adaptation. For both stages 1 and 2, we optimize the framework with a single diffusion loss, enabling unified learning for improved instructional video editing performance while maintaining simplicity and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.615, + 0.487, + 0.736 + ], + "angle": 0, + "content": "Classifier-Free Guidance during Testing. We employ classifier-free guidance to balance quality and diversity in diffusion-generated samples. Following prior work [3, 17], we apply classifier-free guidance to instructional visual editing considering two conditions: the grounded task tokens and the original video. To obtain unconditional guidance, we set null values \\((\\varnothing)\\) for both task tokens and input video. In this case, our score estimate is:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.751, + 0.473, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {e _ {\\theta}} (z _ {t}, c _ {T}, c _ {V}) = e _ {\\theta} (z _ {t}, \\varnothing , \\varnothing) \\\\ + g _ {T} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, c _ {T}\\right) - e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right)\\right) \\\\ + g _ {V} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right) - e _ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\theta\\) represents the model parameters, \\(C_T\\) and \\(C_V\\) denote the task tokens and video conditions, \\(\\varnothing\\) is the null value, \\(z_t\\) is the noised latent at timestamp \\(t\\), and \\(g_T\\) and \\(g_V\\) are the task guidance and video guidance scales, respectively. More training details are included later in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.729, + 0.108 + ], + "angle": 0, + "content": "3.3. Data Curation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.113, + 0.907, + 0.295 + ], + "angle": 0, + "content": "Existing video editing models, both instructional and non-instructional, struggle with diverse editing skills due to the lack of high-quality multitasking fine-tuning data. In this section, we introduce our data curation strategy to support VEGGIE in achieving versatile video editing skills. As listed in Tab. 1, we collect 3.4M image and 133.9K video data from diverse sources to support our VEGGIE curriculum learning as discussed in Sec. 3.2. We create our training dataset from two sources: (1) collecting existing image and video data and converting it into an instructional editing format, and (2) synthesizing new instructional video editing samples using existing datasets and generative models." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.296, + 0.909, + 0.614 + ], + "angle": 0, + "content": "Collecting Diverse Multitask Image and Video Data. We bring together instructional editing data from both image (Seed-Data-Edit [18], MagicBrush [83], EraseDraw [5]) and video (InstructV2V [9], VPLM [75]) sources. These datasets provide pairs of original and edited visual contents with user instructions. The tasks include adding, removing, and changing objects, stylizing, and performing global/local edits. Beyond editing datasets, we incorporate segmentation data at both the image level (gRefCoCo [44] and Phrase-Cut [65]) and the video level (RVoS and MeViS). These segmentation tasks are reformulated as color-filling challenges, which guide the model in learning referring grounding (i.e., understanding which object or region to edit) and strengthen its conceptual learning. To further unlock complex instruction understanding via MLLM, we include data that requires more advanced reasoning and implicit referencing. Specifically, we include: reasoning segmentation (LISA [37]), reasoning editing (SmartEdit [29]), interactive video inpainting (LGVI [66]), and motion-grounded video reasoning (GroundMoRe [12]). These tasks help VEGGIE learn implicit references and reasoning." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.614, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Synthesizing Instructional Video Editing Data via Image-to-Video Animation. Recent methods [3, 54] generate synthetic instructional video-editing data by first creating text instructions with LLM, then getting edited videos via T2V models and prompt-to-prompt editing [23]. While these methods adapt image-based editing pipelines [9] to videos, the generated data suffer from temporal-consistency issues. To address this gap, we propose a novel image-to-video animation strategy that leverages the abundance of high-quality image-level instructional editing datasets [64, 83], which provide well-annotated instructions, paired edited images, and well-organized editing skill categories. As illustrated in Fig. 5, given an original image \\( I \\), an edited image \\( \\bar{I} \\), and an instruction from an instructional image editing dataset [64], our approach involves three key steps. First, we use an offline MLLM [61, 72] to generate an image caption and an animation prompt that describes plausible motion within the image. Next, an image-to-video (I2V) model animates the image into a video \\( V \\). Finally," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15151" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.088, + 0.451, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.362, + 0.486, + 0.418 + ], + "angle": 0, + "content": "Figure 5. Our data generation pipeline for synthetic instructional video editing data. It injects dynamics into well-constructed instructional image editing datasets via the Image-to-Video (I2V) Model, and generates paired video data for instruction editing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.485, + 0.658 + ], + "angle": 0, + "content": "we generate the corresponding edited video \\(\\bar{V}\\) using a first-frame-conditioned video editing model [35], leveraging \\(\\bar{I}\\) as a strong prior to ensure consistent edits across frames. Finally, to ensure data quality, we evaluate each original-edited video pair with automatic video quality evaluation metrics [30], which assess the generated videos from diverse dimensions, e.g., motion smoothness, image quality, and background consistency. This pipeline transforms carefully curated image-based datasets into instructional video-editing resources while preserving the precision of the original edits. As a result, our data method expands the availability of high-quality synthetic video-editing data, supporting a wider range of editing tasks in our end-to-end unified framework. More details on data generation, prompting, examples, and pre/post-processing are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.674, + 0.224, + 0.691 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.486, + 0.762 + ], + "angle": 0, + "content": "We first introduce the VEG-Bench Benchmark and then demonstrate the superiority of VEGGIE across diverse video instructional editing skills. More experiments, visualization, and implementation details are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.314, + 0.789 + ], + "angle": 0, + "content": "4.1. VEG-Bench and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "As no existing benchmark is designed for fine-grained instructional video editing skills, we manually collect and annotate VEG-Bench, containing 132 video-instruction pairs that balanced cover 8 different video generative skills (15-20 for each). Beyond standard metrics, including text-to-video alignment (CLIP-Text [55]), video smoothness (CLIP-F [55]), and image quality (MUSIQ [32]), we also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.289 + ], + "angle": 0, + "content": "first introduce MLLM-as-a-Judge to give a holistic evaluation score according to the given original video, edited video, and user instruction. It is achieved by prompting GPT4o [21] to evaluate whether the requested semantic change has been fulfilled, using a scale from 1 to 10. For addition and removal, we also introduce an object detector (GroundingDiNo [45]) to detect if the object is added/removed faithfully. For grounding and reasoning segmentation, we following video grounding tasks [12, 14, 33, 58] and adopt the Jaccard index \\((\\mathcal{I})\\) [31], F-measure \\((\\mathcal{F})\\) [13], and their mean \\((\\mathcal{J} \\& \\mathcal{F})\\). We also compute SSIM between the generated video and the original video masked with GT masks. More evaluation/metrics details are included in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.301, + 0.719, + 0.317 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.909, + 0.596 + ], + "angle": 0, + "content": "Instructional Video Editing over Diverse Skills. As shown in Tab. 2, we evaluate 7 different models on VEG-Bench across 8 distinct editing skills. Overall, VEGGIE demonstrates the best performance among instructional video editing models. Compared to VEGGIE, non-instructional models often struggle with concept removal and addition. This limitation arises because these models rely on attention control or additional conditions (e.g., depth maps) that impose strong priors, constraining the model and making object addition or removal challenging. We also observe that InsV2V achieves high scores in quality and smoothness metrics, but underperforms in alignment and MLLM judgment, which demand faithful semantic changes. Qualitative examples in Fig. 6 illustrate that InsV2V often makes minimal changes to the input video, resulting in high video quality but unfaithful outputs. In contrast, VEGGIE strikes a better balance, delivering both high-quality visuals and accurate semantic alignment with the intended edits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Can Multi-Task Help Each Other? To test the previous hypothesis, we train our model on the VPLM [75] dataset, which includes paired grounding and removal tasks (approximately 5.5K samples for each task). We focus on these tasks as representative examples due to their straightforward evaluation against ground truth. As shown in Table 3, multitask training yields a lower FVD score and a higher SSIM score, demonstrating that learning to locate and remove a video concept can mutually reinforce performance. We show an example in Fig. 7. However, this conclusion only holds with a balanced data combination. We also observe that an excessive amount of grounding data can introduce more artifacts and negatively impact visual editing skills." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Emergent Zero-shot Multimodal Instruction Following. We also highlight the emergent behavior of VEGGIE on multi-modal instruction following, even without dedicated training data for this specific editing instruction. Notably, VEGGIE demonstrates the ability to perform zero-shot multimodal instructional video editing. As illustrated in Fig. 2, VEGGIE can transfer styles or add objects from a reference" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15152" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.091, + 0.35, + 0.1 + ], + "angle": 0, + "content": "[Addition] Please add a ball in the given video frames." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.1, + 0.368, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.092, + 0.627, + 0.1 + ], + "angle": 0, + "content": "[Removal] Please remove the man in black in given video frames." + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.1, + 0.626, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.092, + 0.865, + 0.1 + ], + "angle": 0, + "content": "[Swap] Replace golden building with a white mountain." + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.101, + 0.882, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.343, + 0.321, + 0.352 + ], + "angle": 0, + "content": "[Environment] Make it on the beach." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.351, + 0.369, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.343, + 0.562, + 0.352 + ], + "angle": 0, + "content": "[Color] Make the swan white." + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.351, + 0.626, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.343, + 0.831, + 0.351 + ], + "angle": 0, + "content": "[Texture] Make the rhinocero furry." + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.351, + 0.882, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.598, + 0.312, + 0.607 + ], + "angle": 0, + "content": "[Style] Make it chinese ink style." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.607, + 0.369, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.592, + 0.589, + 0.601 + ], + "angle": 0, + "content": "[Grounding] Could you locate the knife" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.601, + 0.626, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.592, + 0.863, + 0.601 + ], + "angle": 0, + "content": "[Reasoning] What can be used for heating food?" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.601, + 0.882, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.858, + 0.905, + 0.886 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of editing results across 8 different abilities (splitting visual features into color and texture). We provide zoom-in details for a more detailed comparison. Best viewed in color. More in Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15153" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.089, + 0.842, + 0.513 + ], + "angle": 0, + "content": "
Methods\nGeneratorNon-Instructional Editing ModelInstructional Editing Model
VidToMe [41] (SD1.5)TokenFlow [20] (SD2.1)Flatten [10] (SD2.1)InstructDiff [19] (SD1.5)LGVI [66] (SD1.5)InsV2V [9] (SD1.5)VEGGIE (Ours) (SD1.5)
Concept Addition
MLLM-Judge (↑)5.005.806.627.262.735.697.44
Alignment (↑)27.8029.3028.2228.1025.0628.2729.27
Smoothness (↑)96.7997.2695.7493.6696.0996.9494.93
Quality (↑)62.8165.6255.4554.0841.5956.2461.31
Detection (↑)47.9849.5349.7455.3614.4248.0157.96
Concept Removal
MLLM-Judge (↑)2.603.734.466.126.592.785.07
Alignment (↑)75.0175.9978.4075.5175.6774.4175.63
Smoothness (↑)96.1396.4795.8291.8397.0396.9995.04
Quality (↑)66.3271.5250.7755.0842.3158.7950.99
Detection (↑)34.3155.1670.9164.8178.4025.6470.22
Object Changing
MLLM-Judge (↑)5.006.537.377.002.066.606.63
Alignment (↑)25.6928.7627.0627.3622.1726.6027.77
Smoothness (↑)96.2397.2196.1392.0795.6696.7495.44
Quality (↑)64.0669.9759.3755.0138.2060.9058.15
Environment & Background Changing
MLLM-Judge (↑)5.817.357.376.052.376.607.18
Alignment (↑)28.1730.0030.0428.0321.9428.2729.15
Smoothness (↑)95.7696.9695.9089.8595.6696.0394.58
Quality (↑)61.9567.0654.5853.0638.9754.9454.25
Visual Feature Changing (Color & Texture)
MLLM-Judge (↑)5.866.856.606.432.147.537.33
Alignment (↑)27.9929.2529.4627.5423.1828.8828.69
Smoothness (↑)95.9397.1095.8391.7194.7596.6694.52
Quality (↑)65.8069.3153.3258.2936.2759.3657.91
Stylization
MLLM-Judge (↑)7.237.628.317.413.718.078.26
Alignment (↑)29.8430.2529.0027.7422.8029.1429.38
Smoothness (↑)96.3197.2396.7188.9795.6296.5095.69
Quality (↑)64.0568.2253.1854.1535.7662.5957.00
Object Grounding
SSIM (↑)40.4750.4647.2137.9866.8449.6570.90
Jaccard Index J (↑)13.8519.2925.6219.881.5213.8937.74
F-measure F (↑)15.5016.8617.6012.813.0717.3721.83
Reasoning Segmentation
SSIM (↑)---32.3944.4759.8668.41
Jaccard Index J (↑)---14.0210.1216.8922.53
F-measure F (↑)---8.079.0610.4515.97
Avg. Ranking2.611.411.963.003.212.001.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.517, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Table 2. Comparison of video editing task with instructional / non-instructional models on VEG-Bench. -: the task is not capable of non-instructional models. We gray out numbers of non-instructional models that are in different categories." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.557, + 0.465, + 0.622 + ], + "angle": 0, + "content": "
SettingsRemoval (FVD ↓)Grounding (SSIM ↑)
Grd.-only-52.34
Rmv.-only1098.52-
Mixed987.8055.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Table 3. An ablation study on whether multi-task learning provides transferable benefits that enhance performance across tasks. We focus on removal and grounding tasks as representative examples." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.688, + 0.462, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Figure 7. Comparison between single- and multi-skill models with different data training. We find tasks can help each other." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.56, + 0.835, + 0.574 + ], + "angle": 0, + "content": "image into the input video based on instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.681 + ], + "angle": 0, + "content": "Emergent Few-shot In-Context Editing. As shown in Fig. 3, VEGGIE can effectively utilize a few example image pairs to transfer the intended editing changes seamlessly to the input video. We observe that VEGGIE exhibits in-context learning for image editing without the need for language instructions. Instead, it uses image pairs as examples to infer and apply the desired editing intention directly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.888 + ], + "angle": 0, + "content": "We present VEGGIE, a unified end-to-end model for instructional video editing that handles diverse pixel-level tasks. VEGGIE leverages MLLM for robust instruction understanding and employs a video diffusion model to execute pixel-level edits. Our framework uses a single diffusion loss for end-to-end optimization across varied tasks/skills. We also introduce a novel synthetic data generation pipeline and VEG-Bench, a benchmark that assesses a broad range of editing skills. Our VEGGIE outperforms previous methods as a versatile, all-in-one solution. We hope our model, data, and benchmark to advance research on instructional generative video models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15154" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI, https://openai.com/research/video-generation-models-as-world-simulators, 2024.2.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[5] Alper Canberk, Maksym Bondarenko, Ege Ozguroglu, Ruoshi Liu, and Carl Vondrick. Erasedraw: Learning to insert objects by erasing them from images. arXiv preprint arXiv:2409.00522, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[6] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[7] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.484, + 0.676 + ], + "angle": 0, + "content": "[9] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In The Twelfth International Conference on Learning Representations, 2024. 5, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[10] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENtion for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[11] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Mohit Bansal, and Chen Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. arXiv preprint arXiv:2411.09921, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Bansal Mohit, and Chen. Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. 2024. 5, 6, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "[13] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.121, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[14] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In CVPR, pages 7346-7356, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[16] Hao Fei, Shengqiong Wu, Hanwang Zhang, Tat-Seng Chua, and Shuicheng Yan. Vitron: A unified pixel-level vision llm for understanding, generating, segmenting, editing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.908, + 0.383 + ], + "angle": 0, + "content": "[18] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.907, + 0.469 + ], + "angle": 0, + "content": "[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.907, + 0.511 + ], + "angle": 0, + "content": "[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2, 3, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.908, + 0.537 + ], + "angle": 0, + "content": "[21] gpt 4o. https://openai.com/index/hello-gpt-4o/.2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.908, + 0.607 + ], + "angle": 0, + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.908, + 0.663 + ], + "angle": 0, + "content": "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.907, + 0.734 + ], + "angle": 0, + "content": "[24] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15155" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[28] Jiahao Hu, Tianxiong Zhong, Xuebo Wang, Boyuan Jiang, Xingye Tian, Fei Yang, Pengfei Wan, and Di Zhang. Vivid-10m: A dataset and baseline for versatile and interactive video local editing. arXiv preprint arXiv:2411.15260, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.486, + 0.246 + ], + "angle": 0, + "content": "[29] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.331 + ], + "angle": 0, + "content": "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.485, + 0.36 + ], + "angle": 0, + "content": "[31] Paul Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37-50, 1912. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.485, + 0.417 + ], + "angle": 0, + "content": "[32] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.485, + 0.501 + ], + "angle": 0, + "content": "[33] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.485, + 0.559 + ], + "angle": 0, + "content": "[34] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.485, + 0.615 + ], + "angle": 0, + "content": "[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.485, + 0.672 + ], + "angle": 0, + "content": "[36] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.485, + 0.744 + ], + "angle": 0, + "content": "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.485, + 0.801 + ], + "angle": 0, + "content": "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.485, + 0.857 + ], + "angle": 0, + "content": "[39] Jialu Li, Shoubin Yu, Han Lin, Jaemin Cho, Jaehong Yoon, and Mohit Bansal. Training-free guidance in text-to-video generation via multimodal planning and structured noise initialization. arXiv preprint arXiv:2504.08641, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.858, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[40] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.134 + ], + "angle": 0, + "content": "image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.205 + ], + "angle": 0, + "content": "[41] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 2, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[42] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.303 + ], + "angle": 0, + "content": "[43] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.908, + 0.362 + ], + "angle": 0, + "content": "[44] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[45] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.49 + ], + "angle": 0, + "content": "[46] Shaoteng Liu, Tianyu Wang, Jui-Hsien Wang, Qing Liu, Zhifei Zhang, Joon-Young Lee, Yijun Li, Bei Yu, Zhe Lin, Soo Ye Kim, et al. Generative video propagation. arXiv preprint arXiv:2412.19761, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[47] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.908, + 0.631 + ], + "angle": 0, + "content": "[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.907, + 0.674 + ], + "angle": 0, + "content": "[49] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 9117-9125, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.908, + 0.731 + ], + "angle": 0, + "content": "[50] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, Mubarak Shah, and Ajmal Mian. Towards temporally consistent referring video object segmentation. https://arxiv.org/abs/2403.19407, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.908, + 0.8 + ], + "angle": 0, + "content": "[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[52] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. Detgpt: Detect what you need via reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14172-14189, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[53] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15156" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": " attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[54] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[58] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[59] Shangkun Sun, Xiaoyu Liang, Songlin Fan, Wenxu Gao, and Wei Gao. Ve-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[60] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[61] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[62] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.484, + 0.869 + ], + "angle": 0, + "content": "[63] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal IIm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[64] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image edit" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "ing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[65] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[66] Jianzong Wu, Xiangtai Li, Chenyang Si, Shangchen Zhou, Jingkang Yang, Jiangning Zhang, Yining Li, Kai Chen, Yunhai Tong, Ziwei Liu, et al. Towards language-driven video inpainting via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12501-12511, 2024. 2, 3, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[67] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.405 + ], + "angle": 0, + "content": "[68] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[69] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.907, + 0.533 + ], + "angle": 0, + "content": "[70] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.535, + 0.907, + 0.576 + ], + "angle": 0, + "content": "[71] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.579, + 0.907, + 0.772 + ], + "angle": 0, + "content": "[72] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[73] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[74] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15157" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[75] Jaehong Yoon, Shoubin Yu, and Mohit Bansal. Raccoon: Remove, add, and change video content with auto-generated narratives. arXiv preprint arXiv:2405.18406, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[77] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.369 + ], + "angle": 0, + "content": "[78] Shoubin Yu, Jacob Zhiyuan Fang, Jian Zheng, Gunnar Sigurdsson, Vicente Ordonez, Robinson Piramuthu, and Mohit Bansal. Zero-shot controllable image-to-video animation via motion decomposition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3332-3341, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[79] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.484, + 0.536 + ], + "angle": 0, + "content": "[81] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[82] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferret-v2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[83] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.733 + ], + "angle": 0, + "content": "[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[85] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. arXiv preprint arXiv:2312.03816, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[86] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.484, + 0.898 + ], + "angle": 0, + "content": "[87] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15158" + } + ] +] \ No newline at end of file diff --git a/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_origin.pdf b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8428bcac28e5266dfcc25a7bf7e3f6140ce6a8c0 --- /dev/null +++ b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/d727aeb7-1f89-498a-ac85-9bcc05bddf8b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb04cfcd242d4fca0335d8c3779b2a738b08925e46803f0c40b19d7eec2207ab +size 29056000 diff --git a/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/full.md b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e54c408bc73cfdf101fb6341dae50ae97bb4113 --- /dev/null +++ b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/full.md @@ -0,0 +1,278 @@ +# VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation + +Shoubin Yu $^{1,3*}$ Difan Liu $^{1*}$ Ziqiao Ma $^{1,2*}$ Yicong Hong $^{1}$ Yang Zhou $^{1}$ Hao Tan $^{1}$ Joyce Chai $^{2}$ Mohit Bansal $^{3}$ + +$^{1}$ Adobe Research $^{2}$ University of Michigan $^{3}$ UNC Chapel Hill + +https://veggie-gen.github.io/ + +![](images/6de461b48922321aa8d110e2c86339f81fccd8681d7979d3daf8439499923d75.jpg) +Figure 1. We propose VEGGIE, a unified and versatile video generative model that handles various tasks for both video concept grounding and editing according to user instructions. With VEGGIE, users can locate, add, delete, and change concepts in a given video through diverse instruction formats (direct referring instruction or reasoning-demanding questions). Users can also edit videos with multimodal instruction empowered by MLLM, enabling applications like video editing from a reference image. + +# Abstract + +While recent video diffusion models enable video editing, unifying diverse instructional editing tasks (e.g., add, remove, modify) under a single framework remains a significant challenge. In this paper, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions, a simple end-to-end framework that unifies video concept editing, grounding, and reasoning based on diverse user instructions. Specifically, given a video and text query, VEGGIE first utilizes an MLLM to interpret user intentions in instructions and ground them to the video contexts, generating framespecific grounded task queries for pixel-space responses. A diffusion model then renders these plans and generates edited videos that align with user intent. To support diverse + +tasks and complex instructions, we employ a curriculum learning strategy: first aligning the MLLM and video diffusion model with large-scale instructional image editing data, followed by end-to-end fine-tuning on high-quality multitask video data. Additionally, we introduce a novel data synthesis pipeline to generate paired instructional video editing data for model training. It transforms static image data into diverse, high-quality video editing samples by leveraging Image-to-Video models to inject dynamics. VEGGIE shows strong performance in instructional video editing with different editing skills, outperforming the best instructional baseline as a versatile model, while other models struggle with multi-tasking. VEGGIE also excels in video object grounding and reasoning segmentation, where other baselines fail. We further reveal how the multiple tasks help each other and highlight promising applications like zero-shot multimodal instructional and in-context video editing. + +# 1. Introduction + +Building on the advances in Video Diffusion Models (VIDDMs) [2, 4, 24, 25, 74], video editing methods have emerged as video design tools, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 86] have been developed, using triples of text prompts, source videos, and target videos for training. Due to their limited performance in understanding user intent and multimodal semantics [17], several methods have incorporated multimodal large language models (MLLMs) to handle complex instructions/reasoning [17, 29, 75]. + +However, existing methods fall short of the goal of a simple, versatile video concept editor, facing three primary challenges. First, most methods are not end-to-end, requiring intermediate layout/mask/human or model caption guidance [35, 53, 63, 75], which adds workload on users and disrupts a seamless editing experience. Second, existing pipelines connecting MLLMs to VidDMs require multiple training objectives beyond simple pixel-space diffusion loss, such as language loss [75] or mask losses [66]. This increases optimization difficulty and often requires additional hyperparameter tuning or annotations. Third, existing video editing models, both instructional and non-instructional, struggle with handling other diverse editing tasks, ranging from addition, and deletion to stylization. For example, LGVI [66] fails in global edits such as stylization and color change, while VidToMe [41] struggles with local edits such as adding or removing objects. These methods also struggle with input videos that contain multiple objects or when user instructions require complex reasoning. + +These challenges result from two limitations: First, there is a lack of multitasking fine-tuning on well-curated instructional video editing datasets that span a broad range of skills. Second, models often lack two critical capabilities needed to interpret user intentions and accurately locate concepts: multimodal reasoning to infer the intended modification from the user's instruction; and grounding language to the input video to precisely identify the region or object to be edited. For example, in Figure 1, one can effortlessly locate the girl given "identify the little girl." When asked to "add a hat to the little girl," we intuitively imagine the hat placed on her head from commonsense, even without seeing an actual hat. + +To address these challenges, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions. VEGGIE unifies video concept grounding and editing without relying on additional layout, mask guidance, or intermediate caption [35, 39, 42, 75, 78, 79]. Instead, we formulate the problem as end-to-end grounded generation in pixel space, using only a diffusion loss. Specifically, given a video and a text query, VEGGIE first leverages an MLLM to interpret complex instructions, generating frame-wise con + +ditions. Unlike prior methods [43, 75] that use discrete text tokens as conditions, which disconnect the pipeline and block gradient propagation, VEGGIE employs continuous, learnable task query embeddings per frame. This enables end-to-end training and effectively captures grounded task representations for diffusion model conditioning. To handle diverse tasks and accurately interpret complex queries, we employ a curriculum learning strategy that begins by aligning MLLMs with diffusion models using massive paired instructional image editing data and then fine-tuning the model end-to-end on high-quality multitask video data to adapt video. Unlike tool-use methods [16, 35, 75], VEGGIE formulates both video grounding and instructional editing in the same video-to-video task formulation, enabling efficient handling through a unified single model. To further support end-to-end training, we introduce a novel automatic instructional video data generation pipeline that lifts high-quality instructional image editing data into the video domain using image-to-video and video evaluation tools. + +Existing video editing benchmarks do not provide wide and uniform coverage of diverse editing skills [59, 68]. To address this gap, we contribute VEG-Bench, an instructional video editing benchmark that spans 8 editing skills: concept addition, removal, object changing, environment background changing, visual feature changing, stylization, object grounding, and reasoning segmentation. Each skill is evaluated using a dedicated suite of metrics. We assess the proposed VEGGIE alongside 6 baselines on VEG-Bench. VEGGIE demonstrates strong performance across diverse editing skills, outperforming the best instructional baseline as a versatile, all-in-one model, while other models struggle with multi-tasking. Additionally, VEGGIE excels in video object grounding and reasoning segmentation tasks, where other baselines fall short. We show further analysis of how multi-task learning enhances our framework and highlight applications such as zero-shot multimodal instructional following (Fig. 2) and few-shot in-context editing (Fig. 3). Our contributions are summarized as follows: + +- We propose VEGGIE, an end-to-end model that integrates an MLLM and a VidDM. VEGGIE is a versatile framework that handles diverse instructional requests for editing and grounding various video concepts. Unlike existing work that achieves multitasking via tool use, VEGGIE unifies diverse tasks in a single model, thus simplifying the training with only diffusion loss. +- We propose a data synthesis pipeline, scaling high-quality instructional video editing data for future work. +- We propose VEG-Bench, an instructional video editing benchmark that spans 8 editing skills with dedicated metrics for each skill. +- VEGGIE achieves strong performance across diverse editing skills compared with SoTA methods, and shows potentials for multimodal instruction and in-context following. + +![](images/e0d151b937adc7407e14b52cd875ca0bfb673a4ac4838074cdae58f7cce26412.jpg) +Reference + +![](images/4540fb3623e1042b72cdc51d40fae52db70bda03718617114a98d44bd5cf175c.jpg) +Transfer the style in the reference image. +Add the object in the reference image on the women. + +![](images/a83a9ab1de823a8e4032e1dedd9635d6aa3ee84abd0b1589dedcb0a1c5522e76.jpg) +Reference + +![](images/22d931f8d1d6d4cd46b69a5e91dbd7bdc650d603c3c3e56bcf3baecacfd0144e.jpg) +Figure 2. Multimodal instruction following emerges in VEGGIE, allowing for style transfer or object addition from reference images. + +![](images/b5f1369b96b80acff89b8832473d68767bd364283daa8b18fad10e9742759d4e.jpg) +Figure 3. In-context editing emerges in VEGGIE, allowing for few-shot learning of editing tasks with paired image demonstrations. + +# 2. Related Work + +Instructional Video Editing Video Diffusion Models (VidDMs) [2, 4, 6, 7, 15, 26, 49, 67, 71], enable high-quality video generation across a wide range of video concepts. Building on these advances, video editing methods have emerged as tools for video design, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 70, 86] have been developed, using triples of video instructions, source videos, and target videos for training. These methods demonstrate limited performance when complex multimodal reasoning is required, as noted by previous research on instructional image editing [17]. Moreover, they struggle with diverse editing tasks, from addition and deletion to stylization. For example, LGVI [66] is primarily designed for removal tasks, while TokenFlow [20] struggles with local edits such as adding, removing, or changing objects. We address this limitation with pixel-level multitasking fine-tuning on well-curated instructional video editing datasets covering various grounding and editing skills. + +Video Grounding and Segmentation Visual grounding requires models to connect language to its corresponding visual concept in the visual context [40, 47]. This is commonly evaluated via the language-guided semantic localization tasks, ranging from simple referring expressions in RefCOCO series [48, 77] and their generalized variant [44] that takes no-target and multi-target into account. Recently, grounded multimodal large lan + +guage models (MLLMs) are trained for object grounding to bounding boxes [8, 51, 52, 76, 82, 87] and segmentation masks [56, 69, 81, 84] using text-image pairs with fine-grained annotations linking phrases to entities. These models unlock the potential of reasoning segmentation [11, 36], bringing language-informed reasoning into semantic segmentation. Instead of using dedicated object detection or segmentation modules, we achieve video grounding through end-to-end training with only diffusion loss. + +# 3. Our Method: VEGGIE + +In this paper, we introduce a Video Editor with Grounded Generation from Instructions (VEGGIE), a unified and versatile generative video model. It combines the complex instruction understanding and reasoning capabilities of MLLMs with the generative capacity of VidDMs. The model is trained end-to-end with diffusion loss only. VEGGIE efficiently handles diverse user inputs, including direct instructions, complex questions requiring in-depth reasoning, and multimodal conditioning. It performs various pixel-level manipulations, enabling tasks such as video concept addition, removal, changing, stylization, grounding, and reasoning segmentation based on user instructions. We elaborate on the model design (Sec. 3.1), training and inference process (Sec. 3.2), and data curation (Sec. 3.3). + +# 3.1. Model Architecture + +VEGGIE consists of four main components (see Fig. 4): (1) a multimodal large language model, (2) a set of learnable + +![](images/419c74734a879ff33d66c796ab67f7d5591f5073932ab3c43c6916f3395226b5.jpg) + +![](images/4614bdaab383a2d7be34477abb6dd43be30e9fc13b804982c124645ab6ae1578.jpg) +Figure 4. Overview of our proposed end-to-end VEGGIE framework. Our Multimodal Large Language Model first understands input video frames and diverse user instructions, then it generates frame-wise reasoning queries that maintain per-frame editing conditions for the video diffusion model. The video diffusion model will render the MLLM-generated conditions to the pixel space for diverse tasks, including video editing, video grounding, and video reasoning segmentation with questions. We only apply diffusion loss for the whole pipeline training. + +grounded task queries, (3) an alignment network (single-layer MLP) that projects the MLLM output into the condition space of the diffusion model, and (4) a video diffusion model initialized from an instructional image editing model [83]. Our model first generates latent conditions for target video frames by querying multimodal context using an MLLM, then renders these conditions at the pixel level through a video diffusion model, as detailed below. + +MLLM for Generating Grounded Task Guidance. As illustrated in the left of Fig. 4, given a video consisting of a sequence of frames $V = [f_1, \dots, f_n]$ , where $n$ is the frame number of the given video, a user instruction/question $I$ , our goal is to obtain the response $\widehat{V} = [\widehat{f}_1, \dots, \widehat{f}_n]$ at pixel space that faithfully reflects user instruction about the given video. The MLLM module processes both the input video $V$ and a user instruction $I$ to generate a sequence of grounded task tokens per frame: $C = [c_1, \dots, c_n]$ , which are input and output in parallel. These tokens serve as task guidance and implicitly encode the target manipulation, such as object attributes, spatial relationships, or style transfer parameters. The MLLM ensures the model captures both explicit user instructions and implicit reasoning needs. + +VidDM for Rendering MLLM Guidance at Pixel Space. As illustrated in the right of Table 4, the VidDM takes the original video $V$ and the grounded task tokens $C$ as conditions to synthesize the target video $\widehat{V}$ . The original video is concatenated with the noise volume, and the task tokens are input to the cross-attention. With grounded task guidance in denoising steps, the generation process ensures that the output faithfully follows user instructions while preserving the + +video's structure and motion dynamics. Through iterative denoising, it refines each frame while maintaining temporal consistency, applying pixel modifications coherently for a smooth and visually consistent output video $\widehat{V}$ . + +# 3.2. Curriculum Learning from Images to Videos + +Training the model directly on video tasks presents two key challenges: (1) misalignment between MLLM and diffusion model representations, making it difficult for the diffusion model to interpret MLLM-generated task queries with limited fine-tuning data, and (2) the diffusion model's lack of multitasking capability, even for image tasks, due to insufficient training on diverse tasks. Our initial experiments also found the model collapsed when the whole pipeline was directly trained with all data. These challenges/observations underscore the need for pre-alignment between MLLM and the diffusion model to enable seamless adaptation from language-space task queries to pixel-space modifications. To this end, we adopt a two-stage curriculum learning strategy for the proposed VEGGIE framework. + +Stage 1: Aligning Diffusion and Language Spaces. In the first stage, we align the diffusion model with the MLLM using large-scale image-level instructional editing data. The MLLM remains frozen while we update the alignment network, grounded task queries, and diffusion UNet. This process fine-tunes the diffusion model weights to align with the language space, enabling the model to interpret MLLM-generated guidance and translate user instructions into pixel-level edits while preserving the MLLM's strong ability to understand instructions and user intentions. + +
TypeSourceR.E.G.# Img./Vid.# Ins.
VideoROVI [66]✓✓✓4.3K27.4K
VPLM [75]✓4.3K5.5K
GroundMoRe [12]✓✓1.3K5.5K
RVoS [60]✓1.9K6.1K
MeViS [14]✓1.5K17.1K
InstructV2V [9]✓68.3K68.3K
VEG-Edit (Ours)✓4.0K6.2K
Total136.1K
ImageSeed-Data-Edit [18]✓3M3M
LISA [37]✓✓0.2K1.3K
gRefCoCo [44]✓13.6K73.4K
PhraseCut [65]✓310.8K310.8K
EraseDraw [5]✓64.9K42.4K
MagicBrush [83]✓9.3K9.3K
SmartEdit [29]✓0.5K0.9K
Total3438.1K
+ +Table 1. Summary of our data for training. R.: Reasoning, E.: Editing, G.: Grounding. #Img/Vid: the number of images/videos, and #Ins.: the number of instruction-image/video pairs. + +Stage 2: Enhancing Temporal Consistency and Dynamics. With the MLLM and diffusion model aligned, fine-tuning diverse instructional video editing data becomes more effective for improved instruction following at pixel-space including temporal consistency, dynamic coherence, and editing faithfulness. In this stage, we fine-tune the framework with the MLLM, including the alignment network, grounded task queries, and all 3 dimensions in diffusion UNet, end-to-end with carefully curated multitasking instructional video editing data. Following prior work [22, 67], we inflated the 2D UNet from Stage 1 with temporal attention layers for video adaptation. For both stages 1 and 2, we optimize the framework with a single diffusion loss, enabling unified learning for improved instructional video editing performance while maintaining simplicity and efficiency. + +Classifier-Free Guidance during Testing. We employ classifier-free guidance to balance quality and diversity in diffusion-generated samples. Following prior work [3, 17], we apply classifier-free guidance to instructional visual editing considering two conditions: the grounded task tokens and the original video. To obtain unconditional guidance, we set null values $(\varnothing)$ for both task tokens and input video. In this case, our score estimate is: + +$$ +\begin{array}{l} \tilde {e _ {\theta}} (z _ {t}, c _ {T}, c _ {V}) = e _ {\theta} (z _ {t}, \varnothing , \varnothing) \\ + g _ {T} \cdot \left(e _ {\theta} \left(z _ {t}, c _ {V}, c _ {T}\right) - e _ {\theta} \left(z _ {t}, c _ {V}, \varnothing\right)\right) \\ + g _ {V} \cdot \left(e _ {\theta} \left(z _ {t}, c _ {V}, \varnothing\right) - e _ {\theta} \left(z _ {t}, \varnothing , \varnothing\right)\right), \\ \end{array} +$$ + +where $\theta$ represents the model parameters, $C_T$ and $C_V$ denote the task tokens and video conditions, $\varnothing$ is the null value, $z_t$ is the noised latent at timestamp $t$ , and $g_T$ and $g_V$ are the task guidance and video guidance scales, respectively. More training details are included later in Appendix. + +# 3.3. Data Curation Pipeline + +Existing video editing models, both instructional and non-instructional, struggle with diverse editing skills due to the lack of high-quality multitasking fine-tuning data. In this section, we introduce our data curation strategy to support VEGGIE in achieving versatile video editing skills. As listed in Tab. 1, we collect 3.4M image and 133.9K video data from diverse sources to support our VEGGIE curriculum learning as discussed in Sec. 3.2. We create our training dataset from two sources: (1) collecting existing image and video data and converting it into an instructional editing format, and (2) synthesizing new instructional video editing samples using existing datasets and generative models. + +Collecting Diverse Multitask Image and Video Data. We bring together instructional editing data from both image (Seed-Data-Edit [18], MagicBrush [83], EraseDraw [5]) and video (InstructV2V [9], VPLM [75]) sources. These datasets provide pairs of original and edited visual contents with user instructions. The tasks include adding, removing, and changing objects, stylizing, and performing global/local edits. Beyond editing datasets, we incorporate segmentation data at both the image level (gRefCoCo [44] and Phrase-Cut [65]) and the video level (RVoS and MeViS). These segmentation tasks are reformulated as color-filling challenges, which guide the model in learning referring grounding (i.e., understanding which object or region to edit) and strengthen its conceptual learning. To further unlock complex instruction understanding via MLLM, we include data that requires more advanced reasoning and implicit referencing. Specifically, we include: reasoning segmentation (LISA [37]), reasoning editing (SmartEdit [29]), interactive video inpainting (LGVI [66]), and motion-grounded video reasoning (GroundMoRe [12]). These tasks help VEGGIE learn implicit references and reasoning. + +Synthesizing Instructional Video Editing Data via Image-to-Video Animation. Recent methods [3, 54] generate synthetic instructional video-editing data by first creating text instructions with LLM, then getting edited videos via T2V models and prompt-to-prompt editing [23]. While these methods adapt image-based editing pipelines [9] to videos, the generated data suffer from temporal-consistency issues. To address this gap, we propose a novel image-to-video animation strategy that leverages the abundance of high-quality image-level instructional editing datasets [64, 83], which provide well-annotated instructions, paired edited images, and well-organized editing skill categories. As illustrated in Fig. 5, given an original image $I$ , an edited image $\bar{I}$ , and an instruction from an instructional image editing dataset [64], our approach involves three key steps. First, we use an offline MLLM [61, 72] to generate an image caption and an animation prompt that describes plausible motion within the image. Next, an image-to-video (I2V) model animates the image into a video $V$ . Finally, + +![](images/519aa3cbb726c3f46600125ec3400403ec4db3182b9494a57667d8cbec97125a.jpg) +Figure 5. Our data generation pipeline for synthetic instructional video editing data. It injects dynamics into well-constructed instructional image editing datasets via the Image-to-Video (I2V) Model, and generates paired video data for instruction editing. + +we generate the corresponding edited video $\bar{V}$ using a first-frame-conditioned video editing model [35], leveraging $\bar{I}$ as a strong prior to ensure consistent edits across frames. Finally, to ensure data quality, we evaluate each original-edited video pair with automatic video quality evaluation metrics [30], which assess the generated videos from diverse dimensions, e.g., motion smoothness, image quality, and background consistency. This pipeline transforms carefully curated image-based datasets into instructional video-editing resources while preserving the precision of the original edits. As a result, our data method expands the availability of high-quality synthetic video-editing data, supporting a wider range of editing tasks in our end-to-end unified framework. More details on data generation, prompting, examples, and pre/post-processing are in the Appendix. + +# 4. Experiments + +We first introduce the VEG-Bench Benchmark and then demonstrate the superiority of VEGGIE across diverse video instructional editing skills. More experiments, visualization, and implementation details are in the Appendix. + +# 4.1. VEG-Bench and Metrics + +As no existing benchmark is designed for fine-grained instructional video editing skills, we manually collect and annotate VEG-Bench, containing 132 video-instruction pairs that balanced cover 8 different video generative skills (15-20 for each). Beyond standard metrics, including text-to-video alignment (CLIP-Text [55]), video smoothness (CLIP-F [55]), and image quality (MUSIQ [32]), we also + +first introduce MLLM-as-a-Judge to give a holistic evaluation score according to the given original video, edited video, and user instruction. It is achieved by prompting GPT4o [21] to evaluate whether the requested semantic change has been fulfilled, using a scale from 1 to 10. For addition and removal, we also introduce an object detector (GroundingDiNo [45]) to detect if the object is added/removed faithfully. For grounding and reasoning segmentation, we following video grounding tasks [12, 14, 33, 58] and adopt the Jaccard index $(\mathcal{I})$ [31], F-measure $(\mathcal{F})$ [13], and their mean $(\mathcal{J} \& \mathcal{F})$ . We also compute SSIM between the generated video and the original video masked with GT masks. More evaluation/metrics details are included in Appendix. + +# 4.2. Experimental Results + +Instructional Video Editing over Diverse Skills. As shown in Tab. 2, we evaluate 7 different models on VEG-Bench across 8 distinct editing skills. Overall, VEGGIE demonstrates the best performance among instructional video editing models. Compared to VEGGIE, non-instructional models often struggle with concept removal and addition. This limitation arises because these models rely on attention control or additional conditions (e.g., depth maps) that impose strong priors, constraining the model and making object addition or removal challenging. We also observe that InsV2V achieves high scores in quality and smoothness metrics, but underperforms in alignment and MLLM judgment, which demand faithful semantic changes. Qualitative examples in Fig. 6 illustrate that InsV2V often makes minimal changes to the input video, resulting in high video quality but unfaithful outputs. In contrast, VEGGIE strikes a better balance, delivering both high-quality visuals and accurate semantic alignment with the intended edits. + +Can Multi-Task Help Each Other? To test the previous hypothesis, we train our model on the VPLM [75] dataset, which includes paired grounding and removal tasks (approximately 5.5K samples for each task). We focus on these tasks as representative examples due to their straightforward evaluation against ground truth. As shown in Table 3, multitask training yields a lower FVD score and a higher SSIM score, demonstrating that learning to locate and remove a video concept can mutually reinforce performance. We show an example in Fig. 7. However, this conclusion only holds with a balanced data combination. We also observe that an excessive amount of grounding data can introduce more artifacts and negatively impact visual editing skills. + +Emergent Zero-shot Multimodal Instruction Following. We also highlight the emergent behavior of VEGGIE on multi-modal instruction following, even without dedicated training data for this specific editing instruction. Notably, VEGGIE demonstrates the ability to perform zero-shot multimodal instructional video editing. As illustrated in Fig. 2, VEGGIE can transfer styles or add objects from a reference + +![](images/448541813648206c696b8060be6836952b32cd5e03312d900e585ca11af2f430.jpg) +[Addition] Please add a ball in the given video frames. + +![](images/67afd19caa3fbcb6b392c5ad751d2b0f2e3161441709a94940152b6d86723f7e.jpg) +[Removal] Please remove the man in black in given video frames. + +![](images/373ec45a7f7d566ba4d3a78d8e06b09d7bad2b84179f4cdf62f91a27bbbf530a.jpg) +[Swap] Replace golden building with a white mountain. + +![](images/41198c72d6faf22421eb58cefcc01dfd03a7dbb2d35994a01bd56dc226535d45.jpg) +[Environment] Make it on the beach. + +![](images/2aaed0e07fec7c8843e49de92130946b5b89bbd966dcb8e209bd72feffe8789e.jpg) +[Color] Make the swan white. + +![](images/a238051c85810f38dffc43dcef81eb1d2021aded79c1a77a3508a3bafc592ab6.jpg) +[Texture] Make the rhinocero furry. + +![](images/a41b7f9d6e96f7f28878ef89d70835e2d12c0c4a44228a64db4c7f7f34d6c0d5.jpg) +[Style] Make it chinese ink style. +Figure 6. Qualitative comparison of editing results across 8 different abilities (splitting visual features into color and texture). We provide zoom-in details for a more detailed comparison. Best viewed in color. More in Appendix. + +![](images/6fecf891dfe80f439d7ac4732ec66bb78527fab1d6f6d18a6235978f3c780c18.jpg) +[Grounding] Could you locate the knife + +![](images/b7c113e68f1b140f6b2fd89134382a5e2e6aa73949f6da185815fa51ef01fdde.jpg) +[Reasoning] What can be used for heating food? + +
Methods +GeneratorNon-Instructional Editing ModelInstructional Editing Model
VidToMe [41] (SD1.5)TokenFlow [20] (SD2.1)Flatten [10] (SD2.1)InstructDiff [19] (SD1.5)LGVI [66] (SD1.5)InsV2V [9] (SD1.5)VEGGIE (Ours) (SD1.5)
Concept Addition
MLLM-Judge (↑)5.005.806.627.262.735.697.44
Alignment (↑)27.8029.3028.2228.1025.0628.2729.27
Smoothness (↑)96.7997.2695.7493.6696.0996.9494.93
Quality (↑)62.8165.6255.4554.0841.5956.2461.31
Detection (↑)47.9849.5349.7455.3614.4248.0157.96
Concept Removal
MLLM-Judge (↑)2.603.734.466.126.592.785.07
Alignment (↑)75.0175.9978.4075.5175.6774.4175.63
Smoothness (↑)96.1396.4795.8291.8397.0396.9995.04
Quality (↑)66.3271.5250.7755.0842.3158.7950.99
Detection (↑)34.3155.1670.9164.8178.4025.6470.22
Object Changing
MLLM-Judge (↑)5.006.537.377.002.066.606.63
Alignment (↑)25.6928.7627.0627.3622.1726.6027.77
Smoothness (↑)96.2397.2196.1392.0795.6696.7495.44
Quality (↑)64.0669.9759.3755.0138.2060.9058.15
Environment & Background Changing
MLLM-Judge (↑)5.817.357.376.052.376.607.18
Alignment (↑)28.1730.0030.0428.0321.9428.2729.15
Smoothness (↑)95.7696.9695.9089.8595.6696.0394.58
Quality (↑)61.9567.0654.5853.0638.9754.9454.25
Visual Feature Changing (Color & Texture)
MLLM-Judge (↑)5.866.856.606.432.147.537.33
Alignment (↑)27.9929.2529.4627.5423.1828.8828.69
Smoothness (↑)95.9397.1095.8391.7194.7596.6694.52
Quality (↑)65.8069.3153.3258.2936.2759.3657.91
Stylization
MLLM-Judge (↑)7.237.628.317.413.718.078.26
Alignment (↑)29.8430.2529.0027.7422.8029.1429.38
Smoothness (↑)96.3197.2396.7188.9795.6296.5095.69
Quality (↑)64.0568.2253.1854.1535.7662.5957.00
Object Grounding
SSIM (↑)40.4750.4647.2137.9866.8449.6570.90
Jaccard Index J (↑)13.8519.2925.6219.881.5213.8937.74
F-measure F (↑)15.5016.8617.6012.813.0717.3721.83
Reasoning Segmentation
SSIM (↑)---32.3944.4759.8668.41
Jaccard Index J (↑)---14.0210.1216.8922.53
F-measure F (↑)---8.079.0610.4515.97
Avg. Ranking2.611.411.963.003.212.001.78
+ +Table 2. Comparison of video editing task with instructional / non-instructional models on VEG-Bench. -: the task is not capable of non-instructional models. We gray out numbers of non-instructional models that are in different categories. + +
SettingsRemoval (FVD ↓)Grounding (SSIM ↑)
Grd.-only-52.34
Rmv.-only1098.52-
Mixed987.8055.21
+ +Table 3. An ablation study on whether multi-task learning provides transferable benefits that enhance performance across tasks. We focus on removal and grounding tasks as representative examples. + +![](images/16dd7320589fd8681a64e1346642e6554bace0aed5cea65f14131a25d7187071.jpg) +Figure 7. Comparison between single- and multi-skill models with different data training. We find tasks can help each other. + +image into the input video based on instructions. + +Emergent Few-shot In-Context Editing. As shown in Fig. 3, VEGGIE can effectively utilize a few example image pairs to transfer the intended editing changes seamlessly to the input video. We observe that VEGGIE exhibits in-context learning for image editing without the need for language instructions. Instead, it uses image pairs as examples to infer and apply the desired editing intention directly. + +# 5. Conclusion + +We present VEGGIE, a unified end-to-end model for instructional video editing that handles diverse pixel-level tasks. VEGGIE leverages MLLM for robust instruction understanding and employs a video diffusion model to execute pixel-level edits. Our framework uses a single diffusion loss for end-to-end optimization across varied tasks/skills. We also introduce a novel synthetic data generation pipeline and VEG-Bench, a benchmark that assesses a broad range of editing skills. Our VEGGIE outperforms previous methods as a versatile, all-in-one solution. We hope our model, data, and benchmark to advance research on instructional generative video models. + +# References + +[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 2 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 5 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI, https://openai.com/research/video-generation-models-as-world-simulators, 2024.2.3 +[5] Alper Canberk, Maksym Bondarenko, Ege Ozguroglu, Ruoshi Liu, and Carl Vondrick. Erasedraw: Learning to insert objects by erasing them from images. arXiv preprint arXiv:2409.00522, 2024. 5 +[6] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3 +[7] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 3 +[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3 +[9] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In The Twelfth International Conference on Learning Representations, 2024. 5, 8, 1, 2 +[10] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENtion for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 8, 1 +[11] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Mohit Bansal, and Chen Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. arXiv preprint arXiv:2411.09921, 2024. 3 +[12] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Bansal Mohit, and Chen. Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. 2024. 5, 6, 2 + +[13] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 6 +[14] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 5, 6 +[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In CVPR, pages 7346-7356, 2023. 3 +[16] Hao Fei, Shengqiong Wu, Hanwang Zhang, Tat-Seng Chua, and Shuicheng Yan. Vitron: A unified pixel-level vision llm for understanding, generating, segmenting, editing, 2024. 2 +[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 3, 5 +[18] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 5 +[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 8, 1, 2 +[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2, 3, 8, 1 +[21] gpt 4o. https://openai.com/index/hello-gpt-4o/.2024.6 +[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 5, 1 +[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.5 +[24] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 2022. 2 +[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, 2023. 3 +[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 1 + +[28] Jiahao Hu, Tianxiong Zhong, Xuebo Wang, Boyuan Jiang, Xingye Tian, Fei Yang, Pengfei Wan, and Di Zhang. Vivid-10m: A dataset and baseline for versatile and interactive video local editing. arXiv preprint arXiv:2411.15260, 2024. 2 +[29] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 5 +[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 2 +[31] Paul Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37-50, 1912. 6 +[32] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[33] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 6 +[34] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2 +[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 6 +[36] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3 +[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 5 +[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1 +[39] Jialu Li, Shoubin Yu, Han Lin, Jaemin Cho, Jaehong Yoon, and Mohit Bansal. Training-free guidance in text-to-video generation via multimodal planning and structured noise initialization. arXiv preprint arXiv:2504.08641, 2025. 2 +[40] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language + +image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3 +[41] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 2, 8, 1 +[42] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2 +[43] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2 +[44] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. 3, 5 +[45] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6 +[46] Shaoteng Liu, Tianyu Wang, Jui-Hsien Wang, Qing Liu, Zhifei Zhang, Joon-Young Lee, Yijun Li, Bei Yu, Zhe Lin, Soo Ye Kim, et al. Generative video propagation. arXiv preprint arXiv:2412.19761, 2024. 2 +[47] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 3 +[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3 +[49] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 9117-9125, 2023. 3 +[50] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, Mubarak Shah, and Ajmal Mian. Towards temporally consistent referring video object segmentation. https://arxiv.org/abs/2403.19407, 2024. 2 +[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In The Twelfth International Conference on Learning Representations, 2024. 3 +[52] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. Detgpt: Detect what you need via reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14172-14189, 2023. 3 +[53] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing + +attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3 +[54] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 3, 5 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6 +[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[58] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6 +[59] Shangkun Sun, Xiaoyu Liang, Songlin Fan, Wenxu Gao, and Wei Gao. Ve-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2 +[60] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 5 +[61] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 2 +[62] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[63] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal IIm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 2 +[64] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image edit + +ing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 5, 2 +[65] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5 +[66] Jianzong Wu, Xiangtai Li, Chenyang Si, Shangchen Zhou, Jingkang Yang, Jiangning Zhang, Yining Li, Kai Chen, Yunhai Tong, Ziwei Liu, et al. Towards language-driven video inpainting via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12501-12511, 2024. 2, 3, 5, 8, 1 +[67] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 5 +[68] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2 +[69] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[70] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3 +[71] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3 +[72] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5 +[73] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1 +[74] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2 + +[75] Jaehong Yoon, Shoubin Yu, and Mohit Bansal. Raccoon: Remove, add, and change video content with auto-generated narratives. arXiv preprint arXiv:2405.18406, 2024. 2, 5, 6 +[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations, 2023. 3 +[77] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3 +[78] Shoubin Yu, Jacob Zhiyuan Fang, Jian Zheng, Gunnar Sigurdsson, Vicente Ordonez, Robinson Piramuthu, and Mohit Bansal. Zero-shot controllable image-to-video animation via motion decomposition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3332-3341, 2024. 2 +[79] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2, 3 +[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 1 +[81] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949, 2023. 3 +[82] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferret-v2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 3 +[83] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 1 +[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[85] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. arXiv preprint arXiv:2312.03816, 2023. 2, 3 +[86] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 2, 3 +[87] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 3 \ No newline at end of file diff --git a/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/images.zip b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..59961f21745331845b4556fb490c791e6482cfea --- /dev/null +++ b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78c227e9126fa5888a5e552a39c6477995db227dc770bbca52b8786616e40c28 +size 1338865 diff --git a/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/layout.json b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..129e7fc1afc445bd22f455ebef126eb64c03b04b --- /dev/null +++ b/2025/VEGGIE_ Instructional Editing and Reasoning Video Concepts with Grounded Generation/layout.json @@ -0,0 +1,8152 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 102, + 458, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 102, + 458, + 140 + ], + "spans": [ + { + "bbox": [ + 153, + 102, + 458, + 140 + ], + "type": "text", + "content": "VEGGIE: Instructional Editing and Reasoning Video Concepts with Grounded Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "spans": [ + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": "Shoubin Yu" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": " Difan Liu" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": " Ziqiao Ma" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": " Yicong Hong" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": " Yang Zhou" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": " Hao Tan" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": " Joyce Chai" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "text", + "content": " Mohit Bansal" + }, + { + "bbox": [ + 124, + 160, + 486, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "spans": [ + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "type": "text", + "content": "Adobe Research " + }, + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "type": "text", + "content": "University of Michigan " + }, + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 202, + 466, + 217 + ], + "type": "text", + "content": "UNC Chapel Hill" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 219, + 392, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 219, + 392, + 232 + ], + "spans": [ + { + "bbox": [ + 215, + 219, + 392, + 232 + ], + "type": "text", + "content": "https://veggie-gen.github.io/" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 82, + 245, + 525, + 440 + ], + "blocks": [ + { + "bbox": [ + 82, + 245, + 525, + 440 + ], + "lines": [ + { + "bbox": [ + 82, + 245, + 525, + 440 + ], + "spans": [ + { + "bbox": [ + 82, + 245, + 525, + 440 + ], + "type": "image", + "image_path": "6de461b48922321aa8d110e2c86339f81fccd8681d7979d3daf8439499923d75.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 447, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 54, + 447, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 555, + 492 + ], + "type": "text", + "content": "Figure 1. We propose VEGGIE, a unified and versatile video generative model that handles various tasks for both video concept grounding and editing according to user instructions. With VEGGIE, users can locate, add, delete, and change concepts in a given video through diverse instruction formats (direct referring instruction or reasoning-demanding questions). Users can also edit videos with multimodal instruction empowered by MLLM, enabling applications like video editing from a reference image." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 503, + 200, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 503, + 200, + 516 + ], + "spans": [ + { + "bbox": [ + 151, + 503, + 200, + 516 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 529, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 529, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 529, + 297, + 685 + ], + "type": "text", + "content": "While recent video diffusion models enable video editing, unifying diverse instructional editing tasks (e.g., add, remove, modify) under a single framework remains a significant challenge. In this paper, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions, a simple end-to-end framework that unifies video concept editing, grounding, and reasoning based on diverse user instructions. Specifically, given a video and text query, VEGGIE first utilizes an MLLM to interpret user intentions in instructions and ground them to the video contexts, generating framespecific grounded task queries for pixel-space responses. A diffusion model then renders these plans and generates edited videos that align with user intent. To support diverse" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 505, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 556, + 708 + ], + "type": "text", + "content": "tasks and complex instructions, we employ a curriculum learning strategy: first aligning the MLLM and video diffusion model with large-scale instructional image editing data, followed by end-to-end fine-tuning on high-quality multitask video data. Additionally, we introduce a novel data synthesis pipeline to generate paired instructional video editing data for model training. It transforms static image data into diverse, high-quality video editing samples by leveraging Image-to-Video models to inject dynamics. VEGGIE shows strong performance in instructional video editing with different editing skills, outperforming the best instructional baseline as a versatile model, while other models struggle with multi-tasking. VEGGIE also excels in video object grounding and reasoning segmentation, where other baselines fail. We further reveal how the multiple tasks help each other and highlight promising applications like zero-shot multimodal instructional and in-context video editing." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 228, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 228, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 228, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 703, + 228, + 712 + ], + "type": "text", + "content": " Work done during internship at Adobe Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15147" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 84 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 235 + ], + "type": "text", + "content": "Building on the advances in Video Diffusion Models (VIDDMs) [2, 4, 24, 25, 74], video editing methods have emerged as video design tools, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 86] have been developed, using triples of text prompts, source videos, and target videos for training. Due to their limited performance in understanding user intent and multimodal semantics [17], several methods have incorporated multimodal large language models (MLLMs) to handle complex instructions/reasoning [17, 29, 75]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 235, + 297, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 297, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 297, + 462 + ], + "type": "text", + "content": "However, existing methods fall short of the goal of a simple, versatile video concept editor, facing three primary challenges. First, most methods are not end-to-end, requiring intermediate layout/mask/human or model caption guidance [35, 53, 63, 75], which adds workload on users and disrupts a seamless editing experience. Second, existing pipelines connecting MLLMs to VidDMs require multiple training objectives beyond simple pixel-space diffusion loss, such as language loss [75] or mask losses [66]. This increases optimization difficulty and often requires additional hyperparameter tuning or annotations. Third, existing video editing models, both instructional and non-instructional, struggle with handling other diverse editing tasks, ranging from addition, and deletion to stylization. For example, LGVI [66] fails in global edits such as stylization and color change, while VidToMe [41] struggles with local edits such as adding or removing objects. These methods also struggle with input videos that contain multiple objects or when user instructions require complex reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 297, + 605 + ], + "type": "text", + "content": "These challenges result from two limitations: First, there is a lack of multitasking fine-tuning on well-curated instructional video editing datasets that span a broad range of skills. Second, models often lack two critical capabilities needed to interpret user intentions and accurately locate concepts: multimodal reasoning to infer the intended modification from the user's instruction; and grounding language to the input video to precisely identify the region or object to be edited. For example, in Figure 1, one can effortlessly locate the girl given \"identify the little girl.\" When asked to \"add a hat to the little girl,\" we intuitively imagine the hat placed on her head from commonsense, even without seeing an actual hat." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 606, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 297, + 714 + ], + "type": "text", + "content": "To address these challenges, we introduce VEGGIE, a Video Editor with Grounded Generation from Instructions. VEGGIE unifies video concept grounding and editing without relying on additional layout, mask guidance, or intermediate caption [35, 39, 42, 75, 78, 79]. Instead, we formulate the problem as end-to-end grounded generation in pixel space, using only a diffusion loss. Specifically, given a video and a text query, VEGGIE first leverages an MLLM to interpret complex instructions, generating frame-wise con" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": "ditions. Unlike prior methods [43, 75] that use discrete text tokens as conditions, which disconnect the pipeline and block gradient propagation, VEGGIE employs continuous, learnable task query embeddings per frame. This enables end-to-end training and effectively captures grounded task representations for diffusion model conditioning. To handle diverse tasks and accurately interpret complex queries, we employ a curriculum learning strategy that begins by aligning MLLMs with diffusion models using massive paired instructional image editing data and then fine-tuning the model end-to-end on high-quality multitask video data to adapt video. Unlike tool-use methods [16, 35, 75], VEGGIE formulates both video grounding and instructional editing in the same video-to-video task formulation, enabling efficient handling through a unified single model. To further support end-to-end training, we introduce a novel automatic instructional video data generation pipeline that lifts high-quality instructional image editing data into the video domain using image-to-video and video evaluation tools." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 302, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 556, + 529 + ], + "type": "text", + "content": "Existing video editing benchmarks do not provide wide and uniform coverage of diverse editing skills [59, 68]. To address this gap, we contribute VEG-Bench, an instructional video editing benchmark that spans 8 editing skills: concept addition, removal, object changing, environment background changing, visual feature changing, stylization, object grounding, and reasoning segmentation. Each skill is evaluated using a dedicated suite of metrics. We assess the proposed VEGGIE alongside 6 baselines on VEG-Bench. VEGGIE demonstrates strong performance across diverse editing skills, outperforming the best instructional baseline as a versatile, all-in-one model, while other models struggle with multi-tasking. Additionally, VEGGIE excels in video object grounding and reasoning segmentation tasks, where other baselines fall short. We show further analysis of how multi-task learning enhances our framework and highlight applications such as zero-shot multimodal instructional following (Fig. 2) and few-shot in-context editing (Fig. 3). Our contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 534, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 534, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 534, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 315, + 534, + 555, + 617 + ], + "type": "text", + "content": "- We propose VEGGIE, an end-to-end model that integrates an MLLM and a VidDM. VEGGIE is a versatile framework that handles diverse instructional requests for editing and grounding various video concepts. Unlike existing work that achieves multitasking via tool use, VEGGIE unifies diverse tasks in a single model, thus simplifying the training with only diffusion loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 618, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 618, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 315, + 618, + 553, + 641 + ], + "type": "text", + "content": "- We propose a data synthesis pipeline, scaling high-quality instructional video editing data for future work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 642, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 642, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 642, + 555, + 677 + ], + "type": "text", + "content": "- We propose VEG-Bench, an instructional video editing benchmark that spans 8 editing skills with dedicated metrics for each skill." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 678, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 678, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 678, + 556, + 713 + ], + "type": "text", + "content": "- VEGGIE achieves strong performance across diverse editing skills compared with SoTA methods, and shows potentials for multimodal instruction and in-context following." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15148" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 71, + 149, + 113 + ], + "blocks": [ + { + "bbox": [ + 99, + 71, + 149, + 113 + ], + "lines": [ + { + "bbox": [ + 99, + 71, + 149, + 113 + ], + "spans": [ + { + "bbox": [ + 99, + 71, + 149, + 113 + ], + "type": "image", + "image_path": "e0d151b937adc7407e14b52cd875ca0bfb673a4ac4838074cdae58f7cce26412.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 114, + 146, + 121 + ], + "lines": [ + { + "bbox": [ + 103, + 114, + 146, + 121 + ], + "spans": [ + { + "bbox": [ + 103, + 114, + 146, + 121 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 152, + 72, + 302, + 173 + ], + "blocks": [ + { + "bbox": [ + 102, + 128, + 148, + 166 + ], + "lines": [ + { + "bbox": [ + 102, + 128, + 148, + 166 + ], + "spans": [ + { + "bbox": [ + 102, + 128, + 148, + 166 + ], + "type": "text", + "content": "Transfer the style in the reference image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 152, + 72, + 302, + 173 + ], + "lines": [ + { + "bbox": [ + 152, + 72, + 302, + 173 + ], + "spans": [ + { + "bbox": [ + 152, + 72, + 302, + 173 + ], + "type": "image", + "image_path": "4540fb3623e1042b72cdc51d40fae52db70bda03718617114a98d44bd5cf175c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 122, + 353, + 171 + ], + "lines": [ + { + "bbox": [ + 306, + 122, + 353, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 122, + 353, + 171 + ], + "type": "text", + "content": "Add the object in the reference image on the women." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 80, + 356, + 113 + ], + "blocks": [ + { + "bbox": [ + 306, + 80, + 356, + 113 + ], + "lines": [ + { + "bbox": [ + 306, + 80, + 356, + 113 + ], + "spans": [ + { + "bbox": [ + 306, + 80, + 356, + 113 + ], + "type": "image", + "image_path": "a83a9ab1de823a8e4032e1dedd9635d6aa3ee84abd0b1589dedcb0a1c5522e76.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 113, + 355, + 121 + ], + "lines": [ + { + "bbox": [ + 306, + 113, + 355, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 113, + 355, + 121 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 509, + 172 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 509, + 172 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 509, + 172 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 509, + 172 + ], + "type": "image", + "image_path": "22d931f8d1d6d4cd46b69a5e91dbd7bdc650d603c3c3e56bcf3baecacfd0144e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 182, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 63, + 182, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 63, + 182, + 545, + 194 + ], + "type": "text", + "content": "Figure 2. Multimodal instruction following emerges in VEGGIE, allowing for style transfer or object addition from reference images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 99, + 208, + 509, + 312 + ], + "blocks": [ + { + "bbox": [ + 99, + 208, + 509, + 312 + ], + "lines": [ + { + "bbox": [ + 99, + 208, + 509, + 312 + ], + "spans": [ + { + "bbox": [ + 99, + 208, + 509, + 312 + ], + "type": "image", + "image_path": "b5f1369b96b80acff89b8832473d68767bd364283daa8b18fad10e9742759d4e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 324, + 541, + 335 + ], + "lines": [ + { + "bbox": [ + 67, + 324, + 541, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 324, + 541, + 335 + ], + "type": "text", + "content": "Figure 3. In-context editing emerges in VEGGIE, allowing for few-shot learning of editing tasks with paired image demonstrations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 355, + 142, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 142, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 142, + 367 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 376, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 616 + ], + "type": "text", + "content": "Instructional Video Editing Video Diffusion Models (VidDMs) [2, 4, 6, 7, 15, 26, 49, 67, 71], enable high-quality video generation across a wide range of video concepts. Building on these advances, video editing methods have emerged as tools for video design, allowing users to manipulate video concepts such as adding, removing, altering objects and style translation [20, 53, 62, 79, 85]. To enhance user experiences, instructional video editing methods [54, 70, 86] have been developed, using triples of video instructions, source videos, and target videos for training. These methods demonstrate limited performance when complex multimodal reasoning is required, as noted by previous research on instructional image editing [17]. Moreover, they struggle with diverse editing tasks, from addition and deletion to stylization. For example, LGVI [66] is primarily designed for removal tasks, while TokenFlow [20] struggles with local edits such as adding, removing, or changing objects. We address this limitation with pixel-level multitasking fine-tuning on well-curated instructional video editing datasets covering various grounding and editing skills." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Video Grounding and Segmentation Visual grounding requires models to connect language to its corresponding visual concept in the visual context [40, 47]. This is commonly evaluated via the language-guided semantic localization tasks, ranging from simple referring expressions in RefCOCO series [48, 77] and their generalized variant [44] that takes no-target and multi-target into account. Recently, grounded multimodal large lan" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 464 + ], + "type": "text", + "content": "guage models (MLLMs) are trained for object grounding to bounding boxes [8, 51, 52, 76, 82, 87] and segmentation masks [56, 69, 81, 84] using text-image pairs with fine-grained annotations linking phrases to entities. These models unlock the potential of reasoning segmentation [11, 36], bringing language-informed reasoning into semantic segmentation. Instead of using dedicated object detection or segmentation modules, we achieve video grounding through end-to-end training with only diffusion loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 475, + 444, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 444, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 444, + 487 + ], + "type": "text", + "content": "3. Our Method: VEGGIE" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 496, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 663 + ], + "type": "text", + "content": "In this paper, we introduce a Video Editor with Grounded Generation from Instructions (VEGGIE), a unified and versatile generative video model. It combines the complex instruction understanding and reasoning capabilities of MLLMs with the generative capacity of VidDMs. The model is trained end-to-end with diffusion loss only. VEGGIE efficiently handles diverse user inputs, including direct instructions, complex questions requiring in-depth reasoning, and multimodal conditioning. It performs various pixel-level manipulations, enabling tasks such as video concept addition, removal, changing, stylization, grounding, and reasoning segmentation based on user instructions. We elaborate on the model design (Sec. 3.1), training and inference process (Sec. 3.2), and data curation (Sec. 3.3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 672, + 429, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 429, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 429, + 683 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "VEGGIE consists of four main components (see Fig. 4): (1) a multimodal large language model, (2) a set of learnable" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15149" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 73, + 138, + 87 + ], + "blocks": [ + { + "bbox": [ + 123, + 73, + 138, + 87 + ], + "lines": [ + { + "bbox": [ + 123, + 73, + 138, + 87 + ], + "spans": [ + { + "bbox": [ + 123, + 73, + 138, + 87 + ], + "type": "image", + "image_path": "419c74734a879ff33d66c796ab67f7d5591f5073932ab3c43c6916f3395226b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 94, + 541, + 281 + ], + "blocks": [ + { + "bbox": [ + 69, + 94, + 541, + 281 + ], + "lines": [ + { + "bbox": [ + 69, + 94, + 541, + 281 + ], + "spans": [ + { + "bbox": [ + 69, + 94, + 541, + 281 + ], + "type": "image", + "image_path": "4614bdaab383a2d7be34477abb6dd43be30e9fc13b804982c124645ab6ae1578.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 286, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 286, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 555, + 331 + ], + "type": "text", + "content": "Figure 4. Overview of our proposed end-to-end VEGGIE framework. Our Multimodal Large Language Model first understands input video frames and diverse user instructions, then it generates frame-wise reasoning queries that maintain per-frame editing conditions for the video diffusion model. The video diffusion model will render the MLLM-generated conditions to the pixel space for diverse tasks, including video editing, video grounding, and video reasoning segmentation with questions. We only apply diffusion loss for the whole pipeline training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 351, + 297, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 297, + 446 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 297, + 446 + ], + "type": "text", + "content": "grounded task queries, (3) an alignment network (single-layer MLP) that projects the MLLM output into the condition space of the diffusion model, and (4) a video diffusion model initialized from an instructional image editing model [83]. Our model first generates latent conditions for target video frames by querying multimodal context using an MLLM, then renders these conditions at the pixel level through a video diffusion model, as detailed below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": "MLLM for Generating Grounded Task Guidance. As illustrated in the left of Fig. 4, given a video consisting of a sequence of frames " + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "inline_equation", + "content": "V = [f_1, \\dots, f_n]" + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": " is the frame number of the given video, a user instruction/question " + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": ", our goal is to obtain the response " + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\widehat{V} = [\\widehat{f}_1, \\dots, \\widehat{f}_n]" + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": " at pixel space that faithfully reflects user instruction about the given video. The MLLM module processes both the input video " + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": " and a user instruction " + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": " to generate a sequence of grounded task tokens per frame: " + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "inline_equation", + "content": "C = [c_1, \\dots, c_n]" + }, + { + "bbox": [ + 54, + 448, + 295, + 616 + ], + "type": "text", + "content": ", which are input and output in parallel. These tokens serve as task guidance and implicitly encode the target manipulation, such as object attributes, spatial relationships, or style transfer parameters. The MLLM ensures the model captures both explicit user instructions and implicit reasoning needs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": "VidDM for Rendering MLLM Guidance at Pixel Space. As illustrated in the right of Table 4, the VidDM takes the original video " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": " and the grounded task tokens " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": " as conditions to synthesize the target video " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\widehat{V}" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": ". The original video is concatenated with the noise volume, and the task tokens are input to the cross-attention. With grounded task guidance in denoising steps, the generation process ensures that the output faithfully follows user instructions while preserving the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 351, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 399 + ], + "type": "text", + "content": "video's structure and motion dynamics. Through iterative denoising, it refines each frame while maintaining temporal consistency, applying pixel modifications coherently for a smooth and visually consistent output video " + }, + { + "bbox": [ + 313, + 351, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\widehat{V}" + }, + { + "bbox": [ + 313, + 351, + 555, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 408, + 547, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 547, + 421 + ], + "type": "text", + "content": "3.2. Curriculum Learning from Images to Videos" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 593 + ], + "type": "text", + "content": "Training the model directly on video tasks presents two key challenges: (1) misalignment between MLLM and diffusion model representations, making it difficult for the diffusion model to interpret MLLM-generated task queries with limited fine-tuning data, and (2) the diffusion model's lack of multitasking capability, even for image tasks, due to insufficient training on diverse tasks. Our initial experiments also found the model collapsed when the whole pipeline was directly trained with all data. These challenges/observations underscore the need for pre-alignment between MLLM and the diffusion model to enable seamless adaptation from language-space task queries to pixel-space modifications. To this end, we adopt a two-stage curriculum learning strategy for the proposed VEGGIE framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Stage 1: Aligning Diffusion and Language Spaces. In the first stage, we align the diffusion model with the MLLM using large-scale image-level instructional editing data. The MLLM remains frozen while we update the alignment network, grounded task queries, and diffusion UNet. This process fine-tunes the diffusion model weights to align with the language space, enabling the model to interpret MLLM-generated guidance and translate user instructions into pixel-level edits while preserving the MLLM's strong ability to understand instructions and user intentions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 138, + 73, + 466, + 86 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 73, + 466, + 86 + ], + "spans": [ + { + "bbox": [ + 138, + 73, + 466, + 86 + ], + "type": "text", + "content": "VEGGIE: A Unified and Versatile Instructional Video Generative Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15150" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 293, + 247 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 293, + 247 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 293, + 247 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 293, + 247 + ], + "type": "table", + "html": "
TypeSourceR.E.G.# Img./Vid.# Ins.
VideoROVI [66]✓✓✓4.3K27.4K
VPLM [75]✓4.3K5.5K
GroundMoRe [12]✓✓1.3K5.5K
RVoS [60]✓1.9K6.1K
MeViS [14]✓1.5K17.1K
InstructV2V [9]✓68.3K68.3K
VEG-Edit (Ours)✓4.0K6.2K
Total136.1K
ImageSeed-Data-Edit [18]✓3M3M
LISA [37]✓✓0.2K1.3K
gRefCoCo [44]✓13.6K73.4K
PhraseCut [65]✓310.8K310.8K
EraseDraw [5]✓64.9K42.4K
MagicBrush [83]✓9.3K9.3K
SmartEdit [29]✓0.5K0.9K
Total3438.1K
", + "image_path": "c6400e6808829343ebe8bff1b8ea156dad0faa011c49ed265b4bc83db3d484f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 251, + 296, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 285 + ], + "type": "text", + "content": "Table 1. Summary of our data for training. R.: Reasoning, E.: Editing, G.: Grounding. #Img/Vid: the number of images/videos, and #Ins.: the number of instruction-image/video pairs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 304, + 297, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 297, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 297, + 483 + ], + "type": "text", + "content": "Stage 2: Enhancing Temporal Consistency and Dynamics. With the MLLM and diffusion model aligned, fine-tuning diverse instructional video editing data becomes more effective for improved instruction following at pixel-space including temporal consistency, dynamic coherence, and editing faithfulness. In this stage, we fine-tune the framework with the MLLM, including the alignment network, grounded task queries, and all 3 dimensions in diffusion UNet, end-to-end with carefully curated multitasking instructional video editing data. Following prior work [22, 67], we inflated the 2D UNet from Stage 1 with temporal attention layers for video adaptation. For both stages 1 and 2, we optimize the framework with a single diffusion loss, enabling unified learning for improved instructional video editing performance while maintaining simplicity and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 487, + 298, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 298, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 298, + 582 + ], + "type": "text", + "content": "Classifier-Free Guidance during Testing. We employ classifier-free guidance to balance quality and diversity in diffusion-generated samples. Following prior work [3, 17], we apply classifier-free guidance to instructional visual editing considering two conditions: the grounded task tokens and the original video. To obtain unconditional guidance, we set null values " + }, + { + "bbox": [ + 54, + 487, + 298, + 582 + ], + "type": "inline_equation", + "content": "(\\varnothing)" + }, + { + "bbox": [ + 54, + 487, + 298, + 582 + ], + "type": "text", + "content": " for both task tokens and input video. In this case, our score estimate is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 594, + 289, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 594, + 289, + 638 + ], + "spans": [ + { + "bbox": [ + 61, + 594, + 289, + 638 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {e _ {\\theta}} (z _ {t}, c _ {T}, c _ {V}) = e _ {\\theta} (z _ {t}, \\varnothing , \\varnothing) \\\\ + g _ {T} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, c _ {T}\\right) - e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right)\\right) \\\\ + g _ {V} \\cdot \\left(e _ {\\theta} \\left(z _ {t}, c _ {V}, \\varnothing\\right) - e _ {\\theta} \\left(z _ {t}, \\varnothing , \\varnothing\\right)\\right), \\\\ \\end{array}", + "image_path": "d574a8c0f7021ea88b10f2cd0f55fa710e69b0e0df08baf1f3c010b568df75b7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " represents the model parameters, " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_T" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_V" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " denote the task tokens and video conditions, " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " is the null value, " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " is the noised latent at timestamp " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_T" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_V" + }, + { + "bbox": [ + 54, + 653, + 296, + 715 + ], + "type": "text", + "content": " are the task guidance and video guidance scales, respectively. More training details are included later in Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "type": "text", + "content": "3.3. Data Curation Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 89, + 555, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 89, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 312, + 89, + 555, + 233 + ], + "type": "text", + "content": "Existing video editing models, both instructional and non-instructional, struggle with diverse editing skills due to the lack of high-quality multitasking fine-tuning data. In this section, we introduce our data curation strategy to support VEGGIE in achieving versatile video editing skills. As listed in Tab. 1, we collect 3.4M image and 133.9K video data from diverse sources to support our VEGGIE curriculum learning as discussed in Sec. 3.2. We create our training dataset from two sources: (1) collecting existing image and video data and converting it into an instructional editing format, and (2) synthesizing new instructional video editing samples using existing datasets and generative models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 234, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 234, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 312, + 234, + 556, + 486 + ], + "type": "text", + "content": "Collecting Diverse Multitask Image and Video Data. We bring together instructional editing data from both image (Seed-Data-Edit [18], MagicBrush [83], EraseDraw [5]) and video (InstructV2V [9], VPLM [75]) sources. These datasets provide pairs of original and edited visual contents with user instructions. The tasks include adding, removing, and changing objects, stylizing, and performing global/local edits. Beyond editing datasets, we incorporate segmentation data at both the image level (gRefCoCo [44] and Phrase-Cut [65]) and the video level (RVoS and MeViS). These segmentation tasks are reformulated as color-filling challenges, which guide the model in learning referring grounding (i.e., understanding which object or region to edit) and strengthen its conceptual learning. To further unlock complex instruction understanding via MLLM, we include data that requires more advanced reasoning and implicit referencing. Specifically, we include: reasoning segmentation (LISA [37]), reasoning editing (SmartEdit [29]), interactive video inpainting (LGVI [66]), and motion-grounded video reasoning (GroundMoRe [12]). These tasks help VEGGIE learn implicit references and reasoning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "text", + "content": "Synthesizing Instructional Video Editing Data via Image-to-Video Animation. Recent methods [3, 54] generate synthetic instructional video-editing data by first creating text instructions with LLM, then getting edited videos via T2V models and prompt-to-prompt editing [23]. While these methods adapt image-based editing pipelines [9] to videos, the generated data suffer from temporal-consistency issues. To address this gap, we propose a novel image-to-video animation strategy that leverages the abundance of high-quality image-level instructional editing datasets [64, 83], which provide well-annotated instructions, paired edited images, and well-organized editing skill categories. As illustrated in Fig. 5, given an original image " + }, + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "text", + "content": ", an edited image " + }, + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\bar{I}" + }, + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "text", + "content": ", and an instruction from an instructional image editing dataset [64], our approach involves three key steps. First, we use an offline MLLM [61, 72] to generate an image caption and an animation prompt that describes plausible motion within the image. Next, an image-to-video (I2V) model animates the image into a video " + }, + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 312, + 486, + 556, + 714 + ], + "type": "text", + "content": ". Finally," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15151" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 69, + 276, + 282 + ], + "blocks": [ + { + "bbox": [ + 73, + 69, + 276, + 282 + ], + "lines": [ + { + "bbox": [ + 73, + 69, + 276, + 282 + ], + "spans": [ + { + "bbox": [ + 73, + 69, + 276, + 282 + ], + "type": "image", + "image_path": "519aa3cbb726c3f46600125ec3400403ec4db3182b9494a57667d8cbec97125a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 286, + 297, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 297, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 297, + 331 + ], + "type": "text", + "content": "Figure 5. Our data generation pipeline for synthetic instructional video editing data. It injects dynamics into well-constructed instructional image editing datasets via the Image-to-Video (I2V) Model, and generates paired video data for instruction editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 341, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 296, + 521 + ], + "type": "text", + "content": "we generate the corresponding edited video " + }, + { + "bbox": [ + 54, + 341, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\bar{V}" + }, + { + "bbox": [ + 54, + 341, + 296, + 521 + ], + "type": "text", + "content": " using a first-frame-conditioned video editing model [35], leveraging " + }, + { + "bbox": [ + 54, + 341, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\bar{I}" + }, + { + "bbox": [ + 54, + 341, + 296, + 521 + ], + "type": "text", + "content": " as a strong prior to ensure consistent edits across frames. Finally, to ensure data quality, we evaluate each original-edited video pair with automatic video quality evaluation metrics [30], which assess the generated videos from diverse dimensions, e.g., motion smoothness, image quality, and background consistency. This pipeline transforms carefully curated image-based datasets into instructional video-editing resources while preserving the precision of the original edits. As a result, our data method expands the availability of high-quality synthetic video-editing data, supporting a wider range of editing tasks in our end-to-end unified framework. More details on data generation, prompting, examples, and pre/post-processing are in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 137, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 137, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 137, + 547 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 554, + 297, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 297, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 297, + 603 + ], + "type": "text", + "content": "We first introduce the VEG-Bench Benchmark and then demonstrate the superiority of VEGGIE across diverse video instructional editing skills. More experiments, visualization, and implementation details are in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 192, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 192, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 192, + 624 + ], + "type": "text", + "content": "4.1. VEG-Bench and Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": "As no existing benchmark is designed for fine-grained instructional video editing skills, we manually collect and annotate VEG-Bench, containing 132 video-instruction pairs that balanced cover 8 different video generative skills (15-20 for each). Beyond standard metrics, including text-to-video alignment (CLIP-Text [55]), video smoothness (CLIP-F [55]), and image quality (MUSIQ [32]), we also" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "text", + "content": "first introduce MLLM-as-a-Judge to give a holistic evaluation score according to the given original video, edited video, and user instruction. It is achieved by prompting GPT4o [21] to evaluate whether the requested semantic change has been fulfilled, using a scale from 1 to 10. For addition and removal, we also introduce an object detector (GroundingDiNo [45]) to detect if the object is added/removed faithfully. For grounding and reasoning segmentation, we following video grounding tasks [12, 14, 33, 58] and adopt the Jaccard index " + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "text", + "content": " [31], F-measure " + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "text", + "content": " [13], and their mean " + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "inline_equation", + "content": "(\\mathcal{J} \\& \\mathcal{F})" + }, + { + "bbox": [ + 313, + 72, + 556, + 228 + ], + "type": "text", + "content": ". We also compute SSIM between the generated video and the original video masked with GT masks. More evaluation/metrics details are included in Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 238, + 440, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 440, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 440, + 251 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 256, + 556, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 556, + 472 + ], + "type": "text", + "content": "Instructional Video Editing over Diverse Skills. As shown in Tab. 2, we evaluate 7 different models on VEG-Bench across 8 distinct editing skills. Overall, VEGGIE demonstrates the best performance among instructional video editing models. Compared to VEGGIE, non-instructional models often struggle with concept removal and addition. This limitation arises because these models rely on attention control or additional conditions (e.g., depth maps) that impose strong priors, constraining the model and making object addition or removal challenging. We also observe that InsV2V achieves high scores in quality and smoothness metrics, but underperforms in alignment and MLLM judgment, which demand faithful semantic changes. Qualitative examples in Fig. 6 illustrate that InsV2V often makes minimal changes to the input video, resulting in high video quality but unfaithful outputs. In contrast, VEGGIE strikes a better balance, delivering both high-quality visuals and accurate semantic alignment with the intended edits." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 628 + ], + "type": "text", + "content": "Can Multi-Task Help Each Other? To test the previous hypothesis, we train our model on the VPLM [75] dataset, which includes paired grounding and removal tasks (approximately 5.5K samples for each task). We focus on these tasks as representative examples due to their straightforward evaluation against ground truth. As shown in Table 3, multitask training yields a lower FVD score and a higher SSIM score, demonstrating that learning to locate and remove a video concept can mutually reinforce performance. We show an example in Fig. 7. However, this conclusion only holds with a balanced data combination. We also observe that an excessive amount of grounding data can introduce more artifacts and negatively impact visual editing skills." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "Emergent Zero-shot Multimodal Instruction Following. We also highlight the emergent behavior of VEGGIE on multi-modal instruction following, even without dedicated training data for this specific editing instruction. Notably, VEGGIE demonstrates the ability to perform zero-shot multimodal instructional video editing. As illustrated in Fig. 2, VEGGIE can transfer styles or add objects from a reference" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15152" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 79, + 225, + 269 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 214, + 79 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 214, + 79 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 214, + 79 + ], + "type": "text", + "content": "[Addition] Please add a ball in the given video frames." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 79, + 225, + 269 + ], + "lines": [ + { + "bbox": [ + 70, + 79, + 225, + 269 + ], + "spans": [ + { + "bbox": [ + 70, + 79, + 225, + 269 + ], + "type": "image", + "image_path": "448541813648206c696b8060be6836952b32cd5e03312d900e585ca11af2f430.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 79, + 383, + 269 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 383, + 79 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 383, + 79 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 383, + 79 + ], + "type": "text", + "content": "[Removal] Please remove the man in black in given video frames." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 79, + 383, + 269 + ], + "lines": [ + { + "bbox": [ + 227, + 79, + 383, + 269 + ], + "spans": [ + { + "bbox": [ + 227, + 79, + 383, + 269 + ], + "type": "image", + "image_path": "67afd19caa3fbcb6b392c5ad751d2b0f2e3161441709a94940152b6d86723f7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 383, + 79, + 539, + 269 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 529, + 79 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 529, + 79 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 529, + 79 + ], + "type": "text", + "content": "[Swap] Replace golden building with a white mountain." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 79, + 539, + 269 + ], + "lines": [ + { + "bbox": [ + 383, + 79, + 539, + 269 + ], + "spans": [ + { + "bbox": [ + 383, + 79, + 539, + 269 + ], + "type": "image", + "image_path": "373ec45a7f7d566ba4d3a78d8e06b09d7bad2b84179f4cdf62f91a27bbbf530a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 277, + 225, + 467 + ], + "blocks": [ + { + "bbox": [ + 99, + 271, + 196, + 278 + ], + "lines": [ + { + "bbox": [ + 99, + 271, + 196, + 278 + ], + "spans": [ + { + "bbox": [ + 99, + 271, + 196, + 278 + ], + "type": "text", + "content": "[Environment] Make it on the beach." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 277, + 225, + 467 + ], + "lines": [ + { + "bbox": [ + 70, + 277, + 225, + 467 + ], + "spans": [ + { + "bbox": [ + 70, + 277, + 225, + 467 + ], + "type": "image", + "image_path": "41198c72d6faf22421eb58cefcc01dfd03a7dbb2d35994a01bd56dc226535d45.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 227, + 277, + 383, + 467 + ], + "blocks": [ + { + "bbox": [ + 266, + 271, + 343, + 278 + ], + "lines": [ + { + "bbox": [ + 266, + 271, + 343, + 278 + ], + "spans": [ + { + "bbox": [ + 266, + 271, + 343, + 278 + ], + "type": "text", + "content": "[Color] Make the swan white." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 277, + 383, + 467 + ], + "lines": [ + { + "bbox": [ + 227, + 277, + 383, + 467 + ], + "spans": [ + { + "bbox": [ + 227, + 277, + 383, + 467 + ], + "type": "image", + "image_path": "2aaed0e07fec7c8843e49de92130946b5b89bbd966dcb8e209bd72feffe8789e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 383, + 277, + 539, + 467 + ], + "blocks": [ + { + "bbox": [ + 416, + 271, + 508, + 277 + ], + "lines": [ + { + "bbox": [ + 416, + 271, + 508, + 277 + ], + "spans": [ + { + "bbox": [ + 416, + 271, + 508, + 277 + ], + "type": "text", + "content": "[Texture] Make the rhinocero furry." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 277, + 539, + 467 + ], + "lines": [ + { + "bbox": [ + 383, + 277, + 539, + 467 + ], + "spans": [ + { + "bbox": [ + 383, + 277, + 539, + 467 + ], + "type": "image", + "image_path": "a238051c85810f38dffc43dcef81eb1d2021aded79c1a77a3508a3bafc592ab6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 70, + 480, + 225, + 670 + ], + "blocks": [ + { + "bbox": [ + 105, + 473, + 190, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 190, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 190, + 480 + ], + "type": "text", + "content": "[Style] Make it chinese ink style." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 480, + 225, + 670 + ], + "lines": [ + { + "bbox": [ + 70, + 480, + 225, + 670 + ], + "spans": [ + { + "bbox": [ + 70, + 480, + 225, + 670 + ], + "type": "image", + "image_path": "a41b7f9d6e96f7f28878ef89d70835e2d12c0c4a44228a64db4c7f7f34d6c0d5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 679, + 553, + 701 + ], + "lines": [ + { + "bbox": [ + 55, + 679, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 679, + 553, + 701 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of editing results across 8 different abilities (splitting visual features into color and texture). We provide zoom-in details for a more detailed comparison. Best viewed in color. More in Appendix." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 227, + 475, + 383, + 670 + ], + "blocks": [ + { + "bbox": [ + 252, + 468, + 360, + 475 + ], + "lines": [ + { + "bbox": [ + 252, + 468, + 360, + 475 + ], + "spans": [ + { + "bbox": [ + 252, + 468, + 360, + 475 + ], + "type": "text", + "content": "[Grounding] Could you locate the knife" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 475, + 383, + 670 + ], + "lines": [ + { + "bbox": [ + 227, + 475, + 383, + 670 + ], + "spans": [ + { + "bbox": [ + 227, + 475, + 383, + 670 + ], + "type": "image", + "image_path": "6fecf891dfe80f439d7ac4732ec66bb78527fab1d6f6d18a6235978f3c780c18.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 383, + 475, + 539, + 670 + ], + "blocks": [ + { + "bbox": [ + 398, + 468, + 528, + 475 + ], + "lines": [ + { + "bbox": [ + 398, + 468, + 528, + 475 + ], + "spans": [ + { + "bbox": [ + 398, + 468, + 528, + 475 + ], + "type": "text", + "content": "[Reasoning] What can be used for heating food?" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 475, + 539, + 670 + ], + "lines": [ + { + "bbox": [ + 383, + 475, + 539, + 670 + ], + "spans": [ + { + "bbox": [ + 383, + 475, + 539, + 670 + ], + "type": "image", + "image_path": "b7c113e68f1b140f6b2fd89134382a5e2e6aa73949f6da185815fa51ef01fdde.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15153" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 70, + 515, + 406 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 515, + 406 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 515, + 406 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 515, + 406 + ], + "type": "table", + "html": "
Methods\nGeneratorNon-Instructional Editing ModelInstructional Editing Model
VidToMe [41] (SD1.5)TokenFlow [20] (SD2.1)Flatten [10] (SD2.1)InstructDiff [19] (SD1.5)LGVI [66] (SD1.5)InsV2V [9] (SD1.5)VEGGIE (Ours) (SD1.5)
Concept Addition
MLLM-Judge (↑)5.005.806.627.262.735.697.44
Alignment (↑)27.8029.3028.2228.1025.0628.2729.27
Smoothness (↑)96.7997.2695.7493.6696.0996.9494.93
Quality (↑)62.8165.6255.4554.0841.5956.2461.31
Detection (↑)47.9849.5349.7455.3614.4248.0157.96
Concept Removal
MLLM-Judge (↑)2.603.734.466.126.592.785.07
Alignment (↑)75.0175.9978.4075.5175.6774.4175.63
Smoothness (↑)96.1396.4795.8291.8397.0396.9995.04
Quality (↑)66.3271.5250.7755.0842.3158.7950.99
Detection (↑)34.3155.1670.9164.8178.4025.6470.22
Object Changing
MLLM-Judge (↑)5.006.537.377.002.066.606.63
Alignment (↑)25.6928.7627.0627.3622.1726.6027.77
Smoothness (↑)96.2397.2196.1392.0795.6696.7495.44
Quality (↑)64.0669.9759.3755.0138.2060.9058.15
Environment & Background Changing
MLLM-Judge (↑)5.817.357.376.052.376.607.18
Alignment (↑)28.1730.0030.0428.0321.9428.2729.15
Smoothness (↑)95.7696.9695.9089.8595.6696.0394.58
Quality (↑)61.9567.0654.5853.0638.9754.9454.25
Visual Feature Changing (Color & Texture)
MLLM-Judge (↑)5.866.856.606.432.147.537.33
Alignment (↑)27.9929.2529.4627.5423.1828.8828.69
Smoothness (↑)95.9397.1095.8391.7194.7596.6694.52
Quality (↑)65.8069.3153.3258.2936.2759.3657.91
Stylization
MLLM-Judge (↑)7.237.628.317.413.718.078.26
Alignment (↑)29.8430.2529.0027.7422.8029.1429.38
Smoothness (↑)96.3197.2396.7188.9795.6296.5095.69
Quality (↑)64.0568.2253.1854.1535.7662.5957.00
Object Grounding
SSIM (↑)40.4750.4647.2137.9866.8449.6570.90
Jaccard Index J (↑)13.8519.2925.6219.881.5213.8937.74
F-measure F (↑)15.5016.8617.6012.813.0717.3721.83
Reasoning Segmentation
SSIM (↑)---32.3944.4759.8668.41
Jaccard Index J (↑)---14.0210.1216.8922.53
F-measure F (↑)---8.079.0610.4515.97
Avg. Ranking2.611.411.963.003.212.001.78
", + "image_path": "24831cb68b0c27cf147e918fb58dc3a60803aa169edf93235ef0f987dc167bca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 441, + 284, + 492 + ], + "blocks": [ + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "type": "text", + "content": "Table 2. Comparison of video editing task with instructional / non-instructional models on VEG-Bench. -: the task is not capable of non-instructional models. We gray out numbers of non-instructional models that are in different categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 441, + 284, + 492 + ], + "lines": [ + { + "bbox": [ + 72, + 441, + 284, + 492 + ], + "spans": [ + { + "bbox": [ + 72, + 441, + 284, + 492 + ], + "type": "table", + "html": "
SettingsRemoval (FVD ↓)Grounding (SSIM ↑)
Grd.-only-52.34
Rmv.-only1098.52-
Mixed987.8055.21
", + "image_path": "9b625bcc9addcaa022aa6854cf74a6df0f31f087476d9e9b0190a15cb855a15d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 501, + 295, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 534 + ], + "type": "text", + "content": "Table 3. An ablation study on whether multi-task learning provides transferable benefits that enhance performance across tasks. We focus on removal and grounding tasks as representative examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 544, + 282, + 656 + ], + "blocks": [ + { + "bbox": [ + 70, + 544, + 282, + 656 + ], + "lines": [ + { + "bbox": [ + 70, + 544, + 282, + 656 + ], + "spans": [ + { + "bbox": [ + 70, + 544, + 282, + 656 + ], + "type": "image", + "image_path": "16dd7320589fd8681a64e1346642e6554bace0aed5cea65f14131a25d7187071.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "Figure 7. Comparison between single- and multi-skill models with different data training. We find tasks can help each other." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 443, + 511, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 511, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 511, + 454 + ], + "type": "text", + "content": "image into the input video based on instructions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 455, + 555, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 539 + ], + "type": "text", + "content": "Emergent Few-shot In-Context Editing. As shown in Fig. 3, VEGGIE can effectively utilize a few example image pairs to transfer the intended editing changes seamlessly to the input video. We observe that VEGGIE exhibits in-context learning for image editing without the need for language instructions. Instead, it uses image pairs as examples to infer and apply the desired editing intention directly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 703 + ], + "type": "text", + "content": "We present VEGGIE, a unified end-to-end model for instructional video editing that handles diverse pixel-level tasks. VEGGIE leverages MLLM for robust instruction understanding and employs a video diffusion model to execute pixel-level edits. Our framework uses a single diffusion loss for end-to-end optimization across varied tasks/skills. We also introduce a novel synthetic data generation pipeline and VEG-Bench, a benchmark that assesses a broad range of editing skills. Our VEGGIE outperforms previous methods as a versatile, all-in-one solution. We hope our model, data, and benchmark to advance research on instructional generative video models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15154" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Zechen Bai, Tong He, Haiyang Mei, Pichao Wang, Ziteng Gao, Joya Chen, Lei Liu, Zheng Zhang, and Mike Zheng Shou. One token to seg them all: Language instructed reasoning segmentation in videos. arXiv preprint arXiv:2409.19603, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 312 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. OpenAI, https://openai.com/research/video-generation-models-as-world-simulators, 2024.2.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 296, + 356 + ], + "type": "text", + "content": "[5] Alper Canberk, Maksym Bondarenko, Ege Ozguroglu, Ruoshi Liu, and Carl Vondrick. Erasedraw: Learning to insert objects by erasing them from images. arXiv preprint arXiv:2409.00522, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 402 + ], + "type": "text", + "content": "[6] Duygu Ceylan, Chun-Hao P Huang, and Niloy J Mitra. Pix2video: Video editing using image diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23206-23217, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 446 + ], + "type": "text", + "content": "[7] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stablevideo: Text-driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23040-23050, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "text", + "content": "[8] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 296, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 535 + ], + "type": "text", + "content": "[9] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In The Twelfth International Conference on Learning Representations, 2024. 5, 8, 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 296, + 602 + ], + "type": "text", + "content": "[10] Yuren Cong, Mengmeng Xu, christian simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. FLATTEN: optical FLOW-guided ATTENtion for consistent text-to-video editing. In The Twelfth International Conference on Learning Representations, 2024. 8, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 603, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 658 + ], + "type": "text", + "content": "[11] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Mohit Bansal, and Chen Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. arXiv preprint arXiv:2411.09921, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[12] Andong Deng, Tongjia Chen, Shoubin Yu, Taojiannan Yang, Lincoln Spencer, Yapeng Tian, Ajmal Saeed Mian, Bansal Mohit, and Chen. Chen. Motion-grounded video reasoning: Understanding and perceiving motion at pixel level. 2024. 5, 6, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[13] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 95, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 95, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 555, + 139 + ], + "type": "text", + "content": "[14] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. Mevis: A large-scale benchmark for video segmentation with motion expressions. In ICCV, pages 2694-2703, 2023. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "text", + "content": "[15] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In CVPR, pages 7346-7356, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 554, + 217 + ], + "type": "text", + "content": "[16] Hao Fei, Shengqiong Wu, Hanwang Zhang, Tat-Seng Chua, and Shuicheng Yan. Vitron: A unified pixel-level vision llm for understanding, generating, segmenting, editing, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 259 + ], + "type": "text", + "content": "[17] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 555, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 303 + ], + "type": "text", + "content": "[18] Yuying Ge, Sijie Zhao, Chen Li, Yixiao Ge, and Ying Shan. Seed-data-edit technical report: A hybrid dataset for instructional image editing. arXiv preprint arXiv:2405.04007, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 555, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 371 + ], + "type": "text", + "content": "[19] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12709-12720, 2024. 8, 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 404 + ], + "type": "text", + "content": "[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2, 3, 8, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 405, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 555, + 425 + ], + "type": "text", + "content": "[21] gpt 4o. https://openai.com/index/hello-gpt-4o/.2024.6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 426, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 555, + 480 + ], + "type": "text", + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 5, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 482, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 555, + 525 + ], + "type": "text", + "content": "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022.5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 526, + 555, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 555, + 581 + ], + "type": "text", + "content": "[24] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "text", + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "text", + "content": "[26] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15155" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[28] Jiahao Hu, Tianxiong Zhong, Xuebo Wang, Boyuan Jiang, Xingye Tian, Fei Yang, Pengfei Wan, and Di Zhang. Vivid-10m: A dataset and baseline for versatile and interactive video local editing. arXiv preprint arXiv:2411.15260, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 297, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 194 + ], + "type": "text", + "content": "[29] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. Smartedit: Exploring complex instruction-based image editing with multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8362-8371, 2024. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 262 + ], + "type": "text", + "content": "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 6, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 262, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 296, + 285 + ], + "type": "text", + "content": "[31] Paul Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37-50, 1912. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 330 + ], + "type": "text", + "content": "[32] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 396 + ], + "type": "text", + "content": "[33] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 398, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 296, + 442 + ], + "type": "text", + "content": "[34] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 443, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 296, + 487 + ], + "type": "text", + "content": "[35] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 296, + 532 + ], + "type": "text", + "content": "[36] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 533, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 296, + 589 + ], + "type": "text", + "content": "[37] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "text", + "content": "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "text", + "content": "[39] Jialu Li, Shoubin Yu, Han Lin, Jaemin Cho, Jaehong Yoon, and Mohit Bansal. Training-free guidance in text-to-video generation via multimodal planning and structured noise initialization. arXiv preprint arXiv:2504.08641, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 679, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 679, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 679, + 296, + 714 + ], + "type": "text", + "content": "[40] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "text", + "content": "image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[41] Xirui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidthome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 2, 8, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 195 + ], + "type": "text", + "content": "[42] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "text", + "content": "[43] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 555, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 286 + ], + "type": "text", + "content": "[44] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23592-23601, 2023. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 342 + ], + "type": "text", + "content": "[45] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 555, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 388 + ], + "type": "text", + "content": "[46] Shaoteng Liu, Tianyu Wang, Jui-Hsien Wang, Qing Liu, Zhifei Zhang, Joon-Young Lee, Yijun Li, Bei Yu, Zhe Lin, Soo Ye Kim, et al. Generative video propagation. arXiv preprint arXiv:2412.19761, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 443 + ], + "type": "text", + "content": "[47] Ziqiao Ma, Jiayi Pan, and Joyce Chai. World-to-words: Grounded open vocabulary acquisition through fast mapping in vision-language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 524–544, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 555, + 499 + ], + "type": "text", + "content": "[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 555, + 533 + ], + "type": "text", + "content": "[49] Kangfu Mei and Vishal Patel. Vidm: Video implicit diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 9117-9125, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "text", + "content": "[50] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, Mubarak Shah, and Ajmal Mian. Towards temporally consistent referring video object segmentation. https://arxiv.org/abs/2403.19407, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 579, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 555, + 633 + ], + "type": "text", + "content": "[51] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 690 + ], + "type": "text", + "content": "[52] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. Detgpt: Detect what you need via reasoning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14172-14189, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "text", + "content": "[53] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15156" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "text", + "content": " attentions for zero-shot text-based video editing. In ICCV, pages 15932-15942, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[54] Bosheng Qin, Juncheng Li, Siliang Tang, Tat-Seng Chua, and Yueting Zhuang. Instructvid2vid: Controllable video editing with natural language instructions. In 2024 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2024. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 285 + ], + "type": "text", + "content": "[56] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 342 + ], + "type": "text", + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 407 + ], + "type": "text", + "content": "[58] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "text", + "content": "[59] Shangkun Sun, Xiaoyu Liang, Songlin Fan, Wenxu Gao, and Wei Gao. Ve-bench: Subjective-aligned benchmark suite for text-driven video editing quality assessment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "text", + "content": "[60] Carles Ventura, Miriam Bellver, Andreu Girbau, Amaia Salvador, Ferran Marques, and Xavier Giro-i Nieto. Rvos: End-to-end recurrent network for video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5277-5286, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 586 + ], + "type": "text", + "content": "[61] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 590, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 644 + ], + "type": "text", + "content": "[62] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 646, + 296, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 296, + 688 + ], + "type": "text", + "content": "[63] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal IIm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[64] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image edit" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "content": "ing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 5, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[65] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasecut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 218 + ], + "type": "text", + "content": "[66] Jianzong Wu, Xiangtai Li, Chenyang Si, Shangchen Zhou, Jingkang Yang, Jiangning Zhang, Yining Li, Kai Chen, Yunhai Tong, Ziwei Liu, et al. Towards language-driven video inpainting via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12501-12511, 2024. 2, 3, 5, 8, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "text", + "content": "[67] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In ICCV, pages 7623-7633, 2023. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 555, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 320 + ], + "type": "text", + "content": "[68] Jay Zhangjie Wu, Xiuyu Li, Difei Gao, Zhen Dong, Jinbin Bai, Aishani Singh, Xiaoyu Xiang, Youzeng Li, Zuwei Huang, Yuanxi Sun, et al. Cvpr 2023 text guided video editing competition. arXiv preprint arXiv:2310.16003, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 322, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 555, + 376 + ], + "type": "text", + "content": "[69] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 378, + 555, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 555, + 422 + ], + "type": "text", + "content": "[70] Zhen Xing, Qi Dai, Zihao Zhang, Hui Zhang, Han Hu, Zuxuan Wu, and Yu-Gang Jiang. Vidiff: Translating videos via multi-modal instructions with diffusion models. arXiv preprint arXiv:2311.18837, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 423, + 555, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 423, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 423, + 555, + 456 + ], + "type": "text", + "content": "[71] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 458, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 458, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 458, + 555, + 611 + ], + "type": "text", + "content": "[72] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 613, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 656 + ], + "type": "text", + "content": "[73] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[74] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15157" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 711 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[75] Jaehong Yoon, Shoubin Yu, and Mohit Bansal. Raccoon: Remove, add, and change video content with auto-generated narratives. arXiv preprint arXiv:2405.18406, 2024. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 161 + ], + "type": "text", + "content": "[76] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 225 + ], + "type": "text", + "content": "[77] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 292 + ], + "type": "text", + "content": "[78] Shoubin Yu, Jacob Zhiyuan Fang, Jian Zheng, Gunnar Sigurdsson, Vicente Ordonez, Robinson Piramuthu, and Mohit Bansal. Zero-shot controllable image-to-video animation via motion decomposition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3332-3341, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 337 + ], + "type": "text", + "content": "[79] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "type": "text", + "content": "[80] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 296, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 296, + 424 + ], + "type": "text", + "content": "[81] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Lei Zhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. arXiv preprint arXiv:2312.02949, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 480 + ], + "type": "text", + "content": "[82] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferret-v2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 525 + ], + "type": "text", + "content": "[83] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36, 2024. 4, 5, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 580 + ], + "type": "text", + "content": "[84] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. Groundhog: Grounding large language models to holistic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "text", + "content": "[85] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. arXiv preprint arXiv:2312.03816, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "text", + "content": "[86] Zhenghao Zhang, Zuozhuo Dai, Long Qin, and Weizhi Wang. Effived: Efficient video editing via text-instruction diffusion models. arXiv preprint arXiv:2403.11568, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 711 + ], + "type": "text", + "content": "[87] Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt: Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15158" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_content_list.json b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8300327f7161d8c452260f95e296c817ca03f11b --- /dev/null +++ b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_content_list.json @@ -0,0 +1,1631 @@ +[ + { + "type": "text", + "text": "VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sihan Yang1 Runsen Xu1,2 Chenhang Cui3 Tai Wang1† Dahua Lin1,2 Jiangmiao Pang1†", + "bbox": [ + 120, + 200, + 875, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong $^{3}$ National University of Singapore", + "bbox": [ + 184, + 223, + 813, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† Corresponding Author taiwang.me@gmail.com pangjiangmiao@gmail.com", + "bbox": [ + 217, + 239, + 772, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) excel in visual-language tasks by leveraging numerous visual tokens for fine-grained visual information, but this token redundancy results in significant computational costs. Previous research aimed at reducing visual tokens during inference typically leverages importance maps derived from attention scores among vision-only tokens or vision-language tokens to prune tokens across one or multiple pruning stages. Despite this progress, pruning frameworks and strategies remain simplistic and insufficiently explored, often resulting in substantial performance degradation. In this paper, we propose VFlowOpt, a token pruning framework that introduces an importance map derivation process and a progressive pruning module with a recycling mechanism. The hyperparameters of its pruning strategy are further optimized by a visual information flow-guided method. Specifically, we compute an importance map for image tokens based on their attention-derived context relevance and patch-level information entropy. We then decide which tokens to retain or prune and aggregate the pruned ones as recycled tokens to avoid potential information loss. Finally, we apply a visual information flow-guided method that regards the last token in the LMM as the most representative signal of text-visual interactions. This method minimizes the discrepancy between token representations in LMMs with and without pruning, thereby enabling superior pruning strategies tailored to different LMMs. Experiments demonstrate that VFlowOpt can prune $90\\%$ of visual tokens while maintaining comparable performance, leading to an $89\\%$ reduction in KV-Cache memory and $3.8 \\times$ faster inference.", + "bbox": [ + 88, + 324, + 485, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 811, + 220, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) [3, 8, 21, 22] have attained remarkable performance in tasks such as visual question answering [2, 12] and multimodal reasoning [44, 45], making them indispensable for applications like au", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4bd6a78824a2c043ef30859d3263d0397b752420a0d40c549f030e7784e45242.jpg", + "image_caption": [ + "Figure 1. VFlowOpt Performance and Efficiency. Our VFlowOpt significantly outperforms other LMM token reduction methods across three general benchmarks (MME, MMStar, and MMBench) on LLaVA-OneVision-7B. It achieves zero performance degradation when reducing $50\\%$ of visual tokens. Furthermore, when retaining only $10\\%$ of the tokens, it achieves $90\\%$ of the original performance while reducing KV-Cache memory usage by $89\\%$ and accelerating inference by $3.8\\times$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 280, + 903, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tonomous driving [9, 31, 35] and robotics [18, 20, 24]. To capture fine-grained visual details, modern LMMs treat images as token sequences, with models such as LLaVA-1.5 [21] processing hundreds of tokens and LLaVA-OneVision [17] handling up to several thousand tokens. Nonetheless, the increasing quantity of visual tokens significantly magnifies computational cost, memory usage, and inference time, creating a critical bottleneck for LMM deployment, particularly in resource-constrained environments or latency-sensitive scenarios.", + "bbox": [ + 509, + 594, + 906, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate these constraints, recent work explores various methods for reducing the number of visual tokens during inference. Most existing approaches evaluate the importance of visual tokens in LMMs using attention-based importance maps derived either from vision-only tokens [40, 49] or from text-visual interactions [36, 52]. Following a predefined pruning ratio, visual tokens are pruned in a single stage [33] or progressively across multiple stages [42], often with heuristic strategies applied uniformly across different LMMs. Although promising, these methods are sim", + "bbox": [ + 509, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23924", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plastic and underexplored, with coarse-grained pruning often causing significant performance drops due to lacking fitness for different LMMs models' characteristics.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these limitations, we propose a novel token pruning framework, VFlowOpt, that more accurately estimates token importance by integrating attention calibration and entropy, and reduces information loss through token recycling and progressive pruning strategy with fine-grained pruning ratios. Furthermore, by customizing pruning strategies for different models, our framework better preserves model performance. Specifically, the pruning framework consists of importance map computation and a progressive pruning module with a recycling mechanism. We decompose the importance map of image tokens into two aspects: the importance of tokens in the visual context, reflected by their attention maps, and the richness of visual information in each image patch, captured by its information entropy. Based on the weighted summation of these two factors, we determine which tokens to retain or prune across multiple pruning stages and introduce a recycling mechanism to aggregate pruned tokens to avoid potential information loss. Built with this pruning strategy with several hyperparameters, a key ingredient is the visual information flow-guided method to optimize the pruning strategy, which treats the last token in each pruning stage as the most representative signal of text-visual interactions during inference. By minimizing the difference between token representations in LMMs with and without pruning, VFlowOpt ultimately delivers superior pruning strategies specifically tailored to different LMMs.", + "bbox": [ + 91, + 138, + 486, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comprehensive experiments on multiple vision-language benchmarks validate the efficacy of VFlowOpt. In particular, it substantially lowers computational cost while preserving competitive performance. As shown in Fig. 1, VFlowOpt can prune $50\\%$ of visual tokens with negligible performance loss. Moreover, it can prune $90\\%$ of visual tokens while maintaining $90\\%$ of the model's performance, resulting in an $89\\%$ reduction in KV-Cache memory usage and a $3.8 \\times$ speedup in inference. These experimental results highlight its effectiveness as a practical and efficient solution for deploying LMMs in resource-constrained environments.", + "bbox": [ + 89, + 547, + 482, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 744, + 232, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Multimodal Models", + "text_level": 1, + "bbox": [ + 89, + 771, + 328, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) [3, 8, 21, 22] extend the reasoning capabilities of Large Language Models (LLMs) [4, 32, 38, 39] to vision-language tasks by integrating a pre-trained vision encoder [29, 46] with a language model, linked by an alignment module such as an MLP, or a query-based network. This design transforms visual inputs into token sequences that the LLM can pro", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cess, facilitating multimodal prompts for tasks like visual question answering [2, 6, 12, 44]. To enhance performance, advanced LMMs, such as LLaVA-OneVision [17] and Qwen2-VL [34], can encode higher-resolution images into more image tokens, thereby capturing more granular visual details. However, as image resolution grows, the number of visual tokens rises exponentially, leading to substantially higher computational costs. For example, LLaVA-1.5 [21] processes $336 \\times 336$ images into 576 tokens, whereas LLaVA-OneVision can handle $1152 \\times 1152$ images, producing 7,290 tokens. This challenge becomes even more pronounced in video-based models like LongVA [48], which must process tokens across numerous frames. Fine-grained visual tokenization boosts LMM performance but poses an inference bottleneck, driving efforts to balance performance and cost through token reduction.", + "bbox": [ + 511, + 90, + 906, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Token Reduction for LMMs", + "text_level": 1, + "bbox": [ + 511, + 340, + 764, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Token reduction has become a key strategy for improving the efficiency of LMMs by mitigating the computational cost associated with extensive visual token sequences. Existing methods can be broadly classified into training-based and training-free approaches. Training-based methods, such as LLaVA-Mini [50] and LLaVolta [5], introduce additional modules during model training to compress visual tokens and preserve critical information, while approaches like ATP-LLaVA [43] and p-MoD [47] train pruning modules to dynamically retain important tokens across LLM layers. However, these methods demand substantial computational resources to retrain the models, limiting their real-world applicability. In contrast, training-free methods prune tokens without additional training, often leveraging attention mechanisms to identify and discard redundant tokens. For instance, FastV[7] and SparseVLM exploit text-visual attention to rank token importance, whereas FasterVLM [49], VisionZip [40], and VTC-CLS [33] rely on [CLS] token attention in the vision encoder to evaluate token importance and prune redundant tokens. FitPrune [42] and PDrop [36] propose a progressive multistage pruning strategy in LMMs to fully utilize visual information. However, the pruning strategies in these methods often lack adaptability to different LMMs, which frequently leads to significant performance drops.", + "bbox": [ + 511, + 364, + 908, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 753, + 604, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce VFlowOpt, a framework that employs a pruning strategy to reduce redundant visual tokens during LMM inference while preserving essential visual information. This strategy involves multiple hyperparameters that our VFlowOpt framework automatically optimizes for proper configuration—an essential factor for maintaining model performance. Specifically, Section 3.1 explains how our method evaluates the importance of visual", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23925", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11f3eec939c6bc6f8cac28c20264c1deb05bddf83b8a2d827d4fc8af5103a5c8.jpg", + "image_caption": [ + "(a) Token Importance Estimation" + ], + "image_footnote": [], + "bbox": [ + 99, + 109, + 429, + 262 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bd996f6af1ded17ff73c3562764b3550a45b47a57cd004df5ca48fe7e66614ac.jpg", + "image_caption": [ + "(b) Token Recycling" + ], + "image_footnote": [], + "bbox": [ + 447, + 113, + 604, + 256 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ac77e9c733e7b164f80cd00a8207b0782e0c95b091775067933bce49f861a589.jpg", + "image_caption": [ + "(1) Inference Stage" + ], + "image_footnote": [], + "bbox": [ + 98, + 290, + 609, + 467 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3c405a4b6ddb48846a13a61c0ebe6a3c45543abb3506bdd88281f3b31d79216c.jpg", + "image_caption": [ + "(2) Optimization Stage" + ], + "image_footnote": [], + "bbox": [ + 617, + 88, + 893, + 406 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08f38546afb6798412810532e22ad42dbf3845ad1e93985beeff65b10ec97b23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 450, + 784, + 464 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9c4ac0f406f77133d4fb72a2bffd19daf953d57a3f3ef903f97d9543a1ade086.jpg", + "image_caption": [ + "Figure 2. Overview of VFlowOpt. (1) During inference, VFlowOpt first assesses the importance of visual tokens, based on which progressive token pruning is performed. After the initial pruning stage, the pruned tokens are merged and recycled. The pruning strategy used in this process is defined by the (2) Optimization Stage. (a) The importance map is computed by combining the attention of relatively important tokens with the entropy of image patches. (b) The pruned tokens are grouped into grid cells, where each cell has a side length of a. Within each grid cell, the pruned tokens are fused using a weighted average, with their importance values as weights, and then recycled. (2) VFlowOpt optimizes the pruning strategy by minimizing the discrepancy of the last token in the final layer of the LMM with and without applying the pruning strategy.", + "Figure 3. The attention of redundant tokens (marked in red) fails to reflect the importance of other tokens and instead focuses on similar tokens, such as background elements." + ], + "image_footnote": [], + "bbox": [ + 107, + 575, + 464, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tokens, Section 3.2 presents progressive pruning and token recycling strategies, and finally, Section 3.3 details how VFlowOpt customizes the superior pruning strategy (i.e., selects the pruning hyperparameters) for different LMMs.", + "bbox": [ + 89, + 779, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Visual Token Importance Estimation", + "text_level": 1, + "bbox": [ + 89, + 848, + 410, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a general approach to evaluate visual token importance, the previous works [14, 41] propose that the importance of", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "visual tokens can be estimated using the average attention from all other tokens in ViTs. Although it has achieved some success, we find that redundant tokens (e.g., tokens corresponding to background regions) often assign disproportionately high attention to other similar redundant tokens (shown in Fig. 3), reducing the reliability of the importance estimation. To address this, we first identify relatively important tokens based on the attention they receive from all tokens (tokens receiving higher attention are considered relatively important), as shown in Fig. 2 (a). We then exclude redundant tokens and use the attention from these relatively important tokens as a more robust metric. Additionally, we incorporate the information entropy of image patches into the importance score to prioritize tokens corresponding to visually informative regions.", + "bbox": [ + 511, + 577, + 906, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To identify relatively important visual tokens, we define a threshold:", + "bbox": [ + 511, + 805, + 906, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = t \\cdot \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} A _ {i j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 834, + 906, + 877 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the threshold, $A_{ij}$ is the attention weight from", + "bbox": [ + 513, + 885, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23926", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "token $i$ to token $j$ within a ViT layer, $N$ is the total number of tokens, and $t$ is a sensitivity hyperparameter. We then treat tokens with a total attention exceeding $\\tau$ as relatively important:", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {K} = \\{j \\mid \\sum_ {i = 1} ^ {N} A _ {i j} > \\tau \\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 151, + 480, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{K}$ denotes the set of indices of these relatively important tokens.", + "bbox": [ + 89, + 196, + 480, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The information entropy of an image patch corresponding to token $i$ is defined as:", + "bbox": [ + 89, + 226, + 480, + 255 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH \\left(V _ {i}\\right) = - \\sum_ {k = 0} ^ {L - 1} p _ {k} \\log p _ {k}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 266, + 480, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{i}$ is the image patch corresponding to token $i$ , $p_{k}$ is the proportion of pixels in the patch with gray level $k$ , and $L$ is the total number of possible gray levels (256 for 8-bit images). Here, the gray level is defined as the average of the RGB channel values, providing a scalar representation of pixel brightness. Higher entropy reflects greater diversity in pixel intensities and richer visual information.", + "bbox": [ + 89, + 316, + 482, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The importance score of token $i$ is computed by combining the attention it receives from tokens in $\\mathcal{K}$ with the entropy of its corresponding image patch, normalized via the softmax function:", + "bbox": [ + 89, + 422, + 483, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} = \\sum_ {k \\in \\mathcal {K}} A _ {k i} + \\alpha \\cdot \\frac {\\exp (H (V _ {i}))}{\\sum_ {j = 1} ^ {N} \\exp (H (V _ {j}))}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 489, + 480, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{i}$ denotes the importance score of token $i$ , $A_{ki}$ is the attention weight from token $k$ to token $i$ , $H(V_{i})$ is the entropy of the image patch corresponding to token $i$ , $\\alpha$ is a hyperparameter that determines the entropy term's contribution, and $N$ is the total number of tokens. This approach mitigates biases from redundant tokens and leverages the visual information of image patches, thereby facilitating effective estimation of visual token importance.", + "bbox": [ + 89, + 537, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Progressive Pruning and Token Recycling", + "text_level": 1, + "bbox": [ + 89, + 667, + 446, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous studies [37, 51] indicate that visual tokens play a more critical role in the shallower layers of LMMs, whereas redundancy tends to increase in deeper layers. Based on this, we adopt a progressive pruning strategy. To balance the pruning process's simplicity with its fine-grained configuration, we evenly divide the LMM into three stages. At the beginning of each stage, a predetermined fraction of visual tokens with higher importance scores (as defined in Sec. 3.1) is retained according to the stage-specific retention ratios $R = [R_{1}, R_{2}, R_{3}]$ , while the rest are pruned. Notably, the position IDs of visual tokens remain unchanged following pruning, preserving the original spatial structure of the visual input. Following the previous work [37], we compute the average visual token retention rate across the entire", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMM as follows:", + "bbox": [ + 513, + 90, + 633, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {R} = \\frac {R _ {1} \\cdot L _ {1} + R _ {1} \\cdot R _ {2} \\cdot L _ {2} + R _ {1} \\cdot R _ {2} \\cdot R _ {3} \\cdot L _ {3}}{L}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 114, + 903, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{1}, L_{2}$ , and $L_{3}$ represent the number of layers in the three stages, respectively, and $L$ denotes the total number of layers in the LMM.", + "bbox": [ + 511, + 156, + 903, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent the loss of any small but potentially significant information during the initial pruning process (before visual tokens are fed into the LLM), we propose a token merging and recycling strategy to compactly represent redundant visual information. Specifically, We use a square grid (with each cell having side length $a$ ) to group pruned tokens that fall into the same grid cell, as shown in Fig. 2 (b). Within each cell, the pruned tokens are fused into a single token by computing a weighted average of their representations, using their importance scores as weights. This fused token then replaces the pruned token with the highest importance score in that cell and is incorporated into the set of retained tokens.", + "bbox": [ + 511, + 202, + 903, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, a token $t_i$ with spatial coordinates $(x_i, y_i)$ belongs to a grid cell $\\mathcal{G}_{p,q}$ if", + "bbox": [ + 511, + 398, + 903, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lfloor x _ {i} / a \\rfloor = p \\quad \\text {a n d} \\quad \\lfloor y _ {i} / a \\rfloor = q, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 446, + 903, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ and $q$ are the row and column indices of the grid cell, respectively. Suppose there are $k$ pruned tokens in $\\mathcal{G}_{p,q}$ with corresponding importance scores $I_{1}, I_{2}, \\ldots, I_{k}$ and token representations $\\mathbf{t}_1, \\mathbf{t}_2, \\ldots, \\mathbf{t}_k$ . The fused token $\\mathbf{t}_{\\mathrm{merged}}^{p,q}$ is computed as:", + "bbox": [ + 511, + 472, + 903, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {\\text {m e r g e d}} ^ {p, q} = \\frac {\\sum_ {i = 1} ^ {k} I _ {i} \\cdot \\mathbf {t} _ {i}}{\\sum_ {i = 1} ^ {k} I _ {i}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 561, + 903, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then assign $\\mathbf{t}_{\\mathrm{merged}}^{p,q}$ to the position of the pruned token with the highest importance score $I_{\\mathrm{max}}$ in $\\mathcal{G}_{p,q}$ , counting it among the retained tokens. By combining token pruning with this recycling strategy, we reduce token redundancy while avoiding potential information loss.", + "bbox": [ + 511, + 608, + 903, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pruning Strategy Optimization", + "text_level": 1, + "bbox": [ + 511, + 696, + 787, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pruning strategies described in Sec. 3.1 and Sec. 3.2 involve several key hyperparameters that directly affect the performance of LMMs after token pruning. Properly defining these hyperparameters is crucial for preserving model performance. However, existing approaches often rely on manually designed pruning strategies and apply the same strategy across different LMMs, without considering the unique characteristics of each model. This coarse-grained approach can lead to significant performance degradation due to its lack of fitness for different LMMs, leaving the task of designing pruning strategies tailored to specific LMMs as a formidable challenge.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23927", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous interpretability studies of LMMs [15, 53] offer crucial insights into their internal mechanisms. Specifically, these works reveal that in the lower and middle layers of LMMs, visual information from visual tokens is aggregated into the corresponding query text tokens. In the higher layers, the multimodal representation encoded in the query text tokens is further progressively propagated to the final position of the input sequence, ultimately influencing the subsequent inference process.", + "bbox": [ + 89, + 90, + 483, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by this insight, we propose a framework that requires only a small amount of unlabeled data and leverages the internal flow of visual information within LMMs to search for the superior pruning strategy tailored to the characteristics of different LMMs, as shown in Fig. 2 (2). Specifically, we recast the task of designing the superior pruning strategy as an optimization problem, aiming to minimize the discrepancy in visual information flow with and without applying a pruning strategy to perform visual token pruning. In this framework, we treat the final token in the last layer as the representative outcome of the visual information flow. We use cosine similarity to measure the similarity of the final token representation with and without visual token pruning. A higher similarity indicates that the discrepancy between the visual information flow with and without pruning is smaller. Accordingly, we define the following optimization objective:", + "bbox": [ + 91, + 227, + 485, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {s \\in \\mathcal {S}} f (s) = \\operatorname {C o s i n e S i m} \\left(h _ {f}, g _ {s} \\left(h _ {f}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 496, + 482, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $h_f$ represents the representation of the final token in the last layer before pruning, and $g_s(h_f)$ represents the final token representation after applying the pruning strategy $s$ (with $g_s$ modeling its effect on the token feature), and $S$ is the solution space.", + "bbox": [ + 89, + 529, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To efficiently search for the superior pruning strategy $s$ , we employ Bayesian optimization, which systematically explores the hyperparameter space—including the threshold sensitivity $t$ , the entropy weight $\\alpha$ , the grid size $a$ , and the pruning ratios $R_{1}, R_{2}$ , and $R_{3}$ —to maximize the target function $f(s)$ . In short, Bayesian optimization constructs a surrogate model to approximate the objective function and employs an acquisition function to balance exploration and exploitation, thereby efficiently guiding the search for promising hyperparameter settings. The details of this optimization process are presented in Algorithm 1.", + "bbox": [ + 89, + 604, + 483, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 784, + 223, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our approach on various LMMs across diverse image and video benchmarks, followed by an efficiency analysis and an ablation study of each component. Finally, we illustrate how our method affects discrepancies in the visual information flow with and without pruning, offering deeper insights into our approach.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 VFlowOpt with Bayesian Optimization", + "bbox": [ + 514, + 90, + 864, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: LMM $\\theta$ , data samples $\\mathcal{D}$ , number of Bayesian optimization iterations $T$ , computation budget $\\overline{R}$ , target function $f(\\cdot)$ : the sum of cosine similarities between the last token in the last layer of the LMM $\\theta$ with and without visual token pruning, computed over all data samples $\\mathcal{D}$ . \nOutput: retention rates $R_{1}, R_{2}, R_{3}$ ; threshold sensitivity $t$ ; entropy weight in importance score $\\alpha$ ; grid size $a$ .", + "bbox": [ + 514, + 109, + 906, + 229 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize a Gaussian Process model $\\mathcal{GP}$", + "2: Define the acquisition function $A(\\cdot)$ (Expected Improvement is adopted)", + "3: Uniformly sample initial points:" + ], + "bbox": [ + 524, + 231, + 903, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {0} = \\left\\{\\left(R _ {1}, R _ {2}, t, \\alpha , a\\right) \\mid R _ {v}, R _ {1}, t, \\alpha , a \\in \\text {v a l i d r a n g e s} \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 303, + 915, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4: for all $x \\in X_0$ do", + "bbox": [ + 524, + 332, + 663, + 345 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5: Calculate $R_{3}$ using the constraint: $\\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L$", + "6: Form pruning strategy: $s = (R_1, R_2, R_3, t, \\alpha, a)$", + "7: Evaluate the target function $f(s \\mid \\theta, \\mathcal{D})$" + ], + "bbox": [ + 524, + 347, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "8: end for", + "bbox": [ + 524, + 407, + 596, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "9: for $n = 0$ to $T - 1$ do", + "bbox": [ + 524, + 422, + 692, + 436 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: Fit $\\mathcal{GP}$ to the observed data $(S_{n},f(S_{n}\\mid \\theta ,\\mathcal{D}))$", + "11: Select the next point: $x_{n + 1}\\gets \\arg \\max_xA(x;\\mathcal{GP})$", + "12: Calculate $R_{3}$ using the constraint: $\\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L$", + "13: Form pruning strategy: $s_{n+1} = (R_1, R_2, R_3, t, \\alpha, a)_{n+1}$", + "14: Update $S_{n + 1}\\gets S_n\\cup \\{s_{n + 1}\\}$", + "15: end for", + "16: return $(R_{1}, R_{2}, R_{3}, t, \\alpha, a)^{*}$ that maximize $f$" + ], + "bbox": [ + 519, + 436, + 903, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setting", + "text_level": 1, + "bbox": [ + 513, + 590, + 699, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on ten image-based multimodal benchmarks: GQA [1], VizWiz [13], ScienceQA-IMG [26], TextVQA [30], ChartQA [27], POPE [19], MME [10], MMBench [25], MMStar [6], and DocVQA [28]. For video understanding, we adopt two datasets—SeedBench (video) [16] and VideoMME [11], where VideoMME is partitioned by video length into short, medium, and long subsets. Further details are presented in the Appendix.", + "bbox": [ + 511, + 613, + 906, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Architectures. We integrate VFlowOpt into multiple LMMs, including LLaVA-OneVision-7B, LLaVA-NeXT-7B [23], and Qwen2-VL-7B [34]. These models employ various vision encoders (SigLIP, CLIP, and a ViT designed for Qwen2-VL) and LLM backbones (Qwen2-7B and Vicuna-7B). To prevent out-of-memory issues in Qwen2-VL, we set maxPixels = 3000000.", + "bbox": [ + 511, + 750, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Methods. We compare VFlowOpt against five baseline methods: FastV [7], FitPrune [42], PDrop [36], SparseVLM [52], and VisionZip [40]. FastV,", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23928", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c08e4252dfd62ec3e56072a83d39363f69a64d4c10b6a680c0e600cc68ee8b4b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMMStarMMEMMBSQAPOPEGQAVizWiz\\(VQA^{Text}\\)ChartQADocVQAAvg.
Upper Bound, 100% Tokens
Vanilla61.7158180.895.889.162.260.476.080.087.5100%
Retain 50% Tokens (↓ 50%)
FastV (ECCV24)58.9154979.492.887.961.561.172.568.684.096.5%
FitPrune (AAAI25)4.596.61.08.618.816.26.837.637.730.721.4%
PDrop (CVPR25)37.1628.632.164.881.344.248.963.762.762.271.3%
SparseVLM (2024.10)59.8157780.594.188.161.960.473.970.580.897.1%
VisionZip (CVPR25)60.4158780.394.689.362.759.874.275.488.498.9%
VFlowOpt61.3159181.195.489.462.460.075.177.890.099.9%
Retain 25% Tokens (↓ 75%)
FastV (ECCV24)54.0153977.088.683.858.261.058.342.762.986.4%
FitPrune (AAAI25)5.292.61.15.920.214.55.129.727.323.217.6%
PDrop (CVPR25)36.8574.431.362.878.042.247.156.347.547.065.2%
SparseVLM (2024.10)56.6152078.790.387.259.760.866.354.066.690.5%
VisionZip (CVPR25)54.6156278.990.488.861.060.470.066.379.694.3%
VFlowOpt57.8157079.992.389.161.260.472.569.182.396.3%
Retain 10% Tokens (↓ 90%)
FastV (ECCV24)46.0120970.181.777.051.556.435.621.333.269.7%
FitPrune (AAAI25)5.427.00.93.52.34.60.76.95.96.95.1%
PDrop (CVPR25)31.8456.322.559.065.935.743.333.719.618.549.9%
SparseVLM (2024.10)45.1119171.883.780.054.656.439.837.639.674.0%
VisionZip (CVPR25)49.5138974.886.286.157.256.856.446.149.082.1%
VFlowOpt52.0146475.188.485.257.357.360.253.656.185.5%
", + "bbox": [ + 99, + 88, + 898, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparison on LLaVA-OneVision-7B under different token retention conditions. \"Avg.\" refers to average accuracy on 10 benchmarks. For each reduction ratio, the best average performance is shown in bold.", + "bbox": [ + 89, + 424, + 906, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FitPrune, PDrop, and SparseVLM rely on text-visual attention in the LLM to prune visual tokens but differ as follows: FastV performs a single pruning step after the second LLM layer; FitPrune prunes tokens in each layer according to predefined, layer-specific ratios; PDrop conducts four pruning steps within the LLM, using the same retention ratio each time; and SparseVLM uses the attention weights of preselected text tokens to evaluate the importance of visual tokens. VisionZip, in contrast, determines token importance based on the [cls] token's attention; for models lacking a [cls] token, we follow VisionZip's original procedure by computing the average attention each token receives from every other token in the ViTs.", + "bbox": [ + 88, + 474, + 482, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For LLaVA-OneVision-7B and Qwen2-VL-7B, we perform token pruning at three distinct points: before the LLM, and after the 9th and 18th layers. For LLaVA-NeXT-7B, pruning is conducted before the LLM, and again after the 10th and 20th layers. During optimization, we sample 30 unlabeled instances from each model's training datasets; for models without publicly available training dataset, we instead use random samples from the LLaVA-OneVision training set. The optimization is performed for a total of 50 iterations. For LMMs that modify visual tokens output by the vision encoder (e.g., unpadding and interpolation in LLaVA-OneVision, and unpadding in LLaVA-NeXT), we apply corresponding transformations to the importance maps so they remain fully aligned with the final visual tokens.", + "bbox": [ + 89, + 671, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Understanding Tasks", + "text_level": 1, + "bbox": [ + 511, + 474, + 764, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed VFlowOpt on image understanding benchmarks with LLaVA-OneVision-7B using various pruning ratios, and present the results in Tab. 1. Compared to other baselines, VFlowOpt consistently maintains superior accuracy across different levels of token pruning. With $50\\%$ token retention, VFlowOpt achieves $99.9\\%$ of the original performance, exceeding the second-best approach by $1.0\\%$ . This negligible performance drop underscores the method's strong potential for practical deployment. Under more extreme pruning conditions with very few retained tokens, VFlowOpt's advantage becomes more pronounced. When only $10\\%$ of the visual tokens are retained, VFlowOpt preserves $85.5\\%$ of the original performance, surpassing the second-best approach by $3.3\\%$ . This finding suggests that VFlowOpt can effectively leverage limited visual information to maintain high performance under strict computational budgets. When applied to LLaVA-OneVision, both FitPrune and PDrop disrupt the model's normal output, indicating their inability to devise effective pruning strategies for different architectures. In contrast, VFlowOpt demonstrates a clear advantage by tailoring its pruning strategy to each model's unique characteristics.", + "bbox": [ + 509, + 502, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further validate VFlowOpt's generalization capability, we evaluate it on LLaVA-NeXT-7B and Qwen2-VL-7B. As shown in Tab. 2 and Tab. 3, VFlowOpt once again delivers the top results on these LMMs. When retaining $25\\%$", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23929", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89770b1edab563629f5768444b850141eae77d84f4414fa572579c778afb191b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla37.6151967.470.186.564.2100%
Retain 25% Tokens (↓ 75%)
FastV35.1147765.667.483.160.495.7%
VisionZip35.8150165.467.986.761.597.3%
VFlowOpt37.0151467.067.787.662.698.8%
Retain 10% Tokens (↓ 90%)
FastV29.2128261.663.871.755.985.7%
VisionZip32.6137861.567.183.557.091.6%
VFlowOpt35.1139362.967.483.657.393.4%
", + "bbox": [ + 96, + 89, + 486, + 243 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36abdb835487ebf46093e346e5ae6ca0c9428afb043cb14d96319d2e489f57f2.jpg", + "table_caption": [ + "Table 2. Comparative experiments on LLaVA-NeXT-7B." + ], + "table_footnote": [], + "table_body": "
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla57.5168080.384.788.462.2100%
Retain 25% Tokens (↓ 75%)
FastV54.5159776.379.081.857.293.8
VisionZip55.2161878.981.386.960.096.8
VFlowOpt55.9165979.881.587.160.497.8
Retain 10% Tokens (↓ 90%)
FastV44.9140570.075.575.851.684.4
VisionZip49.3151876.578.284.354.190.9
VFlowOpt51.0159178.078.684.554.892.8
", + "bbox": [ + 96, + 271, + 482, + 426 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/40d07dfea5d2b525bc1b5d6e804cc5d1b171db9c191f8b647fb6c7255033c3b0.jpg", + "table_caption": [ + "Table 3. Comparative experiments on Qwen2-VL-7B." + ], + "table_footnote": [], + "table_body": "
MethodSeedBench (video)VideoMMEAvg.
SML
Upper Bound, 100% Tokens
Vanilla56.970.554.649.5100%
Retain 25% Tokens (↓75%)
FastV54.766.953.247.796.2%
VisionZip56.468.355.349.099.0%
VFlowOpt56.868.955.849.2100%
Retain 10% Tokens (↓90%)
FastV48.753.747.642.383.6%
VisionZip55.059.751.846.392.4%
VFlowOpt55.563.352.748.695.5%
", + "bbox": [ + 96, + 452, + 480, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparative experiments on video understanding tasks.", + "bbox": [ + 94, + 627, + 477, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the tokens, VFlowOpt preserves $98.8\\%$ of the original performance on LLaVA-NeXT-7B and $97.8\\%$ on Qwen2-VL-7B. Even under the stringent condition of retaining only $10\\%$ of the tokens, VFlowOpt maintains $93.4\\%$ of the performance on LLaVA-NeXT-7B and $92.8\\%$ on Qwen2-VL-7B, demonstrating its strong generalizability.", + "bbox": [ + 89, + 657, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Generalization to Video Tasks", + "text_level": 1, + "bbox": [ + 89, + 757, + 356, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further explore VFlowOpt's generalization across different modalities by evaluating it on video benchmarks using LLaVA-OneVision-7B. As shown in Tab. 4, VFlowOpt surpasses other baselines under various token retention ratios. Notably, when only $25\\%$ of the tokens are preserved, VFlowOpt incurs virtually no performance loss on LLaVA-OneVision-7B, maintaining $100\\%$ of its original performance. Even under the more stringent condition of retain", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing just $10\\%$ of the tokens, VFlowOpt still retains $95.5\\%$ of the original performance, underscoring its robust generalization in the video domain.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Efficiency Analysis", + "text_level": 1, + "bbox": [ + 511, + 147, + 694, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the efficiency of VFlowOpt by conducting a comparative study on LLaVA-OneVision-7B running on a single NVIDIA A100-SXM4-80GB GPU, focusing on FLOPs, KV-Cache memory usage, and inference latency. The efficiency analysis of other baseline methods is provided in the Appendix.", + "bbox": [ + 511, + 170, + 903, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We assess various token pruning ratios on the MME benchmark, measuring overall performance and average efficiency metrics across all samples. As shown in Tab. 5, VFlowOpt significantly enhances LLaVA-OneVision-7B's computational efficiency by reducing FLOPs, shrinking KV-Cache memory usage, and accelerating inference speed. Notably, at a $50\\%$ token pruning ratio, VFlowOpt achieves a $49.5\\%$ reduction in KV-Cache memory and a $1.8\\times$ speedup in inference, while boosting performance by $0.7\\%$ . At a more aggressive $75\\%$ pruning ratio, we observe a $74.2\\%$ reduction in KV-Cache memory and a $3.1\\times$ speedup, with only a $0.7\\%$ performance drop. Under extremely tight computational budgets, pruning $90\\%$ of visual tokens reduces the KV-Cache memory footprint by $89\\%$ , accelerates inference by $3.8\\times$ , and degrades performance by only $7.4\\%$ .", + "bbox": [ + 511, + 261, + 906, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Such reductions in inference latency substantially benefit user-facing applications demanding real-time performance, such as autonomous driving and robotics. Moreover, by significantly shrinking the KV-Cache, large-batch inference on LMMs can accommodate more user requests simultaneously, thereby substantially reducing overall inference costs. In summary, VFlowOpt uses far less GPU memory and delivers faster inference while preserving model performance, offering a highly practical solution for efficiently deploying LMMs in real-world scenarios.", + "bbox": [ + 511, + 488, + 908, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 650, + 666, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To confirm the contributions of each component in our token pruning strategy, we conduct ablation experiments on LLaVA-OneVision-7B under a computational budget of retaining $25\\%$ of tokens. As shown in Tab. 6, removing importance calibration (i.e., directly using the mean attention over all tokens received by each visual token as its importance score), omitting token recycling, or discarding progressive pruning (i.e., maintaining the same number of tokens at each layer) leads to noticeable performance degradation across MMStar, MMBench, and SQA. By contrast, the complete VFlowOpt method consistently exhibits minimal performance loss, highlighting the effectiveness of all its components in pruning redundant tokens while preserving essential visual information.", + "bbox": [ + 511, + 672, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further examine how the number of samples and the", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23930", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3af5b2956741a6de8945bd4b3d799f7aa67ca963a5c5c8b6c2c586d111305f97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsToken ReductionFLOPs ↓ (T)ΔLatency ↓ (ms)ΔKV Cache ↓ (MB)ΔPerformance ↑Δ
LLaVA-OneVision-7B-71.4-1040.1-1786.4-1581-
+ VFlowOpt50%37.2-48.0%584.2-43.8%902.8-49.5%1591+0.6%
75%19.1-73.2%328.5-68.4%460.6-74.2%1570-0.7%
90%7.7-89.2%272.1-73.8%197.1-89.0%1464-7.4%
", + "bbox": [ + 143, + 88, + 854, + 180 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7df6f58322eeff0d18022495fcc373c29148d80151a63306e98a2dbebc3eff21.jpg", + "table_caption": [ + "Table 5. Efficiency analysis of LLaVA-OneVision-7B with VFlowOpt. The detailed metric includes computation (FLOPs), latency, and KV-Cache memory. $(\\Delta)$ denotes the reduction ratio." + ], + "table_footnote": [], + "table_body": "
Pruning StrategyMMStarMMBenchSQA
VFlowOpt57.879.992.3
w/o Importance Calibration56.279.491.8
w/o Token Merging57.679.891.9
w/o Progressive Pruning56.079.191.0
", + "bbox": [ + 117, + 232, + 455, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation studies of pruning strategy components.", + "bbox": [ + 114, + 319, + 457, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2747e4b8f4cda23aeadfcdea5c5eba1512cdc543528df07152841897b6e5eb9c.jpg", + "image_caption": [ + "Figure 4. Relationships of sample size (left) and optimization steps (right) with optimization time and final performance." + ], + "image_footnote": [], + "bbox": [ + 94, + 343, + 475, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "number of optimization iterations in the Bayesian optimization procedure influence both optimization time and final performance. As illustrated in Fig. 4, the optimization time increases linearly with the number of samples and steps. Using 30 samples and 50 steps strikes an effective balance between time efficiency and performance, requiring only about 30 minutes to reach an exceptional result. More ablation results are provided in the supplementary material.", + "bbox": [ + 89, + 506, + 483, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Visual Information Flow Analysis", + "text_level": 1, + "bbox": [ + 89, + 636, + 385, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further investigate how Bayesian optimization formulates the superior pruning strategy by examining its impact on visual information flow. Specifically, we compare the randomly initialized pruning strategy derived from the Bayesian optimization process with its final optimized strategy, measuring their respective effects on the text tokens that follow the visual tokens at each LLM layer. Because these text tokens can receive information from the visual tokens via the LLM's unidirectional attention mechanism, examining differences in these tokens with and without pruning effectively reveals how visual information flow is altered. As shown in Fig. 5, under the same computational budget, the initial strategy causes a pronounced discrepancy between text token representations with and without pruning, indicating that it significantly disrupts the visual information flow. In contrast, when using the optimized strategy,", + "bbox": [ + 89, + 659, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2a427acb1d1889fcb27ddbe93caecc35d46160cebc60341d0b621e7434e3262.jpg", + "image_caption": [ + "Figure 5. Visualization of the discrepancies in visual information flow corresponding to the pruning strategies before and after optimization, along with an example of visual question answering." + ], + "image_footnote": [], + "bbox": [ + 517, + 232, + 898, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the text tokens remain much more similar to those observed without visual token pruning. This important finding suggests that the optimization process yields a pruning strategy that preserves crucial visual information by minimizing the difference in visual information flow with and without pruning, thereby retaining LMM performance to the greatest extent possible.", + "bbox": [ + 511, + 542, + 906, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 661, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we address the pressing challenge of computational inefficiency in LMMs by introducing a general framework for visual token pruning. We reformulate pruning as an optimization problem focused on minimizing the divergence of visual information flow with and without pruning, thereby facilitating a tailored pruning strategy for diverse LMMs. By leveraging calibrated attention in ViTs to evaluate token significance and merging pruned tokens using importance-based weighting, we preserve critical information during inference without requiring retraining or manual tuning. Extensive experiments underscore the efficiency and generalizability of our approach, making it a practical solution for deploying LMMs in real-world settings, particularly in resource-constrained environments.", + "bbox": [ + 511, + 686, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23931", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 116, + 480, + 148 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 165, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 5", + "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 1, 2", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2", + "[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 2", + "[5] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Llavolta: Efficient multi-modal models via stage-wise visual context compression. arXiv preprint arXiv:2406.20092, 2024. 2", + "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 5", + "[7] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for" + ], + "bbox": [ + 99, + 190, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5", + "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 2", + "[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving, 2023. 1", + "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5", + "[11] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5", + "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering, 2017. 1, 2", + "[13] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5", + "[14] Lianyu Hu, Fanhua Shang, Liang Wan, and Wei Feng. illava: An image is worth fewer than 1/3 input tokens in large multimodal models. arXiv preprint arXiv:2412.06263, 2024. 3", + "[15] Omri Kaduri, Shai Bagon, and Tali Dekel. What's in the image? a deep-dive into the vision of vision language models. arXiv preprint arXiv:2411.17491, 2024. 5", + "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5", + "[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 2", + "[18] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation, 2023. 1", + "[19] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5", + "[20] Yuelei Li, Ge Yan, Annabella Macaluso, Mazeyu Ji, Xueyan Zou, and Xiaolong Wang. Integrating lmm planners and 3d" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23932", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "skill policies for generalizable manipulation. arXiv preprint arXiv:2501.18733, 2025. 1", + "[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2", + "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2", + "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5", + "[24] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation, 2024. 1", + "[25] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5", + "[26] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5", + "[27] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5", + "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2", + "[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5", + "[31] Haoxiang Tian, Xingshuo Han, Guoquan Wu, Yuan Zhou, Shuo Li, Jun Wei, Dan Ye, Wei Wang, and Tianwei Zhang. An llm-enhanced multi-objective evolutionary search for autonomous driving test scenario generation. arXiv preprint arXiv:2406.10857, 2024. 1", + "[32] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn," + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 2", + "[33] Ao Wang, Fengyuan Sun, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. [cls] token tells everything needed for training-free efficient mllms. arXiv preprint arXiv:2412.05819, 2024. 1, 2", + "[34] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5", + "[35] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. Drivemlm: Aligning multimodal large language models with behavioral planning states for autonomous driving, 2023. 1", + "[36] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 1, 2, 5", + "[37] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 4", + "[38] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2", + "[39] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin," + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23933", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2", + "[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 1, 2, 5", + "[41] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 3", + "[42] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 5", + "[43] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. arXiv preprint arXiv:2412.00447, 2024. 2", + "[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 1, 2", + "[45] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark, 2024. 1", + "[46] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 2", + "[47] Jun Zhang, Desen Meng, Ji Qi, Zhenpeng Huang, Tao Wu, and Limin Wang. p-mod: Building mixture-of-depths mllms via progressive ratio decay. arXiv preprint arXiv:2412.04449, 2024. 2", + "[48] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2", + "[49] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster, 2024. 1, 2", + "[50] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. In The Thirteenth International Conference on Learning Representations. 2", + "[51] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hao Tang, and Jieping Ye. From redundancy to relevance: Information flow in lvlms across reasoning tasks. arXiv preprint arXiv:2406.06579, 2024. 4", + "[52] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 1, 5", + "[53] Zhi Zhang, Srishti Yadav, Fengze Han, and Ekaterina Shutova. Cross-modal information flow in multimodal large language models. arXiv preprint arXiv:2411.18620, 2024. 5" + ], + "bbox": [ + 516, + 92, + 903, + 247 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23934", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_model.json b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1dbf9cc9aaff240e8ffb6a21158b6a816113e8cd --- /dev/null +++ b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.89, + 0.178 + ], + "angle": 0, + "content": "VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.202, + 0.877, + 0.223 + ], + "angle": 0, + "content": "Sihan Yang1 Runsen Xu1,2 Chenhang Cui3 Tai Wang1† Dahua Lin1,2 Jiangmiao Pang1†" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.224, + 0.815, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai AI Laboratory \\(^{2}\\) The Chinese University of Hong Kong \\(^{3}\\) National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.241, + 0.774, + 0.257 + ], + "angle": 0, + "content": "† Corresponding Author taiwang.me@gmail.com pangjiangmiao@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.78 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) excel in visual-language tasks by leveraging numerous visual tokens for fine-grained visual information, but this token redundancy results in significant computational costs. Previous research aimed at reducing visual tokens during inference typically leverages importance maps derived from attention scores among vision-only tokens or vision-language tokens to prune tokens across one or multiple pruning stages. Despite this progress, pruning frameworks and strategies remain simplistic and insufficiently explored, often resulting in substantial performance degradation. In this paper, we propose VFlowOpt, a token pruning framework that introduces an importance map derivation process and a progressive pruning module with a recycling mechanism. The hyperparameters of its pruning strategy are further optimized by a visual information flow-guided method. Specifically, we compute an importance map for image tokens based on their attention-derived context relevance and patch-level information entropy. We then decide which tokens to retain or prune and aggregate the pruned ones as recycled tokens to avoid potential information loss. Finally, we apply a visual information flow-guided method that regards the last token in the LMM as the most representative signal of text-visual interactions. This method minimizes the discrepancy between token representations in LMMs with and without pruning, thereby enabling superior pruning strategies tailored to different LMMs. Experiments demonstrate that VFlowOpt can prune \\(90\\%\\) of visual tokens while maintaining comparable performance, leading to an \\(89\\%\\) reduction in KV-Cache memory and \\(3.8 \\times\\) faster inference." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.813, + 0.222, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) [3, 8, 21, 22] have attained remarkable performance in tasks such as visual question answering [2, 12] and multimodal reasoning [44, 45], making them indispensable for applications like au" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.281, + 0.905, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.467, + 0.909, + 0.58 + ], + "angle": 0, + "content": "Figure 1. VFlowOpt Performance and Efficiency. Our VFlowOpt significantly outperforms other LMM token reduction methods across three general benchmarks (MME, MMStar, and MMBench) on LLaVA-OneVision-7B. It achieves zero performance degradation when reducing \\(50\\%\\) of visual tokens. Furthermore, when retaining only \\(10\\%\\) of the tokens, it achieves \\(90\\%\\) of the original performance while reducing KV-Cache memory usage by \\(89\\%\\) and accelerating inference by \\(3.8\\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.595, + 0.907, + 0.746 + ], + "angle": 0, + "content": "tonomous driving [9, 31, 35] and robotics [18, 20, 24]. To capture fine-grained visual details, modern LMMs treat images as token sequences, with models such as LLaVA-1.5 [21] processing hundreds of tokens and LLaVA-OneVision [17] handling up to several thousand tokens. Nonetheless, the increasing quantity of visual tokens significantly magnifies computational cost, memory usage, and inference time, creating a critical bottleneck for LMM deployment, particularly in resource-constrained environments or latency-sensitive scenarios." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To alleviate these constraints, recent work explores various methods for reducing the number of visual tokens during inference. Most existing approaches evaluate the importance of visual tokens in LMMs using attention-based importance maps derived either from vision-only tokens [40, 49] or from text-visual interactions [36, 52]. Following a predefined pruning ratio, visual tokens are pruned in a single stage [33] or progressively across multiple stages [42], often with heuristic strategies applied uniformly across different LMMs. Although promising, these methods are sim" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23924" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "plastic and underexplored, with coarse-grained pruning often causing significant performance drops due to lacking fitness for different LMMs models' characteristics." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.139, + 0.487, + 0.545 + ], + "angle": 0, + "content": "In response to these limitations, we propose a novel token pruning framework, VFlowOpt, that more accurately estimates token importance by integrating attention calibration and entropy, and reduces information loss through token recycling and progressive pruning strategy with fine-grained pruning ratios. Furthermore, by customizing pruning strategies for different models, our framework better preserves model performance. Specifically, the pruning framework consists of importance map computation and a progressive pruning module with a recycling mechanism. We decompose the importance map of image tokens into two aspects: the importance of tokens in the visual context, reflected by their attention maps, and the richness of visual information in each image patch, captured by its information entropy. Based on the weighted summation of these two factors, we determine which tokens to retain or prune across multiple pruning stages and introduce a recycling mechanism to aggregate pruned tokens to avoid potential information loss. Built with this pruning strategy with several hyperparameters, a key ingredient is the visual information flow-guided method to optimize the pruning strategy, which treats the last token in each pruning stage as the most representative signal of text-visual interactions during inference. By minimizing the difference between token representations in LMMs with and without pruning, VFlowOpt ultimately delivers superior pruning strategies specifically tailored to different LMMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Comprehensive experiments on multiple vision-language benchmarks validate the efficacy of VFlowOpt. In particular, it substantially lowers computational cost while preserving competitive performance. As shown in Fig. 1, VFlowOpt can prune \\(50\\%\\) of visual tokens with negligible performance loss. Moreover, it can prune \\(90\\%\\) of visual tokens while maintaining \\(90\\%\\) of the model's performance, resulting in an \\(89\\%\\) reduction in KV-Cache memory usage and a \\(3.8 \\times\\) speedup in inference. These experimental results highlight its effectiveness as a practical and efficient solution for deploying LMMs in resource-constrained environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.233, + 0.761 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.33, + 0.787 + ], + "angle": 0, + "content": "2.1. Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) [3, 8, 21, 22] extend the reasoning capabilities of Large Language Models (LLMs) [4, 32, 38, 39] to vision-language tasks by integrating a pre-trained vision encoder [29, 46] with a language model, linked by an alignment module such as an MLP, or a query-based network. This design transforms visual inputs into token sequences that the LLM can pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.334 + ], + "angle": 0, + "content": "cess, facilitating multimodal prompts for tasks like visual question answering [2, 6, 12, 44]. To enhance performance, advanced LMMs, such as LLaVA-OneVision [17] and Qwen2-VL [34], can encode higher-resolution images into more image tokens, thereby capturing more granular visual details. However, as image resolution grows, the number of visual tokens rises exponentially, leading to substantially higher computational costs. For example, LLaVA-1.5 [21] processes \\(336 \\times 336\\) images into 576 tokens, whereas LLaVA-OneVision can handle \\(1152 \\times 1152\\) images, producing 7,290 tokens. This challenge becomes even more pronounced in video-based models like LongVA [48], which must process tokens across numerous frames. Fine-grained visual tokenization boosts LMM performance but poses an inference bottleneck, driving efforts to balance performance and cost through token reduction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.342, + 0.766, + 0.357 + ], + "angle": 0, + "content": "2.2. Token Reduction for LMMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.909, + 0.742 + ], + "angle": 0, + "content": "Token reduction has become a key strategy for improving the efficiency of LMMs by mitigating the computational cost associated with extensive visual token sequences. Existing methods can be broadly classified into training-based and training-free approaches. Training-based methods, such as LLaVA-Mini [50] and LLaVolta [5], introduce additional modules during model training to compress visual tokens and preserve critical information, while approaches like ATP-LLaVA [43] and p-MoD [47] train pruning modules to dynamically retain important tokens across LLM layers. However, these methods demand substantial computational resources to retrain the models, limiting their real-world applicability. In contrast, training-free methods prune tokens without additional training, often leveraging attention mechanisms to identify and discard redundant tokens. For instance, FastV[7] and SparseVLM exploit text-visual attention to rank token importance, whereas FasterVLM [49], VisionZip [40], and VTC-CLS [33] rely on [CLS] token attention in the vision encoder to evaluate token importance and prune redundant tokens. FitPrune [42] and PDrop [36] propose a progressive multistage pruning strategy in LMMs to fully utilize visual information. However, the pruning strategies in these methods often lack adaptability to different LMMs, which frequently leads to significant performance drops." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.605, + 0.77 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this section, we introduce VFlowOpt, a framework that employs a pruning strategy to reduce redundant visual tokens during LMM inference while preserving essential visual information. This strategy involves multiple hyperparameters that our VFlowOpt framework automatically optimizes for proper configuration—an essential factor for maintaining model performance. Specifically, Section 3.1 explains how our method evaluates the importance of visual" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23925" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.093, + 0.381, + 0.107 + ], + "angle": 0, + "content": "(a) Token Importance Estimation" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.111, + 0.43, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.092, + 0.599, + 0.107 + ], + "angle": 0, + "content": "(b) Token Recycling" + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.114, + 0.605, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.272, + 0.271, + 0.289 + ], + "angle": 0, + "content": "(1) Inference Stage" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.291, + 0.611, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.089, + 0.895, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.415, + 0.856, + 0.433 + ], + "angle": 0, + "content": "(2) Optimization Stage" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.451, + 0.785, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Figure 2. Overview of VFlowOpt. (1) During inference, VFlowOpt first assesses the importance of visual tokens, based on which progressive token pruning is performed. After the initial pruning stage, the pruned tokens are merged and recycled. The pruning strategy used in this process is defined by the (2) Optimization Stage. (a) The importance map is computed by combining the attention of relatively important tokens with the entropy of image patches. (b) The pruned tokens are grouped into grid cells, where each cell has a side length of a. Within each grid cell, the pruned tokens are fused using a weighted average, with their importance values as weights, and then recycled. (2) VFlowOpt optimizes the pruning strategy by minimizing the discrepancy of the last token in the final layer of the LMM with and without applying the pruning strategy." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.577, + 0.465, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Figure 3. The attention of redundant tokens (marked in red) fails to reflect the importance of other tokens and instead focuses on similar tokens, such as background elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.84 + ], + "angle": 0, + "content": "tokens, Section 3.2 presents progressive pruning and token recycling strategies, and finally, Section 3.3 details how VFlowOpt customizes the superior pruning strategy (i.e., selects the pruning hyperparameters) for different LMMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.411, + 0.866 + ], + "angle": 0, + "content": "3.1. Visual Token Importance Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For a general approach to evaluate visual token importance, the previous works [14, 41] propose that the importance of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.804 + ], + "angle": 0, + "content": "visual tokens can be estimated using the average attention from all other tokens in ViTs. Although it has achieved some success, we find that redundant tokens (e.g., tokens corresponding to background regions) often assign disproportionately high attention to other similar redundant tokens (shown in Fig. 3), reducing the reliability of the importance estimation. To address this, we first identify relatively important tokens based on the attention they receive from all tokens (tokens receiving higher attention are considered relatively important), as shown in Fig. 2 (a). We then exclude redundant tokens and use the attention from these relatively important tokens as a more robust metric. Additionally, we incorporate the information entropy of image patches into the importance score to prioritize tokens corresponding to visually informative regions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.806, + 0.907, + 0.834 + ], + "angle": 0, + "content": "To identify relatively important visual tokens, we define a threshold:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.835, + 0.907, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\tau = t \\cdot \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} A _ {i j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the threshold, \\(A_{ij}\\) is the attention weight from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23926" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "token \\(i\\) to token \\(j\\) within a ViT layer, \\(N\\) is the total number of tokens, and \\(t\\) is a sensitivity hyperparameter. We then treat tokens with a total attention exceeding \\(\\tau\\) as relatively important:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.152, + 0.482, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathcal {K} = \\{j \\mid \\sum_ {i = 1} ^ {N} A _ {i j} > \\tau \\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.482, + 0.225 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{K}\\) denotes the set of indices of these relatively important tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.482, + 0.256 + ], + "angle": 0, + "content": "The information entropy of an image patch corresponding to token \\( i \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.267, + 0.482, + 0.307 + ], + "angle": 0, + "content": "\\[\nH \\left(V _ {i}\\right) = - \\sum_ {k = 0} ^ {L - 1} p _ {k} \\log p _ {k}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.483, + 0.421 + ], + "angle": 0, + "content": "where \\( V_{i} \\) is the image patch corresponding to token \\( i \\), \\( p_{k} \\) is the proportion of pixels in the patch with gray level \\( k \\), and \\( L \\) is the total number of possible gray levels (256 for 8-bit images). Here, the gray level is defined as the average of the RGB channel values, providing a scalar representation of pixel brightness. Higher entropy reflects greater diversity in pixel intensities and richer visual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.423, + 0.484, + 0.482 + ], + "angle": 0, + "content": "The importance score of token \\( i \\) is computed by combining the attention it receives from tokens in \\( \\mathcal{K} \\) with the entropy of its corresponding image patch, normalized via the softmax function:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.49, + 0.482, + 0.53 + ], + "angle": 0, + "content": "\\[\nI _ {i} = \\sum_ {k \\in \\mathcal {K}} A _ {k i} + \\alpha \\cdot \\frac {\\exp (H (V _ {i}))}{\\sum_ {j = 1} ^ {N} \\exp (H (V _ {j}))}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.659 + ], + "angle": 0, + "content": "where \\( I_{i} \\) denotes the importance score of token \\( i \\), \\( A_{ki} \\) is the attention weight from token \\( k \\) to token \\( i \\), \\( H(V_{i}) \\) is the entropy of the image patch corresponding to token \\( i \\), \\( \\alpha \\) is a hyperparameter that determines the entropy term's contribution, and \\( N \\) is the total number of tokens. This approach mitigates biases from redundant tokens and leverages the visual information of image patches, thereby facilitating effective estimation of visual token importance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.447, + 0.684 + ], + "angle": 0, + "content": "3.2. Progressive Pruning and Token Recycling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Previous studies [37, 51] indicate that visual tokens play a more critical role in the shallower layers of LMMs, whereas redundancy tends to increase in deeper layers. Based on this, we adopt a progressive pruning strategy. To balance the pruning process's simplicity with its fine-grained configuration, we evenly divide the LMM into three stages. At the beginning of each stage, a predetermined fraction of visual tokens with higher importance scores (as defined in Sec. 3.1) is retained according to the stage-specific retention ratios \\( R = [R_{1}, R_{2}, R_{3}] \\), while the rest are pruned. Notably, the position IDs of visual tokens remain unchanged following pruning, preserving the original spatial structure of the visual input. Following the previous work [37], we compute the average visual token retention rate across the entire" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.634, + 0.105 + ], + "angle": 0, + "content": "LMM as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.115, + 0.905, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\bar {R} = \\frac {R _ {1} \\cdot L _ {1} + R _ {1} \\cdot R _ {2} \\cdot L _ {2} + R _ {1} \\cdot R _ {2} \\cdot R _ {3} \\cdot L _ {3}}{L}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.157, + 0.905, + 0.202 + ], + "angle": 0, + "content": "where \\(L_{1}, L_{2}\\), and \\(L_{3}\\) represent the number of layers in the three stages, respectively, and \\(L\\) denotes the total number of layers in the LMM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.905, + 0.398 + ], + "angle": 0, + "content": "To prevent the loss of any small but potentially significant information during the initial pruning process (before visual tokens are fed into the LLM), we propose a token merging and recycling strategy to compactly represent redundant visual information. Specifically, We use a square grid (with each cell having side length \\(a\\)) to group pruned tokens that fall into the same grid cell, as shown in Fig. 2 (b). Within each cell, the pruned tokens are fused into a single token by computing a weighted average of their representations, using their importance scores as weights. This fused token then replaces the pruned token with the highest importance score in that cell and is incorporated into the set of retained tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.905, + 0.432 + ], + "angle": 0, + "content": "Formally, a token \\( t_i \\) with spatial coordinates \\( (x_i, y_i) \\) belongs to a grid cell \\( \\mathcal{G}_{p,q} \\) if" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.448, + 0.905, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\lfloor x _ {i} / a \\rfloor = p \\quad \\text {a n d} \\quad \\lfloor y _ {i} / a \\rfloor = q, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.905, + 0.548 + ], + "angle": 0, + "content": "where \\(p\\) and \\(q\\) are the row and column indices of the grid cell, respectively. Suppose there are \\(k\\) pruned tokens in \\(\\mathcal{G}_{p,q}\\) with corresponding importance scores \\(I_{1}, I_{2}, \\ldots, I_{k}\\) and token representations \\(\\mathbf{t}_1, \\mathbf{t}_2, \\ldots, \\mathbf{t}_k\\). The fused token \\(\\mathbf{t}_{\\mathrm{merged}}^{p,q}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {\\text {m e r g e d}} ^ {p, q} = \\frac {\\sum_ {i = 1} ^ {k} I _ {i} \\cdot \\mathbf {t} _ {i}}{\\sum_ {i = 1} ^ {k} I _ {i}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.905, + 0.686 + ], + "angle": 0, + "content": "We then assign \\(\\mathbf{t}_{\\mathrm{merged}}^{p,q}\\) to the position of the pruned token with the highest importance score \\(I_{\\mathrm{max}}\\) in \\(\\mathcal{G}_{p,q}\\), counting it among the retained tokens. By combining token pruning with this recycling strategy, we reduce token redundancy while avoiding potential information loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.788, + 0.714 + ], + "angle": 0, + "content": "3.3. Pruning Strategy Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The pruning strategies described in Sec. 3.1 and Sec. 3.2 involve several key hyperparameters that directly affect the performance of LMMs after token pruning. Properly defining these hyperparameters is crucial for preserving model performance. However, existing approaches often rely on manually designed pruning strategies and apply the same strategy across different LMMs, without considering the unique characteristics of each model. This coarse-grained approach can lead to significant performance degradation due to its lack of fitness for different LMMs, leaving the task of designing pruning strategies tailored to specific LMMs as a formidable challenge." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23927" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.227 + ], + "angle": 0, + "content": "Previous interpretability studies of LMMs [15, 53] offer crucial insights into their internal mechanisms. Specifically, these works reveal that in the lower and middle layers of LMMs, visual information from visual tokens is aggregated into the corresponding query text tokens. In the higher layers, the multimodal representation encoded in the query text tokens is further progressively propagated to the final position of the input sequence, ultimately influencing the subsequent inference process." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.486, + 0.485 + ], + "angle": 0, + "content": "Inspired by this insight, we propose a framework that requires only a small amount of unlabeled data and leverages the internal flow of visual information within LMMs to search for the superior pruning strategy tailored to the characteristics of different LMMs, as shown in Fig. 2 (2). Specifically, we recast the task of designing the superior pruning strategy as an optimization problem, aiming to minimize the discrepancy in visual information flow with and without applying a pruning strategy to perform visual token pruning. In this framework, we treat the final token in the last layer as the representative outcome of the visual information flow. We use cosine similarity to measure the similarity of the final token representation with and without visual token pruning. A higher similarity indicates that the discrepancy between the visual information flow with and without pruning is smaller. Accordingly, we define the following optimization objective:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.497, + 0.483, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\max _ {s \\in \\mathcal {S}} f (s) = \\operatorname {C o s i n e S i m} \\left(h _ {f}, g _ {s} \\left(h _ {f}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.605 + ], + "angle": 0, + "content": "Here, \\( h_f \\) represents the representation of the final token in the last layer before pruning, and \\( g_s(h_f) \\) represents the final token representation after applying the pruning strategy \\( s \\) (with \\( g_s \\) modeling its effect on the token feature), and \\( S \\) is the solution space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.772 + ], + "angle": 0, + "content": "To efficiently search for the superior pruning strategy \\( s \\), we employ Bayesian optimization, which systematically explores the hyperparameter space—including the threshold sensitivity \\( t \\), the entropy weight \\( \\alpha \\), the grid size \\( a \\), and the pruning ratios \\( R_{1}, R_{2} \\), and \\( R_{3} \\)—to maximize the target function \\( f(s) \\). In short, Bayesian optimization constructs a surrogate model to approximate the objective function and employs an acquisition function to balance exploration and exploitation, thereby efficiently guiding the search for promising hyperparameter settings. The details of this optimization process are presented in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.224, + 0.803 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this section, we evaluate our approach on various LMMs across diverse image and video benchmarks, followed by an efficiency analysis and an ablation study of each component. Finally, we illustrate how our method affects discrepancies in the visual information flow with and without pruning, offering deeper insights into our approach." + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.866, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 VFlowOpt with Bayesian Optimization" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.111, + 0.907, + 0.231 + ], + "angle": 0, + "content": "Input: LMM \\(\\theta\\), data samples \\(\\mathcal{D}\\), number of Bayesian optimization iterations \\(T\\), computation budget \\(\\overline{R}\\), target function \\(f(\\cdot)\\): the sum of cosine similarities between the last token in the last layer of the LMM \\(\\theta\\) with and without visual token pruning, computed over all data samples \\(\\mathcal{D}\\). \nOutput: retention rates \\(R_{1}, R_{2}, R_{3}\\); threshold sensitivity \\(t\\); entropy weight in importance score \\(\\alpha\\); grid size \\(a\\)." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.232, + 0.807, + 0.246 + ], + "angle": 0, + "content": "1: Initialize a Gaussian Process model \\(\\mathcal{GP}\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.248, + 0.905, + 0.277 + ], + "angle": 0, + "content": "2: Define the acquisition function \\( A(\\cdot) \\) (Expected Improvement is adopted)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.278, + 0.756, + 0.292 + ], + "angle": 0, + "content": "3: Uniformly sample initial points:" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.232, + 0.905, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.304, + 0.916, + 0.321 + ], + "angle": 0, + "content": "\\[\nX _ {0} = \\left\\{\\left(R _ {1}, R _ {2}, t, \\alpha , a\\right) \\mid R _ {v}, R _ {1}, t, \\alpha , a \\in \\text {v a l i d r a n g e s} \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.333, + 0.664, + 0.347 + ], + "angle": 0, + "content": "4: for all \\(x \\in X_0\\) do" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.348, + 0.905, + 0.378 + ], + "angle": 0, + "content": "5: Calculate \\( R_{3} \\) using the constraint: \\( \\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L \\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.379, + 0.888, + 0.393 + ], + "angle": 0, + "content": "6: Form pruning strategy: \\( s = (R_1, R_2, R_3, t, \\alpha, a) \\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.394, + 0.826, + 0.408 + ], + "angle": 0, + "content": "7: Evaluate the target function \\( f(s \\mid \\theta, \\mathcal{D}) \\)" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.348, + 0.905, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.409, + 0.597, + 0.421 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.423, + 0.693, + 0.437 + ], + "angle": 0, + "content": "9: for \\( n = 0 \\) to \\( T - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.438, + 0.877, + 0.453 + ], + "angle": 0, + "content": "10: Fit \\(\\mathcal{GP}\\) to the observed data \\((S_{n},f(S_{n}\\mid \\theta ,\\mathcal{D}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.454, + 0.904, + 0.468 + ], + "angle": 0, + "content": "11: Select the next point: \\(x_{n + 1}\\gets \\arg \\max_xA(x;\\mathcal{GP})\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.469, + 0.904, + 0.499 + ], + "angle": 0, + "content": "12: Calculate \\( R_{3} \\) using the constraint: \\( \\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L \\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.5, + 0.901, + 0.514 + ], + "angle": 0, + "content": "13: Form pruning strategy: \\( s_{n+1} = (R_1, R_2, R_3, t, \\alpha, a)_{n+1} \\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.515, + 0.768, + 0.529 + ], + "angle": 0, + "content": "14: Update \\(S_{n + 1}\\gets S_n\\cup \\{s_{n + 1}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.53, + 0.596, + 0.542 + ], + "angle": 0, + "content": "15: end for" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.544, + 0.846, + 0.559 + ], + "angle": 0, + "content": "16: return \\((R_{1}, R_{2}, R_{3}, t, \\alpha, a)^{*}\\) that maximize \\(f\\)" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.438, + 0.904, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.592, + 0.7, + 0.607 + ], + "angle": 0, + "content": "4.1. Experiment Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on ten image-based multimodal benchmarks: GQA [1], VizWiz [13], ScienceQA-IMG [26], TextVQA [30], ChartQA [27], POPE [19], MME [10], MMBench [25], MMStar [6], and DocVQA [28]. For video understanding, we adopt two datasets—SeedBench (video) [16] and VideoMME [11], where VideoMME is partitioned by video length into short, medium, and long subsets. Further details are presented in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Model Architectures. We integrate VFlowOpt into multiple LMMs, including LLaVA-OneVision-7B, LLaVA-NeXT-7B [23], and Qwen2-VL-7B [34]. These models employ various vision encoders (SigLIP, CLIP, and a ViT designed for Qwen2-VL) and LLM backbones (Qwen2-7B and Vicuna-7B). To prevent out-of-memory issues in Qwen2-VL, we set maxPixels = 3000000." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Comparison Methods. We compare VFlowOpt against five baseline methods: FastV [7], FitPrune [42], PDrop [36], SparseVLM [52], and VisionZip [40]. FastV," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23928" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.899, + 0.419 + ], + "angle": 0, + "content": "
MethodMMStarMMEMMBSQAPOPEGQAVizWiz\\(VQA^{Text}\\)ChartQADocVQAAvg.
Upper Bound, 100% Tokens
Vanilla61.7158180.895.889.162.260.476.080.087.5100%
Retain 50% Tokens (↓ 50%)
FastV (ECCV24)58.9154979.492.887.961.561.172.568.684.096.5%
FitPrune (AAAI25)4.596.61.08.618.816.26.837.637.730.721.4%
PDrop (CVPR25)37.1628.632.164.881.344.248.963.762.762.271.3%
SparseVLM (2024.10)59.8157780.594.188.161.960.473.970.580.897.1%
VisionZip (CVPR25)60.4158780.394.689.362.759.874.275.488.498.9%
VFlowOpt61.3159181.195.489.462.460.075.177.890.099.9%
Retain 25% Tokens (↓ 75%)
FastV (ECCV24)54.0153977.088.683.858.261.058.342.762.986.4%
FitPrune (AAAI25)5.292.61.15.920.214.55.129.727.323.217.6%
PDrop (CVPR25)36.8574.431.362.878.042.247.156.347.547.065.2%
SparseVLM (2024.10)56.6152078.790.387.259.760.866.354.066.690.5%
VisionZip (CVPR25)54.6156278.990.488.861.060.470.066.379.694.3%
VFlowOpt57.8157079.992.389.161.260.472.569.182.396.3%
Retain 10% Tokens (↓ 90%)
FastV (ECCV24)46.0120970.181.777.051.556.435.621.333.269.7%
FitPrune (AAAI25)5.427.00.93.52.34.60.76.95.96.95.1%
PDrop (CVPR25)31.8456.322.559.065.935.743.333.719.618.549.9%
SparseVLM (2024.10)45.1119171.883.780.054.656.439.837.639.674.0%
VisionZip (CVPR25)49.5138974.886.286.157.256.856.446.149.082.1%
VFlowOpt52.0146475.188.485.257.357.360.253.656.185.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.425, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Table 1. Performance comparison on LLaVA-OneVision-7B under different token retention conditions. \"Avg.\" refers to average accuracy on 10 benchmarks. For each reduction ratio, the best average performance is shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.672 + ], + "angle": 0, + "content": "FitPrune, PDrop, and SparseVLM rely on text-visual attention in the LLM to prune visual tokens but differ as follows: FastV performs a single pruning step after the second LLM layer; FitPrune prunes tokens in each layer according to predefined, layer-specific ratios; PDrop conducts four pruning steps within the LLM, using the same retention ratio each time; and SparseVLM uses the attention weights of preselected text tokens to evaluate the importance of visual tokens. VisionZip, in contrast, determines token importance based on the [cls] token's attention; for models lacking a [cls] token, we follow VisionZip's original procedure by computing the average attention each token receives from every other token in the ViTs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.899 + ], + "angle": 0, + "content": "Implementation Details. For LLaVA-OneVision-7B and Qwen2-VL-7B, we perform token pruning at three distinct points: before the LLM, and after the 9th and 18th layers. For LLaVA-NeXT-7B, pruning is conducted before the LLM, and again after the 10th and 20th layers. During optimization, we sample 30 unlabeled instances from each model's training datasets; for models without publicly available training dataset, we instead use random samples from the LLaVA-OneVision training set. The optimization is performed for a total of 50 iterations. For LMMs that modify visual tokens output by the vision encoder (e.g., unpadding and interpolation in LLaVA-OneVision, and unpadding in LLaVA-NeXT), we apply corresponding transformations to the importance maps so they remain fully aligned with the final visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.475, + 0.766, + 0.492 + ], + "angle": 0, + "content": "4.2. Image Understanding Tasks" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.503, + 0.907, + 0.836 + ], + "angle": 0, + "content": "We evaluate the proposed VFlowOpt on image understanding benchmarks with LLaVA-OneVision-7B using various pruning ratios, and present the results in Tab. 1. Compared to other baselines, VFlowOpt consistently maintains superior accuracy across different levels of token pruning. With \\(50\\%\\) token retention, VFlowOpt achieves \\(99.9\\%\\) of the original performance, exceeding the second-best approach by \\(1.0\\%\\). This negligible performance drop underscores the method's strong potential for practical deployment. Under more extreme pruning conditions with very few retained tokens, VFlowOpt's advantage becomes more pronounced. When only \\(10\\%\\) of the visual tokens are retained, VFlowOpt preserves \\(85.5\\%\\) of the original performance, surpassing the second-best approach by \\(3.3\\%\\). This finding suggests that VFlowOpt can effectively leverage limited visual information to maintain high performance under strict computational budgets. When applied to LLaVA-OneVision, both FitPrune and PDrop disrupt the model's normal output, indicating their inability to devise effective pruning strategies for different architectures. In contrast, VFlowOpt demonstrates a clear advantage by tailoring its pruning strategy to each model's unique characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To further validate VFlowOpt's generalization capability, we evaluate it on LLaVA-NeXT-7B and Qwen2-VL-7B. As shown in Tab. 2 and Tab. 3, VFlowOpt once again delivers the top results on these LMMs. When retaining \\(25\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23929" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.488, + 0.244 + ], + "angle": 0, + "content": "
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla37.6151967.470.186.564.2100%
Retain 25% Tokens (↓ 75%)
FastV35.1147765.667.483.160.495.7%
VisionZip35.8150165.467.986.761.597.3%
VFlowOpt37.0151467.067.787.662.698.8%
Retain 10% Tokens (↓ 90%)
FastV29.2128261.663.871.755.985.7%
VisionZip32.6137861.567.183.557.091.6%
VFlowOpt35.1139362.967.483.657.393.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.247, + 0.455, + 0.261 + ], + "angle": 0, + "content": "Table 2. Comparative experiments on LLaVA-NeXT-7B." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.272, + 0.483, + 0.427 + ], + "angle": 0, + "content": "
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla57.5168080.384.788.462.2100%
Retain 25% Tokens (↓ 75%)
FastV54.5159776.379.081.857.293.8
VisionZip55.2161878.981.386.960.096.8
VFlowOpt55.9165979.881.587.160.497.8
Retain 10% Tokens (↓ 90%)
FastV44.9140570.075.575.851.684.4
VisionZip49.3151876.578.284.354.190.9
VFlowOpt51.0159178.078.684.554.892.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.429, + 0.446, + 0.442 + ], + "angle": 0, + "content": "Table 3. Comparative experiments on Qwen2-VL-7B." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.453, + 0.481, + 0.626 + ], + "angle": 0, + "content": "
MethodSeedBench (video)VideoMMEAvg.
SML
Upper Bound, 100% Tokens
Vanilla56.970.554.649.5100%
Retain 25% Tokens (↓75%)
FastV54.766.953.247.796.2%
VisionZip56.468.355.349.099.0%
VFlowOpt56.868.955.849.2100%
Retain 10% Tokens (↓90%)
FastV48.753.747.642.383.6%
VisionZip55.059.751.846.392.4%
VFlowOpt55.563.352.748.695.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.628, + 0.478, + 0.642 + ], + "angle": 0, + "content": "Table 4. Comparative experiments on video understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.749 + ], + "angle": 0, + "content": "of the tokens, VFlowOpt preserves \\(98.8\\%\\) of the original performance on LLaVA-NeXT-7B and \\(97.8\\%\\) on Qwen2-VL-7B. Even under the stringent condition of retaining only \\(10\\%\\) of the tokens, VFlowOpt maintains \\(93.4\\%\\) of the performance on LLaVA-NeXT-7B and \\(92.8\\%\\) on Qwen2-VL-7B, demonstrating its strong generalizability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.357, + 0.772 + ], + "angle": 0, + "content": "4.3. Generalization to Video Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We further explore VFlowOpt's generalization across different modalities by evaluating it on video benchmarks using LLaVA-OneVision-7B. As shown in Tab. 4, VFlowOpt surpasses other baselines under various token retention ratios. Notably, when only \\(25\\%\\) of the tokens are preserved, VFlowOpt incurs virtually no performance loss on LLaVA-OneVision-7B, maintaining \\(100\\%\\) of its original performance. Even under the more stringent condition of retain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "ing just \\(10\\%\\) of the tokens, VFlowOpt still retains \\(95.5\\%\\) of the original performance, underscoring its robust generalization in the video domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.696, + 0.166 + ], + "angle": 0, + "content": "4.4. Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.905, + 0.262 + ], + "angle": 0, + "content": "We demonstrate the efficiency of VFlowOpt by conducting a comparative study on LLaVA-OneVision-7B running on a single NVIDIA A100-SXM4-80GB GPU, focusing on FLOPs, KV-Cache memory usage, and inference latency. The efficiency analysis of other baseline methods is provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.907, + 0.488 + ], + "angle": 0, + "content": "We assess various token pruning ratios on the MME benchmark, measuring overall performance and average efficiency metrics across all samples. As shown in Tab. 5, VFlowOpt significantly enhances LLaVA-OneVision-7B's computational efficiency by reducing FLOPs, shrinking KV-Cache memory usage, and accelerating inference speed. Notably, at a \\(50\\%\\) token pruning ratio, VFlowOpt achieves a \\(49.5\\%\\) reduction in KV-Cache memory and a \\(1.8\\times\\) speedup in inference, while boosting performance by \\(0.7\\%\\). At a more aggressive \\(75\\%\\) pruning ratio, we observe a \\(74.2\\%\\) reduction in KV-Cache memory and a \\(3.1\\times\\) speedup, with only a \\(0.7\\%\\) performance drop. Under extremely tight computational budgets, pruning \\(90\\%\\) of visual tokens reduces the KV-Cache memory footprint by \\(89\\%\\), accelerates inference by \\(3.8\\times\\), and degrades performance by only \\(7.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.641 + ], + "angle": 0, + "content": "Such reductions in inference latency substantially benefit user-facing applications demanding real-time performance, such as autonomous driving and robotics. Moreover, by significantly shrinking the KV-Cache, large-batch inference on LMMs can accommodate more user requests simultaneously, thereby substantially reducing overall inference costs. In summary, VFlowOpt uses far less GPU memory and delivers faster inference while preserving model performance, offering a highly practical solution for efficiently deploying LMMs in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.667, + 0.667 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.884 + ], + "angle": 0, + "content": "To confirm the contributions of each component in our token pruning strategy, we conduct ablation experiments on LLaVA-OneVision-7B under a computational budget of retaining \\(25\\%\\) of tokens. As shown in Tab. 6, removing importance calibration (i.e., directly using the mean attention over all tokens received by each visual token as its importance score), omitting token recycling, or discarding progressive pruning (i.e., maintaining the same number of tokens at each layer) leads to noticeable performance degradation across MMStar, MMBench, and SQA. By contrast, the complete VFlowOpt method consistently exhibits minimal performance loss, highlighting the effectiveness of all its components in pruning redundant tokens while preserving essential visual information." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We further examine how the number of samples and the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23930" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.089, + 0.855, + 0.181 + ], + "angle": 0, + "content": "
MethodsToken ReductionFLOPs ↓ (T)ΔLatency ↓ (ms)ΔKV Cache ↓ (MB)ΔPerformance ↑Δ
LLaVA-OneVision-7B-71.4-1040.1-1786.4-1581-
+ VFlowOpt50%37.2-48.0%584.2-43.8%902.8-49.5%1591+0.6%
75%19.1-73.2%328.5-68.4%460.6-74.2%1570-0.7%
90%7.7-89.2%272.1-73.8%197.1-89.0%1464-7.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.188, + 0.907, + 0.216 + ], + "angle": 0, + "content": "Table 5. Efficiency analysis of LLaVA-OneVision-7B with VFlowOpt. The detailed metric includes computation (FLOPs), latency, and KV-Cache memory. \\((\\Delta)\\) denotes the reduction ratio." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.233, + 0.457, + 0.314 + ], + "angle": 0, + "content": "
Pruning StrategyMMStarMMBenchSQA
VFlowOpt57.879.992.3
w/o Importance Calibration56.279.491.8
w/o Token Merging57.679.891.9
w/o Progressive Pruning56.079.191.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.32, + 0.459, + 0.336 + ], + "angle": 0, + "content": "Table 6. Ablation studies of pruning strategy components." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.344, + 0.476, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.491 + ], + "angle": 0, + "content": "Figure 4. Relationships of sample size (left) and optimization steps (right) with optimization time and final performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.629 + ], + "angle": 0, + "content": "number of optimization iterations in the Bayesian optimization procedure influence both optimization time and final performance. As illustrated in Fig. 4, the optimization time increases linearly with the number of samples and steps. Using 30 samples and 50 steps strikes an effective balance between time efficiency and performance, requiring only about 30 minutes to reach an exceptional result. More ablation results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.387, + 0.654 + ], + "angle": 0, + "content": "4.6. Visual Information Flow Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We further investigate how Bayesian optimization formulates the superior pruning strategy by examining its impact on visual information flow. Specifically, we compare the randomly initialized pruning strategy derived from the Bayesian optimization process with its final optimized strategy, measuring their respective effects on the text tokens that follow the visual tokens at each LLM layer. Because these text tokens can receive information from the visual tokens via the LLM's unidirectional attention mechanism, examining differences in these tokens with and without pruning effectively reveals how visual information flow is altered. As shown in Fig. 5, under the same computational budget, the initial strategy causes a pronounced discrepancy between text token representations with and without pruning, indicating that it significantly disrupts the visual information flow. In contrast, when using the optimized strategy," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.233, + 0.9, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.489, + 0.907, + 0.531 + ], + "angle": 0, + "content": "Figure 5. Visualization of the discrepancies in visual information flow corresponding to the pruning strategies before and after optimization, along with an example of visual question answering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.648 + ], + "angle": 0, + "content": "the text tokens remain much more similar to those observed without visual token pruning. This important finding suggests that the optimization process yields a pruning strategy that preserves crucial visual information by minimizing the difference in visual information flow with and without pruning, thereby retaining LMM performance to the greatest extent possible." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.662, + 0.634, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.899 + ], + "angle": 0, + "content": "In this paper, we address the pressing challenge of computational inefficiency in LMMs by introducing a general framework for visual token pruning. We reformulate pruning as an optimization problem focused on minimizing the divergence of visual information flow with and without pruning, thereby facilitating a tailored pruning strategy for diverse LMMs. By leveraging calibrated attention in ViTs to evaluate token significance and merging pruned tokens using importance-based weighting, we preserve critical information during inference without requiring retraining or manual tuning. Extensive experiments underscore the efficiency and generalizability of our approach, making it a practical solution for deploying LMMs in real-world settings, particularly in resource-constrained environments." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23931" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.482, + 0.149 + ], + "angle": 0, + "content": "This research was supported by Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.166, + 0.188, + 0.182 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[1] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.336, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.409, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.729, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[5] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Llavolta: Efficient multi-modal models via stage-wise visual context compression. arXiv preprint arXiv:2406.20092, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.855 + ], + "angle": 0, + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[7] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.191, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.205, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.904, + 0.355 + ], + "angle": 0, + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.426 + ], + "angle": 0, + "content": "[11] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.904, + 0.481 + ], + "angle": 0, + "content": "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[13] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[14] Lianyu Hu, Fanhua Shang, Liang Wan, and Wei Feng. illava: An image is worth fewer than 1/3 input tokens in large multimodal models. arXiv preprint arXiv:2412.06263, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[15] Omri Kaduri, Shai Bagon, and Tali Dekel. What's in the image? a deep-dive into the vision of vision language models. arXiv preprint arXiv:2411.17491, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[18] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[19] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[20] Yuelei Li, Ge Yan, Annabella Macaluso, Mazeyu Ji, Xueyan Zou, and Xiaolong Wang. Integrating lmm planners and 3d" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "skill policies for generalizable manipulation. arXiv preprint arXiv:2501.18733, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.483, + 0.236 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.483, + 0.306 + ], + "angle": 0, + "content": "[24] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.378 + ], + "angle": 0, + "content": "[25] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[26] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.519 + ], + "angle": 0, + "content": "[27] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.578, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[31] Haoxiang Tian, Xingshuo Han, Guoquan Wu, Yuan Zhou, Shuo Li, Jun Wei, Dan Ye, Wei Wang, and Tianwei Zhang. An llm-enhanced multi-objective evolutionary search for autonomous driving test scenario generation. arXiv preprint arXiv:2406.10857, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[32] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.092, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.287, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[33] Ao Wang, Fengyuan Sun, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. [cls] token tells everything needed for training-free efficient mllms. arXiv preprint arXiv:2412.05819, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[34] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[35] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. Drivemlm: Aligning multimodal large language models with behavioral planning states for autonomous driving, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[36] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[37] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[38] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[39] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23933" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.175 + ], + "angle": 0, + "content": "Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[41] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[42] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[43] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. arXiv preprint arXiv:2412.00447, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[45] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[46] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[47] Jun Zhang, Desen Meng, Ji Qi, Zhenpeng Huang, Tao Wu, and Limin Wang. p-mod: Building mixture-of-depths mllms via progressive ratio decay. arXiv preprint arXiv:2412.04449, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[48] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[49] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[50] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. In The Thirteenth International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[51] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Hao Tang, and Jieping Ye. From redundancy to relevance: Information flow in lvlms across reasoning tasks. arXiv preprint arXiv:2406.06579, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[52] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[53] Zhi Zhang, Srishti Yadav, Fengze Han, and Ekaterina Shutova. Cross-modal information flow in multimodal large language models. arXiv preprint arXiv:2411.18620, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23934" + } + ] +] \ No newline at end of file diff --git a/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_origin.pdf b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5d74b15d3ca41f0c6388876d82307cd3bf0b7689 --- /dev/null +++ b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/b08d8e36-4b30-4260-aad4-6cb0fd4d67ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1db02048554fc50466f7266ceb76a5da8a8a203bc05d5c03d8be2d14f3c16f6c +size 4892235 diff --git a/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/full.md b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ee6ea88dfe36ee08b183327097cf691f0cf5242e --- /dev/null +++ b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/full.md @@ -0,0 +1,333 @@ +# VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization + +Sihan Yang1 Runsen Xu1,2 Chenhang Cui3 Tai Wang1† Dahua Lin1,2 Jiangmiao Pang1† + +$^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong $^{3}$ National University of Singapore + +† Corresponding Author taiwang.me@gmail.com pangjiangmiao@gmail.com + +# Abstract + +Large Multimodal Models (LMMs) excel in visual-language tasks by leveraging numerous visual tokens for fine-grained visual information, but this token redundancy results in significant computational costs. Previous research aimed at reducing visual tokens during inference typically leverages importance maps derived from attention scores among vision-only tokens or vision-language tokens to prune tokens across one or multiple pruning stages. Despite this progress, pruning frameworks and strategies remain simplistic and insufficiently explored, often resulting in substantial performance degradation. In this paper, we propose VFlowOpt, a token pruning framework that introduces an importance map derivation process and a progressive pruning module with a recycling mechanism. The hyperparameters of its pruning strategy are further optimized by a visual information flow-guided method. Specifically, we compute an importance map for image tokens based on their attention-derived context relevance and patch-level information entropy. We then decide which tokens to retain or prune and aggregate the pruned ones as recycled tokens to avoid potential information loss. Finally, we apply a visual information flow-guided method that regards the last token in the LMM as the most representative signal of text-visual interactions. This method minimizes the discrepancy between token representations in LMMs with and without pruning, thereby enabling superior pruning strategies tailored to different LMMs. Experiments demonstrate that VFlowOpt can prune $90\%$ of visual tokens while maintaining comparable performance, leading to an $89\%$ reduction in KV-Cache memory and $3.8 \times$ faster inference. + +# 1. Introduction + +Large Multimodal Models (LMMs) [3, 8, 21, 22] have attained remarkable performance in tasks such as visual question answering [2, 12] and multimodal reasoning [44, 45], making them indispensable for applications like au + +![](images/4bd6a78824a2c043ef30859d3263d0397b752420a0d40c549f030e7784e45242.jpg) +Figure 1. VFlowOpt Performance and Efficiency. Our VFlowOpt significantly outperforms other LMM token reduction methods across three general benchmarks (MME, MMStar, and MMBench) on LLaVA-OneVision-7B. It achieves zero performance degradation when reducing $50\%$ of visual tokens. Furthermore, when retaining only $10\%$ of the tokens, it achieves $90\%$ of the original performance while reducing KV-Cache memory usage by $89\%$ and accelerating inference by $3.8\times$ . + +tonomous driving [9, 31, 35] and robotics [18, 20, 24]. To capture fine-grained visual details, modern LMMs treat images as token sequences, with models such as LLaVA-1.5 [21] processing hundreds of tokens and LLaVA-OneVision [17] handling up to several thousand tokens. Nonetheless, the increasing quantity of visual tokens significantly magnifies computational cost, memory usage, and inference time, creating a critical bottleneck for LMM deployment, particularly in resource-constrained environments or latency-sensitive scenarios. + +To alleviate these constraints, recent work explores various methods for reducing the number of visual tokens during inference. Most existing approaches evaluate the importance of visual tokens in LMMs using attention-based importance maps derived either from vision-only tokens [40, 49] or from text-visual interactions [36, 52]. Following a predefined pruning ratio, visual tokens are pruned in a single stage [33] or progressively across multiple stages [42], often with heuristic strategies applied uniformly across different LMMs. Although promising, these methods are sim + +plastic and underexplored, with coarse-grained pruning often causing significant performance drops due to lacking fitness for different LMMs models' characteristics. + +In response to these limitations, we propose a novel token pruning framework, VFlowOpt, that more accurately estimates token importance by integrating attention calibration and entropy, and reduces information loss through token recycling and progressive pruning strategy with fine-grained pruning ratios. Furthermore, by customizing pruning strategies for different models, our framework better preserves model performance. Specifically, the pruning framework consists of importance map computation and a progressive pruning module with a recycling mechanism. We decompose the importance map of image tokens into two aspects: the importance of tokens in the visual context, reflected by their attention maps, and the richness of visual information in each image patch, captured by its information entropy. Based on the weighted summation of these two factors, we determine which tokens to retain or prune across multiple pruning stages and introduce a recycling mechanism to aggregate pruned tokens to avoid potential information loss. Built with this pruning strategy with several hyperparameters, a key ingredient is the visual information flow-guided method to optimize the pruning strategy, which treats the last token in each pruning stage as the most representative signal of text-visual interactions during inference. By minimizing the difference between token representations in LMMs with and without pruning, VFlowOpt ultimately delivers superior pruning strategies specifically tailored to different LMMs. + +Comprehensive experiments on multiple vision-language benchmarks validate the efficacy of VFlowOpt. In particular, it substantially lowers computational cost while preserving competitive performance. As shown in Fig. 1, VFlowOpt can prune $50\%$ of visual tokens with negligible performance loss. Moreover, it can prune $90\%$ of visual tokens while maintaining $90\%$ of the model's performance, resulting in an $89\%$ reduction in KV-Cache memory usage and a $3.8 \times$ speedup in inference. These experimental results highlight its effectiveness as a practical and efficient solution for deploying LMMs in resource-constrained environments. + +# 2. Related Work + +# 2.1. Large Multimodal Models + +Large Multimodal Models (LMMs) [3, 8, 21, 22] extend the reasoning capabilities of Large Language Models (LLMs) [4, 32, 38, 39] to vision-language tasks by integrating a pre-trained vision encoder [29, 46] with a language model, linked by an alignment module such as an MLP, or a query-based network. This design transforms visual inputs into token sequences that the LLM can pro + +cess, facilitating multimodal prompts for tasks like visual question answering [2, 6, 12, 44]. To enhance performance, advanced LMMs, such as LLaVA-OneVision [17] and Qwen2-VL [34], can encode higher-resolution images into more image tokens, thereby capturing more granular visual details. However, as image resolution grows, the number of visual tokens rises exponentially, leading to substantially higher computational costs. For example, LLaVA-1.5 [21] processes $336 \times 336$ images into 576 tokens, whereas LLaVA-OneVision can handle $1152 \times 1152$ images, producing 7,290 tokens. This challenge becomes even more pronounced in video-based models like LongVA [48], which must process tokens across numerous frames. Fine-grained visual tokenization boosts LMM performance but poses an inference bottleneck, driving efforts to balance performance and cost through token reduction. + +# 2.2. Token Reduction for LMMs + +Token reduction has become a key strategy for improving the efficiency of LMMs by mitigating the computational cost associated with extensive visual token sequences. Existing methods can be broadly classified into training-based and training-free approaches. Training-based methods, such as LLaVA-Mini [50] and LLaVolta [5], introduce additional modules during model training to compress visual tokens and preserve critical information, while approaches like ATP-LLaVA [43] and p-MoD [47] train pruning modules to dynamically retain important tokens across LLM layers. However, these methods demand substantial computational resources to retrain the models, limiting their real-world applicability. In contrast, training-free methods prune tokens without additional training, often leveraging attention mechanisms to identify and discard redundant tokens. For instance, FastV[7] and SparseVLM exploit text-visual attention to rank token importance, whereas FasterVLM [49], VisionZip [40], and VTC-CLS [33] rely on [CLS] token attention in the vision encoder to evaluate token importance and prune redundant tokens. FitPrune [42] and PDrop [36] propose a progressive multistage pruning strategy in LMMs to fully utilize visual information. However, the pruning strategies in these methods often lack adaptability to different LMMs, which frequently leads to significant performance drops. + +# 3. Method + +In this section, we introduce VFlowOpt, a framework that employs a pruning strategy to reduce redundant visual tokens during LMM inference while preserving essential visual information. This strategy involves multiple hyperparameters that our VFlowOpt framework automatically optimizes for proper configuration—an essential factor for maintaining model performance. Specifically, Section 3.1 explains how our method evaluates the importance of visual + +![](images/11f3eec939c6bc6f8cac28c20264c1deb05bddf83b8a2d827d4fc8af5103a5c8.jpg) +(a) Token Importance Estimation + +![](images/bd996f6af1ded17ff73c3562764b3550a45b47a57cd004df5ca48fe7e66614ac.jpg) +(b) Token Recycling + +![](images/ac77e9c733e7b164f80cd00a8207b0782e0c95b091775067933bce49f861a589.jpg) +(1) Inference Stage + +![](images/3c405a4b6ddb48846a13a61c0ebe6a3c45543abb3506bdd88281f3b31d79216c.jpg) +(2) Optimization Stage + +![](images/08f38546afb6798412810532e22ad42dbf3845ad1e93985beeff65b10ec97b23.jpg) + +![](images/9c4ac0f406f77133d4fb72a2bffd19daf953d57a3f3ef903f97d9543a1ade086.jpg) +Figure 2. Overview of VFlowOpt. (1) During inference, VFlowOpt first assesses the importance of visual tokens, based on which progressive token pruning is performed. After the initial pruning stage, the pruned tokens are merged and recycled. The pruning strategy used in this process is defined by the (2) Optimization Stage. (a) The importance map is computed by combining the attention of relatively important tokens with the entropy of image patches. (b) The pruned tokens are grouped into grid cells, where each cell has a side length of a. Within each grid cell, the pruned tokens are fused using a weighted average, with their importance values as weights, and then recycled. (2) VFlowOpt optimizes the pruning strategy by minimizing the discrepancy of the last token in the final layer of the LMM with and without applying the pruning strategy. +Figure 3. The attention of redundant tokens (marked in red) fails to reflect the importance of other tokens and instead focuses on similar tokens, such as background elements. + +tokens, Section 3.2 presents progressive pruning and token recycling strategies, and finally, Section 3.3 details how VFlowOpt customizes the superior pruning strategy (i.e., selects the pruning hyperparameters) for different LMMs. + +# 3.1. Visual Token Importance Estimation + +For a general approach to evaluate visual token importance, the previous works [14, 41] propose that the importance of + +visual tokens can be estimated using the average attention from all other tokens in ViTs. Although it has achieved some success, we find that redundant tokens (e.g., tokens corresponding to background regions) often assign disproportionately high attention to other similar redundant tokens (shown in Fig. 3), reducing the reliability of the importance estimation. To address this, we first identify relatively important tokens based on the attention they receive from all tokens (tokens receiving higher attention are considered relatively important), as shown in Fig. 2 (a). We then exclude redundant tokens and use the attention from these relatively important tokens as a more robust metric. Additionally, we incorporate the information entropy of image patches into the importance score to prioritize tokens corresponding to visually informative regions. + +To identify relatively important visual tokens, we define a threshold: + +$$ +\tau = t \cdot \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {N} A _ {i j}, \tag {1} +$$ + +where $\tau$ is the threshold, $A_{ij}$ is the attention weight from + +token $i$ to token $j$ within a ViT layer, $N$ is the total number of tokens, and $t$ is a sensitivity hyperparameter. We then treat tokens with a total attention exceeding $\tau$ as relatively important: + +$$ +\mathcal {K} = \{j \mid \sum_ {i = 1} ^ {N} A _ {i j} > \tau \}. \tag {2} +$$ + +Here, $\mathcal{K}$ denotes the set of indices of these relatively important tokens. + +The information entropy of an image patch corresponding to token $i$ is defined as: + +$$ +H \left(V _ {i}\right) = - \sum_ {k = 0} ^ {L - 1} p _ {k} \log p _ {k}, \tag {3} +$$ + +where $V_{i}$ is the image patch corresponding to token $i$ , $p_{k}$ is the proportion of pixels in the patch with gray level $k$ , and $L$ is the total number of possible gray levels (256 for 8-bit images). Here, the gray level is defined as the average of the RGB channel values, providing a scalar representation of pixel brightness. Higher entropy reflects greater diversity in pixel intensities and richer visual information. + +The importance score of token $i$ is computed by combining the attention it receives from tokens in $\mathcal{K}$ with the entropy of its corresponding image patch, normalized via the softmax function: + +$$ +I _ {i} = \sum_ {k \in \mathcal {K}} A _ {k i} + \alpha \cdot \frac {\exp (H (V _ {i}))}{\sum_ {j = 1} ^ {N} \exp (H (V _ {j}))}, \tag {4} +$$ + +where $I_{i}$ denotes the importance score of token $i$ , $A_{ki}$ is the attention weight from token $k$ to token $i$ , $H(V_{i})$ is the entropy of the image patch corresponding to token $i$ , $\alpha$ is a hyperparameter that determines the entropy term's contribution, and $N$ is the total number of tokens. This approach mitigates biases from redundant tokens and leverages the visual information of image patches, thereby facilitating effective estimation of visual token importance. + +# 3.2. Progressive Pruning and Token Recycling + +Previous studies [37, 51] indicate that visual tokens play a more critical role in the shallower layers of LMMs, whereas redundancy tends to increase in deeper layers. Based on this, we adopt a progressive pruning strategy. To balance the pruning process's simplicity with its fine-grained configuration, we evenly divide the LMM into three stages. At the beginning of each stage, a predetermined fraction of visual tokens with higher importance scores (as defined in Sec. 3.1) is retained according to the stage-specific retention ratios $R = [R_{1}, R_{2}, R_{3}]$ , while the rest are pruned. Notably, the position IDs of visual tokens remain unchanged following pruning, preserving the original spatial structure of the visual input. Following the previous work [37], we compute the average visual token retention rate across the entire + +LMM as follows: + +$$ +\bar {R} = \frac {R _ {1} \cdot L _ {1} + R _ {1} \cdot R _ {2} \cdot L _ {2} + R _ {1} \cdot R _ {2} \cdot R _ {3} \cdot L _ {3}}{L}, \tag {5} +$$ + +where $L_{1}, L_{2}$ , and $L_{3}$ represent the number of layers in the three stages, respectively, and $L$ denotes the total number of layers in the LMM. + +To prevent the loss of any small but potentially significant information during the initial pruning process (before visual tokens are fed into the LLM), we propose a token merging and recycling strategy to compactly represent redundant visual information. Specifically, We use a square grid (with each cell having side length $a$ ) to group pruned tokens that fall into the same grid cell, as shown in Fig. 2 (b). Within each cell, the pruned tokens are fused into a single token by computing a weighted average of their representations, using their importance scores as weights. This fused token then replaces the pruned token with the highest importance score in that cell and is incorporated into the set of retained tokens. + +Formally, a token $t_i$ with spatial coordinates $(x_i, y_i)$ belongs to a grid cell $\mathcal{G}_{p,q}$ if + +$$ +\lfloor x _ {i} / a \rfloor = p \quad \text {a n d} \quad \lfloor y _ {i} / a \rfloor = q, \tag {6} +$$ + +where $p$ and $q$ are the row and column indices of the grid cell, respectively. Suppose there are $k$ pruned tokens in $\mathcal{G}_{p,q}$ with corresponding importance scores $I_{1}, I_{2}, \ldots, I_{k}$ and token representations $\mathbf{t}_1, \mathbf{t}_2, \ldots, \mathbf{t}_k$ . The fused token $\mathbf{t}_{\mathrm{merged}}^{p,q}$ is computed as: + +$$ +\mathbf {t} _ {\text {m e r g e d}} ^ {p, q} = \frac {\sum_ {i = 1} ^ {k} I _ {i} \cdot \mathbf {t} _ {i}}{\sum_ {i = 1} ^ {k} I _ {i}}. \tag {7} +$$ + +We then assign $\mathbf{t}_{\mathrm{merged}}^{p,q}$ to the position of the pruned token with the highest importance score $I_{\mathrm{max}}$ in $\mathcal{G}_{p,q}$ , counting it among the retained tokens. By combining token pruning with this recycling strategy, we reduce token redundancy while avoiding potential information loss. + +# 3.3. Pruning Strategy Optimization + +The pruning strategies described in Sec. 3.1 and Sec. 3.2 involve several key hyperparameters that directly affect the performance of LMMs after token pruning. Properly defining these hyperparameters is crucial for preserving model performance. However, existing approaches often rely on manually designed pruning strategies and apply the same strategy across different LMMs, without considering the unique characteristics of each model. This coarse-grained approach can lead to significant performance degradation due to its lack of fitness for different LMMs, leaving the task of designing pruning strategies tailored to specific LMMs as a formidable challenge. + +Previous interpretability studies of LMMs [15, 53] offer crucial insights into their internal mechanisms. Specifically, these works reveal that in the lower and middle layers of LMMs, visual information from visual tokens is aggregated into the corresponding query text tokens. In the higher layers, the multimodal representation encoded in the query text tokens is further progressively propagated to the final position of the input sequence, ultimately influencing the subsequent inference process. + +Inspired by this insight, we propose a framework that requires only a small amount of unlabeled data and leverages the internal flow of visual information within LMMs to search for the superior pruning strategy tailored to the characteristics of different LMMs, as shown in Fig. 2 (2). Specifically, we recast the task of designing the superior pruning strategy as an optimization problem, aiming to minimize the discrepancy in visual information flow with and without applying a pruning strategy to perform visual token pruning. In this framework, we treat the final token in the last layer as the representative outcome of the visual information flow. We use cosine similarity to measure the similarity of the final token representation with and without visual token pruning. A higher similarity indicates that the discrepancy between the visual information flow with and without pruning is smaller. Accordingly, we define the following optimization objective: + +$$ +\max _ {s \in \mathcal {S}} f (s) = \operatorname {C o s i n e S i m} \left(h _ {f}, g _ {s} \left(h _ {f}\right)\right). \tag {8} +$$ + +Here, $h_f$ represents the representation of the final token in the last layer before pruning, and $g_s(h_f)$ represents the final token representation after applying the pruning strategy $s$ (with $g_s$ modeling its effect on the token feature), and $S$ is the solution space. + +To efficiently search for the superior pruning strategy $s$ , we employ Bayesian optimization, which systematically explores the hyperparameter space—including the threshold sensitivity $t$ , the entropy weight $\alpha$ , the grid size $a$ , and the pruning ratios $R_{1}, R_{2}$ , and $R_{3}$ —to maximize the target function $f(s)$ . In short, Bayesian optimization constructs a surrogate model to approximate the objective function and employs an acquisition function to balance exploration and exploitation, thereby efficiently guiding the search for promising hyperparameter settings. The details of this optimization process are presented in Algorithm 1. + +# 4. Experiments + +In this section, we evaluate our approach on various LMMs across diverse image and video benchmarks, followed by an efficiency analysis and an ablation study of each component. Finally, we illustrate how our method affects discrepancies in the visual information flow with and without pruning, offering deeper insights into our approach. + +Algorithm 1 VFlowOpt with Bayesian Optimization + +Input: LMM $\theta$ , data samples $\mathcal{D}$ , number of Bayesian optimization iterations $T$ , computation budget $\overline{R}$ , target function $f(\cdot)$ : the sum of cosine similarities between the last token in the last layer of the LMM $\theta$ with and without visual token pruning, computed over all data samples $\mathcal{D}$ . +Output: retention rates $R_{1}, R_{2}, R_{3}$ ; threshold sensitivity $t$ ; entropy weight in importance score $\alpha$ ; grid size $a$ . + +1: Initialize a Gaussian Process model $\mathcal{GP}$ +2: Define the acquisition function $A(\cdot)$ (Expected Improvement is adopted) +3: Uniformly sample initial points: + +$$ +X _ {0} = \left\{\left(R _ {1}, R _ {2}, t, \alpha , a\right) \mid R _ {v}, R _ {1}, t, \alpha , a \in \text {v a l i d r a n g e s} \right\} +$$ + +4: for all $x \in X_0$ do + +5: Calculate $R_{3}$ using the constraint: $\overline{R} = (R_{1} \cdot L_{1} + R_{1} \cdot R_{2} \cdot L_{2} + R_{1} \cdot R_{2} \cdot R_{3} \cdot L_{3}) / L$ +6: Form pruning strategy: $s = (R_1, R_2, R_3, t, \alpha, a)$ +7: Evaluate the target function $f(s \mid \theta, \mathcal{D})$ + +8: end for + +9: for $n = 0$ to $T - 1$ do + +10: Fit $\mathcal{GP}$ to the observed data $(S_{n},f(S_{n}\mid \theta ,\mathcal{D}))$ +11: Select the next point: $x_{n + 1}\gets \arg \max_xA(x;\mathcal{GP})$ +12: Calculate $R_{3}$ using the constraint: $\overline{R} = (R_{1} \cdot L_{1} + R_{1} \cdot R_{2} \cdot L_{2} + R_{1} \cdot R_{2} \cdot R_{3} \cdot L_{3}) / L$ +13: Form pruning strategy: $s_{n+1} = (R_1, R_2, R_3, t, \alpha, a)_{n+1}$ +14: Update $S_{n + 1}\gets S_n\cup \{s_{n + 1}\}$ +15: end for +16: return $(R_{1}, R_{2}, R_{3}, t, \alpha, a)^{*}$ that maximize $f$ + +# 4.1. Experiment Setting + +Datasets. We evaluate our method on ten image-based multimodal benchmarks: GQA [1], VizWiz [13], ScienceQA-IMG [26], TextVQA [30], ChartQA [27], POPE [19], MME [10], MMBench [25], MMStar [6], and DocVQA [28]. For video understanding, we adopt two datasets—SeedBench (video) [16] and VideoMME [11], where VideoMME is partitioned by video length into short, medium, and long subsets. Further details are presented in the Appendix. + +Model Architectures. We integrate VFlowOpt into multiple LMMs, including LLaVA-OneVision-7B, LLaVA-NeXT-7B [23], and Qwen2-VL-7B [34]. These models employ various vision encoders (SigLIP, CLIP, and a ViT designed for Qwen2-VL) and LLM backbones (Qwen2-7B and Vicuna-7B). To prevent out-of-memory issues in Qwen2-VL, we set maxPixels = 3000000. + +Comparison Methods. We compare VFlowOpt against five baseline methods: FastV [7], FitPrune [42], PDrop [36], SparseVLM [52], and VisionZip [40]. FastV, + +
MethodMMStarMMEMMBSQAPOPEGQAVizWiz\(VQA^{Text}\)ChartQADocVQAAvg.
Upper Bound, 100% Tokens
Vanilla61.7158180.895.889.162.260.476.080.087.5100%
Retain 50% Tokens (↓ 50%)
FastV (ECCV24)58.9154979.492.887.961.561.172.568.684.096.5%
FitPrune (AAAI25)4.596.61.08.618.816.26.837.637.730.721.4%
PDrop (CVPR25)37.1628.632.164.881.344.248.963.762.762.271.3%
SparseVLM (2024.10)59.8157780.594.188.161.960.473.970.580.897.1%
VisionZip (CVPR25)60.4158780.394.689.362.759.874.275.488.498.9%
VFlowOpt61.3159181.195.489.462.460.075.177.890.099.9%
Retain 25% Tokens (↓ 75%)
FastV (ECCV24)54.0153977.088.683.858.261.058.342.762.986.4%
FitPrune (AAAI25)5.292.61.15.920.214.55.129.727.323.217.6%
PDrop (CVPR25)36.8574.431.362.878.042.247.156.347.547.065.2%
SparseVLM (2024.10)56.6152078.790.387.259.760.866.354.066.690.5%
VisionZip (CVPR25)54.6156278.990.488.861.060.470.066.379.694.3%
VFlowOpt57.8157079.992.389.161.260.472.569.182.396.3%
Retain 10% Tokens (↓ 90%)
FastV (ECCV24)46.0120970.181.777.051.556.435.621.333.269.7%
FitPrune (AAAI25)5.427.00.93.52.34.60.76.95.96.95.1%
PDrop (CVPR25)31.8456.322.559.065.935.743.333.719.618.549.9%
SparseVLM (2024.10)45.1119171.883.780.054.656.439.837.639.674.0%
VisionZip (CVPR25)49.5138974.886.286.157.256.856.446.149.082.1%
VFlowOpt52.0146475.188.485.257.357.360.253.656.185.5%
+ +Table 1. Performance comparison on LLaVA-OneVision-7B under different token retention conditions. "Avg." refers to average accuracy on 10 benchmarks. For each reduction ratio, the best average performance is shown in bold. + +FitPrune, PDrop, and SparseVLM rely on text-visual attention in the LLM to prune visual tokens but differ as follows: FastV performs a single pruning step after the second LLM layer; FitPrune prunes tokens in each layer according to predefined, layer-specific ratios; PDrop conducts four pruning steps within the LLM, using the same retention ratio each time; and SparseVLM uses the attention weights of preselected text tokens to evaluate the importance of visual tokens. VisionZip, in contrast, determines token importance based on the [cls] token's attention; for models lacking a [cls] token, we follow VisionZip's original procedure by computing the average attention each token receives from every other token in the ViTs. + +Implementation Details. For LLaVA-OneVision-7B and Qwen2-VL-7B, we perform token pruning at three distinct points: before the LLM, and after the 9th and 18th layers. For LLaVA-NeXT-7B, pruning is conducted before the LLM, and again after the 10th and 20th layers. During optimization, we sample 30 unlabeled instances from each model's training datasets; for models without publicly available training dataset, we instead use random samples from the LLaVA-OneVision training set. The optimization is performed for a total of 50 iterations. For LMMs that modify visual tokens output by the vision encoder (e.g., unpadding and interpolation in LLaVA-OneVision, and unpadding in LLaVA-NeXT), we apply corresponding transformations to the importance maps so they remain fully aligned with the final visual tokens. + +# 4.2. Image Understanding Tasks + +We evaluate the proposed VFlowOpt on image understanding benchmarks with LLaVA-OneVision-7B using various pruning ratios, and present the results in Tab. 1. Compared to other baselines, VFlowOpt consistently maintains superior accuracy across different levels of token pruning. With $50\%$ token retention, VFlowOpt achieves $99.9\%$ of the original performance, exceeding the second-best approach by $1.0\%$ . This negligible performance drop underscores the method's strong potential for practical deployment. Under more extreme pruning conditions with very few retained tokens, VFlowOpt's advantage becomes more pronounced. When only $10\%$ of the visual tokens are retained, VFlowOpt preserves $85.5\%$ of the original performance, surpassing the second-best approach by $3.3\%$ . This finding suggests that VFlowOpt can effectively leverage limited visual information to maintain high performance under strict computational budgets. When applied to LLaVA-OneVision, both FitPrune and PDrop disrupt the model's normal output, indicating their inability to devise effective pruning strategies for different architectures. In contrast, VFlowOpt demonstrates a clear advantage by tailoring its pruning strategy to each model's unique characteristics. + +To further validate VFlowOpt's generalization capability, we evaluate it on LLaVA-NeXT-7B and Qwen2-VL-7B. As shown in Tab. 2 and Tab. 3, VFlowOpt once again delivers the top results on these LMMs. When retaining $25\%$ + +
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla37.6151967.470.186.564.2100%
Retain 25% Tokens (↓ 75%)
FastV35.1147765.667.483.160.495.7%
VisionZip35.8150165.467.986.761.597.3%
VFlowOpt37.0151467.067.787.662.698.8%
Retain 10% Tokens (↓ 90%)
FastV29.2128261.663.871.755.985.7%
VisionZip32.6137861.567.183.557.091.6%
VFlowOpt35.1139362.967.483.657.393.4%
+ +Table 2. Comparative experiments on LLaVA-NeXT-7B. + +
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla57.5168080.384.788.462.2100%
Retain 25% Tokens (↓ 75%)
FastV54.5159776.379.081.857.293.8
VisionZip55.2161878.981.386.960.096.8
VFlowOpt55.9165979.881.587.160.497.8
Retain 10% Tokens (↓ 90%)
FastV44.9140570.075.575.851.684.4
VisionZip49.3151876.578.284.354.190.9
VFlowOpt51.0159178.078.684.554.892.8
+ +Table 3. Comparative experiments on Qwen2-VL-7B. + +
MethodSeedBench (video)VideoMMEAvg.
SML
Upper Bound, 100% Tokens
Vanilla56.970.554.649.5100%
Retain 25% Tokens (↓75%)
FastV54.766.953.247.796.2%
VisionZip56.468.355.349.099.0%
VFlowOpt56.868.955.849.2100%
Retain 10% Tokens (↓90%)
FastV48.753.747.642.383.6%
VisionZip55.059.751.846.392.4%
VFlowOpt55.563.352.748.695.5%
+ +Table 4. Comparative experiments on video understanding tasks. + +of the tokens, VFlowOpt preserves $98.8\%$ of the original performance on LLaVA-NeXT-7B and $97.8\%$ on Qwen2-VL-7B. Even under the stringent condition of retaining only $10\%$ of the tokens, VFlowOpt maintains $93.4\%$ of the performance on LLaVA-NeXT-7B and $92.8\%$ on Qwen2-VL-7B, demonstrating its strong generalizability. + +# 4.3. Generalization to Video Tasks + +We further explore VFlowOpt's generalization across different modalities by evaluating it on video benchmarks using LLaVA-OneVision-7B. As shown in Tab. 4, VFlowOpt surpasses other baselines under various token retention ratios. Notably, when only $25\%$ of the tokens are preserved, VFlowOpt incurs virtually no performance loss on LLaVA-OneVision-7B, maintaining $100\%$ of its original performance. Even under the more stringent condition of retain + +ing just $10\%$ of the tokens, VFlowOpt still retains $95.5\%$ of the original performance, underscoring its robust generalization in the video domain. + +# 4.4. Efficiency Analysis + +We demonstrate the efficiency of VFlowOpt by conducting a comparative study on LLaVA-OneVision-7B running on a single NVIDIA A100-SXM4-80GB GPU, focusing on FLOPs, KV-Cache memory usage, and inference latency. The efficiency analysis of other baseline methods is provided in the Appendix. + +We assess various token pruning ratios on the MME benchmark, measuring overall performance and average efficiency metrics across all samples. As shown in Tab. 5, VFlowOpt significantly enhances LLaVA-OneVision-7B's computational efficiency by reducing FLOPs, shrinking KV-Cache memory usage, and accelerating inference speed. Notably, at a $50\%$ token pruning ratio, VFlowOpt achieves a $49.5\%$ reduction in KV-Cache memory and a $1.8\times$ speedup in inference, while boosting performance by $0.7\%$ . At a more aggressive $75\%$ pruning ratio, we observe a $74.2\%$ reduction in KV-Cache memory and a $3.1\times$ speedup, with only a $0.7\%$ performance drop. Under extremely tight computational budgets, pruning $90\%$ of visual tokens reduces the KV-Cache memory footprint by $89\%$ , accelerates inference by $3.8\times$ , and degrades performance by only $7.4\%$ . + +Such reductions in inference latency substantially benefit user-facing applications demanding real-time performance, such as autonomous driving and robotics. Moreover, by significantly shrinking the KV-Cache, large-batch inference on LMMs can accommodate more user requests simultaneously, thereby substantially reducing overall inference costs. In summary, VFlowOpt uses far less GPU memory and delivers faster inference while preserving model performance, offering a highly practical solution for efficiently deploying LMMs in real-world scenarios. + +# 4.5. Ablation Study + +To confirm the contributions of each component in our token pruning strategy, we conduct ablation experiments on LLaVA-OneVision-7B under a computational budget of retaining $25\%$ of tokens. As shown in Tab. 6, removing importance calibration (i.e., directly using the mean attention over all tokens received by each visual token as its importance score), omitting token recycling, or discarding progressive pruning (i.e., maintaining the same number of tokens at each layer) leads to noticeable performance degradation across MMStar, MMBench, and SQA. By contrast, the complete VFlowOpt method consistently exhibits minimal performance loss, highlighting the effectiveness of all its components in pruning redundant tokens while preserving essential visual information. + +We further examine how the number of samples and the + +
MethodsToken ReductionFLOPs ↓ (T)ΔLatency ↓ (ms)ΔKV Cache ↓ (MB)ΔPerformance ↑Δ
LLaVA-OneVision-7B-71.4-1040.1-1786.4-1581-
+ VFlowOpt50%37.2-48.0%584.2-43.8%902.8-49.5%1591+0.6%
75%19.1-73.2%328.5-68.4%460.6-74.2%1570-0.7%
90%7.7-89.2%272.1-73.8%197.1-89.0%1464-7.4%
+ +Table 5. Efficiency analysis of LLaVA-OneVision-7B with VFlowOpt. The detailed metric includes computation (FLOPs), latency, and KV-Cache memory. $(\Delta)$ denotes the reduction ratio. + +
Pruning StrategyMMStarMMBenchSQA
VFlowOpt57.879.992.3
w/o Importance Calibration56.279.491.8
w/o Token Merging57.679.891.9
w/o Progressive Pruning56.079.191.0
+ +Table 6. Ablation studies of pruning strategy components. + +![](images/2747e4b8f4cda23aeadfcdea5c5eba1512cdc543528df07152841897b6e5eb9c.jpg) +Figure 4. Relationships of sample size (left) and optimization steps (right) with optimization time and final performance. + +number of optimization iterations in the Bayesian optimization procedure influence both optimization time and final performance. As illustrated in Fig. 4, the optimization time increases linearly with the number of samples and steps. Using 30 samples and 50 steps strikes an effective balance between time efficiency and performance, requiring only about 30 minutes to reach an exceptional result. More ablation results are provided in the supplementary material. + +# 4.6. Visual Information Flow Analysis + +We further investigate how Bayesian optimization formulates the superior pruning strategy by examining its impact on visual information flow. Specifically, we compare the randomly initialized pruning strategy derived from the Bayesian optimization process with its final optimized strategy, measuring their respective effects on the text tokens that follow the visual tokens at each LLM layer. Because these text tokens can receive information from the visual tokens via the LLM's unidirectional attention mechanism, examining differences in these tokens with and without pruning effectively reveals how visual information flow is altered. As shown in Fig. 5, under the same computational budget, the initial strategy causes a pronounced discrepancy between text token representations with and without pruning, indicating that it significantly disrupts the visual information flow. In contrast, when using the optimized strategy, + +![](images/d2a427acb1d1889fcb27ddbe93caecc35d46160cebc60341d0b621e7434e3262.jpg) +Figure 5. Visualization of the discrepancies in visual information flow corresponding to the pruning strategies before and after optimization, along with an example of visual question answering. + +the text tokens remain much more similar to those observed without visual token pruning. This important finding suggests that the optimization process yields a pruning strategy that preserves crucial visual information by minimizing the difference in visual information flow with and without pruning, thereby retaining LMM performance to the greatest extent possible. + +# 5. Conclusion + +In this paper, we address the pressing challenge of computational inefficiency in LMMs by introducing a general framework for visual token pruning. We reformulate pruning as an optimization problem focused on minimizing the divergence of visual information flow with and without pruning, thereby facilitating a tailored pruning strategy for diverse LMMs. By leveraging calibrated attention in ViTs to evaluate token significance and merging pruned tokens using importance-based weighting, we preserve critical information during inference without requiring retraining or manual tuning. Extensive experiments underscore the efficiency and generalizability of our approach, making it a practical solution for deploying LMMs in real-world settings, particularly in resource-constrained environments. + +# Acknowledgements + +This research was supported by Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 5 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 1, 2 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2 +[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 2 +[5] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Llavolta: Efficient multi-modal models via stage-wise visual context compression. arXiv preprint arXiv:2406.20092, 2024. 2 +[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 5 +[7] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for + +large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5 +[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 2 +[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving, 2023. 1 +[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5 +[11] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5 +[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering, 2017. 1, 2 +[13] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5 +[14] Lianyu Hu, Fanhua Shang, Liang Wan, and Wei Feng. illava: An image is worth fewer than 1/3 input tokens in large multimodal models. arXiv preprint arXiv:2412.06263, 2024. 3 +[15] Omri Kaduri, Shai Bagon, and Tali Dekel. What's in the image? a deep-dive into the vision of vision language models. arXiv preprint arXiv:2411.17491, 2024. 5 +[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5 +[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 2 +[18] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation, 2023. 1 +[19] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5 +[20] Yuelei Li, Ge Yan, Annabella Macaluso, Mazeyu Ji, Xueyan Zou, and Xiaolong Wang. Integrating lmm planners and 3d + +skill policies for generalizable manipulation. arXiv preprint arXiv:2501.18733, 2025. 1 +[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2 +[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2 +[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5 +[24] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation, 2024. 1 +[25] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5 +[26] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5 +[27] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5 +[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2 +[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5 +[31] Haoxiang Tian, Xingshuo Han, Guoquan Wu, Yuan Zhou, Shuo Li, Jun Wei, Dan Ye, Wei Wang, and Tianwei Zhang. An llm-enhanced multi-objective evolutionary search for autonomous driving test scenario generation. arXiv preprint arXiv:2406.10857, 2024. 1 +[32] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, + +Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 2 +[33] Ao Wang, Fengyuan Sun, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. [cls] token tells everything needed for training-free efficient mllms. arXiv preprint arXiv:2412.05819, 2024. 1, 2 +[34] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5 +[35] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. Drivemlm: Aligning multimodal large language models with behavioral planning states for autonomous driving, 2023. 1 +[36] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 1, 2, 5 +[37] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 4 +[38] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2 +[39] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, + +Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2 +[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 1, 2, 5 +[41] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 3 +[42] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 5 +[43] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. arXiv preprint arXiv:2412.00447, 2024. 2 +[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 1, 2 +[45] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark, 2024. 1 +[46] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 2 +[47] Jun Zhang, Desen Meng, Ji Qi, Zhenpeng Huang, Tao Wu, and Limin Wang. p-mod: Building mixture-of-depths mllms via progressive ratio decay. arXiv preprint arXiv:2412.04449, 2024. 2 +[48] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2 +[49] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster, 2024. 1, 2 +[50] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. In The Thirteenth International Conference on Learning Representations. 2 +[51] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, + +Hao Tang, and Jieping Ye. From redundancy to relevance: Information flow in lvlms across reasoning tasks. arXiv preprint arXiv:2406.06579, 2024. 4 +[52] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 1, 5 +[53] Zhi Zhang, Srishti Yadav, Fengze Han, and Ekaterina Shutova. Cross-modal information flow in multimodal large language models. arXiv preprint arXiv:2411.18620, 2024. 5 \ No newline at end of file diff --git a/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/images.zip b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e7cf8d49a1a41abcf3f75725e479088ec90d6a60 --- /dev/null +++ b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4fb6be8870c5452c86ce6bf00456960ec966e6551d96f2a4d445ea182d9284b +size 649996 diff --git a/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/layout.json b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..417755c9e63bbc9d1f2130b303d02c22d598300e --- /dev/null +++ b/2025/VFlowOpt_ A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization/layout.json @@ -0,0 +1,9766 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "text", + "content": "VFlowOpt: A Token Pruning Framework for LMMs with Visual Information Flow-Guided Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 159, + 536, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 159, + 536, + 176 + ], + "spans": [ + { + "bbox": [ + 74, + 159, + 536, + 176 + ], + "type": "text", + "content": "Sihan Yang1 Runsen Xu1,2 Chenhang Cui3 Tai Wang1† Dahua Lin1,2 Jiangmiao Pang1†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "type": "text", + "content": " Shanghai AI Laboratory " + }, + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "type": "text", + "content": " The Chinese University of Hong Kong " + }, + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 177, + 498, + 190 + ], + "type": "text", + "content": " National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 190, + 473, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 473, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 473, + 203 + ], + "type": "text", + "content": "† Corresponding Author taiwang.me@gmail.com pangjiangmiao@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) excel in visual-language tasks by leveraging numerous visual tokens for fine-grained visual information, but this token redundancy results in significant computational costs. Previous research aimed at reducing visual tokens during inference typically leverages importance maps derived from attention scores among vision-only tokens or vision-language tokens to prune tokens across one or multiple pruning stages. Despite this progress, pruning frameworks and strategies remain simplistic and insufficiently explored, often resulting in substantial performance degradation. In this paper, we propose VFlowOpt, a token pruning framework that introduces an importance map derivation process and a progressive pruning module with a recycling mechanism. The hyperparameters of its pruning strategy are further optimized by a visual information flow-guided method. Specifically, we compute an importance map for image tokens based on their attention-derived context relevance and patch-level information entropy. We then decide which tokens to retain or prune and aggregate the pruned ones as recycled tokens to avoid potential information loss. Finally, we apply a visual information flow-guided method that regards the last token in the LMM as the most representative signal of text-visual interactions. This method minimizes the discrepancy between token representations in LMMs with and without pruning, thereby enabling superior pruning strategies tailored to different LMMs. Experiments demonstrate that VFlowOpt can prune " + }, + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "text", + "content": " of visual tokens while maintaining comparable performance, leading to an " + }, + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "text", + "content": " reduction in KV-Cache memory and " + }, + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "inline_equation", + "content": "3.8 \\times" + }, + { + "bbox": [ + 54, + 257, + 297, + 617 + ], + "type": "text", + "content": " faster inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 643, + 135, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 643, + 135, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 643, + 135, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) [3, 8, 21, 22] have attained remarkable performance in tasks such as visual question answering [2, 12] and multimodal reasoning [44, 45], making them indispensable for applications like au" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 222, + 553, + 366 + ], + "blocks": [ + { + "bbox": [ + 317, + 222, + 553, + 366 + ], + "lines": [ + { + "bbox": [ + 317, + 222, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 317, + 222, + 553, + 366 + ], + "type": "image", + "image_path": "4bd6a78824a2c043ef30859d3263d0397b752420a0d40c549f030e7784e45242.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "text", + "content": "Figure 1. VFlowOpt Performance and Efficiency. Our VFlowOpt significantly outperforms other LMM token reduction methods across three general benchmarks (MME, MMStar, and MMBench) on LLaVA-OneVision-7B. It achieves zero performance degradation when reducing " + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "text", + "content": " of visual tokens. Furthermore, when retaining only " + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "text", + "content": " of the tokens, it achieves " + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "text", + "content": " of the original performance while reducing KV-Cache memory usage by " + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "text", + "content": " and accelerating inference by " + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "inline_equation", + "content": "3.8\\times" + }, + { + "bbox": [ + 313, + 369, + 556, + 459 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 471, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 471, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 312, + 471, + 555, + 590 + ], + "type": "text", + "content": "tonomous driving [9, 31, 35] and robotics [18, 20, 24]. To capture fine-grained visual details, modern LMMs treat images as token sequences, with models such as LLaVA-1.5 [21] processing hundreds of tokens and LLaVA-OneVision [17] handling up to several thousand tokens. Nonetheless, the increasing quantity of visual tokens significantly magnifies computational cost, memory usage, and inference time, creating a critical bottleneck for LMM deployment, particularly in resource-constrained environments or latency-sensitive scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "content": "To alleviate these constraints, recent work explores various methods for reducing the number of visual tokens during inference. Most existing approaches evaluate the importance of visual tokens in LMMs using attention-based importance maps derived either from vision-only tokens [40, 49] or from text-visual interactions [36, 52]. Following a predefined pruning ratio, visual tokens are pruned in a single stage [33] or progressively across multiple stages [42], often with heuristic strategies applied uniformly across different LMMs. Although promising, these methods are sim" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23924" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "plastic and underexplored, with coarse-grained pruning often causing significant performance drops due to lacking fitness for different LMMs models' characteristics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 298, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 298, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 298, + 431 + ], + "type": "text", + "content": "In response to these limitations, we propose a novel token pruning framework, VFlowOpt, that more accurately estimates token importance by integrating attention calibration and entropy, and reduces information loss through token recycling and progressive pruning strategy with fine-grained pruning ratios. Furthermore, by customizing pruning strategies for different models, our framework better preserves model performance. Specifically, the pruning framework consists of importance map computation and a progressive pruning module with a recycling mechanism. We decompose the importance map of image tokens into two aspects: the importance of tokens in the visual context, reflected by their attention maps, and the richness of visual information in each image patch, captured by its information entropy. Based on the weighted summation of these two factors, we determine which tokens to retain or prune across multiple pruning stages and introduce a recycling mechanism to aggregate pruned tokens to avoid potential information loss. Built with this pruning strategy with several hyperparameters, a key ingredient is the visual information flow-guided method to optimize the pruning strategy, which treats the last token in each pruning stage as the most representative signal of text-visual interactions during inference. By minimizing the difference between token representations in LMMs with and without pruning, VFlowOpt ultimately delivers superior pruning strategies specifically tailored to different LMMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "content": "Comprehensive experiments on multiple vision-language benchmarks validate the efficacy of VFlowOpt. In particular, it substantially lowers computational cost while preserving competitive performance. As shown in Fig. 1, VFlowOpt can prune " + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "content": " of visual tokens with negligible performance loss. Moreover, it can prune " + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "content": " of visual tokens while maintaining " + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "content": " of the model's performance, resulting in an " + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "content": " reduction in KV-Cache memory usage and a " + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "inline_equation", + "content": "3.8 \\times" + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "content": " speedup in inference. These experimental results highlight its effectiveness as a practical and efficient solution for deploying LMMs in resource-constrained environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 590, + 142, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 142, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 142, + 602 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 611, + 201, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 201, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 201, + 623 + ], + "type": "text", + "content": "2.1. Large Multimodal Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) [3, 8, 21, 22] extend the reasoning capabilities of Large Language Models (LLMs) [4, 32, 38, 39] to vision-language tasks by integrating a pre-trained vision encoder [29, 46] with a language model, linked by an alignment module such as an MLP, or a query-based network. This design transforms visual inputs into token sequences that the LLM can pro" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "content": "cess, facilitating multimodal prompts for tasks like visual question answering [2, 6, 12, 44]. To enhance performance, advanced LMMs, such as LLaVA-OneVision [17] and Qwen2-VL [34], can encode higher-resolution images into more image tokens, thereby capturing more granular visual details. However, as image resolution grows, the number of visual tokens rises exponentially, leading to substantially higher computational costs. For example, LLaVA-1.5 [21] processes " + }, + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "content": " images into 576 tokens, whereas LLaVA-OneVision can handle " + }, + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "inline_equation", + "content": "1152 \\times 1152" + }, + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "content": " images, producing 7,290 tokens. This challenge becomes even more pronounced in video-based models like LongVA [48], which must process tokens across numerous frames. Fine-grained visual tokenization boosts LMM performance but poses an inference bottleneck, driving efforts to balance performance and cost through token reduction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 270, + 468, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 468, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 468, + 282 + ], + "type": "text", + "content": "2.2. Token Reduction for LMMs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 556, + 587 + ], + "type": "text", + "content": "Token reduction has become a key strategy for improving the efficiency of LMMs by mitigating the computational cost associated with extensive visual token sequences. Existing methods can be broadly classified into training-based and training-free approaches. Training-based methods, such as LLaVA-Mini [50] and LLaVolta [5], introduce additional modules during model training to compress visual tokens and preserve critical information, while approaches like ATP-LLaVA [43] and p-MoD [47] train pruning modules to dynamically retain important tokens across LLM layers. However, these methods demand substantial computational resources to retrain the models, limiting their real-world applicability. In contrast, training-free methods prune tokens without additional training, often leveraging attention mechanisms to identify and discard redundant tokens. For instance, FastV[7] and SparseVLM exploit text-visual attention to rank token importance, whereas FasterVLM [49], VisionZip [40], and VTC-CLS [33] rely on [CLS] token attention in the vision encoder to evaluate token importance and prune redundant tokens. FitPrune [42] and PDrop [36] propose a progressive multistage pruning strategy in LMMs to fully utilize visual information. However, the pruning strategies in these methods often lack adaptability to different LMMs, which frequently leads to significant performance drops." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 597, + 370, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 370, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 370, + 609 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "In this section, we introduce VFlowOpt, a framework that employs a pruning strategy to reduce redundant visual tokens during LMM inference while preserving essential visual information. This strategy involves multiple hyperparameters that our VFlowOpt framework automatically optimizes for proper configuration—an essential factor for maintaining model performance. Specifically, Section 3.1 explains how our method evaluates the importance of visual" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23925" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 87, + 263, + 208 + ], + "blocks": [ + { + "bbox": [ + 86, + 73, + 233, + 84 + ], + "lines": [ + { + "bbox": [ + 86, + 73, + 233, + 84 + ], + "spans": [ + { + "bbox": [ + 86, + 73, + 233, + 84 + ], + "type": "text", + "content": "(a) Token Importance Estimation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 87, + 263, + 208 + ], + "lines": [ + { + "bbox": [ + 61, + 87, + 263, + 208 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 263, + 208 + ], + "type": "image", + "image_path": "11f3eec939c6bc6f8cac28c20264c1deb05bddf83b8a2d827d4fc8af5103a5c8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 274, + 90, + 370, + 203 + ], + "blocks": [ + { + "bbox": [ + 276, + 72, + 366, + 84 + ], + "lines": [ + { + "bbox": [ + 276, + 72, + 366, + 84 + ], + "spans": [ + { + "bbox": [ + 276, + 72, + 366, + 84 + ], + "type": "text", + "content": "(b) Token Recycling" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 274, + 90, + 370, + 203 + ], + "lines": [ + { + "bbox": [ + 274, + 90, + 370, + 203 + ], + "spans": [ + { + "bbox": [ + 274, + 90, + 370, + 203 + ], + "type": "image", + "image_path": "bd996f6af1ded17ff73c3562764b3550a45b47a57cd004df5ca48fe7e66614ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 230, + 373, + 370 + ], + "blocks": [ + { + "bbox": [ + 62, + 215, + 165, + 228 + ], + "lines": [ + { + "bbox": [ + 62, + 215, + 165, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 165, + 228 + ], + "type": "text", + "content": "(1) Inference Stage" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 230, + 373, + 370 + ], + "lines": [ + { + "bbox": [ + 60, + 230, + 373, + 370 + ], + "spans": [ + { + "bbox": [ + 60, + 230, + 373, + 370 + ], + "type": "image", + "image_path": "ac77e9c733e7b164f80cd00a8207b0782e0c95b091775067933bce49f861a589.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 547, + 322 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 547, + 322 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 547, + 322 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 547, + 322 + ], + "type": "image", + "image_path": "3c405a4b6ddb48846a13a61c0ebe6a3c45543abb3506bdd88281f3b31d79216c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 328, + 523, + 342 + ], + "lines": [ + { + "bbox": [ + 402, + 328, + 523, + 342 + ], + "spans": [ + { + "bbox": [ + 402, + 328, + 523, + 342 + ], + "type": "text", + "content": "(2) Optimization Stage" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 357, + 480, + 368 + ], + "blocks": [ + { + "bbox": [ + 427, + 357, + 480, + 368 + ], + "lines": [ + { + "bbox": [ + 427, + 357, + 480, + 368 + ], + "spans": [ + { + "bbox": [ + 427, + 357, + 480, + 368 + ], + "type": "image", + "image_path": "08f38546afb6798412810532e22ad42dbf3845ad1e93985beeff65b10ec97b23.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 66, + 456, + 284, + 565 + ], + "blocks": [ + { + "bbox": [ + 55, + 373, + 555, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 555, + 452 + ], + "type": "text", + "content": "Figure 2. Overview of VFlowOpt. (1) During inference, VFlowOpt first assesses the importance of visual tokens, based on which progressive token pruning is performed. After the initial pruning stage, the pruned tokens are merged and recycled. The pruning strategy used in this process is defined by the (2) Optimization Stage. (a) The importance map is computed by combining the attention of relatively important tokens with the entropy of image patches. (b) The pruned tokens are grouped into grid cells, where each cell has a side length of a. Within each grid cell, the pruned tokens are fused using a weighted average, with their importance values as weights, and then recycled. (2) VFlowOpt optimizes the pruning strategy by minimizing the discrepancy of the last token in the final layer of the LMM with and without applying the pruning strategy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 456, + 284, + 565 + ], + "lines": [ + { + "bbox": [ + 66, + 456, + 284, + 565 + ], + "spans": [ + { + "bbox": [ + 66, + 456, + 284, + 565 + ], + "type": "image", + "image_path": "9c4ac0f406f77133d4fb72a2bffd19daf953d57a3f3ef903f97d9543a1ade086.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 571, + 295, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 605 + ], + "type": "text", + "content": "Figure 3. The attention of redundant tokens (marked in red) fails to reflect the importance of other tokens and instead focuses on similar tokens, such as background elements." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "type": "text", + "content": "tokens, Section 3.2 presents progressive pruning and token recycling strategies, and finally, Section 3.3 details how VFlowOpt customizes the superior pruning strategy (i.e., selects the pruning hyperparameters) for different LMMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 251, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 251, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 251, + 685 + ], + "type": "text", + "content": "3.1. Visual Token Importance Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "For a general approach to evaluate visual token importance, the previous works [14, 41] propose that the importance of" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 457, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 636 + ], + "type": "text", + "content": "visual tokens can be estimated using the average attention from all other tokens in ViTs. Although it has achieved some success, we find that redundant tokens (e.g., tokens corresponding to background regions) often assign disproportionately high attention to other similar redundant tokens (shown in Fig. 3), reducing the reliability of the importance estimation. To address this, we first identify relatively important tokens based on the attention they receive from all tokens (tokens receiving higher attention are considered relatively important), as shown in Fig. 2 (a). We then exclude redundant tokens and use the attention from these relatively important tokens as a more robust metric. Additionally, we incorporate the information entropy of image patches into the importance score to prioritize tokens corresponding to visually informative regions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 638, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 555, + 660 + ], + "type": "text", + "content": "To identify relatively important visual tokens, we define a threshold:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 386, + 661, + 555, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 661, + 555, + 695 + ], + "spans": [ + { + "bbox": [ + 386, + 661, + 555, + 695 + ], + "type": "interline_equation", + "content": "\\tau = t \\cdot \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {N} A _ {i j}, \\tag {1}", + "image_path": "7d50fcacb6fc5fd3fabad5ec5aae19c37ee8be9c3a1bc8dc51dd8c5ea19d3507.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": " is the threshold, " + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "A_{ij}" + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": " is the attention weight from" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23926" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "token " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " to token " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " within a ViT layer, " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " is the total number of tokens, and " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " is a sensitivity hyperparameter. We then treat tokens with a total attention exceeding " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " as relatively important:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 120, + 294, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 120, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 125, + 120, + 294, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {K} = \\{j \\mid \\sum_ {i = 1} ^ {N} A _ {i j} > \\tau \\}. \\tag {2}", + "image_path": "ae1067e6124002b4c0800cc917b1b11097596ee219df0d9cee2d6f896c6c0ae7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 178 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 156, + 294, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 55, + 156, + 294, + 178 + ], + "type": "text", + "content": " denotes the set of indices of these relatively important tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 179, + 294, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 294, + 202 + ], + "type": "text", + "content": "The information entropy of an image patch corresponding to token " + }, + { + "bbox": [ + 55, + 179, + 294, + 202 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 179, + 294, + 202 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 211, + 294, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 211, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 121, + 211, + 294, + 243 + ], + "type": "interline_equation", + "content": "H \\left(V _ {i}\\right) = - \\sum_ {k = 0} ^ {L - 1} p _ {k} \\log p _ {k}, \\tag {3}", + "image_path": "c60ed71ae86e9bba8479526f5a6c7f0891ea548b84c88b3a9b77f8de596d80b0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "text", + "content": " is the image patch corresponding to token " + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "text", + "content": " is the proportion of pixels in the patch with gray level " + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 251, + 295, + 333 + ], + "type": "text", + "content": " is the total number of possible gray levels (256 for 8-bit images). Here, the gray level is defined as the average of the RGB channel values, providing a scalar representation of pixel brightness. Higher entropy reflects greater diversity in pixel intensities and richer visual information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 335, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 381 + ], + "type": "text", + "content": "The importance score of token " + }, + { + "bbox": [ + 55, + 335, + 296, + 381 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 335, + 296, + 381 + ], + "type": "text", + "content": " is computed by combining the attention it receives from tokens in " + }, + { + "bbox": [ + 55, + 335, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 55, + 335, + 296, + 381 + ], + "type": "text", + "content": " with the entropy of its corresponding image patch, normalized via the softmax function:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 388, + 294, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 388, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 89, + 388, + 294, + 419 + ], + "type": "interline_equation", + "content": "I _ {i} = \\sum_ {k \\in \\mathcal {K}} A _ {k i} + \\alpha \\cdot \\frac {\\exp (H (V _ {i}))}{\\sum_ {j = 1} ^ {N} \\exp (H (V _ {j}))}, \\tag {4}", + "image_path": "3a263cca687d2c9c5cb8e368ad1291c69871d9158abe36c77b8aa335455d36a1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": " denotes the importance score of token " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "A_{ki}" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": " is the attention weight from token " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": " to token " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "H(V_{i})" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": " is the entropy of the image patch corresponding to token " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": " is a hyperparameter that determines the entropy term's contribution, and " + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": " is the total number of tokens. This approach mitigates biases from redundant tokens and leverages the visual information of image patches, thereby facilitating effective estimation of visual token importance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 529, + 273, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 273, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 273, + 541 + ], + "type": "text", + "content": "3.2. Progressive Pruning and Token Recycling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Previous studies [37, 51] indicate that visual tokens play a more critical role in the shallower layers of LMMs, whereas redundancy tends to increase in deeper layers. Based on this, we adopt a progressive pruning strategy. To balance the pruning process's simplicity with its fine-grained configuration, we evenly divide the LMM into three stages. At the beginning of each stage, a predetermined fraction of visual tokens with higher importance scores (as defined in Sec. 3.1) is retained according to the stage-specific retention ratios " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "R = [R_{1}, R_{2}, R_{3}]" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", while the rest are pruned. Notably, the position IDs of visual tokens remain unchanged following pruning, preserving the original spatial structure of the visual input. Following the previous work [37], we compute the average visual token retention rate across the entire" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 388, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 388, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 388, + 83 + ], + "type": "text", + "content": "LMM as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 91, + 553, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 91, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 326, + 91, + 553, + 114 + ], + "type": "interline_equation", + "content": "\\bar {R} = \\frac {R _ {1} \\cdot L _ {1} + R _ {1} \\cdot R _ {2} \\cdot L _ {2} + R _ {1} \\cdot R _ {2} \\cdot R _ {3} \\cdot L _ {3}}{L}, \\tag {5}", + "image_path": "f3cc9c615509a46ae5ec82618e19c7a193c9375424e3532ce285c14ce97d4cbf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "inline_equation", + "content": "L_{1}, L_{2}" + }, + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "inline_equation", + "content": "L_{3}" + }, + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "text", + "content": " represent the number of layers in the three stages, respectively, and " + }, + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 124, + 553, + 159 + ], + "type": "text", + "content": " denotes the total number of layers in the LMM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 160, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 315 + ], + "type": "text", + "content": "To prevent the loss of any small but potentially significant information during the initial pruning process (before visual tokens are fed into the LLM), we propose a token merging and recycling strategy to compactly represent redundant visual information. Specifically, We use a square grid (with each cell having side length " + }, + { + "bbox": [ + 313, + 160, + 553, + 315 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 160, + 553, + 315 + ], + "type": "text", + "content": ") to group pruned tokens that fall into the same grid cell, as shown in Fig. 2 (b). Within each cell, the pruned tokens are fused into a single token by computing a weighted average of their representations, using their importance scores as weights. This fused token then replaces the pruned token with the highest importance score in that cell and is incorporated into the set of retained tokens." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "content": "Formally, a token " + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "content": " with spatial coordinates " + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "inline_equation", + "content": "(x_i, y_i)" + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "content": " belongs to a grid cell " + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{p,q}" + }, + { + "bbox": [ + 313, + 316, + 553, + 342 + ], + "type": "text", + "content": " if" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 368, + 354, + 553, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 354, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 368, + 354, + 553, + 368 + ], + "type": "interline_equation", + "content": "\\lfloor x _ {i} / a \\rfloor = p \\quad \\text {a n d} \\quad \\lfloor y _ {i} / a \\rfloor = q, \\tag {6}", + "image_path": "fafad2e0cb5936e0172042a45c624ff1ad875146ebc33aa7915016c1a80e98f8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": " are the row and column indices of the grid cell, respectively. Suppose there are " + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": " pruned tokens in " + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{p,q}" + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": " with corresponding importance scores " + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "inline_equation", + "content": "I_{1}, I_{2}, \\ldots, I_{k}" + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": " and token representations " + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_1, \\mathbf{t}_2, \\ldots, \\mathbf{t}_k" + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": ". The fused token " + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{merged}}^{p,q}" + }, + { + "bbox": [ + 313, + 374, + 553, + 434 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 445, + 553, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 386, + 445, + 553, + 476 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {\\text {m e r g e d}} ^ {p, q} = \\frac {\\sum_ {i = 1} ^ {k} I _ {i} \\cdot \\mathbf {t} _ {i}}{\\sum_ {i = 1} ^ {k} I _ {i}}. \\tag {7}", + "image_path": "b02869541267d6f1f474e91f6f4661fccd47fc4133201b5007e4c43ce142cc30.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "text", + "content": "We then assign " + }, + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{merged}}^{p,q}" + }, + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "text", + "content": " to the position of the pruned token with the highest importance score " + }, + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{p,q}" + }, + { + "bbox": [ + 313, + 482, + 553, + 543 + ], + "type": "text", + "content": ", counting it among the retained tokens. By combining token pruning with this recycling strategy, we reduce token redundancy while avoiding potential information loss." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 552, + 482, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 482, + 565 + ], + "type": "text", + "content": "3.3. Pruning Strategy Optimization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "The pruning strategies described in Sec. 3.1 and Sec. 3.2 involve several key hyperparameters that directly affect the performance of LMMs after token pruning. Properly defining these hyperparameters is crucial for preserving model performance. However, existing approaches often rely on manually designed pruning strategies and apply the same strategy across different LMMs, without considering the unique characteristics of each model. This coarse-grained approach can lead to significant performance degradation due to its lack of fitness for different LMMs, leaving the task of designing pruning strategies tailored to specific LMMs as a formidable challenge." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23927" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": "Previous interpretability studies of LMMs [15, 53] offer crucial insights into their internal mechanisms. Specifically, these works reveal that in the lower and middle layers of LMMs, visual information from visual tokens is aggregated into the corresponding query text tokens. In the higher layers, the multimodal representation encoded in the query text tokens is further progressively propagated to the final position of the input sequence, ultimately influencing the subsequent inference process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 297, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 297, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 297, + 384 + ], + "type": "text", + "content": "Inspired by this insight, we propose a framework that requires only a small amount of unlabeled data and leverages the internal flow of visual information within LMMs to search for the superior pruning strategy tailored to the characteristics of different LMMs, as shown in Fig. 2 (2). Specifically, we recast the task of designing the superior pruning strategy as an optimization problem, aiming to minimize the discrepancy in visual information flow with and without applying a pruning strategy to perform visual token pruning. In this framework, we treat the final token in the last layer as the representative outcome of the visual information flow. We use cosine similarity to measure the similarity of the final token representation with and without visual token pruning. A higher similarity indicates that the discrepancy between the visual information flow with and without pruning is smaller. Accordingly, we define the following optimization objective:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 393, + 295, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 393, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 97, + 393, + 295, + 410 + ], + "type": "interline_equation", + "content": "\\max _ {s \\in \\mathcal {S}} f (s) = \\operatorname {C o s i n e S i m} \\left(h _ {f}, g _ {s} \\left(h _ {f}\\right)\\right). \\tag {8}", + "image_path": "cb3b702778c9336f27d7bfe235b68a73352a72dbf19333e968433ae9a33c2d90.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "inline_equation", + "content": "h_f" + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": " represents the representation of the final token in the last layer before pruning, and " + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "inline_equation", + "content": "g_s(h_f)" + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": " represents the final token representation after applying the pruning strategy " + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "inline_equation", + "content": "g_s" + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": " modeling its effect on the token feature), and " + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": " is the solution space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": "To efficiently search for the superior pruning strategy " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", we employ Bayesian optimization, which systematically explores the hyperparameter space—including the threshold sensitivity " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", the entropy weight " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", the grid size " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", and the pruning ratios " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "R_{1}, R_{2}" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "R_{3}" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": "—to maximize the target function " + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "inline_equation", + "content": "f(s)" + }, + { + "bbox": [ + 55, + 479, + 296, + 611 + ], + "type": "text", + "content": ". In short, Bayesian optimization constructs a surrogate model to approximate the objective function and employs an acquisition function to balance exploration and exploitation, thereby efficiently guiding the search for promising hyperparameter settings. The details of this optimization process are presented in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In this section, we evaluate our approach on various LMMs across diverse image and video benchmarks, followed by an efficiency analysis and an ablation study of each component. Finally, we illustrate how our method affects discrepancies in the visual information flow with and without pruning, offering deeper insights into our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 72, + 529, + 84 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 529, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 529, + 84 + ], + "type": "text", + "content": "Algorithm 1 VFlowOpt with Bayesian Optimization" + } + ] + } + ], + "index": 7, + "type": "text" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": "Input: LMM " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": ", data samples " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": ", number of Bayesian optimization iterations " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": ", computation budget " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\overline{R}" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": ", target function " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": ": the sum of cosine similarities between the last token in the last layer of the LMM " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": " with and without visual token pruning, computed over all data samples " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": ". \nOutput: retention rates " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "R_{1}, R_{2}, R_{3}" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": "; threshold sensitivity " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": "; entropy weight in importance score " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": "; grid size " + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 315, + 87, + 555, + 182 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 183, + 553, + 231 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 321, + 183, + 493, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 183, + 493, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 183, + 493, + 194 + ], + "type": "text", + "content": "1: Initialize a Gaussian Process model " + }, + { + "bbox": [ + 321, + 183, + 493, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{GP}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 196, + 553, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 196, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 321, + 196, + 553, + 219 + ], + "type": "text", + "content": "2: Define the acquisition function " + }, + { + "bbox": [ + 321, + 196, + 553, + 219 + ], + "type": "inline_equation", + "content": "A(\\cdot)" + }, + { + "bbox": [ + 321, + 196, + 553, + 219 + ], + "type": "text", + "content": " (Expected Improvement is adopted)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 220, + 462, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 220, + 462, + 231 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 462, + 231 + ], + "type": "text", + "content": "3: Uniformly sample initial points:" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 331, + 240, + 560, + 254 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 240, + 560, + 254 + ], + "spans": [ + { + "bbox": [ + 331, + 240, + 560, + 254 + ], + "type": "interline_equation", + "content": "X _ {0} = \\left\\{\\left(R _ {1}, R _ {2}, t, \\alpha , a\\right) \\mid R _ {v}, R _ {1}, t, \\alpha , a \\in \\text {v a l i d r a n g e s} \\right\\}", + "image_path": "8658a1e6e9a97056c0fa5d2b5ae6c36cd011b31ac2b30f97590b5ef5b770651c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 263, + 406, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 263, + 406, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 263, + 406, + 274 + ], + "type": "text", + "content": "4: for all " + }, + { + "bbox": [ + 321, + 263, + 406, + 274 + ], + "type": "inline_equation", + "content": "x \\in X_0" + }, + { + "bbox": [ + 321, + 263, + 406, + 274 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 275, + 553, + 323 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 321, + 275, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 275, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 321, + 275, + 553, + 299 + ], + "type": "text", + "content": "5: Calculate " + }, + { + "bbox": [ + 321, + 275, + 553, + 299 + ], + "type": "inline_equation", + "content": "R_{3}" + }, + { + "bbox": [ + 321, + 275, + 553, + 299 + ], + "type": "text", + "content": " using the constraint: " + }, + { + "bbox": [ + 321, + 275, + 553, + 299 + ], + "type": "inline_equation", + "content": "\\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 300, + 543, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 300, + 543, + 311 + ], + "spans": [ + { + "bbox": [ + 321, + 300, + 543, + 311 + ], + "type": "text", + "content": "6: Form pruning strategy: " + }, + { + "bbox": [ + 321, + 300, + 543, + 311 + ], + "type": "inline_equation", + "content": "s = (R_1, R_2, R_3, t, \\alpha, a)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 312, + 505, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 312, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 321, + 312, + 505, + 323 + ], + "type": "text", + "content": "7: Evaluate the target function " + }, + { + "bbox": [ + 321, + 312, + 505, + 323 + ], + "type": "inline_equation", + "content": "f(s \\mid \\theta, \\mathcal{D})" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 323, + 365, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 323, + 365, + 333 + ], + "spans": [ + { + "bbox": [ + 321, + 323, + 365, + 333 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 335, + 424, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 335, + 424, + 346 + ], + "spans": [ + { + "bbox": [ + 321, + 335, + 424, + 346 + ], + "type": "text", + "content": "9: for " + }, + { + "bbox": [ + 321, + 335, + 424, + 346 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 321, + 335, + 424, + 346 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 335, + 424, + 346 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 321, + 335, + 424, + 346 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 346, + 553, + 442 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 318, + 346, + 536, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 346, + 536, + 358 + ], + "spans": [ + { + "bbox": [ + 318, + 346, + 536, + 358 + ], + "type": "text", + "content": "10: Fit " + }, + { + "bbox": [ + 318, + 346, + 536, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{GP}" + }, + { + "bbox": [ + 318, + 346, + 536, + 358 + ], + "type": "text", + "content": " to the observed data " + }, + { + "bbox": [ + 318, + 346, + 536, + 358 + ], + "type": "inline_equation", + "content": "(S_{n},f(S_{n}\\mid \\theta ,\\mathcal{D}))" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 359, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 359, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 318, + 359, + 553, + 370 + ], + "type": "text", + "content": "11: Select the next point: " + }, + { + "bbox": [ + 318, + 359, + 553, + 370 + ], + "type": "inline_equation", + "content": "x_{n + 1}\\gets \\arg \\max_xA(x;\\mathcal{GP})" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 371, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 371, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 318, + 371, + 553, + 395 + ], + "type": "text", + "content": "12: Calculate " + }, + { + "bbox": [ + 318, + 371, + 553, + 395 + ], + "type": "inline_equation", + "content": "R_{3}" + }, + { + "bbox": [ + 318, + 371, + 553, + 395 + ], + "type": "text", + "content": " using the constraint: " + }, + { + "bbox": [ + 318, + 371, + 553, + 395 + ], + "type": "inline_equation", + "content": "\\overline{R} = (R_{1} \\cdot L_{1} + R_{1} \\cdot R_{2} \\cdot L_{2} + R_{1} \\cdot R_{2} \\cdot R_{3} \\cdot L_{3}) / L" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 396, + 551, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 396, + 551, + 407 + ], + "spans": [ + { + "bbox": [ + 318, + 396, + 551, + 407 + ], + "type": "text", + "content": "13: Form pruning strategy: " + }, + { + "bbox": [ + 318, + 396, + 551, + 407 + ], + "type": "inline_equation", + "content": "s_{n+1} = (R_1, R_2, R_3, t, \\alpha, a)_{n+1}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 407, + 470, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 407, + 470, + 418 + ], + "spans": [ + { + "bbox": [ + 318, + 407, + 470, + 418 + ], + "type": "text", + "content": "14: Update " + }, + { + "bbox": [ + 318, + 407, + 470, + 418 + ], + "type": "inline_equation", + "content": "S_{n + 1}\\gets S_n\\cup \\{s_{n + 1}\\}" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 419, + 364, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 419, + 364, + 429 + ], + "spans": [ + { + "bbox": [ + 318, + 419, + 364, + 429 + ], + "type": "text", + "content": "15: end for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 430, + 517, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 430, + 517, + 442 + ], + "spans": [ + { + "bbox": [ + 318, + 430, + 517, + 442 + ], + "type": "text", + "content": "16: return " + }, + { + "bbox": [ + 318, + 430, + 517, + 442 + ], + "type": "inline_equation", + "content": "(R_{1}, R_{2}, R_{3}, t, \\alpha, a)^{*}" + }, + { + "bbox": [ + 318, + 430, + 517, + 442 + ], + "type": "text", + "content": " that maximize " + }, + { + "bbox": [ + 318, + 430, + 517, + 442 + ], + "type": "inline_equation", + "content": "f" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 468, + 428, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 428, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 428, + 480 + ], + "type": "text", + "content": "4.1. Experiment Setting" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": "Datasets. We evaluate our method on ten image-based multimodal benchmarks: GQA [1], VizWiz [13], ScienceQA-IMG [26], TextVQA [30], ChartQA [27], POPE [19], MME [10], MMBench [25], MMStar [6], and DocVQA [28]. For video understanding, we adopt two datasets—SeedBench (video) [16] and VideoMME [11], where VideoMME is partitioned by video length into short, medium, and long subsets. Further details are presented in the Appendix." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 594, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 677 + ], + "type": "text", + "content": "Model Architectures. We integrate VFlowOpt into multiple LMMs, including LLaVA-OneVision-7B, LLaVA-NeXT-7B [23], and Qwen2-VL-7B [34]. These models employ various vision encoders (SigLIP, CLIP, and a ViT designed for Qwen2-VL) and LLM backbones (Qwen2-7B and Vicuna-7B). To prevent out-of-memory issues in Qwen2-VL, we set maxPixels = 3000000." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "Comparison Methods. We compare VFlowOpt against five baseline methods: FastV [7], FitPrune [42], PDrop [36], SparseVLM [52], and VisionZip [40]. FastV," + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23928" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 331 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 331 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 331 + ], + "type": "table", + "html": "
MethodMMStarMMEMMBSQAPOPEGQAVizWiz\\(VQA^{Text}\\)ChartQADocVQAAvg.
Upper Bound, 100% Tokens
Vanilla61.7158180.895.889.162.260.476.080.087.5100%
Retain 50% Tokens (↓ 50%)
FastV (ECCV24)58.9154979.492.887.961.561.172.568.684.096.5%
FitPrune (AAAI25)4.596.61.08.618.816.26.837.637.730.721.4%
PDrop (CVPR25)37.1628.632.164.881.344.248.963.762.762.271.3%
SparseVLM (2024.10)59.8157780.594.188.161.960.473.970.580.897.1%
VisionZip (CVPR25)60.4158780.394.689.362.759.874.275.488.498.9%
VFlowOpt61.3159181.195.489.462.460.075.177.890.099.9%
Retain 25% Tokens (↓ 75%)
FastV (ECCV24)54.0153977.088.683.858.261.058.342.762.986.4%
FitPrune (AAAI25)5.292.61.15.920.214.55.129.727.323.217.6%
PDrop (CVPR25)36.8574.431.362.878.042.247.156.347.547.065.2%
SparseVLM (2024.10)56.6152078.790.387.259.760.866.354.066.690.5%
VisionZip (CVPR25)54.6156278.990.488.861.060.470.066.379.694.3%
VFlowOpt57.8157079.992.389.161.260.472.569.182.396.3%
Retain 10% Tokens (↓ 90%)
FastV (ECCV24)46.0120970.181.777.051.556.435.621.333.269.7%
FitPrune (AAAI25)5.427.00.93.52.34.60.76.95.96.95.1%
PDrop (CVPR25)31.8456.322.559.065.935.743.333.719.618.549.9%
SparseVLM (2024.10)45.1119171.883.780.054.656.439.837.639.674.0%
VisionZip (CVPR25)49.5138974.886.286.157.256.856.446.149.082.1%
VFlowOpt52.0146475.188.485.257.357.360.253.656.185.5%
", + "image_path": "c08e4252dfd62ec3e56072a83d39363f69a64d4c10b6a680c0e600cc68ee8b4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 336, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 555, + 360 + ], + "type": "text", + "content": "Table 1. Performance comparison on LLaVA-OneVision-7B under different token retention conditions. \"Avg.\" refers to average accuracy on 10 benchmarks. For each reduction ratio, the best average performance is shown in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 376, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 532 + ], + "type": "text", + "content": "FitPrune, PDrop, and SparseVLM rely on text-visual attention in the LLM to prune visual tokens but differ as follows: FastV performs a single pruning step after the second LLM layer; FitPrune prunes tokens in each layer according to predefined, layer-specific ratios; PDrop conducts four pruning steps within the LLM, using the same retention ratio each time; and SparseVLM uses the attention weights of preselected text tokens to evaluate the importance of visual tokens. VisionZip, in contrast, determines token importance based on the [cls] token's attention; for models lacking a [cls] token, we follow VisionZip's original procedure by computing the average attention each token receives from every other token in the ViTs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 532, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 712 + ], + "type": "text", + "content": "Implementation Details. For LLaVA-OneVision-7B and Qwen2-VL-7B, we perform token pruning at three distinct points: before the LLM, and after the 9th and 18th layers. For LLaVA-NeXT-7B, pruning is conducted before the LLM, and again after the 10th and 20th layers. During optimization, we sample 30 unlabeled instances from each model's training datasets; for models without publicly available training dataset, we instead use random samples from the LLaVA-OneVision training set. The optimization is performed for a total of 50 iterations. For LMMs that modify visual tokens output by the vision encoder (e.g., unpadding and interpolation in LLaVA-OneVision, and unpadding in LLaVA-NeXT), we apply corresponding transformations to the importance maps so they remain fully aligned with the final visual tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 376, + 468, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 468, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 468, + 389 + ], + "type": "text", + "content": "4.2. Image Understanding Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "content": "We evaluate the proposed VFlowOpt on image understanding benchmarks with LLaVA-OneVision-7B using various pruning ratios, and present the results in Tab. 1. Compared to other baselines, VFlowOpt consistently maintains superior accuracy across different levels of token pruning. With " + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "content": " token retention, VFlowOpt achieves " + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "content": " of the original performance, exceeding the second-best approach by " + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "content": ". This negligible performance drop underscores the method's strong potential for practical deployment. Under more extreme pruning conditions with very few retained tokens, VFlowOpt's advantage becomes more pronounced. When only " + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "content": " of the visual tokens are retained, VFlowOpt preserves " + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "inline_equation", + "content": "85.5\\%" + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "content": " of the original performance, surpassing the second-best approach by " + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 312, + 398, + 555, + 662 + ], + "type": "text", + "content": ". This finding suggests that VFlowOpt can effectively leverage limited visual information to maintain high performance under strict computational budgets. When applied to LLaVA-OneVision, both FitPrune and PDrop disrupt the model's normal output, indicating their inability to devise effective pruning strategies for different architectures. In contrast, VFlowOpt demonstrates a clear advantage by tailoring its pruning strategy to each model's unique characteristics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "To further validate VFlowOpt's generalization capability, we evaluate it on LLaVA-NeXT-7B and Qwen2-VL-7B. As shown in Tab. 2 and Tab. 3, VFlowOpt once again delivers the top results on these LMMs. When retaining " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "25\\%" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23929" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 298, + 193 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 298, + 193 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 298, + 193 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 298, + 193 + ], + "type": "table", + "html": "
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla37.6151967.470.186.564.2100%
Retain 25% Tokens (↓ 75%)
FastV35.1147765.667.483.160.495.7%
VisionZip35.8150165.467.986.761.597.3%
VFlowOpt37.0151467.067.787.662.698.8%
Retain 10% Tokens (↓ 90%)
FastV29.2128261.663.871.755.985.7%
VisionZip32.6137861.567.183.557.091.6%
VFlowOpt35.1139362.967.483.657.393.4%
", + "image_path": "89770b1edab563629f5768444b850141eae77d84f4414fa572579c778afb191b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 215, + 295, + 338 + ], + "blocks": [ + { + "bbox": [ + 72, + 195, + 278, + 206 + ], + "lines": [ + { + "bbox": [ + 72, + 195, + 278, + 206 + ], + "spans": [ + { + "bbox": [ + 72, + 195, + 278, + 206 + ], + "type": "text", + "content": "Table 2. Comparative experiments on LLaVA-NeXT-7B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 215, + 295, + 338 + ], + "lines": [ + { + "bbox": [ + 59, + 215, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 215, + 295, + 338 + ], + "type": "table", + "html": "
MethodMMStarMMEMMBSQAPOPEGQAAvg.
Upper Bound, 100% Tokens
Vanilla57.5168080.384.788.462.2100%
Retain 25% Tokens (↓ 75%)
FastV54.5159776.379.081.857.293.8
VisionZip55.2161878.981.386.960.096.8
VFlowOpt55.9165979.881.587.160.497.8
Retain 10% Tokens (↓ 90%)
FastV44.9140570.075.575.851.684.4
VisionZip49.3151876.578.284.354.190.9
VFlowOpt51.0159178.078.684.554.892.8
", + "image_path": "36abdb835487ebf46093e346e5ae6ca0c9428afb043cb14d96319d2e489f57f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 358, + 294, + 495 + ], + "blocks": [ + { + "bbox": [ + 77, + 339, + 272, + 350 + ], + "lines": [ + { + "bbox": [ + 77, + 339, + 272, + 350 + ], + "spans": [ + { + "bbox": [ + 77, + 339, + 272, + 350 + ], + "type": "text", + "content": "Table 3. Comparative experiments on Qwen2-VL-7B." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 358, + 294, + 495 + ], + "lines": [ + { + "bbox": [ + 59, + 358, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 59, + 358, + 294, + 495 + ], + "type": "table", + "html": "
MethodSeedBench (video)VideoMMEAvg.
SML
Upper Bound, 100% Tokens
Vanilla56.970.554.649.5100%
Retain 25% Tokens (↓75%)
FastV54.766.953.247.796.2%
VisionZip56.468.355.349.099.0%
VFlowOpt56.868.955.849.2100%
Retain 10% Tokens (↓90%)
FastV48.753.747.642.383.6%
VisionZip55.059.751.846.392.4%
VFlowOpt55.563.352.748.695.5%
", + "image_path": "40d07dfea5d2b525bc1b5d6e804cc5d1b171db9c191f8b647fb6c7255033c3b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 497, + 292, + 508 + ], + "lines": [ + { + "bbox": [ + 58, + 497, + 292, + 508 + ], + "spans": [ + { + "bbox": [ + 58, + 497, + 292, + 508 + ], + "type": "text", + "content": "Table 4. Comparative experiments on video understanding tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": "of the tokens, VFlowOpt preserves " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "98.8\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": " of the original performance on LLaVA-NeXT-7B and " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "97.8\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": " on Qwen2-VL-7B. Even under the stringent condition of retaining only " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": " of the tokens, VFlowOpt maintains " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "93.4\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": " of the performance on LLaVA-NeXT-7B and " + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "inline_equation", + "content": "92.8\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": " on Qwen2-VL-7B, demonstrating its strong generalizability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 218, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 218, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 218, + 611 + ], + "type": "text", + "content": "4.3. Generalization to Video Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "We further explore VFlowOpt's generalization across different modalities by evaluating it on video benchmarks using LLaVA-OneVision-7B. As shown in Tab. 4, VFlowOpt surpasses other baselines under various token retention ratios. Notably, when only " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " of the tokens are preserved, VFlowOpt incurs virtually no performance loss on LLaVA-OneVision-7B, maintaining " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " of its original performance. Even under the more stringent condition of retain" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "ing just " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": " of the tokens, VFlowOpt still retains " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "95.5\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": " of the original performance, underscoring its robust generalization in the video domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 117, + 425, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 425, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 425, + 131 + ], + "type": "text", + "content": "4.4. Efficiency Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 135, + 553, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 553, + 207 + ], + "type": "text", + "content": "We demonstrate the efficiency of VFlowOpt by conducting a comparative study on LLaVA-OneVision-7B running on a single NVIDIA A100-SXM4-80GB GPU, focusing on FLOPs, KV-Cache memory usage, and inference latency. The efficiency analysis of other baseline methods is provided in the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": "We assess various token pruning ratios on the MME benchmark, measuring overall performance and average efficiency metrics across all samples. As shown in Tab. 5, VFlowOpt significantly enhances LLaVA-OneVision-7B's computational efficiency by reducing FLOPs, shrinking KV-Cache memory usage, and accelerating inference speed. Notably, at a " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " token pruning ratio, VFlowOpt achieves a " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "49.5\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " reduction in KV-Cache memory and a " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "1.8\\times" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " speedup in inference, while boosting performance by " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": ". At a more aggressive " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " pruning ratio, we observe a " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "74.2\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " reduction in KV-Cache memory and a " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "3.1\\times" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " speedup, with only a " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " performance drop. Under extremely tight computational budgets, pruning " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": " of visual tokens reduces the KV-Cache memory footprint by " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": ", accelerates inference by " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "3.8\\times" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": ", and degrades performance by only " + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 313, + 207, + 555, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 387, + 556, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 507 + ], + "type": "text", + "content": "Such reductions in inference latency substantially benefit user-facing applications demanding real-time performance, such as autonomous driving and robotics. Moreover, by significantly shrinking the KV-Cache, large-batch inference on LMMs can accommodate more user requests simultaneously, thereby substantially reducing overall inference costs. In summary, VFlowOpt uses far less GPU memory and delivers faster inference while preserving model performance, offering a highly practical solution for efficiently deploying LMMs in real-world scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 515, + 408, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 408, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 408, + 528 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": "To confirm the contributions of each component in our token pruning strategy, we conduct ablation experiments on LLaVA-OneVision-7B under a computational budget of retaining " + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 700 + ], + "type": "text", + "content": " of tokens. As shown in Tab. 6, removing importance calibration (i.e., directly using the mean attention over all tokens received by each visual token as its importance score), omitting token recycling, or discarding progressive pruning (i.e., maintaining the same number of tokens at each layer) leads to noticeable performance degradation across MMStar, MMBench, and SQA. By contrast, the complete VFlowOpt method consistently exhibits minimal performance loss, highlighting the effectiveness of all its components in pruning redundant tokens while preserving essential visual information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "We further examine how the number of samples and the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23930" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 523, + 143 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 523, + 143 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 523, + 143 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 523, + 143 + ], + "type": "table", + "html": "
MethodsToken ReductionFLOPs ↓ (T)ΔLatency ↓ (ms)ΔKV Cache ↓ (MB)ΔPerformance ↑Δ
LLaVA-OneVision-7B-71.4-1040.1-1786.4-1581-
+ VFlowOpt50%37.2-48.0%584.2-43.8%902.8-49.5%1591+0.6%
75%19.1-73.2%328.5-68.4%460.6-74.2%1570-0.7%
90%7.7-89.2%272.1-73.8%197.1-89.0%1464-7.4%
", + "image_path": "3af5b2956741a6de8945bd4b3d799f7aa67ca963a5c5c8b6c2c586d111305f97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 184, + 279, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 148, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 555, + 171 + ], + "type": "text", + "content": "Table 5. Efficiency analysis of LLaVA-OneVision-7B with VFlowOpt. The detailed metric includes computation (FLOPs), latency, and KV-Cache memory. " + }, + { + "bbox": [ + 55, + 148, + 555, + 171 + ], + "type": "inline_equation", + "content": "(\\Delta)" + }, + { + "bbox": [ + 55, + 148, + 555, + 171 + ], + "type": "text", + "content": " denotes the reduction ratio." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 184, + 279, + 248 + ], + "lines": [ + { + "bbox": [ + 72, + 184, + 279, + 248 + ], + "spans": [ + { + "bbox": [ + 72, + 184, + 279, + 248 + ], + "type": "table", + "html": "
Pruning StrategyMMStarMMBenchSQA
VFlowOpt57.879.992.3
w/o Importance Calibration56.279.491.8
w/o Token Merging57.679.891.9
w/o Progressive Pruning56.079.191.0
", + "image_path": "7df6f58322eeff0d18022495fcc373c29148d80151a63306e98a2dbebc3eff21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 253, + 280, + 266 + ], + "lines": [ + { + "bbox": [ + 70, + 253, + 280, + 266 + ], + "spans": [ + { + "bbox": [ + 70, + 253, + 280, + 266 + ], + "type": "text", + "content": "Table 6. Ablation studies of pruning strategy components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 272, + 291, + 358 + ], + "blocks": [ + { + "bbox": [ + 58, + 272, + 291, + 358 + ], + "lines": [ + { + "bbox": [ + 58, + 272, + 291, + 358 + ], + "spans": [ + { + "bbox": [ + 58, + 272, + 291, + 358 + ], + "type": "image", + "image_path": "2747e4b8f4cda23aeadfcdea5c5eba1512cdc543528df07152841897b6e5eb9c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "type": "text", + "content": "Figure 4. Relationships of sample size (left) and optimization steps (right) with optimization time and final performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 498 + ], + "type": "text", + "content": "number of optimization iterations in the Bayesian optimization procedure influence both optimization time and final performance. As illustrated in Fig. 4, the optimization time increases linearly with the number of samples and steps. Using 30 samples and 50 steps strikes an effective balance between time efficiency and performance, requiring only about 30 minutes to reach an exceptional result. More ablation results are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 504, + 236, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 236, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 236, + 517 + ], + "type": "text", + "content": "4.6. Visual Information Flow Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "content": "We further investigate how Bayesian optimization formulates the superior pruning strategy by examining its impact on visual information flow. Specifically, we compare the randomly initialized pruning strategy derived from the Bayesian optimization process with its final optimized strategy, measuring their respective effects on the text tokens that follow the visual tokens at each LLM layer. Because these text tokens can receive information from the visual tokens via the LLM's unidirectional attention mechanism, examining differences in these tokens with and without pruning effectively reveals how visual information flow is altered. As shown in Fig. 5, under the same computational budget, the initial strategy causes a pronounced discrepancy between text token representations with and without pruning, indicating that it significantly disrupts the visual information flow. In contrast, when using the optimized strategy," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 184, + 550, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 184, + 550, + 384 + ], + "lines": [ + { + "bbox": [ + 317, + 184, + 550, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 550, + 384 + ], + "type": "image", + "image_path": "d2a427acb1d1889fcb27ddbe93caecc35d46160cebc60341d0b621e7434e3262.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 555, + 420 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 420 + ], + "type": "text", + "content": "Figure 5. Visualization of the discrepancies in visual information flow corresponding to the pruning strategies before and after optimization, along with an example of visual question answering." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 513 + ], + "type": "text", + "content": "the text tokens remain much more similar to those observed without visual token pruning. This important finding suggests that the optimization process yields a pruning strategy that preserves crucial visual information by minimizing the difference in visual information flow with and without pruning, thereby retaining LMM performance to the greatest extent possible." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 524, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 388, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we address the pressing challenge of computational inefficiency in LMMs by introducing a general framework for visual token pruning. We reformulate pruning as an optimization problem focused on minimizing the divergence of visual information flow with and without pruning, thereby facilitating a tailored pruning strategy for diverse LMMs. By leveraging calibrated attention in ViTs to evaluate token significance and merging pruned tokens using importance-based weighting, we preserve critical information during inference without requiring retraining or manual tuning. Extensive experiments underscore the efficiency and generalizability of our approach, making it a practical solution for deploying LMMs in real-world settings, particularly in resource-constrained environments." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23931" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 294, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 294, + 118 + ], + "type": "text", + "content": "This research was supported by Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 151, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 151, + 294, + 205 + ], + "type": "text", + "content": "[1] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 294, + 262 + ], + "type": "text", + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 266, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 294, + 320 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 323, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 574 + ], + "type": "text", + "content": "[4] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, Shuaibin Li, Wei Li, Yinling Li, Hongwei Liu, Jiangning Liu, Jiawei Hong, Kaiwen Liu, Kuikun Liu, Xiaoran Liu, Chengqi Lv, Haijun Lv, Kai Lv, Li Ma, Runyuan Ma, Zerun Ma, Wenchang Ning, Linke Ouyang, Jiantao Qiu, Yuan Qu, Fukai Shang, Yunfan Shao, Demin Song, Zifan Song, Zhihao Sui, Peng Sun, Yu Sun, Huanze Tang, Bin Wang, Guoteng Wang, Jiaqi Wang, Jiayu Wang, Rui Wang, Yudong Wang, Ziyi Wang, Xingjian Wei, Qizhen Weng, Fan Wu, Yingtong Xiong, Chao Xu, Ruiliang Xu, Hang Yan, Yirong Yan, Xiaogui Yang, Haochen Ye, Huaiyuan Ying, Jia Yu, Jing Yu, Yuhang Zang, Chuyu Zhang, Li Zhang, Pan Zhang, Peng Zhang, Ruijie Zhang, Shuo Zhang, Songyang Zhang, Wenjian Zhang, Wenwei Zhang, Xingcheng Zhang, Xinyue Zhang, Hui Zhao, Qian Zhao, Xiaomeng Zhao, Fengzhe Zhou, Zaida Zhou, Jingming Zhuo, Yicheng Zou, Xipeng Qiu, Yu Qiao and Dahua Lin. Internl m2 technical report, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 577, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 577, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 577, + 294, + 621 + ], + "type": "text", + "content": "[5] Jieneng Chen, Luoxin Ye, Ju He, Zhao-Yang Wang, Daniel Khashabi, and Alan Yuille. Llavolta: Efficient multi-modal models via stage-wise visual context compression. arXiv preprint arXiv:2406.20092, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 677 + ], + "type": "text", + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "text", + "content": "[7] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 161 + ], + "type": "text", + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 162, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 553, + 227 + ], + "type": "text", + "content": "[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A survey on multimodal large language models for autonomous driving, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 283, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 337 + ], + "type": "text", + "content": "[11] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "text", + "content": "[12] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering, 2017. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 383, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 446 + ], + "type": "text", + "content": "[13] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 449, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 480 + ], + "type": "text", + "content": "[14] Lianyu Hu, Fanhua Shang, Liang Wan, and Wei Feng. illava: An image is worth fewer than 1/3 input tokens in large multimodal models. arXiv preprint arXiv:2412.06263, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 514 + ], + "type": "text", + "content": "[15] Omri Kaduri, Shai Bagon, and Tali Dekel. What's in the image? a deep-dive into the vision of vision language models. arXiv preprint arXiv:2411.17491, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 557 + ], + "type": "text", + "content": "[16] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 601 + ], + "type": "text", + "content": "[17] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 646 + ], + "type": "text", + "content": "[18] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[19] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[20] Yuelei Li, Ge Yan, Annabella Macaluso, Mazeyu Ji, Xueyan Zou, and Xiaolong Wang. Integrating lmm planners and 3d" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23932" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "skill policies for generalizable manipulation. arXiv preprint arXiv:2501.18733, 2025. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 152 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 295, + 186 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 188, + 295, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 295, + 242 + ], + "type": "text", + "content": "[24] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 295, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 299 + ], + "type": "text", + "content": "[25] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 300, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 295, + 365 + ], + "type": "text", + "content": "[26] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 411 + ], + "type": "text", + "content": "[27] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 412, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 296, + 456 + ], + "type": "text", + "content": "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 457, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 295, + 523 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 524, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 295, + 580 + ], + "type": "text", + "content": "[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "text", + "content": "[31] Haoxiang Tian, Xingshuo Han, Guoquan Wu, Yuan Zhou, Shuo Li, Jun Wei, Dan Ye, Wei Wang, and Tianwei Zhang. An llm-enhanced multi-objective evolutionary search for autonomous driving test scenario generation. arXiv preprint arXiv:2406.10857, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 714 + ], + "type": "text", + "content": "[32] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 335, + 72, + 553, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 72, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 335, + 72, + 553, + 225 + ], + "type": "text", + "content": "Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 227, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 227, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 553, + 270 + ], + "type": "text", + "content": "[33] Ao Wang, Fengyuan Sun, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. [cls] token tells everything needed for training-free efficient mllms. arXiv preprint arXiv:2412.05819, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "type": "text", + "content": "[34] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "text", + "content": "[35] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. Drivemlm: Aligning multimodal large language models with behavioral planning states for autonomous driving, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "type": "text", + "content": "[36] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "text", + "content": "[37] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 505, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 668 + ], + "type": "text", + "content": "[38] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 714 + ], + "type": "text", + "content": "[39] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin," + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23933" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 138 + ], + "type": "text", + "content": "Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[40] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[41] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "text", + "content": "[42] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 1, 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 314 + ], + "type": "text", + "content": "[43] Xubing Ye, Yukang Gan, Yixiao Ge, Xiao-Ping Zhang, and Yansong Tang. Atp-llava: Adaptive token pruning for large vision language models. arXiv preprint arXiv:2412.00447, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 403 + ], + "type": "text", + "content": "[44] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[45] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[46] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "text", + "content": "[47] Jun Zhang, Desen Meng, Ji Qi, Zhenpeng Huang, Tao Wu, and Limin Wang. p-mod: Building mixture-of-depths mllms via progressive ratio decay. arXiv preprint arXiv:2412.04449, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "type": "text", + "content": "[48] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[49] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[50] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. In The Thirteenth International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[51] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 196 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Hao Tang, and Jieping Ye. From redundancy to relevance: Information flow in lvlms across reasoning tasks. arXiv preprint arXiv:2406.06579, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[52] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[53] Zhi Zhang, Srishti Yadav, Fengze Han, and Ekaterina Shutova. Cross-modal information flow in multimodal large language models. arXiv preprint arXiv:2411.18620, 2024. 5" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23934" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_content_list.json b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc5a9af888024d0f39863826ce81579f2b136fb0 --- /dev/null +++ b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "image", + "img_path": "images/68614e85b835a66db0667fcd63c2b5cc77b7b39f4a49010171318974b8b9b39b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 156, + 126, + 192, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VGGSounder: Audio-Visual Evaluations for Foundation Models", + "text_level": 1, + "bbox": [ + 194, + 130, + 841, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniil Zverev1,* Thaddäus Wiedemer2,3,4,* Ameya Prabhu2,3", + "bbox": [ + 251, + 179, + 743, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Bethge $^{2,3}$ Wieland Brendel $^{3,4}$ A. Sophia Koepke $^{1,2,3}$", + "bbox": [ + 243, + 202, + 751, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Technical University of Munich, MCML $^{2}$ University of Tübingen $^{3}$ Tübingen AI Center $^{4}$ MPI for Intelligent Systems, ELLIS Institute Tübingen", + "bbox": [ + 129, + 229, + 869, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{daniil.zverev,a-sophia.koepke}@tum.de,{thaddaeus.wiedemer,ameya.prabhu}@uni-tuebingen.de", + "bbox": [ + 102, + 279, + 887, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 329, + 326, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of audio-visual foundation models underscores the importance of reliably assessing their multimodal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric. Our dataset and project page are available at https://vqgsounder.github.io/", + "bbox": [ + 89, + 362, + 483, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rigorous evaluation benchmarks have been instrumental in assessing the effectiveness of audio-visual models [33, 43, 49, 57]. Specifically, multi-modal foundation models integrating visual and auditory data aim to achieve a holistic understanding of audio-visual content. However, the field lacks large-scale modality-aware classification benchmarks with ground-truth annotations indicating whether each label is visible, audible, or both. Such annotations would allow detailed evaluations of multi-modal model capabilities. To address this gap, we introduce VGGSounder, an enhanced version of the widely-used audio-visual classification dataset VGGSound [13], which facilitates modality-", + "bbox": [ + 88, + 691, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b489317aa6ca2aff9ddd7b4f939ceccbb63d7e5eb5c60b2e953cee0770ab8f8.jpg", + "image_caption": [ + "Figure 1. We introduce VGGSounder, a multi-label audio-visual classification benchmark with modality annotations. We extend the original VGGSound test set with human-annotated audible, visible, and visible+audible labels. We add meta labels for common confounders, such as background music. We benchmark eleven recent audio-visual models on VGGSounder. It enables selective analysis of a model's auditory and visual capabilities on classes relevant for the queried modality." + ], + "image_footnote": [], + "bbox": [ + 517, + 329, + 903, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aware evaluation of audio-visual foundation models.", + "bbox": [ + 511, + 715, + 861, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VGGSound has several notable limitations. First, its data is inherently multi-label; for instance, a single sample may simultaneously include labels such as playing drum kit and playing acoustic guitar when multiple instruments are present. Additionally, evaluating how different modalities contribute to model performance becomes difficult without explicit modality annotations, as some labels are either not visually present or not audible (e.g., certain instruments might only be audible but not visible in advertisements). Moreover, overlapping label classes present another challenge; for example,", + "bbox": [ + 509, + 734, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution", + "bbox": [ + 109, + 887, + 215, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1027", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the orchestra label often coincides with labels for individual instruments. These issues result in systematic under-evaluation of multi-modal audio-visual foundation models.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we present VGGSounder, an improved benchmark inspired by similar advancements in other domains [12, 28]. We re-announce the dataset to create a comprehensive multi-label classification setting by collecting detailed annotations for each sample, including (1) additional classes present, (2) explicit modality annotations to label modality misalignment, (3) metadata indicating the presence of background music, voice-over, or static images, and (4) merging of classes to address overlapping classes. Consequently, VGGSounder provides a robust, foundation-model-ready benchmark enabling structured analysis of whether models rely on audio or visual cues. Furthermore, we include meta-labels (e.g., background music, voice-over, or static images) to easily filter out unreliable labels during evaluation. Utilising VGGSounder, we evaluate audio-visual foundation models, demonstrating their poor performance on our benchmark. We find that the state-of-the-art, closed-source Gemini models consistently rely exclusively on the visual modality. In addition to that, we measure the modality confusion, i.e. when models get distracted by an additional input modality, which exposes the unsuccessful merging of modalities. These findings highlight the importance of the audio-focused VGGSounder benchmark as a critical tool for accurately assessing audio-visual foundation models.", + "bbox": [ + 91, + 138, + 483, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We make the following contributions:", + "bbox": [ + 102, + 532, + 352, + 547 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We illustrate limitations of VGGSound in Sec. 3", + "2. We curate VGGSounder with multi-modal human annotations for multi-label classification in Sec. 4", + "3. We evaluate state-of-the-art audio-visual models, observing differences between embedding models and autoregressive foundation models in Sec. 5.", + "4. We propose new metrics to quantify the negative impact of using multiple input modalities in Sec. 5." + ], + "bbox": [ + 101, + 547, + 470, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 689, + 227, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-visual learning Many prior works consider audio-visual tasks that include sound source localisation and separation [3, 5, 9, 15, 27, 56, 62, 67, 75, 80, 85, 86, 90], event localisation [50, 51, 74, 78], audio-visual question answering [48, 54, 83, 84], audio-visual synchronisation [14, 23, 25, 38, 39, 42], audio synthesis using visual information [19, 26, 31, 44, 45, 61, 69-71, 87], or audio-driven face image synthesis [7, 40, 77]. Audio-visual data has also been leveraged for speech-related tasks, including speech and speaker recognition [2, 4, 59], or the spotting of spoken keywords [58, 66].", + "bbox": [ + 89, + 719, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, the natural alignment between audio and", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/3086e0733e283efca0637873a1cab8fd54187d2014223a4b780e6becda7f20a0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# Clips# ClassesMulti- labelModality labelsAnnotation pipeline
Flickr-SoundNet [10]2M-✘✘-
Kinetics-Sound [7]18.8k34✘✘MTurk
AudioSet [26]2.1M537✓✘Manual
AVE [62]4k28✓✘Manual
VEGSAS [76]132k10✘✘MTurk
Visually Aligned Sounds [15]13k8✘✘MTurk
VGGSound [12]200k309✘✘Classifiers+Manual
VGGSound-Sparse [32]7.1k12✘✘Manual
Visual Sound [66]91k309✘✘ImageBind [30]
VGGSounder15.4k309✓✓MTurk
", + "bbox": [ + 517, + 88, + 903, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of audio-visual classification benchmarks.", + "bbox": [ + 519, + 208, + 897, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "video has been exploited to learn improved audio-visual embeddings for downstream tasks [6, 10, 11, 18, 20, 21, 46, 60, 63, 65, 79]. Using both modalities jointly generally leads to performance boosts over using one modality in isolation. We examine this observation closely and aim to evaluate the effective use of multiple input modalities for the video classification task. To enable this, we propose — to the best of our knowledge, the first multi-label video classification benchmark that includes per-modality annotations for every sample (see Tab. 1).", + "bbox": [ + 511, + 241, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-visual foundation models Recently, multi-modal general-purpose models have emerged that can handle diverse downstream tasks without task-specific finetuning - also referred to as multi-modal foundation models. For instance, images or language were used as the bridge between modalities including audio, image, and text [30, 89]. Building on this, PandaGPT [72] leverages Vicuna [22] and ImageBind's embedding space to train a general multi-modal model exclusively on image-text pairs. Unified-IO 2 [53] employs universal tokenisation to process audio, video, and text. VideoLLaMA2 [21] uses a Spatial-Temporal Convolution connector in the visual branch before projecting audio and visual information into the LLM input space. The recently introduced Ola model [52] advances omnimodal processing through progressive modality alignment, using video to bridge audio and visual information. The Gemini models [73] are closed-source multi-modal models that achieve impressive performance on diverse downstream tasks. We use VGGSounder to benchmark the audio-visual capabilities of the aforementioned models.", + "bbox": [ + 511, + 397, + 906, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-visual classification benchmarks Audio-visual classification is distinct from general video classification (e.g. on YouTube-8M [1]), as classes typically cover both audible and visible actions or events. Commonly used datasets for audio-visual classification include KineticsSound [8] sourced from the Kinetics dataset [41], Flickr-SoundNet [11] scraped from Flickr, and AudioSet [29] and VGGSound [13], both sourced from YouTube.", + "bbox": [ + 511, + 703, + 903, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Kinetics-Sound features manual labels of human actions, but covers only 34 classes. Flickr-SoundNet is much larger, but only a small subset is labelled. Similarly, only a small fraction of the roughly 2M AudioSet samples are annotated and have aligned audio and video.", + "bbox": [ + 511, + 825, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1028", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 2. Limitations of VGGSound. We show video frames from videos in the VGGSound test set along with their annotated label (grey) to demonstrate various limitations. A. VGGSound samples are labelled with a single class, yet many videos contain multiple distinct classes. B. Additionally, many classes partially overlap or are ambiguous. C. Some samples are labelled with classes that are not present in one of the modalities (i.e., the labelled class is not visible or audible).", + "bbox": [ + 89, + 292, + 906, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, VGGSound ensures audio-visual correspondence for around 200 000 samples and was curated with an automatic pipeline involving class-list generation, and auditory and visual content verification. The visual verification step ensures that a class is represented in the centre frame.", + "bbox": [ + 88, + 369, + 480, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The VEGAS dataset [88] provides better quality assurances for a small subset of AudioSet with only 10 classes. Visually Aligned Sounds [16] subsamples VEGAS and AudioSet after human verification, and Visual Sound [76] subsamples VGGSound using a multi-modal embedding model, both aiming for high audio-visual correspondence. Similarly, VGGSound-Sparse [37] is a subset of VGGSound with a focus on temporally and spatially sparse synchronisation signals (e.g., short loud noises).", + "bbox": [ + 88, + 444, + 482, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, VGGSound strikes the best balance between size, generality, and annotations, making it a common benchmark for audio-visual classification. We update VGGSound to sustain its usability for the development of the next generation of multi-modal foundation models.", + "bbox": [ + 89, + 580, + 482, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Limitations of VGGSound", + "text_level": 1, + "bbox": [ + 89, + 667, + 336, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since we are interested in the VGGSound dataset for benchmarking audio-visual multi-modal models, our analysis focuses on the VGGSound test set, which consists of 15446 video clips, each 10s long and labelled with one of 309 classes. We qualitatively identify several limitations of the VGGSound annotations outlined below and in Fig. 2.", + "bbox": [ + 89, + 693, + 482, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Co-occurring classes While VGGSound's visual verification aimed to minimise multiple classes co-occurring in a clip, we find that most samples nevertheless clearly contain multiple classes, see Fig. 2A. In some cases, classes are temporally separated, e.g., showing male speech, man speaking and then cutting to", + "bbox": [ + 89, + 787, + 482, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "footage of firing cannon. Most often, classes co-occur at the same time, sometimes for the entire duration of the video clip. Overlapping classes are often related, such as different instruments in a band or orchestra, but can also be entirely unrelated, e.g., donkey, ass braying co-occurring with playing violin. As additional empirical evidence, we provide a co-occurrence matrix computed using CAV-MAE [33], a state-of-the-art audio-visual model, in Appendix D.", + "bbox": [ + 511, + 369, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overlapping classes The issue of co-occurring classes is exacerbated by many of the 309 automatically generated classes partially overlapping in their definition, as illustrated in Fig. 2B. We found two pairs of synonymous classes: timpani and tympani and dog barking and dog bow-wow. Additionally, some classes are strict subclasses of others, such as the gender-specific versions of cattle moing: cow lowing and bull bellowing; or the more specific variants of people eating: people eating noodle and people eating apple. Finally, several classes commonly appear together, such as playing snare drum which is often played as part of a drum kit, or semantically similar concepts: running electric fan and air conditioning noise, and sloshing water and splashing water.", + "bbox": [ + 511, + 513, + 908, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modality misalignment Despite VGGSound's auditory and visual content verification, we find that many of the annotated classes are not visible or not audible, as shown in Fig. 2C. A large fraction of videos contains background music, voice-over and narration, or other background sounds like bird chirping, tweeting or cricket chirping without a visible source. Similarly, some videos contain visible but inaudible cues for classes like sea waves. Static images and slide shows accompa", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 Although these issues most likely also apply to the training set.", + "bbox": [ + 107, + 886, + 447, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1029", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c9ba699608f529a550d998b95fefef92562d56bd4f6c77fefb12df0bb131ed6.jpg", + "image_caption": [ + "Figure 3. Overview of VGGSounder. A. Most samples contain more than one label. B. More than a quarter of labels are audible but not visible. In contrast, only a tiny fraction is visible but not audible. C. Speech and bird sounds are the most common classes; more details can be found in Appendix B. D. Forty percent of the samples contain some combination of background music, voice over, and static image(s), making the classification task significantly harder." + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 339, + 270 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/94fbb8f345cb4997d6a8c853c6e957bfcd248d0cf8614eb593fba6ff5bc8f081.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 123, + 514, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3c1a4923a33d3f9b62441bb900fbf905856b1e68189ceeca9616911ae16440e.jpg", + "image_caption": [ + "B Label modality distribution", + "C Most common labels" + ], + "image_footnote": [], + "bbox": [ + 531, + 104, + 707, + 169 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c63d15d87efbc62ba4c4c351ed68e88dbdc145a6c81095ccf49958c67c05d52b.jpg", + "image_caption": [ + "D Meta labels" + ], + "image_footnote": [], + "bbox": [ + 722, + 104, + 897, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nied by music or other sounds are other frequent sources of misaligned modalities. Finally, some classes are misaligned by definition: wind noise is only audible and not visible. Overall, $48.43\\%$ of the original VGGSound test samples have misaligned modalities. This finding challenges the widely held assumption that VGGSound has strong modality alignment [32, 47].", + "bbox": [ + 88, + 356, + 482, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Other datasets, such as Visually Aligned Sounds, Visual Sound, and VGGSound-Sparse (see Sec. 2) omitted samples with misaligned modalities. In contrast, we contend that inaudible or invisible cues are common in natural videos and should be considered when benchmarking multi-modal models. We, therefore, place particular emphasis on crafting reliable modality annotations for all samples, allowing users to evaluate models on samples that guarantee modality alignment, and on those where classes are only visible or only audible (see Tab. 1).", + "bbox": [ + 88, + 462, + 482, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Takeaway 1 VGGSound suffers from several issues: class co-occurrence not captured by single labels, overlapping class definitions, and modality misalignment, see Fig. 2.", + "bbox": [ + 99, + 628, + 473, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Building VGGSounder", + "text_level": 1, + "bbox": [ + 89, + 709, + 305, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a series of fixes for VGGSound's issues, ultimately resulting in the updated VGGSounder benchmark. We are not the first aiming to future-proof an existing benchmark: [12] analysed shortcomings of ImageNet [24], ultimately proposing switching to a multi-label classification task with additional manual labels. [28] similarly reannotated samples in MMLU [35, 36] to fix labelling errors. Both works inspire our approach to improve VGGSound. To deal with co-occurring classes, we switch to a multi-label classification setting. This effectively handles most overlapping class definitions: a strong model can assign a", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "high probability to multiple classes, even if they partially overlap. This also allows us to ensure that synonymous classes, as well as subclasses and their superclass, always appear together in the ground-truth labels.", + "bbox": [ + 511, + 356, + 903, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To deal with modality misalignment, we add a modality annotation to each label. For example, we can label a video as containing people clapping in the audio and containing playing volleyball in the audio and video data. We also add meta-labels to indicate whether a sample contains background music, voice over, or static image(s) to optionally treat these cases separately during evaluation.", + "bbox": [ + 511, + 417, + 905, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a pipeline similar to [12] to annotate multiple labels per sample, which we outline below.", + "bbox": [ + 511, + 541, + 903, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Collecting proposals We create a gold standard reference set by labelling a small, randomly selected subset of VGGSound test samples with four in-house computer vision experts. The interface used for this first labelling step is shown in Appendix A. We extend the subset until each class is covered at least once, leading to a final size of 417 samples. Labels from different annotators are merged via a simple majority vote.", + "bbox": [ + 511, + 575, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the gold standard set, we want to find a solid strategy for automatically generating label proposals which are shown to the humans labelling the test set. This should have a recall greater than $90\\%$ while maximising precision compared to the gold standard labels to produce a small set of proposals with good label coverage. Our final strategy combines predictions from several state-of-the-art models with a manual heuristic to obtain $93\\%$ recall for an average of 30 proposals per sample, see Appendix A.", + "bbox": [ + 511, + 699, + 905, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human labelling We use Amazon Mechanical Turk to re-associate the entire VGGSound test set. For each sample, we first ask annotators to indicate whether the video contains background music, voice over, or", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1030", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "static image(s). Then, annotators are asked to indicate for each label proposal whether the class is audible and/or visible. Finally, annotators can add a class if it is missing from the proposals. Annotators were paid the US minimum wage; the interface used is shown in Appendix A. We let annotators label the samples in batches of 20, each containing two gold standard samples as catch trials. We reject and re-associate all batches with a catch trial $F_{1}$ -score below $25\\%$ . In addition, we obtain modality annotations for the original VGGSound labels and meta-classes. Further details are provided in Appendix A.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Final labels We merge all obtained annotations using majority voting. Additionally, we automatically add synonymous classes and superclasses for a given subclass, e.g., we add cattle moing whenever cow lowing is in the set of labels. The full set of classes added this way is described in Appendix A.", + "bbox": [ + 89, + 260, + 480, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Takeaway 2 We develop VGGSounder: A multi-label video classification benchmark extending VGGSound with human-annotated multi-labels, modality annotations, and meta-labels as summarised in Fig. 3", + "bbox": [ + 99, + 366, + 473, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Benchmarking audio-visual models", + "text_level": 1, + "bbox": [ + 89, + 445, + 410, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use VGGSounder to benchmark four popular audiovisual embedding models and seven foundation models, and analyse their auditory and visual capabilities.", + "bbox": [ + 89, + 470, + 482, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models We evaluate the audio-visual embedding models CAV-MAE [33], DeepAVFusion [57], AV-Siam [49], and Equi-AV [43]. Those were finetuned on VGGSound.", + "bbox": [ + 89, + 518, + 482, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We benchmark several models from the closed-source Gemini family [73] in a zero-shot evaluation protocol. Furthermore, we use LLM-assisted evaluation to evaluate the following four open-source autoregressive foundation models: VideoLLaMA-2 [21], Unified-IO-2 [53], Panda-GPT [72], and Ola [52]. All models are evaluated in three modes: using unimodal-audio, unimodal-visual, or multi-modal (audio and visual) inputs. Further details about models and their evaluation are provided in Appendix C.1", + "bbox": [ + 89, + 564, + 482, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics To benchmark the models on VGGSounder, we use multi-label classification metrics. For embedding models, all metrics are computed for the top- $k$ predictions, with $k \\in \\{1, 3, 5, 10\\}$ . In contrast, prompting foundation models yields an unordered set of class predictions of varying size, and we compute only a single metric using the entire set. As a result, metrics are not directly comparable between embedding and foundation models. To get a sense of their relative performance, we report metrics for embedding models for $k = 1$ in the main text, matching the median number of predictions per sample for the foundation models.", + "bbox": [ + 89, + 704, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For open-source models such as VideoLLaMA-2, Ola, Unified-I0-2, and Panda-GPT, we employ LLM-assisted", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "evaluation [55, 81], in which the Qwen-3 model [82] is tasked to assess the correspondence between model outputs and target classes. Closed-source models from the Gemini family are evaluated by providing the full list of 309 classes as input. Further details on the evaluation procedures and exact prompts are provided in Appendix C", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subset accuracy compares the predicted label set to the ground-truth label set and reports the fraction of samples for which they match. This is our strictest metric.", + "bbox": [ + 511, + 181, + 903, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$F_{1}$ -score is the harmonic mean of precision and recall. It is strictly larger than the subset accuracy.", + "bbox": [ + 513, + 227, + 903, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hit reports the fraction of samples for which any of the predicted labels are part of the ground-truth label set. This is the most lenient metric which is strictly larger than the $F_{1}$ score. We include this metric for ease of comparison to the \"ReaL-Accuracy\" used in [12].", + "bbox": [ + 511, + 257, + 903, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All metrics are computed separately for each input modality (audio, video, and audio-visual) and label modality. We use lowercase symbols a, v, and av to indicate the input modality: audio-only, visual-only, or audio-visual inputs, respectively. For label modality, we use uppercase symbols $A$ , $V$ , and $AV$ , referring to the subsets of the benchmark with audible, visible, and audio-visual labels. We further include $A \\neg V$ (audible but not visible) and $V \\neg A$ (visible but not audible) to analyze unaligned cues. For clarity, we define shorthand notations such as $a = a(A)$ to denote the model's performance on the audible subset $A$ using only audio input. Analogously, $v = v(V)$ and $av = av(AV)$ refer to video-only and audio-visual performance on their respective label subsets. Furthermore, we use micro-aggregation to balance the contribution from each class.", + "bbox": [ + 511, + 333, + 905, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We additionally measure the negative impact of using multimodal inputs. In particular, $\\mu$ is a new metric we propose to measure a model's modality confusion $(\\mu)$ . We define it as", + "bbox": [ + 511, + 559, + 903, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\mathrm {M}} = 1 0 0 \\cdot \\frac {\\sum_ {x \\in M} \\mathcal {I} [ m (x) - \\text {c o r r e c t} \\cap a v (x) - \\text {w r o n g} ]}{N _ {\\text {t o t a l}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 611, + 903, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M \\in [A, V, A \\cap V]$ and their associated modality inputs are $m \\in [a, v]$ , correct/wrong is determined as in the Hit score (with $k = 1$ for embedding models). $N_{total}$ refers to the total number of samples. $\\mu$ measures the fraction of samples a model correctly classified given an input modality but got wrong when using both modalities simultaneously. We additionally report $\\mu_{A \\cap V}$ as the percentage of samples a model could solve in either modality unimodally but could not solve multi-modally. In other words, the modality confusion $\\mu$ captures how frequently a model is distracted by an additional input modality, which can indicate the unsuccessful merging of modalities.", + "bbox": [ + 511, + 651, + 906, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Takeaway 3 We propose a new metric, modality confusion $\\mu$ , that measures how frequently a model is distracted by an additional input modality; see Eq. (1).", + "bbox": [ + 521, + 847, + 897, + 895 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1031", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7f3c2bebd34a8528a23fab9c42cd1cf3cda07f540249bab807c8564bc3c9e197.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSubset Accuracy ↑F1↑Hit ↑μ↓
avavavava(A-√V)v(V-√A)avavμAμVμA∩V
Embedding Models
CAV-MAE13.1919.2324.4934.4634.9142.6213.9419.0062.2953.4464.173.586.430.77
DeepAVFusion10.1911.1021.5325.3121.2937.3510.3710.5545.7732.6156.273.743.930.17
Equi-AV11.6010.5220.0029.3920.4234.6912.5510.6553.1231.2652.246.977.131.38
AV-Siam12.7919.7522.8333.3035.4139.4312.9018.2160.1954.2059.369.368.803.58
Closed-source Foundation Models
Gemini 1.5 Flash1.7814.4416.4414.4936.9842.5215.6121.6132.7347.3659.1010.224.250.77
Gemini 1.5 Pro3.0520.8622.5319.2649.7353.7417.7322.9035.0369.2375.422.094.850.57
Gemini 2.0 Flash1.8512.5412.6911.8034.0836.456.1918.9018.5143.8347.722.395.431.00
Open-source Foundation Models
VideoLLaMA 212.8619.8524.4738.8747.8252.3520.3428.0858.9152.0259.8012.725.462.95
Unified-IO 211.9411.5625.6135.3127.9248.8921.3816.5354.3931.0565.118.705.161.79
PandaGPT3.194.195.4618.7318.5620.8516.8214.4021.0817.0118.827.595.902.47
OLA14.118.6918.1947.7024.8546.4840.4413.4559.0524.5751.5115.476.802.49
", + "bbox": [ + 94, + 88, + 903, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Audio-visual video classification results on VGGSounder. We report multi-label classification metrics (subset accuracy, $F_{1}$ -score, Hit accuracy, modality confusion $(\\mu)$ for audio- $a(A)$ , visual - $v(V)$ , audio-visual - $av(AV)$ , audio-only - $a(A\\neg V)$ and video-only - $v(V\\neg A)$ inputs. The embedding models CAV-MAE, DeepAVFusion, and Equi-AV were finetuned on the VGGSound training set. We report metrics for $k = 1$ here and for other $k$ in Appendix D The closed sourced multi-modal foundation models Gemini and open-sourced models use a zero-shot evaluation protocol and LLM-assisted protocol respectively.", + "bbox": [ + 89, + 338, + 906, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Re-evaluating the state of the art", + "text_level": 1, + "bbox": [ + 89, + 424, + 377, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the benchmark performance of state-of-the-art audio-visual models in Tab. 2.", + "bbox": [ + 89, + 446, + 482, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall performance Unsurprisingly, all models perform best with access to both input modalities (AV). Across all metrics, both open- and closed-source general-purpose models perform comparably to the purpose-built embedding models CAV-MAE and AV-Siam. This indicates that foundation models have reached — and for some modalities exceeded — the performance of specialised models. However, the embedding models finetuned on VGGSound generally have stronger unimodal performance with audio inputs $(A)$ compared to visual inputs $(V)$ , which is in line with their pretraining on AudioSet. Interestingly, this trend is reversed for most foundation models, which seem to be biased towards visual inputs, with unimodal video performance $(V)$ being substantially higher than unimodal audio performance $(A)$ .", + "bbox": [ + 88, + 479, + 482, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Takeaway 4 Foundation models perform comparably to finetuned embedding models. Embedding models more heavily rely on audio cues than on visual ones, while foundation models exploit visible cues rather than audible ones, see Tab. 2", + "bbox": [ + 99, + 724, + 473, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Modality confusion The modality confusion score $\\mu$ shows that, for all models, a notable fraction of test samples $(4 - 11\\%)$ were misclassified when an additional modality was included—despite being correctly classified with unimodal input. Furthermore, for all models, a small portion of", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "test samples is not solvable multi-modally even though they were solvable in both modalities alone $(\\mu_{A\\cap V})$ . This insight is made possible by VGGSounder's per-label modality annotations and shows that all models are susceptible to being distracted given an additional modality. This is a concerning issue for multi-modal models since they should preserve unimodal capabilities when adding a second modality. Being able to evaluate this behaviour on the VGGSounder benchmark is a first step towards enabling the development of mitigation strategies, eventually resulting in stronger audio-visual models.", + "bbox": [ + 511, + 425, + 906, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Takeaway 5 Our modality confusion score reveals that all models are negatively impacted by additional modalities for a substantial amount of samples (see Tab. ②).", + "bbox": [ + 521, + 606, + 898, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance across modalities Fig. 4 shows the performance profiles across modalities. At first glance, we can see that VideoLLaMA-2's performance is well balanced for different input modalities, while models from the Gemini family distinctly underperform on audio inputs. In contrast, embedding models exhibit a moderate balance across modalities, with DeepAVFusion and EquiAV showing slight underperformance for visual input.", + "bbox": [ + 511, + 672, + 906, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As Fig. 4 also illustrates, profiling of this kind is enabled through the modality annotations in VGGSounder. In contrast, VGGSound assumed that all classes are perceivable in both modalities, and did not account for background sounds. This resulted in consistent under-evaluation of foundation models (that were not finetuned on VGGSound) for audio inputs.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1032", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the radar plot in Fig. 4, we provide results on VGGSound in Appendix D showing that all models have substantially lower performance than their hit scores in Tab. 2. This confirms that many model predictions were incorrectly flagged as false positives in VGGSound due to the incomplete ground-truth labels, painting a distorted picture of models' limitations.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc5cecfd4791f27111ac0dae1e53c38773b7a815acb789f7c7016fb2e18df776.jpg", + "image_caption": [ + "Figure 4. VGGSounder more accurately captures model performance across input modalities. We show the Hit score on VGGSounder and accuracy on VGGSound, normalised by the per-model maximum performance on each benchmark. Specifically for foundation models, we observe a significant difference in performance between VGGSound and VGGSounder." + ], + "image_footnote": [], + "bbox": [ + 106, + 210, + 473, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Takeaway 6 VGGSounder's more complete ground-truth labels allow for more accurate, modality-specific profiling of model performance (see Fig.4).", + "bbox": [ + 99, + 584, + 475, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Performance analysis using meta-classes", + "text_level": 1, + "bbox": [ + 89, + 651, + 436, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VGGSounder includes annotations of three meta-classes for each sample: background music indicates whether samples contain music without a visible source, voice over similarly marks speech without a visible source, and static image(s) flags that the visual stream consists of one or only a few static visual frames.", + "bbox": [ + 89, + 672, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These new meta-classes allow us to evaluate the model behaviour in challenging scenarios where information from one modality dominates. We consider the performance difference between samples that do contain a meta-label and samples that do not contain it. Positive numbers indicate that the models perform better on the subset with metal-labels. In Tab.3, we summarise the main findings, focussing on $F_{1}$ -score as the most balanced multi-label metric. Additional results are provided in Appendix D", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background music All models perform worse on video samples containing background music. This indicates that it is challenging to decouple background audio from the rest of the video. The evaluated models are not good at differentiating between sound sources without visual cues, e.g. predicting different instruments in the background music.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Voice over In contrast to background music, we observe a clear difference between embedding models and foundation models for samples with voice-over. While the audio classification performance of embedding models drops substantially. This drop is only slight for VideoLLaMA-2 and Unified-IO 2, and the performance of other foundation models even improves. This indicates that the foundation models are less distracted by voice-over.", + "bbox": [ + 511, + 185, + 906, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Static image(s) The impact of static images is more nuanced: First, audio classification performance improves for embedding models while it decreases for the foundation models. This shows that the purpose-built, VGGSoundfinetuned embedding models can more accurately predict specific sounds in the absence of other cues.", + "bbox": [ + 511, + 309, + 903, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, visual classification performance on static images drops for all models, suggesting that models rely on rich temporal cues to make accurate predictions.", + "bbox": [ + 511, + 400, + 903, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Third, the subset accuracy (i.e. exact label set matches) for uni-modal predictions shows that foundation models perform better in visual than audio classification with and without static images. In contrast, embedding models perform better in audio classification on static images but worse in visual classification for other samples. This suggests that non-static samples form a challenging subset for embedding models, where a model needs to favour one modality over another one to make a correct prediction.", + "bbox": [ + 511, + 446, + 905, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Samples without any meta-label When comparing the model performance on samples without background music, static images, or voice-over annotations (column neither in Tab. 3) to the performance on the entire test set, we see a performance gain (here negative numbers indicate an increased $\\mathbf{F}_1$ score). This finding concludes that these three categories form challenging subsets of the dataset.", + "bbox": [ + 511, + 585, + 905, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Takeaway 7 Samples with background music, static images(s), and voice over provide distinct challenges for each model (see Tab.3). This highlights VGGSounder's value for comprehensive model evaluation.", + "bbox": [ + 521, + 707, + 898, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Impact of VGGSounder labels", + "text_level": 1, + "bbox": [ + 511, + 787, + 781, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our relabelling pipeline adds two types of labels to those in VGGSound: (1) automatically generated labels based on synonymous classes and sub-/superclasses, and (2) human-curated labels. In Tab.4, we ablate the impact of each type of added labels in terms of the relative performance gains (Hit score). A complete breakdown of the effects across all", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1033", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e96621888bbfcfb87649dd88334bc9a38904778b7eead2dce93dcf0fa39dfbb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelbackground musicvoice over ΔF1static image(s)neither ΔF1
ΔF1ΔF1ΔF1Sub. Acc. w/Sub. Acc. w/oavav
avavaavavavavav
Embedding Models
CAV-MAE-0.66-0.59-0.65-1.130.31-0.2812.7319.2211.9819.21-1.00-1.27-1.24
DeepAVFusion-0.71-0.80-0.70-1.250.28-0.189.7910.819.3010.81-1.27-1.01-1.36
Equi-AV-0.79-0.42-0.41-0.970.28-0.2311.1210.4510.4910.45-0.97-0.97-0.89
AV-Siam-0.73-0.72-0.82-1.080.34-0.2612.3419.5311.5719.52-1.01-1.34-1.39
Closed-source Foundation Models
Gemini 1.5 Flash-0.22-0.41-0.692.08-0.36-0.251.6614.351.6814.391.40-1.10-1.62
Gemini 1.5 Pro-0.36-0.61-0.942.48-0.31-0.242.8320.852.8720.801.64-0.89-1.09
Gemini 2.0 Flash-0.09-0.32-0.580.030.13-0.261.7012.331.5312.40-0.03-1.07-1.22
Open-source Foundation Models
VideoLLaMA 2-0.47-0.72-0.84-0.540.27-0.3112.5519.6412.0419.70-0.47-1.26-1.33
Unified-IO 2-1.200.18-0.65-0.680.12-0.2111.3911.8910.8812.00-1.07-0.15-1.17
PandaGPT-1.13-0.15-0.420.56-0.22-0.162.944.272.914.24-0.60-0.15-0.65
OLA-2.160.09-0.441.50-0.48-0.1813.038.8812.888.89-0.80-0.13-0.06
", + "bbox": [ + 94, + 88, + 906, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61fa65244ee5c430de75afa7adb150130023403062ea9ab4ae8143a8f510ad6b.jpg", + "table_caption": [ + "Table 3. Summary of performance differences in the presence/absence of meta-classes. Difference in $\\mathbf{F}_1$ scores ( $\\Delta \\mathbf{F}_1$ ) for audio-visual video classification on VGGSounder between videos with a meta-class and those without it. Positive numbers ( $\\Delta$ ) indicate better performance when the meta-class is present. Additional results are provided in Appendix D" + ], + "table_footnote": [], + "table_body": "
ModelHuman labels ↑Auto labels ↑
AVAVAVAV
Gemini 1.5 Flash29.2814.5916.360.480.931.51
Gemini 1.5 Pro28.6125.5227.630.311.992.10
Gemini 2.0 Flash8.8012.1611.130.221.121.28
", + "bbox": [ + 93, + 372, + 480, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Impact of added labels using different strategies in VGGSounder. We show the change in multi-label classification performance ( $\\Delta$ Hit) when adding automatically added (Auto) or human-annotated (Human) labels to VGGSound, and compare to the original VGGSound data.", + "bbox": [ + 89, + 454, + 482, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models and metrics is provided in Appendix D. While performance is consistently higher with automatically added labels (Auto), the increase is noticeably smaller than that for human-curated labels. Paired with the observation that models do frequently predict correct classes that were not part of the original VGGSound label set, this indicates that human-curated labels better cover the ground truth.", + "bbox": [ + 89, + 542, + 483, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Takeaway 8 Automatically added labels are an important step, but human-curated labels have a bigger effect on eliminating incorrectly flagged false positives, underscoring the value of accurate human annotation.", + "bbox": [ + 99, + 667, + 473, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 750, + 205, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of VGGSound as base dataset VGGSound is commonly used to evaluate audio-visual models on the multimodal video classification task. As it is currently the most suitable testbed for audio-visual classification tasks (due to its size, diversity of categories, non-constrained setting, and relatively strong audio-visual correspondence), it serves as an optimal starting point for our substantially improved VGGSounder benchmark with a multi-label evaluation pro", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tocol for foundation models that makes the benchmark suitable for meaningful evaluation.", + "bbox": [ + 511, + 375, + 903, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-label vs single-label classification Video content is inherently complex, often containing multiple co-occurring objects and actions both within and across modalities. This makes it unlikely that a given clip belongs to just one class as is the case in the single-label classification task. Therefore, our VGGSounder extends the VGGSound test set to multi-label classification. Unlike models trained on a narrow single-label dataset, foundation models develop versatile representations. Multi-label evaluation thus also aligns very naturally with this capability.", + "bbox": [ + 511, + 409, + 906, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 573, + 633, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced an modality-aware evaluation set for audiovisual foundation models. VGGSounder builds on the widely used VGGSound dataset by adding: (a) comprehensive human annotations for missing classes, (b) specifying modality information per label, (c) introducing specialised meta-labels for frequently occurring real-world challenges, and (d) using heuristic methods to improve label quality.", + "bbox": [ + 511, + 598, + 906, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through our newly introduced metric, modality confusion, we observe that incorporating additional modalities does not necessarily yield better results. Models often become more confused on a substantial subset of test samples. Furthermore, finetuned embedding models tend to rely heavily on audio cues, while foundation models depend more on visual information. Additionally, our meta-label analysis highlights distinct challenges across various specialised yet commonly occurring scenarios such as background music, static images, and voice-overs. Overall, we hope that the VGGSounder benchmark will advance the evaluation and development of foundational audio-visual models.", + "bbox": [ + 511, + 705, + 908, + 885 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1034", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors would like to thank Felix Förster, Sayak Mallick, and Prasanna Mayilvahananan for their help with data annotation, Felix Förster and Monica Riedler for proofreading the paper, and Thomas Klein and Shyamgopal Karthik for their help in setting up MTurk. They also thank numerous MTurk workers for labelling.", + "bbox": [ + 89, + 114, + 480, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was in part supported by the BMBF (FKZ: 01IS24060, 01I524085B, 01IS18039A), the DFG (SFB 1233, project number: 276693517), and the Open Philanthropy Foundation funded by the Good Ventures Foundation. WB acknowledges financial support via an Emmy Noether Grant funded by the German Research Foundation (DFG) under grant no. BR 6382/1-1. WB is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This research utilised compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG. The authors thank the IMPRS-IS for supporting TW.", + "bbox": [ + 89, + 205, + 482, + 387 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 402, + 187, + 417 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016.", + "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE TPAMI, 2018.", + "[3] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In ECCV, 2020.", + "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP, 2020.", + "[5] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In ECCV, 2020.", + "[6] Humam Alwassel, Dhruv Mahajan, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020.", + "[7] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In CVPR, 2024.", + "[8] Relja Arandjelović and Andrew Zisserman. Look, listen and learn. In ICCV, 2017.", + "[9] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV, 2018.", + "[10] Yuki M. Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020.", + "[11] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. In NeurIPS, 2016." + ], + "bbox": [ + 94, + 426, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020.", + "[13] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020.", + "[14] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Audio-visual synchronisation in the wild. In BMVC, 2021.", + "[15] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In CVPR, 2021.", + "[16] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 2020.", + "[17] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022.", + "[18] Yanbei Chen, Yongqin Xian, A. Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021.", + "[19] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. In CVPR, 2025.", + "[20] Ying Cheng, Ruize Wang, Zhihao Pan, Rui Feng, and Yuejie Zhang. Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In ACM MM, 2020.", + "[21] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024.", + "[22] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. 2023.", + "[23] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016.", + "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.", + "[25] Joshua P Ebeneze, Yongjun Wu, Hai Wei, Sriram Sethuraman, and Zongyi Liu. Detection of audio-video synchronization errors via event detection. In ICASSP, 2021.", + "[26] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In ECCV, 2020.", + "[27] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In ICCV, 2019.", + "[28] Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1035", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Reza Ghasemi Madani, Claire Barale, Robert McHardy, Joshua Harris, Jean Kaddour, Emile van Krieken, and Pasquale Minervini. Are we done with mmlu? arXiv preprint arXiv:2406.04127, 2024.", + "[29] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017.", + "[30] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023.", + "[31] Shir Goldstein and Yael Moses. Guitar music transcription from silent video. In BMVC, 2018.", + "[32] Yuan Gong, Alexander H Liu, Andrew Rouditchenko, and James Glass. Uavm: Towards unifying audio and visual models. IEEE Signal Processing Letters, 2022.", + "[33] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James Glass. Contrastive audio-visual masked autoencoder. In ICLR, 2023.", + "[34] Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594, 2024.", + "[35] Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. Aligning ai with shared human values. In ICLR, 2021.", + "[36] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021.", + "[37] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022.", + "[38] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022.", + "[39] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, 2024.", + "[40] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. IJCV, 2019.", + "[41] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.", + "[42] Naji Khosravan, Shervin Ardeshir, and Rohit Puri. On attention modules for audio-visual synchronization. In CVPR Workshop, 2019.", + "[43] Jongsuk Kim, Hyeongkeun Lee, Kyeongha Rho, Junmo Kim, and Joon Son Chung. Equipav: Leveraging equivariance for audio-visual contrastive learning. In ICML, 2024.", + "[44] A. Sophia Koepke, Olivia Wiles, and Andrew Zisserman. Visual pitch estimation. In SMC, 2019." + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] A Sophia Koepke, Olivia Wiles, Yael Moses, and Andrew Zisserman. Sight to sound: An end-to-end approach for visual piano transcription. In ICASSP, 2020.", + "[46] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurlPS, 2018.", + "[47] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song. Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning. In ICCV, 2021.", + "[48] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022.", + "[49] Yan-Bo Lin and Gedas Bertasius. Siamese vision transformers are scalable audio-visual learners. In ECCV, 2024.", + "[50] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In ACCV, 2020.", + "[51] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP, 2019.", + "[52] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025.", + "[53] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In CVPR, 2024.", + "[54] Jie Ma, Min Hu, Pinghui Wang, Wangchun Sun, Lingyun Song, Hongbin Pei, Jun Liu, and Youtian Du. Look, listen, and answer: Overcoming biases for audio-visual question answering. NeurIPS, 2024.", + "[55] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023.", + "[56] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In ECCV, 2022.", + "[57] Shentong Mo and Pedro Morgado. Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. In CVPR, 2024.", + "[58] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. In BMVC, 2020.", + "[59] Arsha Nagrani, Joon Son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision. In ICASSP, 2020.", + "[60] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In NeurIPS, 2021.", + "[61] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. arXiv preprint arXiv:2104.02687, 2021." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1036", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018.", + "[63] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In ECCV, 2016.", + "[64] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Learning sight from sound: Ambient sound provides supervision for visual learning. IJCV, 2018.", + "[65] Mandela Patrick, Yuki M Asano, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In NeurIPS, 2020.", + "[66] KR Prajwal, Liliane Momeni, Triantafyllos Afouras, and Andrew Zisserman. Visual keyword spotting with attention. In BMVC, 2021.", + "[67] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In ECCV, 2020.", + "[68] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICLR, 2023.", + "[69] Kun Su, Xiulong Liu, and Eli Shlizerman. Multi-instrumentalist net: Unsupervised generation of music from body movements. arXiv preprint arXiv:2012.03478, 2020.", + "[70] Kun Su, Xiulong Liu, and Eli Shlizerman. How does it sound? In NeurIPS, 2021.", + "[71] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In ICML, 2024.", + "[72] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023.", + "[73] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023.", + "[74] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018.", + "[75] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel PW Ellis, and John R Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In ICLR, 2021.", + "[76] Ilpo Viertola, Vladimir Iashin, and Esa Rahtu. Temporally aligned audio for video with autoregression. In ICASSP, 2025.", + "[77] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018.", + "[78] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In CVPR, 2019.", + "[79] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020." + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Haoming Xu, Runhao Zeng, Qingyao Wu, Mingkui Tan, and Chuang Gan. Cross-modal relation-aware networks for audio-visual event localization. In ACM MM, 2020.", + "[81] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In ECCV, 2024.", + "[82] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025.", + "[83] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audio-visual question answering on videos. In ACM MM, 2022.", + "[84] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021.", + "[85] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In ECCV, 2018.", + "[86] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In ICCV, 2019.", + "[87] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019.", + "[88] Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L. Berg. Visual to sound: Generating natural sound for videos in the wild. In CVPR, 2018.", + "[89] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023.", + "[90] Lingyu Zhu and Esa Rahtu. V-slowfast network for efficient visual sound separation. In WACV, 2022." + ], + "bbox": [ + 516, + 92, + 903, + 561 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1037", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_model.json b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ff0356d4c1e965fa0574e69eefa728ebf7437813 --- /dev/null +++ b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_model.json @@ -0,0 +1,2598 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.157, + 0.127, + 0.194, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.843, + 0.151 + ], + "angle": 0, + "content": "VGGSounder: Audio-Visual Evaluations for Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.18, + 0.744, + 0.2 + ], + "angle": 0, + "content": "Daniil Zverev1,* Thaddäus Wiedemer2,3,4,* Ameya Prabhu2,3" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.203, + 0.753, + 0.223 + ], + "angle": 0, + "content": "Matthias Bethge\\(^{2,3}\\) Wieland Brendel\\(^{3,4}\\) A. Sophia Koepke\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.231, + 0.87, + 0.274 + ], + "angle": 0, + "content": "\\(^{1}\\)Technical University of Munich, MCML \\(^{2}\\)University of Tübingen \\(^{3}\\)Tübingen AI Center \\(^{4}\\)MPI for Intelligent Systems, ELLIS Institute Tübingen" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.28, + 0.888, + 0.296 + ], + "angle": 0, + "content": "{daniil.zverev,a-sophia.koepke}@tum.de,{thaddaeus.wiedemer,ameya.prabhu}@uni-tuebingen.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.33, + 0.327, + 0.346 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.485, + 0.635 + ], + "angle": 0, + "content": "The emergence of audio-visual foundation models underscores the importance of reliably assessing their multimodal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric. Our dataset and project page are available at https://vqgsounder.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.222, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.693, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Rigorous evaluation benchmarks have been instrumental in assessing the effectiveness of audio-visual models [33, 43, 49, 57]. Specifically, multi-modal foundation models integrating visual and auditory data aim to achieve a holistic understanding of audio-visual content. However, the field lacks large-scale modality-aware classification benchmarks with ground-truth annotations indicating whether each label is visible, audible, or both. Such annotations would allow detailed evaluations of multi-modal model capabilities. To address this gap, we introduce VGGSounder, an enhanced version of the widely-used audio-visual classification dataset VGGSound [13], which facilitates modality-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.33, + 0.905, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.696 + ], + "angle": 0, + "content": "Figure 1. We introduce VGGSounder, a multi-label audio-visual classification benchmark with modality annotations. We extend the original VGGSound test set with human-annotated audible, visible, and visible+audible labels. We add meta labels for common confounders, such as background music. We benchmark eleven recent audio-visual models on VGGSounder. It enables selective analysis of a model's auditory and visual capabilities on classes relevant for the queried modality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.862, + 0.732 + ], + "angle": 0, + "content": "aware evaluation of audio-visual foundation models." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "VGGSound has several notable limitations. First, its data is inherently multi-label; for instance, a single sample may simultaneously include labels such as playing drum kit and playing acoustic guitar when multiple instruments are present. Additionally, evaluating how different modalities contribute to model performance becomes difficult without explicit modality annotations, as some labels are either not visually present or not audible (e.g., certain instruments might only be audible but not visible in advertisements). Moreover, overlapping label classes present another challenge; for example," + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.216, + 0.9 + ], + "angle": 0, + "content": "*equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1027" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "the orchestra label often coincides with labels for individual instruments. These issues result in systematic under-evaluation of multi-modal audio-visual foundation models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.139, + 0.484, + 0.515 + ], + "angle": 0, + "content": "To overcome these limitations, we present VGGSounder, an improved benchmark inspired by similar advancements in other domains [12, 28]. We re-announce the dataset to create a comprehensive multi-label classification setting by collecting detailed annotations for each sample, including (1) additional classes present, (2) explicit modality annotations to label modality misalignment, (3) metadata indicating the presence of background music, voice-over, or static images, and (4) merging of classes to address overlapping classes. Consequently, VGGSounder provides a robust, foundation-model-ready benchmark enabling structured analysis of whether models rely on audio or visual cues. Furthermore, we include meta-labels (e.g., background music, voice-over, or static images) to easily filter out unreliable labels during evaluation. Utilising VGGSounder, we evaluate audio-visual foundation models, demonstrating their poor performance on our benchmark. We find that the state-of-the-art, closed-source Gemini models consistently rely exclusively on the visual modality. In addition to that, we measure the modality confusion, i.e. when models get distracted by an additional input modality, which exposes the unsuccessful merging of modalities. These findings highlight the importance of the audio-focused VGGSounder benchmark as a critical tool for accurately assessing audio-visual foundation models." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.533, + 0.353, + 0.548 + ], + "angle": 0, + "content": "We make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.548, + 0.446, + 0.563 + ], + "angle": 0, + "content": "1. We illustrate limitations of VGGSound in Sec. 3" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.564, + 0.472, + 0.593 + ], + "angle": 0, + "content": "2. We curate VGGSounder with multi-modal human annotations for multi-label classification in Sec. 4" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.594, + 0.472, + 0.639 + ], + "angle": 0, + "content": "3. We evaluate state-of-the-art audio-visual models, observing differences between embedding models and autoregressive foundation models in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.639, + 0.472, + 0.671 + ], + "angle": 0, + "content": "4. We propose new metrics to quantify the negative impact of using multiple input modalities in Sec. 5." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.548, + 0.472, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.69, + 0.228, + 0.706 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Audio-visual learning Many prior works consider audio-visual tasks that include sound source localisation and separation [3, 5, 9, 15, 27, 56, 62, 67, 75, 80, 85, 86, 90], event localisation [50, 51, 74, 78], audio-visual question answering [48, 54, 83, 84], audio-visual synchronisation [14, 23, 25, 38, 39, 42], audio synthesis using visual information [19, 26, 31, 44, 45, 61, 69-71, 87], or audio-driven face image synthesis [7, 40, 77]. Audio-visual data has also been leveraged for speech-related tasks, including speech and speaker recognition [2, 4, 59], or the spotting of spoken keywords [58, 66]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Furthermore, the natural alignment between audio and" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.199 + ], + "angle": 0, + "content": "
Dataset# Clips# ClassesMulti- labelModality labelsAnnotation pipeline
Flickr-SoundNet [10]2M-✘✘-
Kinetics-Sound [7]18.8k34✘✘MTurk
AudioSet [26]2.1M537✓✘Manual
AVE [62]4k28✓✘Manual
VEGSAS [76]132k10✘✘MTurk
Visually Aligned Sounds [15]13k8✘✘MTurk
VGGSound [12]200k309✘✘Classifiers+Manual
VGGSound-Sparse [32]7.1k12✘✘Manual
Visual Sound [66]91k309✘✘ImageBind [30]
VGGSounder15.4k309✓✓MTurk
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.209, + 0.898, + 0.222 + ], + "angle": 0, + "content": "Table 1. Comparison of audio-visual classification benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.905, + 0.394 + ], + "angle": 0, + "content": "video has been exploited to learn improved audio-visual embeddings for downstream tasks [6, 10, 11, 18, 20, 21, 46, 60, 63, 65, 79]. Using both modalities jointly generally leads to performance boosts over using one modality in isolation. We examine this observation closely and aim to evaluate the effective use of multiple input modalities for the video classification task. To enable this, we propose — to the best of our knowledge, the first multi-label video classification benchmark that includes per-modality annotations for every sample (see Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Audio-visual foundation models Recently, multi-modal general-purpose models have emerged that can handle diverse downstream tasks without task-specific finetuning - also referred to as multi-modal foundation models. For instance, images or language were used as the bridge between modalities including audio, image, and text [30, 89]. Building on this, PandaGPT [72] leverages Vicuna [22] and ImageBind's embedding space to train a general multi-modal model exclusively on image-text pairs. Unified-IO 2 [53] employs universal tokenisation to process audio, video, and text. VideoLLaMA2 [21] uses a Spatial-Temporal Convolution connector in the visual branch before projecting audio and visual information into the LLM input space. The recently introduced Ola model [52] advances omnimodal processing through progressive modality alignment, using video to bridge audio and visual information. The Gemini models [73] are closed-source multi-modal models that achieve impressive performance on diverse downstream tasks. We use VGGSounder to benchmark the audio-visual capabilities of the aforementioned models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.824 + ], + "angle": 0, + "content": "Audio-visual classification benchmarks Audio-visual classification is distinct from general video classification (e.g. on YouTube-8M [1]), as classes typically cover both audible and visible actions or events. Commonly used datasets for audio-visual classification include KineticsSound [8] sourced from the Kinetics dataset [41], Flickr-SoundNet [11] scraped from Flickr, and AudioSet [29] and VGGSound [13], both sourced from YouTube." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Kinetics-Sound features manual labels of human actions, but covers only 34 classes. Flickr-SoundNet is much larger, but only a small subset is labelled. Similarly, only a small fraction of the roughly 2M AudioSet samples are annotated and have aligned audio and video." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1028" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Limitations of VGGSound. We show video frames from videos in the VGGSound test set along with their annotated label (grey) to demonstrate various limitations. A. VGGSound samples are labelled with a single class, yet many videos contain multiple distinct classes. B. Additionally, many classes partially overlap or are ambiguous. C. Some samples are labelled with classes that are not present in one of the modalities (i.e., the labelled class is not visible or audible)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.482, + 0.445 + ], + "angle": 0, + "content": "In contrast, VGGSound ensures audio-visual correspondence for around 200 000 samples and was curated with an automatic pipeline involving class-list generation, and auditory and visual content verification. The visual verification step ensures that a class is represented in the centre frame." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.483, + 0.581 + ], + "angle": 0, + "content": "The VEGAS dataset [88] provides better quality assurances for a small subset of AudioSet with only 10 classes. Visually Aligned Sounds [16] subsamples VEGAS and AudioSet after human verification, and Visual Sound [76] subsamples VGGSound using a multi-modal embedding model, both aiming for high audio-visual correspondence. Similarly, VGGSound-Sparse [37] is a subset of VGGSound with a focus on temporally and spatially sparse synchronisation signals (e.g., short loud noises)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Overall, VGGSound strikes the best balance between size, generality, and annotations, making it a common benchmark for audio-visual classification. We update VGGSound to sustain its usability for the development of the next generation of multi-modal foundation models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.338, + 0.685 + ], + "angle": 0, + "content": "3. Limitations of VGGSound" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Since we are interested in the VGGSound dataset for benchmarking audio-visual multi-modal models, our analysis focuses on the VGGSound test set, which consists of 15446 video clips, each 10s long and labelled with one of 309 classes. We qualitatively identify several limitations of the VGGSound annotations outlined below and in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Co-occurring classes While VGGSound's visual verification aimed to minimise multiple classes co-occurring in a clip, we find that most samples nevertheless clearly contain multiple classes, see Fig. 2A. In some cases, classes are temporally separated, e.g., showing male speech, man speaking and then cutting to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.508 + ], + "angle": 0, + "content": "footage of firing cannon. Most often, classes co-occur at the same time, sometimes for the entire duration of the video clip. Overlapping classes are often related, such as different instruments in a band or orchestra, but can also be entirely unrelated, e.g., donkey, ass braying co-occurring with playing violin. As additional empirical evidence, we provide a co-occurrence matrix computed using CAV-MAE [33], a state-of-the-art audio-visual model, in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.909, + 0.756 + ], + "angle": 0, + "content": "Overlapping classes The issue of co-occurring classes is exacerbated by many of the 309 automatically generated classes partially overlapping in their definition, as illustrated in Fig. 2B. We found two pairs of synonymous classes: timpani and tympani and dog barking and dog bow-wow. Additionally, some classes are strict subclasses of others, such as the gender-specific versions of cattle moing: cow lowing and bull bellowing; or the more specific variants of people eating: people eating noodle and people eating apple. Finally, several classes commonly appear together, such as playing snare drum which is often played as part of a drum kit, or semantically similar concepts: running electric fan and air conditioning noise, and sloshing water and splashing water." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Modality misalignment Despite VGGSound's auditory and visual content verification, we find that many of the annotated classes are not visible or not audible, as shown in Fig. 2C. A large fraction of videos contains background music, voice-over and narration, or other background sounds like bird chirping, tweeting or cricket chirping without a visible source. Similarly, some videos contain visible but inaudible cues for classes like sea waves. Static images and slide shows accompa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.448, + 0.902 + ], + "angle": 0, + "content": "1 Although these issues most likely also apply to the training set." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.34, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.092, + 0.522, + 0.105 + ], + "angle": 0, + "content": "B Label modality distribution" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.124, + 0.515, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.094, + 0.684, + 0.104 + ], + "angle": 0, + "content": "C Most common labels" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.106, + 0.709, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.095, + 0.848, + 0.104 + ], + "angle": 0, + "content": "D Meta labels" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.106, + 0.898, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Overview of VGGSounder. A. Most samples contain more than one label. B. More than a quarter of labels are audible but not visible. In contrast, only a tiny fraction is visible but not audible. C. Speech and bird sounds are the most common classes; more details can be found in Appendix B. D. Forty percent of the samples contain some combination of background music, voice over, and static image(s), making the classification task significantly harder." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.357, + 0.483, + 0.462 + ], + "angle": 0, + "content": "nied by music or other sounds are other frequent sources of misaligned modalities. Finally, some classes are misaligned by definition: wind noise is only audible and not visible. Overall, \\(48.43\\%\\) of the original VGGSound test samples have misaligned modalities. This finding challenges the widely held assumption that VGGSound has strong modality alignment [32, 47]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Other datasets, such as Visually Aligned Sounds, Visual Sound, and VGGSound-Sparse (see Sec. 2) omitted samples with misaligned modalities. In contrast, we contend that inaudible or invisible cues are common in natural videos and should be considered when benchmarking multi-modal models. We, therefore, place particular emphasis on crafting reliable modality annotations for all samples, allowing users to evaluate models on samples that guarantee modality alignment, and on those where classes are only visible or only audible (see Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.629, + 0.475, + 0.692 + ], + "angle": 0, + "content": "Takeaway 1 VGGSound suffers from several issues: class co-occurrence not captured by single labels, overlapping class definitions, and modality misalignment, see Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.307, + 0.726 + ], + "angle": 0, + "content": "4. Building VGGSounder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We propose a series of fixes for VGGSound's issues, ultimately resulting in the updated VGGSounder benchmark. We are not the first aiming to future-proof an existing benchmark: [12] analysed shortcomings of ImageNet [24], ultimately proposing switching to a multi-label classification task with additional manual labels. [28] similarly reannotated samples in MMLU [35, 36] to fix labelling errors. Both works inspire our approach to improve VGGSound. To deal with co-occurring classes, we switch to a multi-label classification setting. This effectively handles most overlapping class definitions: a strong model can assign a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.905, + 0.417 + ], + "angle": 0, + "content": "high probability to multiple classes, even if they partially overlap. This also allows us to ensure that synonymous classes, as well as subclasses and their superclass, always appear together in the ground-truth labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.906, + 0.54 + ], + "angle": 0, + "content": "To deal with modality misalignment, we add a modality annotation to each label. For example, we can label a video as containing people clapping in the audio and containing playing volleyball in the audio and video data. We also add meta-labels to indicate whether a sample contains background music, voice over, or static image(s) to optionally treat these cases separately during evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.905, + 0.571 + ], + "angle": 0, + "content": "We employ a pipeline similar to [12] to annotate multiple labels per sample, which we outline below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.906, + 0.698 + ], + "angle": 0, + "content": "Collecting proposals We create a gold standard reference set by labelling a small, randomly selected subset of VGGSound test samples with four in-house computer vision experts. The interface used for this first labelling step is shown in Appendix A. We extend the subset until each class is covered at least once, leading to a final size of 417 samples. Labels from different annotators are merged via a simple majority vote." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Given the gold standard set, we want to find a solid strategy for automatically generating label proposals which are shown to the humans labelling the test set. This should have a recall greater than \\(90\\%\\) while maximising precision compared to the gold standard labels to produce a small set of proposals with good label coverage. Our final strategy combines predictions from several state-of-the-art models with a manual heuristic to obtain \\(93\\%\\) recall for an average of 30 proposals per sample, see Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Human labelling We use Amazon Mechanical Turk to re-associate the entire VGGSound test set. For each sample, we first ask annotators to indicate whether the video contains background music, voice over, or" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1030" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "static image(s). Then, annotators are asked to indicate for each label proposal whether the class is audible and/or visible. Finally, annotators can add a class if it is missing from the proposals. Annotators were paid the US minimum wage; the interface used is shown in Appendix A. We let annotators label the samples in batches of 20, each containing two gold standard samples as catch trials. We reject and re-associate all batches with a catch trial \\( F_{1} \\)-score below \\( 25\\% \\). In addition, we obtain modality annotations for the original VGGSound labels and meta-classes. Further details are provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.482, + 0.352 + ], + "angle": 0, + "content": "Final labels We merge all obtained annotations using majority voting. Additionally, we automatically add synonymous classes and superclasses for a given subclass, e.g., we add cattle moing whenever cow lowing is in the set of labels. The full set of classes added this way is described in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.367, + 0.474, + 0.43 + ], + "angle": 0, + "content": "Takeaway 2 We develop VGGSounder: A multi-label video classification benchmark extending VGGSound with human-annotated multi-labels, modality annotations, and meta-labels as summarised in Fig. 3" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.411, + 0.463 + ], + "angle": 0, + "content": "5. Benchmarking audio-visual models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.517 + ], + "angle": 0, + "content": "We use VGGSounder to benchmark four popular audiovisual embedding models and seven foundation models, and analyse their auditory and visual capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.564 + ], + "angle": 0, + "content": "Models We evaluate the audio-visual embedding models CAV-MAE [33], DeepAVFusion [57], AV-Siam [49], and Equi-AV [43]. Those were finetuned on VGGSound." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.702 + ], + "angle": 0, + "content": "We benchmark several models from the closed-source Gemini family [73] in a zero-shot evaluation protocol. Furthermore, we use LLM-assisted evaluation to evaluate the following four open-source autoregressive foundation models: VideoLLaMA-2 [21], Unified-IO-2 [53], Panda-GPT [72], and Ola [52]. All models are evaluated in three modes: using unimodal-audio, unimodal-visual, or multi-modal (audio and visual) inputs. Further details about models and their evaluation are provided in Appendix C.1" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Metrics To benchmark the models on VGGSounder, we use multi-label classification metrics. For embedding models, all metrics are computed for the top-\\(k\\) predictions, with \\(k \\in \\{1, 3, 5, 10\\}\\). In contrast, prompting foundation models yields an unordered set of class predictions of varying size, and we compute only a single metric using the entire set. As a result, metrics are not directly comparable between embedding and foundation models. To get a sense of their relative performance, we report metrics for embedding models for \\(k = 1\\) in the main text, matching the median number of predictions per sample for the foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For open-source models such as VideoLLaMA-2, Ola, Unified-I0-2, and Panda-GPT, we employ LLM-assisted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "evaluation [55, 81], in which the Qwen-3 model [82] is tasked to assess the correspondence between model outputs and target classes. Closed-source models from the Gemini family are evaluated by providing the full list of 309 classes as input. Further details on the evaluation procedures and exact prompts are provided in Appendix C" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Subset accuracy compares the predicted label set to the ground-truth label set and reports the fraction of samples for which they match. This is our strictest metric." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.905, + 0.258 + ], + "angle": 0, + "content": "\\(F_{1}\\)-score is the harmonic mean of precision and recall. It is strictly larger than the subset accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.333 + ], + "angle": 0, + "content": "Hit reports the fraction of samples for which any of the predicted labels are part of the ground-truth label set. This is the most lenient metric which is strictly larger than the \\(F_{1}\\) score. We include this metric for ease of comparison to the \"ReaL-Accuracy\" used in [12]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.906, + 0.559 + ], + "angle": 0, + "content": "All metrics are computed separately for each input modality (audio, video, and audio-visual) and label modality. We use lowercase symbols a, v, and av to indicate the input modality: audio-only, visual-only, or audio-visual inputs, respectively. For label modality, we use uppercase symbols \\( A \\), \\( V \\), and \\( AV \\), referring to the subsets of the benchmark with audible, visible, and audio-visual labels. We further include \\( A \\neg V \\) (audible but not visible) and \\( V \\neg A \\) (visible but not audible) to analyze unaligned cues. For clarity, we define shorthand notations such as \\( a = a(A) \\) to denote the model's performance on the audible subset \\( A \\) using only audio input. Analogously, \\( v = v(V) \\) and \\( av = av(AV) \\) refer to video-only and audio-visual performance on their respective label subsets. Furthermore, we use micro-aggregation to balance the contribution from each class." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.56, + 0.905, + 0.605 + ], + "angle": 0, + "content": "We additionally measure the negative impact of using multimodal inputs. In particular, \\(\\mu\\) is a new metric we propose to measure a model's modality confusion \\((\\mu)\\). We define it as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.612, + 0.905, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\mathrm {M}} = 1 0 0 \\cdot \\frac {\\sum_ {x \\in M} \\mathcal {I} [ m (x) - \\text {c o r r e c t} \\cap a v (x) - \\text {w r o n g} ]}{N _ {\\text {t o t a l}}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.835 + ], + "angle": 0, + "content": "where \\( M \\in [A, V, A \\cap V] \\) and their associated modality inputs are \\( m \\in [a, v] \\), correct/wrong is determined as in the Hit score (with \\( k = 1 \\) for embedding models). \\( N_{total} \\) refers to the total number of samples. \\( \\mu \\) measures the fraction of samples a model correctly classified given an input modality but got wrong when using both modalities simultaneously. We additionally report \\( \\mu_{A \\cap V} \\) as the percentage of samples a model could solve in either modality unimodally but could not solve multi-modally. In other words, the modality confusion \\( \\mu \\) captures how frequently a model is distracted by an additional input modality, which can indicate the unsuccessful merging of modalities." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.848, + 0.898, + 0.896 + ], + "angle": 0, + "content": "Takeaway 3 We propose a new metric, modality confusion \\(\\mu\\), that measures how frequently a model is distracted by an additional input modality; see Eq. (1)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.328 + ], + "angle": 0, + "content": "
ModelSubset Accuracy ↑F1↑Hit ↑μ↓
avavavava(A-√V)v(V-√A)avavμAμVμA∩V
Embedding Models
CAV-MAE13.1919.2324.4934.4634.9142.6213.9419.0062.2953.4464.173.586.430.77
DeepAVFusion10.1911.1021.5325.3121.2937.3510.3710.5545.7732.6156.273.743.930.17
Equi-AV11.6010.5220.0029.3920.4234.6912.5510.6553.1231.2652.246.977.131.38
AV-Siam12.7919.7522.8333.3035.4139.4312.9018.2160.1954.2059.369.368.803.58
Closed-source Foundation Models
Gemini 1.5 Flash1.7814.4416.4414.4936.9842.5215.6121.6132.7347.3659.1010.224.250.77
Gemini 1.5 Pro3.0520.8622.5319.2649.7353.7417.7322.9035.0369.2375.422.094.850.57
Gemini 2.0 Flash1.8512.5412.6911.8034.0836.456.1918.9018.5143.8347.722.395.431.00
Open-source Foundation Models
VideoLLaMA 212.8619.8524.4738.8747.8252.3520.3428.0858.9152.0259.8012.725.462.95
Unified-IO 211.9411.5625.6135.3127.9248.8921.3816.5354.3931.0565.118.705.161.79
PandaGPT3.194.195.4618.7318.5620.8516.8214.4021.0817.0118.827.595.902.47
OLA14.118.6918.1947.7024.8546.4840.4413.4559.0524.5751.5115.476.802.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.339, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Table 2. Audio-visual video classification results on VGGSounder. We report multi-label classification metrics (subset accuracy, \\(F_{1}\\)-score, Hit accuracy, modality confusion \\((\\mu)\\) for audio- \\(a(A)\\), visual - \\(v(V)\\), audio-visual - \\(av(AV)\\), audio-only - \\(a(A\\neg V)\\) and video-only - \\(v(V\\neg A)\\) inputs. The embedding models CAV-MAE, DeepAVFusion, and Equi-AV were finetuned on the VGGSound training set. We report metrics for \\(k = 1\\) here and for other \\(k\\) in Appendix D The closed sourced multi-modal foundation models Gemini and open-sourced models use a zero-shot evaluation protocol and LLM-assisted protocol respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.379, + 0.442 + ], + "angle": 0, + "content": "5.1. Re-evaluating the state of the art" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.478 + ], + "angle": 0, + "content": "We present the benchmark performance of state-of-the-art audio-visual models in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.481, + 0.483, + 0.709 + ], + "angle": 0, + "content": "Overall performance Unsurprisingly, all models perform best with access to both input modalities (AV). Across all metrics, both open- and closed-source general-purpose models perform comparably to the purpose-built embedding models CAV-MAE and AV-Siam. This indicates that foundation models have reached — and for some modalities exceeded — the performance of specialised models. However, the embedding models finetuned on VGGSound generally have stronger unimodal performance with audio inputs \\((A)\\) compared to visual inputs \\((V)\\), which is in line with their pretraining on AudioSet. Interestingly, this trend is reversed for most foundation models, which seem to be biased towards visual inputs, with unimodal video performance \\((V)\\) being substantially higher than unimodal audio performance \\((A)\\)." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.726, + 0.475, + 0.805 + ], + "angle": 0, + "content": "Takeaway 4 Foundation models perform comparably to finetuned embedding models. Embedding models more heavily rely on audio cues than on visual ones, while foundation models exploit visible cues rather than audible ones, see Tab. 2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Modality confusion The modality confusion score \\(\\mu\\) shows that, for all models, a notable fraction of test samples \\((4 - 11\\%)\\) were misclassified when an additional modality was included—despite being correctly classified with unimodal input. Furthermore, for all models, a small portion of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.593 + ], + "angle": 0, + "content": "test samples is not solvable multi-modally even though they were solvable in both modalities alone \\((\\mu_{A\\cap V})\\). This insight is made possible by VGGSounder's per-label modality annotations and shows that all models are susceptible to being distracted given an additional modality. This is a concerning issue for multi-modal models since they should preserve unimodal capabilities when adding a second modality. Being able to evaluate this behaviour on the VGGSounder benchmark is a first step towards enabling the development of mitigation strategies, eventually resulting in stronger audio-visual models." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.607, + 0.899, + 0.656 + ], + "angle": 0, + "content": "Takeaway 5 Our modality confusion score reveals that all models are negatively impacted by additional modalities for a substantial amount of samples (see Tab. ②)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Performance across modalities Fig. 4 shows the performance profiles across modalities. At first glance, we can see that VideoLLaMA-2's performance is well balanced for different input modalities, while models from the Gemini family distinctly underperform on audio inputs. In contrast, embedding models exhibit a moderate balance across modalities, with DeepAVFusion and EquiAV showing slight underperformance for visual input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As Fig. 4 also illustrates, profiling of this kind is enabled through the modality annotations in VGGSounder. In contrast, VGGSound assumed that all classes are perceivable in both modalities, and did not account for background sounds. This resulted in consistent under-evaluation of foundation models (that were not finetuned on VGGSound) for audio inputs." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1032" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "In addition to the radar plot in Fig. 4, we provide results on VGGSound in Appendix D showing that all models have substantially lower performance than their hit scores in Tab. 2. This confirms that many model predictions were incorrectly flagged as false positives in VGGSound due to the incomplete ground-truth labels, painting a distorted picture of models' limitations." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.211, + 0.474, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.558 + ], + "angle": 0, + "content": "Figure 4. VGGSounder more accurately captures model performance across input modalities. We show the Hit score on VGGSounder and accuracy on VGGSound, normalised by the per-model maximum performance on each benchmark. Specifically for foundation models, we observe a significant difference in performance between VGGSound and VGGSounder." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.585, + 0.476, + 0.634 + ], + "angle": 0, + "content": "Takeaway 6 VGGSounder's more complete ground-truth labels allow for more accurate, modality-specific profiling of model performance (see Fig.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.437, + 0.669 + ], + "angle": 0, + "content": "5.2. Performance analysis using meta-classes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.764 + ], + "angle": 0, + "content": "VGGSounder includes annotations of three meta-classes for each sample: background music indicates whether samples contain music without a visible source, voice over similarly marks speech without a visible source, and static image(s) flags that the visual stream consists of one or only a few static visual frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "These new meta-classes allow us to evaluate the model behaviour in challenging scenarios where information from one modality dominates. We consider the performance difference between samples that do contain a meta-label and samples that do not contain it. Positive numbers indicate that the models perform better on the subset with metal-labels. In Tab.3, we summarise the main findings, focussing on \\(F_{1}\\)-score as the most balanced multi-label metric. Additional results are provided in Appendix D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "Background music All models perform worse on video samples containing background music. This indicates that it is challenging to decouple background audio from the rest of the video. The evaluated models are not good at differentiating between sound sources without visual cues, e.g. predicting different instruments in the background music." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Voice over In contrast to background music, we observe a clear difference between embedding models and foundation models for samples with voice-over. While the audio classification performance of embedding models drops substantially. This drop is only slight for VideoLLaMA-2 and Unified-IO 2, and the performance of other foundation models even improves. This indicates that the foundation models are less distracted by voice-over." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.4 + ], + "angle": 0, + "content": "Static image(s) The impact of static images is more nuanced: First, audio classification performance improves for embedding models while it decreases for the foundation models. This shows that the purpose-built, VGGSoundfinetuned embedding models can more accurately predict specific sounds in the absence of other cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Second, visual classification performance on static images drops for all models, suggesting that models rely on rich temporal cues to make accurate predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Third, the subset accuracy (i.e. exact label set matches) for uni-modal predictions shows that foundation models perform better in visual than audio classification with and without static images. In contrast, embedding models perform better in audio classification on static images but worse in visual classification for other samples. This suggests that non-static samples form a challenging subset for embedding models, where a model needs to favour one modality over another one to make a correct prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.906, + 0.692 + ], + "angle": 0, + "content": "Samples without any meta-label When comparing the model performance on samples without background music, static images, or voice-over annotations (column neither in Tab. 3) to the performance on the entire test set, we see a performance gain (here negative numbers indicate an increased \\(\\mathbf{F}_1\\) score). This finding concludes that these three categories form challenging subsets of the dataset." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.708, + 0.899, + 0.77 + ], + "angle": 0, + "content": "Takeaway 7 Samples with background music, static images(s), and voice over provide distinct challenges for each model (see Tab.3). This highlights VGGSounder's value for comprehensive model evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.782, + 0.804 + ], + "angle": 0, + "content": "5.3. Impact of VGGSounder labels" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our relabelling pipeline adds two types of labels to those in VGGSound: (1) automatically generated labels based on synonymous classes and sub-/superclasses, and (2) human-curated labels. In Tab.4, we ablate the impact of each type of added labels in terms of the relative performance gains (Hit score). A complete breakdown of the effects across all" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1033" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.306 + ], + "angle": 0, + "content": "
Modelbackground musicvoice over ΔF1static image(s)neither ΔF1
ΔF1ΔF1ΔF1Sub. Acc. w/Sub. Acc. w/oavav
avavaavavavavav
Embedding Models
CAV-MAE-0.66-0.59-0.65-1.130.31-0.2812.7319.2211.9819.21-1.00-1.27-1.24
DeepAVFusion-0.71-0.80-0.70-1.250.28-0.189.7910.819.3010.81-1.27-1.01-1.36
Equi-AV-0.79-0.42-0.41-0.970.28-0.2311.1210.4510.4910.45-0.97-0.97-0.89
AV-Siam-0.73-0.72-0.82-1.080.34-0.2612.3419.5311.5719.52-1.01-1.34-1.39
Closed-source Foundation Models
Gemini 1.5 Flash-0.22-0.41-0.692.08-0.36-0.251.6614.351.6814.391.40-1.10-1.62
Gemini 1.5 Pro-0.36-0.61-0.942.48-0.31-0.242.8320.852.8720.801.64-0.89-1.09
Gemini 2.0 Flash-0.09-0.32-0.580.030.13-0.261.7012.331.5312.40-0.03-1.07-1.22
Open-source Foundation Models
VideoLLaMA 2-0.47-0.72-0.84-0.540.27-0.3112.5519.6412.0419.70-0.47-1.26-1.33
Unified-IO 2-1.200.18-0.65-0.680.12-0.2111.3911.8910.8812.00-1.07-0.15-1.17
PandaGPT-1.13-0.15-0.420.56-0.22-0.162.944.272.914.24-0.60-0.15-0.65
OLA-2.160.09-0.441.50-0.48-0.1813.038.8812.888.89-0.80-0.13-0.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.316, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 3. Summary of performance differences in the presence/absence of meta-classes. Difference in \\( \\mathbf{F}_1 \\) scores (\\( \\Delta \\mathbf{F}_1 \\)) for audio-visual video classification on VGGSounder between videos with a meta-class and those without it. Positive numbers (\\( \\Delta \\)) indicate better performance when the meta-class is present. Additional results are provided in Appendix D" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.373, + 0.481, + 0.445 + ], + "angle": 0, + "content": "
ModelHuman labels ↑Auto labels ↑
AVAVAVAV
Gemini 1.5 Flash29.2814.5916.360.480.931.51
Gemini 1.5 Pro28.6125.5227.630.311.992.10
Gemini 2.0 Flash8.8012.1611.130.221.121.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Table 4. Impact of added labels using different strategies in VGGSounder. We show the change in multi-label classification performance (\\(\\Delta\\) Hit) when adding automatically added (Auto) or human-annotated (Human) labels to VGGSound, and compare to the original VGGSound data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.484, + 0.651 + ], + "angle": 0, + "content": "models and metrics is provided in Appendix D. While performance is consistently higher with automatically added labels (Auto), the increase is noticeably smaller than that for human-curated labels. Paired with the observation that models do frequently predict correct classes that were not part of the original VGGSound label set, this indicates that human-curated labels better cover the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.668, + 0.475, + 0.73 + ], + "angle": 0, + "content": "Takeaway 8 Automatically added labels are an important step, but human-curated labels have a bigger effect on eliminating incorrectly flagged false positives, underscoring the value of accurate human annotation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.751, + 0.206, + 0.767 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Choice of VGGSound as base dataset VGGSound is commonly used to evaluate audio-visual models on the multimodal video classification task. As it is currently the most suitable testbed for audio-visual classification tasks (due to its size, diversity of categories, non-constrained setting, and relatively strong audio-visual correspondence), it serves as an optimal starting point for our substantially improved VGGSounder benchmark with a multi-label evaluation pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.905, + 0.406 + ], + "angle": 0, + "content": "tocol for foundation models that makes the benchmark suitable for meaningful evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.56 + ], + "angle": 0, + "content": "Multi-label vs single-label classification Video content is inherently complex, often containing multiple co-occurring objects and actions both within and across modalities. This makes it unlikely that a given clip belongs to just one class as is the case in the single-label classification task. Therefore, our VGGSounder extends the VGGSound test set to multi-label classification. Unlike models trained on a narrow single-label dataset, foundation models develop versatile representations. Multi-label evaluation thus also aligns very naturally with this capability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.634, + 0.59 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.705 + ], + "angle": 0, + "content": "We introduced an modality-aware evaluation set for audiovisual foundation models. VGGSounder builds on the widely used VGGSound dataset by adding: (a) comprehensive human annotations for missing classes, (b) specifying modality information per label, (c) introducing specialised meta-labels for frequently occurring real-world challenges, and (d) using heuristic methods to improve label quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.909, + 0.886 + ], + "angle": 0, + "content": "Through our newly introduced metric, modality confusion, we observe that incorporating additional modalities does not necessarily yield better results. Models often become more confused on a substantial subset of test samples. Furthermore, finetuned embedding models tend to rely heavily on audio cues, while foundation models depend more on visual information. Additionally, our meta-label analysis highlights distinct challenges across various specialised yet commonly occurring scenarios such as background music, static images, and voice-overs. Overall, we hope that the VGGSounder benchmark will advance the evaluation and development of foundational audio-visual models." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1034" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.482, + 0.206 + ], + "angle": 0, + "content": "The authors would like to thank Felix Förster, Sayak Mallick, and Prasanna Mayilvahananan for their help with data annotation, Felix Förster and Monica Riedler for proofreading the paper, and Thomas Klein and Shyamgopal Karthik for their help in setting up MTurk. They also thank numerous MTurk workers for labelling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.483, + 0.388 + ], + "angle": 0, + "content": "This work was in part supported by the BMBF (FKZ: 01IS24060, 01I524085B, 01IS18039A), the DFG (SFB 1233, project number: 276693517), and the Open Philanthropy Foundation funded by the Good Ventures Foundation. WB acknowledges financial support via an Emmy Noether Grant funded by the German Research Foundation (DFG) under grant no. BR 6382/1-1. WB is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This research utilised compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG. The authors thank the IMPRS-IS for supporting TW." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.403, + 0.188, + 0.418 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.428, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.499, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE TPAMI, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[3] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.629, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[5] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[6] Humam Alwassel, Dhruv Mahajan, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.715, + 0.482, + 0.756 + ], + "angle": 0, + "content": "[7] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.758, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[8] Relja Arandjelović and Andrew Zisserman. Look, listen and learn. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[9] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.817, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Yuki M. Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[11] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. In NeurIPS, 2016." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.428, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[12] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Audio-visual synchronisation in the wild. In BMVC, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[15] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[16] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[17] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[18] Yanbei Chen, Yongqin Xian, A. Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[19] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. In CVPR, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[20] Ying Cheng, Ruize Wang, Zhihao Pan, Rui Feng, and Yuejie Zhang. Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In ACM MM, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[21] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[22] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[23] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[25] Joshua P Ebeneze, Yongjun Wu, Hai Wei, Sriram Sethuraman, and Zongyi Liu. Detection of audio-video synchronization errors via event detection. In ICASSP, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[26] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[27] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1035" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Reza Ghasemi Madani, Claire Barale, Robert McHardy, Joshua Harris, Jean Kaddour, Emile van Krieken, and Pasquale Minervini. Are we done with mmlu? arXiv preprint arXiv:2406.04127, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[29] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[30] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[31] Shir Goldstein and Yael Moses. Guitar music transcription from silent video. In BMVC, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[32] Yuan Gong, Alexander H Liu, Andrew Rouditchenko, and James Glass. Uavm: Towards unifying audio and visual models. IEEE Signal Processing Letters, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[33] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James Glass. Contrastive audio-visual masked autoencoder. In ICLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[34] Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[35] Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. Aligning ai with shared human values. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[36] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[37] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[38] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[39] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[40] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. IJCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[41] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[42] Naji Khosravan, Shervin Ardeshir, and Rohit Puri. On attention modules for audio-visual synchronization. In CVPR Workshop, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[43] Jongsuk Kim, Hyeongkeun Lee, Kyeongha Rho, Junmo Kim, and Joon Son Chung. Equipav: Leveraging equivariance for audio-visual contrastive learning. In ICML, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] A. Sophia Koepke, Olivia Wiles, and Andrew Zisserman. Visual pitch estimation. In SMC, 2019." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[45] A Sophia Koepke, Olivia Wiles, Yael Moses, and Andrew Zisserman. Sight to sound: An end-to-end approach for visual piano transcription. In ICASSP, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[46] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurlPS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[47] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song. Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.905, + 0.278 + ], + "angle": 0, + "content": "[48] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.904, + 0.308 + ], + "angle": 0, + "content": "[49] Yan-Bo Lin and Gedas Bertasius. Siamese vision transformers are scalable audio-visual learners. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[50] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In ACCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.904, + 0.394 + ], + "angle": 0, + "content": "[51] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[52] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[53] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[54] Jie Ma, Min Hu, Pinghui Wang, Wangchun Sun, Lingyun Song, Hongbin Pei, Jun Liu, and Youtian Du. Look, listen, and answer: Overcoming biases for audio-visual question answering. NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[55] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[56] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[57] Shentong Mo and Pedro Morgado. Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[58] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. In BMVC, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[59] Arsha Nagrani, Joon Son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision. In ICASSP, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[60] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[61] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. arXiv preprint arXiv:2104.02687, 2021." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1036" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[62] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[63] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In ECCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[64] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Learning sight from sound: Ambient sound provides supervision for visual learning. IJCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[65] Mandela Patrick, Yuki M Asano, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[66] KR Prajwal, Liliane Momeni, Triantafyllos Afouras, and Andrew Zisserman. Visual keyword spotting with attention. In BMVC, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[67] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[68] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[69] Kun Su, Xiulong Liu, and Eli Shlizerman. Multi-instrumentalist net: Unsupervised generation of music from body movements. arXiv preprint arXiv:2012.03478, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[70] Kun Su, Xiulong Liu, and Eli Shlizerman. How does it sound? In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[71] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In ICML, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[72] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[73] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[74] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[75] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel PW Ellis, and John R Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[76] Ilpo Viertola, Vladimir Iashin, and Esa Rahtu. Temporally aligned audio for video with autoregression. In ICASSP, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[77] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[78] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[79] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[80] Haoming Xu, Runhao Zeng, Qingyao Wu, Mingkui Tan, and Chuang Gan. Cross-modal relation-aware networks for audio-visual event localization. In ACM MM, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[81] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[82] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[83] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audio-visual question answering on videos. In ACM MM, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[84] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[85] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[86] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[87] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[88] Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L. Berg. Visual to sound: Generating natural sound for videos in the wild. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.533 + ], + "angle": 0, + "content": "[89] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.562 + ], + "angle": 0, + "content": "[90] Lingyu Zhu and Esa Rahtu. V-slowfast network for efficient visual sound separation. In WACV, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1037" + } + ] +] \ No newline at end of file diff --git a/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_origin.pdf b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3bac496b514a38ea0b5ee02f475723257c7b09f9 --- /dev/null +++ b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/a6042e74-fb7f-4488-a331-f3b644cd7772_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:701a3a3a50894a72b2fd7312938fc0aae508ecf2d8a28624711d615b85da6514 +size 938871 diff --git a/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/full.md b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3d6e62ba6888a83b484d0463a20faf35294fdb7b --- /dev/null +++ b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/full.md @@ -0,0 +1,331 @@ +![](images/68614e85b835a66db0667fcd63c2b5cc77b7b39f4a49010171318974b8b9b39b.jpg) + +# VGGSounder: Audio-Visual Evaluations for Foundation Models + +Daniil Zverev1,* Thaddäus Wiedemer2,3,4,* Ameya Prabhu2,3 + +Matthias Bethge $^{2,3}$ Wieland Brendel $^{3,4}$ A. Sophia Koepke $^{1,2,3}$ + +$^{1}$ Technical University of Munich, MCML $^{2}$ University of Tübingen $^{3}$ Tübingen AI Center $^{4}$ MPI for Intelligent Systems, ELLIS Institute Tübingen + +{daniil.zverev,a-sophia.koepke}@tum.de,{thaddaeus.wiedemer,ameya.prabhu}@uni-tuebingen.de + +# Abstract + +The emergence of audio-visual foundation models underscores the importance of reliably assessing their multimodal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric. Our dataset and project page are available at https://vqgsounder.github.io/ + +# 1. Introduction + +Rigorous evaluation benchmarks have been instrumental in assessing the effectiveness of audio-visual models [33, 43, 49, 57]. Specifically, multi-modal foundation models integrating visual and auditory data aim to achieve a holistic understanding of audio-visual content. However, the field lacks large-scale modality-aware classification benchmarks with ground-truth annotations indicating whether each label is visible, audible, or both. Such annotations would allow detailed evaluations of multi-modal model capabilities. To address this gap, we introduce VGGSounder, an enhanced version of the widely-used audio-visual classification dataset VGGSound [13], which facilitates modality- + +![](images/2b489317aa6ca2aff9ddd7b4f939ceccbb63d7e5eb5c60b2e953cee0770ab8f8.jpg) +Figure 1. We introduce VGGSounder, a multi-label audio-visual classification benchmark with modality annotations. We extend the original VGGSound test set with human-annotated audible, visible, and visible+audible labels. We add meta labels for common confounders, such as background music. We benchmark eleven recent audio-visual models on VGGSounder. It enables selective analysis of a model's auditory and visual capabilities on classes relevant for the queried modality. + +aware evaluation of audio-visual foundation models. + +VGGSound has several notable limitations. First, its data is inherently multi-label; for instance, a single sample may simultaneously include labels such as playing drum kit and playing acoustic guitar when multiple instruments are present. Additionally, evaluating how different modalities contribute to model performance becomes difficult without explicit modality annotations, as some labels are either not visually present or not audible (e.g., certain instruments might only be audible but not visible in advertisements). Moreover, overlapping label classes present another challenge; for example, + +the orchestra label often coincides with labels for individual instruments. These issues result in systematic under-evaluation of multi-modal audio-visual foundation models. + +To overcome these limitations, we present VGGSounder, an improved benchmark inspired by similar advancements in other domains [12, 28]. We re-announce the dataset to create a comprehensive multi-label classification setting by collecting detailed annotations for each sample, including (1) additional classes present, (2) explicit modality annotations to label modality misalignment, (3) metadata indicating the presence of background music, voice-over, or static images, and (4) merging of classes to address overlapping classes. Consequently, VGGSounder provides a robust, foundation-model-ready benchmark enabling structured analysis of whether models rely on audio or visual cues. Furthermore, we include meta-labels (e.g., background music, voice-over, or static images) to easily filter out unreliable labels during evaluation. Utilising VGGSounder, we evaluate audio-visual foundation models, demonstrating their poor performance on our benchmark. We find that the state-of-the-art, closed-source Gemini models consistently rely exclusively on the visual modality. In addition to that, we measure the modality confusion, i.e. when models get distracted by an additional input modality, which exposes the unsuccessful merging of modalities. These findings highlight the importance of the audio-focused VGGSounder benchmark as a critical tool for accurately assessing audio-visual foundation models. + +We make the following contributions: + +1. We illustrate limitations of VGGSound in Sec. 3 +2. We curate VGGSounder with multi-modal human annotations for multi-label classification in Sec. 4 +3. We evaluate state-of-the-art audio-visual models, observing differences between embedding models and autoregressive foundation models in Sec. 5. +4. We propose new metrics to quantify the negative impact of using multiple input modalities in Sec. 5. + +# 2. Related work + +Audio-visual learning Many prior works consider audio-visual tasks that include sound source localisation and separation [3, 5, 9, 15, 27, 56, 62, 67, 75, 80, 85, 86, 90], event localisation [50, 51, 74, 78], audio-visual question answering [48, 54, 83, 84], audio-visual synchronisation [14, 23, 25, 38, 39, 42], audio synthesis using visual information [19, 26, 31, 44, 45, 61, 69-71, 87], or audio-driven face image synthesis [7, 40, 77]. Audio-visual data has also been leveraged for speech-related tasks, including speech and speaker recognition [2, 4, 59], or the spotting of spoken keywords [58, 66]. + +Furthermore, the natural alignment between audio and + +
Dataset# Clips# ClassesMulti- labelModality labelsAnnotation pipeline
Flickr-SoundNet [10]2M-✘✘-
Kinetics-Sound [7]18.8k34✘✘MTurk
AudioSet [26]2.1M537✓✘Manual
AVE [62]4k28✓✘Manual
VEGSAS [76]132k10✘✘MTurk
Visually Aligned Sounds [15]13k8✘✘MTurk
VGGSound [12]200k309✘✘Classifiers+Manual
VGGSound-Sparse [32]7.1k12✘✘Manual
Visual Sound [66]91k309✘✘ImageBind [30]
VGGSounder15.4k309✓✓MTurk
+ +Table 1. Comparison of audio-visual classification benchmarks. + +video has been exploited to learn improved audio-visual embeddings for downstream tasks [6, 10, 11, 18, 20, 21, 46, 60, 63, 65, 79]. Using both modalities jointly generally leads to performance boosts over using one modality in isolation. We examine this observation closely and aim to evaluate the effective use of multiple input modalities for the video classification task. To enable this, we propose — to the best of our knowledge, the first multi-label video classification benchmark that includes per-modality annotations for every sample (see Tab. 1). + +Audio-visual foundation models Recently, multi-modal general-purpose models have emerged that can handle diverse downstream tasks without task-specific finetuning - also referred to as multi-modal foundation models. For instance, images or language were used as the bridge between modalities including audio, image, and text [30, 89]. Building on this, PandaGPT [72] leverages Vicuna [22] and ImageBind's embedding space to train a general multi-modal model exclusively on image-text pairs. Unified-IO 2 [53] employs universal tokenisation to process audio, video, and text. VideoLLaMA2 [21] uses a Spatial-Temporal Convolution connector in the visual branch before projecting audio and visual information into the LLM input space. The recently introduced Ola model [52] advances omnimodal processing through progressive modality alignment, using video to bridge audio and visual information. The Gemini models [73] are closed-source multi-modal models that achieve impressive performance on diverse downstream tasks. We use VGGSounder to benchmark the audio-visual capabilities of the aforementioned models. + +Audio-visual classification benchmarks Audio-visual classification is distinct from general video classification (e.g. on YouTube-8M [1]), as classes typically cover both audible and visible actions or events. Commonly used datasets for audio-visual classification include KineticsSound [8] sourced from the Kinetics dataset [41], Flickr-SoundNet [11] scraped from Flickr, and AudioSet [29] and VGGSound [13], both sourced from YouTube. + +Kinetics-Sound features manual labels of human actions, but covers only 34 classes. Flickr-SoundNet is much larger, but only a small subset is labelled. Similarly, only a small fraction of the roughly 2M AudioSet samples are annotated and have aligned audio and video. + +Figure 2. Limitations of VGGSound. We show video frames from videos in the VGGSound test set along with their annotated label (grey) to demonstrate various limitations. A. VGGSound samples are labelled with a single class, yet many videos contain multiple distinct classes. B. Additionally, many classes partially overlap or are ambiguous. C. Some samples are labelled with classes that are not present in one of the modalities (i.e., the labelled class is not visible or audible). + +In contrast, VGGSound ensures audio-visual correspondence for around 200 000 samples and was curated with an automatic pipeline involving class-list generation, and auditory and visual content verification. The visual verification step ensures that a class is represented in the centre frame. + +The VEGAS dataset [88] provides better quality assurances for a small subset of AudioSet with only 10 classes. Visually Aligned Sounds [16] subsamples VEGAS and AudioSet after human verification, and Visual Sound [76] subsamples VGGSound using a multi-modal embedding model, both aiming for high audio-visual correspondence. Similarly, VGGSound-Sparse [37] is a subset of VGGSound with a focus on temporally and spatially sparse synchronisation signals (e.g., short loud noises). + +Overall, VGGSound strikes the best balance between size, generality, and annotations, making it a common benchmark for audio-visual classification. We update VGGSound to sustain its usability for the development of the next generation of multi-modal foundation models. + +# 3. Limitations of VGGSound + +Since we are interested in the VGGSound dataset for benchmarking audio-visual multi-modal models, our analysis focuses on the VGGSound test set, which consists of 15446 video clips, each 10s long and labelled with one of 309 classes. We qualitatively identify several limitations of the VGGSound annotations outlined below and in Fig. 2. + +Co-occurring classes While VGGSound's visual verification aimed to minimise multiple classes co-occurring in a clip, we find that most samples nevertheless clearly contain multiple classes, see Fig. 2A. In some cases, classes are temporally separated, e.g., showing male speech, man speaking and then cutting to + +footage of firing cannon. Most often, classes co-occur at the same time, sometimes for the entire duration of the video clip. Overlapping classes are often related, such as different instruments in a band or orchestra, but can also be entirely unrelated, e.g., donkey, ass braying co-occurring with playing violin. As additional empirical evidence, we provide a co-occurrence matrix computed using CAV-MAE [33], a state-of-the-art audio-visual model, in Appendix D. + +Overlapping classes The issue of co-occurring classes is exacerbated by many of the 309 automatically generated classes partially overlapping in their definition, as illustrated in Fig. 2B. We found two pairs of synonymous classes: timpani and tympani and dog barking and dog bow-wow. Additionally, some classes are strict subclasses of others, such as the gender-specific versions of cattle moing: cow lowing and bull bellowing; or the more specific variants of people eating: people eating noodle and people eating apple. Finally, several classes commonly appear together, such as playing snare drum which is often played as part of a drum kit, or semantically similar concepts: running electric fan and air conditioning noise, and sloshing water and splashing water. + +Modality misalignment Despite VGGSound's auditory and visual content verification, we find that many of the annotated classes are not visible or not audible, as shown in Fig. 2C. A large fraction of videos contains background music, voice-over and narration, or other background sounds like bird chirping, tweeting or cricket chirping without a visible source. Similarly, some videos contain visible but inaudible cues for classes like sea waves. Static images and slide shows accompa + +![](images/7c9ba699608f529a550d998b95fefef92562d56bd4f6c77fefb12df0bb131ed6.jpg) +Figure 3. Overview of VGGSounder. A. Most samples contain more than one label. B. More than a quarter of labels are audible but not visible. In contrast, only a tiny fraction is visible but not audible. C. Speech and bird sounds are the most common classes; more details can be found in Appendix B. D. Forty percent of the samples contain some combination of background music, voice over, and static image(s), making the classification task significantly harder. + +![](images/94fbb8f345cb4997d6a8c853c6e957bfcd248d0cf8614eb593fba6ff5bc8f081.jpg) + +![](images/d3c1a4923a33d3f9b62441bb900fbf905856b1e68189ceeca9616911ae16440e.jpg) +B Label modality distribution +C Most common labels + +![](images/c63d15d87efbc62ba4c4c351ed68e88dbdc145a6c81095ccf49958c67c05d52b.jpg) +D Meta labels + +nied by music or other sounds are other frequent sources of misaligned modalities. Finally, some classes are misaligned by definition: wind noise is only audible and not visible. Overall, $48.43\%$ of the original VGGSound test samples have misaligned modalities. This finding challenges the widely held assumption that VGGSound has strong modality alignment [32, 47]. + +Other datasets, such as Visually Aligned Sounds, Visual Sound, and VGGSound-Sparse (see Sec. 2) omitted samples with misaligned modalities. In contrast, we contend that inaudible or invisible cues are common in natural videos and should be considered when benchmarking multi-modal models. We, therefore, place particular emphasis on crafting reliable modality annotations for all samples, allowing users to evaluate models on samples that guarantee modality alignment, and on those where classes are only visible or only audible (see Tab. 1). + +Takeaway 1 VGGSound suffers from several issues: class co-occurrence not captured by single labels, overlapping class definitions, and modality misalignment, see Fig. 2. + +# 4. Building VGGSounder + +We propose a series of fixes for VGGSound's issues, ultimately resulting in the updated VGGSounder benchmark. We are not the first aiming to future-proof an existing benchmark: [12] analysed shortcomings of ImageNet [24], ultimately proposing switching to a multi-label classification task with additional manual labels. [28] similarly reannotated samples in MMLU [35, 36] to fix labelling errors. Both works inspire our approach to improve VGGSound. To deal with co-occurring classes, we switch to a multi-label classification setting. This effectively handles most overlapping class definitions: a strong model can assign a + +high probability to multiple classes, even if they partially overlap. This also allows us to ensure that synonymous classes, as well as subclasses and their superclass, always appear together in the ground-truth labels. + +To deal with modality misalignment, we add a modality annotation to each label. For example, we can label a video as containing people clapping in the audio and containing playing volleyball in the audio and video data. We also add meta-labels to indicate whether a sample contains background music, voice over, or static image(s) to optionally treat these cases separately during evaluation. + +We employ a pipeline similar to [12] to annotate multiple labels per sample, which we outline below. + +Collecting proposals We create a gold standard reference set by labelling a small, randomly selected subset of VGGSound test samples with four in-house computer vision experts. The interface used for this first labelling step is shown in Appendix A. We extend the subset until each class is covered at least once, leading to a final size of 417 samples. Labels from different annotators are merged via a simple majority vote. + +Given the gold standard set, we want to find a solid strategy for automatically generating label proposals which are shown to the humans labelling the test set. This should have a recall greater than $90\%$ while maximising precision compared to the gold standard labels to produce a small set of proposals with good label coverage. Our final strategy combines predictions from several state-of-the-art models with a manual heuristic to obtain $93\%$ recall for an average of 30 proposals per sample, see Appendix A. + +Human labelling We use Amazon Mechanical Turk to re-associate the entire VGGSound test set. For each sample, we first ask annotators to indicate whether the video contains background music, voice over, or + +static image(s). Then, annotators are asked to indicate for each label proposal whether the class is audible and/or visible. Finally, annotators can add a class if it is missing from the proposals. Annotators were paid the US minimum wage; the interface used is shown in Appendix A. We let annotators label the samples in batches of 20, each containing two gold standard samples as catch trials. We reject and re-associate all batches with a catch trial $F_{1}$ -score below $25\%$ . In addition, we obtain modality annotations for the original VGGSound labels and meta-classes. Further details are provided in Appendix A. + +Final labels We merge all obtained annotations using majority voting. Additionally, we automatically add synonymous classes and superclasses for a given subclass, e.g., we add cattle moing whenever cow lowing is in the set of labels. The full set of classes added this way is described in Appendix A. + +Takeaway 2 We develop VGGSounder: A multi-label video classification benchmark extending VGGSound with human-annotated multi-labels, modality annotations, and meta-labels as summarised in Fig. 3 + +# 5. Benchmarking audio-visual models + +We use VGGSounder to benchmark four popular audiovisual embedding models and seven foundation models, and analyse their auditory and visual capabilities. + +Models We evaluate the audio-visual embedding models CAV-MAE [33], DeepAVFusion [57], AV-Siam [49], and Equi-AV [43]. Those were finetuned on VGGSound. + +We benchmark several models from the closed-source Gemini family [73] in a zero-shot evaluation protocol. Furthermore, we use LLM-assisted evaluation to evaluate the following four open-source autoregressive foundation models: VideoLLaMA-2 [21], Unified-IO-2 [53], Panda-GPT [72], and Ola [52]. All models are evaluated in three modes: using unimodal-audio, unimodal-visual, or multi-modal (audio and visual) inputs. Further details about models and their evaluation are provided in Appendix C.1 + +Metrics To benchmark the models on VGGSounder, we use multi-label classification metrics. For embedding models, all metrics are computed for the top- $k$ predictions, with $k \in \{1, 3, 5, 10\}$ . In contrast, prompting foundation models yields an unordered set of class predictions of varying size, and we compute only a single metric using the entire set. As a result, metrics are not directly comparable between embedding and foundation models. To get a sense of their relative performance, we report metrics for embedding models for $k = 1$ in the main text, matching the median number of predictions per sample for the foundation models. + +For open-source models such as VideoLLaMA-2, Ola, Unified-I0-2, and Panda-GPT, we employ LLM-assisted + +evaluation [55, 81], in which the Qwen-3 model [82] is tasked to assess the correspondence between model outputs and target classes. Closed-source models from the Gemini family are evaluated by providing the full list of 309 classes as input. Further details on the evaluation procedures and exact prompts are provided in Appendix C + +Subset accuracy compares the predicted label set to the ground-truth label set and reports the fraction of samples for which they match. This is our strictest metric. + +$F_{1}$ -score is the harmonic mean of precision and recall. It is strictly larger than the subset accuracy. + +Hit reports the fraction of samples for which any of the predicted labels are part of the ground-truth label set. This is the most lenient metric which is strictly larger than the $F_{1}$ score. We include this metric for ease of comparison to the "ReaL-Accuracy" used in [12]. + +All metrics are computed separately for each input modality (audio, video, and audio-visual) and label modality. We use lowercase symbols a, v, and av to indicate the input modality: audio-only, visual-only, or audio-visual inputs, respectively. For label modality, we use uppercase symbols $A$ , $V$ , and $AV$ , referring to the subsets of the benchmark with audible, visible, and audio-visual labels. We further include $A \neg V$ (audible but not visible) and $V \neg A$ (visible but not audible) to analyze unaligned cues. For clarity, we define shorthand notations such as $a = a(A)$ to denote the model's performance on the audible subset $A$ using only audio input. Analogously, $v = v(V)$ and $av = av(AV)$ refer to video-only and audio-visual performance on their respective label subsets. Furthermore, we use micro-aggregation to balance the contribution from each class. + +We additionally measure the negative impact of using multimodal inputs. In particular, $\mu$ is a new metric we propose to measure a model's modality confusion $(\mu)$ . We define it as + +$$ +\mu_ {\mathrm {M}} = 1 0 0 \cdot \frac {\sum_ {x \in M} \mathcal {I} [ m (x) - \text {c o r r e c t} \cap a v (x) - \text {w r o n g} ]}{N _ {\text {t o t a l}}}, \tag {1} +$$ + +where $M \in [A, V, A \cap V]$ and their associated modality inputs are $m \in [a, v]$ , correct/wrong is determined as in the Hit score (with $k = 1$ for embedding models). $N_{total}$ refers to the total number of samples. $\mu$ measures the fraction of samples a model correctly classified given an input modality but got wrong when using both modalities simultaneously. We additionally report $\mu_{A \cap V}$ as the percentage of samples a model could solve in either modality unimodally but could not solve multi-modally. In other words, the modality confusion $\mu$ captures how frequently a model is distracted by an additional input modality, which can indicate the unsuccessful merging of modalities. + +Takeaway 3 We propose a new metric, modality confusion $\mu$ , that measures how frequently a model is distracted by an additional input modality; see Eq. (1). + +
ModelSubset Accuracy ↑F1↑Hit ↑μ↓
avavavava(A-√V)v(V-√A)avavμAμVμA∩V
Embedding Models
CAV-MAE13.1919.2324.4934.4634.9142.6213.9419.0062.2953.4464.173.586.430.77
DeepAVFusion10.1911.1021.5325.3121.2937.3510.3710.5545.7732.6156.273.743.930.17
Equi-AV11.6010.5220.0029.3920.4234.6912.5510.6553.1231.2652.246.977.131.38
AV-Siam12.7919.7522.8333.3035.4139.4312.9018.2160.1954.2059.369.368.803.58
Closed-source Foundation Models
Gemini 1.5 Flash1.7814.4416.4414.4936.9842.5215.6121.6132.7347.3659.1010.224.250.77
Gemini 1.5 Pro3.0520.8622.5319.2649.7353.7417.7322.9035.0369.2375.422.094.850.57
Gemini 2.0 Flash1.8512.5412.6911.8034.0836.456.1918.9018.5143.8347.722.395.431.00
Open-source Foundation Models
VideoLLaMA 212.8619.8524.4738.8747.8252.3520.3428.0858.9152.0259.8012.725.462.95
Unified-IO 211.9411.5625.6135.3127.9248.8921.3816.5354.3931.0565.118.705.161.79
PandaGPT3.194.195.4618.7318.5620.8516.8214.4021.0817.0118.827.595.902.47
OLA14.118.6918.1947.7024.8546.4840.4413.4559.0524.5751.5115.476.802.49
+ +Table 2. Audio-visual video classification results on VGGSounder. We report multi-label classification metrics (subset accuracy, $F_{1}$ -score, Hit accuracy, modality confusion $(\mu)$ for audio- $a(A)$ , visual - $v(V)$ , audio-visual - $av(AV)$ , audio-only - $a(A\neg V)$ and video-only - $v(V\neg A)$ inputs. The embedding models CAV-MAE, DeepAVFusion, and Equi-AV were finetuned on the VGGSound training set. We report metrics for $k = 1$ here and for other $k$ in Appendix D The closed sourced multi-modal foundation models Gemini and open-sourced models use a zero-shot evaluation protocol and LLM-assisted protocol respectively. + +# 5.1. Re-evaluating the state of the art + +We present the benchmark performance of state-of-the-art audio-visual models in Tab. 2. + +Overall performance Unsurprisingly, all models perform best with access to both input modalities (AV). Across all metrics, both open- and closed-source general-purpose models perform comparably to the purpose-built embedding models CAV-MAE and AV-Siam. This indicates that foundation models have reached — and for some modalities exceeded — the performance of specialised models. However, the embedding models finetuned on VGGSound generally have stronger unimodal performance with audio inputs $(A)$ compared to visual inputs $(V)$ , which is in line with their pretraining on AudioSet. Interestingly, this trend is reversed for most foundation models, which seem to be biased towards visual inputs, with unimodal video performance $(V)$ being substantially higher than unimodal audio performance $(A)$ . + +Takeaway 4 Foundation models perform comparably to finetuned embedding models. Embedding models more heavily rely on audio cues than on visual ones, while foundation models exploit visible cues rather than audible ones, see Tab. 2 + +Modality confusion The modality confusion score $\mu$ shows that, for all models, a notable fraction of test samples $(4 - 11\%)$ were misclassified when an additional modality was included—despite being correctly classified with unimodal input. Furthermore, for all models, a small portion of + +test samples is not solvable multi-modally even though they were solvable in both modalities alone $(\mu_{A\cap V})$ . This insight is made possible by VGGSounder's per-label modality annotations and shows that all models are susceptible to being distracted given an additional modality. This is a concerning issue for multi-modal models since they should preserve unimodal capabilities when adding a second modality. Being able to evaluate this behaviour on the VGGSounder benchmark is a first step towards enabling the development of mitigation strategies, eventually resulting in stronger audio-visual models. + +Takeaway 5 Our modality confusion score reveals that all models are negatively impacted by additional modalities for a substantial amount of samples (see Tab. ②). + +Performance across modalities Fig. 4 shows the performance profiles across modalities. At first glance, we can see that VideoLLaMA-2's performance is well balanced for different input modalities, while models from the Gemini family distinctly underperform on audio inputs. In contrast, embedding models exhibit a moderate balance across modalities, with DeepAVFusion and EquiAV showing slight underperformance for visual input. + +As Fig. 4 also illustrates, profiling of this kind is enabled through the modality annotations in VGGSounder. In contrast, VGGSound assumed that all classes are perceivable in both modalities, and did not account for background sounds. This resulted in consistent under-evaluation of foundation models (that were not finetuned on VGGSound) for audio inputs. + +In addition to the radar plot in Fig. 4, we provide results on VGGSound in Appendix D showing that all models have substantially lower performance than their hit scores in Tab. 2. This confirms that many model predictions were incorrectly flagged as false positives in VGGSound due to the incomplete ground-truth labels, painting a distorted picture of models' limitations. + +![](images/cc5cecfd4791f27111ac0dae1e53c38773b7a815acb789f7c7016fb2e18df776.jpg) +Figure 4. VGGSounder more accurately captures model performance across input modalities. We show the Hit score on VGGSounder and accuracy on VGGSound, normalised by the per-model maximum performance on each benchmark. Specifically for foundation models, we observe a significant difference in performance between VGGSound and VGGSounder. + +Takeaway 6 VGGSounder's more complete ground-truth labels allow for more accurate, modality-specific profiling of model performance (see Fig.4). + +# 5.2. Performance analysis using meta-classes + +VGGSounder includes annotations of three meta-classes for each sample: background music indicates whether samples contain music without a visible source, voice over similarly marks speech without a visible source, and static image(s) flags that the visual stream consists of one or only a few static visual frames. + +These new meta-classes allow us to evaluate the model behaviour in challenging scenarios where information from one modality dominates. We consider the performance difference between samples that do contain a meta-label and samples that do not contain it. Positive numbers indicate that the models perform better on the subset with metal-labels. In Tab.3, we summarise the main findings, focussing on $F_{1}$ -score as the most balanced multi-label metric. Additional results are provided in Appendix D + +Background music All models perform worse on video samples containing background music. This indicates that it is challenging to decouple background audio from the rest of the video. The evaluated models are not good at differentiating between sound sources without visual cues, e.g. predicting different instruments in the background music. + +Voice over In contrast to background music, we observe a clear difference between embedding models and foundation models for samples with voice-over. While the audio classification performance of embedding models drops substantially. This drop is only slight for VideoLLaMA-2 and Unified-IO 2, and the performance of other foundation models even improves. This indicates that the foundation models are less distracted by voice-over. + +Static image(s) The impact of static images is more nuanced: First, audio classification performance improves for embedding models while it decreases for the foundation models. This shows that the purpose-built, VGGSoundfinetuned embedding models can more accurately predict specific sounds in the absence of other cues. + +Second, visual classification performance on static images drops for all models, suggesting that models rely on rich temporal cues to make accurate predictions. + +Third, the subset accuracy (i.e. exact label set matches) for uni-modal predictions shows that foundation models perform better in visual than audio classification with and without static images. In contrast, embedding models perform better in audio classification on static images but worse in visual classification for other samples. This suggests that non-static samples form a challenging subset for embedding models, where a model needs to favour one modality over another one to make a correct prediction. + +Samples without any meta-label When comparing the model performance on samples without background music, static images, or voice-over annotations (column neither in Tab. 3) to the performance on the entire test set, we see a performance gain (here negative numbers indicate an increased $\mathbf{F}_1$ score). This finding concludes that these three categories form challenging subsets of the dataset. + +Takeaway 7 Samples with background music, static images(s), and voice over provide distinct challenges for each model (see Tab.3). This highlights VGGSounder's value for comprehensive model evaluation. + +# 5.3. Impact of VGGSounder labels + +Our relabelling pipeline adds two types of labels to those in VGGSound: (1) automatically generated labels based on synonymous classes and sub-/superclasses, and (2) human-curated labels. In Tab.4, we ablate the impact of each type of added labels in terms of the relative performance gains (Hit score). A complete breakdown of the effects across all + +
Modelbackground musicvoice over ΔF1static image(s)neither ΔF1
ΔF1ΔF1ΔF1Sub. Acc. w/Sub. Acc. w/oavav
avavaavavavavav
Embedding Models
CAV-MAE-0.66-0.59-0.65-1.130.31-0.2812.7319.2211.9819.21-1.00-1.27-1.24
DeepAVFusion-0.71-0.80-0.70-1.250.28-0.189.7910.819.3010.81-1.27-1.01-1.36
Equi-AV-0.79-0.42-0.41-0.970.28-0.2311.1210.4510.4910.45-0.97-0.97-0.89
AV-Siam-0.73-0.72-0.82-1.080.34-0.2612.3419.5311.5719.52-1.01-1.34-1.39
Closed-source Foundation Models
Gemini 1.5 Flash-0.22-0.41-0.692.08-0.36-0.251.6614.351.6814.391.40-1.10-1.62
Gemini 1.5 Pro-0.36-0.61-0.942.48-0.31-0.242.8320.852.8720.801.64-0.89-1.09
Gemini 2.0 Flash-0.09-0.32-0.580.030.13-0.261.7012.331.5312.40-0.03-1.07-1.22
Open-source Foundation Models
VideoLLaMA 2-0.47-0.72-0.84-0.540.27-0.3112.5519.6412.0419.70-0.47-1.26-1.33
Unified-IO 2-1.200.18-0.65-0.680.12-0.2111.3911.8910.8812.00-1.07-0.15-1.17
PandaGPT-1.13-0.15-0.420.56-0.22-0.162.944.272.914.24-0.60-0.15-0.65
OLA-2.160.09-0.441.50-0.48-0.1813.038.8812.888.89-0.80-0.13-0.06
+ +Table 3. Summary of performance differences in the presence/absence of meta-classes. Difference in $\mathbf{F}_1$ scores ( $\Delta \mathbf{F}_1$ ) for audio-visual video classification on VGGSounder between videos with a meta-class and those without it. Positive numbers ( $\Delta$ ) indicate better performance when the meta-class is present. Additional results are provided in Appendix D + +
ModelHuman labels ↑Auto labels ↑
AVAVAVAV
Gemini 1.5 Flash29.2814.5916.360.480.931.51
Gemini 1.5 Pro28.6125.5227.630.311.992.10
Gemini 2.0 Flash8.8012.1611.130.221.121.28
+ +Table 4. Impact of added labels using different strategies in VGGSounder. We show the change in multi-label classification performance ( $\Delta$ Hit) when adding automatically added (Auto) or human-annotated (Human) labels to VGGSound, and compare to the original VGGSound data. + +models and metrics is provided in Appendix D. While performance is consistently higher with automatically added labels (Auto), the increase is noticeably smaller than that for human-curated labels. Paired with the observation that models do frequently predict correct classes that were not part of the original VGGSound label set, this indicates that human-curated labels better cover the ground truth. + +Takeaway 8 Automatically added labels are an important step, but human-curated labels have a bigger effect on eliminating incorrectly flagged false positives, underscoring the value of accurate human annotation. + +# 6. Discussion + +Choice of VGGSound as base dataset VGGSound is commonly used to evaluate audio-visual models on the multimodal video classification task. As it is currently the most suitable testbed for audio-visual classification tasks (due to its size, diversity of categories, non-constrained setting, and relatively strong audio-visual correspondence), it serves as an optimal starting point for our substantially improved VGGSounder benchmark with a multi-label evaluation pro + +tocol for foundation models that makes the benchmark suitable for meaningful evaluation. + +Multi-label vs single-label classification Video content is inherently complex, often containing multiple co-occurring objects and actions both within and across modalities. This makes it unlikely that a given clip belongs to just one class as is the case in the single-label classification task. Therefore, our VGGSounder extends the VGGSound test set to multi-label classification. Unlike models trained on a narrow single-label dataset, foundation models develop versatile representations. Multi-label evaluation thus also aligns very naturally with this capability. + +# 7. Conclusion + +We introduced an modality-aware evaluation set for audiovisual foundation models. VGGSounder builds on the widely used VGGSound dataset by adding: (a) comprehensive human annotations for missing classes, (b) specifying modality information per label, (c) introducing specialised meta-labels for frequently occurring real-world challenges, and (d) using heuristic methods to improve label quality. + +Through our newly introduced metric, modality confusion, we observe that incorporating additional modalities does not necessarily yield better results. Models often become more confused on a substantial subset of test samples. Furthermore, finetuned embedding models tend to rely heavily on audio cues, while foundation models depend more on visual information. Additionally, our meta-label analysis highlights distinct challenges across various specialised yet commonly occurring scenarios such as background music, static images, and voice-overs. Overall, we hope that the VGGSounder benchmark will advance the evaluation and development of foundational audio-visual models. + +# Acknowledgements + +The authors would like to thank Felix Förster, Sayak Mallick, and Prasanna Mayilvahananan for their help with data annotation, Felix Förster and Monica Riedler for proofreading the paper, and Thomas Klein and Shyamgopal Karthik for their help in setting up MTurk. They also thank numerous MTurk workers for labelling. + +This work was in part supported by the BMBF (FKZ: 01IS24060, 01I524085B, 01IS18039A), the DFG (SFB 1233, project number: 276693517), and the Open Philanthropy Foundation funded by the Good Ventures Foundation. WB acknowledges financial support via an Emmy Noether Grant funded by the German Research Foundation (DFG) under grant no. BR 6382/1-1. WB is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This research utilised compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG. The authors thank the IMPRS-IS for supporting TW. + +# References + +[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. +[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE TPAMI, 2018. +[3] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In ECCV, 2020. +[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP, 2020. +[5] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In ECCV, 2020. +[6] Humam Alwassel, Dhruv Mahajan, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. +[7] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In CVPR, 2024. +[8] Relja Arandjelović and Andrew Zisserman. Look, listen and learn. In ICCV, 2017. +[9] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV, 2018. +[10] Yuki M. Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020. +[11] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. In NeurIPS, 2016. + +[12] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020. +[13] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020. +[14] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Audio-visual synchronisation in the wild. In BMVC, 2021. +[15] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In CVPR, 2021. +[16] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 2020. +[17] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022. +[18] Yanbei Chen, Yongqin Xian, A. Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021. +[19] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. In CVPR, 2025. +[20] Ying Cheng, Ruize Wang, Zhihao Pan, Rui Feng, and Yuejie Zhang. Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In ACM MM, 2020. +[21] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. +[22] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. 2023. +[23] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016. +[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. +[25] Joshua P Ebeneze, Yongjun Wu, Hai Wei, Sriram Sethuraman, and Zongyi Liu. Detection of audio-video synchronization errors via event detection. In ICASSP, 2021. +[26] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In ECCV, 2020. +[27] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In ICCV, 2019. +[28] Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad + +Reza Ghasemi Madani, Claire Barale, Robert McHardy, Joshua Harris, Jean Kaddour, Emile van Krieken, and Pasquale Minervini. Are we done with mmlu? arXiv preprint arXiv:2406.04127, 2024. +[29] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017. +[30] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. +[31] Shir Goldstein and Yael Moses. Guitar music transcription from silent video. In BMVC, 2018. +[32] Yuan Gong, Alexander H Liu, Andrew Rouditchenko, and James Glass. Uavm: Towards unifying audio and visual models. IEEE Signal Processing Letters, 2022. +[33] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James Glass. Contrastive audio-visual masked autoencoder. In ICLR, 2023. +[34] Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594, 2024. +[35] Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. Aligning ai with shared human values. In ICLR, 2021. +[36] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021. +[37] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022. +[38] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022. +[39] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, 2024. +[40] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. IJCV, 2019. +[41] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. +[42] Naji Khosravan, Shervin Ardeshir, and Rohit Puri. On attention modules for audio-visual synchronization. In CVPR Workshop, 2019. +[43] Jongsuk Kim, Hyeongkeun Lee, Kyeongha Rho, Junmo Kim, and Joon Son Chung. Equipav: Leveraging equivariance for audio-visual contrastive learning. In ICML, 2024. +[44] A. Sophia Koepke, Olivia Wiles, and Andrew Zisserman. Visual pitch estimation. In SMC, 2019. + +[45] A Sophia Koepke, Olivia Wiles, Yael Moses, and Andrew Zisserman. Sight to sound: An end-to-end approach for visual piano transcription. In ICASSP, 2020. +[46] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurlPS, 2018. +[47] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song. Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning. In ICCV, 2021. +[48] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022. +[49] Yan-Bo Lin and Gedas Bertasius. Siamese vision transformers are scalable audio-visual learners. In ECCV, 2024. +[50] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In ACCV, 2020. +[51] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP, 2019. +[52] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025. +[53] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In CVPR, 2024. +[54] Jie Ma, Min Hu, Pinghui Wang, Wangchun Sun, Lingyun Song, Hongbin Pei, Jun Liu, and Youtian Du. Look, listen, and answer: Overcoming biases for audio-visual question answering. NeurIPS, 2024. +[55] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. +[56] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In ECCV, 2022. +[57] Shentong Mo and Pedro Morgado. Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. In CVPR, 2024. +[58] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. In BMVC, 2020. +[59] Arsha Nagrani, Joon Son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision. In ICASSP, 2020. +[60] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In NeurIPS, 2021. +[61] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. arXiv preprint arXiv:2104.02687, 2021. + +[62] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018. +[63] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In ECCV, 2016. +[64] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Learning sight from sound: Ambient sound provides supervision for visual learning. IJCV, 2018. +[65] Mandela Patrick, Yuki M Asano, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In NeurIPS, 2020. +[66] KR Prajwal, Liliane Momeni, Triantafyllos Afouras, and Andrew Zisserman. Visual keyword spotting with attention. In BMVC, 2021. +[67] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In ECCV, 2020. +[68] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICLR, 2023. +[69] Kun Su, Xiulong Liu, and Eli Shlizerman. Multi-instrumentalist net: Unsupervised generation of music from body movements. arXiv preprint arXiv:2012.03478, 2020. +[70] Kun Su, Xiulong Liu, and Eli Shlizerman. How does it sound? In NeurIPS, 2021. +[71] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In ICML, 2024. +[72] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. +[73] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. +[74] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. +[75] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel PW Ellis, and John R Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In ICLR, 2021. +[76] Ilpo Viertola, Vladimir Iashin, and Esa Rahtu. Temporally aligned audio for video with autoregression. In ICASSP, 2025. +[77] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018. +[78] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In CVPR, 2019. +[79] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020. + +[80] Haoming Xu, Runhao Zeng, Qingyao Wu, Mingkui Tan, and Chuang Gan. Cross-modal relation-aware networks for audio-visual event localization. In ACM MM, 2020. +[81] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In ECCV, 2024. +[82] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. +[83] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audio-visual question answering on videos. In ACM MM, 2022. +[84] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021. +[85] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In ECCV, 2018. +[86] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In ICCV, 2019. +[87] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019. +[88] Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L. Berg. Visual to sound: Generating natural sound for videos in the wild. In CVPR, 2018. +[89] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. +[90] Lingyu Zhu and Esa Rahtu. V-slowfast network for efficient visual sound separation. In WACV, 2022. \ No newline at end of file diff --git a/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/images.zip b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8d6722430406cb17fca92dbcc0fcfc3503f9c50c --- /dev/null +++ b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d629bdb083d5224ee972240e7290c242ece71373e6064c758490cfd4f1ff9a5b +size 345399 diff --git a/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/layout.json b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7bdb070312546cb070cb3789e71e41000ce19a46 --- /dev/null +++ b/2025/VGGSounder_ Audio-Visual Evaluations for Foundation Models/layout.json @@ -0,0 +1,9003 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 100, + 118, + 120 + ], + "blocks": [ + { + "bbox": [ + 96, + 100, + 118, + 120 + ], + "lines": [ + { + "bbox": [ + 96, + 100, + 118, + 120 + ], + "spans": [ + { + "bbox": [ + 96, + 100, + 118, + 120 + ], + "type": "image", + "image_path": "68614e85b835a66db0667fcd63c2b5cc77b7b39f4a49010171318974b8b9b39b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 103, + 515, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 515, + 119 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 515, + 119 + ], + "type": "text", + "content": "VGGSounder: Audio-Visual Evaluations for Foundation Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 142, + 455, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 142, + 455, + 158 + ], + "spans": [ + { + "bbox": [ + 154, + 142, + 455, + 158 + ], + "type": "text", + "content": "Daniil Zverev1,* Thaddäus Wiedemer2,3,4,* Ameya Prabhu2,3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "type": "text", + "content": "Matthias Bethge" + }, + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "type": "text", + "content": " Wieland Brendel" + }, + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "type": "text", + "content": " A. Sophia Koepke" + }, + { + "bbox": [ + 149, + 160, + 460, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "spans": [ + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "text", + "content": "Technical University of Munich, MCML " + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "text", + "content": "University of Tübingen " + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "text", + "content": "Tübingen AI Center " + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 79, + 182, + 532, + 217 + ], + "type": "text", + "content": "MPI for Intelligent Systems, ELLIS Institute Tübingen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 221, + 543, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 221, + 543, + 234 + ], + "spans": [ + { + "bbox": [ + 63, + 221, + 543, + 234 + ], + "type": "text", + "content": "{daniil.zverev,a-sophia.koepke}@tum.de,{thaddaeus.wiedemer,ameya.prabhu}@uni-tuebingen.de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 261, + 200, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 261, + 200, + 274 + ], + "spans": [ + { + "bbox": [ + 151, + 261, + 200, + 274 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "content": "The emergence of audio-visual foundation models underscores the importance of reliably assessing their multimodal understanding. The VGGSound dataset is commonly used as a benchmark for evaluation audio-visual classification. However, our analysis identifies several limitations of VGGSound, including incomplete labelling, partially overlapping classes, and misaligned modalities. These lead to distorted evaluations of auditory and visual capabilities. To address these limitations, we introduce VGGSounder, a comprehensively re-annotated, multi-label test set that extends VGGSound and is specifically designed to evaluate audio-visual foundation models. VGGSounder features detailed modality annotations, enabling precise analyses of modality-specific performance. Furthermore, we reveal model limitations by analysing performance degradation when adding another input modality with our new modality confusion metric. Our dataset and project page are available at https://vqgsounder.github.io/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 548, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 548, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 548, + 295, + 693 + ], + "type": "text", + "content": "Rigorous evaluation benchmarks have been instrumental in assessing the effectiveness of audio-visual models [33, 43, 49, 57]. Specifically, multi-modal foundation models integrating visual and auditory data aim to achieve a holistic understanding of audio-visual content. However, the field lacks large-scale modality-aware classification benchmarks with ground-truth annotations indicating whether each label is visible, audible, or both. Such annotations would allow detailed evaluations of multi-modal model capabilities. To address this gap, we introduce VGGSounder, an enhanced version of the widely-used audio-visual classification dataset VGGSound [13], which facilitates modality-" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 261, + 553, + 452 + ], + "blocks": [ + { + "bbox": [ + 317, + 261, + 553, + 452 + ], + "lines": [ + { + "bbox": [ + 317, + 261, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 553, + 452 + ], + "type": "image", + "image_path": "2b489317aa6ca2aff9ddd7b4f939ceccbb63d7e5eb5c60b2e953cee0770ab8f8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 461, + 555, + 551 + ], + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 551 + ], + "type": "text", + "content": "Figure 1. We introduce VGGSounder, a multi-label audio-visual classification benchmark with modality annotations. We extend the original VGGSound test set with human-annotated audible, visible, and visible+audible labels. We add meta labels for common confounders, such as background music. We benchmark eleven recent audio-visual models on VGGSounder. It enables selective analysis of a model's auditory and visual capabilities on classes relevant for the queried modality." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 567, + 527, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 527, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 527, + 579 + ], + "type": "text", + "content": "aware evaluation of audio-visual foundation models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "content": "VGGSound has several notable limitations. First, its data is inherently multi-label; for instance, a single sample may simultaneously include labels such as playing drum kit and playing acoustic guitar when multiple instruments are present. Additionally, evaluating how different modalities contribute to model performance becomes difficult without explicit modality annotations, as some labels are either not visually present or not audible (e.g., certain instruments might only be audible but not visible in advertisements). Moreover, overlapping label classes present another challenge; for example," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 132, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 132, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 132, + 712 + ], + "type": "text", + "content": "*equal contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1027" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "the orchestra label often coincides with labels for individual instruments. These issues result in systematic under-evaluation of multi-modal audio-visual foundation models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 296, + 407 + ], + "type": "text", + "content": "To overcome these limitations, we present VGGSounder, an improved benchmark inspired by similar advancements in other domains [12, 28]. We re-announce the dataset to create a comprehensive multi-label classification setting by collecting detailed annotations for each sample, including (1) additional classes present, (2) explicit modality annotations to label modality misalignment, (3) metadata indicating the presence of background music, voice-over, or static images, and (4) merging of classes to address overlapping classes. Consequently, VGGSounder provides a robust, foundation-model-ready benchmark enabling structured analysis of whether models rely on audio or visual cues. Furthermore, we include meta-labels (e.g., background music, voice-over, or static images) to easily filter out unreliable labels during evaluation. Utilising VGGSounder, we evaluate audio-visual foundation models, demonstrating their poor performance on our benchmark. We find that the state-of-the-art, closed-source Gemini models consistently rely exclusively on the visual modality. In addition to that, we measure the modality confusion, i.e. when models get distracted by an additional input modality, which exposes the unsuccessful merging of modalities. These findings highlight the importance of the audio-focused VGGSounder benchmark as a critical tool for accurately assessing audio-visual foundation models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 422, + 216, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 422, + 216, + 434 + ], + "spans": [ + { + "bbox": [ + 63, + 422, + 216, + 434 + ], + "type": "text", + "content": "We make the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 434, + 288, + 531 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 63, + 434, + 272, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 434, + 272, + 445 + ], + "spans": [ + { + "bbox": [ + 63, + 434, + 272, + 445 + ], + "type": "text", + "content": "1. We illustrate limitations of VGGSound in Sec. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 446, + 288, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 288, + 469 + ], + "type": "text", + "content": "2. We curate VGGSounder with multi-modal human annotations for multi-label classification in Sec. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 470, + 288, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 288, + 506 + ], + "type": "text", + "content": "3. We evaluate state-of-the-art audio-visual models, observing differences between embedding models and autoregressive foundation models in Sec. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 506, + 288, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 506, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 506, + 288, + 531 + ], + "type": "text", + "content": "4. We propose new metrics to quantify the negative impact of using multiple input modalities in Sec. 5." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 546, + 139, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 139, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 139, + 559 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 700 + ], + "type": "text", + "content": "Audio-visual learning Many prior works consider audio-visual tasks that include sound source localisation and separation [3, 5, 9, 15, 27, 56, 62, 67, 75, 80, 85, 86, 90], event localisation [50, 51, 74, 78], audio-visual question answering [48, 54, 83, 84], audio-visual synchronisation [14, 23, 25, 38, 39, 42], audio synthesis using visual information [19, 26, 31, 44, 45, 61, 69-71, 87], or audio-driven face image synthesis [7, 40, 77]. Audio-visual data has also been leveraged for speech-related tasks, including speech and speaker recognition [2, 4, 59], or the spotting of spoken keywords [58, 66]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Furthermore, the natural alignment between audio and" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 157 + ], + "type": "table", + "html": "
Dataset# Clips# ClassesMulti- labelModality labelsAnnotation pipeline
Flickr-SoundNet [10]2M-✘✘-
Kinetics-Sound [7]18.8k34✘✘MTurk
AudioSet [26]2.1M537✓✘Manual
AVE [62]4k28✓✘Manual
VEGSAS [76]132k10✘✘MTurk
Visually Aligned Sounds [15]13k8✘✘MTurk
VGGSound [12]200k309✘✘Classifiers+Manual
VGGSound-Sparse [32]7.1k12✘✘Manual
Visual Sound [66]91k309✘✘ImageBind [30]
VGGSounder15.4k309✓✓MTurk
", + "image_path": "3086e0733e283efca0637873a1cab8fd54187d2014223a4b780e6becda7f20a0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 165, + 549, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 165, + 549, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 165, + 549, + 175 + ], + "type": "text", + "content": "Table 1. Comparison of audio-visual classification benchmarks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 191, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 553, + 312 + ], + "type": "text", + "content": "video has been exploited to learn improved audio-visual embeddings for downstream tasks [6, 10, 11, 18, 20, 21, 46, 60, 63, 65, 79]. Using both modalities jointly generally leads to performance boosts over using one modality in isolation. We examine this observation closely and aim to evaluate the effective use of multiple input modalities for the video classification task. To enable this, we propose — to the best of our knowledge, the first multi-label video classification benchmark that includes per-modality annotations for every sample (see Tab. 1)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 315, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 553 + ], + "type": "text", + "content": "Audio-visual foundation models Recently, multi-modal general-purpose models have emerged that can handle diverse downstream tasks without task-specific finetuning - also referred to as multi-modal foundation models. For instance, images or language were used as the bridge between modalities including audio, image, and text [30, 89]. Building on this, PandaGPT [72] leverages Vicuna [22] and ImageBind's embedding space to train a general multi-modal model exclusively on image-text pairs. Unified-IO 2 [53] employs universal tokenisation to process audio, video, and text. VideoLLaMA2 [21] uses a Spatial-Temporal Convolution connector in the visual branch before projecting audio and visual information into the LLM input space. The recently introduced Ola model [52] advances omnimodal processing through progressive modality alignment, using video to bridge audio and visual information. The Gemini models [73] are closed-source multi-modal models that achieve impressive performance on diverse downstream tasks. We use VGGSounder to benchmark the audio-visual capabilities of the aforementioned models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 557, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 652 + ], + "type": "text", + "content": "Audio-visual classification benchmarks Audio-visual classification is distinct from general video classification (e.g. on YouTube-8M [1]), as classes typically cover both audible and visible actions or events. Commonly used datasets for audio-visual classification include KineticsSound [8] sourced from the Kinetics dataset [41], Flickr-SoundNet [11] scraped from Flickr, and AudioSet [29] and VGGSound [13], both sourced from YouTube." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "Kinetics-Sound features manual labels of human actions, but covers only 34 classes. Flickr-SoundNet is much larger, but only a small subset is labelled. Similarly, only a small fraction of the roughly 2M AudioSet samples are annotated and have aligned audio and video." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1028" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Limitations of VGGSound. We show video frames from videos in the VGGSound test set along with their annotated label (grey) to demonstrate various limitations. A. VGGSound samples are labelled with a single class, yet many videos contain multiple distinct classes. B. Additionally, many classes partially overlap or are ambiguous. C. Some samples are labelled with classes that are not present in one of the modalities (i.e., the labelled class is not visible or audible)." + } + ] + } + ], + "index": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 293, + 294, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 294, + 352 + ], + "type": "text", + "content": "In contrast, VGGSound ensures audio-visual correspondence for around 200 000 samples and was curated with an automatic pipeline involving class-list generation, and auditory and visual content verification. The visual verification step ensures that a class is represented in the centre frame." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 352, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 295, + 460 + ], + "type": "text", + "content": "The VEGAS dataset [88] provides better quality assurances for a small subset of AudioSet with only 10 classes. Visually Aligned Sounds [16] subsamples VEGAS and AudioSet after human verification, and Visual Sound [76] subsamples VGGSound using a multi-modal embedding model, both aiming for high audio-visual correspondence. Similarly, VGGSound-Sparse [37] is a subset of VGGSound with a focus on temporally and spatially sparse synchronisation signals (e.g., short loud noises)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 519 + ], + "type": "text", + "content": "Overall, VGGSound strikes the best balance between size, generality, and annotations, making it a common benchmark for audio-visual classification. We update VGGSound to sustain its usability for the development of the next generation of multi-modal foundation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 206, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 206, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 206, + 542 + ], + "type": "text", + "content": "3. Limitations of VGGSound" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 549, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 295, + 622 + ], + "type": "text", + "content": "Since we are interested in the VGGSound dataset for benchmarking audio-visual multi-modal models, our analysis focuses on the VGGSound test set, which consists of 15446 video clips, each 10s long and labelled with one of 309 classes. We qualitatively identify several limitations of the VGGSound annotations outlined below and in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "text", + "content": "Co-occurring classes While VGGSound's visual verification aimed to minimise multiple classes co-occurring in a clip, we find that most samples nevertheless clearly contain multiple classes, see Fig. 2A. In some cases, classes are temporally separated, e.g., showing male speech, man speaking and then cutting to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 293, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 402 + ], + "type": "text", + "content": "footage of firing cannon. Most often, classes co-occur at the same time, sometimes for the entire duration of the video clip. Overlapping classes are often related, such as different instruments in a band or orchestra, but can also be entirely unrelated, e.g., donkey, ass braying co-occurring with playing violin. As additional empirical evidence, we provide a co-occurrence matrix computed using CAV-MAE [33], a state-of-the-art audio-visual model, in Appendix D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 598 + ], + "type": "text", + "content": "Overlapping classes The issue of co-occurring classes is exacerbated by many of the 309 automatically generated classes partially overlapping in their definition, as illustrated in Fig. 2B. We found two pairs of synonymous classes: timpani and tympani and dog barking and dog bow-wow. Additionally, some classes are strict subclasses of others, such as the gender-specific versions of cattle moing: cow lowing and bull bellowing; or the more specific variants of people eating: people eating noodle and people eating apple. Finally, several classes commonly appear together, such as playing snare drum which is often played as part of a drum kit, or semantically similar concepts: running electric fan and air conditioning noise, and sloshing water and splashing water." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Modality misalignment Despite VGGSound's auditory and visual content verification, we find that many of the annotated classes are not visible or not audible, as shown in Fig. 2C. A large fraction of videos contains background music, voice-over and narration, or other background sounds like bird chirping, tweeting or cricket chirping without a visible source. Similarly, some videos contain visible but inaudible cues for classes like sea waves. Static images and slide shows accompa" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 274, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 274, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 274, + 714 + ], + "type": "text", + "content": "1 Although these issues most likely also apply to the training set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1029" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 208, + 214 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 208, + 214 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 208, + 214 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 208, + 214 + ], + "type": "image", + "image_path": "7c9ba699608f529a550d998b95fefef92562d56bd4f6c77fefb12df0bb131ed6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 269 + ], + "type": "text", + "content": "Figure 3. Overview of VGGSounder. A. Most samples contain more than one label. B. More than a quarter of labels are audible but not visible. In contrast, only a tiny fraction is visible but not audible. C. Speech and bird sounds are the most common classes; more details can be found in Appendix B. D. Forty percent of the samples contain some combination of background music, voice over, and static image(s), making the classification task significantly harder." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 98, + 315, + 190 + ], + "blocks": [ + { + "bbox": [ + 222, + 98, + 315, + 190 + ], + "lines": [ + { + "bbox": [ + 222, + 98, + 315, + 190 + ], + "spans": [ + { + "bbox": [ + 222, + 98, + 315, + 190 + ], + "type": "image", + "image_path": "94fbb8f345cb4997d6a8c853c6e957bfcd248d0cf8614eb593fba6ff5bc8f081.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 325, + 83, + 433, + 134 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 319, + 83 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 319, + 83 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 319, + 83 + ], + "type": "text", + "content": "B Label modality distribution" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 74, + 418, + 82 + ], + "lines": [ + { + "bbox": [ + 342, + 74, + 418, + 82 + ], + "spans": [ + { + "bbox": [ + 342, + 74, + 418, + 82 + ], + "type": "text", + "content": "C Most common labels" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 83, + 433, + 134 + ], + "lines": [ + { + "bbox": [ + 325, + 83, + 433, + 134 + ], + "spans": [ + { + "bbox": [ + 325, + 83, + 433, + 134 + ], + "type": "image", + "image_path": "d3c1a4923a33d3f9b62441bb900fbf905856b1e68189ceeca9616911ae16440e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 442, + 83, + 549, + 206 + ], + "blocks": [ + { + "bbox": [ + 472, + 75, + 518, + 82 + ], + "lines": [ + { + "bbox": [ + 472, + 75, + 518, + 82 + ], + "spans": [ + { + "bbox": [ + 472, + 75, + 518, + 82 + ], + "type": "text", + "content": "D Meta labels" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 83, + 549, + 206 + ], + "lines": [ + { + "bbox": [ + 442, + 83, + 549, + 206 + ], + "spans": [ + { + "bbox": [ + 442, + 83, + 549, + 206 + ], + "type": "image", + "image_path": "c63d15d87efbc62ba4c4c351ed68e88dbdc145a6c81095ccf49958c67c05d52b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 282, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 282, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 295, + 365 + ], + "type": "text", + "content": "nied by music or other sounds are other frequent sources of misaligned modalities. Finally, some classes are misaligned by definition: wind noise is only audible and not visible. Overall, " + }, + { + "bbox": [ + 54, + 282, + 295, + 365 + ], + "type": "inline_equation", + "content": "48.43\\%" + }, + { + "bbox": [ + 54, + 282, + 295, + 365 + ], + "type": "text", + "content": " of the original VGGSound test samples have misaligned modalities. This finding challenges the widely held assumption that VGGSound has strong modality alignment [32, 47]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 366, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 295, + 486 + ], + "type": "text", + "content": "Other datasets, such as Visually Aligned Sounds, Visual Sound, and VGGSound-Sparse (see Sec. 2) omitted samples with misaligned modalities. In contrast, we contend that inaudible or invisible cues are common in natural videos and should be considered when benchmarking multi-modal models. We, therefore, place particular emphasis on crafting reliable modality annotations for all samples, allowing users to evaluate models on samples that guarantee modality alignment, and on those where classes are only visible or only audible (see Tab. 1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 498, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 498, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 61, + 498, + 290, + 548 + ], + "type": "text", + "content": "Takeaway 1 VGGSound suffers from several issues: class co-occurrence not captured by single labels, overlapping class definitions, and modality misalignment, see Fig. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 562, + 187, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 187, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 187, + 574 + ], + "type": "text", + "content": "4. Building VGGSounder" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "We propose a series of fixes for VGGSound's issues, ultimately resulting in the updated VGGSounder benchmark. We are not the first aiming to future-proof an existing benchmark: [12] analysed shortcomings of ImageNet [24], ultimately proposing switching to a multi-label classification task with additional manual labels. [28] similarly reannotated samples in MMLU [35, 36] to fix labelling errors. Both works inspire our approach to improve VGGSound. To deal with co-occurring classes, we switch to a multi-label classification setting. This effectively handles most overlapping class definitions: a strong model can assign a" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 282, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 553, + 330 + ], + "type": "text", + "content": "high probability to multiple classes, even if they partially overlap. This also allows us to ensure that synonymous classes, as well as subclasses and their superclass, always appear together in the ground-truth labels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 331, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 427 + ], + "type": "text", + "content": "To deal with modality misalignment, we add a modality annotation to each label. For example, we can label a video as containing people clapping in the audio and containing playing volleyball in the audio and video data. We also add meta-labels to indicate whether a sample contains background music, voice over, or static image(s) to optionally treat these cases separately during evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 429, + 553, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 553, + 452 + ], + "type": "text", + "content": "We employ a pipeline similar to [12] to annotate multiple labels per sample, which we outline below." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 552 + ], + "type": "text", + "content": "Collecting proposals We create a gold standard reference set by labelling a small, randomly selected subset of VGGSound test samples with four in-house computer vision experts. The interface used for this first labelling step is shown in Appendix A. We extend the subset until each class is covered at least once, leading to a final size of 417 samples. Labels from different annotators are merged via a simple majority vote." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 554, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 554, + 662 + ], + "type": "text", + "content": "Given the gold standard set, we want to find a solid strategy for automatically generating label proposals which are shown to the humans labelling the test set. This should have a recall greater than " + }, + { + "bbox": [ + 313, + 554, + 554, + 662 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 554, + 554, + 662 + ], + "type": "text", + "content": " while maximising precision compared to the gold standard labels to produce a small set of proposals with good label coverage. Our final strategy combines predictions from several state-of-the-art models with a manual heuristic to obtain " + }, + { + "bbox": [ + 313, + 554, + 554, + 662 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 313, + 554, + 554, + 662 + ], + "type": "text", + "content": " recall for an average of 30 proposals per sample, see Appendix A." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Human labelling We use Amazon Mechanical Turk to re-associate the entire VGGSound test set. For each sample, we first ask annotators to indicate whether the video contains background music, voice over, or" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1030" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "static image(s). Then, annotators are asked to indicate for each label proposal whether the class is audible and/or visible. Finally, annotators can add a class if it is missing from the proposals. Annotators were paid the US minimum wage; the interface used is shown in Appendix A. We let annotators label the samples in batches of 20, each containing two gold standard samples as catch trials. We reject and re-associate all batches with a catch trial " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "-score below " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ". In addition, we obtain modality annotations for the original VGGSound labels and meta-classes. Further details are provided in Appendix A." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 294, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 294, + 278 + ], + "type": "text", + "content": "Final labels We merge all obtained annotations using majority voting. Additionally, we automatically add synonymous classes and superclasses for a given subclass, e.g., we add cattle moing whenever cow lowing is in the set of labels. The full set of classes added this way is described in Appendix A." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 290, + 290, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 290, + 290, + 340 + ], + "spans": [ + { + "bbox": [ + 61, + 290, + 290, + 340 + ], + "type": "text", + "content": "Takeaway 2 We develop VGGSounder: A multi-label video classification benchmark extending VGGSound with human-annotated multi-labels, modality annotations, and meta-labels as summarised in Fig. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 251, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 251, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 251, + 366 + ], + "type": "text", + "content": "5. Benchmarking audio-visual models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 409 + ], + "type": "text", + "content": "We use VGGSounder to benchmark four popular audiovisual embedding models and seven foundation models, and analyse their auditory and visual capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 446 + ], + "type": "text", + "content": "Models We evaluate the audio-visual embedding models CAV-MAE [33], DeepAVFusion [57], AV-Siam [49], and Equi-AV [43]. Those were finetuned on VGGSound." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 447, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 555 + ], + "type": "text", + "content": "We benchmark several models from the closed-source Gemini family [73] in a zero-shot evaluation protocol. Furthermore, we use LLM-assisted evaluation to evaluate the following four open-source autoregressive foundation models: VideoLLaMA-2 [21], Unified-IO-2 [53], Panda-GPT [72], and Ola [52]. All models are evaluated in three modes: using unimodal-audio, unimodal-visual, or multi-modal (audio and visual) inputs. Further details about models and their evaluation are provided in Appendix C.1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": "Metrics To benchmark the models on VGGSounder, we use multi-label classification metrics. For embedding models, all metrics are computed for the top-" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " predictions, with " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "k \\in \\{1, 3, 5, 10\\}" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": ". In contrast, prompting foundation models yields an unordered set of class predictions of varying size, and we compute only a single metric using the entire set. As a result, metrics are not directly comparable between embedding and foundation models. To get a sense of their relative performance, we report metrics for embedding models for " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " in the main text, matching the median number of predictions per sample for the foundation models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "For open-source models such as VideoLLaMA-2, Ola, Unified-I0-2, and Panda-GPT, we employ LLM-assisted" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "evaluation [55, 81], in which the Qwen-3 model [82] is tasked to assess the correspondence between model outputs and target classes. Closed-source models from the Gemini family are evaluated by providing the full list of 309 classes as input. Further details on the evaluation procedures and exact prompts are provided in Appendix C" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "content": "Subset accuracy compares the predicted label set to the ground-truth label set and reports the fraction of samples for which they match. This is our strictest metric." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 180, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 553, + 204 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 314, + 180, + 553, + 204 + ], + "type": "text", + "content": "-score is the harmonic mean of precision and recall. It is strictly larger than the subset accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 204, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 263 + ], + "type": "text", + "content": "Hit reports the fraction of samples for which any of the predicted labels are part of the ground-truth label set. This is the most lenient metric which is strictly larger than the " + }, + { + "bbox": [ + 313, + 204, + 553, + 263 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 313, + 204, + 553, + 263 + ], + "type": "text", + "content": " score. We include this metric for ease of comparison to the \"ReaL-Accuracy\" used in [12]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": "All metrics are computed separately for each input modality (audio, video, and audio-visual) and label modality. We use lowercase symbols a, v, and av to indicate the input modality: audio-only, visual-only, or audio-visual inputs, respectively. For label modality, we use uppercase symbols " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "AV" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": ", referring to the subsets of the benchmark with audible, visible, and audio-visual labels. We further include " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "A \\neg V" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": " (audible but not visible) and " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "V \\neg A" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": " (visible but not audible) to analyze unaligned cues. For clarity, we define shorthand notations such as " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "a = a(A)" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": " to denote the model's performance on the audible subset " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": " using only audio input. Analogously, " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "v = v(V)" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "inline_equation", + "content": "av = av(AV)" + }, + { + "bbox": [ + 313, + 264, + 554, + 442 + ], + "type": "text", + "content": " refer to video-only and audio-visual performance on their respective label subsets. Furthermore, we use micro-aggregation to balance the contribution from each class." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "text", + "content": "We additionally measure the negative impact of using multimodal inputs. In particular, " + }, + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "text", + "content": " is a new metric we propose to measure a model's modality confusion " + }, + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "inline_equation", + "content": "(\\mu)" + }, + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "text", + "content": ". We define it as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 484, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 484, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 318, + 484, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\mu_ {\\mathrm {M}} = 1 0 0 \\cdot \\frac {\\sum_ {x \\in M} \\mathcal {I} [ m (x) - \\text {c o r r e c t} \\cap a v (x) - \\text {w r o n g} ]}{N _ {\\text {t o t a l}}}, \\tag {1}", + "image_path": "7e4e4ebc03be7ecffde5712bebd5647a3e33c5d59fbe1f4343cea9d9fce1d813.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "inline_equation", + "content": "M \\in [A, V, A \\cap V]" + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": " and their associated modality inputs are " + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "inline_equation", + "content": "m \\in [a, v]" + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": ", correct/wrong is determined as in the Hit score (with " + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": " for embedding models). " + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "inline_equation", + "content": "N_{total}" + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": " refers to the total number of samples. " + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": " measures the fraction of samples a model correctly classified given an input modality but got wrong when using both modalities simultaneously. We additionally report " + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mu_{A \\cap V}" + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": " as the percentage of samples a model could solve in either modality unimodally but could not solve multi-modally. In other words, the modality confusion " + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 516, + 555, + 661 + ], + "type": "text", + "content": " captures how frequently a model is distracted by an additional input modality, which can indicate the unsuccessful merging of modalities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 671, + 549, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 671, + 549, + 709 + ], + "spans": [ + { + "bbox": [ + 319, + 671, + 549, + 709 + ], + "type": "text", + "content": "Takeaway 3 We propose a new metric, modality confusion " + }, + { + "bbox": [ + 319, + 671, + 549, + 709 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 319, + 671, + 549, + 709 + ], + "type": "text", + "content": ", that measures how frequently a model is distracted by an additional input modality; see Eq. (1)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1031" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 259 + ], + "type": "table", + "html": "
ModelSubset Accuracy ↑F1↑Hit ↑μ↓
avavavava(A-√V)v(V-√A)avavμAμVμA∩V
Embedding Models
CAV-MAE13.1919.2324.4934.4634.9142.6213.9419.0062.2953.4464.173.586.430.77
DeepAVFusion10.1911.1021.5325.3121.2937.3510.3710.5545.7732.6156.273.743.930.17
Equi-AV11.6010.5220.0029.3920.4234.6912.5510.6553.1231.2652.246.977.131.38
AV-Siam12.7919.7522.8333.3035.4139.4312.9018.2160.1954.2059.369.368.803.58
Closed-source Foundation Models
Gemini 1.5 Flash1.7814.4416.4414.4936.9842.5215.6121.6132.7347.3659.1010.224.250.77
Gemini 1.5 Pro3.0520.8622.5319.2649.7353.7417.7322.9035.0369.2375.422.094.850.57
Gemini 2.0 Flash1.8512.5412.6911.8034.0836.456.1918.9018.5143.8347.722.395.431.00
Open-source Foundation Models
VideoLLaMA 212.8619.8524.4738.8747.8252.3520.3428.0858.9152.0259.8012.725.462.95
Unified-IO 211.9411.5625.6135.3127.9248.8921.3816.5354.3931.0565.118.705.161.79
PandaGPT3.194.195.4618.7318.5620.8516.8214.4021.0817.0118.827.595.902.47
OLA14.118.6918.1947.7024.8546.4840.4413.4559.0524.5751.5115.476.802.49
", + "image_path": "7f3c2bebd34a8528a23fab9c42cd1cf3cda07f540249bab807c8564bc3c9e197.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": "Table 2. Audio-visual video classification results on VGGSounder. We report multi-label classification metrics (subset accuracy, " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": "-score, Hit accuracy, modality confusion " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "(\\mu)" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": " for audio- " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "a(A)" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": ", visual - " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "v(V)" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": ", audio-visual - " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "av(AV)" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": ", audio-only - " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "a(A\\neg V)" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": " and video-only - " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "v(V\\neg A)" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": " inputs. The embedding models CAV-MAE, DeepAVFusion, and Equi-AV were finetuned on the VGGSound training set. We report metrics for " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": " here and for other " + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 268, + 555, + 324 + ], + "type": "text", + "content": " in Appendix D The closed sourced multi-modal foundation models Gemini and open-sourced models use a zero-shot evaluation protocol and LLM-assisted protocol respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 336, + 231, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 231, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 231, + 350 + ], + "type": "text", + "content": "5.1. Re-evaluating the state of the art" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": "We present the benchmark performance of state-of-the-art audio-visual models in Tab. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "text", + "content": "Overall performance Unsurprisingly, all models perform best with access to both input modalities (AV). Across all metrics, both open- and closed-source general-purpose models perform comparably to the purpose-built embedding models CAV-MAE and AV-Siam. This indicates that foundation models have reached — and for some modalities exceeded — the performance of specialised models. However, the embedding models finetuned on VGGSound generally have stronger unimodal performance with audio inputs " + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "text", + "content": " compared to visual inputs " + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "text", + "content": ", which is in line with their pretraining on AudioSet. Interestingly, this trend is reversed for most foundation models, which seem to be biased towards visual inputs, with unimodal video performance " + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "text", + "content": " being substantially higher than unimodal audio performance " + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 54, + 380, + 295, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 574, + 290, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 574, + 290, + 637 + ], + "spans": [ + { + "bbox": [ + 61, + 574, + 290, + 637 + ], + "type": "text", + "content": "Takeaway 4 Foundation models perform comparably to finetuned embedding models. Embedding models more heavily rely on audio cues than on visual ones, while foundation models exploit visible cues rather than audible ones, see Tab. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Modality confusion The modality confusion score " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " shows that, for all models, a notable fraction of test samples " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "(4 - 11\\%)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " were misclassified when an additional modality was included—despite being correctly classified with unimodal input. Furthermore, for all models, a small portion of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 337, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 469 + ], + "type": "text", + "content": "test samples is not solvable multi-modally even though they were solvable in both modalities alone " + }, + { + "bbox": [ + 313, + 337, + 555, + 469 + ], + "type": "inline_equation", + "content": "(\\mu_{A\\cap V})" + }, + { + "bbox": [ + 313, + 337, + 555, + 469 + ], + "type": "text", + "content": ". This insight is made possible by VGGSounder's per-label modality annotations and shows that all models are susceptible to being distracted given an additional modality. This is a concerning issue for multi-modal models since they should preserve unimodal capabilities when adding a second modality. Being able to evaluate this behaviour on the VGGSounder benchmark is a first step towards enabling the development of mitigation strategies, eventually resulting in stronger audio-visual models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 480, + 550, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 480, + 550, + 519 + ], + "spans": [ + { + "bbox": [ + 319, + 480, + 550, + 519 + ], + "type": "text", + "content": "Takeaway 5 Our modality confusion score reveals that all models are negatively impacted by additional modalities for a substantial amount of samples (see Tab. ②)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": "Performance across modalities Fig. 4 shows the performance profiles across modalities. At first glance, we can see that VideoLLaMA-2's performance is well balanced for different input modalities, while models from the Gemini family distinctly underperform on audio inputs. In contrast, embedding models exhibit a moderate balance across modalities, with DeepAVFusion and EquiAV showing slight underperformance for visual input." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "As Fig. 4 also illustrates, profiling of this kind is enabled through the modality annotations in VGGSounder. In contrast, VGGSound assumed that all classes are perceivable in both modalities, and did not account for background sounds. This resulted in consistent under-evaluation of foundation models (that were not finetuned on VGGSound) for audio inputs." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1032" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "In addition to the radar plot in Fig. 4, we provide results on VGGSound in Appendix D showing that all models have substantially lower performance than their hit scores in Tab. 2. This confirms that many model predictions were incorrectly flagged as false positives in VGGSound due to the incomplete ground-truth labels, painting a distorted picture of models' limitations." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 167, + 290, + 369 + ], + "blocks": [ + { + "bbox": [ + 65, + 167, + 290, + 369 + ], + "lines": [ + { + "bbox": [ + 65, + 167, + 290, + 369 + ], + "spans": [ + { + "bbox": [ + 65, + 167, + 290, + 369 + ], + "type": "image", + "image_path": "cc5cecfd4791f27111ac0dae1e53c38773b7a815acb789f7c7016fb2e18df776.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 375, + 296, + 441 + ], + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 441 + ], + "type": "text", + "content": "Figure 4. VGGSounder more accurately captures model performance across input modalities. We show the Hit score on VGGSounder and accuracy on VGGSound, normalised by the per-model maximum performance on each benchmark. Specifically for foundation models, we observe a significant difference in performance between VGGSound and VGGSounder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 463, + 291, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 463, + 291, + 502 + ], + "spans": [ + { + "bbox": [ + 61, + 463, + 291, + 502 + ], + "type": "text", + "content": "Takeaway 6 VGGSounder's more complete ground-truth labels allow for more accurate, modality-specific profiling of model performance (see Fig.4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 267, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 267, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 267, + 529 + ], + "type": "text", + "content": "5.2. Performance analysis using meta-classes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": "VGGSounder includes annotations of three meta-classes for each sample: background music indicates whether samples contain music without a visible source, voice over similarly marks speech without a visible source, and static image(s) flags that the visual stream consists of one or only a few static visual frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "These new meta-classes allow us to evaluate the model behaviour in challenging scenarios where information from one modality dominates. We consider the performance difference between samples that do contain a meta-label and samples that do not contain it. Positive numbers indicate that the models perform better on the subset with metal-labels. In Tab.3, we summarise the main findings, focussing on " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "-score as the most balanced multi-label metric. Additional results are provided in Appendix D" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "Background music All models perform worse on video samples containing background music. This indicates that it is challenging to decouple background audio from the rest of the video. The evaluated models are not good at differentiating between sound sources without visual cues, e.g. predicting different instruments in the background music." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 147, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 243 + ], + "type": "text", + "content": "Voice over In contrast to background music, we observe a clear difference between embedding models and foundation models for samples with voice-over. While the audio classification performance of embedding models drops substantially. This drop is only slight for VideoLLaMA-2 and Unified-IO 2, and the performance of other foundation models even improves. This indicates that the foundation models are less distracted by voice-over." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 245, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 316 + ], + "type": "text", + "content": "Static image(s) The impact of static images is more nuanced: First, audio classification performance improves for embedding models while it decreases for the foundation models. This shows that the purpose-built, VGGSoundfinetuned embedding models can more accurately predict specific sounds in the absence of other cues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 317, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 553, + 353 + ], + "type": "text", + "content": "Second, visual classification performance on static images drops for all models, suggesting that models rely on rich temporal cues to make accurate predictions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 461 + ], + "type": "text", + "content": "Third, the subset accuracy (i.e. exact label set matches) for uni-modal predictions shows that foundation models perform better in visual than audio classification with and without static images. In contrast, embedding models perform better in audio classification on static images but worse in visual classification for other samples. This suggests that non-static samples form a challenging subset for embedding models, where a model needs to favour one modality over another one to make a correct prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 464, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 548 + ], + "type": "text", + "content": "Samples without any meta-label When comparing the model performance on samples without background music, static images, or voice-over annotations (column neither in Tab. 3) to the performance on the entire test set, we see a performance gain (here negative numbers indicate an increased " + }, + { + "bbox": [ + 313, + 464, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1" + }, + { + "bbox": [ + 313, + 464, + 554, + 548 + ], + "type": "text", + "content": " score). This finding concludes that these three categories form challenging subsets of the dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 560, + 550, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 560, + 550, + 609 + ], + "spans": [ + { + "bbox": [ + 319, + 560, + 550, + 609 + ], + "type": "text", + "content": "Takeaway 7 Samples with background music, static images(s), and voice over provide distinct challenges for each model (see Tab.3). This highlights VGGSounder's value for comprehensive model evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 624, + 478, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 478, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 478, + 636 + ], + "type": "text", + "content": "5.3. Impact of VGGSounder labels" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Our relabelling pipeline adds two types of labels to those in VGGSound: (1) automatically generated labels based on synonymous classes and sub-/superclasses, and (2) human-curated labels. In Tab.4, we ablate the impact of each type of added labels in terms of the relative performance gains (Hit score). A complete breakdown of the effects across all" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1033" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 242 + ], + "type": "table", + "html": "
Modelbackground musicvoice over ΔF1static image(s)neither ΔF1
ΔF1ΔF1ΔF1Sub. Acc. w/Sub. Acc. w/oavav
avavaavavavavav
Embedding Models
CAV-MAE-0.66-0.59-0.65-1.130.31-0.2812.7319.2211.9819.21-1.00-1.27-1.24
DeepAVFusion-0.71-0.80-0.70-1.250.28-0.189.7910.819.3010.81-1.27-1.01-1.36
Equi-AV-0.79-0.42-0.41-0.970.28-0.2311.1210.4510.4910.45-0.97-0.97-0.89
AV-Siam-0.73-0.72-0.82-1.080.34-0.2612.3419.5311.5719.52-1.01-1.34-1.39
Closed-source Foundation Models
Gemini 1.5 Flash-0.22-0.41-0.692.08-0.36-0.251.6614.351.6814.391.40-1.10-1.62
Gemini 1.5 Pro-0.36-0.61-0.942.48-0.31-0.242.8320.852.8720.801.64-0.89-1.09
Gemini 2.0 Flash-0.09-0.32-0.580.030.13-0.261.7012.331.5312.40-0.03-1.07-1.22
Open-source Foundation Models
VideoLLaMA 2-0.47-0.72-0.84-0.540.27-0.3112.5519.6412.0419.70-0.47-1.26-1.33
Unified-IO 2-1.200.18-0.65-0.680.12-0.2111.3911.8910.8812.00-1.07-0.15-1.17
PandaGPT-1.13-0.15-0.420.56-0.22-0.162.944.272.914.24-0.60-0.15-0.65
OLA-2.160.09-0.441.50-0.48-0.1813.038.8812.888.89-0.80-0.13-0.06
", + "image_path": "6e96621888bbfcfb87649dd88334bc9a38904778b7eead2dce93dcf0fa39dfbb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 295, + 294, + 352 + ], + "blocks": [ + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "type": "text", + "content": "Table 3. Summary of performance differences in the presence/absence of meta-classes. Difference in " + }, + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_1" + }, + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "type": "text", + "content": " scores (" + }, + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{F}_1" + }, + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "type": "text", + "content": ") for audio-visual video classification on VGGSounder between videos with a meta-class and those without it. Positive numbers (" + }, + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 250, + 555, + 285 + ], + "type": "text", + "content": ") indicate better performance when the meta-class is present. Additional results are provided in Appendix D" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 295, + 294, + 352 + ], + "lines": [ + { + "bbox": [ + 57, + 295, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 57, + 295, + 294, + 352 + ], + "type": "table", + "html": "
ModelHuman labels ↑Auto labels ↑
AVAVAVAV
Gemini 1.5 Flash29.2814.5916.360.480.931.51
Gemini 1.5 Pro28.6125.5227.630.311.992.10
Gemini 2.0 Flash8.8012.1611.130.221.121.28
", + "image_path": "61fa65244ee5c430de75afa7adb150130023403062ea9ab4ae8143a8f510ad6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 415 + ], + "type": "text", + "content": "Table 4. Impact of added labels using different strategies in VGGSounder. We show the change in multi-label classification performance (" + }, + { + "bbox": [ + 55, + 360, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 360, + 295, + 415 + ], + "type": "text", + "content": " Hit) when adding automatically added (Auto) or human-annotated (Human) labels to VGGSound, and compare to the original VGGSound data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 430, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 515 + ], + "type": "text", + "content": "models and metrics is provided in Appendix D. While performance is consistently higher with automatically added labels (Auto), the increase is noticeably smaller than that for human-curated labels. Paired with the observation that models do frequently predict correct classes that were not part of the original VGGSound label set, this indicates that human-curated labels better cover the ground truth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 529, + 290, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 529, + 290, + 578 + ], + "spans": [ + { + "bbox": [ + 61, + 529, + 290, + 578 + ], + "type": "text", + "content": "Takeaway 8 Automatically added labels are an important step, but human-curated labels have a bigger effect on eliminating incorrectly flagged false positives, underscoring the value of accurate human annotation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 126, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 126, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 126, + 607 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Choice of VGGSound as base dataset VGGSound is commonly used to evaluate audio-visual models on the multimodal video classification task. As it is currently the most suitable testbed for audio-visual classification tasks (due to its size, diversity of categories, non-constrained setting, and relatively strong audio-visual correspondence), it serves as an optimal starting point for our substantially improved VGGSounder benchmark with a multi-label evaluation pro" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 553, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 553, + 321 + ], + "type": "text", + "content": "tocol for foundation models that makes the benchmark suitable for meaningful evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 324, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 443 + ], + "type": "text", + "content": "Multi-label vs single-label classification Video content is inherently complex, often containing multiple co-occurring objects and actions both within and across modalities. This makes it unlikely that a given clip belongs to just one class as is the case in the single-label classification task. Therefore, our VGGSounder extends the VGGSound test set to multi-label classification. Unlike models trained on a narrow single-label dataset, foundation models develop versatile representations. Multi-label evaluation thus also aligns very naturally with this capability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 388, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 388, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 388, + 467 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 558 + ], + "type": "text", + "content": "We introduced an modality-aware evaluation set for audiovisual foundation models. VGGSounder builds on the widely used VGGSound dataset by adding: (a) comprehensive human annotations for missing classes, (b) specifying modality information per label, (c) introducing specialised meta-labels for frequently occurring real-world challenges, and (d) using heuristic methods to improve label quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 559, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 556, + 701 + ], + "type": "text", + "content": "Through our newly introduced metric, modality confusion, we observe that incorporating additional modalities does not necessarily yield better results. Models often become more confused on a substantial subset of test samples. Furthermore, finetuned embedding models tend to rely heavily on audio cues, while foundation models depend more on visual information. Additionally, our meta-label analysis highlights distinct challenges across various specialised yet commonly occurring scenarios such as background music, static images, and voice-overs. Overall, we hope that the VGGSounder benchmark will advance the evaluation and development of foundational audio-visual models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1034" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 163 + ], + "type": "text", + "content": "The authors would like to thank Felix Förster, Sayak Mallick, and Prasanna Mayilvahananan for their help with data annotation, Felix Förster and Monica Riedler for proofreading the paper, and Thomas Klein and Shyamgopal Karthik for their help in setting up MTurk. They also thank numerous MTurk workers for labelling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 295, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 307 + ], + "type": "text", + "content": "This work was in part supported by the BMBF (FKZ: 01IS24060, 01I524085B, 01IS18039A), the DFG (SFB 1233, project number: 276693517), and the Open Philanthropy Foundation funded by the Good Ventures Foundation. WB acknowledges financial support via an Emmy Noether Grant funded by the German Research Foundation (DFG) under grant no. BR 6382/1-1. WB is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. This research utilised compute resources at the Tübingen Machine Learning Cloud, DFG FKZ INST 37/1057-1 FUGG. The authors thank the IMPRS-IS for supporting TW." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 319, + 115, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 115, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 115, + 331 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 338, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 295, + 392 + ], + "type": "text", + "content": "[1] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 395, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 395, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 61, + 395, + 295, + 427 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Deep audio-visual speech recognition. IEEE TPAMI, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 430, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 462 + ], + "type": "text", + "content": "[3] Triantafyllos Afouras, Yuki M Asano, Francois Fagan, Andrea Vedaldi, and Florian Metze. Self-supervised object detection from audio-visual correspondence. In ECCV, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 464, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 495 + ], + "type": "text", + "content": "[4] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Asr is all you need: Cross-modal distillation for lip reading. In ICASSP, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 498, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 294, + 529 + ], + "type": "text", + "content": "[5] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In ECCV, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 532, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 294, + 564 + ], + "type": "text", + "content": "[6] Humam Alwassel, Dhruv Mahajan, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 566, + 294, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 566, + 294, + 598 + ], + "spans": [ + { + "bbox": [ + 62, + 566, + 294, + 598 + ], + "type": "text", + "content": "[7] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Facetalk: Audio-driven motion diffusion for neural parametric head models. In CVPR, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 600, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 295, + 621 + ], + "type": "text", + "content": "[8] Relja Arandjelović and Andrew Zisserman. Look, listen and learn. In ICCV, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 624, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 295, + 644 + ], + "type": "text", + "content": "[9] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In ECCV, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 647, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 647, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 58, + 647, + 295, + 678 + ], + "type": "text", + "content": "[10] Yuki M. Asano, Mandela Patrick, Christian Rupprecht, and Andrea Vedaldi. Labelling unlabelled videos from scratch with multi-modal self-supervision. In NeurIPS, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 294, + 712 + ], + "type": "text", + "content": "[11] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Sound-net: Learning sound representations from unlabeled video. In NeurIPS, 2016." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "text", + "content": "[13] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[14] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Audio-visual synchronisation in the wild. In BMVC, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[15] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In CVPR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "text", + "content": "[16] Peihao Chen, Yang Zhang, Mingkui Tan, Hongdong Xiao, Deng Huang, and Chuang Gan. Generating visually aligned sound from videos. IEEE Transactions on Image Processing, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[17] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 299, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 331 + ], + "type": "text", + "content": "[18] Yanbei Chen, Yongqin Xian, A. Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "type": "text", + "content": "[19] Ziyang Chen, Prem Seetharaman, Bryan Russell, Oriol Nieto, David Bourgin, Andrew Owens, and Justin Salamon. Video-guided foley sound generation with multimodal controls. In CVPR, 2025." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 419 + ], + "type": "text", + "content": "[20] Ying Cheng, Ruize Wang, Zhihao Pan, Rui Feng, and Yuejie Zhang. Look, listen, and attend: Co-attention network for self-supervised audio-visual representation learning. In ACM MM, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[21] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "text", + "content": "[22] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "text", + "content": " * chatgpt quality. 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 533, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 555 + ], + "type": "text", + "content": "[23] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In ACCV, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 557, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 588 + ], + "type": "text", + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 590, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 621 + ], + "type": "text", + "content": "[25] Joshua P Ebeneze, Yongjun Wu, Hai Wei, Sriram Sethuraman, and Zongyi Liu. Detection of audio-video synchronization errors via event detection. In ICASSP, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 655 + ], + "type": "text", + "content": "[26] Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley music: Learning to generate music from videos. In ECCV, 2020." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 658, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 678 + ], + "type": "text", + "content": "[27] Ruohan Gao and Kristen Grauman. Co-separating sounds of visual objects. In ICCV, 2019." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "text", + "content": "[28] Aryo Pradipta Gema, Joshua Ong Jun Leang, Giwon Hong, Alessio Devoto, Alberto Carlo Maria Mancino, Rohit Saxena, Xuanli He, Yu Zhao, Xiaotang Du, Mohammad" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1035" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Reza Ghasemi Madani, Claire Barale, Robert McHardy, Joshua Harris, Jean Kaddour, Emile van Krieken, and Pasquale Minervini. Are we done with mmlu? arXiv preprint arXiv:2406.04127, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[29] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[30] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 294, + 228 + ], + "type": "text", + "content": "[31] Shir Goldstein and Yael Moses. Guitar music transcription from silent video. In BMVC, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 294, + 263 + ], + "type": "text", + "content": "[32] Yuan Gong, Alexander H Liu, Andrew Rouditchenko, and James Glass. Uavm: Towards unifying audio and visual models. IEEE Signal Processing Letters, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 307 + ], + "type": "text", + "content": "[33] Yuan Gong, Andrew Rouditchenko, Alexander H Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James Glass. Contrastive audio-visual masked autoencoder. In ICLR, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 352 + ], + "type": "text", + "content": "[34] Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "text", + "content": "[35] Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. Aligning ai with shared human values. In ICLR, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[36] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In ICLR, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "text", + "content": "[37] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 466, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 498 + ], + "type": "text", + "content": "[38] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Sparse in space and time: Audio-visual synchronisation with trainable selectors. In BMVC, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "text", + "content": "[39] Vladimir Iashin, Weidi Xie, Esa Rahtu, and Andrew Zisserman. Synchformer: Efficient synchronization from sparse cues. In ICASSP, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 534, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 566 + ], + "type": "text", + "content": "[40] Amir Jamaludin, Joon Son Chung, and Andrew Zisserman. You said that?: Synthesising talking faces from audio. IJCV, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 621 + ], + "type": "text", + "content": "[41] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "text", + "content": "[42] Naji Khosravan, Shervin Ardeshir, and Rohit Puri. On attention modules for audio-visual synchronization. In CVPR Workshop, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[43] Jongsuk Kim, Hyeongkeun Lee, Kyeongha Rho, Junmo Kim, and Joon Son Chung. Equipav: Leveraging equivariance for audio-visual contrastive learning. In ICML, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[44] A. Sophia Koepke, Olivia Wiles, and Andrew Zisserman. Visual pitch estimation. In SMC, 2019." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[45] A Sophia Koepke, Olivia Wiles, Yael Moses, and Andrew Zisserman. Sight to sound: An end-to-end approach for visual piano transcription. In ICASSP, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "text", + "content": "[46] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurlPS, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 142, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 553, + 186 + ], + "type": "text", + "content": "[47] Sangho Lee, Jiwan Chung, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, and Yale Song. Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning. In ICCV, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 188, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 553, + 220 + ], + "type": "text", + "content": "[48] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 222, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 243 + ], + "type": "text", + "content": "[49] Yan-Bo Lin and Gedas Bertasius. Siamese vision transformers are scalable audio-visual learners. In ECCV, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 246, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 277 + ], + "type": "text", + "content": "[50] Yan-Bo Lin and Yu-Chiang Frank Wang. Audiovisual transformer with instance attention for audio-visual event localization. In ACCV, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "text", + "content": "[51] Yan-Bo Lin, Yu-Jhe Li, and Yu-Chiang Frank Wang. Dual-modality seq2seq network for audio-visual event localization. In ICASSP, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "type": "text", + "content": "[52] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 360, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 414 + ], + "type": "text", + "content": "[53] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In CVPR, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 416, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 460 + ], + "type": "text", + "content": "[54] Jie Ma, Min Hu, Pinghui Wang, Wangchun Sun, Lingyun Song, Hongbin Pei, Jun Liu, and Youtian Du. Look, listen, and answer: Overcoming biases for audio-visual question answering. NeurIPS, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 462, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 506 + ], + "type": "text", + "content": "[55] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 508, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 553, + 529 + ], + "type": "text", + "content": "[56] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In ECCV, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 531, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 564 + ], + "type": "text", + "content": "[57] Shentong Mo and Pedro Morgado. Unveiling the power of audio-visual early fusion transformers with dense interactions through masked modeling. In CVPR, 2024." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "text", + "content": "[58] Liliane Momeni, Triantafyllos Afouras, Themos Stafylakis, Samuel Albanie, and Andrew Zisserman. Seeing wake words: Audio-visual keyword spotting. In BMVC, 2020." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[59] Arsha Nagrani, Joon Son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision. In ICASSP, 2020." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[60] Arsha Nagrani, Shan Yang, Anurag Arnab, Aren Jansen, Cordelia Schmid, and Chen Sun. Attention bottlenecks for multimodal fusion. In NeurIPS, 2021." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[61] Medhini Narasimhan, Shiry Ginosar, Andrew Owens, Alexei A Efros, and Trevor Darrell. Strumming to the beat: Audio-conditioned contrastive video textures. arXiv preprint arXiv:2104.02687, 2021." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1036" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[62] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "text", + "content": "[63] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Ambient sound provides supervision for visual learning. In ECCV, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[64] Andrew Owens, Jiajun Wu, Josh H McDermott, William T Freeman, and Antonio Torralba. Learning sight from sound: Ambient sound provides supervision for visual learning. IJCV, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[65] Mandela Patrick, Yuki M Asano, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multi-modal self-supervision from generalized data transformations. In NeurIPS, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "text", + "content": "[66] KR Prajwal, Liliane Momeni, Triantafyllos Afouras, and Andrew Zisserman. Visual keyword spotting with attention. In BMVC, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[67] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from coarse to fine. In ECCV, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 327 + ], + "type": "text", + "content": "[68] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICLR, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "text", + "content": "[69] Kun Su, Xiulong Liu, and Eli Shlizerman. Multi-instrumentalist net: Unsupervised generation of music from body movements. arXiv preprint arXiv:2012.03478, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "type": "text", + "content": "[70] Kun Su, Xiulong Liu, and Eli Shlizerman. How does it sound? In NeurIPS, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 415 + ], + "type": "text", + "content": "[71] Kun Su, Xiulong Liu, and Eli Shlizerman. From vision to audio and beyond: A unified model for audio-visual representation and generation. In ICML, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "text", + "content": "[72] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[73] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "text", + "content": "[74] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "text", + "content": "[75] Efthymios Tzinis, Scott Wisdom, Aren Jansen, Shawn Hershey, Tal Remez, Daniel PW Ellis, and John R Hershey. Into the wild with audioscope: Unsupervised audio-visual separation of on-screen sounds. In ICLR, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 581, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 613 + ], + "type": "text", + "content": "[76] Ilpo Viertola, Vladimir Iashin, and Esa Rahtu. Temporally aligned audio for video with autoregression. In ICASSP, 2025." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "text", + "content": "[77] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In ECCV, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "text", + "content": "[78] Yu Wu, Linchao Zhu, Yan Yan, and Yi Yang. Dual attention matching for audio-visual event localization. In CVPR, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[79] Fanyi Xiao, Yong Jae Lee, Kristen Grauman, Jitendra Malik, and Christoph Feichtenhofer. Audiovisual slowfast networks for video recognition. arXiv preprint arXiv:2001.08740, 2020." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 445 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[80] Haoming Xu, Runhao Zeng, Qingyao Wu, Mingkui Tan, and Chuang Gan. Cross-modal relation-aware networks for audio-visual event localization. In ACM MM, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[81] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In ECCV, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[82] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 185, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 218 + ], + "type": "text", + "content": "[83] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audio-visual question answering on videos. In ACM MM, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "text", + "content": "[84] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[85] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In ECCV, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 287, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 308 + ], + "type": "text", + "content": "[86] Hang Zhao, Chuang Gan, Wei-Chiu Ma, and Antonio Torralba. The sound of motions. In ICCV, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 310, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 332 + ], + "type": "text", + "content": "[87] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "text", + "content": "[88] Yipin Zhou, Zhaowen Wang, Chen Fang, Trung Bui, and Tamara L. Berg. Visual to sound: Generating natural sound for videos in the wild. In CVPR, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 366, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 422 + ], + "type": "text", + "content": "[89] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 422, + 553, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 445 + ], + "type": "text", + "content": "[90] Lingyu Zhu and Esa Rahtu. V-slowfast network for efficient visual sound separation. In WACV, 2022." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1037" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_content_list.json b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a65c9ff95e91018cfa15d1ef28b403ee849eec74 --- /dev/null +++ b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_content_list.json @@ -0,0 +1,1928 @@ +[ + { + "type": "text", + "text": "VGMamba: Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 233, + 138, + 767, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yihang Zhu, Jinhao Zhang, Yuxuan Wang, Among Wu, Cheng Deng* Xidian University, China", + "bbox": [ + 236, + 210, + 767, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhuyihang, zhangjinhao, wangyuxuan}@stu.xidian.edu.cn amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn", + "bbox": [ + 264, + 248, + 730, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 314, + 331, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As an important direction of embodied intelligence, 3D Visual Grounding has attracted much attention, aiming to identify 3D objects matching the given language description. Most existing methods often follow a two-stage process, i.e., first detecting proposal objects and identifying the right objects based on the relevance to the given query. However, when the query is complex, it is difficult to leverage an abstract language representation to lock the corresponding objects accurately, affecting the grounding performance. In general, given a specific object, humans usually follow two clues to finish the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, attribute-to-location clue reasoning, to conduct accurate grounding. Particularly, we propose a VGMamba network that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Taking a 3D point cloud scene and language query as the input, we first exploit SVD to make a decomposition of the extracted features. Then, a slidingwindow operation is conducted to capture attribute characteristics. Next, a location mamba is presented to obtain the corresponding location information. Finally, by means of multi-modal mamba fusion, the model could effectively localize the object that matches the given query. In the experiment, our method is verified on four datasets. Extensive experimental results demonstrate the superiority of our method.", + "bbox": [ + 99, + 344, + 485, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 101, + 753, + 232, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Visual Grounding aims to identify and localize objects in a 3D scene based on natural language descriptions. This task has become a key challenge at the intersection of computer vision and natural language processing, with significant applications in areas such as human-robot interaction and vision-language navigation [5, 7, 20]. In the context", + "bbox": [ + 99, + 777, + 483, + 868 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6711d085304180f206b02272e05fd46b312d1eff8323d3907a2af354f9e0adfd.jpg", + "image_caption": [ + "Figure 1. To achieve accurate 3D Visual Grounding, we explore two clues. First, we leverage the attribute clue to capture proposal objects. Then, a location clue is exploited to focus on the right grounding object. Experimental results on multiple datasets verify the effectiveness of this mechanism." + ], + "image_footnote": [], + "bbox": [ + 526, + 314, + 681, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/daa327c8f74bd1dc16ca2f51ef2e19984d8c47ae7bd3d5e76eb416e0549f5de8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 314, + 880, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of 3D environments, this task becomes even more complex due to the sparsity and irregularity of 3D point cloud data, as well as the variability in object shapes, scales, and the complexity of the language descriptions. Therefore, accurately localizing objects in 3D scene based on natural language descriptions is crucial for robots to understand and interact with their surroundings.", + "bbox": [ + 509, + 575, + 895, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies [2, 6, 38, 40, 41] have attempted to address this challenge by incorporating both object attributes (e.g., color, shape, texture) and spatial relationships (e.g., relative positions) to improve grounding accuracy. These methods typically follow a two-stage paradigm: first detecting candidate objects in the scene, and then assessing relevance to the given query based on feature similarity. While these methods have demonstrated effectiveness in certain scenarios, they still exhibit some limitations.", + "bbox": [ + 509, + 680, + 895, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, most existing works [6, 38, 40] treat object attributes and spatial relationships in isolation, neglecting the inherent dependencies between them. In human perception, object attributes often serve as priors for location reasoning. For example, when searching for \"The red chair next to the", + "bbox": [ + 509, + 815, + 895, + 891 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 117, + 876, + 243, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5295", + "bbox": [ + 482, + 926, + 514, + 939 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "table,\" humans first identify objects matching the attribute \"red chair\" and then refine their search based on location cues. Existing approaches lack a mechanism to systematically leverage this reasoning process, resulting in suboptimal performance in complex scenes. Furthermore, in complex 3D environments, a query may refer to multiple objects. However, existing Transformer-based methods often fail to effectively capture global dependencies among multiple objects with shared attributes and spatial relations due to their quadratic complexity in self-attention, which limits scalability, and their tendency to focus on local interactions rather than modeling long-range dependencies.", + "bbox": [ + 99, + 102, + 480, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, as shown in Fig. 1, we separately explore reasoning attribute-based information and location-based characteristic to ground corresponding 3D objects. To be specific, we propose VGMamba, a novel architecture that systematically models attribute-to-location dependencies while efficiently capturing long-range interactions. VGMamba consists of three key modules: SVD-based Attribute Mamba, which leverages Singular Value Decomposition (SVD) to decompose the extracted visual features into multiple subspaces corresponding to different attribute scales. However, SVD alone lacks long-range modeling capability, which is critical for capturing hierarchical dependencies among attributes. To overcome this, we employ Mamba [11] to sequentially integrate multi-scale attribute features, preserving their interdependencies over long sequences. Compared to Transformers, Mamba maintains linear computational complexity while effectively modeling long-range dependencies among attribute components; the Location Mamba, which models both global and local positional relationships among objects by establishing cross-channel communication through SSM, effectively enhancing object-related features; and the Instructive Dual-Mamba block, which fuses object-related and text features through cross-modal guidance of text query. By replacing computationally expensive attention mechanisms with structured state-space modeling, VGMamba significantly improves efficiency without sacrificing performance. Extensive experiments demonstrate that our method achieves state-of-the-art performance, particularly excelling in scenarios involving multiple objects and ambiguous descriptions. Our chief contributions are threefold:", + "bbox": [ + 99, + 277, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We explore a novel mechanism, i.e., attribute-to-location clue reasoning, for performing 3D visual grounding.", + "- We propose a novel framework VGMamba, comprising three core modules: the SVD-based Attribute Mamba for disentangling attributes, the Location Mamba for modeling location relations, and the Instructive Dual-Mamba block for text and object-related feature fusion.", + "- We achieve new state-of-the-art performance on multiple 3D Visual Grounding benchmarks, demonstrating superior accuracy and computational efficiency." + ], + "bbox": [ + 101, + 737, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 101, + 650, + 116 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Visual Grounding. Recent advancements incorporate self-attention-based scene encoding [2], viewpoint-aware spatial reasoning [29], and multi-attribute Transformers for fine-grained language-visual alignment [38]. However, these attention-based approaches struggle with long-sequence feature modeling and computational efficiency. For multi-object grounding, M3DRef-CLIP [42] introduced the first multi-object dataset and an attention-based fusion module integrating 3D object and 2D rendered image features. Subsequent works [41] refine proposal selection, camera pose optimization, and fusion strategies. However, these methods often embed global 2D features, introducing redundancy. In contrast, we focus on fine-grained attribute extraction as priors to enhance spatial reasoning.", + "bbox": [ + 509, + 125, + 893, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models. State Space Models (SSMs) [12, 30, 34] have become a prominent approach in sequence modeling, particularly for capturing long-range dependencies with linear time complexity. Furthermore, Mamba [11] introduced a selective state-space mechanism, providing efficient sequence modeling and surpassing traditional Transformers in certain tasks. This advancement led to various extensions across vision applications, including Vision Mamba [43] and VMamba [21], which extends scanning in four directions to fully capture interrelations among image patches. Besides, Mamba-based models have also been applied in video understanding [19, 22, 25, 39] and medical image segmentation [24, 28, 36, 37], demonstrating the versatility and efficiency of SSMs in diverse visual domains.", + "bbox": [ + 511, + 330, + 893, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce Mamba into the object grounding task for the first time, exploring its potential in fine-grained feature extraction. We also propose a novel cross-modal fusion framework for language and object-related features, which achieves outstanding performance and significantly reduces computational costs.", + "bbox": [ + 511, + 537, + 893, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Overview of State Space Models", + "text_level": 1, + "bbox": [ + 511, + 637, + 797, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, state space models (SSMs) [9, 12, 30, 32] have attracted much attention for their ability to model continuous systems, constructing the foundation for the popular Mamba architecture [11]. Particularly, SSMs generally take an input sequence $x(t) \\in \\mathbb{R}^L$ as the input and output the corresponding sequence $y(t) \\in \\mathbb{R}^L$ through hidden states $h(t) \\in \\mathbb{R}^N$ , where $N$ is the number of hidden states. The system is governed by differential equations that describe how the hidden state evolves over time:", + "bbox": [ + 511, + 662, + 893, + 792 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 804, + 893, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathbf{A}$ , $\\mathbf{B}$ , and $\\mathbf{C}$ are matrices that project the input into the hidden states and subsequently to the output.", + "bbox": [ + 511, + 830, + 893, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to apply SSMs in practice tasks, a discretization process is employed, typically using a time scale parameter", + "bbox": [ + 511, + 859, + 893, + 890 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5296", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e778ec5bc5e10d21d9e714ed850c870048656c05a56dbc6c33c6138fcceacb16.jpg", + "image_caption": [ + "Figure 2. VGMamba for quantity-agnostic 3D Visual Grounding. Based on the given language description, we could utilize two clues to localize corresponding objects accurately, i.e., attribute and location information. To this end, we first design an SVD-based attribute mamba to discover multiple attribute-related objects. Then, a location mamba is further designed to select location-relevant objects. Finally, we present an Instructive Dual-Mamba block to localize the object that matches the given query." + ], + "image_footnote": [], + "bbox": [ + 102, + 99, + 620, + 377 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5352586d5d5e8dd8741d74df41b65fe6d2702b70f49457916439407fe948b6cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 103, + 897, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\Delta$ to convert continuous parameters into discrete ones. The processes are shown as follows:", + "bbox": [ + 101, + 465, + 482, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} (\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}) \\cdot \\Delta \\mathbf {B}. (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 505, + 482, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The discretized state-space output can be represented as:", + "bbox": [ + 101, + 534, + 465, + 549 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} = \\overline {{\\mathbf {A}}} h _ {k - 1} + \\overline {{\\mathbf {B}}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 559, + 482, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the SSMs framework, Mamba [11] employs the Selective Scan Mechanism (S6) to conduct dynamic sequential modeling and capture critical task-related information. Though Mamba-based networks have been verified to be effective in many tasks [19, 24], it is under-explored for 3D Visual Grounding. To this end, we study whether the Mamba could be used to capture the relation between objects and language descriptions. We directly plug Mamba into M3DRef-CLIP [42] and conduct experiments on Multi3DRefer dataset [42]. However, we observe that this operation could not improve the grounding performance. The reason may be that Mamba emphasizes capturing global object information and cannot construct the relation between objects and language descriptions.", + "bbox": [ + 99, + 587, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. VGMamba for 3D Visual Grounding", + "text_level": 1, + "bbox": [ + 99, + 806, + 424, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose a new VGMamba architecture, aiming to reason attribute and location clues based on the given language descriptions and improve the performance of 3D Visual Grounding effectively.", + "bbox": [ + 99, + 830, + 483, + 891 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, in Fig. 2 (a), we follow the work [42] and adopt PointGroup [17] as a 3D detector to generate an object proposal list $o = (o_1,o_2,\\dots,o_n)$ with $n$ proposal boxes and corresponding features $F_{\\mathrm{3D}}\\in \\mathbb{R}^{n\\times 32}$ . Each proposal is then rendered as $m$ multi-view images, which are encoded by CLIP [27] to extract $d$ -dimensional patch features $I\\in \\mathbb{R}^{n\\times m\\times p\\times d}$ and aggregated features $\\hat{I}\\in \\mathbb{R}^{n\\times m\\times d}$ , where $p$ is the patch number of each multiview image. Next, for the given language description, we continue following the work [42] and utilize CLIP to obtain text feature $T\\in \\mathbb{R}^d$ .", + "bbox": [ + 511, + 465, + 893, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, 3D Visual Grounding can be achieved by reasoning two key clues. One clue is to accurately recognize object-related attribute, e.g., dark blue. Another clue is to effectively obtain the location-relevant visual content, e.g., facing the backboard. With this in mind, we separately design an attribute mamba and location mamba to sufficiently capture object-related attribute and location characteristic.", + "bbox": [ + 511, + 628, + 893, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. SVD-based Attribute Mamba", + "text_level": 1, + "bbox": [ + 511, + 748, + 769, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, given a specific object, the extracted representations involve multiple different attribute characteristics, e.g., color, shape, and texture. However, the 3D Visual Grounding task involves localizing objects that possess specific attributes, such as dark blue. Therefore, it is meaningful to disentangle current representations. To this end, we explore using Singular Value Decomposition (SVD) to perform feature-level decomposition and recombination ac", + "bbox": [ + 511, + 772, + 893, + 890 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5297", + "bbox": [ + 483, + 926, + 514, + 939 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cording to the calculated singular values.", + "bbox": [ + 101, + 102, + 364, + 116 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Particularly, as illustrated in Fig. 2 (b), for each patch feature $I_{i}\\in \\mathbb{R}^{d\\times p}$ , SVD is first used to decompose $I_{i}$ into three matrices: $I_{i} = U\\Sigma V^{T}\\in \\mathbb{R}^{d\\times p}$ , where $U\\in \\mathbb{R}^{d\\times d}$ and $V\\in \\mathbb{R}^{p\\times p}$ are orthogonal matrices, and $\\Sigma$ contains singular values sorted along its main diagonal [10]. Subsequently, inspired by Multi-Scale Attention [18], we introduce Sliding Window to recombine the singular vectors using three sliding windows with varying size $r$ :", + "bbox": [ + 99, + 117, + 483, + 234 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{F _ {r} \\right\\} = \\text {S l i d i n g W i n d o w} (\\Sigma , r), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 244, + 482, + 261 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S l i d i n g W i n d o w} (\\Sigma , r) = U _ {d \\times r} \\Sigma_ {r \\times r} V _ {r \\times p} ^ {T}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 263, + 482, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r \\in \\{r_1, r_2, r_3\\}$ . SlidingWindow refers to combining r consecutive singular vectors into a new multi-scale feature $F_r$ . Since the singular vectors corresponding large singular values and those corresponding small singular values separately contain generalized coarse-grained information and discriminative fine-grained information, through multiscale operation, we can sufficiently capture attribute information involved in visual representations. Following this, $F_r$ is processed by three independent Mamba to model attribute feature. Finally, a concatenation followed by a $1 \\times 1$ convolution is applied to merge the multi-scale features, yielding the final feature representations $F_{i-att} \\in \\mathbb{R}^d$ of the single-view image. For $n$ objects with $m$ rendered images, we can obtain the feature representation $F_{att} \\in \\mathbb{R}^{n \\times m \\times d}$ .", + "bbox": [ + 99, + 292, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Location Mamba", + "text_level": 1, + "bbox": [ + 101, + 507, + 266, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (a), only relying on the attribute clue could not guarantee accurately identifying the objects corresponding to the given language description. To this end, based on the captured attribute features, we design a cross-channel communication mechanism to facilitate the learning of multi-object location relationships, enforcing the locking of the object matching the given description.", + "bbox": [ + 99, + 527, + 483, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, to capture spatial relationships from a consistent viewpoint, for each rendered object image, we define an encoding $pos \\in \\mathbb{R}^{n \\times m \\times 6}$ that includes object center coordinates $(x,y,z)$ and bounding box dimensions $(h,w,l)$ . Following this, we utilize a linear layer to perform a transformation on $pos$ :", + "bbox": [ + 99, + 631, + 483, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\hat {o} s = \\operatorname {L i n e a r} (p o s) \\in \\mathbb {R} ^ {n \\times m \\times d}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 729, + 482, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the summation result $I_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times m \\times d}$ of $p \\hat{o} s$ , $\\hat{I} \\in \\mathbb{R}^{n \\times m \\times d}$ , and $F_{\\mathrm{att}} \\in \\mathbb{R}^{n \\times m \\times d}$ is taken to perform the following operations: exploring a dedicated mamba to respectively conduct communication from global and local perspectives. For global communication, the channels of $n$ proposal objects are treated as a sequence to enable cross-object interaction:", + "bbox": [ + 99, + 757, + 483, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {g l o b a l}} = \\operatorname {M a m b a} _ {\\text {g l o b a l}} \\left(I _ {\\text {s p a}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 875, + 482, + 891 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34e446207457583f5f4981041dfca6d2d5c3db36ed02820ded22c0b98f669096.jpg", + "image_caption": [ + "Figure 3. The architecture of Instructive Dual-Mamba (IDM) for conducting multi-modal fusion." + ], + "image_footnote": [], + "bbox": [ + 514, + 101, + 893, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{\\mathrm{global}} \\in \\mathbb{R}^{n \\times m \\times d}$ denotes the SSM output.", + "bbox": [ + 511, + 253, + 831, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for the local communication, the $u$ nearest neighbors of each object are selected based on Euclidean distance, obtaining multi-neighbor feature $\\hat{I}_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times m \\times u \\times d}$ . Cross-channel scanning is then applied to $\\hat{I}_{\\mathrm{spa}}$ and outputs central object features:", + "bbox": [ + 511, + 270, + 893, + 345 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {l o c a l}} = \\operatorname {S e l e c t C e n t e r} \\left(\\mathrm {M a m b a} _ {\\text {l o c a l}} \\left(\\hat {I} _ {\\mathrm {s p a}}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 354, + 893, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{\\mathrm{local}} \\in \\mathbb{R}^{n \\times m \\times d}$ denotes the current SSM output. Subsequently, by summing $I_{\\mathrm{global}}$ and $I_{\\mathrm{local}}$ , and then averaging over the $m$ images, we obtain the feature $F_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times d}$ that contains plentiful location information. Finally, $F_{\\mathrm{spa}}$ is concatenated with the proposal feature $F_{3\\mathrm{D}} \\in \\mathbb{R}^{n \\times 32}$ output by the 3D detector, followed by a $1 \\times 1$ convolution layer for dimensionality reduction, yielding $F_{\\mathrm{obj}} \\in \\mathbb{R}^{n \\times d}$ . We represent the feature of each object in $F_{\\mathrm{obj}}$ as $G \\in \\mathbb{R}^d$ .", + "bbox": [ + 511, + 381, + 893, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Multi-modal Mamba Fusion", + "text_level": 1, + "bbox": [ + 511, + 508, + 759, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further enhance the grounding accuracy, we present a multi-modal mamba fusion module to align the proposals to the corresponding language descriptions. Specifically, as shown in Fig. 3, the Instructive Dual-Mamba (IDM) block is proposed to use descriptions as guidance to capture matched object-related feature. Taking the language feature $T \\in \\mathbb{R}^d$ , object-related feature $G \\in \\mathbb{R}^d$ as the input, we first divide $T$ into a forward sequence and a reversed backward sequence, which are then processed by two independent SSMs (Eq. (1)) respectively. The outputs from the two branches are integrated to obtain the updated feature. The processing of input feature $G \\in \\mathbb{R}^d$ is similar to $T \\in \\mathbb{R}^d$ . According to Eq. (1), the state transition matrices $A$ retain critical historical information for updating hidden states, while $B$ and $C$ primarily influence input and output. Therefore, we propose replacing the state transition matrices $\\vec{\\mathbf{A}}_G, \\vec{\\mathbf{A}}_G$ with their counterparts $\\vec{\\mathbf{A}}_T, \\vec{\\mathbf{A}}_T$ derived from language features, enabling cross-modal guidance in extracting object-related features. The remaining parameter matrices remain unchanged. The process of above could be represented as:", + "bbox": [ + 511, + 530, + 893, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overleftarrow {h} _ {\\mathrm {T}} ^ {t} = \\overleftarrow {\\mathbf {A}} _ {\\mathrm {T}} \\overleftarrow {h} _ {\\mathrm {T}} ^ {t - 1} + \\overleftarrow {B} _ {\\mathrm {T}} \\overleftarrow {x} _ {\\mathrm {T}} ^ {t}, \\quad \\overrightarrow {h} _ {\\mathrm {T}} ^ {t} = \\overrightarrow {\\mathbf {A}} _ {\\mathrm {T}} \\overrightarrow {h} _ {\\mathrm {T}} ^ {t - 1} + \\overrightarrow {B} _ {\\mathrm {T}} \\overrightarrow {x} _ {\\mathrm {T}} ^ {t}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 845, + 893, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overleftarrow {h} _ {\\mathsf {G}} ^ {t} = \\overleftarrow {\\mathbf {A}} _ {\\mathsf {T}} \\overleftarrow {h} _ {\\mathsf {G}} ^ {t - 1} + \\overleftarrow {B} _ {\\mathsf {G}} \\overleftarrow {x} _ {\\mathsf {G}} ^ {t}, \\quad \\overrightarrow {h} _ {\\mathsf {G}} ^ {t} = \\overrightarrow {\\mathbf {A}} _ {\\mathsf {T}} \\overrightarrow {h} _ {\\mathsf {G}} ^ {t - 1} + \\overrightarrow {B} _ {\\mathsf {G}} \\overrightarrow {x} _ {\\mathsf {G}} ^ {t}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 866, + 893, + 886 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5298", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2650de171ada3752c8c5741d94de92d2601f4b852ca2e3d3fb53a46bd0ddf722.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [3]ECCV2067.646.232.121.340.026.1
TGNN [14]AAAI2168.656.829.823.237.429.7
MVT [15]CVPR2277.766.531.925.340.833.3
3D-SPS [23]CVPR2284.166.740.329.848.837.0
EDA [35]CVPR2385.868.649.137.654.642.3
VPP-Net [29]CVPR2486.167.150.339.055.743.3
M3DRef-CLIP [42]ICCV23-77.2-36.8-44.7
G3-LQ [33]CVPR2488.673.350.239.756.044.7
MCLN [26]ECCV2486.972.752.040.857.245.5
MA2TransVG [38]CVPR2486.374.153.841.457.945.7
ConcreteNet [31]ECCV2486.482.142.438.450.646.5
D-LISA [41]NeurIPS24-75.5-40.0-46.9
GPS [16]ECCV24-77.9-42.7-48.1
Chat-Scene [13]NeurIPS2489.682.547.842.955.550.2
PQ3D [44]ECCV2486.778.351.546.257.051.2
VGMamba-91.979.654.847.660.053.9
", + "bbox": [ + 205, + 99, + 792, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $x_{\\mathrm{T / G}}^t$ represents the input at time step $t$ . $\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{T}}$ and $\\overrightarrow{\\mathbf{A}}_{\\mathrm{T}}$ are the state transition matrices used for retaining essential historical information and updating hidden states, while matrices $B_{\\mathrm{T / G}}$ primarily affect input.", + "bbox": [ + 99, + 446, + 482, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we share the gating mechanism parameters $z$ of the language features to construct cross-modal fusion in the hidden state space, ultimately outputting the average multi-view attribute features of each object in the scene, represented as $F_{\\mathrm{att - spa}} \\in \\mathbb{R}^{n \\times d}$ . Finally, a convolution operation is applied to output confidence scores $O \\in \\mathbb{R}^{n \\times 1}$ for each object proposal.", + "bbox": [ + 99, + 513, + 483, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Training Objectives", + "text_level": 1, + "bbox": [ + 101, + 638, + 285, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on previous work [42], the loss of VGMamba consists of the 3D Visual Grounding loss $L_{\\mathrm{ref}}$ , text-object contrastive loss $L_{\\mathrm{con}}$ , and object detection loss $L_{\\mathrm{det}}$ , which is formulated as:", + "bbox": [ + 99, + 664, + 483, + 723 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {r e f}} + L _ {\\text {c o n}} + L _ {\\text {d e t}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 744, + 482, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 101, + 786, + 230, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our proposed VGMamba on 3D Visual Grounding tasks with varying numbers of target objects, we conducted experiments on both single-object datasets (ScanRefer [3], Nr3D and Sr3D [1]) and multi-object dataset Multi3DRefer [42].", + "bbox": [ + 99, + 815, + 483, + 888 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1778e7ad0b94046d2457b2225ff4d498fb49fb905eeaead2263ab530df51b240.jpg", + "table_caption": [ + "Table 1. Comparison results on the ScanRefer [3] dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. Our proposed VGMamba performs favorably over current state-of-the-art approach PQ3D [44]." + ], + "table_footnote": [], + "table_body": "
MethodNr3DSr3D
OverallHardOverallHard
ReferIt3D [?]35.627.940.831.5
TGNN [14]37.330.645.036.9
M3DRef-CLIP [42]49.443.4--
3D-SPS [23]51.545.162.665.4
EDA [35]52.146.168.162.9
D-LISA [41]53.146.2--
MVT [15]55.449.164.558.8
VPP-Net [29]56.9-68.7-
G3-LQ [33]58.450.773.166.3
MCLN [26]59.8-68.4-
MiKASA [2]64.459.475.267.3
GPS [16]64.957.877.571.6
MA2TransVG [38]65.257.673.969.3
PQ3D [44]66.758.779.772.8
VGMamba68.361.481.374.4
", + "bbox": [ + 517, + 446, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons on Nr3D and Sr3D dataset. We have highlighted the top-performing three methods in purple.", + "bbox": [ + 511, + 727, + 893, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Single-object Grounding", + "text_level": 1, + "bbox": [ + 511, + 785, + 732, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 809, + 777, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ScanRefer. The ScanRefer dataset comprises 51,583 natural language descriptions for 11,046 objects across 800 3D scenes from the ScanNet dataset [8]. Each description specifies a target object and its location relations within", + "bbox": [ + 511, + 830, + 893, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5299", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Query Text", + "bbox": [ + 181, + 104, + 243, + 116 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The green chair next to the brown table with a white vase.", + "bbox": [ + 145, + 152, + 302, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The brown chairs under the white rectangular table.", + "bbox": [ + 147, + 222, + 290, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The white pillow closest to the wall with the painting.", + "bbox": [ + 143, + 296, + 292, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "M3DRef-CLIP", + "bbox": [ + 326, + 104, + 408, + 116 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7af92848cd849a159b2c3f9792c387e4c14eda7647d89b5b8a7d49aa75e057e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 130, + 424, + 194 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0c4f4861ee5c83719a71282974d1e73d2ea4f44c629480fc078445010055178.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 199, + 423, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/081c860578023974147507a4f27e639d032bab4b7f33dc71986f57a09378c8d3.jpg", + "image_caption": [ + "Figure 4. Qualitative examples of the Multi3DRefer dataset. The results illustrate the grounding performance of M3DRef-CLIP (blue box), D-LISA (blue box), and our VGMamba(green box), compared to the ground truth labels (red box)." + ], + "image_footnote": [], + "bbox": [ + 305, + 276, + 419, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "D-LISA", + "bbox": [ + 486, + 104, + 529, + 114 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6b018c6ba987e46ce7952b65b2b64175c972dfd434bfee05fc958d84691d748.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 130, + 562, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/93338e41061c3d2a1846cc49f16d4f186e38aa78896d9a1015dfa88b33de1b29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 196, + 563, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1fce43dc233542dee267fcf47ea7ec9e877facdb498f63fdc6d52176622d9a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 277, + 558, + 333 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51b9f339182965f4d04895b49b88ca83d09f9849cd7d173467a0430730045022.jpg", + "image_caption": [ + "VGMamba (ours)" + ], + "image_footnote": [], + "bbox": [ + 612, + 130, + 715, + 194 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ae35d0d9bed1b0d0aad93858f834b02cc8e5759476240eb8f6a292ae1388847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 196, + 709, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4bc2e55ec9184409c94c96d181da8d640898cc3f35dc2575f2a19040bd44527d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 277, + 705, + 333 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f84da7be7687c7062e9369bb8b9dd4e709de942c48c593ff06ddfcb508130193.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 746, + 130, + 851, + 194 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38ac7537dc9007e61f99c1e6ece4f4e7636884e2296c6e9d8949240497093dc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 196, + 851, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7afff4911b1c37a90619362d3399f3e7a841128bdd0f7e723b38cb34838d2fbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 277, + 846, + 333 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5be0d6f0ff3d58577f18901fee964a67d0bca9b8aac0174715bebdfe2fce0656.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodZT w/o DZT w/DST w/o DST w/DMTAll
D3Net [4]81.632.538.623.335.032.2
M3DRef-CLIP [42]81.839.447.830.637.938.4
M3DRef-CLIP w/NMS [41]79.040.567.640.049.149.3
PQ3D [44]85.457.768.543.640.950.1
D-LISA [41]82.443.767.142.551.051.2
VGMamba83.645.270.946.755.154.9
", + "bbox": [ + 106, + 397, + 478, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of F1@0.5 results on the Multi3DRefer [42] validation set, where our method outperforms all baselines.", + "bbox": [ + 101, + 517, + 480, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the scene, enabling the model to understand both object attributes and location context. The dataset is divided into two subsets: Unique, where the target object is the only instance of its class within the scene, and Multiple, where multiple instances of the same class are present. We evaluate our model on validation set using Acc@0.25 and Acc@0.5, which measure the proportion of predicted bounding boxes that achieve an Intersection over Union (IoU) larger than 0.25 and 0.5 with the ground truth box.", + "bbox": [ + 99, + 573, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\text{Nr3D/Sr3D}$ . The $\\text{Nr3D}$ and $\\text{Sr3D}$ datasets, part of the ReferIt3D benchmark [1], are built upon the ScanNet dataset [8] for 3D visual grounding. $\\text{Nr3D}$ (Natural Reference in 3D) consists of 41,503 human-annotated referring expressions across 707 indoor scenes, where each description uniquely identifies a target object among distractors. $\\text{Sr3D}$ (Spatial Reference in 3D) contains 83,572 machine-generated expressions based on predefined templates, focusing on spatial relationships between objects. Both datasets provide ground-truth object masks. Following [1], performance is evaluated using accuracy, measuring the proportion of correctly localized target objects.", + "bbox": [ + 99, + 712, + 482, + 890 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 398, + 715, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement the proposed VGMamba model using PyTorch and train it end-to-end on a single NVIDIA A6000 GPU. The model is optimized using the AdamW optimizer with a batch size of 4 and a learning rate of 5e-4 with cosine decay scheduling. Following prior work [42], we employ a pre-trained PointGroup [17] module as the detector, which is fine-tuned on the ScanNet dataset. The proposal box with the highest matching score is selected as the final prediction. We adopt the official train/val/test splits provided by the ScanRefer and $\\mathrm{Nr3D / Sr3D}$ datasets to ensure fair comparisons with baseline methods.", + "bbox": [ + 509, + 419, + 893, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.3. Baseline Comparison", + "text_level": 1, + "bbox": [ + 511, + 590, + 700, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanRefer. The experimental results on ScanRefer are shown in Tab. 1, with the following key observations: (i) Our method achieves state-of-the-art performance with an overall accuracy of $60.0\\%$ at IoU 0.25 and $53.9\\%$ at IoU 0.5, surpassing the best baseline PQ3D [44] by $3.0\\%$ and $2.7\\%$ , respectively. (ii) In the Multiple case, VGMamba attains $54.8\\%$ at IoU 0.25 and $47.6\\%$ at IoU 0.5, outperforming PQ3D and demonstrating that its attribute-to-location reasoning effectively disambiguates objects under complex scenarios. (iii) Unlike previous methods [13, 44] that show notable performance variations across settings, our VGMamba maintains consistently high accuracy, validating its robust generalization capability in both unique and multiple-object scenarios.", + "bbox": [ + 509, + 609, + 893, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$Nr3D/Sr3D$ . Tab. 2 presents the grounding accuracy of different methods on the $\\mathrm{Nr3D}$ and Sr3D datasets. VGMamba achieves state-of-the-art performance of $68.3\\%$ and $81.3\\%$ on both datasets, surpassing prior state-of-the-art methods. The $\\mathrm{Nr3D}$ dataset presents additional challenges due to", + "bbox": [ + 511, + 816, + 893, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5300", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/209dc3abe4f2db1508be21731d85950122899605fece1cd36b9b4f221cc85352.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModulesMulti3DRefer (F1@0.5 on Val)
RowIDMLMSAMZT w/o DZT w/DST w/o DST w/DMTAll
1XXX81.839.447.830.637.938.4
2✓XX83.943.756.839.743.542.3
3X✓X80.240.761.142.845.143.8
4XX✓81.544.067.643.947.845.2
5✓✓X83.546.165.542.348.648.3
6✓X✓82.948.268.745.849.049.2
7X✓✓85.445.369.244.253.152.5
8✓✓✓83.645.270.946.755.154.9
", + "bbox": [ + 106, + 99, + 480, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study of proposed modules on Multi3DRefer.", + "bbox": [ + 107, + 250, + 475, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "its intricate and free-form textual descriptions, which increase the difficulty of cross-modal alignment. Despite this, our method significantly outperforms alternatives that leverage large-scale 3D vision-language pre-training models [13, 44]. Notably, in the more challenging \"Hard\" subset, VGMamba improves accuracy by $2.7\\%$ in $\\mathrm{Nr3D}$ and $1.6\\%$ in Sr3D, demonstrating its superior capability in distinguishing ambiguous objects by effectively capturing fine-grained attribute and spatial relationships.", + "bbox": [ + 99, + 276, + 483, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Multi-object Grounding", + "text_level": 1, + "bbox": [ + 101, + 417, + 318, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 101, + 439, + 367, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi3DRefer is a dataset derived from ScanRefer [3], comprising a total of 61,926 textual descriptions corresponding to 11,609 distinct objects. Each description may reference zero, one, or multiple target objects, enabling the evaluation of grounding models in a variety of scenarios. Following previous work[42], we compute the F1 score at the Intersection over union (IoU) threshold of 0.5(denoted as F1@0.5) across five distinct scenarios: a) zero target without distractors of the same semantic class (ZT w/o D); b) zero target with distractors (ZT w/D); c) single target without distractors (ST w/o D); d) single target with distractors (ST w/D); and e) multiple targets (MT). Finally, the average across these five evaluation cases is calculated to provide an overall performance score.", + "bbox": [ + 99, + 455, + 483, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2. Implementation Details", + "text_level": 1, + "bbox": [ + 101, + 669, + 305, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental setting for the multi-object dataset is similar to that for the single-object dataset, with a few differences: during training, we treat all predicted bounding boxes with an IoU greater than $\\tau_{\\mathrm{train}} = 0.25$ relative to the ground truth as target bounding boxes. During inference, we consider all predicted bounding boxes with a matching score greater than $\\tau_{\\mathrm{pred}} = 0.1$ as the model's predictions.", + "bbox": [ + 99, + 686, + 483, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.3. Baseline Comparison", + "text_level": 1, + "bbox": [ + 101, + 797, + 290, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our 3D multi-object grounding performance with existing works quantitatively and qualitatively. As illustrated in Tab. 3, our proposed VGMamba achieves competitive performance across all scenarios. In cases with distractors (ST w/D), VGMamba achieves F1 scores of", + "bbox": [ + 99, + 815, + 483, + 888 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a7a8fbc8d47d4bfd1b88ff83e3a7d384d6d0a820b32355db816599b6858a07c.jpg", + "image_caption": [ + "Figure 5. Efficiency-Performance Trade-off on ScanRefer." + ], + "image_footnote": [], + "bbox": [ + 550, + 99, + 857, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$46.7\\%$ , surpassing the second-best competitor by $3.1\\%$ , which highlights its robustness in managing complex scenes with multiple potential matches. In the MT case, VG-Mamba attains an F1 score that is $4.1\\%$ higher than the second-best competitor, underscoring the advantage of the Instructive Dual-Mamba (IDM) in addressing the multi-object grounding task.", + "bbox": [ + 509, + 282, + 893, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the qualitative comparison with M3DRef-CLIP[42] and D-LISA[41] in Fig. 4. The top two rows illustrate examples involving multiple target objects in complex scenes. Our VGMamba accurately locates the objects matching the given textual descriptions, even in the presence of distractors. The last row demonstrates a single-target case with surrounding distractors. VGMamba effectively isolates and identifies the correct object, while baseline models struggle, often mistakenly including distractors in their predictions.", + "bbox": [ + 509, + 386, + 893, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 542, + 671, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of each proposed module within our VGMamba framework, we conduct ablation studies on the Multi3DRefer dataset, as shown in Tab. 4. More ablation results are detailed in supplementary.", + "bbox": [ + 509, + 565, + 893, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of SVD-based Attribute Mamba (SAM). Comparing Row1 with Row4, we observe that adding SAM yields significant improvements across various scenarios, especially in the ST w/o D case, where the F1 score improves by $19.8\\%$ . This demonstrates SAM's ability to decompose complex representations and capture fine-grained attribute features, which is essential for distinguishing target objects based on descriptive characteristics.", + "bbox": [ + 511, + 625, + 893, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Location Mamba (LM). The impact of LM is evident when comparing Row 4 with Row 7. Adding LM enhances performance, particularly in the $MT$ scenario, where the F1 score increases by $5.3\\%$ . This improvement demonstrates the importance of LM in capturing location relationships between objects, allowing the model to better handle cases with distractors.", + "bbox": [ + 511, + 742, + 893, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Instructive Dual-Mamba (IDM). The contribution of IDM is highlighted by comparing Row 7 with Row 8. Incorporating IDM results in consistent per", + "bbox": [ + 511, + 845, + 893, + 890 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5301", + "bbox": [ + 483, + 926, + 513, + 939 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4c88b316b8665d4771a5c03d8d1bfa6093b80f44e3c464c6768755032164cf4.jpg", + "image_caption": [ + "Query Text" + ], + "image_footnote": [], + "bbox": [ + 354, + 127, + 452, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2875bf1191b9d68d0310ad3268f71a9768da104fa981fde662fc69f58cb06107.jpg", + "image_caption": [ + "The red sofa is against the wall with a white cushion." + ], + "image_footnote": [], + "bbox": [ + 354, + 210, + 457, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6ea0d90da02796e5254e24cfd42d3d3570721f420dfe02824829148dc66a6c4.jpg", + "image_caption": [ + "A wooden chair without arms is tucked under the table.", + "A black chair without armrests, back to the window.", + "Figure 6. Visualization of the attribute, location relation, and overall scores in our model." + ], + "image_footnote": [], + "bbox": [ + 352, + 291, + 457, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc5d3d221026f955a46a573647194a13f2008f34fac8ab91a578e66372bc7248.jpg", + "image_caption": [ + "Ground Truth", + "Attribute Score" + ], + "image_footnote": [], + "bbox": [ + 480, + 127, + 581, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f36053b5c8c927fd46cf62b1191df4b46c4304bc9b3c11ac65dc9a363b227926.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 210, + 581, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fdb66e3d5d2449bc1559f9d98a89fd04fe122e2cd9d3631456e74a431522a2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 295, + 584, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0debec5e2c53069c271c33197ca906284d269a6f17478993254f2de2de0c7324.jpg", + "image_caption": [ + "Location Score" + ], + "image_footnote": [], + "bbox": [ + 614, + 127, + 715, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8dbd50435c96bd44c64b9258a01632685b5f654ee4a74d0132038988e64a0bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 210, + 718, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5bf46b0ea7f5098b83cfa8cbb7447b56f7029f3eb2010d1f6499598083d5bc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 294, + 718, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/268055ec478ada3ce8efaf501374ee53beda0ffce44a2b1190f9653684901175.jpg", + "image_caption": [ + "Overall Score" + ], + "image_footnote": [], + "bbox": [ + 746, + 127, + 846, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1649b3a7550b2b3fb94000171c9d50bab1b61d6166df9e6ee6813a661721995b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 212, + 851, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faaec61c9cab9784a95c8c10509a835bc1f04fbb77c71dd48172174430dcba2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 295, + 851, + 359 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fc90ff9ba27eb698b21835d1446c2162a44a5d04cfe532dea3ef5f76348ac2b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodZT w/o DZT w/DST w/o DST w/DMTAll
Location First82.143.968.746.354.753.1
Attribute First83.645.270.946.755.154.9
", + "bbox": [ + 104, + 410, + 480, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison of grounding performance between attribute first and location first model.", + "bbox": [ + 99, + 470, + 482, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance gains across all metrics, with a notable increase in the $MT$ case, where the F1 score improves by $2\\%$ . This shows IDM's effectiveness in handling multi-object grounding, enabling precise language-instructed grounding in complex scenes with multiple target objects.", + "bbox": [ + 99, + 505, + 483, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency-Performance Trade-off. Fig. 5 illustrates the balance between computational complexity (FLOPs), inference speed, and grounding performance on the ScanRefer dataset. VGMamba achieves the highest accuracy with significantly lower FLOPs than methods like Chat-Scene, demonstrating superior efficiency. This indicates that our method achieves a superior balance between efficiency and performance, making it well-suited for real-world applications where both accuracy and inference speed are critical. The lightweight nature of VGMamba stems from its efficient attribute and location feature sequence modeling with Mamba, which enables better feature aggregation without excessive computational overhead.", + "bbox": [ + 99, + 579, + 483, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Why Attribute First? We conducted experiments on the Multi3DRefer val set to compare the performance between the location first and our proposed attribute first model. As shown in Tab. 5, the attribute first approach achieves relatively higher performance. One possible explanation is that attribute cues serve as more distinguishing priors compared to location cues, enabling the model to initially identify a larger and more accurate set of candidate objects. In con", + "bbox": [ + 99, + 771, + 483, + 888 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trast, location information is often less specific and could potentially mislead the model toward irrelevant objects.", + "bbox": [ + 511, + 412, + 893, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Interpretability Analysis", + "text_level": 1, + "bbox": [ + 511, + 450, + 732, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benefiting from the sequential attribute-to-location reasoning process, VGMamba demonstrates a high degree of interpretability. To visualize the model's understanding of attribute and location relationships, we utilize the output features from the SVD-based Attribute Mamba, Location Mamba, and Instructive Dual-Mamba modules to obtain the attribute score, location score, and overall score. Fig. 6 presents qualitative results, where darker colors indicate higher scores. Specifically, our model first identifies objects that match attribute cues, such as \"red sofa,\" \"wooden chair,\" and \"black chair.\" Then, using these attributes as prior, it incorporates location cues to filter out mismatched objects, e.g., \"against the wall,\" \"under the table,\" and \"back to the window\" (highlighted in yellow boxes). The final prediction scores, generated by the Instructive Dual-Mamba block, are highlighted in red boxes.", + "bbox": [ + 509, + 470, + 895, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 718, + 627, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we are concerned with 3D Visual Grounding that aims to identify 3D objects matching the language description. Particularly, we consider that human beings often follow two clues to achieve the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, i.e., attribute-to-location clue reasoning, and propose a new VGMamba method that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Experimental results on multiple datasets demonstrate the effectiveness of our method.", + "bbox": [ + 509, + 742, + 895, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5302", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 102, + 101, + 266, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), the Joint Fund of the Ministry of Education of China (Nos. 8091B022149, 8091B02072404), the National Natural Science Foundation of China (No. 62132016), and the Natural Science Basic Research Program of Shaanxi (No. 2020JC-23).", + "bbox": [ + 99, + 126, + 485, + 214 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 228, + 196, + 243 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 5, 6", + "[2] Chun-Peng Chang, Shaoxiang Wang, Alain Pagani, and Didier Stricker. Mikasa: Multi-key-anchor & scene-aware transformer for 3d visual grounding. In CVPR, pages 14131-14140, 2024. 1, 2, 5", + "[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221, 2020. 5, 7", + "[4] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 6", + "[5] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems, 34:5834-5847, 2021. 1", + "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 1", + "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1", + "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6", + "[9] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2", + "[10] Gene H. Golub and William Kahan. Calculating the singular values and pseudo-inverse of a matrix. In Milestones in Matrix Computation, 2007. 4", + "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3", + "[12] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + ], + "bbox": [ + 102, + 252, + 483, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 5, 6, 7", + "[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 5", + "[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 5", + "[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In ECCV, pages 289-310. Springer, 2024. 5", + "[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 3, 6", + "[18] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE transactions on multimedia, 25:8906-8919, 2023. 4", + "[19] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, pages 237-255. Springer, 2024. 2, 3", + "[20] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1", + "[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2", + "[22] Hui Lu, Albert Ali Salah, and Ronald Poppe. Videomambapro: A leap forward for mamba in video understanding. arXiv preprint arXiv:2406.19006, 2024. 2", + "[23] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 5", + "[24] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 2, 3", + "[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2", + "[26] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381-398. Springer, 2024. 5", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 514, + 103, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5303", + "bbox": [ + 482, + 926, + 514, + 938 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3", + "[28] Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2", + "[29] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 2, 5", + "[30] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2", + "[31] Ozan Unal, Christos Sakaridis, Suman Saha, and Luc Van Gool. Four ways to improve verbo-visual fusion for dense 3d visual grounding. In European Conference on Computer Vision, pages 196–213. Springer, 2024. 5", + "[32] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 2", + "[33] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13917-13926, 2024. 5", + "[34] Yuxuan Wang, Among Wu, Muli Yang, Yukuan Min, Yihang Zhu, and Cheng Deng. Reasoning mamba: Hypergraph-guided region relation calculating for weakly supervised affordance grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27618-27627, 2025. 2", + "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19231-19242, 2023. 5", + "[36] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2", + "[37] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2", + "[38] Can Xu, Yuehui Han, Rui Xu, Le Hui, Jin Xie, and Jian Yang. Multi-attribute interactions matter for 3d visual grounding. In CVPR, pages 17253–17262, 2024. 1, 2, 5", + "[39] Chenghao Xu, Guangtao Lyu, Jiexi Yan, Muli Yang, and Cheng Deng. Llm knows body language, too: Translating speech voices into human gestures. In ACL, pages 5004-5013, 2024. 2", + "[40] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1" + ], + "bbox": [ + 104, + 103, + 482, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Haomeng Zhang, Chiao-An Yang, and Raymond A Yeh. Multi-object 3d grounding with dynamic modules and language-informed spatial attention. arXiv preprint arXiv:2410.22306, 2024. 1, 2, 5, 6, 7", + "[42] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 2, 3, 5, 6, 7", + "[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model, 2024. 2", + "[44] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188-206. Springer, 2024. 5, 6, 7" + ], + "bbox": [ + 514, + 103, + 893, + 321 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5304", + "bbox": [ + 483, + 928, + 514, + 938 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_model.json b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a3c0ce2831a1c8406720390a8a7dad5f5b4c71ab --- /dev/null +++ b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.14, + 0.768, + 0.186 + ], + "angle": 0, + "content": "VGMamba: Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.212, + 0.768, + 0.247 + ], + "angle": 0, + "content": "Yihang Zhu, Jinhao Zhang, Yuxuan Wang, Among Wu, Cheng Deng* Xidian University, China" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.249, + 0.731, + 0.281 + ], + "angle": 0, + "content": "{zhuyihang, zhangjinhao, wangyuxuan}@stu.xidian.edu.cn amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.255, + 0.315, + 0.332, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.345, + 0.486, + 0.729 + ], + "angle": 0, + "content": "As an important direction of embodied intelligence, 3D Visual Grounding has attracted much attention, aiming to identify 3D objects matching the given language description. Most existing methods often follow a two-stage process, i.e., first detecting proposal objects and identifying the right objects based on the relevance to the given query. However, when the query is complex, it is difficult to leverage an abstract language representation to lock the corresponding objects accurately, affecting the grounding performance. In general, given a specific object, humans usually follow two clues to finish the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, attribute-to-location clue reasoning, to conduct accurate grounding. Particularly, we propose a VGMamba network that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Taking a 3D point cloud scene and language query as the input, we first exploit SVD to make a decomposition of the extracted features. Then, a slidingwindow operation is conducted to capture attribute characteristics. Next, a location mamba is presented to obtain the corresponding location information. Finally, by means of multi-modal mamba fusion, the model could effectively localize the object that matches the given query. In the experiment, our method is verified on four datasets. Extensive experimental results demonstrate the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.755, + 0.233, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.779, + 0.485, + 0.869 + ], + "angle": 0, + "content": "3D Visual Grounding aims to identify and localize objects in a 3D scene based on natural language descriptions. This task has become a key challenge at the intersection of computer vision and natural language processing, with significant applications in areas such as human-robot interaction and vision-language navigation [5, 7, 20]. In the context" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.315, + 0.683, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.315, + 0.882, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.495, + 0.896, + 0.563 + ], + "angle": 0, + "content": "Figure 1. To achieve accurate 3D Visual Grounding, we explore two clues. First, we leverage the attribute clue to capture proposal objects. Then, a location clue is exploited to focus on the right grounding object. Experimental results on multiple datasets verify the effectiveness of this mechanism." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.576, + 0.896, + 0.679 + ], + "angle": 0, + "content": "of 3D environments, this task becomes even more complex due to the sparsity and irregularity of 3D point cloud data, as well as the variability in object shapes, scales, and the complexity of the language descriptions. Therefore, accurately localizing objects in 3D scene based on natural language descriptions is crucial for robots to understand and interact with their surroundings." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.681, + 0.896, + 0.814 + ], + "angle": 0, + "content": "Recent studies [2, 6, 38, 40, 41] have attempted to address this challenge by incorporating both object attributes (e.g., color, shape, texture) and spatial relationships (e.g., relative positions) to improve grounding accuracy. These methods typically follow a two-stage paradigm: first detecting candidate objects in the scene, and then assessing relevance to the given query based on feature similarity. While these methods have demonstrated effectiveness in certain scenarios, they still exhibit some limitations." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.816, + 0.897, + 0.892 + ], + "angle": 0, + "content": "First, most existing works [6, 38, 40] treat object attributes and spatial relationships in isolation, neglecting the inherent dependencies between them. In human perception, object attributes often serve as priors for location reasoning. For example, when searching for \"The red chair next to the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.119, + 0.877, + 0.245, + 0.89 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.516, + 0.94 + ], + "angle": 0, + "content": "5295" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.482, + 0.278 + ], + "angle": 0, + "content": "table,\" humans first identify objects matching the attribute \"red chair\" and then refine their search based on location cues. Existing approaches lack a mechanism to systematically leverage this reasoning process, resulting in suboptimal performance in complex scenes. Furthermore, in complex 3D environments, a query may refer to multiple objects. However, existing Transformer-based methods often fail to effectively capture global dependencies among multiple objects with shared attributes and spatial relations due to their quadratic complexity in self-attention, which limits scalability, and their tendency to focus on local interactions rather than modeling long-range dependencies." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.279, + 0.483, + 0.735 + ], + "angle": 0, + "content": "To address these limitations, as shown in Fig. 1, we separately explore reasoning attribute-based information and location-based characteristic to ground corresponding 3D objects. To be specific, we propose VGMamba, a novel architecture that systematically models attribute-to-location dependencies while efficiently capturing long-range interactions. VGMamba consists of three key modules: SVD-based Attribute Mamba, which leverages Singular Value Decomposition (SVD) to decompose the extracted visual features into multiple subspaces corresponding to different attribute scales. However, SVD alone lacks long-range modeling capability, which is critical for capturing hierarchical dependencies among attributes. To overcome this, we employ Mamba [11] to sequentially integrate multi-scale attribute features, preserving their interdependencies over long sequences. Compared to Transformers, Mamba maintains linear computational complexity while effectively modeling long-range dependencies among attribute components; the Location Mamba, which models both global and local positional relationships among objects by establishing cross-channel communication through SSM, effectively enhancing object-related features; and the Instructive Dual-Mamba block, which fuses object-related and text features through cross-modal guidance of text query. By replacing computationally expensive attention mechanisms with structured state-space modeling, VGMamba significantly improves efficiency without sacrificing performance. Extensive experiments demonstrate that our method achieves state-of-the-art performance, particularly excelling in scenarios involving multiple objects and ambiguous descriptions. Our chief contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.738, + 0.483, + 0.767 + ], + "angle": 0, + "content": "- We explore a novel mechanism, i.e., attribute-to-location clue reasoning, for performing 3D visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.768, + 0.483, + 0.84 + ], + "angle": 0, + "content": "- We propose a novel framework VGMamba, comprising three core modules: the SVD-based Attribute Mamba for disentangling attributes, the Location Mamba for modeling location relations, and the Instructive Dual-Mamba block for text and object-related feature fusion." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.84, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We achieve new state-of-the-art performance on multiple 3D Visual Grounding benchmarks, demonstrating superior accuracy and computational efficiency." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.738, + 0.483, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.102, + 0.651, + 0.117 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.126, + 0.895, + 0.331 + ], + "angle": 0, + "content": "3D Visual Grounding. Recent advancements incorporate self-attention-based scene encoding [2], viewpoint-aware spatial reasoning [29], and multi-attribute Transformers for fine-grained language-visual alignment [38]. However, these attention-based approaches struggle with long-sequence feature modeling and computational efficiency. For multi-object grounding, M3DRef-CLIP [42] introduced the first multi-object dataset and an attention-based fusion module integrating 3D object and 2D rendered image features. Subsequent works [41] refine proposal selection, camera pose optimization, and fusion strategies. However, these methods often embed global 2D features, introducing redundancy. In contrast, we focus on fine-grained attribute extraction as priors to enhance spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.895, + 0.537 + ], + "angle": 0, + "content": "State Space Models. State Space Models (SSMs) [12, 30, 34] have become a prominent approach in sequence modeling, particularly for capturing long-range dependencies with linear time complexity. Furthermore, Mamba [11] introduced a selective state-space mechanism, providing efficient sequence modeling and surpassing traditional Transformers in certain tasks. This advancement led to various extensions across vision applications, including Vision Mamba [43] and VMamba [21], which extends scanning in four directions to fully capture interrelations among image patches. Besides, Mamba-based models have also been applied in video understanding [19, 22, 25, 39] and medical image segmentation [24, 28, 36, 37], demonstrating the versatility and efficiency of SSMs in diverse visual domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.894, + 0.625 + ], + "angle": 0, + "content": "In this work, we introduce Mamba into the object grounding task for the first time, exploring its potential in fine-grained feature extraction. We also propose a novel cross-modal fusion framework for language and object-related features, which achieves outstanding performance and significantly reduces computational costs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.799, + 0.656 + ], + "angle": 0, + "content": "3. Overview of State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.894, + 0.794 + ], + "angle": 0, + "content": "Recently, state space models (SSMs) [9, 12, 30, 32] have attracted much attention for their ability to model continuous systems, constructing the foundation for the popular Mamba architecture [11]. Particularly, SSMs generally take an input sequence \\( x(t) \\in \\mathbb{R}^L \\) as the input and output the corresponding sequence \\( y(t) \\in \\mathbb{R}^L \\) through hidden states \\( h(t) \\in \\mathbb{R}^N \\), where \\( N \\) is the number of hidden states. The system is governed by differential equations that describe how the hidden state evolves over time:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.805, + 0.894, + 0.822 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.832, + 0.894, + 0.86 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}\\), \\(\\mathbf{B}\\), and \\(\\mathbf{C}\\) are matrices that project the input into the hidden states and subsequently to the output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.861, + 0.894, + 0.891 + ], + "angle": 0, + "content": "In order to apply SSMs in practice tasks, a discretization process is employed, typically using a time scale parameter" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.1, + 0.621, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.104, + 0.898, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.386, + 0.897, + 0.442 + ], + "angle": 0, + "content": "Figure 2. VGMamba for quantity-agnostic 3D Visual Grounding. Based on the given language description, we could utilize two clues to localize corresponding objects accurately, i.e., attribute and location information. To this end, we first design an SVD-based attribute mamba to discover multiple attribute-related objects. Then, a location mamba is further designed to select location-relevant objects. Finally, we present an Instructive Dual-Mamba block to localize the object that matches the given query." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.497 + ], + "angle": 0, + "content": "\\(\\Delta\\) to convert continuous parameters into discrete ones. The processes are shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.506, + 0.483, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} (\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}) \\cdot \\Delta \\mathbf {B}. (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.535, + 0.466, + 0.55 + ], + "angle": 0, + "content": "The discretized state-space output can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.56, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\nh _ {k} = \\overline {{\\mathbf {A}}} h _ {k - 1} + \\overline {{\\mathbf {B}}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.588, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Based on the SSMs framework, Mamba [11] employs the Selective Scan Mechanism (S6) to conduct dynamic sequential modeling and capture critical task-related information. Though Mamba-based networks have been verified to be effective in many tasks [19, 24], it is under-explored for 3D Visual Grounding. To this end, we study whether the Mamba could be used to capture the relation between objects and language descriptions. We directly plug Mamba into M3DRef-CLIP [42] and conduct experiments on Multi3DRefer dataset [42]. However, we observe that this operation could not improve the grounding performance. The reason may be that Mamba emphasizes capturing global object information and cannot construct the relation between objects and language descriptions." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.807, + 0.426, + 0.825 + ], + "angle": 0, + "content": "4. VGMamba for 3D Visual Grounding" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.832, + 0.484, + 0.892 + ], + "angle": 0, + "content": "In this paper, we propose a new VGMamba architecture, aiming to reason attribute and location clues based on the given language descriptions and improve the performance of 3D Visual Grounding effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Specifically, in Fig. 2 (a), we follow the work [42] and adopt PointGroup [17] as a 3D detector to generate an object proposal list \\(o = (o_1,o_2,\\dots,o_n)\\) with \\(n\\) proposal boxes and corresponding features \\(F_{\\mathrm{3D}}\\in \\mathbb{R}^{n\\times 32}\\). Each proposal is then rendered as \\(m\\) multi-view images, which are encoded by CLIP [27] to extract \\(d\\)-dimensional patch features \\(I\\in \\mathbb{R}^{n\\times m\\times p\\times d}\\) and aggregated features \\(\\hat{I}\\in \\mathbb{R}^{n\\times m\\times d}\\), where \\(p\\) is the patch number of each multiview image. Next, for the given language description, we continue following the work [42] and utilize CLIP to obtain text feature \\(T\\in \\mathbb{R}^d\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.894, + 0.734 + ], + "angle": 0, + "content": "In practice, 3D Visual Grounding can be achieved by reasoning two key clues. One clue is to accurately recognize object-related attribute, e.g., dark blue. Another clue is to effectively obtain the location-relevant visual content, e.g., facing the backboard. With this in mind, we separately design an attribute mamba and location mamba to sufficiently capture object-related attribute and location characteristic." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.77, + 0.765 + ], + "angle": 0, + "content": "4.1. SVD-based Attribute Mamba" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.773, + 0.895, + 0.891 + ], + "angle": 0, + "content": "In general, given a specific object, the extracted representations involve multiple different attribute characteristics, e.g., color, shape, and texture. However, the 3D Visual Grounding task involves localizing objects that possess specific attributes, such as dark blue. Therefore, it is meaningful to disentangle current representations. To this end, we explore using Singular Value Decomposition (SVD) to perform feature-level decomposition and recombination ac" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.94 + ], + "angle": 0, + "content": "5297" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.102, + 0.103, + 0.366, + 0.117 + ], + "angle": 0, + "content": "cording to the calculated singular values." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.118, + 0.484, + 0.236 + ], + "angle": 0, + "content": "Particularly, as illustrated in Fig. 2 (b), for each patch feature \\(I_{i}\\in \\mathbb{R}^{d\\times p}\\), SVD is first used to decompose \\(I_{i}\\) into three matrices: \\(I_{i} = U\\Sigma V^{T}\\in \\mathbb{R}^{d\\times p}\\), where \\(U\\in \\mathbb{R}^{d\\times d}\\) and \\(V\\in \\mathbb{R}^{p\\times p}\\) are orthogonal matrices, and \\(\\Sigma\\) contains singular values sorted along its main diagonal [10]. Subsequently, inspired by Multi-Scale Attention [18], we introduce Sliding Window to recombine the singular vectors using three sliding windows with varying size \\(r\\):" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.246, + 0.483, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\left\\{F _ {r} \\right\\} = \\text {S l i d i n g W i n d o w} (\\Sigma , r), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.265, + 0.483, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\text {S l i d i n g W i n d o w} (\\Sigma , r) = U _ {d \\times r} \\Sigma_ {r \\times r} V _ {r \\times p} ^ {T}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.293, + 0.484, + 0.501 + ], + "angle": 0, + "content": "where \\(r \\in \\{r_1, r_2, r_3\\}\\). SlidingWindow refers to combining r consecutive singular vectors into a new multi-scale feature \\(F_r\\). Since the singular vectors corresponding large singular values and those corresponding small singular values separately contain generalized coarse-grained information and discriminative fine-grained information, through multiscale operation, we can sufficiently capture attribute information involved in visual representations. Following this, \\(F_r\\) is processed by three independent Mamba to model attribute feature. Finally, a concatenation followed by a \\(1 \\times 1\\) convolution is applied to merge the multi-scale features, yielding the final feature representations \\(F_{i-att} \\in \\mathbb{R}^d\\) of the single-view image. For \\(n\\) objects with \\(m\\) rendered images, we can obtain the feature representation \\(F_{att} \\in \\mathbb{R}^{n \\times m \\times d}\\)." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.508, + 0.267, + 0.523 + ], + "angle": 0, + "content": "4.2. Location Mamba" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.529, + 0.484, + 0.632 + ], + "angle": 0, + "content": "As shown in Fig. 2 (a), only relying on the attribute clue could not guarantee accurately identifying the objects corresponding to the given language description. To this end, based on the captured attribute features, we design a cross-channel communication mechanism to facilitate the learning of multi-object location relationships, enforcing the locking of the object matching the given description." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.632, + 0.484, + 0.72 + ], + "angle": 0, + "content": "Specifically, to capture spatial relationships from a consistent viewpoint, for each rendered object image, we define an encoding \\(pos \\in \\mathbb{R}^{n \\times m \\times 6}\\) that includes object center coordinates \\((x,y,z)\\) and bounding box dimensions \\((h,w,l)\\). Following this, we utilize a linear layer to perform a transformation on \\(pos\\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.73, + 0.483, + 0.748 + ], + "angle": 0, + "content": "\\[\np \\hat {o} s = \\operatorname {L i n e a r} (p o s) \\in \\mathbb {R} ^ {n \\times m \\times d}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.758, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Subsequently, the summation result \\( I_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times m \\times d} \\) of \\( p \\hat{o} s \\), \\( \\hat{I} \\in \\mathbb{R}^{n \\times m \\times d} \\), and \\( F_{\\mathrm{att}} \\in \\mathbb{R}^{n \\times m \\times d} \\) is taken to perform the following operations: exploring a dedicated mamba to respectively conduct communication from global and local perspectives. For global communication, the channels of \\( n \\) proposal objects are treated as a sequence to enable cross-object interaction:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.875, + 0.483, + 0.892 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {g l o b a l}} = \\operatorname {M a m b a} _ {\\text {g l o b a l}} \\left(I _ {\\text {s p a}}\\right), \\tag {7}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.102, + 0.895, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.214, + 0.895, + 0.242 + ], + "angle": 0, + "content": "Figure 3. The architecture of Instructive Dual-Mamba (IDM) for conducting multi-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.833, + 0.271 + ], + "angle": 0, + "content": "where \\(I_{\\mathrm{global}} \\in \\mathbb{R}^{n \\times m \\times d}\\) denotes the SSM output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.894, + 0.346 + ], + "angle": 0, + "content": "As for the local communication, the \\(u\\) nearest neighbors of each object are selected based on Euclidean distance, obtaining multi-neighbor feature \\(\\hat{I}_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times m \\times u \\times d}\\). Cross-channel scanning is then applied to \\(\\hat{I}_{\\mathrm{spa}}\\) and outputs central object features:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.356, + 0.894, + 0.373 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {l o c a l}} = \\operatorname {S e l e c t C e n t e r} \\left(\\mathrm {M a m b a} _ {\\text {l o c a l}} \\left(\\hat {I} _ {\\mathrm {s p a}}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.895, + 0.502 + ], + "angle": 0, + "content": "where \\(I_{\\mathrm{local}} \\in \\mathbb{R}^{n \\times m \\times d}\\) denotes the current SSM output. Subsequently, by summing \\(I_{\\mathrm{global}}\\) and \\(I_{\\mathrm{local}}\\), and then averaging over the \\(m\\) images, we obtain the feature \\(F_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times d}\\) that contains plentiful location information. Finally, \\(F_{\\mathrm{spa}}\\) is concatenated with the proposal feature \\(F_{3\\mathrm{D}} \\in \\mathbb{R}^{n \\times 32}\\) output by the 3D detector, followed by a \\(1 \\times 1\\) convolution layer for dimensionality reduction, yielding \\(F_{\\mathrm{obj}} \\in \\mathbb{R}^{n \\times d}\\). We represent the feature of each object in \\(F_{\\mathrm{obj}}\\) as \\(G \\in \\mathbb{R}^d\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.509, + 0.761, + 0.523 + ], + "angle": 0, + "content": "4.3. Multi-modal Mamba Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.895, + 0.83 + ], + "angle": 0, + "content": "To further enhance the grounding accuracy, we present a multi-modal mamba fusion module to align the proposals to the corresponding language descriptions. Specifically, as shown in Fig. 3, the Instructive Dual-Mamba (IDM) block is proposed to use descriptions as guidance to capture matched object-related feature. Taking the language feature \\( T \\in \\mathbb{R}^d \\), object-related feature \\( G \\in \\mathbb{R}^d \\) as the input, we first divide \\( T \\) into a forward sequence and a reversed backward sequence, which are then processed by two independent SSMs (Eq. (1)) respectively. The outputs from the two branches are integrated to obtain the updated feature. The processing of input feature \\( G \\in \\mathbb{R}^d \\) is similar to \\( T \\in \\mathbb{R}^d \\). According to Eq. (1), the state transition matrices \\( A \\) retain critical historical information for updating hidden states, while \\( B \\) and \\( C \\) primarily influence input and output. Therefore, we propose replacing the state transition matrices \\( \\vec{\\mathbf{A}}_G, \\vec{\\mathbf{A}}_G \\) with their counterparts \\( \\vec{\\mathbf{A}}_T, \\vec{\\mathbf{A}}_T \\) derived from language features, enabling cross-modal guidance in extracting object-related features. The remaining parameter matrices remain unchanged. The process of above could be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.846, + 0.894, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\overleftarrow {h} _ {\\mathrm {T}} ^ {t} = \\overleftarrow {\\mathbf {A}} _ {\\mathrm {T}} \\overleftarrow {h} _ {\\mathrm {T}} ^ {t - 1} + \\overleftarrow {B} _ {\\mathrm {T}} \\overleftarrow {x} _ {\\mathrm {T}} ^ {t}, \\quad \\overrightarrow {h} _ {\\mathrm {T}} ^ {t} = \\overrightarrow {\\mathbf {A}} _ {\\mathrm {T}} \\overrightarrow {h} _ {\\mathrm {T}} ^ {t - 1} + \\overrightarrow {B} _ {\\mathrm {T}} \\overrightarrow {x} _ {\\mathrm {T}} ^ {t}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.867, + 0.894, + 0.887 + ], + "angle": 0, + "content": "\\[\n\\overleftarrow {h} _ {\\mathsf {G}} ^ {t} = \\overleftarrow {\\mathbf {A}} _ {\\mathsf {T}} \\overleftarrow {h} _ {\\mathsf {G}} ^ {t - 1} + \\overleftarrow {B} _ {\\mathsf {G}} \\overleftarrow {x} _ {\\mathsf {G}} ^ {t}, \\quad \\overrightarrow {h} _ {\\mathsf {G}} ^ {t} = \\overrightarrow {\\mathbf {A}} _ {\\mathsf {T}} \\overrightarrow {h} _ {\\mathsf {G}} ^ {t - 1} + \\overrightarrow {B} _ {\\mathsf {G}} \\overrightarrow {x} _ {\\mathsf {G}} ^ {t}, \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5298" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.207, + 0.1, + 0.794, + 0.387 + ], + "angle": 0, + "content": "
MethodVenueUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [3]ECCV2067.646.232.121.340.026.1
TGNN [14]AAAI2168.656.829.823.237.429.7
MVT [15]CVPR2277.766.531.925.340.833.3
3D-SPS [23]CVPR2284.166.740.329.848.837.0
EDA [35]CVPR2385.868.649.137.654.642.3
VPP-Net [29]CVPR2486.167.150.339.055.743.3
M3DRef-CLIP [42]ICCV23-77.2-36.8-44.7
G3-LQ [33]CVPR2488.673.350.239.756.044.7
MCLN [26]ECCV2486.972.752.040.857.245.5
MA2TransVG [38]CVPR2486.374.153.841.457.945.7
ConcreteNet [31]ECCV2486.482.142.438.450.646.5
D-LISA [41]NeurIPS24-75.5-40.0-46.9
GPS [16]ECCV24-77.9-42.7-48.1
Chat-Scene [13]NeurIPS2489.682.547.842.955.550.2
PQ3D [44]ECCV2486.778.351.546.257.051.2
VGMamba-91.979.654.847.660.053.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.397, + 0.897, + 0.425 + ], + "angle": 0, + "content": "Table 1. Comparison results on the ScanRefer [3] dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. Our proposed VGMamba performs favorably over current state-of-the-art approach PQ3D [44]." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.447, + 0.483, + 0.512 + ], + "angle": 0, + "content": "Here, \\(x_{\\mathrm{T / G}}^t\\) represents the input at time step \\(t\\). \\(\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{T}}\\) and \\(\\overrightarrow{\\mathbf{A}}_{\\mathrm{T}}\\) are the state transition matrices used for retaining essential historical information and updating hidden states, while matrices \\(B_{\\mathrm{T / G}}\\) primarily affect input." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.514, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Additionally, we share the gating mechanism parameters \\(z\\) of the language features to construct cross-modal fusion in the hidden state space, ultimately outputting the average multi-view attribute features of each object in the scene, represented as \\(F_{\\mathrm{att - spa}} \\in \\mathbb{R}^{n \\times d}\\). Finally, a convolution operation is applied to output confidence scores \\(O \\in \\mathbb{R}^{n \\times 1}\\) for each object proposal." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.639, + 0.287, + 0.656 + ], + "angle": 0, + "content": "4.4. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.665, + 0.484, + 0.724 + ], + "angle": 0, + "content": "Building on previous work [42], the loss of VGMamba consists of the 3D Visual Grounding loss \\( L_{\\mathrm{ref}} \\), text-object contrastive loss \\( L_{\\mathrm{con}} \\), and object detection loss \\( L_{\\mathrm{det}} \\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.745, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {r e f}} + L _ {\\text {c o n}} + L _ {\\text {d e t}}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.787, + 0.231, + 0.804 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.816, + 0.484, + 0.89 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our proposed VGMamba on 3D Visual Grounding tasks with varying numbers of target objects, we conducted experiments on both single-object datasets (ScanRefer [3], Nr3D and Sr3D [1]) and multi-object dataset Multi3DRefer [42]." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.447, + 0.892, + 0.719 + ], + "angle": 0, + "content": "
MethodNr3DSr3D
OverallHardOverallHard
ReferIt3D [?]35.627.940.831.5
TGNN [14]37.330.645.036.9
M3DRef-CLIP [42]49.443.4--
3D-SPS [23]51.545.162.665.4
EDA [35]52.146.168.162.9
D-LISA [41]53.146.2--
MVT [15]55.449.164.558.8
VPP-Net [29]56.9-68.7-
G3-LQ [33]58.450.773.166.3
MCLN [26]59.8-68.4-
MiKASA [2]64.459.475.267.3
GPS [16]64.957.877.571.6
MA2TransVG [38]65.257.673.969.3
PQ3D [44]66.758.779.772.8
VGMamba68.361.481.374.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.728, + 0.895, + 0.757 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on Nr3D and Sr3D dataset. We have highlighted the top-performing three methods in purple." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.733, + 0.803 + ], + "angle": 0, + "content": "5.1. Single-object Grounding" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.81, + 0.779, + 0.824 + ], + "angle": 0, + "content": "5.1.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.895, + 0.89 + ], + "angle": 0, + "content": "ScanRefer. The ScanRefer dataset comprises 51,583 natural language descriptions for 11,046 objects across 800 3D scenes from the ScanNet dataset [8]. Each description specifies a target object and its location relations within" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.182, + 0.106, + 0.245, + 0.117 + ], + "angle": 0, + "content": "Query Text" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.153, + 0.303, + 0.175 + ], + "angle": 0, + "content": "The green chair next to the brown table with a white vase." + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.223, + 0.291, + 0.246 + ], + "angle": 0, + "content": "The brown chairs under the white rectangular table." + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.297, + 0.293, + 0.321 + ], + "angle": 0, + "content": "The white pillow closest to the wall with the painting." + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.106, + 0.409, + 0.117 + ], + "angle": 0, + "content": "M3DRef-CLIP" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.131, + 0.425, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.2, + 0.424, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.277, + 0.421, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.487, + 0.106, + 0.53, + 0.116 + ], + "angle": 0, + "content": "D-LISA" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.131, + 0.563, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.198, + 0.565, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.278, + 0.559, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.101, + 0.684, + 0.123 + ], + "angle": 0, + "content": "VGMamba (ours)" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.131, + 0.716, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.198, + 0.71, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.279, + 0.706, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.108, + 0.833, + 0.116 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.131, + 0.852, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.197, + 0.852, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.278, + 0.847, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.347, + 0.894, + 0.374 + ], + "angle": 0, + "content": "Figure 4. Qualitative examples of the Multi3DRefer dataset. The results illustrate the grounding performance of M3DRef-CLIP (blue box), D-LISA (blue box), and our VGMamba(green box), compared to the ground truth labels (red box)." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.398, + 0.48, + 0.508 + ], + "angle": 0, + "content": "
MethodZT w/o DZT w/DST w/o DST w/DMTAll
D3Net [4]81.632.538.623.335.032.2
M3DRef-CLIP [42]81.839.447.830.637.938.4
M3DRef-CLIP w/NMS [41]79.040.567.640.049.149.3
PQ3D [44]85.457.768.543.640.950.1
D-LISA [41]82.443.767.142.551.051.2
VGMamba83.645.270.946.755.154.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.518, + 0.482, + 0.545 + ], + "angle": 0, + "content": "Table 3. Comparison of F1@0.5 results on the Multi3DRefer [42] validation set, where our method outperforms all baselines." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.574, + 0.483, + 0.707 + ], + "angle": 0, + "content": "the scene, enabling the model to understand both object attributes and location context. The dataset is divided into two subsets: Unique, where the target object is the only instance of its class within the scene, and Multiple, where multiple instances of the same class are present. We evaluate our model on validation set using Acc@0.25 and Acc@0.5, which measure the proportion of predicted bounding boxes that achieve an Intersection over Union (IoU) larger than 0.25 and 0.5 with the ground truth box." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.713, + 0.483, + 0.891 + ], + "angle": 0, + "content": "\\( \\text{Nr3D/Sr3D} \\). The \\( \\text{Nr3D} \\) and \\( \\text{Sr3D} \\) datasets, part of the ReferIt3D benchmark [1], are built upon the ScanNet dataset [8] for 3D visual grounding. \\( \\text{Nr3D} \\) (Natural Reference in 3D) consists of 41,503 human-annotated referring expressions across 707 indoor scenes, where each description uniquely identifies a target object among distractors. \\( \\text{Sr3D} \\) (Spatial Reference in 3D) contains 83,572 machine-generated expressions based on predefined templates, focusing on spatial relationships between objects. Both datasets provide ground-truth object masks. Following [1], performance is evaluated using accuracy, measuring the proportion of correctly localized target objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.4, + 0.717, + 0.415 + ], + "angle": 0, + "content": "5.1.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.42, + 0.895, + 0.581 + ], + "angle": 0, + "content": "We implement the proposed VGMamba model using PyTorch and train it end-to-end on a single NVIDIA A6000 GPU. The model is optimized using the AdamW optimizer with a batch size of 4 and a learning rate of 5e-4 with cosine decay scheduling. Following prior work [42], we employ a pre-trained PointGroup [17] module as the detector, which is fine-tuned on the ScanNet dataset. The proposal box with the highest matching score is selected as the final prediction. We adopt the official train/val/test splits provided by the ScanRefer and \\(\\mathrm{Nr3D / Sr3D}\\) datasets to ensure fair comparisons with baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.702, + 0.606 + ], + "angle": 0, + "content": "5.1.3. Baseline Comparison" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.61, + 0.895, + 0.815 + ], + "angle": 0, + "content": "ScanRefer. The experimental results on ScanRefer are shown in Tab. 1, with the following key observations: (i) Our method achieves state-of-the-art performance with an overall accuracy of \\(60.0\\%\\) at IoU 0.25 and \\(53.9\\%\\) at IoU 0.5, surpassing the best baseline PQ3D [44] by \\(3.0\\%\\) and \\(2.7\\%\\), respectively. (ii) In the Multiple case, VGMamba attains \\(54.8\\%\\) at IoU 0.25 and \\(47.6\\%\\) at IoU 0.5, outperforming PQ3D and demonstrating that its attribute-to-location reasoning effectively disambiguates objects under complex scenarios. (iii) Unlike previous methods [13, 44] that show notable performance variations across settings, our VGMamba maintains consistently high accuracy, validating its robust generalization capability in both unique and multiple-object scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.894, + 0.89 + ], + "angle": 0, + "content": "\\(Nr3D/Sr3D\\). Tab. 2 presents the grounding accuracy of different methods on the \\(\\mathrm{Nr3D}\\) and Sr3D datasets. VGMamba achieves state-of-the-art performance of \\(68.3\\%\\) and \\(81.3\\%\\) on both datasets, surpassing prior state-of-the-art methods. The \\(\\mathrm{Nr3D}\\) dataset presents additional challenges due to" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5300" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.101, + 0.482, + 0.233 + ], + "angle": 0, + "content": "
ModulesMulti3DRefer (F1@0.5 on Val)
RowIDMLMSAMZT w/o DZT w/DST w/o DST w/DMTAll
1XXX81.839.447.830.637.938.4
2✓XX83.943.756.839.743.542.3
3X✓X80.240.761.142.845.143.8
4XX✓81.544.067.643.947.845.2
5✓✓X83.546.165.542.348.648.3
6✓X✓82.948.268.745.849.049.2
7X✓✓85.445.369.244.253.152.5
8✓✓✓83.645.270.946.755.154.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.25, + 0.477, + 0.264 + ], + "angle": 0, + "content": "Table 4. Ablation study of proposed modules on Multi3DRefer." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.41 + ], + "angle": 0, + "content": "its intricate and free-form textual descriptions, which increase the difficulty of cross-modal alignment. Despite this, our method significantly outperforms alternatives that leverage large-scale 3D vision-language pre-training models [13, 44]. Notably, in the more challenging \"Hard\" subset, VGMamba improves accuracy by \\(2.7\\%\\) in \\(\\mathrm{Nr3D}\\) and \\(1.6\\%\\) in Sr3D, demonstrating its superior capability in distinguishing ambiguous objects by effectively capturing fine-grained attribute and spatial relationships." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.418, + 0.319, + 0.433 + ], + "angle": 0, + "content": "5.2. Multi-object Grounding" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.44, + 0.368, + 0.453 + ], + "angle": 0, + "content": "5.2.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.457, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Multi3DRefer is a dataset derived from ScanRefer [3], comprising a total of 61,926 textual descriptions corresponding to 11,609 distinct objects. Each description may reference zero, one, or multiple target objects, enabling the evaluation of grounding models in a variety of scenarios. Following previous work[42], we compute the F1 score at the Intersection over union (IoU) threshold of 0.5(denoted as F1@0.5) across five distinct scenarios: a) zero target without distractors of the same semantic class (ZT w/o D); b) zero target with distractors (ZT w/D); c) single target without distractors (ST w/o D); d) single target with distractors (ST w/D); and e) multiple targets (MT). Finally, the average across these five evaluation cases is calculated to provide an overall performance score." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.67, + 0.307, + 0.684 + ], + "angle": 0, + "content": "5.2.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.688, + 0.484, + 0.792 + ], + "angle": 0, + "content": "The experimental setting for the multi-object dataset is similar to that for the single-object dataset, with a few differences: during training, we treat all predicted bounding boxes with an IoU greater than \\(\\tau_{\\mathrm{train}} = 0.25\\) relative to the ground truth as target bounding boxes. During inference, we consider all predicted bounding boxes with a matching score greater than \\(\\tau_{\\mathrm{pred}} = 0.1\\) as the model's predictions." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.798, + 0.292, + 0.813 + ], + "angle": 0, + "content": "5.2.3. Baseline Comparison" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.816, + 0.485, + 0.89 + ], + "angle": 0, + "content": "We compare our 3D multi-object grounding performance with existing works quantitatively and qualitatively. As illustrated in Tab. 3, our proposed VGMamba achieves competitive performance across all scenarios. In cases with distractors (ST w/D), VGMamba achieves F1 scores of" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.1, + 0.858, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.255, + 0.872, + 0.269 + ], + "angle": 0, + "content": "Figure 5. Efficiency-Performance Trade-off on ScanRefer." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.283, + 0.895, + 0.386 + ], + "angle": 0, + "content": "\\(46.7\\%\\), surpassing the second-best competitor by \\(3.1\\%\\), which highlights its robustness in managing complex scenes with multiple potential matches. In the MT case, VG-Mamba attains an F1 score that is \\(4.1\\%\\) higher than the second-best competitor, underscoring the advantage of the Instructive Dual-Mamba (IDM) in addressing the multi-object grounding task." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.387, + 0.895, + 0.533 + ], + "angle": 0, + "content": "We show the qualitative comparison with M3DRef-CLIP[42] and D-LISA[41] in Fig. 4. The top two rows illustrate examples involving multiple target objects in complex scenes. Our VGMamba accurately locates the objects matching the given textual descriptions, even in the presence of distractors. The last row demonstrates a single-target case with surrounding distractors. VGMamba effectively isolates and identifies the correct object, while baseline models struggle, often mistakenly including distractors in their predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.673, + 0.558 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.566, + 0.895, + 0.624 + ], + "angle": 0, + "content": "To validate the effectiveness of each proposed module within our VGMamba framework, we conduct ablation studies on the Multi3DRefer dataset, as shown in Tab. 4. More ablation results are detailed in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.895, + 0.742 + ], + "angle": 0, + "content": "Effectiveness of SVD-based Attribute Mamba (SAM). Comparing Row1 with Row4, we observe that adding SAM yields significant improvements across various scenarios, especially in the ST w/o D case, where the F1 score improves by \\(19.8\\%\\). This demonstrates SAM's ability to decompose complex representations and capture fine-grained attribute features, which is essential for distinguishing target objects based on descriptive characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.895, + 0.844 + ], + "angle": 0, + "content": "Effectiveness of Location Mamba (LM). The impact of LM is evident when comparing Row 4 with Row 7. Adding LM enhances performance, particularly in the \\(MT\\) scenario, where the F1 score increases by \\(5.3\\%\\). This improvement demonstrates the importance of LM in capturing location relationships between objects, allowing the model to better handle cases with distractors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.846, + 0.895, + 0.891 + ], + "angle": 0, + "content": "Effectiveness of Instructive Dual-Mamba (IDM). The contribution of IDM is highlighted by comparing Row 7 with Row 8. Incorporating IDM results in consistent per" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.514, + 0.94 + ], + "angle": 0, + "content": "5301" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.103, + 0.272, + 0.114 + ], + "angle": 0, + "content": "Query Text" + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.152, + 0.329, + 0.179 + ], + "angle": 0, + "content": "The red sofa is against the wall with a white cushion." + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.235, + 0.333, + 0.261 + ], + "angle": 0, + "content": "A wooden chair without arms is tucked under the table." + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.314, + 0.329, + 0.34 + ], + "angle": 0, + "content": "A black chair without armrests, back to the window." + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.102, + 0.444, + 0.113 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.128, + 0.454, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.212, + 0.458, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.292, + 0.458, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.103, + 0.576, + 0.113 + ], + "angle": 0, + "content": "Attribute Score" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.128, + 0.583, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.211, + 0.582, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.296, + 0.586, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.102, + 0.708, + 0.113 + ], + "angle": 0, + "content": "Location Score" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.128, + 0.716, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.212, + 0.719, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.295, + 0.72, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.102, + 0.831, + 0.113 + ], + "angle": 0, + "content": "Overall Score" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.128, + 0.848, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.213, + 0.852, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.296, + 0.852, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.374, + 0.755, + 0.388 + ], + "angle": 0, + "content": "Figure 6. Visualization of the attribute, location relation, and overall scores in our model." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.411, + 0.481, + 0.46 + ], + "angle": 0, + "content": "
MethodZT w/o DZT w/DST w/o DST w/DMTAll
Location First82.143.968.746.354.753.1
Attribute First83.645.270.946.755.154.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.471, + 0.483, + 0.498 + ], + "angle": 0, + "content": "Table 5. Comparison of grounding performance between attribute first and location first model." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.506, + 0.484, + 0.579 + ], + "angle": 0, + "content": "formance gains across all metrics, with a notable increase in the \\(MT\\) case, where the F1 score improves by \\(2\\%\\). This shows IDM's effectiveness in handling multi-object grounding, enabling precise language-instructed grounding in complex scenes with multiple target objects." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.58, + 0.484, + 0.771 + ], + "angle": 0, + "content": "Efficiency-Performance Trade-off. Fig. 5 illustrates the balance between computational complexity (FLOPs), inference speed, and grounding performance on the ScanRefer dataset. VGMamba achieves the highest accuracy with significantly lower FLOPs than methods like Chat-Scene, demonstrating superior efficiency. This indicates that our method achieves a superior balance between efficiency and performance, making it well-suited for real-world applications where both accuracy and inference speed are critical. The lightweight nature of VGMamba stems from its efficient attribute and location feature sequence modeling with Mamba, which enables better feature aggregation without excessive computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.772, + 0.484, + 0.89 + ], + "angle": 0, + "content": "Why Attribute First? We conducted experiments on the Multi3DRefer val set to compare the performance between the location first and our proposed attribute first model. As shown in Tab. 5, the attribute first approach achieves relatively higher performance. One possible explanation is that attribute cues serve as more distinguishing priors compared to location cues, enabling the model to initially identify a larger and more accurate set of candidate objects. In con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.895, + 0.444 + ], + "angle": 0, + "content": "trast, location information is often less specific and could potentially mislead the model toward irrelevant objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.733, + 0.466 + ], + "angle": 0, + "content": "5.4. Interpretability Analysis" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.472, + 0.896, + 0.707 + ], + "angle": 0, + "content": "Benefiting from the sequential attribute-to-location reasoning process, VGMamba demonstrates a high degree of interpretability. To visualize the model's understanding of attribute and location relationships, we utilize the output features from the SVD-based Attribute Mamba, Location Mamba, and Instructive Dual-Mamba modules to obtain the attribute score, location score, and overall score. Fig. 6 presents qualitative results, where darker colors indicate higher scores. Specifically, our model first identifies objects that match attribute cues, such as \"red sofa,\" \"wooden chair,\" and \"black chair.\" Then, using these attributes as prior, it incorporates location cues to filter out mismatched objects, e.g., \"against the wall,\" \"under the table,\" and \"back to the window\" (highlighted in yellow boxes). The final prediction scores, generated by the Instructive Dual-Mamba block, are highlighted in red boxes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.719, + 0.629, + 0.733 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.743, + 0.896, + 0.89 + ], + "angle": 0, + "content": "In this paper, we are concerned with 3D Visual Grounding that aims to identify 3D objects matching the language description. Particularly, we consider that human beings often follow two clues to achieve the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, i.e., attribute-to-location clue reasoning, and propose a new VGMamba method that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Experimental results on multiple datasets demonstrate the effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5302" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.103, + 0.102, + 0.267, + 0.119 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.127, + 0.486, + 0.215 + ], + "angle": 0, + "content": "Our work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), the Joint Fund of the Ministry of Education of China (Nos. 8091B022149, 8091B02072404), the National Natural Science Foundation of China (No. 62132016), and the Natural Science Basic Research Program of Shaanxi (No. 2020JC-23)." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.229, + 0.197, + 0.244 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.253, + 0.483, + 0.307 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.309, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[2] Chun-Peng Chang, Shaoxiang Wang, Alain Pagani, and Didier Stricker. Mikasa: Multi-key-anchor & scene-aware transformer for 3d visual grounding. In CVPR, pages 14131-14140, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.364, + 0.483, + 0.405 + ], + "angle": 0, + "content": "[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.407, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[4] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.462, + 0.483, + 0.515 + ], + "angle": 0, + "content": "[5] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems, 34:5834-5847, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.517, + 0.483, + 0.583 + ], + "angle": 0, + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.586, + 0.483, + 0.64 + ], + "angle": 0, + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.642, + 0.483, + 0.708 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.71, + 0.483, + 0.763 + ], + "angle": 0, + "content": "[9] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.766, + 0.483, + 0.806 + ], + "angle": 0, + "content": "[10] Gene H. Golub and William Kahan. Calculating the singular values and pseudo-inverse of a matrix. In Milestones in Matrix Computation, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.808, + 0.483, + 0.847 + ], + "angle": 0, + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.849, + 0.483, + 0.89 + ], + "angle": 0, + "content": "[12] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.253, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.17 + ], + "angle": 0, + "content": "[13] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.172, + 0.894, + 0.224 + ], + "angle": 0, + "content": "[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.226, + 0.894, + 0.265 + ], + "angle": 0, + "content": "[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.267, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In ECCV, pages 289-310. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.335, + 0.894, + 0.386 + ], + "angle": 0, + "content": "[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.388, + 0.893, + 0.442 + ], + "angle": 0, + "content": "[18] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE transactions on multimedia, 25:8906-8919, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.444, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[19] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, pages 237-255. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.498, + 0.893, + 0.551 + ], + "angle": 0, + "content": "[20] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.552, + 0.894, + 0.605 + ], + "angle": 0, + "content": "[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.606, + 0.893, + 0.646 + ], + "angle": 0, + "content": "[22] Hui Lu, Albert Ali Salah, and Ronald Poppe. Videomambapro: A leap forward for mamba in video understanding. arXiv preprint arXiv:2406.19006, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.647, + 0.894, + 0.699 + ], + "angle": 0, + "content": "[23] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.701, + 0.893, + 0.74 + ], + "angle": 0, + "content": "[24] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.741, + 0.894, + 0.807 + ], + "angle": 0, + "content": "[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.809, + 0.894, + 0.862 + ], + "angle": 0, + "content": "[26] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381-398. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.863, + 0.894, + 0.89 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.515, + 0.939 + ], + "angle": 0, + "content": "5303" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.104, + 0.483, + 0.157 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.16, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[28] Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.202, + 0.483, + 0.24 + ], + "angle": 0, + "content": "[29] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.243, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[30] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.285, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[31] Ozan Unal, Christos Sakaridis, Suman Saha, and Luc Van Gool. Four ways to improve verbo-visual fusion for dense 3d visual grounding. In European Conference on Computer Vision, pages 196–213. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.341, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[32] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.409, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[33] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13917-13926, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.478, + 0.483, + 0.558 + ], + "angle": 0, + "content": "[34] Yuxuan Wang, Among Wu, Muli Yang, Yukuan Min, Yihang Zhu, and Cheng Deng. Reasoning mamba: Hypergraph-guided region relation calculating for weakly supervised affordance grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27618-27627, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.561, + 0.483, + 0.626 + ], + "angle": 0, + "content": "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19231-19242, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.629, + 0.483, + 0.681 + ], + "angle": 0, + "content": "[36] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.684, + 0.483, + 0.75 + ], + "angle": 0, + "content": "[37] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.753, + 0.483, + 0.792 + ], + "angle": 0, + "content": "[38] Can Xu, Yuehui Han, Rui Xu, Le Hui, Jin Xie, and Jian Yang. Multi-attribute interactions matter for 3d visual grounding. In CVPR, pages 17253–17262, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.795, + 0.483, + 0.847 + ], + "angle": 0, + "content": "[39] Chenghao Xu, Guangtao Lyu, Jiexi Yan, Muli Yang, and Cheng Deng. Llm knows body language, too: Translating speech voices into human gestures. In ACL, pages 5004-5013, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.85, + 0.483, + 0.89 + ], + "angle": 0, + "content": "[40] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.104, + 0.483, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.157 + ], + "angle": 0, + "content": "[41] Haomeng Zhang, Chiao-An Yang, and Raymond A Yeh. Multi-object 3d grounding with dynamic modules and language-informed spatial attention. arXiv preprint arXiv:2410.22306, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.16, + 0.894, + 0.199 + ], + "angle": 0, + "content": "[42] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.201, + 0.894, + 0.253 + ], + "angle": 0, + "content": "[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.256, + 0.894, + 0.323 + ], + "angle": 0, + "content": "[44] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188-206. Springer, 2024. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.929, + 0.515, + 0.939 + ], + "angle": 0, + "content": "5304" + } + ] +] \ No newline at end of file diff --git a/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_origin.pdf b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f3cfe69100f99edeaebc0ab7133adc6258300950 --- /dev/null +++ b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/db8ed93c-5bcb-43cd-a62c-cc51ec338aab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba5ff9a31828a5c117ceaa4452a14ac55408f467c1fff271e71e622f0a3cfa4e +size 664479 diff --git a/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/full.md b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..add4e2a0222ef54b4820aadf4f0b165d794ff96b --- /dev/null +++ b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/full.md @@ -0,0 +1,370 @@ +# VGMamba: Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding + +Yihang Zhu, Jinhao Zhang, Yuxuan Wang, Among Wu, Cheng Deng* Xidian University, China + +{zhuyihang, zhangjinhao, wangyuxuan}@stu.xidian.edu.cn amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn + +# Abstract + +As an important direction of embodied intelligence, 3D Visual Grounding has attracted much attention, aiming to identify 3D objects matching the given language description. Most existing methods often follow a two-stage process, i.e., first detecting proposal objects and identifying the right objects based on the relevance to the given query. However, when the query is complex, it is difficult to leverage an abstract language representation to lock the corresponding objects accurately, affecting the grounding performance. In general, given a specific object, humans usually follow two clues to finish the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, attribute-to-location clue reasoning, to conduct accurate grounding. Particularly, we propose a VGMamba network that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Taking a 3D point cloud scene and language query as the input, we first exploit SVD to make a decomposition of the extracted features. Then, a slidingwindow operation is conducted to capture attribute characteristics. Next, a location mamba is presented to obtain the corresponding location information. Finally, by means of multi-modal mamba fusion, the model could effectively localize the object that matches the given query. In the experiment, our method is verified on four datasets. Extensive experimental results demonstrate the superiority of our method. + +# 1. Introduction + +3D Visual Grounding aims to identify and localize objects in a 3D scene based on natural language descriptions. This task has become a key challenge at the intersection of computer vision and natural language processing, with significant applications in areas such as human-robot interaction and vision-language navigation [5, 7, 20]. In the context + +![](images/6711d085304180f206b02272e05fd46b312d1eff8323d3907a2af354f9e0adfd.jpg) +Figure 1. To achieve accurate 3D Visual Grounding, we explore two clues. First, we leverage the attribute clue to capture proposal objects. Then, a location clue is exploited to focus on the right grounding object. Experimental results on multiple datasets verify the effectiveness of this mechanism. + +![](images/daa327c8f74bd1dc16ca2f51ef2e19984d8c47ae7bd3d5e76eb416e0549f5de8.jpg) + +of 3D environments, this task becomes even more complex due to the sparsity and irregularity of 3D point cloud data, as well as the variability in object shapes, scales, and the complexity of the language descriptions. Therefore, accurately localizing objects in 3D scene based on natural language descriptions is crucial for robots to understand and interact with their surroundings. + +Recent studies [2, 6, 38, 40, 41] have attempted to address this challenge by incorporating both object attributes (e.g., color, shape, texture) and spatial relationships (e.g., relative positions) to improve grounding accuracy. These methods typically follow a two-stage paradigm: first detecting candidate objects in the scene, and then assessing relevance to the given query based on feature similarity. While these methods have demonstrated effectiveness in certain scenarios, they still exhibit some limitations. + +First, most existing works [6, 38, 40] treat object attributes and spatial relationships in isolation, neglecting the inherent dependencies between them. In human perception, object attributes often serve as priors for location reasoning. For example, when searching for "The red chair next to the + +table," humans first identify objects matching the attribute "red chair" and then refine their search based on location cues. Existing approaches lack a mechanism to systematically leverage this reasoning process, resulting in suboptimal performance in complex scenes. Furthermore, in complex 3D environments, a query may refer to multiple objects. However, existing Transformer-based methods often fail to effectively capture global dependencies among multiple objects with shared attributes and spatial relations due to their quadratic complexity in self-attention, which limits scalability, and their tendency to focus on local interactions rather than modeling long-range dependencies. + +To address these limitations, as shown in Fig. 1, we separately explore reasoning attribute-based information and location-based characteristic to ground corresponding 3D objects. To be specific, we propose VGMamba, a novel architecture that systematically models attribute-to-location dependencies while efficiently capturing long-range interactions. VGMamba consists of three key modules: SVD-based Attribute Mamba, which leverages Singular Value Decomposition (SVD) to decompose the extracted visual features into multiple subspaces corresponding to different attribute scales. However, SVD alone lacks long-range modeling capability, which is critical for capturing hierarchical dependencies among attributes. To overcome this, we employ Mamba [11] to sequentially integrate multi-scale attribute features, preserving their interdependencies over long sequences. Compared to Transformers, Mamba maintains linear computational complexity while effectively modeling long-range dependencies among attribute components; the Location Mamba, which models both global and local positional relationships among objects by establishing cross-channel communication through SSM, effectively enhancing object-related features; and the Instructive Dual-Mamba block, which fuses object-related and text features through cross-modal guidance of text query. By replacing computationally expensive attention mechanisms with structured state-space modeling, VGMamba significantly improves efficiency without sacrificing performance. Extensive experiments demonstrate that our method achieves state-of-the-art performance, particularly excelling in scenarios involving multiple objects and ambiguous descriptions. Our chief contributions are threefold: + +- We explore a novel mechanism, i.e., attribute-to-location clue reasoning, for performing 3D visual grounding. +- We propose a novel framework VGMamba, comprising three core modules: the SVD-based Attribute Mamba for disentangling attributes, the Location Mamba for modeling location relations, and the Instructive Dual-Mamba block for text and object-related feature fusion. +- We achieve new state-of-the-art performance on multiple 3D Visual Grounding benchmarks, demonstrating superior accuracy and computational efficiency. + +# 2. Related Work + +3D Visual Grounding. Recent advancements incorporate self-attention-based scene encoding [2], viewpoint-aware spatial reasoning [29], and multi-attribute Transformers for fine-grained language-visual alignment [38]. However, these attention-based approaches struggle with long-sequence feature modeling and computational efficiency. For multi-object grounding, M3DRef-CLIP [42] introduced the first multi-object dataset and an attention-based fusion module integrating 3D object and 2D rendered image features. Subsequent works [41] refine proposal selection, camera pose optimization, and fusion strategies. However, these methods often embed global 2D features, introducing redundancy. In contrast, we focus on fine-grained attribute extraction as priors to enhance spatial reasoning. + +State Space Models. State Space Models (SSMs) [12, 30, 34] have become a prominent approach in sequence modeling, particularly for capturing long-range dependencies with linear time complexity. Furthermore, Mamba [11] introduced a selective state-space mechanism, providing efficient sequence modeling and surpassing traditional Transformers in certain tasks. This advancement led to various extensions across vision applications, including Vision Mamba [43] and VMamba [21], which extends scanning in four directions to fully capture interrelations among image patches. Besides, Mamba-based models have also been applied in video understanding [19, 22, 25, 39] and medical image segmentation [24, 28, 36, 37], demonstrating the versatility and efficiency of SSMs in diverse visual domains. + +In this work, we introduce Mamba into the object grounding task for the first time, exploring its potential in fine-grained feature extraction. We also propose a novel cross-modal fusion framework for language and object-related features, which achieves outstanding performance and significantly reduces computational costs. + +# 3. Overview of State Space Models + +Recently, state space models (SSMs) [9, 12, 30, 32] have attracted much attention for their ability to model continuous systems, constructing the foundation for the popular Mamba architecture [11]. Particularly, SSMs generally take an input sequence $x(t) \in \mathbb{R}^L$ as the input and output the corresponding sequence $y(t) \in \mathbb{R}^L$ through hidden states $h(t) \in \mathbb{R}^N$ , where $N$ is the number of hidden states. The system is governed by differential equations that describe how the hidden state evolves over time: + +$$ +h ^ {\prime} (t) = \mathbf {A} h (t) + \mathbf {B} x (t), \quad y (t) = \mathbf {C} h (t), \tag {1} +$$ + +where $\mathbf{A}$ , $\mathbf{B}$ , and $\mathbf{C}$ are matrices that project the input into the hidden states and subsequently to the output. + +In order to apply SSMs in practice tasks, a discretization process is employed, typically using a time scale parameter + +![](images/e778ec5bc5e10d21d9e714ed850c870048656c05a56dbc6c33c6138fcceacb16.jpg) +Figure 2. VGMamba for quantity-agnostic 3D Visual Grounding. Based on the given language description, we could utilize two clues to localize corresponding objects accurately, i.e., attribute and location information. To this end, we first design an SVD-based attribute mamba to discover multiple attribute-related objects. Then, a location mamba is further designed to select location-relevant objects. Finally, we present an Instructive Dual-Mamba block to localize the object that matches the given query. + +![](images/5352586d5d5e8dd8741d74df41b65fe6d2702b70f49457916439407fe948b6cb.jpg) + +$\Delta$ to convert continuous parameters into discrete ones. The processes are shown as follows: + +$$ +\overline {{\mathbf {A}}} = \exp (\Delta \mathbf {A}), \overline {{\mathbf {B}}} = (\Delta \mathbf {A}) ^ {- 1} (\exp (\Delta \mathbf {A}) - \mathbf {I}) \cdot \Delta \mathbf {B}. (2) +$$ + +The discretized state-space output can be represented as: + +$$ +h _ {k} = \overline {{\mathbf {A}}} h _ {k - 1} + \overline {{\mathbf {B}}} x _ {k}, \quad y _ {k} = \mathbf {C} h _ {k}. \tag {3} +$$ + +Based on the SSMs framework, Mamba [11] employs the Selective Scan Mechanism (S6) to conduct dynamic sequential modeling and capture critical task-related information. Though Mamba-based networks have been verified to be effective in many tasks [19, 24], it is under-explored for 3D Visual Grounding. To this end, we study whether the Mamba could be used to capture the relation between objects and language descriptions. We directly plug Mamba into M3DRef-CLIP [42] and conduct experiments on Multi3DRefer dataset [42]. However, we observe that this operation could not improve the grounding performance. The reason may be that Mamba emphasizes capturing global object information and cannot construct the relation between objects and language descriptions. + +# 4. VGMamba for 3D Visual Grounding + +In this paper, we propose a new VGMamba architecture, aiming to reason attribute and location clues based on the given language descriptions and improve the performance of 3D Visual Grounding effectively. + +Specifically, in Fig. 2 (a), we follow the work [42] and adopt PointGroup [17] as a 3D detector to generate an object proposal list $o = (o_1,o_2,\dots,o_n)$ with $n$ proposal boxes and corresponding features $F_{\mathrm{3D}}\in \mathbb{R}^{n\times 32}$ . Each proposal is then rendered as $m$ multi-view images, which are encoded by CLIP [27] to extract $d$ -dimensional patch features $I\in \mathbb{R}^{n\times m\times p\times d}$ and aggregated features $\hat{I}\in \mathbb{R}^{n\times m\times d}$ , where $p$ is the patch number of each multiview image. Next, for the given language description, we continue following the work [42] and utilize CLIP to obtain text feature $T\in \mathbb{R}^d$ . + +In practice, 3D Visual Grounding can be achieved by reasoning two key clues. One clue is to accurately recognize object-related attribute, e.g., dark blue. Another clue is to effectively obtain the location-relevant visual content, e.g., facing the backboard. With this in mind, we separately design an attribute mamba and location mamba to sufficiently capture object-related attribute and location characteristic. + +# 4.1. SVD-based Attribute Mamba + +In general, given a specific object, the extracted representations involve multiple different attribute characteristics, e.g., color, shape, and texture. However, the 3D Visual Grounding task involves localizing objects that possess specific attributes, such as dark blue. Therefore, it is meaningful to disentangle current representations. To this end, we explore using Singular Value Decomposition (SVD) to perform feature-level decomposition and recombination ac + +cording to the calculated singular values. + +Particularly, as illustrated in Fig. 2 (b), for each patch feature $I_{i}\in \mathbb{R}^{d\times p}$ , SVD is first used to decompose $I_{i}$ into three matrices: $I_{i} = U\Sigma V^{T}\in \mathbb{R}^{d\times p}$ , where $U\in \mathbb{R}^{d\times d}$ and $V\in \mathbb{R}^{p\times p}$ are orthogonal matrices, and $\Sigma$ contains singular values sorted along its main diagonal [10]. Subsequently, inspired by Multi-Scale Attention [18], we introduce Sliding Window to recombine the singular vectors using three sliding windows with varying size $r$ : + +$$ +\left\{F _ {r} \right\} = \text {S l i d i n g W i n d o w} (\Sigma , r), \tag {4} +$$ + +$$ +\text {S l i d i n g W i n d o w} (\Sigma , r) = U _ {d \times r} \Sigma_ {r \times r} V _ {r \times p} ^ {T}, \tag {5} +$$ + +where $r \in \{r_1, r_2, r_3\}$ . SlidingWindow refers to combining r consecutive singular vectors into a new multi-scale feature $F_r$ . Since the singular vectors corresponding large singular values and those corresponding small singular values separately contain generalized coarse-grained information and discriminative fine-grained information, through multiscale operation, we can sufficiently capture attribute information involved in visual representations. Following this, $F_r$ is processed by three independent Mamba to model attribute feature. Finally, a concatenation followed by a $1 \times 1$ convolution is applied to merge the multi-scale features, yielding the final feature representations $F_{i-att} \in \mathbb{R}^d$ of the single-view image. For $n$ objects with $m$ rendered images, we can obtain the feature representation $F_{att} \in \mathbb{R}^{n \times m \times d}$ . + +# 4.2. Location Mamba + +As shown in Fig. 2 (a), only relying on the attribute clue could not guarantee accurately identifying the objects corresponding to the given language description. To this end, based on the captured attribute features, we design a cross-channel communication mechanism to facilitate the learning of multi-object location relationships, enforcing the locking of the object matching the given description. + +Specifically, to capture spatial relationships from a consistent viewpoint, for each rendered object image, we define an encoding $pos \in \mathbb{R}^{n \times m \times 6}$ that includes object center coordinates $(x,y,z)$ and bounding box dimensions $(h,w,l)$ . Following this, we utilize a linear layer to perform a transformation on $pos$ : + +$$ +p \hat {o} s = \operatorname {L i n e a r} (p o s) \in \mathbb {R} ^ {n \times m \times d}. \tag {6} +$$ + +Subsequently, the summation result $I_{\mathrm{spa}} \in \mathbb{R}^{n \times m \times d}$ of $p \hat{o} s$ , $\hat{I} \in \mathbb{R}^{n \times m \times d}$ , and $F_{\mathrm{att}} \in \mathbb{R}^{n \times m \times d}$ is taken to perform the following operations: exploring a dedicated mamba to respectively conduct communication from global and local perspectives. For global communication, the channels of $n$ proposal objects are treated as a sequence to enable cross-object interaction: + +$$ +I _ {\text {g l o b a l}} = \operatorname {M a m b a} _ {\text {g l o b a l}} \left(I _ {\text {s p a}}\right), \tag {7} +$$ + +![](images/34e446207457583f5f4981041dfca6d2d5c3db36ed02820ded22c0b98f669096.jpg) +Figure 3. The architecture of Instructive Dual-Mamba (IDM) for conducting multi-modal fusion. + +where $I_{\mathrm{global}} \in \mathbb{R}^{n \times m \times d}$ denotes the SSM output. + +As for the local communication, the $u$ nearest neighbors of each object are selected based on Euclidean distance, obtaining multi-neighbor feature $\hat{I}_{\mathrm{spa}} \in \mathbb{R}^{n \times m \times u \times d}$ . Cross-channel scanning is then applied to $\hat{I}_{\mathrm{spa}}$ and outputs central object features: + +$$ +I _ {\text {l o c a l}} = \operatorname {S e l e c t C e n t e r} \left(\mathrm {M a m b a} _ {\text {l o c a l}} \left(\hat {I} _ {\mathrm {s p a}}\right)\right), \tag {8} +$$ + +where $I_{\mathrm{local}} \in \mathbb{R}^{n \times m \times d}$ denotes the current SSM output. Subsequently, by summing $I_{\mathrm{global}}$ and $I_{\mathrm{local}}$ , and then averaging over the $m$ images, we obtain the feature $F_{\mathrm{spa}} \in \mathbb{R}^{n \times d}$ that contains plentiful location information. Finally, $F_{\mathrm{spa}}$ is concatenated with the proposal feature $F_{3\mathrm{D}} \in \mathbb{R}^{n \times 32}$ output by the 3D detector, followed by a $1 \times 1$ convolution layer for dimensionality reduction, yielding $F_{\mathrm{obj}} \in \mathbb{R}^{n \times d}$ . We represent the feature of each object in $F_{\mathrm{obj}}$ as $G \in \mathbb{R}^d$ . + +# 4.3. Multi-modal Mamba Fusion + +To further enhance the grounding accuracy, we present a multi-modal mamba fusion module to align the proposals to the corresponding language descriptions. Specifically, as shown in Fig. 3, the Instructive Dual-Mamba (IDM) block is proposed to use descriptions as guidance to capture matched object-related feature. Taking the language feature $T \in \mathbb{R}^d$ , object-related feature $G \in \mathbb{R}^d$ as the input, we first divide $T$ into a forward sequence and a reversed backward sequence, which are then processed by two independent SSMs (Eq. (1)) respectively. The outputs from the two branches are integrated to obtain the updated feature. The processing of input feature $G \in \mathbb{R}^d$ is similar to $T \in \mathbb{R}^d$ . According to Eq. (1), the state transition matrices $A$ retain critical historical information for updating hidden states, while $B$ and $C$ primarily influence input and output. Therefore, we propose replacing the state transition matrices $\vec{\mathbf{A}}_G, \vec{\mathbf{A}}_G$ with their counterparts $\vec{\mathbf{A}}_T, \vec{\mathbf{A}}_T$ derived from language features, enabling cross-modal guidance in extracting object-related features. The remaining parameter matrices remain unchanged. The process of above could be represented as: + +$$ +\overleftarrow {h} _ {\mathrm {T}} ^ {t} = \overleftarrow {\mathbf {A}} _ {\mathrm {T}} \overleftarrow {h} _ {\mathrm {T}} ^ {t - 1} + \overleftarrow {B} _ {\mathrm {T}} \overleftarrow {x} _ {\mathrm {T}} ^ {t}, \quad \overrightarrow {h} _ {\mathrm {T}} ^ {t} = \overrightarrow {\mathbf {A}} _ {\mathrm {T}} \overrightarrow {h} _ {\mathrm {T}} ^ {t - 1} + \overrightarrow {B} _ {\mathrm {T}} \overrightarrow {x} _ {\mathrm {T}} ^ {t}, \tag {9} +$$ + +$$ +\overleftarrow {h} _ {\mathsf {G}} ^ {t} = \overleftarrow {\mathbf {A}} _ {\mathsf {T}} \overleftarrow {h} _ {\mathsf {G}} ^ {t - 1} + \overleftarrow {B} _ {\mathsf {G}} \overleftarrow {x} _ {\mathsf {G}} ^ {t}, \quad \overrightarrow {h} _ {\mathsf {G}} ^ {t} = \overrightarrow {\mathbf {A}} _ {\mathsf {T}} \overrightarrow {h} _ {\mathsf {G}} ^ {t - 1} + \overrightarrow {B} _ {\mathsf {G}} \overrightarrow {x} _ {\mathsf {G}} ^ {t}, \tag {10} +$$ + +
MethodVenueUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [3]ECCV2067.646.232.121.340.026.1
TGNN [14]AAAI2168.656.829.823.237.429.7
MVT [15]CVPR2277.766.531.925.340.833.3
3D-SPS [23]CVPR2284.166.740.329.848.837.0
EDA [35]CVPR2385.868.649.137.654.642.3
VPP-Net [29]CVPR2486.167.150.339.055.743.3
M3DRef-CLIP [42]ICCV23-77.2-36.8-44.7
G3-LQ [33]CVPR2488.673.350.239.756.044.7
MCLN [26]ECCV2486.972.752.040.857.245.5
MA2TransVG [38]CVPR2486.374.153.841.457.945.7
ConcreteNet [31]ECCV2486.482.142.438.450.646.5
D-LISA [41]NeurIPS24-75.5-40.0-46.9
GPS [16]ECCV24-77.9-42.7-48.1
Chat-Scene [13]NeurIPS2489.682.547.842.955.550.2
PQ3D [44]ECCV2486.778.351.546.257.051.2
VGMamba-91.979.654.847.660.053.9
+ +Here, $x_{\mathrm{T / G}}^t$ represents the input at time step $t$ . $\overleftrightarrow{\mathbf{A}}_{\mathrm{T}}$ and $\overrightarrow{\mathbf{A}}_{\mathrm{T}}$ are the state transition matrices used for retaining essential historical information and updating hidden states, while matrices $B_{\mathrm{T / G}}$ primarily affect input. + +Additionally, we share the gating mechanism parameters $z$ of the language features to construct cross-modal fusion in the hidden state space, ultimately outputting the average multi-view attribute features of each object in the scene, represented as $F_{\mathrm{att - spa}} \in \mathbb{R}^{n \times d}$ . Finally, a convolution operation is applied to output confidence scores $O \in \mathbb{R}^{n \times 1}$ for each object proposal. + +# 4.4. Training Objectives + +Building on previous work [42], the loss of VGMamba consists of the 3D Visual Grounding loss $L_{\mathrm{ref}}$ , text-object contrastive loss $L_{\mathrm{con}}$ , and object detection loss $L_{\mathrm{det}}$ , which is formulated as: + +$$ +L = L _ {\text {r e f}} + L _ {\text {c o n}} + L _ {\text {d e t}}. \tag {11} +$$ + +# 5. Experiments + +To evaluate the effectiveness of our proposed VGMamba on 3D Visual Grounding tasks with varying numbers of target objects, we conducted experiments on both single-object datasets (ScanRefer [3], Nr3D and Sr3D [1]) and multi-object dataset Multi3DRefer [42]. + +Table 1. Comparison results on the ScanRefer [3] dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. Our proposed VGMamba performs favorably over current state-of-the-art approach PQ3D [44]. + +
MethodNr3DSr3D
OverallHardOverallHard
ReferIt3D [?]35.627.940.831.5
TGNN [14]37.330.645.036.9
M3DRef-CLIP [42]49.443.4--
3D-SPS [23]51.545.162.665.4
EDA [35]52.146.168.162.9
D-LISA [41]53.146.2--
MVT [15]55.449.164.558.8
VPP-Net [29]56.9-68.7-
G3-LQ [33]58.450.773.166.3
MCLN [26]59.8-68.4-
MiKASA [2]64.459.475.267.3
GPS [16]64.957.877.571.6
MA2TransVG [38]65.257.673.969.3
PQ3D [44]66.758.779.772.8
VGMamba68.361.481.374.4
+ +Table 2. Quantitative comparisons on Nr3D and Sr3D dataset. We have highlighted the top-performing three methods in purple. + +# 5.1. Single-object Grounding + +# 5.1.1. Datasets and Evaluation Metrics + +ScanRefer. The ScanRefer dataset comprises 51,583 natural language descriptions for 11,046 objects across 800 3D scenes from the ScanNet dataset [8]. Each description specifies a target object and its location relations within + +Query Text + +The green chair next to the brown table with a white vase. + +The brown chairs under the white rectangular table. + +The white pillow closest to the wall with the painting. + +M3DRef-CLIP + +![](images/7af92848cd849a159b2c3f9792c387e4c14eda7647d89b5b8a7d49aa75e057e9.jpg) + +![](images/d0c4f4861ee5c83719a71282974d1e73d2ea4f44c629480fc078445010055178.jpg) + +![](images/081c860578023974147507a4f27e639d032bab4b7f33dc71986f57a09378c8d3.jpg) +Figure 4. Qualitative examples of the Multi3DRefer dataset. The results illustrate the grounding performance of M3DRef-CLIP (blue box), D-LISA (blue box), and our VGMamba(green box), compared to the ground truth labels (red box). + +D-LISA + +![](images/f6b018c6ba987e46ce7952b65b2b64175c972dfd434bfee05fc958d84691d748.jpg) + +![](images/93338e41061c3d2a1846cc49f16d4f186e38aa78896d9a1015dfa88b33de1b29.jpg) + +![](images/b1fce43dc233542dee267fcf47ea7ec9e877facdb498f63fdc6d52176622d9a0.jpg) + +![](images/51b9f339182965f4d04895b49b88ca83d09f9849cd7d173467a0430730045022.jpg) +VGMamba (ours) + +![](images/3ae35d0d9bed1b0d0aad93858f834b02cc8e5759476240eb8f6a292ae1388847.jpg) + +![](images/4bc2e55ec9184409c94c96d181da8d640898cc3f35dc2575f2a19040bd44527d.jpg) + +![](images/f84da7be7687c7062e9369bb8b9dd4e709de942c48c593ff06ddfcb508130193.jpg) +Ground Truth + +![](images/38ac7537dc9007e61f99c1e6ece4f4e7636884e2296c6e9d8949240497093dc0.jpg) + +![](images/7afff4911b1c37a90619362d3399f3e7a841128bdd0f7e723b38cb34838d2fbe.jpg) + +
MethodZT w/o DZT w/DST w/o DST w/DMTAll
D3Net [4]81.632.538.623.335.032.2
M3DRef-CLIP [42]81.839.447.830.637.938.4
M3DRef-CLIP w/NMS [41]79.040.567.640.049.149.3
PQ3D [44]85.457.768.543.640.950.1
D-LISA [41]82.443.767.142.551.051.2
VGMamba83.645.270.946.755.154.9
+ +Table 3. Comparison of F1@0.5 results on the Multi3DRefer [42] validation set, where our method outperforms all baselines. + +the scene, enabling the model to understand both object attributes and location context. The dataset is divided into two subsets: Unique, where the target object is the only instance of its class within the scene, and Multiple, where multiple instances of the same class are present. We evaluate our model on validation set using Acc@0.25 and Acc@0.5, which measure the proportion of predicted bounding boxes that achieve an Intersection over Union (IoU) larger than 0.25 and 0.5 with the ground truth box. + +$\text{Nr3D/Sr3D}$ . The $\text{Nr3D}$ and $\text{Sr3D}$ datasets, part of the ReferIt3D benchmark [1], are built upon the ScanNet dataset [8] for 3D visual grounding. $\text{Nr3D}$ (Natural Reference in 3D) consists of 41,503 human-annotated referring expressions across 707 indoor scenes, where each description uniquely identifies a target object among distractors. $\text{Sr3D}$ (Spatial Reference in 3D) contains 83,572 machine-generated expressions based on predefined templates, focusing on spatial relationships between objects. Both datasets provide ground-truth object masks. Following [1], performance is evaluated using accuracy, measuring the proportion of correctly localized target objects. + +# 5.1.2. Implementation Details + +We implement the proposed VGMamba model using PyTorch and train it end-to-end on a single NVIDIA A6000 GPU. The model is optimized using the AdamW optimizer with a batch size of 4 and a learning rate of 5e-4 with cosine decay scheduling. Following prior work [42], we employ a pre-trained PointGroup [17] module as the detector, which is fine-tuned on the ScanNet dataset. The proposal box with the highest matching score is selected as the final prediction. We adopt the official train/val/test splits provided by the ScanRefer and $\mathrm{Nr3D / Sr3D}$ datasets to ensure fair comparisons with baseline methods. + +# 5.1.3. Baseline Comparison + +ScanRefer. The experimental results on ScanRefer are shown in Tab. 1, with the following key observations: (i) Our method achieves state-of-the-art performance with an overall accuracy of $60.0\%$ at IoU 0.25 and $53.9\%$ at IoU 0.5, surpassing the best baseline PQ3D [44] by $3.0\%$ and $2.7\%$ , respectively. (ii) In the Multiple case, VGMamba attains $54.8\%$ at IoU 0.25 and $47.6\%$ at IoU 0.5, outperforming PQ3D and demonstrating that its attribute-to-location reasoning effectively disambiguates objects under complex scenarios. (iii) Unlike previous methods [13, 44] that show notable performance variations across settings, our VGMamba maintains consistently high accuracy, validating its robust generalization capability in both unique and multiple-object scenarios. + +$Nr3D/Sr3D$ . Tab. 2 presents the grounding accuracy of different methods on the $\mathrm{Nr3D}$ and Sr3D datasets. VGMamba achieves state-of-the-art performance of $68.3\%$ and $81.3\%$ on both datasets, surpassing prior state-of-the-art methods. The $\mathrm{Nr3D}$ dataset presents additional challenges due to + +
ModulesMulti3DRefer (F1@0.5 on Val)
RowIDMLMSAMZT w/o DZT w/DST w/o DST w/DMTAll
1XXX81.839.447.830.637.938.4
2✓XX83.943.756.839.743.542.3
3X✓X80.240.761.142.845.143.8
4XX✓81.544.067.643.947.845.2
5✓✓X83.546.165.542.348.648.3
6✓X✓82.948.268.745.849.049.2
7X✓✓85.445.369.244.253.152.5
8✓✓✓83.645.270.946.755.154.9
+ +Table 4. Ablation study of proposed modules on Multi3DRefer. + +its intricate and free-form textual descriptions, which increase the difficulty of cross-modal alignment. Despite this, our method significantly outperforms alternatives that leverage large-scale 3D vision-language pre-training models [13, 44]. Notably, in the more challenging "Hard" subset, VGMamba improves accuracy by $2.7\%$ in $\mathrm{Nr3D}$ and $1.6\%$ in Sr3D, demonstrating its superior capability in distinguishing ambiguous objects by effectively capturing fine-grained attribute and spatial relationships. + +# 5.2. Multi-object Grounding + +# 5.2.1. Datasets and Evaluation Metrics + +Multi3DRefer is a dataset derived from ScanRefer [3], comprising a total of 61,926 textual descriptions corresponding to 11,609 distinct objects. Each description may reference zero, one, or multiple target objects, enabling the evaluation of grounding models in a variety of scenarios. Following previous work[42], we compute the F1 score at the Intersection over union (IoU) threshold of 0.5(denoted as F1@0.5) across five distinct scenarios: a) zero target without distractors of the same semantic class (ZT w/o D); b) zero target with distractors (ZT w/D); c) single target without distractors (ST w/o D); d) single target with distractors (ST w/D); and e) multiple targets (MT). Finally, the average across these five evaluation cases is calculated to provide an overall performance score. + +# 5.2.2. Implementation Details + +The experimental setting for the multi-object dataset is similar to that for the single-object dataset, with a few differences: during training, we treat all predicted bounding boxes with an IoU greater than $\tau_{\mathrm{train}} = 0.25$ relative to the ground truth as target bounding boxes. During inference, we consider all predicted bounding boxes with a matching score greater than $\tau_{\mathrm{pred}} = 0.1$ as the model's predictions. + +# 5.2.3. Baseline Comparison + +We compare our 3D multi-object grounding performance with existing works quantitatively and qualitatively. As illustrated in Tab. 3, our proposed VGMamba achieves competitive performance across all scenarios. In cases with distractors (ST w/D), VGMamba achieves F1 scores of + +![](images/3a7a8fbc8d47d4bfd1b88ff83e3a7d384d6d0a820b32355db816599b6858a07c.jpg) +Figure 5. Efficiency-Performance Trade-off on ScanRefer. + +$46.7\%$ , surpassing the second-best competitor by $3.1\%$ , which highlights its robustness in managing complex scenes with multiple potential matches. In the MT case, VG-Mamba attains an F1 score that is $4.1\%$ higher than the second-best competitor, underscoring the advantage of the Instructive Dual-Mamba (IDM) in addressing the multi-object grounding task. + +We show the qualitative comparison with M3DRef-CLIP[42] and D-LISA[41] in Fig. 4. The top two rows illustrate examples involving multiple target objects in complex scenes. Our VGMamba accurately locates the objects matching the given textual descriptions, even in the presence of distractors. The last row demonstrates a single-target case with surrounding distractors. VGMamba effectively isolates and identifies the correct object, while baseline models struggle, often mistakenly including distractors in their predictions. + +# 5.3. Ablation Studies + +To validate the effectiveness of each proposed module within our VGMamba framework, we conduct ablation studies on the Multi3DRefer dataset, as shown in Tab. 4. More ablation results are detailed in supplementary. + +Effectiveness of SVD-based Attribute Mamba (SAM). Comparing Row1 with Row4, we observe that adding SAM yields significant improvements across various scenarios, especially in the ST w/o D case, where the F1 score improves by $19.8\%$ . This demonstrates SAM's ability to decompose complex representations and capture fine-grained attribute features, which is essential for distinguishing target objects based on descriptive characteristics. + +Effectiveness of Location Mamba (LM). The impact of LM is evident when comparing Row 4 with Row 7. Adding LM enhances performance, particularly in the $MT$ scenario, where the F1 score increases by $5.3\%$ . This improvement demonstrates the importance of LM in capturing location relationships between objects, allowing the model to better handle cases with distractors. + +Effectiveness of Instructive Dual-Mamba (IDM). The contribution of IDM is highlighted by comparing Row 7 with Row 8. Incorporating IDM results in consistent per + +![](images/e4c88b316b8665d4771a5c03d8d1bfa6093b80f44e3c464c6768755032164cf4.jpg) +Query Text + +![](images/2875bf1191b9d68d0310ad3268f71a9768da104fa981fde662fc69f58cb06107.jpg) +The red sofa is against the wall with a white cushion. + +![](images/f6ea0d90da02796e5254e24cfd42d3d3570721f420dfe02824829148dc66a6c4.jpg) +A wooden chair without arms is tucked under the table. +A black chair without armrests, back to the window. +Figure 6. Visualization of the attribute, location relation, and overall scores in our model. + +![](images/fc5d3d221026f955a46a573647194a13f2008f34fac8ab91a578e66372bc7248.jpg) +Ground Truth +Attribute Score + +![](images/f36053b5c8c927fd46cf62b1191df4b46c4304bc9b3c11ac65dc9a363b227926.jpg) + +![](images/3fdb66e3d5d2449bc1559f9d98a89fd04fe122e2cd9d3631456e74a431522a2d.jpg) + +![](images/0debec5e2c53069c271c33197ca906284d269a6f17478993254f2de2de0c7324.jpg) +Location Score + +![](images/f8dbd50435c96bd44c64b9258a01632685b5f654ee4a74d0132038988e64a0bf.jpg) + +![](images/c5bf46b0ea7f5098b83cfa8cbb7447b56f7029f3eb2010d1f6499598083d5bc2.jpg) + +![](images/268055ec478ada3ce8efaf501374ee53beda0ffce44a2b1190f9653684901175.jpg) +Overall Score + +![](images/1649b3a7550b2b3fb94000171c9d50bab1b61d6166df9e6ee6813a661721995b.jpg) + +![](images/faaec61c9cab9784a95c8c10509a835bc1f04fbb77c71dd48172174430dcba2a.jpg) + +
MethodZT w/o DZT w/DST w/o DST w/DMTAll
Location First82.143.968.746.354.753.1
Attribute First83.645.270.946.755.154.9
+ +Table 5. Comparison of grounding performance between attribute first and location first model. + +formance gains across all metrics, with a notable increase in the $MT$ case, where the F1 score improves by $2\%$ . This shows IDM's effectiveness in handling multi-object grounding, enabling precise language-instructed grounding in complex scenes with multiple target objects. + +Efficiency-Performance Trade-off. Fig. 5 illustrates the balance between computational complexity (FLOPs), inference speed, and grounding performance on the ScanRefer dataset. VGMamba achieves the highest accuracy with significantly lower FLOPs than methods like Chat-Scene, demonstrating superior efficiency. This indicates that our method achieves a superior balance between efficiency and performance, making it well-suited for real-world applications where both accuracy and inference speed are critical. The lightweight nature of VGMamba stems from its efficient attribute and location feature sequence modeling with Mamba, which enables better feature aggregation without excessive computational overhead. + +Why Attribute First? We conducted experiments on the Multi3DRefer val set to compare the performance between the location first and our proposed attribute first model. As shown in Tab. 5, the attribute first approach achieves relatively higher performance. One possible explanation is that attribute cues serve as more distinguishing priors compared to location cues, enabling the model to initially identify a larger and more accurate set of candidate objects. In con + +trast, location information is often less specific and could potentially mislead the model toward irrelevant objects. + +# 5.4. Interpretability Analysis + +Benefiting from the sequential attribute-to-location reasoning process, VGMamba demonstrates a high degree of interpretability. To visualize the model's understanding of attribute and location relationships, we utilize the output features from the SVD-based Attribute Mamba, Location Mamba, and Instructive Dual-Mamba modules to obtain the attribute score, location score, and overall score. Fig. 6 presents qualitative results, where darker colors indicate higher scores. Specifically, our model first identifies objects that match attribute cues, such as "red sofa," "wooden chair," and "black chair." Then, using these attributes as prior, it incorporates location cues to filter out mismatched objects, e.g., "against the wall," "under the table," and "back to the window" (highlighted in yellow boxes). The final prediction scores, generated by the Instructive Dual-Mamba block, are highlighted in red boxes. + +# 6. Conclusion + +In this paper, we are concerned with 3D Visual Grounding that aims to identify 3D objects matching the language description. Particularly, we consider that human beings often follow two clues to achieve the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, i.e., attribute-to-location clue reasoning, and propose a new VGMamba method that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Experimental results on multiple datasets demonstrate the effectiveness of our method. + +# Acknowledgements + +Our work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), the Joint Fund of the Ministry of Education of China (Nos. 8091B022149, 8091B02072404), the National Natural Science Foundation of China (No. 62132016), and the Natural Science Basic Research Program of Shaanxi (No. 2020JC-23). + +# References + +[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 5, 6 +[2] Chun-Peng Chang, Shaoxiang Wang, Alain Pagani, and Didier Stricker. Mikasa: Multi-key-anchor & scene-aware transformer for 3d visual grounding. In CVPR, pages 14131-14140, 2024. 1, 2, 5 +[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221, 2020. 5, 7 +[4] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 6 +[5] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems, 34:5834-5847, 2021. 1 +[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 1 +[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1 +[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6 +[9] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2 +[10] Gene H. Golub and William Kahan. Calculating the singular values and pseudo-inverse of a matrix. In Milestones in Matrix Computation, 2007. 4 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3 +[12] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2 + +[13] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 5, 6, 7 +[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 5 +[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 5 +[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In ECCV, pages 289-310. Springer, 2024. 5 +[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 3, 6 +[18] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE transactions on multimedia, 25:8906-8919, 2023. 4 +[19] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, pages 237-255. Springer, 2024. 2, 3 +[20] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1 +[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2 +[22] Hui Lu, Albert Ali Salah, and Ronald Poppe. Videomambapro: A leap forward for mamba in video understanding. arXiv preprint arXiv:2406.19006, 2024. 2 +[23] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 5 +[24] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 2, 3 +[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2 +[26] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381-398. Springer, 2024. 5 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3 +[28] Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2 +[29] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 2, 5 +[30] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2 +[31] Ozan Unal, Christos Sakaridis, Suman Saha, and Luc Van Gool. Four ways to improve verbo-visual fusion for dense 3d visual grounding. In European Conference on Computer Vision, pages 196–213. Springer, 2024. 5 +[32] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 2 +[33] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13917-13926, 2024. 5 +[34] Yuxuan Wang, Among Wu, Muli Yang, Yukuan Min, Yihang Zhu, and Cheng Deng. Reasoning mamba: Hypergraph-guided region relation calculating for weakly supervised affordance grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27618-27627, 2025. 2 +[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19231-19242, 2023. 5 +[36] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2 +[37] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2 +[38] Can Xu, Yuehui Han, Rui Xu, Le Hui, Jin Xie, and Jian Yang. Multi-attribute interactions matter for 3d visual grounding. In CVPR, pages 17253–17262, 2024. 1, 2, 5 +[39] Chenghao Xu, Guangtao Lyu, Jiexi Yan, Muli Yang, and Cheng Deng. Llm knows body language, too: Translating speech voices into human gestures. In ACL, pages 5004-5013, 2024. 2 +[40] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1 + +[41] Haomeng Zhang, Chiao-An Yang, and Raymond A Yeh. Multi-object 3d grounding with dynamic modules and language-informed spatial attention. arXiv preprint arXiv:2410.22306, 2024. 1, 2, 5, 6, 7 +[42] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 2, 3, 5, 6, 7 +[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model, 2024. 2 +[44] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188-206. Springer, 2024. 5, 6, 7 \ No newline at end of file diff --git a/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/images.zip b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..95ae2c64c472f6c60d174fb9eea19abfd9713023 --- /dev/null +++ b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a85204bc00a133ad3ac59a947df257dd405ac0e690a699e59406ea780dce2a8 +size 552011 diff --git a/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/layout.json b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..910ddc4af8d27912ec54ac1f0373179929c2938c --- /dev/null +++ b/2025/VGMamba_ Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding/layout.json @@ -0,0 +1,9883 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 110, + 470, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 110, + 470, + 147 + ], + "spans": [ + { + "bbox": [ + 143, + 110, + 470, + 147 + ], + "type": "text", + "content": "VGMamba: Attribute-to-Location Clue Reasoning for Quantity-Agnostic 3D Visual Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 167, + 470, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 167, + 470, + 195 + ], + "spans": [ + { + "bbox": [ + 145, + 167, + 470, + 195 + ], + "type": "text", + "content": "Yihang Zhu, Jinhao Zhang, Yuxuan Wang, Among Wu, Cheng Deng* Xidian University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 197, + 447, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 197, + 447, + 222 + ], + "spans": [ + { + "bbox": [ + 162, + 197, + 447, + 222 + ], + "type": "text", + "content": "{zhuyihang, zhangjinhao, wangyuxuan}@stu.xidian.edu.cn amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 249, + 203, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 249, + 203, + 261 + ], + "spans": [ + { + "bbox": [ + 156, + 249, + 203, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 273, + 297, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 273, + 297, + 577 + ], + "spans": [ + { + "bbox": [ + 61, + 273, + 297, + 577 + ], + "type": "text", + "content": "As an important direction of embodied intelligence, 3D Visual Grounding has attracted much attention, aiming to identify 3D objects matching the given language description. Most existing methods often follow a two-stage process, i.e., first detecting proposal objects and identifying the right objects based on the relevance to the given query. However, when the query is complex, it is difficult to leverage an abstract language representation to lock the corresponding objects accurately, affecting the grounding performance. In general, given a specific object, humans usually follow two clues to finish the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, attribute-to-location clue reasoning, to conduct accurate grounding. Particularly, we propose a VGMamba network that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Taking a 3D point cloud scene and language query as the input, we first exploit SVD to make a decomposition of the extracted features. Then, a slidingwindow operation is conducted to capture attribute characteristics. Next, a location mamba is presented to obtain the corresponding location information. Finally, by means of multi-modal mamba fusion, the model could effectively localize the object that matches the given query. In the experiment, our method is verified on four datasets. Extensive experimental results demonstrate the superiority of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 597, + 142, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 597, + 142, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 597, + 142, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 616, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 616, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 61, + 616, + 296, + 688 + ], + "type": "text", + "content": "3D Visual Grounding aims to identify and localize objects in a 3D scene based on natural language descriptions. This task has become a key challenge at the intersection of computer vision and natural language processing, with significant applications in areas such as human-robot interaction and vision-language navigation [5, 7, 20]. In the context" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 249, + 417, + 381 + ], + "blocks": [ + { + "bbox": [ + 322, + 249, + 417, + 381 + ], + "lines": [ + { + "bbox": [ + 322, + 249, + 417, + 381 + ], + "spans": [ + { + "bbox": [ + 322, + 249, + 417, + 381 + ], + "type": "image", + "image_path": "6711d085304180f206b02272e05fd46b312d1eff8323d3907a2af354f9e0adfd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 548, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 548, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 548, + 445 + ], + "type": "text", + "content": "Figure 1. To achieve accurate 3D Visual Grounding, we explore two clues. First, we leverage the attribute clue to capture proposal objects. Then, a location clue is exploited to focus on the right grounding object. Experimental results on multiple datasets verify the effectiveness of this mechanism." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 418, + 249, + 539, + 381 + ], + "blocks": [ + { + "bbox": [ + 418, + 249, + 539, + 381 + ], + "lines": [ + { + "bbox": [ + 418, + 249, + 539, + 381 + ], + "spans": [ + { + "bbox": [ + 418, + 249, + 539, + 381 + ], + "type": "image", + "image_path": "daa327c8f74bd1dc16ca2f51ef2e19984d8c47ae7bd3d5e76eb416e0549f5de8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 456, + 548, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 548, + 537 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 548, + 537 + ], + "type": "text", + "content": "of 3D environments, this task becomes even more complex due to the sparsity and irregularity of 3D point cloud data, as well as the variability in object shapes, scales, and the complexity of the language descriptions. Therefore, accurately localizing objects in 3D scene based on natural language descriptions is crucial for robots to understand and interact with their surroundings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 539, + 548, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 539, + 548, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 539, + 548, + 644 + ], + "type": "text", + "content": "Recent studies [2, 6, 38, 40, 41] have attempted to address this challenge by incorporating both object attributes (e.g., color, shape, texture) and spatial relationships (e.g., relative positions) to improve grounding accuracy. These methods typically follow a two-stage paradigm: first detecting candidate objects in the scene, and then assessing relevance to the given query based on feature similarity. While these methods have demonstrated effectiveness in certain scenarios, they still exhibit some limitations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 646, + 548, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 548, + 706 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 548, + 706 + ], + "type": "text", + "content": "First, most existing works [6, 38, 40] treat object attributes and spatial relationships in isolation, neglecting the inherent dependencies between them. In human perception, object attributes often serve as priors for location reasoning. For example, when searching for \"The red chair next to the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 694, + 149, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 694, + 149, + 704 + ], + "spans": [ + { + "bbox": [ + 72, + 694, + 149, + 704 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 734, + 315, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 744 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 744 + ], + "type": "text", + "content": "5295" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 294, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 294, + 220 + ], + "type": "text", + "content": "table,\" humans first identify objects matching the attribute \"red chair\" and then refine their search based on location cues. Existing approaches lack a mechanism to systematically leverage this reasoning process, resulting in suboptimal performance in complex scenes. Furthermore, in complex 3D environments, a query may refer to multiple objects. However, existing Transformer-based methods often fail to effectively capture global dependencies among multiple objects with shared attributes and spatial relations due to their quadratic complexity in self-attention, which limits scalability, and their tendency to focus on local interactions rather than modeling long-range dependencies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 220, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 295, + 582 + ], + "type": "text", + "content": "To address these limitations, as shown in Fig. 1, we separately explore reasoning attribute-based information and location-based characteristic to ground corresponding 3D objects. To be specific, we propose VGMamba, a novel architecture that systematically models attribute-to-location dependencies while efficiently capturing long-range interactions. VGMamba consists of three key modules: SVD-based Attribute Mamba, which leverages Singular Value Decomposition (SVD) to decompose the extracted visual features into multiple subspaces corresponding to different attribute scales. However, SVD alone lacks long-range modeling capability, which is critical for capturing hierarchical dependencies among attributes. To overcome this, we employ Mamba [11] to sequentially integrate multi-scale attribute features, preserving their interdependencies over long sequences. Compared to Transformers, Mamba maintains linear computational complexity while effectively modeling long-range dependencies among attribute components; the Location Mamba, which models both global and local positional relationships among objects by establishing cross-channel communication through SSM, effectively enhancing object-related features; and the Instructive Dual-Mamba block, which fuses object-related and text features through cross-modal guidance of text query. By replacing computationally expensive attention mechanisms with structured state-space modeling, VGMamba significantly improves efficiency without sacrificing performance. Extensive experiments demonstrate that our method achieves state-of-the-art performance, particularly excelling in scenarios involving multiple objects and ambiguous descriptions. Our chief contributions are threefold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 584, + 295, + 700 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 62, + 584, + 295, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 584, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 62, + 584, + 295, + 607 + ], + "type": "text", + "content": "- We explore a novel mechanism, i.e., attribute-to-location clue reasoning, for performing 3D visual grounding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 608, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 608, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 62, + 608, + 295, + 665 + ], + "type": "text", + "content": "- We propose a novel framework VGMamba, comprising three core modules: the SVD-based Attribute Mamba for disentangling attributes, the Location Mamba for modeling location relations, and the Instructive Dual-Mamba block for text and object-related feature fusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 665, + 295, + 700 + ], + "type": "text", + "content": "- We achieve new state-of-the-art performance on multiple 3D Visual Grounding benchmarks, demonstrating superior accuracy and computational efficiency." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 80, + 398, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 80, + 398, + 92 + ], + "spans": [ + { + "bbox": [ + 314, + 80, + 398, + 92 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 99, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 99, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 99, + 547, + 262 + ], + "type": "text", + "content": "3D Visual Grounding. Recent advancements incorporate self-attention-based scene encoding [2], viewpoint-aware spatial reasoning [29], and multi-attribute Transformers for fine-grained language-visual alignment [38]. However, these attention-based approaches struggle with long-sequence feature modeling and computational efficiency. For multi-object grounding, M3DRef-CLIP [42] introduced the first multi-object dataset and an attention-based fusion module integrating 3D object and 2D rendered image features. Subsequent works [41] refine proposal selection, camera pose optimization, and fusion strategies. However, these methods often embed global 2D features, introducing redundancy. In contrast, we focus on fine-grained attribute extraction as priors to enhance spatial reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 547, + 425 + ], + "type": "text", + "content": "State Space Models. State Space Models (SSMs) [12, 30, 34] have become a prominent approach in sequence modeling, particularly for capturing long-range dependencies with linear time complexity. Furthermore, Mamba [11] introduced a selective state-space mechanism, providing efficient sequence modeling and surpassing traditional Transformers in certain tasks. This advancement led to various extensions across vision applications, including Vision Mamba [43] and VMamba [21], which extends scanning in four directions to fully capture interrelations among image patches. Besides, Mamba-based models have also been applied in video understanding [19, 22, 25, 39] and medical image segmentation [24, 28, 36, 37], demonstrating the versatility and efficiency of SSMs in diverse visual domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 547, + 495 + ], + "type": "text", + "content": "In this work, we introduce Mamba into the object grounding task for the first time, exploring its potential in fine-grained feature extraction. We also propose a novel cross-modal fusion framework for language and object-related features, which achieves outstanding performance and significantly reduces computational costs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 505, + 488, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 488, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 488, + 519 + ], + "type": "text", + "content": "3. Overview of State Space Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "text", + "content": "Recently, state space models (SSMs) [9, 12, 30, 32] have attracted much attention for their ability to model continuous systems, constructing the foundation for the popular Mamba architecture [11]. Particularly, SSMs generally take an input sequence " + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}^L" + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "text", + "content": " as the input and output the corresponding sequence " + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "inline_equation", + "content": "y(t) \\in \\mathbb{R}^L" + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "text", + "content": " through hidden states " + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 525, + 547, + 628 + ], + "type": "text", + "content": " is the number of hidden states. The system is governed by differential equations that describe how the hidden state evolves over time:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 637, + 547, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 637, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 347, + 637, + 547, + 651 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t), \\tag {1}", + "image_path": "b4e8374a01024bb807807279416679a6ab63c811800dfb76b395fb0388e0d5ab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 658, + 547, + 681 + ], + "type": "text", + "content": " are matrices that project the input into the hidden states and subsequently to the output." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 681, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 681, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 681, + 547, + 705 + ], + "type": "text", + "content": "In order to apply SSMs in practice tasks, a discretization process is employed, typically using a time scale parameter" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5296" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 79, + 380, + 299 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 380, + 299 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 380, + 299 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 380, + 299 + ], + "type": "image", + "image_path": "e778ec5bc5e10d21d9e714ed850c870048656c05a56dbc6c33c6138fcceacb16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 305, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 61, + 305, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 305, + 548, + 350 + ], + "type": "text", + "content": "Figure 2. VGMamba for quantity-agnostic 3D Visual Grounding. Based on the given language description, we could utilize two clues to localize corresponding objects accurately, i.e., attribute and location information. To this end, we first design an SVD-based attribute mamba to discover multiple attribute-related objects. Then, a location mamba is further designed to select location-relevant objects. Finally, we present an Instructive Dual-Mamba block to localize the object that matches the given query." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 381, + 82, + 549, + 297 + ], + "blocks": [ + { + "bbox": [ + 381, + 82, + 549, + 297 + ], + "lines": [ + { + "bbox": [ + 381, + 82, + 549, + 297 + ], + "spans": [ + { + "bbox": [ + 381, + 82, + 549, + 297 + ], + "type": "image", + "image_path": "5352586d5d5e8dd8741d74df41b65fe6d2702b70f49457916439407fe948b6cb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 369, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 393 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 62, + 369, + 295, + 393 + ], + "type": "text", + "content": " to convert continuous parameters into discrete ones. The processes are shown as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 400, + 295, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 400, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 67, + 400, + 295, + 415 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} (\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}) \\cdot \\Delta \\mathbf {B}. (2)", + "image_path": "adad6fcddb2b0d7fd7b080c190c855efa7fd65c60c6dde83edea93ec32437213.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 423, + 285, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 285, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 285, + 435 + ], + "type": "text", + "content": "The discretized state-space output can be represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 443, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 443, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 107, + 443, + 295, + 456 + ], + "type": "interline_equation", + "content": "h _ {k} = \\overline {{\\mathbf {A}}} h _ {k - 1} + \\overline {{\\mathbf {B}}} x _ {k}, \\quad y _ {k} = \\mathbf {C} h _ {k}. \\tag {3}", + "image_path": "0efd4b1a109483b0cc236eed1d2bdd783766aca6513b96fe0a2a89e695da12cf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 465, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 465, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 61, + 465, + 295, + 628 + ], + "type": "text", + "content": "Based on the SSMs framework, Mamba [11] employs the Selective Scan Mechanism (S6) to conduct dynamic sequential modeling and capture critical task-related information. Though Mamba-based networks have been verified to be effective in many tasks [19, 24], it is under-explored for 3D Visual Grounding. To this end, we study whether the Mamba could be used to capture the relation between objects and language descriptions. We directly plug Mamba into M3DRef-CLIP [42] and conduct experiments on Multi3DRefer dataset [42]. However, we observe that this operation could not improve the grounding performance. The reason may be that Mamba emphasizes capturing global object information and cannot construct the relation between objects and language descriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 639, + 260, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 639, + 260, + 653 + ], + "spans": [ + { + "bbox": [ + 61, + 639, + 260, + 653 + ], + "type": "text", + "content": "4. VGMamba for 3D Visual Grounding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "content": "In this paper, we propose a new VGMamba architecture, aiming to reason attribute and location clues based on the given language descriptions and improve the performance of 3D Visual Grounding effectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": "Specifically, in Fig. 2 (a), we follow the work [42] and adopt PointGroup [17] as a 3D detector to generate an object proposal list " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "o = (o_1,o_2,\\dots,o_n)" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": " proposal boxes and corresponding features " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{3D}}\\in \\mathbb{R}^{n\\times 32}" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": ". Each proposal is then rendered as " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": " multi-view images, which are encoded by CLIP [27] to extract " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": "-dimensional patch features " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "I\\in \\mathbb{R}^{n\\times m\\times p\\times d}" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": " and aggregated features " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "\\hat{I}\\in \\mathbb{R}^{n\\times m\\times d}" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": " is the patch number of each multiview image. Next, for the given language description, we continue following the work [42] and utilize CLIP to obtain text feature " + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "inline_equation", + "content": "T\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 369, + 547, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 547, + 581 + ], + "type": "text", + "content": "In practice, 3D Visual Grounding can be achieved by reasoning two key clues. One clue is to accurately recognize object-related attribute, e.g., dark blue. Another clue is to effectively obtain the location-relevant visual content, e.g., facing the backboard. With this in mind, we separately design an attribute mamba and location mamba to sufficiently capture object-related attribute and location characteristic." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 593, + 471, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 471, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 471, + 605 + ], + "type": "text", + "content": "4.1. SVD-based Attribute Mamba" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 612, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 547, + 705 + ], + "type": "text", + "content": "In general, given a specific object, the extracted representations involve multiple different attribute characteristics, e.g., color, shape, and texture. However, the 3D Visual Grounding task involves localizing objects that possess specific attributes, such as dark blue. Therefore, it is meaningful to disentangle current representations. To this end, we explore using Singular Value Decomposition (SVD) to perform feature-level decomposition and recombination ac" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 744 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 744 + ], + "type": "text", + "content": "5297" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 81, + 223, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 81, + 223, + 92 + ], + "spans": [ + { + "bbox": [ + 62, + 81, + 223, + 92 + ], + "type": "text", + "content": "cording to the calculated singular values." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": "Particularly, as illustrated in Fig. 2 (b), for each patch feature " + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "inline_equation", + "content": "I_{i}\\in \\mathbb{R}^{d\\times p}" + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": ", SVD is first used to decompose " + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": " into three matrices: " + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "inline_equation", + "content": "I_{i} = U\\Sigma V^{T}\\in \\mathbb{R}^{d\\times p}" + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "inline_equation", + "content": "U\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "inline_equation", + "content": "V\\in \\mathbb{R}^{p\\times p}" + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": " are orthogonal matrices, and " + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": " contains singular values sorted along its main diagonal [10]. Subsequently, inspired by Multi-Scale Attention [18], we introduce Sliding Window to recombine the singular vectors using three sliding windows with varying size " + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 61, + 93, + 296, + 186 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 194, + 295, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 194, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 117, + 194, + 295, + 207 + ], + "type": "interline_equation", + "content": "\\left\\{F _ {r} \\right\\} = \\text {S l i d i n g W i n d o w} (\\Sigma , r), \\tag {4}", + "image_path": "8bbf4ec26dcc5d3b0dadf9bdad6c83b9371f9718016061a82b6813b3a66b87fc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 209, + 295, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 209, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 96, + 209, + 295, + 224 + ], + "type": "interline_equation", + "content": "\\text {S l i d i n g W i n d o w} (\\Sigma , r) = U _ {d \\times r} \\Sigma_ {r \\times r} V _ {r \\times p} ^ {T}, \\tag {5}", + "image_path": "eb70730ba2000760d6e5e4bf15882b3519b1fca9e0eea16a085a762ad91483c3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "r \\in \\{r_1, r_2, r_3\\}" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": ". SlidingWindow refers to combining r consecutive singular vectors into a new multi-scale feature " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "F_r" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": ". Since the singular vectors corresponding large singular values and those corresponding small singular values separately contain generalized coarse-grained information and discriminative fine-grained information, through multiscale operation, we can sufficiently capture attribute information involved in visual representations. Following this, " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "F_r" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": " is processed by three independent Mamba to model attribute feature. Finally, a concatenation followed by a " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": " convolution is applied to merge the multi-scale features, yielding the final feature representations " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "F_{i-att} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": " of the single-view image. For " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": " objects with " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": " rendered images, we can obtain the feature representation " + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "inline_equation", + "content": "F_{att} \\in \\mathbb{R}^{n \\times m \\times d}" + }, + { + "bbox": [ + 61, + 232, + 296, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 402, + 163, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 163, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 163, + 414 + ], + "type": "text", + "content": "4.2. Location Mamba" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 418, + 296, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 418, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 61, + 418, + 296, + 500 + ], + "type": "text", + "content": "As shown in Fig. 2 (a), only relying on the attribute clue could not guarantee accurately identifying the objects corresponding to the given language description. To this end, based on the captured attribute features, we design a cross-channel communication mechanism to facilitate the learning of multi-object location relationships, enforcing the locking of the object matching the given description." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "text", + "content": "Specifically, to capture spatial relationships from a consistent viewpoint, for each rendered object image, we define an encoding " + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "inline_equation", + "content": "pos \\in \\mathbb{R}^{n \\times m \\times 6}" + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "text", + "content": " that includes object center coordinates " + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "text", + "content": " and bounding box dimensions " + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "inline_equation", + "content": "(h,w,l)" + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "text", + "content": ". Following this, we utilize a linear layer to perform a transformation on " + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "inline_equation", + "content": "pos" + }, + { + "bbox": [ + 61, + 500, + 296, + 570 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 578, + 295, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 578, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 115, + 578, + 295, + 592 + ], + "type": "interline_equation", + "content": "p \\hat {o} s = \\operatorname {L i n e a r} (p o s) \\in \\mathbb {R} ^ {n \\times m \\times d}. \\tag {6}", + "image_path": "32d62cdccc7ac5cb09b8473cafafd18ee6ba8523622eb8896fa8fb86ec8be4ad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "text", + "content": "Subsequently, the summation result " + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times m \\times d}" + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "inline_equation", + "content": "p \\hat{o} s" + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\hat{I} \\in \\mathbb{R}^{n \\times m \\times d}" + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{att}} \\in \\mathbb{R}^{n \\times m \\times d}" + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "text", + "content": " is taken to perform the following operations: exploring a dedicated mamba to respectively conduct communication from global and local perspectives. For global communication, the channels of " + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 600, + 296, + 684 + ], + "type": "text", + "content": " proposal objects are treated as a sequence to enable cross-object interaction:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 124, + 693, + 295, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 693, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 124, + 693, + 295, + 706 + ], + "type": "interline_equation", + "content": "I _ {\\text {g l o b a l}} = \\operatorname {M a m b a} _ {\\text {g l o b a l}} \\left(I _ {\\text {s p a}}\\right), \\tag {7}", + "image_path": "ec3ec3f7c074e474c112fbe8adacb8ab9d19f01bbbeafe99abf9e91027452663.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 80, + 547, + 164 + ], + "blocks": [ + { + "bbox": [ + 315, + 80, + 547, + 164 + ], + "lines": [ + { + "bbox": [ + 315, + 80, + 547, + 164 + ], + "spans": [ + { + "bbox": [ + 315, + 80, + 547, + 164 + ], + "type": "image", + "image_path": "34e446207457583f5f4981041dfca6d2d5c3db36ed02820ded22c0b98f669096.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 169, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 169, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 547, + 191 + ], + "type": "text", + "content": "Figure 3. The architecture of Instructive Dual-Mamba (IDM) for conducting multi-modal fusion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 201, + 509, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 509, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 509, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 201, + 509, + 214 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{global}} \\in \\mathbb{R}^{n \\times m \\times d}" + }, + { + "bbox": [ + 313, + 201, + 509, + 214 + ], + "type": "text", + "content": " denotes the SSM output." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "text", + "content": "As for the local communication, the " + }, + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "text", + "content": " nearest neighbors of each object are selected based on Euclidean distance, obtaining multi-neighbor feature " + }, + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times m \\times u \\times d}" + }, + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "text", + "content": ". Cross-channel scanning is then applied to " + }, + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "inline_equation", + "content": "\\hat{I}_{\\mathrm{spa}}" + }, + { + "bbox": [ + 313, + 214, + 547, + 274 + ], + "type": "text", + "content": " and outputs central object features:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 281, + 547, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 281, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 351, + 281, + 547, + 295 + ], + "type": "interline_equation", + "content": "I _ {\\text {l o c a l}} = \\operatorname {S e l e c t C e n t e r} \\left(\\mathrm {M a m b a} _ {\\text {l o c a l}} \\left(\\hat {I} _ {\\mathrm {s p a}}\\right)\\right), \\tag {8}", + "image_path": "42e229821787da6616e2bf0675695c891ce34adb0acb829ee666f28d293ba491.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{local}} \\in \\mathbb{R}^{n \\times m \\times d}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " denotes the current SSM output. Subsequently, by summing " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{global}}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{local}}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": ", and then averaging over the " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " images, we obtain the feature " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{spa}} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " that contains plentiful location information. Finally, " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{spa}}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " is concatenated with the proposal feature " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "F_{3\\mathrm{D}} \\in \\mathbb{R}^{n \\times 32}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " output by the 3D detector, followed by a " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " convolution layer for dimensionality reduction, yielding " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{obj}} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": ". We represent the feature of each object in " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{obj}}" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 302, + 547, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 403, + 465, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 465, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 465, + 414 + ], + "type": "text", + "content": "4.3. Multi-modal Mamba Fusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": "To further enhance the grounding accuracy, we present a multi-modal mamba fusion module to align the proposals to the corresponding language descriptions. Specifically, as shown in Fig. 3, the Instructive Dual-Mamba (IDM) block is proposed to use descriptions as guidance to capture matched object-related feature. Taking the language feature " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": ", object-related feature " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " as the input, we first divide " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " into a forward sequence and a reversed backward sequence, which are then processed by two independent SSMs (Eq. (1)) respectively. The outputs from the two branches are integrated to obtain the updated feature. The processing of input feature " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " is similar to " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": ". According to Eq. (1), the state transition matrices " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " retain critical historical information for updating hidden states, while " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " primarily influence input and output. Therefore, we propose replacing the state transition matrices " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{A}}_G, \\vec{\\mathbf{A}}_G" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " with their counterparts " + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "inline_equation", + "content": "\\vec{\\mathbf{A}}_T, \\vec{\\mathbf{A}}_T" + }, + { + "bbox": [ + 313, + 420, + 547, + 657 + ], + "type": "text", + "content": " derived from language features, enabling cross-modal guidance in extracting object-related features. The remaining parameter matrices remain unchanged. The process of above could be represented as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 324, + 670, + 547, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 670, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 324, + 670, + 547, + 685 + ], + "type": "interline_equation", + "content": "\\overleftarrow {h} _ {\\mathrm {T}} ^ {t} = \\overleftarrow {\\mathbf {A}} _ {\\mathrm {T}} \\overleftarrow {h} _ {\\mathrm {T}} ^ {t - 1} + \\overleftarrow {B} _ {\\mathrm {T}} \\overleftarrow {x} _ {\\mathrm {T}} ^ {t}, \\quad \\overrightarrow {h} _ {\\mathrm {T}} ^ {t} = \\overrightarrow {\\mathbf {A}} _ {\\mathrm {T}} \\overrightarrow {h} _ {\\mathrm {T}} ^ {t - 1} + \\overrightarrow {B} _ {\\mathrm {T}} \\overrightarrow {x} _ {\\mathrm {T}} ^ {t}, \\tag {9}", + "image_path": "89a5ef8d05f3dd54703dcdf9d90e442178e1f489d812099ddac32cbf98d9f712.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 324, + 686, + 547, + 702 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 686, + 547, + 702 + ], + "spans": [ + { + "bbox": [ + 324, + 686, + 547, + 702 + ], + "type": "interline_equation", + "content": "\\overleftarrow {h} _ {\\mathsf {G}} ^ {t} = \\overleftarrow {\\mathbf {A}} _ {\\mathsf {T}} \\overleftarrow {h} _ {\\mathsf {G}} ^ {t - 1} + \\overleftarrow {B} _ {\\mathsf {G}} \\overleftarrow {x} _ {\\mathsf {G}} ^ {t}, \\quad \\overrightarrow {h} _ {\\mathsf {G}} ^ {t} = \\overrightarrow {\\mathbf {A}} _ {\\mathsf {T}} \\overrightarrow {h} _ {\\mathsf {G}} ^ {t - 1} + \\overrightarrow {B} _ {\\mathsf {G}} \\overrightarrow {x} _ {\\mathsf {G}} ^ {t}, \\tag {10}", + "image_path": "2780427425557c5965ac3567f20f0a071a8d6d3a4d16c98ca799b28635082e75.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5298" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 126, + 79, + 485, + 306 + ], + "blocks": [ + { + "bbox": [ + 126, + 79, + 485, + 306 + ], + "lines": [ + { + "bbox": [ + 126, + 79, + 485, + 306 + ], + "spans": [ + { + "bbox": [ + 126, + 79, + 485, + 306 + ], + "type": "table", + "html": "
MethodVenueUnique (~19%)Multiple (~81%)Overall
0.250.50.250.50.250.5
ScanRefer [3]ECCV2067.646.232.121.340.026.1
TGNN [14]AAAI2168.656.829.823.237.429.7
MVT [15]CVPR2277.766.531.925.340.833.3
3D-SPS [23]CVPR2284.166.740.329.848.837.0
EDA [35]CVPR2385.868.649.137.654.642.3
VPP-Net [29]CVPR2486.167.150.339.055.743.3
M3DRef-CLIP [42]ICCV23-77.2-36.8-44.7
G3-LQ [33]CVPR2488.673.350.239.756.044.7
MCLN [26]ECCV2486.972.752.040.857.245.5
MA2TransVG [38]CVPR2486.374.153.841.457.945.7
ConcreteNet [31]ECCV2486.482.142.438.450.646.5
D-LISA [41]NeurIPS24-75.5-40.0-46.9
GPS [16]ECCV24-77.9-42.7-48.1
Chat-Scene [13]NeurIPS2489.682.547.842.955.550.2
PQ3D [44]ECCV2486.778.351.546.257.051.2
VGMamba-91.979.654.847.660.053.9
", + "image_path": "2650de171ada3752c8c5741d94de92d2601f4b852ca2e3d3fb53a46bd0ddf722.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{T / G}}^t" + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "text", + "content": " represents the input at time step " + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\overleftrightarrow{\\mathbf{A}}_{\\mathrm{T}}" + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\overrightarrow{\\mathbf{A}}_{\\mathrm{T}}" + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "text", + "content": " are the state transition matrices used for retaining essential historical information and updating hidden states, while matrices " + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{T / G}}" + }, + { + "bbox": [ + 61, + 354, + 295, + 405 + ], + "type": "text", + "content": " primarily affect input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "text", + "content": "Additionally, we share the gating mechanism parameters " + }, + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "text", + "content": " of the language features to construct cross-modal fusion in the hidden state space, ultimately outputting the average multi-view attribute features of each object in the scene, represented as " + }, + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{att - spa}} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "text", + "content": ". Finally, a convolution operation is applied to output confidence scores " + }, + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{n \\times 1}" + }, + { + "bbox": [ + 61, + 407, + 296, + 489 + ], + "type": "text", + "content": " for each object proposal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 506, + 175, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 506, + 175, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 506, + 175, + 519 + ], + "type": "text", + "content": "4.4. Training Objectives" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "text", + "content": "Building on previous work [42], the loss of VGMamba consists of the 3D Visual Grounding loss " + }, + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{ref}}" + }, + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "text", + "content": ", text-object contrastive loss " + }, + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{con}}" + }, + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "text", + "content": ", and object detection loss " + }, + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{det}}" + }, + { + "bbox": [ + 61, + 526, + 296, + 573 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 590, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 590, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 131, + 590, + 295, + 602 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {r e f}} + L _ {\\text {c o n}} + L _ {\\text {d e t}}. \\tag {11}", + "image_path": "a21c5d3a2d5fe061f0f02f44281f4942c8893545cc123b7cd49e7ec92383def4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 623, + 141, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 623, + 141, + 636 + ], + "spans": [ + { + "bbox": [ + 62, + 623, + 141, + 636 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 646, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 646, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 646, + 296, + 704 + ], + "type": "text", + "content": "To evaluate the effectiveness of our proposed VGMamba on 3D Visual Grounding tasks with varying numbers of target objects, we conducted experiments on both single-object datasets (ScanRefer [3], Nr3D and Sr3D [1]) and multi-object dataset Multi3DRefer [42]." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 354, + 545, + 569 + ], + "blocks": [ + { + "bbox": [ + 61, + 314, + 548, + 336 + ], + "lines": [ + { + "bbox": [ + 61, + 314, + 548, + 336 + ], + "spans": [ + { + "bbox": [ + 61, + 314, + 548, + 336 + ], + "type": "text", + "content": "Table 1. Comparison results on the ScanRefer [3] dataset, in terms of the accuracy evaluated by IoU 0.25 and IoU 0.5. Our proposed VGMamba performs favorably over current state-of-the-art approach PQ3D [44]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 354, + 545, + 569 + ], + "lines": [ + { + "bbox": [ + 317, + 354, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 545, + 569 + ], + "type": "table", + "html": "
MethodNr3DSr3D
OverallHardOverallHard
ReferIt3D [?]35.627.940.831.5
TGNN [14]37.330.645.036.9
M3DRef-CLIP [42]49.443.4--
3D-SPS [23]51.545.162.665.4
EDA [35]52.146.168.162.9
D-LISA [41]53.146.2--
MVT [15]55.449.164.558.8
VPP-Net [29]56.9-68.7-
G3-LQ [33]58.450.773.166.3
MCLN [26]59.8-68.4-
MiKASA [2]64.459.475.267.3
GPS [16]64.957.877.571.6
MA2TransVG [38]65.257.673.969.3
PQ3D [44]66.758.779.772.8
VGMamba68.361.481.374.4
", + "image_path": "1778e7ad0b94046d2457b2225ff4d498fb49fb905eeaead2263ab530df51b240.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 576, + 547, + 599 + ], + "lines": [ + { + "bbox": [ + 313, + 576, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 547, + 599 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on Nr3D and Sr3D dataset. We have highlighted the top-performing three methods in purple." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 622, + 448, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 448, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 448, + 635 + ], + "type": "text", + "content": "5.1. Single-object Grounding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 476, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 476, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 476, + 652 + ], + "type": "text", + "content": "5.1.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 658, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 547, + 704 + ], + "type": "text", + "content": "ScanRefer. The ScanRefer dataset comprises 51,583 natural language descriptions for 11,046 objects across 800 3D scenes from the ScanNet dataset [8]. Each description specifies a target object and its location relations within" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5299" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 111, + 83, + 149, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 83, + 149, + 92 + ], + "spans": [ + { + "bbox": [ + 111, + 83, + 149, + 92 + ], + "type": "text", + "content": "Query Text" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 121, + 185, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 121, + 185, + 138 + ], + "spans": [ + { + "bbox": [ + 89, + 121, + 185, + 138 + ], + "type": "text", + "content": "The green chair next to the brown table with a white vase." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 90, + 176, + 178, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 176, + 178, + 194 + ], + "spans": [ + { + "bbox": [ + 90, + 176, + 178, + 194 + ], + "type": "text", + "content": "The brown chairs under the white rectangular table." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 235, + 179, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 235, + 179, + 254 + ], + "spans": [ + { + "bbox": [ + 88, + 235, + 179, + 254 + ], + "type": "text", + "content": "The white pillow closest to the wall with the painting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 200, + 83, + 250, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 83, + 250, + 92 + ], + "spans": [ + { + "bbox": [ + 200, + 83, + 250, + 92 + ], + "type": "text", + "content": "M3DRef-CLIP" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 103, + 260, + 154 + ], + "blocks": [ + { + "bbox": [ + 194, + 103, + 260, + 154 + ], + "lines": [ + { + "bbox": [ + 194, + 103, + 260, + 154 + ], + "spans": [ + { + "bbox": [ + 194, + 103, + 260, + 154 + ], + "type": "image", + "image_path": "7af92848cd849a159b2c3f9792c387e4c14eda7647d89b5b8a7d49aa75e057e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 192, + 158, + 259, + 209 + ], + "blocks": [ + { + "bbox": [ + 192, + 158, + 259, + 209 + ], + "lines": [ + { + "bbox": [ + 192, + 158, + 259, + 209 + ], + "spans": [ + { + "bbox": [ + 192, + 158, + 259, + 209 + ], + "type": "image", + "image_path": "d0c4f4861ee5c83719a71282974d1e73d2ea4f44c629480fc078445010055178.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 187, + 219, + 257, + 263 + ], + "blocks": [ + { + "bbox": [ + 187, + 219, + 257, + 263 + ], + "lines": [ + { + "bbox": [ + 187, + 219, + 257, + 263 + ], + "spans": [ + { + "bbox": [ + 187, + 219, + 257, + 263 + ], + "type": "image", + "image_path": "081c860578023974147507a4f27e639d032bab4b7f33dc71986f57a09378c8d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 274, + 547, + 296 + ], + "lines": [ + { + "bbox": [ + 61, + 274, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 547, + 296 + ], + "type": "text", + "content": "Figure 4. Qualitative examples of the Multi3DRefer dataset. The results illustrate the grounding performance of M3DRef-CLIP (blue box), D-LISA (blue box), and our VGMamba(green box), compared to the ground truth labels (red box)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 298, + 83, + 324, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 83, + 324, + 91 + ], + "spans": [ + { + "bbox": [ + 298, + 83, + 324, + 91 + ], + "type": "text", + "content": "D-LISA" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 281, + 103, + 344, + 153 + ], + "blocks": [ + { + "bbox": [ + 281, + 103, + 344, + 153 + ], + "lines": [ + { + "bbox": [ + 281, + 103, + 344, + 153 + ], + "spans": [ + { + "bbox": [ + 281, + 103, + 344, + 153 + ], + "type": "image", + "image_path": "f6b018c6ba987e46ce7952b65b2b64175c972dfd434bfee05fc958d84691d748.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 277, + 156, + 345, + 210 + ], + "blocks": [ + { + "bbox": [ + 277, + 156, + 345, + 210 + ], + "lines": [ + { + "bbox": [ + 277, + 156, + 345, + 210 + ], + "spans": [ + { + "bbox": [ + 277, + 156, + 345, + 210 + ], + "type": "image", + "image_path": "93338e41061c3d2a1846cc49f16d4f186e38aa78896d9a1015dfa88b33de1b29.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 271, + 220, + 342, + 264 + ], + "blocks": [ + { + "bbox": [ + 271, + 220, + 342, + 264 + ], + "lines": [ + { + "bbox": [ + 271, + 220, + 342, + 264 + ], + "spans": [ + { + "bbox": [ + 271, + 220, + 342, + 264 + ], + "type": "image", + "image_path": "b1fce43dc233542dee267fcf47ea7ec9e877facdb498f63fdc6d52176622d9a0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 375, + 103, + 438, + 154 + ], + "blocks": [ + { + "bbox": [ + 381, + 79, + 418, + 97 + ], + "lines": [ + { + "bbox": [ + 381, + 79, + 418, + 97 + ], + "spans": [ + { + "bbox": [ + 381, + 79, + 418, + 97 + ], + "type": "text", + "content": "VGMamba (ours)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 375, + 103, + 438, + 154 + ], + "lines": [ + { + "bbox": [ + 375, + 103, + 438, + 154 + ], + "spans": [ + { + "bbox": [ + 375, + 103, + 438, + 154 + ], + "type": "image", + "image_path": "51b9f339182965f4d04895b49b88ca83d09f9849cd7d173467a0430730045022.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 366, + 156, + 434, + 210 + ], + "blocks": [ + { + "bbox": [ + 366, + 156, + 434, + 210 + ], + "lines": [ + { + "bbox": [ + 366, + 156, + 434, + 210 + ], + "spans": [ + { + "bbox": [ + 366, + 156, + 434, + 210 + ], + "type": "image", + "image_path": "3ae35d0d9bed1b0d0aad93858f834b02cc8e5759476240eb8f6a292ae1388847.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 362, + 220, + 432, + 264 + ], + "blocks": [ + { + "bbox": [ + 362, + 220, + 432, + 264 + ], + "lines": [ + { + "bbox": [ + 362, + 220, + 432, + 264 + ], + "spans": [ + { + "bbox": [ + 362, + 220, + 432, + 264 + ], + "type": "image", + "image_path": "4bc2e55ec9184409c94c96d181da8d640898cc3f35dc2575f2a19040bd44527d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 457, + 103, + 521, + 154 + ], + "blocks": [ + { + "bbox": [ + 463, + 85, + 509, + 91 + ], + "lines": [ + { + "bbox": [ + 463, + 85, + 509, + 91 + ], + "spans": [ + { + "bbox": [ + 463, + 85, + 509, + 91 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 103, + 521, + 154 + ], + "lines": [ + { + "bbox": [ + 457, + 103, + 521, + 154 + ], + "spans": [ + { + "bbox": [ + 457, + 103, + 521, + 154 + ], + "type": "image", + "image_path": "f84da7be7687c7062e9369bb8b9dd4e709de942c48c593ff06ddfcb508130193.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 453, + 156, + 521, + 213 + ], + "blocks": [ + { + "bbox": [ + 453, + 156, + 521, + 213 + ], + "lines": [ + { + "bbox": [ + 453, + 156, + 521, + 213 + ], + "spans": [ + { + "bbox": [ + 453, + 156, + 521, + 213 + ], + "type": "image", + "image_path": "38ac7537dc9007e61f99c1e6ece4f4e7636884e2296c6e9d8949240497093dc0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 448, + 220, + 518, + 264 + ], + "blocks": [ + { + "bbox": [ + 448, + 220, + 518, + 264 + ], + "lines": [ + { + "bbox": [ + 448, + 220, + 518, + 264 + ], + "spans": [ + { + "bbox": [ + 448, + 220, + 518, + 264 + ], + "type": "image", + "image_path": "7afff4911b1c37a90619362d3399f3e7a841128bdd0f7e723b38cb34838d2fbe.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 65, + 315, + 293, + 402 + ], + "blocks": [ + { + "bbox": [ + 65, + 315, + 293, + 402 + ], + "lines": [ + { + "bbox": [ + 65, + 315, + 293, + 402 + ], + "spans": [ + { + "bbox": [ + 65, + 315, + 293, + 402 + ], + "type": "table", + "html": "
MethodZT w/o DZT w/DST w/o DST w/DMTAll
D3Net [4]81.632.538.623.335.032.2
M3DRef-CLIP [42]81.839.447.830.637.938.4
M3DRef-CLIP w/NMS [41]79.040.567.640.049.149.3
PQ3D [44]85.457.768.543.640.950.1
D-LISA [41]82.443.767.142.551.051.2
VGMamba83.645.270.946.755.154.9
", + "image_path": "5be0d6f0ff3d58577f18901fee964a67d0bca9b8aac0174715bebdfe2fce0656.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 62, + 410, + 294, + 431 + ], + "lines": [ + { + "bbox": [ + 62, + 410, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 294, + 431 + ], + "type": "text", + "content": "Table 3. Comparison of F1@0.5 results on the Multi3DRefer [42] validation set, where our method outperforms all baselines." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 454, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 454, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 61, + 454, + 295, + 559 + ], + "type": "text", + "content": "the scene, enabling the model to understand both object attributes and location context. The dataset is divided into two subsets: Unique, where the target object is the only instance of its class within the scene, and Multiple, where multiple instances of the same class are present. We evaluate our model on validation set using Acc@0.25 and Acc@0.5, which measure the proportion of predicted bounding boxes that achieve an Intersection over Union (IoU) larger than 0.25 and 0.5 with the ground truth box." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "inline_equation", + "content": "\\text{Nr3D/Sr3D}" + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "inline_equation", + "content": "\\text{Nr3D}" + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "inline_equation", + "content": "\\text{Sr3D}" + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "text", + "content": " datasets, part of the ReferIt3D benchmark [1], are built upon the ScanNet dataset [8] for 3D visual grounding. " + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "inline_equation", + "content": "\\text{Nr3D}" + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "text", + "content": " (Natural Reference in 3D) consists of 41,503 human-annotated referring expressions across 707 indoor scenes, where each description uniquely identifies a target object among distractors. " + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "inline_equation", + "content": "\\text{Sr3D}" + }, + { + "bbox": [ + 61, + 564, + 295, + 705 + ], + "type": "text", + "content": " (Spatial Reference in 3D) contains 83,572 machine-generated expressions based on predefined templates, focusing on spatial relationships between objects. Both datasets provide ground-truth object masks. Following [1], performance is evaluated using accuracy, measuring the proportion of correctly localized target objects." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 316, + 438, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 438, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 438, + 328 + ], + "type": "text", + "content": "5.1.2. Implementation Details" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 332, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 332, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 547, + 460 + ], + "type": "text", + "content": "We implement the proposed VGMamba model using PyTorch and train it end-to-end on a single NVIDIA A6000 GPU. The model is optimized using the AdamW optimizer with a batch size of 4 and a learning rate of 5e-4 with cosine decay scheduling. Following prior work [42], we employ a pre-trained PointGroup [17] module as the detector, which is fine-tuned on the ScanNet dataset. The proposal box with the highest matching score is selected as the final prediction. We adopt the official train/val/test splits provided by the ScanRefer and " + }, + { + "bbox": [ + 312, + 332, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D / Sr3D}" + }, + { + "bbox": [ + 312, + 332, + 547, + 460 + ], + "type": "text", + "content": " datasets to ensure fair comparisons with baseline methods." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 468, + 429, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 429, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 429, + 479 + ], + "type": "text", + "content": "5.1.3. Baseline Comparison" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "content": "ScanRefer. The experimental results on ScanRefer are shown in Tab. 1, with the following key observations: (i) Our method achieves state-of-the-art performance with an overall accuracy of " + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "inline_equation", + "content": "60.0\\%" + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "content": " at IoU 0.25 and " + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "inline_equation", + "content": "53.9\\%" + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "content": " at IoU 0.5, surpassing the best baseline PQ3D [44] by " + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "content": ", respectively. (ii) In the Multiple case, VGMamba attains " + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "inline_equation", + "content": "54.8\\%" + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "content": " at IoU 0.25 and " + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "inline_equation", + "content": "47.6\\%" + }, + { + "bbox": [ + 312, + 483, + 547, + 645 + ], + "type": "text", + "content": " at IoU 0.5, outperforming PQ3D and demonstrating that its attribute-to-location reasoning effectively disambiguates objects under complex scenarios. (iii) Unlike previous methods [13, 44] that show notable performance variations across settings, our VGMamba maintains consistently high accuracy, validating its robust generalization capability in both unique and multiple-object scenarios." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "inline_equation", + "content": "Nr3D/Sr3D" + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "content": ". Tab. 2 presents the grounding accuracy of different methods on the " + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "content": " and Sr3D datasets. VGMamba achieves state-of-the-art performance of " + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "inline_equation", + "content": "68.3\\%" + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "inline_equation", + "content": "81.3\\%" + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "content": " on both datasets, surpassing prior state-of-the-art methods. The " + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "content": " dataset presents additional challenges due to" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5300" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 79, + 294, + 184 + ], + "blocks": [ + { + "bbox": [ + 65, + 79, + 294, + 184 + ], + "lines": [ + { + "bbox": [ + 65, + 79, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 65, + 79, + 294, + 184 + ], + "type": "table", + "html": "
ModulesMulti3DRefer (F1@0.5 on Val)
RowIDMLMSAMZT w/o DZT w/DST w/o DST w/DMTAll
1XXX81.839.447.830.637.938.4
2✓XX83.943.756.839.743.542.3
3X✓X80.240.761.142.845.143.8
4XX✓81.544.067.643.947.845.2
5✓✓X83.546.165.542.348.648.3
6✓X✓82.948.268.745.849.049.2
7X✓✓85.445.369.244.253.152.5
8✓✓✓83.645.270.946.755.154.9
", + "image_path": "209dc3abe4f2db1508be21731d85950122899605fece1cd36b9b4f221cc85352.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 198, + 291, + 209 + ], + "lines": [ + { + "bbox": [ + 66, + 198, + 291, + 209 + ], + "spans": [ + { + "bbox": [ + 66, + 198, + 291, + 209 + ], + "type": "text", + "content": "Table 4. Ablation study of proposed modules on Multi3DRefer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "text", + "content": "its intricate and free-form textual descriptions, which increase the difficulty of cross-modal alignment. Despite this, our method significantly outperforms alternatives that leverage large-scale 3D vision-language pre-training models [13, 44]. Notably, in the more challenging \"Hard\" subset, VGMamba improves accuracy by " + }, + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 61, + 219, + 296, + 324 + ], + "type": "text", + "content": " in Sr3D, demonstrating its superior capability in distinguishing ambiguous objects by effectively capturing fine-grained attribute and spatial relationships." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 331, + 195, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 195, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 195, + 342 + ], + "type": "text", + "content": "5.2. Multi-object Grounding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 348, + 225, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 225, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 225, + 358 + ], + "type": "text", + "content": "5.2.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 361, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 296, + 525 + ], + "type": "text", + "content": "Multi3DRefer is a dataset derived from ScanRefer [3], comprising a total of 61,926 textual descriptions corresponding to 11,609 distinct objects. Each description may reference zero, one, or multiple target objects, enabling the evaluation of grounding models in a variety of scenarios. Following previous work[42], we compute the F1 score at the Intersection over union (IoU) threshold of 0.5(denoted as F1@0.5) across five distinct scenarios: a) zero target without distractors of the same semantic class (ZT w/o D); b) zero target with distractors (ZT w/D); c) single target without distractors (ST w/o D); d) single target with distractors (ST w/D); and e) multiple targets (MT). Finally, the average across these five evaluation cases is calculated to provide an overall performance score." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 530, + 187, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 530, + 187, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 530, + 187, + 541 + ], + "type": "text", + "content": "5.2.2. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 544, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 544, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 61, + 544, + 296, + 627 + ], + "type": "text", + "content": "The experimental setting for the multi-object dataset is similar to that for the single-object dataset, with a few differences: during training, we treat all predicted bounding boxes with an IoU greater than " + }, + { + "bbox": [ + 61, + 544, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{train}} = 0.25" + }, + { + "bbox": [ + 61, + 544, + 296, + 627 + ], + "type": "text", + "content": " relative to the ground truth as target bounding boxes. During inference, we consider all predicted bounding boxes with a matching score greater than " + }, + { + "bbox": [ + 61, + 544, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{pred}} = 0.1" + }, + { + "bbox": [ + 61, + 544, + 296, + 627 + ], + "type": "text", + "content": " as the model's predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 632, + 178, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 632, + 178, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 632, + 178, + 643 + ], + "type": "text", + "content": "5.2.3. Baseline Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 646, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 646, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 646, + 296, + 704 + ], + "type": "text", + "content": "We compare our 3D multi-object grounding performance with existing works quantitatively and qualitatively. As illustrated in Tab. 3, our proposed VGMamba achieves competitive performance across all scenarios. In cases with distractors (ST w/D), VGMamba achieves F1 scores of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 337, + 79, + 525, + 194 + ], + "blocks": [ + { + "bbox": [ + 337, + 79, + 525, + 194 + ], + "lines": [ + { + "bbox": [ + 337, + 79, + 525, + 194 + ], + "spans": [ + { + "bbox": [ + 337, + 79, + 525, + 194 + ], + "type": "image", + "image_path": "3a7a8fbc8d47d4bfd1b88ff83e3a7d384d6d0a820b32355db816599b6858a07c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 201, + 533, + 213 + ], + "lines": [ + { + "bbox": [ + 326, + 201, + 533, + 213 + ], + "spans": [ + { + "bbox": [ + 326, + 201, + 533, + 213 + ], + "type": "text", + "content": "Figure 5. Efficiency-Performance Trade-off on ScanRefer." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "type": "inline_equation", + "content": "46.7\\%" + }, + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "type": "text", + "content": ", surpassing the second-best competitor by " + }, + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "type": "text", + "content": ", which highlights its robustness in managing complex scenes with multiple potential matches. In the MT case, VG-Mamba attains an F1 score that is " + }, + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 312, + 224, + 547, + 305 + ], + "type": "text", + "content": " higher than the second-best competitor, underscoring the advantage of the Instructive Dual-Mamba (IDM) in addressing the multi-object grounding task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 306, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 306, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 312, + 306, + 547, + 422 + ], + "type": "text", + "content": "We show the qualitative comparison with M3DRef-CLIP[42] and D-LISA[41] in Fig. 4. The top two rows illustrate examples involving multiple target objects in complex scenes. Our VGMamba accurately locates the objects matching the given textual descriptions, even in the presence of distractors. The last row demonstrates a single-target case with surrounding distractors. VGMamba effectively isolates and identifies the correct object, while baseline models struggle, often mistakenly including distractors in their predictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 430, + 411, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 411, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 411, + 441 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 448, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 448, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 312, + 448, + 547, + 494 + ], + "type": "text", + "content": "To validate the effectiveness of each proposed module within our VGMamba framework, we conduct ablation studies on the Multi3DRefer dataset, as shown in Tab. 4. More ablation results are detailed in supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 495, + 547, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 547, + 587 + ], + "type": "text", + "content": "Effectiveness of SVD-based Attribute Mamba (SAM). Comparing Row1 with Row4, we observe that adding SAM yields significant improvements across various scenarios, especially in the ST w/o D case, where the F1 score improves by " + }, + { + "bbox": [ + 313, + 495, + 547, + 587 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 313, + 495, + 547, + 587 + ], + "type": "text", + "content": ". This demonstrates SAM's ability to decompose complex representations and capture fine-grained attribute features, which is essential for distinguishing target objects based on descriptive characteristics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 588, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 547, + 668 + ], + "type": "text", + "content": "Effectiveness of Location Mamba (LM). The impact of LM is evident when comparing Row 4 with Row 7. Adding LM enhances performance, particularly in the " + }, + { + "bbox": [ + 313, + 588, + 547, + 668 + ], + "type": "inline_equation", + "content": "MT" + }, + { + "bbox": [ + 313, + 588, + 547, + 668 + ], + "type": "text", + "content": " scenario, where the F1 score increases by " + }, + { + "bbox": [ + 313, + 588, + 547, + 668 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 588, + 547, + 668 + ], + "type": "text", + "content": ". This improvement demonstrates the importance of LM in capturing location relationships between objects, allowing the model to better handle cases with distractors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 670, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 547, + 705 + ], + "type": "text", + "content": "Effectiveness of Instructive Dual-Mamba (IDM). The contribution of IDM is highlighted by comparing Row 7 with Row 8. Incorporating IDM results in consistent per" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 314, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 314, + 744 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 314, + 744 + ], + "type": "text", + "content": "5301" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 217, + 101, + 277, + 159 + ], + "blocks": [ + { + "bbox": [ + 126, + 81, + 166, + 90 + ], + "lines": [ + { + "bbox": [ + 126, + 81, + 166, + 90 + ], + "spans": [ + { + "bbox": [ + 126, + 81, + 166, + 90 + ], + "type": "text", + "content": "Query Text" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 101, + 277, + 159 + ], + "lines": [ + { + "bbox": [ + 217, + 101, + 277, + 159 + ], + "spans": [ + { + "bbox": [ + 217, + 101, + 277, + 159 + ], + "type": "image", + "image_path": "e4c88b316b8665d4771a5c03d8d1bfa6093b80f44e3c464c6768755032164cf4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 217, + 167, + 280, + 228 + ], + "blocks": [ + { + "bbox": [ + 102, + 120, + 201, + 141 + ], + "lines": [ + { + "bbox": [ + 102, + 120, + 201, + 141 + ], + "spans": [ + { + "bbox": [ + 102, + 120, + 201, + 141 + ], + "type": "text", + "content": "The red sofa is against the wall with a white cushion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 167, + 280, + 228 + ], + "lines": [ + { + "bbox": [ + 217, + 167, + 280, + 228 + ], + "spans": [ + { + "bbox": [ + 217, + 167, + 280, + 228 + ], + "type": "image", + "image_path": "2875bf1191b9d68d0310ad3268f71a9768da104fa981fde662fc69f58cb06107.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 216, + 231, + 280, + 285 + ], + "blocks": [ + { + "bbox": [ + 88, + 186, + 203, + 206 + ], + "lines": [ + { + "bbox": [ + 88, + 186, + 203, + 206 + ], + "spans": [ + { + "bbox": [ + 88, + 186, + 203, + 206 + ], + "type": "text", + "content": "A wooden chair without arms is tucked under the table." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 248, + 201, + 269 + ], + "lines": [ + { + "bbox": [ + 90, + 248, + 201, + 269 + ], + "spans": [ + { + "bbox": [ + 90, + 248, + 201, + 269 + ], + "type": "text", + "content": "A black chair without armrests, back to the window." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 231, + 280, + 285 + ], + "lines": [ + { + "bbox": [ + 216, + 231, + 280, + 285 + ], + "spans": [ + { + "bbox": [ + 216, + 231, + 280, + 285 + ], + "type": "image", + "image_path": "f6ea0d90da02796e5254e24cfd42d3d3570721f420dfe02824829148dc66a6c4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 296, + 462, + 307 + ], + "lines": [ + { + "bbox": [ + 147, + 296, + 462, + 307 + ], + "spans": [ + { + "bbox": [ + 147, + 296, + 462, + 307 + ], + "type": "text", + "content": "Figure 6. Visualization of the attribute, location relation, and overall scores in our model." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 294, + 101, + 356, + 159 + ], + "blocks": [ + { + "bbox": [ + 222, + 80, + 271, + 89 + ], + "lines": [ + { + "bbox": [ + 222, + 80, + 271, + 89 + ], + "spans": [ + { + "bbox": [ + 222, + 80, + 271, + 89 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 81, + 352, + 89 + ], + "lines": [ + { + "bbox": [ + 299, + 81, + 352, + 89 + ], + "spans": [ + { + "bbox": [ + 299, + 81, + 352, + 89 + ], + "type": "text", + "content": "Attribute Score" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 101, + 356, + 159 + ], + "lines": [ + { + "bbox": [ + 294, + 101, + 356, + 159 + ], + "spans": [ + { + "bbox": [ + 294, + 101, + 356, + 159 + ], + "type": "image", + "image_path": "fc5d3d221026f955a46a573647194a13f2008f34fac8ab91a578e66372bc7248.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 296, + 167, + 356, + 228 + ], + "blocks": [ + { + "bbox": [ + 296, + 167, + 356, + 228 + ], + "lines": [ + { + "bbox": [ + 296, + 167, + 356, + 228 + ], + "spans": [ + { + "bbox": [ + 296, + 167, + 356, + 228 + ], + "type": "image", + "image_path": "f36053b5c8c927fd46cf62b1191df4b46c4304bc9b3c11ac65dc9a363b227926.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 296, + 234, + 358, + 285 + ], + "blocks": [ + { + "bbox": [ + 296, + 234, + 358, + 285 + ], + "lines": [ + { + "bbox": [ + 296, + 234, + 358, + 285 + ], + "spans": [ + { + "bbox": [ + 296, + 234, + 358, + 285 + ], + "type": "image", + "image_path": "3fdb66e3d5d2449bc1559f9d98a89fd04fe122e2cd9d3631456e74a431522a2d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 101, + 438, + 160 + ], + "blocks": [ + { + "bbox": [ + 381, + 80, + 433, + 89 + ], + "lines": [ + { + "bbox": [ + 381, + 80, + 433, + 89 + ], + "spans": [ + { + "bbox": [ + 381, + 80, + 433, + 89 + ], + "type": "text", + "content": "Location Score" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 101, + 438, + 160 + ], + "lines": [ + { + "bbox": [ + 376, + 101, + 438, + 160 + ], + "spans": [ + { + "bbox": [ + 376, + 101, + 438, + 160 + ], + "type": "image", + "image_path": "0debec5e2c53069c271c33197ca906284d269a6f17478993254f2de2de0c7324.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 377, + 167, + 440, + 228 + ], + "blocks": [ + { + "bbox": [ + 377, + 167, + 440, + 228 + ], + "lines": [ + { + "bbox": [ + 377, + 167, + 440, + 228 + ], + "spans": [ + { + "bbox": [ + 377, + 167, + 440, + 228 + ], + "type": "image", + "image_path": "f8dbd50435c96bd44c64b9258a01632685b5f654ee4a74d0132038988e64a0bf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 378, + 233, + 440, + 285 + ], + "blocks": [ + { + "bbox": [ + 378, + 233, + 440, + 285 + ], + "lines": [ + { + "bbox": [ + 378, + 233, + 440, + 285 + ], + "spans": [ + { + "bbox": [ + 378, + 233, + 440, + 285 + ], + "type": "image", + "image_path": "c5bf46b0ea7f5098b83cfa8cbb7447b56f7029f3eb2010d1f6499598083d5bc2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 457, + 101, + 518, + 162 + ], + "blocks": [ + { + "bbox": [ + 461, + 80, + 508, + 89 + ], + "lines": [ + { + "bbox": [ + 461, + 80, + 508, + 89 + ], + "spans": [ + { + "bbox": [ + 461, + 80, + 508, + 89 + ], + "type": "text", + "content": "Overall Score" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 101, + 518, + 162 + ], + "lines": [ + { + "bbox": [ + 457, + 101, + 518, + 162 + ], + "spans": [ + { + "bbox": [ + 457, + 101, + 518, + 162 + ], + "type": "image", + "image_path": "268055ec478ada3ce8efaf501374ee53beda0ffce44a2b1190f9653684901175.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 458, + 168, + 521, + 228 + ], + "blocks": [ + { + "bbox": [ + 458, + 168, + 521, + 228 + ], + "lines": [ + { + "bbox": [ + 458, + 168, + 521, + 228 + ], + "spans": [ + { + "bbox": [ + 458, + 168, + 521, + 228 + ], + "type": "image", + "image_path": "1649b3a7550b2b3fb94000171c9d50bab1b61d6166df9e6ee6813a661721995b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 234, + 521, + 285 + ], + "blocks": [ + { + "bbox": [ + 459, + 234, + 521, + 285 + ], + "lines": [ + { + "bbox": [ + 459, + 234, + 521, + 285 + ], + "spans": [ + { + "bbox": [ + 459, + 234, + 521, + 285 + ], + "type": "image", + "image_path": "faaec61c9cab9784a95c8c10509a835bc1f04fbb77c71dd48172174430dcba2a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 64, + 325, + 294, + 364 + ], + "blocks": [ + { + "bbox": [ + 64, + 325, + 294, + 364 + ], + "lines": [ + { + "bbox": [ + 64, + 325, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 64, + 325, + 294, + 364 + ], + "type": "table", + "html": "
MethodZT w/o DZT w/DST w/o DST w/DMTAll
Location First82.143.968.746.354.753.1
Attribute First83.645.270.946.755.154.9
", + "image_path": "fc90ff9ba27eb698b21835d1446c2162a44a5d04cfe532dea3ef5f76348ac2b0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 61, + 373, + 295, + 394 + ], + "lines": [ + { + "bbox": [ + 61, + 373, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 61, + 373, + 295, + 394 + ], + "type": "text", + "content": "Table 5. Comparison of grounding performance between attribute first and location first model." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 400, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 400, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 61, + 400, + 296, + 458 + ], + "type": "text", + "content": "formance gains across all metrics, with a notable increase in the " + }, + { + "bbox": [ + 61, + 400, + 296, + 458 + ], + "type": "inline_equation", + "content": "MT" + }, + { + "bbox": [ + 61, + 400, + 296, + 458 + ], + "type": "text", + "content": " case, where the F1 score improves by " + }, + { + "bbox": [ + 61, + 400, + 296, + 458 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 61, + 400, + 296, + 458 + ], + "type": "text", + "content": ". This shows IDM's effectiveness in handling multi-object grounding, enabling precise language-instructed grounding in complex scenes with multiple target objects." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 61, + 459, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 459, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 61, + 459, + 296, + 610 + ], + "type": "text", + "content": "Efficiency-Performance Trade-off. Fig. 5 illustrates the balance between computational complexity (FLOPs), inference speed, and grounding performance on the ScanRefer dataset. VGMamba achieves the highest accuracy with significantly lower FLOPs than methods like Chat-Scene, demonstrating superior efficiency. This indicates that our method achieves a superior balance between efficiency and performance, making it well-suited for real-world applications where both accuracy and inference speed are critical. The lightweight nature of VGMamba stems from its efficient attribute and location feature sequence modeling with Mamba, which enables better feature aggregation without excessive computational overhead." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 61, + 611, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 611, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 611, + 296, + 704 + ], + "type": "text", + "content": "Why Attribute First? We conducted experiments on the Multi3DRefer val set to compare the performance between the location first and our proposed attribute first model. As shown in Tab. 5, the attribute first approach achieves relatively higher performance. One possible explanation is that attribute cues serve as more distinguishing priors compared to location cues, enabling the model to initially identify a larger and more accurate set of candidate objects. In con" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 327, + 547, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 547, + 351 + ], + "type": "text", + "content": "trast, location information is often less specific and could potentially mislead the model toward irrelevant objects." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 357, + 448, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 448, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 448, + 369 + ], + "type": "text", + "content": "5.4. Interpretability Analysis" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 373, + 548, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 548, + 559 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 548, + 559 + ], + "type": "text", + "content": "Benefiting from the sequential attribute-to-location reasoning process, VGMamba demonstrates a high degree of interpretability. To visualize the model's understanding of attribute and location relationships, we utilize the output features from the SVD-based Attribute Mamba, Location Mamba, and Instructive Dual-Mamba modules to obtain the attribute score, location score, and overall score. Fig. 6 presents qualitative results, where darker colors indicate higher scores. Specifically, our model first identifies objects that match attribute cues, such as \"red sofa,\" \"wooden chair,\" and \"black chair.\" Then, using these attributes as prior, it incorporates location cues to filter out mismatched objects, e.g., \"against the wall,\" \"under the table,\" and \"back to the window\" (highlighted in yellow boxes). The final prediction scores, generated by the Instructive Dual-Mamba block, are highlighted in red boxes." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 569, + 384, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 384, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 384, + 580 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 588, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 588, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 588, + 548, + 704 + ], + "type": "text", + "content": "In this paper, we are concerned with 3D Visual Grounding that aims to identify 3D objects matching the language description. Particularly, we consider that human beings often follow two clues to achieve the corresponding grounding, i.e., attribute and location clues. To this end, we explore a new mechanism, i.e., attribute-to-location clue reasoning, and propose a new VGMamba method that consists of an SVD-based attribute mamba, location mamba, and multi-modal fusion mamba. Experimental results on multiple datasets demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5302" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 80, + 163, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 80, + 163, + 94 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 163, + 94 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 100, + 297, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 100, + 297, + 170 + ], + "spans": [ + { + "bbox": [ + 61, + 100, + 297, + 170 + ], + "type": "text", + "content": "Our work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), the Joint Fund of the Ministry of Education of China (Nos. 8091B022149, 8091B02072404), the National Natural Science Foundation of China (No. 62132016), and the Natural Science Basic Research Program of Shaanxi (No. 2020JC-23)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 181, + 120, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 181, + 120, + 193 + ], + "spans": [ + { + "bbox": [ + 63, + 181, + 120, + 193 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 200, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 200, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 200, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 67, + 200, + 295, + 243 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 244, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 244, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 67, + 244, + 296, + 286 + ], + "type": "text", + "content": "[2] Chun-Peng Chang, Shaoxiang Wang, Alain Pagani, and Didier Stricker. Mikasa: Multi-key-anchor & scene-aware transformer for 3d visual grounding. In CVPR, pages 14131-14140, 2024. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 288, + 295, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 288, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 67, + 288, + 295, + 320 + ], + "type": "text", + "content": "[3] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221, 2020. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 322, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 322, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 67, + 322, + 295, + 364 + ], + "type": "text", + "content": "[4] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In ECCV, pages 487-505. Springer, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 365, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 365, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 365, + 295, + 407 + ], + "type": "text", + "content": "[5] Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. History aware multimodal transformer for vision-and-language navigation. Advances in neural information processing systems, 34:5834-5847, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 409, + 295, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 409, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 67, + 409, + 295, + 461 + ], + "type": "text", + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 464, + 295, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 464, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 67, + 464, + 295, + 506 + ], + "type": "text", + "content": "[7] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Think global, act local: Dual-scale graph transformer for vision-and-language navigation. In CVPR, pages 16537-16547, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 508, + 295, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 508, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 67, + 508, + 295, + 560 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 562, + 295, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 562, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 67, + 562, + 295, + 604 + ], + "type": "text", + "content": "[9] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 606, + 295, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 606, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 63, + 606, + 295, + 638 + ], + "type": "text", + "content": "[10] Gene H. Golub and William Kahan. Calculating the singular values and pseudo-inverse of a matrix. In Milestones in Matrix Computation, 2007. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 639, + 295, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 639, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 63, + 639, + 295, + 670 + ], + "type": "text", + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 63, + 672, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 672, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 672, + 295, + 704 + ], + "type": "text", + "content": "[12] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 82, + 547, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 82, + 547, + 134 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 547, + 134 + ], + "type": "text", + "content": "[13] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 136, + 547, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 136, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 315, + 136, + 547, + 177 + ], + "type": "text", + "content": "[14] Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text-guided graph neural networks for referring 3d instance segmentation. In AAAI, pages 1610-1618, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 178, + 547, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 178, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 547, + 209 + ], + "type": "text", + "content": "[15] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 211, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 211, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 315, + 211, + 547, + 263 + ], + "type": "text", + "content": "[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In ECCV, pages 289-310. Springer, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 265, + 547, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 265, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 315, + 265, + 547, + 305 + ], + "type": "text", + "content": "[17] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, pages 4867-4876, 2020. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 307, + 546, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 307, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 315, + 307, + 546, + 350 + ], + "type": "text", + "content": "[18] Jiayu Jiao, Yu-Ming Tang, Kun-Yu Lin, Yipeng Gao, Andy J Ma, Yaowei Wang, and Wei-Shi Zheng. Dilateformer: Multiscale dilated transformer for visual recognition. IEEE transactions on multimedia, 25:8906-8919, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 351, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 351, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 547, + 392 + ], + "type": "text", + "content": "[19] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, pages 237-255. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 394, + 546, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 394, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 315, + 394, + 546, + 436 + ], + "type": "text", + "content": "[20] Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. Bird's-eye-view scene graph for vision-language navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10968-10980, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 437, + 547, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 437, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 437, + 547, + 479 + ], + "type": "text", + "content": "[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 479, + 546, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 479, + 546, + 511 + ], + "spans": [ + { + "bbox": [ + 315, + 479, + 546, + 511 + ], + "type": "text", + "content": "[22] Hui Lu, Albert Ali Salah, and Ronald Poppe. Videomambapro: A leap forward for mamba in video understanding. arXiv preprint arXiv:2406.19006, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 512, + 547, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 512, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 315, + 512, + 547, + 553 + ], + "type": "text", + "content": "[23] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16454-16463, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 555, + 546, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 546, + 586 + ], + "type": "text", + "content": "[24] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 586, + 547, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 586, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 315, + 586, + 547, + 639 + ], + "type": "text", + "content": "[25] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 640, + 547, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 640, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 315, + 640, + 547, + 682 + ], + "type": "text", + "content": "[26] Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Multi-branch collaborative learning network for 3d visual grounding. In ECCV, pages 381-398. Springer, 2024. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 683, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 683, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 315, + 683, + 547, + 704 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 743 + ], + "type": "text", + "content": "5303" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 82, + 295, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 82, + 82, + 295, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 82, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 82, + 82, + 295, + 124 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 126, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 126, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 64, + 126, + 295, + 158 + ], + "type": "text", + "content": "[28] Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 159, + 295, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 159, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 295, + 190 + ], + "type": "text", + "content": "[29] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Viewpoint-aware visual grounding in 3d scenes. In CVPR, pages 14056-14065, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 192, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 192, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 64, + 192, + 295, + 223 + ], + "type": "text", + "content": "[30] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 225, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 225, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 64, + 225, + 295, + 267 + ], + "type": "text", + "content": "[31] Ozan Unal, Christos Sakaridis, Suman Saha, and Luc Van Gool. Four ways to improve verbo-visual fusion for dense 3d visual grounding. In European Conference on Computer Vision, pages 196–213. Springer, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 270, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 270, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 64, + 270, + 295, + 322 + ], + "type": "text", + "content": "[32] Jue Wang, Wentao Zhu, Pichao Wang, Xiang Yu, Linda Liu, Mohamed Omar, and Raffay Hamid. Selective structured state-spaces for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6387-6397, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 323, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 323, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 64, + 323, + 295, + 376 + ], + "type": "text", + "content": "[33] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13917-13926, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 378, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 378, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 64, + 378, + 295, + 441 + ], + "type": "text", + "content": "[34] Yuxuan Wang, Among Wu, Muli Yang, Yukuan Min, Yihang Zhu, and Cheng Deng. Reasoning mamba: Hypergraph-guided region relation calculating for weakly supervised affordance grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27618-27627, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 444, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 444, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 64, + 444, + 295, + 495 + ], + "type": "text", + "content": "[35] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19231-19242, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 498, + 295, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 498, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 64, + 498, + 295, + 539 + ], + "type": "text", + "content": "[36] Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for multimodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 541, + 295, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 541, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 64, + 541, + 295, + 594 + ], + "type": "text", + "content": "[37] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 596, + 295, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 596, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 64, + 596, + 295, + 627 + ], + "type": "text", + "content": "[38] Can Xu, Yuehui Han, Rui Xu, Le Hui, Jin Xie, and Jian Yang. Multi-attribute interactions matter for 3d visual grounding. In CVPR, pages 17253–17262, 2024. 1, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 629, + 295, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 629, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 64, + 629, + 295, + 670 + ], + "type": "text", + "content": "[39] Chenghao Xu, Guangtao Lyu, Jiexi Yan, Muli Yang, and Cheng Deng. Llm knows body language, too: Translating speech voices into human gestures. In ACL, pages 5004-5013, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 64, + 673, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 673, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 64, + 673, + 295, + 704 + ], + "type": "text", + "content": "[40] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 255 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 82, + 547, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 82, + 547, + 124 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 547, + 124 + ], + "type": "text", + "content": "[41] Haomeng Zhang, Chiao-An Yang, and Raymond A Yeh. Multi-object 3d grounding with dynamic modules and language-informed spatial attention. arXiv preprint arXiv:2410.22306, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 126, + 547, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 126, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 316, + 126, + 547, + 157 + ], + "type": "text", + "content": "[42] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In ICCV, pages 15225-15236, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 159, + 547, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 159, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 159, + 547, + 200 + ], + "type": "text", + "content": "[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 202, + 547, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 202, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 202, + 547, + 255 + ], + "type": "text", + "content": "[44] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188-206. Springer, 2024. 5, 6, 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "type": "text", + "content": "5304" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_content_list.json b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c766ee0898e8ae4e91da84af4e1ef51aa6049cb7 --- /dev/null +++ b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_content_list.json @@ -0,0 +1,1502 @@ +[ + { + "type": "text", + "text": "VIGFace: Virtual Identity Generation for Privacy-Free Face Recognition dataset", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minsoo Kim $^{1,2}$ Min-Cheol Sagong $^{1}$ Gi Pyo Nam $^{1,2}$ Junghyun Cho $^{1,2,3}$ Ig-Jae Kim $^{1,2}$", + "bbox": [ + 129, + 179, + 867, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Korea Institute of Science and Technology(KIST)", + "bbox": [ + 297, + 198, + 699, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Korea National University of Science and Technology", + "bbox": [ + 282, + 215, + 714, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Yonsei-KIST Convergence Research Institute", + "bbox": [ + 315, + 233, + 683, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kim1102,mcsagong,gpnam,jhcho,drjay}@kist.re.kr", + "bbox": [ + 289, + 253, + 702, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning-based face recognition continues to face challenges due to its reliance on huge datasets obtained from web crawling, which can be costly to gather and raise significant real-world privacy concerns. To address this issue, we propose VIGFace, a novel framework capable of generating synthetic facial images. Our idea originates from pre-assigning virtual identities in the feature space. Initially, we train the face recognition model using a real face dataset and create a feature space for both real and virtual identities, where virtual prototypes are orthogonal to other prototypes. Subsequently, we train the diffusion model based on the established feature space, enabling it to generate authentic human face images from real prototypes and synthesize virtual face images from virtual prototypes. Our proposed framework provides two significant benefits. Firstly, it shows clear separability between existing individuals and virtual face images, allowing one to create synthetic images with confidence and without concerns about privacy and portrait rights. Secondly, it ensures improved performance through data augmentation by incorporating real existing images. Extensive experiments demonstrate the superiority of our virtual face dataset and framework, outperforming the previous state-of-the-art on various face recognition benchmarks. https://github.com/kim1102/VIGFace", + "bbox": [ + 89, + 335, + 485, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 727, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning-based Face Recognition (FR) models have significantly improved their performance due to recent advances in network architectures [17, 18, 21, 26, 45, 47] and enhancements in loss functions [9, 28, 32, 42, 43, 54]. The latest FR models utilize the softmax-variant loss for training to reduce intra-class variance and increase inter-class separability in the embedding space. This necessitates a very large dataset with numerous distinct individuals, signifi", + "bbox": [ + 89, + 753, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cant variations within each individual for intra-class variance, and precise labels of subject identities (IDs). However, datasets are typically collected through web-crawling and then refined using automatic techniques that employ FR logits [11, 60]. Although this approach is successful in eliminating mislabeled data, it still struggles with persistent issues of small intra-class variance. Moreover, in the face recognition field, unlike other ordinary image classifications, there are practical issues such as portrait rights, making it even more difficult to collect training data. For example, data sets such as those referenced in [16, 56, 60] consist of images of celebrities collected from the internet without consent. Furthermore, the datasets mentioned in [27, 38] include facial images of the general population, including children. The privacy sensitivity of such data poses significant challenges for face recognition research.", + "bbox": [ + 511, + 305, + 906, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthetic datasets have been used to address the limitations caused by the scarcity of real datasets [8, 25, 49] and biases present in the available real datasets [20, 50]. In the realm of face recognition, artificial faces show promise in addressing the aforementioned issues related to real face datasets. Generated faces are at low risk for label noise due to conditional generation. Bias problems, such as long-tailed distributions, which lead to class imbalances, can be mitigated by data augmentation. Importantly, there are no privacy concerns with virtual identities facial images if the method is effective. Therefore, when creating artificial datasets, it is essential to consider: 1) ensuring that the generated data mirror the real data distribution, 2) the ability to generate new subjects separated from real data, and 3) maintaining ID consistency for each subject.", + "bbox": [ + 511, + 549, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous attempts to create artificial face datasets have addressed some of the three aspects individually, but to the best of our knowledge, none have simultaneously taken into account all three aspects[2, 5, 29, 40]. SynFace [40] introduces a high-quality synthetic face image dataset, closely reminiscent of real face images, using DiscoFaceGAN [13]. However, DiscoFaceGAN is only capable of generating a limited number of unique subjects, fewer than 500 [29]. On", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work", + "bbox": [ + 109, + 886, + 362, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10043", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6fc969961fa4e39eacfade113c5cc0a22fb6a00e7a5626d4f3f72c95b52c3a00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 92, + 903, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d819eacfc352f6cc1d4054a1391b661b8d4656347a54fde7f2b7cc07959c2683.jpg", + "image_caption": [ + "Figure 1. T-distributed Stochastic Neighbor Embedding (T-SNE) [51] plot of embeddings from real and synthesized images. The filled and lined stars represent the real and virtual prototypes, while filled and lined circles indicate the embeddings of real and synthesized images, respectively. The bottom of the figure shows the face images included in the cluster, and the dotted outlined images represent the face images generated using our method." + ], + "image_footnote": [], + "bbox": [ + 106, + 232, + 303, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0eb5df0cbc394491258ccb41da6775eb48bccd7eebe237385f3727e11f36ca39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 232, + 503, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a5903228c2906cfd2ad202515e58cd723c174f6cd93af2b5d3d05f33ce5f5ef0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 232, + 700, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6b4b0b13e19c97ea681a4a19be14469eceae1dba0d8193022065b58ed38325c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 232, + 900, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the other hand, DigiFace [2] utilizes 3D parametric modeling to create synthetic face images of various subjects. However, it faces difficulties in accurately replicating the quality and style distribution observed in real face images. DCFace [29] proposes a diffusion-based method to generate data that maintains the style of real datasets and ensures label consistency. However, DCFace primarily focuses on creating synthetic data rather than enhancing the capabilities of data augmentation. IDiffFace [5] utilizes identity-conditioned latent diffusion to synthesize facial images from FR feature representations. However, it fails to guarantee the uniqueness of the synthesized identities. CemiFace [48] and HSFace [55] demonstrate high performance on the various FR benchmarks. However, we found identity leakage of the training dataset from both CemiFace and HSFace, which is critical from a privacy standpoint.", + "bbox": [ + 88, + 371, + 483, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The virtual data generated by our method possesses all three essential properties of virtual data mentioned above. The main concept of our paper is to incorporate virtual prototypes into the FR model. Virtual prototypes are trained simultaneously with real prototypes so that they pre-assign on the same feature space. The diffusion model is trained to preserve the identity of the individuals when generating face images based on the FR embedding. As virtual prototypes are allocated within the same embedding space as real prototypes, the produced face images are ensured to reflect the true data distribution. In addition, since prototypes are orthogonal to others, virtual subjects are guaranteed to be distinguished from the original data. Fig. 1 shows a toy example that visualizes the embeddings of real and virtual subjects to demonstrate the effectiveness of the suggested method. The virtual embedding can be distinct from the real individual clusters, whereas the images generated from virtual prototypes form unique clusters. The visualization", + "bbox": [ + 89, + 628, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrates that our framework generates distinctive virtual human faces with high consistency in ID. The proposed method effectively addresses privacy concerns by creating datasets of non-existent individuals and achieving state-of-the-art performance compared to models trained with previous virtual face generation methods. Additionally, the FR model, which was trained using a combination of real and synthetic images together, achieves better performance than the model trained using only real images. This shows that the proposed method is also superior from the perspective of a data augmentation method.", + "bbox": [ + 511, + 372, + 906, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution can be summarized as follows:", + "bbox": [ + 531, + 537, + 851, + 551 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Introducing VIGFace, a method designed to generate virtual identities with realistic appearance, guided by three main criteria: generating novel subjects that are not found in the real-world, consistently preserving the identity, but ensuring diverse characteristics for each individual.", + "- Showing that the synthetic data generated by VIGFace can leverage intra-class diversity and inter-class variance, achieving SOTA performance in face recognition.", + "- Releasing the virtual-only face dataset that can fully substitute the real dataset, helping alleviate privacy concerns." + ], + "bbox": [ + 511, + 553, + 905, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 718, + 651, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Face Recognition Models", + "text_level": 1, + "bbox": [ + 511, + 742, + 741, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current state-of-the-arts (SOTA) FR methods, such as [3, 9, 12, 24, 28, 32, 36, 53], are designed based on softmax loss, but particularly utilize the angular/cosine distance instead of the Euclidean distance. The goal of these methods is to maximize the similarity between the embeddings and the ground-truth (GT) prototype, while minimizing the similarity between the embedding and prototypes of other classes. As a result, all embeddings in the same class, including prototypes, converge while maintaining their maxi", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10044", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/32cca86e77a26c1dd7fe611425fe8ab10da25f53c42bf38953a159929143ef5e.jpg", + "image_caption": [ + "Stage 1", + "Stage 2" + ], + "image_footnote": [], + "bbox": [ + 99, + 109, + 496, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef7405955034e53b21402fc9ee04575c8d1e2172a1180975db261f930f7bd042.jpg", + "image_caption": [ + "Figure 2. Pipeline for the proposed method. Conventional FR training includes prototypes for only real individuals, indicated as $W_{R}$ . We add $k$ prototypes for virtual IDs, denoted as $W_{V}$ . The virtual embedding $f_{\\mathrm{FR}}'(x_j)$ corresponding to the virtual person ID: $j$ is generated to follow distribution of the real embeddings. To synthesize the facial image from virtual prototypes, we adopt the DiT architecture [39], following the design approach of the Vision Transformer (ViT) [14]. Additionally, we adjust the DiT model to utilize 5-point landmark images to handle pose variations." + ], + "image_footnote": [], + "bbox": [ + 506, + 118, + 895, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mum distance from other classes. Therefore, if the embedding dimension is large enough, the clusters of each class will be nearly orthogonal [6, 9]. Naturally, a larger dataset that contains a greater variety of images allows for the training of better-performing FR models. Therefore, over time, larger and more diverse datasets [1, 16, 22, 60] have been collected and published in the academic field. However, issues such as portrait rights and the high cost of building large-scale data remain unresolved.", + "bbox": [ + 88, + 386, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Synthetic Face Image Generation", + "text_level": 1, + "bbox": [ + 89, + 534, + 382, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SynFace [40] proposed a method that uses DiscoFaceGAN[13] to synthesize virtual faces. DigiFace-1M [2] proposes a 3D model-based face rendering method to generate virtual face data. DCFace [29] proposes a diffusion-based face generator combining subject appearance (ID) and external factor (style) conditions. GanDiffFace [35] uses GANs to generate identity features. Vec2Face [55] introduces a masked autoencoder to control the identity of face images and their attributes. CemiFace [48] produces facial samples with various levels of similarity to the center of the subject. IDiffFace [5] uses latent diffusion conditioned on FR identity to generate synthetic images. Although these methods succeed in generating separable identities, some still require additional networks or thresholding to sample synthetic identities. Furthermore, unlike [5], we propose a novel method for the pre-assignment of FR representations to maintain latent space alignment and introduce pose control of facial images using facial landmarks.", + "bbox": [ + 89, + 556, + 483, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 844, + 189, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework comprises two stages. First, we train the FR model using the real face dataset and design the feature", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "space for both real and virtual IDs. Second, synthetic images are generated using the diffusion model based on the feature space of the pre-trained FR model that was trained in the first stage. This section provides a detailed explanation of each stage of the proposed framework.", + "bbox": [ + 511, + 386, + 905, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Stage 1: FR Model Training", + "text_level": 1, + "bbox": [ + 511, + 469, + 766, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed framework begins with training the FR model using real face images. This stage serves two purposes: 1) Training the FR model to achieve prototype features from face images, which are necessary for training the diffusion model in the second stage, and 2) Simultaneously assigning the positions of both the real ID and the virtual ID on the feature space. We choose ArcFace [9] to train the FR model in this stage. The ArcFace loss used to train the FR backbone $f_{\\mathrm{FR}}$ and the prototype $W = [w_{1}, w_{2}, \\dots, w_{n}]$ can be described as follows:", + "bbox": [ + 511, + 492, + 906, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {a r c}} = - \\log \\frac {e ^ {s \\cos \\left(\\theta_ {y _ {i}} + m\\right)}}{e ^ {s \\cos \\left(\\theta_ {y _ {i}} + m\\right)} + \\sum_ {j = 1 , j \\neq y _ {i}} ^ {n} e ^ {s \\cos \\theta_ {j}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 648, + 906, + 689 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\cos \\theta_ {j} = \\frac {w _ {j} ^ {T} f _ {\\mathrm {F R}} \\left(x _ {i}\\right)}{\\| w _ {j} \\| \\| f _ {\\mathrm {F R}} \\left(x _ {i}\\right) \\|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 695, + 903, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n$ denotes the total number of real IDs, while $m$ and $s$ represent the margin and scale hyperparameters, respectively. With conventional FR training methods, only prototypes for real individuals, which can be denoted as $W_{R} = [w_{r}^{1},w_{r}^{2},\\dots,w_{r}^{n}]$ , are necessary for training. In contrast, we include additional $k$ prototypes for virtual IDs, denoted as $W_{V} = [w_{v}^{1},w_{v}^{2},\\dots,w_{v}^{k}]$ , which are used to generate facial images of nonexistent individuals in the real world. As a result, the prototype is defined as a linear transformation matrix $W\\in \\mathbb{R}^{(n + k)\\times D}$ , where $D$ refers to the dimension of the embedding features.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10045", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43bd49bd2bc2e5d82b88ffca18bc08f6beecd074260540f272dd54e911c5c1f7.jpg", + "image_caption": [ + "(a) Prototypes that are trained without virtual embeddings.", + "Figure 3. Changes in the similarity matrix of the prototypes from our method. Similarity values were min-max normalized." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 282, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0437c452840701ea3dd8d6ffde051f60d69aa442a3b6fa9965447d2ae05dbb9.jpg", + "image_caption": [ + "(b) Prototypes that are trained with both real and virtual embeddings." + ], + "image_footnote": [], + "bbox": [ + 292, + 87, + 483, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, due to the absence of a face image for virtual IDs, their prototypes cannot be updated to maximize the distance from each other with $L_{arc}$ . Consequently, all virtual prototypes converge to a single point in the feature space when trained with the original $L_{arc}$ as shown in Fig. 3a. To address this problem, we propose to use virtual feature embedding $f_{\\mathrm{FR}}'(x_j)$ to update the virtual prototypes $w_v^j$ . The virtual embedding $f_{\\mathrm{FR}}'(x_j)$ corresponding to the virtual person ID: $j$ was generated as follows:", + "bbox": [ + 89, + 330, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {F R}} ^ {\\prime} (x _ {j}) = w _ {v} ^ {j} + \\mathcal {N} (0, 1) \\cdot \\sigma , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 479, + 482, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {2} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\left(f _ {\\mathrm {F R}} \\left(x _ {i}\\right) - w _ {r} ^ {i}\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 508, + 482, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $b$ refers the mini-batch size. As can be seen from the equations, the virtual embedding $f_{\\mathrm{FR}}'(x_j)$ follows a distribution in which the standard deviation matches that of the real embeddings. The aggregate loss $L_{\\mathrm{arc}}$ is calculated simultaneously using both the real embedding $f_{\\mathrm{FR}}(x)$ and the virtual embedding $f_{\\mathrm{FR}}'(x)$ , allowing the virtual prototype $w_v^j$ to maintain minimal similarities with other prototypes, as illustrated in Fig. 3. Since batch configuration affects the calculated standard deviation, we utilize the exponential moving average (EMA) to reduce this influence. The corrected standard deviation $\\sigma$ for the current iteration $t$ is calculated as follows:", + "bbox": [ + 89, + 556, + 483, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\sigma_ {(t)} \\cdot \\alpha + \\sigma_ {(t - 1)} \\cdot (1 - \\alpha). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 751, + 482, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The hyperparameter $\\alpha$ is set to 0.9. The number of virtual embedding, $b_{v}$ is determined based on the mini-batch size $b_{r}$ , the number of virtual ID prototypes $k$ , and the number of real ID prototypes $n$ . In our study, we set $b_{v} = (k\\times b_{r}) / n$ so that both real and virtual prototypes can be updated evenly.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline of this stage is illustrated in Fig. 2. Note that, as can be seen in the figure, gradients for virtual embeddings have no effect on the backbone. We confirm", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that the similarity distribution between the virtual prototypes closely resembles the similarity distribution between the real prototypes, as shown in Fig. 3b.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stage 2: Face Generation with Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 165, + 906, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The next step is to synthesize face images using the diffusion model. To obtain the training dataset for the diffusion model, we utilize the FR model, which involves collecting pairs of images $x_0$ and their corresponding prototypes $w_r$ . The input to our diffusion model includes the timestep $t$ , the FR prototype vector $w_r$ , the five facial landmark image $y$ , and the noisy image $x_t$ . In line with the approach proposed in the previous method [31], our model predicts velocity $v_t$ rather than noise $\\epsilon$ injected into $x_t$ . We adopt the DiT architecture [39] using the design approach of the visual transformer (ViT) [14]. We modify the DiT model by incorporating the five facial landmarks (including the left eye, the right eye, the nose, the left mouth corner, and the right mouth corner) [57] image. The five facial landmark images are acquired using RetinaFace [10], and employed as conditions to account for pose variations. Fig. 4 shows the synthesized face images conditioned by five facial landmarks. The first column lists input landmark conditions in different pose environments. As illustrated in the figure, the proposed model is capable of producing a wide range of styles (like hair, glasses, and lighting) while also delivering various pose variations, all with a high degree of consistency. In order to achieve both the generation of facial images and the synchronization of their feature space on the FR model, our diffusion model incorporates a constraint which aims to minimize the feature distance between the original image and the input prototypes as follows:", + "bbox": [ + 511, + 188, + 906, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\epsilon , t} \\| f _ {\\mathrm {F R}} (\\hat {x} _ {\\theta} (x _ {t}, t, w _ {r}, y))) - w _ {r} \\| _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 608, + 906, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt classifier-free guidance [19] by randomly assigning zero values to condition embeddings, $w_{r}$ , $10\\%$ of time. Sampling is performed as follows:", + "bbox": [ + 511, + 642, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {x} _ {\\theta} \\left(x _ {t}, t, W, y\\right) = g \\cdot x _ {\\theta} \\left(x _ {t}, t, W, y\\right) \\\\ + (1 - g) \\cdot x _ {\\theta} (x _ {t}, t, y), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 705, + 903, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{\\theta}(x_t, t, w_r, y)$ and $x_{\\theta}(x_t, t, y)$ are conditional and unconditional $x_0$ -prediction, respectively and $g$ is guidance weight.", + "bbox": [ + 511, + 753, + 906, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 814, + 645, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our implementation details for the full framework are described in Appendix A. We analyze our framework from two perspectives. Firstly, by training the FR model exclusively with facial images generated for virtual IDs, we", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10046", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d369e7e0eb34aed5072b34e6e75537d7dbf96fc545a00cc7c6aa741f0e5cf5e.jpg", + "image_caption": [ + "Figure 4. Virtual face generated in VIGFace (B). Each row lists facial images in different pose environments created using five facial landmarks. Our method can generate various conditions of face images, such as illumination, occlusion by accessories, and facial expressions, while controlling the pose variations of the face images." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 406, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/09cd79ebeface17a8f98d200536d9ba5a88e0f98d308e421cbcadd9ae7374381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 89, + 656, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/910c7a759608f38d4e511cb8105f360d78e14c165b3968a9e1e032c6fe672044.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 90, + 905, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "demonstrated the capability of our model to serve as a viable alternative to real face datasets, addressing concerns such as label noise or privacy. Secondly, we train the FR model using both real images and generated face images simultaneously, showcasing the potential of our model as a data augmentation framework for face recognition.", + "bbox": [ + 89, + 381, + 483, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Virtual Identity Generation", + "text_level": 1, + "bbox": [ + 89, + 483, + 339, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since diffusion models focus on optimizing the visual aspect with input conditions, they tend to produce images that are high in consistency but lack of variance. One simple way to achieve a balance between consistency and variance is by adjusting the classifier-free guidance weight scale. Higher guidance enforces stronger conditioning of input labels. In other words, it results in more constant images but slightly similar samples. We observe that the proposed model performs best as the scale becomes $w = 4.0$ .", + "bbox": [ + 89, + 506, + 483, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct a toy example visualizing the feature space consists of three real persons and five virtual persons to demonstrate the efficacy of the synthesized images in Fig. 1. As seen in the figure on the left, the virtual embedding optimization obviously provides separable virtual prototypes from real individual clusters. The figure on the right illustrates that our synthesized images enforce high variance to the dataset. Synthesized images of real individuals filling the gap in the cluster provide intra-class variance to the subjects. Additionally, the cluster of generated images from virtual prototypes, while staying separate from other subjects, demonstrates the effectiveness of our framework for generating unique virtual human faces in high consistency.", + "bbox": [ + 89, + 643, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 5 shows the qualitative comparison of conventional synthetic datasets and the proposed VIGFace. As shown in the figure, the facial images generated using the VIGFace model show remarkable uniformity in generating consistent", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "virtual individuals, while also incorporating variations such as hair styles, accessories, makeup, and expressions. It is important to emphasize that the key characteristic of a training dataset to achieve a high-performance FR model is not the high quality of the images themselves, but rather the high consistency and variance of the generated facial images belonging to the same person.", + "bbox": [ + 511, + 381, + 906, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Property of generated Dataset To compare the properties of the synthetic face datasets, we measured the 1) class consistency, 2) class separability, and 3) intra-class diversity of generated images. The class consistency reflects the uniformity of the samples in the same label condition. Consequently, the consistency of class $k$ ( $C_k$ ) was measured as follows:", + "bbox": [ + 511, + 505, + 905, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {k} = \\frac {1}{N ^ {2}} \\sum_ {j = 1} ^ {N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(x _ {i}\\right) \\cdot f \\left(x _ {j}\\right)}{\\| f \\left(x _ {i}\\right) \\| \\| f \\left(x _ {j}\\right) \\|}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 618, + 905, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of images in a class $k$ . Higher class consistency means that the samples are more uniform under the same label.", + "bbox": [ + 511, + 671, + 905, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also measured the class separability to assess the integrity of the dataset; in other words, to ensure that all subjects in the dataset are unique. The class separability for a class $k$ ( $S_k$ ) is measured as the average distance between the center of class $k$ and the centers of the negative classes, as follows:", + "bbox": [ + 511, + 715, + 905, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} _ {k} = \\frac {1}{K - 1} \\sum_ {i = 1, i \\neq k} ^ {K} 1 - \\frac {\\overline {{f _ {k}}} \\cdot \\overline {{f _ {i}}}}{\\| \\overline {{f _ {k}}} \\| \\| \\overline {{f _ {i}}} \\|}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 815, + 903, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K$ is the number of classes. $\\overline{f_k}$ represents the class center obtained by averaging the embedding vectors of the", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10047", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33feea5ec5acd2ac0a96374e0954de8ffb24046050a5198d485078224ec50a6f.jpg", + "image_caption": [ + "Figure 5. Comparison of randomly selected virtual ID images generated by conventional methods [2, 5, 29, 35, 40, 48, 55] and by our method, all trained on CASIA-WebFace. For each synthetic dataset, we present two subjects in a single row." + ], + "image_footnote": [], + "bbox": [ + 101, + 94, + 477, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images that belong to class $k$ . As shown in the formula, a high $S_{k}$ indicates that the images are distinct from the negative classes.", + "bbox": [ + 89, + 489, + 483, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The intra-class diversity measures how various the conditions of samples are under the same label. In particular, we focus on challenging scenarios that directly impact FR performance, such as pose variations, occlusions, or lighting conditions, rather than image style. Higher diversity indicates that the dataset covers a broad spectrum of cases, from easy to hard. Motivated by this observation, we calculated the intra-class diversity of a class $k(\\mathrm{D}_k)$ using the variance of Face Image Quality Assessment (FIQA) scores as follows:", + "bbox": [ + 89, + 536, + 483, + 685 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathrm {F I Q A} _ {k}}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathrm {F I Q A} \\left(x _ {i}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 695, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D} _ {k} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\operatorname {F I Q A} \\left(x _ {i}\\right) - \\overline {{\\operatorname {F I Q A} _ {k}}}\\right) ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 747, + 482, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{FIQA}(x_i)$ indicates the normalized CR-FIQA [4] score of the generated image $x_{i}$ .", + "bbox": [ + 89, + 794, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The scores $C_k$ and $S_k$ are derived from a pre-trained ArcFace model that was trained using the Glint-360K dataset. Fig. 6a shows the average values of the properties for all classes, normalized by the average values achieved with CASIA-WebFace. VIGFace achieves remarkable scores in", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbece17d284ceb0522e4bdb2bd1ea3224fb0e884f519d7c15b5936df81388791.jpg", + "image_caption": [ + "(a) The normalized properties of various (b) Identity leakage test on the synthetic datasets. The sizes of the cir- synthetic face dataset and its cor-cles indicate the intra-class diversity. responding training dataset." + ], + "image_footnote": [], + "bbox": [ + 514, + 92, + 727, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00623ab2a7fa511910c0a76de64cf27de0e23be7051e412b45ab70afa25afaa8.jpg", + "image_caption": [ + "Figure 6. Properties of VIGFace compared with those of synthesized datasets generated by previous methods." + ], + "image_footnote": [], + "bbox": [ + 730, + 95, + 901, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "aspects of class consistency and class separability compared to other methods. SynFace exhibits the lowest separability due to its mix-up generation method. This fact indicates that SynFace can generate only a limited number of subjects as mentioned in the introduction. DigiFace achieves strong consistency, due to its 3D rendering methods, but its image style may differ significantly from real datasets, leading to lower accuracy in real-world applications. GANDiffFace demonstrates inadequate FIQA diversity, indicating limited variations in factors such as pose, lighting, and occlusion. DCFace, which uses a diffusion model, shows separability similar to that of VIGFace but with lower consistency.", + "bbox": [ + 511, + 363, + 906, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In comparisons of SOTA models, HSFace, despite using a more refined dataset, comparatively lacks consistency. CemiFace, trained on the same dataset as VIGFace, shows low consistency and separability. Both methods also exhibit inferior qualitative uniformity within the same class, as illustrated in Fig. 5. Our approach uses prototypes instead of FR embedding vectors to disentangle identity features from other characteristics, allowing VIGFace to generate subjects with greater consistency by focusing on identity features. Moreover, it enhances separability by leveraging the pre-computation of feature orthogonality. We included further analysis on the properties of the datasets and provided examples in Appendix H.", + "bbox": [ + 511, + 545, + 908, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Identity leakage from real human To claim complete privacy-free, it is necessary to prove that no real IDs or training images are included in the synthesized dataset. In this reason, we demonstrate that the generated face images represent non-existent humans by querying the most similar face in the CASIA-WebFace dataset. For comparison, we also present the similarity values of the nearest negative class in CASIA-WebFace itself. In Fig. 6b, we count the number of synthetic face images whose similarity with", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10048", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/90faf2b729efadf545d320fef6bcc4720b9dad36148f99a5d54561d73c39cc7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTraining Dataset# of Images (classes × variations)Verification Benchmarks
LFWCFP-FPCPLFWAgeDBCALFWAvg.
CASIA-Webface (Real)-0.49M (≈10.5K×47)99.4096.6390.2394.6893.7094.93
SynFaceFFHQ0.5M (10K × 50)91.9375.0370.4361.6374.7374.75
DigiFace3D modeling0.5M (10K × 50)95.4087.4078.8776.9778.6283.45
DCFaceFFHQ+CASIA0.5M (10K × 50)98.5585.3382.6289.7091.6089.56
IDiffFaceFFHQ0.5M (10K × 50)98.0085.4780.4586.4390.6588.20
GANDiffFaceFFHQ0.5M (10K × 50)90.7773.2772.3266.3574.6875.48
ID3CASIA0.5M (10K × 50)97.6886.8482.7791.0090.7389.80
CemiFaceCASIA+WebFace4M0.5M (10K × 50)99.0391.0687.6291.3392.4292.30
Arc2FaceWebFace42M0.5M (10K × 50)98.8191.8785.1690.1892.6391.73
HyperFaceWebFace42M0.5M (10K × 50)98.5088.8384.2386.5389.4089.50
HSFace10KWebFace4M0.5M (10K × 50)98.8788.9785.4793.1293.5792.00
VIGFace(S), OursCASIA0.5M (10K × 50)99.0295.0987.7290.9590.0092.56
DCFaceFFHQ+CASIA1.2M (20K × 50 + 40K × 5)98.5888.6185.0790.9792.8291.21
CemiFaceCASIA+WebFace4M1.2M (20K × 50 + 40K × 5)99.2292.8488.8692.1393.0393.22
Arc2FaceWebFace42M1.2M (20K × 50 + 40K × 5)98.9294.5886.4592.4593.3393.14
HSFace20KWebFace4M1.0M (20K × 50)98.8789.8786.1393.8593.6592.47
HSFace300KWebFace4M15M (300K × 50)99.3091.5487.7094.4594.5893.52
VIGFace(B), OursCASIA1.2M (60K × 20)99.4897.0790.1593.6292.8894.64
VIGFace(L), OursCASIA6.0M (120K × 50)99.3397.3191.1293.8292.9594.91
", + "bbox": [ + 91, + 87, + 903, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. FR benchmark results trained with various virtual face datasets. All results except for CASIA-WebFace and VIGFace are obtained from the original paper. Our method is specified by the number of identities as small (S), base (B) and large (L). FR backbone is IR-SE50 + AdaFace [28]. Bold and underline indicates the best and the second best, respectively.", + "bbox": [ + 89, + 376, + 906, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the average feature vector of each class in the training data exceeds a given threshold. As shown in the figure, face images in VIGFace have low similarity to the subjects in CASIA-WebFace. This supports the claim that each synthesized image depicts a non-existent human face and there is no identity leakage from the training data. In contrast, we observed that some of the conventional SOTA methods exhibit identity leakage. As CemiFace [48] samples identity embeddings from the WebFace4M dataset, it generates individuals that exhibit resemblance to those in WebFace4M. Vec2Face [55] fails in creating novel identities, so that it produces almost same persons in WebFace4M. Further analysis and examples are reported in Appendix I. In this paper, we avoid thresholding or selective sampling, which could harm fair comparisons with previous methods.", + "bbox": [ + 88, + 435, + 485, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 672, + 212, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we train the FR model with VIGFace to compare it to conventional methods. Detailed hyperparameters for training the FR network can be found in Appendix A.", + "bbox": [ + 89, + 696, + 483, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VIGFace as Virtual Dataset We compare the performance of the FR model trained with generated facial images for virtual IDs using VIGFace with conventional methods. For the experiment, we set the number of virtual IDs to $10K$ , $60K$ and $120K$ . To ensure a fair comparison, we established image counts of 0.5M, 1.2M, and 6.0M, which reflects the scale of the real CASIA-WebFace dataset and previous synthetic approaches, respectively. In Tab. 1, we present the 1:1 verification accuracy (\\%) on five bench", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "marks [23, 37, 44, 58, 59].", + "bbox": [ + 511, + 435, + 691, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the table, VIGFace outperforms conventional methods in the average accuracy of five verification benchmarks. In particular, VIGFace achieves equal or even better performance on CFP-FP and CPLFW compared to the model trained with the real dataset, CASIA-WebFace. This indicates that our method benefits from its pose-variable generation capabilities. CemiFace shows notable performance on the AgeDB and CALFW benchmarks, but they sampled identity embeddings from the WebFace4M [60] dataset. Consequently, the individuals included in the CemiFace dataset may still have privacy concerns. HSFace10K also delivers outstanding results among conventional methods. However, it is important to note that HSFace benefits from training on the extensive WebFace4M dataset, which boasts twice as many images and quadruples the number of identities when compared to CASIA-WebFace. VIGFace(L) outperforms HSFace300K even with a smaller number of images. As a result, when the FR model was trained using VIGFace, it achieved performance comparable to a model trained using a real dataset in terms of average accuracy without any external identity sampling. This suggests that our virtual dataset can serve as a full replacement for the CASIA-WebFace dataset to train the FR model, while avoiding privacy concerns.", + "bbox": [ + 509, + 450, + 906, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VIGFace as Data Augmentation To demonstrate the efficacy of VIGFace as an augmentation framework, we evaluate the accuracy of the FR model trained on real and synthetic images. We utilized the dataset that was uploaded", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10049", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/15e02a97e2f8b3b5b5abcbcc7fbeefc5f5c41fd8a5a96d6b7ba34a064472ad2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConditionVerification Benchmarks
MethodReal ImageSynthetic ImageLFWCFP-FPCPLFWAgeDBCALFWAvg.Δ
Real IDVirtual ID
CASIA-WebFace✓99.4096.6390.2394.6893.7094.93-
DigiFace✓✓99.3797.5190.9294.9593.7795.30+0.37
DCFace (1.2M)✓✓99.4396.9790.3395.2094.3895.26+0.33
iDiffFace✓✓99.5897.0490.4094.7894.0095.16+0.23
GANDiffFace✓✓99.5296.6190.3093.9893.5794.80-0.13
HSFace20K✓✓99.6097.4191.0795.4894.4095.59+0.66
VIGFace(B)✓✓99.4597.2390.7895.2593.2295.19+0.26
✓✓99.5598.0391.8095.7394.1295.85+0.92
✓✓✓99.7098.1091.5795.8594.3895.92+0.99
", + "bbox": [ + 91, + 88, + 903, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Augmented FR performance results for various conditions of synthetic dataset. The accuracies (%) for LFW, CFP-FP, CPLFW, AgeDB-30, CALFW, and the average benchmark accuracies are reported. FR backbone is IR-SE50 + AdaFace [28].", + "bbox": [ + 89, + 280, + 906, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the official repository. Tab. 2 shows the performance change of the trained FR model using various combinations of datasets. DigiFace, which utilizes unique 3D modeling technology, has demonstrated improved benchmark performance and offers advantages from a data-augmentation perspective. However, DigiFace struggles to perform well on its own, devoid of real data [2]. This is due to their failure to precisely replicate the appearance of real-world facial images, making it impractical to present a privacy-free synthetic dataset. HSFace shows a notable improvement among conventional methods. However, it utilized WebFace4M for training, which may make the comparison with others unfair. The boosted accuracy using VIGFace, which outperforms conventional methods, demonstrates that VIGFace can accurately mirror real facial data and achieve synergy by blending them. In particular, augmented images of real IDs improve the results of the CFP-FP and CPLFW benchmarks. This observation suggests that the use of five-point facial landmarks in the conditioning method can create a variety of posed facial images, substantially improving the FR model's ability to understand pose variations. Consequently, VIGFace not only effectively solves the privacy issue, but can also be used with real data as part of the augmentation process. When trained with the VIGFace dataset, the FR backbone achieves better generalization and higher performance, in contrast to conventional methods that achieved only marginal performance improvements.", + "bbox": [ + 88, + 325, + 485, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 illustrates the results that synthesize variations of real and virtual IDs. Since the training process is significantly biased not just by a lack of subject but also by inner-class variance, increasing the variance within the class is crucial. The results demonstrate that the proposed approach can generate a variety of conditions for the existing real ID without any help from external data. Since VIGFace can generate varied images of real individuals, we performed experiments on each type of augmentation, i.e. synthetic images of real and virtual ID. In particular, we extend the long-tailed real-ID class, which contains fewer than 50 im", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ages, to 50 images. With this method, additional 0.15M images were added as real ID synthesis. As can be seen in Tab. 2, the augmented dataset for real ID shows improved FR performance on most benchmarks. This indicates that our method can increase the intra-class diversity of dataset, a critical factor in achieving high FR performance. As a result, the FR model that utilizes the entire set of augmented data exhibits the best performance, benefiting from enhanced intra-class diversity and inter-class variance.", + "bbox": [ + 511, + 325, + 908, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 476, + 640, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents an effective method for creating a synthetic face dataset that guarantees unique virtual identities. Synthesized facial data can serve as a solution for various challenges faced by traditional real datasets, such as high expenses, inaccuracies and biases in labeling, and concerns regarding privacy. To this end, we propose the Virtual Identity Generation framework and demonstrate that it can generate not only realistic but also diverse facial images of virtual individuals, significantly narrowing the performance gap with FR models trained on real data. Furthermore, the model exhibits superior performance when trained on a combination of VIGFace and existing real data compared to models trained solely on real data. This confirms that our proposed method has potential as an effective augmentation technique. We will publicly distribute the virtual face dataset created by VIGFace and believe that these new virtual data will contribute to resolving the privacy issues inherent in face recognition training dataset.", + "bbox": [ + 511, + 503, + 908, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This work was partially supported by the Institutional Programs of the Korea Institute of Science and Technology (KIST) (Project No. 2E33612, $50\\%$ ) and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2023-00227592, Development of Robust 3D Object Identification Technology for Viewpoint Changes, $50\\%$", + "bbox": [ + 511, + 787, + 910, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10050", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 3, 2", + "[2] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 1, 2, 3, 6, 8, 4", + "[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1578-1587, 2022. 2", + "[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5836-5845, 2023. 6, 3", + "[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 1, 2, 3, 6, 4", + "[6] Johann S Brauchart, Alexander B Reznikov, Edward B Saff, Ian H Sloan, Yu Guang Wang, and Robert S Womersley. Random point sets on the sphere—hole radii, covering, and separation. Experimental Mathematics, 27(1):62-81, 2018. 3, 2", + "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 1", + "[8] Hyunwoo Cho, Haesol Park, Ig-Jae Kim, and Junghyun Cho. Data augmentation of backscatter x-ray images for deep learning-based automatic cargo inspection. Sensors, 2021. 1", + "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 2, 3", + "[10] Jiankang Deng, Jia Guo, Yuxiang Zhou, Jinke Yu, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641, 2019. 4", + "[11] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 741-757. Springer, 2020. 1", + "[12] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep" + ], + "bbox": [ + 93, + 143, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 2", + "bbox": [ + 547, + 93, + 903, + 132 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 1, 3", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2010.3, 4", + "[15] Papantoniou et al. Arc2face: A foundation model for id-consistent human faces. In ECCV, 2024. 1", + "[16] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 3", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 1", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4, 1", + "[20] Je Hyeong Hong, Hanjo Kim, Minsoo Kim, Gi Pyo Nam, Junghyun Cho, Hyeong-Seok Ko, and Ig-Jae Kim. A 3d model-based approach for fitting masks to faces in the wild. In 2021 IEEE International Conference on Image Processing (ICIP), 2021. 1", + "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 1", + "[22] Gary Huang, Marwan Mattar, Honglak Lee, and Erik Learned-Miller. Learning to align from scratch. Advances in neural information processing systems, 25, 2012. 3, 1", + "[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 7", + "[24] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 2", + "[25] Hochul Hwang, Cheongjae Jang, Geonwoo Park, Junghyun Cho, and Ig-Jae Kim. Eldersim: A synthetic data generation" + ], + "bbox": [ + 93, + 143, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "platform for human action recognition in eldercare applications. IEEE Access, 2023. 1", + "[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 1", + "[27] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 1", + "[28] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7, 8", + "[29] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. Dface: Synthetic face generation with dual condition diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12715-12725, 2023. 1, 2, 3, 6, 4", + "[30] D Kinga, Jimmy Ba Adam, et al. A method for stochastic optimization. In International conference on learning representations (ICLR), page 6. San Diego, California.; 2015. 1", + "[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 4, 1", + "[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1, 2", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1", + "[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 2", + "[35] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 3, 6, 4", + "[36] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14225-14234, 2021. 2", + "[37] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 7", + "[38] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7044-7053, 2017. 1", + "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4", + "[40] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 1, 3, 6, 4", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2", + "[42] Swami Sankaranarayanan, Azadeh Alavi, Carlos D Castillo, and Rama Chellappa. Triplet probabilistic embedding for face verification and clustering. In 2016 IEEE 8th international conference on biometrics theory, applications and systems (BTAS), pages 1-8. IEEE, 2016. 1", + "[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1", + "[44] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 7", + "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1", + "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1", + "[47] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 1", + "[48] Zhonglin Sun, Siyang Song, Ioannis Patras, and Georgios Tzimiropoulos. Cemifice: Center-based semi-hard synthetic face generation for face recognition. Advances in Neural Information Processing Systems, 37:35612-35638, 2025. 2, 3, 6, 7, 4", + "[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10052", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 1", + "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2", + "[52] Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, and Chen Change Loy. The devil of face recognition is in the noise. In Proceedings of the European Conference on Computer Vision (ECCV), pages 765-780, 2018. 1", + "[53] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2", + "[54] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 499-515. Springer, 2016. 1", + "[55] Haiyu Wu, Jaskirat Singh, Sicong Tian, Liang Zheng, and Kevin W Bowyer. Vec2face: Scaling face dataset generation with loosely constrained vectors. arXiv preprint arXiv:2409.02979, 2024. 2, 3, 6, 7, 1, 4", + "[56] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 1", + "[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 4", + "[58] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 7", + "[59] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 7", + "[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10492-10502, 2021. 1, 3, 7" + ], + "bbox": [ + 91, + 92, + 480, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_model.json b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f956ee87830a0009ddb4e0b634e3668fca78111b --- /dev/null +++ b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.908, + 0.153 + ], + "angle": 0, + "content": "VIGFace: Virtual Identity Generation for Privacy-Free Face Recognition dataset" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.18, + 0.868, + 0.2 + ], + "angle": 0, + "content": "Minsoo Kim\\(^{1,2}\\) Min-Cheol Sagong\\(^{1}\\) Gi Pyo Nam\\(^{1,2}\\) Junghyun Cho\\(^{1,2,3}\\) Ig-Jae Kim\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.199, + 0.7, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)Korea Institute of Science and Technology(KIST)" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.217, + 0.715, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Korea National University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.234, + 0.684, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)Yonsei-KIST Convergence Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.255, + 0.704, + 0.27 + ], + "angle": 0, + "content": "{kim1102,mcsagong,gpnam,jhcho,drjay}@kist.re.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.715 + ], + "angle": 0, + "content": "Deep learning-based face recognition continues to face challenges due to its reliance on huge datasets obtained from web crawling, which can be costly to gather and raise significant real-world privacy concerns. To address this issue, we propose VIGFace, a novel framework capable of generating synthetic facial images. Our idea originates from pre-assigning virtual identities in the feature space. Initially, we train the face recognition model using a real face dataset and create a feature space for both real and virtual identities, where virtual prototypes are orthogonal to other prototypes. Subsequently, we train the diffusion model based on the established feature space, enabling it to generate authentic human face images from real prototypes and synthesize virtual face images from virtual prototypes. Our proposed framework provides two significant benefits. Firstly, it shows clear separability between existing individuals and virtual face images, allowing one to create synthetic images with confidence and without concerns about privacy and portrait rights. Secondly, it ensures improved performance through data augmentation by incorporating real existing images. Extensive experiments demonstrate the superiority of our virtual face dataset and framework, outperforming the previous state-of-the-art on various face recognition benchmarks. https://github.com/kim1102/VIGFace" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.728, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Deep learning-based Face Recognition (FR) models have significantly improved their performance due to recent advances in network architectures [17, 18, 21, 26, 45, 47] and enhancements in loss functions [9, 28, 32, 42, 43, 54]. The latest FR models utilize the softmax-variant loss for training to reduce intra-class variance and increase inter-class separability in the embedding space. This necessitates a very large dataset with numerous distinct individuals, signifi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.908, + 0.548 + ], + "angle": 0, + "content": "cant variations within each individual for intra-class variance, and precise labels of subject identities (IDs). However, datasets are typically collected through web-crawling and then refined using automatic techniques that employ FR logits [11, 60]. Although this approach is successful in eliminating mislabeled data, it still struggles with persistent issues of small intra-class variance. Moreover, in the face recognition field, unlike other ordinary image classifications, there are practical issues such as portrait rights, making it even more difficult to collect training data. For example, data sets such as those referenced in [16, 56, 60] consist of images of celebrities collected from the internet without consent. Furthermore, the datasets mentioned in [27, 38] include facial images of the general population, including children. The privacy sensitivity of such data poses significant challenges for face recognition research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Synthetic datasets have been used to address the limitations caused by the scarcity of real datasets [8, 25, 49] and biases present in the available real datasets [20, 50]. In the realm of face recognition, artificial faces show promise in addressing the aforementioned issues related to real face datasets. Generated faces are at low risk for label noise due to conditional generation. Bias problems, such as long-tailed distributions, which lead to class imbalances, can be mitigated by data augmentation. Importantly, there are no privacy concerns with virtual identities facial images if the method is effective. Therefore, when creating artificial datasets, it is essential to consider: 1) ensuring that the generated data mirror the real data distribution, 2) the ability to generate new subjects separated from real data, and 3) maintaining ID consistency for each subject." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Previous attempts to create artificial face datasets have addressed some of the three aspects individually, but to the best of our knowledge, none have simultaneously taken into account all three aspects[2, 5, 29, 40]. SynFace [40] introduces a high-quality synthetic face image dataset, closely reminiscent of real face images, using DiscoFaceGAN [13]. However, DiscoFaceGAN is only capable of generating a limited number of unique subjects, fewer than 500 [29]. On" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.363, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10043" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.093, + 0.905, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.233, + 0.305, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.233, + 0.504, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.233, + 0.702, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.233, + 0.901, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 1. T-distributed Stochastic Neighbor Embedding (T-SNE) [51] plot of embeddings from real and synthesized images. The filled and lined stars represent the real and virtual prototypes, while filled and lined circles indicate the embeddings of real and synthesized images, respectively. The bottom of the figure shows the face images included in the cluster, and the dotted outlined images represent the face images generated using our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.485, + 0.614 + ], + "angle": 0, + "content": "the other hand, DigiFace [2] utilizes 3D parametric modeling to create synthetic face images of various subjects. However, it faces difficulties in accurately replicating the quality and style distribution observed in real face images. DCFace [29] proposes a diffusion-based method to generate data that maintains the style of real datasets and ensures label consistency. However, DCFace primarily focuses on creating synthetic data rather than enhancing the capabilities of data augmentation. IDiffFace [5] utilizes identity-conditioned latent diffusion to synthesize facial images from FR feature representations. However, it fails to guarantee the uniqueness of the synthesized identities. CemiFace [48] and HSFace [55] demonstrate high performance on the various FR benchmarks. However, we found identity leakage of the training dataset from both CemiFace and HSFace, which is critical from a privacy standpoint." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The virtual data generated by our method possesses all three essential properties of virtual data mentioned above. The main concept of our paper is to incorporate virtual prototypes into the FR model. Virtual prototypes are trained simultaneously with real prototypes so that they pre-assign on the same feature space. The diffusion model is trained to preserve the identity of the individuals when generating face images based on the FR embedding. As virtual prototypes are allocated within the same embedding space as real prototypes, the produced face images are ensured to reflect the true data distribution. In addition, since prototypes are orthogonal to others, virtual subjects are guaranteed to be distinguished from the original data. Fig. 1 shows a toy example that visualizes the embeddings of real and virtual subjects to demonstrate the effectiveness of the suggested method. The virtual embedding can be distinct from the real individual clusters, whereas the images generated from virtual prototypes form unique clusters. The visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.538 + ], + "angle": 0, + "content": "demonstrates that our framework generates distinctive virtual human faces with high consistency in ID. The proposed method effectively addresses privacy concerns by creating datasets of non-existent individuals and achieving state-of-the-art performance compared to models trained with previous virtual face generation methods. Additionally, the FR model, which was trained using a combination of real and synthetic images together, achieves better performance than the model trained using only real images. This shows that the proposed method is also superior from the perspective of a data augmentation method." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.539, + 0.852, + 0.552 + ], + "angle": 0, + "content": "Our contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.629 + ], + "angle": 0, + "content": "- Introducing VIGFace, a method designed to generate virtual identities with realistic appearance, guided by three main criteria: generating novel subjects that are not found in the real-world, consistently preserving the identity, but ensuring diverse characteristics for each individual." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.905, + 0.674 + ], + "angle": 0, + "content": "- Showing that the synthetic data generated by VIGFace can leverage intra-class diversity and inter-class variance, achieving SOTA performance in face recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.706 + ], + "angle": 0, + "content": "- Releasing the virtual-only face dataset that can fully substitute the real dataset, helping alleviate privacy concerns." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.719, + 0.652, + 0.733 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.742, + 0.759 + ], + "angle": 0, + "content": "2.1. Face Recognition Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Current state-of-the-arts (SOTA) FR methods, such as [3, 9, 12, 24, 28, 32, 36, 53], are designed based on softmax loss, but particularly utilize the angular/cosine distance instead of the Euclidean distance. The goal of these methods is to maximize the similarity between the embeddings and the ground-truth (GT) prototype, while minimizing the similarity between the embedding and prototypes of other classes. As a result, all embeddings in the same class, including prototypes, converge while maintaining their maxi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10044" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.098, + 0.157, + 0.107 + ], + "angle": 0, + "content": "Stage 1" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.11, + 0.498, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.097, + 0.55, + 0.106 + ], + "angle": 0, + "content": "Stage 2" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.119, + 0.897, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.301, + 0.906, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Pipeline for the proposed method. Conventional FR training includes prototypes for only real individuals, indicated as \\( W_{R} \\). We add \\( k \\) prototypes for virtual IDs, denoted as \\( W_{V} \\). The virtual embedding \\( f_{\\mathrm{FR}}'(x_j) \\) corresponding to the virtual person ID: \\( j \\) is generated to follow distribution of the real embeddings. To synthesize the facial image from virtual prototypes, we adopt the DiT architecture [39], following the design approach of the Vision Transformer (ViT) [14]. Additionally, we adjust the DiT model to utilize 5-point landmark images to handle pose variations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.484, + 0.522 + ], + "angle": 0, + "content": "mum distance from other classes. Therefore, if the embedding dimension is large enough, the clusters of each class will be nearly orthogonal [6, 9]. Naturally, a larger dataset that contains a greater variety of images allows for the training of better-performing FR models. Therefore, over time, larger and more diverse datasets [1, 16, 22, 60] have been collected and published in the academic field. However, issues such as portrait rights and the high cost of building large-scale data remain unresolved." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.383, + 0.55 + ], + "angle": 0, + "content": "2.2. Synthetic Face Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.83 + ], + "angle": 0, + "content": "SynFace [40] proposed a method that uses DiscoFaceGAN[13] to synthesize virtual faces. DigiFace-1M [2] proposes a 3D model-based face rendering method to generate virtual face data. DCFace [29] proposes a diffusion-based face generator combining subject appearance (ID) and external factor (style) conditions. GanDiffFace [35] uses GANs to generate identity features. Vec2Face [55] introduces a masked autoencoder to control the identity of face images and their attributes. CemiFace [48] produces facial samples with various levels of similarity to the center of the subject. IDiffFace [5] uses latent diffusion conditioned on FR identity to generate synthetic images. Although these methods succeed in generating separable identities, some still require additional networks or thresholding to sample synthetic identities. Furthermore, unlike [5], we propose a novel method for the pre-assignment of FR representations to maintain latent space alignment and introduce pose control of facial images using facial landmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.19, + 0.86 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our framework comprises two stages. First, we train the FR model using the real face dataset and design the feature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.463 + ], + "angle": 0, + "content": "space for both real and virtual IDs. Second, synthetic images are generated using the diffusion model based on the feature space of the pre-trained FR model that was trained in the first stage. This section provides a detailed explanation of each stage of the proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.767, + 0.487 + ], + "angle": 0, + "content": "3.1. Stage 1: FR Model Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.643 + ], + "angle": 0, + "content": "The proposed framework begins with training the FR model using real face images. This stage serves two purposes: 1) Training the FR model to achieve prototype features from face images, which are necessary for training the diffusion model in the second stage, and 2) Simultaneously assigning the positions of both the real ID and the virtual ID on the feature space. We choose ArcFace [9] to train the FR model in this stage. The ArcFace loss used to train the FR backbone \\( f_{\\mathrm{FR}} \\) and the prototype \\( W = [w_{1}, w_{2}, \\dots, w_{n}] \\) can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.65, + 0.907, + 0.69 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {a r c}} = - \\log \\frac {e ^ {s \\cos \\left(\\theta_ {y _ {i}} + m\\right)}}{e ^ {s \\cos \\left(\\theta_ {y _ {i}} + m\\right)} + \\sum_ {j = 1 , j \\neq y _ {i}} ^ {n} e ^ {s \\cos \\theta_ {j}}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.696, + 0.905, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\cos \\theta_ {j} = \\frac {w _ {j} ^ {T} f _ {\\mathrm {F R}} \\left(x _ {i}\\right)}{\\| w _ {j} \\| \\| f _ {\\mathrm {F R}} \\left(x _ {i}\\right) \\|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(n\\) denotes the total number of real IDs, while \\(m\\) and \\(s\\) represent the margin and scale hyperparameters, respectively. With conventional FR training methods, only prototypes for real individuals, which can be denoted as \\(W_{R} = [w_{r}^{1},w_{r}^{2},\\dots,w_{r}^{n}]\\), are necessary for training. In contrast, we include additional \\(k\\) prototypes for virtual IDs, denoted as \\(W_{V} = [w_{v}^{1},w_{v}^{2},\\dots,w_{v}^{k}]\\), which are used to generate facial images of nonexistent individuals in the real world. As a result, the prototype is defined as a linear transformation matrix \\(W\\in \\mathbb{R}^{(n + k)\\times D}\\), where \\(D\\) refers to the dimension of the embedding features." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10045" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.283, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.24, + 0.281, + 0.265 + ], + "angle": 0, + "content": "(a) Prototypes that are trained without virtual embeddings." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.088, + 0.485, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.24, + 0.482, + 0.266 + ], + "angle": 0, + "content": "(b) Prototypes that are trained with both real and virtual embeddings." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.277, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Figure 3. Changes in the similarity matrix of the prototypes from our method. Similarity values were min-max normalized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.468 + ], + "angle": 0, + "content": "However, due to the absence of a face image for virtual IDs, their prototypes cannot be updated to maximize the distance from each other with \\( L_{arc} \\). Consequently, all virtual prototypes converge to a single point in the feature space when trained with the original \\( L_{arc} \\) as shown in Fig. 3a. To address this problem, we propose to use virtual feature embedding \\( f_{\\mathrm{FR}}'(x_j) \\) to update the virtual prototypes \\( w_v^j \\). The virtual embedding \\( f_{\\mathrm{FR}}'(x_j) \\) corresponding to the virtual person ID: \\( j \\) was generated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.48, + 0.483, + 0.499 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {F R}} ^ {\\prime} (x _ {j}) = w _ {v} ^ {j} + \\mathcal {N} (0, 1) \\cdot \\sigma , \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.51, + 0.483, + 0.552 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {2} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\left(f _ {\\mathrm {F R}} \\left(x _ {i}\\right) - w _ {r} ^ {i}\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.738 + ], + "angle": 0, + "content": "where \\(b\\) refers the mini-batch size. As can be seen from the equations, the virtual embedding \\(f_{\\mathrm{FR}}'(x_j)\\) follows a distribution in which the standard deviation matches that of the real embeddings. The aggregate loss \\(L_{\\mathrm{arc}}\\) is calculated simultaneously using both the real embedding \\(f_{\\mathrm{FR}}(x)\\) and the virtual embedding \\(f_{\\mathrm{FR}}'(x)\\), allowing the virtual prototype \\(w_v^j\\) to maintain minimal similarities with other prototypes, as illustrated in Fig. 3. Since batch configuration affects the calculated standard deviation, we utilize the exponential moving average (EMA) to reduce this influence. The corrected standard deviation \\(\\sigma\\) for the current iteration \\(t\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.752, + 0.483, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\sigma_ {(t)} \\cdot \\alpha + \\sigma_ {(t - 1)} \\cdot (1 - \\alpha). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.856 + ], + "angle": 0, + "content": "The hyperparameter \\(\\alpha\\) is set to 0.9. The number of virtual embedding, \\(b_{v}\\) is determined based on the mini-batch size \\(b_{r}\\), the number of virtual ID prototypes \\(k\\), and the number of real ID prototypes \\(n\\). In our study, we set \\(b_{v} = (k\\times b_{r}) / n\\) so that both real and virtual prototypes can be updated evenly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The overall pipeline of this stage is illustrated in Fig. 2. Note that, as can be seen in the figure, gradients for virtual embeddings have no effect on the backbone. We confirm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "that the similarity distribution between the virtual prototypes closely resembles the similarity distribution between the real prototypes, as shown in Fig. 3b." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.166, + 0.907, + 0.183 + ], + "angle": 0, + "content": "3.2. Stage 2: Face Generation with Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.908, + 0.597 + ], + "angle": 0, + "content": "The next step is to synthesize face images using the diffusion model. To obtain the training dataset for the diffusion model, we utilize the FR model, which involves collecting pairs of images \\( x_0 \\) and their corresponding prototypes \\( w_r \\). The input to our diffusion model includes the timestep \\( t \\), the FR prototype vector \\( w_r \\), the five facial landmark image \\( y \\), and the noisy image \\( x_t \\). In line with the approach proposed in the previous method [31], our model predicts velocity \\( v_t \\) rather than noise \\( \\epsilon \\) injected into \\( x_t \\). We adopt the DiT architecture [39] using the design approach of the visual transformer (ViT) [14]. We modify the DiT model by incorporating the five facial landmarks (including the left eye, the right eye, the nose, the left mouth corner, and the right mouth corner) [57] image. The five facial landmark images are acquired using RetinaFace [10], and employed as conditions to account for pose variations. Fig. 4 shows the synthesized face images conditioned by five facial landmarks. The first column lists input landmark conditions in different pose environments. As illustrated in the figure, the proposed model is capable of producing a wide range of styles (like hair, glasses, and lighting) while also delivering various pose variations, all with a high degree of consistency. In order to achieve both the generation of facial images and the synchronization of their feature space on the FR model, our diffusion model incorporates a constraint which aims to minimize the feature distance between the original image and the input prototypes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.609, + 0.907, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\epsilon , t} \\| f _ {\\mathrm {F R}} (\\hat {x} _ {\\theta} (x _ {t}, t, w _ {r}, y))) - w _ {r} \\| _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.689 + ], + "angle": 0, + "content": "We adopt classifier-free guidance [19] by randomly assigning zero values to condition embeddings, \\( w_{r} \\), \\( 10\\% \\) of time. Sampling is performed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.706, + 0.905, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {x} _ {\\theta} \\left(x _ {t}, t, W, y\\right) = g \\cdot x _ {\\theta} \\left(x _ {t}, t, W, y\\right) \\\\ + (1 - g) \\cdot x _ {\\theta} (x _ {t}, t, y), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.754, + 0.907, + 0.8 + ], + "angle": 0, + "content": "where \\( x_{\\theta}(x_t, t, w_r, y) \\) and \\( x_{\\theta}(x_t, t, y) \\) are conditional and unconditional \\( x_0 \\)-prediction, respectively and \\( g \\) is guidance weight." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.646, + 0.832 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our implementation details for the full framework are described in Appendix A. We analyze our framework from two perspectives. Firstly, by training the FR model exclusively with facial images generated for virtual IDs, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10046" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.408, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.09, + 0.657, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.092, + 0.906, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 4. Virtual face generated in VIGFace (B). Each row lists facial images in different pose environments created using five facial landmarks. Our method can generate various conditions of face images, such as illumination, occlusion by accessories, and facial expressions, while controlling the pose variations of the face images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.484, + 0.473 + ], + "angle": 0, + "content": "demonstrated the capability of our model to serve as a viable alternative to real face datasets, addressing concerns such as label noise or privacy. Secondly, we train the FR model using both real images and generated face images simultaneously, showcasing the potential of our model as a data augmentation framework for face recognition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.34, + 0.5 + ], + "angle": 0, + "content": "4.1. Virtual Identity Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Since diffusion models focus on optimizing the visual aspect with input conditions, they tend to produce images that are high in consistency but lack of variance. One simple way to achieve a balance between consistency and variance is by adjusting the classifier-free guidance weight scale. Higher guidance enforces stronger conditioning of input labels. In other words, it results in more constant images but slightly similar samples. We observe that the proposed model performs best as the scale becomes \\( w = 4.0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.84 + ], + "angle": 0, + "content": "We construct a toy example visualizing the feature space consists of three real persons and five virtual persons to demonstrate the efficacy of the synthesized images in Fig. 1. As seen in the figure on the left, the virtual embedding optimization obviously provides separable virtual prototypes from real individual clusters. The figure on the right illustrates that our synthesized images enforce high variance to the dataset. Synthesized images of real individuals filling the gap in the cluster provide intra-class variance to the subjects. Additionally, the cluster of generated images from virtual prototypes, while staying separate from other subjects, demonstrates the effectiveness of our framework for generating unique virtual human faces in high consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fig. 5 shows the qualitative comparison of conventional synthetic datasets and the proposed VIGFace. As shown in the figure, the facial images generated using the VIGFace model show remarkable uniformity in generating consistent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.489 + ], + "angle": 0, + "content": "virtual individuals, while also incorporating variations such as hair styles, accessories, makeup, and expressions. It is important to emphasize that the key characteristic of a training dataset to achieve a high-performance FR model is not the high quality of the images themselves, but rather the high consistency and variance of the generated facial images belonging to the same person." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.61 + ], + "angle": 0, + "content": "Property of generated Dataset To compare the properties of the synthetic face datasets, we measured the 1) class consistency, 2) class separability, and 3) intra-class diversity of generated images. The class consistency reflects the uniformity of the samples in the same label condition. Consequently, the consistency of class \\( k \\) (\\( C_k \\)) was measured as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.619, + 0.906, + 0.662 + ], + "angle": 0, + "content": "\\[\nC _ {k} = \\frac {1}{N ^ {2}} \\sum_ {j = 1} ^ {N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(x _ {i}\\right) \\cdot f \\left(x _ {j}\\right)}{\\| f \\left(x _ {i}\\right) \\| \\| f \\left(x _ {j}\\right) \\|}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.716 + ], + "angle": 0, + "content": "where \\(N\\) is the number of images in a class \\(k\\). Higher class consistency means that the samples are more uniform under the same label." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.807 + ], + "angle": 0, + "content": "We also measured the class separability to assess the integrity of the dataset; in other words, to ensure that all subjects in the dataset are unique. The class separability for a class \\( k \\) (\\( S_k \\)) is measured as the average distance between the center of class \\( k \\) and the centers of the negative classes, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.816, + 0.905, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} _ {k} = \\frac {1}{K - 1} \\sum_ {i = 1, i \\neq k} ^ {K} 1 - \\frac {\\overline {{f _ {k}}} \\cdot \\overline {{f _ {i}}}}{\\| \\overline {{f _ {k}}} \\| \\| \\overline {{f _ {i}}} \\|}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(K\\) is the number of classes. \\(\\overline{f_k}\\) represents the class center obtained by averaging the embedding vectors of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10047" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.095, + 0.478, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.476 + ], + "angle": 0, + "content": "Figure 5. Comparison of randomly selected virtual ID images generated by conventional methods [2, 5, 29, 35, 40, 48, 55] and by our method, all trained on CASIA-WebFace. For each synthetic dataset, we present two subjects in a single row." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.536 + ], + "angle": 0, + "content": "images that belong to class \\( k \\). As shown in the formula, a high \\( S_{k} \\) indicates that the images are distinct from the negative classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.686 + ], + "angle": 0, + "content": "The intra-class diversity measures how various the conditions of samples are under the same label. In particular, we focus on challenging scenarios that directly impact FR performance, such as pose variations, occlusions, or lighting conditions, rather than image style. Higher diversity indicates that the dataset covers a broad spectrum of cases, from easy to hard. Motivated by this observation, we calculated the intra-class diversity of a class \\( k(\\mathrm{D}_k) \\) using the variance of Face Image Quality Assessment (FIQA) scores as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.696, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathrm {F I Q A} _ {k}}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathrm {F I Q A} \\left(x _ {i}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.748, + 0.483, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D} _ {k} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\operatorname {F I Q A} \\left(x _ {i}\\right) - \\overline {{\\operatorname {F I Q A} _ {k}}}\\right) ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathrm{FIQA}(x_i)\\) indicates the normalized CR-FIQA [4] score of the generated image \\(x_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "The scores \\(C_k\\) and \\(S_k\\) are derived from a pre-trained ArcFace model that was trained using the Glint-360K dataset. Fig. 6a shows the average values of the properties for all classes, normalized by the average values achieved with CASIA-WebFace. VIGFace achieves remarkable scores in" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.093, + 0.728, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.096, + 0.903, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.268, + 0.905, + 0.307 + ], + "angle": 0, + "content": "(a) The normalized properties of various (b) Identity leakage test on the synthetic datasets. The sizes of the cir- synthetic face dataset and its cor-cles indicate the intra-class diversity. responding training dataset." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.346 + ], + "angle": 0, + "content": "Figure 6. Properties of VIGFace compared with those of synthesized datasets generated by previous methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.544 + ], + "angle": 0, + "content": "aspects of class consistency and class separability compared to other methods. SynFace exhibits the lowest separability due to its mix-up generation method. This fact indicates that SynFace can generate only a limited number of subjects as mentioned in the introduction. DigiFace achieves strong consistency, due to its 3D rendering methods, but its image style may differ significantly from real datasets, leading to lower accuracy in real-world applications. GANDiffFace demonstrates inadequate FIQA diversity, indicating limited variations in factors such as pose, lighting, and occlusion. DCFace, which uses a diffusion model, shows separability similar to that of VIGFace but with lower consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.909, + 0.743 + ], + "angle": 0, + "content": "In comparisons of SOTA models, HSFace, despite using a more refined dataset, comparatively lacks consistency. CemiFace, trained on the same dataset as VIGFace, shows low consistency and separability. Both methods also exhibit inferior qualitative uniformity within the same class, as illustrated in Fig. 5. Our approach uses prototypes instead of FR embedding vectors to disentangle identity features from other characteristics, allowing VIGFace to generate subjects with greater consistency by focusing on identity features. Moreover, it enhances separability by leveraging the pre-computation of feature orthogonality. We included further analysis on the properties of the datasets and provided examples in Appendix H." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Identity leakage from real human To claim complete privacy-free, it is necessary to prove that no real IDs or training images are included in the synthesized dataset. In this reason, we demonstrate that the generated face images represent non-existent humans by querying the most similar face in the CASIA-WebFace dataset. For comparison, we also present the similarity values of the nearest negative class in CASIA-WebFace itself. In Fig. 6b, we count the number of synthetic face images whose similarity with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10048" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.368 + ], + "angle": 0, + "content": "
MethodTraining Dataset# of Images (classes × variations)Verification Benchmarks
LFWCFP-FPCPLFWAgeDBCALFWAvg.
CASIA-Webface (Real)-0.49M (≈10.5K×47)99.4096.6390.2394.6893.7094.93
SynFaceFFHQ0.5M (10K × 50)91.9375.0370.4361.6374.7374.75
DigiFace3D modeling0.5M (10K × 50)95.4087.4078.8776.9778.6283.45
DCFaceFFHQ+CASIA0.5M (10K × 50)98.5585.3382.6289.7091.6089.56
IDiffFaceFFHQ0.5M (10K × 50)98.0085.4780.4586.4390.6588.20
GANDiffFaceFFHQ0.5M (10K × 50)90.7773.2772.3266.3574.6875.48
ID3CASIA0.5M (10K × 50)97.6886.8482.7791.0090.7389.80
CemiFaceCASIA+WebFace4M0.5M (10K × 50)99.0391.0687.6291.3392.4292.30
Arc2FaceWebFace42M0.5M (10K × 50)98.8191.8785.1690.1892.6391.73
HyperFaceWebFace42M0.5M (10K × 50)98.5088.8384.2386.5389.4089.50
HSFace10KWebFace4M0.5M (10K × 50)98.8788.9785.4793.1293.5792.00
VIGFace(S), OursCASIA0.5M (10K × 50)99.0295.0987.7290.9590.0092.56
DCFaceFFHQ+CASIA1.2M (20K × 50 + 40K × 5)98.5888.6185.0790.9792.8291.21
CemiFaceCASIA+WebFace4M1.2M (20K × 50 + 40K × 5)99.2292.8488.8692.1393.0393.22
Arc2FaceWebFace42M1.2M (20K × 50 + 40K × 5)98.9294.5886.4592.4593.3393.14
HSFace20KWebFace4M1.0M (20K × 50)98.8789.8786.1393.8593.6592.47
HSFace300KWebFace4M15M (300K × 50)99.3091.5487.7094.4594.5893.52
VIGFace(B), OursCASIA1.2M (60K × 20)99.4897.0790.1593.6292.8894.64
VIGFace(L), OursCASIA6.0M (120K × 50)99.3397.3191.1293.8292.9594.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.377, + 0.908, + 0.42 + ], + "angle": 0, + "content": "Table 1. FR benchmark results trained with various virtual face datasets. All results except for CASIA-WebFace and VIGFace are obtained from the original paper. Our method is specified by the number of identities as small (S), base (B) and large (L). FR backbone is IR-SE50 + AdaFace [28]. Bold and underline indicates the best and the second best, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.436, + 0.486, + 0.664 + ], + "angle": 0, + "content": "the average feature vector of each class in the training data exceeds a given threshold. As shown in the figure, face images in VIGFace have low similarity to the subjects in CASIA-WebFace. This supports the claim that each synthesized image depicts a non-existent human face and there is no identity leakage from the training data. In contrast, we observed that some of the conventional SOTA methods exhibit identity leakage. As CemiFace [48] samples identity embeddings from the WebFace4M dataset, it generates individuals that exhibit resemblance to those in WebFace4M. Vec2Face [55] fails in creating novel identities, so that it produces almost same persons in WebFace4M. Further analysis and examples are reported in Appendix I. In this paper, we avoid thresholding or selective sampling, which could harm fair comparisons with previous methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.674, + 0.214, + 0.689 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.485, + 0.743 + ], + "angle": 0, + "content": "In this section, we train the FR model with VIGFace to compare it to conventional methods. Detailed hyperparameters for training the FR network can be found in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "VIGFace as Virtual Dataset We compare the performance of the FR model trained with generated facial images for virtual IDs using VIGFace with conventional methods. For the experiment, we set the number of virtual IDs to \\(10K\\), \\(60K\\) and \\(120K\\). To ensure a fair comparison, we established image counts of 0.5M, 1.2M, and 6.0M, which reflects the scale of the real CASIA-WebFace dataset and previous synthetic approaches, respectively. In Tab. 1, we present the 1:1 verification accuracy (\\%) on five bench" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.436, + 0.692, + 0.451 + ], + "angle": 0, + "content": "marks [23, 37, 44, 58, 59]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.452, + 0.908, + 0.816 + ], + "angle": 0, + "content": "As shown in the table, VIGFace outperforms conventional methods in the average accuracy of five verification benchmarks. In particular, VIGFace achieves equal or even better performance on CFP-FP and CPLFW compared to the model trained with the real dataset, CASIA-WebFace. This indicates that our method benefits from its pose-variable generation capabilities. CemiFace shows notable performance on the AgeDB and CALFW benchmarks, but they sampled identity embeddings from the WebFace4M [60] dataset. Consequently, the individuals included in the CemiFace dataset may still have privacy concerns. HSFace10K also delivers outstanding results among conventional methods. However, it is important to note that HSFace benefits from training on the extensive WebFace4M dataset, which boasts twice as many images and quadruples the number of identities when compared to CASIA-WebFace. VIGFace(L) outperforms HSFace300K even with a smaller number of images. As a result, when the FR model was trained using VIGFace, it achieved performance comparable to a model trained using a real dataset in terms of average accuracy without any external identity sampling. This suggests that our virtual dataset can serve as a full replacement for the CASIA-WebFace dataset to train the FR model, while avoiding privacy concerns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "VIGFace as Data Augmentation To demonstrate the efficacy of VIGFace as an augmentation framework, we evaluate the accuracy of the FR model trained on real and synthetic images. We utilized the dataset that was uploaded" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10049" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.273 + ], + "angle": 0, + "content": "
ConditionVerification Benchmarks
MethodReal ImageSynthetic ImageLFWCFP-FPCPLFWAgeDBCALFWAvg.Δ
Real IDVirtual ID
CASIA-WebFace✓99.4096.6390.2394.6893.7094.93-
DigiFace✓✓99.3797.5190.9294.9593.7795.30+0.37
DCFace (1.2M)✓✓99.4396.9790.3395.2094.3895.26+0.33
iDiffFace✓✓99.5897.0490.4094.7894.0095.16+0.23
GANDiffFace✓✓99.5296.6190.3093.9893.5794.80-0.13
HSFace20K✓✓99.6097.4191.0795.4894.4095.59+0.66
VIGFace(B)✓✓99.4597.2390.7895.2593.2295.19+0.26
✓✓99.5598.0391.8095.7394.1295.85+0.92
✓✓✓99.7098.1091.5795.8594.3895.92+0.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.281, + 0.907, + 0.311 + ], + "angle": 0, + "content": "Table 2. Augmented FR performance results for various conditions of synthetic dataset. The accuracies (%) for LFW, CFP-FP, CPLFW, AgeDB-30, CALFW, and the average benchmark accuracies are reported. FR backbone is IR-SE50 + AdaFace [28]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.734 + ], + "angle": 0, + "content": "to the official repository. Tab. 2 shows the performance change of the trained FR model using various combinations of datasets. DigiFace, which utilizes unique 3D modeling technology, has demonstrated improved benchmark performance and offers advantages from a data-augmentation perspective. However, DigiFace struggles to perform well on its own, devoid of real data [2]. This is due to their failure to precisely replicate the appearance of real-world facial images, making it impractical to present a privacy-free synthetic dataset. HSFace shows a notable improvement among conventional methods. However, it utilized WebFace4M for training, which may make the comparison with others unfair. The boosted accuracy using VIGFace, which outperforms conventional methods, demonstrates that VIGFace can accurately mirror real facial data and achieve synergy by blending them. In particular, augmented images of real IDs improve the results of the CFP-FP and CPLFW benchmarks. This observation suggests that the use of five-point facial landmarks in the conditioning method can create a variety of posed facial images, substantially improving the FR model's ability to understand pose variations. Consequently, VIGFace not only effectively solves the privacy issue, but can also be used with real data as part of the augmentation process. When trained with the VIGFace dataset, the FR backbone achieves better generalization and higher performance, in contrast to conventional methods that achieved only marginal performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Fig. 5 illustrates the results that synthesize variations of real and virtual IDs. Since the training process is significantly biased not just by a lack of subject but also by inner-class variance, increasing the variance within the class is crucial. The results demonstrate that the proposed approach can generate a variety of conditions for the existing real ID without any help from external data. Since VIGFace can generate varied images of real individuals, we performed experiments on each type of augmentation, i.e. synthetic images of real and virtual ID. In particular, we extend the long-tailed real-ID class, which contains fewer than 50 im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.909, + 0.465 + ], + "angle": 0, + "content": "ages, to 50 images. With this method, additional 0.15M images were added as real ID synthesis. As can be seen in Tab. 2, the augmented dataset for real ID shows improved FR performance on most benchmarks. This indicates that our method can increase the intra-class diversity of dataset, a critical factor in achieving high FR performance. As a result, the FR model that utilizes the entire set of augmented data exhibits the best performance, benefiting from enhanced intra-class diversity and inter-class variance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.642, + 0.494 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.777 + ], + "angle": 0, + "content": "This paper presents an effective method for creating a synthetic face dataset that guarantees unique virtual identities. Synthesized facial data can serve as a solution for various challenges faced by traditional real datasets, such as high expenses, inaccuracies and biases in labeling, and concerns regarding privacy. To this end, we propose the Virtual Identity Generation framework and demonstrate that it can generate not only realistic but also diverse facial images of virtual individuals, significantly narrowing the performance gap with FR models trained on real data. Furthermore, the model exhibits superior performance when trained on a combination of VIGFace and existing real data compared to models trained solely on real data. This confirms that our proposed method has potential as an effective augmentation technique. We will publicly distribute the virtual face dataset created by VIGFace and believe that these new virtual data will contribute to resolving the privacy issues inherent in face recognition training dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment This work was partially supported by the Institutional Programs of the Korea Institute of Science and Technology (KIST) (Project No. 2E33612, \\(50\\%\\) ) and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2023-00227592, Development of Robust 3D Object Identification Technology for Viewpoint Changes, \\(50\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10050" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.214 + ], + "angle": 0, + "content": "[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[2] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 1, 2, 3, 6, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.482, + 0.366 + ], + "angle": 0, + "content": "[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1578-1587, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5836-5845, 2023. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 1, 2, 3, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[6] Johann S Brauchart, Alexander B Reznikov, Edward B Saff, Ian H Sloan, Yu Guang Wang, and Robert S Womersley. Random point sets on the sphere—hole radii, covering, and separation. Experimental Mathematics, 27(1):62-81, 2018. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[8] Hyunwoo Cho, Haesol Park, Ig-Jae Kim, and Junghyun Cho. Data augmentation of backscatter x-ray images for deep learning-based automatic cargo inspection. Sensors, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[10] Jiankang Deng, Jia Guo, Yuxiang Zhou, Jinke Yu, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[11] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 741-757. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.133 + ], + "angle": 0, + "content": "face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2010.3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[15] Papantoniou et al. Arc2face: A foundation model for id-consistent human faces. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[16] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.524 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[20] Je Hyeong Hong, Hanjo Kim, Minsoo Kim, Gi Pyo Nam, Junghyun Cho, Hyeong-Seok Ko, and Ig-Jae Kim. A 3d model-based approach for fitting masks to faces in the wild. In 2021 IEEE International Conference on Image Processing (ICIP), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[22] Gary Huang, Marwan Mattar, Honglak Lee, and Erik Learned-Miller. Learning to align from scratch. Advances in neural information processing systems, 25, 2012. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Hochul Hwang, Cheongjae Jang, Geonwoo Park, Junghyun Cho, and Ig-Jae Kim. Eldersim: A synthetic data generation" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10051" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "platform for human action recognition in eldercare applications. IEEE Access, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[27] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[28] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[29] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. Dface: Synthetic face generation with dual condition diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12715-12725, 2023. 1, 2, 3, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[30] D Kinga, Jimmy Ba Adam, et al. A method for stochastic optimization. In International conference on learning representations (ICLR), page 6. San Diego, California.; 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[35] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 3, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[36] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14225-14234, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[38] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7044-7053, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[40] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 1, 3, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[42] Swami Sankaranarayanan, Azadeh Alavi, Carlos D Castillo, and Rama Chellappa. Triplet probabilistic embedding for face verification and clustering. In 2016 IEEE 8th international conference on biometrics theory, applications and systems (BTAS), pages 1-8. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[44] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[47] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[48] Zhonglin Sun, Siyang Song, Ioannis Patras, and Georgios Tzimiropoulos. Cemifice: Center-based semi-hard synthetic face generation for face recognition. Advances in Neural Information Processing Systems, 37:35612-35638, 2025. 2, 3, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10052" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[50] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[52] Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, and Chen Change Loy. The devil of face recognition is in the noise. In Proceedings of the European Conference on Computer Vision (ECCV), pages 765-780, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[53] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[54] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 499-515. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[55] Haiyu Wu, Jaskirat Singh, Sicong Tian, Liang Zheng, and Kevin W Bowyer. Vec2face: Scaling face dataset generation with loosely constrained vectors. arXiv preprint arXiv:2409.02979, 2024. 2, 3, 6, 7, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[56] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[58] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[59] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10492-10502, 2021. 1, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10053" + } + ] +] \ No newline at end of file diff --git a/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_origin.pdf b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7a4735fd2ccdf811712322373b01660455cca9ba --- /dev/null +++ b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/570e97bc-508f-4feb-946d-7b35db443845_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84a7893f17382b8ebece2a92f6385bc6b07dc107f4c11f26c51ffa9506553092 +size 3794715 diff --git a/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/full.md b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ef3a017f035e40267b08702933d9c7f3c4a82c07 --- /dev/null +++ b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/full.md @@ -0,0 +1,307 @@ +# VIGFace: Virtual Identity Generation for Privacy-Free Face Recognition dataset + +Minsoo Kim $^{1,2}$ Min-Cheol Sagong $^{1}$ Gi Pyo Nam $^{1,2}$ Junghyun Cho $^{1,2,3}$ Ig-Jae Kim $^{1,2}$ + +$^{1}$ Korea Institute of Science and Technology(KIST) + +$^{2}$ Korea National University of Science and Technology + +$^{3}$ Yonsei-KIST Convergence Research Institute + +{kim1102,mcsagong,gpnam,jhcho,drjay}@kist.re.kr + +# Abstract + +Deep learning-based face recognition continues to face challenges due to its reliance on huge datasets obtained from web crawling, which can be costly to gather and raise significant real-world privacy concerns. To address this issue, we propose VIGFace, a novel framework capable of generating synthetic facial images. Our idea originates from pre-assigning virtual identities in the feature space. Initially, we train the face recognition model using a real face dataset and create a feature space for both real and virtual identities, where virtual prototypes are orthogonal to other prototypes. Subsequently, we train the diffusion model based on the established feature space, enabling it to generate authentic human face images from real prototypes and synthesize virtual face images from virtual prototypes. Our proposed framework provides two significant benefits. Firstly, it shows clear separability between existing individuals and virtual face images, allowing one to create synthetic images with confidence and without concerns about privacy and portrait rights. Secondly, it ensures improved performance through data augmentation by incorporating real existing images. Extensive experiments demonstrate the superiority of our virtual face dataset and framework, outperforming the previous state-of-the-art on various face recognition benchmarks. https://github.com/kim1102/VIGFace + +# 1. Introduction + +Deep learning-based Face Recognition (FR) models have significantly improved their performance due to recent advances in network architectures [17, 18, 21, 26, 45, 47] and enhancements in loss functions [9, 28, 32, 42, 43, 54]. The latest FR models utilize the softmax-variant loss for training to reduce intra-class variance and increase inter-class separability in the embedding space. This necessitates a very large dataset with numerous distinct individuals, signifi + +cant variations within each individual for intra-class variance, and precise labels of subject identities (IDs). However, datasets are typically collected through web-crawling and then refined using automatic techniques that employ FR logits [11, 60]. Although this approach is successful in eliminating mislabeled data, it still struggles with persistent issues of small intra-class variance. Moreover, in the face recognition field, unlike other ordinary image classifications, there are practical issues such as portrait rights, making it even more difficult to collect training data. For example, data sets such as those referenced in [16, 56, 60] consist of images of celebrities collected from the internet without consent. Furthermore, the datasets mentioned in [27, 38] include facial images of the general population, including children. The privacy sensitivity of such data poses significant challenges for face recognition research. + +Synthetic datasets have been used to address the limitations caused by the scarcity of real datasets [8, 25, 49] and biases present in the available real datasets [20, 50]. In the realm of face recognition, artificial faces show promise in addressing the aforementioned issues related to real face datasets. Generated faces are at low risk for label noise due to conditional generation. Bias problems, such as long-tailed distributions, which lead to class imbalances, can be mitigated by data augmentation. Importantly, there are no privacy concerns with virtual identities facial images if the method is effective. Therefore, when creating artificial datasets, it is essential to consider: 1) ensuring that the generated data mirror the real data distribution, 2) the ability to generate new subjects separated from real data, and 3) maintaining ID consistency for each subject. + +Previous attempts to create artificial face datasets have addressed some of the three aspects individually, but to the best of our knowledge, none have simultaneously taken into account all three aspects[2, 5, 29, 40]. SynFace [40] introduces a high-quality synthetic face image dataset, closely reminiscent of real face images, using DiscoFaceGAN [13]. However, DiscoFaceGAN is only capable of generating a limited number of unique subjects, fewer than 500 [29]. On + +![](images/6fc969961fa4e39eacfade113c5cc0a22fb6a00e7a5626d4f3f72c95b52c3a00.jpg) + +![](images/d819eacfc352f6cc1d4054a1391b661b8d4656347a54fde7f2b7cc07959c2683.jpg) +Figure 1. T-distributed Stochastic Neighbor Embedding (T-SNE) [51] plot of embeddings from real and synthesized images. The filled and lined stars represent the real and virtual prototypes, while filled and lined circles indicate the embeddings of real and synthesized images, respectively. The bottom of the figure shows the face images included in the cluster, and the dotted outlined images represent the face images generated using our method. + +![](images/0eb5df0cbc394491258ccb41da6775eb48bccd7eebe237385f3727e11f36ca39.jpg) + +![](images/a5903228c2906cfd2ad202515e58cd723c174f6cd93af2b5d3d05f33ce5f5ef0.jpg) + +![](images/b6b4b0b13e19c97ea681a4a19be14469eceae1dba0d8193022065b58ed38325c.jpg) + +the other hand, DigiFace [2] utilizes 3D parametric modeling to create synthetic face images of various subjects. However, it faces difficulties in accurately replicating the quality and style distribution observed in real face images. DCFace [29] proposes a diffusion-based method to generate data that maintains the style of real datasets and ensures label consistency. However, DCFace primarily focuses on creating synthetic data rather than enhancing the capabilities of data augmentation. IDiffFace [5] utilizes identity-conditioned latent diffusion to synthesize facial images from FR feature representations. However, it fails to guarantee the uniqueness of the synthesized identities. CemiFace [48] and HSFace [55] demonstrate high performance on the various FR benchmarks. However, we found identity leakage of the training dataset from both CemiFace and HSFace, which is critical from a privacy standpoint. + +The virtual data generated by our method possesses all three essential properties of virtual data mentioned above. The main concept of our paper is to incorporate virtual prototypes into the FR model. Virtual prototypes are trained simultaneously with real prototypes so that they pre-assign on the same feature space. The diffusion model is trained to preserve the identity of the individuals when generating face images based on the FR embedding. As virtual prototypes are allocated within the same embedding space as real prototypes, the produced face images are ensured to reflect the true data distribution. In addition, since prototypes are orthogonal to others, virtual subjects are guaranteed to be distinguished from the original data. Fig. 1 shows a toy example that visualizes the embeddings of real and virtual subjects to demonstrate the effectiveness of the suggested method. The virtual embedding can be distinct from the real individual clusters, whereas the images generated from virtual prototypes form unique clusters. The visualization + +demonstrates that our framework generates distinctive virtual human faces with high consistency in ID. The proposed method effectively addresses privacy concerns by creating datasets of non-existent individuals and achieving state-of-the-art performance compared to models trained with previous virtual face generation methods. Additionally, the FR model, which was trained using a combination of real and synthetic images together, achieves better performance than the model trained using only real images. This shows that the proposed method is also superior from the perspective of a data augmentation method. + +Our contribution can be summarized as follows: + +- Introducing VIGFace, a method designed to generate virtual identities with realistic appearance, guided by three main criteria: generating novel subjects that are not found in the real-world, consistently preserving the identity, but ensuring diverse characteristics for each individual. +- Showing that the synthetic data generated by VIGFace can leverage intra-class diversity and inter-class variance, achieving SOTA performance in face recognition. +- Releasing the virtual-only face dataset that can fully substitute the real dataset, helping alleviate privacy concerns. + +# 2. Related work + +# 2.1. Face Recognition Models + +Current state-of-the-arts (SOTA) FR methods, such as [3, 9, 12, 24, 28, 32, 36, 53], are designed based on softmax loss, but particularly utilize the angular/cosine distance instead of the Euclidean distance. The goal of these methods is to maximize the similarity between the embeddings and the ground-truth (GT) prototype, while minimizing the similarity between the embedding and prototypes of other classes. As a result, all embeddings in the same class, including prototypes, converge while maintaining their maxi + +![](images/32cca86e77a26c1dd7fe611425fe8ab10da25f53c42bf38953a159929143ef5e.jpg) +Stage 1 +Stage 2 + +![](images/ef7405955034e53b21402fc9ee04575c8d1e2172a1180975db261f930f7bd042.jpg) +Figure 2. Pipeline for the proposed method. Conventional FR training includes prototypes for only real individuals, indicated as $W_{R}$ . We add $k$ prototypes for virtual IDs, denoted as $W_{V}$ . The virtual embedding $f_{\mathrm{FR}}'(x_j)$ corresponding to the virtual person ID: $j$ is generated to follow distribution of the real embeddings. To synthesize the facial image from virtual prototypes, we adopt the DiT architecture [39], following the design approach of the Vision Transformer (ViT) [14]. Additionally, we adjust the DiT model to utilize 5-point landmark images to handle pose variations. + +mum distance from other classes. Therefore, if the embedding dimension is large enough, the clusters of each class will be nearly orthogonal [6, 9]. Naturally, a larger dataset that contains a greater variety of images allows for the training of better-performing FR models. Therefore, over time, larger and more diverse datasets [1, 16, 22, 60] have been collected and published in the academic field. However, issues such as portrait rights and the high cost of building large-scale data remain unresolved. + +# 2.2. Synthetic Face Image Generation + +SynFace [40] proposed a method that uses DiscoFaceGAN[13] to synthesize virtual faces. DigiFace-1M [2] proposes a 3D model-based face rendering method to generate virtual face data. DCFace [29] proposes a diffusion-based face generator combining subject appearance (ID) and external factor (style) conditions. GanDiffFace [35] uses GANs to generate identity features. Vec2Face [55] introduces a masked autoencoder to control the identity of face images and their attributes. CemiFace [48] produces facial samples with various levels of similarity to the center of the subject. IDiffFace [5] uses latent diffusion conditioned on FR identity to generate synthetic images. Although these methods succeed in generating separable identities, some still require additional networks or thresholding to sample synthetic identities. Furthermore, unlike [5], we propose a novel method for the pre-assignment of FR representations to maintain latent space alignment and introduce pose control of facial images using facial landmarks. + +# 3. Methods + +Our framework comprises two stages. First, we train the FR model using the real face dataset and design the feature + +space for both real and virtual IDs. Second, synthetic images are generated using the diffusion model based on the feature space of the pre-trained FR model that was trained in the first stage. This section provides a detailed explanation of each stage of the proposed framework. + +# 3.1. Stage 1: FR Model Training + +The proposed framework begins with training the FR model using real face images. This stage serves two purposes: 1) Training the FR model to achieve prototype features from face images, which are necessary for training the diffusion model in the second stage, and 2) Simultaneously assigning the positions of both the real ID and the virtual ID on the feature space. We choose ArcFace [9] to train the FR model in this stage. The ArcFace loss used to train the FR backbone $f_{\mathrm{FR}}$ and the prototype $W = [w_{1}, w_{2}, \dots, w_{n}]$ can be described as follows: + +$$ +L _ {\mathrm {a r c}} = - \log \frac {e ^ {s \cos \left(\theta_ {y _ {i}} + m\right)}}{e ^ {s \cos \left(\theta_ {y _ {i}} + m\right)} + \sum_ {j = 1 , j \neq y _ {i}} ^ {n} e ^ {s \cos \theta_ {j}}}, \tag {1} +$$ + +$$ +\cos \theta_ {j} = \frac {w _ {j} ^ {T} f _ {\mathrm {F R}} \left(x _ {i}\right)}{\| w _ {j} \| \| f _ {\mathrm {F R}} \left(x _ {i}\right) \|}, \tag {2} +$$ + +where $n$ denotes the total number of real IDs, while $m$ and $s$ represent the margin and scale hyperparameters, respectively. With conventional FR training methods, only prototypes for real individuals, which can be denoted as $W_{R} = [w_{r}^{1},w_{r}^{2},\dots,w_{r}^{n}]$ , are necessary for training. In contrast, we include additional $k$ prototypes for virtual IDs, denoted as $W_{V} = [w_{v}^{1},w_{v}^{2},\dots,w_{v}^{k}]$ , which are used to generate facial images of nonexistent individuals in the real world. As a result, the prototype is defined as a linear transformation matrix $W\in \mathbb{R}^{(n + k)\times D}$ , where $D$ refers to the dimension of the embedding features. + +![](images/43bd49bd2bc2e5d82b88ffca18bc08f6beecd074260540f272dd54e911c5c1f7.jpg) +(a) Prototypes that are trained without virtual embeddings. +Figure 3. Changes in the similarity matrix of the prototypes from our method. Similarity values were min-max normalized. + +![](images/f0437c452840701ea3dd8d6ffde051f60d69aa442a3b6fa9965447d2ae05dbb9.jpg) +(b) Prototypes that are trained with both real and virtual embeddings. + +However, due to the absence of a face image for virtual IDs, their prototypes cannot be updated to maximize the distance from each other with $L_{arc}$ . Consequently, all virtual prototypes converge to a single point in the feature space when trained with the original $L_{arc}$ as shown in Fig. 3a. To address this problem, we propose to use virtual feature embedding $f_{\mathrm{FR}}'(x_j)$ to update the virtual prototypes $w_v^j$ . The virtual embedding $f_{\mathrm{FR}}'(x_j)$ corresponding to the virtual person ID: $j$ was generated as follows: + +$$ +f _ {\mathrm {F R}} ^ {\prime} (x _ {j}) = w _ {v} ^ {j} + \mathcal {N} (0, 1) \cdot \sigma , \tag {3} +$$ + +$$ +\sigma^ {2} = \frac {1}{b} \sum_ {i = 1} ^ {b} \left(f _ {\mathrm {F R}} \left(x _ {i}\right) - w _ {r} ^ {i}\right) ^ {2}, \tag {4} +$$ + +where $b$ refers the mini-batch size. As can be seen from the equations, the virtual embedding $f_{\mathrm{FR}}'(x_j)$ follows a distribution in which the standard deviation matches that of the real embeddings. The aggregate loss $L_{\mathrm{arc}}$ is calculated simultaneously using both the real embedding $f_{\mathrm{FR}}(x)$ and the virtual embedding $f_{\mathrm{FR}}'(x)$ , allowing the virtual prototype $w_v^j$ to maintain minimal similarities with other prototypes, as illustrated in Fig. 3. Since batch configuration affects the calculated standard deviation, we utilize the exponential moving average (EMA) to reduce this influence. The corrected standard deviation $\sigma$ for the current iteration $t$ is calculated as follows: + +$$ +\sigma = \sigma_ {(t)} \cdot \alpha + \sigma_ {(t - 1)} \cdot (1 - \alpha). \tag {5} +$$ + +The hyperparameter $\alpha$ is set to 0.9. The number of virtual embedding, $b_{v}$ is determined based on the mini-batch size $b_{r}$ , the number of virtual ID prototypes $k$ , and the number of real ID prototypes $n$ . In our study, we set $b_{v} = (k\times b_{r}) / n$ so that both real and virtual prototypes can be updated evenly. + +The overall pipeline of this stage is illustrated in Fig. 2. Note that, as can be seen in the figure, gradients for virtual embeddings have no effect on the backbone. We confirm + +that the similarity distribution between the virtual prototypes closely resembles the similarity distribution between the real prototypes, as shown in Fig. 3b. + +# 3.2. Stage 2: Face Generation with Diffusion Model + +The next step is to synthesize face images using the diffusion model. To obtain the training dataset for the diffusion model, we utilize the FR model, which involves collecting pairs of images $x_0$ and their corresponding prototypes $w_r$ . The input to our diffusion model includes the timestep $t$ , the FR prototype vector $w_r$ , the five facial landmark image $y$ , and the noisy image $x_t$ . In line with the approach proposed in the previous method [31], our model predicts velocity $v_t$ rather than noise $\epsilon$ injected into $x_t$ . We adopt the DiT architecture [39] using the design approach of the visual transformer (ViT) [14]. We modify the DiT model by incorporating the five facial landmarks (including the left eye, the right eye, the nose, the left mouth corner, and the right mouth corner) [57] image. The five facial landmark images are acquired using RetinaFace [10], and employed as conditions to account for pose variations. Fig. 4 shows the synthesized face images conditioned by five facial landmarks. The first column lists input landmark conditions in different pose environments. As illustrated in the figure, the proposed model is capable of producing a wide range of styles (like hair, glasses, and lighting) while also delivering various pose variations, all with a high degree of consistency. In order to achieve both the generation of facial images and the synchronization of their feature space on the FR model, our diffusion model incorporates a constraint which aims to minimize the feature distance between the original image and the input prototypes as follows: + +$$ +\min _ {\theta} \mathbb {E} _ {\epsilon , t} \| f _ {\mathrm {F R}} (\hat {x} _ {\theta} (x _ {t}, t, w _ {r}, y))) - w _ {r} \| _ {2} ^ {2}. \tag {6} +$$ + +We adopt classifier-free guidance [19] by randomly assigning zero values to condition embeddings, $w_{r}$ , $10\%$ of time. Sampling is performed as follows: + +$$ +\begin{array}{l} \tilde {x} _ {\theta} \left(x _ {t}, t, W, y\right) = g \cdot x _ {\theta} \left(x _ {t}, t, W, y\right) \\ + (1 - g) \cdot x _ {\theta} (x _ {t}, t, y), \tag {7} \\ \end{array} +$$ + +where $x_{\theta}(x_t, t, w_r, y)$ and $x_{\theta}(x_t, t, y)$ are conditional and unconditional $x_0$ -prediction, respectively and $g$ is guidance weight. + +# 4. Experiments + +Our implementation details for the full framework are described in Appendix A. We analyze our framework from two perspectives. Firstly, by training the FR model exclusively with facial images generated for virtual IDs, we + +![](images/4d369e7e0eb34aed5072b34e6e75537d7dbf96fc545a00cc7c6aa741f0e5cf5e.jpg) +Figure 4. Virtual face generated in VIGFace (B). Each row lists facial images in different pose environments created using five facial landmarks. Our method can generate various conditions of face images, such as illumination, occlusion by accessories, and facial expressions, while controlling the pose variations of the face images. + +![](images/09cd79ebeface17a8f98d200536d9ba5a88e0f98d308e421cbcadd9ae7374381.jpg) + +![](images/910c7a759608f38d4e511cb8105f360d78e14c165b3968a9e1e032c6fe672044.jpg) + +demonstrated the capability of our model to serve as a viable alternative to real face datasets, addressing concerns such as label noise or privacy. Secondly, we train the FR model using both real images and generated face images simultaneously, showcasing the potential of our model as a data augmentation framework for face recognition. + +# 4.1. Virtual Identity Generation + +Since diffusion models focus on optimizing the visual aspect with input conditions, they tend to produce images that are high in consistency but lack of variance. One simple way to achieve a balance between consistency and variance is by adjusting the classifier-free guidance weight scale. Higher guidance enforces stronger conditioning of input labels. In other words, it results in more constant images but slightly similar samples. We observe that the proposed model performs best as the scale becomes $w = 4.0$ . + +We construct a toy example visualizing the feature space consists of three real persons and five virtual persons to demonstrate the efficacy of the synthesized images in Fig. 1. As seen in the figure on the left, the virtual embedding optimization obviously provides separable virtual prototypes from real individual clusters. The figure on the right illustrates that our synthesized images enforce high variance to the dataset. Synthesized images of real individuals filling the gap in the cluster provide intra-class variance to the subjects. Additionally, the cluster of generated images from virtual prototypes, while staying separate from other subjects, demonstrates the effectiveness of our framework for generating unique virtual human faces in high consistency. + +Fig. 5 shows the qualitative comparison of conventional synthetic datasets and the proposed VIGFace. As shown in the figure, the facial images generated using the VIGFace model show remarkable uniformity in generating consistent + +virtual individuals, while also incorporating variations such as hair styles, accessories, makeup, and expressions. It is important to emphasize that the key characteristic of a training dataset to achieve a high-performance FR model is not the high quality of the images themselves, but rather the high consistency and variance of the generated facial images belonging to the same person. + +Property of generated Dataset To compare the properties of the synthetic face datasets, we measured the 1) class consistency, 2) class separability, and 3) intra-class diversity of generated images. The class consistency reflects the uniformity of the samples in the same label condition. Consequently, the consistency of class $k$ ( $C_k$ ) was measured as follows: + +$$ +C _ {k} = \frac {1}{N ^ {2}} \sum_ {j = 1} ^ {N} \sum_ {i = 1} ^ {N} \frac {f \left(x _ {i}\right) \cdot f \left(x _ {j}\right)}{\| f \left(x _ {i}\right) \| \| f \left(x _ {j}\right) \|}, \tag {8} +$$ + +where $N$ is the number of images in a class $k$ . Higher class consistency means that the samples are more uniform under the same label. + +We also measured the class separability to assess the integrity of the dataset; in other words, to ensure that all subjects in the dataset are unique. The class separability for a class $k$ ( $S_k$ ) is measured as the average distance between the center of class $k$ and the centers of the negative classes, as follows: + +$$ +\mathrm {S} _ {k} = \frac {1}{K - 1} \sum_ {i = 1, i \neq k} ^ {K} 1 - \frac {\overline {{f _ {k}}} \cdot \overline {{f _ {i}}}}{\| \overline {{f _ {k}}} \| \| \overline {{f _ {i}}} \|}, \tag {9} +$$ + +where $K$ is the number of classes. $\overline{f_k}$ represents the class center obtained by averaging the embedding vectors of the + +![](images/33feea5ec5acd2ac0a96374e0954de8ffb24046050a5198d485078224ec50a6f.jpg) +Figure 5. Comparison of randomly selected virtual ID images generated by conventional methods [2, 5, 29, 35, 40, 48, 55] and by our method, all trained on CASIA-WebFace. For each synthetic dataset, we present two subjects in a single row. + +images that belong to class $k$ . As shown in the formula, a high $S_{k}$ indicates that the images are distinct from the negative classes. + +The intra-class diversity measures how various the conditions of samples are under the same label. In particular, we focus on challenging scenarios that directly impact FR performance, such as pose variations, occlusions, or lighting conditions, rather than image style. Higher diversity indicates that the dataset covers a broad spectrum of cases, from easy to hard. Motivated by this observation, we calculated the intra-class diversity of a class $k(\mathrm{D}_k)$ using the variance of Face Image Quality Assessment (FIQA) scores as follows: + +$$ +\overline {{\mathrm {F I Q A} _ {k}}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathrm {F I Q A} \left(x _ {i}\right), \tag {10} +$$ + +$$ +\mathrm {D} _ {k} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\operatorname {F I Q A} \left(x _ {i}\right) - \overline {{\operatorname {F I Q A} _ {k}}}\right) ^ {2}, \tag {11} +$$ + +where $\mathrm{FIQA}(x_i)$ indicates the normalized CR-FIQA [4] score of the generated image $x_{i}$ . + +The scores $C_k$ and $S_k$ are derived from a pre-trained ArcFace model that was trained using the Glint-360K dataset. Fig. 6a shows the average values of the properties for all classes, normalized by the average values achieved with CASIA-WebFace. VIGFace achieves remarkable scores in + +![](images/bbece17d284ceb0522e4bdb2bd1ea3224fb0e884f519d7c15b5936df81388791.jpg) +(a) The normalized properties of various (b) Identity leakage test on the synthetic datasets. The sizes of the cir- synthetic face dataset and its cor-cles indicate the intra-class diversity. responding training dataset. + +![](images/00623ab2a7fa511910c0a76de64cf27de0e23be7051e412b45ab70afa25afaa8.jpg) +Figure 6. Properties of VIGFace compared with those of synthesized datasets generated by previous methods. + +aspects of class consistency and class separability compared to other methods. SynFace exhibits the lowest separability due to its mix-up generation method. This fact indicates that SynFace can generate only a limited number of subjects as mentioned in the introduction. DigiFace achieves strong consistency, due to its 3D rendering methods, but its image style may differ significantly from real datasets, leading to lower accuracy in real-world applications. GANDiffFace demonstrates inadequate FIQA diversity, indicating limited variations in factors such as pose, lighting, and occlusion. DCFace, which uses a diffusion model, shows separability similar to that of VIGFace but with lower consistency. + +In comparisons of SOTA models, HSFace, despite using a more refined dataset, comparatively lacks consistency. CemiFace, trained on the same dataset as VIGFace, shows low consistency and separability. Both methods also exhibit inferior qualitative uniformity within the same class, as illustrated in Fig. 5. Our approach uses prototypes instead of FR embedding vectors to disentangle identity features from other characteristics, allowing VIGFace to generate subjects with greater consistency by focusing on identity features. Moreover, it enhances separability by leveraging the pre-computation of feature orthogonality. We included further analysis on the properties of the datasets and provided examples in Appendix H. + +Identity leakage from real human To claim complete privacy-free, it is necessary to prove that no real IDs or training images are included in the synthesized dataset. In this reason, we demonstrate that the generated face images represent non-existent humans by querying the most similar face in the CASIA-WebFace dataset. For comparison, we also present the similarity values of the nearest negative class in CASIA-WebFace itself. In Fig. 6b, we count the number of synthetic face images whose similarity with + +
MethodTraining Dataset# of Images (classes × variations)Verification Benchmarks
LFWCFP-FPCPLFWAgeDBCALFWAvg.
CASIA-Webface (Real)-0.49M (≈10.5K×47)99.4096.6390.2394.6893.7094.93
SynFaceFFHQ0.5M (10K × 50)91.9375.0370.4361.6374.7374.75
DigiFace3D modeling0.5M (10K × 50)95.4087.4078.8776.9778.6283.45
DCFaceFFHQ+CASIA0.5M (10K × 50)98.5585.3382.6289.7091.6089.56
IDiffFaceFFHQ0.5M (10K × 50)98.0085.4780.4586.4390.6588.20
GANDiffFaceFFHQ0.5M (10K × 50)90.7773.2772.3266.3574.6875.48
ID3CASIA0.5M (10K × 50)97.6886.8482.7791.0090.7389.80
CemiFaceCASIA+WebFace4M0.5M (10K × 50)99.0391.0687.6291.3392.4292.30
Arc2FaceWebFace42M0.5M (10K × 50)98.8191.8785.1690.1892.6391.73
HyperFaceWebFace42M0.5M (10K × 50)98.5088.8384.2386.5389.4089.50
HSFace10KWebFace4M0.5M (10K × 50)98.8788.9785.4793.1293.5792.00
VIGFace(S), OursCASIA0.5M (10K × 50)99.0295.0987.7290.9590.0092.56
DCFaceFFHQ+CASIA1.2M (20K × 50 + 40K × 5)98.5888.6185.0790.9792.8291.21
CemiFaceCASIA+WebFace4M1.2M (20K × 50 + 40K × 5)99.2292.8488.8692.1393.0393.22
Arc2FaceWebFace42M1.2M (20K × 50 + 40K × 5)98.9294.5886.4592.4593.3393.14
HSFace20KWebFace4M1.0M (20K × 50)98.8789.8786.1393.8593.6592.47
HSFace300KWebFace4M15M (300K × 50)99.3091.5487.7094.4594.5893.52
VIGFace(B), OursCASIA1.2M (60K × 20)99.4897.0790.1593.6292.8894.64
VIGFace(L), OursCASIA6.0M (120K × 50)99.3397.3191.1293.8292.9594.91
+ +Table 1. FR benchmark results trained with various virtual face datasets. All results except for CASIA-WebFace and VIGFace are obtained from the original paper. Our method is specified by the number of identities as small (S), base (B) and large (L). FR backbone is IR-SE50 + AdaFace [28]. Bold and underline indicates the best and the second best, respectively. + +the average feature vector of each class in the training data exceeds a given threshold. As shown in the figure, face images in VIGFace have low similarity to the subjects in CASIA-WebFace. This supports the claim that each synthesized image depicts a non-existent human face and there is no identity leakage from the training data. In contrast, we observed that some of the conventional SOTA methods exhibit identity leakage. As CemiFace [48] samples identity embeddings from the WebFace4M dataset, it generates individuals that exhibit resemblance to those in WebFace4M. Vec2Face [55] fails in creating novel identities, so that it produces almost same persons in WebFace4M. Further analysis and examples are reported in Appendix I. In this paper, we avoid thresholding or selective sampling, which could harm fair comparisons with previous methods. + +# 4.2. Evaluation + +In this section, we train the FR model with VIGFace to compare it to conventional methods. Detailed hyperparameters for training the FR network can be found in Appendix A. + +VIGFace as Virtual Dataset We compare the performance of the FR model trained with generated facial images for virtual IDs using VIGFace with conventional methods. For the experiment, we set the number of virtual IDs to $10K$ , $60K$ and $120K$ . To ensure a fair comparison, we established image counts of 0.5M, 1.2M, and 6.0M, which reflects the scale of the real CASIA-WebFace dataset and previous synthetic approaches, respectively. In Tab. 1, we present the 1:1 verification accuracy (\%) on five bench + +marks [23, 37, 44, 58, 59]. + +As shown in the table, VIGFace outperforms conventional methods in the average accuracy of five verification benchmarks. In particular, VIGFace achieves equal or even better performance on CFP-FP and CPLFW compared to the model trained with the real dataset, CASIA-WebFace. This indicates that our method benefits from its pose-variable generation capabilities. CemiFace shows notable performance on the AgeDB and CALFW benchmarks, but they sampled identity embeddings from the WebFace4M [60] dataset. Consequently, the individuals included in the CemiFace dataset may still have privacy concerns. HSFace10K also delivers outstanding results among conventional methods. However, it is important to note that HSFace benefits from training on the extensive WebFace4M dataset, which boasts twice as many images and quadruples the number of identities when compared to CASIA-WebFace. VIGFace(L) outperforms HSFace300K even with a smaller number of images. As a result, when the FR model was trained using VIGFace, it achieved performance comparable to a model trained using a real dataset in terms of average accuracy without any external identity sampling. This suggests that our virtual dataset can serve as a full replacement for the CASIA-WebFace dataset to train the FR model, while avoiding privacy concerns. + +VIGFace as Data Augmentation To demonstrate the efficacy of VIGFace as an augmentation framework, we evaluate the accuracy of the FR model trained on real and synthetic images. We utilized the dataset that was uploaded + +
ConditionVerification Benchmarks
MethodReal ImageSynthetic ImageLFWCFP-FPCPLFWAgeDBCALFWAvg.Δ
Real IDVirtual ID
CASIA-WebFace✓99.4096.6390.2394.6893.7094.93-
DigiFace✓✓99.3797.5190.9294.9593.7795.30+0.37
DCFace (1.2M)✓✓99.4396.9790.3395.2094.3895.26+0.33
iDiffFace✓✓99.5897.0490.4094.7894.0095.16+0.23
GANDiffFace✓✓99.5296.6190.3093.9893.5794.80-0.13
HSFace20K✓✓99.6097.4191.0795.4894.4095.59+0.66
VIGFace(B)✓✓99.4597.2390.7895.2593.2295.19+0.26
✓✓99.5598.0391.8095.7394.1295.85+0.92
✓✓✓99.7098.1091.5795.8594.3895.92+0.99
+ +Table 2. Augmented FR performance results for various conditions of synthetic dataset. The accuracies (%) for LFW, CFP-FP, CPLFW, AgeDB-30, CALFW, and the average benchmark accuracies are reported. FR backbone is IR-SE50 + AdaFace [28]. + +to the official repository. Tab. 2 shows the performance change of the trained FR model using various combinations of datasets. DigiFace, which utilizes unique 3D modeling technology, has demonstrated improved benchmark performance and offers advantages from a data-augmentation perspective. However, DigiFace struggles to perform well on its own, devoid of real data [2]. This is due to their failure to precisely replicate the appearance of real-world facial images, making it impractical to present a privacy-free synthetic dataset. HSFace shows a notable improvement among conventional methods. However, it utilized WebFace4M for training, which may make the comparison with others unfair. The boosted accuracy using VIGFace, which outperforms conventional methods, demonstrates that VIGFace can accurately mirror real facial data and achieve synergy by blending them. In particular, augmented images of real IDs improve the results of the CFP-FP and CPLFW benchmarks. This observation suggests that the use of five-point facial landmarks in the conditioning method can create a variety of posed facial images, substantially improving the FR model's ability to understand pose variations. Consequently, VIGFace not only effectively solves the privacy issue, but can also be used with real data as part of the augmentation process. When trained with the VIGFace dataset, the FR backbone achieves better generalization and higher performance, in contrast to conventional methods that achieved only marginal performance improvements. + +Fig. 5 illustrates the results that synthesize variations of real and virtual IDs. Since the training process is significantly biased not just by a lack of subject but also by inner-class variance, increasing the variance within the class is crucial. The results demonstrate that the proposed approach can generate a variety of conditions for the existing real ID without any help from external data. Since VIGFace can generate varied images of real individuals, we performed experiments on each type of augmentation, i.e. synthetic images of real and virtual ID. In particular, we extend the long-tailed real-ID class, which contains fewer than 50 im + +ages, to 50 images. With this method, additional 0.15M images were added as real ID synthesis. As can be seen in Tab. 2, the augmented dataset for real ID shows improved FR performance on most benchmarks. This indicates that our method can increase the intra-class diversity of dataset, a critical factor in achieving high FR performance. As a result, the FR model that utilizes the entire set of augmented data exhibits the best performance, benefiting from enhanced intra-class diversity and inter-class variance. + +# 5. Conclusions + +This paper presents an effective method for creating a synthetic face dataset that guarantees unique virtual identities. Synthesized facial data can serve as a solution for various challenges faced by traditional real datasets, such as high expenses, inaccuracies and biases in labeling, and concerns regarding privacy. To this end, we propose the Virtual Identity Generation framework and demonstrate that it can generate not only realistic but also diverse facial images of virtual individuals, significantly narrowing the performance gap with FR models trained on real data. Furthermore, the model exhibits superior performance when trained on a combination of VIGFace and existing real data compared to models trained solely on real data. This confirms that our proposed method has potential as an effective augmentation technique. We will publicly distribute the virtual face dataset created by VIGFace and believe that these new virtual data will contribute to resolving the privacy issues inherent in face recognition training dataset. + +Acknowledgment This work was partially supported by the Institutional Programs of the Korea Institute of Science and Technology (KIST) (Project No. 2E33612, $50\%$ ) and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2023-00227592, Development of Robust 3D Object Identification Technology for Viewpoint Changes, $50\%$ + +# References + +[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 3, 2 +[2] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 1, 2, 3, 6, 8, 4 +[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1578-1587, 2022. 2 +[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5836-5845, 2023. 6, 3 +[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 1, 2, 3, 6, 4 +[6] Johann S Brauchart, Alexander B Reznikov, Edward B Saff, Ian H Sloan, Yu Guang Wang, and Robert S Womersley. Random point sets on the sphere—hole radii, covering, and separation. Experimental Mathematics, 27(1):62-81, 2018. 3, 2 +[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 1 +[8] Hyunwoo Cho, Haesol Park, Ig-Jae Kim, and Junghyun Cho. Data augmentation of backscatter x-ray images for deep learning-based automatic cargo inspection. Sensors, 2021. 1 +[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 2, 3 +[10] Jiankang Deng, Jia Guo, Yuxiang Zhou, Jinke Yu, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641, 2019. 4 +[11] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 741-757. Springer, 2020. 1 +[12] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep + +face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 2 + +[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 1, 3 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2010.3, 4 +[15] Papantoniou et al. Arc2face: A foundation model for id-consistent human faces. In ECCV, 2024. 1 +[16] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 3 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 1 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4, 1 +[20] Je Hyeong Hong, Hanjo Kim, Minsoo Kim, Gi Pyo Nam, Junghyun Cho, Hyeong-Seok Ko, and Ig-Jae Kim. A 3d model-based approach for fitting masks to faces in the wild. In 2021 IEEE International Conference on Image Processing (ICIP), 2021. 1 +[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 1 +[22] Gary Huang, Marwan Mattar, Honglak Lee, and Erik Learned-Miller. Learning to align from scratch. Advances in neural information processing systems, 25, 2012. 3, 1 +[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 7 +[24] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 2 +[25] Hochul Hwang, Cheongjae Jang, Geonwoo Park, Junghyun Cho, and Ig-Jae Kim. Eldersim: A synthetic data generation + +platform for human action recognition in eldercare applications. IEEE Access, 2023. 1 +[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 1 +[27] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 1 +[28] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7, 8 +[29] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. Dface: Synthetic face generation with dual condition diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12715-12725, 2023. 1, 2, 3, 6, 4 +[30] D Kinga, Jimmy Ba Adam, et al. A method for stochastic optimization. In International conference on learning representations (ICLR), page 6. San Diego, California.; 2015. 1 +[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 4, 1 +[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1, 2 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1 +[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 2 +[35] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 3, 6, 4 +[36] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14225-14234, 2021. 2 +[37] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild + +age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 7 +[38] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7044-7053, 2017. 1 +[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4 +[40] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 1, 3, 6, 4 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2 +[42] Swami Sankaranarayanan, Azadeh Alavi, Carlos D Castillo, and Rama Chellappa. Triplet probabilistic embedding for face verification and clustering. In 2016 IEEE 8th international conference on biometrics theory, applications and systems (BTAS), pages 1-8. IEEE, 2016. 1 +[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1 +[44] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 7 +[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1 +[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1 +[47] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 1 +[48] Zhonglin Sun, Siyang Song, Ioannis Patras, and Georgios Tzimiropoulos. Cemifice: Center-based semi-hard synthetic face generation for face recognition. Advances in Neural Information Processing Systems, 37:35612-35638, 2025. 2, 3, 6, 7, 4 +[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 1 + +[50] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 1 +[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2 +[52] Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, and Chen Change Loy. The devil of face recognition is in the noise. In Proceedings of the European Conference on Computer Vision (ECCV), pages 765-780, 2018. 1 +[53] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2 +[54] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 499-515. Springer, 2016. 1 +[55] Haiyu Wu, Jaskirat Singh, Sicong Tian, Liang Zheng, and Kevin W Bowyer. Vec2face: Scaling face dataset generation with loosely constrained vectors. arXiv preprint arXiv:2409.02979, 2024. 2, 3, 6, 7, 1, 4 +[56] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 1 +[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 4 +[58] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 7 +[59] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 7 +[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10492-10502, 2021. 1, 3, 7 \ No newline at end of file diff --git a/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/images.zip b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a84891135737296bf80379b827b5f057f6b9f25 --- /dev/null +++ b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f72f62beac8b827915e055e592d825c9345bb4c1097d8f31be45817533167a9 +size 688912 diff --git a/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/layout.json b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d9fb9bd6ff1b73a0b2094d242b53425e5f4a58b2 --- /dev/null +++ b/2025/VIGFace_ Virtual Identity Generation for Privacy-Free Face Recognition Dataset/layout.json @@ -0,0 +1,8314 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 121 + ], + "type": "text", + "content": "VIGFace: Virtual Identity Generation for Privacy-Free Face Recognition dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "spans": [ + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "text", + "content": "Minsoo Kim" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "text", + "content": " Min-Cheol Sagong" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "text", + "content": " Gi Pyo Nam" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "text", + "content": " Junghyun Cho" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "text", + "content": " Ig-Jae Kim" + }, + { + "bbox": [ + 79, + 142, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 157, + 428, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 157, + 428, + 172 + ], + "spans": [ + { + "bbox": [ + 182, + 157, + 428, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 157, + 428, + 172 + ], + "type": "text", + "content": "Korea Institute of Science and Technology(KIST)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 173, + 171, + 437, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 171, + 437, + 186 + ], + "spans": [ + { + "bbox": [ + 173, + 171, + 437, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 171, + 437, + 186 + ], + "type": "text", + "content": "Korea National University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 193, + 185, + 418, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 185, + 418, + 200 + ], + "spans": [ + { + "bbox": [ + 193, + 185, + 418, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 193, + 185, + 418, + 200 + ], + "type": "text", + "content": "Yonsei-KIST Convergence Research Institute" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 177, + 201, + 430, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 201, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 177, + 201, + 430, + 213 + ], + "type": "text", + "content": "{kim1102,mcsagong,gpnam,jhcho,drjay}@kist.re.kr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 266, + 297, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 566 + ], + "type": "text", + "content": "Deep learning-based face recognition continues to face challenges due to its reliance on huge datasets obtained from web crawling, which can be costly to gather and raise significant real-world privacy concerns. To address this issue, we propose VIGFace, a novel framework capable of generating synthetic facial images. Our idea originates from pre-assigning virtual identities in the feature space. Initially, we train the face recognition model using a real face dataset and create a feature space for both real and virtual identities, where virtual prototypes are orthogonal to other prototypes. Subsequently, we train the diffusion model based on the established feature space, enabling it to generate authentic human face images from real prototypes and synthesize virtual face images from virtual prototypes. Our proposed framework provides two significant benefits. Firstly, it shows clear separability between existing individuals and virtual face images, allowing one to create synthetic images with confidence and without concerns about privacy and portrait rights. Secondly, it ensures improved performance through data augmentation by incorporating real existing images. Extensive experiments demonstrate the superiority of our virtual face dataset and framework, outperforming the previous state-of-the-art on various face recognition benchmarks. https://github.com/kim1102/VIGFace" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "content": "Deep learning-based Face Recognition (FR) models have significantly improved their performance due to recent advances in network architectures [17, 18, 21, 26, 45, 47] and enhancements in loss functions [9, 28, 32, 42, 43, 54]. The latest FR models utilize the softmax-variant loss for training to reduce intra-class variance and increase inter-class separability in the embedding space. This necessitates a very large dataset with numerous distinct individuals, signifi" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 242, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 434 + ], + "type": "text", + "content": "cant variations within each individual for intra-class variance, and precise labels of subject identities (IDs). However, datasets are typically collected through web-crawling and then refined using automatic techniques that employ FR logits [11, 60]. Although this approach is successful in eliminating mislabeled data, it still struggles with persistent issues of small intra-class variance. Moreover, in the face recognition field, unlike other ordinary image classifications, there are practical issues such as portrait rights, making it even more difficult to collect training data. For example, data sets such as those referenced in [16, 56, 60] consist of images of celebrities collected from the internet without consent. Furthermore, the datasets mentioned in [27, 38] include facial images of the general population, including children. The privacy sensitivity of such data poses significant challenges for face recognition research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": "Synthetic datasets have been used to address the limitations caused by the scarcity of real datasets [8, 25, 49] and biases present in the available real datasets [20, 50]. In the realm of face recognition, artificial faces show promise in addressing the aforementioned issues related to real face datasets. Generated faces are at low risk for label noise due to conditional generation. Bias problems, such as long-tailed distributions, which lead to class imbalances, can be mitigated by data augmentation. Importantly, there are no privacy concerns with virtual identities facial images if the method is effective. Therefore, when creating artificial datasets, it is essential to consider: 1) ensuring that the generated data mirror the real data distribution, 2) the ability to generate new subjects separated from real data, and 3) maintaining ID consistency for each subject." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "Previous attempts to create artificial face datasets have addressed some of the three aspects individually, but to the best of our knowledge, none have simultaneously taken into account all three aspects[2, 5, 29, 40]. SynFace [40] introduces a high-quality synthetic face image dataset, closely reminiscent of real face images, using DiscoFaceGAN [13]. However, DiscoFaceGAN is only capable of generating a limited number of unique subjects, fewer than 500 [29]. On" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 222, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10043" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 73, + 553, + 183 + ], + "blocks": [ + { + "bbox": [ + 65, + 73, + 553, + 183 + ], + "lines": [ + { + "bbox": [ + 65, + 73, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 65, + 73, + 553, + 183 + ], + "type": "image", + "image_path": "6fc969961fa4e39eacfade113c5cc0a22fb6a00e7a5626d4f3f72c95b52c3a00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 184, + 186, + 228 + ], + "blocks": [ + { + "bbox": [ + 65, + 184, + 186, + 228 + ], + "lines": [ + { + "bbox": [ + 65, + 184, + 186, + 228 + ], + "spans": [ + { + "bbox": [ + 65, + 184, + 186, + 228 + ], + "type": "image", + "image_path": "d819eacfc352f6cc1d4054a1391b661b8d4656347a54fde7f2b7cc07959c2683.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 282 + ], + "type": "text", + "content": "Figure 1. T-distributed Stochastic Neighbor Embedding (T-SNE) [51] plot of embeddings from real and synthesized images. The filled and lined stars represent the real and virtual prototypes, while filled and lined circles indicate the embeddings of real and synthesized images, respectively. The bottom of the figure shows the face images included in the cluster, and the dotted outlined images represent the face images generated using our method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 187, + 184, + 308, + 228 + ], + "blocks": [ + { + "bbox": [ + 187, + 184, + 308, + 228 + ], + "lines": [ + { + "bbox": [ + 187, + 184, + 308, + 228 + ], + "spans": [ + { + "bbox": [ + 187, + 184, + 308, + 228 + ], + "type": "image", + "image_path": "0eb5df0cbc394491258ccb41da6775eb48bccd7eebe237385f3727e11f36ca39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 308, + 184, + 429, + 228 + ], + "blocks": [ + { + "bbox": [ + 308, + 184, + 429, + 228 + ], + "lines": [ + { + "bbox": [ + 308, + 184, + 429, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 429, + 228 + ], + "type": "image", + "image_path": "a5903228c2906cfd2ad202515e58cd723c174f6cd93af2b5d3d05f33ce5f5ef0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 430, + 184, + 551, + 228 + ], + "blocks": [ + { + "bbox": [ + 430, + 184, + 551, + 228 + ], + "lines": [ + { + "bbox": [ + 430, + 184, + 551, + 228 + ], + "spans": [ + { + "bbox": [ + 430, + 184, + 551, + 228 + ], + "type": "image", + "image_path": "b6b4b0b13e19c97ea681a4a19be14469eceae1dba0d8193022065b58ed38325c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 294, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 296, + 486 + ], + "type": "text", + "content": "the other hand, DigiFace [2] utilizes 3D parametric modeling to create synthetic face images of various subjects. However, it faces difficulties in accurately replicating the quality and style distribution observed in real face images. DCFace [29] proposes a diffusion-based method to generate data that maintains the style of real datasets and ensures label consistency. However, DCFace primarily focuses on creating synthetic data rather than enhancing the capabilities of data augmentation. IDiffFace [5] utilizes identity-conditioned latent diffusion to synthesize facial images from FR feature representations. However, it fails to guarantee the uniqueness of the synthesized identities. CemiFace [48] and HSFace [55] demonstrate high performance on the various FR benchmarks. However, we found identity leakage of the training dataset from both CemiFace and HSFace, which is critical from a privacy standpoint." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": "The virtual data generated by our method possesses all three essential properties of virtual data mentioned above. The main concept of our paper is to incorporate virtual prototypes into the FR model. Virtual prototypes are trained simultaneously with real prototypes so that they pre-assign on the same feature space. The diffusion model is trained to preserve the identity of the individuals when generating face images based on the FR embedding. As virtual prototypes are allocated within the same embedding space as real prototypes, the produced face images are ensured to reflect the true data distribution. In addition, since prototypes are orthogonal to others, virtual subjects are guaranteed to be distinguished from the original data. Fig. 1 shows a toy example that visualizes the embeddings of real and virtual subjects to demonstrate the effectiveness of the suggested method. The virtual embedding can be distinct from the real individual clusters, whereas the images generated from virtual prototypes form unique clusters. The visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 295, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 426 + ], + "type": "text", + "content": "demonstrates that our framework generates distinctive virtual human faces with high consistency in ID. The proposed method effectively addresses privacy concerns by creating datasets of non-existent individuals and achieving state-of-the-art performance compared to models trained with previous virtual face generation methods. Additionally, the FR model, which was trained using a combination of real and synthetic images together, achieves better performance than the model trained using only real images. This shows that the proposed method is also superior from the perspective of a data augmentation method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 426, + 521, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 426, + 521, + 437 + ], + "spans": [ + { + "bbox": [ + 325, + 426, + 521, + 437 + ], + "type": "text", + "content": "Our contribution can be summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 554, + 559 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "content": "- Introducing VIGFace, a method designed to generate virtual identities with realistic appearance, guided by three main criteria: generating novel subjects that are not found in the real-world, consistently preserving the identity, but ensuring diverse characteristics for each individual." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "content": "- Showing that the synthetic data generated by VIGFace can leverage intra-class diversity and inter-class variance, achieving SOTA performance in face recognition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 559 + ], + "type": "text", + "content": "- Releasing the virtual-only face dataset that can fully substitute the real dataset, helping alleviate privacy concerns." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 569, + 399, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 399, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 399, + 580 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 588, + 454, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 454, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 454, + 601 + ], + "type": "text", + "content": "2.1. Face Recognition Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Current state-of-the-arts (SOTA) FR methods, such as [3, 9, 12, 24, 28, 32, 36, 53], are designed based on softmax loss, but particularly utilize the angular/cosine distance instead of the Euclidean distance. The goal of these methods is to maximize the similarity between the embeddings and the ground-truth (GT) prototype, while minimizing the similarity between the embedding and prototypes of other classes. As a result, all embeddings in the same class, including prototypes, converge while maintaining their maxi" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10044" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 87, + 304, + 232 + ], + "blocks": [ + { + "bbox": [ + 72, + 77, + 96, + 84 + ], + "lines": [ + { + "bbox": [ + 72, + 77, + 96, + 84 + ], + "spans": [ + { + "bbox": [ + 72, + 77, + 96, + 84 + ], + "type": "text", + "content": "Stage 1" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 87, + 304, + 232 + ], + "lines": [ + { + "bbox": [ + 61, + 87, + 304, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 304, + 232 + ], + "type": "image", + "image_path": "32cca86e77a26c1dd7fe611425fe8ab10da25f53c42bf38953a159929143ef5e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 76, + 336, + 83 + ], + "lines": [ + { + "bbox": [ + 313, + 76, + 336, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 76, + 336, + 83 + ], + "type": "text", + "content": "Stage 2" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 310, + 94, + 548, + 225 + ], + "blocks": [ + { + "bbox": [ + 310, + 94, + 548, + 225 + ], + "lines": [ + { + "bbox": [ + 310, + 94, + 548, + 225 + ], + "spans": [ + { + "bbox": [ + 310, + 94, + 548, + 225 + ], + "type": "image", + "image_path": "ef7405955034e53b21402fc9ee04575c8d1e2172a1180975db261f930f7bd042.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "text", + "content": "Figure 2. Pipeline for the proposed method. Conventional FR training includes prototypes for only real individuals, indicated as " + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "inline_equation", + "content": "W_{R}" + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "text", + "content": ". We add " + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "text", + "content": " prototypes for virtual IDs, denoted as " + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "text", + "content": ". The virtual embedding " + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FR}}'(x_j)" + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "text", + "content": " corresponding to the virtual person ID: " + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 238, + 554, + 293 + ], + "type": "text", + "content": " is generated to follow distribution of the real embeddings. To synthesize the facial image from virtual prototypes, we adopt the DiT architecture [39], following the design approach of the Vision Transformer (ViT) [14]. Additionally, we adjust the DiT model to utilize 5-point landmark images to handle pose variations." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 306, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 296, + 413 + ], + "type": "text", + "content": "mum distance from other classes. Therefore, if the embedding dimension is large enough, the clusters of each class will be nearly orthogonal [6, 9]. Naturally, a larger dataset that contains a greater variety of images allows for the training of better-performing FR models. Therefore, over time, larger and more diverse datasets [1, 16, 22, 60] have been collected and published in the academic field. However, issues such as portrait rights and the high cost of building large-scale data remain unresolved." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 423, + 234, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 234, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 234, + 435 + ], + "type": "text", + "content": "2.2. Synthetic Face Image Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 657 + ], + "type": "text", + "content": "SynFace [40] proposed a method that uses DiscoFaceGAN[13] to synthesize virtual faces. DigiFace-1M [2] proposes a 3D model-based face rendering method to generate virtual face data. DCFace [29] proposes a diffusion-based face generator combining subject appearance (ID) and external factor (style) conditions. GanDiffFace [35] uses GANs to generate identity features. Vec2Face [55] introduces a masked autoencoder to control the identity of face images and their attributes. CemiFace [48] produces facial samples with various levels of similarity to the center of the subject. IDiffFace [5] uses latent diffusion conditioned on FR identity to generate synthetic images. Although these methods succeed in generating separable identities, some still require additional networks or thresholding to sample synthetic identities. Furthermore, unlike [5], we propose a novel method for the pre-assignment of FR representations to maintain latent space alignment and introduce pose control of facial images using facial landmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 669, + 116, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 116, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 116, + 681 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Our framework comprises two stages. First, we train the FR model using the real face dataset and design the feature" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 366 + ], + "type": "text", + "content": "space for both real and virtual IDs. Second, synthetic images are generated using the diffusion model based on the feature space of the pre-trained FR model that was trained in the first stage. This section provides a detailed explanation of each stage of the proposed framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 372, + 469, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 469, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 469, + 385 + ], + "type": "text", + "content": "3.1. Stage 1: FR Model Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 390, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 509 + ], + "type": "text", + "content": "The proposed framework begins with training the FR model using real face images. This stage serves two purposes: 1) Training the FR model to achieve prototype features from face images, which are necessary for training the diffusion model in the second stage, and 2) Simultaneously assigning the positions of both the real ID and the virtual ID on the feature space. We choose ArcFace [9] to train the FR model in this stage. The ArcFace loss used to train the FR backbone " + }, + { + "bbox": [ + 313, + 390, + 555, + 509 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FR}}" + }, + { + "bbox": [ + 313, + 390, + 555, + 509 + ], + "type": "text", + "content": " and the prototype " + }, + { + "bbox": [ + 313, + 390, + 555, + 509 + ], + "type": "inline_equation", + "content": "W = [w_{1}, w_{2}, \\dots, w_{n}]" + }, + { + "bbox": [ + 313, + 390, + 555, + 509 + ], + "type": "text", + "content": " can be described as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 514, + 555, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 514, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 325, + 514, + 555, + 546 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {a r c}} = - \\log \\frac {e ^ {s \\cos \\left(\\theta_ {y _ {i}} + m\\right)}}{e ^ {s \\cos \\left(\\theta_ {y _ {i}} + m\\right)} + \\sum_ {j = 1 , j \\neq y _ {i}} ^ {n} e ^ {s \\cos \\theta_ {j}}}, \\tag {1}", + "image_path": "ea356b737792edd096fbeca47c428f849009fec77b13017c37cf504b21f820fe.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 551, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 551, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 378, + 551, + 553, + 578 + ], + "type": "interline_equation", + "content": "\\cos \\theta_ {j} = \\frac {w _ {j} ^ {T} f _ {\\mathrm {F R}} \\left(x _ {i}\\right)}{\\| w _ {j} \\| \\| f _ {\\mathrm {F R}} \\left(x _ {i}\\right) \\|}, \\tag {2}", + "image_path": "a8390c5adfddb62a7d76431b8ab98bb3c6b117f36f3d4ae40fe70652bef614ed.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " denotes the total number of real IDs, while " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " represent the margin and scale hyperparameters, respectively. With conventional FR training methods, only prototypes for real individuals, which can be denoted as " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "W_{R} = [w_{r}^{1},w_{r}^{2},\\dots,w_{r}^{n}]" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", are necessary for training. In contrast, we include additional " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " prototypes for virtual IDs, denoted as " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "W_{V} = [w_{v}^{1},w_{v}^{2},\\dots,w_{v}^{k}]" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", which are used to generate facial images of nonexistent individuals in the real world. As a result, the prototype is defined as a linear transformation matrix " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "W\\in \\mathbb{R}^{(n + k)\\times D}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " refers to the dimension of the embedding features." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10045" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 173, + 188 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 173, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 173, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 173, + 188 + ], + "type": "image", + "image_path": "43bd49bd2bc2e5d82b88ffca18bc08f6beecd074260540f272dd54e911c5c1f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 190, + 171, + 209 + ], + "lines": [ + { + "bbox": [ + 56, + 190, + 171, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 171, + 209 + ], + "type": "text", + "content": "(a) Prototypes that are trained without virtual embeddings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 219, + 295, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 295, + 242 + ], + "type": "text", + "content": "Figure 3. Changes in the similarity matrix of the prototypes from our method. Similarity values were min-max normalized." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 69, + 296, + 188 + ], + "blocks": [ + { + "bbox": [ + 179, + 69, + 296, + 188 + ], + "lines": [ + { + "bbox": [ + 179, + 69, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 179, + 69, + 296, + 188 + ], + "type": "image", + "image_path": "f0437c452840701ea3dd8d6ffde051f60d69aa442a3b6fa9965447d2ae05dbb9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 190, + 294, + 210 + ], + "lines": [ + { + "bbox": [ + 179, + 190, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 179, + 190, + 294, + 210 + ], + "type": "text", + "content": "(b) Prototypes that are trained with both real and virtual embeddings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "content": "However, due to the absence of a face image for virtual IDs, their prototypes cannot be updated to maximize the distance from each other with " + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "inline_equation", + "content": "L_{arc}" + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "content": ". Consequently, all virtual prototypes converge to a single point in the feature space when trained with the original " + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "inline_equation", + "content": "L_{arc}" + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "content": " as shown in Fig. 3a. To address this problem, we propose to use virtual feature embedding " + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FR}}'(x_j)" + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "content": " to update the virtual prototypes " + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "inline_equation", + "content": "w_v^j" + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "content": ". The virtual embedding " + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FR}}'(x_j)" + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "content": " corresponding to the virtual person ID: " + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 262, + 296, + 370 + ], + "type": "text", + "content": " was generated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 380, + 295, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 380, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 113, + 380, + 295, + 395 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {F R}} ^ {\\prime} (x _ {j}) = w _ {v} ^ {j} + \\mathcal {N} (0, 1) \\cdot \\sigma , \\tag {3}", + "image_path": "1f2e32ce3ec1bdf101d5c5982eca4bab05274552be5fcb81016e4729b1c2b672.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 403, + 295, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 403, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 113, + 403, + 295, + 437 + ], + "type": "interline_equation", + "content": "\\sigma^ {2} = \\frac {1}{b} \\sum_ {i = 1} ^ {b} \\left(f _ {\\mathrm {F R}} \\left(x _ {i}\\right) - w _ {r} ^ {i}\\right) ^ {2}, \\tag {4}", + "image_path": "07b4219e1dac81aabde8c9160a30ebb49e7d41b418183af79675f6e7fa81f7fa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": " refers the mini-batch size. As can be seen from the equations, the virtual embedding " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FR}}'(x_j)" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": " follows a distribution in which the standard deviation matches that of the real embeddings. The aggregate loss " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{arc}}" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": " is calculated simultaneously using both the real embedding " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FR}}(x)" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": " and the virtual embedding " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FR}}'(x)" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": ", allowing the virtual prototype " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "w_v^j" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": " to maintain minimal similarities with other prototypes, as illustrated in Fig. 3. Since batch configuration affects the calculated standard deviation, we utilize the exponential moving average (EMA) to reduce this influence. The corrected standard deviation " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": " for the current iteration " + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 441, + 296, + 584 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 595, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 595, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 110, + 595, + 295, + 609 + ], + "type": "interline_equation", + "content": "\\sigma = \\sigma_ {(t)} \\cdot \\alpha + \\sigma_ {(t - 1)} \\cdot (1 - \\alpha). \\tag {5}", + "image_path": "4aeb871a3433dee82f3c5499bc2c04c583c261647a5d22947fbf7c84405f3f81.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "The hyperparameter " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " is set to 0.9. The number of virtual embedding, " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "b_{v}" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " is determined based on the mini-batch size " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "b_{r}" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": ", the number of virtual ID prototypes " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": ", and the number of real ID prototypes " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": ". In our study, we set " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "b_{v} = (k\\times b_{r}) / n" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " so that both real and virtual prototypes can be updated evenly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The overall pipeline of this stage is illustrated in Fig. 2. Note that, as can be seen in the figure, gradients for virtual embeddings have no effect on the backbone. We confirm" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "that the similarity distribution between the virtual prototypes closely resembles the similarity distribution between the real prototypes, as shown in Fig. 3b." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 131, + 555, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 131, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 555, + 144 + ], + "type": "text", + "content": "3.2. Stage 2: Face Generation with Diffusion Model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": "The next step is to synthesize face images using the diffusion model. To obtain the training dataset for the diffusion model, we utilize the FR model, which involves collecting pairs of images " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": " and their corresponding prototypes " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "w_r" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": ". The input to our diffusion model includes the timestep " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": ", the FR prototype vector " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "w_r" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": ", the five facial landmark image " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": ", and the noisy image " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": ". In line with the approach proposed in the previous method [31], our model predicts velocity " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": " rather than noise " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": " injected into " + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 149, + 555, + 472 + ], + "type": "text", + "content": ". We adopt the DiT architecture [39] using the design approach of the visual transformer (ViT) [14]. We modify the DiT model by incorporating the five facial landmarks (including the left eye, the right eye, the nose, the left mouth corner, and the right mouth corner) [57] image. The five facial landmark images are acquired using RetinaFace [10], and employed as conditions to account for pose variations. Fig. 4 shows the synthesized face images conditioned by five facial landmarks. The first column lists input landmark conditions in different pose environments. As illustrated in the figure, the proposed model is capable of producing a wide range of styles (like hair, glasses, and lighting) while also delivering various pose variations, all with a high degree of consistency. In order to achieve both the generation of facial images and the synchronization of their feature space on the FR model, our diffusion model incorporates a constraint which aims to minimize the feature distance between the original image and the input prototypes as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 482, + 555, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 482, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 350, + 482, + 555, + 500 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\epsilon , t} \\| f _ {\\mathrm {F R}} (\\hat {x} _ {\\theta} (x _ {t}, t, w _ {r}, y))) - w _ {r} \\| _ {2} ^ {2}. \\tag {6}", + "image_path": "6c9e3989d9189eb4c475a19a477d71409e191ec24c7f1c5e25dbb116beb599a2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 509, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 545 + ], + "type": "text", + "content": "We adopt classifier-free guidance [19] by randomly assigning zero values to condition embeddings, " + }, + { + "bbox": [ + 313, + 509, + 554, + 545 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 313, + 509, + 554, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 509, + 554, + 545 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 509, + 554, + 545 + ], + "type": "text", + "content": " of time. Sampling is performed as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 559, + 553, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 559, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 323, + 559, + 553, + 587 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {x} _ {\\theta} \\left(x _ {t}, t, W, y\\right) = g \\cdot x _ {\\theta} \\left(x _ {t}, t, W, y\\right) \\\\ + (1 - g) \\cdot x _ {\\theta} (x _ {t}, t, y), \\tag {7} \\\\ \\end{array}", + "image_path": "ce55c8a312384de4af381193fb619e531256ec5dce5fd162dec0a25d5a6be323.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "inline_equation", + "content": "x_{\\theta}(x_t, t, w_r, y)" + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "inline_equation", + "content": "x_{\\theta}(x_t, t, y)" + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "text", + "content": " are conditional and unconditional " + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "text", + "content": "-prediction, respectively and " + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 597, + 555, + 633 + ], + "type": "text", + "content": " is guidance weight." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 645, + 395, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 395, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 395, + 658 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Our implementation details for the full framework are described in Appendix A. We analyze our framework from two perspectives. Firstly, by training the FR model exclusively with facial images generated for virtual IDs, we" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10046" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 249, + 240 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 249, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 249, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 249, + 240 + ], + "type": "image", + "image_path": "4d369e7e0eb34aed5072b34e6e75537d7dbf96fc545a00cc7c6aa741f0e5cf5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 282 + ], + "type": "text", + "content": "Figure 4. Virtual face generated in VIGFace (B). Each row lists facial images in different pose environments created using five facial landmarks. Our method can generate various conditions of face images, such as illumination, occlusion by accessories, and facial expressions, while controlling the pose variations of the face images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 256, + 71, + 402, + 240 + ], + "blocks": [ + { + "bbox": [ + 256, + 71, + 402, + 240 + ], + "lines": [ + { + "bbox": [ + 256, + 71, + 402, + 240 + ], + "spans": [ + { + "bbox": [ + 256, + 71, + 402, + 240 + ], + "type": "image", + "image_path": "09cd79ebeface17a8f98d200536d9ba5a88e0f98d308e421cbcadd9ae7374381.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 411, + 72, + 554, + 240 + ], + "blocks": [ + { + "bbox": [ + 411, + 72, + 554, + 240 + ], + "lines": [ + { + "bbox": [ + 411, + 72, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 411, + 72, + 554, + 240 + ], + "type": "image", + "image_path": "910c7a759608f38d4e511cb8105f360d78e14c165b3968a9e1e032c6fe672044.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "type": "text", + "content": "demonstrated the capability of our model to serve as a viable alternative to real face datasets, addressing concerns such as label noise or privacy. Secondly, we train the FR model using both real images and generated face images simultaneously, showcasing the potential of our model as a data augmentation framework for face recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 208, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 208, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 208, + 396 + ], + "type": "text", + "content": "4.1. Virtual Identity Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": "Since diffusion models focus on optimizing the visual aspect with input conditions, they tend to produce images that are high in consistency but lack of variance. One simple way to achieve a balance between consistency and variance is by adjusting the classifier-free guidance weight scale. Higher guidance enforces stronger conditioning of input labels. In other words, it results in more constant images but slightly similar samples. We observe that the proposed model performs best as the scale becomes " + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "inline_equation", + "content": "w = 4.0" + }, + { + "bbox": [ + 55, + 401, + 296, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "text", + "content": "We construct a toy example visualizing the feature space consists of three real persons and five virtual persons to demonstrate the efficacy of the synthesized images in Fig. 1. As seen in the figure on the left, the virtual embedding optimization obviously provides separable virtual prototypes from real individual clusters. The figure on the right illustrates that our synthesized images enforce high variance to the dataset. Synthesized images of real individuals filling the gap in the cluster provide intra-class variance to the subjects. Additionally, the cluster of generated images from virtual prototypes, while staying separate from other subjects, demonstrates the effectiveness of our framework for generating unique virtual human faces in high consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Fig. 5 shows the qualitative comparison of conventional synthetic datasets and the proposed VIGFace. As shown in the figure, the facial images generated using the VIGFace model show remarkable uniformity in generating consistent" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 387 + ], + "type": "text", + "content": "virtual individuals, while also incorporating variations such as hair styles, accessories, makeup, and expressions. It is important to emphasize that the key characteristic of a training dataset to achieve a high-performance FR model is not the high quality of the images themselves, but rather the high consistency and variance of the generated facial images belonging to the same person." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "text", + "content": "Property of generated Dataset To compare the properties of the synthetic face datasets, we measured the 1) class consistency, 2) class separability, and 3) intra-class diversity of generated images. The class consistency reflects the uniformity of the samples in the same label condition. Consequently, the consistency of class " + }, + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 313, + 400, + 554, + 483 + ], + "type": "text", + "content": ") was measured as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 490, + 554, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 490, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 359, + 490, + 554, + 524 + ], + "type": "interline_equation", + "content": "C _ {k} = \\frac {1}{N ^ {2}} \\sum_ {j = 1} ^ {N} \\sum_ {i = 1} ^ {N} \\frac {f \\left(x _ {i}\\right) \\cdot f \\left(x _ {j}\\right)}{\\| f \\left(x _ {i}\\right) \\| \\| f \\left(x _ {j}\\right) \\|}, \\tag {8}", + "image_path": "e88d535e695401492034ea9928dcfaaff70d1978da32f7487065dc99cf602806.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": " is the number of images in a class " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": ". Higher class consistency means that the samples are more uniform under the same label." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "content": "We also measured the class separability to assess the integrity of the dataset; in other words, to ensure that all subjects in the dataset are unique. The class separability for a class " + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "inline_equation", + "content": "S_k" + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "content": ") is measured as the average distance between the center of class " + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "content": " and the centers of the negative classes, as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 646, + 553, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 646, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 359, + 646, + 553, + 680 + ], + "type": "interline_equation", + "content": "\\mathrm {S} _ {k} = \\frac {1}{K - 1} \\sum_ {i = 1, i \\neq k} ^ {K} 1 - \\frac {\\overline {{f _ {k}}} \\cdot \\overline {{f _ {i}}}}{\\| \\overline {{f _ {k}}} \\| \\| \\overline {{f _ {i}}} \\|}, \\tag {9}", + "image_path": "fd03e5b8b9f46bcae391bfe1005434a53d179cf0b42d6a53f76ecabffbf87d53.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " is the number of classes. " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\overline{f_k}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " represents the class center obtained by averaging the embedding vectors of the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10047" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 75, + 292, + 327 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 292, + 327 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 292, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 292, + 327 + ], + "type": "image", + "image_path": "33feea5ec5acd2ac0a96374e0954de8ffb24046050a5198d485078224ec50a6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 296, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 376 + ], + "type": "text", + "content": "Figure 5. Comparison of randomly selected virtual ID images generated by conventional methods [2, 5, 29, 35, 40, 48, 55] and by our method, all trained on CASIA-WebFace. For each synthetic dataset, we present two subjects in a single row." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 388, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 424 + ], + "type": "text", + "content": "images that belong to class " + }, + { + "bbox": [ + 55, + 388, + 296, + 424 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 388, + 296, + 424 + ], + "type": "text", + "content": ". As shown in the formula, a high " + }, + { + "bbox": [ + 55, + 388, + 296, + 424 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 55, + 388, + 296, + 424 + ], + "type": "text", + "content": " indicates that the images are distinct from the negative classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 543 + ], + "type": "text", + "content": "The intra-class diversity measures how various the conditions of samples are under the same label. In particular, we focus on challenging scenarios that directly impact FR performance, such as pose variations, occlusions, or lighting conditions, rather than image style. Higher diversity indicates that the dataset covers a broad spectrum of cases, from easy to hard. Motivated by this observation, we calculated the intra-class diversity of a class " + }, + { + "bbox": [ + 55, + 425, + 296, + 543 + ], + "type": "inline_equation", + "content": "k(\\mathrm{D}_k)" + }, + { + "bbox": [ + 55, + 425, + 296, + 543 + ], + "type": "text", + "content": " using the variance of Face Image Quality Assessment (FIQA) scores as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 551, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 551, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 295, + 583 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathrm {F I Q A} _ {k}}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathrm {F I Q A} \\left(x _ {i}\\right), \\tag {10}", + "image_path": "808d8530af552050313b869be93aaa8dc4d3e64012e5e79625c7f116df9965ce.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 592, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 97, + 592, + 295, + 624 + ], + "type": "interline_equation", + "content": "\\mathrm {D} _ {k} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\operatorname {F I Q A} \\left(x _ {i}\\right) - \\overline {{\\operatorname {F I Q A} _ {k}}}\\right) ^ {2}, \\tag {11}", + "image_path": "b73c5e93cf56c34c6dc85f7a8fd30d262dc589e510c3d8cab9c426e61e60c6dc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{FIQA}(x_i)" + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": " indicates the normalized CR-FIQA [4] score of the generated image " + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "The scores " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "S_k" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " are derived from a pre-trained ArcFace model that was trained using the Glint-360K dataset. Fig. 6a shows the average values of the properties for all classes, normalized by the average values achieved with CASIA-WebFace. VIGFace achieves remarkable scores in" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 73, + 445, + 209 + ], + "blocks": [ + { + "bbox": [ + 315, + 73, + 445, + 209 + ], + "lines": [ + { + "bbox": [ + 315, + 73, + 445, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 445, + 209 + ], + "type": "image", + "image_path": "bbece17d284ceb0522e4bdb2bd1ea3224fb0e884f519d7c15b5936df81388791.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 243 + ], + "type": "text", + "content": "(a) The normalized properties of various (b) Identity leakage test on the synthetic datasets. The sizes of the cir- synthetic face dataset and its cor-cles indicate the intra-class diversity. responding training dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 447, + 76, + 552, + 209 + ], + "blocks": [ + { + "bbox": [ + 447, + 76, + 552, + 209 + ], + "lines": [ + { + "bbox": [ + 447, + 76, + 552, + 209 + ], + "spans": [ + { + "bbox": [ + 447, + 76, + 552, + 209 + ], + "type": "image", + "image_path": "00623ab2a7fa511910c0a76de64cf27de0e23be7051e412b45ab70afa25afaa8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "type": "text", + "content": "Figure 6. Properties of VIGFace compared with those of synthesized datasets generated by previous methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 430 + ], + "type": "text", + "content": "aspects of class consistency and class separability compared to other methods. SynFace exhibits the lowest separability due to its mix-up generation method. This fact indicates that SynFace can generate only a limited number of subjects as mentioned in the introduction. DigiFace achieves strong consistency, due to its 3D rendering methods, but its image style may differ significantly from real datasets, leading to lower accuracy in real-world applications. GANDiffFace demonstrates inadequate FIQA diversity, indicating limited variations in factors such as pose, lighting, and occlusion. DCFace, which uses a diffusion model, shows separability similar to that of VIGFace but with lower consistency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 432, + 556, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 588 + ], + "type": "text", + "content": "In comparisons of SOTA models, HSFace, despite using a more refined dataset, comparatively lacks consistency. CemiFace, trained on the same dataset as VIGFace, shows low consistency and separability. Both methods also exhibit inferior qualitative uniformity within the same class, as illustrated in Fig. 5. Our approach uses prototypes instead of FR embedding vectors to disentangle identity features from other characteristics, allowing VIGFace to generate subjects with greater consistency by focusing on identity features. Moreover, it enhances separability by leveraging the pre-computation of feature orthogonality. We included further analysis on the properties of the datasets and provided examples in Appendix H." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Identity leakage from real human To claim complete privacy-free, it is necessary to prove that no real IDs or training images are included in the synthesized dataset. In this reason, we demonstrate that the generated face images represent non-existent humans by querying the most similar face in the CASIA-WebFace dataset. For comparison, we also present the similarity values of the nearest negative class in CASIA-WebFace itself. In Fig. 6b, we count the number of synthetic face images whose similarity with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10048" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 553, + 291 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 291 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 291 + ], + "type": "table", + "html": "
MethodTraining Dataset# of Images (classes × variations)Verification Benchmarks
LFWCFP-FPCPLFWAgeDBCALFWAvg.
CASIA-Webface (Real)-0.49M (≈10.5K×47)99.4096.6390.2394.6893.7094.93
SynFaceFFHQ0.5M (10K × 50)91.9375.0370.4361.6374.7374.75
DigiFace3D modeling0.5M (10K × 50)95.4087.4078.8776.9778.6283.45
DCFaceFFHQ+CASIA0.5M (10K × 50)98.5585.3382.6289.7091.6089.56
IDiffFaceFFHQ0.5M (10K × 50)98.0085.4780.4586.4390.6588.20
GANDiffFaceFFHQ0.5M (10K × 50)90.7773.2772.3266.3574.6875.48
ID3CASIA0.5M (10K × 50)97.6886.8482.7791.0090.7389.80
CemiFaceCASIA+WebFace4M0.5M (10K × 50)99.0391.0687.6291.3392.4292.30
Arc2FaceWebFace42M0.5M (10K × 50)98.8191.8785.1690.1892.6391.73
HyperFaceWebFace42M0.5M (10K × 50)98.5088.8384.2386.5389.4089.50
HSFace10KWebFace4M0.5M (10K × 50)98.8788.9785.4793.1293.5792.00
VIGFace(S), OursCASIA0.5M (10K × 50)99.0295.0987.7290.9590.0092.56
DCFaceFFHQ+CASIA1.2M (20K × 50 + 40K × 5)98.5888.6185.0790.9792.8291.21
CemiFaceCASIA+WebFace4M1.2M (20K × 50 + 40K × 5)99.2292.8488.8692.1393.0393.22
Arc2FaceWebFace42M1.2M (20K × 50 + 40K × 5)98.9294.5886.4592.4593.3393.14
HSFace20KWebFace4M1.0M (20K × 50)98.8789.8786.1393.8593.6592.47
HSFace300KWebFace4M15M (300K × 50)99.3091.5487.7094.4594.5893.52
VIGFace(B), OursCASIA1.2M (60K × 20)99.4897.0790.1593.6292.8894.64
VIGFace(L), OursCASIA6.0M (120K × 50)99.3397.3191.1293.8292.9594.91
", + "image_path": "90faf2b729efadf545d320fef6bcc4720b9dad36148f99a5d54561d73c39cc7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 298, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 332 + ], + "type": "text", + "content": "Table 1. FR benchmark results trained with various virtual face datasets. All results except for CASIA-WebFace and VIGFace are obtained from the original paper. Our method is specified by the number of identities as small (S), base (B) and large (L). FR backbone is IR-SE50 + AdaFace [28]. Bold and underline indicates the best and the second best, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 345, + 297, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 345, + 297, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 297, + 525 + ], + "type": "text", + "content": "the average feature vector of each class in the training data exceeds a given threshold. As shown in the figure, face images in VIGFace have low similarity to the subjects in CASIA-WebFace. This supports the claim that each synthesized image depicts a non-existent human face and there is no identity leakage from the training data. In contrast, we observed that some of the conventional SOTA methods exhibit identity leakage. As CemiFace [48] samples identity embeddings from the WebFace4M dataset, it generates individuals that exhibit resemblance to those in WebFace4M. Vec2Face [55] fails in creating novel identities, so that it produces almost same persons in WebFace4M. Further analysis and examples are reported in Appendix I. In this paper, we avoid thresholding or selective sampling, which could harm fair comparisons with previous methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 130, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 130, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 130, + 545 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 588 + ], + "type": "text", + "content": "In this section, we train the FR model with VIGFace to compare it to conventional methods. Detailed hyperparameters for training the FR network can be found in Appendix A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "VIGFace as Virtual Dataset We compare the performance of the FR model trained with generated facial images for virtual IDs using VIGFace with conventional methods. For the experiment, we set the number of virtual IDs to " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "10K" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "60K" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "inline_equation", + "content": "120K" + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": ". To ensure a fair comparison, we established image counts of 0.5M, 1.2M, and 6.0M, which reflects the scale of the real CASIA-WebFace dataset and previous synthetic approaches, respectively. In Tab. 1, we present the 1:1 verification accuracy (\\%) on five bench" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 345, + 423, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 423, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 423, + 357 + ], + "type": "text", + "content": "marks [23, 37, 44, 58, 59]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 357, + 555, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 357, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 312, + 357, + 555, + 646 + ], + "type": "text", + "content": "As shown in the table, VIGFace outperforms conventional methods in the average accuracy of five verification benchmarks. In particular, VIGFace achieves equal or even better performance on CFP-FP and CPLFW compared to the model trained with the real dataset, CASIA-WebFace. This indicates that our method benefits from its pose-variable generation capabilities. CemiFace shows notable performance on the AgeDB and CALFW benchmarks, but they sampled identity embeddings from the WebFace4M [60] dataset. Consequently, the individuals included in the CemiFace dataset may still have privacy concerns. HSFace10K also delivers outstanding results among conventional methods. However, it is important to note that HSFace benefits from training on the extensive WebFace4M dataset, which boasts twice as many images and quadruples the number of identities when compared to CASIA-WebFace. VIGFace(L) outperforms HSFace300K even with a smaller number of images. As a result, when the FR model was trained using VIGFace, it achieved performance comparable to a model trained using a real dataset in terms of average accuracy without any external identity sampling. This suggests that our virtual dataset can serve as a full replacement for the CASIA-WebFace dataset to train the FR model, while avoiding privacy concerns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "VIGFace as Data Augmentation To demonstrate the efficacy of VIGFace as an augmentation framework, we evaluate the accuracy of the FR model trained on real and synthetic images. We utilized the dataset that was uploaded" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10049" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 216 + ], + "type": "table", + "html": "
ConditionVerification Benchmarks
MethodReal ImageSynthetic ImageLFWCFP-FPCPLFWAgeDBCALFWAvg.Δ
Real IDVirtual ID
CASIA-WebFace✓99.4096.6390.2394.6893.7094.93-
DigiFace✓✓99.3797.5190.9294.9593.7795.30+0.37
DCFace (1.2M)✓✓99.4396.9790.3395.2094.3895.26+0.33
iDiffFace✓✓99.5897.0490.4094.7894.0095.16+0.23
GANDiffFace✓✓99.5296.6190.3093.9893.5794.80-0.13
HSFace20K✓✓99.6097.4191.0795.4894.4095.59+0.66
VIGFace(B)✓✓99.4597.2390.7895.2593.2295.19+0.26
✓✓99.5598.0391.8095.7394.1295.85+0.92
✓✓✓99.7098.1091.5795.8594.3895.92+0.99
", + "image_path": "15e02a97e2f8b3b5b5abcbcc7fbeefc5f5c41fd8a5a96d6b7ba34a064472ad2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 222, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 555, + 246 + ], + "type": "text", + "content": "Table 2. Augmented FR performance results for various conditions of synthetic dataset. The accuracies (%) for LFW, CFP-FP, CPLFW, AgeDB-30, CALFW, and the average benchmark accuracies are reported. FR backbone is IR-SE50 + AdaFace [28]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 258, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 581 + ], + "type": "text", + "content": "to the official repository. Tab. 2 shows the performance change of the trained FR model using various combinations of datasets. DigiFace, which utilizes unique 3D modeling technology, has demonstrated improved benchmark performance and offers advantages from a data-augmentation perspective. However, DigiFace struggles to perform well on its own, devoid of real data [2]. This is due to their failure to precisely replicate the appearance of real-world facial images, making it impractical to present a privacy-free synthetic dataset. HSFace shows a notable improvement among conventional methods. However, it utilized WebFace4M for training, which may make the comparison with others unfair. The boosted accuracy using VIGFace, which outperforms conventional methods, demonstrates that VIGFace can accurately mirror real facial data and achieve synergy by blending them. In particular, augmented images of real IDs improve the results of the CFP-FP and CPLFW benchmarks. This observation suggests that the use of five-point facial landmarks in the conditioning method can create a variety of posed facial images, substantially improving the FR model's ability to understand pose variations. Consequently, VIGFace not only effectively solves the privacy issue, but can also be used with real data as part of the augmentation process. When trained with the VIGFace dataset, the FR backbone achieves better generalization and higher performance, in contrast to conventional methods that achieved only marginal performance improvements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Fig. 5 illustrates the results that synthesize variations of real and virtual IDs. Since the training process is significantly biased not just by a lack of subject but also by inner-class variance, increasing the variance within the class is crucial. The results demonstrate that the proposed approach can generate a variety of conditions for the existing real ID without any help from external data. Since VIGFace can generate varied images of real individuals, we performed experiments on each type of augmentation, i.e. synthetic images of real and virtual ID. In particular, we extend the long-tailed real-ID class, which contains fewer than 50 im" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 556, + 368 + ], + "type": "text", + "content": "ages, to 50 images. With this method, additional 0.15M images were added as real ID synthesis. As can be seen in Tab. 2, the augmented dataset for real ID shows improved FR performance on most benchmarks. This indicates that our method can increase the intra-class diversity of dataset, a critical factor in achieving high FR performance. As a result, the FR model that utilizes the entire set of augmented data exhibits the best performance, benefiting from enhanced intra-class diversity and inter-class variance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 377, + 392, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 392, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 392, + 391 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 399, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 615 + ], + "type": "text", + "content": "This paper presents an effective method for creating a synthetic face dataset that guarantees unique virtual identities. Synthesized facial data can serve as a solution for various challenges faced by traditional real datasets, such as high expenses, inaccuracies and biases in labeling, and concerns regarding privacy. To this end, we propose the Virtual Identity Generation framework and demonstrate that it can generate not only realistic but also diverse facial images of virtual individuals, significantly narrowing the performance gap with FR models trained on real data. Furthermore, the model exhibits superior performance when trained on a combination of VIGFace and existing real data compared to models trained solely on real data. This confirms that our proposed method has potential as an effective augmentation technique. We will publicly distribute the virtual face dataset created by VIGFace and believe that these new virtual data will contribute to resolving the privacy issues inherent in face recognition training dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 624, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 557, + 715 + ], + "type": "text", + "content": "Acknowledgment This work was partially supported by the Institutional Programs of the Korea Institute of Science and Technology (KIST) (Project No. 2E33612, " + }, + { + "bbox": [ + 313, + 624, + 557, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 624, + 557, + 715 + ], + "type": "text", + "content": " ) and the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2023-00227592, Development of Robust 3D Object Identification Technology for Viewpoint Changes, " + }, + { + "bbox": [ + 313, + 624, + 557, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10050" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 114, + 295, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 169 + ], + "type": "text", + "content": "[1] Xiang An, Xuhan Zhu, Yuan Gao, Yang Xiao, Yongle Zhao, Ziyong Feng, Lan Wu, Bin Qin, Ming Zhang, Debing Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445-1449, 2021. 3, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 295, + 235 + ], + "type": "text", + "content": "[2] Gwangbin Bae, Martin de La Gorce, Tadas Baltrusaitis, Charlie Hewitt, Dong Chen, Julien Valentin, Roberto Cipolla, and Jingjing Shen. Digiface-1m: 1 million digital face images for face recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3526-3535, 2023. 1, 2, 3, 6, 8, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "type": "text", + "content": "[3] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1578-1587, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 292, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 295, + 346 + ], + "type": "text", + "content": "[4] Fadi Boutros, Meiling Fang, Marcel Klemt, Biying Fu, and Naser Damer. Cr-fiqa: face image quality assessment by learning sample relative classifiability. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5836-5845, 2023. 6, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 347, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 401 + ], + "type": "text", + "content": "[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 1, 2, 3, 6, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 403, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 456 + ], + "type": "text", + "content": "[6] Johann S Brauchart, Alexander B Reznikov, Edward B Saff, Ian H Sloan, Yu Guang Wang, and Robert S Womersley. Random point sets on the sphere—hole radii, covering, and separation. Experimental Mathematics, 27(1):62-81, 2018. 3, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "type": "text", + "content": "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 481, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 523 + ], + "type": "text", + "content": "[8] Hyunwoo Cho, Haesol Park, Ig-Jae Kim, and Junghyun Cho. Data augmentation of backscatter x-ray images for deep learning-based automatic cargo inspection. Sensors, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "type": "text", + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 624 + ], + "type": "text", + "content": "[10] Jiankang Deng, Jia Guo, Yuxiang Zhou, Jinke Yu, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv:1905.00641, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 690 + ], + "type": "text", + "content": "[11] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI 16, pages 741-757. Springer, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Jiankang Deng, Jia Guo, Jing Yang, Alexandros Lattas, and Stefanos Zafeiriou. Variational prototype learning for deep" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "text", + "content": "face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11906-11915, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 114, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 317, + 106, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 161 + ], + "type": "text", + "content": "[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2010.3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 249 + ], + "type": "text", + "content": "[15] Papantoniou et al. Arc2face: A foundation model for id-consistent human faces. In ECCV, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 251, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 316 + ], + "type": "text", + "content": "[16] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 87-102. Springer, 2016. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 317, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 369 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 4, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "text", + "content": "[20] Je Hyeong Hong, Hanjo Kim, Minsoo Kim, Gi Pyo Nam, Junghyun Cho, Hyeong-Seok Ko, and Ig-Jae Kim. A 3d model-based approach for fitting masks to faces in the wild. In 2021 IEEE International Conference on Image Processing (ICIP), 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "type": "text", + "content": "[22] Gary Huang, Marwan Mattar, Honglak Lee, and Erik Learned-Miller. Learning to align from scratch. Advances in neural information processing systems, 25, 2012. 3, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[23] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 689 + ], + "type": "text", + "content": "[24] Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Hochul Hwang, Cheongjae Jang, Geonwoo Park, Junghyun Cho, and Ig-Jae Kim. Eldersim: A synthetic data generation" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10051" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "platform for human action recognition in eldercare applications. IEEE Access, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "text", + "content": "[27] Ira Kemelmacher-Shlizerman, Steven M Seitz, Daniel Miller, and Evan Brossard. The megaface benchmark: 1 million faces for recognition at scale. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4873-4882, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "text", + "content": "[28] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18750-18759, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "text", + "content": "[29] Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. Dface: Synthetic face generation with dual condition diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12715-12725, 2023. 1, 2, 3, 6, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 342 + ], + "type": "text", + "content": "[30] D Kinga, Jimmy Ba Adam, et al. A method for stochastic optimization. In International conference on learning representations (ICLR), page 6. San Diego, California.; 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 398 + ], + "type": "text", + "content": "[31] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 4, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "text", + "content": "[32] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherical face: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 478 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 544 + ], + "type": "text", + "content": "[34] Brianna Maze, Jocelyn Adams, James A Duncan, Nathan Kalka, Tim Miller, Charles Otto, Anil K Jain, W Tyler Niggel, Janet Anderson, Jordan Cheney, et al. Iarpa janus benchmark-c: Face dataset and protocol. In 2018 international conference on biometrics (ICB), pages 158-165. IEEE, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 621 + ], + "type": "text", + "content": "[35] Pietro Melzi, Christian Rathgeb, Ruben Tolosana, Ruben Vera-Rodriguez, Dominik Lawatsch, Florian Domin, and Maxim Schaubert. Gandifface: Controllable generation of synthetic datasets for face recognition with realistic variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3086-3095, 2023. 3, 6, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "text", + "content": "[36] Qiang Meng, Shichao Zhao, Zhida Huang, and Feng Zhou. Magface: A universal representation for face recognition and quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14225-14234, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[37] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "age database. In proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 51-59, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[38] Aaron Nech and Ira Kemelmacher-Shlizerman. Level playing field for million scale face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7044-7053, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "text", + "content": "[40] Haibo Qiu, Baosheng Yu, Dihong Gong, Zhifeng Li, Wei Liu, and Dacheng Tao. Synface: Face recognition with synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10880-10890, 2021. 1, 3, 6, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "type": "text", + "content": "[42] Swami Sankaranarayanan, Azadeh Alavi, Carlos D Castillo, and Rama Chellappa. Triplet probabilistic embedding for face verification and clustering. In 2016 IEEE 8th international conference on biometrics theory, applications and systems (BTAS), pages 1-8. IEEE, 2016. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[43] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "text", + "content": "[44] Soumyadip Sengupta, Jun-Cheng Chen, Carlos Castillo, Vishal M Patel, Rama Chellappa, and David W Jacobs. Frontal to profile face verification in the wild. In 2016 IEEE winter conference on applications of computer vision (WACV), pages 1-9. IEEE, 2016. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 472, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 504 + ], + "type": "text", + "content": "[45] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 536 + ], + "type": "text", + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[47] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[48] Zhonglin Sun, Siyang Song, Ioannis Patras, and Georgios Tzimiropoulos. Cemifice: Center-based semi-hard synthetic face generation for face recognition. Advances in Neural Information Processing Systems, 37:35612-35638, 2025. 2, 3, 6, 7, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 713 + ], + "type": "text", + "content": "[49] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 969-977, 2018. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10052" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 620 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[50] Boris van Breugel, Trent Kyono, Jeroen Berrevoets, and Michaela van der Schaar. Decaf: Generating fair synthetic data using causally-aware generative networks. Advances in Neural Information Processing Systems, 34:22221-22233, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "type": "text", + "content": "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 216 + ], + "type": "text", + "content": "[52] Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, and Chen Change Loy. The devil of face recognition is in the noise. In Proceedings of the European Conference on Computer Vision (ECCV), pages 765-780, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "text", + "content": "[53] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 339 + ], + "type": "text", + "content": "[54] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A discriminative feature learning approach for deep face recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VII 14, pages 499-515. Springer, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 384 + ], + "type": "text", + "content": "[55] Haiyu Wu, Jaskirat Singh, Sicong Tian, Liang Zheng, and Kevin W Bowyer. Vec2face: Scaling face dataset generation with loosely constrained vectors. arXiv preprint arXiv:2409.02979, 2024. 2, 3, 6, 7, 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 387, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 418 + ], + "type": "text", + "content": "[56] Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z Li. Learning face representation from scratch. arXiv preprint arXiv:1411.7923, 2014. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 463 + ], + "type": "text", + "content": "[57] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE signal processing letters, 23 (10):1499-1503, 2016. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "text", + "content": "[58] Tianyue Zheng and Weihong Deng. Cross-pose lfw: A database for studying cross-posed face recognition in unconstrained environments. Beijing University of Posts and Telecommunications, Tech. Rep, 5(7), 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 552 + ], + "type": "text", + "content": "[59] Tianyue Zheng, Weihong Deng, and Jiani Hu. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 620 + ], + "type": "text", + "content": "[60] Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Jiwen Lu, Dalong Du, et al. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10492-10502, 2021. 1, 3, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10053" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_content_list.json b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..293a73ee91acf12fe04bdc551df4fac132e874ad --- /dev/null +++ b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_content_list.json @@ -0,0 +1,1550 @@ +[ + { + "type": "text", + "text": "VIPerson: Flexibly Generating Virtual Identity for Person Re-Identification", + "text_level": 1, + "bbox": [ + 116, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao-Wen Zhang $^{1,3*}$ \nZhi Ouyang $^{1,3}$", + "bbox": [ + 245, + 179, + 410, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Delong Zhang $^{1,3*}$ Jingke Meng $^{1,3\\dagger}$", + "bbox": [ + 439, + 179, + 584, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi-Xing Peng $^{1,3}$ Wei-Shi Zheng $^{1,2,3\\dagger}$", + "bbox": [ + 599, + 179, + 759, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China;", + "bbox": [ + 186, + 215, + 807, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Peng Cheng Laboratory, Shenzhen, China;", + "bbox": [ + 326, + 233, + 671, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China.", + "bbox": [ + 106, + 250, + 887, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangxw75,zhangdlong5,pengyx23,}@mail2.sysu.edu.cn wszheng@ieee.org", + "bbox": [ + 178, + 271, + 813, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) is to match the person images under different camera views. Training ReID models necessitates a substantial amount of labeled real-world data, leading to high labeling costs and privacy issues. Although several ReID data synthetic methods are proposed to address these issues, they fail to generate images with new identities or real-world camera styles. In this paper, we propose a novel pedestrian generation pipeline, VIPerson, to generate camera-realistic pedestrian images with flexible Virtual Identities for the Person ReID task. VIPerson focuses on three key factors in data synthesis: (I) Virtual identity diversity: Enhancing the latent diffusion model with our proposed dropout text embedding, we flexibly generate random and hard identities. (II) Scalable cross-camera variations: VIPerson introduces scalable variations of scenes and poses within each identity. (III) Camera-realistic style: Adopting an identity-agnostic approach to transfer realistic styles, we avoid privacy exposure of real identities. Extensive experimental results across a broad range of downstream ReID tasks demonstrate the superiority of our generated dataset over existing methods. In addition, VIPerson can be adapted to the identity expansion scenario, which widens the application of our pipeline. The dataset and code of VIPerson are available at https://isee-laboratory.github.io/VIPerson/.", + "bbox": [ + 88, + 354, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The person re-identification (ReID) task aims to match pedestrians from different camera views. With the rise of deep learning[52], ReID methods have been developed based on real annotated pedestrian data [44, 59]. These real-world ReID datasets are often limited in both scale and", + "bbox": [ + 89, + 789, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0a50b0054e4d39ec4508e89ee9098d9fd5e45b16c576a0016a805b657a754c6.jpg", + "image_caption": [ + "Figure 1. Comparison of synthetic data. Left: Images of engine-based methods have a style gap compared with real images. Right: Learning-based methods expose privacy by generating images that match real identities in real-world datasets." + ], + "image_footnote": [], + "bbox": [ + 558, + 321, + 859, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "diversity, primarily due to the high cost of data collection and manual annotation. Moreover, pedestrian images containing sensitive information often expose real identities, raising serious privacy concerns. Strict data privacy regulations [2, 4, 17, 35], such as EU's General Data Protection Regulation, restrict the ability to gather and share sensitive visual data. This further impedes the collection of public ReID datasets, with some, like DukeMTMC [61], even taken down. The lack of high-quality annotated data without privacy concerns significantly limits the performance and generalization ability of ReID models.", + "bbox": [ + 511, + 580, + 906, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, recent studies propose to synthesize pedestrian images to improve ReID models. These methods can be divided into two research streams. Engine-based methods [1, 41, 42, 57] are developed to leverage simulation engines to synthesize virtual pedestrian images. However, these synthetic images exhibit distinct virtual characteristics (see 'Left' in Figure 1), resulting in a style gap compared to real-world images that showcase camera-realistic styles (e.g. illumination, motion blur, camera bias, and environmental conditions). These synthetic", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23374", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ed0c6fb776e314454d32e8ebb09d663ae2549bde02f9b4d9fa2e4fd35062331.jpg", + "image_caption": [ + "Figure 2. Image examples from our VIPerson dataset. Our data pipeline can generate virtual identities, containing random identities, hard identities, and intra-identity (intra-ID) groups that contain camera-realistic images with rich cross-camera variations. More images are provided in section A of the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 473, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data hinder ReID models' transferability to real-world scenarios. Learning-based synthetic approaches [7, 19, 50], propose using generative models[13, 16], to generate realistic pedestrian images. However, these models face privacy concerns as they are only capable of generating images that are consistent with the real identities present in real-world ReID datasets (see 'Right' in Figure 1), which limits their application in privacy-sensitive scenarios. Moreover, both approaches struggle with limited cross-camera variations (e.g. poses and scenes) due to constraints imposed by costs of scene design or limited variations in training data.", + "bbox": [ + 89, + 412, + 482, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to propose a novel pipeline to generate a camera-realistic pedestrian dataset composed entirely of virtual identities for the person ReID task. However, achieving this is non-trivial due to three core challenges in pedestrian data synthesis: (1) How to flexibly generate diverse virtual identities, containing new random identities and hard identities? New random identities exhibit diverse appearances, while hard identities, often referred to as hard samples, share similar visual characteristics despite belonging to different individuals. Both are crucial for enhancing identity diversity, aiding ReID models to learn discriminative features across a broad range of individuals. This necessitates a flexible generative model for skillfully controlling the generation of diverse identities. (2) How to introduce scalable cross-camera variations within each identity? Abundant cross-camera variations of generated images can facilitate the generalization ability of ReID models. How to adapt a specific identity to extensive pose and scene variations is very challenging. (3) How to produce images with camera-realistic styles in an identity-agnostic way? Engine-based methods struggle", + "bbox": [ + 89, + 584, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to generalize effectively on real data due to insufficient realism, whereas realistic styles introduced by learning-based methods are often entangled with real identity information.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present a pedestrian generation pipeline, entitled VIPerson, to flexibly generate camera-realistic images of pedestrians with Virtual Identities for the Person ReID task. Our pipeline comprises three components: flexible identity generator, cross-camera diversification, and style realization. First, we propose a flexible identity generator, which enhances the stable diffusion model through our proposed text embedding dropout. This enhancement allows the model to generate hard identities based on dropout text embeddings, while still being flexibly capable of infinitely creating new virtual identities using random embeddings. Second, for each generated identity image, we introduce cross-camera diversification, enabling diverse and scalable changes in poses and scenes. Leveraging the IP-adapter [51, 56] to vary poses and our proposed segment-and-paste strategy to alter scenes, we seamlessly facilitate cross-camera variations for each identity. Third, we present style realization to extract and transfer identity-agnostic camera-realistic style in real images to virtual images using style transfer techniques [9]. This process effectively reduces the style gap between generated and real images, while preserving the virtual identities. To capture comprehensive poses, scenes, and styles, we leverage pedestrian images from the large-scale unlabeled dataset LUPerson [12].", + "bbox": [ + 511, + 138, + 906, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We perform in-depth experimental analysis on three key factors, i.e. virtual identity diversity, cross-camera variations, and camera-realistic style in data synthesis. These factors are verified to be crucial for enhancing the generalization ability of the ReID model. Furthermore, when knowledge of scenes and styles from specific scenarios is available, our approach can be extended to incorporate scenario-specific patterns, thereby enhancing ReID performances in target applications. The effectiveness of our generated dataset is validated in two groups of experiments. First, we verify the effectiveness of generated data across a broad range of downstream ReID tasks, including direct transfer, supervised fine-tuning, and unsupervised domain adaptation. Second, we adapt our pipeline to the identity expansion scenario, i.e., there exist few labeled identities for training under the target scenario. Our method consistently achieves state-of-the-art results across these settings, demonstrating its superiority and applicability.", + "bbox": [ + 511, + 516, + 908, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 805, + 653, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Person Re-identification", + "text_level": 1, + "bbox": [ + 513, + 830, + 733, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Person re-identification [21, 24, 26, 31, 33, 47, 52, 60] aims to retrieve pedestrians of the same identity from different camera views. This task presents challenges in distinguish", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23375", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing between similar identities and retrieving same-identity samples with diversity cross-camera variations [30, 32, 52]. With the rise of deep learning[14, 25, 40, 46], various methods are applied to ReID tasks, and these models achieve excellent results when trained and tested on the same datasets. However, when transferred to other datasets, they tend to suffer performance degradation due to distributional differences. There is an urgent need for a dataset with a wide distribution to enhance the generalization ability of ReID models across diverse scenarios and environments. Unfortunately, collecting a labeled dataset based on test scenarios is costly. For example, the large-scale ReID dataset MSMT17[44], consisting of 4,101 identities captured from 15 cameras, required two months of annotation work by three labelers. Moreover, it also leads to concerns about privacy leakage of pedestrians, which limits the application of ReID models trained on these datasets.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Synthesized ReID Dataset", + "text_level": 1, + "bbox": [ + 89, + 361, + 326, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the challenges associated with real-world data collection, a few synthesized datasets for ReID have emerged as a promising solution.", + "bbox": [ + 89, + 383, + 482, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Engine-based Methods. These datasets[1, 22, 41-43, 48, 49, 57] are typically generated using computer graphics techniques or simulation engines, such as Unity [36] and MakeHuman [10], which enable the creation of virtual labeled ReID data. SyRI[1] created pedestrians in a virtual scene, introducing cross-camera diversity by varying lighting and shooting perspectives. PersonX[41] utilized virtual hand-crafted 3D models as pedestrians and adjusted the angle and scene of the shot to acquire images. RandPerson[42] generated random UV texture maps to create 3D clothing person models. UnrealPerson[57] sampled real clothing to form the virtual pedestrian models and obtained images in virtual scenarios. ReID models trained on these virtual datasets can be effectively generalized to real datasets. However, they are limited by the manually designed nature of the virtual pedestrians, hard identities, scenes, lighting conditions, and video shooting, which increases synthesis costs and restricts scalability. Furthermore, the synthetic pedestrians and scenes still exhibit significant discrepancies compared to real-world data, affecting their transferability to real datasets.", + "bbox": [ + 89, + 431, + 482, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based Methods. With the development of generative models, including generative adversary network (GAN) [13] and diffusion[37], some methods [7, 19, 50] leveraged them to synthesize realistic images. GCL [7] and Pose-dIVE [19] modify pedestrian poses and synthesize diverse viewpoint images using GANs and diffusion models, respectively, which effectively expand the real-world datasets. However, these models require training on real ReID datasets with a limited number of cameras, which makes the distribution of the model restricted to insuffi", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cient cross-camera variations as shown in the right of Figure 1. Moreover, while the models learn realistic styles, they inadvertently embed identity-related features, making the generated images highly consistent with the real identities present in real-world datasets, which raises privacy concerns[17, 27-29, 35, 54, 55, 62]. MALS[50] is a synthetic text-based person-retrieval dataset that contains 1,510,330 image-text pairs. It used textual descriptions of real pedestrians to guide the off-the-shelf diffusion models to synthesize person images. Although its identities are virtual, each pedestrian has only a single image and lacks cross-camera diversity, such as poses, camera viewpoints, scenes, and lighting condition variations. What's more, it lacks the ability to synthesize diverse virtual identities with hard identities, which is not suitable for person ReID tasks.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Data Generation Pipeline", + "text_level": 1, + "bbox": [ + 511, + 333, + 751, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. What makes a person dataset?", + "text_level": 1, + "bbox": [ + 511, + 359, + 784, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the person re-identification (ReID) task, the dataset used for supervised model training is generally denoted as $D = \\{X_{1}, X_{2}, \\dots, X_{P}\\}$ , where $P$ is the number of identities, $X = \\{x_{1}, x_{2}, \\dots, x_{K}\\}$ is a collection of images $x$ of the same identity, which denoted as an intra-identity (intra-ID) group, and $K$ is the number of samples in each identity.", + "bbox": [ + 511, + 381, + 903, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In real-world surveillance systems, there are numerous pedestrian identities, each appearing in different scenes captured by different cameras. Collecting real pedestrian data presents significant challenges, primarily due to the high cost of manual annotation and privacy concerns. To address these challenges, recent studies propose synthetic methods to generate pedestrians images to improve ReID models. However, they struggle to fully satisfy or scale the following three key factors: virtual identity diversity, cross-camera variations, and camera-realistic styles. In the following, we will introduce the flexible identity generator, cross-camera diversification, and style realization to address the above challenges.", + "bbox": [ + 511, + 474, + 903, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Flexible Identity Generator", + "text_level": 1, + "bbox": [ + 511, + 680, + 759, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our first task is to generate diverse virtual identities in the dataset, which plays a critical role in aiding ReID models to learn discriminative features. In particular, hard identities, which share similar appearances, enable ReID models to capture fine-grained features. To flexibly generate new random identities and hard identities, we present the flexible identity generator which is upgraded based on the latent diffusion model [37]. Next, we will elaborate on the training and generation process.", + "bbox": [ + 511, + 703, + 903, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training of Identity Generator. Previous learning-based methods are troubled by the limitation of generating pedestrian images that are strictly consistent with real identities, which raises concerns about privacy exposure. To", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23376", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/05e7a99d11b2112ad3a5bc48fc8f6ccfc670b3896371a50e6e9c977c3694e8bb.jpg", + "image_caption": [ + "Figure 3. Illustration of pedestrian generation pipeline VIPerson. The flexible identity generator flexibly generates new random identities and hard identities. Afterward for each identity, cross-camera diversification generates intra-identity groups that contain images with rich cross-camera variations. Finally, style realization effectively reduces the style gap between generated and real images, while preserving virtual identities." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 898, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "solve this, we train the generator on a large-scale dataset, MALS[50], which contains 1.5 million person images of virtual identities. In this way, the generator learns a rich distribution of virtual identities, enabling it to not only generate virtual identities but also to expand the diversity of identities with varying characteristics. Note that we empirically adjust all training images to the same body pose with IP-adapter[51, 56] beforehand to mitigate training difficulties caused by pose variations.", + "bbox": [ + 88, + 376, + 483, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our flexible identity generator is based on the latent diffusion model[3, 37] (LDM) conditioned on the identity description. The LDM is a generative model that learns to reverse a gradual noising process in the latent space to generate data. The architecture of our flexible identity generator is a modified UNet [38] that includes residual and attention blocks. To reduce the computation during training, we first compress the input data into a latent space through an autoencoder, denoted as $z = AE(x)$ . During the forward diffusion process, at each timestep $t$ , the latent data $z$ is transformed into $z_{t}$ by adding Gaussian noise $\\varepsilon$ . During training, the model learns to predict and reverse the noise at time step $t$ based on the current estimate $z_{t}$ and the additional condition. As the condition, we employ the text embedding of the pedestrian image description $d$ . To provide an accurate $d$ for each pedestrian image, we generate the text description containing four main dressing attributes, i.e. genders/ages, up dress, down dress, and shoes, using BLIP[20, 23], a powerful model for visual question answering. Then we extract the text embedding of $d$ with the text encoder $\\varphi$ of CLIP[34], and inject the text embedding $\\varphi(d)$ into the cross-attention layer of our flexible identity generator. The training loss is defined as:", + "bbox": [ + 91, + 515, + 483, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L D M} = \\mathbb {E} _ {z, \\varepsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\varepsilon - \\varepsilon_ {\\theta} (z _ {t}, t, \\varphi (d)) \\| _ {2} ^ {2} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 877, + 483, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation of New Random Identity. After our flexible identity generator learns the diverse distribution of virtual pedestrians, we leverage a Gaussian-sampled embedding as the condition in the diffusion model's reverse denoising process to generate new random identities. Specifically, we sample a random embedding $c_{r} \\sim \\mathcal{N}(0,1)$ . This random embedding serves as a unique identifier that guides the generation process, ensuring diversity in the synthesized outputs. The reverse denoising process is modified to incorporate the $c_{r}$ at each step $t$ . The iterative update for the reverse process is given by:", + "bbox": [ + 511, + 375, + 906, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t - 1} = \\mu_ {\\theta} (z _ {t}, t, c _ {r}) + \\sigma_ {t} \\zeta_ {t} \\\\ = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\varepsilon_ {\\theta} \\left(z _ {t}, t, c _ {r}\\right)\\right) + \\sigma_ {t} \\zeta_ {t}, \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 551, + 906, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ and $\\sigma_{t}$ are parameters to control the noise schedule, and $\\zeta_{t}$ is an additional noise vector to maintain stochasticity. At the final iteration, the estimated $z_{0}$ is mapped back to the pixel space using a pre-trained decoder, resulting in a high-quality image that embodies the new random identity defined by the $c_{r}$ .", + "bbox": [ + 511, + 613, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation of Hard Identity. To generate hard identities, we propose leveraging dropout text embedding $c_{d}$ as conditions in the diffusion process. Given an identity image $I$ , our objective is to generate a hard identity $I'$ that shares a similar appearance to $I$ . We first use BLIP[20, 23] to generate the text descriptions $d_{I}$ of $I$ . The description is then encoded into 512-dimensional text embeddings using CLIP. To introduce hard identities, we apply a dropout strategy by randomly setting some of the dimensions of the text embedding to zero, resulting in $c_{d} = dropout(\\varphi(d_{I}))$ . This process randomly alter subtle identity-related features while preserving core semantic information. $c_{d}$ is then used as conditional inputs for the identity generator to generate", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23377", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/360ed29545afbd4e472c2ea63be7506a968528ecf10552a57883c2d6d8413f39.jpg", + "table_caption": [ + "Table 1. Details of generated datasets. The \"Generalization Ability\" represents the rank-1 performance of direct transfer results on MSMT17 with BoT[26] model trained on these synthesized datasets. MALS* is a synthetic text-based person-retrieval dataset, which is challenging to generalize to real-world person ReID tasks." + ], + "table_footnote": [], + "table_body": "
TypeDatasetsVirtual IdentityCamera-realisticCross-camera variationsHard Identity#Identity#BBoxesGeneralization Ability
Engine-basedSyRI [1]YesNoLimitedNo10056,0007.8
PersonX [41]YesNoLimitedNo4109,8407.6
RandPerson [42]YesNoLimitedNo8,000132,14514.5
UnrealPerson [57]YesNoLimitedLimited3,000120,00026.8
Learning-basedMALS* [50]YesNoLimitedNo1,510,3301,510,3300.2
GCL [7]NoYesLimitedLimited---
Pose-dIVE [19]NoYesLimitedLimited---
VIPersonYesYesScalableScalable5,000114,73832.5
", + "bbox": [ + 135, + 132, + 864, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "new, hard identities with subtle differences in appearance. The iterative process can be formally described as:", + "bbox": [ + 89, + 268, + 480, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\varepsilon_ {\\theta} \\left(z _ {t}, t, c _ {d}\\right)\\right) + \\sigma_ {t} \\zeta_ {t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 306, + 482, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The \"flexible\" in the flexible identity generator lies in the ability to flexibly control conditions during inference, enabling the generation of versatile and diverse identities with varying characteristics. We provide more ablation studies on designing hard identities in section B of supplementary material.", + "bbox": [ + 89, + 356, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Cross-camera Diversification", + "text_level": 1, + "bbox": [ + 89, + 455, + 349, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Enhancing cross-camera variations within each identity is crucial for improving the generalization ability of ReID models across different camera views. However, previous synthetic methods struggle to scale up cross-camera variations, due to the high manual effort required in designing diverse scenes or limited variation in training data. In this subsection, we aim to expand the intra-identity (intra-ID) group $X$ for each identity with scalable cross-camera variations, specifically in terms of poses and scenes.", + "bbox": [ + 89, + 477, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Collection. First, we collect real-world pedestrian images, ensuring a diverse range of poses and scenes that reflect the variations encountered in different camera setups. Unless otherwise specified, our poses and scenes come from LUPerson [12], a large-scale unlabeled pedestrian dataset with over 4.18 million images collected across 46,260 scenes.", + "bbox": [ + 89, + 614, + 483, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Variations. Second, we extract person poses in the collected images with openpose[5, 6, 39, 45, 56]. Then we utilize ControlNet-based IP-adapter[51, 56] to modify each virtual identity's pose with the randomly collected poses as the condition. However, significant pose variations may lead to identity degradation. To enhance identity consistency, we generate a corresponding identity description for each individual using BLIP. This description, along with the collected pose, is then used as control conditions for the IP-Adapter to guide the generation of the new image.", + "bbox": [ + 89, + 719, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene Variations. Lastly, we propose segment-and-paste strategy to modify the scene of images. For the collected", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "real-world images, we first apply the inpainting diffusion technique [53] to remove the original pedestrian, creating clean scene images. We then utilize the semantic segmentation method[8] to segment our generated pose-changed persons and patch them into the scene images.", + "bbox": [ + 511, + 268, + 906, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Style Realization", + "text_level": 1, + "bbox": [ + 511, + 357, + 679, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera-realistic styles, such as illumination, motion blur, camera bias, and environmental conditions, provide crucial contextual information that reflect real-world scenarios, playing a key role in improving the transferability of ReID models. Engine-based synthetic images inherently lack realistic styles, as they are created using virtual engines that do not capture real-world environmental nuances. Previous learning-based generative models learn realistic styles while embedding identity-related features, leading to generated images highly consistent with the real identities of the real-world ReID dataset, which raises privacy concerns. To address these, unlike learning the realistic distribution during training, we propose a post-processing approach, leveraging the neural style transfer method [9] to inject the style. Specifically, we randomly select a pedestrian image from LUPerson as the reference image, and leverage [9] to separate the style and content of the real-world image. Then, for each generated virtual image, we transfer the identity-agnostic style from the real-world image to the virtual image while preserving its original virtual identity.", + "bbox": [ + 511, + 380, + 906, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Summary of the VIPerson Dataset", + "text_level": 1, + "bbox": [ + 511, + 696, + 815, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The entire data generation pipeline is illustrated in Figure 3, which can generate a ReID dataset composed entirely of virtual identities without manual annotation. We generate the virtual dataset VIPerson, which serves as the primary dataset in all our experiments unless otherwise specified. Refer to Figure 3 and section A of supplementary material for examples of generated images. The detailed parameters of the datasets are shown in Table 1. Furthermore, we extend our approach to incorporate scenario-specific patterns, generating VIPerson* with available pose, scene and style information from target scenarios to enhance ReID performance in the targeted applications.", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23378", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9d01b3e8a6ad9386754c23bdd82f51a35004e611debd3857c2f4ba4adee31368.jpg", + "image_caption": [ + "Figure 4. Effect of different numbers of identities. \"Ours w/o Hard\" does not contain hard identities generated by $c_d$ ." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 294, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/609ba3a32a9901921eca21f31e3fe2d5b18f108ed27de4780d72a2dbc9d38894.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 90, + 480, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Analysis of Key Factors in Data Synthesis", + "text_level": 1, + "bbox": [ + 89, + 292, + 464, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The synthesized pedestrian data is a fundamental component of downstream ReID tasks. In this study, we conduct an in-depth experimental analysis of several key factors in the generated dataset VIPerson: virtual identity diversity (including identity diversity and hard identity), cross-camera variations (including intra-ID group scale, pose variations, and scene variations), and camera-realistic style. We employ direct transfer performance on real datasets with BoT[26] to assess the quality of generated data, as direct transfer serves as the foundation for all other tasks.", + "bbox": [ + 88, + 319, + 483, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Virtual Identity Diversity", + "text_level": 1, + "bbox": [ + 89, + 500, + 321, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Identity Diversity. We explore the effect of the identity diversity by changing identity scale as shown in Figure 4. As the number of identities increases from 1000 to 3000, the rank-1 on MSMT17 and Market1501 has increased rapidly. It demonstrates that our flexible identity generator can generate scalable identities with diverse information, facilitating the model to better discriminate between unseen persons in the test set. However, as the identity scale increases further, the model performance grows slowly, especially on Market1501. This limitation may contribute to the scaling law, which is summarised in some studies[11, 15, 18] in many tasks: Even with an infinite amount of data, the learning capacity of ReID model is still constrained by other factors such as model capacity. More analysis between data scale and model size refers to section D in the supplementary material.", + "bbox": [ + 91, + 521, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hard Identity. We explore the impact of hard identities on model performance as shown in Figure 4. For example, at an identity count of 5000, 'ours' outperforms 'ours-w/o Hard' on both test datasets. This suggests that hard identities in datasets allow the ReID model to learn to focus on more detailed discriminative features, improving the ability to distinguish between similar persons. In contrast to previous synthetic datasets that require manual design of hard identities, our approach generates hard identities automati-", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae141a1f85e1ce399c8920b3f9c3d5125d49054cefd6bf615db886288919e8c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 715, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55a7e2fd7353f6ae8e17ef713ae5f915ac90ab417c39242b7b90471746d352f7.jpg", + "image_caption": [ + "Figure 5. Effect of different Intra-ID group scales." + ], + "image_footnote": [], + "bbox": [ + 718, + 90, + 903, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f01b1a54a66d3d0f2a5fef843851fc9b34142c05005725c5221892d607bbd97.jpg", + "image_caption": [ + "Figure 6. Ablation study of modules in VIPerson. 'PV' indicates pose variations, 'SV' refers to scene variations, and 'SR' is style realization. Each point reports the rank-1 performance." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 901, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cally, making the process more cost-effective and scalable.", + "bbox": [ + 511, + 503, + 898, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Cross-camera Variations", + "text_level": 1, + "bbox": [ + 511, + 530, + 740, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intra-ID Group Scale. Cross-camera variations are positively correlated with the scale of intra-ID group. This is because as the intra-ID group increases, there is a larger variety of viewpoints, poses, and scenes for each identity, which is introduced via our random sampling process. Here, we demonstrate how the scale of the intra-ID group affects the performance of the ReID model. As the results shown in Figure 5, with the increment of our intra-ID group scale, the performance continues to increase on both test datasets. Besides, our VIPerson is consistently superior to UnrealPerson, which demonstrates that our cross-camera variations are more diverse than UnrealPerson. Benefited from our proposed cross-camera diversification that introduces scalable and diverse poses and scenes without relying on manual design, we can expand the intra-ID group to any scale for better performance at a low cost.", + "bbox": [ + 511, + 551, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose Variations. To investigate the impact of pose variations, we minimise the variation of pose introduced through cross-camera diversification, as shown in the 'SV+SR' column of Figure 6. The rank-1 performance drops from $32.5\\%$ to $20.5\\%$ on MSMT17, and from $66.5\\%$ to $39.8\\%$ on Market1501. In the absence of pose diversity, the model may overfit to specific pose-related features, which hin", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23379", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ders its ability to generalize across different viewing angles. Since our pose set is randomly collected from real pedestrian images, the poses of generated persons are natural and varied. Our results demonstrate that pose variations in training data plays a crucial role in improving the model's robustness and performance in real-world scenarios, where pedestrians are often captured from various angles and perspectives.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Variations. As shown in the column 'PV', Without scene variations, the performance is limited to $0.9\\%$ on MSMT17 and $11.7\\%$ on Market1501. Introducing scene variations significantly boosts performance to $7.5\\%$ on MSMT17 and $33.6\\%$ on Market1501 as shown in the column 'PV+SV', indicating that scene variations are crucial for the model's effectiveness. Since the background in each image of the same pedestrian varies in the test data, incorporating diverse backgrounds into the training set can substantially enhance performance. Previous approaches, which rely on both synthetic and real datasets sampled from sequential camera videos, are limited in terms of scene variations within each intra-ID group. It would be costly for them to capture more videos to expand the variation in scenes. In contrast, our segment-and-paste strategy allows the Intra-ID group to exhibit richer background variations with a significantly lower cost. More analysis of scenes refer to the section E in supplementary material.", + "bbox": [ + 91, + 212, + 483, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Camera-realistic Style", + "text_level": 1, + "bbox": [ + 89, + 496, + 299, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Style Realization. As shown in column 'PV+SV+SR' of Figure 6, introducing camera-realistic styles significantly improves rank-1 from $33.6\\%$ to $66.5\\%$ on Market1501 and from $7.5\\%$ to $32.5\\%$ on MSMT17. This highlights the importance of style realization, as it bridges the gap between synthetic and real-world images. By transferring illumination, motion blur, camera bias, and environmental conditions from real reference images, our style realization process makes the synthetic images more consistent with real-world data distributions. This not only enhances model performance but also improves generalization when deployed in diverse real-world environments.", + "bbox": [ + 89, + 518, + 482, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Downstream ReID Tasks", + "text_level": 1, + "bbox": [ + 89, + 715, + 323, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 742, + 308, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Details of Synthetic Data. Existing synthetic ReID datasets include SyRI[1], PersonX[41], RandPerson[42], and UnrealPerson[57], and their details are shown in Table 1. Our dataset is named VIPerson in this paper. We provide synthetic datasets for two scenarios, including the dataset without target scenario information (denoted as 'VIPerson') and the datasets based on target scenario information (denoted as 'VIPerson*'). We test the model on MSMT17 [44] and Market1501 [59]. For VIPerson, pose,", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b71ea0572362015b0f34f7ad40ba76d2dc032b46ab8dff6fea57341b2bedfba4.jpg", + "table_caption": [ + "Table 2. Direct transfer results with BoT[26]. The fully-supervised results are shown in italics, the best direct transfer results are indicated in bold, and the second-best results are underlined. More experiments on different ReID models are provided in section C of supplementary material." + ], + "table_footnote": [], + "table_body": "
Dataset TypeTraining DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealMSMT1773.949.956.229.8
Market15019.23.094.686.3
SyntheticMALS0.20.64.91.5
SyRI7.81.615.24.9
PersonX7.62.436.815.5
RandPerson14.54.855.929.9
UnrealPerson26.89.964.037.2
VIPerson32.510.366.538.3
VIPerson*34.310.867.537.6
", + "bbox": [ + 531, + 160, + 885, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9faeae57951eca6e0d6ebb2b55a53d3b6ddddbeb6737d9a332ff9d68cc27cbd1.jpg", + "table_caption": [ + "Table 3. Supervised fine-tuning results with BoT[26] method." + ], + "table_footnote": [], + "table_body": "
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet72.542.491.476.8
MSMT17--93.782.5
Market150174.144.5--
SyntheticMALS28.114.474.854.7
UnrealPerson74.546.094.084.7
VIPerson76.551.594.085.3
VIPerson*76.751.295.186.9
", + "bbox": [ + 522, + 324, + 893, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "background, and style information is consistently from LUPerson [12] across all tested datasets. VIPerson consists of 5000 identities including hard identities. For VIPerson*, when testing on MSMT17, we obtain pose, scene and style information from MSMT17 training set. When testing on Market1501, we obtain pose and scene information from LUPerson and MSMT17, and transfer style from Market1501 training set.", + "bbox": [ + 511, + 455, + 905, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Details. We conduct our experiments and analysis primarily using ResNet50[14] as the backbone model. Specifically, for the direct transfer experiment, we adopt BoT[26] and resize the image to $256 \\times 128$ and set the batch size as 128. We adopt the Adam optimizer for training, with the base learning rate set to $3.5 \\times 10^{-4}$ and decreased by a factor of 0.1 at the $10^{th}$ and $40^{th}$ epochs. Details about fine-tuning and domain adaptation experiments refer to section F in supplementary material.", + "bbox": [ + 511, + 579, + 905, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Direct Transfer", + "text_level": 1, + "bbox": [ + 513, + 726, + 669, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the quality of generated data, we train the models only on generated data and test them on real data. As shown in Table 2, our VIPerson surpass all other datasets. Unlike training data from real-world, VIPerson consists entirely of virtual identities, thus avoiding privacy concerns. Nevertheless, VIPerson still exceeds Market1501 by $23.3\\%$ on MSMT17 and outperforms MSMT17 by $10.3\\%$ on Market1501. Compared to the previous synthetic datasets, our VIPerson significantly improves performance, as it achieve three key factors in synthetic person data as analysis in Sec-", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23380", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/340ee758020e16a1e17defb004bd47670cc5540f81afef741d742a369f549bc0.jpg", + "table_caption": [ + "Table 4. Unsupervised domain adaptation results with ISE[58]." + ], + "table_footnote": [], + "table_body": "
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet64.735.094.084.7
MSMT17--93.884.2
Market150160.431.2--
SyntheticRandPerson59.030.293.483.8
UnrealPerson65.936.194.586.3
VIPerson70.439.594.986.4
", + "bbox": [ + 99, + 104, + 470, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion 4, allowing the ReID model to generalize better to unseen scenarios.", + "bbox": [ + 89, + 222, + 482, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, our cross-camera diversification and style realization are flexible to adapt to the target scenarios. Given scene and style information from the target scenario, we leverage them to develop VIPerson*, which further enhances performance on target datasets. This makes our method more practical and adaptable for real-world applications under diverse scenarios.", + "bbox": [ + 89, + 251, + 483, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Supervised Fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 364, + 307, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When sufficient labeled data is available, pretraining the model on synthetic data and fine-tuning it on target datasets can enhance performance. As shown in Table 3, with pretrained on our datasets, the rank-1 performance reaches $76.7\\%$ on MSMT17 and $95.1\\%$ on Market1501, suppressing all the other datasets. These results demonstrate the diversity, strong feature representation, and ample scale of our dataset, underscoring its value as a powerful resource for enhancing fine-tuning performance.", + "bbox": [ + 89, + 387, + 483, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Unsupervised Domain Adaptation", + "text_level": 1, + "bbox": [ + 89, + 532, + 387, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In most application scenarios, pedestrian data is usually unlabeled due to the limitations of annotation costs. In these cases, leveraging other labeled datasets for pretraining and fine-tuning on the unlabeled data can also improve performance. From Table 4, we see that the model pre-train on our datasets suppresses all the other datasets, achieving $70.4\\%$ on MSMT17 and $94.9\\%$ on Market1501. Our VIPerson effectively enhance the performance without relying on any manual annotations, which underscores its significance for reducing annotation dependency in real-world applications.", + "bbox": [ + 89, + 553, + 483, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Identity Expansion Scenarios", + "text_level": 1, + "bbox": [ + 89, + 712, + 352, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As rising privacy concerns, we consider a scenario with few label identities, where most individuals are unwilling to share their personal image as training data, leading to a scarcity of labeled identities for training ReID models. Fortunately, we can add generated data to enhance the identity diversity. To this end, we simulate environments where only a small fraction of labeled data (20 or 40 identities) is available, exploring how ReID models perform with limited labeled identities and expended synthetic data.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 7, when only 20 real identities are available, adding real data from other domains (MK) only", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10b788ab478551c3d2c5348ff66a4289980f9e5c5ae02ae172d79d33178d8915.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 715, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce09a8d147a203610f97064c6f472475047543475f1000659456f2ba8aeab11b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 90, + 898, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0cd81c33b169fdb4a3ecbbc44dbc0c10eeb2699452dba4eeba48f1a32cb0f35.jpg", + "image_caption": [ + "Market1501" + ], + "image_footnote": [], + "bbox": [ + 517, + 202, + 715, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5769d1002c807a0dd3531487f99fc4604edb7852e00369e1d2485b532d30428b.jpg", + "image_caption": [ + "Figure 7. Results under identity expansion scenarios. 'R' indicates training with real images from the testing scenarios, 'MK' refers to training set of Market1501, 'MS' corresponds to training set of MSMT17, 'UP' denotes UnrealPerson, and 'VIP' represents our VIPerson dataset." + ], + "image_footnote": [], + "bbox": [ + 720, + 202, + 900, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "improves rank-1 by $15\\%$ on MSMT17 test set. Although collected from the real world, it is constrained by limited identity and cross-camera variations, which hinders its ability to effectively enhance model performance. As a comparison, adding VIPerson significantly improves rank-1 performance on MSMT17 by $30.4\\%$ , which effectively bridge the gap caused by the lack of real-world data. These results underscore the practical applicability of our approach in scenarios with few label identities, making it possible to train effective ReID models with minimal privacy exposure to real identity. In environments where privacy is a critical concern, such as surveillance systems and smart cities, this approach can be a game-changer for identity verification and tracking systems.", + "bbox": [ + 511, + 440, + 906, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 675, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce VIPerson, a novel pedestrian generation pipeline for flexibly creating virtual identities for the person ReID task to address privacy concerns and annotations consumptions in real ReID dataset. Our work identifies three key factors that enhance the effectiveness of the synthesized dataset, offering valuable insights for future directions in data synthesis. Moreover, we validated the scaling law, which shows that performance improves with an increase in the scale of synthesized data and model parameters. However, we also observed a performance bottleneck at a certain data scale, which suggests that future work should focus on how to effectively leverage synthetic data to enhance ReID models.", + "bbox": [ + 511, + 703, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23381", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Slawomir Bak, Peter Carr, and Jean-Francois Lalonde. Domain adaptation through synthesis for unsupervised person re-identification. In Proceedings of the European conference on computer vision (ECCV), pages 189–205, 2018. 1, 3, 5, 7", + "[2] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1", + "[3] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 4", + "[4] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, pages 1092-1104. PMLR, 2021. 1", + "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 5", + "[6] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5", + "[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2004–2013, 2021. 2, 3, 5", + "[8] Xiangguang Chen, Ye Zhu, Yu Li, Bingtao Fu, Lei Sun, Ying Shan, and Shan Liu. Robust human matting via semantic guidance. In Proceedings of the Asian Conference on Computer Vision, pages 2984-2999, 2022. 5", + "[9] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 5", + "[10] Makehuman community. Makehuman. http://www.makehumancommunity.org/. 2024.3", + "[11] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 6", + "[12] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceed" + ], + "bbox": [ + 93, + 213, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750-14759, 2021. 2, 5, 7", + "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7", + "[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 6", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[17] Erin Illman and Paul Temple. California consumer privacy act. The Business Lawyer, 75(1):1637-1646, 2019. 1, 3", + "[18] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 6", + "[19] Inès Hyeonsu Kim, JoungBin Lee, Woojeong Jin, Soowon Son, Kyusun Cho, Junyoung Seo, Min-Seop Kwak, Seokju Cho, JeongYeol Baek, Byeongwon Lee, et al. Pose-dive: Pose-diversified augmentation with diffusion model for person re-identification. arXiv preprint arXiv:2406.16042, 2024. 2, 3, 5", + "[20] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 31-41, Toronto, Canada, 2023. Association for Computational Linguistics. 4", + "[21] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6729–6738, 2021. 2", + "[22] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In Proceedings of the 29th ACM international conference on multimedia, pages 3115-3123, 2021. 3", + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4", + "[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI conference on artificial intelligence, pages 1405–1413, 2023. 2", + "[25] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23382", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 3", + "[26] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 5, 6, 7", + "[27] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Jihong Guan, and Shuigeng Zhou. Fine-grained zero-shot object detection. arXiv preprint arXiv:2507.10358, 2025. 3", + "[28] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024.", + "[29] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21203-21214, 2025. 3", + "[30] Yi-Xing Peng, Jile Jiao, Xuetao Feng, and Wei-Shi Zheng. Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia, 25:2393–2403, 2022. 3", + "[31] Yi-Xing Peng, Yuanxun Li, and Wei-Shi Zheng. Revisiting person re-identification by camera selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5): 2692-2708, 2023. 2", + "[32] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, and Wei-Shi Zheng. Generalized intra-camera supervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4516-4527, 2023. 3", + "[33] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, Qize Yang, Jingke Meng, Xihan Wei, and Wei-Shi Zheng. Person deidentification: A variation-guided identity shift modeling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29331-29341, 2025. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4", + "[35] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1, 3", + "[36] John Ricciello. John ricciello sets out to identify the engine of growth for unity technologies (interview). VentureBeat. Interview with Dean Takahashi. Retrieved January, 18 (3), 2015. 3", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4", + "[39] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 5", + "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3", + "[41] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 608-617, 2019. 1, 3, 5, 7", + "[42] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In Proceedings of the 28th ACM international conference on multimedia, pages 3422–3430, 2020. 1, 3, 5, 7", + "[43] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4900–4909, 2022. 3", + "[44] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 1, 3, 7", + "[45] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 5", + "[46] Xiao-Ming Wu, Dian Zheng, Zuhao Liu, and Wei-Shi Zheng. Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17055-17064, 2023. 3", + "[47] Yuqiao Xian, Yi-Xing Peng, Xing Sun, and Wei-Shi Zheng. Distilling consistent relations for multi-source domain adaptive person re-identification. Pattern Recognition, 157: 110821, 2025. 2", + "[48] Suncheng Xiang, Yuzhuo Fu, Guanjie You, and Ting Liu. Unsupervised domain adaptation through synthesis for person re-identification. In 2020 IEEE international conference on multimedia and expo (ICME), pages 1–6. IEEE, 2020. 3", + "[49] Suncheng Xiang, Dahong Qian, Mengyuan Guan, Binjie Yan, Ting Liu, Yuzhuo Fu, and Guanjie You. Less is more: Learning from synthetic data with fine-grained attributes for person re-identification. ACM Transactions on Multimedia Computing, Communications and Applications, 19(5s): 1-20, 2023. 3", + "[50] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 2, 3, 4, 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4, 5", + "[52] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 2, 3", + "[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 5", + "[54] Delong Zhang, Yixing Peng, Ancong Wu, and Wei-shi Zheng. Privacy-preserving face recognition with adaptive generative perturbations. In International Conference on Pattern Recognition, pages 210-227. Springer, 2024. 3", + "[55] Delong Zhang, Yi-Xing Peng, Xiao-Ming Wu, Ancong Wu, and Wei-Shi Zheng. Pixelfade: Privacy-preserving person re-identification with noise-guided progressive replacement. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6326–6334, 2024. 3", + "[56] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5", + "[57] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11506–11515, 2021. 1, 3, 5, 7", + "[58] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7369–7378, 2022. 8", + "[59] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 1, 7", + "[60] Wei-Shi Zheng, Junkai Yan, and Yi-Xing Peng. A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[61] Zhedong Zheng, Liang Zheng, and Yi Yang. Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In Proceedings of the IEEE international conference on computer vision, pages 3754–3762, 2017. 1", + "[62] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo, Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3" + ], + "bbox": [ + 91, + 92, + 480, + 852 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23384", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_model.json b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_model.json new file mode 100644 index 0000000000000000000000000000000000000000..838f17ce62e52e491bbf81c329fc96ec9cccde7e --- /dev/null +++ b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "VIPerson: Flexibly Generating Virtual Identity for Person Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.18, + 0.411, + 0.217 + ], + "angle": 0, + "content": "Xiao-Wen Zhang\\(^{1,3*}\\) \nZhi Ouyang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.18, + 0.586, + 0.217 + ], + "angle": 0, + "content": "Delong Zhang\\(^{1,3*}\\) Jingke Meng\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.6, + 0.18, + 0.761, + 0.217 + ], + "angle": 0, + "content": "Yi-Xing Peng\\(^{1,3}\\) Wei-Shi Zheng\\(^{1,2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.217, + 0.808, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Sun Yat-sen University, China;" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.234, + 0.672, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Peng Cheng Laboratory, Shenzhen, China;" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.251, + 0.888, + 0.27 + ], + "angle": 0, + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China." + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.272, + 0.815, + 0.287 + ], + "angle": 0, + "content": "{zhangxw75,zhangdlong5,pengyx23,}@mail2.sysu.edu.cn wszheng@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.338 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Person re-identification (ReID) is to match the person images under different camera views. Training ReID models necessitates a substantial amount of labeled real-world data, leading to high labeling costs and privacy issues. Although several ReID data synthetic methods are proposed to address these issues, they fail to generate images with new identities or real-world camera styles. In this paper, we propose a novel pedestrian generation pipeline, VIPerson, to generate camera-realistic pedestrian images with flexible Virtual Identities for the Person ReID task. VIPerson focuses on three key factors in data synthesis: (I) Virtual identity diversity: Enhancing the latent diffusion model with our proposed dropout text embedding, we flexibly generate random and hard identities. (II) Scalable cross-camera variations: VIPerson introduces scalable variations of scenes and poses within each identity. (III) Camera-realistic style: Adopting an identity-agnostic approach to transfer realistic styles, we avoid privacy exposure of real identities. Extensive experimental results across a broad range of downstream ReID tasks demonstrate the superiority of our generated dataset over existing methods. In addition, VIPerson can be adapted to the identity expansion scenario, which widens the application of our pipeline. The dataset and code of VIPerson are available at https://isee-laboratory.github.io/VIPerson/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.484, + 0.864 + ], + "angle": 0, + "content": "The person re-identification (ReID) task aims to match pedestrians from different camera views. With the rise of deep learning[52], ReID methods have been developed based on real annotated pedestrian data [44, 59]. These real-world ReID datasets are often limited in both scale and" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.323, + 0.86, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.494, + 0.907, + 0.549 + ], + "angle": 0, + "content": "Figure 1. Comparison of synthetic data. Left: Images of engine-based methods have a style gap compared with real images. Right: Learning-based methods expose privacy by generating images that match real identities in real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.747 + ], + "angle": 0, + "content": "diversity, primarily due to the high cost of data collection and manual annotation. Moreover, pedestrian images containing sensitive information often expose real identities, raising serious privacy concerns. Strict data privacy regulations [2, 4, 17, 35], such as EU's General Data Protection Regulation, restrict the ability to gather and share sensitive visual data. This further impedes the collection of public ReID datasets, with some, like DukeMTMC [61], even taken down. The lack of high-quality annotated data without privacy concerns significantly limits the performance and generalization ability of ReID models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address these challenges, recent studies propose to synthesize pedestrian images to improve ReID models. These methods can be divided into two research streams. Engine-based methods [1, 41, 42, 57] are developed to leverage simulation engines to synthesize virtual pedestrian images. However, these synthetic images exhibit distinct virtual characteristics (see 'Left' in Figure 1), resulting in a style gap compared to real-world images that showcase camera-realistic styles (e.g. illumination, motion blur, camera bias, and environmental conditions). These synthetic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23374" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.094, + 0.475, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.405 + ], + "angle": 0, + "content": "Figure 2. Image examples from our VIPerson dataset. Our data pipeline can generate virtual identities, containing random identities, hard identities, and intra-identity (intra-ID) groups that contain camera-realistic images with rich cross-camera variations. More images are provided in section A of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.581 + ], + "angle": 0, + "content": "data hinder ReID models' transferability to real-world scenarios. Learning-based synthetic approaches [7, 19, 50], propose using generative models[13, 16], to generate realistic pedestrian images. However, these models face privacy concerns as they are only capable of generating images that are consistent with the real identities present in real-world ReID datasets (see 'Right' in Figure 1), which limits their application in privacy-sensitive scenarios. Moreover, both approaches struggle with limited cross-camera variations (e.g. poses and scenes) due to constraints imposed by costs of scene design or limited variations in training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we aim to propose a novel pipeline to generate a camera-realistic pedestrian dataset composed entirely of virtual identities for the person ReID task. However, achieving this is non-trivial due to three core challenges in pedestrian data synthesis: (1) How to flexibly generate diverse virtual identities, containing new random identities and hard identities? New random identities exhibit diverse appearances, while hard identities, often referred to as hard samples, share similar visual characteristics despite belonging to different individuals. Both are crucial for enhancing identity diversity, aiding ReID models to learn discriminative features across a broad range of individuals. This necessitates a flexible generative model for skillfully controlling the generation of diverse identities. (2) How to introduce scalable cross-camera variations within each identity? Abundant cross-camera variations of generated images can facilitate the generalization ability of ReID models. How to adapt a specific identity to extensive pose and scene variations is very challenging. (3) How to produce images with camera-realistic styles in an identity-agnostic way? Engine-based methods struggle" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "to generalize effectively on real data due to insufficient realism, whereas realistic styles introduced by learning-based methods are often entangled with real identity information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.515 + ], + "angle": 0, + "content": "To address these challenges, we present a pedestrian generation pipeline, entitled VIPerson, to flexibly generate camera-realistic images of pedestrians with Virtual Identities for the Person ReID task. Our pipeline comprises three components: flexible identity generator, cross-camera diversification, and style realization. First, we propose a flexible identity generator, which enhances the stable diffusion model through our proposed text embedding dropout. This enhancement allows the model to generate hard identities based on dropout text embeddings, while still being flexibly capable of infinitely creating new virtual identities using random embeddings. Second, for each generated identity image, we introduce cross-camera diversification, enabling diverse and scalable changes in poses and scenes. Leveraging the IP-adapter [51, 56] to vary poses and our proposed segment-and-paste strategy to alter scenes, we seamlessly facilitate cross-camera variations for each identity. Third, we present style realization to extract and transfer identity-agnostic camera-realistic style in real images to virtual images using style transfer techniques [9]. This process effectively reduces the style gap between generated and real images, while preserving the virtual identities. To capture comprehensive poses, scenes, and styles, we leverage pedestrian images from the large-scale unlabeled dataset LUPerson [12]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.909, + 0.791 + ], + "angle": 0, + "content": "We perform in-depth experimental analysis on three key factors, i.e. virtual identity diversity, cross-camera variations, and camera-realistic style in data synthesis. These factors are verified to be crucial for enhancing the generalization ability of the ReID model. Furthermore, when knowledge of scenes and styles from specific scenarios is available, our approach can be extended to incorporate scenario-specific patterns, thereby enhancing ReID performances in target applications. The effectiveness of our generated dataset is validated in two groups of experiments. First, we verify the effectiveness of generated data across a broad range of downstream ReID tasks, including direct transfer, supervised fine-tuning, and unsupervised domain adaptation. Second, we adapt our pipeline to the identity expansion scenario, i.e., there exist few labeled identities for training under the target scenario. Our method consistently achieves state-of-the-art results across these settings, demonstrating its superiority and applicability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.806, + 0.655, + 0.822 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.832, + 0.735, + 0.848 + ], + "angle": 0, + "content": "2.1. Person Re-identification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Person re-identification [21, 24, 26, 31, 33, 47, 52, 60] aims to retrieve pedestrians of the same identity from different camera views. This task presents challenges in distinguish" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23375" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "ing between similar identities and retrieving same-identity samples with diversity cross-camera variations [30, 32, 52]. With the rise of deep learning[14, 25, 40, 46], various methods are applied to ReID tasks, and these models achieve excellent results when trained and tested on the same datasets. However, when transferred to other datasets, they tend to suffer performance degradation due to distributional differences. There is an urgent need for a dataset with a wide distribution to enhance the generalization ability of ReID models across diverse scenarios and environments. Unfortunately, collecting a labeled dataset based on test scenarios is costly. For example, the large-scale ReID dataset MSMT17[44], consisting of 4,101 identities captured from 15 cameras, required two months of annotation work by three labelers. Moreover, it also leads to concerns about privacy leakage of pedestrians, which limits the application of ReID models trained on these datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.327, + 0.377 + ], + "angle": 0, + "content": "2.2. Synthesized ReID Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.431 + ], + "angle": 0, + "content": "To alleviate the challenges associated with real-world data collection, a few synthesized datasets for ReID have emerged as a promising solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Engine-based Methods. These datasets[1, 22, 41-43, 48, 49, 57] are typically generated using computer graphics techniques or simulation engines, such as Unity [36] and MakeHuman [10], which enable the creation of virtual labeled ReID data. SyRI[1] created pedestrians in a virtual scene, introducing cross-camera diversity by varying lighting and shooting perspectives. PersonX[41] utilized virtual hand-crafted 3D models as pedestrians and adjusted the angle and scene of the shot to acquire images. RandPerson[42] generated random UV texture maps to create 3D clothing person models. UnrealPerson[57] sampled real clothing to form the virtual pedestrian models and obtained images in virtual scenarios. ReID models trained on these virtual datasets can be effectively generalized to real datasets. However, they are limited by the manually designed nature of the virtual pedestrians, hard identities, scenes, lighting conditions, and video shooting, which increases synthesis costs and restricts scalability. Furthermore, the synthetic pedestrians and scenes still exhibit significant discrepancies compared to real-world data, affecting their transferability to real datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Learning-based Methods. With the development of generative models, including generative adversary network (GAN) [13] and diffusion[37], some methods [7, 19, 50] leveraged them to synthesize realistic images. GCL [7] and Pose-dIVE [19] modify pedestrian poses and synthesize diverse viewpoint images using GANs and diffusion models, respectively, which effectively expand the real-world datasets. However, these models require training on real ReID datasets with a limited number of cameras, which makes the distribution of the model restricted to insuffi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "cient cross-camera variations as shown in the right of Figure 1. Moreover, while the models learn realistic styles, they inadvertently embed identity-related features, making the generated images highly consistent with the real identities present in real-world datasets, which raises privacy concerns[17, 27-29, 35, 54, 55, 62]. MALS[50] is a synthetic text-based person-retrieval dataset that contains 1,510,330 image-text pairs. It used textual descriptions of real pedestrians to guide the off-the-shelf diffusion models to synthesize person images. Although its identities are virtual, each pedestrian has only a single image and lacks cross-camera diversity, such as poses, camera viewpoints, scenes, and lighting condition variations. What's more, it lacks the ability to synthesize diverse virtual identities with hard identities, which is not suitable for person ReID tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.334, + 0.753, + 0.351 + ], + "angle": 0, + "content": "3. Data Generation Pipeline" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.785, + 0.375 + ], + "angle": 0, + "content": "3.1. What makes a person dataset?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.382, + 0.905, + 0.473 + ], + "angle": 0, + "content": "In the person re-identification (ReID) task, the dataset used for supervised model training is generally denoted as \\( D = \\{X_{1}, X_{2}, \\dots, X_{P}\\} \\), where \\( P \\) is the number of identities, \\( X = \\{x_{1}, x_{2}, \\dots, x_{K}\\} \\) is a collection of images \\( x \\) of the same identity, which denoted as an intra-identity (intra-ID) group, and \\( K \\) is the number of samples in each identity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.905, + 0.67 + ], + "angle": 0, + "content": "In real-world surveillance systems, there are numerous pedestrian identities, each appearing in different scenes captured by different cameras. Collecting real pedestrian data presents significant challenges, primarily due to the high cost of manual annotation and privacy concerns. To address these challenges, recent studies propose synthetic methods to generate pedestrians images to improve ReID models. However, they struggle to fully satisfy or scale the following three key factors: virtual identity diversity, cross-camera variations, and camera-realistic styles. In the following, we will introduce the flexible identity generator, cross-camera diversification, and style realization to address the above challenges." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.761, + 0.697 + ], + "angle": 0, + "content": "3.2. Flexible Identity Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Our first task is to generate diverse virtual identities in the dataset, which plays a critical role in aiding ReID models to learn discriminative features. In particular, hard identities, which share similar appearances, enable ReID models to capture fine-grained features. To flexibly generate new random identities and hard identities, we present the flexible identity generator which is upgraded based on the latent diffusion model [37]. Next, we will elaborate on the training and generation process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Training of Identity Generator. Previous learning-based methods are troubled by the limitation of generating pedestrian images that are strictly consistent with real identities, which raises concerns about privacy exposure. To" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23376" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.899, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.304, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Illustration of pedestrian generation pipeline VIPerson. The flexible identity generator flexibly generates new random identities and hard identities. Afterward for each identity, cross-camera diversification generates intra-identity groups that contain images with rich cross-camera variations. Finally, style realization effectively reduces the style gap between generated and real images, while preserving virtual identities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.484, + 0.513 + ], + "angle": 0, + "content": "solve this, we train the generator on a large-scale dataset, MALS[50], which contains 1.5 million person images of virtual identities. In this way, the generator learns a rich distribution of virtual identities, enabling it to not only generate virtual identities but also to expand the diversity of identities with varying characteristics. Note that we empirically adjust all training images to the same body pose with IP-adapter[51, 56] beforehand to mitigate training difficulties caused by pose variations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.516, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Our flexible identity generator is based on the latent diffusion model[3, 37] (LDM) conditioned on the identity description. The LDM is a generative model that learns to reverse a gradual noising process in the latent space to generate data. The architecture of our flexible identity generator is a modified UNet [38] that includes residual and attention blocks. To reduce the computation during training, we first compress the input data into a latent space through an autoencoder, denoted as \\( z = AE(x) \\). During the forward diffusion process, at each timestep \\( t \\), the latent data \\( z \\) is transformed into \\( z_{t} \\) by adding Gaussian noise \\( \\varepsilon \\). During training, the model learns to predict and reverse the noise at time step \\( t \\) based on the current estimate \\( z_{t} \\) and the additional condition. As the condition, we employ the text embedding of the pedestrian image description \\( d \\). To provide an accurate \\( d \\) for each pedestrian image, we generate the text description containing four main dressing attributes, i.e. genders/ages, up dress, down dress, and shoes, using BLIP[20, 23], a powerful model for visual question answering. Then we extract the text embedding of \\( d \\) with the text encoder \\( \\varphi \\) of CLIP[34], and inject the text embedding \\( \\varphi(d) \\) into the cross-attention layer of our flexible identity generator. The training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.878, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\nL _ {L D M} = \\mathbb {E} _ {z, \\varepsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\varepsilon - \\varepsilon_ {\\theta} (z _ {t}, t, \\varphi (d)) \\| _ {2} ^ {2} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Generation of New Random Identity. After our flexible identity generator learns the diverse distribution of virtual pedestrians, we leverage a Gaussian-sampled embedding as the condition in the diffusion model's reverse denoising process to generate new random identities. Specifically, we sample a random embedding \\( c_{r} \\sim \\mathcal{N}(0,1) \\). This random embedding serves as a unique identifier that guides the generation process, ensuring diversity in the synthesized outputs. The reverse denoising process is modified to incorporate the \\( c_{r} \\) at each step \\( t \\). The iterative update for the reverse process is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.553, + 0.907, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t - 1} = \\mu_ {\\theta} (z _ {t}, t, c _ {r}) + \\sigma_ {t} \\zeta_ {t} \\\\ = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\varepsilon_ {\\theta} \\left(z _ {t}, t, c _ {r}\\right)\\right) + \\sigma_ {t} \\zeta_ {t}, \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.704 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are parameters to control the noise schedule, and \\(\\zeta_{t}\\) is an additional noise vector to maintain stochasticity. At the final iteration, the estimated \\(z_{0}\\) is mapped back to the pixel space using a pre-trained decoder, resulting in a high-quality image that embodies the new random identity defined by the \\(c_{r}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Generation of Hard Identity. To generate hard identities, we propose leveraging dropout text embedding \\( c_{d} \\) as conditions in the diffusion process. Given an identity image \\( I \\), our objective is to generate a hard identity \\( I' \\) that shares a similar appearance to \\( I \\). We first use BLIP[20, 23] to generate the text descriptions \\( d_{I} \\) of \\( I \\). The description is then encoded into 512-dimensional text embeddings using CLIP. To introduce hard identities, we apply a dropout strategy by randomly setting some of the dimensions of the text embedding to zero, resulting in \\( c_{d} = dropout(\\varphi(d_{I})) \\). This process randomly alter subtle identity-related features while preserving core semantic information. \\( c_{d} \\) is then used as conditional inputs for the identity generator to generate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23377" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Details of generated datasets. The \"Generalization Ability\" represents the rank-1 performance of direct transfer results on MSMT17 with BoT[26] model trained on these synthesized datasets. MALS* is a synthetic text-based person-retrieval dataset, which is challenging to generalize to real-world person ReID tasks." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.133, + 0.865, + 0.254 + ], + "angle": 0, + "content": "
TypeDatasetsVirtual IdentityCamera-realisticCross-camera variationsHard Identity#Identity#BBoxesGeneralization Ability
Engine-basedSyRI [1]YesNoLimitedNo10056,0007.8
PersonX [41]YesNoLimitedNo4109,8407.6
RandPerson [42]YesNoLimitedNo8,000132,14514.5
UnrealPerson [57]YesNoLimitedLimited3,000120,00026.8
Learning-basedMALS* [50]YesNoLimitedNo1,510,3301,510,3300.2
GCL [7]NoYesLimitedLimited---
Pose-dIVE [19]NoYesLimitedLimited---
VIPersonYesYesScalableScalable5,000114,73832.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.482, + 0.3 + ], + "angle": 0, + "content": "new, hard identities with subtle differences in appearance. The iterative process can be formally described as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.308, + 0.483, + 0.343 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\varepsilon_ {\\theta} \\left(z _ {t}, t, c _ {d}\\right)\\right) + \\sigma_ {t} \\zeta_ {t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.448 + ], + "angle": 0, + "content": "The \"flexible\" in the flexible identity generator lies in the ability to flexibly control conditions during inference, enabling the generation of versatile and diverse identities with varying characteristics. We provide more ablation studies on designing hard identities in section B of supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.351, + 0.471 + ], + "angle": 0, + "content": "3.3. Cross-camera Diversification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Enhancing cross-camera variations within each identity is crucial for improving the generalization ability of ReID models across different camera views. However, previous synthetic methods struggle to scale up cross-camera variations, due to the high manual effort required in designing diverse scenes or limited variation in training data. In this subsection, we aim to expand the intra-identity (intra-ID) group \\( X \\) for each identity with scalable cross-camera variations, specifically in terms of poses and scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Data Collection. First, we collect real-world pedestrian images, ensuring a diverse range of poses and scenes that reflect the variations encountered in different camera setups. Unless otherwise specified, our poses and scenes come from LUPerson [12], a large-scale unlabeled pedestrian dataset with over 4.18 million images collected across 46,260 scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Pose Variations. Second, we extract person poses in the collected images with openpose[5, 6, 39, 45, 56]. Then we utilize ControlNet-based IP-adapter[51, 56] to modify each virtual identity's pose with the randomly collected poses as the condition. However, significant pose variations may lead to identity degradation. To enhance identity consistency, we generate a corresponding identity description for each individual using BLIP. This description, along with the collected pose, is then used as control conditions for the IP-Adapter to guide the generation of the new image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Scene Variations. Lastly, we propose segment-and-paste strategy to modify the scene of images. For the collected" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.346 + ], + "angle": 0, + "content": "real-world images, we first apply the inpainting diffusion technique [53] to remove the original pedestrian, creating clean scene images. We then utilize the semantic segmentation method[8] to segment our generated pose-changed persons and patch them into the scene images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.358, + 0.681, + 0.374 + ], + "angle": 0, + "content": "3.4. Style Realization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.908, + 0.685 + ], + "angle": 0, + "content": "Camera-realistic styles, such as illumination, motion blur, camera bias, and environmental conditions, provide crucial contextual information that reflect real-world scenarios, playing a key role in improving the transferability of ReID models. Engine-based synthetic images inherently lack realistic styles, as they are created using virtual engines that do not capture real-world environmental nuances. Previous learning-based generative models learn realistic styles while embedding identity-related features, leading to generated images highly consistent with the real identities of the real-world ReID dataset, which raises privacy concerns. To address these, unlike learning the realistic distribution during training, we propose a post-processing approach, leveraging the neural style transfer method [9] to inject the style. Specifically, we randomly select a pedestrian image from LUPerson as the reference image, and leverage [9] to separate the style and content of the real-world image. Then, for each generated virtual image, we transfer the identity-agnostic style from the real-world image to the virtual image while preserving its original virtual identity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.816, + 0.713 + ], + "angle": 0, + "content": "3.5. Summary of the VIPerson Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The entire data generation pipeline is illustrated in Figure 3, which can generate a ReID dataset composed entirely of virtual identities without manual annotation. We generate the virtual dataset VIPerson, which serves as the primary dataset in all our experiments unless otherwise specified. Refer to Figure 3 and section A of supplementary material for examples of generated images. The detailed parameters of the datasets are shown in Table 1. Furthermore, we extend our approach to incorporate scenario-specific patterns, generating VIPerson* with available pose, scene and style information from target scenarios to enhance ReID performance in the targeted applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23378" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.295, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.091, + 0.482, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.239, + 0.483, + 0.267 + ], + "angle": 0, + "content": "Figure 4. Effect of different numbers of identities. \"Ours w/o Hard\" does not contain hard identities generated by \\( c_d \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.294, + 0.465, + 0.312 + ], + "angle": 0, + "content": "4. Analysis of Key Factors in Data Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.485, + 0.486 + ], + "angle": 0, + "content": "The synthesized pedestrian data is a fundamental component of downstream ReID tasks. In this study, we conduct an in-depth experimental analysis of several key factors in the generated dataset VIPerson: virtual identity diversity (including identity diversity and hard identity), cross-camera variations (including intra-ID group scale, pose variations, and scene variations), and camera-realistic style. We employ direct transfer performance on real datasets with BoT[26] to assess the quality of generated data, as direct transfer serves as the foundation for all other tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.5, + 0.322, + 0.516 + ], + "angle": 0, + "content": "4.1. Virtual Identity Diversity" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.522, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Identity Diversity. We explore the effect of the identity diversity by changing identity scale as shown in Figure 4. As the number of identities increases from 1000 to 3000, the rank-1 on MSMT17 and Market1501 has increased rapidly. It demonstrates that our flexible identity generator can generate scalable identities with diverse information, facilitating the model to better discriminate between unseen persons in the test set. However, as the identity scale increases further, the model performance grows slowly, especially on Market1501. This limitation may contribute to the scaling law, which is summarised in some studies[11, 15, 18] in many tasks: Even with an infinite amount of data, the learning capacity of ReID model is still constrained by other factors such as model capacity. More analysis between data scale and model size refers to section D in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Hard Identity. We explore the impact of hard identities on model performance as shown in Figure 4. For example, at an identity count of 5000, 'ours' outperforms 'ours-w/o Hard' on both test datasets. This suggests that hard identities in datasets allow the ReID model to learn to focus on more detailed discriminative features, improving the ability to distinguish between similar persons. In contrast to previous synthetic datasets that require manual design of hard identities, our approach generates hard identities automati-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.716, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.092, + 0.904, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.237, + 0.859, + 0.251 + ], + "angle": 0, + "content": "Figure 5. Effect of different Intra-ID group scales." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.903, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.435, + 0.906, + 0.478 + ], + "angle": 0, + "content": "Figure 6. Ablation study of modules in VIPerson. 'PV' indicates pose variations, 'SV' refers to scene variations, and 'SR' is style realization. Each point reports the rank-1 performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.9, + 0.52 + ], + "angle": 0, + "content": "cally, making the process more cost-effective and scalable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.741, + 0.546 + ], + "angle": 0, + "content": "4.2. Cross-camera Variations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Intra-ID Group Scale. Cross-camera variations are positively correlated with the scale of intra-ID group. This is because as the intra-ID group increases, there is a larger variety of viewpoints, poses, and scenes for each identity, which is introduced via our random sampling process. Here, we demonstrate how the scale of the intra-ID group affects the performance of the ReID model. As the results shown in Figure 5, with the increment of our intra-ID group scale, the performance continues to increase on both test datasets. Besides, our VIPerson is consistently superior to UnrealPerson, which demonstrates that our cross-camera variations are more diverse than UnrealPerson. Benefited from our proposed cross-camera diversification that introduces scalable and diverse poses and scenes without relying on manual design, we can expand the intra-ID group to any scale for better performance at a low cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Pose Variations. To investigate the impact of pose variations, we minimise the variation of pose introduced through cross-camera diversification, as shown in the 'SV+SR' column of Figure 6. The rank-1 performance drops from \\(32.5\\%\\) to \\(20.5\\%\\) on MSMT17, and from \\(66.5\\%\\) to \\(39.8\\%\\) on Market1501. In the absence of pose diversity, the model may overfit to specific pose-related features, which hin" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23379" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "ders its ability to generalize across different viewing angles. Since our pose set is randomly collected from real pedestrian images, the poses of generated persons are natural and varied. Our results demonstrate that pose variations in training data plays a crucial role in improving the model's robustness and performance in real-world scenarios, where pedestrians are often captured from various angles and perspectives." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.213, + 0.485, + 0.486 + ], + "angle": 0, + "content": "Scene Variations. As shown in the column 'PV', Without scene variations, the performance is limited to \\(0.9\\%\\) on MSMT17 and \\(11.7\\%\\) on Market1501. Introducing scene variations significantly boosts performance to \\(7.5\\%\\) on MSMT17 and \\(33.6\\%\\) on Market1501 as shown in the column 'PV+SV', indicating that scene variations are crucial for the model's effectiveness. Since the background in each image of the same pedestrian varies in the test data, incorporating diverse backgrounds into the training set can substantially enhance performance. Previous approaches, which rely on both synthetic and real datasets sampled from sequential camera videos, are limited in terms of scene variations within each intra-ID group. It would be costly for them to capture more videos to expand the variation in scenes. In contrast, our segment-and-paste strategy allows the Intra-ID group to exhibit richer background variations with a significantly lower cost. More analysis of scenes refer to the section E in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.497, + 0.3, + 0.513 + ], + "angle": 0, + "content": "4.3. Camera-realistic Style" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.701 + ], + "angle": 0, + "content": "Style Realization. As shown in column 'PV+SV+SR' of Figure 6, introducing camera-realistic styles significantly improves rank-1 from \\(33.6\\%\\) to \\(66.5\\%\\) on Market1501 and from \\(7.5\\%\\) to \\(32.5\\%\\) on MSMT17. This highlights the importance of style realization, as it bridges the gap between synthetic and real-world images. By transferring illumination, motion blur, camera bias, and environmental conditions from real reference images, our style realization process makes the synthetic images more consistent with real-world data distributions. This not only enhances model performance but also improves generalization when deployed in diverse real-world environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.325, + 0.733 + ], + "angle": 0, + "content": "5. Downstream ReID Tasks" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.31, + 0.759 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Details of Synthetic Data. Existing synthetic ReID datasets include SyRI[1], PersonX[41], RandPerson[42], and UnrealPerson[57], and their details are shown in Table 1. Our dataset is named VIPerson in this paper. We provide synthetic datasets for two scenarios, including the dataset without target scenario information (denoted as 'VIPerson') and the datasets based on target scenario information (denoted as 'VIPerson*'). We test the model on MSMT17 [44] and Market1501 [59]. For VIPerson, pose," + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.159 + ], + "angle": 0, + "content": "Table 2. Direct transfer results with BoT[26]. The fully-supervised results are shown in italics, the best direct transfer results are indicated in bold, and the second-best results are underlined. More experiments on different ReID models are provided in section C of supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.161, + 0.887, + 0.294 + ], + "angle": 0, + "content": "
Dataset TypeTraining DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealMSMT1773.949.956.229.8
Market15019.23.094.686.3
SyntheticMALS0.20.64.91.5
SyRI7.81.615.24.9
PersonX7.62.436.815.5
RandPerson14.54.855.929.9
UnrealPerson26.89.964.037.2
VIPerson32.510.366.538.3
VIPerson*34.310.867.537.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.309, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Table 3. Supervised fine-tuning results with BoT[26] method." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.325, + 0.894, + 0.435 + ], + "angle": 0, + "content": "
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet72.542.491.476.8
MSMT17--93.782.5
Market150174.144.5--
SyntheticMALS28.114.474.854.7
UnrealPerson74.546.094.084.7
VIPerson76.551.594.085.3
VIPerson*76.751.295.186.9
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.578 + ], + "angle": 0, + "content": "background, and style information is consistently from LUPerson [12] across all tested datasets. VIPerson consists of 5000 identities including hard identities. For VIPerson*, when testing on MSMT17, we obtain pose, scene and style information from MSMT17 training set. When testing on Market1501, we obtain pose and scene information from LUPerson and MSMT17, and transfer style from Market1501 training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Training Details. We conduct our experiments and analysis primarily using ResNet50[14] as the backbone model. Specifically, for the direct transfer experiment, we adopt BoT[26] and resize the image to \\(256 \\times 128\\) and set the batch size as 128. We adopt the Adam optimizer for training, with the base learning rate set to \\(3.5 \\times 10^{-4}\\) and decreased by a factor of 0.1 at the \\(10^{th}\\) and \\(40^{th}\\) epochs. Details about fine-tuning and domain adaptation experiments refer to section F in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.727, + 0.671, + 0.742 + ], + "angle": 0, + "content": "5.2. Direct Transfer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To evaluate the quality of generated data, we train the models only on generated data and test them on real data. As shown in Table 2, our VIPerson surpass all other datasets. Unlike training data from real-world, VIPerson consists entirely of virtual identities, thus avoiding privacy concerns. Nevertheless, VIPerson still exceeds Market1501 by \\(23.3\\%\\) on MSMT17 and outperforms MSMT17 by \\(10.3\\%\\) on Market1501. Compared to the previous synthetic datasets, our VIPerson significantly improves performance, as it achieve three key factors in synthetic person data as analysis in Sec-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23380" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.09, + 0.474, + 0.104 + ], + "angle": 0, + "content": "Table 4. Unsupervised domain adaptation results with ISE[58]." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.106, + 0.472, + 0.205 + ], + "angle": 0, + "content": "
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet64.735.094.084.7
MSMT17--93.884.2
Market150160.431.2--
SyntheticRandPerson59.030.293.483.8
UnrealPerson65.936.194.586.3
VIPerson70.439.594.986.4
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.223, + 0.483, + 0.251 + ], + "angle": 0, + "content": "tion 4, allowing the ReID model to generalize better to unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.484, + 0.357 + ], + "angle": 0, + "content": "Moreover, our cross-camera diversification and style realization are flexible to adapt to the target scenarios. Given scene and style information from the target scenario, we leverage them to develop VIPerson*, which further enhances performance on target datasets. This makes our method more practical and adaptable for real-world applications under diverse scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.366, + 0.308, + 0.383 + ], + "angle": 0, + "content": "5.3. Supervised Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.525 + ], + "angle": 0, + "content": "When sufficient labeled data is available, pretraining the model on synthetic data and fine-tuning it on target datasets can enhance performance. As shown in Table 3, with pretrained on our datasets, the rank-1 performance reaches \\(76.7\\%\\) on MSMT17 and \\(95.1\\%\\) on Market1501, suppressing all the other datasets. These results demonstrate the diversity, strong feature representation, and ample scale of our dataset, underscoring its value as a powerful resource for enhancing fine-tuning performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.533, + 0.388, + 0.549 + ], + "angle": 0, + "content": "5.4. Unsupervised Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.706 + ], + "angle": 0, + "content": "In most application scenarios, pedestrian data is usually unlabeled due to the limitations of annotation costs. In these cases, leveraging other labeled datasets for pretraining and fine-tuning on the unlabeled data can also improve performance. From Table 4, we see that the model pre-train on our datasets suppresses all the other datasets, achieving \\(70.4\\%\\) on MSMT17 and \\(94.9\\%\\) on Market1501. Our VIPerson effectively enhance the performance without relying on any manual annotations, which underscores its significance for reducing annotation dependency in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.353, + 0.729 + ], + "angle": 0, + "content": "5.5. Identity Expansion Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "As rising privacy concerns, we consider a scenario with few label identities, where most individuals are unwilling to share their personal image as training data, leading to a scarcity of labeled identities for training ReID models. Fortunately, we can add generated data to enhance the identity diversity. To this end, we simulate environments where only a small fraction of labeled data (20 or 40 identities) is available, exploring how ReID models perform with limited labeled identities and expended synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 7, when only 20 real identities are available, adding real data from other domains (MK) only" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.716, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.092, + 0.9, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.203, + 0.716, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.203, + 0.901, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.317, + 0.763, + 0.328 + ], + "angle": 0, + "content": "Market1501" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.339, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Figure 7. Results under identity expansion scenarios. 'R' indicates training with real images from the testing scenarios, 'MK' refers to training set of Market1501, 'MS' corresponds to training set of MSMT17, 'UP' denotes UnrealPerson, and 'VIP' represents our VIPerson dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.653 + ], + "angle": 0, + "content": "improves rank-1 by \\(15\\%\\) on MSMT17 test set. Although collected from the real world, it is constrained by limited identity and cross-camera variations, which hinders its ability to effectively enhance model performance. As a comparison, adding VIPerson significantly improves rank-1 performance on MSMT17 by \\(30.4\\%\\), which effectively bridge the gap caused by the lack of real-world data. These results underscore the practical applicability of our approach in scenarios with few label identities, making it possible to train effective ReID models with minimal privacy exposure to real identity. In environments where privacy is a critical concern, such as surveillance systems and smart cities, this approach can be a game-changer for identity verification and tracking systems." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.676, + 0.634, + 0.692 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this work, we introduce VIPerson, a novel pedestrian generation pipeline for flexibly creating virtual identities for the person ReID task to address privacy concerns and annotations consumptions in real ReID dataset. Our work identifies three key factors that enhance the effectiveness of the synthesized dataset, offering valuable insights for future directions in data synthesis. Moreover, we validated the scaling law, which shows that performance improves with an increase in the scale of synthesized data and model parameters. However, we also observed a performance bottleneck at a certain data scale, which suggests that future work should focus on how to effectively leverage synthetic data to enhance ReID models." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23381" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[1] Slawomir Bak, Peter Carr, and Jean-Francois Lalonde. Domain adaptation through synthesis for unsupervised person re-identification. In Proceedings of the European conference on computer vision (ECCV), pages 189–205, 2018. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.271, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[2] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.341, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[3] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.484, + 0.451 + ], + "angle": 0, + "content": "[4] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, pages 1092-1104. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[6] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2004–2013, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[8] Xiangguang Chen, Ye Zhu, Yu Li, Bingtao Fu, Lei Sun, Ying Shan, and Shan Liu. Robust human matting via semantic guidance. In Proceedings of the Asian Conference on Computer Vision, pages 2984-2999, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[9] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.481, + 0.788 + ], + "angle": 0, + "content": "[10] Makehuman community. Makehuman. http://www.makehumancommunity.org/. 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[11] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceed" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750-14759, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.906, + 0.348 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[17] Erin Illman and Paul Temple. California consumer privacy act. The Business Lawyer, 75(1):1637-1646, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[18] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.45, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[19] Inès Hyeonsu Kim, JoungBin Lee, Woojeong Jin, Soowon Son, Kyusun Cho, Junyoung Seo, Min-Seop Kwak, Seokju Cho, JeongYeol Baek, Byeongwon Lee, et al. Pose-dive: Pose-diversified augmentation with diffusion model for person re-identification. arXiv preprint arXiv:2406.16042, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[20] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 31-41, Toronto, Canada, 2023. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.688 + ], + "angle": 0, + "content": "[21] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6729–6738, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[22] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In Proceedings of the 29th ACM international conference on multimedia, pages 3115-3123, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI conference on artificial intelligence, pages 1405–1413, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23382" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[26] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[27] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Jihong Guan, and Shuigeng Zhou. Fine-grained zero-shot object detection. arXiv preprint arXiv:2507.10358, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.484, + 0.333 + ], + "angle": 0, + "content": "[28] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[29] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21203-21214, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[30] Yi-Xing Peng, Jile Jiao, Xuetao Feng, and Wei-Shi Zheng. Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia, 25:2393–2403, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[31] Yi-Xing Peng, Yuanxun Li, and Wei-Shi Zheng. Revisiting person re-identification by camera selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5): 2692-2708, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[32] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, and Wei-Shi Zheng. Generalized intra-camera supervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4516-4527, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[33] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, Qize Yang, Jingke Meng, Xihan Wei, and Wei-Shi Zheng. Person deidentification: A variation-guided identity shift modeling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29331-29341, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[35] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[36] John Ricciello. John ricciello sets out to identify the engine of growth for unity technologies (interview). VentureBeat. Interview with Dean Takahashi. Retrieved January, 18 (3), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[39] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.907, + 0.316 + ], + "angle": 0, + "content": "[41] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 608-617, 2019. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.907, + 0.385 + ], + "angle": 0, + "content": "[42] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In Proceedings of the 28th ACM international conference on multimedia, pages 3422–3430, 2020. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.907, + 0.455 + ], + "angle": 0, + "content": "[43] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4900–4909, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[44] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.553 + ], + "angle": 0, + "content": "[45] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.907, + 0.622 + ], + "angle": 0, + "content": "[46] Xiao-Ming Wu, Dian Zheng, Zuhao Liu, and Wei-Shi Zheng. Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17055-17064, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.907, + 0.677 + ], + "angle": 0, + "content": "[47] Yuqiao Xian, Yi-Xing Peng, Xing Sun, and Wei-Shi Zheng. Distilling consistent relations for multi-source domain adaptive person re-identification. Pattern Recognition, 157: 110821, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.907, + 0.735 + ], + "angle": 0, + "content": "[48] Suncheng Xiang, Yuzhuo Fu, Guanjie You, and Ting Liu. Unsupervised domain adaptation through synthesis for person re-identification. In 2020 IEEE international conference on multimedia and expo (ICME), pages 1–6. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.736, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[49] Suncheng Xiang, Dahong Qian, Mengyuan Guan, Binjie Yan, Ting Liu, Yuzhuo Fu, and Guanjie You. Less is more: Learning from synthetic data with fine-grained attributes for person re-identification. ACM Transactions on Multimedia Computing, Communications and Applications, 19(5s): 1-20, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[50] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23383" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[52] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[54] Delong Zhang, Yixing Peng, Ancong Wu, and Wei-shi Zheng. Privacy-preserving face recognition with adaptive generative perturbations. In International Conference on Pattern Recognition, pages 210-227. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[55] Delong Zhang, Yi-Xing Peng, Xiao-Ming Wu, Ancong Wu, and Wei-Shi Zheng. Pixelfade: Privacy-preserving person re-identification with noise-guided progressive replacement. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6326–6334, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[56] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[57] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11506–11515, 2021. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[58] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7369–7378, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[59] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[60] Wei-Shi Zheng, Junkai Yan, and Yi-Xing Peng. A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.769 + ], + "angle": 0, + "content": "[61] Zhedong Zheng, Liang Zheng, and Yi Yang. Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In Proceedings of the IEEE international conference on computer vision, pages 3754–3762, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.482, + 0.853 + ], + "angle": 0, + "content": "[62] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo, Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "23384" + } + ] +] \ No newline at end of file diff --git a/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_origin.pdf b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dd37bd8553967a8862edd4a62dd70eee806598b7 --- /dev/null +++ b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/9c97a79e-c7c9-4628-ab6c-7d4d09ccbb66_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:202609ce99902baa28926e500fb4eac78bf1385ff27955691bfe1ef3732133e5 +size 1428242 diff --git a/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/full.md b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a93dc012fa6c72884d29012621d0e9122a9cdc7e --- /dev/null +++ b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/full.md @@ -0,0 +1,303 @@ +# VIPerson: Flexibly Generating Virtual Identity for Person Re-Identification + +Xiao-Wen Zhang $^{1,3*}$ +Zhi Ouyang $^{1,3}$ + +Delong Zhang $^{1,3*}$ Jingke Meng $^{1,3\dagger}$ + +Yi-Xing Peng $^{1,3}$ Wei-Shi Zheng $^{1,2,3\dagger}$ + +$^{1}$ School of Computer Science and Engineering, Sun Yat-sen University, China; + +$^{2}$ Peng Cheng Laboratory, Shenzhen, China; + +3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China. + +{zhangxw75,zhangdlong5,pengyx23,}@mail2.sysu.edu.cn wszheng@ieee.org + +# Abstract + +Person re-identification (ReID) is to match the person images under different camera views. Training ReID models necessitates a substantial amount of labeled real-world data, leading to high labeling costs and privacy issues. Although several ReID data synthetic methods are proposed to address these issues, they fail to generate images with new identities or real-world camera styles. In this paper, we propose a novel pedestrian generation pipeline, VIPerson, to generate camera-realistic pedestrian images with flexible Virtual Identities for the Person ReID task. VIPerson focuses on three key factors in data synthesis: (I) Virtual identity diversity: Enhancing the latent diffusion model with our proposed dropout text embedding, we flexibly generate random and hard identities. (II) Scalable cross-camera variations: VIPerson introduces scalable variations of scenes and poses within each identity. (III) Camera-realistic style: Adopting an identity-agnostic approach to transfer realistic styles, we avoid privacy exposure of real identities. Extensive experimental results across a broad range of downstream ReID tasks demonstrate the superiority of our generated dataset over existing methods. In addition, VIPerson can be adapted to the identity expansion scenario, which widens the application of our pipeline. The dataset and code of VIPerson are available at https://isee-laboratory.github.io/VIPerson/. + +# 1. Introduction + +The person re-identification (ReID) task aims to match pedestrians from different camera views. With the rise of deep learning[52], ReID methods have been developed based on real annotated pedestrian data [44, 59]. These real-world ReID datasets are often limited in both scale and + +![](images/e0a50b0054e4d39ec4508e89ee9098d9fd5e45b16c576a0016a805b657a754c6.jpg) +Figure 1. Comparison of synthetic data. Left: Images of engine-based methods have a style gap compared with real images. Right: Learning-based methods expose privacy by generating images that match real identities in real-world datasets. + +diversity, primarily due to the high cost of data collection and manual annotation. Moreover, pedestrian images containing sensitive information often expose real identities, raising serious privacy concerns. Strict data privacy regulations [2, 4, 17, 35], such as EU's General Data Protection Regulation, restrict the ability to gather and share sensitive visual data. This further impedes the collection of public ReID datasets, with some, like DukeMTMC [61], even taken down. The lack of high-quality annotated data without privacy concerns significantly limits the performance and generalization ability of ReID models. + +To address these challenges, recent studies propose to synthesize pedestrian images to improve ReID models. These methods can be divided into two research streams. Engine-based methods [1, 41, 42, 57] are developed to leverage simulation engines to synthesize virtual pedestrian images. However, these synthetic images exhibit distinct virtual characteristics (see 'Left' in Figure 1), resulting in a style gap compared to real-world images that showcase camera-realistic styles (e.g. illumination, motion blur, camera bias, and environmental conditions). These synthetic + +![](images/3ed0c6fb776e314454d32e8ebb09d663ae2549bde02f9b4d9fa2e4fd35062331.jpg) +Figure 2. Image examples from our VIPerson dataset. Our data pipeline can generate virtual identities, containing random identities, hard identities, and intra-identity (intra-ID) groups that contain camera-realistic images with rich cross-camera variations. More images are provided in section A of the supplementary material. + +data hinder ReID models' transferability to real-world scenarios. Learning-based synthetic approaches [7, 19, 50], propose using generative models[13, 16], to generate realistic pedestrian images. However, these models face privacy concerns as they are only capable of generating images that are consistent with the real identities present in real-world ReID datasets (see 'Right' in Figure 1), which limits their application in privacy-sensitive scenarios. Moreover, both approaches struggle with limited cross-camera variations (e.g. poses and scenes) due to constraints imposed by costs of scene design or limited variations in training data. + +In this paper, we aim to propose a novel pipeline to generate a camera-realistic pedestrian dataset composed entirely of virtual identities for the person ReID task. However, achieving this is non-trivial due to three core challenges in pedestrian data synthesis: (1) How to flexibly generate diverse virtual identities, containing new random identities and hard identities? New random identities exhibit diverse appearances, while hard identities, often referred to as hard samples, share similar visual characteristics despite belonging to different individuals. Both are crucial for enhancing identity diversity, aiding ReID models to learn discriminative features across a broad range of individuals. This necessitates a flexible generative model for skillfully controlling the generation of diverse identities. (2) How to introduce scalable cross-camera variations within each identity? Abundant cross-camera variations of generated images can facilitate the generalization ability of ReID models. How to adapt a specific identity to extensive pose and scene variations is very challenging. (3) How to produce images with camera-realistic styles in an identity-agnostic way? Engine-based methods struggle + +to generalize effectively on real data due to insufficient realism, whereas realistic styles introduced by learning-based methods are often entangled with real identity information. + +To address these challenges, we present a pedestrian generation pipeline, entitled VIPerson, to flexibly generate camera-realistic images of pedestrians with Virtual Identities for the Person ReID task. Our pipeline comprises three components: flexible identity generator, cross-camera diversification, and style realization. First, we propose a flexible identity generator, which enhances the stable diffusion model through our proposed text embedding dropout. This enhancement allows the model to generate hard identities based on dropout text embeddings, while still being flexibly capable of infinitely creating new virtual identities using random embeddings. Second, for each generated identity image, we introduce cross-camera diversification, enabling diverse and scalable changes in poses and scenes. Leveraging the IP-adapter [51, 56] to vary poses and our proposed segment-and-paste strategy to alter scenes, we seamlessly facilitate cross-camera variations for each identity. Third, we present style realization to extract and transfer identity-agnostic camera-realistic style in real images to virtual images using style transfer techniques [9]. This process effectively reduces the style gap between generated and real images, while preserving the virtual identities. To capture comprehensive poses, scenes, and styles, we leverage pedestrian images from the large-scale unlabeled dataset LUPerson [12]. + +We perform in-depth experimental analysis on three key factors, i.e. virtual identity diversity, cross-camera variations, and camera-realistic style in data synthesis. These factors are verified to be crucial for enhancing the generalization ability of the ReID model. Furthermore, when knowledge of scenes and styles from specific scenarios is available, our approach can be extended to incorporate scenario-specific patterns, thereby enhancing ReID performances in target applications. The effectiveness of our generated dataset is validated in two groups of experiments. First, we verify the effectiveness of generated data across a broad range of downstream ReID tasks, including direct transfer, supervised fine-tuning, and unsupervised domain adaptation. Second, we adapt our pipeline to the identity expansion scenario, i.e., there exist few labeled identities for training under the target scenario. Our method consistently achieves state-of-the-art results across these settings, demonstrating its superiority and applicability. + +# 2. Related Work + +# 2.1. Person Re-identification + +Person re-identification [21, 24, 26, 31, 33, 47, 52, 60] aims to retrieve pedestrians of the same identity from different camera views. This task presents challenges in distinguish + +ing between similar identities and retrieving same-identity samples with diversity cross-camera variations [30, 32, 52]. With the rise of deep learning[14, 25, 40, 46], various methods are applied to ReID tasks, and these models achieve excellent results when trained and tested on the same datasets. However, when transferred to other datasets, they tend to suffer performance degradation due to distributional differences. There is an urgent need for a dataset with a wide distribution to enhance the generalization ability of ReID models across diverse scenarios and environments. Unfortunately, collecting a labeled dataset based on test scenarios is costly. For example, the large-scale ReID dataset MSMT17[44], consisting of 4,101 identities captured from 15 cameras, required two months of annotation work by three labelers. Moreover, it also leads to concerns about privacy leakage of pedestrians, which limits the application of ReID models trained on these datasets. + +# 2.2. Synthesized ReID Dataset + +To alleviate the challenges associated with real-world data collection, a few synthesized datasets for ReID have emerged as a promising solution. + +Engine-based Methods. These datasets[1, 22, 41-43, 48, 49, 57] are typically generated using computer graphics techniques or simulation engines, such as Unity [36] and MakeHuman [10], which enable the creation of virtual labeled ReID data. SyRI[1] created pedestrians in a virtual scene, introducing cross-camera diversity by varying lighting and shooting perspectives. PersonX[41] utilized virtual hand-crafted 3D models as pedestrians and adjusted the angle and scene of the shot to acquire images. RandPerson[42] generated random UV texture maps to create 3D clothing person models. UnrealPerson[57] sampled real clothing to form the virtual pedestrian models and obtained images in virtual scenarios. ReID models trained on these virtual datasets can be effectively generalized to real datasets. However, they are limited by the manually designed nature of the virtual pedestrians, hard identities, scenes, lighting conditions, and video shooting, which increases synthesis costs and restricts scalability. Furthermore, the synthetic pedestrians and scenes still exhibit significant discrepancies compared to real-world data, affecting their transferability to real datasets. + +Learning-based Methods. With the development of generative models, including generative adversary network (GAN) [13] and diffusion[37], some methods [7, 19, 50] leveraged them to synthesize realistic images. GCL [7] and Pose-dIVE [19] modify pedestrian poses and synthesize diverse viewpoint images using GANs and diffusion models, respectively, which effectively expand the real-world datasets. However, these models require training on real ReID datasets with a limited number of cameras, which makes the distribution of the model restricted to insuffi + +cient cross-camera variations as shown in the right of Figure 1. Moreover, while the models learn realistic styles, they inadvertently embed identity-related features, making the generated images highly consistent with the real identities present in real-world datasets, which raises privacy concerns[17, 27-29, 35, 54, 55, 62]. MALS[50] is a synthetic text-based person-retrieval dataset that contains 1,510,330 image-text pairs. It used textual descriptions of real pedestrians to guide the off-the-shelf diffusion models to synthesize person images. Although its identities are virtual, each pedestrian has only a single image and lacks cross-camera diversity, such as poses, camera viewpoints, scenes, and lighting condition variations. What's more, it lacks the ability to synthesize diverse virtual identities with hard identities, which is not suitable for person ReID tasks. + +# 3. Data Generation Pipeline + +# 3.1. What makes a person dataset? + +In the person re-identification (ReID) task, the dataset used for supervised model training is generally denoted as $D = \{X_{1}, X_{2}, \dots, X_{P}\}$ , where $P$ is the number of identities, $X = \{x_{1}, x_{2}, \dots, x_{K}\}$ is a collection of images $x$ of the same identity, which denoted as an intra-identity (intra-ID) group, and $K$ is the number of samples in each identity. + +In real-world surveillance systems, there are numerous pedestrian identities, each appearing in different scenes captured by different cameras. Collecting real pedestrian data presents significant challenges, primarily due to the high cost of manual annotation and privacy concerns. To address these challenges, recent studies propose synthetic methods to generate pedestrians images to improve ReID models. However, they struggle to fully satisfy or scale the following three key factors: virtual identity diversity, cross-camera variations, and camera-realistic styles. In the following, we will introduce the flexible identity generator, cross-camera diversification, and style realization to address the above challenges. + +# 3.2. Flexible Identity Generator + +Our first task is to generate diverse virtual identities in the dataset, which plays a critical role in aiding ReID models to learn discriminative features. In particular, hard identities, which share similar appearances, enable ReID models to capture fine-grained features. To flexibly generate new random identities and hard identities, we present the flexible identity generator which is upgraded based on the latent diffusion model [37]. Next, we will elaborate on the training and generation process. + +Training of Identity Generator. Previous learning-based methods are troubled by the limitation of generating pedestrian images that are strictly consistent with real identities, which raises concerns about privacy exposure. To + +![](images/05e7a99d11b2112ad3a5bc48fc8f6ccfc670b3896371a50e6e9c977c3694e8bb.jpg) +Figure 3. Illustration of pedestrian generation pipeline VIPerson. The flexible identity generator flexibly generates new random identities and hard identities. Afterward for each identity, cross-camera diversification generates intra-identity groups that contain images with rich cross-camera variations. Finally, style realization effectively reduces the style gap between generated and real images, while preserving virtual identities. + +solve this, we train the generator on a large-scale dataset, MALS[50], which contains 1.5 million person images of virtual identities. In this way, the generator learns a rich distribution of virtual identities, enabling it to not only generate virtual identities but also to expand the diversity of identities with varying characteristics. Note that we empirically adjust all training images to the same body pose with IP-adapter[51, 56] beforehand to mitigate training difficulties caused by pose variations. + +Our flexible identity generator is based on the latent diffusion model[3, 37] (LDM) conditioned on the identity description. The LDM is a generative model that learns to reverse a gradual noising process in the latent space to generate data. The architecture of our flexible identity generator is a modified UNet [38] that includes residual and attention blocks. To reduce the computation during training, we first compress the input data into a latent space through an autoencoder, denoted as $z = AE(x)$ . During the forward diffusion process, at each timestep $t$ , the latent data $z$ is transformed into $z_{t}$ by adding Gaussian noise $\varepsilon$ . During training, the model learns to predict and reverse the noise at time step $t$ based on the current estimate $z_{t}$ and the additional condition. As the condition, we employ the text embedding of the pedestrian image description $d$ . To provide an accurate $d$ for each pedestrian image, we generate the text description containing four main dressing attributes, i.e. genders/ages, up dress, down dress, and shoes, using BLIP[20, 23], a powerful model for visual question answering. Then we extract the text embedding of $d$ with the text encoder $\varphi$ of CLIP[34], and inject the text embedding $\varphi(d)$ into the cross-attention layer of our flexible identity generator. The training loss is defined as: + +$$ +L _ {L D M} = \mathbb {E} _ {z, \varepsilon \sim \mathcal {N} (0, 1), t} \left[ \| \varepsilon - \varepsilon_ {\theta} (z _ {t}, t, \varphi (d)) \| _ {2} ^ {2} \right]. \tag {1} +$$ + +Generation of New Random Identity. After our flexible identity generator learns the diverse distribution of virtual pedestrians, we leverage a Gaussian-sampled embedding as the condition in the diffusion model's reverse denoising process to generate new random identities. Specifically, we sample a random embedding $c_{r} \sim \mathcal{N}(0,1)$ . This random embedding serves as a unique identifier that guides the generation process, ensuring diversity in the synthesized outputs. The reverse denoising process is modified to incorporate the $c_{r}$ at each step $t$ . The iterative update for the reverse process is given by: + +$$ +\begin{array}{l} z _ {t - 1} = \mu_ {\theta} (z _ {t}, t, c _ {r}) + \sigma_ {t} \zeta_ {t} \\ = \frac {1}{\sqrt {\alpha_ {t}}} \left(z _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \varepsilon_ {\theta} \left(z _ {t}, t, c _ {r}\right)\right) + \sigma_ {t} \zeta_ {t}, \tag {2} \\ \end{array} +$$ + +where $\alpha_{t}$ and $\sigma_{t}$ are parameters to control the noise schedule, and $\zeta_{t}$ is an additional noise vector to maintain stochasticity. At the final iteration, the estimated $z_{0}$ is mapped back to the pixel space using a pre-trained decoder, resulting in a high-quality image that embodies the new random identity defined by the $c_{r}$ . + +Generation of Hard Identity. To generate hard identities, we propose leveraging dropout text embedding $c_{d}$ as conditions in the diffusion process. Given an identity image $I$ , our objective is to generate a hard identity $I'$ that shares a similar appearance to $I$ . We first use BLIP[20, 23] to generate the text descriptions $d_{I}$ of $I$ . The description is then encoded into 512-dimensional text embeddings using CLIP. To introduce hard identities, we apply a dropout strategy by randomly setting some of the dimensions of the text embedding to zero, resulting in $c_{d} = dropout(\varphi(d_{I}))$ . This process randomly alter subtle identity-related features while preserving core semantic information. $c_{d}$ is then used as conditional inputs for the identity generator to generate + +Table 1. Details of generated datasets. The "Generalization Ability" represents the rank-1 performance of direct transfer results on MSMT17 with BoT[26] model trained on these synthesized datasets. MALS* is a synthetic text-based person-retrieval dataset, which is challenging to generalize to real-world person ReID tasks. + +
TypeDatasetsVirtual IdentityCamera-realisticCross-camera variationsHard Identity#Identity#BBoxesGeneralization Ability
Engine-basedSyRI [1]YesNoLimitedNo10056,0007.8
PersonX [41]YesNoLimitedNo4109,8407.6
RandPerson [42]YesNoLimitedNo8,000132,14514.5
UnrealPerson [57]YesNoLimitedLimited3,000120,00026.8
Learning-basedMALS* [50]YesNoLimitedNo1,510,3301,510,3300.2
GCL [7]NoYesLimitedLimited---
Pose-dIVE [19]NoYesLimitedLimited---
VIPersonYesYesScalableScalable5,000114,73832.5
+ +new, hard identities with subtle differences in appearance. The iterative process can be formally described as: + +$$ +z _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(z _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \varepsilon_ {\theta} \left(z _ {t}, t, c _ {d}\right)\right) + \sigma_ {t} \zeta_ {t}. \tag {3} +$$ + +The "flexible" in the flexible identity generator lies in the ability to flexibly control conditions during inference, enabling the generation of versatile and diverse identities with varying characteristics. We provide more ablation studies on designing hard identities in section B of supplementary material. + +# 3.3. Cross-camera Diversification + +Enhancing cross-camera variations within each identity is crucial for improving the generalization ability of ReID models across different camera views. However, previous synthetic methods struggle to scale up cross-camera variations, due to the high manual effort required in designing diverse scenes or limited variation in training data. In this subsection, we aim to expand the intra-identity (intra-ID) group $X$ for each identity with scalable cross-camera variations, specifically in terms of poses and scenes. + +Data Collection. First, we collect real-world pedestrian images, ensuring a diverse range of poses and scenes that reflect the variations encountered in different camera setups. Unless otherwise specified, our poses and scenes come from LUPerson [12], a large-scale unlabeled pedestrian dataset with over 4.18 million images collected across 46,260 scenes. + +Pose Variations. Second, we extract person poses in the collected images with openpose[5, 6, 39, 45, 56]. Then we utilize ControlNet-based IP-adapter[51, 56] to modify each virtual identity's pose with the randomly collected poses as the condition. However, significant pose variations may lead to identity degradation. To enhance identity consistency, we generate a corresponding identity description for each individual using BLIP. This description, along with the collected pose, is then used as control conditions for the IP-Adapter to guide the generation of the new image. + +Scene Variations. Lastly, we propose segment-and-paste strategy to modify the scene of images. For the collected + +real-world images, we first apply the inpainting diffusion technique [53] to remove the original pedestrian, creating clean scene images. We then utilize the semantic segmentation method[8] to segment our generated pose-changed persons and patch them into the scene images. + +# 3.4. Style Realization + +Camera-realistic styles, such as illumination, motion blur, camera bias, and environmental conditions, provide crucial contextual information that reflect real-world scenarios, playing a key role in improving the transferability of ReID models. Engine-based synthetic images inherently lack realistic styles, as they are created using virtual engines that do not capture real-world environmental nuances. Previous learning-based generative models learn realistic styles while embedding identity-related features, leading to generated images highly consistent with the real identities of the real-world ReID dataset, which raises privacy concerns. To address these, unlike learning the realistic distribution during training, we propose a post-processing approach, leveraging the neural style transfer method [9] to inject the style. Specifically, we randomly select a pedestrian image from LUPerson as the reference image, and leverage [9] to separate the style and content of the real-world image. Then, for each generated virtual image, we transfer the identity-agnostic style from the real-world image to the virtual image while preserving its original virtual identity. + +# 3.5. Summary of the VIPerson Dataset + +The entire data generation pipeline is illustrated in Figure 3, which can generate a ReID dataset composed entirely of virtual identities without manual annotation. We generate the virtual dataset VIPerson, which serves as the primary dataset in all our experiments unless otherwise specified. Refer to Figure 3 and section A of supplementary material for examples of generated images. The detailed parameters of the datasets are shown in Table 1. Furthermore, we extend our approach to incorporate scenario-specific patterns, generating VIPerson* with available pose, scene and style information from target scenarios to enhance ReID performance in the targeted applications. + +![](images/9d01b3e8a6ad9386754c23bdd82f51a35004e611debd3857c2f4ba4adee31368.jpg) +Figure 4. Effect of different numbers of identities. "Ours w/o Hard" does not contain hard identities generated by $c_d$ . + +![](images/609ba3a32a9901921eca21f31e3fe2d5b18f108ed27de4780d72a2dbc9d38894.jpg) + +# 4. Analysis of Key Factors in Data Synthesis + +The synthesized pedestrian data is a fundamental component of downstream ReID tasks. In this study, we conduct an in-depth experimental analysis of several key factors in the generated dataset VIPerson: virtual identity diversity (including identity diversity and hard identity), cross-camera variations (including intra-ID group scale, pose variations, and scene variations), and camera-realistic style. We employ direct transfer performance on real datasets with BoT[26] to assess the quality of generated data, as direct transfer serves as the foundation for all other tasks. + +# 4.1. Virtual Identity Diversity + +Identity Diversity. We explore the effect of the identity diversity by changing identity scale as shown in Figure 4. As the number of identities increases from 1000 to 3000, the rank-1 on MSMT17 and Market1501 has increased rapidly. It demonstrates that our flexible identity generator can generate scalable identities with diverse information, facilitating the model to better discriminate between unseen persons in the test set. However, as the identity scale increases further, the model performance grows slowly, especially on Market1501. This limitation may contribute to the scaling law, which is summarised in some studies[11, 15, 18] in many tasks: Even with an infinite amount of data, the learning capacity of ReID model is still constrained by other factors such as model capacity. More analysis between data scale and model size refers to section D in the supplementary material. + +Hard Identity. We explore the impact of hard identities on model performance as shown in Figure 4. For example, at an identity count of 5000, 'ours' outperforms 'ours-w/o Hard' on both test datasets. This suggests that hard identities in datasets allow the ReID model to learn to focus on more detailed discriminative features, improving the ability to distinguish between similar persons. In contrast to previous synthetic datasets that require manual design of hard identities, our approach generates hard identities automati- + +![](images/ae141a1f85e1ce399c8920b3f9c3d5125d49054cefd6bf615db886288919e8c5.jpg) + +![](images/55a7e2fd7353f6ae8e17ef713ae5f915ac90ab417c39242b7b90471746d352f7.jpg) +Figure 5. Effect of different Intra-ID group scales. + +![](images/0f01b1a54a66d3d0f2a5fef843851fc9b34142c05005725c5221892d607bbd97.jpg) +Figure 6. Ablation study of modules in VIPerson. 'PV' indicates pose variations, 'SV' refers to scene variations, and 'SR' is style realization. Each point reports the rank-1 performance. + +cally, making the process more cost-effective and scalable. + +# 4.2. Cross-camera Variations + +Intra-ID Group Scale. Cross-camera variations are positively correlated with the scale of intra-ID group. This is because as the intra-ID group increases, there is a larger variety of viewpoints, poses, and scenes for each identity, which is introduced via our random sampling process. Here, we demonstrate how the scale of the intra-ID group affects the performance of the ReID model. As the results shown in Figure 5, with the increment of our intra-ID group scale, the performance continues to increase on both test datasets. Besides, our VIPerson is consistently superior to UnrealPerson, which demonstrates that our cross-camera variations are more diverse than UnrealPerson. Benefited from our proposed cross-camera diversification that introduces scalable and diverse poses and scenes without relying on manual design, we can expand the intra-ID group to any scale for better performance at a low cost. + +Pose Variations. To investigate the impact of pose variations, we minimise the variation of pose introduced through cross-camera diversification, as shown in the 'SV+SR' column of Figure 6. The rank-1 performance drops from $32.5\%$ to $20.5\%$ on MSMT17, and from $66.5\%$ to $39.8\%$ on Market1501. In the absence of pose diversity, the model may overfit to specific pose-related features, which hin + +ders its ability to generalize across different viewing angles. Since our pose set is randomly collected from real pedestrian images, the poses of generated persons are natural and varied. Our results demonstrate that pose variations in training data plays a crucial role in improving the model's robustness and performance in real-world scenarios, where pedestrians are often captured from various angles and perspectives. + +Scene Variations. As shown in the column 'PV', Without scene variations, the performance is limited to $0.9\%$ on MSMT17 and $11.7\%$ on Market1501. Introducing scene variations significantly boosts performance to $7.5\%$ on MSMT17 and $33.6\%$ on Market1501 as shown in the column 'PV+SV', indicating that scene variations are crucial for the model's effectiveness. Since the background in each image of the same pedestrian varies in the test data, incorporating diverse backgrounds into the training set can substantially enhance performance. Previous approaches, which rely on both synthetic and real datasets sampled from sequential camera videos, are limited in terms of scene variations within each intra-ID group. It would be costly for them to capture more videos to expand the variation in scenes. In contrast, our segment-and-paste strategy allows the Intra-ID group to exhibit richer background variations with a significantly lower cost. More analysis of scenes refer to the section E in supplementary material. + +# 4.3. Camera-realistic Style + +Style Realization. As shown in column 'PV+SV+SR' of Figure 6, introducing camera-realistic styles significantly improves rank-1 from $33.6\%$ to $66.5\%$ on Market1501 and from $7.5\%$ to $32.5\%$ on MSMT17. This highlights the importance of style realization, as it bridges the gap between synthetic and real-world images. By transferring illumination, motion blur, camera bias, and environmental conditions from real reference images, our style realization process makes the synthetic images more consistent with real-world data distributions. This not only enhances model performance but also improves generalization when deployed in diverse real-world environments. + +# 5. Downstream ReID Tasks + +# 5.1. Implementation Details + +Details of Synthetic Data. Existing synthetic ReID datasets include SyRI[1], PersonX[41], RandPerson[42], and UnrealPerson[57], and their details are shown in Table 1. Our dataset is named VIPerson in this paper. We provide synthetic datasets for two scenarios, including the dataset without target scenario information (denoted as 'VIPerson') and the datasets based on target scenario information (denoted as 'VIPerson*'). We test the model on MSMT17 [44] and Market1501 [59]. For VIPerson, pose, + +Table 2. Direct transfer results with BoT[26]. The fully-supervised results are shown in italics, the best direct transfer results are indicated in bold, and the second-best results are underlined. More experiments on different ReID models are provided in section C of supplementary material. + +
Dataset TypeTraining DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealMSMT1773.949.956.229.8
Market15019.23.094.686.3
SyntheticMALS0.20.64.91.5
SyRI7.81.615.24.9
PersonX7.62.436.815.5
RandPerson14.54.855.929.9
UnrealPerson26.89.964.037.2
VIPerson32.510.366.538.3
VIPerson*34.310.867.537.6
+ +Table 3. Supervised fine-tuning results with BoT[26] method. + +
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet72.542.491.476.8
MSMT17--93.782.5
Market150174.144.5--
SyntheticMALS28.114.474.854.7
UnrealPerson74.546.094.084.7
VIPerson76.551.594.085.3
VIPerson*76.751.295.186.9
+ +background, and style information is consistently from LUPerson [12] across all tested datasets. VIPerson consists of 5000 identities including hard identities. For VIPerson*, when testing on MSMT17, we obtain pose, scene and style information from MSMT17 training set. When testing on Market1501, we obtain pose and scene information from LUPerson and MSMT17, and transfer style from Market1501 training set. + +Training Details. We conduct our experiments and analysis primarily using ResNet50[14] as the backbone model. Specifically, for the direct transfer experiment, we adopt BoT[26] and resize the image to $256 \times 128$ and set the batch size as 128. We adopt the Adam optimizer for training, with the base learning rate set to $3.5 \times 10^{-4}$ and decreased by a factor of 0.1 at the $10^{th}$ and $40^{th}$ epochs. Details about fine-tuning and domain adaptation experiments refer to section F in supplementary material. + +# 5.2. Direct Transfer + +To evaluate the quality of generated data, we train the models only on generated data and test them on real data. As shown in Table 2, our VIPerson surpass all other datasets. Unlike training data from real-world, VIPerson consists entirely of virtual identities, thus avoiding privacy concerns. Nevertheless, VIPerson still exceeds Market1501 by $23.3\%$ on MSMT17 and outperforms MSMT17 by $10.3\%$ on Market1501. Compared to the previous synthetic datasets, our VIPerson significantly improves performance, as it achieve three key factors in synthetic person data as analysis in Sec- + +Table 4. Unsupervised domain adaptation results with ISE[58]. + +
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet64.735.094.084.7
MSMT17--93.884.2
Market150160.431.2--
SyntheticRandPerson59.030.293.483.8
UnrealPerson65.936.194.586.3
VIPerson70.439.594.986.4
+ +tion 4, allowing the ReID model to generalize better to unseen scenarios. + +Moreover, our cross-camera diversification and style realization are flexible to adapt to the target scenarios. Given scene and style information from the target scenario, we leverage them to develop VIPerson*, which further enhances performance on target datasets. This makes our method more practical and adaptable for real-world applications under diverse scenarios. + +# 5.3. Supervised Fine-tuning + +When sufficient labeled data is available, pretraining the model on synthetic data and fine-tuning it on target datasets can enhance performance. As shown in Table 3, with pretrained on our datasets, the rank-1 performance reaches $76.7\%$ on MSMT17 and $95.1\%$ on Market1501, suppressing all the other datasets. These results demonstrate the diversity, strong feature representation, and ample scale of our dataset, underscoring its value as a powerful resource for enhancing fine-tuning performance. + +# 5.4. Unsupervised Domain Adaptation + +In most application scenarios, pedestrian data is usually unlabeled due to the limitations of annotation costs. In these cases, leveraging other labeled datasets for pretraining and fine-tuning on the unlabeled data can also improve performance. From Table 4, we see that the model pre-train on our datasets suppresses all the other datasets, achieving $70.4\%$ on MSMT17 and $94.9\%$ on Market1501. Our VIPerson effectively enhance the performance without relying on any manual annotations, which underscores its significance for reducing annotation dependency in real-world applications. + +# 5.5. Identity Expansion Scenarios + +As rising privacy concerns, we consider a scenario with few label identities, where most individuals are unwilling to share their personal image as training data, leading to a scarcity of labeled identities for training ReID models. Fortunately, we can add generated data to enhance the identity diversity. To this end, we simulate environments where only a small fraction of labeled data (20 or 40 identities) is available, exploring how ReID models perform with limited labeled identities and expended synthetic data. + +As shown in Figure 7, when only 20 real identities are available, adding real data from other domains (MK) only + +![](images/10b788ab478551c3d2c5348ff66a4289980f9e5c5ae02ae172d79d33178d8915.jpg) + +![](images/ce09a8d147a203610f97064c6f472475047543475f1000659456f2ba8aeab11b.jpg) + +![](images/f0cd81c33b169fdb4a3ecbbc44dbc0c10eeb2699452dba4eeba48f1a32cb0f35.jpg) +Market1501 + +![](images/5769d1002c807a0dd3531487f99fc4604edb7852e00369e1d2485b532d30428b.jpg) +Figure 7. Results under identity expansion scenarios. 'R' indicates training with real images from the testing scenarios, 'MK' refers to training set of Market1501, 'MS' corresponds to training set of MSMT17, 'UP' denotes UnrealPerson, and 'VIP' represents our VIPerson dataset. + +improves rank-1 by $15\%$ on MSMT17 test set. Although collected from the real world, it is constrained by limited identity and cross-camera variations, which hinders its ability to effectively enhance model performance. As a comparison, adding VIPerson significantly improves rank-1 performance on MSMT17 by $30.4\%$ , which effectively bridge the gap caused by the lack of real-world data. These results underscore the practical applicability of our approach in scenarios with few label identities, making it possible to train effective ReID models with minimal privacy exposure to real identity. In environments where privacy is a critical concern, such as surveillance systems and smart cities, this approach can be a game-changer for identity verification and tracking systems. + +# 6. Conclusion + +In this work, we introduce VIPerson, a novel pedestrian generation pipeline for flexibly creating virtual identities for the person ReID task to address privacy concerns and annotations consumptions in real ReID dataset. Our work identifies three key factors that enhance the effectiveness of the synthesized dataset, offering valuable insights for future directions in data synthesis. Moreover, we validated the scaling law, which shows that performance improves with an increase in the scale of synthesized data and model parameters. However, we also observed a performance bottleneck at a certain data scale, which suggests that future work should focus on how to effectively leverage synthetic data to enhance ReID models. + +# Acknowledgements + +This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025). + +# References + +[1] Slawomir Bak, Peter Carr, and Jean-Francois Lalonde. Domain adaptation through synthesis for unsupervised person re-identification. In Proceedings of the European conference on computer vision (ECCV), pages 189–205, 2018. 1, 3, 5, 7 +[2] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1 +[3] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 4 +[4] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, pages 1092-1104. PMLR, 2021. 1 +[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 5 +[6] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5 +[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2004–2013, 2021. 2, 3, 5 +[8] Xiangguang Chen, Ye Zhu, Yu Li, Bingtao Fu, Lei Sun, Ying Shan, and Shan Liu. Robust human matting via semantic guidance. In Proceedings of the Asian Conference on Computer Vision, pages 2984-2999, 2022. 5 +[9] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 5 +[10] Makehuman community. Makehuman. http://www.makehumancommunity.org/. 2024.3 +[11] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 6 +[12] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceed + +ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750-14759, 2021. 2, 5, 7 +[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7 +[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 6 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[17] Erin Illman and Paul Temple. California consumer privacy act. The Business Lawyer, 75(1):1637-1646, 2019. 1, 3 +[18] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 6 +[19] Inès Hyeonsu Kim, JoungBin Lee, Woojeong Jin, Soowon Son, Kyusun Cho, Junyoung Seo, Min-Seop Kwak, Seokju Cho, JeongYeol Baek, Byeongwon Lee, et al. Pose-dive: Pose-diversified augmentation with diffusion model for person re-identification. arXiv preprint arXiv:2406.16042, 2024. 2, 3, 5 +[20] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 31-41, Toronto, Canada, 2023. Association for Computational Linguistics. 4 +[21] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6729–6738, 2021. 2 +[22] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In Proceedings of the 29th ACM international conference on multimedia, pages 3115-3123, 2021. 3 +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4 +[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI conference on artificial intelligence, pages 1405–1413, 2023. 2 +[25] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei + +Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 3 +[26] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 5, 6, 7 +[27] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Jihong Guan, and Shuigeng Zhou. Fine-grained zero-shot object detection. arXiv preprint arXiv:2507.10358, 2025. 3 +[28] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024. +[29] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21203-21214, 2025. 3 +[30] Yi-Xing Peng, Jile Jiao, Xuetao Feng, and Wei-Shi Zheng. Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia, 25:2393–2403, 2022. 3 +[31] Yi-Xing Peng, Yuanxun Li, and Wei-Shi Zheng. Revisiting person re-identification by camera selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5): 2692-2708, 2023. 2 +[32] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, and Wei-Shi Zheng. Generalized intra-camera supervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4516-4527, 2023. 3 +[33] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, Qize Yang, Jingke Meng, Xihan Wei, and Wei-Shi Zheng. Person deidentification: A variation-guided identity shift modeling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29331-29341, 2025. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4 +[35] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1, 3 +[36] John Ricciello. John ricciello sets out to identify the engine of growth for unity technologies (interview). VentureBeat. Interview with Dean Takahashi. Retrieved January, 18 (3), 2015. 3 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4 + +[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[39] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 5 +[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3 +[41] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 608-617, 2019. 1, 3, 5, 7 +[42] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In Proceedings of the 28th ACM international conference on multimedia, pages 3422–3430, 2020. 1, 3, 5, 7 +[43] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4900–4909, 2022. 3 +[44] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 1, 3, 7 +[45] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 5 +[46] Xiao-Ming Wu, Dian Zheng, Zuhao Liu, and Wei-Shi Zheng. Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17055-17064, 2023. 3 +[47] Yuqiao Xian, Yi-Xing Peng, Xing Sun, and Wei-Shi Zheng. Distilling consistent relations for multi-source domain adaptive person re-identification. Pattern Recognition, 157: 110821, 2025. 2 +[48] Suncheng Xiang, Yuzhuo Fu, Guanjie You, and Ting Liu. Unsupervised domain adaptation through synthesis for person re-identification. In 2020 IEEE international conference on multimedia and expo (ICME), pages 1–6. IEEE, 2020. 3 +[49] Suncheng Xiang, Dahong Qian, Mengyuan Guan, Binjie Yan, Ting Liu, Yuzhuo Fu, and Guanjie You. Less is more: Learning from synthetic data with fine-grained attributes for person re-identification. ACM Transactions on Multimedia Computing, Communications and Applications, 19(5s): 1-20, 2023. 3 +[50] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 2, 3, 4, 5 + +[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4, 5 +[52] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 2, 3 +[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 5 +[54] Delong Zhang, Yixing Peng, Ancong Wu, and Wei-shi Zheng. Privacy-preserving face recognition with adaptive generative perturbations. In International Conference on Pattern Recognition, pages 210-227. Springer, 2024. 3 +[55] Delong Zhang, Yi-Xing Peng, Xiao-Ming Wu, Ancong Wu, and Wei-Shi Zheng. Pixelfade: Privacy-preserving person re-identification with noise-guided progressive replacement. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6326–6334, 2024. 3 +[56] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5 +[57] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11506–11515, 2021. 1, 3, 5, 7 +[58] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7369–7378, 2022. 8 +[59] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 1, 7 +[60] Wei-Shi Zheng, Junkai Yan, and Yi-Xing Peng. A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[61] Zhedong Zheng, Liang Zheng, and Yi Yang. Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In Proceedings of the IEEE international conference on computer vision, pages 3754–3762, 2017. 1 +[62] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo, Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3 \ No newline at end of file diff --git a/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/images.zip b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a34bc0e401792e7561df28871a00c45f25681665 --- /dev/null +++ b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93e9eee9c3b9526d5399096905b6028511c718a9f118f381aef556923263961e +size 490330 diff --git a/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/layout.json b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b4623cfdb949c75f1a3762340af13833e3427342 --- /dev/null +++ b/2025/VIPerson_ Flexibly Generating Virtual Identity for Person Re-Identification/layout.json @@ -0,0 +1,8187 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 71, + 103, + 539, + 121 + ], + "type": "text", + "content": "VIPerson: Flexibly Generating Virtual Identity for Person Re-Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 150, + 142, + 251, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 142, + 251, + 171 + ], + "spans": [ + { + "bbox": [ + 150, + 142, + 251, + 171 + ], + "type": "text", + "content": "Xiao-Wen Zhang" + }, + { + "bbox": [ + 150, + 142, + 251, + 171 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 150, + 142, + 251, + 171 + ], + "type": "text", + "content": " \nZhi Ouyang" + }, + { + "bbox": [ + 150, + 142, + 251, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 269, + 142, + 358, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 142, + 358, + 171 + ], + "spans": [ + { + "bbox": [ + 269, + 142, + 358, + 171 + ], + "type": "text", + "content": "Delong Zhang" + }, + { + "bbox": [ + 269, + 142, + 358, + 171 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 269, + 142, + 358, + 171 + ], + "type": "text", + "content": " Jingke Meng" + }, + { + "bbox": [ + 269, + 142, + 358, + 171 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 367, + 142, + 465, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 142, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 367, + 142, + 465, + 171 + ], + "type": "text", + "content": "Yi-Xing Peng" + }, + { + "bbox": [ + 367, + 142, + 465, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 367, + 142, + 465, + 171 + ], + "type": "text", + "content": " Wei-Shi Zheng" + }, + { + "bbox": [ + 367, + 142, + 465, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 171, + 494, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 171, + 494, + 185 + ], + "spans": [ + { + "bbox": [ + 114, + 171, + 494, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 171, + 494, + 185 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University, China;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 185, + 411, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 185, + 411, + 199 + ], + "spans": [ + { + "bbox": [ + 200, + 185, + 411, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 200, + 185, + 411, + 199 + ], + "type": "text", + "content": "Peng Cheng Laboratory, Shenzhen, China;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 198, + 543, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 198, + 543, + 213 + ], + "spans": [ + { + "bbox": [ + 65, + 198, + 543, + 213 + ], + "type": "text", + "content": "3Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 215, + 498, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 215, + 498, + 227 + ], + "spans": [ + { + "bbox": [ + 109, + 215, + 498, + 227 + ], + "type": "text", + "content": "{zhangxw75,zhangdlong5,pengyx23,}@mail2.sysu.edu.cn wszheng@ieee.org" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 255, + 200, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 267 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 267 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 281, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 296, + 579 + ], + "type": "text", + "content": "Person re-identification (ReID) is to match the person images under different camera views. Training ReID models necessitates a substantial amount of labeled real-world data, leading to high labeling costs and privacy issues. Although several ReID data synthetic methods are proposed to address these issues, they fail to generate images with new identities or real-world camera styles. In this paper, we propose a novel pedestrian generation pipeline, VIPerson, to generate camera-realistic pedestrian images with flexible Virtual Identities for the Person ReID task. VIPerson focuses on three key factors in data synthesis: (I) Virtual identity diversity: Enhancing the latent diffusion model with our proposed dropout text embedding, we flexibly generate random and hard identities. (II) Scalable cross-camera variations: VIPerson introduces scalable variations of scenes and poses within each identity. (III) Camera-realistic style: Adopting an identity-agnostic approach to transfer realistic styles, we avoid privacy exposure of real identities. Extensive experimental results across a broad range of downstream ReID tasks demonstrate the superiority of our generated dataset over existing methods. In addition, VIPerson can be adapted to the identity expansion scenario, which widens the application of our pipeline. The dataset and code of VIPerson are available at https://isee-laboratory.github.io/VIPerson/." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 625, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 684 + ], + "type": "text", + "content": "The person re-identification (ReID) task aims to match pedestrians from different camera views. With the rise of deep learning[52], ReID methods have been developed based on real annotated pedestrian data [44, 59]. These real-world ReID datasets are often limited in both scale and" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 342, + 255, + 526, + 384 + ], + "blocks": [ + { + "bbox": [ + 342, + 255, + 526, + 384 + ], + "lines": [ + { + "bbox": [ + 342, + 255, + 526, + 384 + ], + "spans": [ + { + "bbox": [ + 342, + 255, + 526, + 384 + ], + "type": "image", + "image_path": "e0a50b0054e4d39ec4508e89ee9098d9fd5e45b16c576a0016a805b657a754c6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 391, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 434 + ], + "type": "text", + "content": "Figure 1. Comparison of synthetic data. Left: Images of engine-based methods have a style gap compared with real images. Right: Learning-based methods expose privacy by generating images that match real identities in real-world datasets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 591 + ], + "type": "text", + "content": "diversity, primarily due to the high cost of data collection and manual annotation. Moreover, pedestrian images containing sensitive information often expose real identities, raising serious privacy concerns. Strict data privacy regulations [2, 4, 17, 35], such as EU's General Data Protection Regulation, restrict the ability to gather and share sensitive visual data. This further impedes the collection of public ReID datasets, with some, like DukeMTMC [61], even taken down. The lack of high-quality annotated data without privacy concerns significantly limits the performance and generalization ability of ReID models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "To address these challenges, recent studies propose to synthesize pedestrian images to improve ReID models. These methods can be divided into two research streams. Engine-based methods [1, 41, 42, 57] are developed to leverage simulation engines to synthesize virtual pedestrian images. However, these synthetic images exhibit distinct virtual characteristics (see 'Left' in Figure 1), resulting in a style gap compared to real-world images that showcase camera-realistic styles (e.g. illumination, motion blur, camera bias, and environmental conditions). These synthetic" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23374" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 290, + 255 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 290, + 255 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 290, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 290, + 255 + ], + "type": "image", + "image_path": "3ed0c6fb776e314454d32e8ebb09d663ae2549bde02f9b4d9fa2e4fd35062331.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 320 + ], + "type": "text", + "content": "Figure 2. Image examples from our VIPerson dataset. Our data pipeline can generate virtual identities, containing random identities, hard identities, and intra-identity (intra-ID) groups that contain camera-realistic images with rich cross-camera variations. More images are provided in section A of the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 460 + ], + "type": "text", + "content": "data hinder ReID models' transferability to real-world scenarios. Learning-based synthetic approaches [7, 19, 50], propose using generative models[13, 16], to generate realistic pedestrian images. However, these models face privacy concerns as they are only capable of generating images that are consistent with the real identities present in real-world ReID datasets (see 'Right' in Figure 1), which limits their application in privacy-sensitive scenarios. Moreover, both approaches struggle with limited cross-camera variations (e.g. poses and scenes) due to constraints imposed by costs of scene design or limited variations in training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we aim to propose a novel pipeline to generate a camera-realistic pedestrian dataset composed entirely of virtual identities for the person ReID task. However, achieving this is non-trivial due to three core challenges in pedestrian data synthesis: (1) How to flexibly generate diverse virtual identities, containing new random identities and hard identities? New random identities exhibit diverse appearances, while hard identities, often referred to as hard samples, share similar visual characteristics despite belonging to different individuals. Both are crucial for enhancing identity diversity, aiding ReID models to learn discriminative features across a broad range of individuals. This necessitates a flexible generative model for skillfully controlling the generation of diverse identities. (2) How to introduce scalable cross-camera variations within each identity? Abundant cross-camera variations of generated images can facilitate the generalization ability of ReID models. How to adapt a specific identity to extensive pose and scene variations is very challenging. (3) How to produce images with camera-realistic styles in an identity-agnostic way? Engine-based methods struggle" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "to generalize effectively on real data due to insufficient realism, whereas realistic styles introduced by learning-based methods are often entangled with real identity information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 110, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 407 + ], + "type": "text", + "content": "To address these challenges, we present a pedestrian generation pipeline, entitled VIPerson, to flexibly generate camera-realistic images of pedestrians with Virtual Identities for the Person ReID task. Our pipeline comprises three components: flexible identity generator, cross-camera diversification, and style realization. First, we propose a flexible identity generator, which enhances the stable diffusion model through our proposed text embedding dropout. This enhancement allows the model to generate hard identities based on dropout text embeddings, while still being flexibly capable of infinitely creating new virtual identities using random embeddings. Second, for each generated identity image, we introduce cross-camera diversification, enabling diverse and scalable changes in poses and scenes. Leveraging the IP-adapter [51, 56] to vary poses and our proposed segment-and-paste strategy to alter scenes, we seamlessly facilitate cross-camera variations for each identity. Third, we present style realization to extract and transfer identity-agnostic camera-realistic style in real images to virtual images using style transfer techniques [9]. This process effectively reduces the style gap between generated and real images, while preserving the virtual identities. To capture comprehensive poses, scenes, and styles, we leverage pedestrian images from the large-scale unlabeled dataset LUPerson [12]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 409, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 556, + 626 + ], + "type": "text", + "content": "We perform in-depth experimental analysis on three key factors, i.e. virtual identity diversity, cross-camera variations, and camera-realistic style in data synthesis. These factors are verified to be crucial for enhancing the generalization ability of the ReID model. Furthermore, when knowledge of scenes and styles from specific scenarios is available, our approach can be extended to incorporate scenario-specific patterns, thereby enhancing ReID performances in target applications. The effectiveness of our generated dataset is validated in two groups of experiments. First, we verify the effectiveness of generated data across a broad range of downstream ReID tasks, including direct transfer, supervised fine-tuning, and unsupervised domain adaptation. Second, we adapt our pipeline to the identity expansion scenario, i.e., there exist few labeled identities for training under the target scenario. Our method consistently achieves state-of-the-art results across these settings, demonstrating its superiority and applicability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 638, + 400, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 638, + 400, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 638, + 400, + 651 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 658, + 449, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 658, + 449, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 658, + 449, + 671 + ], + "type": "text", + "content": "2.1. Person Re-identification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Person re-identification [21, 24, 26, 31, 33, 47, 52, 60] aims to retrieve pedestrians of the same identity from different camera views. This task presents challenges in distinguish" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23375" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "ing between similar identities and retrieving same-identity samples with diversity cross-camera variations [30, 32, 52]. With the rise of deep learning[14, 25, 40, 46], various methods are applied to ReID tasks, and these models achieve excellent results when trained and tested on the same datasets. However, when transferred to other datasets, they tend to suffer performance degradation due to distributional differences. There is an urgent need for a dataset with a wide distribution to enhance the generalization ability of ReID models across diverse scenarios and environments. Unfortunately, collecting a labeled dataset based on test scenarios is costly. For example, the large-scale ReID dataset MSMT17[44], consisting of 4,101 identities captured from 15 cameras, required two months of annotation work by three labelers. Moreover, it also leads to concerns about privacy leakage of pedestrians, which limits the application of ReID models trained on these datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 286, + 200, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 200, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 200, + 298 + ], + "type": "text", + "content": "2.2. Synthesized ReID Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 341 + ], + "type": "text", + "content": "To alleviate the challenges associated with real-world data collection, a few synthesized datasets for ReID have emerged as a promising solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 592 + ], + "type": "text", + "content": "Engine-based Methods. These datasets[1, 22, 41-43, 48, 49, 57] are typically generated using computer graphics techniques or simulation engines, such as Unity [36] and MakeHuman [10], which enable the creation of virtual labeled ReID data. SyRI[1] created pedestrians in a virtual scene, introducing cross-camera diversity by varying lighting and shooting perspectives. PersonX[41] utilized virtual hand-crafted 3D models as pedestrians and adjusted the angle and scene of the shot to acquire images. RandPerson[42] generated random UV texture maps to create 3D clothing person models. UnrealPerson[57] sampled real clothing to form the virtual pedestrian models and obtained images in virtual scenarios. ReID models trained on these virtual datasets can be effectively generalized to real datasets. However, they are limited by the manually designed nature of the virtual pedestrians, hard identities, scenes, lighting conditions, and video shooting, which increases synthesis costs and restricts scalability. Furthermore, the synthetic pedestrians and scenes still exhibit significant discrepancies compared to real-world data, affecting their transferability to real datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Learning-based Methods. With the development of generative models, including generative adversary network (GAN) [13] and diffusion[37], some methods [7, 19, 50] leveraged them to synthesize realistic images. GCL [7] and Pose-dIVE [19] modify pedestrian poses and synthesize diverse viewpoint images using GANs and diffusion models, respectively, which effectively expand the real-world datasets. However, these models require training on real ReID datasets with a limited number of cameras, which makes the distribution of the model restricted to insuffi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "cient cross-camera variations as shown in the right of Figure 1. Moreover, while the models learn realistic styles, they inadvertently embed identity-related features, making the generated images highly consistent with the real identities present in real-world datasets, which raises privacy concerns[17, 27-29, 35, 54, 55, 62]. MALS[50] is a synthetic text-based person-retrieval dataset that contains 1,510,330 image-text pairs. It used textual descriptions of real pedestrians to guide the off-the-shelf diffusion models to synthesize person images. Although its identities are virtual, each pedestrian has only a single image and lacks cross-camera diversity, such as poses, camera viewpoints, scenes, and lighting condition variations. What's more, it lacks the ability to synthesize diverse virtual identities with hard identities, which is not suitable for person ReID tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 264, + 460, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 460, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 460, + 277 + ], + "type": "text", + "content": "3. Data Generation Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 285, + 480, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 480, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 480, + 297 + ], + "type": "text", + "content": "3.1. What makes a person dataset?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": "In the person re-identification (ReID) task, the dataset used for supervised model training is generally denoted as " + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "inline_equation", + "content": "D = \\{X_{1}, X_{2}, \\dots, X_{P}\\}" + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": " is the number of identities, " + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "inline_equation", + "content": "X = \\{x_{1}, x_{2}, \\dots, x_{K}\\}" + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": " is a collection of images " + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": " of the same identity, which denoted as an intra-identity (intra-ID) group, and " + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 302, + 553, + 374 + ], + "type": "text", + "content": " is the number of samples in each identity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 553, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 530 + ], + "type": "text", + "content": "In real-world surveillance systems, there are numerous pedestrian identities, each appearing in different scenes captured by different cameras. Collecting real pedestrian data presents significant challenges, primarily due to the high cost of manual annotation and privacy concerns. To address these challenges, recent studies propose synthetic methods to generate pedestrians images to improve ReID models. However, they struggle to fully satisfy or scale the following three key factors: virtual identity diversity, cross-camera variations, and camera-realistic styles. In the following, we will introduce the flexible identity generator, cross-camera diversification, and style realization to address the above challenges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 539, + 465, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 465, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 465, + 552 + ], + "type": "text", + "content": "3.2. Flexible Identity Generator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 665 + ], + "type": "text", + "content": "Our first task is to generate diverse virtual identities in the dataset, which plays a critical role in aiding ReID models to learn discriminative features. In particular, hard identities, which share similar appearances, enable ReID models to capture fine-grained features. To flexibly generate new random identities and hard identities, we present the flexible identity generator which is upgraded based on the latent diffusion model [37]. Next, we will elaborate on the training and generation process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Training of Identity Generator. Previous learning-based methods are troubled by the limitation of generating pedestrian images that are strictly consistent with real identities, which raises concerns about privacy exposure. To" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23376" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 550, + 239 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 550, + 239 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 550, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 550, + 239 + ], + "type": "image", + "image_path": "05e7a99d11b2112ad3a5bc48fc8f6ccfc670b3896371a50e6e9c977c3694e8bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 240, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 240, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Illustration of pedestrian generation pipeline VIPerson. The flexible identity generator flexibly generates new random identities and hard identities. Afterward for each identity, cross-camera diversification generates intra-identity groups that contain images with rich cross-camera variations. Finally, style realization effectively reduces the style gap between generated and real images, while preserving virtual identities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 298, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 296, + 406 + ], + "type": "text", + "content": "solve this, we train the generator on a large-scale dataset, MALS[50], which contains 1.5 million person images of virtual identities. In this way, the generator learns a rich distribution of virtual identities, enabling it to not only generate virtual identities but also to expand the diversity of identities with varying characteristics. Note that we empirically adjust all training images to the same body pose with IP-adapter[51, 56] beforehand to mitigate training difficulties caused by pose variations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": "Our flexible identity generator is based on the latent diffusion model[3, 37] (LDM) conditioned on the identity description. The LDM is a generative model that learns to reverse a gradual noising process in the latent space to generate data. The architecture of our flexible identity generator is a modified UNet [38] that includes residual and attention blocks. To reduce the computation during training, we first compress the input data into a latent space through an autoencoder, denoted as " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "z = AE(x)" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": ". During the forward diffusion process, at each timestep " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": ", the latent data " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " is transformed into " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " by adding Gaussian noise " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": ". During training, the model learns to predict and reverse the noise at time step " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " based on the current estimate " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " and the additional condition. As the condition, we employ the text embedding of the pedestrian image description " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": ". To provide an accurate " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " for each pedestrian image, we generate the text description containing four main dressing attributes, i.e. genders/ages, up dress, down dress, and shoes, using BLIP[20, 23], a powerful model for visual question answering. Then we extract the text embedding of " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " with the text encoder " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " of CLIP[34], and inject the text embedding " + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\varphi(d)" + }, + { + "bbox": [ + 56, + 408, + 296, + 684 + ], + "type": "text", + "content": " into the cross-attention layer of our flexible identity generator. The training loss is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 695, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 695, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 66, + 695, + 296, + 717 + ], + "type": "interline_equation", + "content": "L _ {L D M} = \\mathbb {E} _ {z, \\varepsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\varepsilon - \\varepsilon_ {\\theta} (z _ {t}, t, \\varphi (d)) \\| _ {2} ^ {2} \\right]. \\tag {1}", + "image_path": "0c90d9d9d6a4d855d9a78e4e67db5a90b37aa2554dfbcba687a11a2af82b492b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "text", + "content": "Generation of New Random Identity. After our flexible identity generator learns the diverse distribution of virtual pedestrians, we leverage a Gaussian-sampled embedding as the condition in the diffusion model's reverse denoising process to generate new random identities. Specifically, we sample a random embedding " + }, + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "inline_equation", + "content": "c_{r} \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "text", + "content": ". This random embedding serves as a unique identifier that guides the generation process, ensuring diversity in the synthesized outputs. The reverse denoising process is modified to incorporate the " + }, + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "text", + "content": " at each step " + }, + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 297, + 555, + 430 + ], + "type": "text", + "content": ". The iterative update for the reverse process is given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 437, + 555, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 437, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 324, + 437, + 555, + 480 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t - 1} = \\mu_ {\\theta} (z _ {t}, t, c _ {r}) + \\sigma_ {t} \\zeta_ {t} \\\\ = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\varepsilon_ {\\theta} \\left(z _ {t}, t, c _ {r}\\right)\\right) + \\sigma_ {t} \\zeta_ {t}, \\tag {2} \\\\ \\end{array}", + "image_path": "51b606a977109d4e15ebe50e3bbd0807ed58224175c16b54a0f2930e7ad605aa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " are parameters to control the noise schedule, and " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\zeta_{t}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " is an additional noise vector to maintain stochasticity. At the final iteration, the estimated " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " is mapped back to the pixel space using a pre-trained decoder, resulting in a high-quality image that embodies the new random identity defined by the " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Generation of Hard Identity. To generate hard identities, we propose leveraging dropout text embedding " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "c_{d}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " as conditions in the diffusion process. Given an identity image " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ", our objective is to generate a hard identity " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " that shares a similar appearance to " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ". We first use BLIP[20, 23] to generate the text descriptions " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "d_{I}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ". The description is then encoded into 512-dimensional text embeddings using CLIP. To introduce hard identities, we apply a dropout strategy by randomly setting some of the dimensions of the text embedding to zero, resulting in " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "c_{d} = dropout(\\varphi(d_{I}))" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ". This process randomly alter subtle identity-related features while preserving core semantic information. " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "c_{d}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " is then used as conditional inputs for the identity generator to generate" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23377" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 105, + 529, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Details of generated datasets. The \"Generalization Ability\" represents the rank-1 performance of direct transfer results on MSMT17 with BoT[26] model trained on these synthesized datasets. MALS* is a synthetic text-based person-retrieval dataset, which is challenging to generalize to real-world person ReID tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 105, + 529, + 201 + ], + "lines": [ + { + "bbox": [ + 83, + 105, + 529, + 201 + ], + "spans": [ + { + "bbox": [ + 83, + 105, + 529, + 201 + ], + "type": "table", + "html": "
TypeDatasetsVirtual IdentityCamera-realisticCross-camera variationsHard Identity#Identity#BBoxesGeneralization Ability
Engine-basedSyRI [1]YesNoLimitedNo10056,0007.8
PersonX [41]YesNoLimitedNo4109,8407.6
RandPerson [42]YesNoLimitedNo8,000132,14514.5
UnrealPerson [57]YesNoLimitedLimited3,000120,00026.8
Learning-basedMALS* [50]YesNoLimitedNo1,510,3301,510,3300.2
GCL [7]NoYesLimitedLimited---
Pose-dIVE [19]NoYesLimitedLimited---
VIPersonYesYesScalableScalable5,000114,73832.5
", + "image_path": "360ed29545afbd4e472c2ea63be7506a968528ecf10552a57883c2d6d8413f39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 294, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 294, + 237 + ], + "type": "text", + "content": "new, hard identities with subtle differences in appearance. The iterative process can be formally described as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 243, + 295, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 243, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 65, + 243, + 295, + 271 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(z _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\varepsilon_ {\\theta} \\left(z _ {t}, t, c _ {d}\\right)\\right) + \\sigma_ {t} \\zeta_ {t}. \\tag {3}", + "image_path": "0317157aec59a7d625deabc580cf3f01c08786bcc9f1e95de04e2366d19b5448.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 282, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 354 + ], + "type": "text", + "content": "The \"flexible\" in the flexible identity generator lies in the ability to flexibly control conditions during inference, enabling the generation of versatile and diverse identities with varying characteristics. We provide more ablation studies on designing hard identities in section B of supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 214, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 214, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 214, + 373 + ], + "type": "text", + "content": "3.3. Cross-camera Diversification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "text", + "content": "Enhancing cross-camera variations within each identity is crucial for improving the generalization ability of ReID models across different camera views. However, previous synthetic methods struggle to scale up cross-camera variations, due to the high manual effort required in designing diverse scenes or limited variation in training data. In this subsection, we aim to expand the intra-identity (intra-ID) group " + }, + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "text", + "content": " for each identity with scalable cross-camera variations, specifically in terms of poses and scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 487, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 569 + ], + "type": "text", + "content": "Data Collection. First, we collect real-world pedestrian images, ensuring a diverse range of poses and scenes that reflect the variations encountered in different camera setups. Unless otherwise specified, our poses and scenes come from LUPerson [12], a large-scale unlabeled pedestrian dataset with over 4.18 million images collected across 46,260 scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": "Pose Variations. Second, we extract person poses in the collected images with openpose[5, 6, 39, 45, 56]. Then we utilize ControlNet-based IP-adapter[51, 56] to modify each virtual identity's pose with the randomly collected poses as the condition. However, significant pose variations may lead to identity degradation. To enhance identity consistency, we generate a corresponding identity description for each individual using BLIP. This description, along with the collected pose, is then used as control conditions for the IP-Adapter to guide the generation of the new image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Scene Variations. Lastly, we propose segment-and-paste strategy to modify the scene of images. For the collected" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 555, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 274 + ], + "type": "text", + "content": "real-world images, we first apply the inpainting diffusion technique [53] to remove the original pedestrian, creating clean scene images. We then utilize the semantic segmentation method[8] to segment our generated pose-changed persons and patch them into the scene images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 283, + 416, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 416, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 416, + 296 + ], + "type": "text", + "content": "3.4. Style Realization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 542 + ], + "type": "text", + "content": "Camera-realistic styles, such as illumination, motion blur, camera bias, and environmental conditions, provide crucial contextual information that reflect real-world scenarios, playing a key role in improving the transferability of ReID models. Engine-based synthetic images inherently lack realistic styles, as they are created using virtual engines that do not capture real-world environmental nuances. Previous learning-based generative models learn realistic styles while embedding identity-related features, leading to generated images highly consistent with the real identities of the real-world ReID dataset, which raises privacy concerns. To address these, unlike learning the realistic distribution during training, we propose a post-processing approach, leveraging the neural style transfer method [9] to inject the style. Specifically, we randomly select a pedestrian image from LUPerson as the reference image, and leverage [9] to separate the style and content of the real-world image. Then, for each generated virtual image, we transfer the identity-agnostic style from the real-world image to the virtual image while preserving its original virtual identity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 499, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 499, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 499, + 564 + ], + "type": "text", + "content": "3.5. Summary of the VIPerson Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "The entire data generation pipeline is illustrated in Figure 3, which can generate a ReID dataset composed entirely of virtual identities without manual annotation. We generate the virtual dataset VIPerson, which serves as the primary dataset in all our experiments unless otherwise specified. Refer to Figure 3 and section A of supplementary material for examples of generated images. The detailed parameters of the datasets are shown in Table 1. Furthermore, we extend our approach to incorporate scenario-specific patterns, generating VIPerson* with available pose, scene and style information from target scenarios to enhance ReID performance in the targeted applications." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23378" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 180, + 186 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 180, + 186 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 180, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 180, + 186 + ], + "type": "image", + "image_path": "9d01b3e8a6ad9386754c23bdd82f51a35004e611debd3857c2f4ba4adee31368.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "lines": [ + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "type": "text", + "content": "Figure 4. Effect of different numbers of identities. \"Ours w/o Hard\" does not contain hard identities generated by " + }, + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "type": "inline_equation", + "content": "c_d" + }, + { + "bbox": [ + 55, + 189, + 295, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 72, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 181, + 72, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 181, + 72, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 181, + 72, + 294, + 187 + ], + "type": "image", + "image_path": "609ba3a32a9901921eca21f31e3fe2d5b18f108ed27de4780d72a2dbc9d38894.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 284, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 284, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 284, + 247 + ], + "type": "text", + "content": "4. Analysis of Key Factors in Data Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 253, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 296, + 384 + ], + "type": "text", + "content": "The synthesized pedestrian data is a fundamental component of downstream ReID tasks. In this study, we conduct an in-depth experimental analysis of several key factors in the generated dataset VIPerson: virtual identity diversity (including identity diversity and hard identity), cross-camera variations (including intra-ID group scale, pose variations, and scene variations), and camera-realistic style. We employ direct transfer performance on real datasets with BoT[26] to assess the quality of generated data, as direct transfer serves as the foundation for all other tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 197, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 197, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 197, + 408 + ], + "type": "text", + "content": "4.1. Virtual Identity Diversity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 413, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 413, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 413, + 295, + 604 + ], + "type": "text", + "content": "Identity Diversity. We explore the effect of the identity diversity by changing identity scale as shown in Figure 4. As the number of identities increases from 1000 to 3000, the rank-1 on MSMT17 and Market1501 has increased rapidly. It demonstrates that our flexible identity generator can generate scalable identities with diverse information, facilitating the model to better discriminate between unseen persons in the test set. However, as the identity scale increases further, the model performance grows slowly, especially on Market1501. This limitation may contribute to the scaling law, which is summarised in some studies[11, 15, 18] in many tasks: Even with an infinite amount of data, the learning capacity of ReID model is still constrained by other factors such as model capacity. More analysis between data scale and model size refers to section D in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Hard Identity. We explore the impact of hard identities on model performance as shown in Figure 4. For example, at an identity count of 5000, 'ours' outperforms 'ours-w/o Hard' on both test datasets. This suggests that hard identities in datasets allow the ReID model to learn to focus on more detailed discriminative features, improving the ability to distinguish between similar persons. In contrast to previous synthetic datasets that require manual design of hard identities, our approach generates hard identities automati-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 438, + 186 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 438, + 186 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 438, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 438, + 186 + ], + "type": "image", + "image_path": "ae141a1f85e1ce399c8920b3f9c3d5125d49054cefd6bf615db886288919e8c5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 553, + 187 + ], + "type": "image", + "image_path": "55a7e2fd7353f6ae8e17ef713ae5f915ac90ab417c39242b7b90471746d352f7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 187, + 525, + 198 + ], + "lines": [ + { + "bbox": [ + 342, + 187, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 342, + 187, + 525, + 198 + ], + "type": "text", + "content": "Figure 5. Effect of different Intra-ID group scales." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 552, + 335 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 552, + 335 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 552, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 552, + 335 + ], + "type": "image", + "image_path": "0f01b1a54a66d3d0f2a5fef843851fc9b34142c05005725c5221892d607bbd97.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 344, + 554, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 344, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 554, + 378 + ], + "type": "text", + "content": "Figure 6. Ablation study of modules in VIPerson. 'PV' indicates pose variations, 'SV' refers to scene variations, and 'SR' is style realization. Each point reports the rank-1 performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 550, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 550, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 550, + 411 + ], + "type": "text", + "content": "cally, making the process more cost-effective and scalable." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 420, + 453, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 453, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 453, + 432 + ], + "type": "text", + "content": "4.2. Cross-camera Variations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": "Intra-ID Group Scale. Cross-camera variations are positively correlated with the scale of intra-ID group. This is because as the intra-ID group increases, there is a larger variety of viewpoints, poses, and scenes for each identity, which is introduced via our random sampling process. Here, we demonstrate how the scale of the intra-ID group affects the performance of the ReID model. As the results shown in Figure 5, with the increment of our intra-ID group scale, the performance continues to increase on both test datasets. Besides, our VIPerson is consistently superior to UnrealPerson, which demonstrates that our cross-camera variations are more diverse than UnrealPerson. Benefited from our proposed cross-camera diversification that introduces scalable and diverse poses and scenes without relying on manual design, we can expand the intra-ID group to any scale for better performance at a low cost." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Pose Variations. To investigate the impact of pose variations, we minimise the variation of pose introduced through cross-camera diversification, as shown in the 'SV+SR' column of Figure 6. The rank-1 performance drops from " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "32.5\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "20.5\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " on MSMT17, and from " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "66.5\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "39.8\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " on Market1501. In the absence of pose diversity, the model may overfit to specific pose-related features, which hin" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23379" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "ders its ability to generalize across different viewing angles. Since our pose set is randomly collected from real pedestrian images, the poses of generated persons are natural and varied. Our results demonstrate that pose variations in training data plays a crucial role in improving the model's robustness and performance in real-world scenarios, where pedestrians are often captured from various angles and perspectives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "text", + "content": "Scene Variations. As shown in the column 'PV', Without scene variations, the performance is limited to " + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "text", + "content": " on MSMT17 and " + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "text", + "content": " on Market1501. Introducing scene variations significantly boosts performance to " + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "text", + "content": " on MSMT17 and " + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "inline_equation", + "content": "33.6\\%" + }, + { + "bbox": [ + 56, + 168, + 296, + 384 + ], + "type": "text", + "content": " on Market1501 as shown in the column 'PV+SV', indicating that scene variations are crucial for the model's effectiveness. Since the background in each image of the same pedestrian varies in the test data, incorporating diverse backgrounds into the training set can substantially enhance performance. Previous approaches, which rely on both synthetic and real datasets sampled from sequential camera videos, are limited in terms of scene variations within each intra-ID group. It would be costly for them to capture more videos to expand the variation in scenes. In contrast, our segment-and-paste strategy allows the Intra-ID group to exhibit richer background variations with a significantly lower cost. More analysis of scenes refer to the section E in supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 393, + 183, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 183, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 183, + 406 + ], + "type": "text", + "content": "4.3. Camera-realistic Style" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "text", + "content": "Style Realization. As shown in column 'PV+SV+SR' of Figure 6, introducing camera-realistic styles significantly improves rank-1 from " + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "inline_equation", + "content": "33.6\\%" + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "inline_equation", + "content": "66.5\\%" + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "text", + "content": " on Market1501 and from " + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "inline_equation", + "content": "32.5\\%" + }, + { + "bbox": [ + 55, + 411, + 295, + 555 + ], + "type": "text", + "content": " on MSMT17. This highlights the importance of style realization, as it bridges the gap between synthetic and real-world images. By transferring illumination, motion blur, camera bias, and environmental conditions from real reference images, our style realization process makes the synthetic images more consistent with real-world data distributions. This not only enhances model performance but also improves generalization when deployed in diverse real-world environments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 567, + 198, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 198, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 198, + 580 + ], + "type": "text", + "content": "5. Downstream ReID Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 588, + 189, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 189, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 189, + 601 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Details of Synthetic Data. Existing synthetic ReID datasets include SyRI[1], PersonX[41], RandPerson[42], and UnrealPerson[57], and their details are shown in Table 1. Our dataset is named VIPerson in this paper. We provide synthetic datasets for two scenarios, including the dataset without target scenario information (denoted as 'VIPerson') and the datasets based on target scenario information (denoted as 'VIPerson*'). We test the model on MSMT17 [44] and Market1501 [59]. For VIPerson, pose," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 325, + 127, + 542, + 232 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 125 + ], + "type": "text", + "content": "Table 2. Direct transfer results with BoT[26]. The fully-supervised results are shown in italics, the best direct transfer results are indicated in bold, and the second-best results are underlined. More experiments on different ReID models are provided in section C of supplementary material." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 127, + 542, + 232 + ], + "lines": [ + { + "bbox": [ + 325, + 127, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 325, + 127, + 542, + 232 + ], + "type": "table", + "html": "
Dataset TypeTraining DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealMSMT1773.949.956.229.8
Market15019.23.094.686.3
SyntheticMALS0.20.64.91.5
SyRI7.81.615.24.9
PersonX7.62.436.815.5
RandPerson14.54.855.929.9
UnrealPerson26.89.964.037.2
VIPerson32.510.366.538.3
VIPerson*34.310.867.537.6
", + "image_path": "b71ea0572362015b0f34f7ad40ba76d2dc032b46ab8dff6fea57341b2bedfba4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 320, + 257, + 547, + 344 + ], + "blocks": [ + { + "bbox": [ + 321, + 244, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 321, + 244, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 321, + 244, + 545, + 255 + ], + "type": "text", + "content": "Table 3. Supervised fine-tuning results with BoT[26] method." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 257, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 320, + 257, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 320, + 257, + 547, + 344 + ], + "type": "table", + "html": "
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet72.542.491.476.8
MSMT17--93.782.5
Market150174.144.5--
SyntheticMALS28.114.474.854.7
UnrealPerson74.546.094.084.7
VIPerson76.551.594.085.3
VIPerson*76.751.295.186.9
", + "image_path": "9faeae57951eca6e0d6ebb2b55a53d3b6ddddbeb6737d9a332ff9d68cc27cbd1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 457 + ], + "type": "text", + "content": "background, and style information is consistently from LUPerson [12] across all tested datasets. VIPerson consists of 5000 identities including hard identities. For VIPerson*, when testing on MSMT17, we obtain pose, scene and style information from MSMT17 training set. When testing on Market1501, we obtain pose and scene information from LUPerson and MSMT17, and transfer style from Market1501 training set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "content": "Training Details. We conduct our experiments and analysis primarily using ResNet50[14] as the backbone model. Specifically, for the direct transfer experiment, we adopt BoT[26] and resize the image to " + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "inline_equation", + "content": "256 \\times 128" + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "content": " and set the batch size as 128. We adopt the Adam optimizer for training, with the base learning rate set to " + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "inline_equation", + "content": "3.5 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "content": " and decreased by a factor of 0.1 at the " + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "inline_equation", + "content": "10^{th}" + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "inline_equation", + "content": "40^{th}" + }, + { + "bbox": [ + 313, + 459, + 554, + 566 + ], + "type": "text", + "content": " epochs. Details about fine-tuning and domain adaptation experiments refer to section F in supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 575, + 410, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 575, + 410, + 587 + ], + "spans": [ + { + "bbox": [ + 314, + 575, + 410, + 587 + ], + "type": "text", + "content": "5.2. Direct Transfer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "To evaluate the quality of generated data, we train the models only on generated data and test them on real data. As shown in Table 2, our VIPerson surpass all other datasets. Unlike training data from real-world, VIPerson consists entirely of virtual identities, thus avoiding privacy concerns. Nevertheless, VIPerson still exceeds Market1501 by " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "23.3\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " on MSMT17 and outperforms MSMT17 by " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " on Market1501. Compared to the previous synthetic datasets, our VIPerson significantly improves performance, as it achieve three key factors in synthetic person data as analysis in Sec-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23380" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 83, + 288, + 162 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 290, + 82 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 290, + 82 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 290, + 82 + ], + "type": "text", + "content": "Table 4. Unsupervised domain adaptation results with ISE[58]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 83, + 288, + 162 + ], + "lines": [ + { + "bbox": [ + 61, + 83, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 83, + 288, + 162 + ], + "type": "table", + "html": "
Dataset TypePre-training DatasetMSMT17Market1501
rank-1mAPrank-1mAP
RealImageNet64.735.094.084.7
MSMT17--93.884.2
Market150160.431.2--
SyntheticRandPerson59.030.293.483.8
UnrealPerson65.936.194.586.3
VIPerson70.439.594.986.4
", + "image_path": "340ee758020e16a1e17defb004bd47670cc5540f81afef741d742a369f549bc0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 176, + 295, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 176, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 295, + 198 + ], + "type": "text", + "content": "tion 4, allowing the ReID model to generalize better to unseen scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 199, + 296, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 296, + 282 + ], + "type": "text", + "content": "Moreover, our cross-camera diversification and style realization are flexible to adapt to the target scenarios. Given scene and style information from the target scenario, we leverage them to develop VIPerson*, which further enhances performance on target datasets. This makes our method more practical and adaptable for real-world applications under diverse scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 188, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 188, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 188, + 303 + ], + "type": "text", + "content": "5.3. Supervised Fine-tuning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": "When sufficient labeled data is available, pretraining the model on synthetic data and fine-tuning it on target datasets can enhance performance. As shown in Table 3, with pretrained on our datasets, the rank-1 performance reaches " + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "inline_equation", + "content": "76.7\\%" + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": " on MSMT17 and " + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "inline_equation", + "content": "95.1\\%" + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": " on Market1501, suppressing all the other datasets. These results demonstrate the diversity, strong feature representation, and ample scale of our dataset, underscoring its value as a powerful resource for enhancing fine-tuning performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 422, + 237, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 237, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 237, + 434 + ], + "type": "text", + "content": "5.4. Unsupervised Domain Adaptation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": "In most application scenarios, pedestrian data is usually unlabeled due to the limitations of annotation costs. In these cases, leveraging other labeled datasets for pretraining and fine-tuning on the unlabeled data can also improve performance. From Table 4, we see that the model pre-train on our datasets suppresses all the other datasets, achieving " + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "inline_equation", + "content": "70.4\\%" + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": " on MSMT17 and " + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "inline_equation", + "content": "94.9\\%" + }, + { + "bbox": [ + 55, + 438, + 296, + 559 + ], + "type": "text", + "content": " on Market1501. Our VIPerson effectively enhance the performance without relying on any manual annotations, which underscores its significance for reducing annotation dependency in real-world applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 564, + 216, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 216, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 216, + 577 + ], + "type": "text", + "content": "5.5. Identity Expansion Scenarios" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "As rising privacy concerns, we consider a scenario with few label identities, where most individuals are unwilling to share their personal image as training data, leading to a scarcity of labeled identities for training ReID models. Fortunately, we can add generated data to enhance the identity diversity. To this end, we simulate environments where only a small fraction of labeled data (20 or 40 identities) is available, exploring how ReID models perform with limited labeled identities and expended synthetic data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "As shown in Figure 7, when only 20 real identities are available, adding real data from other domains (MK) only" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 438, + 160 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 438, + 160 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 438, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 438, + 160 + ], + "type": "image", + "image_path": "10b788ab478551c3d2c5348ff66a4289980f9e5c5ae02ae172d79d33178d8915.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 550, + 160 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 550, + 160 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 550, + 160 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 550, + 160 + ], + "type": "image", + "image_path": "ce09a8d147a203610f97064c6f472475047543475f1000659456f2ba8aeab11b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 160, + 438, + 251 + ], + "blocks": [ + { + "bbox": [ + 317, + 160, + 438, + 251 + ], + "lines": [ + { + "bbox": [ + 317, + 160, + 438, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 160, + 438, + 251 + ], + "type": "image", + "image_path": "f0cd81c33b169fdb4a3ecbbc44dbc0c10eeb2699452dba4eeba48f1a32cb0f35.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 251, + 466, + 259 + ], + "lines": [ + { + "bbox": [ + 422, + 251, + 466, + 259 + ], + "spans": [ + { + "bbox": [ + 422, + 251, + 466, + 259 + ], + "type": "text", + "content": "Market1501" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 441, + 160, + 551, + 250 + ], + "blocks": [ + { + "bbox": [ + 441, + 160, + 551, + 250 + ], + "lines": [ + { + "bbox": [ + 441, + 160, + 551, + 250 + ], + "spans": [ + { + "bbox": [ + 441, + 160, + 551, + 250 + ], + "type": "image", + "image_path": "5769d1002c807a0dd3531487f99fc4604edb7852e00369e1d2485b532d30428b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 268, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 322 + ], + "type": "text", + "content": "Figure 7. Results under identity expansion scenarios. 'R' indicates training with real images from the testing scenarios, 'MK' refers to training set of Market1501, 'MS' corresponds to training set of MSMT17, 'UP' denotes UnrealPerson, and 'VIP' represents our VIPerson dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 349, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 517 + ], + "type": "text", + "content": "improves rank-1 by " + }, + { + "bbox": [ + 313, + 349, + 555, + 517 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 349, + 555, + 517 + ], + "type": "text", + "content": " on MSMT17 test set. Although collected from the real world, it is constrained by limited identity and cross-camera variations, which hinders its ability to effectively enhance model performance. As a comparison, adding VIPerson significantly improves rank-1 performance on MSMT17 by " + }, + { + "bbox": [ + 313, + 349, + 555, + 517 + ], + "type": "inline_equation", + "content": "30.4\\%" + }, + { + "bbox": [ + 313, + 349, + 555, + 517 + ], + "type": "text", + "content": ", which effectively bridge the gap caused by the lack of real-world data. These results underscore the practical applicability of our approach in scenarios with few label identities, making it possible to train effective ReID models with minimal privacy exposure to real identity. In environments where privacy is a critical concern, such as surveillance systems and smart cities, this approach can be a game-changer for identity verification and tracking systems." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 535, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 388, + 548 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "content": "In this work, we introduce VIPerson, a novel pedestrian generation pipeline for flexibly creating virtual identities for the person ReID task to address privacy concerns and annotations consumptions in real ReID dataset. Our work identifies three key factors that enhance the effectiveness of the synthesized dataset, offering valuable insights for future directions in data synthesis. Moreover, we validated the scaling law, which shows that performance improves with an increase in the scale of synthesized data and model parameters. However, we also observed a performance bottleneck at a certain data scale, which suggests that future work should focus on how to effectively leverage synthetic data to enhance ReID models." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23381" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported partially by NSFC(U21A20471, 92470202), by the National Key Research and Development Program of China (2023YFA1008503), Guangdong NSF Project (No. 2023B1515040025)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "text", + "content": "[1] Slawomir Bak, Peter Carr, and Jean-Francois Lalonde. Domain adaptation through synthesis for unsupervised person re-identification. In Proceedings of the European conference on computer vision (ECCV), pages 189–205, 2018. 1, 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 214, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 296, + 267 + ], + "type": "text", + "content": "[2] Lucas Bourtoule, Varun Chandrasekaran, Christopher A Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine unlearning. In 2021 IEEE symposium on security and privacy (SP), pages 141-159. IEEE, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 270, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 296, + 323 + ], + "type": "text", + "content": "[3] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 296, + 357 + ], + "type": "text", + "content": "[4] Jonathan Brophy and Daniel Lowd. Machine unlearning for random forests. In International Conference on Machine Learning, pages 1092-1104. PMLR, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 390 + ], + "type": "text", + "content": "[5] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 435 + ], + "type": "text", + "content": "[6] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 500 + ], + "type": "text", + "content": "[7] Hao Chen, Yaohui Wang, Benoit Lagadec, Antitza Dantcheva, and Francois Bremond. Joint generative and contrastive learning for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2004–2013, 2021. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "type": "text", + "content": "[8] Xiangguang Chen, Ye Zhu, Yu Li, Bingtao Fu, Lei Sun, Ying Shan, and Shan Liu. Robust human matting via semantic guidance. In Proceedings of the Asian Conference on Computer Vision, pages 2984-2999, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 547, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 296, + 601 + ], + "type": "text", + "content": "[9] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 624 + ], + "type": "text", + "content": "[10] Makehuman community. Makehuman. http://www.makehumancommunity.org/. 2024.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 679 + ], + "type": "text", + "content": "[11] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[12] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceed" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750-14759, 2021. 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 554, + 140 + ], + "type": "text", + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 142, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 554, + 185 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 239 + ], + "type": "text", + "content": "[15] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 554, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 554, + 275 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 277, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 554, + 297 + ], + "type": "text", + "content": "[17] Erin Illman and Paul Temple. California consumer privacy act. The Business Lawyer, 75(1):1637-1646, 2019. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 300, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 554, + 354 + ], + "type": "text", + "content": "[18] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 356, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 356, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 356, + 554, + 421 + ], + "type": "text", + "content": "[19] Inès Hyeonsu Kim, JoungBin Lee, Woojeong Jin, Soowon Son, Kyusun Cho, Junyoung Seo, Min-Seop Kwak, Seokju Cho, JeongYeol Baek, Byeongwon Lee, et al. Pose-dive: Pose-diversified augmentation with diffusion model for person re-identification. arXiv preprint arXiv:2406.16042, 2024. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 422, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 554, + 498 + ], + "type": "text", + "content": "[20] Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C.H. Hoi. LAVIS: A one-stop library for language-vision intelligence. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 31-41, Toronto, Canada, 2023. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "text", + "content": "[21] Hanjun Li, Gaojie Wu, and Wei-Shi Zheng. Combined depth space based architecture search for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6729–6738, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 589 + ], + "type": "text", + "content": "[22] He Li, Mang Ye, and Bo Du. Weperson: Learning a generalized re-identification model from all-weather virtual data. In Proceedings of the 29th ACM international conference on multimedia, pages 3115-3123, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI conference on artificial intelligence, pages 1405–1413, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[25] Kun-Yu Lin, Henghui Ding, Jiaming Zhou, Yu-Ming Tang, Yi-Xing Peng, Zhilin Zhao, Chen Change Loy, and Wei" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23382" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "Shi Zheng. Rethinking clip-based video learners in cross-domain open-vocabulary action recognition. arXiv preprint arXiv:2403.01560, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[26] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 196 + ], + "type": "text", + "content": "[27] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Jihong Guan, and Shuigeng Zhou. Fine-grained zero-shot object detection. arXiv preprint arXiv:2507.10358, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 263 + ], + "type": "text", + "content": "[28] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Jiazhen Ji, Jianqing Xu, Jun Wang, Shaoming Wang, Shouhong Ding, and Shuigeng Zhou. Privacy-preserving face recognition using trainable feature subtraction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 297-307, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 296, + 331 + ], + "type": "text", + "content": "[29] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21203-21214, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "type": "text", + "content": "[30] Yi-Xing Peng, Jile Jiao, Xuetao Feng, and Wei-Shi Zheng. Consistent discrepancy learning for intra-camera supervised person re-identification. IEEE Transactions on Multimedia, 25:2393–2403, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 421 + ], + "type": "text", + "content": "[31] Yi-Xing Peng, Yuanxun Li, and Wei-Shi Zheng. Revisiting person re-identification by camera selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5): 2692-2708, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 466 + ], + "type": "text", + "content": "[32] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, and Wei-Shi Zheng. Generalized intra-camera supervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4516-4527, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 296, + 521 + ], + "type": "text", + "content": "[33] Yi-Xing Peng, Yu-Ming Tang, Kun-Yu Lin, Qize Yang, Jingke Meng, Xihan Wei, and Wei-Shi Zheng. Person deidentification: A variation-guided identity shift modeling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29331-29341, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 588 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 612 + ], + "type": "text", + "content": "[35] Protection Regulation. General data protection regulation. Intouch, 25:1-5, 2018. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 657 + ], + "type": "text", + "content": "[36] John Ricciello. John ricciello sets out to identify the engine of growth for unity technologies (interview). VentureBeat. Interview with Dean Takahashi. Retrieved January, 18 (3), 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 714 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 172 + ], + "type": "text", + "content": "[39] Tomas Simon, Hanbyul Joo, Iain Matthews, and Yaser Sheikh. Hand keypoint detection in single images using multiview bootstrapping. In CVPR, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 205 + ], + "type": "text", + "content": "[40] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 250 + ], + "type": "text", + "content": "[41] Xiaoxiao Sun and Liang Zheng. Dissecting person re-identification from the viewpoint of viewpoint. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 608-617, 2019. 1, 3, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "type": "text", + "content": "[42] Yanan Wang, Shengcai Liao, and Ling Shao. Surpassing real-world source training data: Random 3d characters for generalizable person re-identification. In Proceedings of the 28th ACM international conference on multimedia, pages 3422–3430, 2020. 1, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 555, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 555, + 360 + ], + "type": "text", + "content": "[43] Yanan Wang, Xuezhi Liang, and Shengcai Liao. Cloning outfits from real-world images to 3d characters for generalizable person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4900–4909, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 555, + 415 + ], + "type": "text", + "content": "[44] Longhui Wei, Shiliang Zhang, Wen Gao, and Qi Tian. Person transfer gan to bridge domain gap for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 79–88, 2018. 1, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 555, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 437 + ], + "type": "text", + "content": "[45] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 555, + 492 + ], + "type": "text", + "content": "[46] Xiao-Ming Wu, Dian Zheng, Zuhao Liu, and Wei-Shi Zheng. Estimator meets equilibrium perspective: A rectified straight through estimator for binary neural networks training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17055-17064, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 555, + 536 + ], + "type": "text", + "content": "[47] Yuqiao Xian, Yi-Xing Peng, Xing Sun, and Wei-Shi Zheng. Distilling consistent relations for multi-source domain adaptive person re-identification. Pattern Recognition, 157: 110821, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 555, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 582 + ], + "type": "text", + "content": "[48] Suncheng Xiang, Yuzhuo Fu, Guanjie You, and Ting Liu. Unsupervised domain adaptation through synthesis for person re-identification. In 2020 IEEE international conference on multimedia and expo (ICME), pages 1–6. IEEE, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 646 + ], + "type": "text", + "content": "[49] Suncheng Xiang, Dahong Qian, Mengyuan Guan, Binjie Yan, Ting Liu, Yuzhuo Fu, and Guanjie You. Less is more: Learning from synthetic data with fine-grained attributes for person re-identification. ACM Transactions on Multimedia Computing, Communications and Applications, 19(5s): 1-20, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "text", + "content": "[50] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23383" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 675 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[51] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[52] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 216 + ], + "type": "text", + "content": "[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[54] Delong Zhang, Yixing Peng, Ancong Wu, and Wei-shi Zheng. Privacy-preserving face recognition with adaptive generative perturbations. In International Conference on Pattern Recognition, pages 210-227. Springer, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[55] Delong Zhang, Yi-Xing Peng, Xiao-Ming Wu, Ancong Wu, and Wei-Shi Zheng. Pixelfade: Privacy-preserving person re-identification with noise-guided progressive replacement. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6326–6334, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 341 + ], + "type": "text", + "content": "[56] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 407 + ], + "type": "text", + "content": "[57] Tianyu Zhang, Lingxi Xie, Longhui Wei, Zijie Zhuang, Yongfei Zhang, Bo Li, and Qi Tian. Unrealperson: An adaptive pipeline towards costless person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11506–11515, 2021. 1, 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 473 + ], + "type": "text", + "content": "[58] Xinyu Zhang, Dongdong Li, Zhigang Wang, Jian Wang, Errui Ding, Javen Qinfeng Shi, Zhaoxiang Zhang, and Jingdong Wang. Implicit sample extension for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7369–7378, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "type": "text", + "content": "[59] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 521, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 563 + ], + "type": "text", + "content": "[60] Wei-Shi Zheng, Junkai Yan, and Yi-Xing Peng. A versatile framework for multi-scene person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "text", + "content": "[61] Zhedong Zheng, Liang Zheng, and Yi Yang. Unlabeled samples generated by gan improve the person re-identification baseline in vitro. In Proceedings of the IEEE international conference on computer vision, pages 3754–3762, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 294, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 294, + 675 + ], + "type": "text", + "content": "[62] Zhizhou Zhong, Yuxi Mi, Yuge Huang, Jianqing Xu, Guodong Mu, Shouhong Ding, Jingyun Zhang, Rizen Guo, Yunsheng Wu, and Shuigeng Zhou. Slerpface: Face template protection via spherical linear interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10698-10706, 2025. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23384" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_content_list.json b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..92def7b4abb56da9e80fce8b8756ec2ddf146fd4 --- /dev/null +++ b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_content_list.json @@ -0,0 +1,1514 @@ +[ + { + "type": "text", + "text": "VISION-XL: High Definition Video Inverse Problem Solver using Latent Image Diffusion Models", + "text_level": 1, + "bbox": [ + 99, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taesung Kwon KAIST", + "bbox": [ + 316, + 204, + 439, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "star.kwon@kaist.ac.kr", + "bbox": [ + 284, + 242, + 472, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jong Chul Ye KAIST", + "bbox": [ + 570, + 204, + 679, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jong.ye@kaist.ac.kr", + "bbox": [ + 539, + 242, + 707, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4947b8efca9a41483d9f71670b69062e99f2294903e7bffe05d68a5d4a7b6fe1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 280, + 254, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78fcfa979d5a60d395e874e72e311c5410babb0219ba2b2f7445241fc337c2e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 280, + 580, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4ad9f802ad006a053e6cd5238ca3d4b21170db5e6c4a1181f70a41ac55b99cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 280, + 903, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25eb6fc1b7c688902cfb16896f58fb94ff38248e61afa69997a398367ba0678b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 421, + 598, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df863e580e5b2d7e34f66403acdbb80adaab488c926ee9b4c09a4ace5b14e32c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 421, + 903, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2dd0fdd5b8bb0736056be69ea844372f7ef9c3800d4474874e0a28182658e281.jpg", + "image_caption": [ + "Figure 1. Representative video reconstruction by VISION-XL: $\\mathrm{SR + }$ (frame averaging with $\\times 4$ super-resolution, top), Deblur+ (frame averaging with deblurring, $\\sigma = 3.0$ , bottom-left), and Inpaint+ (frame averaging with $50\\%$ random inpainting, bottom-right)." + ], + "image_footnote": [], + "bbox": [ + 93, + 501, + 598, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87fd66dd26cfa1170791243348284b967be8a511e4925c853c23727c8c3bc4b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 501, + 903, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 627, + 326, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models. Building on recent advancements in spatio-temporal optimization for video inverse problems using image diffusion models, our approach leverages latent-space diffusion models to achieve enhanced video quality and resolution. To address the high computational demands of processing high-resolution frames, we introduce a pseudo-batch consistent sampling strategy, allowing efficient operation on a single GPU. Additionally, to improve temporal consistency, we present pseudo-batch inversion, an initialization technique that incorporates informative latents from the measurement. By integrating with SDXL, our framework achieves state-of-the-art video reconstruction across a wide range of spatio-temporal inverse problems, including complex combinations of frame averaging", + "bbox": [ + 88, + 659, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and various spatial degradations, such as deblurring, super-resolution, and inpainting. Unlike previous methods, our approach supports multiple aspect ratios (landscape, vertical, and square) and delivers HD-resolution reconstructions (exceeding $1280 \\times 720$ ) in under 6 seconds per frame on a single NVIDIA 4090 GPU. Project page: https://vision-x1.github.io/.", + "bbox": [ + 511, + 628, + 906, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 753, + 645, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [6, 9, 10, 16, 18, 20, 22] have set a new benchmark in generative modeling, enabling the generation of high-quality samples. These models have become the foundation for advancements in various fields, such as controllable image editing [33], image personalization [8], synthetic data augmentation [23], and even reconstructing images from brain signals [14, 24].", + "bbox": [ + 509, + 777, + 906, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, diffusion model-based inverse problem", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10465", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/5a52b3ed258d1244730990fda8dfe133323d4684b10e2549b8858ff7ce07004f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSpatio-temporal degradationLatent space
Optical flow-based [5, 32]X✓
SVI [13]✓X
Ours✓✓
", + "bbox": [ + 94, + 87, + 478, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "solvers (DIS) [2, 4, 11, 21, 27, 29] address a variety of image restoration tasks, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on. A key feature of DIS is its plug-and-play capability, allowing diffusion models to be applied flexibly across different inverse problems without requiring task-specific training or fine-tuning.", + "bbox": [ + 88, + 205, + 482, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, several extensions [5, 13, 32] from the DIS have been proposed to solve video inverse problems using image diffusion models. Naive application of image diffusion models to videos may break temporal consistency. To address this problem, these methods preserve temporal consistency by utilizing a batch-consistent sampling strategy [13] and applying optical flow guidance to warp either latent representations [32] or the noise prior [5].", + "bbox": [ + 88, + 311, + 482, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these innovative approaches enable powerful image generative models [6, 16, 18] to solve video inverse problems with significantly reduced computational requirements, there is still room for improvement in these methods. Optical flow-based methods [5, 32] have reported a key limitation: their performance is highly dependent on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes problematic when extreme degradations complicate the estimation process, restricting their applicability to a wider range of restoration tasks. Additionally, these methods require task-specific restoration modules [32] or fine-tuning of the diffusion model [5]. In this perspective, batch consistent sampling strategy [13] successfully addressed various spatio-temporal degradations without requiring task-specific training or fine-tuning. However, we empirically found that extending SVI [13] to latent diffusion models leads to unsatisfactory reconstruction, particularly in terms of FVD [26], as shown in Table 2, despite the fact that most modern latent diffusion models are essential for scaling up to large-size video inverse problems.", + "bbox": [ + 91, + 433, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, here we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models that involves no additional training. To address the high computational demands of batch processing (e.g., 16-frame batch processing used in [13]) with high-resolution latent diffusion models [16], we introduce pseudo-batch consistent sampling. This strategy enables multi-frame video processing while requiring only the memory needed for a single frame, making it feasible on a single GPU. Furthermore, we propose pseudo-batch inversion, which initializes the process with", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/66e6936327aaba5fafeafd638112e60a1cd98b765a8b2152e23c89aea453cdff.jpg", + "table_caption": [ + "Table 1. Comparison of image diffusion-based video inverse problem solvers. Unlike other methods, our approach leverages latent image diffusion models to address spatio-temporal degradation, enabling more effective restoration." + ], + "table_footnote": [], + "table_body": "
InitializationFVD ↓LPIPS ↓PSNR ↑
Random noise10470.25129.43
Batch synchronized noise (SVI [13])707.70.24830.10
Pseudo-batch inversion (Ours)184.80.23630.74
", + "bbox": [ + 535, + 88, + 880, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 2. Impact of our initialization method on SR+ video restoration using SDXL [16]. Our method significantly improves performance, reducing FVD [26] by more than $3\\times$", + "bbox": [ + 511, + 152, + 903, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "informative latents derived from the measurement. This initialization enhances temporal consistency and improves the efficiency of solving spatio-temporal inverse problems as shown in Table. 2.", + "bbox": [ + 511, + 205, + 903, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By integrating these components, our framework achieves state-of-the-art video reconstruction performance using SDXL [16]. We name the method integrating these components as VISION-XL, short for Video Inverse-problem Solver using latent diffusION models (with stable diffusion XL). It supports various aspect ratios, including landscape, vertical, and square formats. Thanks to its efficiency, our framework can reconstruct $1280 \\times 768$ (exceeding HD resolution) videos in under 6 seconds per frame on a single NVIDIA 4090 GPU. Our contribution can be summarized as follows:", + "bbox": [ + 511, + 267, + 906, + 431 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a high-definition video inverse problem solver integrated with SDXL, supporting multiple aspect ratios and achieving state-of-the-art reconstruction.", + "- We introduce a novel pseudo-batch consistent sampling and inversion strategy for efficient and effective video reconstruction across diverse inverse problems." + ], + "bbox": [ + 511, + 434, + 903, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 541, + 653, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion model-based inverse problem solvers (DIS). Diffusion models [9, 20, 22] attempt to model the data distribution $p_{\\theta}(\\pmb{x})$ based on the Gaussian transitions. In the geometric view of diffusion models [1], the transitions are typically described as iterative manifold transitions $\\mathcal{M}_t \\to \\mathcal{M}_{t-1}, t = T, \\dots, 1$ , moving from the noisy manifold $\\mathcal{M}_T$ to the clean manifold $\\mathcal{M}_0$ .", + "bbox": [ + 511, + 566, + 903, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion model-based inverse problem solvers (DIS) [2, 4, 11, 21, 27] aim to guide manifold transitions to sample from the posterior distribution $p_{\\theta}(\\boldsymbol{x}|\\boldsymbol{y})$ , which represents sampling $\\boldsymbol{x}$ from the measurement $\\boldsymbol{y}$ obtained from the forward model $\\mathcal{A}(\\boldsymbol{x})$ . In Bayesian inference, the posterior distribution, $p_{\\theta}(\\boldsymbol{x}|\\boldsymbol{y}) \\propto p_{\\theta}(\\boldsymbol{x})p(\\boldsymbol{y}|\\boldsymbol{x})$ is decomposed into the likelihood $p(\\boldsymbol{y}|\\boldsymbol{x})$ , representing the probability of observing $\\boldsymbol{y}$ given $\\boldsymbol{x}$ , and the prior data distribution $p_{\\theta}(\\boldsymbol{x})$ . This decomposition enables posterior sampling by combining diffusion sampling with iterative guidance using the forward model $\\mathcal{A}$ and measurement $\\boldsymbol{y}$ . This approach provides sophisticated, precise solutions to complex inverse problems, leveraging the power and flexibility of diffusion models in practical applications, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10466", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DIS using latent diffusion models (LDIS). Most DIS [2, 4, 11, 21, 27] use pixel-space diffusion models, which facilitate easy integration of the forward model $\\mathcal{A}$ and measurement $\\pmb{y}$ , as both are defined in pixel space. Integrating forward models in latent space presents more challenges. Latent-space methods [3, 12, 19] calculate data consistency terms after decoding the denoised latent representation, then update these guidances within the latent space.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During this process, VAE mapping errors accumulate in iterative sampling, causing the representation to drift from the clean manifold $\\mathcal{M}_0$ . Additionally, most latent diffusion models provide a text-conditioned prior distribution $p_{\\theta}(\\boldsymbol{x}|c_{\\mathrm{text}})$ , which is challenging to implement in cases where text conditioning ( $c_{\\mathrm{text}}$ ) is unavailable. As a result, latent-space methods prioritize two main goals: (i) managing text embeddings effectively and (ii) preserving the updated latent close to the clean manifold $\\mathcal{M}_0$ .", + "bbox": [ + 89, + 219, + 480, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For text embeddings, PSLD [19] uses only null-text, while TReg [12] and P2L [3] apply either null-text optimization or text optimization to enhance the reconstruction performance. To maintain the updated latent representation quality, the regularization term for aligning pixel and latent spaces is used to enforce latent feasibility [3, 12, 19].", + "bbox": [ + 89, + 363, + 480, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Solving video inverse problems using DIS. Recently, several extensions [5, 13, 32] from DIS have been introduced to address video inverse problems. A straightforward application of image diffusion models to video, processing frames individually, often disrupts temporal consistency. These approaches maintain temporal coherence by employing a batch-consistent sampling strategy [13] and leveraging optical flow guidance to warp either latent representations [32] or the noise prior [5].", + "bbox": [ + 89, + 462, + 480, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these innovative approaches allow powerful image generative models [6, 16, 18] to address video inverse problems with reduced computational demands, there is still room for improvement. A key limitation of optical flow-based methods [5, 32] is their heavy reliance on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes particularly problematic in scenarios involving severe degradations, which can hinder the estimation process and limit the methods' applicability to broader restoration tasks. Moreover, these approaches often require task-specific restoration modules [32] or fine-tuning of the diffusion model [5].", + "bbox": [ + 89, + 604, + 480, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, the batch-consistent sampling strategy [13] has demonstrated effectiveness in addressing various spatiotemporal degradations without the need for task-specific training or model fine-tuning. However, this method utilizes the unconditional pixel-space diffusion model provided by ADM [6], and its extension to latent diffusion models remains unexplored.", + "bbox": [ + 89, + 794, + 480, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. High Definition Video Inverse Solver Using Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 513, + 89, + 903, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces a novel approach for reconstructing high-definition videos that include various spatio-temporal degradations. The overall pipeline of the algorithm is illustrated in Fig. 2.", + "bbox": [ + 511, + 132, + 903, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider the spatio-temporal degradation process is formulated as:", + "bbox": [ + 511, + 193, + 903, + 222 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {Y} = \\mathcal {A} (\\boldsymbol {X}) = \\mathcal {A} ([ \\boldsymbol {X} [ 1 ], \\dots , \\boldsymbol {X} [ N ]) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 232, + 903, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{Y}$ denotes the measurement, $X[n]$ denotes the $n$ -th frame ground-truth frame, $N$ is the number of video frames, and $\\mathcal{A}$ refers to the operator describing the spatio-temporal degradation process. The detailed spatio-temporal degradation operator $\\mathcal{A}$ is described in Sec. 4.1.", + "bbox": [ + 511, + 258, + 903, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach begins by inverting the measurement frames, denoted as $\\bar{Y}$ , to initialize the informative latents $z_{\\tau}$ , enhancing batch-wise consistency (Step 1). Next, we construct the corresponding denoised batch $\\hat{\\mathbf{X}}_{\\tau}$ by sampling each latent sequentially using Tweedie's formula [7], followed by decoding (Step 2). In Step 3, the corresponding denoised batch $\\hat{\\mathbf{X}}_{\\tau}$ is further refined by applying $l$ -step conjugate gradient (CG) optimization [4, 13] to enforce the data consistency from spatio-temporal degradation $\\mathcal{A}$ . In Step 4, we apply a scheduled low-pass filter to the updated batch $\\bar{\\mathbf{X}}_{\\tau}$ inspired by the frequency-based analysis of spectral diffusion [31]. $\\bar{\\mathbf{X}}_{\\tau}$ is then re-encoded into latent space to form $\\bar{z}_{\\tau}$ . Finally, we obtain the one-step denoised latents $z_{\\tau - 1}$ by adding noise to the encoded latents $\\bar{z}_{\\tau}$ (Step 5). In the following, we provide a detailed description of each of these steps.", + "bbox": [ + 511, + 335, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 1: Initialize informative latents. One of our key insights is to initialize the informative latents by inverting the measurement frames, thereby enhancing batch-wise consistent initialization. Although these latents cannot restore the ground-truth frames directly, inverted latent variables can inherit information from the measurement frame, providing good initializations [29]. Different from SVI [13], which initializes sampling from a batch-wise synchronized uninformative Gaussian prior $z_{T} \\sim \\mathcal{N}(0, I)$ as the initial sampling point, we replace $z_{T}$ with the informative prior $z_{\\tau}$ , defined as:", + "bbox": [ + 511, + 575, + 905, + 741 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {0} = \\boldsymbol {E} _ {\\theta} (\\boldsymbol {Y}), \\quad \\boldsymbol {z} _ {\\tau} = \\mathrm {D D I M} ^ {- 1} \\left(\\boldsymbol {z} _ {0}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 751, + 903, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{\\theta}(\\cdot)$ and $\\mathrm{DDIM}^{-1}(\\cdot)$ denote frame-wise encoding from pretrained VAE and DDIM inversion of timestep $\\tau$ respectively.", + "bbox": [ + 511, + 779, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 2: Pseudo-batch sampling. After initialization, we guide the sampling path to ensure the data consistency condition. At timestep $0 < t \\leq \\tau$ , we sample denoise batch $\\hat{z}_t$ from given latents $z_t \\coloneqq [z_t[1] \\cdots z_t[N]]$ by using Tweedie's formula [7]. Unlike SVI [13], we split latent", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10467", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19e9d2e13fc3ea1b81ae162bd267bdf3b7536d1ccc29225bf614217e236133f8.jpg", + "image_caption": [ + "Figure 2. Illustration of VISION-XL sampling at timestep $t$ : $z_{t}$ is split into individual frames and denoised sequentially using Tweedie's formula. The denoised latents $\\hat{z}_{t}$ are then merged and decoded. The decoded batch $\\hat{\\mathbf{X}}_t$ is optimized to enforce the data consistency, followed by low-pass filtered encoding and re-noising to obtain $z_{t-1}$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frames to construct a pseudo-batch and sequentially sample each frame, requiring memory for only a single frame during the sampling, as shown in Fig. 2. Similarly, inversion is also conducted within the pseudo-batch framework. This enables the recent advanced latent diffusion model to operate in this framework without a frame limit. As a proof-of-concept, we conduct experiments on 25-frame videos.", + "bbox": [ + 88, + 310, + 480, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, consider a sequential sampling of latent diffusion models $\\epsilon_{\\theta}$ along the temporal direction:", + "bbox": [ + 89, + 415, + 480, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {E}} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t}) := \\left[ \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t} [ 1 ]) \\dots \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t} [ N ]) \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 453, + 482, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The denoised latents $\\hat{Z}_t$ are computed using Tweedie's formula [7]:", + "bbox": [ + 89, + 484, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {z}} _ {t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {z} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\mathcal {E}} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t})\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 523, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ is the noise schedule defined in the Gaussian process of diffusion models [9, 15]. Then denoised batch $\\hat{X}_t$ is decoded from the denoised latents $\\hat{z}_t$ using VAE decoder $D_{\\theta}$ :", + "bbox": [ + 89, + 561, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {X}} _ {t} = \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t}) := [ \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t} [ 1 ]) \\dots \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t} [ N ]) ]. \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 631, + 482, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: Batch optimization in pixel-space. The denoised batch $\\hat{\\pmb{X}}_t$ is then refined as a whole by applying the $l$ -step CG optimization to enhance the data consistency from the measurement $\\pmb{Y}$ and spatio-temporal degradation $\\mathcal{A}$ . This can be formally represented by", + "bbox": [ + 89, + 657, + 483, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {X}} _ {t} := \\underset {\\boldsymbol {X} \\in \\hat {\\boldsymbol {X}} _ {t} + \\boldsymbol {\\mathcal {K}} _ {l}} {\\arg \\min } \\| \\boldsymbol {Y} - \\mathcal {A} (\\boldsymbol {X}) \\| ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 742, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{K}_l$ denotes the $l$ -dimensional Kyrlov subspace associated with the given inverse problem [4]. The multistep CG allows each temporal frame to be diversified, enhancing data consistency and achieving faster convergence without requiring memory-intensive gradient calculations [2].", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 4: Low-pass filtered encoding. Recent frequency-based analyses of diffusion models [10, 31] suggest that optimal denoisers first recover low-frequency components in", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "early denoising stages, while high-frequency details added progressively in later stages. Building on these findings, we observed that applying a scheduled low-pass filter to the updated batch $\\bar{X}_t$ in early stages effectively removes undesired artifact caused by VAE error accumulation, resulting in more natural and refined outputs.", + "bbox": [ + 511, + 309, + 905, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the observation that the denoiser restores high-frequency details as the noise scale $\\sqrt{1 - \\bar{\\alpha}_t}$ decreases, we set the filter width $\\sigma_t$ to be proportional to the noise scale [15], defined as $\\sigma_t \\coloneqq \\lambda \\sqrt{1 - \\bar{\\alpha}_t}$ , which goes to zero at $t \\to 0$ . After applying the low-pass filter $h_{\\sigma_t}$ , we re-encode $\\bar{\\boldsymbol{X}}_t$ into the latent space. Specifically, the re-encoded latents are given by:", + "bbox": [ + 511, + 404, + 905, + 510 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {X}} _ {t} \\leftarrow \\bar {\\boldsymbol {X}} _ {t} * \\boldsymbol {h} _ {\\sigma_ {t}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 527, + 903, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {z} _ {t} = \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t}) := [ \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t} [ 1 ]) \\dots \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t} [ N ]) ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 547, + 903, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{\\theta}$ denotes the VAE encoder.", + "bbox": [ + 511, + 582, + 756, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 5: Renoising. After encoding, updated latents $\\bar{z}_t$ are renoised as:", + "bbox": [ + 511, + 601, + 905, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\bar {\\boldsymbol {z}} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\boldsymbol {\\mathcal {E}} _ {t}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 650, + 903, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\varepsilon}_{t}$ is composed of batch-consistent stochastic noise [13] and deterministic noise [20].", + "bbox": [ + 511, + 685, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the geometric view of diffusion models [1], the sampling path evolves as illustrated in Fig. 3. The initialized latent $z_{\\tau}$ is projected onto the clean manifold $\\mathcal{M}_0$ using Tweedie's formula. The projected latent is then decoded into the pixel space and refined through multi-step CG to satisfy the data consistency constraint $Y = \\mathcal{A}(X)$ . Next, a scheduled low-pass filter is applied to reduce VAE error accumulation and keep the encoding close to $\\mathcal{M}_0$ . Finally, the encoded latents are re-noised to transition back to $\\mathcal{M}_{\\tau -1}$ , and this process iterates until the final state converges to the clean manifold $\\mathcal{M}_0$ . The complete algorithm is provided in Algorithm 1.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10468", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/914006bdc2f45db2ebb1f074074d053b35372efb9aabca0c78d018f2dc80a405.jpg", + "image_caption": [ + "Figure 3. Geometric illustration of the sampling path evolution. (Step 1) Initialize latent $\\pmb{z}_{\\tau}$ . (Step 2) Project onto $\\mathcal{M}_0$ via pseudobatch sampling and decode to pixel space. (Step 3) Optimize for measurement consistency $\\pmb{Y} = \\mathcal{A}(\\pmb{X})$ . (Step 4) Apply a scheduled low-pass filter and encode back to latent space. (Step 5) Renoise to $\\mathcal{M}_{\\tau -1}$ ." + ], + "image_footnote": [], + "bbox": [ + 135, + 78, + 444, + 318 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 High-definition video inverse problem solver using latent diffusion models" + ], + "code_body": "Require: $\\pmb{\\mathcal{E}}_{\\theta}^{(t)},\\pmb {E}_{\\theta},\\pmb {D}_{\\theta},\\pmb {Y},\\mathcal{A},\\tau ,l,\\sigma_{t},\\{\\bar{\\alpha}_{t}\\}_{t = 1}^{T}$ \n1: $z_0\\gets E_\\theta (Y)$ \n2: $z_{\\tau}\\gets \\mathrm{DDIM}^{-1}(z_0)$ ▷ Step 1 \n3: for $t = \\tau :2$ do \n4: $\\hat{z}_t\\gets \\left(z_t - \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\mathcal{E}}_\\theta^{(t)}(z_t)\\right) / \\sqrt{\\bar{\\alpha}_t}$ ▷ Step 2 \n5: $\\hat{\\pmb{X}}_t\\gets \\pmb {D}_\\theta (\\hat{\\pmb{z}}_t)$ \n6: $\\bar{\\pmb{X}}_t\\coloneqq \\arg \\min_{\\pmb {X}\\in \\hat{\\pmb{X}}_t + \\pmb {\\kappa}_l}\\| \\pmb {Y} - \\mathcal{A}(\\pmb {X})\\| ^2$ ▷ Step 3 \n7: $\\bar{\\pmb{X}}_t\\gets \\bar{\\pmb{X}}_t*\\pmb {h}_{\\sigma_t}$ ▷ Step 4 \n8: $\\bar{z}_t = E_\\theta (\\bar{X}_t)$ \n9: $z_{t - 1} = \\sqrt{\\bar{\\alpha}_{t - 1}}\\bar{z}_t + \\sqrt{1 - \\bar{\\alpha}_{t - 1}}\\pmb {\\mathcal{E}}_t$ ▷ Step 5 \n10: end for \n11: $z_0\\gets \\left(z_1 - \\sqrt{1 - \\bar{\\alpha}_1}\\pmb {\\mathcal{E}}_\\theta^{(1)}(z_1)\\right) / \\sqrt{\\bar{\\alpha}_1}$ \n12: return $Z_{0}$", + "bbox": [ + 91, + 452, + 483, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 694, + 295, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 719, + 277, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We used four high-resolution (with resolutions exceeding 1080p) video datasets for evaluation, sourced from the DAVIS dataset [17] and the Pexels dataset1. A subset of 100 videos from the DAVIS dataset is resized to $768 \\times 1280$ resolution and consists of 25 frames, originally provided in landscape orientation. The Pexels dataset is a large, open-source collection of high-resolution stock videos and images, widely used for creative and research purposes. For the Pexels subset, we collect a total of 120 videos: 45", + "bbox": [ + 89, + 741, + 483, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in landscape orientation (Pexels (landscape)), 45 in vertical orientation (Pexels (vertical)), and 30 in square orientation (Pexels (square)). These subsets are resized to resolutions of $768 \\times 1280$ for landscape, $1280 \\times 768$ for vertical, and $1024 \\times 1024$ for square orientations, with each video consisting of 25 frames.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inverse problems. We test our method on the following spatial degradations: 1) Deblur: Gaussian deblurring from an image convolved with a $61 \\times 61$ size Gaussian kernel with $\\sigma = 3.0$ , 2) SR: Super-resolution from $\\times 4$ average pooling, 3) Inpaint: Inpainting from $50\\%$ random masking. Furthermore, test our method on the following spatio-temporal degradations: 4) Deblur+: Deblur + 7-frame averaging using temporal uniform blur kernel as used in [13], 5) SR+: SR + 7-frame averaging, and 6) Inpaint+: Inpaint + 7-frame averaging.", + "bbox": [ + 511, + 181, + 906, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline comparison. The primary objective of this study is to improve the performance of video inverse problem solvers through latent image diffusion models. Thus, our evaluation primarily compares video inverse problem solvers using image diffusion models. As a recently emerging field, only a few methods are available: SVI [13], DiffIR2VR [32], and Warped Diffusion [5]. Notably, DiffIR2VR and Warped Diffusion cannot address spatiotemporal degradations, and DiffIR2VR only supports SR among the inverse problems we address in this paper. We conducted comparisons with SVI and DiffIR2VR but excluded Warped Diffusion, as it is not currently open-source. SVI officially supports a resolution of $256 \\times 256$ , while DiffIR2VR supports $480 \\times 854$ . To ensure fair comparisons with identical resolutions, we used patch reconstruction. Additionally, we included a comparison with the classical optimization method ADMM-TV, following the protocol established by SVI [13].", + "bbox": [ + 511, + 334, + 908, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For quantitative comparison, we focus on two widely used standard metrics: peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) [28]. Additionally, we evaluate two perceptual metrics: Learned Perceptual Image Patch Similarity (LPIPS) [34] and Fréchet Video Distance (FVD) [26]. For computing the metrics, we follow the protocol from the open-source project2.", + "bbox": [ + 511, + 606, + 905, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. While our method is applicable to general latent diffusion models, we use Stable Diffusion XL 1.0 (SDXL) [16]—the current state-of-the-art text-to-image diffusion model—as a proof of concept in this paper. For all experiments, we employ $T = 25$ , $\\tau = 0.3T$ , $\\lambda = 2$ , and $l = 10$ . These optimal values are obtained through extensive ablation studies and the results are described in Sec. 4.3. To reduce undesired guidance from the text condition $c_{\\mathrm{text}}$ , we use a null-text condition, $c_{\\emptyset}$ . All experiments were done on a single NVIDIA 4090 GPU.", + "bbox": [ + 511, + 712, + 905, + 862 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://www.pexels.com/", + "bbox": [ + 107, + 886, + 297, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2https://github.com/JunyaoHu/common_metricson Video_quality", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10469", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cbaf10e908d1663342b1250c5042a5bed893b8e53668bbed55d9b3c0f863554e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMethodDAVISPexels (landscape)Pexels (vertical)Pexels (square)
FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑
Deblur+ADMM-TV15120.39724.300.742671.80.28429.420.818709.20.23730.370.847569.50.22430.680.856
SVI [13]638.90.32228.040.799830.50.26530.420.831656.20.23930.400.856499.40.22131.590.862
DiffIR2VR [32]----------------
Ours228.60.29228.760.807196.40.24931.110.839209.90.22431.870.860157.90.21732.400.864
SR+ADMM-TV14290.35924.230.740634.10.27929.010.820669.20.32229.710.836545.90.30630.080.838
SVI [13]223.40.23429.000.812386.80.26530.700.831558.90.26130.480.842313.20.26531.180.847
DiffIR2VR [32]----------------
Ours158.50.24429.180.818166.20.24630.820.832173.10.22931.380.847138.80.22031.900.856
Inpaint+ADMM-TV18480.33924.160.762797.80.29229.150.778805.40.25829.360.805652.40.26830.390.804
SVI [13]208.60.23829.600.848269.30.25029.920.826428.10.24630.270.838206.90.23831.120.847
DiffIR2VR [32]----------------
Ours241.10.24228.810.815222.40.21630.130.828240.40.20130.980.845164.90.23031.440.853
DeblurADMM-TV169.00.23230.490.873192.20.26331.670.842185.70.18132.190.861199.60.25730.980.875
SVI [13]99.500.17631.700.875153.80.21232.440.862160.20.17433.350.866116.60.20733.510.885
DiffIR2VR [32]----------------
Ours72.030.17131.940.88096.160.18432.830.87689.390.17033.550.88891.550.15734.100.896
SRADMM-TV285.60.22126.650.788301.50.22227.910.769298.70.21327.120.786209.40.25728.740.798
SVI [13]176.10.17629.280.815201.60.20231.000.836219.30.20030.590.847167.90.20631.830.851
DiffIR2VR [32]319.60.23826.290.738412.60.24427.180.742511.60.21426.910.752412.00.29627.760.741
Ours81.100.19030.270.848104.30.18531.840.85898.570.15832.540.87698.410.13833.020.886
InpaintADMM-TV212.00.31528.420.797270.90.31629.140.794270.30.21229.840.803264.70.31030.460.809
SVI [13]143.50.17730.200.858159.40.23330.090.827164.30.13231.190.847139.30.22531.570.855
DiffIR2VR [32]----------------
Ours143.60.20929.740.835158.70.20830.410.834174.00.19531.150.847125.40.21631.630.859
", + "bbox": [ + 94, + 77, + 903, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative evaluation (FVD, LPIPS, PSNR, SSIM) of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios (landscape, vertical, square). Bold denotes the best results and underline indicates the runner-up. Notably, DiffIR2VR [32] is only capable of restoring SR among our experimental tasks, highlighting the broader task generalizability of our approach.", + "bbox": [ + 89, + 367, + 906, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebc344f22dcbc44e9bc03683f2a303babc26e2833978cca0bfe83af7a792ec48.jpg", + "image_caption": [ + "Figure 4. Qualitative evaluation of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios. Notably, ADMM-TV fails to remove ghosting artifacts caused by temporal degradation (red arrows), while SVI produces excessive intensity fluctuations (red box) or blurred information restoration (green and blue boxes)." + ], + "image_footnote": [], + "bbox": [ + 91, + 435, + 906, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 791, + 184, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents a quantitative comparison across various spatio-temporal inverse problems. The proposed method consistently outperforms baseline approaches in most metrics, particularly in addressing spatio-temporal degradations. Notably, it achieves a significant reduction in FVD,", + "bbox": [ + 89, + 825, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "indicating superior perceptual video quality compared to the runner-up across all datasets. This improvement is also evident in the qualitative results shown in Fig. 4. While SVI [13] effectively handles spatio-temporal degradations, it often struggles with temporal consistency, leading to issues such as inaccurate intensity restoration (first row) and loss of details (second and third rows). This suggests", + "bbox": [ + 511, + 792, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10470", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1b7dbc97f9b4c9aace0c561ad3cdb3e609f28b5a2d36dd59e1e711e0e0c446d3.jpg", + "image_caption": [ + "Figure 5. Qualitative evaluation of SR $(\\times 4)$ performance across multiple aspect ratios (landscape, vertical). DiffIR2VR often produces unwanted artifacts in the background (red and blue boxes), while SVI inaccurately restores intensity (red box), leading to frame-wise fluctuations." + ], + "image_footnote": [], + "bbox": [ + 112, + 80, + 885, + 273 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce7479669ed8273676975d9b9159ebd51ed54ae6af560ef30751631bc57694d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTime [min]Time per frame [sec]Memory (GB)
SVI [13]153618.5
DiffIR2VR [32]4.71113.6
Ours2.5612.7
", + "bbox": [ + 130, + 325, + 439, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that batch-consistent noise initialization [13], further examined in our ablation study (Table 5), may be insufficient to fully preserve temporal consistency. Additionally, its patch reconstruction for high-resolution videos can introduce patch-wise inconsistencies, degrading overall performance. The classical optimization method, ADMM-TV, effectively reconstructs static backgrounds and stationary objects but fails to remove ghosting artifacts caused by temporal deformations, as shown in Fig. 4. This limitation is reflected in its lower performance metrics. For the SR task (Fig.5), DiffIR2VR[32] often introduces unwanted artifacts in backgrounds or over-generates object details, likely due to inaccuracies in optical flow estimation. Notably, in pixel-wise random inpainting, latent diffusion methods may lose fine pixel-level details due to their encoded representations. However, leveraging the strong SDXL prior, our method achieves competitive inpainting performance with SVI, which employs a pixel-space diffusion model.", + "bbox": [ + 88, + 445, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While baseline methods encounter various challenges across different inverse problems, our approach demonstrates stable, high-quality reconstructions, as evidenced by the overall results. Further comparisons of total sampling time and memory consumption are shown in Table 4, where our method achieves the highest efficiency in both sampling time and memory usage.", + "bbox": [ + 89, + 718, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further experimental results on real-world denoising and blind video deconvolution are available in the supplementary material. Additional visualizations, including reconstruction results for deblurring, inpainting, and other tasks, are provided in video format for further evaluation: https:", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "//vision-x1.github.io/supple/.", + "bbox": [ + 514, + 329, + 777, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 513, + 351, + 673, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze the key components of our method. To highlight their impact, we conduct an ablation study on the $\\mathrm{SR + }$ task in the Pexels (landscape) dataset, which involves significant spatio-temporal degradation.", + "bbox": [ + 511, + 375, + 906, + 434 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d57d314c399f06e02b82fd1a5ec212f620fe482903edcb2f2e23a71119ea3a7.jpg", + "table_caption": [ + "Table 4. Comparison of total sampling time and memory efficiency for solving video inverse problems on a single 25-frame video at $768 \\times 1280$ resolution. **Bold** denotes the best results and **underline** indicates the runner-up." + ], + "table_footnote": [], + "table_body": "
InitializationInitialization Time [sec]Total Time [min]FVD↓PSNR↑
Random noise08.3104729.43
Batch-consistent noise [13]08.3707.730.10
Pseudo-batch inversion (τ: 0.15T)151.3229.530.00
Pseudo-batch inversion (τ: 0.30T)302.5184.830.74
Pseudo-batch inversion (τ: 0.45T)453.8288.730.70
", + "bbox": [ + 517, + 445, + 903, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Comparison of initialization strategies by computational cost and video quality for SR+ on the Pexels dataset.", + "bbox": [ + 511, + 521, + 905, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of initialization (in Step 1). In Table 5, we conduct an ablation study to see the effect of pseudo-batch inversion for initialization. From the table, we confirm that pseudobatch inversion effectively extracts informative latents to reconstruct video evidenced by about 0.6dB and 1.3dB PSNR increase compared with batch-synchronized noise initialization used in SVI [13] and random noise initialization, respectively. Notably, pseudo-batch inversion achieves $\\times 3$ lower FVD compared to the batch-synchronized noise initialization which indicates a significant improvement in temporal consistency. It is also evident in the visualization of the ablation study shown in Fig. 6. The reconstruction results from random noise and batch-synchronized noise initialization fail to reconstruct the color of the cloud and are temporally inconsistent. In contrast, our method successfully reconstructs the color of the cloud and temporally consistent results. Furthermore, as shown in Table 5, pseudobatch inversion significantly improves sampling time efficiency. This is because it reduces the total sampling steps, and inversion does not require the measurement update process, which involves encoding and decoding. As a result, our method reconstructs high-definition video in under 6 seconds per frame on a single NVIDIA 4090 GPU.", + "bbox": [ + 511, + 553, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51a94a435cc7029a6530a15131f8518c88ff159de6b77e49c44ebc502ec8d0ba.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of initialization strategies. Our approach improves temporal consistency and precise restoration." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 467, + 354 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa0477c48cac455c1b06c988af625b812a7d8808180d83f05b7844d8889612ae.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Update step lFVD↓LPIPS↓PSNR↑SSIM↑
111500.28127.550.799
5241.00.19730.690.839
10184.80.23630.740.826
20486.10.47028.160.690
", + "bbox": [ + 163, + 402, + 413, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of optimization step $l$ (in Step 3). In Table 6, we present an ablation study on the effect of the CG update step $l$ . The table confirms that the CG update is essential for enhancing data consistency. We found that at least 5 iterations of CG updates yield satisfactory results, and 10 iterations produce the best results.", + "bbox": [ + 89, + 523, + 483, + 613 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/327bfc4a0bb7904887188ede6f615ffbe3e0926844a529c72d4b6ede32750028.jpg", + "table_caption": [ + "Table 6. Ablation study on the effect of optimization step $l$ . Ten update steps yield the best temporal consistency." + ], + "table_footnote": [], + "table_body": "
LPF λFVD↓LPIPS↓PSNR↑SSIM↑
No LPF209.70.27329.920.797
1186.30.23930.590.819
√2184.80.23630.740.826
2179.30.24530.810.832
2√2191.40.26230.890.837
", + "bbox": [ + 173, + 626, + 401, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on the effect of low-pass filtering parameter $\\lambda$ . $\\lambda = 2$ yields the best temporal consistency.", + "bbox": [ + 89, + 715, + 482, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of LPF $\\lambda$ (in Step 4). Table 7 presents an ablation study on the effect of low-pass filtering. The results confirm that low-pass filtering enhances the reconstruction quality as evidenced by all metrics. Specifically, low-pass filtering results in approximately a 30-point decrease in FVD and a 1.0dB increase in PSNR compared to the absence of low-pass filtering. This improvement is also evident in the visualizations in Fig. 7. In the second row of the figure, undesired artifacts appear when low-pass filtering is not applied. In contrast, as shown in the third and fourth rows,", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/685af7bf18c39cab7723baa4af91bf19e62d4c7509ea244827039cdbf1231341.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of low-pass filtering parameter $\\lambda$ . $\\lambda = 2$ effectively removes artifacts from VAE errors." + ], + "image_footnote": [], + "bbox": [ + 537, + 92, + 885, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "these artifacts are effectively removed as the parameter $\\lambda$ increases. From a frequency-based perspective, we believe that low-pass filtering effectively guides the updated latents to remain within the desired denoised manifold, $\\mathcal{M}_0$ , and helps to mitigate error accumulation from the VAE.", + "bbox": [ + 511, + 407, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 497, + 633, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel framework for addressing high-definition video inverse problems using latent diffusion models that introduce two new strategies. First, a pseudo-batch consistent sampling strategy to manage intensive batch memory consumption with advanced latent diffusion models (e.g., SDXL). To acquire a denoised batch, we conduct sequential sampling of each latents rather than batch sampling to efficiently manage the high memory consumption of advanced latent diffusion models. Second, a pseudo-batch inversion for leveraging informative latent as initialization is proposed. We confirmed that pseudo-batch inversion significantly improves reconstruction performance in both traditional and perceptual quality metrics. Leveraging the powerful SDXL, our method achieves state-of-the-art performance across diverse spatio-temporal inverse problems, including challenging tasks such as the combination of frame averaging with deblurring, super-resolution, and inpainting. Importantly, our method supports multiple aspect ratios (landscape, vertical, and square), making it versatile for different video formats and delivering HD reconstructions in under 6 seconds per frame on a single NVIDIA 4090 GPU. Overall, our framework not only enhances video reconstruction quality but also sets new standards for efficiency and flexibility in solving high-definition video inverse problems.", + "bbox": [ + 511, + 523, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10472", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Research Foundation of Korea under Grant RS-2024-00336454, and by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. RS-2024-00457882, AI Research Hub Project; No. RS-2025-02304967, AI Star Fellowship (KAIST))", + "bbox": [ + 89, + 116, + 483, + 223 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 2, 4", + "[2] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4", + "[3] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 3", + "[4] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 4", + "[5] Giannis Daras, Weili Nie, Karsten Kreis, Alex Dimakis, Morteza Mardani, Nikola Borislavov Kovachki, and Arash Vahdat. Warped diffusion: Solving video inverse problems with image diffusion models. arXiv preprint arXiv:2410.16152, 2024. 2, 3, 5", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2, 3", + "[7] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 3, 4", + "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4", + "[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 4", + "[11] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2, 3" + ], + "bbox": [ + 93, + 263, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jeongsol Kim, Geon Yeong Park, Hyungjin Chung, and Jong Chul Ye. Regularization by texts for latent diffusion inverse solvers. arXiv preprint arXiv:2311.15658, 2023. 3", + "[13] Taesung Kwon and Jong Chul Ye. Solving video inverse problems using image diffusion models. In The Thirteenth International Conference on Learning Representations, 2025. 2, 3, 4, 5, 6, 7", + "[14] Dongyang Li, Chen Wei, Shiying Li, Jiachen Zou, and Quanying Liu. Visual decoding and reconstruction via eeg embeddings with guided diffusion. arXiv preprint arXiv:2403.07721, 2024. 1", + "[15] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4", + "[16] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 5", + "[17] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5", + "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3", + "[19] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[20] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 4", + "[21] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3", + "[22] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1, 2", + "[23] Andreas Stöckl. Evaluating a synthetic image dataset generated with stable diffusion. In International Congress on Information and Communication Technology, pages 805-818. Springer, 2023. 1", + "[24] Yu Takagi and Shinji Nishimoto. High-resolution image reconstruction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14453-14463, 2023. 1", + "[25] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3", + "[26] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 2, 5", + "[27] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2023. 2, 3", + "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[29] Jie Xiao, Ruili Feng, Han Zhang, Zhiheng Liu, Zhantao Yang, Yurui Zhu, Xueyang Fu, Kai Zhu, Yu Liu, and Zheng-Jun Zha. Dreamclean: Restoring clean image using deep diffusion prior. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[30] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 2, 3", + "[31] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 3, 4", + "[32] Chang-Han Yeh, Chin-Yang Lin, Zhixiang Wang, Chi-Wei Hsiao, Ting-Hsuan Chen, and Yu-Lun Liu. Diffir2vr-zero: Zero-shot video restoration with diffusion-based image restoration models. arXiv preprint arXiv:2407.01519, 2024. 2, 3, 5, 6, 7", + "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1", + "[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + ], + "bbox": [ + 91, + 90, + 480, + 656 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10474", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_model.json b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e17b5f7342f7b4f9c2780b3e2f3014a6dce3198b --- /dev/null +++ b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_model.json @@ -0,0 +1,1925 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "VISION-XL: High Definition Video Inverse Problem Solver using Latent Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.205, + 0.44, + 0.238 + ], + "angle": 0, + "content": "Taesung Kwon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.243, + 0.473, + 0.256 + ], + "angle": 0, + "content": "star.kwon@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.571, + 0.205, + 0.681, + 0.238 + ], + "angle": 0, + "content": "Jong Chul Ye KAIST" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.243, + 0.708, + 0.257 + ], + "angle": 0, + "content": "jong.ye@kaist.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.281, + 0.256, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.281, + 0.581, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.281, + 0.905, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.422, + 0.599, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.422, + 0.904, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.502, + 0.599, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.502, + 0.904, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.586, + 0.908, + 0.616 + ], + "angle": 0, + "content": "Figure 1. Representative video reconstruction by VISION-XL: \\(\\mathrm{SR + }\\) (frame averaging with \\(\\times 4\\) super-resolution, top), Deblur+ (frame averaging with deblurring, \\(\\sigma = 3.0\\), bottom-left), and Inpaint+ (frame averaging with \\(50\\%\\) random inpainting, bottom-right)." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.628, + 0.327, + 0.643 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.66, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models. Building on recent advancements in spatio-temporal optimization for video inverse problems using image diffusion models, our approach leverages latent-space diffusion models to achieve enhanced video quality and resolution. To address the high computational demands of processing high-resolution frames, we introduce a pseudo-batch consistent sampling strategy, allowing efficient operation on a single GPU. Additionally, to improve temporal consistency, we present pseudo-batch inversion, an initialization technique that incorporates informative latents from the measurement. By integrating with SDXL, our framework achieves state-of-the-art video reconstruction across a wide range of spatio-temporal inverse problems, including complex combinations of frame averaging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.735 + ], + "angle": 0, + "content": "ing and various spatial degradations, such as deblurring, super-resolution, and inpainting. Unlike previous methods, our approach supports multiple aspect ratios (landscape, vertical, and square) and delivers HD-resolution reconstructions (exceeding \\(1280 \\times 720\\)) in under 6 seconds per frame on a single NVIDIA 4090 GPU. Project page: https://vision-x1.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.754, + 0.646, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.779, + 0.907, + 0.886 + ], + "angle": 0, + "content": "Diffusion models [6, 9, 10, 16, 18, 20, 22] have set a new benchmark in generative modeling, enabling the generation of high-quality samples. These models have become the foundation for advancements in various fields, such as controllable image editing [33], image personalization [8], synthetic data augmentation [23], and even reconstructing images from brain signals [14, 24]." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Furthermore, diffusion model-based inverse problem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10465" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.48, + 0.138 + ], + "angle": 0, + "content": "
MethodsSpatio-temporal degradationLatent space
Optical flow-based [5, 32]X✓
SVI [13]✓X
Ours✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.142, + 0.483, + 0.198 + ], + "angle": 0, + "content": "Table 1. Comparison of image diffusion-based video inverse problem solvers. Unlike other methods, our approach leverages latent image diffusion models to address spatio-temporal degradation, enabling more effective restoration." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.206, + 0.483, + 0.311 + ], + "angle": 0, + "content": "solvers (DIS) [2, 4, 11, 21, 27, 29] address a variety of image restoration tasks, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on. A key feature of DIS is its plug-and-play capability, allowing diffusion models to be applied flexibly across different inverse problems without requiring task-specific training or fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.312, + 0.483, + 0.432 + ], + "angle": 0, + "content": "Recently, several extensions [5, 13, 32] from the DIS have been proposed to solve video inverse problems using image diffusion models. Naive application of image diffusion models to videos may break temporal consistency. To address this problem, these methods preserve temporal consistency by utilizing a batch-consistent sampling strategy [13] and applying optical flow guidance to warp either latent representations [32] or the noise prior [5]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Although these innovative approaches enable powerful image generative models [6, 16, 18] to solve video inverse problems with significantly reduced computational requirements, there is still room for improvement in these methods. Optical flow-based methods [5, 32] have reported a key limitation: their performance is highly dependent on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes problematic when extreme degradations complicate the estimation process, restricting their applicability to a wider range of restoration tasks. Additionally, these methods require task-specific restoration modules [32] or fine-tuning of the diffusion model [5]. In this perspective, batch consistent sampling strategy [13] successfully addressed various spatio-temporal degradations without requiring task-specific training or fine-tuning. However, we empirically found that extending SVI [13] to latent diffusion models leads to unsatisfactory reconstruction, particularly in terms of FVD [26], as shown in Table 2, despite the fact that most modern latent diffusion models are essential for scaling up to large-size video inverse problems." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.736, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To overcome these limitations, here we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models that involves no additional training. To address the high computational demands of batch processing (e.g., 16-frame batch processing used in [13]) with high-resolution latent diffusion models [16], we introduce pseudo-batch consistent sampling. This strategy enables multi-frame video processing while requiring only the memory needed for a single frame, making it feasible on a single GPU. Furthermore, we propose pseudo-batch inversion, which initializes the process with" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.089, + 0.882, + 0.149 + ], + "angle": 0, + "content": "
InitializationFVD ↓LPIPS ↓PSNR ↑
Random noise10470.25129.43
Batch synchronized noise (SVI [13])707.70.24830.10
Pseudo-batch inversion (Ours)184.80.23630.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.195 + ], + "angle": 0, + "content": "Table 2. Impact of our initialization method on SR+ video restoration using SDXL [16]. Our method significantly improves performance, reducing FVD [26] by more than \\(3\\times\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.905, + 0.266 + ], + "angle": 0, + "content": "informative latents derived from the measurement. This initialization enhances temporal consistency and improves the efficiency of solving spatio-temporal inverse problems as shown in Table. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.432 + ], + "angle": 0, + "content": "By integrating these components, our framework achieves state-of-the-art video reconstruction performance using SDXL [16]. We name the method integrating these components as VISION-XL, short for Video Inverse-problem Solver using latent diffusION models (with stable diffusion XL). It supports various aspect ratios, including landscape, vertical, and square formats. Thanks to its efficiency, our framework can reconstruct \\( 1280 \\times 768 \\) (exceeding HD resolution) videos in under 6 seconds per frame on a single NVIDIA 4090 GPU. Our contribution can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.435, + 0.905, + 0.478 + ], + "angle": 0, + "content": "- We propose a high-definition video inverse problem solver integrated with SDXL, supporting multiple aspect ratios and achieving state-of-the-art reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.48, + 0.905, + 0.525 + ], + "angle": 0, + "content": "- We introduce a novel pseudo-batch consistent sampling and inversion strategy for efficient and effective video reconstruction across diverse inverse problems." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.435, + 0.905, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.655, + 0.557 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.905, + 0.673 + ], + "angle": 0, + "content": "Diffusion model-based inverse problem solvers (DIS). Diffusion models [9, 20, 22] attempt to model the data distribution \\( p_{\\theta}(\\pmb{x}) \\) based on the Gaussian transitions. In the geometric view of diffusion models [1], the transitions are typically described as iterative manifold transitions \\( \\mathcal{M}_t \\to \\mathcal{M}_{t-1}, t = T, \\dots, 1 \\), moving from the noisy manifold \\( \\mathcal{M}_T \\) to the clean manifold \\( \\mathcal{M}_0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Diffusion model-based inverse problem solvers (DIS) [2, 4, 11, 21, 27] aim to guide manifold transitions to sample from the posterior distribution \\( p_{\\theta}(\\boldsymbol{x}|\\boldsymbol{y}) \\), which represents sampling \\( \\boldsymbol{x} \\) from the measurement \\( \\boldsymbol{y} \\) obtained from the forward model \\( \\mathcal{A}(\\boldsymbol{x}) \\). In Bayesian inference, the posterior distribution, \\( p_{\\theta}(\\boldsymbol{x}|\\boldsymbol{y}) \\propto p_{\\theta}(\\boldsymbol{x})p(\\boldsymbol{y}|\\boldsymbol{x}) \\) is decomposed into the likelihood \\( p(\\boldsymbol{y}|\\boldsymbol{x}) \\), representing the probability of observing \\( \\boldsymbol{y} \\) given \\( \\boldsymbol{x} \\), and the prior data distribution \\( p_{\\theta}(\\boldsymbol{x}) \\). This decomposition enables posterior sampling by combining diffusion sampling with iterative guidance using the forward model \\( \\mathcal{A} \\) and measurement \\( \\boldsymbol{y} \\). This approach provides sophisticated, precise solutions to complex inverse problems, leveraging the power and flexibility of diffusion models in practical applications, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10466" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.213 + ], + "angle": 0, + "content": "DIS using latent diffusion models (LDIS). Most DIS [2, 4, 11, 21, 27] use pixel-space diffusion models, which facilitate easy integration of the forward model \\(\\mathcal{A}\\) and measurement \\(\\pmb{y}\\), as both are defined in pixel space. Integrating forward models in latent space presents more challenges. Latent-space methods [3, 12, 19] calculate data consistency terms after decoding the denoised latent representation, then update these guidances within the latent space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.482, + 0.356 + ], + "angle": 0, + "content": "During this process, VAE mapping errors accumulate in iterative sampling, causing the representation to drift from the clean manifold \\(\\mathcal{M}_0\\). Additionally, most latent diffusion models provide a text-conditioned prior distribution \\(p_{\\theta}(\\boldsymbol{x}|c_{\\mathrm{text}})\\), which is challenging to implement in cases where text conditioning (\\(c_{\\mathrm{text}}\\)) is unavailable. As a result, latent-space methods prioritize two main goals: (i) managing text embeddings effectively and (ii) preserving the updated latent close to the clean manifold \\(\\mathcal{M}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.482, + 0.456 + ], + "angle": 0, + "content": "For text embeddings, PSLD [19] uses only null-text, while TReg [12] and P2L [3] apply either null-text optimization or text optimization to enhance the reconstruction performance. To maintain the updated latent representation quality, the regularization term for aligning pixel and latent spaces is used to enforce latent feasibility [3, 12, 19]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.482, + 0.599 + ], + "angle": 0, + "content": "Solving video inverse problems using DIS. Recently, several extensions [5, 13, 32] from DIS have been introduced to address video inverse problems. A straightforward application of image diffusion models to video, processing frames individually, often disrupts temporal consistency. These approaches maintain temporal coherence by employing a batch-consistent sampling strategy [13] and leveraging optical flow guidance to warp either latent representations [32] or the noise prior [5]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.482, + 0.789 + ], + "angle": 0, + "content": "While these innovative approaches allow powerful image generative models [6, 16, 18] to address video inverse problems with reduced computational demands, there is still room for improvement. A key limitation of optical flow-based methods [5, 32] is their heavy reliance on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes particularly problematic in scenarios involving severe degradations, which can hinder the estimation process and limit the methods' applicability to broader restoration tasks. Moreover, these approaches often require task-specific restoration modules [32] or fine-tuning of the diffusion model [5]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.902 + ], + "angle": 0, + "content": "In contrast, the batch-consistent sampling strategy [13] has demonstrated effectiveness in addressing various spatiotemporal degradations without the need for task-specific training or model fine-tuning. However, this method utilizes the unconditional pixel-space diffusion model provided by ADM [6], and its extension to latent diffusion models remains unexplored." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.905, + 0.123 + ], + "angle": 0, + "content": "3. High Definition Video Inverse Solver Using Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.133, + 0.905, + 0.193 + ], + "angle": 0, + "content": "This section introduces a novel approach for reconstructing high-definition videos that include various spatio-temporal degradations. The overall pipeline of the algorithm is illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.194, + 0.905, + 0.223 + ], + "angle": 0, + "content": "Consider the spatio-temporal degradation process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.233, + 0.905, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {Y} = \\mathcal {A} (\\boldsymbol {X}) = \\mathcal {A} ([ \\boldsymbol {X} [ 1 ], \\dots , \\boldsymbol {X} [ N ]) ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.334 + ], + "angle": 0, + "content": "where \\(\\mathbf{Y}\\) denotes the measurement, \\(X[n]\\) denotes the \\(n\\)-th frame ground-truth frame, \\(N\\) is the number of video frames, and \\(\\mathcal{A}\\) refers to the operator describing the spatio-temporal degradation process. The detailed spatio-temporal degradation operator \\(\\mathcal{A}\\) is described in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.336, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Our approach begins by inverting the measurement frames, denoted as \\(\\bar{Y}\\), to initialize the informative latents \\(z_{\\tau}\\), enhancing batch-wise consistency (Step 1). Next, we construct the corresponding denoised batch \\(\\hat{\\mathbf{X}}_{\\tau}\\) by sampling each latent sequentially using Tweedie's formula [7], followed by decoding (Step 2). In Step 3, the corresponding denoised batch \\(\\hat{\\mathbf{X}}_{\\tau}\\) is further refined by applying \\(l\\)-step conjugate gradient (CG) optimization [4, 13] to enforce the data consistency from spatio-temporal degradation \\(\\mathcal{A}\\). In Step 4, we apply a scheduled low-pass filter to the updated batch \\(\\bar{\\mathbf{X}}_{\\tau}\\) inspired by the frequency-based analysis of spectral diffusion [31]. \\(\\bar{\\mathbf{X}}_{\\tau}\\) is then re-encoded into latent space to form \\(\\bar{z}_{\\tau}\\). Finally, we obtain the one-step denoised latents \\(z_{\\tau - 1}\\) by adding noise to the encoded latents \\(\\bar{z}_{\\tau}\\) (Step 5). In the following, we provide a detailed description of each of these steps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.577, + 0.906, + 0.742 + ], + "angle": 0, + "content": "Step 1: Initialize informative latents. One of our key insights is to initialize the informative latents by inverting the measurement frames, thereby enhancing batch-wise consistent initialization. Although these latents cannot restore the ground-truth frames directly, inverted latent variables can inherit information from the measurement frame, providing good initializations [29]. Different from SVI [13], which initializes sampling from a batch-wise synchronized uninformative Gaussian prior \\( z_{T} \\sim \\mathcal{N}(0, I) \\) as the initial sampling point, we replace \\( z_{T} \\) with the informative prior \\( z_{\\tau} \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.752, + 0.905, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {0} = \\boldsymbol {E} _ {\\theta} (\\boldsymbol {Y}), \\quad \\boldsymbol {z} _ {\\tau} = \\mathrm {D D I M} ^ {- 1} \\left(\\boldsymbol {z} _ {0}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.905, + 0.825 + ], + "angle": 0, + "content": "where \\(E_{\\theta}(\\cdot)\\) and \\(\\mathrm{DDIM}^{-1}(\\cdot)\\) denote frame-wise encoding from pretrained VAE and DDIM inversion of timestep \\(\\tau\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Step 2: Pseudo-batch sampling. After initialization, we guide the sampling path to ensure the data consistency condition. At timestep \\(0 < t \\leq \\tau\\), we sample denoise batch \\(\\hat{z}_t\\) from given latents \\(z_t \\coloneqq [z_t[1] \\cdots z_t[N]]\\) by using Tweedie's formula [7]. Unlike SVI [13], we split latent" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10467" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.912, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.259, + 0.908, + 0.303 + ], + "angle": 0, + "content": "Figure 2. Illustration of VISION-XL sampling at timestep \\( t \\): \\( z_{t} \\) is split into individual frames and denoised sequentially using Tweedie's formula. The denoised latents \\( \\hat{z}_{t} \\) are then merged and decoded. The decoded batch \\( \\hat{\\mathbf{X}}_t \\) is optimized to enforce the data consistency, followed by low-pass filtered encoding and re-noising to obtain \\( z_{t-1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.311, + 0.482, + 0.415 + ], + "angle": 0, + "content": "frames to construct a pseudo-batch and sequentially sample each frame, requiring memory for only a single frame during the sampling, as shown in Fig. 2. Similarly, inversion is also conducted within the pseudo-batch framework. This enables the recent advanced latent diffusion model to operate in this framework without a frame limit. As a proof-of-concept, we conduct experiments on 25-frame videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.482, + 0.448 + ], + "angle": 0, + "content": "Specifically, consider a sequential sampling of latent diffusion models \\(\\epsilon_{\\theta}\\) along the temporal direction:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.454, + 0.483, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {E}} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t}) := \\left[ \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t} [ 1 ]) \\dots \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t} [ N ]) \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.517 + ], + "angle": 0, + "content": "The denoised latents \\(\\hat{Z}_t\\) are computed using Tweedie's formula [7]:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.524, + 0.483, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {z}} _ {t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {z} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\mathcal {E}} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t})\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.625 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) is the noise schedule defined in the Gaussian process of diffusion models [9, 15]. Then denoised batch \\(\\hat{X}_t\\) is decoded from the denoised latents \\(\\hat{z}_t\\) using VAE decoder \\(D_{\\theta}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.632, + 0.483, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {X}} _ {t} = \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t}) := [ \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t} [ 1 ]) \\dots \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t} [ N ]) ]. \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Step 3: Batch optimization in pixel-space. The denoised batch \\(\\hat{\\pmb{X}}_t\\) is then refined as a whole by applying the \\(l\\)-step CG optimization to enhance the data consistency from the measurement \\(\\pmb{Y}\\) and spatio-temporal degradation \\(\\mathcal{A}\\). This can be formally represented by" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.743, + 0.483, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {X}} _ {t} := \\underset {\\boldsymbol {X} \\in \\hat {\\boldsymbol {X}} _ {t} + \\boldsymbol {\\mathcal {K}} _ {l}} {\\arg \\min } \\| \\boldsymbol {Y} - \\mathcal {A} (\\boldsymbol {X}) \\| ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathcal{K}_l\\) denotes the \\(l\\)-dimensional Kyrlov subspace associated with the given inverse problem [4]. The multistep CG allows each temporal frame to be diversified, enhancing data consistency and achieving faster convergence without requiring memory-intensive gradient calculations [2]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Step 4: Low-pass filtered encoding. Recent frequency-based analyses of diffusion models [10, 31] suggest that optimal denoisers first recover low-frequency components in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.906, + 0.401 + ], + "angle": 0, + "content": "early denoising stages, while high-frequency details added progressively in later stages. Building on these findings, we observed that applying a scheduled low-pass filter to the updated batch \\(\\bar{X}_t\\) in early stages effectively removes undesired artifact caused by VAE error accumulation, resulting in more natural and refined outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.906, + 0.511 + ], + "angle": 0, + "content": "Based on the observation that the denoiser restores high-frequency details as the noise scale \\(\\sqrt{1 - \\bar{\\alpha}_t}\\) decreases, we set the filter width \\(\\sigma_t\\) to be proportional to the noise scale [15], defined as \\(\\sigma_t \\coloneqq \\lambda \\sqrt{1 - \\bar{\\alpha}_t}\\), which goes to zero at \\(t \\to 0\\). After applying the low-pass filter \\(h_{\\sigma_t}\\), we re-encode \\(\\bar{\\boldsymbol{X}}_t\\) into the latent space. Specifically, the re-encoded latents are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.529, + 0.905, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {X}} _ {t} \\leftarrow \\bar {\\boldsymbol {X}} _ {t} * \\boldsymbol {h} _ {\\sigma_ {t}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.548, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\bar {z} _ {t} = \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t}) := [ \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t} [ 1 ]) \\dots \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t} [ N ]) ], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.583, + 0.757, + 0.598 + ], + "angle": 0, + "content": "where \\(E_{\\theta}\\) denotes the VAE encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.906, + 0.631 + ], + "angle": 0, + "content": "Step 5: Renoising. After encoding, updated latents \\(\\bar{z}_t\\) are renoised as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.651, + 0.905, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\bar {\\boldsymbol {z}} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\boldsymbol {\\mathcal {E}} _ {t}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.717 + ], + "angle": 0, + "content": "where \\(\\pmb{\\varepsilon}_{t}\\) is composed of batch-consistent stochastic noise [13] and deterministic noise [20]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the geometric view of diffusion models [1], the sampling path evolves as illustrated in Fig. 3. The initialized latent \\( z_{\\tau} \\) is projected onto the clean manifold \\( \\mathcal{M}_0 \\) using Tweedie's formula. The projected latent is then decoded into the pixel space and refined through multi-step CG to satisfy the data consistency constraint \\( Y = \\mathcal{A}(X) \\). Next, a scheduled low-pass filter is applied to reduce VAE error accumulation and keep the encoding close to \\( \\mathcal{M}_0 \\). Finally, the encoded latents are re-noised to transition back to \\( \\mathcal{M}_{\\tau -1} \\), and this process iterates until the final state converges to the clean manifold \\( \\mathcal{M}_0 \\). The complete algorithm is provided in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10468" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.079, + 0.445, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.406 + ], + "angle": 0, + "content": "Figure 3. Geometric illustration of the sampling path evolution. (Step 1) Initialize latent \\( \\pmb{z}_{\\tau} \\). (Step 2) Project onto \\( \\mathcal{M}_0 \\) via pseudobatch sampling and decode to pixel space. (Step 3) Optimize for measurement consistency \\( \\pmb{Y} = \\mathcal{A}(\\pmb{X}) \\). (Step 4) Apply a scheduled low-pass filter and encode back to latent space. (Step 5) Renoise to \\( \\mathcal{M}_{\\tau -1} \\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.45 + ], + "angle": 0, + "content": "Algorithm 1 High-definition video inverse problem solver using latent diffusion models" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.453, + 0.484, + 0.667 + ], + "angle": 0, + "content": "Require: \\(\\pmb{\\mathcal{E}}_{\\theta}^{(t)},\\pmb {E}_{\\theta},\\pmb {D}_{\\theta},\\pmb {Y},\\mathcal{A},\\tau ,l,\\sigma_{t},\\{\\bar{\\alpha}_{t}\\}_{t = 1}^{T}\\) \n1: \\(z_0\\gets E_\\theta (Y)\\) \n2: \\(z_{\\tau}\\gets \\mathrm{DDIM}^{-1}(z_0)\\) ▷ Step 1 \n3: for \\(t = \\tau :2\\) do \n4: \\(\\hat{z}_t\\gets \\left(z_t - \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\mathcal{E}}_\\theta^{(t)}(z_t)\\right) / \\sqrt{\\bar{\\alpha}_t}\\) ▷ Step 2 \n5: \\(\\hat{\\pmb{X}}_t\\gets \\pmb {D}_\\theta (\\hat{\\pmb{z}}_t)\\) \n6: \\(\\bar{\\pmb{X}}_t\\coloneqq \\arg \\min_{\\pmb {X}\\in \\hat{\\pmb{X}}_t + \\pmb {\\kappa}_l}\\| \\pmb {Y} - \\mathcal{A}(\\pmb {X})\\| ^2\\) ▷ Step 3 \n7: \\(\\bar{\\pmb{X}}_t\\gets \\bar{\\pmb{X}}_t*\\pmb {h}_{\\sigma_t}\\) ▷ Step 4 \n8: \\(\\bar{z}_t = E_\\theta (\\bar{X}_t)\\) \n9: \\(z_{t - 1} = \\sqrt{\\bar{\\alpha}_{t - 1}}\\bar{z}_t + \\sqrt{1 - \\bar{\\alpha}_{t - 1}}\\pmb {\\mathcal{E}}_t\\) ▷ Step 5 \n10: end for \n11: \\(z_0\\gets \\left(z_1 - \\sqrt{1 - \\bar{\\alpha}_1}\\pmb {\\mathcal{E}}_\\theta^{(1)}(z_1)\\right) / \\sqrt{\\bar{\\alpha}_1}\\) \n12: return \\(Z_{0}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.296, + 0.712 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.72, + 0.279, + 0.737 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Dataset. We used four high-resolution (with resolutions exceeding 1080p) video datasets for evaluation, sourced from the DAVIS dataset [17] and the Pexels dataset1. A subset of 100 videos from the DAVIS dataset is resized to \\(768 \\times 1280\\) resolution and consists of 25 frames, originally provided in landscape orientation. The Pexels dataset is a large, open-source collection of high-resolution stock videos and images, widely used for creative and research purposes. For the Pexels subset, we collect a total of 120 videos: 45" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "in landscape orientation (Pexels (landscape)), 45 in vertical orientation (Pexels (vertical)), and 30 in square orientation (Pexels (square)). These subsets are resized to resolutions of \\(768 \\times 1280\\) for landscape, \\(1280 \\times 768\\) for vertical, and \\(1024 \\times 1024\\) for square orientations, with each video consisting of 25 frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.333 + ], + "angle": 0, + "content": "Inverse problems. We test our method on the following spatial degradations: 1) Deblur: Gaussian deblurring from an image convolved with a \\(61 \\times 61\\) size Gaussian kernel with \\(\\sigma = 3.0\\), 2) SR: Super-resolution from \\(\\times 4\\) average pooling, 3) Inpaint: Inpainting from \\(50\\%\\) random masking. Furthermore, test our method on the following spatio-temporal degradations: 4) Deblur+: Deblur + 7-frame averaging using temporal uniform blur kernel as used in [13], 5) SR+: SR + 7-frame averaging, and 6) Inpaint+: Inpaint + 7-frame averaging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.909, + 0.606 + ], + "angle": 0, + "content": "Baseline comparison. The primary objective of this study is to improve the performance of video inverse problem solvers through latent image diffusion models. Thus, our evaluation primarily compares video inverse problem solvers using image diffusion models. As a recently emerging field, only a few methods are available: SVI [13], DiffIR2VR [32], and Warped Diffusion [5]. Notably, DiffIR2VR and Warped Diffusion cannot address spatiotemporal degradations, and DiffIR2VR only supports SR among the inverse problems we address in this paper. We conducted comparisons with SVI and DiffIR2VR but excluded Warped Diffusion, as it is not currently open-source. SVI officially supports a resolution of \\(256 \\times 256\\), while DiffIR2VR supports \\(480 \\times 854\\). To ensure fair comparisons with identical resolutions, we used patch reconstruction. Additionally, we included a comparison with the classical optimization method ADMM-TV, following the protocol established by SVI [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.713 + ], + "angle": 0, + "content": "For quantitative comparison, we focus on two widely used standard metrics: peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) [28]. Additionally, we evaluate two perceptual metrics: Learned Perceptual Image Patch Similarity (LPIPS) [34] and Fréchet Video Distance (FVD) [26]. For computing the metrics, we follow the protocol from the open-source project2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.863 + ], + "angle": 0, + "content": "Implementation details. While our method is applicable to general latent diffusion models, we use Stable Diffusion XL 1.0 (SDXL) [16]—the current state-of-the-art text-to-image diffusion model—as a proof of concept in this paper. For all experiments, we employ \\( T = 25 \\), \\( \\tau = 0.3T \\), \\( \\lambda = 2 \\), and \\( l = 10 \\). These optimal values are obtained through extensive ablation studies and the results are described in Sec. 4.3. To reduce undesired guidance from the text condition \\( c_{\\mathrm{text}} \\), we use a null-text condition, \\( c_{\\emptyset} \\). All experiments were done on a single NVIDIA 4090 GPU." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.298, + 0.9 + ], + "angle": 0, + "content": "1https://www.pexels.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2https://github.com/JunyaoHu/common_metricson Video_quality" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10469" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.078, + 0.904, + 0.362 + ], + "angle": 0, + "content": "
TaskMethodDAVISPexels (landscape)Pexels (vertical)Pexels (square)
FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑
Deblur+ADMM-TV15120.39724.300.742671.80.28429.420.818709.20.23730.370.847569.50.22430.680.856
SVI [13]638.90.32228.040.799830.50.26530.420.831656.20.23930.400.856499.40.22131.590.862
DiffIR2VR [32]----------------
Ours228.60.29228.760.807196.40.24931.110.839209.90.22431.870.860157.90.21732.400.864
SR+ADMM-TV14290.35924.230.740634.10.27929.010.820669.20.32229.710.836545.90.30630.080.838
SVI [13]223.40.23429.000.812386.80.26530.700.831558.90.26130.480.842313.20.26531.180.847
DiffIR2VR [32]----------------
Ours158.50.24429.180.818166.20.24630.820.832173.10.22931.380.847138.80.22031.900.856
Inpaint+ADMM-TV18480.33924.160.762797.80.29229.150.778805.40.25829.360.805652.40.26830.390.804
SVI [13]208.60.23829.600.848269.30.25029.920.826428.10.24630.270.838206.90.23831.120.847
DiffIR2VR [32]----------------
Ours241.10.24228.810.815222.40.21630.130.828240.40.20130.980.845164.90.23031.440.853
DeblurADMM-TV169.00.23230.490.873192.20.26331.670.842185.70.18132.190.861199.60.25730.980.875
SVI [13]99.500.17631.700.875153.80.21232.440.862160.20.17433.350.866116.60.20733.510.885
DiffIR2VR [32]----------------
Ours72.030.17131.940.88096.160.18432.830.87689.390.17033.550.88891.550.15734.100.896
SRADMM-TV285.60.22126.650.788301.50.22227.910.769298.70.21327.120.786209.40.25728.740.798
SVI [13]176.10.17629.280.815201.60.20231.000.836219.30.20030.590.847167.90.20631.830.851
DiffIR2VR [32]319.60.23826.290.738412.60.24427.180.742511.60.21426.910.752412.00.29627.760.741
Ours81.100.19030.270.848104.30.18531.840.85898.570.15832.540.87698.410.13833.020.886
InpaintADMM-TV212.00.31528.420.797270.90.31629.140.794270.30.21229.840.803264.70.31030.460.809
SVI [13]143.50.17730.200.858159.40.23330.090.827164.30.13231.190.847139.30.22531.570.855
DiffIR2VR [32]----------------
Ours143.60.20929.740.835158.70.20830.410.834174.00.19531.150.847125.40.21631.630.859
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.368, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation (FVD, LPIPS, PSNR, SSIM) of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios (landscape, vertical, square). Bold denotes the best results and underline indicates the runner-up. Notably, DiffIR2VR [32] is only capable of restoring SR among our experimental tasks, highlighting the broader task generalizability of our approach." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.436, + 0.907, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.735, + 0.908, + 0.778 + ], + "angle": 0, + "content": "Figure 4. Qualitative evaluation of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios. Notably, ADMM-TV fails to remove ghosting artifacts caused by temporal degradation (red arrows), while SVI produces excessive intensity fluctuations (red box) or blurred information restoration (green and blue boxes)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.185, + 0.807 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Table 3 presents a quantitative comparison across various spatio-temporal inverse problems. The proposed method consistently outperforms baseline approaches in most metrics, particularly in addressing spatio-temporal degradations. Notably, it achieves a significant reduction in FVD," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.909, + 0.901 + ], + "angle": 0, + "content": "indicating superior perceptual video quality compared to the runner-up across all datasets. This improvement is also evident in the qualitative results shown in Fig. 4. While SVI [13] effectively handles spatio-temporal degradations, it often struggles with temporal consistency, leading to issues such as inaccurate intensity restoration (first row) and loss of details (second and third rows). This suggests" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.081, + 0.887, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.279, + 0.908, + 0.32 + ], + "angle": 0, + "content": "Figure 5. Qualitative evaluation of SR \\((\\times 4)\\) performance across multiple aspect ratios (landscape, vertical). DiffIR2VR often produces unwanted artifacts in the background (red and blue boxes), while SVI inaccurately restores intensity (red box), leading to frame-wise fluctuations." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.326, + 0.44, + 0.376 + ], + "angle": 0, + "content": "
MethodTime [min]Time per frame [sec]Memory (GB)
SVI [13]153618.5
DiffIR2VR [32]4.71113.6
Ours2.5612.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.435 + ], + "angle": 0, + "content": "Table 4. Comparison of total sampling time and memory efficiency for solving video inverse problems on a single 25-frame video at \\(768 \\times 1280\\) resolution. **Bold** denotes the best results and **underline** indicates the runner-up." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.483, + 0.717 + ], + "angle": 0, + "content": "that batch-consistent noise initialization [13], further examined in our ablation study (Table 5), may be insufficient to fully preserve temporal consistency. Additionally, its patch reconstruction for high-resolution videos can introduce patch-wise inconsistencies, degrading overall performance. The classical optimization method, ADMM-TV, effectively reconstructs static backgrounds and stationary objects but fails to remove ghosting artifacts caused by temporal deformations, as shown in Fig. 4. This limitation is reflected in its lower performance metrics. For the SR task (Fig.5), DiffIR2VR[32] often introduces unwanted artifacts in backgrounds or over-generates object details, likely due to inaccuracies in optical flow estimation. Notably, in pixel-wise random inpainting, latent diffusion methods may lose fine pixel-level details due to their encoded representations. However, leveraging the strong SDXL prior, our method achieves competitive inpainting performance with SVI, which employs a pixel-space diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.825 + ], + "angle": 0, + "content": "While baseline methods encounter various challenges across different inverse problems, our approach demonstrates stable, high-quality reconstructions, as evidenced by the overall results. Further comparisons of total sampling time and memory consumption are shown in Table 4, where our method achieves the highest efficiency in both sampling time and memory usage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Further experimental results on real-world denoising and blind video deconvolution are available in the supplementary material. Additional visualizations, including reconstruction results for deblurring, inpainting, and other tasks, are provided in video format for further evaluation: https:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.33, + 0.779, + 0.344 + ], + "angle": 0, + "content": "//vision-x1.github.io/supple/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.352, + 0.674, + 0.367 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.375, + 0.907, + 0.435 + ], + "angle": 0, + "content": "In this section, we analyze the key components of our method. To highlight their impact, we conduct an ablation study on the \\(\\mathrm{SR + }\\) task in the Pexels (landscape) dataset, which involves significant spatio-temporal degradation." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.446, + 0.905, + 0.511 + ], + "angle": 0, + "content": "
InitializationInitialization Time [sec]Total Time [min]FVD↓PSNR↑
Random noise08.3104729.43
Batch-consistent noise [13]08.3707.730.10
Pseudo-batch inversion (τ: 0.15T)151.3229.530.00
Pseudo-batch inversion (τ: 0.30T)302.5184.830.74
Pseudo-batch inversion (τ: 0.45T)453.8288.730.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.522, + 0.906, + 0.549 + ], + "angle": 0, + "content": "Table 5. Comparison of initialization strategies by computational cost and video quality for SR+ on the Pexels dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effect of initialization (in Step 1). In Table 5, we conduct an ablation study to see the effect of pseudo-batch inversion for initialization. From the table, we confirm that pseudobatch inversion effectively extracts informative latents to reconstruct video evidenced by about 0.6dB and 1.3dB PSNR increase compared with batch-synchronized noise initialization used in SVI [13] and random noise initialization, respectively. Notably, pseudo-batch inversion achieves \\(\\times 3\\) lower FVD compared to the batch-synchronized noise initialization which indicates a significant improvement in temporal consistency. It is also evident in the visualization of the ablation study shown in Fig. 6. The reconstruction results from random noise and batch-synchronized noise initialization fail to reconstruct the color of the cloud and are temporally inconsistent. In contrast, our method successfully reconstructs the color of the cloud and temporally consistent results. Furthermore, as shown in Table 5, pseudobatch inversion significantly improves sampling time efficiency. This is because it reduces the total sampling steps, and inversion does not require the measurement update process, which involves encoding and decoding. As a result, our method reconstructs high-definition video in under 6 seconds per frame on a single NVIDIA 4090 GPU." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.468, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of initialization strategies. Our approach improves temporal consistency and precise restoration." + }, + { + "type": "table", + "bbox": [ + 0.165, + 0.404, + 0.415, + 0.466 + ], + "angle": 0, + "content": "
Update step lFVD↓LPIPS↓PSNR↑SSIM↑
111500.28127.550.799
5241.00.19730.690.839
10184.80.23630.740.826
20486.10.47028.160.690
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.507 + ], + "angle": 0, + "content": "Table 6. Ablation study on the effect of optimization step \\(l\\). Ten update steps yield the best temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Effect of optimization step \\( l \\) (in Step 3). In Table 6, we present an ablation study on the effect of the CG update step \\( l \\). The table confirms that the CG update is essential for enhancing data consistency. We found that at least 5 iterations of CG updates yield satisfactory results, and 10 iterations produce the best results." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.627, + 0.403, + 0.706 + ], + "angle": 0, + "content": "
LPF λFVD↓LPIPS↓PSNR↑SSIM↑
No LPF209.70.27329.920.797
1186.30.23930.590.819
√2184.80.23630.740.826
2179.30.24530.810.832
2√2191.40.26230.890.837
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Table 7. Ablation study on the effect of low-pass filtering parameter \\( \\lambda \\) . \\( \\lambda = 2 \\) yields the best temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of LPF \\(\\lambda\\) (in Step 4). Table 7 presents an ablation study on the effect of low-pass filtering. The results confirm that low-pass filtering enhances the reconstruction quality as evidenced by all metrics. Specifically, low-pass filtering results in approximately a 30-point decrease in FVD and a 1.0dB increase in PSNR compared to the absence of low-pass filtering. This improvement is also evident in the visualizations in Fig. 7. In the second row of the figure, undesired artifacts appear when low-pass filtering is not applied. In contrast, as shown in the third and fourth rows," + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.093, + 0.887, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.365, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of low-pass filtering parameter \\(\\lambda\\). \\(\\lambda = 2\\) effectively removes artifacts from VAE errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.485 + ], + "angle": 0, + "content": "these artifacts are effectively removed as the parameter \\(\\lambda\\) increases. From a frequency-based perspective, we believe that low-pass filtering effectively guides the updated latents to remain within the desired denoised manifold, \\(\\mathcal{M}_0\\), and helps to mitigate error accumulation from the VAE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.634, + 0.514 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we proposed a novel framework for addressing high-definition video inverse problems using latent diffusion models that introduce two new strategies. First, a pseudo-batch consistent sampling strategy to manage intensive batch memory consumption with advanced latent diffusion models (e.g., SDXL). To acquire a denoised batch, we conduct sequential sampling of each latents rather than batch sampling to efficiently manage the high memory consumption of advanced latent diffusion models. Second, a pseudo-batch inversion for leveraging informative latent as initialization is proposed. We confirmed that pseudo-batch inversion significantly improves reconstruction performance in both traditional and perceptual quality metrics. Leveraging the powerful SDXL, our method achieves state-of-the-art performance across diverse spatio-temporal inverse problems, including challenging tasks such as the combination of frame averaging with deblurring, super-resolution, and inpainting. Importantly, our method supports multiple aspect ratios (landscape, vertical, and square), making it versatile for different video formats and delivering HD reconstructions in under 6 seconds per frame on a single NVIDIA 4090 GPU. Overall, our framework not only enhances video reconstruction quality but also sets new standards for efficiency and flexibility in solving high-definition video inverse problems." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10472" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.224 + ], + "angle": 0, + "content": "This work was supported by the National Research Foundation of Korea under Grant RS-2024-00336454, and by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. RS-2024-00457882, AI Research Hub Project; No. RS-2025-02304967, AI Star Fellowship (KAIST))" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.24, + 0.188, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.265, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[1] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.323, + 0.484, + 0.392 + ], + "angle": 0, + "content": "[2] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.395, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[3] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.453, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[4] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[5] Giannis Daras, Weili Nie, Karsten Kreis, Alex Dimakis, Morteza Mardani, Nikola Borislavov Kovachki, and Arash Vahdat. Warped diffusion: Solving video inverse problems with image diffusion models. arXiv preprint arXiv:2410.16152, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.583, + 0.484, + 0.625 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.628, + 0.484, + 0.668 + ], + "angle": 0, + "content": "[7] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.743, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.265, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[12] Jeongsol Kim, Geon Yeong Park, Hyungjin Chung, and Jong Chul Ye. Regularization by texts for latent diffusion inverse solvers. arXiv preprint arXiv:2311.15658, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Taesung Kwon and Jong Chul Ye. Solving video inverse problems using image diffusion models. In The Thirteenth International Conference on Learning Representations, 2025. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Dongyang Li, Chen Wei, Shiying Li, Jiachen Zou, and Quanying Liu. Visual decoding and reconstruction via eeg embeddings with guided diffusion. arXiv preprint arXiv:2403.07721, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[15] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[17] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[19] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[20] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[21] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[22] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Andreas Stöckl. Evaluating a synthetic image dataset generated with stable diffusion. In International Congress on Information and Communication Technology, pages 805-818. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Yu Takagi and Shinji Nishimoto. High-resolution image reconstruction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14453-14463, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10473" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[26] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[27] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[29] Jie Xiao, Ruili Feng, Han Zhang, Zhiheng Liu, Zhantao Yang, Yurui Zhu, Xueyang Fu, Kai Zhu, Yu Liu, and Zheng-Jun Zha. Dreamclean: Restoring clean image using deep diffusion prior. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[30] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[31] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[32] Chang-Han Yeh, Chin-Yang Lin, Zhixiang Wang, Chi-Wei Hsiao, Ting-Hsuan Chen, and Yu-Lun Liu. Diffir2vr-zero: Zero-shot video restoration with diffusion-based image restoration models. arXiv preprint arXiv:2407.01519, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10474" + } + ] +] \ No newline at end of file diff --git a/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_origin.pdf b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bace33e2e14b63a9bcb28c430a7986dbdb9c871f --- /dev/null +++ b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/339912e1-b9b8-4ac6-a699-6e214154b6e0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50142677037f70d4b0197b30ab2976fd04539999d57378933bb11fbff025a719 +size 1867062 diff --git a/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/full.md b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50df866661f1d5e568e7196d497b7713887f0937 --- /dev/null +++ b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/full.md @@ -0,0 +1,298 @@ +# VISION-XL: High Definition Video Inverse Problem Solver using Latent Image Diffusion Models + +Taesung Kwon KAIST + +star.kwon@kaist.ac.kr + +Jong Chul Ye KAIST + +jong.ye@kaist.ac.kr + +![](images/4947b8efca9a41483d9f71670b69062e99f2294903e7bffe05d68a5d4a7b6fe1.jpg) + +![](images/78fcfa979d5a60d395e874e72e311c5410babb0219ba2b2f7445241fc337c2e7.jpg) + +![](images/d4ad9f802ad006a053e6cd5238ca3d4b21170db5e6c4a1181f70a41ac55b99cd.jpg) + +![](images/25eb6fc1b7c688902cfb16896f58fb94ff38248e61afa69997a398367ba0678b.jpg) + +![](images/df863e580e5b2d7e34f66403acdbb80adaab488c926ee9b4c09a4ace5b14e32c.jpg) + +![](images/2dd0fdd5b8bb0736056be69ea844372f7ef9c3800d4474874e0a28182658e281.jpg) +Figure 1. Representative video reconstruction by VISION-XL: $\mathrm{SR + }$ (frame averaging with $\times 4$ super-resolution, top), Deblur+ (frame averaging with deblurring, $\sigma = 3.0$ , bottom-left), and Inpaint+ (frame averaging with $50\%$ random inpainting, bottom-right). + +![](images/87fd66dd26cfa1170791243348284b967be8a511e4925c853c23727c8c3bc4b4.jpg) + +# Abstract + +In this paper, we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models. Building on recent advancements in spatio-temporal optimization for video inverse problems using image diffusion models, our approach leverages latent-space diffusion models to achieve enhanced video quality and resolution. To address the high computational demands of processing high-resolution frames, we introduce a pseudo-batch consistent sampling strategy, allowing efficient operation on a single GPU. Additionally, to improve temporal consistency, we present pseudo-batch inversion, an initialization technique that incorporates informative latents from the measurement. By integrating with SDXL, our framework achieves state-of-the-art video reconstruction across a wide range of spatio-temporal inverse problems, including complex combinations of frame averaging + +ing and various spatial degradations, such as deblurring, super-resolution, and inpainting. Unlike previous methods, our approach supports multiple aspect ratios (landscape, vertical, and square) and delivers HD-resolution reconstructions (exceeding $1280 \times 720$ ) in under 6 seconds per frame on a single NVIDIA 4090 GPU. Project page: https://vision-x1.github.io/. + +# 1. Introduction + +Diffusion models [6, 9, 10, 16, 18, 20, 22] have set a new benchmark in generative modeling, enabling the generation of high-quality samples. These models have become the foundation for advancements in various fields, such as controllable image editing [33], image personalization [8], synthetic data augmentation [23], and even reconstructing images from brain signals [14, 24]. + +Furthermore, diffusion model-based inverse problem + +
MethodsSpatio-temporal degradationLatent space
Optical flow-based [5, 32]X✓
SVI [13]✓X
Ours✓✓
+ +solvers (DIS) [2, 4, 11, 21, 27, 29] address a variety of image restoration tasks, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on. A key feature of DIS is its plug-and-play capability, allowing diffusion models to be applied flexibly across different inverse problems without requiring task-specific training or fine-tuning. + +Recently, several extensions [5, 13, 32] from the DIS have been proposed to solve video inverse problems using image diffusion models. Naive application of image diffusion models to videos may break temporal consistency. To address this problem, these methods preserve temporal consistency by utilizing a batch-consistent sampling strategy [13] and applying optical flow guidance to warp either latent representations [32] or the noise prior [5]. + +Although these innovative approaches enable powerful image generative models [6, 16, 18] to solve video inverse problems with significantly reduced computational requirements, there is still room for improvement in these methods. Optical flow-based methods [5, 32] have reported a key limitation: their performance is highly dependent on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes problematic when extreme degradations complicate the estimation process, restricting their applicability to a wider range of restoration tasks. Additionally, these methods require task-specific restoration modules [32] or fine-tuning of the diffusion model [5]. In this perspective, batch consistent sampling strategy [13] successfully addressed various spatio-temporal degradations without requiring task-specific training or fine-tuning. However, we empirically found that extending SVI [13] to latent diffusion models leads to unsatisfactory reconstruction, particularly in terms of FVD [26], as shown in Table 2, despite the fact that most modern latent diffusion models are essential for scaling up to large-size video inverse problems. + +To overcome these limitations, here we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models that involves no additional training. To address the high computational demands of batch processing (e.g., 16-frame batch processing used in [13]) with high-resolution latent diffusion models [16], we introduce pseudo-batch consistent sampling. This strategy enables multi-frame video processing while requiring only the memory needed for a single frame, making it feasible on a single GPU. Furthermore, we propose pseudo-batch inversion, which initializes the process with + +Table 1. Comparison of image diffusion-based video inverse problem solvers. Unlike other methods, our approach leverages latent image diffusion models to address spatio-temporal degradation, enabling more effective restoration. + +
InitializationFVD ↓LPIPS ↓PSNR ↑
Random noise10470.25129.43
Batch synchronized noise (SVI [13])707.70.24830.10
Pseudo-batch inversion (Ours)184.80.23630.74
+ +Table 2. Impact of our initialization method on SR+ video restoration using SDXL [16]. Our method significantly improves performance, reducing FVD [26] by more than $3\times$ + +informative latents derived from the measurement. This initialization enhances temporal consistency and improves the efficiency of solving spatio-temporal inverse problems as shown in Table. 2. + +By integrating these components, our framework achieves state-of-the-art video reconstruction performance using SDXL [16]. We name the method integrating these components as VISION-XL, short for Video Inverse-problem Solver using latent diffusION models (with stable diffusion XL). It supports various aspect ratios, including landscape, vertical, and square formats. Thanks to its efficiency, our framework can reconstruct $1280 \times 768$ (exceeding HD resolution) videos in under 6 seconds per frame on a single NVIDIA 4090 GPU. Our contribution can be summarized as follows: + +- We propose a high-definition video inverse problem solver integrated with SDXL, supporting multiple aspect ratios and achieving state-of-the-art reconstruction. +- We introduce a novel pseudo-batch consistent sampling and inversion strategy for efficient and effective video reconstruction across diverse inverse problems. + +# 2. Related Work + +Diffusion model-based inverse problem solvers (DIS). Diffusion models [9, 20, 22] attempt to model the data distribution $p_{\theta}(\pmb{x})$ based on the Gaussian transitions. In the geometric view of diffusion models [1], the transitions are typically described as iterative manifold transitions $\mathcal{M}_t \to \mathcal{M}_{t-1}, t = T, \dots, 1$ , moving from the noisy manifold $\mathcal{M}_T$ to the clean manifold $\mathcal{M}_0$ . + +Diffusion model-based inverse problem solvers (DIS) [2, 4, 11, 21, 27] aim to guide manifold transitions to sample from the posterior distribution $p_{\theta}(\boldsymbol{x}|\boldsymbol{y})$ , which represents sampling $\boldsymbol{x}$ from the measurement $\boldsymbol{y}$ obtained from the forward model $\mathcal{A}(\boldsymbol{x})$ . In Bayesian inference, the posterior distribution, $p_{\theta}(\boldsymbol{x}|\boldsymbol{y}) \propto p_{\theta}(\boldsymbol{x})p(\boldsymbol{y}|\boldsymbol{x})$ is decomposed into the likelihood $p(\boldsymbol{y}|\boldsymbol{x})$ , representing the probability of observing $\boldsymbol{y}$ given $\boldsymbol{x}$ , and the prior data distribution $p_{\theta}(\boldsymbol{x})$ . This decomposition enables posterior sampling by combining diffusion sampling with iterative guidance using the forward model $\mathcal{A}$ and measurement $\boldsymbol{y}$ . This approach provides sophisticated, precise solutions to complex inverse problems, leveraging the power and flexibility of diffusion models in practical applications, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on. + +DIS using latent diffusion models (LDIS). Most DIS [2, 4, 11, 21, 27] use pixel-space diffusion models, which facilitate easy integration of the forward model $\mathcal{A}$ and measurement $\pmb{y}$ , as both are defined in pixel space. Integrating forward models in latent space presents more challenges. Latent-space methods [3, 12, 19] calculate data consistency terms after decoding the denoised latent representation, then update these guidances within the latent space. + +During this process, VAE mapping errors accumulate in iterative sampling, causing the representation to drift from the clean manifold $\mathcal{M}_0$ . Additionally, most latent diffusion models provide a text-conditioned prior distribution $p_{\theta}(\boldsymbol{x}|c_{\mathrm{text}})$ , which is challenging to implement in cases where text conditioning ( $c_{\mathrm{text}}$ ) is unavailable. As a result, latent-space methods prioritize two main goals: (i) managing text embeddings effectively and (ii) preserving the updated latent close to the clean manifold $\mathcal{M}_0$ . + +For text embeddings, PSLD [19] uses only null-text, while TReg [12] and P2L [3] apply either null-text optimization or text optimization to enhance the reconstruction performance. To maintain the updated latent representation quality, the regularization term for aligning pixel and latent spaces is used to enforce latent feasibility [3, 12, 19]. + +Solving video inverse problems using DIS. Recently, several extensions [5, 13, 32] from DIS have been introduced to address video inverse problems. A straightforward application of image diffusion models to video, processing frames individually, often disrupts temporal consistency. These approaches maintain temporal coherence by employing a batch-consistent sampling strategy [13] and leveraging optical flow guidance to warp either latent representations [32] or the noise prior [5]. + +While these innovative approaches allow powerful image generative models [6, 16, 18] to address video inverse problems with reduced computational demands, there is still room for improvement. A key limitation of optical flow-based methods [5, 32] is their heavy reliance on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes particularly problematic in scenarios involving severe degradations, which can hinder the estimation process and limit the methods' applicability to broader restoration tasks. Moreover, these approaches often require task-specific restoration modules [32] or fine-tuning of the diffusion model [5]. + +In contrast, the batch-consistent sampling strategy [13] has demonstrated effectiveness in addressing various spatiotemporal degradations without the need for task-specific training or model fine-tuning. However, this method utilizes the unconditional pixel-space diffusion model provided by ADM [6], and its extension to latent diffusion models remains unexplored. + +# 3. High Definition Video Inverse Solver Using Latent Diffusion Models + +This section introduces a novel approach for reconstructing high-definition videos that include various spatio-temporal degradations. The overall pipeline of the algorithm is illustrated in Fig. 2. + +Consider the spatio-temporal degradation process is formulated as: + +$$ +\boldsymbol {Y} = \mathcal {A} (\boldsymbol {X}) = \mathcal {A} ([ \boldsymbol {X} [ 1 ], \dots , \boldsymbol {X} [ N ]) ] \tag {1} +$$ + +where $\mathbf{Y}$ denotes the measurement, $X[n]$ denotes the $n$ -th frame ground-truth frame, $N$ is the number of video frames, and $\mathcal{A}$ refers to the operator describing the spatio-temporal degradation process. The detailed spatio-temporal degradation operator $\mathcal{A}$ is described in Sec. 4.1. + +Our approach begins by inverting the measurement frames, denoted as $\bar{Y}$ , to initialize the informative latents $z_{\tau}$ , enhancing batch-wise consistency (Step 1). Next, we construct the corresponding denoised batch $\hat{\mathbf{X}}_{\tau}$ by sampling each latent sequentially using Tweedie's formula [7], followed by decoding (Step 2). In Step 3, the corresponding denoised batch $\hat{\mathbf{X}}_{\tau}$ is further refined by applying $l$ -step conjugate gradient (CG) optimization [4, 13] to enforce the data consistency from spatio-temporal degradation $\mathcal{A}$ . In Step 4, we apply a scheduled low-pass filter to the updated batch $\bar{\mathbf{X}}_{\tau}$ inspired by the frequency-based analysis of spectral diffusion [31]. $\bar{\mathbf{X}}_{\tau}$ is then re-encoded into latent space to form $\bar{z}_{\tau}$ . Finally, we obtain the one-step denoised latents $z_{\tau - 1}$ by adding noise to the encoded latents $\bar{z}_{\tau}$ (Step 5). In the following, we provide a detailed description of each of these steps. + +Step 1: Initialize informative latents. One of our key insights is to initialize the informative latents by inverting the measurement frames, thereby enhancing batch-wise consistent initialization. Although these latents cannot restore the ground-truth frames directly, inverted latent variables can inherit information from the measurement frame, providing good initializations [29]. Different from SVI [13], which initializes sampling from a batch-wise synchronized uninformative Gaussian prior $z_{T} \sim \mathcal{N}(0, I)$ as the initial sampling point, we replace $z_{T}$ with the informative prior $z_{\tau}$ , defined as: + +$$ +\boldsymbol {z} _ {0} = \boldsymbol {E} _ {\theta} (\boldsymbol {Y}), \quad \boldsymbol {z} _ {\tau} = \mathrm {D D I M} ^ {- 1} \left(\boldsymbol {z} _ {0}\right), \tag {2} +$$ + +where $E_{\theta}(\cdot)$ and $\mathrm{DDIM}^{-1}(\cdot)$ denote frame-wise encoding from pretrained VAE and DDIM inversion of timestep $\tau$ respectively. + +Step 2: Pseudo-batch sampling. After initialization, we guide the sampling path to ensure the data consistency condition. At timestep $0 < t \leq \tau$ , we sample denoise batch $\hat{z}_t$ from given latents $z_t \coloneqq [z_t[1] \cdots z_t[N]]$ by using Tweedie's formula [7]. Unlike SVI [13], we split latent + +![](images/19e9d2e13fc3ea1b81ae162bd267bdf3b7536d1ccc29225bf614217e236133f8.jpg) +Figure 2. Illustration of VISION-XL sampling at timestep $t$ : $z_{t}$ is split into individual frames and denoised sequentially using Tweedie's formula. The denoised latents $\hat{z}_{t}$ are then merged and decoded. The decoded batch $\hat{\mathbf{X}}_t$ is optimized to enforce the data consistency, followed by low-pass filtered encoding and re-noising to obtain $z_{t-1}$ . + +frames to construct a pseudo-batch and sequentially sample each frame, requiring memory for only a single frame during the sampling, as shown in Fig. 2. Similarly, inversion is also conducted within the pseudo-batch framework. This enables the recent advanced latent diffusion model to operate in this framework without a frame limit. As a proof-of-concept, we conduct experiments on 25-frame videos. + +Specifically, consider a sequential sampling of latent diffusion models $\epsilon_{\theta}$ along the temporal direction: + +$$ +\boldsymbol {\mathcal {E}} _ {\theta} ^ {(t)} (\boldsymbol {z} _ {t}) := \left[ \boldsymbol {\epsilon} _ {\theta} ^ {(t)} (\boldsymbol {z} _ {t} [ 1 ]) \dots \boldsymbol {\epsilon} _ {\theta} ^ {(t)} (\boldsymbol {z} _ {t} [ N ]) \right]. \tag {3} +$$ + +The denoised latents $\hat{Z}_t$ are computed using Tweedie's formula [7]: + +$$ +\hat {\boldsymbol {z}} _ {t} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\boldsymbol {z} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\mathcal {E}} _ {\theta} ^ {(t)} (\boldsymbol {z} _ {t})\right), \tag {4} +$$ + +where $\bar{\alpha}_t$ is the noise schedule defined in the Gaussian process of diffusion models [9, 15]. Then denoised batch $\hat{X}_t$ is decoded from the denoised latents $\hat{z}_t$ using VAE decoder $D_{\theta}$ : + +$$ +\hat {\boldsymbol {X}} _ {t} = \boldsymbol {D} _ {\theta} (\hat {\boldsymbol {z}} _ {t}) := [ \boldsymbol {D} _ {\theta} (\hat {\boldsymbol {z}} _ {t} [ 1 ]) \dots \boldsymbol {D} _ {\theta} (\hat {\boldsymbol {z}} _ {t} [ N ]) ]. \quad (5) +$$ + +Step 3: Batch optimization in pixel-space. The denoised batch $\hat{\pmb{X}}_t$ is then refined as a whole by applying the $l$ -step CG optimization to enhance the data consistency from the measurement $\pmb{Y}$ and spatio-temporal degradation $\mathcal{A}$ . This can be formally represented by + +$$ +\bar {\boldsymbol {X}} _ {t} := \underset {\boldsymbol {X} \in \hat {\boldsymbol {X}} _ {t} + \boldsymbol {\mathcal {K}} _ {l}} {\arg \min } \| \boldsymbol {Y} - \mathcal {A} (\boldsymbol {X}) \| ^ {2} \tag {6} +$$ + +where $\mathcal{K}_l$ denotes the $l$ -dimensional Kyrlov subspace associated with the given inverse problem [4]. The multistep CG allows each temporal frame to be diversified, enhancing data consistency and achieving faster convergence without requiring memory-intensive gradient calculations [2]. + +Step 4: Low-pass filtered encoding. Recent frequency-based analyses of diffusion models [10, 31] suggest that optimal denoisers first recover low-frequency components in + +early denoising stages, while high-frequency details added progressively in later stages. Building on these findings, we observed that applying a scheduled low-pass filter to the updated batch $\bar{X}_t$ in early stages effectively removes undesired artifact caused by VAE error accumulation, resulting in more natural and refined outputs. + +Based on the observation that the denoiser restores high-frequency details as the noise scale $\sqrt{1 - \bar{\alpha}_t}$ decreases, we set the filter width $\sigma_t$ to be proportional to the noise scale [15], defined as $\sigma_t \coloneqq \lambda \sqrt{1 - \bar{\alpha}_t}$ , which goes to zero at $t \to 0$ . After applying the low-pass filter $h_{\sigma_t}$ , we re-encode $\bar{\boldsymbol{X}}_t$ into the latent space. Specifically, the re-encoded latents are given by: + +$$ +\bar {\boldsymbol {X}} _ {t} \leftarrow \bar {\boldsymbol {X}} _ {t} * \boldsymbol {h} _ {\sigma_ {t}}, \tag {7} +$$ + +$$ +\bar {z} _ {t} = \boldsymbol {E} _ {\theta} (\bar {\boldsymbol {X}} _ {t}) := [ \boldsymbol {E} _ {\theta} (\bar {\boldsymbol {X}} _ {t} [ 1 ]) \dots \boldsymbol {E} _ {\theta} (\bar {\boldsymbol {X}} _ {t} [ N ]) ], \tag {8} +$$ + +where $E_{\theta}$ denotes the VAE encoder. + +Step 5: Renoising. After encoding, updated latents $\bar{z}_t$ are renoised as: + +$$ +\boldsymbol {z} _ {t - 1} = \sqrt {\bar {\alpha} _ {t - 1}} \bar {\boldsymbol {z}} _ {t} + \sqrt {1 - \bar {\alpha} _ {t - 1}} \boldsymbol {\mathcal {E}} _ {t}, \tag {9} +$$ + +where $\pmb{\varepsilon}_{t}$ is composed of batch-consistent stochastic noise [13] and deterministic noise [20]. + +In the geometric view of diffusion models [1], the sampling path evolves as illustrated in Fig. 3. The initialized latent $z_{\tau}$ is projected onto the clean manifold $\mathcal{M}_0$ using Tweedie's formula. The projected latent is then decoded into the pixel space and refined through multi-step CG to satisfy the data consistency constraint $Y = \mathcal{A}(X)$ . Next, a scheduled low-pass filter is applied to reduce VAE error accumulation and keep the encoding close to $\mathcal{M}_0$ . Finally, the encoded latents are re-noised to transition back to $\mathcal{M}_{\tau -1}$ , and this process iterates until the final state converges to the clean manifold $\mathcal{M}_0$ . The complete algorithm is provided in Algorithm 1. + +![](images/914006bdc2f45db2ebb1f074074d053b35372efb9aabca0c78d018f2dc80a405.jpg) +Figure 3. Geometric illustration of the sampling path evolution. (Step 1) Initialize latent $\pmb{z}_{\tau}$ . (Step 2) Project onto $\mathcal{M}_0$ via pseudobatch sampling and decode to pixel space. (Step 3) Optimize for measurement consistency $\pmb{Y} = \mathcal{A}(\pmb{X})$ . (Step 4) Apply a scheduled low-pass filter and encode back to latent space. (Step 5) Renoise to $\mathcal{M}_{\tau -1}$ . + +Algorithm 1 High-definition video inverse problem solver using latent diffusion models +Require: $\pmb{\mathcal{E}}_{\theta}^{(t)},\pmb {E}_{\theta},\pmb {D}_{\theta},\pmb {Y},\mathcal{A},\tau ,l,\sigma_{t},\{\bar{\alpha}_{t}\}_{t = 1}^{T}$ +1: $z_0\gets E_\theta (Y)$ +2: $z_{\tau}\gets \mathrm{DDIM}^{-1}(z_0)$ ▷ Step 1 +3: for $t = \tau :2$ do +4: $\hat{z}_t\gets \left(z_t - \sqrt{1 - \bar{\alpha}_t}\pmb {\mathcal{E}}_\theta^{(t)}(z_t)\right) / \sqrt{\bar{\alpha}_t}$ ▷ Step 2 +5: $\hat{\pmb{X}}_t\gets \pmb {D}_\theta (\hat{\pmb{z}}_t)$ +6: $\bar{\pmb{X}}_t\coloneqq \arg \min_{\pmb {X}\in \hat{\pmb{X}}_t + \pmb {\kappa}_l}\| \pmb {Y} - \mathcal{A}(\pmb {X})\| ^2$ ▷ Step 3 +7: $\bar{\pmb{X}}_t\gets \bar{\pmb{X}}_t*\pmb {h}_{\sigma_t}$ ▷ Step 4 +8: $\bar{z}_t = E_\theta (\bar{X}_t)$ +9: $z_{t - 1} = \sqrt{\bar{\alpha}_{t - 1}}\bar{z}_t + \sqrt{1 - \bar{\alpha}_{t - 1}}\pmb {\mathcal{E}}_t$ ▷ Step 5 +10: end for +11: $z_0\gets \left(z_1 - \sqrt{1 - \bar{\alpha}_1}\pmb {\mathcal{E}}_\theta^{(1)}(z_1)\right) / \sqrt{\bar{\alpha}_1}$ +12: return $Z_{0}$ + +# 4. Experimental Results + +# 4.1. Experimental setup + +Dataset. We used four high-resolution (with resolutions exceeding 1080p) video datasets for evaluation, sourced from the DAVIS dataset [17] and the Pexels dataset1. A subset of 100 videos from the DAVIS dataset is resized to $768 \times 1280$ resolution and consists of 25 frames, originally provided in landscape orientation. The Pexels dataset is a large, open-source collection of high-resolution stock videos and images, widely used for creative and research purposes. For the Pexels subset, we collect a total of 120 videos: 45 + +in landscape orientation (Pexels (landscape)), 45 in vertical orientation (Pexels (vertical)), and 30 in square orientation (Pexels (square)). These subsets are resized to resolutions of $768 \times 1280$ for landscape, $1280 \times 768$ for vertical, and $1024 \times 1024$ for square orientations, with each video consisting of 25 frames. + +Inverse problems. We test our method on the following spatial degradations: 1) Deblur: Gaussian deblurring from an image convolved with a $61 \times 61$ size Gaussian kernel with $\sigma = 3.0$ , 2) SR: Super-resolution from $\times 4$ average pooling, 3) Inpaint: Inpainting from $50\%$ random masking. Furthermore, test our method on the following spatio-temporal degradations: 4) Deblur+: Deblur + 7-frame averaging using temporal uniform blur kernel as used in [13], 5) SR+: SR + 7-frame averaging, and 6) Inpaint+: Inpaint + 7-frame averaging. + +Baseline comparison. The primary objective of this study is to improve the performance of video inverse problem solvers through latent image diffusion models. Thus, our evaluation primarily compares video inverse problem solvers using image diffusion models. As a recently emerging field, only a few methods are available: SVI [13], DiffIR2VR [32], and Warped Diffusion [5]. Notably, DiffIR2VR and Warped Diffusion cannot address spatiotemporal degradations, and DiffIR2VR only supports SR among the inverse problems we address in this paper. We conducted comparisons with SVI and DiffIR2VR but excluded Warped Diffusion, as it is not currently open-source. SVI officially supports a resolution of $256 \times 256$ , while DiffIR2VR supports $480 \times 854$ . To ensure fair comparisons with identical resolutions, we used patch reconstruction. Additionally, we included a comparison with the classical optimization method ADMM-TV, following the protocol established by SVI [13]. + +For quantitative comparison, we focus on two widely used standard metrics: peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) [28]. Additionally, we evaluate two perceptual metrics: Learned Perceptual Image Patch Similarity (LPIPS) [34] and Fréchet Video Distance (FVD) [26]. For computing the metrics, we follow the protocol from the open-source project2. + +Implementation details. While our method is applicable to general latent diffusion models, we use Stable Diffusion XL 1.0 (SDXL) [16]—the current state-of-the-art text-to-image diffusion model—as a proof of concept in this paper. For all experiments, we employ $T = 25$ , $\tau = 0.3T$ , $\lambda = 2$ , and $l = 10$ . These optimal values are obtained through extensive ablation studies and the results are described in Sec. 4.3. To reduce undesired guidance from the text condition $c_{\mathrm{text}}$ , we use a null-text condition, $c_{\emptyset}$ . All experiments were done on a single NVIDIA 4090 GPU. + +
TaskMethodDAVISPexels (landscape)Pexels (vertical)Pexels (square)
FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑
Deblur+ADMM-TV15120.39724.300.742671.80.28429.420.818709.20.23730.370.847569.50.22430.680.856
SVI [13]638.90.32228.040.799830.50.26530.420.831656.20.23930.400.856499.40.22131.590.862
DiffIR2VR [32]----------------
Ours228.60.29228.760.807196.40.24931.110.839209.90.22431.870.860157.90.21732.400.864
SR+ADMM-TV14290.35924.230.740634.10.27929.010.820669.20.32229.710.836545.90.30630.080.838
SVI [13]223.40.23429.000.812386.80.26530.700.831558.90.26130.480.842313.20.26531.180.847
DiffIR2VR [32]----------------
Ours158.50.24429.180.818166.20.24630.820.832173.10.22931.380.847138.80.22031.900.856
Inpaint+ADMM-TV18480.33924.160.762797.80.29229.150.778805.40.25829.360.805652.40.26830.390.804
SVI [13]208.60.23829.600.848269.30.25029.920.826428.10.24630.270.838206.90.23831.120.847
DiffIR2VR [32]----------------
Ours241.10.24228.810.815222.40.21630.130.828240.40.20130.980.845164.90.23031.440.853
DeblurADMM-TV169.00.23230.490.873192.20.26331.670.842185.70.18132.190.861199.60.25730.980.875
SVI [13]99.500.17631.700.875153.80.21232.440.862160.20.17433.350.866116.60.20733.510.885
DiffIR2VR [32]----------------
Ours72.030.17131.940.88096.160.18432.830.87689.390.17033.550.88891.550.15734.100.896
SRADMM-TV285.60.22126.650.788301.50.22227.910.769298.70.21327.120.786209.40.25728.740.798
SVI [13]176.10.17629.280.815201.60.20231.000.836219.30.20030.590.847167.90.20631.830.851
DiffIR2VR [32]319.60.23826.290.738412.60.24427.180.742511.60.21426.910.752412.00.29627.760.741
Ours81.100.19030.270.848104.30.18531.840.85898.570.15832.540.87698.410.13833.020.886
InpaintADMM-TV212.00.31528.420.797270.90.31629.140.794270.30.21229.840.803264.70.31030.460.809
SVI [13]143.50.17730.200.858159.40.23330.090.827164.30.13231.190.847139.30.22531.570.855
DiffIR2VR [32]----------------
Ours143.60.20929.740.835158.70.20830.410.834174.00.19531.150.847125.40.21631.630.859
+ +Table 3. Quantitative evaluation (FVD, LPIPS, PSNR, SSIM) of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios (landscape, vertical, square). Bold denotes the best results and underline indicates the runner-up. Notably, DiffIR2VR [32] is only capable of restoring SR among our experimental tasks, highlighting the broader task generalizability of our approach. + +![](images/ebc344f22dcbc44e9bc03683f2a303babc26e2833978cca0bfe83af7a792ec48.jpg) +Figure 4. Qualitative evaluation of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios. Notably, ADMM-TV fails to remove ghosting artifacts caused by temporal degradation (red arrows), while SVI produces excessive intensity fluctuations (red box) or blurred information restoration (green and blue boxes). + +# 4.2. Results + +Table 3 presents a quantitative comparison across various spatio-temporal inverse problems. The proposed method consistently outperforms baseline approaches in most metrics, particularly in addressing spatio-temporal degradations. Notably, it achieves a significant reduction in FVD, + +indicating superior perceptual video quality compared to the runner-up across all datasets. This improvement is also evident in the qualitative results shown in Fig. 4. While SVI [13] effectively handles spatio-temporal degradations, it often struggles with temporal consistency, leading to issues such as inaccurate intensity restoration (first row) and loss of details (second and third rows). This suggests + +![](images/1b7dbc97f9b4c9aace0c561ad3cdb3e609f28b5a2d36dd59e1e711e0e0c446d3.jpg) +Figure 5. Qualitative evaluation of SR $(\times 4)$ performance across multiple aspect ratios (landscape, vertical). DiffIR2VR often produces unwanted artifacts in the background (red and blue boxes), while SVI inaccurately restores intensity (red box), leading to frame-wise fluctuations. + +
MethodTime [min]Time per frame [sec]Memory (GB)
SVI [13]153618.5
DiffIR2VR [32]4.71113.6
Ours2.5612.7
+ +that batch-consistent noise initialization [13], further examined in our ablation study (Table 5), may be insufficient to fully preserve temporal consistency. Additionally, its patch reconstruction for high-resolution videos can introduce patch-wise inconsistencies, degrading overall performance. The classical optimization method, ADMM-TV, effectively reconstructs static backgrounds and stationary objects but fails to remove ghosting artifacts caused by temporal deformations, as shown in Fig. 4. This limitation is reflected in its lower performance metrics. For the SR task (Fig.5), DiffIR2VR[32] often introduces unwanted artifacts in backgrounds or over-generates object details, likely due to inaccuracies in optical flow estimation. Notably, in pixel-wise random inpainting, latent diffusion methods may lose fine pixel-level details due to their encoded representations. However, leveraging the strong SDXL prior, our method achieves competitive inpainting performance with SVI, which employs a pixel-space diffusion model. + +While baseline methods encounter various challenges across different inverse problems, our approach demonstrates stable, high-quality reconstructions, as evidenced by the overall results. Further comparisons of total sampling time and memory consumption are shown in Table 4, where our method achieves the highest efficiency in both sampling time and memory usage. + +Further experimental results on real-world denoising and blind video deconvolution are available in the supplementary material. Additional visualizations, including reconstruction results for deblurring, inpainting, and other tasks, are provided in video format for further evaluation: https: + +//vision-x1.github.io/supple/. + +# 4.3. Ablation studies + +In this section, we analyze the key components of our method. To highlight their impact, we conduct an ablation study on the $\mathrm{SR + }$ task in the Pexels (landscape) dataset, which involves significant spatio-temporal degradation. + +Table 4. Comparison of total sampling time and memory efficiency for solving video inverse problems on a single 25-frame video at $768 \times 1280$ resolution. **Bold** denotes the best results and **underline** indicates the runner-up. + +
InitializationInitialization Time [sec]Total Time [min]FVD↓PSNR↑
Random noise08.3104729.43
Batch-consistent noise [13]08.3707.730.10
Pseudo-batch inversion (τ: 0.15T)151.3229.530.00
Pseudo-batch inversion (τ: 0.30T)302.5184.830.74
Pseudo-batch inversion (τ: 0.45T)453.8288.730.70
+ +Table 5. Comparison of initialization strategies by computational cost and video quality for SR+ on the Pexels dataset. + +Effect of initialization (in Step 1). In Table 5, we conduct an ablation study to see the effect of pseudo-batch inversion for initialization. From the table, we confirm that pseudobatch inversion effectively extracts informative latents to reconstruct video evidenced by about 0.6dB and 1.3dB PSNR increase compared with batch-synchronized noise initialization used in SVI [13] and random noise initialization, respectively. Notably, pseudo-batch inversion achieves $\times 3$ lower FVD compared to the batch-synchronized noise initialization which indicates a significant improvement in temporal consistency. It is also evident in the visualization of the ablation study shown in Fig. 6. The reconstruction results from random noise and batch-synchronized noise initialization fail to reconstruct the color of the cloud and are temporally inconsistent. In contrast, our method successfully reconstructs the color of the cloud and temporally consistent results. Furthermore, as shown in Table 5, pseudobatch inversion significantly improves sampling time efficiency. This is because it reduces the total sampling steps, and inversion does not require the measurement update process, which involves encoding and decoding. As a result, our method reconstructs high-definition video in under 6 seconds per frame on a single NVIDIA 4090 GPU. + +![](images/51a94a435cc7029a6530a15131f8518c88ff159de6b77e49c44ebc502ec8d0ba.jpg) +Figure 6. Qualitative comparison of initialization strategies. Our approach improves temporal consistency and precise restoration. + +
Update step lFVD↓LPIPS↓PSNR↑SSIM↑
111500.28127.550.799
5241.00.19730.690.839
10184.80.23630.740.826
20486.10.47028.160.690
+ +Effect of optimization step $l$ (in Step 3). In Table 6, we present an ablation study on the effect of the CG update step $l$ . The table confirms that the CG update is essential for enhancing data consistency. We found that at least 5 iterations of CG updates yield satisfactory results, and 10 iterations produce the best results. + +Table 6. Ablation study on the effect of optimization step $l$ . Ten update steps yield the best temporal consistency. + +
LPF λFVD↓LPIPS↓PSNR↑SSIM↑
No LPF209.70.27329.920.797
1186.30.23930.590.819
√2184.80.23630.740.826
2179.30.24530.810.832
2√2191.40.26230.890.837
+ +Table 7. Ablation study on the effect of low-pass filtering parameter $\lambda$ . $\lambda = 2$ yields the best temporal consistency. + +Effect of LPF $\lambda$ (in Step 4). Table 7 presents an ablation study on the effect of low-pass filtering. The results confirm that low-pass filtering enhances the reconstruction quality as evidenced by all metrics. Specifically, low-pass filtering results in approximately a 30-point decrease in FVD and a 1.0dB increase in PSNR compared to the absence of low-pass filtering. This improvement is also evident in the visualizations in Fig. 7. In the second row of the figure, undesired artifacts appear when low-pass filtering is not applied. In contrast, as shown in the third and fourth rows, + +![](images/685af7bf18c39cab7723baa4af91bf19e62d4c7509ea244827039cdbf1231341.jpg) +Figure 7. Qualitative comparison of low-pass filtering parameter $\lambda$ . $\lambda = 2$ effectively removes artifacts from VAE errors. + +these artifacts are effectively removed as the parameter $\lambda$ increases. From a frequency-based perspective, we believe that low-pass filtering effectively guides the updated latents to remain within the desired denoised manifold, $\mathcal{M}_0$ , and helps to mitigate error accumulation from the VAE. + +# 5. Conclusion + +In this paper, we proposed a novel framework for addressing high-definition video inverse problems using latent diffusion models that introduce two new strategies. First, a pseudo-batch consistent sampling strategy to manage intensive batch memory consumption with advanced latent diffusion models (e.g., SDXL). To acquire a denoised batch, we conduct sequential sampling of each latents rather than batch sampling to efficiently manage the high memory consumption of advanced latent diffusion models. Second, a pseudo-batch inversion for leveraging informative latent as initialization is proposed. We confirmed that pseudo-batch inversion significantly improves reconstruction performance in both traditional and perceptual quality metrics. Leveraging the powerful SDXL, our method achieves state-of-the-art performance across diverse spatio-temporal inverse problems, including challenging tasks such as the combination of frame averaging with deblurring, super-resolution, and inpainting. Importantly, our method supports multiple aspect ratios (landscape, vertical, and square), making it versatile for different video formats and delivering HD reconstructions in under 6 seconds per frame on a single NVIDIA 4090 GPU. Overall, our framework not only enhances video reconstruction quality but also sets new standards for efficiency and flexibility in solving high-definition video inverse problems. + +# Acknowledgements + +This work was supported by the National Research Foundation of Korea under Grant RS-2024-00336454, and by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. RS-2024-00457882, AI Research Hub Project; No. RS-2025-02304967, AI Star Fellowship (KAIST)) + +# References + +[1] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 2, 4 +[2] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4 +[3] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 3 +[4] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 4 +[5] Giannis Daras, Weili Nie, Karsten Kreis, Alex Dimakis, Morteza Mardani, Nikola Borislavov Kovachki, and Arash Vahdat. Warped diffusion: Solving video inverse problems with image diffusion models. arXiv preprint arXiv:2410.16152, 2024. 2, 3, 5 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2, 3 +[7] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 3, 4 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4 +[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 4 +[11] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2, 3 + +[12] Jeongsol Kim, Geon Yeong Park, Hyungjin Chung, and Jong Chul Ye. Regularization by texts for latent diffusion inverse solvers. arXiv preprint arXiv:2311.15658, 2023. 3 +[13] Taesung Kwon and Jong Chul Ye. Solving video inverse problems using image diffusion models. In The Thirteenth International Conference on Learning Representations, 2025. 2, 3, 4, 5, 6, 7 +[14] Dongyang Li, Chen Wei, Shiying Li, Jiachen Zou, and Quanying Liu. Visual decoding and reconstruction via eeg embeddings with guided diffusion. arXiv preprint arXiv:2403.07721, 2024. 1 +[15] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4 +[16] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 5 +[17] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5 +[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3 +[19] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[20] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 4 +[21] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3 +[22] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1, 2 +[23] Andreas Stöckl. Evaluating a synthetic image dataset generated with stable diffusion. In International Congress on Information and Communication Technology, pages 805-818. Springer, 2023. 1 +[24] Yu Takagi and Shinji Nishimoto. High-resolution image reconstruction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14453-14463, 2023. 1 +[25] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV + +2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3 +[26] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 2, 5 +[27] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2023. 2, 3 +[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[29] Jie Xiao, Ruili Feng, Han Zhang, Zhiheng Liu, Zhantao Yang, Yurui Zhu, Xueyang Fu, Kai Zhu, Yu Liu, and Zheng-Jun Zha. Dreamclean: Restoring clean image using deep diffusion prior. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[30] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 2, 3 +[31] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 3, 4 +[32] Chang-Han Yeh, Chin-Yang Lin, Zhixiang Wang, Chi-Wei Hsiao, Ting-Hsuan Chen, and Yu-Lun Liu. Diffir2vr-zero: Zero-shot video restoration with diffusion-based image restoration models. arXiv preprint arXiv:2407.01519, 2024. 2, 3, 5, 6, 7 +[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 \ No newline at end of file diff --git a/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/images.zip b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1234631596404c991fa36a0af1eaf3dd65751e99 --- /dev/null +++ b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b259f2ed0d089989d579626845c624f0479ec9cd5aa059e39af02bc367bf9b80 +size 734293 diff --git a/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/layout.json b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..786e0d97bbd57a225c1cfd486ee2d12e5ecdbf8d --- /dev/null +++ b/2025/VISION-XL_ High Definition Video Inverse Problem Solver using Latent Image Diffusion Models/layout.json @@ -0,0 +1,8658 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "type": "text", + "content": "VISION-XL: High Definition Video Inverse Problem Solver using Latent Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 162, + 269, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 162, + 269, + 188 + ], + "spans": [ + { + "bbox": [ + 194, + 162, + 269, + 188 + ], + "type": "text", + "content": "Taesung Kwon KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 174, + 192, + 289, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 192, + 289, + 202 + ], + "spans": [ + { + "bbox": [ + 174, + 192, + 289, + 202 + ], + "type": "text", + "content": "star.kwon@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 349, + 162, + 416, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 162, + 416, + 188 + ], + "spans": [ + { + "bbox": [ + 349, + 162, + 416, + 188 + ], + "type": "text", + "content": "Jong Chul Ye KAIST" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 330, + 192, + 433, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 192, + 433, + 203 + ], + "spans": [ + { + "bbox": [ + 330, + 192, + 433, + 203 + ], + "type": "text", + "content": "jong.ye@kaist.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 222, + 156, + 328 + ], + "blocks": [ + { + "bbox": [ + 59, + 222, + 156, + 328 + ], + "lines": [ + { + "bbox": [ + 59, + 222, + 156, + 328 + ], + "spans": [ + { + "bbox": [ + 59, + 222, + 156, + 328 + ], + "type": "image", + "image_path": "4947b8efca9a41483d9f71670b69062e99f2294903e7bffe05d68a5d4a7b6fe1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 158, + 222, + 355, + 333 + ], + "blocks": [ + { + "bbox": [ + 158, + 222, + 355, + 333 + ], + "lines": [ + { + "bbox": [ + 158, + 222, + 355, + 333 + ], + "spans": [ + { + "bbox": [ + 158, + 222, + 355, + 333 + ], + "type": "image", + "image_path": "78fcfa979d5a60d395e874e72e311c5410babb0219ba2b2f7445241fc337c2e7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 222, + 553, + 333 + ], + "blocks": [ + { + "bbox": [ + 356, + 222, + 553, + 333 + ], + "lines": [ + { + "bbox": [ + 356, + 222, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 356, + 222, + 553, + 333 + ], + "type": "image", + "image_path": "d4ad9f802ad006a053e6cd5238ca3d4b21170db5e6c4a1181f70a41ac55b99cd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 334, + 366, + 397 + ], + "blocks": [ + { + "bbox": [ + 58, + 334, + 366, + 397 + ], + "lines": [ + { + "bbox": [ + 58, + 334, + 366, + 397 + ], + "spans": [ + { + "bbox": [ + 58, + 334, + 366, + 397 + ], + "type": "image", + "image_path": "25eb6fc1b7c688902cfb16896f58fb94ff38248e61afa69997a398367ba0678b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 367, + 334, + 553, + 397 + ], + "blocks": [ + { + "bbox": [ + 367, + 334, + 553, + 397 + ], + "lines": [ + { + "bbox": [ + 367, + 334, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 367, + 334, + 553, + 397 + ], + "type": "image", + "image_path": "df863e580e5b2d7e34f66403acdbb80adaab488c926ee9b4c09a4ace5b14e32c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 57, + 397, + 366, + 460 + ], + "blocks": [ + { + "bbox": [ + 57, + 397, + 366, + 460 + ], + "lines": [ + { + "bbox": [ + 57, + 397, + 366, + 460 + ], + "spans": [ + { + "bbox": [ + 57, + 397, + 366, + 460 + ], + "type": "image", + "image_path": "2dd0fdd5b8bb0736056be69ea844372f7ef9c3800d4474874e0a28182658e281.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "text", + "content": "Figure 1. Representative video reconstruction by VISION-XL: " + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathrm{SR + }" + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "text", + "content": " (frame averaging with " + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "text", + "content": " super-resolution, top), Deblur+ (frame averaging with deblurring, " + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\sigma = 3.0" + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "text", + "content": ", bottom-left), and Inpaint+ (frame averaging with " + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 464, + 555, + 487 + ], + "type": "text", + "content": " random inpainting, bottom-right)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 367, + 397, + 553, + 460 + ], + "blocks": [ + { + "bbox": [ + 367, + 397, + 553, + 460 + ], + "lines": [ + { + "bbox": [ + 367, + 397, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 367, + 397, + 553, + 460 + ], + "type": "image", + "image_path": "87fd66dd26cfa1170791243348284b967be8a511e4925c853c23727c8c3bc4b4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 152, + 497, + 200, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 497, + 200, + 509 + ], + "spans": [ + { + "bbox": [ + 152, + 497, + 200, + 509 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 522, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 522, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 522, + 297, + 714 + ], + "type": "text", + "content": "In this paper, we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models. Building on recent advancements in spatio-temporal optimization for video inverse problems using image diffusion models, our approach leverages latent-space diffusion models to achieve enhanced video quality and resolution. To address the high computational demands of processing high-resolution frames, we introduce a pseudo-batch consistent sampling strategy, allowing efficient operation on a single GPU. Additionally, to improve temporal consistency, we present pseudo-batch inversion, an initialization technique that incorporates informative latents from the measurement. By integrating with SDXL, our framework achieves state-of-the-art video reconstruction across a wide range of spatio-temporal inverse problems, including complex combinations of frame averaging" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "content": "ing and various spatial degradations, such as deblurring, super-resolution, and inpainting. Unlike previous methods, our approach supports multiple aspect ratios (landscape, vertical, and square) and delivers HD-resolution reconstructions (exceeding " + }, + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "content": ") in under 6 seconds per frame on a single NVIDIA 4090 GPU. Project page: https://vision-x1.github.io/." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 395, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 616, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 616, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 312, + 616, + 555, + 701 + ], + "type": "text", + "content": "Diffusion models [6, 9, 10, 16, 18, 20, 22] have set a new benchmark in generative modeling, enabling the generation of high-quality samples. These models have become the foundation for advancements in various fields, such as controllable image editing [33], image personalization [8], synthetic data augmentation [23], and even reconstructing images from brain signals [14, 24]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Furthermore, diffusion model-based inverse problem" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10465" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 293, + 109 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 293, + 109 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 293, + 109 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 293, + 109 + ], + "type": "table", + "html": "
MethodsSpatio-temporal degradationLatent space
Optical flow-based [5, 32]X✓
SVI [13]✓X
Ours✓✓
", + "image_path": "5a52b3ed258d1244730990fda8dfe133323d4684b10e2549b8858ff7ce07004f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 163, + 295, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 163, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 295, + 246 + ], + "type": "text", + "content": "solvers (DIS) [2, 4, 11, 21, 27, 29] address a variety of image restoration tasks, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on. A key feature of DIS is its plug-and-play capability, allowing diffusion models to be applied flexibly across different inverse problems without requiring task-specific training or fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 247, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 295, + 342 + ], + "type": "text", + "content": "Recently, several extensions [5, 13, 32] from the DIS have been proposed to solve video inverse problems using image diffusion models. Naive application of image diffusion models to videos may break temporal consistency. To address this problem, these methods preserve temporal consistency by utilizing a batch-consistent sampling strategy [13] and applying optical flow guidance to warp either latent representations [32] or the noise prior [5]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 343, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 582 + ], + "type": "text", + "content": "Although these innovative approaches enable powerful image generative models [6, 16, 18] to solve video inverse problems with significantly reduced computational requirements, there is still room for improvement in these methods. Optical flow-based methods [5, 32] have reported a key limitation: their performance is highly dependent on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes problematic when extreme degradations complicate the estimation process, restricting their applicability to a wider range of restoration tasks. Additionally, these methods require task-specific restoration modules [32] or fine-tuning of the diffusion model [5]. In this perspective, batch consistent sampling strategy [13] successfully addressed various spatio-temporal degradations without requiring task-specific training or fine-tuning. However, we empirically found that extending SVI [13] to latent diffusion models leads to unsatisfactory reconstruction, particularly in terms of FVD [26], as shown in Table 2, despite the fact that most modern latent diffusion models are essential for scaling up to large-size video inverse problems." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "To overcome these limitations, here we propose a novel framework for solving high-definition video inverse problems using latent image diffusion models that involves no additional training. To address the high computational demands of batch processing (e.g., 16-frame batch processing used in [13]) with high-resolution latent diffusion models [16], we introduce pseudo-batch consistent sampling. This strategy enables multi-frame video processing while requiring only the memory needed for a single frame, making it feasible on a single GPU. Furthermore, we propose pseudo-batch inversion, which initializes the process with" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 70, + 539, + 118 + ], + "blocks": [ + { + "bbox": [ + 55, + 112, + 295, + 156 + ], + "lines": [ + { + "bbox": [ + 55, + 112, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 295, + 156 + ], + "type": "text", + "content": "Table 1. Comparison of image diffusion-based video inverse problem solvers. Unlike other methods, our approach leverages latent image diffusion models to address spatio-temporal degradation, enabling more effective restoration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 70, + 539, + 118 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 539, + 118 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 539, + 118 + ], + "type": "table", + "html": "
InitializationFVD ↓LPIPS ↓PSNR ↑
Random noise10470.25129.43
Batch synchronized noise (SVI [13])707.70.24830.10
Pseudo-batch inversion (Ours)184.80.23630.74
", + "image_path": "66e6936327aaba5fafeafd638112e60a1cd98b765a8b2152e23c89aea453cdff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 553, + 154 + ], + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 154 + ], + "type": "text", + "content": "Table 2. Impact of our initialization method on SR+ video restoration using SDXL [16]. Our method significantly improves performance, reducing FVD [26] by more than " + }, + { + "bbox": [ + 313, + 121, + 553, + 154 + ], + "type": "inline_equation", + "content": "3\\times" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 163, + 553, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 210 + ], + "type": "text", + "content": "informative latents derived from the measurement. This initialization enhances temporal consistency and improves the efficiency of solving spatio-temporal inverse problems as shown in Table. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 212, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 342 + ], + "type": "text", + "content": "By integrating these components, our framework achieves state-of-the-art video reconstruction performance using SDXL [16]. We name the method integrating these components as VISION-XL, short for Video Inverse-problem Solver using latent diffusION models (with stable diffusion XL). It supports various aspect ratios, including landscape, vertical, and square formats. Thanks to its efficiency, our framework can reconstruct " + }, + { + "bbox": [ + 313, + 212, + 555, + 342 + ], + "type": "inline_equation", + "content": "1280 \\times 768" + }, + { + "bbox": [ + 313, + 212, + 555, + 342 + ], + "type": "text", + "content": " (exceeding HD resolution) videos in under 6 seconds per frame on a single NVIDIA 4090 GPU. Our contribution can be summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 344, + 553, + 415 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 344, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 553, + 378 + ], + "type": "text", + "content": "- We propose a high-definition video inverse problem solver integrated with SDXL, supporting multiple aspect ratios and achieving state-of-the-art reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 415 + ], + "type": "text", + "content": "- We introduce a novel pseudo-batch consistent sampling and inversion strategy for efficient and effective video reconstruction across diverse inverse problems." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 429, + 400, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 400, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 400, + 441 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "text", + "content": "Diffusion model-based inverse problem solvers (DIS). Diffusion models [9, 20, 22] attempt to model the data distribution " + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "text", + "content": " based on the Gaussian transitions. In the geometric view of diffusion models [1], the transitions are typically described as iterative manifold transitions " + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t \\to \\mathcal{M}_{t-1}, t = T, \\dots, 1" + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "text", + "content": ", moving from the noisy manifold " + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_T" + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "text", + "content": " to the clean manifold " + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 449, + 553, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Diffusion model-based inverse problem solvers (DIS) [2, 4, 11, 21, 27] aim to guide manifold transitions to sample from the posterior distribution " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}|\\boldsymbol{y})" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", which represents sampling " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " from the measurement " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " obtained from the forward model " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". In Bayesian inference, the posterior distribution, " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}|\\boldsymbol{y}) \\propto p_{\\theta}(\\boldsymbol{x})p(\\boldsymbol{y}|\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " is decomposed into the likelihood " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{y}|\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", representing the probability of observing " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", and the prior data distribution " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". This decomposition enables posterior sampling by combining diffusion sampling with iterative guidance using the forward model " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " and measurement " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". This approach provides sophisticated, precise solutions to complex inverse problems, leveraging the power and flexibility of diffusion models in practical applications, such as deblurring, super-resolution, inpainting, colorization, compressed sensing, and so on." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10466" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "DIS using latent diffusion models (LDIS). Most DIS [2, 4, 11, 21, 27] use pixel-space diffusion models, which facilitate easy integration of the forward model " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " and measurement " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": ", as both are defined in pixel space. Integrating forward models in latent space presents more challenges. Latent-space methods [3, 12, 19] calculate data consistency terms after decoding the denoised latent representation, then update these guidances within the latent space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "text", + "content": "During this process, VAE mapping errors accumulate in iterative sampling, causing the representation to drift from the clean manifold " + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "text", + "content": ". Additionally, most latent diffusion models provide a text-conditioned prior distribution " + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}|c_{\\mathrm{text}})" + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "text", + "content": ", which is challenging to implement in cases where text conditioning (" + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "text", + "content": ") is unavailable. As a result, latent-space methods prioritize two main goals: (i) managing text embeddings effectively and (ii) preserving the updated latent close to the clean manifold " + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 174, + 294, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 294, + 361 + ], + "type": "text", + "content": "For text embeddings, PSLD [19] uses only null-text, while TReg [12] and P2L [3] apply either null-text optimization or text optimization to enhance the reconstruction performance. To maintain the updated latent representation quality, the regularization term for aligning pixel and latent spaces is used to enforce latent feasibility [3, 12, 19]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 294, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 294, + 474 + ], + "type": "text", + "content": "Solving video inverse problems using DIS. Recently, several extensions [5, 13, 32] from DIS have been introduced to address video inverse problems. A straightforward application of image diffusion models to video, processing frames individually, often disrupts temporal consistency. These approaches maintain temporal coherence by employing a batch-consistent sampling strategy [13] and leveraging optical flow guidance to warp either latent representations [32] or the noise prior [5]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 294, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 294, + 624 + ], + "type": "text", + "content": "While these innovative approaches allow powerful image generative models [6, 16, 18] to address video inverse problems with reduced computational demands, there is still room for improvement. A key limitation of optical flow-based methods [5, 32] is their heavy reliance on the accuracy of the optical flow estimation module [25, 30]. This dependency becomes particularly problematic in scenarios involving severe degradations, which can hinder the estimation process and limit the methods' applicability to broader restoration tasks. Moreover, these approaches often require task-specific restoration modules [32] or fine-tuning of the diffusion model [5]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 714 + ], + "type": "text", + "content": "In contrast, the batch-consistent sampling strategy [13] has demonstrated effectiveness in addressing various spatiotemporal degradations without the need for task-specific training or model fine-tuning. However, this method utilizes the unconditional pixel-space diffusion model provided by ADM [6], and its extension to latent diffusion models remains unexplored." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 71, + 553, + 97 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 553, + 97 + ], + "type": "text", + "content": "3. High Definition Video Inverse Solver Using Latent Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 105, + 553, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 553, + 152 + ], + "type": "text", + "content": "This section introduces a novel approach for reconstructing high-definition videos that include various spatio-temporal degradations. The overall pipeline of the algorithm is illustrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 153, + 553, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 553, + 176 + ], + "type": "text", + "content": "Consider the spatio-temporal degradation process is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 184, + 553, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 184, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 356, + 184, + 553, + 198 + ], + "type": "interline_equation", + "content": "\\boldsymbol {Y} = \\mathcal {A} (\\boldsymbol {X}) = \\mathcal {A} ([ \\boldsymbol {X} [ 1 ], \\dots , \\boldsymbol {X} [ N ]) ] \\tag {1}", + "image_path": "4dd67fd285b52e32d95c74977d806a051168c654ce4586f84aefee8792d4ea08.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": " denotes the measurement, " + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "inline_equation", + "content": "X[n]" + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": "-th frame ground-truth frame, " + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": " is the number of video frames, and " + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": " refers to the operator describing the spatio-temporal degradation process. The detailed spatio-temporal degradation operator " + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 205, + 553, + 264 + ], + "type": "text", + "content": " is described in Sec. 4.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": "Our approach begins by inverting the measurement frames, denoted as " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\bar{Y}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": ", to initialize the informative latents " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "z_{\\tau}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": ", enhancing batch-wise consistency (Step 1). Next, we construct the corresponding denoised batch " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\tau}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": " by sampling each latent sequentially using Tweedie's formula [7], followed by decoding (Step 2). In Step 3, the corresponding denoised batch " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\tau}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": " is further refined by applying " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": "-step conjugate gradient (CG) optimization [4, 13] to enforce the data consistency from spatio-temporal degradation " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": ". In Step 4, we apply a scheduled low-pass filter to the updated batch " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}_{\\tau}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": " inspired by the frequency-based analysis of spectral diffusion [31]. " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}_{\\tau}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": " is then re-encoded into latent space to form " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\bar{z}_{\\tau}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": ". Finally, we obtain the one-step denoised latents " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "z_{\\tau - 1}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": " by adding noise to the encoded latents " + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\bar{z}_{\\tau}" + }, + { + "bbox": [ + 313, + 266, + 555, + 456 + ], + "type": "text", + "content": " (Step 5). In the following, we provide a detailed description of each of these steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "text", + "content": "Step 1: Initialize informative latents. One of our key insights is to initialize the informative latents by inverting the measurement frames, thereby enhancing batch-wise consistent initialization. Although these latents cannot restore the ground-truth frames directly, inverted latent variables can inherit information from the measurement frame, providing good initializations [29]. Different from SVI [13], which initializes sampling from a batch-wise synchronized uninformative Gaussian prior " + }, + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "inline_equation", + "content": "z_{T} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "text", + "content": " as the initial sampling point, we replace " + }, + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "text", + "content": " with the informative prior " + }, + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "inline_equation", + "content": "z_{\\tau}" + }, + { + "bbox": [ + 313, + 456, + 554, + 587 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 595, + 553, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 595, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 356, + 595, + 553, + 609 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {0} = \\boldsymbol {E} _ {\\theta} (\\boldsymbol {Y}), \\quad \\boldsymbol {z} _ {\\tau} = \\mathrm {D D I M} ^ {- 1} \\left(\\boldsymbol {z} _ {0}\\right), \\tag {2}", + "image_path": "97f14cdecca4b36b0e7f5419e74749b2a7f75f9b185816e6da08b95f61d2f755.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "inline_equation", + "content": "E_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{DDIM}^{-1}(\\cdot)" + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "text", + "content": " denote frame-wise encoding from pretrained VAE and DDIM inversion of timestep " + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Step 2: Pseudo-batch sampling. After initialization, we guide the sampling path to ensure the data consistency condition. At timestep " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "0 < t \\leq \\tau" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", we sample denoise batch " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " from given latents " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "z_t \\coloneqq [z_t[1] \\cdots z_t[N]]" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " by using Tweedie's formula [7]. Unlike SVI [13], we split latent" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10467" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 202 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 202 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 202 + ], + "type": "image", + "image_path": "19e9d2e13fc3ea1b81ae162bd267bdf3b7536d1ccc29225bf614217e236133f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "text", + "content": "Figure 2. Illustration of VISION-XL sampling at timestep " + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "text", + "content": " is split into individual frames and denoised sequentially using Tweedie's formula. The denoised latents " + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t}" + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "text", + "content": " are then merged and decoded. The decoded batch " + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_t" + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "text", + "content": " is optimized to enforce the data consistency, followed by low-pass filtered encoding and re-noising to obtain " + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 246, + 294, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 246, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 294, + 328 + ], + "type": "text", + "content": "frames to construct a pseudo-batch and sequentially sample each frame, requiring memory for only a single frame during the sampling, as shown in Fig. 2. Similarly, inversion is also conducted within the pseudo-batch framework. This enables the recent advanced latent diffusion model to operate in this framework without a frame limit. As a proof-of-concept, we conduct experiments on 25-frame videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 294, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 294, + 354 + ], + "type": "text", + "content": "Specifically, consider a sequential sampling of latent diffusion models " + }, + { + "bbox": [ + 55, + 329, + 294, + 354 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 329, + 294, + 354 + ], + "type": "text", + "content": " along the temporal direction:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 359, + 295, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 359, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 92, + 359, + 295, + 377 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {E}} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t}) := \\left[ \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t} [ 1 ]) \\dots \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t} [ N ]) \\right]. \\tag {3}", + "image_path": "614a1c3ef96887f31f7b801eafec35c4da14b90117648a70a4060d22c7ef1eb6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "type": "text", + "content": "The denoised latents " + }, + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "type": "inline_equation", + "content": "\\hat{Z}_t" + }, + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "type": "text", + "content": " are computed using Tweedie's formula [7]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 415, + 295, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 415, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 97, + 415, + 295, + 441 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {z}} _ {t} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\boldsymbol {z} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\mathcal {E}} _ {\\theta} ^ {(t)} (\\boldsymbol {z} _ {t})\\right), \\tag {4}", + "image_path": "7124228892015a1d14143a7e0294d474f279971fe313a48a1f9df648b340d6d7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "text", + "content": " is the noise schedule defined in the Gaussian process of diffusion models [9, 15]. Then denoised batch " + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\hat{X}_t" + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "text", + "content": " is decoded from the denoised latents " + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "text", + "content": " using VAE decoder " + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 55, + 445, + 296, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 81, + 500, + 295, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 500, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 81, + 500, + 295, + 516 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {X}} _ {t} = \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t}) := [ \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t} [ 1 ]) \\dots \\boldsymbol {D} _ {\\theta} (\\hat {\\boldsymbol {z}} _ {t} [ N ]) ]. \\quad (5)", + "image_path": "f20def75871b2b0ba33979986d0051281b0353ed4bd14ce207d3d521bd70bf31.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": "Step 3: Batch optimization in pixel-space. The denoised batch " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{X}}_t" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": " is then refined as a whole by applying the " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": "-step CG optimization to enhance the data consistency from the measurement " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\pmb{Y}" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": " and spatio-temporal degradation " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": ". This can be formally represented by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 588, + 295, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 588, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 107, + 588, + 295, + 612 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {X}} _ {t} := \\underset {\\boldsymbol {X} \\in \\hat {\\boldsymbol {X}} _ {t} + \\boldsymbol {\\mathcal {K}} _ {l}} {\\arg \\min } \\| \\boldsymbol {Y} - \\mathcal {A} (\\boldsymbol {X}) \\| ^ {2} \\tag {6}", + "image_path": "2991994ef639057aa0008723d7c2c85a47d24b126f4d0d0202936bad7b7ff01d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{K}_l" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "-dimensional Kyrlov subspace associated with the given inverse problem [4]. The multistep CG allows each temporal frame to be diversified, enhancing data consistency and achieving faster convergence without requiring memory-intensive gradient calculations [2]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Step 4: Low-pass filtered encoding. Recent frequency-based analyses of diffusion models [10, 31] suggest that optimal denoisers first recover low-frequency components in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "text", + "content": "early denoising stages, while high-frequency details added progressively in later stages. Building on these findings, we observed that applying a scheduled low-pass filter to the updated batch " + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "inline_equation", + "content": "\\bar{X}_t" + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "text", + "content": " in early stages effectively removes undesired artifact caused by VAE error accumulation, resulting in more natural and refined outputs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "content": "Based on the observation that the denoiser restores high-frequency details as the noise scale " + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\sqrt{1 - \\bar{\\alpha}_t}" + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "content": " decreases, we set the filter width " + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "content": " to be proportional to the noise scale [15], defined as " + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\sigma_t \\coloneqq \\lambda \\sqrt{1 - \\bar{\\alpha}_t}" + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "content": ", which goes to zero at " + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "inline_equation", + "content": "t \\to 0" + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "content": ". After applying the low-pass filter " + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "inline_equation", + "content": "h_{\\sigma_t}" + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "content": ", we re-encode " + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\bar{\\boldsymbol{X}}_t" + }, + { + "bbox": [ + 313, + 320, + 554, + 404 + ], + "type": "text", + "content": " into the latent space. Specifically, the re-encoded latents are given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 398, + 418, + 553, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 418, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 398, + 418, + 553, + 432 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {X}} _ {t} \\leftarrow \\bar {\\boldsymbol {X}} _ {t} * \\boldsymbol {h} _ {\\sigma_ {t}}, \\tag {7}", + "image_path": "ce90f8fe764f06922ba681ee45d41bd624e69f157acb788f328257d11c35321d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 434, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 434, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 331, + 434, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\bar {z} _ {t} = \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t}) := [ \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t} [ 1 ]) \\dots \\boldsymbol {E} _ {\\theta} (\\bar {\\boldsymbol {X}} _ {t} [ N ]) ], \\tag {8}", + "image_path": "6fad9017cc4f2337601aec10d33a45954eb79046735e3d01be00a1812638c850.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 461, + 463, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 463, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 463, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 461, + 463, + 473 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 313, + 461, + 463, + 473 + ], + "type": "text", + "content": " denotes the VAE encoder." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 476, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 554, + 499 + ], + "type": "text", + "content": "Step 5: Renoising. After encoding, updated latents " + }, + { + "bbox": [ + 313, + 476, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\bar{z}_t" + }, + { + "bbox": [ + 313, + 476, + 554, + 499 + ], + "type": "text", + "content": " are renoised as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 515, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 515, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 361, + 515, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t - 1} = \\sqrt {\\bar {\\alpha} _ {t - 1}} \\bar {\\boldsymbol {z}} _ {t} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\boldsymbol {\\mathcal {E}} _ {t}, \\tag {9}", + "image_path": "66f0a024163aa8c15cf8e2c56d019de255b071f30e155144960872a8323ebe61.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 543, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 543, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{\\varepsilon}_{t}" + }, + { + "bbox": [ + 313, + 543, + 554, + 567 + ], + "type": "text", + "content": " is composed of batch-consistent stochastic noise [13] and deterministic noise [20]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In the geometric view of diffusion models [1], the sampling path evolves as illustrated in Fig. 3. The initialized latent " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{\\tau}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " is projected onto the clean manifold " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " using Tweedie's formula. The projected latent is then decoded into the pixel space and refined through multi-step CG to satisfy the data consistency constraint " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "Y = \\mathcal{A}(X)" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". Next, a scheduled low-pass filter is applied to reduce VAE error accumulation and keep the encoding close to " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". Finally, the encoded latents are re-noised to transition back to " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\tau -1}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ", and this process iterates until the final state converges to the clean manifold " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". The complete algorithm is provided in Algorithm 1." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10468" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 62, + 272, + 252 + ], + "blocks": [ + { + "bbox": [ + 83, + 62, + 272, + 252 + ], + "lines": [ + { + "bbox": [ + 83, + 62, + 272, + 252 + ], + "spans": [ + { + "bbox": [ + 83, + 62, + 272, + 252 + ], + "type": "image", + "image_path": "914006bdc2f45db2ebb1f074074d053b35372efb9aabca0c78d018f2dc80a405.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "text", + "content": "Figure 3. Geometric illustration of the sampling path evolution. (Step 1) Initialize latent " + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{\\tau}" + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "text", + "content": ". (Step 2) Project onto " + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "text", + "content": " via pseudobatch sampling and decode to pixel space. (Step 3) Optimize for measurement consistency " + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "inline_equation", + "content": "\\pmb{Y} = \\mathcal{A}(\\pmb{X})" + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "text", + "content": ". (Step 4) Apply a scheduled low-pass filter and encode back to latent space. (Step 5) Renoise to " + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\tau -1}" + }, + { + "bbox": [ + 55, + 255, + 296, + 321 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 56, + 358, + 296, + 528 + ], + "blocks": [ + { + "bbox": [ + 56, + 332, + 296, + 356 + ], + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 356 + ], + "type": "text", + "content": "Algorithm 1 High-definition video inverse problem solver using latent diffusion models" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "lines": [ + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{E}}_{\\theta}^{(t)},\\pmb {E}_{\\theta},\\pmb {D}_{\\theta},\\pmb {Y},\\mathcal{A},\\tau ,l,\\sigma_{t},\\{\\bar{\\alpha}_{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "z_0\\gets E_\\theta (Y)" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "z_{\\tau}\\gets \\mathrm{DDIM}^{-1}(z_0)" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " ▷ Step 1 \n3: for " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "t = \\tau :2" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\hat{z}_t\\gets \\left(z_t - \\sqrt{1 - \\bar{\\alpha}_t}\\pmb {\\mathcal{E}}_\\theta^{(t)}(z_t)\\right) / \\sqrt{\\bar{\\alpha}_t}" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " ▷ Step 2 \n5: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{X}}_t\\gets \\pmb {D}_\\theta (\\hat{\\pmb{z}}_t)" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{X}}_t\\coloneqq \\arg \\min_{\\pmb {X}\\in \\hat{\\pmb{X}}_t + \\pmb {\\kappa}_l}\\| \\pmb {Y} - \\mathcal{A}(\\pmb {X})\\| ^2" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " ▷ Step 3 \n7: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{X}}_t\\gets \\bar{\\pmb{X}}_t*\\pmb {h}_{\\sigma_t}" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " ▷ Step 4 \n8: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\bar{z}_t = E_\\theta (\\bar{X}_t)" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "z_{t - 1} = \\sqrt{\\bar{\\alpha}_{t - 1}}\\bar{z}_t + \\sqrt{1 - \\bar{\\alpha}_{t - 1}}\\pmb {\\mathcal{E}}_t" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " ▷ Step 5 \n10: end for \n11: " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "z_0\\gets \\left(z_1 - \\sqrt{1 - \\bar{\\alpha}_1}\\pmb {\\mathcal{E}}_\\theta^{(1)}(z_1)\\right) / \\sqrt{\\bar{\\alpha}_1}" + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "text", + "content": " \n12: return " + }, + { + "bbox": [ + 56, + 358, + 296, + 528 + ], + "type": "inline_equation", + "content": "Z_{0}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 550, + 181, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 181, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 181, + 563 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 170, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 170, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 170, + 583 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 587, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 694 + ], + "type": "text", + "content": "Dataset. We used four high-resolution (with resolutions exceeding 1080p) video datasets for evaluation, sourced from the DAVIS dataset [17] and the Pexels dataset1. A subset of 100 videos from the DAVIS dataset is resized to " + }, + { + "bbox": [ + 55, + 587, + 296, + 694 + ], + "type": "inline_equation", + "content": "768 \\times 1280" + }, + { + "bbox": [ + 55, + 587, + 296, + 694 + ], + "type": "text", + "content": " resolution and consists of 25 frames, originally provided in landscape orientation. The Pexels dataset is a large, open-source collection of high-resolution stock videos and images, widely used for creative and research purposes. For the Pexels subset, we collect a total of 120 videos: 45" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "in landscape orientation (Pexels (landscape)), 45 in vertical orientation (Pexels (vertical)), and 30 in square orientation (Pexels (square)). These subsets are resized to resolutions of " + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "768 \\times 1280" + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": " for landscape, " + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "1280 \\times 768" + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": " for vertical, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": " for square orientations, with each video consisting of 25 frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": "Inverse problems. We test our method on the following spatial degradations: 1) Deblur: Gaussian deblurring from an image convolved with a " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "61 \\times 61" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " size Gaussian kernel with " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\sigma = 3.0" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": ", 2) SR: Super-resolution from " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " average pooling, 3) Inpaint: Inpainting from " + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 144, + 555, + 263 + ], + "type": "text", + "content": " random masking. Furthermore, test our method on the following spatio-temporal degradations: 4) Deblur+: Deblur + 7-frame averaging using temporal uniform blur kernel as used in [13], 5) SR+: SR + 7-frame averaging, and 6) Inpaint+: Inpaint + 7-frame averaging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 265, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 556, + 479 + ], + "type": "text", + "content": "Baseline comparison. The primary objective of this study is to improve the performance of video inverse problem solvers through latent image diffusion models. Thus, our evaluation primarily compares video inverse problem solvers using image diffusion models. As a recently emerging field, only a few methods are available: SVI [13], DiffIR2VR [32], and Warped Diffusion [5]. Notably, DiffIR2VR and Warped Diffusion cannot address spatiotemporal degradations, and DiffIR2VR only supports SR among the inverse problems we address in this paper. We conducted comparisons with SVI and DiffIR2VR but excluded Warped Diffusion, as it is not currently open-source. SVI officially supports a resolution of " + }, + { + "bbox": [ + 313, + 265, + 556, + 479 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 265, + 556, + 479 + ], + "type": "text", + "content": ", while DiffIR2VR supports " + }, + { + "bbox": [ + 313, + 265, + 556, + 479 + ], + "type": "inline_equation", + "content": "480 \\times 854" + }, + { + "bbox": [ + 313, + 265, + 556, + 479 + ], + "type": "text", + "content": ". To ensure fair comparisons with identical resolutions, we used patch reconstruction. Additionally, we included a comparison with the classical optimization method ADMM-TV, following the protocol established by SVI [13]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": "For quantitative comparison, we focus on two widely used standard metrics: peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) [28]. Additionally, we evaluate two perceptual metrics: Learned Perceptual Image Patch Similarity (LPIPS) [34] and Fréchet Video Distance (FVD) [26]. For computing the metrics, we follow the protocol from the open-source project2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "content": "Implementation details. While our method is applicable to general latent diffusion models, we use Stable Diffusion XL 1.0 (SDXL) [16]—the current state-of-the-art text-to-image diffusion model—as a proof of concept in this paper. For all experiments, we employ " + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "inline_equation", + "content": "T = 25" + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "inline_equation", + "content": "\\tau = 0.3T" + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "inline_equation", + "content": "\\lambda = 2" + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "inline_equation", + "content": "l = 10" + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "content": ". These optimal values are obtained through extensive ablation studies and the results are described in Sec. 4.3. To reduce undesired guidance from the text condition " + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "content": ", we use a null-text condition, " + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "inline_equation", + "content": "c_{\\emptyset}" + }, + { + "bbox": [ + 313, + 564, + 554, + 683 + ], + "type": "text", + "content": ". All experiments were done on a single NVIDIA 4090 GPU." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 182, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 182, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 182, + 712 + ], + "type": "text", + "content": "1https://www.pexels.com/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "2https://github.com/JunyaoHu/common_metricson Video_quality" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10469" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 61, + 553, + 286 + ], + "blocks": [ + { + "bbox": [ + 58, + 61, + 553, + 286 + ], + "lines": [ + { + "bbox": [ + 58, + 61, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 58, + 61, + 553, + 286 + ], + "type": "table", + "html": "
TaskMethodDAVISPexels (landscape)Pexels (vertical)Pexels (square)
FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑FVD↓LPIPS↓PSNR↑SSIM↑
Deblur+ADMM-TV15120.39724.300.742671.80.28429.420.818709.20.23730.370.847569.50.22430.680.856
SVI [13]638.90.32228.040.799830.50.26530.420.831656.20.23930.400.856499.40.22131.590.862
DiffIR2VR [32]----------------
Ours228.60.29228.760.807196.40.24931.110.839209.90.22431.870.860157.90.21732.400.864
SR+ADMM-TV14290.35924.230.740634.10.27929.010.820669.20.32229.710.836545.90.30630.080.838
SVI [13]223.40.23429.000.812386.80.26530.700.831558.90.26130.480.842313.20.26531.180.847
DiffIR2VR [32]----------------
Ours158.50.24429.180.818166.20.24630.820.832173.10.22931.380.847138.80.22031.900.856
Inpaint+ADMM-TV18480.33924.160.762797.80.29229.150.778805.40.25829.360.805652.40.26830.390.804
SVI [13]208.60.23829.600.848269.30.25029.920.826428.10.24630.270.838206.90.23831.120.847
DiffIR2VR [32]----------------
Ours241.10.24228.810.815222.40.21630.130.828240.40.20130.980.845164.90.23031.440.853
DeblurADMM-TV169.00.23230.490.873192.20.26331.670.842185.70.18132.190.861199.60.25730.980.875
SVI [13]99.500.17631.700.875153.80.21232.440.862160.20.17433.350.866116.60.20733.510.885
DiffIR2VR [32]----------------
Ours72.030.17131.940.88096.160.18432.830.87689.390.17033.550.88891.550.15734.100.896
SRADMM-TV285.60.22126.650.788301.50.22227.910.769298.70.21327.120.786209.40.25728.740.798
SVI [13]176.10.17629.280.815201.60.20231.000.836219.30.20030.590.847167.90.20631.830.851
DiffIR2VR [32]319.60.23826.290.738412.60.24427.180.742511.60.21426.910.752412.00.29627.760.741
Ours81.100.19030.270.848104.30.18531.840.85898.570.15832.540.87698.410.13833.020.886
InpaintADMM-TV212.00.31528.420.797270.90.31629.140.794270.30.21229.840.803264.70.31030.460.809
SVI [13]143.50.17730.200.858159.40.23330.090.827164.30.13231.190.847139.30.22531.570.855
DiffIR2VR [32]----------------
Ours143.60.20929.740.835158.70.20830.410.834174.00.19531.150.847125.40.21631.630.859
", + "image_path": "cbaf10e908d1663342b1250c5042a5bed893b8e53668bbed55d9b3c0f863554e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 555, + 335 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation (FVD, LPIPS, PSNR, SSIM) of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios (landscape, vertical, square). Bold denotes the best results and underline indicates the runner-up. Notably, DiffIR2VR [32] is only capable of restoring SR among our experimental tasks, highlighting the broader task generalizability of our approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 345, + 555, + 578 + ], + "blocks": [ + { + "bbox": [ + 56, + 345, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 56, + 345, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 345, + 555, + 578 + ], + "type": "image", + "image_path": "ebc344f22dcbc44e9bc03683f2a303babc26e2833978cca0bfe83af7a792ec48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 582, + 555, + 616 + ], + "lines": [ + { + "bbox": [ + 55, + 582, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 555, + 616 + ], + "type": "text", + "content": "Figure 4. Qualitative evaluation of solving spatio-temporal inverse problems across DAVIS, Pexels dataset with multiple aspect ratios. Notably, ADMM-TV fails to remove ghosting artifacts caused by temporal degradation (red arrows), while SVI produces excessive intensity fluctuations (red box) or blurred information restoration (green and blue boxes)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 627, + 113, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 113, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 113, + 639 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "content": "Table 3 presents a quantitative comparison across various spatio-temporal inverse problems. The proposed method consistently outperforms baseline approaches in most metrics, particularly in addressing spatio-temporal degradations. Notably, it achieves a significant reduction in FVD," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 628, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 713 + ], + "type": "text", + "content": "indicating superior perceptual video quality compared to the runner-up across all datasets. This improvement is also evident in the qualitative results shown in Fig. 4. While SVI [13] effectively handles spatio-temporal degradations, it often struggles with temporal consistency, leading to issues such as inaccurate intensity restoration (first row) and loss of details (second and third rows). This suggests" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10470" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 64, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 69, + 64, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 69, + 64, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 69, + 64, + 542, + 217 + ], + "type": "image", + "image_path": "1b7dbc97f9b4c9aace0c561ad3cdb3e609f28b5a2d36dd59e1e711e0e0c446d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 253 + ], + "type": "text", + "content": "Figure 5. Qualitative evaluation of SR " + }, + { + "bbox": [ + 55, + 220, + 555, + 253 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 55, + 220, + 555, + 253 + ], + "type": "text", + "content": " performance across multiple aspect ratios (landscape, vertical). DiffIR2VR often produces unwanted artifacts in the background (red and blue boxes), while SVI inaccurately restores intensity (red box), leading to frame-wise fluctuations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 258, + 269, + 297 + ], + "blocks": [ + { + "bbox": [ + 80, + 258, + 269, + 297 + ], + "lines": [ + { + "bbox": [ + 80, + 258, + 269, + 297 + ], + "spans": [ + { + "bbox": [ + 80, + 258, + 269, + 297 + ], + "type": "table", + "html": "
MethodTime [min]Time per frame [sec]Memory (GB)
SVI [13]153618.5
DiffIR2VR [32]4.71113.6
Ours2.5612.7
", + "image_path": "ce7479669ed8273676975d9b9159ebd51ed54ae6af560ef30751631bc57694d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 353, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 295, + 567 + ], + "type": "text", + "content": "that batch-consistent noise initialization [13], further examined in our ablation study (Table 5), may be insufficient to fully preserve temporal consistency. Additionally, its patch reconstruction for high-resolution videos can introduce patch-wise inconsistencies, degrading overall performance. The classical optimization method, ADMM-TV, effectively reconstructs static backgrounds and stationary objects but fails to remove ghosting artifacts caused by temporal deformations, as shown in Fig. 4. This limitation is reflected in its lower performance metrics. For the SR task (Fig.5), DiffIR2VR[32] often introduces unwanted artifacts in backgrounds or over-generates object details, likely due to inaccuracies in optical flow estimation. Notably, in pixel-wise random inpainting, latent diffusion methods may lose fine pixel-level details due to their encoded representations. However, leveraging the strong SDXL prior, our method achieves competitive inpainting performance with SVI, which employs a pixel-space diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 569, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 653 + ], + "type": "text", + "content": "While baseline methods encounter various challenges across different inverse problems, our approach demonstrates stable, high-quality reconstructions, as evidenced by the overall results. Further comparisons of total sampling time and memory consumption are shown in Table 4, where our method achieves the highest efficiency in both sampling time and memory usage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Further experimental results on real-world denoising and blind video deconvolution are available in the supplementary material. Additional visualizations, including reconstruction results for deblurring, inpainting, and other tasks, are provided in video format for further evaluation: https:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 261, + 476, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 261, + 476, + 272 + ], + "spans": [ + { + "bbox": [ + 315, + 261, + 476, + 272 + ], + "type": "text", + "content": "//vision-x1.github.io/supple/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 278, + 412, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 278, + 412, + 290 + ], + "spans": [ + { + "bbox": [ + 314, + 278, + 412, + 290 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "type": "text", + "content": "In this section, we analyze the key components of our method. To highlight their impact, we conduct an ablation study on the " + }, + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathrm{SR + }" + }, + { + "bbox": [ + 313, + 297, + 555, + 344 + ], + "type": "text", + "content": " task in the Pexels (landscape) dataset, which involves significant spatio-temporal degradation." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 353, + 553, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 296, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 344 + ], + "type": "text", + "content": "Table 4. Comparison of total sampling time and memory efficiency for solving video inverse problems on a single 25-frame video at " + }, + { + "bbox": [ + 55, + 300, + 296, + 344 + ], + "type": "inline_equation", + "content": "768 \\times 1280" + }, + { + "bbox": [ + 55, + 300, + 296, + 344 + ], + "type": "text", + "content": " resolution. **Bold** denotes the best results and **underline** indicates the runner-up." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 353, + 553, + 404 + ], + "lines": [ + { + "bbox": [ + 317, + 353, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 353, + 553, + 404 + ], + "type": "table", + "html": "
InitializationInitialization Time [sec]Total Time [min]FVD↓PSNR↑
Random noise08.3104729.43
Batch-consistent noise [13]08.3707.730.10
Pseudo-batch inversion (τ: 0.15T)151.3229.530.00
Pseudo-batch inversion (τ: 0.30T)302.5184.830.74
Pseudo-batch inversion (τ: 0.45T)453.8288.730.70
", + "image_path": "3d57d314c399f06e02b82fd1a5ec212f620fe482903edcb2f2e23a71119ea3a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 413, + 554, + 434 + ], + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 434 + ], + "type": "text", + "content": "Table 5. Comparison of initialization strategies by computational cost and video quality for SR+ on the Pexels dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "type": "text", + "content": "Effect of initialization (in Step 1). In Table 5, we conduct an ablation study to see the effect of pseudo-batch inversion for initialization. From the table, we confirm that pseudobatch inversion effectively extracts informative latents to reconstruct video evidenced by about 0.6dB and 1.3dB PSNR increase compared with batch-synchronized noise initialization used in SVI [13] and random noise initialization, respectively. Notably, pseudo-batch inversion achieves " + }, + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "type": "text", + "content": " lower FVD compared to the batch-synchronized noise initialization which indicates a significant improvement in temporal consistency. It is also evident in the visualization of the ablation study shown in Fig. 6. The reconstruction results from random noise and batch-synchronized noise initialization fail to reconstruct the color of the cloud and are temporally inconsistent. In contrast, our method successfully reconstructs the color of the cloud and temporally consistent results. Furthermore, as shown in Table 5, pseudobatch inversion significantly improves sampling time efficiency. This is because it reduces the total sampling steps, and inversion does not require the measurement update process, which involves encoding and decoding. As a result, our method reconstructs high-definition video in under 6 seconds per frame on a single NVIDIA 4090 GPU." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10471" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 286, + 281 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 286, + 281 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 286, + 281 + ], + "type": "image", + "image_path": "51a94a435cc7029a6530a15131f8518c88ff159de6b77e49c44ebc502ec8d0ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 286, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 310 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of initialization strategies. Our approach improves temporal consistency and precise restoration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 100, + 319, + 253, + 369 + ], + "blocks": [ + { + "bbox": [ + 100, + 319, + 253, + 369 + ], + "lines": [ + { + "bbox": [ + 100, + 319, + 253, + 369 + ], + "spans": [ + { + "bbox": [ + 100, + 319, + 253, + 369 + ], + "type": "table", + "html": "
Update step lFVD↓LPIPS↓PSNR↑SSIM↑
111500.28127.550.799
5241.00.19730.690.839
10184.80.23630.740.826
20486.10.47028.160.690
", + "image_path": "fa0477c48cac455c1b06c988af625b812a7d8808180d83f05b7844d8889612ae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "content": "Effect of optimization step " + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "content": " (in Step 3). In Table 6, we present an ablation study on the effect of the CG update step " + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "content": ". The table confirms that the CG update is essential for enhancing data consistency. We found that at least 5 iterations of CG updates yield satisfactory results, and 10 iterations produce the best results." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 106, + 496, + 246, + 559 + ], + "blocks": [ + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "type": "text", + "content": "Table 6. Ablation study on the effect of optimization step " + }, + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "type": "text", + "content": ". Ten update steps yield the best temporal consistency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 496, + 246, + 559 + ], + "lines": [ + { + "bbox": [ + 106, + 496, + 246, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 246, + 559 + ], + "type": "table", + "html": "
LPF λFVD↓LPIPS↓PSNR↑SSIM↑
No LPF209.70.27329.920.797
1186.30.23930.590.819
√2184.80.23630.740.826
2179.30.24530.810.832
2√2191.40.26230.890.837
", + "image_path": "327bfc4a0bb7904887188ede6f615ffbe3e0926844a529c72d4b6ede32750028.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 295, + 589 + ], + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 589 + ], + "type": "text", + "content": "Table 7. Ablation study on the effect of low-pass filtering parameter " + }, + { + "bbox": [ + 55, + 567, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 567, + 295, + 589 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 55, + 567, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\lambda = 2" + }, + { + "bbox": [ + 55, + 567, + 295, + 589 + ], + "type": "text", + "content": " yields the best temporal consistency." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Effect of LPF " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " (in Step 4). Table 7 presents an ablation study on the effect of low-pass filtering. The results confirm that low-pass filtering enhances the reconstruction quality as evidenced by all metrics. Specifically, low-pass filtering results in approximately a 30-point decrease in FVD and a 1.0dB increase in PSNR compared to the absence of low-pass filtering. This improvement is also evident in the visualizations in Fig. 7. In the second row of the figure, undesired artifacts appear when low-pass filtering is not applied. In contrast, as shown in the third and fourth rows," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 73, + 542, + 281 + ], + "blocks": [ + { + "bbox": [ + 329, + 73, + 542, + 281 + ], + "lines": [ + { + "bbox": [ + 329, + 73, + 542, + 281 + ], + "spans": [ + { + "bbox": [ + 329, + 73, + 542, + 281 + ], + "type": "image", + "image_path": "685af7bf18c39cab7723baa4af91bf19e62d4c7509ea244827039cdbf1231341.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of low-pass filtering parameter " + }, + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\lambda = 2" + }, + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "type": "text", + "content": " effectively removes artifacts from VAE errors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 323, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 384 + ], + "type": "text", + "content": "these artifacts are effectively removed as the parameter " + }, + { + "bbox": [ + 313, + 323, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 323, + 555, + 384 + ], + "type": "text", + "content": " increases. From a frequency-based perspective, we believe that low-pass filtering effectively guides the updated latents to remain within the desired denoised manifold, " + }, + { + "bbox": [ + 313, + 323, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 323, + 555, + 384 + ], + "type": "text", + "content": ", and helps to mitigate error accumulation from the VAE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 394, + 388, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 388, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 388, + 407 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we proposed a novel framework for addressing high-definition video inverse problems using latent diffusion models that introduce two new strategies. First, a pseudo-batch consistent sampling strategy to manage intensive batch memory consumption with advanced latent diffusion models (e.g., SDXL). To acquire a denoised batch, we conduct sequential sampling of each latents rather than batch sampling to efficiently manage the high memory consumption of advanced latent diffusion models. Second, a pseudo-batch inversion for leveraging informative latent as initialization is proposed. We confirmed that pseudo-batch inversion significantly improves reconstruction performance in both traditional and perceptual quality metrics. Leveraging the powerful SDXL, our method achieves state-of-the-art performance across diverse spatio-temporal inverse problems, including challenging tasks such as the combination of frame averaging with deblurring, super-resolution, and inpainting. Importantly, our method supports multiple aspect ratios (landscape, vertical, and square), making it versatile for different video formats and delivering HD reconstructions in under 6 seconds per frame on a single NVIDIA 4090 GPU. Overall, our framework not only enhances video reconstruction quality but also sets new standards for efficiency and flexibility in solving high-definition video inverse problems." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10472" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 177 + ], + "type": "text", + "content": "This work was supported by the National Research Foundation of Korea under Grant RS-2024-00336454, and by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. RS-2024-00457882, AI Research Hub Project; No. RS-2025-02304967, AI Star Fellowship (KAIST))" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 209, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 253 + ], + "type": "text", + "content": "[1] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 35:25683-25696, 2022. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 255, + 296, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 296, + 310 + ], + "type": "text", + "content": "[2] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In International Conference on Learning Representations, 2023. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 312, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 312, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 61, + 312, + 296, + 356 + ], + "type": "text", + "content": "[3] Hyungjin Chung, Jong Chul Ye, Peyman Milanfar, and Mauricio Delbracio. Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 358, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 358, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 61, + 358, + 295, + 403 + ], + "type": "text", + "content": "[4] Hyungjin Chung, Suhyeon Lee, and Jong Chul Ye. Decomposed diffusion sampler for accelerating large-scale inverse problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 296, + 459 + ], + "type": "text", + "content": "[5] Giannis Daras, Weili Nie, Karsten Kreis, Alex Dimakis, Morteza Mardani, Nikola Borislavov Kovachki, and Arash Vahdat. Warped diffusion: Solving video inverse problems with image diffusion models. arXiv preprint arXiv:2410.16152, 2024. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 461, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 461, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 461, + 296, + 495 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 497, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 497, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 497, + 296, + 529 + ], + "type": "text", + "content": "[7] Bradley Efron. Tweedie's formula and selection bias. Journal of the American Statistical Association, 106(496):1602-1614, 2011. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "type": "text", + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 588, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 588, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 588, + 296, + 621 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "text", + "content": "[10] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[11] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. Advances in Neural Information Processing Systems, 35:23593-23606, 2022. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Jeongsol Kim, Geon Yeong Park, Hyungjin Chung, and Jong Chul Ye. Regularization by texts for latent diffusion inverse solvers. arXiv preprint arXiv:2311.15658, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[13] Taesung Kwon and Jong Chul Ye. Solving video inverse problems using image diffusion models. In The Thirteenth International Conference on Learning Representations, 2025. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[14] Dongyang Li, Chen Wei, Shiying Li, Jiachen Zou, and Quanying Liu. Visual decoding and reconstruction via eeg embeddings with guided diffusion. arXiv preprint arXiv:2403.07721, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[15] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "text", + "content": "[16] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[17] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "text", + "content": "[18] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[19] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[20] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1, 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "text", + "content": "[21] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[22] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[23] Andreas Stöckl. Evaluating a synthetic image dataset generated with stable diffusion. In International Congress on Information and Communication Technology, pages 805-818. Springer, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Yu Takagi and Shinji Nishimoto. High-resolution image reconstruction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14453-14463, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[25] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10473" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 520 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "text", + "content": "2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[26] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[27] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[29] Jie Xiao, Ruili Feng, Han Zhang, Zhiheng Liu, Zhantao Yang, Yurui Zhu, Xueyang Fu, Kai Zhu, Yu Liu, and Zheng-Jun Zha. Dreamclean: Restoring clean image using deep diffusion prior. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 340 + ], + "type": "text", + "content": "[30] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "text", + "content": "[31] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "text", + "content": "[32] Chang-Han Yeh, Chin-Yang Lin, Zhixiang Wang, Chi-Wei Hsiao, Ting-Hsuan Chen, and Yu-Lun Liu. Diffir2vr-zero: Zero-shot video restoration with diffusion-based image restoration models. arXiv preprint arXiv:2407.01519, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "text", + "content": "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 520 + ], + "type": "text", + "content": "[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10474" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_content_list.json b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c62690ff7b2cc9f66760a395964651be77167b2c --- /dev/null +++ b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_content_list.json @@ -0,0 +1,1819 @@ +[ + { + "type": "text", + "text": "VISO: Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference", + "text_level": 1, + "bbox": [ + 117, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meiqi Wang Tsinghua University wang-mq22@mails.tsinghua.edu.cn", + "bbox": [ + 228, + 203, + 508, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Han Qiu Tsinghua University qiuhan@tsinghua.edu.cn", + "bbox": [ + 563, + 203, + 763, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In-orbit object detection is essential for Earth observation missions on satellites equipped with GPUs. A promising approach is to use pre-trained vision-language modeling (VLM) to enhance its open-vocabulary capability. However, adopting it on satellites poses two challenges: (1) satellite imagery differs substantially from natural images, and (2) satellites' embedded GPUs are insufficient for complex models' inference. We reveal their lack of a crucial prior: in-orbit detection involves identifying a set of known objects within a cluttered yet monotonous background. Motivated by this observation, we propose VISO, a Vision-language Instructed Satellite Object detection model that focuses on object-specific features while suppressing irrelevant regions through language-guided mask learning. After pre-training on a large-scale satellite dataset with 3.4M region-text pairs, VISO enhances object-text alignment and object-centric features to improve detection accuracy. Also, VISO suppresses irrelevant regions, enabling highly sparse inference to accelerate speed on satellites. Extensive experiments show that VISO without sparsity outperforms state-of-the-art (SOTA) VLMs in zero-shot detection by increasing $34.1\\%$ AP and reducing $27\\times$ FLOPs, and surpasses specialist models in supervised object detection and object referring by improving $2.3\\%$ AP. When sparsifying VISO to a comparable AP, FLOPs can be greatly reduced by up to $8.5\\times$ . Real-world tests reveal that VISO achieves a $2.8-4.8\\times$ FPS speed-up on satellites' embedded GPUs1.", + "bbox": [ + 89, + 323, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 747, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, embedded GPUs are deployed in space [15], enabling in-orbit object detection to support latency-sensitive Earth observation tasks such as disaster monitoring [34] and wildfire detection [35]. In-orbit detection requires detecting flexible in-orbit tasks and responding to unforeseen emergencies beyond predefined closed-set categories. We observe that vision-language models (VLMs) [22] facilitate", + "bbox": [ + 89, + 772, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df6a0c1c5dffa0f4fc8bd879315f523f76ac7691532ef304a520b645f9e50e9c.jpg", + "image_caption": [ + "Figure 1. Earth observation images and natural images are different. Earth observation images typically contain very small objects. On the right is the size ratio of objects to images. We choose 5 categories in DOTA: plane (pl), storage tank (st), roundabout (rb), small vehicle (sv), baseball diamond (bd), and 5 categories in COCO: person (pr), cat (ct), car (cr), bed (bd), bottle (btl)." + ], + "image_footnote": [], + "bbox": [ + 514, + 290, + 694, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/812175ec29974e644e001667763d006274e13eecfd8385ee5e82290621cab7e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 291, + 901, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "open-vocabulary detection (OVD) [7], allowing for detecting novel objects with the guidance of natural language.", + "bbox": [ + 511, + 520, + 905, + 550 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, most existing VLMs [22, 25, 48] are designed for natural images and suffer from high computational overhead due to complex backbones (e.g., Swin [27] used in GLIP [22]). Although YOLO-World [7] provides a practical solution for time-critical applications, directly applying it to satellite data remains challenging due to the disparities between terrestrial natural images and satellite images, as shown in Figure 1. Satellite images are typically captured from a bird's eye view with a long distance from the ground. They often contain a few tiny objects (e.g., ships) relevant to referring detection tasks in orbit (e.g., detect the ship on the lower left), surrounded by vast irrelevant objects and backgrounds (e.g., buildings, roads, etc.), which makes detection difficult [24, 44]. We consider these priors and address two important goals. (1) Accurate detection requires distinguishing object-centric visual features from irrelevant features while incorporating useful contextual information. (2) Fast inference can be achieved by omitting large irrelevant regions, consisting of non-essential backgrounds and objects, to reduce redundant computation.", + "bbox": [ + 511, + 551, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we achieve the above two goals by proposing VISO, a Vision-language Instructed Satellite Object detection model. Our key insight is that rich semantic textual", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our code: https://github.com/joannahuadu/VISO.", + "bbox": [ + 107, + 886, + 452, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23300", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "embeddings of Earth observation tasks can help a detector distinguish different objects at its shallow layers (e.g., Figure 5a). Thus, we design a multi-scale mask learning to embed strongly aligned knowledge of Earth objects into multiscale visual features by pretraining on a large-scale satellite region-text pair. The learnable masks let VISO focus only on relevant areas at an early inference stage, ignoring irrelevant objects or cluttered backgrounds, thereby generating highly sparse features, which forward can be optimized by sparse computation (e.g., sparse convolution [14, 41]).", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to validate the effectiveness and efficiency of VISO. VISO surpasses 4 VLM baselines in zero-shot open-vocabulary detection on 3 benchmarks, obtaining $34.1\\%$ AP increase and $27\\times$ FLOPs reduction. Furthermore, VISO outperforms 3 SOTA specialist models on 2 object detection benchmarks by $2.3\\%$ AP and 4 grounding models on one object referring benchmark by $0.4\\%$ precision. By applying sparsity and aligning AP with baselines, VISO can achieve up to $8.5\\times$ reduction in FLOPs. Real-world tests further demonstrate a $2.8 - 4.8\\times$ FPS speed-up on satellites' embedded GPUs.", + "bbox": [ + 89, + 242, + 480, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Work", + "text_level": 1, + "bbox": [ + 89, + 424, + 375, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. GPUs on Satellites", + "text_level": 1, + "bbox": [ + 89, + 450, + 269, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Earth observation missions [40] rely on a pipeline in which satellites take photos and then send them back to ground stations for analysis like forestry [1] or agriculture [10, 31]. This leads to a substantial requirement for data transmission from satellites to the ground [47] and cannot fulfill latency-sensitive tasks like disaster detection [35].", + "bbox": [ + 89, + 472, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, embedded GPUs are deployed on satellites for Earth observation missions [47]. In-orbit DNN inference can process real-time tasks, especially object detection, and send only results back to ground stations. For example, OroraTech has launched AI nanosatellites with NVIDIA Jetson Xavier NX for in-orbit wildfire detection [35].", + "bbox": [ + 89, + 563, + 482, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although deploying GPUs in orbit is a promising approach, satellites are still constrained by size and energy supply to support sophisticated GPUs. For instance, NVIDIA Jetson Xavier NX, a widely-used in-orbit GPU, can only support 10-20 Watts [35] such that accelerating the in-orbit DNN inference becomes critical.", + "bbox": [ + 89, + 655, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Remote Sensing Object Detection", + "text_level": 1, + "bbox": [ + 89, + 756, + 382, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research [24, 29, 41] has been dedicated to object detection specified for remote sensing, which typically addresses the issue of small objects in satellite images. They propose two types of approaches: utilizing large convolution kernels to introduce a large receptive field and leveraging high-resolution images or features to preserve small objects, both of which lead to high computational overhead.", + "bbox": [ + 89, + 777, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, these methods pursue a lightweight solution.", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LSKNet [24] employs a spatial selective mechanism to decompose the large kernel into a sequence of depthwise kernels. RTMDet [29] uses large-kernel depth-wise convolutions in the backbone and neck while reducing the number of building blocks as compensation to reduce model depth. QueryDet [41] introduces a higher-resolution feature scale in feature pyramids and sparsifies the high-resolution heads through cascaded queries to speed up. However, we find that these methods either aim to compensate for the extra computation or have limited acceleration potential only within heads. In this paper, we emphasize fully releasing the acceleration capability of the detection of small objects.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Vision-language Model for Object Detection", + "text_level": 1, + "bbox": [ + 511, + 281, + 885, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary detection. Vision-language model (VLM) is a promising paradigm to empower open-vocabulary detection (OVD) according to open-set demands. Existing VLMs [22, 25, 48] unify OVD and visual grounding (or object referring) by reformulating the classification head as a text-contrastive head and designing vision-language fusion mode. However, these VLMs rely on complex backbones (e.g., Swin [27]), which are hard to deploy on resource-constrained devices. YOLO-World [7] address this by using the efficient YOLOv8 [17] as the backbone to enable real-time OVD. Despite this, these methods are pre-trained on natural images and the fusion lacks prior knowledge specific to satellite images, e.g., small objects.", + "bbox": [ + 511, + 303, + 903, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual grounding. Specifically for visual grounding, some research establishes a neat multi-modal correspondence based on a cross-modal transformer and directly regresses coordinates. TransVG [11] replaces the complex fusion modules with a simple stack of transformer encoder layers. VLTVG [42] introduces an additional bidirectional visual-linguistic encoder before the cross-model decoder. MGVLF [44] considers scale variations in remote sensing and takes advantage of multi-scale visual features and multi-granularity textual embeddings. LPVA [21] makes better multi-modal interaction in the visual feature extraction process. However, these transformer-based grounding models will increase computational costs and will somehow degrade object detection without bounding box regression.", + "bbox": [ + 511, + 500, + 903, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Region-text alignment. Various supervision strategies like [3, 16, 36] are proposed to enhance object-centric vision-language (i.e., region-text) alignment instead of coarse-grained image-level supervision [32] to make localization more precise. In this paper, we leverage prior knowledge from the Earth observation domain to enhance object-centric alignment and aim to improve efficiency in orbit.", + "bbox": [ + 511, + 710, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 830, + 604, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the aforementioned insights, in this section, we propose VISO, an efficient VLM tailored for Earth observation tasks. VISO can be seamlessly deployed on in", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23301", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e90e046a39a6a409c2dc24cab67d57a6dccb2c8b00fe8f57b4b01fb1545cbe2.jpg", + "image_caption": [ + "Figure 2. Overview of VISO. The black dashed border represents pre-training with additional supervised mask learning. The red dashed border represents satellite deployment with sparse inference in VL-PAN and heads, along with an offline vocabulary omitting the text encoder like [7]. The language-guided attention module inputs textual embeddings and visual features and outputs multi-scale masks. The masks are learned in the training phase through supervised mask learning and serve as guidance to convert dense features into sparse ones." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 883, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "orbit satellites with GPUs for Earth observation tasks like open-vocabulary detection and object referring.", + "bbox": [ + 88, + 340, + 482, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we introduce a multi-scale language-guided attention module to effectively enhance object-centric features relevant to language expressions while suppressing the irrelevant features (§3.1). Second, we perform a large-scale pre-training on satellite region-text pairs with deep supervision to enrich the knowledge of Earth (§3.2). Third, we convert the pre-trained dense models into highly sparse ones through multi-scale mask aggregation, ensuring minimal accuracy loss while improving inference efficiency (§3.3).", + "bbox": [ + 88, + 373, + 482, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 523, + 277, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of VISO is illustrated in Figure 2. The proposed language-guided attention module is integrated with multi-scale visual features at an early stage. Though the current implementation is built upon the lightweight open-vocabulary detector, YOLO-World [7], the integration can be generalized to any other detector.", + "bbox": [ + 88, + 546, + 483, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-vocabulary detector. VISO uses YOLO-World [7] as its base detector. The detector contains a backbone (i.e., CSPDarkNet, details can refer to [4]) to extract multi-scale visual features $C_l \\in \\mathbb{R}^{H_l \\times W_l \\times D_l}, l \\in 3, 4, 5$ from the image $I$ , a pre-trained text encoder (i.e., CLIP [32]) to encode the text input $T$ into textual embeddings $\\mathbf{W}_{j \\in \\{1, \\dots, N\\}} \\in \\mathbb{R}^{N \\times D}$ , a vision-language path aggregation network (VL-PAN) for multi-scale feature pyramids $\\{P_3, P_4, P_5\\}$ to fuse textual embeddings and visual features, a decoupled head to regress bounding box $\\{B_k\\}_{k=1}^K$ and object embeddings, and a text contrastive head to obtain object-text (region-text) similarity $\\{s_{k,j}\\}_{k=1}^K$ between object embeddings and textual embeddings. $N$ denotes the number of categories or expressions. $D$ denotes the dimension of textual embeddings. $K$ denotes the number of object predictions. We modify the regression output dimension from 4 to 5 to add a rotation angle prediction, enabling the commonly used ori", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ented object detection in Earth observation [29]. More details on VL-PAN and text contrastive head can refer to [7].", + "bbox": [ + 511, + 340, + 903, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-guided attention module. We observe that the textual embeddings $\\mathbf{W}$ possess rich semantic prior knowledge and exhibit strong alignment with visual features due to large-scale pre-training [2]. They can be leveraged to differentiate the visual features of different objects and backgrounds early in the detector, rather than relying solely on the final text contrastive head. Thus, we introduce a language-guided attention module immediately following the backbone. Specifically, we adopt one projection layer $l_{p}$ and the max-sigmoid attention to multi-scale visual features $C_l\\in \\mathbb{R}^{H_l\\times W_l\\times D_l},l\\in 3,4,5.$ The projection layer projects textual embeddings to the dimension of each scale. The max-sigmoid attention generates multi-scale masks $X_{l}\\in$ $\\mathbb{R}^{H_l\\times W_l\\times 1}(l\\in \\{3,4,5\\})$ by the cross-modal aggregation between textual and visual features:", + "bbox": [ + 511, + 375, + 906, + 602 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} M _ {l j} = C _ {l} \\mathbf {W} _ {\\mathbf {j}} ^ {T}, M _ {l j} \\in \\mathbb {R} ^ {H _ {l} \\times W _ {l} \\times N} \\\\ X _ {l} = \\delta \\left(\\max _ {j \\in \\{1, \\dots , N \\}} M _ {l j}\\right) \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 608, + 906, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\delta$ indicates the sigmoid function. We note that accurately detecting objects in satellite images often requires a wide range of contextual information due to the scale variations of Earth objects [29]. In general, low-level (high-resolution) features benefit small objects, while high-level features with large receptive fields benefit larger objects. Thus, incorporating multi-scale features to generate masks is crucial for preserving diverse receptive fields.", + "bbox": [ + 511, + 657, + 908, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pre-training Formulation", + "text_level": 1, + "bbox": [ + 511, + 787, + 746, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Region-text contrastive learning. We follow [7] and reformulate the ground-truth annotations as $M$ region-text pairs $\\{B_i, t_i\\}_{i=1}^M$ , where $t_i$ is the text for the object region $B_i$ . $t_i$ can be the category name for object detection or the referring expression for object referring. VISO inputs image $I$ and texts $T$ and outputs K object predictions $\\{B_k, s_k\\}_{k=1}^K$ .", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23302", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow [29] and leverage the dynamic soft label assignment to match predictions with ground-truth annotations. The text index of $t_i$ in $T$ is assigned for each positive prediction as a classification label. The assigned label and the predicted similarity are used to construct the region-text contrastive loss $\\mathcal{L}_{\\mathrm{con}}$ using quality focal loss [23]. In addition, rotated IoU loss [29] is adopted for oriented bounding box regression, and the basic training loss is defined as: $\\mathcal{L}_1 = \\mathcal{L}_{\\mathrm{con}} + \\lambda \\mathcal{L}_{\\mathrm{iou}}$ while $\\lambda$ is a balancing coefficient.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-grained semantic deep supervision. To learn masks $X_{l}$ , we explore a simple fine-grained deep supervision strategy. We assign the ground truth bitmap $\\mathcal{B}_{lj} \\in \\{0,1\\}^{H_l \\times W_l \\times 1} (l \\in \\{3,4,5\\}, j \\in \\{1,\\dots,N\\})$ for each category's $M_{lj}$ . The bitmap of each category is generated from all the object regions belonging to this category and scaled across multiple levels to represent diverse receptive fields. Specifically, the pixels located within the regions are valued at 1, and others are 0. We define the deep supervision's loss function as:", + "bbox": [ + 89, + 231, + 482, + 381 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\mathcal {H} \\left(M _ {l j}, \\mathcal {B} _ {l j}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 390, + 482, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{H}$ denotes the binary cross entropy function. Training the learnable masks is consistent with the basic training, and the total training loss is defined as $\\mathcal{L} = \\mathcal{L}_1 + \\alpha \\cdot \\mathcal{L}_2$ . $\\alpha$ is a balancing coefficient. We visualize the masks in Figure 5.", + "bbox": [ + 89, + 445, + 482, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sparse Accelerated Inference", + "text_level": 1, + "bbox": [ + 89, + 515, + 352, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multi-scale masks generated at the early stage can serve as effective guidance for enabling sparse computation in subsequent layers, e.g., neck and head layers. For instance, when detecting small vehicles in Figure 5a, the masks are centered on the top right, indicating that the remaining areas can be sparsified to avoid redundant further inference. Note that we only consider lightweight CNN-based detectors here. Offline sparse deployment involves three steps.", + "bbox": [ + 89, + 537, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the real-value masks $X_{l}$ need to be converted to binary masks $M_{l} \\in \\{0,1\\}^{H_{l} \\times W_{l} \\times 1} (l \\in \\{3,4,5\\})$ by hard threshold $\\theta$ , in which values greater than $\\theta$ are set to 1 and others are set to 0. Therefore, adjusting $\\theta$ can optimize the trade-off between accuracy and efficiency. Greater sparsity with larger $\\theta$ can lead to increased efficiency and decreased accuracy. We will give more validation in §4.4.", + "bbox": [ + 89, + 659, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, non-sparse locations of visual features $\\{C_3,C_4,C_5\\}$ need to be aggregated across multiple scales to compensate for contexts. We consider the location of 1 in each scale's mask $M_{l}$ as $I_{l} = \\{(x_{l}^{o},y_{l}^{o})\\} (o\\in \\{1,\\ldots ,S_{l}\\})$ in which $S_{l}$ is the non-sparsity of $M_{l}$ , i.e., total number of 1. In general, high-level features with low resolution and low-level features with high resolution are concatenated through up- or down-sampling in the feature pyramid. Thus, we use cross-scale projection to make up for the", + "bbox": [ + 89, + 765, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ce1128fc88bfca58ce1093b47f48692aafdcd5f8232ccc6bc961c9d63e05c34c.jpg", + "image_caption": [ + "Figure 3. The sparse computation in the sparse convolution operates by performing the convolution only when the center of a $3 \\times 3$ kernel (with 1 padding) covers a non-zero value (e.g., only two non-zero values are in the sparse input features)." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 898, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "non-sparse location. Specifically, when using a higher-level location $I_{l}$ to query the lower-level feature $C_{l-1}$ , the additional non-sparse location is denoted as:", + "bbox": [ + 511, + 258, + 903, + 304 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l - 1} ^ {\\prime} = \\left\\{\\left(2 x _ {l} ^ {o} + i, 2 y _ {l} ^ {o} + j\\right) \\mid \\forall i, j \\in \\{0, 1 \\} \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 311, + 903, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, update $I_{l-1}$ as $I_{l-1} \\cup I_{l-1}'$ . Querying higher-level features using lower-level locations is reversed by dividing.", + "bbox": [ + 511, + 340, + 903, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Third, based on $I_{l}$ , dense features $C_l \\in \\mathbb{R}^{H_l \\times W_l \\times D_l}$ are converted into sparse features $\\tilde{C}_l \\in \\mathbb{R}^{S_l \\times D_l}$ , which are then fed into the pre-converted sparse CNN layers for inference. The convolution in VL-PAN and heads are converted into sparse convolution (Figure 2). The computation of sparse convolution is illustrated in Figure 3. Take the $1 \\times 1$ convolution kernel size as an example, the computation complexity with sparse convolution is $S_{l} / HW$ of normal convolution. Thus, greater sparsity leads to lower computational complexity. Note that running sparse CNNs efficiently on general-purpose hardware (e.g., GPU) remains challenging. Thus, we leverage the existing high-performance inference engine (e.g., SpConv [9]) to accelerate sparse convolution computation. Additionally, we configure the sparse conversion as level-wise granularity (e.g., $P_3$ is sparse, while $P_4$ and $P_5$ remain dense), allowing for greater flexibility to adjust based on the hardware performance (\\$5).", + "bbox": [ + 511, + 369, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 638, + 643, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of the VISO pre-trained on large-scale satellite data and the efficiency of the sparse VISO guided by learned masks. We perform VISO in various downstream Earth observation tasks, including zero-shot detection on 3 benchmarks (§4.1), supervised detection on 2 benchmarks (§4.2), and object referring on one benchmark (§4.3). We further evaluate the speed of sparse VISO on the real-world satellite payload (§5).", + "bbox": [ + 511, + 664, + 903, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details. VISO is built upon the MMYOLO [8] and MMDetection [5] toolbox. For sparse inference, we implement a hook (SPHook) in MMDetection to convert VISO into a level-wise sparse model using SPconv [9]. The hook is easy to use, inserted at before_test_epoch $^2$ , with no extra overhead. In line with [7],", + "bbox": [ + 511, + 787, + 903, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2https://mmdetection.readthedocs.io", + "bbox": [ + 529, + 886, + 807, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/093bb8f9b7355abbe814076431055f8d791f4248ee17a737c54d2b09ff12e247.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Pre-training data. A list of datasets for pre-training VISO. HBB and OBB refer to horizontal and oriented bounding boxes, and Ann. denotes the annotations." + ], + "table_body": "
TypeDatasetVocab.ImagesAnno.Format
Object Detection (OD)DOTAv2 [12]18614701100487OBB
DIOR-R [20]2023463192518OBB
FAIR1Mv2 [38]3743611626790OBB
NWPU-VHR-10 [6]108325582HBB
UCAS-AOD [49]2302024610OBB
RSOD [13]4186914604HBB
HRRSD [46]136805593780HBB
VisDrone [50]12298591152646HBB
Visual Grounding (VG)DIOR-RSVG [44]-1740230898HBB
RRSISD [43]-1740213015HBB
GeoChat [18]-3832868563OBB
VRSBench [19]-2961438589OBB
", + "bbox": [ + 94, + 88, + 482, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we provide three variants of VISO to meet different requirements, e.g., small (S), medium (M), and large (L). We utilize the open-source CLIP [32] text encoder with pretrained weights to encode the text input. All satellite images used for pre-training and evaluation are split and resized to $1024 \\times 1024$ . Unless specified, all evaluation results of VISO are presented with two parts: a dense inference result (upper bound) and an optimal sparse inference result, which balances AP and FLOPs by adjusting the sparse threshold $\\theta$ . More selections of $\\theta$ will be discussed in §4.4.", + "bbox": [ + 88, + 311, + 482, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pre-training", + "text_level": 1, + "bbox": [ + 89, + 474, + 223, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training setup. We adopt the AdamW optimizer [28] with an initial learning rate of 0.0002 and weight decay of 0.025. VISO is pre-trained for 20 epochs on $8 \\times$ NVIDIA H800 GPUs with a total batch size of 64. We use vanilla data augmentations to reduce training costs, including random resizing, rotation, and flipping without using mosaic augmentation. The text encoder is fine-tuned on our pretraining data with a learning factor of 0.01.", + "bbox": [ + 88, + 497, + 482, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training data. We establish a large-scale dataset with 3.4M satellite region-text pairs to pre-train VISO. As listed in Table 1, we mainly collect 8 object detection datasets and 4 visual grounding datasets. We exclude the test sets of DOTAv2, FAIR1Mv2, and DIOR-RSVG, which are used for subsequent evaluation. The annotations for each dataset are formatted as region-text pairs, in which all the horizontal and oriented bounding boxes (regions) are represented as the corners of the quadrilateral and are uniformly converted to rotated bounding boxes when loading.", + "bbox": [ + 88, + 623, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot evaluation. After pre-training, we evaluate the pre-trained VISO on MAR20 [39], HRSC2016 [26], and VEDAI [33] benchmarks in a zero-shot manner following [45], as shown in Table 3. The MAR20 and HRSC2016 datasets contain more types of military aircraft and a diverse variety of ships. The VEDAI dataset contains 9 categories of vehicles on Earth. We compare the pre-trained VISO with recent state-of-the-art VLM mod", + "bbox": [ + 88, + 780, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1ac7dd99576fc6ac33934683df9b00901b8e39bbd168bfa030dc06cf0ab7504b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Ablations on pre-training data, frozen CLIP encoder, and language-guided attention module (LAM). We evaluate the zero-shot performance on MAR20, HRSC2016, and VEDAI by pre-training VISO-L with varying amounts of data, using either a frozen or fine-tuned text encoder, with or without LAM." + ], + "table_body": "
Training DataFrozen?LAM?MAR20HRSC2016VEDAI
only DOTAv2✓✓90.76.834.3
OD✗✓90.268.949.9
OD, VG✓✓89.968.845.0
OD, VG✗✗89.270.348.3
OD, VG✗✓90.070.750.2
", + "bbox": [ + 517, + 88, + 905, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "els [7, 22, 25, 48] and report AP and FLOPs. We preserve the pre-trained weights of these baselines (pre-trained on natural data) without any fine-tuning on satellite data. We find that VISO achieves remarkable performance in detecting Earth objects with $90.2\\%$ AP for MAR20, $75.6\\%$ AP for HRSC2016, and $50.2\\%$ AP for VEDAI, while obtaining more than $20\\times$ reduction in FLOPs. The results highlight the disparities between natural and satellite images, showing that existing VLMs cannot be directly applied to the Earth observation domain and incur high computational costs. Furthermore, we test sparse inference with an optimal threshold, which results in a further reduction in FLOPs with $\\leq 1\\%$ AP drop, e.g., $2\\times$ FLOPs reduction of VISO-S on MAR20, with a tiny AP drop, e.g., $0.2\\%$ . Please find more details of fair comparison in our appendix.", + "bbox": [ + 511, + 268, + 906, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ablation study. Table 2 presents ablation studies to analyze VISO from three aspects: pre-training data, fine-tuning the CLIP encoder, and the proposed language-guided attention module. We mainly conduct ablations based on VISO-L and evaluate them on three benchmarks in a zero-shot manner. First, compared to the baseline (grey) trained on all pre-training data, reducing visual grounding (VG) and partial object detection (OD) datasets leads to $0.3 - 63.9\\%$ AP drop. The decrease is due to the richer textual information from visual grounding datasets, enlarging the vocabulary of Earth objects. Second, we exploit two settings for the text encoder, CLIP, i.e., frozen and fine-tuned. Fine-tuning can improve AP by $0.1 - 5.2\\%$ as large-scale satellite data helps align the text encoder with the Earth observation domain, which is not well covered in CLIP's training data. Third, the proposed language-guided attention module improves AP by $0.4 - 1.9\\%$ , highlighting its effectiveness in distinguishing and improving object-centric feature extraction.", + "bbox": [ + 511, + 502, + 908, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Object Detection", + "text_level": 1, + "bbox": [ + 511, + 787, + 679, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We consider 3 specialist detectors as baselines, all of which focus on efficiency and practicality (§2.2). We train the baseline models separately on the training set of DOTAv2 and FAIR1Mv2 for 80 epochs (YOLOv8) and 36 epochs (RTMDet and LSKNet) using 2 NVIDIA H800 GPUs with a total batch size of 16. We adopt the AdamW", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23304", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9251c9b714fc3fd5c15c48a2069581a598c69c406dda8dd1a6600abd8c5060ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBackboneParams(M) ↓MAR20HRSC2016VEDAI
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Pre-trained open-set models
GLIP-TSwin-T23248889.448848.44881.1
GroundingDINO-TSwin-T17246486.746436.84642.6
MM-GroundingDINO-TSwin-T17246486.646445.446416.1
YOLO-World-SYOLOv8-S134454.14410.9440.8
YOLO-World-MYOLOv8-M2911361.111312.51131.2
YOLO-World-LYOLOv8-L4722955.222920.92291.4
Zero-shot transferNo sparsity → Sparsity
VISO-SYOLOv8-S1344→2287.6→87.444→2659.7→59.044→1838.8→37.9
VISO-MYOLOv8-M29113→5990.2→90.1113→6875.6→74.6113→5546.6→46.2
VISO-LYOLOv8-L47229→13090.0→90.0229→13470.7→70.1229→11850.2→49.9
", + "bbox": [ + 94, + 88, + 890, + 265 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/36177831db176a3aa64f0f88c15e44ca61707c66484c22b1a1688c3025eb71df.jpg", + "table_caption": [ + "Table 3. Zero-shot evaluation on MAR20, HRSC2016 and VEDAI test set. We compare VISO with existing VLMs on zero-shot transfer in terms of AP and FLOPs. No sparsity means the max accuracy of VISO and sparsity significantly reduces FLOPs with $\\leq 1\\%$ AP drop." + ], + "table_footnote": [ + "Table 4. Supervised results on DOTAv2 and FAIR1Mv2. We compare VISO with specialist models on object detection in terms of AP and FLOPs. The AP of VISO-L with No sparsity outperforms SOTA baselines (2.3% on DOTAv2 and 0.9% on FAIR1Mv2). We use sparsity to reduce AP until comparable with baselines (AP, AP*, and AP/) and find that FLOPs are significantly reduced." + ], + "table_body": "
ModelDOTAv2 (test)FAIR1Mv2 (test)
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Specialist models
YOLOv8-S4050.84044.8
YOLOv8-M10554.410547.4
YOLOv8-L21756.421749.2
RTMDet-S3858.73847.5
RTMDet-M10160.1/10149.7/
RTMDet-L20660.820651.4*
LSKNet-T12459.512546.9
LSKNet-S16161.1*16249.7/
Open-set modelsNo sparsity → Sparsity
VISO-S44→1956.8→50.644→2446.2→44.8
VISO-M113→6961.8→60.0/113→7650.9→50.0/
VISO-L229→12863.4→61.1*229→12752.3→51.5*
", + "bbox": [ + 94, + 315, + 483, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimizer [28] and set an initial learning rate of 0.0002 for YOLOv8, 0.00025 for RTMDet, and 0.0001 for LSKNet. The same data augmentations as pre-training are applied.", + "bbox": [ + 89, + 619, + 483, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Supervised comparison with specialist models. We compare the proposed VISO with 3 close-set detectors, i.e., YOLOv8 with 3 variants, RTMDet with 3 variants, and LSKNet with 2 variants, ensuring a fair comparison with our 3 variants. Table 4 shows the results in terms of AP and FLOPs on two object detection benchmarks, i.e., the DOTAv2 and FAIR1Mv2 test sets. The results demonstrate that VISO outperforms the SOTA method, i.e., LSKNet-S, by increasing $2.3\\%$ AP for DOTAv2 and $0.9\\%$ AP for FAIR1Mv2. For sparse inference, we adjust the threshold to obtain comparable AP with LSKNet-S, i.e., $61.1\\%$ for DOTAv2. We observe a significant drop in FLOPs of VISO-L from 229G to 128G, which is lower than the FLOPs of LSKNet-S, i.e., 161G. For FAIR1Mv2, we achieve both better FLOPs and AP than baselines. For example, VISO-M", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/32236f6d4f649d6a88671b1347beef143d2196f653bd754891bc8790d94a44e3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParams(M) ↓FLOPs(G) ↓Pr@0.5(%) ↑Pr@0.7(%) ↑
Specialist models
TransVG12620469.755.2
VLTVG12819476.962.3
MGVLF13221976.861.7
LPVA13722875.5*60.8
Open-set modelsNo sparsity → Sparsity
VISO-S1344→2474.4→70.169.1→64.7
VISO-M29113→8476.1→75.5*72.0→71.5
VISO-L48229→16877.3→75.8*72.9→71.6
", + "bbox": [ + 517, + 315, + 905, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Supervised results on DIOR-RSVG test set. We compare VISO with specialist models on object referring in terms of FLOPs, $\\mathrm{Pr}@\\mathrm{0.5}$ , and $\\mathrm{Pr}@\\mathrm{0.7}$ . Specialist models employ HBB and VISO adopts OBB for box format. Input size is $1024 \\times 1024$ .", + "bbox": [ + 511, + 460, + 906, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "obtains a better AP with only 76G FLOPs, reducing $2.1 \\times$ FLOPs than LSKNet-S and $1.3 \\times$ than RTMDet-M. Furthermore, VISO, built upon YOLOv8, achieves extremely lower FLOPs than YOLOv8, e.g., reducing $1.7 - 2.1 \\times$ FLOPs for S with comparable AP. Our approach can be extended to other detectors to attain even greater accuracy.", + "bbox": [ + 511, + 532, + 905, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Object referring", + "text_level": 1, + "bbox": [ + 511, + 632, + 678, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We consider 4 specialist visual grounding models as baselines: TransVG [11], VLTVG [42], MGVLF [44], LPVA [21], all of which are built upon a neat transformer-based framework (§2.3). We train the baseline models on the DIOR-RSVG training set for 90 epochs using 2 NVIDIA H800 GPUs with a total batch size of 64. The learning rate and data augmentation follow the default settings of each method. We use two metrics: $\\mathrm{Pr}@\\mathbf{0.5}$ and $\\mathrm{Pr}@\\mathbf{0.7}$ to measure IOU precision $\\geq 50\\%$ , $70\\%$ .", + "bbox": [ + 511, + 654, + 906, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Supervised comparison with specialist models. We compare our VISO with specialist baselines on the DIOR-RSVG test set. The results are shown in Table 5. We find that VISO outperforms SOTA methods by increasing $0.4\\%$ $\\mathrm{Pr}@\\mathrm{0.5}$ and $10.6\\%$ $\\mathrm{Pr}@\\mathrm{0.7}$ . We adjust the sparse threshold to obtain a comparable precision with baselines and observe a significant drop in FLOPs, e.g., reducing $8.5\\times$ from 204G", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23305", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b22f69b829f2f2db1940f9ade9f9217c374229471b571125ed5eb0b026d32651.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 95, + 235, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/559fac3ca2edbca54a70d25003b81fc09ada96b5f17f504067b4beec52d3a2d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 95, + 348, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88280350efebbe0a16660882d4f19b983774a1c341da05d7948d217e2c672c31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 95, + 478, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/317ccfd6c3cdaeb03161a7023ebb9057e43bf1ea156e40731af4ab037c34a813.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 95, + 604, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4114f4552b353a9f78dc933514fea536918f1d44b375003e3d7445cd5ea4d5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 95, + 741, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a813b5c5d1695e0520d49fb88bb27468c3b6993cd33d28fd6ee3c3fe50e8076a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 95, + 893, + 174 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dababf7fc43c398ea2f07c7f447745047282eaa673fc3686cbc706b9f5f0b353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 175, + 228, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e84e25a0af5b5c7509b5db2b8f37c3c05b59b74e644e5e804ae9b3581407e7e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 175, + 361, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8be960204a44fd272c7d858a77d1194d46883eb65eae3b45af851067b082050d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 175, + 493, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bae24427c1286d72c2e4d5783077467448b5618f788082c3d3cec7092b5a7ee6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 176, + 602, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8687f0cfacce830ac7190916b03bf60a6ea2188382281ceaddb5e3926bb874c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 176, + 756, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/552b35d6d06da7e8bccaa4c2aa7d6d41aef10de5990d05f814d48b707ea96b27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 176, + 893, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d604e734a3b969fdfe389f2f28456b48c65129e31d6e1cfc03ccbcb92e21f2c2.jpg", + "image_caption": [ + "Figure 4. Sparsity analysis of VISO on 6 benchmarks. We conduct extensive experiments on sparse VISO using different sparse thresholds (X-axis) and record AP (left Y-axis) and FLOPs (right Y-axis). As the threshold increases, both AP and FLOPs decline, with FLOPs at a faster rate. We use validation sets of DOTAv2 and FAIR1Mv2 for this evaluation to avoid limits on test submissions." + ], + "image_footnote": [], + "bbox": [ + 101, + 250, + 228, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6762bbdbd3bfcfb978ea23a7b1ab63e89eb11dcf6caf54f3da42f3ccf3488a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 250, + 361, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6624ac58b91a4d22bc66d5e6401bfcfa427d1633294e9c1defd4f237ce1dba04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 250, + 493, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b36a50032efc03113f9a47779b206ba1e51c44fb40ddc581a4a03d8e6a7228bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 250, + 602, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73d6b9b4a31a19386193d428d93791ca9eae40caa6fdafa67b05159528249ef7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 250, + 756, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d071e734b47a7603266512e6b117364fc12625b9586f78f18c39414cb19af1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 250, + 888, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b97d546f21ded74f7b2dd408e3056c1a084efc1da5df4183ce63462df1a58f78.jpg", + "image_caption": [ + "Figure 5. Mask visualizations of VISO. More visualization of detection results and masks is in our appendix." + ], + "image_footnote": [ + "(a) Mask visualization of fine-grained supervision for each target objects: (A) \"baseball diamond\", (B) \"small vehicle\", (C) \"tennis court\", (D) \"helicopter\". From (A) to (D), the multi-scale masks $\\{X_{3}, X_{4}, X_{5}\\}$ are arranged from the top (low-level feature) to the bottom (high-level feature)." + ], + "bbox": [ + 106, + 375, + 261, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c75ecbe5416243520879371c55343e1b24282a9cdabbf5abc96b7f9ae5b5fc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 376, + 318, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b38569b7a54204fad30dc437481bcfaa9ac2fe573b9985aae44f99dda9f2ef6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 320, + 376, + 372, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9bbac551d4c34af3254f0a76fe373862816c0b9ec3089af3f831836aa692990f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 376, + 429, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adb3d547684f1a95ade0db15b24b9da0d521d93ab9c3841226fc0a184fa5ce11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 376, + 486, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eff3fa802b7a44dae9c3fb668e57637e4ac0a7a7a3338030a6040ff6c2dec3d4.jpg", + "image_caption": [], + "image_footnote": [ + "(b) The sparsity of $\\{C_4\\}$ feature under different sparse thresholds. Left is the detection result (yellow-marked planes and white-marked helicopters) in DOTAv2. When the Earth observation task is specified as \"detect the plane\", masks only focus on planes and helicopters are irrelevant objects." + ], + "bbox": [ + 500, + 375, + 651, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b34f1452bba00a5043c7d0ab44e100f5649489dc1950d77efe69129a00e0ed2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 376, + 727, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ceed69ecda4f36001e636cc958e989051b722d978cbc568619f7e16ef38b4bce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 375, + 803, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5850d99058e40945a5c6cba4dff94a1912adb1634f6fb46be752c6f4f1b9042c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 375, + 880, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to 24G when $\\mathrm{Pr}@0.5$ is $70.1\\%$ of VISO-S.", + "bbox": [ + 89, + 580, + 374, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Sparsity analysis", + "text_level": 1, + "bbox": [ + 89, + 606, + 256, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All the above evaluations list only the optimal selection of sparse thresholds and employ a full-level sparse model for sparse inference, in which $P_{3}$ , $P_{4}$ , and $P_{5}$ are all sparse. We claim that adjusting the threshold $\\theta$ and level-wise granularity can address different requirements for in-orbit inference, prioritizing faster or better. Figure 4 illustrates the trade-off between AP and FLOPs with different thresholds from 0.01 to 0.9. The results show a substantial efficiency improvement, i.e., a sharp drop in FLOPs at very low thresholds, which is attributed to the highly sparse model, guided by learned masks of small objects in satellite images, and its ability to distinctly differentiate target objects from others.", + "bbox": [ + 89, + 628, + 483, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, the level-wise granularity also impacts FLOPs and AP. We denote the level-wise sparse conversion for $\\{P_3,P_4,P_5\\}$ as $\\triangleright \\{s_3,s_4,s_5\\} (s_{l\\in \\{3,4,5\\}}\\in \\{0,1\\})$ . For example, $\\triangleright \\{1,0,0\\}$ indicates the sparse $P_{3}$ and dense $P_4$ $P_{5}$ . Figure 4 shows that FLOPs decline more quickly in $\\triangleright \\{1,1,1\\}$ than in $\\triangleright \\{1,1,0\\}$ and $\\triangleright \\{1,0,0\\}$ , however, the AP", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "drop for $\\triangleright \\{1, 1, 1\\}$ and $\\triangleright \\{1, 1, 0\\}$ is nearly identical, indicating that low-resolution features in $P_{5}$ play a minor role in detecting small objects, primarily providing contextual information through a larger receptive field.", + "bbox": [ + 511, + 580, + 905, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualizations", + "text_level": 1, + "bbox": [ + 511, + 652, + 658, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of VISO, we visualize the learned masks in Figure 5. In Figure 5a, the masks focus separately on the relevant objects rather than others. Furthermore, as features go deeper from $C_3$ to $C_5$ , more salient contextual information is captured, more than just the outline of objects. Figure 5b illustrates the sparsity of masked features under different thresholds. As the threshold $\\theta$ increases from 0.01 to 0.9, the focus narrows with minimal loss of essential features. Specifically, the task is to detect the \"plane\", not the \"helicopter\". Thus, the mask gradually becomes more focused around the \"plane\" as the threshold increases. Redundant features (e.g., black areas) are masked out during sparse inference. For example, $\\theta = 0.9$ can result in an extremely high sparsity of $99.7\\%$ , preserving important features while significantly improving speed.", + "bbox": [ + 509, + 674, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "AP FLOPs {1,0,0} {1,1,0} {1,1,1}", + "bbox": [ + 330, + 80, + 669, + 92 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23306", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92faa1698e3e2593547b305ef6c909a93076d8d73c0c56d2a95060238e00c427.jpg", + "image_caption": [ + "Figure 6. The acceleration performance of sparse convolution (SPConv) and normal convolution (Conv). The comparison is across different feature scales of three variants. SPConv provides significant acceleration for large features compared to Conv. However, for small features, Conv is already well-optimized." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Real-world Analysis", + "text_level": 1, + "bbox": [ + 89, + 328, + 284, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following [17], we use frames per second (FPS) to measure the actual speed of detection on a real-world embedded GPU, NVIDIA Jetson Xavier NX [30], which serves as a payload on Chaohu-1 SAR satellite [37] launched in 2022. In this section, we consider a more common scenario in Earth observation missions: single-object detection, i.e., only one target or one category target is of interest.", + "bbox": [ + 89, + 354, + 482, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation details. NX supports three optimized power budgets, e.g., 10W, 15W, and 20W. We pre-confine the mode as 10W with a 4-core CPU of $1900\\mathrm{MHz}$ for our evaluation. The system of our testbed is equipped with Jet-Pack 5.1.3, CUDA 11.4, cuDNN 8.6.0, Pytorch 1.11.0, and SPconv 2.3.6. We test all models without extra acceleration mechanisms, e.g., FP16 or TensorRT.", + "bbox": [ + 89, + 465, + 483, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse configuration. We use SPconv [9] as the high-performance sparse inference engine for onboard deployment. As shown in Figure 6, we test the acceleration performance of SPconv with different feature scales on our testbed. Note that the performance varies across different hardware. We consider the commonly used $3 \\times 3$ kernel and compare the average latency of sparse convolution at different sparsity with that of normal convolution. The sparsity is defined as the ratio of the number of zeros in the feature to the total area of the feature. We observe that the larger the feature scale, the acceleration of SPConv is more pronounced. Especially for $\\mathrm{L}\\{P_3, P_4\\}$ , $\\mathrm{M}\\{P_3, P_4, P_5\\}$ and $\\mathrm{S}\\{P_3\\}$ of VISO, a sparsity of more than $25\\%$ can achieve acceleration, which can be easily attained in our highly sparse model when detecting small objects in satellite images, i.e., the smaller the object, the faster the detection. Thus, we configure the level-wise granularity to allow for more flexible sparse conversion, targeting different feature scales and hardware platforms.", + "bbox": [ + 89, + 577, + 482, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We consider 2 baselines that are widely used in the industrial domain [17, 41], including a commonly used", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/10524d770f15dff6e34b3e866c76a35963b1821e00f7149e23af99b10a1c9303.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFLOPs(G) ↓FPS ↑AP(%) ↑Sparsity(%) ↓
Dense models
YOLOv8-S404.043.6-
YOLOv8-M1052.032.0-
YOLOv8-L2171.125.3-
Sparse models
QueryDet2972.334.838
VISO-S1611.148.388
VISO-M495.433.582
VISO-L1103.038.176
", + "bbox": [ + 526, + 88, + 897, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Real-world test on NX. We compare VISO with YOLOv8 and QueryDet. All variants of VISO are full-level sparse, and we set $\\theta = 0.1$ for S, 0.25 for M, and 0.2 for L. We also list the average sparsity of all sparse features in sparse models.", + "bbox": [ + 511, + 239, + 903, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "YOLOv8 [17] detector mentioned above and a sparse detector QueryDet [41]. QueryDet introduces a higher-resolution $P_{2}$ to improve small object detection and a query mechanism to make high-resolution heads sparse. We reproduce QueryDet using default settings and train on the DOTAv2 training set with the size of $1024 \\times 1024$ .", + "bbox": [ + 511, + 316, + 906, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Speed testing. In Table 6, we detect the \"helicopter\" in DOTAv2 and compare VISO with YOLOv8 and QueryDet in terms of AP, FLOPs, FPS, and sparsity. VISO improves FPS by $2.7 - 2.8 \\times$ than YOLOv8 and $4.8 \\times$ than QueryDet under a better AP. VISO achieves significantly higher sparsity due to two folds: (1) The sparse computation of VISO occurs earlier, after the backbone, with sparse necks and heads, rather than only sparse heads in QueryDet. Also, VISO does not introduce higher-resolution features. (2) When detecting the target in irrelevant images, masks are unfocused, e.g., detect the helicopter in Figure 5a(D), and thus can early exit to avoid further inference.", + "bbox": [ + 511, + 419, + 908, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 628, + 633, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present VISO, an effective and efficient solution for in-orbit real-time object detection. VISO empowers vision-language models with language-guided mask learning and sparse inference. In this paper, we introduce a language-guided attention module to enhance salient object-centric features while suppressing irrelevant regions in satellite images. We pre-train VISO with large-scale satellite data in the form of region-text pairs with contrastive loss and additional mask learning, thereby endowing VISO with a strong capability of in-orbit open-vocabulary detection. Furthermore, the learned masks of VISO can suppress irrelevant regions to enable highly sparse inference and accelerate speed. Extensive experiments can demonstrate the superiority of VISO in both detection accuracy and speed, especially the significant speed-up on real-world satellites' embedded GPUs is insightful for future industrial deployment.", + "bbox": [ + 511, + 659, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23307", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Center of High Performance Computing, Tsinghua University.", + "bbox": [ + 89, + 114, + 482, + 147 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Frédéric Achard, Hans-Jürgen Stibig, Hugh D Eva, Erik J Lindquist, Alexandre Bouvet, Olivier Arino, and Philippe Mayaux. Estimating tropical deforestation from earth observation data. Carbon Management, 1(2):271-287, 2010. 2", + "[2] Ali Bahri, Sina Ghofrani Majelan, Sina Mohammadi, Mehrdad Noori, and Karim Mohammadi. Remote sensing image classification via improved cross-entropy loss and transfer learning strategy based on deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 17 (6):1087–1091, 2020. 3", + "[3] Hanoona Bangalath, Muhammad Maaz, Muhammad Uzair Khattak, Salman H Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. Advances in Neural Information Processing Systems, 35:33781-33794, 2022. 2", + "[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3", + "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 4", + "[6] Gong Cheng, Junwei Han, Peicheng Zhou, and Lei Guo. Multi-class geospatial object detection and geographic image classification based on collection of part detectors. ISPRS Journal of Photogrammetry and Remote Sensing, 98: 119-132, 2014. 5", + "[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xing-gang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 1, 2, 3, 4, 5", + "[8] MMYOLO Contributors. Mmyolo: Openmmlab yolo series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022. Accessed: 2023-03-03. 4", + "[9] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022. 4, 8", + "[10] Laura Crocetti, Matthias Forkel, Milan Fischer, František Jurečka, Aleks Grlj, Andreas Salentinig, Miroslav Trnka, Martha Anderson, Wai-Tim Ng, Žiga Kokalj, et al. Earth observation for agricultural drought monitoring in the pannonian basin (southeastern europe): current state and future directions. Regional Environmental Change, 20:1-17, 2020. 2", + "[11] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 186, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 1769-1779, 2021. 2, 6", + "[12] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44 (11):7778-7796, 2021. 5", + "[13] Qiqi Ding, Peng Li, Xuefeng Yan, Ding Shi, Luming Liang, Weiming Wang, Haoran Xie, Jonathan Li, and Mingqiang Wei. Cf-yolo: Cross fusion yolo for object detection in adverse weather with a high-quality real snow dataset. IEEE Transactions on Intelligent Transportation Systems, 24(10): 10749-10759, 2023. 5", + "[14] Bowei Du, Yecheng Huang, Jiaxin Chen, and Di Huang. Adaptive sparse convolutional networks with global context enhancement for faster object detection on drone images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13435-13444, 2023. 2", + "[15] Max Ghiglione and Vittorio Serra. Opportunities and challenges of ai on satellite processing units. In Proceedings of the 19th ACM international conference on computing Frontiers, pages 221-224, 2022. 1", + "[16] Jiaxing Huang, Jingyi Zhang, Kai Jiang, and Shijian Lu. Open-vocabulary object detection via language hierarchy. arXiv preprint arXiv:2410.20371, 2024. 2", + "[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLOv8, 2023. 2, 8", + "[18] Kartik Kuckreja, Muhammad S. Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad S. Khan. Geochat: Grounded large vision-language model for remote sensing. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5", + "[19] Jian Li, Xiang Ding and Mohamed Elhoseiny. Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding. arXiv:2406.12384, 2024. 5", + "[20] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 5", + "[21] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 6", + "[22] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1, 2, 5", + "[23] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in neural information processing systems, 33:21002-21012, 2020. 4", + "[24] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23308", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF international conference on computer vision, pages 16794-16805, 2023. 1, 2", + "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2, 5", + "[26] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, pages 324-331. SciTePress, 2017. 5", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6", + "[29] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 2, 3, 4", + "[30] NVIDIA. Jetson Xavier NX Developer Kit. https://developer.nvidia.com/embedded/jetson-xavier-nx-devkit, 2024. Online; accessed 14-March-2024. 8", + "[31] George P Petropoulos, Prashant K Srivastava, Maria Piles, and Simon Pearson. Earth observation-based operational estimation of soil moisture and evapotranspiration for agricultural crops in support of sustainable water management. Sustainability, 10(1):181, 2018. 2", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5", + "[33] Sebastien Razakarivony and Frederic Jurie. Vehicle detection in aerial imagery: A small target detection benchmark. Journal of Visual Communication and Image Representation, 34:187-203, 2016. 5", + "[34] Gupta Ritwik et al. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint, 2019. 1", + "[35] Daniel Saaristo. Shoot from the stars: Startup provides early detection of wildfires from space. https://blogs.nvidia.com/blog/ororatech-wildfires-from-space/, 2023.1,2", + "[36] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 2", + "[37] Spacety. Chaohu-1 SAR satellite. https://www.satelliteevolution.com/post/two-new-satellites-launched-including-the-2nd-sar-satellite-with-improved-capabilities,2022.8" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 5", + "[39] YU Wenqi, CHENG Gong, WANG Meijun, YAO Yanqing, XIE Xingxing, YAO Xiwen, and HAN Junwei. Mar20: A benchmark for military aircraft recognition in remote sensing images. National Remote Sensing Bulletin, 27(12):2688-2696, 2024. 5", + "[40] Michael A Wulder, David P Roy, Volker C Radeloff, Thomas R Loveland, Martha C Anderson, David M Johnson, Sean Healey, Zhe Zhu, Theodore A Scambos, Nima Pahlevan, et al. Fifty years of landsat science and impacts. Remote Sensing of Environment, 280:113195, 2022. 2", + "[41] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13668-13677, 2022. 2, 8", + "[42] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 2, 6", + "[43] Zhenghang Yuan, Lichao Mou, Yuansheng Hua, and Xiao Xiang Zhu. Rrasis: Referring remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 5", + "[44] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1, 2, 5, 6", + "[45] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multimodal large language model for multisensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 5", + "[46] Yuanlin Zhang, Yuan Yuan, Yachuang Feng, and Xiaoqiang Lu. Hierarchical and robust convolutional neural network for very high-resolution remote sensing object detection. IEEE Transactions on Geoscience and Remote Sensing, 57 (8):5535-5548, 2019. 5", + "[47] Ziyuan Zhang, Han Qiu, Maosen Zhang, Jun Liu, Bin Chen, Tianwei Zhang, and Hewu Li. COSMIC: Compress satellite images efficiently via diffusion compensation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[48] Xiangyu Zhao, Yicheng Chen, Shilin Xu, Xiangtai Li, Xinjiang Wang, Yining Li, and Haian Huang. An open and comprehensive pipeline for unified object grounding and detection. arXiv preprint arXiv:2401.02361, 2024. 1, 2, 5", + "[49] Haigang Zhu, Xiaogang Chen, Weiqun Dai, Kun Fu, Qixiang Ye, and Jianbin Jiao. Orientation robust object detection in aerial images using deep convolutional neural network. In 2015 IEEE international conference on image processing (ICIP), pages 3735-3739. IEEE, 2015. 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23309", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[50] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 5", + "bbox": [ + 91, + 90, + 482, + 160 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23310", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_model.json b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aebb7a1fb0a2418a898447b55a6596904874d18e --- /dev/null +++ b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "VISO: Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.204, + 0.509, + 0.257 + ], + "angle": 0, + "content": "Meiqi Wang Tsinghua University wang-mq22@mails.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.565, + 0.204, + 0.764, + 0.257 + ], + "angle": 0, + "content": "Han Qiu Tsinghua University qiuhan@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.732 + ], + "angle": 0, + "content": "In-orbit object detection is essential for Earth observation missions on satellites equipped with GPUs. A promising approach is to use pre-trained vision-language modeling (VLM) to enhance its open-vocabulary capability. However, adopting it on satellites poses two challenges: (1) satellite imagery differs substantially from natural images, and (2) satellites' embedded GPUs are insufficient for complex models' inference. We reveal their lack of a crucial prior: in-orbit detection involves identifying a set of known objects within a cluttered yet monotonous background. Motivated by this observation, we propose VISO, a Vision-language Instructed Satellite Object detection model that focuses on object-specific features while suppressing irrelevant regions through language-guided mask learning. After pre-training on a large-scale satellite dataset with 3.4M region-text pairs, VISO enhances object-text alignment and object-centric features to improve detection accuracy. Also, VISO suppresses irrelevant regions, enabling highly sparse inference to accelerate speed on satellites. Extensive experiments show that VISO without sparsity outperforms state-of-the-art (SOTA) VLMs in zero-shot detection by increasing \\(34.1\\%\\) AP and reducing \\(27\\times\\) FLOPs, and surpasses specialist models in supervised object detection and object referring by improving \\(2.3\\%\\) AP. When sparsifying VISO to a comparable AP, FLOPs can be greatly reduced by up to \\(8.5\\times\\). Real-world tests reveal that VISO achieves a \\(2.8-4.8\\times\\) FPS speed-up on satellites' embedded GPUs1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.748, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Recently, embedded GPUs are deployed in space [15], enabling in-orbit object detection to support latency-sensitive Earth observation tasks such as disaster monitoring [34] and wildfire detection [35]. In-orbit detection requires detecting flexible in-orbit tasks and responding to unforeseen emergencies beyond predefined closed-set categories. We observe that vision-language models (VLMs) [22] facilitate" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.291, + 0.695, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.292, + 0.903, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Figure 1. Earth observation images and natural images are different. Earth observation images typically contain very small objects. On the right is the size ratio of objects to images. We choose 5 categories in DOTA: plane (pl), storage tank (st), roundabout (rb), small vehicle (sv), baseball diamond (bd), and 5 categories in COCO: person (pr), cat (ct), car (cr), bed (bd), bottle (btl)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.551 + ], + "angle": 0, + "content": "open-vocabulary detection (OVD) [7], allowing for detecting novel objects with the guidance of natural language." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.854 + ], + "angle": 0, + "content": "However, most existing VLMs [22, 25, 48] are designed for natural images and suffer from high computational overhead due to complex backbones (e.g., Swin [27] used in GLIP [22]). Although YOLO-World [7] provides a practical solution for time-critical applications, directly applying it to satellite data remains challenging due to the disparities between terrestrial natural images and satellite images, as shown in Figure 1. Satellite images are typically captured from a bird's eye view with a long distance from the ground. They often contain a few tiny objects (e.g., ships) relevant to referring detection tasks in orbit (e.g., detect the ship on the lower left), surrounded by vast irrelevant objects and backgrounds (e.g., buildings, roads, etc.), which makes detection difficult [24, 44]. We consider these priors and address two important goals. (1) Accurate detection requires distinguishing object-centric visual features from irrelevant features while incorporating useful contextual information. (2) Fast inference can be achieved by omitting large irrelevant regions, consisting of non-essential backgrounds and objects, to reduce redundant computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we achieve the above two goals by proposing VISO, a Vision-language Instructed Satellite Object detection model. Our key insight is that rich semantic textual" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.454, + 0.901 + ], + "angle": 0, + "content": "1Our code: https://github.com/joannahuadu/VISO." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23300" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "embeddings of Earth observation tasks can help a detector distinguish different objects at its shallow layers (e.g., Figure 5a). Thus, we design a multi-scale mask learning to embed strongly aligned knowledge of Earth objects into multiscale visual features by pretraining on a large-scale satellite region-text pair. The learnable masks let VISO focus only on relevant areas at an early inference stage, ignoring irrelevant objects or cluttered backgrounds, thereby generating highly sparse features, which forward can be optimized by sparse computation (e.g., sparse convolution [14, 41])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.41 + ], + "angle": 0, + "content": "We conduct extensive experiments to validate the effectiveness and efficiency of VISO. VISO surpasses 4 VLM baselines in zero-shot open-vocabulary detection on 3 benchmarks, obtaining \\(34.1\\%\\) AP increase and \\(27\\times\\) FLOPs reduction. Furthermore, VISO outperforms 3 SOTA specialist models on 2 object detection benchmarks by \\(2.3\\%\\) AP and 4 grounding models on one object referring benchmark by \\(0.4\\%\\) precision. By applying sparsity and aligning AP with baselines, VISO can achieve up to \\(8.5\\times\\) reduction in FLOPs. Real-world tests further demonstrate a \\(2.8 - 4.8\\times\\) FPS speed-up on satellites' embedded GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.376, + 0.442 + ], + "angle": 0, + "content": "2. Background and Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.27, + 0.466 + ], + "angle": 0, + "content": "2.1. GPUs on Satellites" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Traditional Earth observation missions [40] rely on a pipeline in which satellites take photos and then send them back to ground stations for analysis like forestry [1] or agriculture [10, 31]. This leads to a substantial requirement for data transmission from satellites to the ground [47] and cannot fulfill latency-sensitive tasks like disaster detection [35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Recently, embedded GPUs are deployed on satellites for Earth observation missions [47]. In-orbit DNN inference can process real-time tasks, especially object detection, and send only results back to ground stations. For example, OroraTech has launched AI nanosatellites with NVIDIA Jetson Xavier NX for in-orbit wildfire detection [35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.745 + ], + "angle": 0, + "content": "Although deploying GPUs in orbit is a promising approach, satellites are still constrained by size and energy supply to support sophisticated GPUs. For instance, NVIDIA Jetson Xavier NX, a widely-used in-orbit GPU, can only support 10-20 Watts [35] such that accelerating the in-orbit DNN inference becomes critical." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.383, + 0.773 + ], + "angle": 0, + "content": "2.2. Remote Sensing Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Recent research [24, 29, 41] has been dedicated to object detection specified for remote sensing, which typically addresses the issue of small objects in satellite images. They propose two types of approaches: utilizing large convolution kernels to introduce a large receptive field and leveraging high-resolution images or features to preserve small objects, both of which lead to high computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Thus, these methods pursue a lightweight solution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "LSKNet [24] employs a spatial selective mechanism to decompose the large kernel into a sequence of depthwise kernels. RTMDet [29] uses large-kernel depth-wise convolutions in the backbone and neck while reducing the number of building blocks as compensation to reduce model depth. QueryDet [41] introduces a higher-resolution feature scale in feature pyramids and sparsifies the high-resolution heads through cascaded queries to speed up. However, we find that these methods either aim to compensate for the extra computation or have limited acceleration potential only within heads. In this paper, we emphasize fully releasing the acceleration capability of the detection of small objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.282, + 0.887, + 0.298 + ], + "angle": 0, + "content": "2.3. Vision-language Model for Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.5 + ], + "angle": 0, + "content": "Open-vocabulary detection. Vision-language model (VLM) is a promising paradigm to empower open-vocabulary detection (OVD) according to open-set demands. Existing VLMs [22, 25, 48] unify OVD and visual grounding (or object referring) by reformulating the classification head as a text-contrastive head and designing vision-language fusion mode. However, these VLMs rely on complex backbones (e.g., Swin [27]), which are hard to deploy on resource-constrained devices. YOLO-World [7] address this by using the efficient YOLOv8 [17] as the backbone to enable real-time OVD. Despite this, these methods are pre-trained on natural images and the fusion lacks prior knowledge specific to satellite images, e.g., small objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.905, + 0.711 + ], + "angle": 0, + "content": "Visual grounding. Specifically for visual grounding, some research establishes a neat multi-modal correspondence based on a cross-modal transformer and directly regresses coordinates. TransVG [11] replaces the complex fusion modules with a simple stack of transformer encoder layers. VLTVG [42] introduces an additional bidirectional visual-linguistic encoder before the cross-model decoder. MGVLF [44] considers scale variations in remote sensing and takes advantage of multi-scale visual features and multi-granularity textual embeddings. LPVA [21] makes better multi-modal interaction in the visual feature extraction process. However, these transformer-based grounding models will increase computational costs and will somehow degrade object detection without bounding box regression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.905, + 0.817 + ], + "angle": 0, + "content": "Region-text alignment. Various supervision strategies like [3, 16, 36] are proposed to enhance object-centric vision-language (i.e., region-text) alignment instead of coarse-grained image-level supervision [32] to make localization more precise. In this paper, we leverage prior knowledge from the Earth observation domain to enhance object-centric alignment and aim to improve efficiency in orbit." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.605, + 0.845 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Motivated by the aforementioned insights, in this section, we propose VISO, an efficient VLM tailored for Earth observation tasks. VISO can be seamlessly deployed on in" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23301" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.091, + 0.885, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.271, + 0.908, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Overview of VISO. The black dashed border represents pre-training with additional supervised mask learning. The red dashed border represents satellite deployment with sparse inference in VL-PAN and heads, along with an offline vocabulary omitting the text encoder like [7]. The language-guided attention module inputs textual embeddings and visual features and outputs multi-scale masks. The masks are learned in the training phase through supervised mask learning and serve as guidance to convert dense features into sparse ones." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.483, + 0.372 + ], + "angle": 0, + "content": "orbit satellites with GPUs for Earth observation tasks like open-vocabulary detection and object referring." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.374, + 0.483, + 0.511 + ], + "angle": 0, + "content": "First, we introduce a multi-scale language-guided attention module to effectively enhance object-centric features relevant to language expressions while suppressing the irrelevant features (§3.1). Second, we perform a large-scale pre-training on satellite region-text pairs with deep supervision to enrich the knowledge of Earth (§3.2). Third, we convert the pre-trained dense models into highly sparse ones through multi-scale mask aggregation, ensuring minimal accuracy loss while improving inference efficiency (§3.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.279, + 0.538 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.484, + 0.639 + ], + "angle": 0, + "content": "The architecture of VISO is illustrated in Figure 2. The proposed language-guided attention module is integrated with multi-scale visual features at an early stage. Though the current implementation is built upon the lightweight open-vocabulary detector, YOLO-World [7], the integration can be generalized to any other detector." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Open-vocabulary detector. VISO uses YOLO-World [7] as its base detector. The detector contains a backbone (i.e., CSPDarkNet, details can refer to [4]) to extract multi-scale visual features \\( C_l \\in \\mathbb{R}^{H_l \\times W_l \\times D_l}, l \\in 3, 4, 5 \\) from the image \\( I \\), a pre-trained text encoder (i.e., CLIP [32]) to encode the text input \\( T \\) into textual embeddings \\( \\mathbf{W}_{j \\in \\{1, \\dots, N\\}} \\in \\mathbb{R}^{N \\times D} \\), a vision-language path aggregation network (VL-PAN) for multi-scale feature pyramids \\( \\{P_3, P_4, P_5\\} \\) to fuse textual embeddings and visual features, a decoupled head to regress bounding box \\( \\{B_k\\}_{k=1}^K \\) and object embeddings, and a text contrastive head to obtain object-text (region-text) similarity \\( \\{s_{k,j}\\}_{k=1}^K \\) between object embeddings and textual embeddings. \\( N \\) denotes the number of categories or expressions. \\( D \\) denotes the dimension of textual embeddings. \\( K \\) denotes the number of object predictions. We modify the regression output dimension from 4 to 5 to add a rotation angle prediction, enabling the commonly used ori" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.372 + ], + "angle": 0, + "content": "ented object detection in Earth observation [29]. More details on VL-PAN and text contrastive head can refer to [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Language-guided attention module. We observe that the textual embeddings \\(\\mathbf{W}\\) possess rich semantic prior knowledge and exhibit strong alignment with visual features due to large-scale pre-training [2]. They can be leveraged to differentiate the visual features of different objects and backgrounds early in the detector, rather than relying solely on the final text contrastive head. Thus, we introduce a language-guided attention module immediately following the backbone. Specifically, we adopt one projection layer \\(l_{p}\\) and the max-sigmoid attention to multi-scale visual features \\(C_l\\in \\mathbb{R}^{H_l\\times W_l\\times D_l},l\\in 3,4,5.\\) The projection layer projects textual embeddings to the dimension of each scale. The max-sigmoid attention generates multi-scale masks \\(X_{l}\\in\\) \\(\\mathbb{R}^{H_l\\times W_l\\times 1}(l\\in \\{3,4,5\\})\\) by the cross-modal aggregation between textual and visual features:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.609, + 0.907, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} M _ {l j} = C _ {l} \\mathbf {W} _ {\\mathbf {j}} ^ {T}, M _ {l j} \\in \\mathbb {R} ^ {H _ {l} \\times W _ {l} \\times N} \\\\ X _ {l} = \\delta \\left(\\max _ {j \\in \\{1, \\dots , N \\}} M _ {l j}\\right) \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.781 + ], + "angle": 0, + "content": "\\(\\delta\\) indicates the sigmoid function. We note that accurately detecting objects in satellite images often requires a wide range of contextual information due to the scale variations of Earth objects [29]. In general, low-level (high-resolution) features benefit small objects, while high-level features with large receptive fields benefit larger objects. Thus, incorporating multi-scale features to generate masks is crucial for preserving diverse receptive fields." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.747, + 0.804 + ], + "angle": 0, + "content": "3.2. Pre-training Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.904 + ], + "angle": 0, + "content": "Region-text contrastive learning. We follow [7] and reformulate the ground-truth annotations as \\(M\\) region-text pairs \\(\\{B_i, t_i\\}_{i=1}^M\\), where \\(t_i\\) is the text for the object region \\(B_i\\). \\(t_i\\) can be the category name for object detection or the referring expression for object referring. VISO inputs image \\(I\\) and texts \\(T\\) and outputs K object predictions \\(\\{B_k, s_k\\}_{k=1}^K\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "We follow [29] and leverage the dynamic soft label assignment to match predictions with ground-truth annotations. The text index of \\( t_i \\) in \\( T \\) is assigned for each positive prediction as a classification label. The assigned label and the predicted similarity are used to construct the region-text contrastive loss \\( \\mathcal{L}_{\\mathrm{con}} \\) using quality focal loss [23]. In addition, rotated IoU loss [29] is adopted for oriented bounding box regression, and the basic training loss is defined as: \\( \\mathcal{L}_1 = \\mathcal{L}_{\\mathrm{con}} + \\lambda \\mathcal{L}_{\\mathrm{iou}} \\) while \\( \\lambda \\) is a balancing coefficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.483, + 0.382 + ], + "angle": 0, + "content": "Fine-grained semantic deep supervision. To learn masks \\( X_{l} \\), we explore a simple fine-grained deep supervision strategy. We assign the ground truth bitmap \\( \\mathcal{B}_{lj} \\in \\{0,1\\}^{H_l \\times W_l \\times 1} (l \\in \\{3,4,5\\}, j \\in \\{1,\\dots,N\\}) \\) for each category's \\( M_{lj} \\). The bitmap of each category is generated from all the object regions belonging to this category and scaled across multiple levels to represent diverse receptive fields. Specifically, the pixels located within the regions are valued at 1, and others are 0. We define the deep supervision's loss function as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.391, + 0.483, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\mathcal {H} \\left(M _ {l j}, \\mathcal {B} _ {l j}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.507 + ], + "angle": 0, + "content": "\\(\\mathcal{H}\\) denotes the binary cross entropy function. Training the learnable masks is consistent with the basic training, and the total training loss is defined as \\(\\mathcal{L} = \\mathcal{L}_1 + \\alpha \\cdot \\mathcal{L}_2\\). \\(\\alpha\\) is a balancing coefficient. We visualize the masks in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.516, + 0.353, + 0.532 + ], + "angle": 0, + "content": "3.3. Sparse Accelerated Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.659 + ], + "angle": 0, + "content": "The multi-scale masks generated at the early stage can serve as effective guidance for enabling sparse computation in subsequent layers, e.g., neck and head layers. For instance, when detecting small vehicles in Figure 5a, the masks are centered on the top right, indicating that the remaining areas can be sparsified to avoid redundant further inference. Note that we only consider lightweight CNN-based detectors here. Offline sparse deployment involves three steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.765 + ], + "angle": 0, + "content": "First, the real-value masks \\( X_{l} \\) need to be converted to binary masks \\( M_{l} \\in \\{0,1\\}^{H_{l} \\times W_{l} \\times 1} (l \\in \\{3,4,5\\}) \\) by hard threshold \\( \\theta \\), in which values greater than \\( \\theta \\) are set to 1 and others are set to 0. Therefore, adjusting \\( \\theta \\) can optimize the trade-off between accuracy and efficiency. Greater sparsity with larger \\( \\theta \\) can lead to increased efficiency and decreased accuracy. We will give more validation in §4.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Second, non-sparse locations of visual features \\(\\{C_3,C_4,C_5\\}\\) need to be aggregated across multiple scales to compensate for contexts. We consider the location of 1 in each scale's mask \\(M_{l}\\) as \\(I_{l} = \\{(x_{l}^{o},y_{l}^{o})\\} (o\\in \\{1,\\ldots ,S_{l}\\})\\) in which \\(S_{l}\\) is the non-sparsity of \\(M_{l}\\), i.e., total number of 1. In general, high-level features with low resolution and low-level features with high resolution are concatenated through up- or down-sampling in the feature pyramid. Thus, we use cross-scale projection to make up for the" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.089, + 0.9, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "Figure 3. The sparse computation in the sparse convolution operates by performing the convolution only when the center of a \\(3 \\times 3\\) kernel (with 1 padding) covers a non-zero value (e.g., only two non-zero values are in the sparse input features)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.305 + ], + "angle": 0, + "content": "non-sparse location. Specifically, when using a higher-level location \\(I_{l}\\) to query the lower-level feature \\(C_{l-1}\\), the additional non-sparse location is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.313, + 0.905, + 0.332 + ], + "angle": 0, + "content": "\\[\nI _ {l - 1} ^ {\\prime} = \\left\\{\\left(2 x _ {l} ^ {o} + i, 2 y _ {l} ^ {o} + j\\right) \\mid \\forall i, j \\in \\{0, 1 \\} \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.905, + 0.371 + ], + "angle": 0, + "content": "Then, update \\(I_{l-1}\\) as \\(I_{l-1} \\cup I_{l-1}'\\). Querying higher-level features using lower-level locations is reversed by dividing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Third, based on \\( I_{l} \\), dense features \\( C_l \\in \\mathbb{R}^{H_l \\times W_l \\times D_l} \\) are converted into sparse features \\( \\tilde{C}_l \\in \\mathbb{R}^{S_l \\times D_l} \\), which are then fed into the pre-converted sparse CNN layers for inference. The convolution in VL-PAN and heads are converted into sparse convolution (Figure 2). The computation of sparse convolution is illustrated in Figure 3. Take the \\( 1 \\times 1 \\) convolution kernel size as an example, the computation complexity with sparse convolution is \\( S_{l} / HW \\) of normal convolution. Thus, greater sparsity leads to lower computational complexity. Note that running sparse CNNs efficiently on general-purpose hardware (e.g., GPU) remains challenging. Thus, we leverage the existing high-performance inference engine (e.g., SpConv [9]) to accelerate sparse convolution computation. Additionally, we configure the sparse conversion as level-wise granularity (e.g., \\( P_3 \\) is sparse, while \\( P_4 \\) and \\( P_5 \\) remain dense), allowing for greater flexibility to adjust based on the hardware performance (\\$5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.64, + 0.645, + 0.657 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.905, + 0.787 + ], + "angle": 0, + "content": "We evaluate the effectiveness of the VISO pre-trained on large-scale satellite data and the efficiency of the sparse VISO guided by learned masks. We perform VISO in various downstream Earth observation tasks, including zero-shot detection on 3 benchmarks (§4.1), supervised detection on 2 benchmarks (§4.2), and object referring on one benchmark (§4.3). We further evaluate the speed of sparse VISO on the real-world satellite payload (§5)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.905, + 0.88 + ], + "angle": 0, + "content": "Implementation details. VISO is built upon the MMYOLO [8] and MMDetection [5] toolbox. For sparse inference, we implement a hook (SPHook) in MMDetection to convert VISO into a level-wise sparse model using SPconv [9]. The hook is easy to use, inserted at before_test_epoch\\(^2\\), with no extra overhead. In line with [7]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.808, + 0.9 + ], + "angle": 0, + "content": "2https://mmdetection.readthedocs.io" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23303" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.25 + ], + "angle": 0, + "content": "
TypeDatasetVocab.ImagesAnno.Format
Object Detection (OD)DOTAv2 [12]18614701100487OBB
DIOR-R [20]2023463192518OBB
FAIR1Mv2 [38]3743611626790OBB
NWPU-VHR-10 [6]108325582HBB
UCAS-AOD [49]2302024610OBB
RSOD [13]4186914604HBB
HRRSD [46]136805593780HBB
VisDrone [50]12298591152646HBB
Visual Grounding (VG)DIOR-RSVG [44]-1740230898HBB
RRSISD [43]-1740213015HBB
GeoChat [18]-3832868563OBB
VRSBench [19]-2961438589OBB
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.302 + ], + "angle": 0, + "content": "Table 1. Pre-training data. A list of datasets for pre-training VISO. HBB and OBB refer to horizontal and oriented bounding boxes, and Ann. denotes the annotations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.483, + 0.465 + ], + "angle": 0, + "content": "we provide three variants of VISO to meet different requirements, e.g., small (S), medium (M), and large (L). We utilize the open-source CLIP [32] text encoder with pretrained weights to encode the text input. All satellite images used for pre-training and evaluation are split and resized to \\(1024 \\times 1024\\). Unless specified, all evaluation results of VISO are presented with two parts: a dense inference result (upper bound) and an optimal sparse inference result, which balances AP and FLOPs by adjusting the sparse threshold \\(\\theta\\). More selections of \\(\\theta\\) will be discussed in §4.4." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.475, + 0.224, + 0.492 + ], + "angle": 0, + "content": "4.1. Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.498, + 0.483, + 0.619 + ], + "angle": 0, + "content": "Pre-training setup. We adopt the AdamW optimizer [28] with an initial learning rate of 0.0002 and weight decay of 0.025. VISO is pre-trained for 20 epochs on \\(8 \\times\\) NVIDIA H800 GPUs with a total batch size of 64. We use vanilla data augmentations to reduce training costs, including random resizing, rotation, and flipping without using mosaic augmentation. The text encoder is fine-tuned on our pretraining data with a learning factor of 0.01." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.624, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Pre-training data. We establish a large-scale dataset with 3.4M satellite region-text pairs to pre-train VISO. As listed in Table 1, we mainly collect 8 object detection datasets and 4 visual grounding datasets. We exclude the test sets of DOTAv2, FAIR1Mv2, and DIOR-RSVG, which are used for subsequent evaluation. The annotations for each dataset are formatted as region-text pairs, in which all the horizontal and oriented bounding boxes (regions) are represented as the corners of the quadrilateral and are uniformly converted to rotated bounding boxes when loading." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Zero-shot evaluation. After pre-training, we evaluate the pre-trained VISO on MAR20 [39], HRSC2016 [26], and VEDAI [33] benchmarks in a zero-shot manner following [45], as shown in Table 3. The MAR20 and HRSC2016 datasets contain more types of military aircraft and a diverse variety of ships. The VEDAI dataset contains 9 categories of vehicles on Earth. We compare the pre-trained VISO with recent state-of-the-art VLM mod" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.179 + ], + "angle": 0, + "content": "
Training DataFrozen?LAM?MAR20HRSC2016VEDAI
only DOTAv2✓✓90.76.834.3
OD✗✓90.268.949.9
OD, VG✓✓89.968.845.0
OD, VG✗✗89.270.348.3
OD, VG✗✓90.070.750.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.189, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Table 2. Ablations on pre-training data, frozen CLIP encoder, and language-guided attention module (LAM). We evaluate the zero-shot performance on MAR20, HRSC2016, and VEDAI by pre-training VISO-L with varying amounts of data, using either a frozen or fine-tuned text encoder, with or without LAM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.498 + ], + "angle": 0, + "content": "els [7, 22, 25, 48] and report AP and FLOPs. We preserve the pre-trained weights of these baselines (pre-trained on natural data) without any fine-tuning on satellite data. We find that VISO achieves remarkable performance in detecting Earth objects with \\(90.2\\%\\) AP for MAR20, \\(75.6\\%\\) AP for HRSC2016, and \\(50.2\\%\\) AP for VEDAI, while obtaining more than \\(20\\times\\) reduction in FLOPs. The results highlight the disparities between natural and satellite images, showing that existing VLMs cannot be directly applied to the Earth observation domain and incur high computational costs. Furthermore, we test sparse inference with an optimal threshold, which results in a further reduction in FLOPs with \\(\\leq 1\\%\\) AP drop, e.g., \\(2\\times\\) FLOPs reduction of VISO-S on MAR20, with a tiny AP drop, e.g., \\(0.2\\%\\). Please find more details of fair comparison in our appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Ablation study. Table 2 presents ablation studies to analyze VISO from three aspects: pre-training data, fine-tuning the CLIP encoder, and the proposed language-guided attention module. We mainly conduct ablations based on VISO-L and evaluate them on three benchmarks in a zero-shot manner. First, compared to the baseline (grey) trained on all pre-training data, reducing visual grounding (VG) and partial object detection (OD) datasets leads to \\(0.3 - 63.9\\%\\) AP drop. The decrease is due to the richer textual information from visual grounding datasets, enlarging the vocabulary of Earth objects. Second, we exploit two settings for the text encoder, CLIP, i.e., frozen and fine-tuned. Fine-tuning can improve AP by \\(0.1 - 5.2\\%\\) as large-scale satellite data helps align the text encoder with the Earth observation domain, which is not well covered in CLIP's training data. Third, the proposed language-guided attention module improves AP by \\(0.4 - 1.9\\%\\), highlighting its effectiveness in distinguishing and improving object-centric feature extraction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.681, + 0.804 + ], + "angle": 0, + "content": "4.2. Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Baselines. We consider 3 specialist detectors as baselines, all of which focus on efficiency and practicality (§2.2). We train the baseline models separately on the training set of DOTAv2 and FAIR1Mv2 for 80 epochs (YOLOv8) and 36 epochs (RTMDet and LSKNet) using 2 NVIDIA H800 GPUs with a total batch size of 16. We adopt the AdamW" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23304" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.891, + 0.266 + ], + "angle": 0, + "content": "
ModelBackboneParams(M) ↓MAR20HRSC2016VEDAI
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Pre-trained open-set models
GLIP-TSwin-T23248889.448848.44881.1
GroundingDINO-TSwin-T17246486.746436.84642.6
MM-GroundingDINO-TSwin-T17246486.646445.446416.1
YOLO-World-SYOLOv8-S134454.14410.9440.8
YOLO-World-MYOLOv8-M2911361.111312.51131.2
YOLO-World-LYOLOv8-L4722955.222920.92291.4
Zero-shot transferNo sparsity → Sparsity
VISO-SYOLOv8-S1344→2287.6→87.444→2659.7→59.044→1838.8→37.9
VISO-MYOLOv8-M29113→5990.2→90.1113→6875.6→74.6113→5546.6→46.2
VISO-LYOLOv8-L47229→13090.0→90.0229→13470.7→70.1229→11850.2→49.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.276, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Table 3. Zero-shot evaluation on MAR20, HRSC2016 and VEDAI test set. We compare VISO with existing VLMs on zero-shot transfer in terms of AP and FLOPs. No sparsity means the max accuracy of VISO and sparsity significantly reduces FLOPs with \\(\\leq 1\\%\\) AP drop." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.316, + 0.484, + 0.507 + ], + "angle": 0, + "content": "
ModelDOTAv2 (test)FAIR1Mv2 (test)
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Specialist models
YOLOv8-S4050.84044.8
YOLOv8-M10554.410547.4
YOLOv8-L21756.421749.2
RTMDet-S3858.73847.5
RTMDet-M10160.1/10149.7/
RTMDet-L20660.820651.4*
LSKNet-T12459.512546.9
LSKNet-S16161.1*16249.7/
Open-set modelsNo sparsity → Sparsity
VISO-S44→1956.8→50.644→2446.2→44.8
VISO-M113→6961.8→60.0/113→7650.9→50.0/
VISO-L229→12863.4→61.1*229→12752.3→51.5*
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Table 4. Supervised results on DOTAv2 and FAIR1Mv2. We compare VISO with specialist models on object detection in terms of AP and FLOPs. The AP of VISO-L with No sparsity outperforms SOTA baselines (2.3% on DOTAv2 and 0.9% on FAIR1Mv2). We use sparsity to reduce AP until comparable with baselines (AP, AP*, and AP/) and find that FLOPs are significantly reduced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.669 + ], + "angle": 0, + "content": "optimizer [28] and set an initial learning rate of 0.0002 for YOLOv8, 0.00025 for RTMDet, and 0.0001 for LSKNet. The same data augmentations as pre-training are applied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Supervised comparison with specialist models. We compare the proposed VISO with 3 close-set detectors, i.e., YOLOv8 with 3 variants, RTMDet with 3 variants, and LSKNet with 2 variants, ensuring a fair comparison with our 3 variants. Table 4 shows the results in terms of AP and FLOPs on two object detection benchmarks, i.e., the DOTAv2 and FAIR1Mv2 test sets. The results demonstrate that VISO outperforms the SOTA method, i.e., LSKNet-S, by increasing \\(2.3\\%\\) AP for DOTAv2 and \\(0.9\\%\\) AP for FAIR1Mv2. For sparse inference, we adjust the threshold to obtain comparable AP with LSKNet-S, i.e., \\(61.1\\%\\) for DOTAv2. We observe a significant drop in FLOPs of VISO-L from 229G to 128G, which is lower than the FLOPs of LSKNet-S, i.e., 161G. For FAIR1Mv2, we achieve both better FLOPs and AP than baselines. For example, VISO-M" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.316, + 0.906, + 0.451 + ], + "angle": 0, + "content": "
ModelParams(M) ↓FLOPs(G) ↓Pr@0.5(%) ↑Pr@0.7(%) ↑
Specialist models
TransVG12620469.755.2
VLTVG12819476.962.3
MGVLF13221976.861.7
LPVA13722875.5*60.8
Open-set modelsNo sparsity → Sparsity
VISO-S1344→2474.4→70.169.1→64.7
VISO-M29113→8476.1→75.5*72.0→71.5
VISO-L48229→16877.3→75.8*72.9→71.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.462, + 0.908, + 0.518 + ], + "angle": 0, + "content": "Table 5. Supervised results on DIOR-RSVG test set. We compare VISO with specialist models on object referring in terms of FLOPs, \\(\\mathrm{Pr}@\\mathrm{0.5}\\), and \\(\\mathrm{Pr}@\\mathrm{0.7}\\). Specialist models employ HBB and VISO adopts OBB for box format. Input size is \\(1024 \\times 1024\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.625 + ], + "angle": 0, + "content": "obtains a better AP with only 76G FLOPs, reducing \\(2.1 \\times\\) FLOPs than LSKNet-S and \\(1.3 \\times\\) than RTMDet-M. Furthermore, VISO, built upon YOLOv8, achieves extremely lower FLOPs than YOLOv8, e.g., reducing \\(1.7 - 2.1 \\times\\) FLOPs for S with comparable AP. Our approach can be extended to other detectors to attain even greater accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.679, + 0.65 + ], + "angle": 0, + "content": "4.3. Object referring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Baselines. We consider 4 specialist visual grounding models as baselines: TransVG [11], VLTVG [42], MGVLF [44], LPVA [21], all of which are built upon a neat transformer-based framework (§2.3). We train the baseline models on the DIOR-RSVG training set for 90 epochs using 2 NVIDIA H800 GPUs with a total batch size of 64. The learning rate and data augmentation follow the default settings of each method. We use two metrics: \\(\\mathrm{Pr}@\\mathbf{0.5}\\) and \\(\\mathrm{Pr}@\\mathbf{0.7}\\) to measure IOU precision \\(\\geq 50\\%\\), \\(70\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Supervised comparison with specialist models. We compare our VISO with specialist baselines on the DIOR-RSVG test set. The results are shown in Table 5. We find that VISO outperforms SOTA methods by increasing \\(0.4\\%\\) \\(\\mathrm{Pr}@\\mathrm{0.5}\\) and \\(10.6\\%\\) \\(\\mathrm{Pr}@\\mathrm{0.7}\\). We adjust the sparse threshold to obtain a comparable precision with baselines and observe a significant drop in FLOPs, e.g., reducing \\(8.5\\times\\) from 204G" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23305" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.331, + 0.081, + 0.671, + 0.093 + ], + "angle": 0, + "content": "AP FLOPs {1,0,0} {1,1,0} {1,1,1}" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.096, + 0.236, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.097, + 0.349, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.097, + 0.48, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.096, + 0.605, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.096, + 0.742, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.096, + 0.895, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.176, + 0.23, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.176, + 0.362, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.176, + 0.495, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.177, + 0.604, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.177, + 0.758, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.177, + 0.894, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.25, + 0.23, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.25, + 0.362, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.25, + 0.495, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.25, + 0.604, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.25, + 0.758, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.25, + 0.89, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 4. Sparsity analysis of VISO on 6 benchmarks. We conduct extensive experiments on sparse VISO using different sparse thresholds (X-axis) and record AP (left Y-axis) and FLOPs (right Y-axis). As the threshold increases, both AP and FLOPs decline, with FLOPs at a faster rate. We use validation sets of DOTAv2 and FAIR1Mv2 for this evaluation to avoid limits on test submissions." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.376, + 0.262, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.377, + 0.319, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.377, + 0.373, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.377, + 0.43, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.377, + 0.487, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.497, + 0.493, + 0.547 + ], + "angle": 0, + "content": "(a) Mask visualization of fine-grained supervision for each target objects: (A) \"baseball diamond\", (B) \"small vehicle\", (C) \"tennis court\", (D) \"helicopter\". From (A) to (D), the multi-scale masks \\(\\{X_{3}, X_{4}, X_{5}\\}\\) are arranged from the top (low-level feature) to the bottom (high-level feature)." + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.376, + 0.653, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.377, + 0.728, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.376, + 0.804, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.376, + 0.881, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.498, + 0.497, + 0.896, + 0.547 + ], + "angle": 0, + "content": "(b) The sparsity of \\(\\{C_4\\}\\) feature under different sparse thresholds. Left is the detection result (yellow-marked planes and white-marked helicopters) in DOTAv2. When the Earth observation task is specified as \"detect the plane\", masks only focus on planes and helicopters are irrelevant objects." + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.552, + 0.82, + 0.567 + ], + "angle": 0, + "content": "Figure 5. Mask visualizations of VISO. More visualization of detection results and masks is in our appendix." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.582, + 0.375, + 0.596 + ], + "angle": 0, + "content": "to 24G when \\(\\mathrm{Pr}@0.5\\) is \\(70.1\\%\\) of VISO-S." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.257, + 0.623 + ], + "angle": 0, + "content": "4.4. Sparsity analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.81 + ], + "angle": 0, + "content": "All the above evaluations list only the optimal selection of sparse thresholds and employ a full-level sparse model for sparse inference, in which \\( P_{3} \\), \\( P_{4} \\), and \\( P_{5} \\) are all sparse. We claim that adjusting the threshold \\( \\theta \\) and level-wise granularity can address different requirements for in-orbit inference, prioritizing faster or better. Figure 4 illustrates the trade-off between AP and FLOPs with different thresholds from 0.01 to 0.9. The results show a substantial efficiency improvement, i.e., a sharp drop in FLOPs at very low thresholds, which is attributed to the highly sparse model, guided by learned masks of small objects in satellite images, and its ability to distinctly differentiate target objects from others." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Besides, the level-wise granularity also impacts FLOPs and AP. We denote the level-wise sparse conversion for \\(\\{P_3,P_4,P_5\\}\\) as \\(\\triangleright \\{s_3,s_4,s_5\\} (s_{l\\in \\{3,4,5\\}}\\in \\{0,1\\})\\) . For example, \\(\\triangleright \\{1,0,0\\}\\) indicates the sparse \\(P_{3}\\) and dense \\(P_4\\) \\(P_{5}\\) . Figure 4 shows that FLOPs decline more quickly in \\(\\triangleright \\{1,1,1\\}\\) than in \\(\\triangleright \\{1,1,0\\}\\) and \\(\\triangleright \\{1,0,0\\}\\) , however, the AP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.643 + ], + "angle": 0, + "content": "drop for \\(\\triangleright \\{1, 1, 1\\}\\) and \\(\\triangleright \\{1, 1, 0\\}\\) is nearly identical, indicating that low-resolution features in \\(P_{5}\\) play a minor role in detecting small objects, primarily providing contextual information through a larger receptive field." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.659, + 0.667 + ], + "angle": 0, + "content": "4.5. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of VISO, we visualize the learned masks in Figure 5. In Figure 5a, the masks focus separately on the relevant objects rather than others. Furthermore, as features go deeper from \\( C_3 \\) to \\( C_5 \\), more salient contextual information is captured, more than just the outline of objects. Figure 5b illustrates the sparsity of masked features under different thresholds. As the threshold \\( \\theta \\) increases from 0.01 to 0.9, the focus narrows with minimal loss of essential features. Specifically, the task is to detect the \"plane\", not the \"helicopter\". Thus, the mask gradually becomes more focused around the \"plane\" as the threshold increases. Redundant features (e.g., black areas) are masked out during sparse inference. For example, \\( \\theta = 0.9 \\) can result in an extremely high sparsity of \\( 99.7\\% \\), preserving important features while significantly improving speed." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23306" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.249, + 0.485, + 0.32 + ], + "angle": 0, + "content": "Figure 6. The acceleration performance of sparse convolution (SPConv) and normal convolution (Conv). The comparison is across different feature scales of three variants. SPConv provides significant acceleration for large features compared to Conv. However, for small features, Conv is already well-optimized." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.329, + 0.285, + 0.347 + ], + "angle": 0, + "content": "5. Real-world Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.462 + ], + "angle": 0, + "content": "Following [17], we use frames per second (FPS) to measure the actual speed of detection on a real-world embedded GPU, NVIDIA Jetson Xavier NX [30], which serves as a payload on Chaohu-1 SAR satellite [37] launched in 2022. In this section, we consider a more common scenario in Earth observation missions: single-object detection, i.e., only one target or one category target is of interest." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.573 + ], + "angle": 0, + "content": "Implementation details. NX supports three optimized power budgets, e.g., 10W, 15W, and 20W. We pre-confine the mode as 10W with a 4-core CPU of \\(1900\\mathrm{MHz}\\) for our evaluation. The system of our testbed is equipped with Jet-Pack 5.1.3, CUDA 11.4, cuDNN 8.6.0, Pytorch 1.11.0, and SPconv 2.3.6. We test all models without extra acceleration mechanisms, e.g., FP16 or TensorRT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Sparse configuration. We use SPconv [9] as the high-performance sparse inference engine for onboard deployment. As shown in Figure 6, we test the acceleration performance of SPconv with different feature scales on our testbed. Note that the performance varies across different hardware. We consider the commonly used \\(3 \\times 3\\) kernel and compare the average latency of sparse convolution at different sparsity with that of normal convolution. The sparsity is defined as the ratio of the number of zeros in the feature to the total area of the feature. We observe that the larger the feature scale, the acceleration of SPConv is more pronounced. Especially for \\(\\mathrm{L}\\{P_3, P_4\\}\\), \\(\\mathrm{M}\\{P_3, P_4, P_5\\}\\) and \\(\\mathrm{S}\\{P_3\\}\\) of VISO, a sparsity of more than \\(25\\%\\) can achieve acceleration, which can be easily attained in our highly sparse model when detecting small objects in satellite images, i.e., the smaller the object, the faster the detection. Thus, we configure the level-wise granularity to allow for more flexible sparse conversion, targeting different feature scales and hardware platforms." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Baselines. We consider 2 baselines that are widely used in the industrial domain [17, 41], including a commonly used" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.898, + 0.231 + ], + "angle": 0, + "content": "
MethodFLOPs(G) ↓FPS ↑AP(%) ↑Sparsity(%) ↓
Dense models
YOLOv8-S404.043.6-
YOLOv8-M1052.032.0-
YOLOv8-L2171.125.3-
Sparse models
QueryDet2972.334.838
VISO-S1611.148.388
VISO-M495.433.582
VISO-L1103.038.176
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.241, + 0.905, + 0.299 + ], + "angle": 0, + "content": "Table 6. Real-world test on NX. We compare VISO with YOLOv8 and QueryDet. All variants of VISO are full-level sparse, and we set \\(\\theta = 0.1\\) for S, 0.25 for M, and 0.2 for L. We also list the average sparsity of all sparse features in sparse models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.409 + ], + "angle": 0, + "content": "YOLOv8 [17] detector mentioned above and a sparse detector QueryDet [41]. QueryDet introduces a higher-resolution \\(P_{2}\\) to improve small object detection and a query mechanism to make high-resolution heads sparse. We reproduce QueryDet using default settings and train on the DOTAv2 training set with the size of \\(1024 \\times 1024\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.909, + 0.603 + ], + "angle": 0, + "content": "Speed testing. In Table 6, we detect the \"helicopter\" in DOTAv2 and compare VISO with YOLOv8 and QueryDet in terms of AP, FLOPs, FPS, and sparsity. VISO improves FPS by \\(2.7 - 2.8 \\times\\) than YOLOv8 and \\(4.8 \\times\\) than QueryDet under a better AP. VISO achieves significantly higher sparsity due to two folds: (1) The sparse computation of VISO occurs earlier, after the backbone, with sparse necks and heads, rather than only sparse heads in QueryDet. Also, VISO does not introduce higher-resolution features. (2) When detecting the target in irrelevant images, masks are unfocused, e.g., detect the helicopter in Figure 5a(D), and thus can early exit to avoid further inference." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.63, + 0.634, + 0.645 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We present VISO, an effective and efficient solution for in-orbit real-time object detection. VISO empowers vision-language models with language-guided mask learning and sparse inference. In this paper, we introduce a language-guided attention module to enhance salient object-centric features while suppressing irrelevant regions in satellite images. We pre-train VISO with large-scale satellite data in the form of region-text pairs with contrastive loss and additional mask learning, thereby endowing VISO with a strong capability of in-orbit open-vocabulary detection. Furthermore, the learned masks of VISO can suppress irrelevant regions to enable highly sparse inference and accelerate speed. Extensive experiments can demonstrate the superiority of VISO in both detection accuracy and speed, especially the significant speed-up on real-world satellites' embedded GPUs is insightful for future industrial deployment." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23307" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.148 + ], + "angle": 0, + "content": "This work is supported by the Center of High Performance Computing, Tsinghua University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.256 + ], + "angle": 0, + "content": "[1] Frédéric Achard, Hans-Jürgen Stibig, Hugh D Eva, Erik J Lindquist, Alexandre Bouvet, Olivier Arino, and Philippe Mayaux. Estimating tropical deforestation from earth observation data. Carbon Management, 1(2):271-287, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.259, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[2] Ali Bahri, Sina Ghofrani Majelan, Sina Mohammadi, Mehrdad Noori, and Karim Mohammadi. Remote sensing image classification via improved cross-entropy loss and transfer learning strategy based on deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 17 (6):1087–1091, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[3] Hanoona Bangalath, Muhammad Maaz, Muhammad Uzair Khattak, Salman H Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. Advances in Neural Information Processing Systems, 35:33781-33794, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.458 + ], + "angle": 0, + "content": "[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[6] Gong Cheng, Junwei Han, Peicheng Zhou, and Lei Guo. Multi-class geospatial object detection and geographic image classification based on collection of part detectors. ISPRS Journal of Photogrammetry and Remote Sensing, 98: 119-132, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xing-gang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[8] MMYOLO Contributors. Mmyolo: Openmmlab yolo series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022. Accessed: 2023-03-03. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[9] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.855 + ], + "angle": 0, + "content": "[10] Laura Crocetti, Matthias Forkel, Milan Fischer, František Jurečka, Aleks Grlj, Andreas Salentinig, Miroslav Trnka, Martha Anderson, Wai-Tim Ng, Žiga Kokalj, et al. Earth observation for agricultural drought monitoring in the pannonian basin (southeastern europe): current state and future directions. Regional Environmental Change, 20:1-17, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.187, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 1769-1779, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[12] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44 (11):7778-7796, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[13] Qiqi Ding, Peng Li, Xuefeng Yan, Ding Shi, Luming Liang, Weiming Wang, Haoran Xie, Jonathan Li, and Mingqiang Wei. Cf-yolo: Cross fusion yolo for object detection in adverse weather with a high-quality real snow dataset. IEEE Transactions on Intelligent Transportation Systems, 24(10): 10749-10759, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[14] Bowei Du, Yecheng Huang, Jiaxin Chen, and Di Huang. Adaptive sparse convolutional networks with global context enhancement for faster object detection on drone images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13435-13444, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[15] Max Ghiglione and Vittorio Serra. Opportunities and challenges of ai on satellite processing units. In Proceedings of the 19th ACM international conference on computing Frontiers, pages 221-224, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[16] Jiaxing Huang, Jingyi Zhang, Kai Jiang, and Shijian Lu. Open-vocabulary object detection via language hierarchy. arXiv preprint arXiv:2410.20371, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLOv8, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[18] Kartik Kuckreja, Muhammad S. Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad S. Khan. Geochat: Grounded large vision-language model for remote sensing. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[19] Jian Li, Xiang Ding and Mohamed Elhoseiny. Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding. arXiv:2406.12384, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.66 + ], + "angle": 0, + "content": "[20] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[21] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[22] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[23] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in neural information processing systems, 33:21002-21012, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23308" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF international conference on computer vision, pages 16794-16805, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[26] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, pages 324-331. SciTePress, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[29] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[30] NVIDIA. Jetson Xavier NX Developer Kit. https://developer.nvidia.com/embedded/jetson-xavier-nx-devkit, 2024. Online; accessed 14-March-2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[31] George P Petropoulos, Prashant K Srivastava, Maria Piles, and Simon Pearson. Earth observation-based operational estimation of soil moisture and evapotranspiration for agricultural crops in support of sustainable water management. Sustainability, 10(1):181, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[33] Sebastien Razakarivony and Frederic Jurie. Vehicle detection in aerial imagery: A small target detection benchmark. Journal of Visual Communication and Image Representation, 34:187-203, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[34] Gupta Ritwik et al. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[35] Daniel Saaristo. Shoot from the stars: Startup provides early detection of wildfires from space. https://blogs.nvidia.com/blog/ororatech-wildfires-from-space/, 2023.1,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[36] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Spacety. Chaohu-1 SAR satellite. https://www.satelliteevolution.com/post/two-new-satellites-launched-including-the-2nd-sar-satellite-with-improved-capabilities,2022.8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[38] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[39] YU Wenqi, CHENG Gong, WANG Meijun, YAO Yanqing, XIE Xingxing, YAO Xiwen, and HAN Junwei. Mar20: A benchmark for military aircraft recognition in remote sensing images. National Remote Sensing Bulletin, 27(12):2688-2696, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[40] Michael A Wulder, David P Roy, Volker C Radeloff, Thomas R Loveland, Martha C Anderson, David M Johnson, Sean Healey, Zhe Zhu, Theodore A Scambos, Nima Pahlevan, et al. Fifty years of landsat science and impacts. Remote Sensing of Environment, 280:113195, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[41] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13668-13677, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[42] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[43] Zhenghang Yuan, Lichao Mou, Yuansheng Hua, and Xiao Xiang Zhu. Rrasis: Referring remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[44] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[45] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multimodal large language model for multisensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[46] Yuanlin Zhang, Yuan Yuan, Yachuang Feng, and Xiaoqiang Lu. Hierarchical and robust convolutional neural network for very high-resolution remote sensing object detection. IEEE Transactions on Geoscience and Remote Sensing, 57 (8):5535-5548, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[47] Ziyuan Zhang, Han Qiu, Maosen Zhang, Jun Liu, Bin Chen, Tianwei Zhang, and Hewu Li. COSMIC: Compress satellite images efficiently via diffusion compensation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[48] Xiangyu Zhao, Yicheng Chen, Shilin Xu, Xiangtai Li, Xinjiang Wang, Yining Li, and Haian Huang. An open and comprehensive pipeline for unified object grounding and detection. arXiv preprint arXiv:2401.02361, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Haigang Zhu, Xiaogang Chen, Weiqun Dai, Kun Fu, Qixiang Ye, and Jianbin Jiao. Orientation robust object detection in aerial images using deep convolutional neural network. In 2015 IEEE international conference on image processing (ICIP), pages 3735-3739. IEEE, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23309" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.161 + ], + "angle": 0, + "content": "[50] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 5" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23310" + } + ] +] \ No newline at end of file diff --git a/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_origin.pdf b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..779a6222c796faa0257688310e8864e0f00eba79 --- /dev/null +++ b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/bca87654-a43f-4442-9368-eee092911b00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59f13224e25147077734781190dd1ad391fddc6e2f5b2f64f5f0c79b40d76f4b +size 7675651 diff --git a/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/full.md b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac0c1245296beb72690ce5962fef5b8fa600f935 --- /dev/null +++ b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/full.md @@ -0,0 +1,332 @@ +# VISO: Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference + +Meiqi Wang Tsinghua University wang-mq22@mails.tsinghua.edu.cn + +Han Qiu Tsinghua University qiuhan@tsinghua.edu.cn + +# Abstract + +In-orbit object detection is essential for Earth observation missions on satellites equipped with GPUs. A promising approach is to use pre-trained vision-language modeling (VLM) to enhance its open-vocabulary capability. However, adopting it on satellites poses two challenges: (1) satellite imagery differs substantially from natural images, and (2) satellites' embedded GPUs are insufficient for complex models' inference. We reveal their lack of a crucial prior: in-orbit detection involves identifying a set of known objects within a cluttered yet monotonous background. Motivated by this observation, we propose VISO, a Vision-language Instructed Satellite Object detection model that focuses on object-specific features while suppressing irrelevant regions through language-guided mask learning. After pre-training on a large-scale satellite dataset with 3.4M region-text pairs, VISO enhances object-text alignment and object-centric features to improve detection accuracy. Also, VISO suppresses irrelevant regions, enabling highly sparse inference to accelerate speed on satellites. Extensive experiments show that VISO without sparsity outperforms state-of-the-art (SOTA) VLMs in zero-shot detection by increasing $34.1\%$ AP and reducing $27\times$ FLOPs, and surpasses specialist models in supervised object detection and object referring by improving $2.3\%$ AP. When sparsifying VISO to a comparable AP, FLOPs can be greatly reduced by up to $8.5\times$ . Real-world tests reveal that VISO achieves a $2.8-4.8\times$ FPS speed-up on satellites' embedded GPUs1. + +# 1. Introduction + +Recently, embedded GPUs are deployed in space [15], enabling in-orbit object detection to support latency-sensitive Earth observation tasks such as disaster monitoring [34] and wildfire detection [35]. In-orbit detection requires detecting flexible in-orbit tasks and responding to unforeseen emergencies beyond predefined closed-set categories. We observe that vision-language models (VLMs) [22] facilitate + +![](images/df6a0c1c5dffa0f4fc8bd879315f523f76ac7691532ef304a520b645f9e50e9c.jpg) +Figure 1. Earth observation images and natural images are different. Earth observation images typically contain very small objects. On the right is the size ratio of objects to images. We choose 5 categories in DOTA: plane (pl), storage tank (st), roundabout (rb), small vehicle (sv), baseball diamond (bd), and 5 categories in COCO: person (pr), cat (ct), car (cr), bed (bd), bottle (btl). + +![](images/812175ec29974e644e001667763d006274e13eecfd8385ee5e82290621cab7e6.jpg) + +open-vocabulary detection (OVD) [7], allowing for detecting novel objects with the guidance of natural language. + +However, most existing VLMs [22, 25, 48] are designed for natural images and suffer from high computational overhead due to complex backbones (e.g., Swin [27] used in GLIP [22]). Although YOLO-World [7] provides a practical solution for time-critical applications, directly applying it to satellite data remains challenging due to the disparities between terrestrial natural images and satellite images, as shown in Figure 1. Satellite images are typically captured from a bird's eye view with a long distance from the ground. They often contain a few tiny objects (e.g., ships) relevant to referring detection tasks in orbit (e.g., detect the ship on the lower left), surrounded by vast irrelevant objects and backgrounds (e.g., buildings, roads, etc.), which makes detection difficult [24, 44]. We consider these priors and address two important goals. (1) Accurate detection requires distinguishing object-centric visual features from irrelevant features while incorporating useful contextual information. (2) Fast inference can be achieved by omitting large irrelevant regions, consisting of non-essential backgrounds and objects, to reduce redundant computation. + +In this paper, we achieve the above two goals by proposing VISO, a Vision-language Instructed Satellite Object detection model. Our key insight is that rich semantic textual + +embeddings of Earth observation tasks can help a detector distinguish different objects at its shallow layers (e.g., Figure 5a). Thus, we design a multi-scale mask learning to embed strongly aligned knowledge of Earth objects into multiscale visual features by pretraining on a large-scale satellite region-text pair. The learnable masks let VISO focus only on relevant areas at an early inference stage, ignoring irrelevant objects or cluttered backgrounds, thereby generating highly sparse features, which forward can be optimized by sparse computation (e.g., sparse convolution [14, 41]). + +We conduct extensive experiments to validate the effectiveness and efficiency of VISO. VISO surpasses 4 VLM baselines in zero-shot open-vocabulary detection on 3 benchmarks, obtaining $34.1\%$ AP increase and $27\times$ FLOPs reduction. Furthermore, VISO outperforms 3 SOTA specialist models on 2 object detection benchmarks by $2.3\%$ AP and 4 grounding models on one object referring benchmark by $0.4\%$ precision. By applying sparsity and aligning AP with baselines, VISO can achieve up to $8.5\times$ reduction in FLOPs. Real-world tests further demonstrate a $2.8 - 4.8\times$ FPS speed-up on satellites' embedded GPUs. + +# 2. Background and Related Work + +# 2.1. GPUs on Satellites + +Traditional Earth observation missions [40] rely on a pipeline in which satellites take photos and then send them back to ground stations for analysis like forestry [1] or agriculture [10, 31]. This leads to a substantial requirement for data transmission from satellites to the ground [47] and cannot fulfill latency-sensitive tasks like disaster detection [35]. + +Recently, embedded GPUs are deployed on satellites for Earth observation missions [47]. In-orbit DNN inference can process real-time tasks, especially object detection, and send only results back to ground stations. For example, OroraTech has launched AI nanosatellites with NVIDIA Jetson Xavier NX for in-orbit wildfire detection [35]. + +Although deploying GPUs in orbit is a promising approach, satellites are still constrained by size and energy supply to support sophisticated GPUs. For instance, NVIDIA Jetson Xavier NX, a widely-used in-orbit GPU, can only support 10-20 Watts [35] such that accelerating the in-orbit DNN inference becomes critical. + +# 2.2. Remote Sensing Object Detection + +Recent research [24, 29, 41] has been dedicated to object detection specified for remote sensing, which typically addresses the issue of small objects in satellite images. They propose two types of approaches: utilizing large convolution kernels to introduce a large receptive field and leveraging high-resolution images or features to preserve small objects, both of which lead to high computational overhead. + +Thus, these methods pursue a lightweight solution. + +LSKNet [24] employs a spatial selective mechanism to decompose the large kernel into a sequence of depthwise kernels. RTMDet [29] uses large-kernel depth-wise convolutions in the backbone and neck while reducing the number of building blocks as compensation to reduce model depth. QueryDet [41] introduces a higher-resolution feature scale in feature pyramids and sparsifies the high-resolution heads through cascaded queries to speed up. However, we find that these methods either aim to compensate for the extra computation or have limited acceleration potential only within heads. In this paper, we emphasize fully releasing the acceleration capability of the detection of small objects. + +# 2.3. Vision-language Model for Object Detection + +Open-vocabulary detection. Vision-language model (VLM) is a promising paradigm to empower open-vocabulary detection (OVD) according to open-set demands. Existing VLMs [22, 25, 48] unify OVD and visual grounding (or object referring) by reformulating the classification head as a text-contrastive head and designing vision-language fusion mode. However, these VLMs rely on complex backbones (e.g., Swin [27]), which are hard to deploy on resource-constrained devices. YOLO-World [7] address this by using the efficient YOLOv8 [17] as the backbone to enable real-time OVD. Despite this, these methods are pre-trained on natural images and the fusion lacks prior knowledge specific to satellite images, e.g., small objects. + +Visual grounding. Specifically for visual grounding, some research establishes a neat multi-modal correspondence based on a cross-modal transformer and directly regresses coordinates. TransVG [11] replaces the complex fusion modules with a simple stack of transformer encoder layers. VLTVG [42] introduces an additional bidirectional visual-linguistic encoder before the cross-model decoder. MGVLF [44] considers scale variations in remote sensing and takes advantage of multi-scale visual features and multi-granularity textual embeddings. LPVA [21] makes better multi-modal interaction in the visual feature extraction process. However, these transformer-based grounding models will increase computational costs and will somehow degrade object detection without bounding box regression. + +Region-text alignment. Various supervision strategies like [3, 16, 36] are proposed to enhance object-centric vision-language (i.e., region-text) alignment instead of coarse-grained image-level supervision [32] to make localization more precise. In this paper, we leverage prior knowledge from the Earth observation domain to enhance object-centric alignment and aim to improve efficiency in orbit. + +# 3. Method + +Motivated by the aforementioned insights, in this section, we propose VISO, an efficient VLM tailored for Earth observation tasks. VISO can be seamlessly deployed on in + +![](images/2e90e046a39a6a409c2dc24cab67d57a6dccb2c8b00fe8f57b4b01fb1545cbe2.jpg) +Figure 2. Overview of VISO. The black dashed border represents pre-training with additional supervised mask learning. The red dashed border represents satellite deployment with sparse inference in VL-PAN and heads, along with an offline vocabulary omitting the text encoder like [7]. The language-guided attention module inputs textual embeddings and visual features and outputs multi-scale masks. The masks are learned in the training phase through supervised mask learning and serve as guidance to convert dense features into sparse ones. + +orbit satellites with GPUs for Earth observation tasks like open-vocabulary detection and object referring. + +First, we introduce a multi-scale language-guided attention module to effectively enhance object-centric features relevant to language expressions while suppressing the irrelevant features (§3.1). Second, we perform a large-scale pre-training on satellite region-text pairs with deep supervision to enrich the knowledge of Earth (§3.2). Third, we convert the pre-trained dense models into highly sparse ones through multi-scale mask aggregation, ensuring minimal accuracy loss while improving inference efficiency (§3.3). + +# 3.1. Model Architecture + +The architecture of VISO is illustrated in Figure 2. The proposed language-guided attention module is integrated with multi-scale visual features at an early stage. Though the current implementation is built upon the lightweight open-vocabulary detector, YOLO-World [7], the integration can be generalized to any other detector. + +Open-vocabulary detector. VISO uses YOLO-World [7] as its base detector. The detector contains a backbone (i.e., CSPDarkNet, details can refer to [4]) to extract multi-scale visual features $C_l \in \mathbb{R}^{H_l \times W_l \times D_l}, l \in 3, 4, 5$ from the image $I$ , a pre-trained text encoder (i.e., CLIP [32]) to encode the text input $T$ into textual embeddings $\mathbf{W}_{j \in \{1, \dots, N\}} \in \mathbb{R}^{N \times D}$ , a vision-language path aggregation network (VL-PAN) for multi-scale feature pyramids $\{P_3, P_4, P_5\}$ to fuse textual embeddings and visual features, a decoupled head to regress bounding box $\{B_k\}_{k=1}^K$ and object embeddings, and a text contrastive head to obtain object-text (region-text) similarity $\{s_{k,j}\}_{k=1}^K$ between object embeddings and textual embeddings. $N$ denotes the number of categories or expressions. $D$ denotes the dimension of textual embeddings. $K$ denotes the number of object predictions. We modify the regression output dimension from 4 to 5 to add a rotation angle prediction, enabling the commonly used ori + +ented object detection in Earth observation [29]. More details on VL-PAN and text contrastive head can refer to [7]. + +Language-guided attention module. We observe that the textual embeddings $\mathbf{W}$ possess rich semantic prior knowledge and exhibit strong alignment with visual features due to large-scale pre-training [2]. They can be leveraged to differentiate the visual features of different objects and backgrounds early in the detector, rather than relying solely on the final text contrastive head. Thus, we introduce a language-guided attention module immediately following the backbone. Specifically, we adopt one projection layer $l_{p}$ and the max-sigmoid attention to multi-scale visual features $C_l\in \mathbb{R}^{H_l\times W_l\times D_l},l\in 3,4,5.$ The projection layer projects textual embeddings to the dimension of each scale. The max-sigmoid attention generates multi-scale masks $X_{l}\in$ $\mathbb{R}^{H_l\times W_l\times 1}(l\in \{3,4,5\})$ by the cross-modal aggregation between textual and visual features: + +$$ +\left\{ \begin{array}{l} M _ {l j} = C _ {l} \mathbf {W} _ {\mathbf {j}} ^ {T}, M _ {l j} \in \mathbb {R} ^ {H _ {l} \times W _ {l} \times N} \\ X _ {l} = \delta \left(\max _ {j \in \{1, \dots , N \}} M _ {l j}\right) \end{array} \right. \tag {1} +$$ + +$\delta$ indicates the sigmoid function. We note that accurately detecting objects in satellite images often requires a wide range of contextual information due to the scale variations of Earth objects [29]. In general, low-level (high-resolution) features benefit small objects, while high-level features with large receptive fields benefit larger objects. Thus, incorporating multi-scale features to generate masks is crucial for preserving diverse receptive fields. + +# 3.2. Pre-training Formulation + +Region-text contrastive learning. We follow [7] and reformulate the ground-truth annotations as $M$ region-text pairs $\{B_i, t_i\}_{i=1}^M$ , where $t_i$ is the text for the object region $B_i$ . $t_i$ can be the category name for object detection or the referring expression for object referring. VISO inputs image $I$ and texts $T$ and outputs K object predictions $\{B_k, s_k\}_{k=1}^K$ . + +We follow [29] and leverage the dynamic soft label assignment to match predictions with ground-truth annotations. The text index of $t_i$ in $T$ is assigned for each positive prediction as a classification label. The assigned label and the predicted similarity are used to construct the region-text contrastive loss $\mathcal{L}_{\mathrm{con}}$ using quality focal loss [23]. In addition, rotated IoU loss [29] is adopted for oriented bounding box regression, and the basic training loss is defined as: $\mathcal{L}_1 = \mathcal{L}_{\mathrm{con}} + \lambda \mathcal{L}_{\mathrm{iou}}$ while $\lambda$ is a balancing coefficient. + +Fine-grained semantic deep supervision. To learn masks $X_{l}$ , we explore a simple fine-grained deep supervision strategy. We assign the ground truth bitmap $\mathcal{B}_{lj} \in \{0,1\}^{H_l \times W_l \times 1} (l \in \{3,4,5\}, j \in \{1,\dots,N\})$ for each category's $M_{lj}$ . The bitmap of each category is generated from all the object regions belonging to this category and scaled across multiple levels to represent diverse receptive fields. Specifically, the pixels located within the regions are valued at 1, and others are 0. We define the deep supervision's loss function as: + +$$ +\mathcal {L} _ {2} = \frac {1}{L} \sum_ {l = 1} ^ {L} \frac {1}{N} \sum_ {j = 1} ^ {N} \mathcal {H} \left(M _ {l j}, \mathcal {B} _ {l j}\right) \tag {2} +$$ + +$\mathcal{H}$ denotes the binary cross entropy function. Training the learnable masks is consistent with the basic training, and the total training loss is defined as $\mathcal{L} = \mathcal{L}_1 + \alpha \cdot \mathcal{L}_2$ . $\alpha$ is a balancing coefficient. We visualize the masks in Figure 5. + +# 3.3. Sparse Accelerated Inference + +The multi-scale masks generated at the early stage can serve as effective guidance for enabling sparse computation in subsequent layers, e.g., neck and head layers. For instance, when detecting small vehicles in Figure 5a, the masks are centered on the top right, indicating that the remaining areas can be sparsified to avoid redundant further inference. Note that we only consider lightweight CNN-based detectors here. Offline sparse deployment involves three steps. + +First, the real-value masks $X_{l}$ need to be converted to binary masks $M_{l} \in \{0,1\}^{H_{l} \times W_{l} \times 1} (l \in \{3,4,5\})$ by hard threshold $\theta$ , in which values greater than $\theta$ are set to 1 and others are set to 0. Therefore, adjusting $\theta$ can optimize the trade-off between accuracy and efficiency. Greater sparsity with larger $\theta$ can lead to increased efficiency and decreased accuracy. We will give more validation in §4.4. + +Second, non-sparse locations of visual features $\{C_3,C_4,C_5\}$ need to be aggregated across multiple scales to compensate for contexts. We consider the location of 1 in each scale's mask $M_{l}$ as $I_{l} = \{(x_{l}^{o},y_{l}^{o})\} (o\in \{1,\ldots ,S_{l}\})$ in which $S_{l}$ is the non-sparsity of $M_{l}$ , i.e., total number of 1. In general, high-level features with low resolution and low-level features with high resolution are concatenated through up- or down-sampling in the feature pyramid. Thus, we use cross-scale projection to make up for the + +![](images/ce1128fc88bfca58ce1093b47f48692aafdcd5f8232ccc6bc961c9d63e05c34c.jpg) +Figure 3. The sparse computation in the sparse convolution operates by performing the convolution only when the center of a $3 \times 3$ kernel (with 1 padding) covers a non-zero value (e.g., only two non-zero values are in the sparse input features). + +non-sparse location. Specifically, when using a higher-level location $I_{l}$ to query the lower-level feature $C_{l-1}$ , the additional non-sparse location is denoted as: + +$$ +I _ {l - 1} ^ {\prime} = \left\{\left(2 x _ {l} ^ {o} + i, 2 y _ {l} ^ {o} + j\right) \mid \forall i, j \in \{0, 1 \} \right\} \tag {3} +$$ + +Then, update $I_{l-1}$ as $I_{l-1} \cup I_{l-1}'$ . Querying higher-level features using lower-level locations is reversed by dividing. + +Third, based on $I_{l}$ , dense features $C_l \in \mathbb{R}^{H_l \times W_l \times D_l}$ are converted into sparse features $\tilde{C}_l \in \mathbb{R}^{S_l \times D_l}$ , which are then fed into the pre-converted sparse CNN layers for inference. The convolution in VL-PAN and heads are converted into sparse convolution (Figure 2). The computation of sparse convolution is illustrated in Figure 3. Take the $1 \times 1$ convolution kernel size as an example, the computation complexity with sparse convolution is $S_{l} / HW$ of normal convolution. Thus, greater sparsity leads to lower computational complexity. Note that running sparse CNNs efficiently on general-purpose hardware (e.g., GPU) remains challenging. Thus, we leverage the existing high-performance inference engine (e.g., SpConv [9]) to accelerate sparse convolution computation. Additionally, we configure the sparse conversion as level-wise granularity (e.g., $P_3$ is sparse, while $P_4$ and $P_5$ remain dense), allowing for greater flexibility to adjust based on the hardware performance (\$5). + +# 4. Experiments + +We evaluate the effectiveness of the VISO pre-trained on large-scale satellite data and the efficiency of the sparse VISO guided by learned masks. We perform VISO in various downstream Earth observation tasks, including zero-shot detection on 3 benchmarks (§4.1), supervised detection on 2 benchmarks (§4.2), and object referring on one benchmark (§4.3). We further evaluate the speed of sparse VISO on the real-world satellite payload (§5). + +Implementation details. VISO is built upon the MMYOLO [8] and MMDetection [5] toolbox. For sparse inference, we implement a hook (SPHook) in MMDetection to convert VISO into a level-wise sparse model using SPconv [9]. The hook is easy to use, inserted at before_test_epoch $^2$ , with no extra overhead. In line with [7], + +
TypeDatasetVocab.ImagesAnno.Format
Object Detection (OD)DOTAv2 [12]18614701100487OBB
DIOR-R [20]2023463192518OBB
FAIR1Mv2 [38]3743611626790OBB
NWPU-VHR-10 [6]108325582HBB
UCAS-AOD [49]2302024610OBB
RSOD [13]4186914604HBB
HRRSD [46]136805593780HBB
VisDrone [50]12298591152646HBB
Visual Grounding (VG)DIOR-RSVG [44]-1740230898HBB
RRSISD [43]-1740213015HBB
GeoChat [18]-3832868563OBB
VRSBench [19]-2961438589OBB
+ +Table 1. Pre-training data. A list of datasets for pre-training VISO. HBB and OBB refer to horizontal and oriented bounding boxes, and Ann. denotes the annotations. + +we provide three variants of VISO to meet different requirements, e.g., small (S), medium (M), and large (L). We utilize the open-source CLIP [32] text encoder with pretrained weights to encode the text input. All satellite images used for pre-training and evaluation are split and resized to $1024 \times 1024$ . Unless specified, all evaluation results of VISO are presented with two parts: a dense inference result (upper bound) and an optimal sparse inference result, which balances AP and FLOPs by adjusting the sparse threshold $\theta$ . More selections of $\theta$ will be discussed in §4.4. + +# 4.1. Pre-training + +Pre-training setup. We adopt the AdamW optimizer [28] with an initial learning rate of 0.0002 and weight decay of 0.025. VISO is pre-trained for 20 epochs on $8 \times$ NVIDIA H800 GPUs with a total batch size of 64. We use vanilla data augmentations to reduce training costs, including random resizing, rotation, and flipping without using mosaic augmentation. The text encoder is fine-tuned on our pretraining data with a learning factor of 0.01. + +Pre-training data. We establish a large-scale dataset with 3.4M satellite region-text pairs to pre-train VISO. As listed in Table 1, we mainly collect 8 object detection datasets and 4 visual grounding datasets. We exclude the test sets of DOTAv2, FAIR1Mv2, and DIOR-RSVG, which are used for subsequent evaluation. The annotations for each dataset are formatted as region-text pairs, in which all the horizontal and oriented bounding boxes (regions) are represented as the corners of the quadrilateral and are uniformly converted to rotated bounding boxes when loading. + +Zero-shot evaluation. After pre-training, we evaluate the pre-trained VISO on MAR20 [39], HRSC2016 [26], and VEDAI [33] benchmarks in a zero-shot manner following [45], as shown in Table 3. The MAR20 and HRSC2016 datasets contain more types of military aircraft and a diverse variety of ships. The VEDAI dataset contains 9 categories of vehicles on Earth. We compare the pre-trained VISO with recent state-of-the-art VLM mod + +
Training DataFrozen?LAM?MAR20HRSC2016VEDAI
only DOTAv2✓✓90.76.834.3
OD✗✓90.268.949.9
OD, VG✓✓89.968.845.0
OD, VG✗✗89.270.348.3
OD, VG✗✓90.070.750.2
+ +Table 2. Ablations on pre-training data, frozen CLIP encoder, and language-guided attention module (LAM). We evaluate the zero-shot performance on MAR20, HRSC2016, and VEDAI by pre-training VISO-L with varying amounts of data, using either a frozen or fine-tuned text encoder, with or without LAM. + +els [7, 22, 25, 48] and report AP and FLOPs. We preserve the pre-trained weights of these baselines (pre-trained on natural data) without any fine-tuning on satellite data. We find that VISO achieves remarkable performance in detecting Earth objects with $90.2\%$ AP for MAR20, $75.6\%$ AP for HRSC2016, and $50.2\%$ AP for VEDAI, while obtaining more than $20\times$ reduction in FLOPs. The results highlight the disparities between natural and satellite images, showing that existing VLMs cannot be directly applied to the Earth observation domain and incur high computational costs. Furthermore, we test sparse inference with an optimal threshold, which results in a further reduction in FLOPs with $\leq 1\%$ AP drop, e.g., $2\times$ FLOPs reduction of VISO-S on MAR20, with a tiny AP drop, e.g., $0.2\%$ . Please find more details of fair comparison in our appendix. + +Ablation study. Table 2 presents ablation studies to analyze VISO from three aspects: pre-training data, fine-tuning the CLIP encoder, and the proposed language-guided attention module. We mainly conduct ablations based on VISO-L and evaluate them on three benchmarks in a zero-shot manner. First, compared to the baseline (grey) trained on all pre-training data, reducing visual grounding (VG) and partial object detection (OD) datasets leads to $0.3 - 63.9\%$ AP drop. The decrease is due to the richer textual information from visual grounding datasets, enlarging the vocabulary of Earth objects. Second, we exploit two settings for the text encoder, CLIP, i.e., frozen and fine-tuned. Fine-tuning can improve AP by $0.1 - 5.2\%$ as large-scale satellite data helps align the text encoder with the Earth observation domain, which is not well covered in CLIP's training data. Third, the proposed language-guided attention module improves AP by $0.4 - 1.9\%$ , highlighting its effectiveness in distinguishing and improving object-centric feature extraction. + +# 4.2. Object Detection + +Baselines. We consider 3 specialist detectors as baselines, all of which focus on efficiency and practicality (§2.2). We train the baseline models separately on the training set of DOTAv2 and FAIR1Mv2 for 80 epochs (YOLOv8) and 36 epochs (RTMDet and LSKNet) using 2 NVIDIA H800 GPUs with a total batch size of 16. We adopt the AdamW + +
ModelBackboneParams(M) ↓MAR20HRSC2016VEDAI
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Pre-trained open-set models
GLIP-TSwin-T23248889.448848.44881.1
GroundingDINO-TSwin-T17246486.746436.84642.6
MM-GroundingDINO-TSwin-T17246486.646445.446416.1
YOLO-World-SYOLOv8-S134454.14410.9440.8
YOLO-World-MYOLOv8-M2911361.111312.51131.2
YOLO-World-LYOLOv8-L4722955.222920.92291.4
Zero-shot transferNo sparsity → Sparsity
VISO-SYOLOv8-S1344→2287.6→87.444→2659.7→59.044→1838.8→37.9
VISO-MYOLOv8-M29113→5990.2→90.1113→6875.6→74.6113→5546.6→46.2
VISO-LYOLOv8-L47229→13090.0→90.0229→13470.7→70.1229→11850.2→49.9
+ +Table 3. Zero-shot evaluation on MAR20, HRSC2016 and VEDAI test set. We compare VISO with existing VLMs on zero-shot transfer in terms of AP and FLOPs. No sparsity means the max accuracy of VISO and sparsity significantly reduces FLOPs with $\leq 1\%$ AP drop. + +
ModelDOTAv2 (test)FAIR1Mv2 (test)
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Specialist models
YOLOv8-S4050.84044.8
YOLOv8-M10554.410547.4
YOLOv8-L21756.421749.2
RTMDet-S3858.73847.5
RTMDet-M10160.1/10149.7/
RTMDet-L20660.820651.4*
LSKNet-T12459.512546.9
LSKNet-S16161.1*16249.7/
Open-set modelsNo sparsity → Sparsity
VISO-S44→1956.8→50.644→2446.2→44.8
VISO-M113→6961.8→60.0/113→7650.9→50.0/
VISO-L229→12863.4→61.1*229→12752.3→51.5*
+ +Table 4. Supervised results on DOTAv2 and FAIR1Mv2. We compare VISO with specialist models on object detection in terms of AP and FLOPs. The AP of VISO-L with No sparsity outperforms SOTA baselines (2.3% on DOTAv2 and 0.9% on FAIR1Mv2). We use sparsity to reduce AP until comparable with baselines (AP, AP*, and AP/) and find that FLOPs are significantly reduced. + +optimizer [28] and set an initial learning rate of 0.0002 for YOLOv8, 0.00025 for RTMDet, and 0.0001 for LSKNet. The same data augmentations as pre-training are applied. + +Supervised comparison with specialist models. We compare the proposed VISO with 3 close-set detectors, i.e., YOLOv8 with 3 variants, RTMDet with 3 variants, and LSKNet with 2 variants, ensuring a fair comparison with our 3 variants. Table 4 shows the results in terms of AP and FLOPs on two object detection benchmarks, i.e., the DOTAv2 and FAIR1Mv2 test sets. The results demonstrate that VISO outperforms the SOTA method, i.e., LSKNet-S, by increasing $2.3\%$ AP for DOTAv2 and $0.9\%$ AP for FAIR1Mv2. For sparse inference, we adjust the threshold to obtain comparable AP with LSKNet-S, i.e., $61.1\%$ for DOTAv2. We observe a significant drop in FLOPs of VISO-L from 229G to 128G, which is lower than the FLOPs of LSKNet-S, i.e., 161G. For FAIR1Mv2, we achieve both better FLOPs and AP than baselines. For example, VISO-M + +
ModelParams(M) ↓FLOPs(G) ↓Pr@0.5(%) ↑Pr@0.7(%) ↑
Specialist models
TransVG12620469.755.2
VLTVG12819476.962.3
MGVLF13221976.861.7
LPVA13722875.5*60.8
Open-set modelsNo sparsity → Sparsity
VISO-S1344→2474.4→70.169.1→64.7
VISO-M29113→8476.1→75.5*72.0→71.5
VISO-L48229→16877.3→75.8*72.9→71.6
+ +Table 5. Supervised results on DIOR-RSVG test set. We compare VISO with specialist models on object referring in terms of FLOPs, $\mathrm{Pr}@\mathrm{0.5}$ , and $\mathrm{Pr}@\mathrm{0.7}$ . Specialist models employ HBB and VISO adopts OBB for box format. Input size is $1024 \times 1024$ . + +obtains a better AP with only 76G FLOPs, reducing $2.1 \times$ FLOPs than LSKNet-S and $1.3 \times$ than RTMDet-M. Furthermore, VISO, built upon YOLOv8, achieves extremely lower FLOPs than YOLOv8, e.g., reducing $1.7 - 2.1 \times$ FLOPs for S with comparable AP. Our approach can be extended to other detectors to attain even greater accuracy. + +# 4.3. Object referring + +Baselines. We consider 4 specialist visual grounding models as baselines: TransVG [11], VLTVG [42], MGVLF [44], LPVA [21], all of which are built upon a neat transformer-based framework (§2.3). We train the baseline models on the DIOR-RSVG training set for 90 epochs using 2 NVIDIA H800 GPUs with a total batch size of 64. The learning rate and data augmentation follow the default settings of each method. We use two metrics: $\mathrm{Pr}@\mathbf{0.5}$ and $\mathrm{Pr}@\mathbf{0.7}$ to measure IOU precision $\geq 50\%$ , $70\%$ . + +Supervised comparison with specialist models. We compare our VISO with specialist baselines on the DIOR-RSVG test set. The results are shown in Table 5. We find that VISO outperforms SOTA methods by increasing $0.4\%$ $\mathrm{Pr}@\mathrm{0.5}$ and $10.6\%$ $\mathrm{Pr}@\mathrm{0.7}$ . We adjust the sparse threshold to obtain a comparable precision with baselines and observe a significant drop in FLOPs, e.g., reducing $8.5\times$ from 204G + +![](images/b22f69b829f2f2db1940f9ade9f9217c374229471b571125ed5eb0b026d32651.jpg) + +![](images/559fac3ca2edbca54a70d25003b81fc09ada96b5f17f504067b4beec52d3a2d8.jpg) + +![](images/88280350efebbe0a16660882d4f19b983774a1c341da05d7948d217e2c672c31.jpg) + +![](images/317ccfd6c3cdaeb03161a7023ebb9057e43bf1ea156e40731af4ab037c34a813.jpg) + +![](images/a4114f4552b353a9f78dc933514fea536918f1d44b375003e3d7445cd5ea4d5c.jpg) + +![](images/a813b5c5d1695e0520d49fb88bb27468c3b6993cd33d28fd6ee3c3fe50e8076a.jpg) + +![](images/dababf7fc43c398ea2f07c7f447745047282eaa673fc3686cbc706b9f5f0b353.jpg) + +![](images/e84e25a0af5b5c7509b5db2b8f37c3c05b59b74e644e5e804ae9b3581407e7e9.jpg) + +![](images/8be960204a44fd272c7d858a77d1194d46883eb65eae3b45af851067b082050d.jpg) + +![](images/bae24427c1286d72c2e4d5783077467448b5618f788082c3d3cec7092b5a7ee6.jpg) + +![](images/e8687f0cfacce830ac7190916b03bf60a6ea2188382281ceaddb5e3926bb874c.jpg) + +![](images/552b35d6d06da7e8bccaa4c2aa7d6d41aef10de5990d05f814d48b707ea96b27.jpg) + +![](images/d604e734a3b969fdfe389f2f28456b48c65129e31d6e1cfc03ccbcb92e21f2c2.jpg) +Figure 4. Sparsity analysis of VISO on 6 benchmarks. We conduct extensive experiments on sparse VISO using different sparse thresholds (X-axis) and record AP (left Y-axis) and FLOPs (right Y-axis). As the threshold increases, both AP and FLOPs decline, with FLOPs at a faster rate. We use validation sets of DOTAv2 and FAIR1Mv2 for this evaluation to avoid limits on test submissions. + +![](images/f6762bbdbd3bfcfb978ea23a7b1ab63e89eb11dcf6caf54f3da42f3ccf3488a4.jpg) + +![](images/6624ac58b91a4d22bc66d5e6401bfcfa427d1633294e9c1defd4f237ce1dba04.jpg) + +![](images/b36a50032efc03113f9a47779b206ba1e51c44fb40ddc581a4a03d8e6a7228bc.jpg) + +![](images/73d6b9b4a31a19386193d428d93791ca9eae40caa6fdafa67b05159528249ef7.jpg) + +![](images/1d071e734b47a7603266512e6b117364fc12625b9586f78f18c39414cb19af1c.jpg) + +Figure 5. Mask visualizations of VISO. More visualization of detection results and masks is in our appendix. +![](images/b97d546f21ded74f7b2dd408e3056c1a084efc1da5df4183ce63462df1a58f78.jpg) +(a) Mask visualization of fine-grained supervision for each target objects: (A) "baseball diamond", (B) "small vehicle", (C) "tennis court", (D) "helicopter". From (A) to (D), the multi-scale masks $\{X_{3}, X_{4}, X_{5}\}$ are arranged from the top (low-level feature) to the bottom (high-level feature). + +![](images/3c75ecbe5416243520879371c55343e1b24282a9cdabbf5abc96b7f9ae5b5fc0.jpg) + +![](images/5b38569b7a54204fad30dc437481bcfaa9ac2fe573b9985aae44f99dda9f2ef6.jpg) + +![](images/9bbac551d4c34af3254f0a76fe373862816c0b9ec3089af3f831836aa692990f.jpg) + +![](images/adb3d547684f1a95ade0db15b24b9da0d521d93ab9c3841226fc0a184fa5ce11.jpg) + +![](images/eff3fa802b7a44dae9c3fb668e57637e4ac0a7a7a3338030a6040ff6c2dec3d4.jpg) +(b) The sparsity of $\{C_4\}$ feature under different sparse thresholds. Left is the detection result (yellow-marked planes and white-marked helicopters) in DOTAv2. When the Earth observation task is specified as "detect the plane", masks only focus on planes and helicopters are irrelevant objects. + +![](images/6b34f1452bba00a5043c7d0ab44e100f5649489dc1950d77efe69129a00e0ed2.jpg) + +![](images/ceed69ecda4f36001e636cc958e989051b722d978cbc568619f7e16ef38b4bce.jpg) + +![](images/5850d99058e40945a5c6cba4dff94a1912adb1634f6fb46be752c6f4f1b9042c.jpg) + +to 24G when $\mathrm{Pr}@0.5$ is $70.1\%$ of VISO-S. + +# 4.4. Sparsity analysis + +All the above evaluations list only the optimal selection of sparse thresholds and employ a full-level sparse model for sparse inference, in which $P_{3}$ , $P_{4}$ , and $P_{5}$ are all sparse. We claim that adjusting the threshold $\theta$ and level-wise granularity can address different requirements for in-orbit inference, prioritizing faster or better. Figure 4 illustrates the trade-off between AP and FLOPs with different thresholds from 0.01 to 0.9. The results show a substantial efficiency improvement, i.e., a sharp drop in FLOPs at very low thresholds, which is attributed to the highly sparse model, guided by learned masks of small objects in satellite images, and its ability to distinctly differentiate target objects from others. + +Besides, the level-wise granularity also impacts FLOPs and AP. We denote the level-wise sparse conversion for $\{P_3,P_4,P_5\}$ as $\triangleright \{s_3,s_4,s_5\} (s_{l\in \{3,4,5\}}\in \{0,1\})$ . For example, $\triangleright \{1,0,0\}$ indicates the sparse $P_{3}$ and dense $P_4$ $P_{5}$ . Figure 4 shows that FLOPs decline more quickly in $\triangleright \{1,1,1\}$ than in $\triangleright \{1,1,0\}$ and $\triangleright \{1,0,0\}$ , however, the AP + +drop for $\triangleright \{1, 1, 1\}$ and $\triangleright \{1, 1, 0\}$ is nearly identical, indicating that low-resolution features in $P_{5}$ play a minor role in detecting small objects, primarily providing contextual information through a larger receptive field. + +# 4.5. Visualizations + +To demonstrate the effectiveness of VISO, we visualize the learned masks in Figure 5. In Figure 5a, the masks focus separately on the relevant objects rather than others. Furthermore, as features go deeper from $C_3$ to $C_5$ , more salient contextual information is captured, more than just the outline of objects. Figure 5b illustrates the sparsity of masked features under different thresholds. As the threshold $\theta$ increases from 0.01 to 0.9, the focus narrows with minimal loss of essential features. Specifically, the task is to detect the "plane", not the "helicopter". Thus, the mask gradually becomes more focused around the "plane" as the threshold increases. Redundant features (e.g., black areas) are masked out during sparse inference. For example, $\theta = 0.9$ can result in an extremely high sparsity of $99.7\%$ , preserving important features while significantly improving speed. + +![](images/92faa1698e3e2593547b305ef6c909a93076d8d73c0c56d2a95060238e00c427.jpg) +Figure 6. The acceleration performance of sparse convolution (SPConv) and normal convolution (Conv). The comparison is across different feature scales of three variants. SPConv provides significant acceleration for large features compared to Conv. However, for small features, Conv is already well-optimized. + +# 5. Real-world Analysis + +Following [17], we use frames per second (FPS) to measure the actual speed of detection on a real-world embedded GPU, NVIDIA Jetson Xavier NX [30], which serves as a payload on Chaohu-1 SAR satellite [37] launched in 2022. In this section, we consider a more common scenario in Earth observation missions: single-object detection, i.e., only one target or one category target is of interest. + +Implementation details. NX supports three optimized power budgets, e.g., 10W, 15W, and 20W. We pre-confine the mode as 10W with a 4-core CPU of $1900\mathrm{MHz}$ for our evaluation. The system of our testbed is equipped with Jet-Pack 5.1.3, CUDA 11.4, cuDNN 8.6.0, Pytorch 1.11.0, and SPconv 2.3.6. We test all models without extra acceleration mechanisms, e.g., FP16 or TensorRT. + +Sparse configuration. We use SPconv [9] as the high-performance sparse inference engine for onboard deployment. As shown in Figure 6, we test the acceleration performance of SPconv with different feature scales on our testbed. Note that the performance varies across different hardware. We consider the commonly used $3 \times 3$ kernel and compare the average latency of sparse convolution at different sparsity with that of normal convolution. The sparsity is defined as the ratio of the number of zeros in the feature to the total area of the feature. We observe that the larger the feature scale, the acceleration of SPConv is more pronounced. Especially for $\mathrm{L}\{P_3, P_4\}$ , $\mathrm{M}\{P_3, P_4, P_5\}$ and $\mathrm{S}\{P_3\}$ of VISO, a sparsity of more than $25\%$ can achieve acceleration, which can be easily attained in our highly sparse model when detecting small objects in satellite images, i.e., the smaller the object, the faster the detection. Thus, we configure the level-wise granularity to allow for more flexible sparse conversion, targeting different feature scales and hardware platforms. + +Baselines. We consider 2 baselines that are widely used in the industrial domain [17, 41], including a commonly used + +
MethodFLOPs(G) ↓FPS ↑AP(%) ↑Sparsity(%) ↓
Dense models
YOLOv8-S404.043.6-
YOLOv8-M1052.032.0-
YOLOv8-L2171.125.3-
Sparse models
QueryDet2972.334.838
VISO-S1611.148.388
VISO-M495.433.582
VISO-L1103.038.176
+ +Table 6. Real-world test on NX. We compare VISO with YOLOv8 and QueryDet. All variants of VISO are full-level sparse, and we set $\theta = 0.1$ for S, 0.25 for M, and 0.2 for L. We also list the average sparsity of all sparse features in sparse models. + +YOLOv8 [17] detector mentioned above and a sparse detector QueryDet [41]. QueryDet introduces a higher-resolution $P_{2}$ to improve small object detection and a query mechanism to make high-resolution heads sparse. We reproduce QueryDet using default settings and train on the DOTAv2 training set with the size of $1024 \times 1024$ . + +Speed testing. In Table 6, we detect the "helicopter" in DOTAv2 and compare VISO with YOLOv8 and QueryDet in terms of AP, FLOPs, FPS, and sparsity. VISO improves FPS by $2.7 - 2.8 \times$ than YOLOv8 and $4.8 \times$ than QueryDet under a better AP. VISO achieves significantly higher sparsity due to two folds: (1) The sparse computation of VISO occurs earlier, after the backbone, with sparse necks and heads, rather than only sparse heads in QueryDet. Also, VISO does not introduce higher-resolution features. (2) When detecting the target in irrelevant images, masks are unfocused, e.g., detect the helicopter in Figure 5a(D), and thus can early exit to avoid further inference. + +# 6. Conclusion + +We present VISO, an effective and efficient solution for in-orbit real-time object detection. VISO empowers vision-language models with language-guided mask learning and sparse inference. In this paper, we introduce a language-guided attention module to enhance salient object-centric features while suppressing irrelevant regions in satellite images. We pre-train VISO with large-scale satellite data in the form of region-text pairs with contrastive loss and additional mask learning, thereby endowing VISO with a strong capability of in-orbit open-vocabulary detection. Furthermore, the learned masks of VISO can suppress irrelevant regions to enable highly sparse inference and accelerate speed. Extensive experiments can demonstrate the superiority of VISO in both detection accuracy and speed, especially the significant speed-up on real-world satellites' embedded GPUs is insightful for future industrial deployment. + +# Acknowledgment + +This work is supported by the Center of High Performance Computing, Tsinghua University. + +# References + +[1] Frédéric Achard, Hans-Jürgen Stibig, Hugh D Eva, Erik J Lindquist, Alexandre Bouvet, Olivier Arino, and Philippe Mayaux. Estimating tropical deforestation from earth observation data. Carbon Management, 1(2):271-287, 2010. 2 +[2] Ali Bahri, Sina Ghofrani Majelan, Sina Mohammadi, Mehrdad Noori, and Karim Mohammadi. Remote sensing image classification via improved cross-entropy loss and transfer learning strategy based on deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 17 (6):1087–1091, 2020. 3 +[3] Hanoona Bangalath, Muhammad Maaz, Muhammad Uzair Khattak, Salman H Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. Advances in Neural Information Processing Systems, 35:33781-33794, 2022. 2 +[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3 +[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 4 +[6] Gong Cheng, Junwei Han, Peicheng Zhou, and Lei Guo. Multi-class geospatial object detection and geographic image classification based on collection of part detectors. ISPRS Journal of Photogrammetry and Remote Sensing, 98: 119-132, 2014. 5 +[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xing-gang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 1, 2, 3, 4, 5 +[8] MMYOLO Contributors. Mmyolo: Openmmlab yolo series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022. Accessed: 2023-03-03. 4 +[9] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022. 4, 8 +[10] Laura Crocetti, Matthias Forkel, Milan Fischer, František Jurečka, Aleks Grlj, Andreas Salentinig, Miroslav Trnka, Martha Anderson, Wai-Tim Ng, Žiga Kokalj, et al. Earth observation for agricultural drought monitoring in the pannonian basin (southeastern europe): current state and future directions. Regional Environmental Change, 20:1-17, 2020. 2 +[11] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF + +International Conference on Computer Vision, pages 1769-1779, 2021. 2, 6 +[12] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44 (11):7778-7796, 2021. 5 +[13] Qiqi Ding, Peng Li, Xuefeng Yan, Ding Shi, Luming Liang, Weiming Wang, Haoran Xie, Jonathan Li, and Mingqiang Wei. Cf-yolo: Cross fusion yolo for object detection in adverse weather with a high-quality real snow dataset. IEEE Transactions on Intelligent Transportation Systems, 24(10): 10749-10759, 2023. 5 +[14] Bowei Du, Yecheng Huang, Jiaxin Chen, and Di Huang. Adaptive sparse convolutional networks with global context enhancement for faster object detection on drone images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13435-13444, 2023. 2 +[15] Max Ghiglione and Vittorio Serra. Opportunities and challenges of ai on satellite processing units. In Proceedings of the 19th ACM international conference on computing Frontiers, pages 221-224, 2022. 1 +[16] Jiaxing Huang, Jingyi Zhang, Kai Jiang, and Shijian Lu. Open-vocabulary object detection via language hierarchy. arXiv preprint arXiv:2410.20371, 2024. 2 +[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLOv8, 2023. 2, 8 +[18] Kartik Kuckreja, Muhammad S. Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad S. Khan. Geochat: Grounded large vision-language model for remote sensing. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5 +[19] Jian Li, Xiang Ding and Mohamed Elhoseiny. Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding. arXiv:2406.12384, 2024. 5 +[20] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 5 +[21] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 6 +[22] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1, 2, 5 +[23] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in neural information processing systems, 33:21002-21012, 2020. 4 +[24] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. In Proceedings of + +the IEEE/CVF international conference on computer vision, pages 16794-16805, 2023. 1, 2 +[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2, 5 +[26] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, pages 324-331. SciTePress, 2017. 5 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6 +[29] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 2, 3, 4 +[30] NVIDIA. Jetson Xavier NX Developer Kit. https://developer.nvidia.com/embedded/jetson-xavier-nx-devkit, 2024. Online; accessed 14-March-2024. 8 +[31] George P Petropoulos, Prashant K Srivastava, Maria Piles, and Simon Pearson. Earth observation-based operational estimation of soil moisture and evapotranspiration for agricultural crops in support of sustainable water management. Sustainability, 10(1):181, 2018. 2 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5 +[33] Sebastien Razakarivony and Frederic Jurie. Vehicle detection in aerial imagery: A small target detection benchmark. Journal of Visual Communication and Image Representation, 34:187-203, 2016. 5 +[34] Gupta Ritwik et al. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint, 2019. 1 +[35] Daniel Saaristo. Shoot from the stars: Startup provides early detection of wildfires from space. https://blogs.nvidia.com/blog/ororatech-wildfires-from-space/, 2023.1,2 +[36] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 2 +[37] Spacety. Chaohu-1 SAR satellite. https://www.satelliteevolution.com/post/two-new-satellites-launched-including-the-2nd-sar-satellite-with-improved-capabilities,2022.8 + +[38] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 5 +[39] YU Wenqi, CHENG Gong, WANG Meijun, YAO Yanqing, XIE Xingxing, YAO Xiwen, and HAN Junwei. Mar20: A benchmark for military aircraft recognition in remote sensing images. National Remote Sensing Bulletin, 27(12):2688-2696, 2024. 5 +[40] Michael A Wulder, David P Roy, Volker C Radeloff, Thomas R Loveland, Martha C Anderson, David M Johnson, Sean Healey, Zhe Zhu, Theodore A Scambos, Nima Pahlevan, et al. Fifty years of landsat science and impacts. Remote Sensing of Environment, 280:113195, 2022. 2 +[41] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13668-13677, 2022. 2, 8 +[42] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 2, 6 +[43] Zhenghang Yuan, Lichao Mou, Yuansheng Hua, and Xiao Xiang Zhu. Rrasis: Referring remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 5 +[44] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1, 2, 5, 6 +[45] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multimodal large language model for multisensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 5 +[46] Yuanlin Zhang, Yuan Yuan, Yachuang Feng, and Xiaoqiang Lu. Hierarchical and robust convolutional neural network for very high-resolution remote sensing object detection. IEEE Transactions on Geoscience and Remote Sensing, 57 (8):5535-5548, 2019. 5 +[47] Ziyuan Zhang, Han Qiu, Maosen Zhang, Jun Liu, Bin Chen, Tianwei Zhang, and Hewu Li. COSMIC: Compress satellite images efficiently via diffusion compensation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[48] Xiangyu Zhao, Yicheng Chen, Shilin Xu, Xiangtai Li, Xinjiang Wang, Yining Li, and Haian Huang. An open and comprehensive pipeline for unified object grounding and detection. arXiv preprint arXiv:2401.02361, 2024. 1, 2, 5 +[49] Haigang Zhu, Xiaogang Chen, Weiqun Dai, Kun Fu, Qixiang Ye, and Jianbin Jiao. Orientation robust object detection in aerial images using deep convolutional neural network. In 2015 IEEE international conference on image processing (ICIP), pages 3735-3739. IEEE, 2015. 5 + +[50] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 5 \ No newline at end of file diff --git a/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/images.zip b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4146e33558a04732c408f12b5a8c9ba4fa51c110 --- /dev/null +++ b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f7c686de016f95500f8b1fb7f92043d9bd0408300fba6364695c224676f5eea +size 673500 diff --git a/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/layout.json b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b29988d14651e33d43e99e929e103d000820697f --- /dev/null +++ b/2025/VISO_ Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference/layout.json @@ -0,0 +1,10219 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "text", + "content": "VISO: Accelerating In-orbit Object Detection with Language-Guided Mask Learning and Sparse Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 161, + 311, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 161, + 311, + 203 + ], + "spans": [ + { + "bbox": [ + 140, + 161, + 311, + 203 + ], + "type": "text", + "content": "Meiqi Wang Tsinghua University wang-mq22@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 345, + 161, + 467, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 161, + 467, + 203 + ], + "spans": [ + { + "bbox": [ + 345, + 161, + 467, + 203 + ], + "type": "text", + "content": "Han Qiu Tsinghua University qiuhan@tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "text", + "content": "In-orbit object detection is essential for Earth observation missions on satellites equipped with GPUs. A promising approach is to use pre-trained vision-language modeling (VLM) to enhance its open-vocabulary capability. However, adopting it on satellites poses two challenges: (1) satellite imagery differs substantially from natural images, and (2) satellites' embedded GPUs are insufficient for complex models' inference. We reveal their lack of a crucial prior: in-orbit detection involves identifying a set of known objects within a cluttered yet monotonous background. Motivated by this observation, we propose VISO, a Vision-language Instructed Satellite Object detection model that focuses on object-specific features while suppressing irrelevant regions through language-guided mask learning. After pre-training on a large-scale satellite dataset with 3.4M region-text pairs, VISO enhances object-text alignment and object-centric features to improve detection accuracy. Also, VISO suppresses irrelevant regions, enabling highly sparse inference to accelerate speed on satellites. Extensive experiments show that VISO without sparsity outperforms state-of-the-art (SOTA) VLMs in zero-shot detection by increasing " + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "inline_equation", + "content": "34.1\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "text", + "content": " AP and reducing " + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "inline_equation", + "content": "27\\times" + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "text", + "content": " FLOPs, and surpasses specialist models in supervised object detection and object referring by improving " + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "text", + "content": " AP. When sparsifying VISO to a comparable AP, FLOPs can be greatly reduced by up to " + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "inline_equation", + "content": "8.5\\times" + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "text", + "content": ". Real-world tests reveal that VISO achieves a " + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "inline_equation", + "content": "2.8-4.8\\times" + }, + { + "bbox": [ + 55, + 256, + 296, + 579 + ], + "type": "text", + "content": " FPS speed-up on satellites' embedded GPUs1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 592, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "type": "text", + "content": "Recently, embedded GPUs are deployed in space [15], enabling in-orbit object detection to support latency-sensitive Earth observation tasks such as disaster monitoring [34] and wildfire detection [35]. In-orbit detection requires detecting flexible in-orbit tasks and responding to unforeseen emergencies beyond predefined closed-set categories. We observe that vision-language models (VLMs) [22] facilitate" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 230, + 425, + 329 + ], + "blocks": [ + { + "bbox": [ + 315, + 230, + 425, + 329 + ], + "lines": [ + { + "bbox": [ + 315, + 230, + 425, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 425, + 329 + ], + "type": "image", + "image_path": "df6a0c1c5dffa0f4fc8bd879315f523f76ac7691532ef304a520b645f9e50e9c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 337, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Earth observation images and natural images are different. Earth observation images typically contain very small objects. On the right is the size ratio of objects to images. We choose 5 categories in DOTA: plane (pl), storage tank (st), roundabout (rb), small vehicle (sv), baseball diamond (bd), and 5 categories in COCO: person (pr), cat (ct), car (cr), bed (bd), bottle (btl)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 231, + 552, + 328 + ], + "blocks": [ + { + "bbox": [ + 425, + 231, + 552, + 328 + ], + "lines": [ + { + "bbox": [ + 425, + 231, + 552, + 328 + ], + "spans": [ + { + "bbox": [ + 425, + 231, + 552, + 328 + ], + "type": "image", + "image_path": "812175ec29974e644e001667763d006274e13eecfd8385ee5e82290621cab7e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 412, + 554, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 436 + ], + "type": "text", + "content": "open-vocabulary detection (OVD) [7], allowing for detecting novel objects with the guidance of natural language." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 676 + ], + "type": "text", + "content": "However, most existing VLMs [22, 25, 48] are designed for natural images and suffer from high computational overhead due to complex backbones (e.g., Swin [27] used in GLIP [22]). Although YOLO-World [7] provides a practical solution for time-critical applications, directly applying it to satellite data remains challenging due to the disparities between terrestrial natural images and satellite images, as shown in Figure 1. Satellite images are typically captured from a bird's eye view with a long distance from the ground. They often contain a few tiny objects (e.g., ships) relevant to referring detection tasks in orbit (e.g., detect the ship on the lower left), surrounded by vast irrelevant objects and backgrounds (e.g., buildings, roads, etc.), which makes detection difficult [24, 44]. We consider these priors and address two important goals. (1) Accurate detection requires distinguishing object-centric visual features from irrelevant features while incorporating useful contextual information. (2) Fast inference can be achieved by omitting large irrelevant regions, consisting of non-essential backgrounds and objects, to reduce redundant computation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we achieve the above two goals by proposing VISO, a Vision-language Instructed Satellite Object detection model. Our key insight is that rich semantic textual" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 277, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 277, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 277, + 713 + ], + "type": "text", + "content": "1Our code: https://github.com/joannahuadu/VISO." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23300" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "embeddings of Earth observation tasks can help a detector distinguish different objects at its shallow layers (e.g., Figure 5a). Thus, we design a multi-scale mask learning to embed strongly aligned knowledge of Earth objects into multiscale visual features by pretraining on a large-scale satellite region-text pair. The learnable masks let VISO focus only on relevant areas at an early inference stage, ignoring irrelevant objects or cluttered backgrounds, thereby generating highly sparse features, which forward can be optimized by sparse computation (e.g., sparse convolution [14, 41])." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "content": "We conduct extensive experiments to validate the effectiveness and efficiency of VISO. VISO surpasses 4 VLM baselines in zero-shot open-vocabulary detection on 3 benchmarks, obtaining " + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "inline_equation", + "content": "34.1\\%" + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "content": " AP increase and " + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "inline_equation", + "content": "27\\times" + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "content": " FLOPs reduction. Furthermore, VISO outperforms 3 SOTA specialist models on 2 object detection benchmarks by " + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "content": " AP and 4 grounding models on one object referring benchmark by " + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "content": " precision. By applying sparsity and aligning AP with baselines, VISO can achieve up to " + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "inline_equation", + "content": "8.5\\times" + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "content": " reduction in FLOPs. Real-world tests further demonstrate a " + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "inline_equation", + "content": "2.8 - 4.8\\times" + }, + { + "bbox": [ + 55, + 192, + 294, + 324 + ], + "type": "text", + "content": " FPS speed-up on satellites' embedded GPUs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 336, + 230, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 230, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 230, + 350 + ], + "type": "text", + "content": "2. Background and Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 165, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 165, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 165, + 369 + ], + "type": "text", + "content": "2.1. GPUs on Satellites" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 374, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 445 + ], + "type": "text", + "content": "Traditional Earth observation missions [40] rely on a pipeline in which satellites take photos and then send them back to ground stations for analysis like forestry [1] or agriculture [10, 31]. This leads to a substantial requirement for data transmission from satellites to the ground [47] and cannot fulfill latency-sensitive tasks like disaster detection [35]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 446, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 518 + ], + "type": "text", + "content": "Recently, embedded GPUs are deployed on satellites for Earth observation missions [47]. In-orbit DNN inference can process real-time tasks, especially object detection, and send only results back to ground stations. For example, OroraTech has launched AI nanosatellites with NVIDIA Jetson Xavier NX for in-orbit wildfire detection [35]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 590 + ], + "type": "text", + "content": "Although deploying GPUs in orbit is a promising approach, satellites are still constrained by size and energy supply to support sophisticated GPUs. For instance, NVIDIA Jetson Xavier NX, a widely-used in-orbit GPU, can only support 10-20 Watts [35] such that accelerating the in-orbit DNN inference becomes critical." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 599, + 234, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 234, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 234, + 612 + ], + "type": "text", + "content": "2.2. Remote Sensing Object Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 700 + ], + "type": "text", + "content": "Recent research [24, 29, 41] has been dedicated to object detection specified for remote sensing, which typically addresses the issue of small objects in satellite images. They propose two types of approaches: utilizing large convolution kernels to introduce a large receptive field and leveraging high-resolution images or features to preserve small objects, both of which lead to high computational overhead." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "Thus, these methods pursue a lightweight solution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "LSKNet [24] employs a spatial selective mechanism to decompose the large kernel into a sequence of depthwise kernels. RTMDet [29] uses large-kernel depth-wise convolutions in the backbone and neck while reducing the number of building blocks as compensation to reduce model depth. QueryDet [41] introduces a higher-resolution feature scale in feature pyramids and sparsifies the high-resolution heads through cascaded queries to speed up. However, we find that these methods either aim to compensate for the extra computation or have limited acceleration potential only within heads. In this paper, we emphasize fully releasing the acceleration capability of the detection of small objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 223, + 542, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 542, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 542, + 236 + ], + "type": "text", + "content": "2.3. Vision-language Model for Object Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 240, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 396 + ], + "type": "text", + "content": "Open-vocabulary detection. Vision-language model (VLM) is a promising paradigm to empower open-vocabulary detection (OVD) according to open-set demands. Existing VLMs [22, 25, 48] unify OVD and visual grounding (or object referring) by reformulating the classification head as a text-contrastive head and designing vision-language fusion mode. However, these VLMs rely on complex backbones (e.g., Swin [27]), which are hard to deploy on resource-constrained devices. YOLO-World [7] address this by using the efficient YOLOv8 [17] as the backbone to enable real-time OVD. Despite this, these methods are pre-trained on natural images and the fusion lacks prior knowledge specific to satellite images, e.g., small objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 553, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 563 + ], + "type": "text", + "content": "Visual grounding. Specifically for visual grounding, some research establishes a neat multi-modal correspondence based on a cross-modal transformer and directly regresses coordinates. TransVG [11] replaces the complex fusion modules with a simple stack of transformer encoder layers. VLTVG [42] introduces an additional bidirectional visual-linguistic encoder before the cross-model decoder. MGVLF [44] considers scale variations in remote sensing and takes advantage of multi-scale visual features and multi-granularity textual embeddings. LPVA [21] makes better multi-modal interaction in the visual feature extraction process. However, these transformer-based grounding models will increase computational costs and will somehow degrade object detection without bounding box regression." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 563, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 553, + 647 + ], + "type": "text", + "content": "Region-text alignment. Various supervision strategies like [3, 16, 36] are proposed to enhance object-centric vision-language (i.e., region-text) alignment instead of coarse-grained image-level supervision [32] to make localization more precise. In this paper, we leverage prior knowledge from the Earth observation domain to enhance object-centric alignment and aim to improve efficiency in orbit." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 658, + 370, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 370, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 370, + 669 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Motivated by the aforementioned insights, in this section, we propose VISO, an efficient VLM tailored for Earth observation tasks. VISO can be seamlessly deployed on in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23301" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 541, + 210 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 541, + 210 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 541, + 210 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 541, + 210 + ], + "type": "image", + "image_path": "2e90e046a39a6a409c2dc24cab67d57a6dccb2c8b00fe8f57b4b01fb1545cbe2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 214, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 54, + 214, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 54, + 214, + 555, + 260 + ], + "type": "text", + "content": "Figure 2. Overview of VISO. The black dashed border represents pre-training with additional supervised mask learning. The red dashed border represents satellite deployment with sparse inference in VL-PAN and heads, along with an offline vocabulary omitting the text encoder like [7]. The language-guided attention module inputs textual embeddings and visual features and outputs multi-scale masks. The masks are learned in the training phase through supervised mask learning and serve as guidance to convert dense features into sparse ones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 270, + 295, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 295, + 294 + ], + "type": "text", + "content": "orbit satellites with GPUs for Earth observation tasks like open-vocabulary detection and object referring." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 296, + 295, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 296, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 295, + 404 + ], + "type": "text", + "content": "First, we introduce a multi-scale language-guided attention module to effectively enhance object-centric features relevant to language expressions while suppressing the irrelevant features (§3.1). Second, we perform a large-scale pre-training on satellite region-text pairs with deep supervision to enrich the knowledge of Earth (§3.2). Third, we convert the pre-trained dense models into highly sparse ones through multi-scale mask aggregation, ensuring minimal accuracy loss while improving inference efficiency (§3.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 170, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 170, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 170, + 426 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 433, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 296, + 506 + ], + "type": "text", + "content": "The architecture of VISO is illustrated in Figure 2. The proposed language-guided attention module is integrated with multi-scale visual features at an early stage. Though the current implementation is built upon the lightweight open-vocabulary detector, YOLO-World [7], the integration can be generalized to any other detector." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Open-vocabulary detector. VISO uses YOLO-World [7] as its base detector. The detector contains a backbone (i.e., CSPDarkNet, details can refer to [4]) to extract multi-scale visual features " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "C_l \\in \\mathbb{R}^{H_l \\times W_l \\times D_l}, l \\in 3, 4, 5" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " from the image " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", a pre-trained text encoder (i.e., CLIP [32]) to encode the text input " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " into textual embeddings " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{j \\in \\{1, \\dots, N\\}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": ", a vision-language path aggregation network (VL-PAN) for multi-scale feature pyramids " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{P_3, P_4, P_5\\}" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " to fuse textual embeddings and visual features, a decoupled head to regress bounding box " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{B_k\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " and object embeddings, and a text contrastive head to obtain object-text (region-text) similarity " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{s_{k,j}\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " between object embeddings and textual embeddings. " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " denotes the number of categories or expressions. " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " denotes the dimension of textual embeddings. " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " denotes the number of object predictions. We modify the regression output dimension from 4 to 5 to add a rotation angle prediction, enabling the commonly used ori" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 270, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 294 + ], + "type": "text", + "content": "ented object detection in Earth observation [29]. More details on VL-PAN and text contrastive head can refer to [7]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": "Language-guided attention module. We observe that the textual embeddings " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": " possess rich semantic prior knowledge and exhibit strong alignment with visual features due to large-scale pre-training [2]. They can be leveraged to differentiate the visual features of different objects and backgrounds early in the detector, rather than relying solely on the final text contrastive head. Thus, we introduce a language-guided attention module immediately following the backbone. Specifically, we adopt one projection layer " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": " and the max-sigmoid attention to multi-scale visual features " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "C_l\\in \\mathbb{R}^{H_l\\times W_l\\times D_l},l\\in 3,4,5." + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": " The projection layer projects textual embeddings to the dimension of each scale. The max-sigmoid attention generates multi-scale masks " + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "X_{l}\\in" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H_l\\times W_l\\times 1}(l\\in \\{3,4,5\\})" + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": " by the cross-modal aggregation between textual and visual features:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 482, + 555, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 482, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 353, + 482, + 555, + 515 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} M _ {l j} = C _ {l} \\mathbf {W} _ {\\mathbf {j}} ^ {T}, M _ {l j} \\in \\mathbb {R} ^ {H _ {l} \\times W _ {l} \\times N} \\\\ X _ {l} = \\delta \\left(\\max _ {j \\in \\{1, \\dots , N \\}} M _ {l j}\\right) \\end{array} \\right. \\tag {1}", + "image_path": "fd1d0eec8663a9196779e9b7a2326860000aa19bcf5a3eaea2584faaa613bb79.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 556, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 618 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 521, + 556, + 618 + ], + "type": "text", + "content": " indicates the sigmoid function. We note that accurately detecting objects in satellite images often requires a wide range of contextual information due to the scale variations of Earth objects [29]. In general, low-level (high-resolution) features benefit small objects, while high-level features with large receptive fields benefit larger objects. Thus, incorporating multi-scale features to generate masks is crucial for preserving diverse receptive fields." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 457, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 457, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 457, + 636 + ], + "type": "text", + "content": "3.2. Pre-training Formulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Region-text contrastive learning. We follow [7] and reformulate the ground-truth annotations as " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " region-text pairs " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{B_i, t_i\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " is the text for the object region " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "B_i" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " can be the category name for object detection or the referring expression for object referring. VISO inputs image " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and texts " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and outputs K object predictions " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{B_k, s_k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23302" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "We follow [29] and leverage the dynamic soft label assignment to match predictions with ground-truth annotations. The text index of " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " is assigned for each positive prediction as a classification label. The assigned label and the predicted similarity are used to construct the region-text contrastive loss " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{con}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " using quality focal loss [23]. In addition, rotated IoU loss [29] is adopted for oriented bounding box regression, and the basic training loss is defined as: " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1 = \\mathcal{L}_{\\mathrm{con}} + \\lambda \\mathcal{L}_{\\mathrm{iou}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " is a balancing coefficient." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "text", + "content": "Fine-grained semantic deep supervision. To learn masks " + }, + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "text", + "content": ", we explore a simple fine-grained deep supervision strategy. We assign the ground truth bitmap " + }, + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{lj} \\in \\{0,1\\}^{H_l \\times W_l \\times 1} (l \\in \\{3,4,5\\}, j \\in \\{1,\\dots,N\\})" + }, + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "text", + "content": " for each category's " + }, + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "inline_equation", + "content": "M_{lj}" + }, + { + "bbox": [ + 55, + 183, + 295, + 302 + ], + "type": "text", + "content": ". The bitmap of each category is generated from all the object regions belonging to this category and scaled across multiple levels to represent diverse receptive fields. Specifically, the pixels located within the regions are valued at 1, and others are 0. We define the deep supervision's loss function as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 309, + 295, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 309, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 107, + 309, + 295, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\mathcal {H} \\left(M _ {l j}, \\mathcal {B} _ {l j}\\right) \\tag {2}", + "image_path": "255b90627498b11a325e209b36c7fb6b56e68b89fde9f4d0210786b6e4124822.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "type": "text", + "content": " denotes the binary cross entropy function. Training the learnable masks is consistent with the basic training, and the total training loss is defined as " + }, + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_1 + \\alpha \\cdot \\mathcal{L}_2" + }, + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 353, + 295, + 401 + ], + "type": "text", + "content": " is a balancing coefficient. We visualize the masks in Figure 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 216, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 216, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 216, + 421 + ], + "type": "text", + "content": "3.3. Sparse Accelerated Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 521 + ], + "type": "text", + "content": "The multi-scale masks generated at the early stage can serve as effective guidance for enabling sparse computation in subsequent layers, e.g., neck and head layers. For instance, when detecting small vehicles in Figure 5a, the masks are centered on the top right, indicating that the remaining areas can be sparsified to avoid redundant further inference. Note that we only consider lightweight CNN-based detectors here. Offline sparse deployment involves three steps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": "First, the real-value masks " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " need to be converted to binary masks " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "M_{l} \\in \\{0,1\\}^{H_{l} \\times W_{l} \\times 1} (l \\in \\{3,4,5\\})" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " by hard threshold " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": ", in which values greater than " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " are set to 1 and others are set to 0. Therefore, adjusting " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " can optimize the trade-off between accuracy and efficiency. Greater sparsity with larger " + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": " can lead to increased efficiency and decreased accuracy. We will give more validation in §4.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": "Second, non-sparse locations of visual features " + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{C_3,C_4,C_5\\}" + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": " need to be aggregated across multiple scales to compensate for contexts. We consider the location of 1 in each scale's mask " + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{l}" + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{l} = \\{(x_{l}^{o},y_{l}^{o})\\} (o\\in \\{1,\\ldots ,S_{l}\\})" + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": " is the non-sparsity of " + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{l}" + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": ", i.e., total number of 1. In general, high-level features with low resolution and low-level features with high resolution are concatenated through up- or down-sampling in the feature pyramid. Thus, we use cross-scale projection to make up for the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 550, + 142 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 550, + 142 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 550, + 142 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 550, + 142 + ], + "type": "image", + "image_path": "ce1128fc88bfca58ce1093b47f48692aafdcd5f8232ccc6bc961c9d63e05c34c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "type": "text", + "content": "Figure 3. The sparse computation in the sparse convolution operates by performing the convolution only when the center of a " + }, + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 152, + 553, + 196 + ], + "type": "text", + "content": " kernel (with 1 padding) covers a non-zero value (e.g., only two non-zero values are in the sparse input features)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "text", + "content": "non-sparse location. Specifically, when using a higher-level location " + }, + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "text", + "content": " to query the lower-level feature " + }, + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "inline_equation", + "content": "C_{l-1}" + }, + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "text", + "content": ", the additional non-sparse location is denoted as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 247, + 553, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 247, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 343, + 247, + 553, + 262 + ], + "type": "interline_equation", + "content": "I _ {l - 1} ^ {\\prime} = \\left\\{\\left(2 x _ {l} ^ {o} + i, 2 y _ {l} ^ {o} + j\\right) \\mid \\forall i, j \\in \\{0, 1 \\} \\right\\} \\tag {3}", + "image_path": "ed255486021fb0205470d4cdb02e2fbb9bd20c287bdcd0a75de28123fe83752f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "text", + "content": "Then, update " + }, + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "inline_equation", + "content": "I_{l-1}" + }, + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "inline_equation", + "content": "I_{l-1} \\cup I_{l-1}'" + }, + { + "bbox": [ + 313, + 270, + 553, + 293 + ], + "type": "text", + "content": ". Querying higher-level features using lower-level locations is reversed by dividing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": "Third, based on " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": ", dense features " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "C_l \\in \\mathbb{R}^{H_l \\times W_l \\times D_l}" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": " are converted into sparse features " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{C}_l \\in \\mathbb{R}^{S_l \\times D_l}" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": ", which are then fed into the pre-converted sparse CNN layers for inference. The convolution in VL-PAN and heads are converted into sparse convolution (Figure 2). The computation of sparse convolution is illustrated in Figure 3. Take the " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": " convolution kernel size as an example, the computation complexity with sparse convolution is " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "S_{l} / HW" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": " of normal convolution. Thus, greater sparsity leads to lower computational complexity. Note that running sparse CNNs efficiently on general-purpose hardware (e.g., GPU) remains challenging. Thus, we leverage the existing high-performance inference engine (e.g., SpConv [9]) to accelerate sparse convolution computation. Additionally, we configure the sparse conversion as level-wise granularity (e.g., " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "P_3" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": " is sparse, while " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "P_4" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "inline_equation", + "content": "P_5" + }, + { + "bbox": [ + 313, + 293, + 554, + 498 + ], + "type": "text", + "content": " remain dense), allowing for greater flexibility to adjust based on the hardware performance (\\$5)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 506, + 394, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 394, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 394, + 520 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 526, + 553, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 553, + 623 + ], + "type": "text", + "content": "We evaluate the effectiveness of the VISO pre-trained on large-scale satellite data and the efficiency of the sparse VISO guided by learned masks. We perform VISO in various downstream Earth observation tasks, including zero-shot detection on 3 benchmarks (§4.1), supervised detection on 2 benchmarks (§4.2), and object referring on one benchmark (§4.3). We further evaluate the speed of sparse VISO on the real-world satellite payload (§5)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 624, + 553, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 553, + 696 + ], + "type": "text", + "content": "Implementation details. VISO is built upon the MMYOLO [8] and MMDetection [5] toolbox. For sparse inference, we implement a hook (SPHook) in MMDetection to convert VISO into a level-wise sparse model using SPconv [9]. The hook is easy to use, inserted at before_test_epoch" + }, + { + "bbox": [ + 313, + 624, + 553, + 696 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 624, + 553, + 696 + ], + "type": "text", + "content": ", with no extra overhead. In line with [7]," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 494, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 494, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 494, + 712 + ], + "type": "text", + "content": "2https://mmdetection.readthedocs.io" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23303" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 198 + ], + "type": "table", + "html": "
TypeDatasetVocab.ImagesAnno.Format
Object Detection (OD)DOTAv2 [12]18614701100487OBB
DIOR-R [20]2023463192518OBB
FAIR1Mv2 [38]3743611626790OBB
NWPU-VHR-10 [6]108325582HBB
UCAS-AOD [49]2302024610OBB
RSOD [13]4186914604HBB
HRRSD [46]136805593780HBB
VisDrone [50]12298591152646HBB
Visual Grounding (VG)DIOR-RSVG [44]-1740230898HBB
RRSISD [43]-1740213015HBB
GeoChat [18]-3832868563OBB
VRSBench [19]-2961438589OBB
", + "image_path": "093bb8f9b7355abbe814076431055f8d791f4248ee17a737c54d2b09ff12e247.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 205, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 239 + ], + "type": "text", + "content": "Table 1. Pre-training data. A list of datasets for pre-training VISO. HBB and OBB refer to horizontal and oriented bounding boxes, and Ann. denotes the annotations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "text", + "content": "we provide three variants of VISO to meet different requirements, e.g., small (S), medium (M), and large (L). We utilize the open-source CLIP [32] text encoder with pretrained weights to encode the text input. All satellite images used for pre-training and evaluation are split and resized to " + }, + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "text", + "content": ". Unless specified, all evaluation results of VISO are presented with two parts: a dense inference result (upper bound) and an optimal sparse inference result, which balances AP and FLOPs by adjusting the sparse threshold " + }, + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "text", + "content": ". More selections of " + }, + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 247, + 295, + 368 + ], + "type": "text", + "content": " will be discussed in §4.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 137, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 137, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 137, + 389 + ], + "type": "text", + "content": "4.1. Pre-training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 394, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 394, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 295, + 490 + ], + "type": "text", + "content": "Pre-training setup. We adopt the AdamW optimizer [28] with an initial learning rate of 0.0002 and weight decay of 0.025. VISO is pre-trained for 20 epochs on " + }, + { + "bbox": [ + 54, + 394, + 295, + 490 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 54, + 394, + 295, + 490 + ], + "type": "text", + "content": " NVIDIA H800 GPUs with a total batch size of 64. We use vanilla data augmentations to reduce training costs, including random resizing, rotation, and flipping without using mosaic augmentation. The text encoder is fine-tuned on our pretraining data with a learning factor of 0.01." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 494, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 494, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 54, + 494, + 295, + 615 + ], + "type": "text", + "content": "Pre-training data. We establish a large-scale dataset with 3.4M satellite region-text pairs to pre-train VISO. As listed in Table 1, we mainly collect 8 object detection datasets and 4 visual grounding datasets. We exclude the test sets of DOTAv2, FAIR1Mv2, and DIOR-RSVG, which are used for subsequent evaluation. The annotations for each dataset are formatted as region-text pairs, in which all the horizontal and oriented bounding boxes (regions) are represented as the corners of the quadrilateral and are uniformly converted to rotated bounding boxes when loading." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "content": "Zero-shot evaluation. After pre-training, we evaluate the pre-trained VISO on MAR20 [39], HRSC2016 [26], and VEDAI [33] benchmarks in a zero-shot manner following [45], as shown in Table 3. The MAR20 and HRSC2016 datasets contain more types of military aircraft and a diverse variety of ships. The VEDAI dataset contains 9 categories of vehicles on Earth. We compare the pre-trained VISO with recent state-of-the-art VLM mod" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 554, + 141 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 141 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 141 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 141 + ], + "type": "table", + "html": "
Training DataFrozen?LAM?MAR20HRSC2016VEDAI
only DOTAv2✓✓90.76.834.3
OD✗✓90.268.949.9
OD, VG✓✓89.968.845.0
OD, VG✗✗89.270.348.3
OD, VG✗✓90.070.750.2
", + "image_path": "1ac7dd99576fc6ac33934683df9b00901b8e39bbd168bfa030dc06cf0ab7504b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": "Table 2. Ablations on pre-training data, frozen CLIP encoder, and language-guided attention module (LAM). We evaluate the zero-shot performance on MAR20, HRSC2016, and VEDAI by pre-training VISO-L with varying amounts of data, using either a frozen or fine-tuned text encoder, with or without LAM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": "els [7, 22, 25, 48] and report AP and FLOPs. We preserve the pre-trained weights of these baselines (pre-trained on natural data) without any fine-tuning on satellite data. We find that VISO achieves remarkable performance in detecting Earth objects with " + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "inline_equation", + "content": "90.2\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": " AP for MAR20, " + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "inline_equation", + "content": "75.6\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": " AP for HRSC2016, and " + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "inline_equation", + "content": "50.2\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": " AP for VEDAI, while obtaining more than " + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": " reduction in FLOPs. The results highlight the disparities between natural and satellite images, showing that existing VLMs cannot be directly applied to the Earth observation domain and incur high computational costs. Furthermore, we test sparse inference with an optimal threshold, which results in a further reduction in FLOPs with " + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\leq 1\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": " AP drop, e.g., " + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": " FLOPs reduction of VISO-S on MAR20, with a tiny AP drop, e.g., " + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 394 + ], + "type": "text", + "content": ". Please find more details of fair comparison in our appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "text", + "content": "Ablation study. Table 2 presents ablation studies to analyze VISO from three aspects: pre-training data, fine-tuning the CLIP encoder, and the proposed language-guided attention module. We mainly conduct ablations based on VISO-L and evaluate them on three benchmarks in a zero-shot manner. First, compared to the baseline (grey) trained on all pre-training data, reducing visual grounding (VG) and partial object detection (OD) datasets leads to " + }, + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "inline_equation", + "content": "0.3 - 63.9\\%" + }, + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "text", + "content": " AP drop. The decrease is due to the richer textual information from visual grounding datasets, enlarging the vocabulary of Earth objects. Second, we exploit two settings for the text encoder, CLIP, i.e., frozen and fine-tuned. Fine-tuning can improve AP by " + }, + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "inline_equation", + "content": "0.1 - 5.2\\%" + }, + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "text", + "content": " as large-scale satellite data helps align the text encoder with the Earth observation domain, which is not well covered in CLIP's training data. Third, the proposed language-guided attention module improves AP by " + }, + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "inline_equation", + "content": "0.4 - 1.9\\%" + }, + { + "bbox": [ + 313, + 398, + 556, + 615 + ], + "type": "text", + "content": ", highlighting its effectiveness in distinguishing and improving object-centric feature extraction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 416, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 416, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 416, + 636 + ], + "type": "text", + "content": "4.2. Object Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Baselines. We consider 3 specialist detectors as baselines, all of which focus on efficiency and practicality (§2.2). We train the baseline models separately on the training set of DOTAv2 and FAIR1Mv2 for 80 epochs (YOLOv8) and 36 epochs (RTMDet and LSKNet) using 2 NVIDIA H800 GPUs with a total batch size of 16. We adopt the AdamW" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23304" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 545, + 210 + ], + "type": "table", + "html": "
ModelBackboneParams(M) ↓MAR20HRSC2016VEDAI
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Pre-trained open-set models
GLIP-TSwin-T23248889.448848.44881.1
GroundingDINO-TSwin-T17246486.746436.84642.6
MM-GroundingDINO-TSwin-T17246486.646445.446416.1
YOLO-World-SYOLOv8-S134454.14410.9440.8
YOLO-World-MYOLOv8-M2911361.111312.51131.2
YOLO-World-LYOLOv8-L4722955.222920.92291.4
Zero-shot transferNo sparsity → Sparsity
VISO-SYOLOv8-S1344→2287.6→87.444→2659.7→59.044→1838.8→37.9
VISO-MYOLOv8-M29113→5990.2→90.1113→6875.6→74.6113→5546.6→46.2
VISO-LYOLOv8-L47229→13090.0→90.0229→13470.7→70.1229→11850.2→49.9
", + "image_path": "9251c9b714fc3fd5c15c48a2069581a598c69c406dda8dd1a6600abd8c5060ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 250, + 296, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 218, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 555, + 242 + ], + "type": "text", + "content": "Table 3. Zero-shot evaluation on MAR20, HRSC2016 and VEDAI test set. We compare VISO with existing VLMs on zero-shot transfer in terms of AP and FLOPs. No sparsity means the max accuracy of VISO and sparsity significantly reduces FLOPs with " + }, + { + "bbox": [ + 55, + 218, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\leq 1\\%" + }, + { + "bbox": [ + 55, + 218, + 555, + 242 + ], + "type": "text", + "content": " AP drop." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 250, + 296, + 401 + ], + "lines": [ + { + "bbox": [ + 58, + 250, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 58, + 250, + 296, + 401 + ], + "type": "table", + "html": "
ModelDOTAv2 (test)FAIR1Mv2 (test)
FLOPs(G) ↓AP(%) ↑FLOPs(G) ↓AP(%) ↑
Specialist models
YOLOv8-S4050.84044.8
YOLOv8-M10554.410547.4
YOLOv8-L21756.421749.2
RTMDet-S3858.73847.5
RTMDet-M10160.1/10149.7/
RTMDet-L20660.820651.4*
LSKNet-T12459.512546.9
LSKNet-S16161.1*16249.7/
Open-set modelsNo sparsity → Sparsity
VISO-S44→1956.8→50.644→2446.2→44.8
VISO-M113→6961.8→60.0/113→7650.9→50.0/
VISO-L229→12863.4→61.1*229→12752.3→51.5*
", + "image_path": "36177831db176a3aa64f0f88c15e44ca61707c66484c22b1a1688c3025eb71df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 409, + 296, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 477 + ], + "type": "text", + "content": "Table 4. Supervised results on DOTAv2 and FAIR1Mv2. We compare VISO with specialist models on object detection in terms of AP and FLOPs. The AP of VISO-L with No sparsity outperforms SOTA baselines (2.3% on DOTAv2 and 0.9% on FAIR1Mv2). We use sparsity to reduce AP until comparable with baselines (AP, AP*, and AP/) and find that FLOPs are significantly reduced." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 529 + ], + "type": "text", + "content": "optimizer [28] and set an initial learning rate of 0.0002 for YOLOv8, 0.00025 for RTMDet, and 0.0001 for LSKNet. The same data augmentations as pre-training are applied." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Supervised comparison with specialist models. We compare the proposed VISO with 3 close-set detectors, i.e., YOLOv8 with 3 variants, RTMDet with 3 variants, and LSKNet with 2 variants, ensuring a fair comparison with our 3 variants. Table 4 shows the results in terms of AP and FLOPs on two object detection benchmarks, i.e., the DOTAv2 and FAIR1Mv2 test sets. The results demonstrate that VISO outperforms the SOTA method, i.e., LSKNet-S, by increasing " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " AP for DOTAv2 and " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " AP for FAIR1Mv2. For sparse inference, we adjust the threshold to obtain comparable AP with LSKNet-S, i.e., " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "61.1\\%" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " for DOTAv2. We observe a significant drop in FLOPs of VISO-L from 229G to 128G, which is lower than the FLOPs of LSKNet-S, i.e., 161G. For FAIR1Mv2, we achieve both better FLOPs and AP than baselines. For example, VISO-M" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 250, + 554, + 357 + ], + "blocks": [ + { + "bbox": [ + 317, + 250, + 554, + 357 + ], + "lines": [ + { + "bbox": [ + 317, + 250, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 317, + 250, + 554, + 357 + ], + "type": "table", + "html": "
ModelParams(M) ↓FLOPs(G) ↓Pr@0.5(%) ↑Pr@0.7(%) ↑
Specialist models
TransVG12620469.755.2
VLTVG12819476.962.3
MGVLF13221976.861.7
LPVA13722875.5*60.8
Open-set modelsNo sparsity → Sparsity
VISO-S1344→2474.4→70.169.1→64.7
VISO-M29113→8476.1→75.5*72.0→71.5
VISO-L48229→16877.3→75.8*72.9→71.6
", + "image_path": "32236f6d4f649d6a88671b1347beef143d2196f653bd754891bc8790d94a44e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "type": "text", + "content": "Table 5. Supervised results on DIOR-RSVG test set. We compare VISO with specialist models on object referring in terms of FLOPs, " + }, + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr}@\\mathrm{0.5}" + }, + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr}@\\mathrm{0.7}" + }, + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "type": "text", + "content": ". Specialist models employ HBB and VISO adopts OBB for box format. Input size is " + }, + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 365, + 555, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": "obtains a better AP with only 76G FLOPs, reducing " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "2.1 \\times" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": " FLOPs than LSKNet-S and " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "1.3 \\times" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": " than RTMDet-M. Furthermore, VISO, built upon YOLOv8, achieves extremely lower FLOPs than YOLOv8, e.g., reducing " + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "inline_equation", + "content": "1.7 - 2.1 \\times" + }, + { + "bbox": [ + 313, + 422, + 554, + 495 + ], + "type": "text", + "content": " FLOPs for S with comparable AP. Our approach can be extended to other detectors to attain even greater accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 501, + 415, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 415, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 415, + 514 + ], + "type": "text", + "content": "4.3. Object referring" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "content": "Baselines. We consider 4 specialist visual grounding models as baselines: TransVG [11], VLTVG [42], MGVLF [44], LPVA [21], all of which are built upon a neat transformer-based framework (§2.3). We train the baseline models on the DIOR-RSVG training set for 90 epochs using 2 NVIDIA H800 GPUs with a total batch size of 64. The learning rate and data augmentation follow the default settings of each method. We use two metrics: " + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr}@\\mathbf{0.5}" + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr}@\\mathbf{0.7}" + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "content": " to measure IOU precision " + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\geq 50\\%" + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 518, + 555, + 626 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Supervised comparison with specialist models. We compare our VISO with specialist baselines on the DIOR-RSVG test set. The results are shown in Table 5. We find that VISO outperforms SOTA methods by increasing " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr}@\\mathrm{0.5}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "10.6\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr}@\\mathrm{0.7}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ". We adjust the sparse threshold to obtain a comparable precision with baselines and observe a significant drop in FLOPs, e.g., reducing " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "8.5\\times" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " from 204G" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23305" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 144, + 138 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 144, + 138 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 144, + 138 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 144, + 138 + ], + "type": "image", + "image_path": "b22f69b829f2f2db1940f9ade9f9217c374229471b571125ed5eb0b026d32651.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 144, + 76, + 213, + 138 + ], + "blocks": [ + { + "bbox": [ + 144, + 76, + 213, + 138 + ], + "lines": [ + { + "bbox": [ + 144, + 76, + 213, + 138 + ], + "spans": [ + { + "bbox": [ + 144, + 76, + 213, + 138 + ], + "type": "image", + "image_path": "559fac3ca2edbca54a70d25003b81fc09ada96b5f17f504067b4beec52d3a2d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 213, + 76, + 293, + 138 + ], + "blocks": [ + { + "bbox": [ + 213, + 76, + 293, + 138 + ], + "lines": [ + { + "bbox": [ + 213, + 76, + 293, + 138 + ], + "spans": [ + { + "bbox": [ + 213, + 76, + 293, + 138 + ], + "type": "image", + "image_path": "88280350efebbe0a16660882d4f19b983774a1c341da05d7948d217e2c672c31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 293, + 76, + 370, + 138 + ], + "blocks": [ + { + "bbox": [ + 293, + 76, + 370, + 138 + ], + "lines": [ + { + "bbox": [ + 293, + 76, + 370, + 138 + ], + "spans": [ + { + "bbox": [ + 293, + 76, + 370, + 138 + ], + "type": "image", + "image_path": "317ccfd6c3cdaeb03161a7023ebb9057e43bf1ea156e40731af4ab037c34a813.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 76, + 454, + 138 + ], + "blocks": [ + { + "bbox": [ + 370, + 76, + 454, + 138 + ], + "lines": [ + { + "bbox": [ + 370, + 76, + 454, + 138 + ], + "spans": [ + { + "bbox": [ + 370, + 76, + 454, + 138 + ], + "type": "image", + "image_path": "a4114f4552b353a9f78dc933514fea536918f1d44b375003e3d7445cd5ea4d5c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 454, + 76, + 547, + 138 + ], + "blocks": [ + { + "bbox": [ + 454, + 76, + 547, + 138 + ], + "lines": [ + { + "bbox": [ + 454, + 76, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 454, + 76, + 547, + 138 + ], + "type": "image", + "image_path": "a813b5c5d1695e0520d49fb88bb27468c3b6993cd33d28fd6ee3c3fe50e8076a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 139, + 140, + 196 + ], + "blocks": [ + { + "bbox": [ + 62, + 139, + 140, + 196 + ], + "lines": [ + { + "bbox": [ + 62, + 139, + 140, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 139, + 140, + 196 + ], + "type": "image", + "image_path": "dababf7fc43c398ea2f07c7f447745047282eaa673fc3686cbc706b9f5f0b353.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 144, + 139, + 221, + 196 + ], + "blocks": [ + { + "bbox": [ + 144, + 139, + 221, + 196 + ], + "lines": [ + { + "bbox": [ + 144, + 139, + 221, + 196 + ], + "spans": [ + { + "bbox": [ + 144, + 139, + 221, + 196 + ], + "type": "image", + "image_path": "e84e25a0af5b5c7509b5db2b8f37c3c05b59b74e644e5e804ae9b3581407e7e9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 221, + 139, + 302, + 196 + ], + "blocks": [ + { + "bbox": [ + 221, + 139, + 302, + 196 + ], + "lines": [ + { + "bbox": [ + 221, + 139, + 302, + 196 + ], + "spans": [ + { + "bbox": [ + 221, + 139, + 302, + 196 + ], + "type": "image", + "image_path": "8be960204a44fd272c7d858a77d1194d46883eb65eae3b45af851067b082050d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 302, + 140, + 369, + 196 + ], + "blocks": [ + { + "bbox": [ + 302, + 140, + 369, + 196 + ], + "lines": [ + { + "bbox": [ + 302, + 140, + 369, + 196 + ], + "spans": [ + { + "bbox": [ + 302, + 140, + 369, + 196 + ], + "type": "image", + "image_path": "bae24427c1286d72c2e4d5783077467448b5618f788082c3d3cec7092b5a7ee6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 370, + 140, + 463, + 196 + ], + "blocks": [ + { + "bbox": [ + 370, + 140, + 463, + 196 + ], + "lines": [ + { + "bbox": [ + 370, + 140, + 463, + 196 + ], + "spans": [ + { + "bbox": [ + 370, + 140, + 463, + 196 + ], + "type": "image", + "image_path": "e8687f0cfacce830ac7190916b03bf60a6ea2188382281ceaddb5e3926bb874c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 465, + 140, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 465, + 140, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 465, + 140, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 465, + 140, + 547, + 196 + ], + "type": "image", + "image_path": "552b35d6d06da7e8bccaa4c2aa7d6d41aef10de5990d05f814d48b707ea96b27.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 62, + 198, + 140, + 255 + ], + "blocks": [ + { + "bbox": [ + 62, + 198, + 140, + 255 + ], + "lines": [ + { + "bbox": [ + 62, + 198, + 140, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 140, + 255 + ], + "type": "image", + "image_path": "d604e734a3b969fdfe389f2f28456b48c65129e31d6e1cfc03ccbcb92e21f2c2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 257, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 555, + 289 + ], + "type": "text", + "content": "Figure 4. Sparsity analysis of VISO on 6 benchmarks. We conduct extensive experiments on sparse VISO using different sparse thresholds (X-axis) and record AP (left Y-axis) and FLOPs (right Y-axis). As the threshold increases, both AP and FLOPs decline, with FLOPs at a faster rate. We use validation sets of DOTAv2 and FAIR1Mv2 for this evaluation to avoid limits on test submissions." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 144, + 198, + 221, + 255 + ], + "blocks": [ + { + "bbox": [ + 144, + 198, + 221, + 255 + ], + "lines": [ + { + "bbox": [ + 144, + 198, + 221, + 255 + ], + "spans": [ + { + "bbox": [ + 144, + 198, + 221, + 255 + ], + "type": "image", + "image_path": "f6762bbdbd3bfcfb978ea23a7b1ab63e89eb11dcf6caf54f3da42f3ccf3488a4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 221, + 198, + 302, + 255 + ], + "blocks": [ + { + "bbox": [ + 221, + 198, + 302, + 255 + ], + "lines": [ + { + "bbox": [ + 221, + 198, + 302, + 255 + ], + "spans": [ + { + "bbox": [ + 221, + 198, + 302, + 255 + ], + "type": "image", + "image_path": "6624ac58b91a4d22bc66d5e6401bfcfa427d1633294e9c1defd4f237ce1dba04.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 302, + 198, + 369, + 255 + ], + "blocks": [ + { + "bbox": [ + 302, + 198, + 369, + 255 + ], + "lines": [ + { + "bbox": [ + 302, + 198, + 369, + 255 + ], + "spans": [ + { + "bbox": [ + 302, + 198, + 369, + 255 + ], + "type": "image", + "image_path": "b36a50032efc03113f9a47779b206ba1e51c44fb40ddc581a4a03d8e6a7228bc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 370, + 198, + 463, + 255 + ], + "blocks": [ + { + "bbox": [ + 370, + 198, + 463, + 255 + ], + "lines": [ + { + "bbox": [ + 370, + 198, + 463, + 255 + ], + "spans": [ + { + "bbox": [ + 370, + 198, + 463, + 255 + ], + "type": "image", + "image_path": "73d6b9b4a31a19386193d428d93791ca9eae40caa6fdafa67b05159528249ef7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 465, + 198, + 544, + 255 + ], + "blocks": [ + { + "bbox": [ + 465, + 198, + 544, + 255 + ], + "lines": [ + { + "bbox": [ + 465, + 198, + 544, + 255 + ], + "spans": [ + { + "bbox": [ + 465, + 198, + 544, + 255 + ], + "type": "image", + "image_path": "1d071e734b47a7603266512e6b117364fc12625b9586f78f18c39414cb19af1c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 65, + 297, + 160, + 392 + ], + "blocks": [ + { + "bbox": [ + 65, + 297, + 160, + 392 + ], + "lines": [ + { + "bbox": [ + 65, + 297, + 160, + 392 + ], + "spans": [ + { + "bbox": [ + 65, + 297, + 160, + 392 + ], + "type": "image", + "image_path": "b97d546f21ded74f7b2dd408e3056c1a084efc1da5df4183ce63462df1a58f78.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 393, + 301, + 433 + ], + "lines": [ + { + "bbox": [ + 59, + 393, + 301, + 433 + ], + "spans": [ + { + "bbox": [ + 59, + 393, + 301, + 433 + ], + "type": "text", + "content": "(a) Mask visualization of fine-grained supervision for each target objects: (A) \"baseball diamond\", (B) \"small vehicle\", (C) \"tennis court\", (D) \"helicopter\". From (A) to (D), the multi-scale masks " + }, + { + "bbox": [ + 59, + 393, + 301, + 433 + ], + "type": "inline_equation", + "content": "\\{X_{3}, X_{4}, X_{5}\\}" + }, + { + "bbox": [ + 59, + 393, + 301, + 433 + ], + "type": "text", + "content": " are arranged from the top (low-level feature) to the bottom (high-level feature)." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 107, + 437, + 501, + 449 + ], + "lines": [ + { + "bbox": [ + 107, + 437, + 501, + 449 + ], + "spans": [ + { + "bbox": [ + 107, + 437, + 501, + 449 + ], + "type": "text", + "content": "Figure 5. Mask visualizations of VISO. More visualization of detection results and masks is in our appendix." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 162, + 298, + 195, + 392 + ], + "blocks": [ + { + "bbox": [ + 162, + 298, + 195, + 392 + ], + "lines": [ + { + "bbox": [ + 162, + 298, + 195, + 392 + ], + "spans": [ + { + "bbox": [ + 162, + 298, + 195, + 392 + ], + "type": "image", + "image_path": "3c75ecbe5416243520879371c55343e1b24282a9cdabbf5abc96b7f9ae5b5fc0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 196, + 298, + 228, + 392 + ], + "blocks": [ + { + "bbox": [ + 196, + 298, + 228, + 392 + ], + "lines": [ + { + "bbox": [ + 196, + 298, + 228, + 392 + ], + "spans": [ + { + "bbox": [ + 196, + 298, + 228, + 392 + ], + "type": "image", + "image_path": "5b38569b7a54204fad30dc437481bcfaa9ac2fe573b9985aae44f99dda9f2ef6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 230, + 298, + 263, + 392 + ], + "blocks": [ + { + "bbox": [ + 230, + 298, + 263, + 392 + ], + "lines": [ + { + "bbox": [ + 230, + 298, + 263, + 392 + ], + "spans": [ + { + "bbox": [ + 230, + 298, + 263, + 392 + ], + "type": "image", + "image_path": "9bbac551d4c34af3254f0a76fe373862816c0b9ec3089af3f831836aa692990f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 265, + 298, + 298, + 392 + ], + "blocks": [ + { + "bbox": [ + 265, + 298, + 298, + 392 + ], + "lines": [ + { + "bbox": [ + 265, + 298, + 298, + 392 + ], + "spans": [ + { + "bbox": [ + 265, + 298, + 298, + 392 + ], + "type": "image", + "image_path": "adb3d547684f1a95ade0db15b24b9da0d521d93ab9c3841226fc0a184fa5ce11.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 306, + 297, + 399, + 392 + ], + "blocks": [ + { + "bbox": [ + 306, + 297, + 399, + 392 + ], + "lines": [ + { + "bbox": [ + 306, + 297, + 399, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 399, + 392 + ], + "type": "image", + "image_path": "eff3fa802b7a44dae9c3fb668e57637e4ac0a7a7a3338030a6040ff6c2dec3d4.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 393, + 548, + 433 + ], + "lines": [ + { + "bbox": [ + 304, + 393, + 548, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 548, + 433 + ], + "type": "text", + "content": "(b) The sparsity of " + }, + { + "bbox": [ + 304, + 393, + 548, + 433 + ], + "type": "inline_equation", + "content": "\\{C_4\\}" + }, + { + "bbox": [ + 304, + 393, + 548, + 433 + ], + "type": "text", + "content": " feature under different sparse thresholds. Left is the detection result (yellow-marked planes and white-marked helicopters) in DOTAv2. When the Earth observation task is specified as \"detect the plane\", masks only focus on planes and helicopters are irrelevant objects." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 399, + 298, + 445, + 392 + ], + "blocks": [ + { + "bbox": [ + 399, + 298, + 445, + 392 + ], + "lines": [ + { + "bbox": [ + 399, + 298, + 445, + 392 + ], + "spans": [ + { + "bbox": [ + 399, + 298, + 445, + 392 + ], + "type": "image", + "image_path": "6b34f1452bba00a5043c7d0ab44e100f5649489dc1950d77efe69129a00e0ed2.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 445, + 297, + 492, + 392 + ], + "blocks": [ + { + "bbox": [ + 445, + 297, + 492, + 392 + ], + "lines": [ + { + "bbox": [ + 445, + 297, + 492, + 392 + ], + "spans": [ + { + "bbox": [ + 445, + 297, + 492, + 392 + ], + "type": "image", + "image_path": "ceed69ecda4f36001e636cc958e989051b722d978cbc568619f7e16ef38b4bce.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 492, + 297, + 539, + 392 + ], + "blocks": [ + { + "bbox": [ + 492, + 297, + 539, + 392 + ], + "lines": [ + { + "bbox": [ + 492, + 297, + 539, + 392 + ], + "spans": [ + { + "bbox": [ + 492, + 297, + 539, + 392 + ], + "type": "image", + "image_path": "5850d99058e40945a5c6cba4dff94a1912adb1634f6fb46be752c6f4f1b9042c.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 460, + 229, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 229, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 229, + 472 + ], + "type": "text", + "content": "to 24G when " + }, + { + "bbox": [ + 55, + 460, + 229, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{Pr}@0.5" + }, + { + "bbox": [ + 55, + 460, + 229, + 472 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 460, + 229, + 472 + ], + "type": "inline_equation", + "content": "70.1\\%" + }, + { + "bbox": [ + 55, + 460, + 229, + 472 + ], + "type": "text", + "content": " of VISO-S." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 480, + 157, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 157, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 157, + 493 + ], + "type": "text", + "content": "4.4. Sparsity analysis" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "content": "All the above evaluations list only the optimal selection of sparse thresholds and employ a full-level sparse model for sparse inference, in which " + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "inline_equation", + "content": "P_{4}" + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "inline_equation", + "content": "P_{5}" + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "content": " are all sparse. We claim that adjusting the threshold " + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "content": " and level-wise granularity can address different requirements for in-orbit inference, prioritizing faster or better. Figure 4 illustrates the trade-off between AP and FLOPs with different thresholds from 0.01 to 0.9. The results show a substantial efficiency improvement, i.e., a sharp drop in FLOPs at very low thresholds, which is attributed to the highly sparse model, guided by learned masks of small objects in satellite images, and its ability to distinctly differentiate target objects from others." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Besides, the level-wise granularity also impacts FLOPs and AP. We denote the level-wise sparse conversion for " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{P_3,P_4,P_5\\}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\triangleright \\{s_3,s_4,s_5\\} (s_{l\\in \\{3,4,5\\}}\\in \\{0,1\\})" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " . For example, " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\triangleright \\{1,0,0\\}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " indicates the sparse " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "P_{3}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and dense " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "P_4" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "P_{5}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " . Figure 4 shows that FLOPs decline more quickly in " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\triangleright \\{1,1,1\\}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " than in " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\triangleright \\{1,1,0\\}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\triangleright \\{1,0,0\\}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " , however, the AP" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "text", + "content": "drop for " + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\triangleright \\{1, 1, 1\\}" + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\triangleright \\{1, 1, 0\\}" + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "text", + "content": " is nearly identical, indicating that low-resolution features in " + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "inline_equation", + "content": "P_{5}" + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "text", + "content": " play a minor role in detecting small objects, primarily providing contextual information through a larger receptive field." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 517, + 403, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 403, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 403, + 528 + ], + "type": "text", + "content": "4.5. Visualizations" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "content": "To demonstrate the effectiveness of VISO, we visualize the learned masks in Figure 5. In Figure 5a, the masks focus separately on the relevant objects rather than others. Furthermore, as features go deeper from " + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "C_5" + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "content": ", more salient contextual information is captured, more than just the outline of objects. Figure 5b illustrates the sparsity of masked features under different thresholds. As the threshold " + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "content": " increases from 0.01 to 0.9, the focus narrows with minimal loss of essential features. Specifically, the task is to detect the \"plane\", not the \"helicopter\". Thus, the mask gradually becomes more focused around the \"plane\" as the threshold increases. Redundant features (e.g., black areas) are masked out during sparse inference. For example, " + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta = 0.9" + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "content": " can result in an extremely high sparsity of " + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "inline_equation", + "content": "99.7\\%" + }, + { + "bbox": [ + 312, + 534, + 555, + 714 + ], + "type": "text", + "content": ", preserving important features while significantly improving speed." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 202, + 64, + 410, + 73 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 64, + 410, + 73 + ], + "spans": [ + { + "bbox": [ + 202, + 64, + 410, + 73 + ], + "type": "text", + "content": "AP FLOPs {1,0,0} {1,1,0} {1,1,1}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23306" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 187 + ], + "type": "image", + "image_path": "92faa1698e3e2593547b305ef6c909a93076d8d73c0c56d2a95060238e00c427.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 197, + 296, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 253 + ], + "type": "text", + "content": "Figure 6. The acceleration performance of sparse convolution (SPConv) and normal convolution (Conv). The comparison is across different feature scales of three variants. SPConv provides significant acceleration for large features compared to Conv. However, for small features, Conv is already well-optimized." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 260, + 174, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 174, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 174, + 274 + ], + "type": "text", + "content": "5. Real-world Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 365 + ], + "type": "text", + "content": "Following [17], we use frames per second (FPS) to measure the actual speed of detection on a real-world embedded GPU, NVIDIA Jetson Xavier NX [30], which serves as a payload on Chaohu-1 SAR satellite [37] launched in 2022. In this section, we consider a more common scenario in Earth observation missions: single-object detection, i.e., only one target or one category target is of interest." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "Implementation details. NX supports three optimized power budgets, e.g., 10W, 15W, and 20W. We pre-confine the mode as 10W with a 4-core CPU of " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "1900\\mathrm{MHz}" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " for our evaluation. The system of our testbed is equipped with Jet-Pack 5.1.3, CUDA 11.4, cuDNN 8.6.0, Pytorch 1.11.0, and SPconv 2.3.6. We test all models without extra acceleration mechanisms, e.g., FP16 or TensorRT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "text", + "content": "Sparse configuration. We use SPconv [9] as the high-performance sparse inference engine for onboard deployment. As shown in Figure 6, we test the acceleration performance of SPconv with different feature scales on our testbed. Note that the performance varies across different hardware. We consider the commonly used " + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "text", + "content": " kernel and compare the average latency of sparse convolution at different sparsity with that of normal convolution. The sparsity is defined as the ratio of the number of zeros in the feature to the total area of the feature. We observe that the larger the feature scale, the acceleration of SPConv is more pronounced. Especially for " + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{L}\\{P_3, P_4\\}" + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{M}\\{P_3, P_4, P_5\\}" + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{S}\\{P_3\\}" + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "text", + "content": " of VISO, a sparsity of more than " + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 457, + 295, + 685 + ], + "type": "text", + "content": " can achieve acceleration, which can be easily attained in our highly sparse model when detecting small objects in satellite images, i.e., the smaller the object, the faster the detection. Thus, we configure the level-wise granularity to allow for more flexible sparse conversion, targeting different feature scales and hardware platforms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Baselines. We consider 2 baselines that are widely used in the industrial domain [17, 41], including a commonly used" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 549, + 182 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 549, + 182 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 549, + 182 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 549, + 182 + ], + "type": "table", + "html": "
MethodFLOPs(G) ↓FPS ↑AP(%) ↑Sparsity(%) ↓
Dense models
YOLOv8-S404.043.6-
YOLOv8-M1052.032.0-
YOLOv8-L2171.125.3-
Sparse models
QueryDet2972.334.838
VISO-S1611.148.388
VISO-M495.433.582
VISO-L1103.038.176
", + "image_path": "10524d770f15dff6e34b3e866c76a35963b1821e00f7149e23af99b10a1c9303.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 190, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 236 + ], + "type": "text", + "content": "Table 6. Real-world test on NX. We compare VISO with YOLOv8 and QueryDet. All variants of VISO are full-level sparse, and we set " + }, + { + "bbox": [ + 313, + 190, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\theta = 0.1" + }, + { + "bbox": [ + 313, + 190, + 553, + 236 + ], + "type": "text", + "content": " for S, 0.25 for M, and 0.2 for L. We also list the average sparsity of all sparse features in sparse models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 251, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 323 + ], + "type": "text", + "content": "YOLOv8 [17] detector mentioned above and a sparse detector QueryDet [41]. QueryDet introduces a higher-resolution " + }, + { + "bbox": [ + 313, + 251, + 555, + 323 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 313, + 251, + 555, + 323 + ], + "type": "text", + "content": " to improve small object detection and a query mechanism to make high-resolution heads sparse. We reproduce QueryDet using default settings and train on the DOTAv2 training set with the size of " + }, + { + "bbox": [ + 313, + 251, + 555, + 323 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 251, + 555, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "text", + "content": "Speed testing. In Table 6, we detect the \"helicopter\" in DOTAv2 and compare VISO with YOLOv8 and QueryDet in terms of AP, FLOPs, FPS, and sparsity. VISO improves FPS by " + }, + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "inline_equation", + "content": "2.7 - 2.8 \\times" + }, + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "text", + "content": " than YOLOv8 and " + }, + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "inline_equation", + "content": "4.8 \\times" + }, + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "text", + "content": " than QueryDet under a better AP. VISO achieves significantly higher sparsity due to two folds: (1) The sparse computation of VISO occurs earlier, after the backbone, with sparse necks and heads, rather than only sparse heads in QueryDet. Also, VISO does not introduce higher-resolution features. (2) When detecting the target in irrelevant images, masks are unfocused, e.g., detect the helicopter in Figure 5a(D), and thus can early exit to avoid further inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 498, + 388, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 498, + 388, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 498, + 388, + 510 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": "We present VISO, an effective and efficient solution for in-orbit real-time object detection. VISO empowers vision-language models with language-guided mask learning and sparse inference. In this paper, we introduce a language-guided attention module to enhance salient object-centric features while suppressing irrelevant regions in satellite images. We pre-train VISO with large-scale satellite data in the form of region-text pairs with contrastive loss and additional mask learning, thereby endowing VISO with a strong capability of in-orbit open-vocabulary detection. Furthermore, the learned masks of VISO can suppress irrelevant regions to enable highly sparse inference and accelerate speed. Extensive experiments can demonstrate the superiority of VISO in both detection accuracy and speed, especially the significant speed-up on real-world satellites' embedded GPUs is insightful for future industrial deployment." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23307" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 117 + ], + "type": "text", + "content": "This work is supported by the Center of High Performance Computing, Tsinghua University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "type": "text", + "content": "[1] Frédéric Achard, Hans-Jürgen Stibig, Hugh D Eva, Erik J Lindquist, Alexandre Bouvet, Olivier Arino, and Philippe Mayaux. Estimating tropical deforestation from earth observation data. Carbon Management, 1(2):271-287, 2010. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 205, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 270 + ], + "type": "text", + "content": "[2] Ali Bahri, Sina Ghofrani Majelan, Sina Mohammadi, Mehrdad Noori, and Karim Mohammadi. Remote sensing image classification via improved cross-entropy loss and transfer learning strategy based on deep convolutional neural networks. IEEE Geoscience and Remote Sensing Letters, 17 (6):1087–1091, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 272, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 295, + 328 + ], + "type": "text", + "content": "[3] Hanoona Bangalath, Muhammad Maaz, Muhammad Uzair Khattak, Salman H Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. Advances in Neural Information Processing Systems, 35:33781-33794, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 296, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 362 + ], + "type": "text", + "content": "[4] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 296, + 418 + ], + "type": "text", + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 474 + ], + "type": "text", + "content": "[6] Gong Cheng, Junwei Han, Peicheng Zhou, and Lei Guo. Multi-class geospatial object detection and geographic image classification based on collection of part detectors. ISPRS Journal of Photogrammetry and Remote Sensing, 98: 119-132, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 532 + ], + "type": "text", + "content": "[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xing-gang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 566 + ], + "type": "text", + "content": "[8] MMYOLO Contributors. Mmyolo: Openmmlab yolo series toolbox and benchmark. https://github.com/open-mmlab/mmyolo, 2022. Accessed: 2023-03-03. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 600 + ], + "type": "text", + "content": "[9] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022. 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 677 + ], + "type": "text", + "content": "[10] Laura Crocetti, Matthias Forkel, Milan Fischer, František Jurečka, Aleks Grlj, Andreas Salentinig, Miroslav Trnka, Martha Anderson, Wai-Tim Ng, Žiga Kokalj, et al. Earth observation for agricultural drought monitoring in the pannonian basin (southeastern europe): current state and future directions. Regional Environmental Change, 20:1-17, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[11] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 1769-1779, 2021. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "text", + "content": "[12] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44 (11):7778-7796, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[13] Qiqi Ding, Peng Li, Xuefeng Yan, Ding Shi, Luming Liang, Weiming Wang, Haoran Xie, Jonathan Li, and Mingqiang Wei. Cf-yolo: Cross fusion yolo for object detection in adverse weather with a high-quality real snow dataset. IEEE Transactions on Intelligent Transportation Systems, 24(10): 10749-10759, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 285 + ], + "type": "text", + "content": "[14] Bowei Du, Yecheng Huang, Jiaxin Chen, and Di Huang. Adaptive sparse convolutional networks with global context enhancement for faster object detection on drone images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13435-13444, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[15] Max Ghiglione and Vittorio Serra. Opportunities and challenges of ai on satellite processing units. In Proceedings of the 19th ACM international conference on computing Frontiers, pages 221-224, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "text", + "content": "[16] Jiaxing Huang, Jingyi Zhang, Kai Jiang, and Shijian Lu. Open-vocabulary object detection via language hierarchy. arXiv preprint arXiv:2410.20371, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 386 + ], + "type": "text", + "content": "[17] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics YOLOv8, 2023. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "text", + "content": "[18] Kartik Kuckreja, Muhammad S. Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad S. Khan. Geochat: Grounded large vision-language model for remote sensing. The IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "text", + "content": "[19] Jian Li, Xiang Ding and Mohamed Elhoseiny. Vrsbench: A versatile vision-language benchmark dataset for remote sensing image understanding. arXiv:2406.12384, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 522 + ], + "type": "text", + "content": "[20] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[21] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "text", + "content": "[22] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1, 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[23] Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in neural information processing systems, 33:21002-21012, 2020. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[24] Yuxuan Li, Qibin Hou, Zhaohui Zheng, Ming-Ming Cheng, Jian Yang, and Xiang Li. Large selective kernel network for remote sensing object detection. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23308" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF international conference on computer vision, pages 16794-16805, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 162 + ], + "type": "text", + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "type": "text", + "content": "[26] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, pages 324-331. SciTePress, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 272 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 294 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "text", + "content": "[29] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[30] NVIDIA. Jetson Xavier NX Developer Kit. https://developer.nvidia.com/embedded/jetson-xavier-nx-devkit, 2024. Online; accessed 14-March-2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[31] George P Petropoulos, Prashant K Srivastava, Maria Piles, and Simon Pearson. Earth observation-based operational estimation of soil moisture and evapotranspiration for agricultural crops in support of sustainable water management. Sustainability, 10(1):181, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "text", + "content": "[33] Sebastien Razakarivony and Frederic Jurie. Vehicle detection in aerial imagery: A small target detection benchmark. Journal of Visual Communication and Image Representation, 34:187-203, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 570 + ], + "type": "text", + "content": "[34] Gupta Ritwik et al. xbd: A dataset for assessing building damage from satellite imagery. arXiv preprint, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[35] Daniel Saaristo. Shoot from the stars: Startup provides early detection of wildfires from space. https://blogs.nvidia.com/blog/ororatech-wildfires-from-space/, 2023.1,2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[36] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[37] Spacety. Chaohu-1 SAR satellite. https://www.satelliteevolution.com/post/two-new-satellites-launched-including-the-2nd-sar-satellite-with-improved-capabilities,2022.8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[38] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[39] YU Wenqi, CHENG Gong, WANG Meijun, YAO Yanqing, XIE Xingxing, YAO Xiwen, and HAN Junwei. Mar20: A benchmark for military aircraft recognition in remote sensing images. National Remote Sensing Bulletin, 27(12):2688-2696, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 194, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 249 + ], + "type": "text", + "content": "[40] Michael A Wulder, David P Roy, Volker C Radeloff, Thomas R Loveland, Martha C Anderson, David M Johnson, Sean Healey, Zhe Zhu, Theodore A Scambos, Nima Pahlevan, et al. Fifty years of landsat science and impacts. Remote Sensing of Environment, 280:113195, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[41] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 13668-13677, 2022. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "text", + "content": "[42] Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499–9508, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "text", + "content": "[43] Zhenghang Yuan, Lichao Mou, Yuansheng Hua, and Xiao Xiang Zhu. Rrasis: Referring remote sensing image segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[44] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-13, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "text", + "content": "[45] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multimodal large language model for multisensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 62:1-20, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 503, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 558 + ], + "type": "text", + "content": "[46] Yuanlin Zhang, Yuan Yuan, Yachuang Feng, and Xiaoqiang Lu. Hierarchical and robust convolutional neural network for very high-resolution remote sensing object detection. IEEE Transactions on Geoscience and Remote Sensing, 57 (8):5535-5548, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[47] Ziyuan Zhang, Han Qiu, Maosen Zhang, Jun Liu, Bin Chen, Tianwei Zhang, and Hewu Li. COSMIC: Compress satellite images efficiently via diffusion compensation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[48] Xiangyu Zhao, Yicheng Chen, Shilin Xu, Xiangtai Li, Xinjiang Wang, Yining Li, and Haian Huang. An open and comprehensive pipeline for unified object grounding and detection. arXiv preprint arXiv:2401.02361, 2024. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[49] Haigang Zhu, Xiaogang Chen, Weiqun Dai, Kun Fu, Qixiang Ye, and Jianbin Jiao. Orientation robust object detection in aerial images using deep convolutional neural network. In 2015 IEEE international conference on image processing (ICIP), pages 3735-3739. IEEE, 2015. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23309" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 127 + ], + "type": "text", + "content": "[50] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 5" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23310" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_content_list.json b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dc456b1c37cc4e4bdf24e2ca520a591c8f0388c0 --- /dev/null +++ b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_content_list.json @@ -0,0 +1,1526 @@ +[ + { + "type": "text", + "text": "VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow", + "text_level": 1, + "bbox": [ + 125, + 128, + 872, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ada Gorgun Bernt Schiele Jonas Fischer", + "bbox": [ + 316, + 203, + 679, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{agoerguen, schiele, jonas.fischer}@mpi-inf.mpg.de Max Planck Institute for Informatics, Saarland Informatics Campus", + "bbox": [ + 228, + 224, + 759, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7062af2ee59342e86cb87f9f3dacaa2efe3f66310c77b20ca873b87ff0dbb1f3.jpg", + "image_caption": [ + "Figure 1. Feature visualization through distribution matching. Unlike traditional feature visualization (FV) methods, which often produce artifacts or repetitive patterns, VITAL generates more understandable visualizations. Our approach scales effectively to modern architectures (rows), generalizes well across diverse classes (columns), and better captures meaningful network representations." + ], + "image_footnote": [], + "bbox": [ + 93, + 290, + 903, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 526, + 326, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural networks are widely adopted to solve complex and challenging tasks. Especially in high-stakes decision-making, understanding their reasoning process is crucial, yet proves challenging for modern deep networks. Feature visualization (FV) is a powerful tool to decode what information neurons are responding to and hence to better understand the reasoning behind such networks. In particular, in FV we generate human-understandable images that reflect the information detected by neurons of interest. However, current methods often yield unrecognizable visualizations, exhibiting repetitive patterns and visual artifacts that are hard to understand for a human. To address these problems, we propose to guide FV through statistics of real image features combined with measures of relevant network flow to generate prototypical images. Our approach yields human-understandable visualizations that both qualitatively and quantitatively improve over state-of-the-art FVs across various architectures. As such, it can be used to decode which information the network uses, complementing mechanistic circuits that identify where it is encoded.", + "bbox": [ + 88, + 558, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 526, + 643, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks have achieved remarkable success across science and industry. Their reasoning process is, however, inherently opaque. Different lines of research in Explainability of Machine Learning have been addressing these issues as understanding of how a neural network arrives at a decision is crucial, especially in safety-critical domains. To understand which input features are used for a decision, both post-hoc, as well as built-in methodologies have been proposed, which give information about input-output relations in the form of attribution or relevance scores [4, 17, 34, 37]. Yet, these offer only a partial understanding of a model's decision-making process. In the scope of Mechanistic Interpretability [3, 31], experts are interested in finding explanations that uncover what the model internally learns. In this line of research, methods that discover paths of information flow – so-called circuits – have been proposed [6, 10, 35], along with approaches to discover network components, such as groups of neurons, that together encode a concept [9]. While these methods reveal where information is propagated towards a particular decision, they fall short in identifying which information is encoded. Feature visualization (FV) bridges this gap by generating images that correspond to what triggers a strong response in", + "bbox": [ + 511, + 553, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at: VITAL Framework", + "bbox": [ + 109, + 886, + 331, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4403", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a given neuron or group of neurons. This is traditionally achieved by optimizing for an image that maximizes the response of the selected neuron(s) through gradient ascent. Complemented with circuits, FV thus help represent the internal reasoning process in human-understandable terms.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature Visualizations have been an integral part of generating explanations for deep neural networks, adopted early on for CNN models [33]. FVs for earlier architectures [18, 38] through explicit regularization [21, 23, 39] or transformations [22, 27] promote robustness against variations within the image. Already in these early works, irrelevant features as well as artificial, repetitive patterns rendered interpreting the visualizations challenging. For subsequent, larger networks, new methods had to be developed and further refined to get towards realistic image visualizations [8, 26, 42], incorporating elaborate regularizations or reparametrizations in the Fourier space to dampen high-frequency features. Still, newer methods also struggle with repetitive patterns and irrelevant features on modern architectures (see Fig. 4) and show deteriorating interpretability when scaling to larger models.", + "bbox": [ + 89, + 169, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we suggest to re-frame FV in terms of optimizing an alignment to realistic feature distributions instead of activation maximization. In particular, we consider the per-channel distribution of activations in intermediate layers across a set of reference images (e.g., training images of a class, top-activating images for a neuron) and optimize our feature visualization to follow a similar distribution of activations. As such, repetitive patterns – visible as a large number of high activations in a channel – are discouraged unless natural in the original data. Similarly, artificial features or colors, which might highly activate a neuron but are far from the input manifold, are discouraged (see Fig. 1). By further incorporating feature relevance for FVs of neurons for adjusting the feature distributions by how relevant they are to the neuron under study, we further remove irrelevant features from the visualizations that impacted the faithfulness of existing FVs. We provide an efficient implementation of our approach, VITAL, making use of a distribution matching algorithm through which we can back-propagate. This enables VITAL to seamlessly scale to modern architectures, including large networks and Vision Transformers (ViTs), but yields more interpretable FVs, which we show in a set of experiments including a human user study. Our contributions are:", + "bbox": [ + 89, + 414, + 482, + 775 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- we propose a novel method, VITAL, to optimize feature visualizations through alignment with feature distributions of real data instead of maximizing activations,", + "- we propose to incorporate feature relevance scores in the optimization to focus on exactly those features that are perceived by the target neuron from the feature visualization, and", + "- we show both qualitatively and quantitatively on a diverse" + ], + "bbox": [ + 91, + 780, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "set of metrics - including a human user study - that VI-TAL yields more understandable and accurate descriptions of what information is encoded in a neuron.", + "bbox": [ + 529, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 162, + 653, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mechanistic Interpretability. Understanding the internal decision-making process of neural networks has regained significant attention as part of Mechanistic Interpretability (MI) of neural networks. While definitions of MI can differ [31], the essence is to find explanations of neural networks that allow an interpretation of how internal components, such as neurons or layers, operate together to give reason to a prediction. We refer to recent reviews for a comprehensive summary [3, 31]. One key area of research is the discovery of circuits, which describe along which network paths information is propagated to a target neuron. We here discuss how to describe which information is encoded in neurons through feature visualization, which also has shown promising results in explaining the information encoded in circuits or feature directions [9, 10].", + "bbox": [ + 511, + 191, + 906, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Inversion. Model inversion techniques reconstruct input images from model outputs by optimizing images to match predictions, revealing learned representations and privacy risks. Recent works have improved inversion fidelity through data augmentations [11], distribution matching [42], reinforcement learning [12], and generative networks [43]. Unlike model inversion, which visualizes class-specific information only, FVs can capture information of any network component, including individual neurons, neuron combinations, and channels, to better understand network representations.", + "bbox": [ + 511, + 425, + 908, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature Visualization. Feature visualization techniques aim to generate an image that maximizes the activation of specific network parts (e.g., a neuron) [33, 44]. To yield understandable visualizations, researchers suggested to impose constraints based on local variance [21], apply blurring [23, 39], or ensure consistency across transformations [22, 26] to promote robustness. One of the most widely used approaches is called preconditioning [27], where optimization is performed on the Fourier basis, ensuring that data is decorrelated and whitened. Most recently, MACO [8] improved this preconditioning, suggesting to only optimize the phase while fixing the magnitude to that of a typical image. Unfortunately, good visualizations for modern architectures such as large ResNets [13] are still hard to achieve. To address this challenge, statistically learned priors based on generative models such as GANs [24, 43], or autoencoders [25, 41] are leveraged. A key drawback of these, compared to traditional FV, is that it is unclear what of the visualization can be attributed to the neural network under study and what is due to the generative model.", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4404", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/299ac9d8c8176995d0aa1de62a2152f6d1a3ae9142149169d99f6868ef6e8ec4.jpg", + "image_caption": [ + "Figure 2. Importance of relevance scores. Given an image selected by its relevance to a neuron detecting an ear, we apply LRP from this neuron back to a preceding building block. We show the activation (first column) and relevance maps (second column) of two neurons from this block, respectively. While the activation is high for the ears (top) and background (bottom), only the ear is relevant for the target neuron. The difference between VITAL with and without LRP (third column) shows that incorporating relevance helps to avoid visualizing irrelevant features." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 473, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VITAL Framework", + "text_level": 1, + "bbox": [ + 89, + 465, + 279, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing FV methods suffer from overly repetitive patterns, artificial, and irrelevant features. To address this, we propose aligning feature distributions between learned and real images across layers, capturing both coarse and fine-grained information (Sec. 3.2). This discourages the use of artificial features and repetitive patterns, which typically spike the activation in classic approaches. We further suggest to take into account relevance scores, matching the distributions of activations weighted by their relevance to the target, which discourages features in the FV that are irrelevant for the target neuron (sec. 3.3). This becomes particularly important for the visualization of inner neurons. For example, while grass should be irrelevant for a bird's beak, it might still coactivate channels across multiple bird images, which misleads interpretations. Incorporating LRP eliminates irrelevant correlations in activation maps, ensuring that only the features contributing to the target neuron's activation are visualized. We provide an example in Fig. 2 and further details about our approach in supp. 1.", + "bbox": [ + 89, + 491, + 483, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notation", + "text_level": 1, + "bbox": [ + 89, + 787, + 194, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a neural network $f(\\cdot ;\\theta): \\mathcal{X} \\to \\mathcal{Y}$ as a composition of $L$ functions corresponding to layers, i.e., $f = f^{(L)}\\circ f^{(L - 1)}\\circ \\ldots \\circ f^{(2)}\\circ f^{(1)}$ from an input space $\\mathcal{X}\\in \\mathbb{R}^{C_0\\times H\\times W}$ to an output space $\\mathcal{Y}\\in \\mathbb{R}^c$ . Here, $\\theta$ are the (trained) network parameters, $C_0$ is the number of channels (typically 1 or 3), and $c$ is the number of classes. For ease", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of notation, we will use $f^{(l)}(x) = f^{(l)} \\circ \\ldots \\circ f^{(2)} \\circ f^{(1)}(x)$ , to indicate the network output of layer $l$ on input $x \\in \\mathcal{X}$ . Without loss of generality, we denote the feature space of layer $l$ flattened over its spatial dimensions as $A^{(l)}(x) = f^{(l)}(x) \\in \\mathbb{R}^{C_l \\times D}$ , where $C_l$ is the number of channels.", + "bbox": [ + 511, + 89, + 908, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Distribution Matching", + "text_level": 1, + "bbox": [ + 511, + 183, + 785, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a target of interest in the network $f$ , e.g. the $i$ -th neuron in layer $l$ , $f_{i}^{(l)}$ , we are after an input $x^{*}$ that captures the information that $f_{i}^{(l)}$ responds to. Instead of finding an image that maximizes this activation, as typical in the literature, we consider the distribution of features $A(l')$ at layers $l' < l$ of both our generated input $x$ and inputs $x' \\in \\mathcal{X}_{ref}$ from reference data. Such references could be training data for which $f_{i}^{(l)}$ responds highly. We would like to have similar feature distributions of $x$ compared to these reference images $x'$ . More formally, we would like to find an image $x^{*}$ such that $\\forall l' < l, x' \\in \\mathcal{X}_{ref}$ :", + "bbox": [ + 511, + 207, + 906, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s t} \\left(A ^ {\\left(l ^ {\\prime}\\right)} \\left(x ^ {*}\\right)\\right) \\approx \\operatorname {d i s t} \\left(A ^ {\\left(l ^ {\\prime}\\right)} \\left(x ^ {\\prime}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 398, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{dist}(\\cdot)$ indicates the distribution of the activations. In Fig. 3, we illustrate the alignment of feature distributions for sample channels and blocks.", + "bbox": [ + 511, + 439, + 906, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Investigating empirical distributions of activations, we observe that they do not consistently follow a typical prior, such as a Gaussian, and hence decided to match empirical distributions directly. In particular, we leverage a result on sort-matching from the style-transfer literature [45], which allows us to compute a matching between two feature vectors, here $A^{(l^{\\prime})}(x^{*})$ and $A^{(l^{\\prime})}(x^{\\prime})$ , that we can back-propagate through. We can, hence, efficiently compute gradients with respect to $x^{*}$ . Accumulating the difference between the (sorted) feature vectors of our current generated image $x^{*}$ and the (sorted) feature vectors for all reference images, and across layers defines a loss function that reflects Statement 1. So, optimizing this loss with respect to the generated image $x^{*}$ , improves the alignment of the features of our generated image with real images.", + "bbox": [ + 511, + 487, + 908, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Sort-Matching Loss", + "text_level": 1, + "bbox": [ + 511, + 728, + 697, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key idea behind the sort-matching loss suggested by Zhang et al. [45] is to compute the difference of the distribution of two arrays, $z = A(l')(x^*)$ and $z' = A(l')(x')$ , by sorting these two arrays that allow for differentiation to $x^*$ . In particular, we first compute indices $\\pi$ and $\\pi'$ that sort the arrays $z$ and $z'$ , respectively. We then compute an indexing $\\bar{\\pi}, \\bar{\\pi}'$ that sorts these indices. Intriguingly, applying both indices yields the original order, i.e., $z = \\bar{\\pi}(\\pi(z))$ . We can now match $z$ , which is a function of $x^*$ , and $z^r = \\bar{\\pi}(\\pi'(z'))$ , which reorders the sorted values of $z'$ based on the indexing", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4405", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from $z$ . Computing the MSE between these two arrays", + "bbox": [ + 86, + 90, + 452, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M S E} \\left(z, z ^ {r}\\right) = \\frac {1}{| z |} \\sum_ {i = 1} ^ {| z |} \\left(z _ {i} - z _ {i} ^ {r}\\right), ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 118, + 482, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we can differentiate for $x^{*}$ , allowing for gradient-based optimization. For multiple real feature vectors, we average the transformed $\\pi'(z')$ vectors to obtain a representative prototype before applying $\\bar{\\pi}$ to obtain $z^r$ .", + "bbox": [ + 89, + 171, + 483, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Selection of Reference Images", + "text_level": 1, + "bbox": [ + 89, + 241, + 346, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically, the reference set should reflect the distribution relevant to the target neuron. For class neurons, we directly use a random selection of images from that class. For intermediate neurons, we select the top- $k$ patches (from distinct images) relevant for this neuron, following Fel et al. [9]. In brief, we crop and resize training images into patches and score each based on the global average pooled activation across spatial dimensions, keeping the top- $k$ patches from unique images. These define our reference $\\mathcal{X}_{ref}$ for feature alignment. We provide additional sampling experiments, including size sensitivity, cluster-based sampling, dataset generalization, and corruption analysis, in the supp. 3.4.", + "bbox": [ + 89, + 260, + 483, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Incorporating Relevance Scores", + "text_level": 1, + "bbox": [ + 89, + 452, + 372, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Especially for visualizations of intermediate neurons, irrelevantly activated features pose a problem. For example, a neuron detecting a facial feature of animals might be correlated with a background feature, such as grass, as this is predominant in training – an animal in nature. Classic approaches of activation maximization do not discourage such activations, which is why irrelevant features can appear in their visualizations. Similarly, matching of distributions of activations would encourage background features in the FV. To overcome this, we incorporate feature relevance scores that describe how much a feature (encoded by an intermediate neuron or channel) is relevant for the downstream target neuron. Methods such as layer-wise relevance propagation (LRP) provide meaningful relevance scores [17, 28]. We provide an example of irrelevant background activation and its actual relevance computed by LRP for the target neuron for ResNet50 trained on ImageNet, in Fig. 2.", + "bbox": [ + 89, + 474, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We thus suggest matching feature distributions based on activation times relevance, i.e., match features $A^{(l^{\\prime})}(x) \\odot R_{n}^{(l^{\\prime})}(x)$ , where $R_{n}$ are the relevance scores for the target neuron $n$ under study. These relevance scores give weight to each neuron (or pixel of a channel) describing how relevant the neuron is for target neuron $n$ on the given sample. They are hence of the same dimension as the activations, $\\odot$ thus indicating the Hadamard product. In principle, any attribution approach can be used as plug-in replacement to obtain relevance scores $R$ . For the remainder of the paper, we use LRP, one of the most prominent attribution methods,", + "bbox": [ + 89, + 732, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d11361a8b5b348d0df4b569a8b70865bcf7c1bb60c4aec4d0dbd845a01a838d.jpg", + "image_caption": [ + "Figure 3. Overview for feature distribution matching. Activations or activation $\\times$ relevance scores (A) are extracted from channels across network blocks for real and optimized images. For feature visualization, the distribution of the generated image $x^{*}$ is optimized to match those of the real reference images $x^{\\prime} \\in \\mathcal{X}_{ref}$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 87, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and provide a comparative result for Guided Backpropagation [36] in supp. 3.5. Besides intermediate neurons, we also applied relevance scores to class neurons in supp. 2.2.", + "bbox": [ + 511, + 508, + 906, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Further Considerations", + "text_level": 1, + "bbox": [ + 511, + 566, + 732, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transparency Map. Irrelevant areas of the generated image stay mostly unchanged during the optimization, essentially representing noise. Analogous to Fel et al. [8], we suggest using transparency maps based on the importance of the image location during optimization to show relevant image parts only. In brief, we accumulate the gradients of our loss (Eq. 2) across each step in the optimization. We thus ensure the identification of the areas that have been most attended to by the network during the generation of the image. In the supp. 3.3, we illustrate the effect of the transparency map for example visualizations.", + "bbox": [ + 511, + 593, + 906, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Auxiliary Regularization. While already working reasonably in practice, existing regularizations further reduce noise and small artifacts in the generated image. Similar to Yin et al. [42], we penalize the total variance (TV) and $\\ell_2$ norm of our FV with parameter $\\alpha_{\\mathrm{TV}}$ and $\\alpha_{\\ell_2}$ as in:", + "bbox": [ + 511, + 763, + 908, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a u x}} \\left(x ^ {*}\\right) = \\alpha_ {\\mathrm {T V}} \\mathcal {L} _ {\\mathrm {T V}} \\left(x ^ {*}\\right) + \\alpha_ {\\ell_ {2}} \\ell_ {2} \\left(x ^ {*}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 854, + 903, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x^{*}$ is the generated image. Accordingly, the total loss", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4406", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to be minimized becomes", + "bbox": [ + 89, + 90, + 264, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {V I T A L}} \\left(x ^ {*}, \\mathcal {X} _ {r e f}\\right) = \\mathcal {L} _ {\\text {S M}} \\left(x ^ {*}, \\mathcal {X} _ {r e f}\\right) + \\mathcal {L} _ {\\text {a u x}} \\left(x ^ {*}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 114, + 482, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_{ref}$ are the set of real reference images, $\\mathcal{L}_{\\mathrm{SM}}(x^*,\\mathcal{X}_{ref})$ is the loss term as specified in Eq. 2 summed over different layers. We ablate on different hyperparameter settings to show their impact on the final visualization in the supp. 3.2.", + "bbox": [ + 89, + 143, + 483, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Choice of Layers for Alignment. While agnostic to the architecture, it is more efficient to focus on aligning features from only few, important layers. To decide on which layers to consider, we conducted a small ablation study (see supp. 3.1) to examine the effects of layers on the visualizations. In a ResNet50, we find that only optimizing for feature alignment of first and last block output is sufficient to produce high-quality images, already capturing the important low-respectively high-level features in the image.", + "bbox": [ + 89, + 222, + 483, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visualization of Concepts. As part of Mechanistic Interpretability, people are interested in finding concept-based explanations of model behavior. These concepts might be feature directions encoded through multiple neurons in a layer, for example, discovered by CRAFT [9]. In VITAL, we modify the initialization of relevances of target neurons in LRP to reflect the weights given by the feature direction. Through this modification, VITAL can give meaning, to these feature directions. We provide technical details and results of this approach in the supp. 2.3.", + "bbox": [ + 89, + 361, + 483, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Runtime. Assuming that feature distributions of the reference images are precomputed, VITAL runs in time $O(|\\mathcal{X}_{ref}| * L * C * P \\log P)$ , for $|\\mathcal{X}_{ref}|$ reference images, $L$ layers considered for matching, $C$ (maximum) number of channels in those layers and $P$ the (maximum) number of pixels in a feature map. As discussed before, $L$ is usually small, in the following, we use $L = 5$ layers for matching. When integrating LRP scores in our matching, we additionally have the cost of one backward pass through the model. In practice, it takes roughly 40 seconds to generate an image including computation of distributions for reference images. For comparison, MACO takes 23-28 seconds and DeepInversion 1-3 minutes, depending on their setting. Similarly expensive as DeepInversion is the generation of inner neurons with VITAL ( $L = 4$ ), with 2-3 minutes including computation of distributions.", + "bbox": [ + 89, + 515, + 483, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 768, + 295, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate VITAL, we compare against Fourier [27] as standard baseline, DeepInversion (DeepInv) [42] as state-of-the-art class visualization, and MACO [8], which is generally considered state-of-the-art for both class and inner neuron visualizations. We implemented VITAL in PyTorch and compared on ResNet50 [13], DenseNet121 [14], ConvNext-base [19], ViT-L-16 [7], ViT-L-32 [7], both qualitatively", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/379a42b7b7e5f82d1504a89e9a4ef0b6b5058ca5698ffe9fa345a90bc898c684.jpg", + "image_caption": [ + "MACO" + ], + "image_footnote": [], + "bbox": [ + 535, + 104, + 606, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/70a81d70bc7c5eff4e641bf777c32c27a245da8481091f2433ca67a4b046992a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 103, + 683, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80913f9c382b2ee4b7aadbcd2e1fe4eac3205b944d7bb7592f3b1cbbfdac941e.jpg", + "image_caption": [ + "Fourier" + ], + "image_footnote": [], + "bbox": [ + 537, + 171, + 604, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/be446db0f208322adef5e425abf54d65e6a443cf84d00f33a969f696ce0a62fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 171, + 681, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bf0db90654135e08658379fca1f79166d520e0f48052025960b05989d7579497.jpg", + "image_caption": [ + "DeepInv", + "VITAL" + ], + "image_footnote": [], + "bbox": [ + 537, + 238, + 609, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/72fd616bd824805d15d55f58c0289ecfdf79d7bf53808759f1c5b656750f15bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 238, + 683, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59a6330d05a020c1e97d18b40f93642b5b57d3c185b1c8a6d3296e9cd1665d75.jpg", + "image_caption": [ + "maltesedog", + "(a) ResNet-50" + ], + "image_footnote": [], + "bbox": [ + 537, + 306, + 607, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c844ded66d007aa5d1d8a574d9bd780748264188fd3c78ceaa8538a9a5191293.jpg", + "image_caption": [ + "candle" + ], + "image_footnote": [], + "bbox": [ + 611, + 305, + 683, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c7f78ca95a4c04752a5a175167f68a79831437f50696ded0f29eb35df3770a5.jpg", + "image_caption": [ + "MACO" + ], + "image_footnote": [], + "bbox": [ + 738, + 104, + 805, + 155 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7be6a7f5a3191cd9be04d958bca41650ca11f7519cb7a98313385c59b5daed6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 104, + 880, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d0739116fe41118abcc312f6fd7b0b070a6e0571394edf30acdba7c886c0db9.jpg", + "image_caption": [ + "Fourier" + ], + "image_footnote": [], + "bbox": [ + 738, + 172, + 807, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0c281c87b9edc834fbc8cb8c74b2f284d22357bf41d37b01ba6b6c0e58168d9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 172, + 880, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af703f34688781888e08d7c0ebe936b718bf4f5990668c133e46e78d6a830b3b.jpg", + "image_caption": [ + "DeepInv", + "VITAL" + ], + "image_footnote": [], + "bbox": [ + 738, + 239, + 810, + 295 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed0cb756abdffa6046933b693a5eda586866192d7a09202e766fcb3b3a403d3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 239, + 883, + 295 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/41fdf855274d14efd093ac4dafb7dd7d1598423a1c33639e7fd57ec4465b7a50.jpg", + "image_caption": [ + "maltesedog" + ], + "image_footnote": [], + "bbox": [ + 738, + 306, + 808, + 361 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f00026c13de96a97ac190662d2fbbec1a493d9c72b23414c66bab74d979e2135.jpg", + "image_caption": [ + "iT-L-32", + "Figure 4. Example class visualizations. We show ImageNet class visualizations for (a) ResNet50 and (b) ViT-L-32.MACO and Fourier-based FV (top rows) produce repetitive, hard-to-interpret patterns. DeepInversion (3rd row) improves interpretability but introduces artifacts and lacks ViT compatibility. VITAL arguably yields much more interpretable visualizations. For additional results, including failure cases, we refer to supp. 2." + ], + "image_footnote": [], + "bbox": [ + 810, + 306, + 882, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and quantitatively. All architectures were pre-trained on ImageNet. We further conducted human user studies assessing human interpretability of the resulting visualizations. We provide details on implementation and setup of competing methods with more results in the supp. 1 and 2, respectively.", + "bbox": [ + 511, + 530, + 906, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 619, + 694, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visually, there is a stark difference between class visualizations of VITAL and existing work (see Fig. 4 for ResNet50 and ViT-L-32 and supp. 2 for all architectures). Where most existing methods, including MACO and Fourier-based activation maximization, show highly repetitive patterns, our approach provides much cleaner, more understandable representations of the class. While DeepInversion shows less repetition of patterns, there are often irrelevant background features, as in the case of the candle, or visual artifacts, as the blue and yellow color for the dog, dominating the image. Yet, we acknowledge that none of the methods is flawless: All approaches, including VITAL, show difficulties with complex spatial arrangements (supp. 2.6). Still, it is usually understandable what class is represented by VITAL. This becomes even more evident when comparing methods across more complex architectures (e.g., ViT, ConvNeXt), where most approaches struggle to remain inter", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4407", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d7cbf370372a58aad0393d5810c13449d0174368316fe3dcaf82c0304e11e5a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSetupAcc. Top1 (↑)FID (↓)Zero-Shot Prediction
Top1 (↑)Top5 (↑)
ResNet50ImageNet--69.1192.23
MACOr: 22429.43360.7412.8729.73
Fourierr: 22421.30422.446.7318.27
DeepInvbs: 64100.0035.7629.9055.20
VITAL99.9058.7966.6292.56
ConvNeXtImageNet--65.6689.80
MACOr: 22466.0762.557.2019.77
Fourierr: 22460.0759.602.778.30
VITAL99.973.9263.5390.30
DenseNet121ImageNet--70.6493.16
MACOr: 2249.201.809.3323.20
Fourierr: 22415.531.634.8712.17
DeepInvbs: 64100.000.2010.0025.47
VITAL99.930.2758.7086.93
ViT-L-16ImageNet--64.7889.31
MACOr: 22444.33946.963.9310.57
Fourierr: 22425.30990.511.675.13
VITAL99.80126.2968.1792.80
ViT-L-32ImageNet--65.8390.03
MACOr: 22424.872318.9017.5337.23
Fourierr: 22417.031983.0910.3028.10
VITAL89.60147.3355.9785.47
", + "bbox": [ + 98, + 89, + 477, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of methods on different architectures trained on Imagenet. We provide FID scores, CLIP Zero-shot prediction scores, and top-1 classification accuracy, indicating the best and second best. In the settings, \"r\" indicates the resolution of the visualization and \"bs\" is the used batch size.", + "bbox": [ + 89, + 467, + 482, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pretable. In contrast, VITAL consistently produces representations where the object and its class remain clear, highlighting its robustness and effectiveness.", + "bbox": [ + 89, + 575, + 482, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Analysis of the Embedding Space", + "text_level": 1, + "bbox": [ + 89, + 648, + 370, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To understand how realistic and distinct the visualizations are, we compare FVs of five different dog classes to original images in terms of their penultimate layer embedding. We show a 2D low-dimensional representation of the embeddings of original and generated images using tSNE [40] in Fig. 6. We find that the real images form five clusters neatly separating the classes. We further observe that fourier-based approaches with their repetitive, often recognizable patterns are collapsed on a single point, away from any real images. While DeepInversion does fall into the correct clusters, it is hard to recognize a dog in the generated images. VITAL is the only method that consistently falls into the center of the cluster, providing distinct dog features we can recognize as known dog breeds. We extend our results for other architectures in supp. 2.5.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Small Circuits", + "text_level": 1, + "bbox": [ + 513, + 90, + 660, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method enables interpreting what inner neurons encode. By selecting the top-3 channels in the last hidden layer most relevant to a class (via LRP), we can visualize local circuits that reveal class-specific patterns (Fig. 5). While most methods, except DeepInversion, support inner neuron visualization, we compare with MACO as the strongest baseline. For example, we see that the most relevant, distinguishing features for a zebra classification are the stripes and texture of their fur. For dogs, there are leg, ear, and specifically colored fur detectors evident in the model. Such specific coloring is necessary to distinguish the plenty of dog breeds available as classes in ImageNet. We also observe signs of overfitting, where the pineapple class is associated with neurons that detect groups of fruits, many of the training images show such a picture. MACO yields less interpretable results, particularly for animals, as confirmed in the next section.", + "bbox": [ + 511, + 111, + 906, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 513, + 367, + 707, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For quantitative evaluation, we consider the classification accuracy and the FID score [32] on the respected model, as well as CLIP zero-shot prediction performance [29] in Tab. 1 (see supp. Tab. 1 for the extended version). We also provide quantitative results for the intermediate neurons with the metrics proposed by [16] in supp. 2.1.", + "bbox": [ + 511, + 390, + 906, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Accuracy. To assess how well visualizations describe their target, we classify each class neuron's visualization (see Tab. 1). Despite explicitly optimizing class logits, most existing methods perform poorly. Only VITAL and DeepInversion are able to produce accurate descriptions, achieving virtually $100\\%$ accuracy. Investigating further, we observe that other methods including MACO seem to produce irrelevant features that mislead the model, producing predictions unrelated to the original class. We give two examples of prediction in supp. 2.4.", + "bbox": [ + 511, + 484, + 908, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FID Score. For all architectures, to quantify visualization realism, we compute FID scores [32] using features from ResNet50's penultimate layer (Tab. 1). While aligned with our objective, the performance gap to other methods is substantial, with VITAL achieving over an order of magnitude better scores. DeepInversion also leverages feature statistics but is highly sensitive to batch size, which impacts both FID and visual quality.", + "bbox": [ + 511, + 640, + 908, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP Zero-shot Prediction. So far, previous metrics relied on the model used for visualization. To assess general perceptual quality, we use a pretrained CLIP ViT-B/32 model to classify FVs via template-based prompts (see supp. 1.2 for more details). We report zero-shot scores in Tab. 1, including original (correctly classified) ImageNet images for reference. As before, we see that DeepInversion outperforms existing work, with standard baselines having overall poor performance even when considering Top-5 accuracy. No", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4408", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3439541d05f3c853dd2a59c106793cd19781c8b25fb7d084f8fcd4d69b07ab6b.jpg", + "image_caption": [ + "Figure 5. Visualizations of small \"circuits\". We provide visualizations of small \"circuits\" for three interesting classes of ImageNet. For each class, we found the three most relevant neurons in the penultimate layer based on LRP relevance scores and indicate the neuron ID on the arrow. For each neuron, we label its likely meaning and provide FVs by MACO and VITAL. The Zebra class shows high dependence on the characteristic pattern of the fur, the dog class a dependence on specific color patterns to be able to distinguish the abundant different dog classes in ImageNet, and Pineapple shows signs of overfitting with an association with collections of fruits." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fd55888b520ffd7ea626b9d8dfd43c7e92791b0eee0631893c9ba0f52a201af.jpg", + "image_caption": [ + "Figure 6. $t$ -SNE projection of embedding. We show a low-dimensional tSNE embedding of the features at the penultimate layer for five dog breeds indicated by color. Transparent circles are original training images and FVs are indicated by symbols:" + ], + "image_footnote": [ + "VITAL, $\\triangleleft$ : MACO, $\\bullet$ : Fourier, $\\star$ : DeepInv." + ], + "bbox": [ + 112, + 343, + 464, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tably, VITAL outperforms all others by a wide margin, approaching real image accuracy.", + "bbox": [ + 89, + 670, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Human Interpretability Study", + "text_level": 1, + "bbox": [ + 89, + 712, + 357, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The primary objective of FVs is to make neural networks more understandable for human users. To address this goal, we conducted a user study (58 participants) to evaluate the interpretability of a FV from the perspective of a human. Specifically, participants (1) given a single word, evaluated how well a FV reflects the provided word, (2) evaluated how well the FVs for an inner neuron reflect the provided reference images (highly activating on the target neuron), and (3) should describe a given generated image with a word or short description. These present settings of increasing difficulty and realism. While (1) is a typical evaluation, it", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "also comes with the largest implicit bias, as the class label already gives away what we want to visualize. In (2), by showing reference images, we encourage a visual match between FV and real image. We compare with MACO, as it is the best performing approach to visualize inner neurons. In (3), we model a real-world scenario, i.e., without any implicit bias, the user is supposed to understand what the network encodes. We provided a simple scoring system from 1(worst) to 5(best) for parts (1) and (2), and evaluated how well the textual description in (3) fits the class using closeness in the embedding space of a text encoder [5].", + "bbox": [ + 511, + 345, + 906, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide an overview of the results of our study in Fig. 7. We observe that our method performs favorable across all three tasks compared to other methods, showing an order of magnitude more high-scoring visualizations (scores 4-5) in both (1) and (2), and showing much higher similarity to the correct class on task (3). In particular, for this challenging task that provides a zero-shot estimate of the interpretability of FVs, the $5\\%$ quantile, i.e., the \"worse\"-performing visualizations, are still close to a similarity of .5 of VITAL. Other methods show only 0.25 similarity for the same quantile. The same stark difference of more than 0.2 in median similarity score shows that VITAL provides more human-understandable visualizations of what neurons encode. We provide more fine-grained analysis, population statistics, and details in supp. 4. We further conducted a human user study following the protocol of [8, 46], which is a simple option-based evaluation similar to setting (2), where VITAL performs best (supp. 4).", + "bbox": [ + 511, + 513, + 908, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 799, + 627, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments show that VITAL produces more understandable visualizations that better capture the unique information encoded by each neuron, outperforming MACO [8], the most recent method in classification accuracy on the target model. While DeepInversion performs similarly, this is", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4409", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7575397bc50f6405f2906b72ae7b7e4c7a3982c0b68e03d6017143da4ea92d2.jpg", + "image_caption": [ + "User Study Section 1 results (reference word)", + "(a) Class visualization with given class label." + ], + "image_footnote": [], + "bbox": [ + 133, + 102, + 433, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a2a72e2396857a4eb1b062b1844e7d6c11a26070357a05c6d71289e70de8559.jpg", + "image_caption": [ + "User Study Section 2 results (reference crops)", + "(b) Inner neuron visualization with given activating images." + ], + "image_footnote": [], + "bbox": [ + 133, + 263, + 433, + 387 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1ffd91c692f3594e4ddd1a772fabfb0d2582adc5fbe9814a23380570ac41465.jpg", + "image_caption": [ + "User Study Section 3 results (free form)", + "(c) Class visualization with free-form labeling.", + "Figure 7. Human Study results. We provide summary statistics on the scores for the different methods obtained for the three-part human $(n = 58)$ evaluation study. In (a), participants were given class visualization and corresponding class name for 10 random classes and scored the visualizations on a scale of 1(worst) to 5(best). In (b) the task was to evaluate the closeness of FVs for an inner neuron to the image crops that it most activates on (same scoring). DeepInversion does not generate visualizations for inner neurons. In (c), participants were asked to label a given FV. We evaluated similarity of their label to the (hidden) ground truth based on Universal Sentence Encoder embeddings. We provide violin plots with median and $5\\%$ and $95\\%$ quantiles of the achieved similarity." + ], + "image_footnote": [], + "bbox": [ + 135, + 420, + 437, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "expected since it directly optimizes the model's logit score. When evaluated by a different model such as CLIP, VITAL surpasses all baselines by a wide margin. Two human user studies further confirm that VITAL yields more interpretable visualizations in describing given labels or images.", + "bbox": [ + 89, + 773, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Realism. In principle, our visualizations, as well as those of other methods, do not reach photo-realism. VITAL visualizations sometimes resemble the style of paintings of", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Édouard Manet, which, however, does not limit their interpretability: Similar to a painting, it is understandable what is expressed, which is also supported by the user studies. Yet, it would be interesting to study how to reach photorealism in the future. Generative models such as diffusion offer a promising direction, though further research is needed to prevent the introduction of model-specific artifacts that reduce faithfulness to the target model.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation. We evaluate our framework using a broad set of metrics, combining standard benchmarks with additional measures like CLIP prompting. Unfortunately, in feature visualization, more broadly in Explainability research, exhaustive benchmarks and clearly defined metrics are actively discussed, but currently still amiss. Through the additional human user studies, we complement the metrics with an evaluation of our original goal of more understandable feature visualizations. As with any user study, the design can be debated, and more refined setups with larger cohorts offer a promising direction for future work.", + "bbox": [ + 511, + 215, + 906, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distribution Matching. Standard approaches to distribution matching include matching statistics, such as mean and variances, or considering prior distributions [15, 20], but these oversimplify the complexity of real images. Activation distributions are often too diverse for such approximations [45]. Hence, directly optimizing an alignment between the empirical distributions showed the most promise, with sort matching [30] as an efficient approach that easily scales to modern networks, but our approach is agnostic to the type of distribution matching algorithm.", + "bbox": [ + 511, + 385, + 905, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work. VITAL provides a substantial improvement over existing work and we anticipate it of great use in downstream applications, such as understanding networks in the medical domain, studying what neurons encode after knowledge transfer, or how pruning affects the representations. Future directions include automatic selection of neurons or neuron sets for circuit visualization, for example using CRP [1], integration of ViT-specific visualization via AttnLRP [2], and extending our analysis to CLIP-based ViT models to better understand multimodal representations.", + "bbox": [ + 511, + 539, + 908, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature visualization is key to understanding neural representations, but existing methods often yield artifacts or irrelevant patterns. We propose a new approach that aligns the feature distributions of generated and real images, and incorporates relevance scores to better capture the true information encoded in intermediate neurons. Across multiple metrics and two user studies, VITAL produces clearer and more interpretable visualizations for both humans and models. We envision VITAL supporting safety-critical fields such as medicine and complementing Mechanistic Interpretability by attributing meaning to neural circuits.", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4410", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 8", + "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. AttnLRP: Attention-aware layer-wise relevance propagation for transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 135-168. PMLR, 2024. 8", + "[3] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research (TMLR), 2024. Survey Certification, Expert Certification. 1, 2", + "[4] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 10329-10338, 2022. 1", + "[5] Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. Universal sentence encoder. CoRR, abs/1803.11175, 2018. 7", + "[6] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. In The Advances in Neural Information Processing Systems (NIPS), pages 16318-16352. Curran Associates, Inc., 2023. 1", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations (ICLR), 2021. 5", + "[8] Thomas Fel, Thibaut Boissin, Victor Boutin, Agustin Picard, Paul Novello, Julien Colin, Drew Linsley, Tom Rousseau, Remi Cadene, Lore Goetschalckx, Laurent Gardes, and Thomas Serre. Unlocking feature visualization for deep network with magnitude constrained optimization. In The Advances in Neural Information Processing Systems (NIPS), pages 37813-37826. Curran Associates, Inc., 2023. 2, 4, 5, 7", + "[9] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadécnc, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2711-2721, 2023. 1, 2, 4, 5", + "[10] Jonas Fischer, Anna Olah, and Jilles Vreeken. What's in the box? exploring the inner life of neural networks with robust rules. In The International Conference on Machine Learning (ICML), pages 3352-3362, 2021. 1, 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Amin Ghiasi, Hamid Kazemi, Steven Reich, Chen Zhu, Micah Goldblum, and Tom Goldstein. Plug-in inversion: Model-agnostic inversion for vision with data augmentations. In The International Conference on Machine Learning (ICML), pages 7484–7512, 2022. 2", + "[12] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 20504-20513, 2023. 2", + "[13] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 5", + "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 5", + "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In The IEEE International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 8", + "[16] Laura Kopf, Pheline Lou Bommer, Anna Hedström, Sebastian Lapuschkin, Marina MC Höhne, and Kirill Bykov. Cosy: Evaluating textual explanations of neurons. In The Advanced Conference on Neural Information Processing Systems (NIPS), 2024. 6", + "[17] Sebastian Lapuschkin, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10:e0130140, 2015. 1, 4", + "[18] Shuying Liu and Weihong Deng. Very deep convolutional neural network based image classification using small training sample size. In IAPRA Asian Conference on Pattern Recognition (ACPR), pages 730-734, 2015. 2", + "[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 5", + "[20] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In The IEEE International Conference on Computer Vision (ICCV), pages 5951-5960, 2019. 8", + "[21] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 2", + "[22] A. Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 2", + "[23] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2015. 2", + "[24] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4411", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In The Advances in Neural Information Processing Systems (NIPS), page 3395-3403, 2016. 2", + "[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 3510-3520, 2017. 2", + "[26] Anh Mai Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. CoRR, abs/1602.03616, 2016. 2", + "[27] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 2, 5", + "[28] Seitaro Otsuki, Tsumugi Iida, Félix Doublet, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi, and Komei Sugiura. Layer-wise relevance propagation with conservation property for resnet. In The European Conference on Computer Vision (ECCV), pages 349-364, 2024. 4", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In The International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 6", + "[30] Jannick P. Rolland, V. Vo, B. Bloss, and Craig K. Abbey. Fast algorithms for histogram matching: application to texture synthesis. Journal of Electronic Imaging, 9(1):39 - 45, 2000. 8", + "[31] Naomi Saphra and Sarah Wiegreff. Mechanistic? In The BlackBoxNLP Workshop at The Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2", + "[32] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6", + "[33] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. CoRR, abs/1312.6034, 2013. 2", + "[34] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. CoRR, abs/1706.03825, 2017. 1", + "[35] Arshia Soltani Moakhar, Eugenia Iofinova, Elias Frantar, and Dan Alistarh. SPADE: Sparsity-guided debugging for deep neural networks. In The International Conference on Machine Learning (ICML), pages 45955-45987, 2024. 1", + "[36] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In International Conference on Learning Representations (ICLR), 2015. 4", + "[37] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In The International Conference on Machine Learning (ICML), page 3319-3328, 2017. 1", + "[38] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 2", + "[39] Mike Tyka. Class visualization with bilateral filters. https://science.miketyka.com/deepdream/2016/02/05/bilateral-class-vis.html, 2016.2", + "[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 6", + "[41] Guangrun Wang and Philip H. S. Torr. Traditional classification neural networks are good generators: They are competitive with ddpms and gans. CoRR, abs/1706.03825, 2022. 2", + "[42] Hongxu Yin, Pavlo Molchanov, Jose M. Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2020. 2, 4, 5", + "[43] Runpeng Yu and Xinchao Wang. Generator born from classifier. In The Advances in Neural Information Processing Systems (NIPS), 2023. 2", + "[44] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In The European Conference on Computer Vision (ECCV), pages 818-833, 2014. 2", + "[45] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2022. 3, 8", + "[46] Roland S. Zimmermann, Judy Borowski, Robert Geirhos, Matthias Bethge, Thomas S. A. Wallis, and Wieland Brendel. How well do feature visualizations support causal understanding of cnn activations? The Advances in Neural Information Processing Systems (NIPS), 34:24369-24381, 2021. 7" + ], + "bbox": [ + 516, + 92, + 903, + 584 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4412", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_model.json b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c0b864b133046b950c56df07c86e1b636cb859c --- /dev/null +++ b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.13, + 0.873, + 0.177 + ], + "angle": 0, + "content": "VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.204, + 0.68, + 0.222 + ], + "angle": 0, + "content": "Ada Gorgun Bernt Schiele Jonas Fischer" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.225, + 0.761, + 0.258 + ], + "angle": 0, + "content": "{agoerguen, schiele, jonas.fischer}@mpi-inf.mpg.de Max Planck Institute for Informatics, Saarland Informatics Campus" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.291, + 0.905, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 1. Feature visualization through distribution matching. Unlike traditional feature visualization (FV) methods, which often produce artifacts or repetitive patterns, VITAL generates more understandable visualizations. Our approach scales effectively to modern architectures (rows), generalizes well across diverse classes (columns), and better captures meaningful network representations." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.527, + 0.327, + 0.542 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.559, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Neural networks are widely adopted to solve complex and challenging tasks. Especially in high-stakes decision-making, understanding their reasoning process is crucial, yet proves challenging for modern deep networks. Feature visualization (FV) is a powerful tool to decode what information neurons are responding to and hence to better understand the reasoning behind such networks. In particular, in FV we generate human-understandable images that reflect the information detected by neurons of interest. However, current methods often yield unrecognizable visualizations, exhibiting repetitive patterns and visual artifacts that are hard to understand for a human. To address these problems, we propose to guide FV through statistics of real image features combined with measures of relevant network flow to generate prototypical images. Our approach yields human-understandable visualizations that both qualitatively and quantitatively improve over state-of-the-art FVs across various architectures. As such, it can be used to decode which information the network uses, complementing mechanistic circuits that identify where it is encoded." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.527, + 0.645, + 0.542 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Deep neural networks have achieved remarkable success across science and industry. Their reasoning process is, however, inherently opaque. Different lines of research in Explainability of Machine Learning have been addressing these issues as understanding of how a neural network arrives at a decision is crucial, especially in safety-critical domains. To understand which input features are used for a decision, both post-hoc, as well as built-in methodologies have been proposed, which give information about input-output relations in the form of attribution or relevance scores [4, 17, 34, 37]. Yet, these offer only a partial understanding of a model's decision-making process. In the scope of Mechanistic Interpretability [3, 31], experts are interested in finding explanations that uncover what the model internally learns. In this line of research, methods that discover paths of information flow – so-called circuits – have been proposed [6, 10, 35], along with approaches to discover network components, such as groups of neurons, that together encode a concept [9]. While these methods reveal where information is propagated towards a particular decision, they fall short in identifying which information is encoded. Feature visualization (FV) bridges this gap by generating images that correspond to what triggers a strong response in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.332, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at: VITAL Framework" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4403" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "a given neuron or group of neurons. This is traditionally achieved by optimizing for an image that maximizes the response of the selected neuron(s) through gradient ascent. Complemented with circuits, FV thus help represent the internal reasoning process in human-understandable terms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Feature Visualizations have been an integral part of generating explanations for deep neural networks, adopted early on for CNN models [33]. FVs for earlier architectures [18, 38] through explicit regularization [21, 23, 39] or transformations [22, 27] promote robustness against variations within the image. Already in these early works, irrelevant features as well as artificial, repetitive patterns rendered interpreting the visualizations challenging. For subsequent, larger networks, new methods had to be developed and further refined to get towards realistic image visualizations [8, 26, 42], incorporating elaborate regularizations or reparametrizations in the Fourier space to dampen high-frequency features. Still, newer methods also struggle with repetitive patterns and irrelevant features on modern architectures (see Fig. 4) and show deteriorating interpretability when scaling to larger models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.776 + ], + "angle": 0, + "content": "In this work, we suggest to re-frame FV in terms of optimizing an alignment to realistic feature distributions instead of activation maximization. In particular, we consider the per-channel distribution of activations in intermediate layers across a set of reference images (e.g., training images of a class, top-activating images for a neuron) and optimize our feature visualization to follow a similar distribution of activations. As such, repetitive patterns – visible as a large number of high activations in a channel – are discouraged unless natural in the original data. Similarly, artificial features or colors, which might highly activate a neuron but are far from the input manifold, are discouraged (see Fig. 1). By further incorporating feature relevance for FVs of neurons for adjusting the feature distributions by how relevant they are to the neuron under study, we further remove irrelevant features from the visualizations that impacted the faithfulness of existing FVs. We provide an efficient implementation of our approach, VITAL, making use of a distribution matching algorithm through which we can back-propagate. This enables VITAL to seamlessly scale to modern architectures, including large networks and Vision Transformers (ViTs), but yields more interpretable FVs, which we show in a set of experiments including a human user study. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- we propose a novel method, VITAL, to optimize feature visualizations through alignment with feature distributions of real data instead of maximizing activations," + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- we propose to incorporate feature relevance scores in the optimization to focus on exactly those features that are perceived by the target neuron from the feature visualization, and" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- we show both qualitatively and quantitatively on a diverse" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "set of metrics - including a human user study - that VI-TAL yields more understandable and accurate descriptions of what information is encoded in a neuron." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.163, + 0.655, + 0.181 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.193, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Mechanistic Interpretability. Understanding the internal decision-making process of neural networks has regained significant attention as part of Mechanistic Interpretability (MI) of neural networks. While definitions of MI can differ [31], the essence is to find explanations of neural networks that allow an interpretation of how internal components, such as neurons or layers, operate together to give reason to a prediction. We refer to recent reviews for a comprehensive summary [3, 31]. One key area of research is the discovery of circuits, which describe along which network paths information is propagated to a target neuron. We here discuss how to describe which information is encoded in neurons through feature visualization, which also has shown promising results in explaining the information encoded in circuits or feature directions [9, 10]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Model Inversion. Model inversion techniques reconstruct input images from model outputs by optimizing images to match predictions, revealing learned representations and privacy risks. Recent works have improved inversion fidelity through data augmentations [11], distribution matching [42], reinforcement learning [12], and generative networks [43]. Unlike model inversion, which visualizes class-specific information only, FVs can capture information of any network component, including individual neurons, neuron combinations, and channels, to better understand network representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Feature Visualization. Feature visualization techniques aim to generate an image that maximizes the activation of specific network parts (e.g., a neuron) [33, 44]. To yield understandable visualizations, researchers suggested to impose constraints based on local variance [21], apply blurring [23, 39], or ensure consistency across transformations [22, 26] to promote robustness. One of the most widely used approaches is called preconditioning [27], where optimization is performed on the Fourier basis, ensuring that data is decorrelated and whitened. Most recently, MACO [8] improved this preconditioning, suggesting to only optimize the phase while fixing the magnitude to that of a typical image. Unfortunately, good visualizations for modern architectures such as large ResNets [13] are still hard to achieve. To address this challenge, statistically learned priors based on generative models such as GANs [24, 43], or autoencoders [25, 41] are leveraged. A key drawback of these, compared to traditional FV, is that it is unclear what of the visualization can be attributed to the neural network under study and what is due to the generative model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4404" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.475, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.487, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Importance of relevance scores. Given an image selected by its relevance to a neuron detecting an ear, we apply LRP from this neuron back to a preceding building block. We show the activation (first column) and relevance maps (second column) of two neurons from this block, respectively. While the activation is high for the ears (top) and background (bottom), only the ear is relevant for the target neuron. The difference between VITAL with and without LRP (third column) shows that incorporating relevance helps to avoid visualizing irrelevant features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.467, + 0.281, + 0.483 + ], + "angle": 0, + "content": "3. VITAL Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.485, + 0.78 + ], + "angle": 0, + "content": "Existing FV methods suffer from overly repetitive patterns, artificial, and irrelevant features. To address this, we propose aligning feature distributions between learned and real images across layers, capturing both coarse and fine-grained information (Sec. 3.2). This discourages the use of artificial features and repetitive patterns, which typically spike the activation in classic approaches. We further suggest to take into account relevance scores, matching the distributions of activations weighted by their relevance to the target, which discourages features in the FV that are irrelevant for the target neuron (sec. 3.3). This becomes particularly important for the visualization of inner neurons. For example, while grass should be irrelevant for a bird's beak, it might still coactivate channels across multiple bird images, which misleads interpretations. Incorporating LRP eliminates irrelevant correlations in activation maps, ensuring that only the features contributing to the target neuron's activation are visualized. We provide an example in Fig. 2 and further details about our approach in supp. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.195, + 0.803 + ], + "angle": 0, + "content": "3.1. Notation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We consider a neural network \\( f(\\cdot ;\\theta): \\mathcal{X} \\to \\mathcal{Y} \\) as a composition of \\( L \\) functions corresponding to layers, i.e., \\( f = f^{(L)}\\circ f^{(L - 1)}\\circ \\ldots \\circ f^{(2)}\\circ f^{(1)} \\) from an input space \\( \\mathcal{X}\\in \\mathbb{R}^{C_0\\times H\\times W} \\) to an output space \\( \\mathcal{Y}\\in \\mathbb{R}^c \\). Here, \\( \\theta \\) are the (trained) network parameters, \\( C_0 \\) is the number of channels (typically 1 or 3), and \\( c \\) is the number of classes. For ease" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.909, + 0.168 + ], + "angle": 0, + "content": "of notation, we will use \\( f^{(l)}(x) = f^{(l)} \\circ \\ldots \\circ f^{(2)} \\circ f^{(1)}(x) \\), to indicate the network output of layer \\( l \\) on input \\( x \\in \\mathcal{X} \\). Without loss of generality, we denote the feature space of layer \\( l \\) flattened over its spatial dimensions as \\( A^{(l)}(x) = f^{(l)}(x) \\in \\mathbb{R}^{C_l \\times D} \\), where \\( C_l \\) is the number of channels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.184, + 0.787, + 0.2 + ], + "angle": 0, + "content": "3.2. Feature Distribution Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Given a target of interest in the network \\( f \\), e.g. the \\( i \\)-th neuron in layer \\( l \\), \\( f_{i}^{(l)} \\), we are after an input \\( x^{*} \\) that captures the information that \\( f_{i}^{(l)} \\) responds to. Instead of finding an image that maximizes this activation, as typical in the literature, we consider the distribution of features \\( A(l') \\) at layers \\( l' < l \\) of both our generated input \\( x \\) and inputs \\( x' \\in \\mathcal{X}_{ref} \\) from reference data. Such references could be training data for which \\( f_{i}^{(l)} \\) responds highly. We would like to have similar feature distributions of \\( x \\) compared to these reference images \\( x' \\). More formally, we would like to find an image \\( x^{*} \\) such that \\( \\forall l' < l, x' \\in \\mathcal{X}_{ref} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.399, + 0.907, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s t} \\left(A ^ {\\left(l ^ {\\prime}\\right)} \\left(x ^ {*}\\right)\\right) \\approx \\operatorname {d i s t} \\left(A ^ {\\left(l ^ {\\prime}\\right)} \\left(x ^ {\\prime}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.485 + ], + "angle": 0, + "content": "where \\(\\mathrm{dist}(\\cdot)\\) indicates the distribution of the activations. In Fig. 3, we illustrate the alignment of feature distributions for sample channels and blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.909, + 0.715 + ], + "angle": 0, + "content": "Investigating empirical distributions of activations, we observe that they do not consistently follow a typical prior, such as a Gaussian, and hence decided to match empirical distributions directly. In particular, we leverage a result on sort-matching from the style-transfer literature [45], which allows us to compute a matching between two feature vectors, here \\( A^{(l^{\\prime})}(x^{*}) \\) and \\( A^{(l^{\\prime})}(x^{\\prime}) \\), that we can back-propagate through. We can, hence, efficiently compute gradients with respect to \\( x^{*} \\). Accumulating the difference between the (sorted) feature vectors of our current generated image \\( x^{*} \\) and the (sorted) feature vectors for all reference images, and across layers defines a loss function that reflects Statement 1. So, optimizing this loss with respect to the generated image \\( x^{*} \\), improves the alignment of the features of our generated image with real images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.729, + 0.698, + 0.744 + ], + "angle": 0, + "content": "3.2.1. Sort-Matching Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The key idea behind the sort-matching loss suggested by Zhang et al. [45] is to compute the difference of the distribution of two arrays, \\( z = A(l')(x^*) \\) and \\( z' = A(l')(x') \\), by sorting these two arrays that allow for differentiation to \\( x^* \\). In particular, we first compute indices \\( \\pi \\) and \\( \\pi' \\) that sort the arrays \\( z \\) and \\( z' \\), respectively. We then compute an indexing \\( \\bar{\\pi}, \\bar{\\pi}' \\) that sorts these indices. Intriguingly, applying both indices yields the original order, i.e., \\( z = \\bar{\\pi}(\\pi(z)) \\). We can now match \\( z \\), which is a function of \\( x^* \\), and \\( z^r = \\bar{\\pi}(\\pi'(z')) \\), which reorders the sorted values of \\( z' \\) based on the indexing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4405" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.088, + 0.092, + 0.454, + 0.108 + ], + "angle": 0, + "content": "from \\( z \\). Computing the MSE between these two arrays" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.119, + 0.483, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M S E} \\left(z, z ^ {r}\\right) = \\frac {1}{| z |} \\sum_ {i = 1} ^ {| z |} \\left(z _ {i} - z _ {i} ^ {r}\\right), ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.485, + 0.234 + ], + "angle": 0, + "content": "we can differentiate for \\( x^{*} \\), allowing for gradient-based optimization. For multiple real feature vectors, we average the transformed \\( \\pi'(z') \\) vectors to obtain a representative prototype before applying \\( \\bar{\\pi} \\) to obtain \\( z^r \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.242, + 0.347, + 0.258 + ], + "angle": 0, + "content": "3.2.2. Selection of Reference Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.444 + ], + "angle": 0, + "content": "Typically, the reference set should reflect the distribution relevant to the target neuron. For class neurons, we directly use a random selection of images from that class. For intermediate neurons, we select the top-\\(k\\) patches (from distinct images) relevant for this neuron, following Fel et al. [9]. In brief, we crop and resize training images into patches and score each based on the global average pooled activation across spatial dimensions, keeping the top-\\(k\\) patches from unique images. These define our reference \\(\\mathcal{X}_{ref}\\) for feature alignment. We provide additional sampling experiments, including size sensitivity, cluster-based sampling, dataset generalization, and corruption analysis, in the supp. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.373, + 0.47 + ], + "angle": 0, + "content": "3.3. Incorporating Relevance Scores" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Especially for visualizations of intermediate neurons, irrelevantly activated features pose a problem. For example, a neuron detecting a facial feature of animals might be correlated with a background feature, such as grass, as this is predominant in training – an animal in nature. Classic approaches of activation maximization do not discourage such activations, which is why irrelevant features can appear in their visualizations. Similarly, matching of distributions of activations would encourage background features in the FV. To overcome this, we incorporate feature relevance scores that describe how much a feature (encoded by an intermediate neuron or channel) is relevant for the downstream target neuron. Methods such as layer-wise relevance propagation (LRP) provide meaningful relevance scores [17, 28]. We provide an example of irrelevant background activation and its actual relevance computed by LRP for the target neuron for ResNet50 trained on ImageNet, in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We thus suggest matching feature distributions based on activation times relevance, i.e., match features \\( A^{(l^{\\prime})}(x) \\odot R_{n}^{(l^{\\prime})}(x) \\), where \\( R_{n} \\) are the relevance scores for the target neuron \\( n \\) under study. These relevance scores give weight to each neuron (or pixel of a channel) describing how relevant the neuron is for target neuron \\( n \\) on the given sample. They are hence of the same dimension as the activations, \\( \\odot \\) thus indicating the Hadamard product. In principle, any attribution approach can be used as plug-in replacement to obtain relevance scores \\( R \\). For the remainder of the paper, we use LRP, one of the most prominent attribution methods," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.088, + 0.905, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.41, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Figure 3. Overview for feature distribution matching. Activations or activation \\(\\times\\) relevance scores (A) are extracted from channels across network blocks for real and optimized images. For feature visualization, the distribution of the generated image \\(x^{*}\\) is optimized to match those of the real reference images \\(x^{\\prime} \\in \\mathcal{X}_{ref}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.556 + ], + "angle": 0, + "content": "and provide a comparative result for Guided Backpropagation [36] in supp. 3.5. Besides intermediate neurons, we also applied relevance scores to class neurons in supp. 2.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.568, + 0.733, + 0.584 + ], + "angle": 0, + "content": "3.4. Further Considerations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Transparency Map. Irrelevant areas of the generated image stay mostly unchanged during the optimization, essentially representing noise. Analogous to Fel et al. [8], we suggest using transparency maps based on the importance of the image location during optimization to show relevant image parts only. In brief, we accumulate the gradients of our loss (Eq. 2) across each step in the optimization. We thus ensure the identification of the areas that have been most attended to by the network during the generation of the image. In the supp. 3.3, we illustrate the effect of the transparency map for example visualizations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.909, + 0.843 + ], + "angle": 0, + "content": "Auxiliary Regularization. While already working reasonably in practice, existing regularizations further reduce noise and small artifacts in the generated image. Similar to Yin et al. [42], we penalize the total variance (TV) and \\(\\ell_2\\) norm of our FV with parameter \\(\\alpha_{\\mathrm{TV}}\\) and \\(\\alpha_{\\ell_2}\\) as in:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.855, + 0.905, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a u x}} \\left(x ^ {*}\\right) = \\alpha_ {\\mathrm {T V}} \\mathcal {L} _ {\\mathrm {T V}} \\left(x ^ {*}\\right) + \\alpha_ {\\ell_ {2}} \\ell_ {2} \\left(x ^ {*}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(x^{*}\\) is the generated image. Accordingly, the total loss" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4406" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.265, + 0.106 + ], + "angle": 0, + "content": "to be minimized becomes" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.116, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {V I T A L}} \\left(x ^ {*}, \\mathcal {X} _ {r e f}\\right) = \\mathcal {L} _ {\\text {S M}} \\left(x ^ {*}, \\mathcal {X} _ {r e f}\\right) + \\mathcal {L} _ {\\text {a u x}} \\left(x ^ {*}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.144, + 0.484, + 0.221 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_{ref}\\) are the set of real reference images, \\(\\mathcal{L}_{\\mathrm{SM}}(x^*,\\mathcal{X}_{ref})\\) is the loss term as specified in Eq. 2 summed over different layers. We ablate on different hyperparameter settings to show their impact on the final visualization in the supp. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.223, + 0.484, + 0.359 + ], + "angle": 0, + "content": "Choice of Layers for Alignment. While agnostic to the architecture, it is more efficient to focus on aligning features from only few, important layers. To decide on which layers to consider, we conducted a small ablation study (see supp. 3.1) to examine the effects of layers on the visualizations. In a ResNet50, we find that only optimizing for feature alignment of first and last block output is sufficient to produce high-quality images, already capturing the important low-respectively high-level features in the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.484, + 0.513 + ], + "angle": 0, + "content": "Visualization of Concepts. As part of Mechanistic Interpretability, people are interested in finding concept-based explanations of model behavior. These concepts might be feature directions encoded through multiple neurons in a layer, for example, discovered by CRAFT [9]. In VITAL, we modify the initialization of relevances of target neurons in LRP to reflect the weights given by the feature direction. Through this modification, VITAL can give meaning, to these feature directions. We provide technical details and results of this approach in the supp. 2.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.757 + ], + "angle": 0, + "content": "Runtime. Assuming that feature distributions of the reference images are precomputed, VITAL runs in time \\( O(|\\mathcal{X}_{ref}| * L * C * P \\log P) \\), for \\( |\\mathcal{X}_{ref}| \\) reference images, \\( L \\) layers considered for matching, \\( C \\) (maximum) number of channels in those layers and \\( P \\) the (maximum) number of pixels in a feature map. As discussed before, \\( L \\) is usually small, in the following, we use \\( L = 5 \\) layers for matching. When integrating LRP scores in our matching, we additionally have the cost of one backward pass through the model. In practice, it takes roughly 40 seconds to generate an image including computation of distributions for reference images. For comparison, MACO takes 23-28 seconds and DeepInversion 1-3 minutes, depending on their setting. Similarly expensive as DeepInversion is the generation of inner neurons with VITAL (\\( L = 4 \\)), with 2-3 minutes including computation of distributions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.296, + 0.788 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To evaluate VITAL, we compare against Fourier [27] as standard baseline, DeepInversion (DeepInv) [42] as state-of-the-art class visualization, and MACO [8], which is generally considered state-of-the-art for both class and inner neuron visualizations. We implemented VITAL in PyTorch and compared on ResNet50 [13], DenseNet121 [14], ConvNext-base [19], ViT-L-16 [7], ViT-L-32 [7], both qualitatively" + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.091, + 0.582, + 0.103 + ], + "angle": 0, + "content": "MACO" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.106, + 0.607, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.104, + 0.684, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.16, + 0.588, + 0.171 + ], + "angle": 0, + "content": "Fourier" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.172, + 0.606, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.172, + 0.682, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.228, + 0.593, + 0.239 + ], + "angle": 0, + "content": "DeepInv" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.239, + 0.61, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.239, + 0.684, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.296, + 0.577, + 0.306 + ], + "angle": 0, + "content": "VITAL" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.307, + 0.609, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.306, + 0.684, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.364, + 0.604, + 0.373 + ], + "angle": 0, + "content": "maltesedog" + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.364, + 0.664, + 0.373 + ], + "angle": 0, + "content": "candle" + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.378, + 0.649, + 0.39 + ], + "angle": 0, + "content": "(a) ResNet-50" + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.094, + 0.783, + 0.105 + ], + "angle": 0, + "content": "MACO" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.106, + 0.807, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.106, + 0.882, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.161, + 0.788, + 0.172 + ], + "angle": 0, + "content": "Fourier" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.173, + 0.808, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.173, + 0.882, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.228, + 0.793, + 0.24 + ], + "angle": 0, + "content": "DeepInv" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.24, + 0.811, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.24, + 0.884, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.296, + 0.779, + 0.306 + ], + "angle": 0, + "content": "VITAL" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.307, + 0.81, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.307, + 0.883, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.364, + 0.805, + 0.373 + ], + "angle": 0, + "content": "maltesedog" + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.378, + 0.846, + 0.389 + ], + "angle": 0, + "content": "iT-L-32" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Figure 4. Example class visualizations. We show ImageNet class visualizations for (a) ResNet50 and (b) ViT-L-32.MACO and Fourier-based FV (top rows) produce repetitive, hard-to-interpret patterns. DeepInversion (3rd row) improves interpretability but introduces artifacts and lacks ViT compatibility. VITAL arguably yields much more interpretable visualizations. For additional results, including failure cases, we refer to supp. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.907, + 0.607 + ], + "angle": 0, + "content": "and quantitatively. All architectures were pre-trained on ImageNet. We further conducted human user studies assessing human interpretability of the resulting visualizations. We provide details on implementation and setup of competing methods with more results in the supp. 1 and 2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.696, + 0.636 + ], + "angle": 0, + "content": "4.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Visually, there is a stark difference between class visualizations of VITAL and existing work (see Fig. 4 for ResNet50 and ViT-L-32 and supp. 2 for all architectures). Where most existing methods, including MACO and Fourier-based activation maximization, show highly repetitive patterns, our approach provides much cleaner, more understandable representations of the class. While DeepInversion shows less repetition of patterns, there are often irrelevant background features, as in the case of the candle, or visual artifacts, as the blue and yellow color for the dog, dominating the image. Yet, we acknowledge that none of the methods is flawless: All approaches, including VITAL, show difficulties with complex spatial arrangements (supp. 2.6). Still, it is usually understandable what class is represented by VITAL. This becomes even more evident when comparing methods across more complex architectures (e.g., ViT, ConvNeXt), where most approaches struggle to remain inter" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4407" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.09, + 0.478, + 0.456 + ], + "angle": 0, + "content": "
MethodSetupAcc. Top1 (↑)FID (↓)Zero-Shot Prediction
Top1 (↑)Top5 (↑)
ResNet50ImageNet--69.1192.23
MACOr: 22429.43360.7412.8729.73
Fourierr: 22421.30422.446.7318.27
DeepInvbs: 64100.0035.7629.9055.20
VITAL99.9058.7966.6292.56
ConvNeXtImageNet--65.6689.80
MACOr: 22466.0762.557.2019.77
Fourierr: 22460.0759.602.778.30
VITAL99.973.9263.5390.30
DenseNet121ImageNet--70.6493.16
MACOr: 2249.201.809.3323.20
Fourierr: 22415.531.634.8712.17
DeepInvbs: 64100.000.2010.0025.47
VITAL99.930.2758.7086.93
ViT-L-16ImageNet--64.7889.31
MACOr: 22444.33946.963.9310.57
Fourierr: 22425.30990.511.675.13
VITAL99.80126.2968.1792.80
ViT-L-32ImageNet--65.8390.03
MACOr: 22424.872318.9017.5337.23
Fourierr: 22417.031983.0910.3028.10
VITAL89.60147.3355.9785.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.468, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Table 1. Comparison of methods on different architectures trained on Imagenet. We provide FID scores, CLIP Zero-shot prediction scores, and top-1 classification accuracy, indicating the best and second best. In the settings, \"r\" indicates the resolution of the visualization and \"bs\" is the used batch size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.624 + ], + "angle": 0, + "content": "pretable. In contrast, VITAL consistently produces representations where the object and its class remain clear, highlighting its robustness and effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.371, + 0.665 + ], + "angle": 0, + "content": "4.1.1. Analysis of the Embedding Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To understand how realistic and distinct the visualizations are, we compare FVs of five different dog classes to original images in terms of their penultimate layer embedding. We show a 2D low-dimensional representation of the embeddings of original and generated images using tSNE [40] in Fig. 6. We find that the real images form five clusters neatly separating the classes. We further observe that fourier-based approaches with their repetitive, often recognizable patterns are collapsed on a single point, away from any real images. While DeepInversion does fall into the correct clusters, it is hard to recognize a dog in the generated images. VITAL is the only method that consistently falls into the center of the cluster, providing distinct dog features we can recognize as known dog breeds. We extend our results for other architectures in supp. 2.5." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.661, + 0.106 + ], + "angle": 0, + "content": "4.1.2. Small Circuits" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.112, + 0.907, + 0.354 + ], + "angle": 0, + "content": "Our method enables interpreting what inner neurons encode. By selecting the top-3 channels in the last hidden layer most relevant to a class (via LRP), we can visualize local circuits that reveal class-specific patterns (Fig. 5). While most methods, except DeepInversion, support inner neuron visualization, we compare with MACO as the strongest baseline. For example, we see that the most relevant, distinguishing features for a zebra classification are the stripes and texture of their fur. For dogs, there are leg, ear, and specifically colored fur detectors evident in the model. Such specific coloring is necessary to distinguish the plenty of dog breeds available as classes in ImageNet. We also observe signs of overfitting, where the pineapple class is associated with neurons that detect groups of fruits, many of the training images show such a picture. MACO yields less interpretable results, particularly for animals, as confirmed in the next section." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.368, + 0.708, + 0.384 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.391, + 0.907, + 0.482 + ], + "angle": 0, + "content": "For quantitative evaluation, we consider the classification accuracy and the FID score [32] on the respected model, as well as CLIP zero-shot prediction performance [29] in Tab. 1 (see supp. Tab. 1 for the extended version). We also provide quantitative results for the intermediate neurons with the metrics proposed by [16] in supp. 2.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.485, + 0.909, + 0.637 + ], + "angle": 0, + "content": "Accuracy. To assess how well visualizations describe their target, we classify each class neuron's visualization (see Tab. 1). Despite explicitly optimizing class logits, most existing methods perform poorly. Only VITAL and DeepInversion are able to produce accurate descriptions, achieving virtually \\(100\\%\\) accuracy. Investigating further, we observe that other methods including MACO seem to produce irrelevant features that mislead the model, producing predictions unrelated to the original class. We give two examples of prediction in supp. 2.4." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.641, + 0.909, + 0.761 + ], + "angle": 0, + "content": "FID Score. For all architectures, to quantify visualization realism, we compute FID scores [32] using features from ResNet50's penultimate layer (Tab. 1). While aligned with our objective, the performance gap to other methods is substantial, with VITAL achieving over an order of magnitude better scores. DeepInversion also leverages feature statistics but is highly sensitive to batch size, which impacts both FID and visual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "CLIP Zero-shot Prediction. So far, previous metrics relied on the model used for visualization. To assess general perceptual quality, we use a pretrained CLIP ViT-B/32 model to classify FVs via template-based prompts (see supp. 1.2 for more details). We report zero-shot scores in Tab. 1, including original (correctly classified) ImageNet images for reference. As before, we see that DeepInversion outperforms existing work, with standard baselines having overall poor performance even when considering Top-5 accuracy. No" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.25, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 5. Visualizations of small \"circuits\". We provide visualizations of small \"circuits\" for three interesting classes of ImageNet. For each class, we found the three most relevant neurons in the penultimate layer based on LRP relevance scores and indicate the neuron ID on the arrow. For each neuron, we label its likely meaning and provide FVs by MACO and VITAL. The Zebra class shows high dependence on the characteristic pattern of the fur, the dog class a dependence on specific color patterns to be able to distinguish the abundant different dog classes in ImageNet, and Pineapple shows signs of overfitting with an association with collections of fruits." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.344, + 0.465, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Figure 6. \\( t \\)-SNE projection of embedding. We show a low-dimensional tSNE embedding of the features at the penultimate layer for five dog breeds indicated by color. Transparent circles are original training images and FVs are indicated by symbols:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.091, + 0.63, + 0.388, + 0.644 + ], + "angle": 0, + "content": "VITAL, \\(\\triangleleft\\) : MACO, \\(\\bullet\\) : Fourier, \\(\\star\\) : DeepInv." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.703 + ], + "angle": 0, + "content": "tably, VITAL outperforms all others by a wide margin, approaching real image accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.358, + 0.729 + ], + "angle": 0, + "content": "4.3. Human Interpretability Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The primary objective of FVs is to make neural networks more understandable for human users. To address this goal, we conducted a user study (58 participants) to evaluate the interpretability of a FV from the perspective of a human. Specifically, participants (1) given a single word, evaluated how well a FV reflects the provided word, (2) evaluated how well the FVs for an inner neuron reflect the provided reference images (highly activating on the target neuron), and (3) should describe a given generated image with a word or short description. These present settings of increasing difficulty and realism. While (1) is a typical evaluation, it" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.513 + ], + "angle": 0, + "content": "also comes with the largest implicit bias, as the class label already gives away what we want to visualize. In (2), by showing reference images, we encourage a visual match between FV and real image. We compare with MACO, as it is the best performing approach to visualize inner neurons. In (3), we model a real-world scenario, i.e., without any implicit bias, the user is supposed to understand what the network encodes. We provided a simple scoring system from 1(worst) to 5(best) for parts (1) and (2), and evaluated how well the textual description in (3) fits the class using closeness in the embedding space of a text encoder [5]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.909, + 0.786 + ], + "angle": 0, + "content": "We provide an overview of the results of our study in Fig. 7. We observe that our method performs favorable across all three tasks compared to other methods, showing an order of magnitude more high-scoring visualizations (scores 4-5) in both (1) and (2), and showing much higher similarity to the correct class on task (3). In particular, for this challenging task that provides a zero-shot estimate of the interpretability of FVs, the \\(5\\%\\) quantile, i.e., the \"worse\"-performing visualizations, are still close to a similarity of .5 of VITAL. Other methods show only 0.25 similarity for the same quantile. The same stark difference of more than 0.2 in median similarity score shows that VITAL provides more human-understandable visualizations of what neurons encode. We provide more fine-grained analysis, population statistics, and details in supp. 4. We further conducted a human user study following the protocol of [8, 46], which is a simple option-based evaluation similar to setting (2), where VITAL performs best (supp. 4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.628, + 0.815 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our experiments show that VITAL produces more understandable visualizations that better capture the unique information encoded by each neuron, outperforming MACO [8], the most recent method in classification accuracy on the target model. While DeepInversion performs similarly, this is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4409" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.091, + 0.347, + 0.102 + ], + "angle": 0, + "content": "User Study Section 1 results (reference word)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.103, + 0.434, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.235, + 0.404, + 0.247 + ], + "angle": 0, + "content": "(a) Class visualization with given class label." + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.25, + 0.349, + 0.26 + ], + "angle": 0, + "content": "User Study Section 2 results (reference crops)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.264, + 0.434, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.394, + 0.443, + 0.407 + ], + "angle": 0, + "content": "(b) Inner neuron visualization with given activating images." + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.409, + 0.327, + 0.419 + ], + "angle": 0, + "content": "User Study Section 3 results (free form)" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.421, + 0.439, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.553, + 0.411, + 0.566 + ], + "angle": 0, + "content": "(c) Class visualization with free-form labeling." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.578, + 0.483, + 0.745 + ], + "angle": 0, + "content": "Figure 7. Human Study results. We provide summary statistics on the scores for the different methods obtained for the three-part human \\((n = 58)\\) evaluation study. In (a), participants were given class visualization and corresponding class name for 10 random classes and scored the visualizations on a scale of 1(worst) to 5(best). In (b) the task was to evaluate the closeness of FVs for an inner neuron to the image crops that it most activates on (same scoring). DeepInversion does not generate visualizations for inner neurons. In (c), participants were asked to label a given FV. We evaluated similarity of their label to the (hidden) ground truth based on Universal Sentence Encoder embeddings. We provide violin plots with median and \\(5\\%\\) and \\(95\\%\\) quantiles of the achieved similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.852 + ], + "angle": 0, + "content": "expected since it directly optimizes the model's logit score. When evaluated by a different model such as CLIP, VITAL surpasses all baselines by a wide margin. Two human user studies further confirm that VITAL yields more interpretable visualizations in describing given labels or images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Realism. In principle, our visualizations, as well as those of other methods, do not reach photo-realism. VITAL visualizations sometimes resemble the style of paintings of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Édouard Manet, which, however, does not limit their interpretability: Similar to a painting, it is understandable what is expressed, which is also supported by the user studies. Yet, it would be interesting to study how to reach photorealism in the future. Generative models such as diffusion offer a promising direction, though further research is needed to prevent the introduction of model-specific artifacts that reduce faithfulness to the target model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Evaluation. We evaluate our framework using a broad set of metrics, combining standard benchmarks with additional measures like CLIP prompting. Unfortunately, in feature visualization, more broadly in Explainability research, exhaustive benchmarks and clearly defined metrics are actively discussed, but currently still amiss. Through the additional human user studies, we complement the metrics with an evaluation of our original goal of more understandable feature visualizations. As with any user study, the design can be debated, and more refined setups with larger cohorts offer a promising direction for future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.537 + ], + "angle": 0, + "content": "Distribution Matching. Standard approaches to distribution matching include matching statistics, such as mean and variances, or considering prior distributions [15, 20], but these oversimplify the complexity of real images. Activation distributions are often too diverse for such approximations [45]. Hence, directly optimizing an alignment between the empirical distributions showed the most promise, with sort matching [30] as an efficient approach that easily scales to modern networks, but our approach is agnostic to the type of distribution matching algorithm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.909, + 0.692 + ], + "angle": 0, + "content": "Future work. VITAL provides a substantial improvement over existing work and we anticipate it of great use in downstream applications, such as understanding networks in the medical domain, studying what neurons encode after knowledge transfer, or how pruning affects the representations. Future directions include automatic selection of neurons or neuron sets for circuit visualization, for example using CRP [1], integration of ViT-specific visualization via AttnLRP [2], and extending our analysis to CLIP-based ViT models to better understand multimodal representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Feature visualization is key to understanding neural representations, but existing methods often yield artifacts or irrelevant patterns. We propose a new approach that aligns the feature distributions of generated and real images, and incorporates relevance scores to better capture the true information encoded in intermediate neurons. Across multiple metrics and two user studies, VITAL produces clearer and more interpretable visualizations for both humans and models. We envision VITAL supporting safety-critical fields such as medicine and complementing Mechanistic Interpretability by attributing meaning to neural circuits." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4410" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. AttnLRP: Attention-aware layer-wise relevance propagation for transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 135-168. PMLR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research (TMLR), 2024. Survey Certification, Expert Certification. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[4] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 10329-10338, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[5] Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. Universal sentence encoder. CoRR, abs/1803.11175, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[6] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. In The Advances in Neural Information Processing Systems (NIPS), pages 16318-16352. Curran Associates, Inc., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations (ICLR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[8] Thomas Fel, Thibaut Boissin, Victor Boutin, Agustin Picard, Paul Novello, Julien Colin, Drew Linsley, Tom Rousseau, Remi Cadene, Lore Goetschalckx, Laurent Gardes, and Thomas Serre. Unlocking feature visualization for deep network with magnitude constrained optimization. In The Advances in Neural Information Processing Systems (NIPS), pages 37813-37826. Curran Associates, Inc., 2023. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[9] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadécnc, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2711-2721, 2023. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Jonas Fischer, Anna Olah, and Jilles Vreeken. What's in the box? exploring the inner life of neural networks with robust rules. In The International Conference on Machine Learning (ICML), pages 3352-3362, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[11] Amin Ghiasi, Hamid Kazemi, Steven Reich, Chen Zhu, Micah Goldblum, and Tom Goldstein. Plug-in inversion: Model-agnostic inversion for vision with data augmentations. In The International Conference on Machine Learning (ICML), pages 7484–7512, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[12] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 20504-20513, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[13] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In The IEEE International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[16] Laura Kopf, Pheline Lou Bommer, Anna Hedström, Sebastian Lapuschkin, Marina MC Höhne, and Kirill Bykov. Cosy: Evaluating textual explanations of neurons. In The Advanced Conference on Neural Information Processing Systems (NIPS), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[17] Sebastian Lapuschkin, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10:e0130140, 2015. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[18] Shuying Liu and Weihong Deng. Very deep convolutional neural network based image classification using small training sample size. In IAPRA Asian Conference on Pattern Recognition (ACPR), pages 730-734, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[20] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In The IEEE International Conference on Computer Vision (ICCV), pages 5951-5960, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[21] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[22] A. Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[23] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4411" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In The Advances in Neural Information Processing Systems (NIPS), page 3395-3403, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 3510-3520, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[26] Anh Mai Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. CoRR, abs/1602.03616, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[27] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[28] Seitaro Otsuki, Tsumugi Iida, Félix Doublet, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi, and Komei Sugiura. Layer-wise relevance propagation with conservation property for resnet. In The European Conference on Computer Vision (ECCV), pages 349-364, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In The International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[30] Jannick P. Rolland, V. Vo, B. Bloss, and Craig K. Abbey. Fast algorithms for histogram matching: application to texture synthesis. Journal of Electronic Imaging, 9(1):39 - 45, 2000. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[31] Naomi Saphra and Sarah Wiegreff. Mechanistic? In The BlackBoxNLP Workshop at The Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[32] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[33] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. CoRR, abs/1312.6034, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[34] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. CoRR, abs/1706.03825, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[35] Arshia Soltani Moakhar, Eugenia Iofinova, Elias Frantar, and Dan Alistarh. SPADE: Sparsity-guided debugging for deep neural networks. In The International Conference on Machine Learning (ICML), pages 45955-45987, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[36] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In International Conference on Learning Representations (ICLR), 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[37] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In The International Conference on Machine Learning (ICML), page 3319-3328, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[39] Mike Tyka. Class visualization with bilateral filters. https://science.miketyka.com/deepdream/2016/02/05/bilateral-class-vis.html, 2016.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[41] Guangrun Wang and Philip H. S. Torr. Traditional classification neural networks are good generators: They are competitive with ddpms and gans. CoRR, abs/1706.03825, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[42] Hongxu Yin, Pavlo Molchanov, Jose M. Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[43] Runpeng Yu and Xinchao Wang. Generator born from classifier. In The Advances in Neural Information Processing Systems (NIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[44] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In The European Conference on Computer Vision (ECCV), pages 818-833, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[45] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[46] Roland S. Zimmermann, Judy Borowski, Robert Geirhos, Matthias Bethge, Thomas S. A. Wallis, and Wieland Brendel. How well do feature visualizations support causal understanding of cnn activations? The Advances in Neural Information Processing Systems (NIPS), 34:24369-24381, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4412" + } + ] +] \ No newline at end of file diff --git a/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_origin.pdf b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71b692b5631495d028b8666378863d9cf2d6ccc4 --- /dev/null +++ b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/65791404-56e7-4d3c-a033-7abfb8aab04e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:591702717aa72160e50d7966138fa373d048e141b0fe3f7cfe868cd87252c0da +size 16045921 diff --git a/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/full.md b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de808ac2db6fdeef1906105340b3846eac7602ad --- /dev/null +++ b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/full.md @@ -0,0 +1,298 @@ +# VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow + +Ada Gorgun Bernt Schiele Jonas Fischer + +{agoerguen, schiele, jonas.fischer}@mpi-inf.mpg.de Max Planck Institute for Informatics, Saarland Informatics Campus + +![](images/7062af2ee59342e86cb87f9f3dacaa2efe3f66310c77b20ca873b87ff0dbb1f3.jpg) +Figure 1. Feature visualization through distribution matching. Unlike traditional feature visualization (FV) methods, which often produce artifacts or repetitive patterns, VITAL generates more understandable visualizations. Our approach scales effectively to modern architectures (rows), generalizes well across diverse classes (columns), and better captures meaningful network representations. + +# Abstract + +Neural networks are widely adopted to solve complex and challenging tasks. Especially in high-stakes decision-making, understanding their reasoning process is crucial, yet proves challenging for modern deep networks. Feature visualization (FV) is a powerful tool to decode what information neurons are responding to and hence to better understand the reasoning behind such networks. In particular, in FV we generate human-understandable images that reflect the information detected by neurons of interest. However, current methods often yield unrecognizable visualizations, exhibiting repetitive patterns and visual artifacts that are hard to understand for a human. To address these problems, we propose to guide FV through statistics of real image features combined with measures of relevant network flow to generate prototypical images. Our approach yields human-understandable visualizations that both qualitatively and quantitatively improve over state-of-the-art FVs across various architectures. As such, it can be used to decode which information the network uses, complementing mechanistic circuits that identify where it is encoded. + +# 1. Introduction + +Deep neural networks have achieved remarkable success across science and industry. Their reasoning process is, however, inherently opaque. Different lines of research in Explainability of Machine Learning have been addressing these issues as understanding of how a neural network arrives at a decision is crucial, especially in safety-critical domains. To understand which input features are used for a decision, both post-hoc, as well as built-in methodologies have been proposed, which give information about input-output relations in the form of attribution or relevance scores [4, 17, 34, 37]. Yet, these offer only a partial understanding of a model's decision-making process. In the scope of Mechanistic Interpretability [3, 31], experts are interested in finding explanations that uncover what the model internally learns. In this line of research, methods that discover paths of information flow – so-called circuits – have been proposed [6, 10, 35], along with approaches to discover network components, such as groups of neurons, that together encode a concept [9]. While these methods reveal where information is propagated towards a particular decision, they fall short in identifying which information is encoded. Feature visualization (FV) bridges this gap by generating images that correspond to what triggers a strong response in + +a given neuron or group of neurons. This is traditionally achieved by optimizing for an image that maximizes the response of the selected neuron(s) through gradient ascent. Complemented with circuits, FV thus help represent the internal reasoning process in human-understandable terms. + +Feature Visualizations have been an integral part of generating explanations for deep neural networks, adopted early on for CNN models [33]. FVs for earlier architectures [18, 38] through explicit regularization [21, 23, 39] or transformations [22, 27] promote robustness against variations within the image. Already in these early works, irrelevant features as well as artificial, repetitive patterns rendered interpreting the visualizations challenging. For subsequent, larger networks, new methods had to be developed and further refined to get towards realistic image visualizations [8, 26, 42], incorporating elaborate regularizations or reparametrizations in the Fourier space to dampen high-frequency features. Still, newer methods also struggle with repetitive patterns and irrelevant features on modern architectures (see Fig. 4) and show deteriorating interpretability when scaling to larger models. + +In this work, we suggest to re-frame FV in terms of optimizing an alignment to realistic feature distributions instead of activation maximization. In particular, we consider the per-channel distribution of activations in intermediate layers across a set of reference images (e.g., training images of a class, top-activating images for a neuron) and optimize our feature visualization to follow a similar distribution of activations. As such, repetitive patterns – visible as a large number of high activations in a channel – are discouraged unless natural in the original data. Similarly, artificial features or colors, which might highly activate a neuron but are far from the input manifold, are discouraged (see Fig. 1). By further incorporating feature relevance for FVs of neurons for adjusting the feature distributions by how relevant they are to the neuron under study, we further remove irrelevant features from the visualizations that impacted the faithfulness of existing FVs. We provide an efficient implementation of our approach, VITAL, making use of a distribution matching algorithm through which we can back-propagate. This enables VITAL to seamlessly scale to modern architectures, including large networks and Vision Transformers (ViTs), but yields more interpretable FVs, which we show in a set of experiments including a human user study. Our contributions are: + +- we propose a novel method, VITAL, to optimize feature visualizations through alignment with feature distributions of real data instead of maximizing activations, +- we propose to incorporate feature relevance scores in the optimization to focus on exactly those features that are perceived by the target neuron from the feature visualization, and +- we show both qualitatively and quantitatively on a diverse + +set of metrics - including a human user study - that VI-TAL yields more understandable and accurate descriptions of what information is encoded in a neuron. + +# 2. Related Work + +Mechanistic Interpretability. Understanding the internal decision-making process of neural networks has regained significant attention as part of Mechanistic Interpretability (MI) of neural networks. While definitions of MI can differ [31], the essence is to find explanations of neural networks that allow an interpretation of how internal components, such as neurons or layers, operate together to give reason to a prediction. We refer to recent reviews for a comprehensive summary [3, 31]. One key area of research is the discovery of circuits, which describe along which network paths information is propagated to a target neuron. We here discuss how to describe which information is encoded in neurons through feature visualization, which also has shown promising results in explaining the information encoded in circuits or feature directions [9, 10]. + +Model Inversion. Model inversion techniques reconstruct input images from model outputs by optimizing images to match predictions, revealing learned representations and privacy risks. Recent works have improved inversion fidelity through data augmentations [11], distribution matching [42], reinforcement learning [12], and generative networks [43]. Unlike model inversion, which visualizes class-specific information only, FVs can capture information of any network component, including individual neurons, neuron combinations, and channels, to better understand network representations. + +Feature Visualization. Feature visualization techniques aim to generate an image that maximizes the activation of specific network parts (e.g., a neuron) [33, 44]. To yield understandable visualizations, researchers suggested to impose constraints based on local variance [21], apply blurring [23, 39], or ensure consistency across transformations [22, 26] to promote robustness. One of the most widely used approaches is called preconditioning [27], where optimization is performed on the Fourier basis, ensuring that data is decorrelated and whitened. Most recently, MACO [8] improved this preconditioning, suggesting to only optimize the phase while fixing the magnitude to that of a typical image. Unfortunately, good visualizations for modern architectures such as large ResNets [13] are still hard to achieve. To address this challenge, statistically learned priors based on generative models such as GANs [24, 43], or autoencoders [25, 41] are leveraged. A key drawback of these, compared to traditional FV, is that it is unclear what of the visualization can be attributed to the neural network under study and what is due to the generative model. + +![](images/299ac9d8c8176995d0aa1de62a2152f6d1a3ae9142149169d99f6868ef6e8ec4.jpg) +Figure 2. Importance of relevance scores. Given an image selected by its relevance to a neuron detecting an ear, we apply LRP from this neuron back to a preceding building block. We show the activation (first column) and relevance maps (second column) of two neurons from this block, respectively. While the activation is high for the ears (top) and background (bottom), only the ear is relevant for the target neuron. The difference between VITAL with and without LRP (third column) shows that incorporating relevance helps to avoid visualizing irrelevant features. + +# 3. VITAL Framework + +Existing FV methods suffer from overly repetitive patterns, artificial, and irrelevant features. To address this, we propose aligning feature distributions between learned and real images across layers, capturing both coarse and fine-grained information (Sec. 3.2). This discourages the use of artificial features and repetitive patterns, which typically spike the activation in classic approaches. We further suggest to take into account relevance scores, matching the distributions of activations weighted by their relevance to the target, which discourages features in the FV that are irrelevant for the target neuron (sec. 3.3). This becomes particularly important for the visualization of inner neurons. For example, while grass should be irrelevant for a bird's beak, it might still coactivate channels across multiple bird images, which misleads interpretations. Incorporating LRP eliminates irrelevant correlations in activation maps, ensuring that only the features contributing to the target neuron's activation are visualized. We provide an example in Fig. 2 and further details about our approach in supp. 1. + +# 3.1. Notation + +We consider a neural network $f(\cdot ;\theta): \mathcal{X} \to \mathcal{Y}$ as a composition of $L$ functions corresponding to layers, i.e., $f = f^{(L)}\circ f^{(L - 1)}\circ \ldots \circ f^{(2)}\circ f^{(1)}$ from an input space $\mathcal{X}\in \mathbb{R}^{C_0\times H\times W}$ to an output space $\mathcal{Y}\in \mathbb{R}^c$ . Here, $\theta$ are the (trained) network parameters, $C_0$ is the number of channels (typically 1 or 3), and $c$ is the number of classes. For ease + +of notation, we will use $f^{(l)}(x) = f^{(l)} \circ \ldots \circ f^{(2)} \circ f^{(1)}(x)$ , to indicate the network output of layer $l$ on input $x \in \mathcal{X}$ . Without loss of generality, we denote the feature space of layer $l$ flattened over its spatial dimensions as $A^{(l)}(x) = f^{(l)}(x) \in \mathbb{R}^{C_l \times D}$ , where $C_l$ is the number of channels. + +# 3.2. Feature Distribution Matching + +Given a target of interest in the network $f$ , e.g. the $i$ -th neuron in layer $l$ , $f_{i}^{(l)}$ , we are after an input $x^{*}$ that captures the information that $f_{i}^{(l)}$ responds to. Instead of finding an image that maximizes this activation, as typical in the literature, we consider the distribution of features $A(l')$ at layers $l' < l$ of both our generated input $x$ and inputs $x' \in \mathcal{X}_{ref}$ from reference data. Such references could be training data for which $f_{i}^{(l)}$ responds highly. We would like to have similar feature distributions of $x$ compared to these reference images $x'$ . More formally, we would like to find an image $x^{*}$ such that $\forall l' < l, x' \in \mathcal{X}_{ref}$ : + +$$ +\operatorname {d i s t} \left(A ^ {\left(l ^ {\prime}\right)} \left(x ^ {*}\right)\right) \approx \operatorname {d i s t} \left(A ^ {\left(l ^ {\prime}\right)} \left(x ^ {\prime}\right)\right), \tag {1} +$$ + +where $\mathrm{dist}(\cdot)$ indicates the distribution of the activations. In Fig. 3, we illustrate the alignment of feature distributions for sample channels and blocks. + +Investigating empirical distributions of activations, we observe that they do not consistently follow a typical prior, such as a Gaussian, and hence decided to match empirical distributions directly. In particular, we leverage a result on sort-matching from the style-transfer literature [45], which allows us to compute a matching between two feature vectors, here $A^{(l^{\prime})}(x^{*})$ and $A^{(l^{\prime})}(x^{\prime})$ , that we can back-propagate through. We can, hence, efficiently compute gradients with respect to $x^{*}$ . Accumulating the difference between the (sorted) feature vectors of our current generated image $x^{*}$ and the (sorted) feature vectors for all reference images, and across layers defines a loss function that reflects Statement 1. So, optimizing this loss with respect to the generated image $x^{*}$ , improves the alignment of the features of our generated image with real images. + +# 3.2.1. Sort-Matching Loss + +The key idea behind the sort-matching loss suggested by Zhang et al. [45] is to compute the difference of the distribution of two arrays, $z = A(l')(x^*)$ and $z' = A(l')(x')$ , by sorting these two arrays that allow for differentiation to $x^*$ . In particular, we first compute indices $\pi$ and $\pi'$ that sort the arrays $z$ and $z'$ , respectively. We then compute an indexing $\bar{\pi}, \bar{\pi}'$ that sorts these indices. Intriguingly, applying both indices yields the original order, i.e., $z = \bar{\pi}(\pi(z))$ . We can now match $z$ , which is a function of $x^*$ , and $z^r = \bar{\pi}(\pi'(z'))$ , which reorders the sorted values of $z'$ based on the indexing + +from $z$ . Computing the MSE between these two arrays + +$$ +\operatorname {M S E} \left(z, z ^ {r}\right) = \frac {1}{| z |} \sum_ {i = 1} ^ {| z |} \left(z _ {i} - z _ {i} ^ {r}\right), ^ {2} \tag {2} +$$ + +we can differentiate for $x^{*}$ , allowing for gradient-based optimization. For multiple real feature vectors, we average the transformed $\pi'(z')$ vectors to obtain a representative prototype before applying $\bar{\pi}$ to obtain $z^r$ . + +# 3.2.2. Selection of Reference Images + +Typically, the reference set should reflect the distribution relevant to the target neuron. For class neurons, we directly use a random selection of images from that class. For intermediate neurons, we select the top- $k$ patches (from distinct images) relevant for this neuron, following Fel et al. [9]. In brief, we crop and resize training images into patches and score each based on the global average pooled activation across spatial dimensions, keeping the top- $k$ patches from unique images. These define our reference $\mathcal{X}_{ref}$ for feature alignment. We provide additional sampling experiments, including size sensitivity, cluster-based sampling, dataset generalization, and corruption analysis, in the supp. 3.4. + +# 3.3. Incorporating Relevance Scores + +Especially for visualizations of intermediate neurons, irrelevantly activated features pose a problem. For example, a neuron detecting a facial feature of animals might be correlated with a background feature, such as grass, as this is predominant in training – an animal in nature. Classic approaches of activation maximization do not discourage such activations, which is why irrelevant features can appear in their visualizations. Similarly, matching of distributions of activations would encourage background features in the FV. To overcome this, we incorporate feature relevance scores that describe how much a feature (encoded by an intermediate neuron or channel) is relevant for the downstream target neuron. Methods such as layer-wise relevance propagation (LRP) provide meaningful relevance scores [17, 28]. We provide an example of irrelevant background activation and its actual relevance computed by LRP for the target neuron for ResNet50 trained on ImageNet, in Fig. 2. + +We thus suggest matching feature distributions based on activation times relevance, i.e., match features $A^{(l^{\prime})}(x) \odot R_{n}^{(l^{\prime})}(x)$ , where $R_{n}$ are the relevance scores for the target neuron $n$ under study. These relevance scores give weight to each neuron (or pixel of a channel) describing how relevant the neuron is for target neuron $n$ on the given sample. They are hence of the same dimension as the activations, $\odot$ thus indicating the Hadamard product. In principle, any attribution approach can be used as plug-in replacement to obtain relevance scores $R$ . For the remainder of the paper, we use LRP, one of the most prominent attribution methods, + +![](images/4d11361a8b5b348d0df4b569a8b70865bcf7c1bb60c4aec4d0dbd845a01a838d.jpg) +Figure 3. Overview for feature distribution matching. Activations or activation $\times$ relevance scores (A) are extracted from channels across network blocks for real and optimized images. For feature visualization, the distribution of the generated image $x^{*}$ is optimized to match those of the real reference images $x^{\prime} \in \mathcal{X}_{ref}$ . + +and provide a comparative result for Guided Backpropagation [36] in supp. 3.5. Besides intermediate neurons, we also applied relevance scores to class neurons in supp. 2.2. + +# 3.4. Further Considerations + +Transparency Map. Irrelevant areas of the generated image stay mostly unchanged during the optimization, essentially representing noise. Analogous to Fel et al. [8], we suggest using transparency maps based on the importance of the image location during optimization to show relevant image parts only. In brief, we accumulate the gradients of our loss (Eq. 2) across each step in the optimization. We thus ensure the identification of the areas that have been most attended to by the network during the generation of the image. In the supp. 3.3, we illustrate the effect of the transparency map for example visualizations. + +Auxiliary Regularization. While already working reasonably in practice, existing regularizations further reduce noise and small artifacts in the generated image. Similar to Yin et al. [42], we penalize the total variance (TV) and $\ell_2$ norm of our FV with parameter $\alpha_{\mathrm{TV}}$ and $\alpha_{\ell_2}$ as in: + +$$ +\mathcal {L} _ {\text {a u x}} \left(x ^ {*}\right) = \alpha_ {\mathrm {T V}} \mathcal {L} _ {\mathrm {T V}} \left(x ^ {*}\right) + \alpha_ {\ell_ {2}} \ell_ {2} \left(x ^ {*}\right), \tag {3} +$$ + +where $x^{*}$ is the generated image. Accordingly, the total loss + +to be minimized becomes + +$$ +\mathcal {L} _ {\text {V I T A L}} \left(x ^ {*}, \mathcal {X} _ {r e f}\right) = \mathcal {L} _ {\text {S M}} \left(x ^ {*}, \mathcal {X} _ {r e f}\right) + \mathcal {L} _ {\text {a u x}} \left(x ^ {*}\right), \tag {4} +$$ + +where $\mathcal{X}_{ref}$ are the set of real reference images, $\mathcal{L}_{\mathrm{SM}}(x^*,\mathcal{X}_{ref})$ is the loss term as specified in Eq. 2 summed over different layers. We ablate on different hyperparameter settings to show their impact on the final visualization in the supp. 3.2. + +Choice of Layers for Alignment. While agnostic to the architecture, it is more efficient to focus on aligning features from only few, important layers. To decide on which layers to consider, we conducted a small ablation study (see supp. 3.1) to examine the effects of layers on the visualizations. In a ResNet50, we find that only optimizing for feature alignment of first and last block output is sufficient to produce high-quality images, already capturing the important low-respectively high-level features in the image. + +Visualization of Concepts. As part of Mechanistic Interpretability, people are interested in finding concept-based explanations of model behavior. These concepts might be feature directions encoded through multiple neurons in a layer, for example, discovered by CRAFT [9]. In VITAL, we modify the initialization of relevances of target neurons in LRP to reflect the weights given by the feature direction. Through this modification, VITAL can give meaning, to these feature directions. We provide technical details and results of this approach in the supp. 2.3. + +Runtime. Assuming that feature distributions of the reference images are precomputed, VITAL runs in time $O(|\mathcal{X}_{ref}| * L * C * P \log P)$ , for $|\mathcal{X}_{ref}|$ reference images, $L$ layers considered for matching, $C$ (maximum) number of channels in those layers and $P$ the (maximum) number of pixels in a feature map. As discussed before, $L$ is usually small, in the following, we use $L = 5$ layers for matching. When integrating LRP scores in our matching, we additionally have the cost of one backward pass through the model. In practice, it takes roughly 40 seconds to generate an image including computation of distributions for reference images. For comparison, MACO takes 23-28 seconds and DeepInversion 1-3 minutes, depending on their setting. Similarly expensive as DeepInversion is the generation of inner neurons with VITAL ( $L = 4$ ), with 2-3 minutes including computation of distributions. + +# 4. Experimental Results + +To evaluate VITAL, we compare against Fourier [27] as standard baseline, DeepInversion (DeepInv) [42] as state-of-the-art class visualization, and MACO [8], which is generally considered state-of-the-art for both class and inner neuron visualizations. We implemented VITAL in PyTorch and compared on ResNet50 [13], DenseNet121 [14], ConvNext-base [19], ViT-L-16 [7], ViT-L-32 [7], both qualitatively + +![](images/379a42b7b7e5f82d1504a89e9a4ef0b6b5058ca5698ffe9fa345a90bc898c684.jpg) +MACO + +![](images/70a81d70bc7c5eff4e641bf777c32c27a245da8481091f2433ca67a4b046992a.jpg) + +![](images/80913f9c382b2ee4b7aadbcd2e1fe4eac3205b944d7bb7592f3b1cbbfdac941e.jpg) +Fourier + +![](images/be446db0f208322adef5e425abf54d65e6a443cf84d00f33a969f696ce0a62fb.jpg) + +![](images/bf0db90654135e08658379fca1f79166d520e0f48052025960b05989d7579497.jpg) +DeepInv +VITAL + +![](images/72fd616bd824805d15d55f58c0289ecfdf79d7bf53808759f1c5b656750f15bf.jpg) + +![](images/59a6330d05a020c1e97d18b40f93642b5b57d3c185b1c8a6d3296e9cd1665d75.jpg) +maltesedog +(a) ResNet-50 + +![](images/c844ded66d007aa5d1d8a574d9bd780748264188fd3c78ceaa8538a9a5191293.jpg) +candle + +![](images/5c7f78ca95a4c04752a5a175167f68a79831437f50696ded0f29eb35df3770a5.jpg) +MACO + +![](images/7be6a7f5a3191cd9be04d958bca41650ca11f7519cb7a98313385c59b5daed6d.jpg) + +![](images/2d0739116fe41118abcc312f6fd7b0b070a6e0571394edf30acdba7c886c0db9.jpg) +Fourier + +![](images/0c281c87b9edc834fbc8cb8c74b2f284d22357bf41d37b01ba6b6c0e58168d9e.jpg) + +![](images/af703f34688781888e08d7c0ebe936b718bf4f5990668c133e46e78d6a830b3b.jpg) +DeepInv +VITAL + +![](images/ed0cb756abdffa6046933b693a5eda586866192d7a09202e766fcb3b3a403d3b.jpg) + +![](images/41fdf855274d14efd093ac4dafb7dd7d1598423a1c33639e7fd57ec4465b7a50.jpg) +maltesedog + +![](images/f00026c13de96a97ac190662d2fbbec1a493d9c72b23414c66bab74d979e2135.jpg) +iT-L-32 +Figure 4. Example class visualizations. We show ImageNet class visualizations for (a) ResNet50 and (b) ViT-L-32.MACO and Fourier-based FV (top rows) produce repetitive, hard-to-interpret patterns. DeepInversion (3rd row) improves interpretability but introduces artifacts and lacks ViT compatibility. VITAL arguably yields much more interpretable visualizations. For additional results, including failure cases, we refer to supp. 2. + +and quantitatively. All architectures were pre-trained on ImageNet. We further conducted human user studies assessing human interpretability of the resulting visualizations. We provide details on implementation and setup of competing methods with more results in the supp. 1 and 2, respectively. + +# 4.1. Qualitative Results + +Visually, there is a stark difference between class visualizations of VITAL and existing work (see Fig. 4 for ResNet50 and ViT-L-32 and supp. 2 for all architectures). Where most existing methods, including MACO and Fourier-based activation maximization, show highly repetitive patterns, our approach provides much cleaner, more understandable representations of the class. While DeepInversion shows less repetition of patterns, there are often irrelevant background features, as in the case of the candle, or visual artifacts, as the blue and yellow color for the dog, dominating the image. Yet, we acknowledge that none of the methods is flawless: All approaches, including VITAL, show difficulties with complex spatial arrangements (supp. 2.6). Still, it is usually understandable what class is represented by VITAL. This becomes even more evident when comparing methods across more complex architectures (e.g., ViT, ConvNeXt), where most approaches struggle to remain inter + +
MethodSetupAcc. Top1 (↑)FID (↓)Zero-Shot Prediction
Top1 (↑)Top5 (↑)
ResNet50ImageNet--69.1192.23
MACOr: 22429.43360.7412.8729.73
Fourierr: 22421.30422.446.7318.27
DeepInvbs: 64100.0035.7629.9055.20
VITAL99.9058.7966.6292.56
ConvNeXtImageNet--65.6689.80
MACOr: 22466.0762.557.2019.77
Fourierr: 22460.0759.602.778.30
VITAL99.973.9263.5390.30
DenseNet121ImageNet--70.6493.16
MACOr: 2249.201.809.3323.20
Fourierr: 22415.531.634.8712.17
DeepInvbs: 64100.000.2010.0025.47
VITAL99.930.2758.7086.93
ViT-L-16ImageNet--64.7889.31
MACOr: 22444.33946.963.9310.57
Fourierr: 22425.30990.511.675.13
VITAL99.80126.2968.1792.80
ViT-L-32ImageNet--65.8390.03
MACOr: 22424.872318.9017.5337.23
Fourierr: 22417.031983.0910.3028.10
VITAL89.60147.3355.9785.47
+ +Table 1. Comparison of methods on different architectures trained on Imagenet. We provide FID scores, CLIP Zero-shot prediction scores, and top-1 classification accuracy, indicating the best and second best. In the settings, "r" indicates the resolution of the visualization and "bs" is the used batch size. + +pretable. In contrast, VITAL consistently produces representations where the object and its class remain clear, highlighting its robustness and effectiveness. + +# 4.1.1. Analysis of the Embedding Space + +To understand how realistic and distinct the visualizations are, we compare FVs of five different dog classes to original images in terms of their penultimate layer embedding. We show a 2D low-dimensional representation of the embeddings of original and generated images using tSNE [40] in Fig. 6. We find that the real images form five clusters neatly separating the classes. We further observe that fourier-based approaches with their repetitive, often recognizable patterns are collapsed on a single point, away from any real images. While DeepInversion does fall into the correct clusters, it is hard to recognize a dog in the generated images. VITAL is the only method that consistently falls into the center of the cluster, providing distinct dog features we can recognize as known dog breeds. We extend our results for other architectures in supp. 2.5. + +# 4.1.2. Small Circuits + +Our method enables interpreting what inner neurons encode. By selecting the top-3 channels in the last hidden layer most relevant to a class (via LRP), we can visualize local circuits that reveal class-specific patterns (Fig. 5). While most methods, except DeepInversion, support inner neuron visualization, we compare with MACO as the strongest baseline. For example, we see that the most relevant, distinguishing features for a zebra classification are the stripes and texture of their fur. For dogs, there are leg, ear, and specifically colored fur detectors evident in the model. Such specific coloring is necessary to distinguish the plenty of dog breeds available as classes in ImageNet. We also observe signs of overfitting, where the pineapple class is associated with neurons that detect groups of fruits, many of the training images show such a picture. MACO yields less interpretable results, particularly for animals, as confirmed in the next section. + +# 4.2. Quantitative Results + +For quantitative evaluation, we consider the classification accuracy and the FID score [32] on the respected model, as well as CLIP zero-shot prediction performance [29] in Tab. 1 (see supp. Tab. 1 for the extended version). We also provide quantitative results for the intermediate neurons with the metrics proposed by [16] in supp. 2.1. + +Accuracy. To assess how well visualizations describe their target, we classify each class neuron's visualization (see Tab. 1). Despite explicitly optimizing class logits, most existing methods perform poorly. Only VITAL and DeepInversion are able to produce accurate descriptions, achieving virtually $100\%$ accuracy. Investigating further, we observe that other methods including MACO seem to produce irrelevant features that mislead the model, producing predictions unrelated to the original class. We give two examples of prediction in supp. 2.4. + +FID Score. For all architectures, to quantify visualization realism, we compute FID scores [32] using features from ResNet50's penultimate layer (Tab. 1). While aligned with our objective, the performance gap to other methods is substantial, with VITAL achieving over an order of magnitude better scores. DeepInversion also leverages feature statistics but is highly sensitive to batch size, which impacts both FID and visual quality. + +CLIP Zero-shot Prediction. So far, previous metrics relied on the model used for visualization. To assess general perceptual quality, we use a pretrained CLIP ViT-B/32 model to classify FVs via template-based prompts (see supp. 1.2 for more details). We report zero-shot scores in Tab. 1, including original (correctly classified) ImageNet images for reference. As before, we see that DeepInversion outperforms existing work, with standard baselines having overall poor performance even when considering Top-5 accuracy. No + +![](images/3439541d05f3c853dd2a59c106793cd19781c8b25fb7d084f8fcd4d69b07ab6b.jpg) +Figure 5. Visualizations of small "circuits". We provide visualizations of small "circuits" for three interesting classes of ImageNet. For each class, we found the three most relevant neurons in the penultimate layer based on LRP relevance scores and indicate the neuron ID on the arrow. For each neuron, we label its likely meaning and provide FVs by MACO and VITAL. The Zebra class shows high dependence on the characteristic pattern of the fur, the dog class a dependence on specific color patterns to be able to distinguish the abundant different dog classes in ImageNet, and Pineapple shows signs of overfitting with an association with collections of fruits. + +Figure 6. $t$ -SNE projection of embedding. We show a low-dimensional tSNE embedding of the features at the penultimate layer for five dog breeds indicated by color. Transparent circles are original training images and FVs are indicated by symbols: +![](images/6fd55888b520ffd7ea626b9d8dfd43c7e92791b0eee0631893c9ba0f52a201af.jpg) +VITAL, $\triangleleft$ : MACO, $\bullet$ : Fourier, $\star$ : DeepInv. + +tably, VITAL outperforms all others by a wide margin, approaching real image accuracy. + +# 4.3. Human Interpretability Study + +The primary objective of FVs is to make neural networks more understandable for human users. To address this goal, we conducted a user study (58 participants) to evaluate the interpretability of a FV from the perspective of a human. Specifically, participants (1) given a single word, evaluated how well a FV reflects the provided word, (2) evaluated how well the FVs for an inner neuron reflect the provided reference images (highly activating on the target neuron), and (3) should describe a given generated image with a word or short description. These present settings of increasing difficulty and realism. While (1) is a typical evaluation, it + +also comes with the largest implicit bias, as the class label already gives away what we want to visualize. In (2), by showing reference images, we encourage a visual match between FV and real image. We compare with MACO, as it is the best performing approach to visualize inner neurons. In (3), we model a real-world scenario, i.e., without any implicit bias, the user is supposed to understand what the network encodes. We provided a simple scoring system from 1(worst) to 5(best) for parts (1) and (2), and evaluated how well the textual description in (3) fits the class using closeness in the embedding space of a text encoder [5]. + +We provide an overview of the results of our study in Fig. 7. We observe that our method performs favorable across all three tasks compared to other methods, showing an order of magnitude more high-scoring visualizations (scores 4-5) in both (1) and (2), and showing much higher similarity to the correct class on task (3). In particular, for this challenging task that provides a zero-shot estimate of the interpretability of FVs, the $5\%$ quantile, i.e., the "worse"-performing visualizations, are still close to a similarity of .5 of VITAL. Other methods show only 0.25 similarity for the same quantile. The same stark difference of more than 0.2 in median similarity score shows that VITAL provides more human-understandable visualizations of what neurons encode. We provide more fine-grained analysis, population statistics, and details in supp. 4. We further conducted a human user study following the protocol of [8, 46], which is a simple option-based evaluation similar to setting (2), where VITAL performs best (supp. 4). + +# 5. Discussion + +Our experiments show that VITAL produces more understandable visualizations that better capture the unique information encoded by each neuron, outperforming MACO [8], the most recent method in classification accuracy on the target model. While DeepInversion performs similarly, this is + +![](images/d7575397bc50f6405f2906b72ae7b7e4c7a3982c0b68e03d6017143da4ea92d2.jpg) +User Study Section 1 results (reference word) +(a) Class visualization with given class label. + +![](images/1a2a72e2396857a4eb1b062b1844e7d6c11a26070357a05c6d71289e70de8559.jpg) +User Study Section 2 results (reference crops) +(b) Inner neuron visualization with given activating images. + +![](images/c1ffd91c692f3594e4ddd1a772fabfb0d2582adc5fbe9814a23380570ac41465.jpg) +User Study Section 3 results (free form) +(c) Class visualization with free-form labeling. +Figure 7. Human Study results. We provide summary statistics on the scores for the different methods obtained for the three-part human $(n = 58)$ evaluation study. In (a), participants were given class visualization and corresponding class name for 10 random classes and scored the visualizations on a scale of 1(worst) to 5(best). In (b) the task was to evaluate the closeness of FVs for an inner neuron to the image crops that it most activates on (same scoring). DeepInversion does not generate visualizations for inner neurons. In (c), participants were asked to label a given FV. We evaluated similarity of their label to the (hidden) ground truth based on Universal Sentence Encoder embeddings. We provide violin plots with median and $5\%$ and $95\%$ quantiles of the achieved similarity. + +expected since it directly optimizes the model's logit score. When evaluated by a different model such as CLIP, VITAL surpasses all baselines by a wide margin. Two human user studies further confirm that VITAL yields more interpretable visualizations in describing given labels or images. + +Realism. In principle, our visualizations, as well as those of other methods, do not reach photo-realism. VITAL visualizations sometimes resemble the style of paintings of + +Édouard Manet, which, however, does not limit their interpretability: Similar to a painting, it is understandable what is expressed, which is also supported by the user studies. Yet, it would be interesting to study how to reach photorealism in the future. Generative models such as diffusion offer a promising direction, though further research is needed to prevent the introduction of model-specific artifacts that reduce faithfulness to the target model. + +Evaluation. We evaluate our framework using a broad set of metrics, combining standard benchmarks with additional measures like CLIP prompting. Unfortunately, in feature visualization, more broadly in Explainability research, exhaustive benchmarks and clearly defined metrics are actively discussed, but currently still amiss. Through the additional human user studies, we complement the metrics with an evaluation of our original goal of more understandable feature visualizations. As with any user study, the design can be debated, and more refined setups with larger cohorts offer a promising direction for future work. + +Distribution Matching. Standard approaches to distribution matching include matching statistics, such as mean and variances, or considering prior distributions [15, 20], but these oversimplify the complexity of real images. Activation distributions are often too diverse for such approximations [45]. Hence, directly optimizing an alignment between the empirical distributions showed the most promise, with sort matching [30] as an efficient approach that easily scales to modern networks, but our approach is agnostic to the type of distribution matching algorithm. + +Future work. VITAL provides a substantial improvement over existing work and we anticipate it of great use in downstream applications, such as understanding networks in the medical domain, studying what neurons encode after knowledge transfer, or how pruning affects the representations. Future directions include automatic selection of neurons or neuron sets for circuit visualization, for example using CRP [1], integration of ViT-specific visualization via AttnLRP [2], and extending our analysis to CLIP-based ViT models to better understand multimodal representations. + +# 6. Conclusion + +Feature visualization is key to understanding neural representations, but existing methods often yield artifacts or irrelevant patterns. We propose a new approach that aligns the feature distributions of generated and real images, and incorporates relevance scores to better capture the true information encoded in intermediate neurons. Across multiple metrics and two user studies, VITAL produces clearer and more interpretable visualizations for both humans and models. We envision VITAL supporting safety-critical fields such as medicine and complementing Mechanistic Interpretability by attributing meaning to neural circuits. + +# References + +[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 8 +[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. AttnLRP: Attention-aware layer-wise relevance propagation for transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 135-168. PMLR, 2024. 8 +[3] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research (TMLR), 2024. Survey Certification, Expert Certification. 1, 2 +[4] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 10329-10338, 2022. 1 +[5] Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. Universal sentence encoder. CoRR, abs/1803.11175, 2018. 7 +[6] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. In The Advances in Neural Information Processing Systems (NIPS), pages 16318-16352. Curran Associates, Inc., 2023. 1 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations (ICLR), 2021. 5 +[8] Thomas Fel, Thibaut Boissin, Victor Boutin, Agustin Picard, Paul Novello, Julien Colin, Drew Linsley, Tom Rousseau, Remi Cadene, Lore Goetschalckx, Laurent Gardes, and Thomas Serre. Unlocking feature visualization for deep network with magnitude constrained optimization. In The Advances in Neural Information Processing Systems (NIPS), pages 37813-37826. Curran Associates, Inc., 2023. 2, 4, 5, 7 +[9] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadécnc, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2711-2721, 2023. 1, 2, 4, 5 +[10] Jonas Fischer, Anna Olah, and Jilles Vreeken. What's in the box? exploring the inner life of neural networks with robust rules. In The International Conference on Machine Learning (ICML), pages 3352-3362, 2021. 1, 2 + +[11] Amin Ghiasi, Hamid Kazemi, Steven Reich, Chen Zhu, Micah Goldblum, and Tom Goldstein. Plug-in inversion: Model-agnostic inversion for vision with data augmentations. In The International Conference on Machine Learning (ICML), pages 7484–7512, 2022. 2 +[12] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 20504-20513, 2023. 2 +[13] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 5 +[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 5 +[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In The IEEE International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 8 +[16] Laura Kopf, Pheline Lou Bommer, Anna Hedström, Sebastian Lapuschkin, Marina MC Höhne, and Kirill Bykov. Cosy: Evaluating textual explanations of neurons. In The Advanced Conference on Neural Information Processing Systems (NIPS), 2024. 6 +[17] Sebastian Lapuschkin, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10:e0130140, 2015. 1, 4 +[18] Shuying Liu and Weihong Deng. Very deep convolutional neural network based image classification using small training sample size. In IAPRA Asian Conference on Pattern Recognition (ACPR), pages 730-734, 2015. 2 +[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 5 +[20] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In The IEEE International Conference on Computer Vision (ICCV), pages 5951-5960, 2019. 8 +[21] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 2 +[22] A. Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 2 +[23] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2015. 2 +[24] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks. + +In The Advances in Neural Information Processing Systems (NIPS), page 3395-3403, 2016. 2 +[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 3510-3520, 2017. 2 +[26] Anh Mai Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. CoRR, abs/1602.03616, 2016. 2 +[27] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 2, 5 +[28] Seitaro Otsuki, Tsumugi Iida, Félix Doublet, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi, and Komei Sugiura. Layer-wise relevance propagation with conservation property for resnet. In The European Conference on Computer Vision (ECCV), pages 349-364, 2024. 4 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In The International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 6 +[30] Jannick P. Rolland, V. Vo, B. Bloss, and Craig K. Abbey. Fast algorithms for histogram matching: application to texture synthesis. Journal of Electronic Imaging, 9(1):39 - 45, 2000. 8 +[31] Naomi Saphra and Sarah Wiegreff. Mechanistic? In The BlackBoxNLP Workshop at The Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2 +[32] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6 +[33] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. CoRR, abs/1312.6034, 2013. 2 +[34] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. CoRR, abs/1706.03825, 2017. 1 +[35] Arshia Soltani Moakhar, Eugenia Iofinova, Elias Frantar, and Dan Alistarh. SPADE: Sparsity-guided debugging for deep neural networks. In The International Conference on Machine Learning (ICML), pages 45955-45987, 2024. 1 +[36] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In International Conference on Learning Representations (ICLR), 2015. 4 +[37] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In The International Conference on Machine Learning (ICML), page 3319-3328, 2017. 1 +[38] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent + +Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 2 +[39] Mike Tyka. Class visualization with bilateral filters. https://science.miketyka.com/deepdream/2016/02/05/bilateral-class-vis.html, 2016.2 +[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 6 +[41] Guangrun Wang and Philip H. S. Torr. Traditional classification neural networks are good generators: They are competitive with ddpms and gans. CoRR, abs/1706.03825, 2022. 2 +[42] Hongxu Yin, Pavlo Molchanov, Jose M. Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2020. 2, 4, 5 +[43] Runpeng Yu and Xinchao Wang. Generator born from classifier. In The Advances in Neural Information Processing Systems (NIPS), 2023. 2 +[44] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In The European Conference on Computer Vision (ECCV), pages 818-833, 2014. 2 +[45] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2022. 3, 8 +[46] Roland S. Zimmermann, Judy Borowski, Robert Geirhos, Matthias Bethge, Thomas S. A. Wallis, and Wieland Brendel. How well do feature visualizations support causal understanding of cnn activations? The Advances in Neural Information Processing Systems (NIPS), 34:24369-24381, 2021. 7 \ No newline at end of file diff --git a/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/images.zip b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ce75b647a75c8903c70130268b495bffde533c9 --- /dev/null +++ b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b529ca25a03cfae233b71ed6b3c4ee3652c0e14cf4061352272795741ff0578c +size 581517 diff --git a/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/layout.json b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b7ef38106e3201d432adf49675c4f15bdffbd8a --- /dev/null +++ b/2025/VITAL_ More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow/layout.json @@ -0,0 +1,8696 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "type": "text", + "content": "VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 161, + 416, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 161, + 416, + 175 + ], + "spans": [ + { + "bbox": [ + 194, + 161, + 416, + 175 + ], + "type": "text", + "content": "Ada Gorgun Bernt Schiele Jonas Fischer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 178, + 465, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 178, + 465, + 204 + ], + "spans": [ + { + "bbox": [ + 140, + 178, + 465, + 204 + ], + "type": "text", + "content": "{agoerguen, schiele, jonas.fischer}@mpi-inf.mpg.de Max Planck Institute for Informatics, Saarland Informatics Campus" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 230, + 553, + 366 + ], + "blocks": [ + { + "bbox": [ + 57, + 230, + 553, + 366 + ], + "lines": [ + { + "bbox": [ + 57, + 230, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 553, + 366 + ], + "type": "image", + "image_path": "7062af2ee59342e86cb87f9f3dacaa2efe3f66310c77b20ca873b87ff0dbb1f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. Feature visualization through distribution matching. Unlike traditional feature visualization (FV) methods, which often produce artifacts or repetitive patterns, VITAL generates more understandable visualizations. Our approach scales effectively to modern architectures (rows), generalizes well across diverse classes (columns), and better captures meaningful network representations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 417, + 200, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 417, + 200, + 429 + ], + "spans": [ + { + "bbox": [ + 152, + 417, + 200, + 429 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 442, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 442, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 442, + 297, + 682 + ], + "type": "text", + "content": "Neural networks are widely adopted to solve complex and challenging tasks. Especially in high-stakes decision-making, understanding their reasoning process is crucial, yet proves challenging for modern deep networks. Feature visualization (FV) is a powerful tool to decode what information neurons are responding to and hence to better understand the reasoning behind such networks. In particular, in FV we generate human-understandable images that reflect the information detected by neurons of interest. However, current methods often yield unrecognizable visualizations, exhibiting repetitive patterns and visual artifacts that are hard to understand for a human. To address these problems, we propose to guide FV through statistics of real image features combined with measures of relevant network flow to generate prototypical images. Our approach yields human-understandable visualizations that both qualitatively and quantitatively improve over state-of-the-art FVs across various architectures. As such, it can be used to decode which information the network uses, complementing mechanistic circuits that identify where it is encoded." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 417, + 394, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 417, + 394, + 429 + ], + "spans": [ + { + "bbox": [ + 315, + 417, + 394, + 429 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": "Deep neural networks have achieved remarkable success across science and industry. Their reasoning process is, however, inherently opaque. Different lines of research in Explainability of Machine Learning have been addressing these issues as understanding of how a neural network arrives at a decision is crucial, especially in safety-critical domains. To understand which input features are used for a decision, both post-hoc, as well as built-in methodologies have been proposed, which give information about input-output relations in the form of attribution or relevance scores [4, 17, 34, 37]. Yet, these offer only a partial understanding of a model's decision-making process. In the scope of Mechanistic Interpretability [3, 31], experts are interested in finding explanations that uncover what the model internally learns. In this line of research, methods that discover paths of information flow – so-called circuits – have been proposed [6, 10, 35], along with approaches to discover network components, such as groups of neurons, that together encode a concept [9]. While these methods reveal where information is propagated towards a particular decision, they fall short in identifying which information is encoded. Feature visualization (FV) bridges this gap by generating images that correspond to what triggers a strong response in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 203, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 203, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 203, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 702, + 203, + 712 + ], + "type": "text", + "content": "Code is available at: VITAL Framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4403" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "a given neuron or group of neurons. This is traditionally achieved by optimizing for an image that maximizes the response of the selected neuron(s) through gradient ascent. Complemented with circuits, FV thus help represent the internal reasoning process in human-understandable terms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 326 + ], + "type": "text", + "content": "Feature Visualizations have been an integral part of generating explanations for deep neural networks, adopted early on for CNN models [33]. FVs for earlier architectures [18, 38] through explicit regularization [21, 23, 39] or transformations [22, 27] promote robustness against variations within the image. Already in these early works, irrelevant features as well as artificial, repetitive patterns rendered interpreting the visualizations challenging. For subsequent, larger networks, new methods had to be developed and further refined to get towards realistic image visualizations [8, 26, 42], incorporating elaborate regularizations or reparametrizations in the Fourier space to dampen high-frequency features. Still, newer methods also struggle with repetitive patterns and irrelevant features on modern architectures (see Fig. 4) and show deteriorating interpretability when scaling to larger models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 614 + ], + "type": "text", + "content": "In this work, we suggest to re-frame FV in terms of optimizing an alignment to realistic feature distributions instead of activation maximization. In particular, we consider the per-channel distribution of activations in intermediate layers across a set of reference images (e.g., training images of a class, top-activating images for a neuron) and optimize our feature visualization to follow a similar distribution of activations. As such, repetitive patterns – visible as a large number of high activations in a channel – are discouraged unless natural in the original data. Similarly, artificial features or colors, which might highly activate a neuron but are far from the input manifold, are discouraged (see Fig. 1). By further incorporating feature relevance for FVs of neurons for adjusting the feature distributions by how relevant they are to the neuron under study, we further remove irrelevant features from the visualizations that impacted the faithfulness of existing FVs. We provide an efficient implementation of our approach, VITAL, making use of a distribution matching algorithm through which we can back-propagate. This enables VITAL to seamlessly scale to modern architectures, including large networks and Vision Transformers (ViTs), but yields more interpretable FVs, which we show in a set of experiments including a human user study. Our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 618, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 618, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 618, + 295, + 653 + ], + "type": "text", + "content": "- we propose a novel method, VITAL, to optimize feature visualizations through alignment with feature distributions of real data instead of maximizing activations," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 654, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 295, + 700 + ], + "type": "text", + "content": "- we propose to incorporate feature relevance scores in the optimization to focus on exactly those features that are perceived by the target neuron from the feature visualization, and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "type": "text", + "content": "- we show both qualitatively and quantitatively on a diverse" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 553, + 108 + ], + "type": "text", + "content": "set of metrics - including a human user study - that VI-TAL yields more understandable and accurate descriptions of what information is encoded in a neuron." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 129, + 400, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 129, + 400, + 143 + ], + "spans": [ + { + "bbox": [ + 314, + 129, + 400, + 143 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 152, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 332 + ], + "type": "text", + "content": "Mechanistic Interpretability. Understanding the internal decision-making process of neural networks has regained significant attention as part of Mechanistic Interpretability (MI) of neural networks. While definitions of MI can differ [31], the essence is to find explanations of neural networks that allow an interpretation of how internal components, such as neurons or layers, operate together to give reason to a prediction. We refer to recent reviews for a comprehensive summary [3, 31]. One key area of research is the discovery of circuits, which describe along which network paths information is propagated to a target neuron. We here discuss how to describe which information is encoded in neurons through feature visualization, which also has shown promising results in explaining the information encoded in circuits or feature directions [9, 10]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 337, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 556, + 469 + ], + "type": "text", + "content": "Model Inversion. Model inversion techniques reconstruct input images from model outputs by optimizing images to match predictions, revealing learned representations and privacy risks. Recent works have improved inversion fidelity through data augmentations [11], distribution matching [42], reinforcement learning [12], and generative networks [43]. Unlike model inversion, which visualizes class-specific information only, FVs can capture information of any network component, including individual neurons, neuron combinations, and channels, to better understand network representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Feature Visualization. Feature visualization techniques aim to generate an image that maximizes the activation of specific network parts (e.g., a neuron) [33, 44]. To yield understandable visualizations, researchers suggested to impose constraints based on local variance [21], apply blurring [23, 39], or ensure consistency across transformations [22, 26] to promote robustness. One of the most widely used approaches is called preconditioning [27], where optimization is performed on the Fourier basis, ensuring that data is decorrelated and whitened. Most recently, MACO [8] improved this preconditioning, suggesting to only optimize the phase while fixing the magnitude to that of a typical image. Unfortunately, good visualizations for modern architectures such as large ResNets [13] are still hard to achieve. To address this challenge, statistically learned priors based on generative models such as GANs [24, 43], or autoencoders [25, 41] are leveraged. A key drawback of these, compared to traditional FV, is that it is unclear what of the visualization can be attributed to the neural network under study and what is due to the generative model." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4404" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 290, + 240 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 290, + 240 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 290, + 240 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 290, + 240 + ], + "type": "image", + "image_path": "299ac9d8c8176995d0aa1de62a2152f6d1a3ae9142149169d99f6868ef6e8ec4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 298, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 298, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 298, + 350 + ], + "type": "text", + "content": "Figure 2. Importance of relevance scores. Given an image selected by its relevance to a neuron detecting an ear, we apply LRP from this neuron back to a preceding building block. We show the activation (first column) and relevance maps (second column) of two neurons from this block, respectively. While the activation is high for the ears (top) and background (bottom), only the ear is relevant for the target neuron. The difference between VITAL with and without LRP (third column) shows that incorporating relevance helps to avoid visualizing irrelevant features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 369, + 171, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 171, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 171, + 382 + ], + "type": "text", + "content": "3. VITAL Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 617 + ], + "type": "text", + "content": "Existing FV methods suffer from overly repetitive patterns, artificial, and irrelevant features. To address this, we propose aligning feature distributions between learned and real images across layers, capturing both coarse and fine-grained information (Sec. 3.2). This discourages the use of artificial features and repetitive patterns, which typically spike the activation in classic approaches. We further suggest to take into account relevance scores, matching the distributions of activations weighted by their relevance to the target, which discourages features in the FV that are irrelevant for the target neuron (sec. 3.3). This becomes particularly important for the visualization of inner neurons. For example, while grass should be irrelevant for a bird's beak, it might still coactivate channels across multiple bird images, which misleads interpretations. Incorporating LRP eliminates irrelevant correlations in activation maps, ensuring that only the features contributing to the target neuron's activation are visualized. We provide an example in Fig. 2 and further details about our approach in supp. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 119, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 119, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 119, + 635 + ], + "type": "text", + "content": "3.1. Notation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We consider a neural network " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "f(\\cdot ;\\theta): \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " as a composition of " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " functions corresponding to layers, i.e., " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "f = f^{(L)}\\circ f^{(L - 1)}\\circ \\ldots \\circ f^{(2)}\\circ f^{(1)}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " from an input space " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}\\in \\mathbb{R}^{C_0\\times H\\times W}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " to an output space " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}\\in \\mathbb{R}^c" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " are the (trained) network parameters, " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "C_0" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the number of channels (typically 1 or 3), and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the number of classes. For ease" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "content": "of notation, we will use " + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "inline_equation", + "content": "f^{(l)}(x) = f^{(l)} \\circ \\ldots \\circ f^{(2)} \\circ f^{(1)}(x)" + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "content": ", to indicate the network output of layer " + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "content": " on input " + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "content": ". Without loss of generality, we denote the feature space of layer " + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "content": " flattened over its spatial dimensions as " + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "inline_equation", + "content": "A^{(l)}(x) = f^{(l)}(x) \\in \\mathbb{R}^{C_l \\times D}" + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 313, + 71, + 556, + 133 + ], + "type": "text", + "content": " is the number of channels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 145, + 481, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 481, + 158 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 481, + 158 + ], + "type": "text", + "content": "3.2. Feature Distribution Matching" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": "Given a target of interest in the network " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": ", e.g. the " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": "-th neuron in layer " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "f_{i}^{(l)}" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": ", we are after an input " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " that captures the information that " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "f_{i}^{(l)}" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " responds to. Instead of finding an image that maximizes this activation, as typical in the literature, we consider the distribution of features " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "A(l')" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " at layers " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "l' < l" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " of both our generated input " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " and inputs " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "x' \\in \\mathcal{X}_{ref}" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " from reference data. Such references could be training data for which " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "f_{i}^{(l)}" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " responds highly. We would like to have similar feature distributions of " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " compared to these reference images " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": ". More formally, we would like to find an image " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\forall l' < l, x' \\in \\mathcal{X}_{ref}" + }, + { + "bbox": [ + 313, + 164, + 555, + 304 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 363, + 316, + 555, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 316, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 363, + 316, + 555, + 338 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s t} \\left(A ^ {\\left(l ^ {\\prime}\\right)} \\left(x ^ {*}\\right)\\right) \\approx \\operatorname {d i s t} \\left(A ^ {\\left(l ^ {\\prime}\\right)} \\left(x ^ {\\prime}\\right)\\right), \\tag {1}", + "image_path": "6c269a1bab3b121009142a8b3561a91804ff034ea10a23dadfe78111ebf3d115.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(\\cdot)" + }, + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "content": " indicates the distribution of the activations. In Fig. 3, we illustrate the alignment of feature distributions for sample channels and blocks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "text", + "content": "Investigating empirical distributions of activations, we observe that they do not consistently follow a typical prior, such as a Gaussian, and hence decided to match empirical distributions directly. In particular, we leverage a result on sort-matching from the style-transfer literature [45], which allows us to compute a matching between two feature vectors, here " + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "inline_equation", + "content": "A^{(l^{\\prime})}(x^{*})" + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "inline_equation", + "content": "A^{(l^{\\prime})}(x^{\\prime})" + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "text", + "content": ", that we can back-propagate through. We can, hence, efficiently compute gradients with respect to " + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "text", + "content": ". Accumulating the difference between the (sorted) feature vectors of our current generated image " + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "text", + "content": " and the (sorted) feature vectors for all reference images, and across layers defines a loss function that reflects Statement 1. So, optimizing this loss with respect to the generated image " + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 313, + 386, + 556, + 566 + ], + "type": "text", + "content": ", improves the alignment of the features of our generated image with real images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 577, + 427, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 427, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 427, + 589 + ], + "type": "text", + "content": "3.2.1. Sort-Matching Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "The key idea behind the sort-matching loss suggested by Zhang et al. [45] is to compute the difference of the distribution of two arrays, " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z = A(l')(x^*)" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z' = A(l')(x')" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", by sorting these two arrays that allow for differentiation to " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "x^*" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". In particular, we first compute indices " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi'" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " that sort the arrays " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", respectively. We then compute an indexing " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\pi}, \\bar{\\pi}'" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " that sorts these indices. Intriguingly, applying both indices yields the original order, i.e., " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z = \\bar{\\pi}(\\pi(z))" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". We can now match " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", which is a function of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "x^*" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z^r = \\bar{\\pi}(\\pi'(z'))" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", which reorders the sorted values of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " based on the indexing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4405" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 277, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 277, + 85 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 277, + 85 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 53, + 72, + 277, + 85 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 53, + 72, + 277, + 85 + ], + "type": "text", + "content": ". Computing the MSE between these two arrays" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 94, + 295, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 94, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 110, + 94, + 295, + 127 + ], + "type": "interline_equation", + "content": "\\operatorname {M S E} \\left(z, z ^ {r}\\right) = \\frac {1}{| z |} \\sum_ {i = 1} ^ {| z |} \\left(z _ {i} - z _ {i} ^ {r}\\right), ^ {2} \\tag {2}", + "image_path": "4267bccdd89c11775a104e8c5a7206ca0898c6e368fc2443e175f31eb00cafa8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "text", + "content": "we can differentiate for " + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "text", + "content": ", allowing for gradient-based optimization. For multiple real feature vectors, we average the transformed " + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\pi'(z')" + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "text", + "content": " vectors to obtain a representative prototype before applying " + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\bar{\\pi}" + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "inline_equation", + "content": "z^r" + }, + { + "bbox": [ + 55, + 136, + 296, + 185 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 191, + 212, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 212, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 212, + 204 + ], + "type": "text", + "content": "3.2.2. Selection of Reference Images" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "text", + "content": "Typically, the reference set should reflect the distribution relevant to the target neuron. For class neurons, we directly use a random selection of images from that class. For intermediate neurons, we select the top-" + }, + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "text", + "content": " patches (from distinct images) relevant for this neuron, following Fel et al. [9]. In brief, we crop and resize training images into patches and score each based on the global average pooled activation across spatial dimensions, keeping the top-" + }, + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "text", + "content": " patches from unique images. These define our reference " + }, + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 55, + 206, + 296, + 351 + ], + "type": "text", + "content": " for feature alignment. We provide additional sampling experiments, including size sensitivity, cluster-based sampling, dataset generalization, and corruption analysis, in the supp. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 358, + 228, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 228, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 228, + 372 + ], + "type": "text", + "content": "3.3. Incorporating Relevance Scores" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 376, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 579 + ], + "type": "text", + "content": "Especially for visualizations of intermediate neurons, irrelevantly activated features pose a problem. For example, a neuron detecting a facial feature of animals might be correlated with a background feature, such as grass, as this is predominant in training – an animal in nature. Classic approaches of activation maximization do not discourage such activations, which is why irrelevant features can appear in their visualizations. Similarly, matching of distributions of activations would encourage background features in the FV. To overcome this, we incorporate feature relevance scores that describe how much a feature (encoded by an intermediate neuron or channel) is relevant for the downstream target neuron. Methods such as layer-wise relevance propagation (LRP) provide meaningful relevance scores [17, 28]. We provide an example of irrelevant background activation and its actual relevance computed by LRP for the target neuron for ResNet50 trained on ImageNet, in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "content": "We thus suggest matching feature distributions based on activation times relevance, i.e., match features " + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "inline_equation", + "content": "A^{(l^{\\prime})}(x) \\odot R_{n}^{(l^{\\prime})}(x)" + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "inline_equation", + "content": "R_{n}" + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "content": " are the relevance scores for the target neuron " + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "content": " under study. These relevance scores give weight to each neuron (or pixel of a channel) describing how relevant the neuron is for target neuron " + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "content": " on the given sample. They are hence of the same dimension as the activations, " + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "content": " thus indicating the Hadamard product. In principle, any attribution approach can be used as plug-in replacement to obtain relevance scores " + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 580, + 296, + 714 + ], + "type": "text", + "content": ". For the remainder of the paper, we use LRP, one of the most prominent attribution methods," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 69, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 318, + 69, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 318, + 69, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 318, + 69, + 553, + 316 + ], + "type": "image", + "image_path": "4d11361a8b5b348d0df4b569a8b70865bcf7c1bb60c4aec4d0dbd845a01a838d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "text", + "content": "Figure 3. Overview for feature distribution matching. Activations or activation " + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "text", + "content": " relevance scores (A) are extracted from channels across network blocks for real and optimized images. For feature visualization, the distribution of the generated image " + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "text", + "content": " is optimized to match those of the real reference images " + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "inline_equation", + "content": "x^{\\prime} \\in \\mathcal{X}_{ref}" + }, + { + "bbox": [ + 313, + 324, + 555, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 440 + ], + "type": "text", + "content": "and provide a comparative result for Guided Backpropagation [36] in supp. 3.5. Besides intermediate neurons, we also applied relevance scores to class neurons in supp. 2.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 448, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 448, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 448, + 462 + ], + "type": "text", + "content": "3.4. Further Considerations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 470, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 601 + ], + "type": "text", + "content": "Transparency Map. Irrelevant areas of the generated image stay mostly unchanged during the optimization, essentially representing noise. Analogous to Fel et al. [8], we suggest using transparency maps based on the importance of the image location during optimization to show relevant image parts only. In brief, we accumulate the gradients of our loss (Eq. 2) across each step in the optimization. We thus ensure the identification of the areas that have been most attended to by the network during the generation of the image. In the supp. 3.3, we illustrate the effect of the transparency map for example visualizations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "text", + "content": "Auxiliary Regularization. While already working reasonably in practice, existing regularizations further reduce noise and small artifacts in the generated image. Similar to Yin et al. [42], we penalize the total variance (TV) and " + }, + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "text", + "content": " norm of our FV with parameter " + }, + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{TV}}" + }, + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "inline_equation", + "content": "\\alpha_{\\ell_2}" + }, + { + "bbox": [ + 313, + 605, + 556, + 667 + ], + "type": "text", + "content": " as in:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 357, + 677, + 553, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 677, + 553, + 692 + ], + "spans": [ + { + "bbox": [ + 357, + 677, + 553, + 692 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a u x}} \\left(x ^ {*}\\right) = \\alpha_ {\\mathrm {T V}} \\mathcal {L} _ {\\mathrm {T V}} \\left(x ^ {*}\\right) + \\alpha_ {\\ell_ {2}} \\ell_ {2} \\left(x ^ {*}\\right), \\tag {3}", + "image_path": "466bd30d4ee985b0abb862a3f93797bf87381186f8dfa0e06948f839bc98d243.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": " is the generated image. Accordingly, the total loss" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4406" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 162, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 162, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 162, + 83 + ], + "type": "text", + "content": "to be minimized becomes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 91, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 91, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 70, + 91, + 295, + 107 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {V I T A L}} \\left(x ^ {*}, \\mathcal {X} _ {r e f}\\right) = \\mathcal {L} _ {\\text {S M}} \\left(x ^ {*}, \\mathcal {X} _ {r e f}\\right) + \\mathcal {L} _ {\\text {a u x}} \\left(x ^ {*}\\right), \\tag {4}", + "image_path": "29dfad2648c2d8e19e19e493f8899e6135a55f49e51d20a530dcd020de37e7be.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 296, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 114, + 296, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{ref}" + }, + { + "bbox": [ + 55, + 114, + 296, + 175 + ], + "type": "text", + "content": " are the set of real reference images, " + }, + { + "bbox": [ + 55, + 114, + 296, + 175 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SM}}(x^*,\\mathcal{X}_{ref})" + }, + { + "bbox": [ + 55, + 114, + 296, + 175 + ], + "type": "text", + "content": " is the loss term as specified in Eq. 2 summed over different layers. We ablate on different hyperparameter settings to show their impact on the final visualization in the supp. 3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 176, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 176, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 296, + 284 + ], + "type": "text", + "content": "Choice of Layers for Alignment. While agnostic to the architecture, it is more efficient to focus on aligning features from only few, important layers. To decide on which layers to consider, we conducted a small ablation study (see supp. 3.1) to examine the effects of layers on the visualizations. In a ResNet50, we find that only optimizing for feature alignment of first and last block output is sufficient to produce high-quality images, already capturing the important low-respectively high-level features in the image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "text", + "content": "Visualization of Concepts. As part of Mechanistic Interpretability, people are interested in finding concept-based explanations of model behavior. These concepts might be feature directions encoded through multiple neurons in a layer, for example, discovered by CRAFT [9]. In VITAL, we modify the initialization of relevances of target neurons in LRP to reflect the weights given by the feature direction. Through this modification, VITAL can give meaning, to these feature directions. We provide technical details and results of this approach in the supp. 2.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": "Runtime. Assuming that feature distributions of the reference images are precomputed, VITAL runs in time " + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "O(|\\mathcal{X}_{ref}| * L * C * P \\log P)" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "|\\mathcal{X}_{ref}|" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": " reference images, " + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": " layers considered for matching, " + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": " (maximum) number of channels in those layers and " + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": " the (maximum) number of pixels in a feature map. As discussed before, " + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": " is usually small, in the following, we use " + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "L = 5" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": " layers for matching. When integrating LRP scores in our matching, we additionally have the cost of one backward pass through the model. In practice, it takes roughly 40 seconds to generate an image including computation of distributions for reference images. For comparison, MACO takes 23-28 seconds and DeepInversion 1-3 minutes, depending on their setting. Similarly expensive as DeepInversion is the generation of inner neurons with VITAL (" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 55, + 408, + 296, + 599 + ], + "type": "text", + "content": "), with 2-3 minutes including computation of distributions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 609, + 181, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 181, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 181, + 624 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "To evaluate VITAL, we compare against Fourier [27] as standard baseline, DeepInversion (DeepInv) [42] as state-of-the-art class visualization, and MACO [8], which is generally considered state-of-the-art for both class and inner neuron visualizations. We implemented VITAL in PyTorch and compared on ResNet50 [13], DenseNet121 [14], ConvNext-base [19], ViT-L-16 [7], ViT-L-32 [7], both qualitatively" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 83, + 371, + 125 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 356, + 81 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 356, + 81 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 356, + 81 + ], + "type": "text", + "content": "MACO" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 83, + 371, + 125 + ], + "lines": [ + { + "bbox": [ + 328, + 83, + 371, + 125 + ], + "spans": [ + { + "bbox": [ + 328, + 83, + 371, + 125 + ], + "type": "image", + "image_path": "379a42b7b7e5f82d1504a89e9a4ef0b6b5058ca5698ffe9fa345a90bc898c684.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 373, + 82, + 418, + 125 + ], + "blocks": [ + { + "bbox": [ + 373, + 82, + 418, + 125 + ], + "lines": [ + { + "bbox": [ + 373, + 82, + 418, + 125 + ], + "spans": [ + { + "bbox": [ + 373, + 82, + 418, + 125 + ], + "type": "image", + "image_path": "70a81d70bc7c5eff4e641bf777c32c27a245da8481091f2433ca67a4b046992a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 136, + 370, + 178 + ], + "blocks": [ + { + "bbox": [ + 329, + 126, + 359, + 135 + ], + "lines": [ + { + "bbox": [ + 329, + 126, + 359, + 135 + ], + "spans": [ + { + "bbox": [ + 329, + 126, + 359, + 135 + ], + "type": "text", + "content": "Fourier" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 136, + 370, + 178 + ], + "lines": [ + { + "bbox": [ + 329, + 136, + 370, + 178 + ], + "spans": [ + { + "bbox": [ + 329, + 136, + 370, + 178 + ], + "type": "image", + "image_path": "80913f9c382b2ee4b7aadbcd2e1fe4eac3205b944d7bb7592f3b1cbbfdac941e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 376, + 136, + 417, + 178 + ], + "blocks": [ + { + "bbox": [ + 376, + 136, + 417, + 178 + ], + "lines": [ + { + "bbox": [ + 376, + 136, + 417, + 178 + ], + "spans": [ + { + "bbox": [ + 376, + 136, + 417, + 178 + ], + "type": "image", + "image_path": "be446db0f208322adef5e425abf54d65e6a443cf84d00f33a969f696ce0a62fb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 329, + 189, + 373, + 233 + ], + "blocks": [ + { + "bbox": [ + 329, + 180, + 362, + 189 + ], + "lines": [ + { + "bbox": [ + 329, + 180, + 362, + 189 + ], + "spans": [ + { + "bbox": [ + 329, + 180, + 362, + 189 + ], + "type": "text", + "content": "DeepInv" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 189, + 373, + 233 + ], + "lines": [ + { + "bbox": [ + 329, + 189, + 373, + 233 + ], + "spans": [ + { + "bbox": [ + 329, + 189, + 373, + 233 + ], + "type": "image", + "image_path": "bf0db90654135e08658379fca1f79166d520e0f48052025960b05989d7579497.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 234, + 353, + 242 + ], + "lines": [ + { + "bbox": [ + 329, + 234, + 353, + 242 + ], + "spans": [ + { + "bbox": [ + 329, + 234, + 353, + 242 + ], + "type": "text", + "content": "VITAL" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 374, + 189, + 418, + 233 + ], + "blocks": [ + { + "bbox": [ + 374, + 189, + 418, + 233 + ], + "lines": [ + { + "bbox": [ + 374, + 189, + 418, + 233 + ], + "spans": [ + { + "bbox": [ + 374, + 189, + 418, + 233 + ], + "type": "image", + "image_path": "72fd616bd824805d15d55f58c0289ecfdf79d7bf53808759f1c5b656750f15bf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 329, + 243, + 372, + 286 + ], + "blocks": [ + { + "bbox": [ + 329, + 243, + 372, + 286 + ], + "lines": [ + { + "bbox": [ + 329, + 243, + 372, + 286 + ], + "spans": [ + { + "bbox": [ + 329, + 243, + 372, + 286 + ], + "type": "image", + "image_path": "59a6330d05a020c1e97d18b40f93642b5b57d3c185b1c8a6d3296e9cd1665d75.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 288, + 369, + 295 + ], + "lines": [ + { + "bbox": [ + 332, + 288, + 369, + 295 + ], + "spans": [ + { + "bbox": [ + 332, + 288, + 369, + 295 + ], + "type": "text", + "content": "maltesedog" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 299, + 397, + 308 + ], + "lines": [ + { + "bbox": [ + 350, + 299, + 397, + 308 + ], + "spans": [ + { + "bbox": [ + 350, + 299, + 397, + 308 + ], + "type": "text", + "content": "(a) ResNet-50" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 374, + 242, + 418, + 286 + ], + "blocks": [ + { + "bbox": [ + 374, + 242, + 418, + 286 + ], + "lines": [ + { + "bbox": [ + 374, + 242, + 418, + 286 + ], + "spans": [ + { + "bbox": [ + 374, + 242, + 418, + 286 + ], + "type": "image", + "image_path": "c844ded66d007aa5d1d8a574d9bd780748264188fd3c78ceaa8538a9a5191293.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 288, + 406, + 295 + ], + "lines": [ + { + "bbox": [ + 385, + 288, + 406, + 295 + ], + "spans": [ + { + "bbox": [ + 385, + 288, + 406, + 295 + ], + "type": "text", + "content": "candle" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 452, + 83, + 493, + 123 + ], + "blocks": [ + { + "bbox": [ + 452, + 74, + 479, + 83 + ], + "lines": [ + { + "bbox": [ + 452, + 74, + 479, + 83 + ], + "spans": [ + { + "bbox": [ + 452, + 74, + 479, + 83 + ], + "type": "text", + "content": "MACO" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 83, + 493, + 123 + ], + "lines": [ + { + "bbox": [ + 452, + 83, + 493, + 123 + ], + "spans": [ + { + "bbox": [ + 452, + 83, + 493, + 123 + ], + "type": "image", + "image_path": "5c7f78ca95a4c04752a5a175167f68a79831437f50696ded0f29eb35df3770a5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 496, + 83, + 539, + 125 + ], + "blocks": [ + { + "bbox": [ + 496, + 83, + 539, + 125 + ], + "lines": [ + { + "bbox": [ + 496, + 83, + 539, + 125 + ], + "spans": [ + { + "bbox": [ + 496, + 83, + 539, + 125 + ], + "type": "image", + "image_path": "7be6a7f5a3191cd9be04d958bca41650ca11f7519cb7a98313385c59b5daed6d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 452, + 137, + 494, + 178 + ], + "blocks": [ + { + "bbox": [ + 452, + 127, + 482, + 136 + ], + "lines": [ + { + "bbox": [ + 452, + 127, + 482, + 136 + ], + "spans": [ + { + "bbox": [ + 452, + 127, + 482, + 136 + ], + "type": "text", + "content": "Fourier" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 137, + 494, + 178 + ], + "lines": [ + { + "bbox": [ + 452, + 137, + 494, + 178 + ], + "spans": [ + { + "bbox": [ + 452, + 137, + 494, + 178 + ], + "type": "image", + "image_path": "2d0739116fe41118abcc312f6fd7b0b070a6e0571394edf30acdba7c886c0db9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 496, + 137, + 539, + 178 + ], + "blocks": [ + { + "bbox": [ + 496, + 137, + 539, + 178 + ], + "lines": [ + { + "bbox": [ + 496, + 137, + 539, + 178 + ], + "spans": [ + { + "bbox": [ + 496, + 137, + 539, + 178 + ], + "type": "image", + "image_path": "0c281c87b9edc834fbc8cb8c74b2f284d22357bf41d37b01ba6b6c0e58168d9e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 452, + 190, + 496, + 234 + ], + "blocks": [ + { + "bbox": [ + 452, + 180, + 485, + 190 + ], + "lines": [ + { + "bbox": [ + 452, + 180, + 485, + 190 + ], + "spans": [ + { + "bbox": [ + 452, + 180, + 485, + 190 + ], + "type": "text", + "content": "DeepInv" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 190, + 496, + 234 + ], + "lines": [ + { + "bbox": [ + 452, + 190, + 496, + 234 + ], + "spans": [ + { + "bbox": [ + 452, + 190, + 496, + 234 + ], + "type": "image", + "image_path": "af703f34688781888e08d7c0ebe936b718bf4f5990668c133e46e78d6a830b3b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 234, + 476, + 242 + ], + "lines": [ + { + "bbox": [ + 452, + 234, + 476, + 242 + ], + "spans": [ + { + "bbox": [ + 452, + 234, + 476, + 242 + ], + "type": "text", + "content": "VITAL" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 496, + 190, + 541, + 234 + ], + "blocks": [ + { + "bbox": [ + 496, + 190, + 541, + 234 + ], + "lines": [ + { + "bbox": [ + 496, + 190, + 541, + 234 + ], + "spans": [ + { + "bbox": [ + 496, + 190, + 541, + 234 + ], + "type": "image", + "image_path": "ed0cb756abdffa6046933b693a5eda586866192d7a09202e766fcb3b3a403d3b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 452, + 243, + 495, + 286 + ], + "blocks": [ + { + "bbox": [ + 452, + 243, + 495, + 286 + ], + "lines": [ + { + "bbox": [ + 452, + 243, + 495, + 286 + ], + "spans": [ + { + "bbox": [ + 452, + 243, + 495, + 286 + ], + "type": "image", + "image_path": "41fdf855274d14efd093ac4dafb7dd7d1598423a1c33639e7fd57ec4465b7a50.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 288, + 492, + 295 + ], + "lines": [ + { + "bbox": [ + 456, + 288, + 492, + 295 + ], + "spans": [ + { + "bbox": [ + 456, + 288, + 492, + 295 + ], + "type": "text", + "content": "maltesedog" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 496, + 243, + 540, + 295 + ], + "blocks": [ + { + "bbox": [ + 496, + 243, + 540, + 295 + ], + "lines": [ + { + "bbox": [ + 496, + 243, + 540, + 295 + ], + "spans": [ + { + "bbox": [ + 496, + 243, + 540, + 295 + ], + "type": "image", + "image_path": "f00026c13de96a97ac190662d2fbbec1a493d9c72b23414c66bab74d979e2135.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 299, + 517, + 308 + ], + "lines": [ + { + "bbox": [ + 492, + 299, + 517, + 308 + ], + "spans": [ + { + "bbox": [ + 492, + 299, + 517, + 308 + ], + "type": "text", + "content": "iT-L-32" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": "Figure 4. Example class visualizations. We show ImageNet class visualizations for (a) ResNet50 and (b) ViT-L-32.MACO and Fourier-based FV (top rows) produce repetitive, hard-to-interpret patterns. DeepInversion (3rd row) improves interpretability but introduces artifacts and lacks ViT compatibility. VITAL arguably yields much more interpretable visualizations. For additional results, including failure cases, we refer to supp. 2." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 420, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 480 + ], + "type": "text", + "content": "and quantitatively. All architectures were pre-trained on ImageNet. We further conducted human user studies assessing human interpretability of the resulting visualizations. We provide details on implementation and setup of competing methods with more results in the supp. 1 and 2, respectively." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 491, + 425, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 425, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 425, + 503 + ], + "type": "text", + "content": "4.1. Qualitative Results" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Visually, there is a stark difference between class visualizations of VITAL and existing work (see Fig. 4 for ResNet50 and ViT-L-32 and supp. 2 for all architectures). Where most existing methods, including MACO and Fourier-based activation maximization, show highly repetitive patterns, our approach provides much cleaner, more understandable representations of the class. While DeepInversion shows less repetition of patterns, there are often irrelevant background features, as in the case of the candle, or visual artifacts, as the blue and yellow color for the dog, dominating the image. Yet, we acknowledge that none of the methods is flawless: All approaches, including VITAL, show difficulties with complex spatial arrangements (supp. 2.6). Still, it is usually understandable what class is represented by VITAL. This becomes even more evident when comparing methods across more complex architectures (e.g., ViT, ConvNeXt), where most approaches struggle to remain inter" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4407" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 71, + 292, + 361 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 292, + 361 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 292, + 361 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 292, + 361 + ], + "type": "table", + "html": "
MethodSetupAcc. Top1 (↑)FID (↓)Zero-Shot Prediction
Top1 (↑)Top5 (↑)
ResNet50ImageNet--69.1192.23
MACOr: 22429.43360.7412.8729.73
Fourierr: 22421.30422.446.7318.27
DeepInvbs: 64100.0035.7629.9055.20
VITAL99.9058.7966.6292.56
ConvNeXtImageNet--65.6689.80
MACOr: 22466.0762.557.2019.77
Fourierr: 22460.0759.602.778.30
VITAL99.973.9263.5390.30
DenseNet121ImageNet--70.6493.16
MACOr: 2249.201.809.3323.20
Fourierr: 22415.531.634.8712.17
DeepInvbs: 64100.000.2010.0025.47
VITAL99.930.2758.7086.93
ViT-L-16ImageNet--64.7889.31
MACOr: 22444.33946.963.9310.57
Fourierr: 22425.30990.511.675.13
VITAL99.80126.2968.1792.80
ViT-L-32ImageNet--65.8390.03
MACOr: 22424.872318.9017.5337.23
Fourierr: 22417.031983.0910.3028.10
VITAL89.60147.3355.9785.47
", + "image_path": "d7cbf370372a58aad0393d5810c13449d0174368316fe3dcaf82c0304e11e5a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 370, + 295, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 295, + 425 + ], + "type": "text", + "content": "Table 1. Comparison of methods on different architectures trained on Imagenet. We provide FID scores, CLIP Zero-shot prediction scores, and top-1 classification accuracy, indicating the best and second best. In the settings, \"r\" indicates the resolution of the visualization and \"bs\" is the used batch size." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 456, + 295, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 494 + ], + "type": "text", + "content": "pretable. In contrast, VITAL consistently produces representations where the object and its class remain clear, highlighting its robustness and effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 227, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 227, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 227, + 526 + ], + "type": "text", + "content": "4.1.1. Analysis of the Embedding Space" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "To understand how realistic and distinct the visualizations are, we compare FVs of five different dog classes to original images in terms of their penultimate layer embedding. We show a 2D low-dimensional representation of the embeddings of original and generated images using tSNE [40] in Fig. 6. We find that the real images form five clusters neatly separating the classes. We further observe that fourier-based approaches with their repetitive, often recognizable patterns are collapsed on a single point, away from any real images. While DeepInversion does fall into the correct clusters, it is hard to recognize a dog in the generated images. VITAL is the only method that consistently falls into the center of the cluster, providing distinct dog features we can recognize as known dog breeds. We extend our results for other architectures in supp. 2.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 404, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 404, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 404, + 83 + ], + "type": "text", + "content": "4.1.2. Small Circuits" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 88, + 555, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 88, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 88, + 555, + 280 + ], + "type": "text", + "content": "Our method enables interpreting what inner neurons encode. By selecting the top-3 channels in the last hidden layer most relevant to a class (via LRP), we can visualize local circuits that reveal class-specific patterns (Fig. 5). While most methods, except DeepInversion, support inner neuron visualization, we compare with MACO as the strongest baseline. For example, we see that the most relevant, distinguishing features for a zebra classification are the stripes and texture of their fur. For dogs, there are leg, ear, and specifically colored fur detectors evident in the model. Such specific coloring is necessary to distinguish the plenty of dog breeds available as classes in ImageNet. We also observe signs of overfitting, where the pineapple class is associated with neurons that detect groups of fruits, many of the training images show such a picture. MACO yields less interpretable results, particularly for animals, as confirmed in the next section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 291, + 433, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 291, + 433, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 291, + 433, + 304 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 381 + ], + "type": "text", + "content": "For quantitative evaluation, we consider the classification accuracy and the FID score [32] on the respected model, as well as CLIP zero-shot prediction performance [29] in Tab. 1 (see supp. Tab. 1 for the extended version). We also provide quantitative results for the intermediate neurons with the metrics proposed by [16] in supp. 2.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": "Accuracy. To assess how well visualizations describe their target, we classify each class neuron's visualization (see Tab. 1). Despite explicitly optimizing class logits, most existing methods perform poorly. Only VITAL and DeepInversion are able to produce accurate descriptions, achieving virtually " + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": " accuracy. Investigating further, we observe that other methods including MACO seem to produce irrelevant features that mislead the model, producing predictions unrelated to the original class. We give two examples of prediction in supp. 2.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 507, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 556, + 602 + ], + "type": "text", + "content": "FID Score. For all architectures, to quantify visualization realism, we compute FID scores [32] using features from ResNet50's penultimate layer (Tab. 1). While aligned with our objective, the performance gap to other methods is substantial, with VITAL achieving over an order of magnitude better scores. DeepInversion also leverages feature statistics but is highly sensitive to batch size, which impacts both FID and visual quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "CLIP Zero-shot Prediction. So far, previous metrics relied on the model used for visualization. To assess general perceptual quality, we use a pretrained CLIP ViT-B/32 model to classify FVs via template-based prompts (see supp. 1.2 for more details). We report zero-shot scores in Tab. 1, including original (correctly classified) ImageNet images for reference. As before, we see that DeepInversion outperforms existing work, with standard baselines having overall poor performance even when considering Top-5 accuracy. No" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4408" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 194 + ], + "type": "image", + "image_path": "3439541d05f3c853dd2a59c106793cd19781c8b25fb7d084f8fcd4d69b07ab6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 198, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 255 + ], + "type": "text", + "content": "Figure 5. Visualizations of small \"circuits\". We provide visualizations of small \"circuits\" for three interesting classes of ImageNet. For each class, we found the three most relevant neurons in the penultimate layer based on LRP relevance scores and indicate the neuron ID on the arrow. For each neuron, we label its likely meaning and provide FVs by MACO and VITAL. The Zebra class shows high dependence on the characteristic pattern of the fur, the dog class a dependence on specific color patterns to be able to distinguish the abundant different dog classes in ImageNet, and Pineapple shows signs of overfitting with an association with collections of fruits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 272, + 284, + 446 + ], + "blocks": [ + { + "bbox": [ + 69, + 272, + 284, + 446 + ], + "lines": [ + { + "bbox": [ + 69, + 272, + 284, + 446 + ], + "spans": [ + { + "bbox": [ + 69, + 272, + 284, + 446 + ], + "type": "image", + "image_path": "6fd55888b520ffd7ea626b9d8dfd43c7e92791b0eee0631893c9ba0f52a201af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "type": "text", + "content": "Figure 6. " + }, + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "type": "text", + "content": "-SNE projection of embedding. We show a low-dimensional tSNE embedding of the features at the penultimate layer for five dog breeds indicated by color. Transparent circles are original training images and FVs are indicated by symbols:" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "type": "text", + "content": "VITAL, " + }, + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "type": "inline_equation", + "content": "\\triangleleft" + }, + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "type": "text", + "content": " : MACO, " + }, + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "type": "text", + "content": " : Fourier, " + }, + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 498, + 237, + 510 + ], + "type": "text", + "content": " : DeepInv." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 531, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 556 + ], + "type": "text", + "content": "tably, VITAL outperforms all others by a wide margin, approaching real image accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 219, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 219, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 219, + 577 + ], + "type": "text", + "content": "4.3. Human Interpretability Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "The primary objective of FVs is to make neural networks more understandable for human users. To address this goal, we conducted a user study (58 participants) to evaluate the interpretability of a FV from the perspective of a human. Specifically, participants (1) given a single word, evaluated how well a FV reflects the provided word, (2) evaluated how well the FVs for an inner neuron reflect the provided reference images (highly activating on the target neuron), and (3) should describe a given generated image with a word or short description. These present settings of increasing difficulty and realism. While (1) is a typical evaluation, it" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 274, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 406 + ], + "type": "text", + "content": "also comes with the largest implicit bias, as the class label already gives away what we want to visualize. In (2), by showing reference images, we encourage a visual match between FV and real image. We compare with MACO, as it is the best performing approach to visualize inner neurons. In (3), we model a real-world scenario, i.e., without any implicit bias, the user is supposed to understand what the network encodes. We provided a simple scoring system from 1(worst) to 5(best) for parts (1) and (2), and evaluated how well the textual description in (3) fits the class using closeness in the embedding space of a text encoder [5]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 407, + 556, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 622 + ], + "type": "text", + "content": "We provide an overview of the results of our study in Fig. 7. We observe that our method performs favorable across all three tasks compared to other methods, showing an order of magnitude more high-scoring visualizations (scores 4-5) in both (1) and (2), and showing much higher similarity to the correct class on task (3). In particular, for this challenging task that provides a zero-shot estimate of the interpretability of FVs, the " + }, + { + "bbox": [ + 313, + 407, + 556, + 622 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 622 + ], + "type": "text", + "content": " quantile, i.e., the \"worse\"-performing visualizations, are still close to a similarity of .5 of VITAL. Other methods show only 0.25 similarity for the same quantile. The same stark difference of more than 0.2 in median similarity score shows that VITAL provides more human-understandable visualizations of what neurons encode. We provide more fine-grained analysis, population statistics, and details in supp. 4. We further conducted a human user study following the protocol of [8, 46], which is a simple option-based evaluation similar to setting (2), where VITAL performs best (supp. 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 633, + 384, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 384, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 384, + 645 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Our experiments show that VITAL produces more understandable visualizations that better capture the unique information encoded by each neuron, outperforming MACO [8], the most recent method in classification accuracy on the target model. While DeepInversion performs similarly, this is" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4409" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 81, + 265, + 181 + ], + "blocks": [ + { + "bbox": [ + 99, + 72, + 212, + 80 + ], + "lines": [ + { + "bbox": [ + 99, + 72, + 212, + 80 + ], + "spans": [ + { + "bbox": [ + 99, + 72, + 212, + 80 + ], + "type": "text", + "content": "User Study Section 1 results (reference word)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 81, + 265, + 181 + ], + "lines": [ + { + "bbox": [ + 82, + 81, + 265, + 181 + ], + "spans": [ + { + "bbox": [ + 82, + 81, + 265, + 181 + ], + "type": "image", + "image_path": "d7575397bc50f6405f2906b72ae7b7e4c7a3982c0b68e03d6017143da4ea92d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 186, + 247, + 195 + ], + "lines": [ + { + "bbox": [ + 102, + 186, + 247, + 195 + ], + "spans": [ + { + "bbox": [ + 102, + 186, + 247, + 195 + ], + "type": "text", + "content": "(a) Class visualization with given class label." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 209, + 265, + 307 + ], + "blocks": [ + { + "bbox": [ + 100, + 198, + 213, + 205 + ], + "lines": [ + { + "bbox": [ + 100, + 198, + 213, + 205 + ], + "spans": [ + { + "bbox": [ + 100, + 198, + 213, + 205 + ], + "type": "text", + "content": "User Study Section 2 results (reference crops)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 209, + 265, + 307 + ], + "lines": [ + { + "bbox": [ + 82, + 209, + 265, + 307 + ], + "spans": [ + { + "bbox": [ + 82, + 209, + 265, + 307 + ], + "type": "image", + "image_path": "1a2a72e2396857a4eb1b062b1844e7d6c11a26070357a05c6d71289e70de8559.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 312, + 271, + 322 + ], + "lines": [ + { + "bbox": [ + 79, + 312, + 271, + 322 + ], + "spans": [ + { + "bbox": [ + 79, + 312, + 271, + 322 + ], + "type": "text", + "content": "(b) Inner neuron visualization with given activating images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 83, + 333, + 268, + 433 + ], + "blocks": [ + { + "bbox": [ + 102, + 323, + 200, + 331 + ], + "lines": [ + { + "bbox": [ + 102, + 323, + 200, + 331 + ], + "spans": [ + { + "bbox": [ + 102, + 323, + 200, + 331 + ], + "type": "text", + "content": "User Study Section 3 results (free form)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 333, + 268, + 433 + ], + "lines": [ + { + "bbox": [ + 83, + 333, + 268, + 433 + ], + "spans": [ + { + "bbox": [ + 83, + 333, + 268, + 433 + ], + "type": "image", + "image_path": "c1ffd91c692f3594e4ddd1a772fabfb0d2582adc5fbe9814a23380570ac41465.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 437, + 251, + 448 + ], + "lines": [ + { + "bbox": [ + 99, + 437, + 251, + 448 + ], + "spans": [ + { + "bbox": [ + 99, + 437, + 251, + 448 + ], + "type": "text", + "content": "(c) Class visualization with free-form labeling." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "lines": [ + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "type": "text", + "content": "Figure 7. Human Study results. We provide summary statistics on the scores for the different methods obtained for the three-part human " + }, + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "type": "inline_equation", + "content": "(n = 58)" + }, + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "type": "text", + "content": " evaluation study. In (a), participants were given class visualization and corresponding class name for 10 random classes and scored the visualizations on a scale of 1(worst) to 5(best). In (b) the task was to evaluate the closeness of FVs for an inner neuron to the image crops that it most activates on (same scoring). DeepInversion does not generate visualizations for inner neurons. In (c), participants were asked to label a given FV. We evaluated similarity of their label to the (hidden) ground truth based on Universal Sentence Encoder embeddings. We provide violin plots with median and " + }, + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 457, + 295, + 590 + ], + "type": "text", + "content": " quantiles of the achieved similarity." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 613, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 674 + ], + "type": "text", + "content": "expected since it directly optimizes the model's logit score. When evaluated by a different model such as CLIP, VITAL surpasses all baselines by a wide margin. Two human user studies further confirm that VITAL yields more interpretable visualizations in describing given labels or images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Realism. In principle, our visualizations, as well as those of other methods, do not reach photo-realism. VITAL visualizations sometimes resemble the style of paintings of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "Édouard Manet, which, however, does not limit their interpretability: Similar to a painting, it is understandable what is expressed, which is also supported by the user studies. Yet, it would be interesting to study how to reach photorealism in the future. Generative models such as diffusion offer a promising direction, though further research is needed to prevent the introduction of model-specific artifacts that reduce faithfulness to the target model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 171, + 555, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 302 + ], + "type": "text", + "content": "Evaluation. We evaluate our framework using a broad set of metrics, combining standard benchmarks with additional measures like CLIP prompting. Unfortunately, in feature visualization, more broadly in Explainability research, exhaustive benchmarks and clearly defined metrics are actively discussed, but currently still amiss. Through the additional human user studies, we complement the metrics with an evaluation of our original goal of more understandable feature visualizations. As with any user study, the design can be debated, and more refined setups with larger cohorts offer a promising direction for future work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 305, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 425 + ], + "type": "text", + "content": "Distribution Matching. Standard approaches to distribution matching include matching statistics, such as mean and variances, or considering prior distributions [15, 20], but these oversimplify the complexity of real images. Activation distributions are often too diverse for such approximations [45]. Hence, directly optimizing an alignment between the empirical distributions showed the most promise, with sort matching [30] as an efficient approach that easily scales to modern networks, but our approach is agnostic to the type of distribution matching algorithm." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 427, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 556, + 548 + ], + "type": "text", + "content": "Future work. VITAL provides a substantial improvement over existing work and we anticipate it of great use in downstream applications, such as understanding networks in the medical domain, studying what neurons encode after knowledge transfer, or how pruning affects the representations. Future directions include automatic selection of neurons or neuron sets for circuit visualization, for example using CRP [1], integration of ViT-specific visualization via AttnLRP [2], and extending our analysis to CLIP-based ViT models to better understand multimodal representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Feature visualization is key to understanding neural representations, but existing methods often yield artifacts or irrelevant patterns. We propose a new approach that aligns the feature distributions of generated and real images, and incorporates relevance scores to better capture the true information encoded in intermediate neurons. Across multiple metrics and two user studies, VITAL produces clearer and more interpretable visualizations for both humans and models. We envision VITAL supporting safety-critical fields such as medicine and complementing Mechanistic Interpretability by attributing meaning to neural circuits." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4410" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 223 + ], + "type": "text", + "content": "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. AttnLRP: Attention-aware layer-wise relevance propagation for transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 135-168. PMLR, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 267 + ], + "type": "text", + "content": "[3] Leonard Bereska and Stratis Gavves. Mechanistic interpretability for AI safety - a review. Transactions on Machine Learning Research (TMLR), 2024. Survey Certification, Expert Certification. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "text", + "content": "[4] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 10329-10338, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "text", + "content": "[5] Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. Universal sentence encoder. CoRR, abs/1803.11175, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 370, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 295, + 434 + ], + "type": "text", + "content": "[6] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. In The Advances in Neural Information Processing Systems (NIPS), pages 16318-16352. Curran Associates, Inc., 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 437, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 513 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In The International Conference on Learning Representations (ICLR), 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 514, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 601 + ], + "type": "text", + "content": "[8] Thomas Fel, Thibaut Boissin, Victor Boutin, Agustin Picard, Paul Novello, Julien Colin, Drew Linsley, Tom Rousseau, Remi Cadene, Lore Goetschalckx, Laurent Gardes, and Thomas Serre. Unlocking feature visualization for deep network with magnitude constrained optimization. In The Advances in Neural Information Processing Systems (NIPS), pages 37813-37826. Curran Associates, Inc., 2023. 2, 4, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 603, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 295, + 668 + ], + "type": "text", + "content": "[9] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadécnc, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2711-2721, 2023. 1, 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[10] Jonas Fischer, Anna Olah, and Jilles Vreeken. What's in the box? exploring the inner life of neural networks with robust rules. In The International Conference on Machine Learning (ICML), pages 3352-3362, 2021. 1, 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[11] Amin Ghiasi, Hamid Kazemi, Steven Reich, Chen Zhu, Micah Goldblum, and Tom Goldstein. Plug-in inversion: Model-agnostic inversion for vision with data augmentations. In The International Conference on Machine Learning (ICML), pages 7484–7512, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[12] Gyojin Han, Jaehyun Choi, Haeil Lee, and Junmo Kim. Reinforcement learning-based black-box model inversion attacks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 20504-20513, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[13] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[15] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In The IEEE International Conference on Computer Vision (ICCV), pages 1510-1519, 2017. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[16] Laura Kopf, Pheline Lou Bommer, Anna Hedström, Sebastian Lapuschkin, Marina MC Höhne, and Kirill Bykov. Cosy: Evaluating textual explanations of neurons. In The Advanced Conference on Neural Information Processing Systems (NIPS), 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "text", + "content": "[17] Sebastian Lapuschkin, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10:e0130140, 2015. 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 421, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 465 + ], + "type": "text", + "content": "[18] Shuying Liu and Weihong Deng. Very deep convolutional neural network based image classification using small training sample size. In IAPRA Asian Conference on Pattern Recognition (ACPR), pages 730-734, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "text", + "content": "[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "text", + "content": "[20] Ming Lu, Hao Zhao, Anbang Yao, Yurong Chen, Feng Xu, and Li Zhang. A closed-form solution to universal style transfer. In The IEEE International Conference on Computer Vision (ICCV), pages 5951-5960, 2019. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 556, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 600 + ], + "type": "text", + "content": "[21] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 5188-5196, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 601, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 624 + ], + "type": "text", + "content": "[22] A. Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 678 + ], + "type": "text", + "content": "[23] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 427-436, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4411" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "In The Advances in Neural Information Processing Systems (NIPS), page 3395-3403, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[25] Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 3510-3520, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[26] Anh Mai Nguyen, Jason Yosinski, and Jeff Clune. Multifaceted feature visualization: Uncovering the different types of features learned by each neuron in deep neural networks. CoRR, abs/1602.03616, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[27] Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization. Distill, 2017. https://distill.pub/2017/feature-visualization. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[28] Seitaro Otsuki, Tsumugi Iida, Félix Doublet, Tsubasa Hirakawa, Takayoshi Yamashita, Hironobu Fujiyoshi, and Komei Sugiura. Layer-wise relevance propagation with conservation property for resnet. In The European Conference on Computer Vision (ECCV), pages 349-364, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 363 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In The International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[30] Jannick P. Rolland, V. Vo, B. Bloss, and Craig K. Abbey. Fast algorithms for histogram matching: application to texture synthesis. Journal of Electronic Imaging, 9(1):39 - 45, 2000. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 442 + ], + "type": "text", + "content": "[31] Naomi Saphra and Sarah Wiegreff. Mechanistic? In The BlackBoxNLP Workshop at The Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 475 + ], + "type": "text", + "content": "[32] Maximilian Seitzer. pytorch-fid: FID Score for PyTorch. https://github.com/mseitzer/pytorch-fid, 2020. Version 0.3.0.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "text", + "content": "[33] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. CoRR, abs/1312.6034, 2013. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "text", + "content": "[34] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise. CoRR, abs/1706.03825, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "type": "text", + "content": "[35] Arshia Soltani Moakhar, Eugenia Iofinova, Elias Frantar, and Dan Alistarh. SPADE: Sparsity-guided debugging for deep neural networks. In The International Conference on Machine Learning (ICML), pages 45955-45987, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "text", + "content": "[36] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. In International Conference on Learning Representations (ICLR), 2015. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[37] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In The International Conference on Machine Learning (ICML), page 3319-3328, 2017. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 463 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[39] Mike Tyka. Class visualization with bilateral filters. https://science.miketyka.com/deepdream/2016/02/05/bilateral-class-vis.html, 2016.2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[40] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[41] Guangrun Wang and Philip H. S. Torr. Traditional classification neural networks are good generators: They are competitive with ddpms and gans. CoRR, abs/1706.03825, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "type": "text", + "content": "[42] Hongxu Yin, Pavlo Molchanov, Jose M. Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2020. 2, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[43] Runpeng Yu and Xinchao Wang. Generator born from classifier. In The Advances in Neural Information Processing Systems (NIPS), 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[44] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In The European Conference on Computer Vision (ECCV), pages 818-833, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "text", + "content": "[45] Yabin Zhang, Minghan Li, Ruihuang Li, Kui Jia, and Lei Zhang. Exact feature distribution matching for arbitrary style transfer and domain generalization. In The IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2022. 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 463 + ], + "type": "text", + "content": "[46] Roland S. Zimmermann, Judy Borowski, Robert Geirhos, Matthias Bethge, Thomas S. A. Wallis, and Wieland Brendel. How well do feature visualizations support causal understanding of cnn activations? The Advances in Neural Information Processing Systems (NIPS), 34:24369-24381, 2021. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4412" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_content_list.json b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0206d5af2672b8db1aad78fb30ff6b04e3b5fa68 --- /dev/null +++ b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks", + "text_level": 1, + "bbox": [ + 117, + 130, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shiduo Zhang $^{1}$ , Zhe Xu $^{1}$ , Peiju Liu $^{1*}$ , Xiaopeng Yu $^{1*}$ , Yuan Li $^{1}$ , Qinghui Gao $^{1}$ , Zhaoye Fei $^{1}$ , Zhangyue Yin $^{1}$ , Zuxuan Wu $^{1}$ , Yu-Gang Jiang $^{1}$ , Xipeng Qiu $^{1\\dagger}$ $^{1}$ Fudan University", + "bbox": [ + 191, + 202, + 805, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2d5b9d02a4369267daa42a154de08e0858272d4a89d1449eefb93d77103a1af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 270, + 251, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/caf0fadd36a258afd4859e02698f1d8efafb2cb934e1808a0914d1aeb705ffde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 271, + 416, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/def1f31efc16ba96dcf5d6586ffd38658bb1ed315966d70a515284b308743162.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 386, + 251, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e97073b7b70ac8f2f07a3822284eb064f9c9f5f37b2ea4cc873deac37cde16ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 387, + 416, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/746e022f6a6837f28187d6f2271c164963d7425ffc48671deb445672556da165.jpg", + "image_caption": [ + "Figure 1. Overview of VLABench. VLABench is a large-scale language-conditioned manipulation benchmark to evaluate the comprehensive skill learning and generalization ability of action policies especially pre-trained vision-language-action models." + ], + "image_footnote": [], + "bbox": [ + 94, + 473, + 254, + 580 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d9d6bd8001aa8ec88ee2192a33ad6ac9502628d5b6a3c016723deccc3b35c76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 472, + 415, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task-Related Comprehensive Ability", + "text_level": 1, + "bbox": [ + 542, + 277, + 776, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4579bc73e65769ac1650ed26881978228912ce4383f267f8e61644907781627e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 301, + 650, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3071cec2797462e47d4a174f32a1306820d137bd8d5a3f1cb4bdf4a76b64beb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 305, + 880, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b262b251d3d39fec7131e9ce6112260537a66252982bd725d5be59653fb259d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 396, + 650, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d28972f469966ffdc8407caba0dd83510eeb5b734e1c9bfa9efc7534a3ade0ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 396, + 880, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fff038d2f224a9d9721e63eaea3f4c556337d0592bde01a0035665fb18fe9748.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 486, + 650, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12f1307e6cfd524303b97c6cbadbd3823073bceffdedde6d41c3bcf7150f1a6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 486, + 880, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 635, + 326, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "General-purposed embodied agents are designed to understand the users' natural instructions or intentions and act precisely to complete universal tasks. Recently, methods based on foundation models especially Vision-Language-Action models (VLAs) have shown a substantial potential to solve language-conditioned manipulation (LCM) tasks well. However, existing benchmarks do not adequately meet the needs of VLAs and relative algorithms. To better define such general-purpose tasks in the context of LLMs and advance the research in VLAs, we present VLABench, an open-source benchmark for evaluating universal LCM task learning. VLABench provides 100 carefully designed categories of tasks, with strong randomization in each category of task and a total of $2000+$ objects. VLABench stands out from previous benchmarks in four key aspects: 1) tasks", + "bbox": [ + 88, + 672, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "requiring world knowledge and common sense transfer, 2) natural language instructions with implicit human intentions rather than templates, 3) long-horizon tasks demanding multi-step reasoning, and 4) evaluation of both action policies and language model capabilities. The benchmark assesses multiple competencies including understanding of mesh&texture, spatial relationship, semantic instruction, physical laws, knowledge transfer and reasoning, etc. To support the downstream finetuning, we provide high-quality training data collected via an automated framework incorporating heuristic skills and prior information. The experimental results indicate that both the current state-of-the-art pretrained VLAs and the workflow based on VLMs face challenges in our tasks.", + "bbox": [ + 511, + 637, + 908, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Codes and more videos are available at https://vlabench.github.io/ \n2Corresponding to: sdzhang23@m.fudan.edu.cn, xpqiu@fudan.edu.cn", + "bbox": [ + 529, + 875, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11142", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-conditioned manipulation represents a fundamental challenge in embodied AI and a stepping stone toward Artificial General Intelligence [1, 5, 14]. Such tasks require agents to master multiple capabilities: interpreting natural language instructions, understanding complex environments, making decisions, formulating plans, and executing precise actions. The rapid advancement of Large Language Models (LLMs) and Vision-Language Models (VLMs) [1, 15] has revolutionized the field with their impressive general abilities in semantic understanding, coding, planning, and reasoning. The strong generalization capabilities has inspired two main approaches in language-conditioned manipulation: pre-training vision-language-action models using large-scale robotics data, as demonstrated by RT-2 and Palm-E [5, 14, 50], and integrating foundation models into agent workflows, like VoxPoser and Copa [24, 25], which combine LLM/VLM outputs with grasp prediction [16, 17] and motion planning algorithms [28].", + "bbox": [ + 89, + 114, + 485, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While real-world robotics experiments provide valuable insights, their complexity and environmental variability often challenge reproducibility. Simulation-based evaluation has emerged as a fair and practical alternative. Existing benchmarks like RLBench, Calvin, and LIBERO [27, 38, 45] offer diverse task sets but fall short in addressing the unique requirements of foundation model-based methods. Tasks designed to align with the capabilities of foundation model-based algorithms should encompass nuanced semantic understanding of user intent, the integration of common-sense knowledge, and a robust ability to interpret diverse visual scenes, as well as require sophisticated multi-step reasoning. Such tasks demand a sophisticated integration of multimodal understanding to effectively interpret and respond to complex, real-world contexts. For example, one of the tasks in RT-2 [5] is \"move the Coke can to Taylor Swift\", while another task in CoPA [24] is \"make me a cup of pour-over coffee\". The first task challenges the robot to use common sense to identify Taylor, a capability of knowledge transfer that previous policies struggled to achieve. The second task further intensifies the difficulty, requiring the robot to decompose the task into subtasks and execute the steps to operate a coffee machine—a long-horizon challenge that has previously been difficult for a single policy to accomplish.", + "bbox": [ + 89, + 402, + 483, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better define the types of language-conditioned manipulation tasks suited for foundation models and provide a standardized evaluation suite to advance robotics research, we introduce VLABench. VLABench is an open-source benchmark specifically designed for methods utilizing foundation models. The tasks in VLABench are carefully divided into several dimensions to evaluate models across various aspects, including 1) Mastery of common", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sense and world knowledge, 2) Understanding of mesh and texture, 3) Comprehension of semantically rich instructions, 4) Spatial understanding, 5) Grasp of physical rules, and 6) Reasoning ability. For benchmarking purposes, VLABench offers 100 task categories with comprehensive evaluations across various methods. With a diverse collection of over 2,000 3D objects and scenes, VLABench creates a wide range of visual contexts and tasks. It enables the assessment of generalization capabilities through learning across multiple skills, providing thorough evaluations spanning visual, linguistic, planning, knowledge transfer, and action dimensions.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To ensure fair comparison and evaluation, we develop an automated data collection framework to construct standardized datasets for each task, supporting model training and fine-tuning. Using this dataset, we conduct extensive experiments to evaluate and benchmark three distinct types of approaches: pre-trained VLA, workflows integrating foundation models, and vision-language models (VLMs). The experimental results indicate that existing VLA methods perform poorly on our tasks and don't exhibit the level of generalization abilities or the \"emergent\" phenomena observed in large models [61]. We summarize contributions as follows:", + "bbox": [ + 511, + 272, + 906, + 450 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose VLABench, the first benchmark designed to comprehensively evaluate the capabilities of VLAs and VLMs in robotics manipulation tasks, covering multiple dimensions such as skills, vision, language, task execution, common sense, and reasoning.", + "- We define 100 novel LCM tasks tailored to the capabilities of foundation models within a standardized evaluation framework. These tasks require a deep understanding of semantics, vision, space, and physical laws, as well as the ability to plan long-horizon tasks and transfer world knowledge and common sense into task execution.", + "- We provide a scalable data construction framework and a standardized evaluation dataset. This automated data construction approach facilitates future research on pretraining robotics data.", + "- Our experiments demonstrate that current pre-trained VLAs have yet to exhibit the strong generalization capabilities observed in LLMs, and existing SOTA VLMs also show limitations in embodied scenarios." + ], + "bbox": [ + 513, + 455, + 906, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 753, + 663, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Benchmarks and Datasets. Numerous benchmarks such as RLBench and LIBERO [27, 35, 38, 45, 66] have been proposed to evaluate language-conditioned manipulation policies. A comparison of these benchmarks is provided in Table 1. However, most of these focus on skill learning and fail to sufficiently address long-horizon planning capabilities. Meanwhile, some benchmarks [52, 55, 63] address room-scale mobile manipulation tasks that require", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11143", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/1232b6c324178c581c0d0b44f6da34ef28a3d18a67473917a4fc34110b35c20d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison of popular benchmarks in robot learning. SemLang: semantically non-template rich language instructions. LogiReason: task logic and relevant information reasoning. Knowledge: tasks require the application of common sense and world knowledge. DL: quantitative task difficulty level. DR: task domain randomization. N-task: total number of tasks. Cate-object: categories of assets used in the simulation. N-object: total number of objects in the asset library. AI-Gen: use of generative AI models for asset library creation. MultiCam: use of multiple cameras. PCD: support point cloud data in 3D methods. Cross Emb: support for cross-embodiment. Auto Traj: supporting automated data collection." + ], + "table_body": "
BenchmarksSemLangLogiReasonKnowledgeDLDRN-taskCate-objN-objAI-GenMultiCamPCDCross EmbAuto Traj
Alfred[52]XXX-X7-3578XXX✓X
Rlbench[27]XXX28.86✓10028168X✓XX✓
Calvin[45]XXX40.00X34530X✓XXX
ManiSkill[21, 46, 56]XXX-✓201002600X✓✓✓✓
LIBERO[38]XXX17.96✓1305175XXXXX
RoboCASA[47]XXX-✓1001532509✓✓X✓✓
ARNOLD[20]XXX--8-40✓✓✓X✓
SimplerEnv[35]XXX27.66✓8-17✓✓✓X✓
Behavior-1K[32]XXX--100022119331XXX✓X
Habitat 2.0[55]XXX--346169XXXXX
VLABench✓✓✓75.96✓1001632164✓✓✓✓✓
", + "bbox": [ + 94, + 88, + 903, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "long memory and reasoning. Yet, these interactions typically occur through interfaces, rather than direct physical manipulation, limiting the transferability of learned policies to real-world scenarios. Additionally, while efforts [21, 32, 38, 47] have made strides in task format, difficulty, and scale, these benchmarks have largely overlooked the guiding role of language in tasks, often relying on template instructions. VLABench is the first to introduce features such as natural human interaction, implicit goal-oriented semantics, and requirements based on common sense into robot manipulation tasks, as shown in Figure 20. In terms of generalization evaluation, previous works [27, 38, 45] typically assess models at the instance level within the same category, which limits their ability to evaluate generalization across diverse object categories or different tasks within the same skill set. In contrast, VLABench is the first benchmark to evaluate generalization capabilities across a wide range of tasks, object types, and task categories, providing a more comprehensive assessment of model versatility.", + "bbox": [ + 88, + 339, + 480, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large-scale datasets have been built in both real and simulation [6, 47, 50, 60] for large-scale imitation learning. However, real-world data faces challenges related to scalability, making it difficult to gather sufficient data at scale [4]. Simulated datasets, while more scalable, often suffer from limited diversity in scenarios and tasks [27, 45], and still require teleoperation [20, 38] for data collection. VLABench addresses these limitations by offering a broader range of tasks that are more closely aligned with real-world conditions, covering diverse aspects of vision, language, tasks, and skills. Furthermore, it introduces an efficient and robust process for the automated generation of simulated data, significantly enhancing task diversity and scalability.", + "bbox": [ + 89, + 628, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pretrained Vision-Language-Action Models. The recent rise of multimodal models [1, 13, 40, 65] and the organization of operational datasets [49, 60], have led to the integration of vision-language-action models (VLAs) [4, 5, 14, 30] into language-conditioned manipulation tasks. While the", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "term VLA generally refers to models that combine visual and language inputs for policy learning, we focus specifically on approaches leveraging pre-trained models. Several works [4, 14, 30] have applied further training to pre-trained vision-language models (VLMs) for language-conditioned manipulation. These models demonstrate impressive generalization to unseen objects and tasks, yet their control precision is somewhat limited by the discretization of actions [51]. To address this limitation, some approaches have explored using diffusion models [9, 51] as policy networks or using diffusion decoders [3, 34, 62]. Pre-trained models based on diffusion models [37, 43] have shown promising advancements in improving continuous space distribution learning. VLABench includes a selection of these representative methods for comprehensive evaluation.", + "bbox": [ + 511, + 339, + 903, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Framework Utilizing Foundation Models. Pre-trained language models [6, 15, 53] and vision-language models [1, 42] have demonstrated strong generalization and versatility. Some researchers [24, 25, 39] combine the general perception and cognitive abilities of these pre-trained models with traditional planning and control algorithms to create agent workflows. These frameworks allow robots to perform complex zero-shot manipulation tasks without requiring additional training. To harness the capabilities of foundation models for manipulation, some works [25, 26, 36] utilize the code comprehension and generation abilities of large language models alongside motion planning optimization algorithms to tackle fundamental manipulation tasks. Additionally, some methods [23, 24] leverage large models to decompose long-horizon tasks into subtasks, then integrate perception and trajectory generation modules to construct the entire manipulation pipeline. However, most zero-shot methods of this kind heavily rely on prompt design [25], the accuracy of each module, and even the specific parameters of the models being invoked [24]. While these methods demonstrate strong generalization, they often face challenges with accuracy. VLABench provides a", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11144", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "zero-shot evaluation framework to assess the performance of such workflows and offers insights into their effectiveness.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. VLABench", + "text_level": 1, + "bbox": [ + 89, + 154, + 210, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Task Description", + "text_level": 1, + "bbox": [ + 89, + 180, + 256, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VLABench is composed of 60 primitive and 40 composite tasks, categorized by task difficulty levels. These tasks are designed to encompass a rich variety of skills while covering ample visual and language semantic information. For skill learning, 100 tasks in VLABench cover a wide range including 1) Pick&place, 2) Open&close door, 3) Open&close drawer, 4) Hang objects on the wall, 5) Use tool e.g. Hammer nail, 6) Press button, 7) Insert, 8) Pour, 9) Twist, and 10) Explore. In addition, VLABench places greater emphasis on real-life scenarios and essential daily tasks, representing more interactive language instructions, a wider variety of task settings, the integration of common sense and societal knowledge, and long-horizon tasks requiring logical planning, as shown in Figure 2. Notably, VLABench adopts a stricter definition of task generalization, which will be elaborated on in the Section 3.3. The whole task list can be found in supplementary material.", + "bbox": [ + 89, + 203, + 483, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Primitive Tasks. Primitive tasks are divided into five dimensions, each corresponding to the assessment of a specific ability dimension.", + "bbox": [ + 89, + 460, + 483, + 507 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Mesh&Texture Understanding. This type of task requires the model to recognize different meshes and understand various texture features. Take the SelectToy task shown in Figure 2 (a) as an example, the robot is directly required to place a specific toy e.g. Aquaman into a receptacle. The model must possess strong visual capabilities to accurately recognize such complex meshes and textures.", + "- Spatial Understanding. Spatial understanding tasks involve various spatial relationships, such as the $n$ -th left/right position, inside/outside of a receptacle, the $m$ -th row and $n$ -th column, near/far, and beside a specific object, representing relative positional relationships. Figure 2 (b) shows one case in task in PullBook. Such a complex relative positional relationship imposes extremely high demands on the model's multimodal understanding [8].", + "- Common Sense & World Knowledge. Tasks relative to common sense/world knowledge require the agent to transfer the knowledge gained in the pre-train stage to solve the problem. The task shown in Figure 2 (c) requires the agent not only to recognize different types of flowers from visual information but also to leverage world knowledge to determine that \"the tulip is the national flower of the Netherlands\".", + "- Semantic Understanding. This type of task emphasizes the complexity, subtlety, and natural interactivity of lan" + ], + "bbox": [ + 89, + 508, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "guage instructions. Task objectives are often implicitly conveyed through a natural conversation. To perform well in GetDrink task as Figure 2 (d) shown, the agent must capture the implied request from a lengthy instruction: to take out a chilled cola from the refrigerator.", + "bbox": [ + 524, + 90, + 906, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Physical Law. This type of task expects the robot to integrate visual information and take correct actions based on physical principles and real-time observation. In UseSeesaw task in Figure 2 (e), the robot is commanded to grasp an object that can not be achieved directly. The agent must recognize the need to apply the principle of leverage by using sufficient weight to lift the target object on the other end.", + "bbox": [ + 513, + 167, + 906, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Composite Tasks. Composite tasks in VLABench involve the combination of multiple skills, long-term task planning, and multi-step logical reasoning from instructions, scenes, and even game rules. Figure 2 (f) showcases a variety of challenging complex tasks. Composite tasks have a significantly longer trajectory horizon, with an average episode length exceeding 500 timesteps—considerably more than the average of 120 timesteps for primitive tasks. In Figure 2 (f1), the agent must not only correctly identify all poker cards from visual information and use world knowledge of poker rules to select the best hand, but also flip face-down cards to acquire complete information. This type of task, requiring the agent to consciously fulfill prerequisite conditions, has never been modeled before. Composite tasks also require extracting the user's implicit needs from natural dialogue. The case in Figure 2 (f2) needs the agent to place the Python textbook on the table and open the laptop, without direct instruction. More cases are shown in Figure 20.", + "bbox": [ + 511, + 287, + 908, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Task Difficulty Level", + "text_level": 1, + "bbox": [ + 511, + 571, + 709, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VLABench stands out among language-conditioned manipulation benchmarks not only for its multi-dimensional task design and evaluation but also for its hierarchical difficulty levels. The difficulty level (DL) is quantified using three metrics: task variants [27, 44], task sub-steps, and reasoning steps. These metrics quantify the overall task difficulty level from the perspectives of generalization difficulty, normalized episode horizon length, and intelligence. The detailed computation of DL is provided in Section 6.3. As shown in Table 1, VLABench's difficulty level surpasses all previous benchmarks and introduces new challenges that a VLA model should overcome.", + "bbox": [ + 511, + 594, + 908, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.Benchmark", + "text_level": 1, + "bbox": [ + 511, + 787, + 640, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation. VLABench organizes evaluations into three categories: assessments of pretrained or fine-tuned vision-language-action (VLA) models, heuristic workflows that integrate foundation models with various algorithms, and multi-dimensional evaluations of vision-language models.", + "bbox": [ + 511, + 809, + 906, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Generalization Ability of VLAs. For pre-trained VLA", + "bbox": [ + 511, + 885, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11145", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/63f654739307941224392488c182ed6f9d04c8e7e800c016b1708463ea988615.jpg", + "image_caption": [ + "(a) Mesh & Texture" + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 254, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0df88616ed1ba70b88e77684ae751625e4823c4092bad9a45ece153918ed947d.jpg", + "image_caption": [ + "Instruction: \"Take out the second book from the right on the top layer of the shelf.\"", + "(b) Spatial Understanding", + "(c) Common Sense & World Knowledge" + ], + "image_footnote": [], + "bbox": [ + 259, + 125, + 415, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d520434d48035729c23df408a27935ac2c17b0c8d1434a4772a523734a5adc0f.jpg", + "image_caption": [ + "Instruction: \"Insert the national flower of the Netherlands into the vase.\"" + ], + "image_footnote": [], + "bbox": [ + 419, + 125, + 576, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18be63a104d2609547b39fecdd9318988aa1b99cedd68998177bcbad48691359.jpg", + "image_caption": [ + "Instruction: \"Ah, I just spent an hour working out at the gym. I'm so thirsty now. I really go for an ice-cold soda!\"", + "(d) Semantic Understanding" + ], + "image_footnote": [], + "bbox": [ + 581, + 125, + 738, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/571d946200018b56978cb39de311371db275f5a04885887bacc1cfdd1708f190.jpg", + "image_caption": [ + "Instruction: \"Get the fruit from the container\"", + "(e) Physical Law" + ], + "image_footnote": [], + "bbox": [ + 741, + 125, + 901, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c63f423c0b2139c629cf11feeab2a04c7edc90e00df750ff9b1bfcd0e555f0f8.jpg", + "image_caption": [ + "Instruction:\"We are playing Texas Hold'em game! Choose your largest combination on the table.\"", + "(f1) Play Texas Hold'em", + "Figure 2. Task examples in each dimension. The first row showcases examples of primitive tasks from Section 3.1, while the second row presents examples of composite tasks." + ], + "image_footnote": [], + "bbox": [ + 94, + 296, + 254, + 417 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4db8d16895c770d823bda986666d596ad56c23c809f48e8c2607e81f18d3755b.jpg", + "image_caption": [ + "Instruction: \"Shortly, I'm going to work on my Python course assignment. Help me set up my desk.\"", + "(f2) Manage Study Table" + ], + "image_footnote": [], + "bbox": [ + 258, + 296, + 415, + 414 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02f7f8bac0f37ab64c3eef82bfdf8ac6eed925f22ac5b9ccfa16aa5f18a8243b.jpg", + "image_caption": [ + "instruction:\"I'd like to make a garlic eggplant dish.Prepare the ingredients I will use,thanks!", + "(f3) Cook Dishes as Menu", + "(f) Type: Composite" + ], + "image_footnote": [], + "bbox": [ + 419, + 296, + 576, + 417 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8607e1c30f4a928917b5e33ad359ee964f1be2fc914afd1b63f1fa3322c806fb.jpg", + "image_caption": [ + "Instruction: \"Could you please show the acid base neutralization experiment?\"", + "(f4) Take Chemistry Experiment" + ], + "image_footnote": [], + "bbox": [ + 581, + 295, + 738, + 417 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1bddbeedd0d2116299958b9c8f642d741c2fc4e1a480d7bd13317fa06a41980e.jpg", + "image_caption": [ + "Instruction: \"I have to go to class this afternoon but I'm so sleepy! I want to have latte.\"", + "(f5) Get Latte Coffee" + ], + "image_footnote": [], + "bbox": [ + 741, + 296, + 901, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models, the evaluation focuses on the model's generalization ability across multiple dimensions. In the standard evaluation, we have set up six evaluation tracks.", + "bbox": [ + 102, + 500, + 482, + 544 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Track 1: Rapid downstream task transferability and skill acquisition ability. The evaluation episodes use objects seen in the training set, with minimal differences.", + "- Track 2: Cross-category visual generalization. The evaluation episodes use unseen categories of objects as the target entity for evaluation.", + "- Track 3: Common sense application. Building upon Track 1, replace parts of the instructions that directly reference the target object with content requiring common sense reasoning in evaluation episodes.", + "- Track 4: Semantically rich instruction adaptation. Building upon Track 1, replace the instructions with unseen and more complex ones.", + "- Track 5: Cross-domain behavior transferability. The evaluation tasks are replaced with ones that differ from those in the training set, but require similar actions.", + "- Track 6: Long-horizon task learning. Training and evaluation using more challenging composite tasks require additional skills and reasoning." + ], + "bbox": [ + 104, + 545, + 482, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Zero-shot Transfer Ability of Heuristic Workflow. Training-free workflow methods are evaluated under sets of tasks in all dimensions mentioned in Section 3.1. We", + "bbox": [ + 89, + 847, + 482, + 891 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "also extend the evaluation to cover various skills and long-horizon tasks to assess the overall capability and execution robustness of the workflow. All agents are evaluated in a zero-shot setting.", + "bbox": [ + 524, + 500, + 903, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Comprehensive Evaluation of VLMs' Capabilities. Similar to heuristic workflows, the evaluation of VLMs is also comprehensive. Since VLMs lack intrinsic action capabilities, we organized a skill library and integrated it into a domain-specific language (DSL) [48, 54], leveraging annotated asset information as prior knowledge. This DSL functions as a straightforward API that VLMs can call, enabling efficient interaction. The evaluation pipeline is shown as Figure 3 and the details will be discussed in Section 9.3.", + "bbox": [ + 513, + 560, + 905, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metric. In addition to the success rate (SR), considering the long-horizon nature and high difficulty level of our tasks, we introduce the intention score (IS) and progress score (PS) for more granular evaluation. Intention score is a useful soft metric measuring the correctness trend in selecting the target object for the operation especially when sub-tasks are not yet completed. The discrete version metric is:", + "bbox": [ + 511, + 710, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI S = \\left\\{ \\begin{array}{l l} 1, & \\text {i f d i s t a n c e} _ {\\min } < \\epsilon \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 824, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $distance_{min}$ is the minimum distance between the end-effector's position and the target entity, and $\\epsilon$ is empiri", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11146", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d9d822663fbed6d68fce1a5811f581e9a193f3368dc74a27d7bdef3ddf74f171.jpg", + "image_caption": [ + "Figure 3. Evaluation pipeline for VLMs. Step 1: Sample the required four-view images, as well as those segmented with numerical information, from the simulation. Meanwhile, save the corresponding instructions and operation sequences. Step 2: Input the original and annotated images and the instructions into the model. Then, obtain the model's output and extract the generated operation sequence. Step 3: The two operation sequences are evaluated using four metrics, which are weighted and summed to produce the final score. The red part represents the error output of the model, where the red solid arrows represent the dependencies generated by the error, and red dashed arrows represent the dependencies lost by the error." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 367, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d81d82de74d5d5819fa74b0e4570f8a6d0e48fa597d833977c5a00114ce82930.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 88, + 660, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/09dfe8349fa35fd6928ebf62a8c1fb5c1e1ffd7527aa94b7e03af43aaabb7f0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 88, + 905, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cally set to 0.15 meters in most cases after experimentation. The progress score refers to the completion level of subtasks in a long-horizon task and serves as a softer process supervision metric compared to the success rate.", + "bbox": [ + 89, + 385, + 482, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP S = N _ {\\text {d o n e}} / N _ {\\text {s u b}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 459, + 482, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{done}$ is the number of completed sub-tasks, and $N_{sub}$ is the total number of task sub-tasks.", + "bbox": [ + 89, + 488, + 483, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Dataset Construction", + "text_level": 1, + "bbox": [ + 89, + 530, + 290, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory Generation. As human teleoperation is time-consuming and not scalable [38, 47], we developed an efficient, scalable automated data collection pipeline based on our custom skill library. Inspired by [20], our data collection framework leverages the prior information including point clouds of the environment, annotated entities' grasppoints, target entity at the current step, etc. The data collection framework includes multiple task-specific motion planners. These motion planners call upon the skills in the skill library based on the current task progress and determine parameters by incorporating prior information. Subsequently, the selected skills generate trajectories using RRT [29], with quaternion interpolation achieved through spherical linear interpolation. The final trajectory is smoothed using a Bezier curve to optimize path quality. To enhance sample efficiency, reject sampling and failure-triggered early termination are applied. We demonstrate the superiority of our data framework through comparative experiments. Detailed analysis can be found in Section 8.3.", + "bbox": [ + 89, + 551, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Diversity and Quality. During the data construction process, we introduced diverse task variants and domain randomization across different episodes of the same task to ensure the diversity of the training data, as discussed", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Sections 6.1 and 6.4. Following the approach of previous benchmarks built on Mujoco [38, 47], the dataset is stored in the same format, with similar visual rendering quality and trajectory accuracy. We further analyze the data quality in Section 8.5.", + "bbox": [ + 511, + 385, + 906, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction Augmentation. We use GPT-4o [1] to generate descriptions that incorporate target-specific characteristics and interactive instructions that encompass a variety of contexts and intentions. Section 8.4 provides more details of the instruction generation.", + "bbox": [ + 511, + 460, + 906, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 551, + 643, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following Section 3.3, we conducted experiments centered on pre-trained VLA models, workflows incorporating multiple algorithmic modules, and various VLMs. The remainder of this section provides a detailed description of the experimental setup.", + "bbox": [ + 511, + 577, + 905, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Generalization Ability of VLAs", + "text_level": 1, + "bbox": [ + 511, + 662, + 790, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretrained VLAs are expected to possess robust generalization and versatility similar to LLMs. Experiments about are set to address the following research questions:", + "bbox": [ + 511, + 686, + 905, + 731 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q1: Do pre-trained VLAs exhibit stronger general abilities with unseen categories of objects?", + "Q2: Can pre-trained VLAs transfer their general knowledge and behavioral abilities to similar but unseen tasks?", + "Q3: Can pre-trained VLAs understand natural user interactions and implicit goal requirements?", + "Q4: Do pre-trained VLAs have the potential to transfer their world knowledge to related tasks?", + "Q5: Can existing VLA architectures accurately support the completion of long-horizon tasks?" + ], + "bbox": [ + 511, + 732, + 903, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment Setup. To investigate the questions outlined", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11147", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f38dfb7afede4937982dbc97f954a8c7a5c9798c63cb48f27537dddc2dd7a079.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTrack 1Track 2Track 3Track 4Track 5Track 6
ISPSSRISPSSRISPSSRISPSSRISPSSRISPSSR
Octo[57]32.05.62.020.42.00.022.02.81.226.84.01.213.60.00.017.23.00.0
OpenVLA[30]50.013.26.435.67.03.238.08.03.238.09.22.816.00.00.028.89.60.4
RDT-1B[43]50.014.36.039.28.23.236.07.24.035.68.04.420.40.00.021.69.20.0
π0-Base[3]59.629.313.249.621.46.060.420.05.260.419.33.622.00.00.034.412.82.0
π0-Fast[3]59.925.59.247.616.52.450.015.43.652.017.26.825.60.00.034.014.41.6
", + "bbox": [ + 96, + 88, + 903, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Overall experiment result of 6 evaluation tracks of fine-tuned VLAs. The detailed result of each task is reported in Table 9.", + "bbox": [ + 96, + 191, + 898, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "above, we fine-tuned various pre-trained VLA architectures [3, 30, 43, 57] on our high-quality dataset, and evaluated them on different tracks mentioned in Section 3.3 to assess their performance across a range of tasks and conditions. Considering the high task difficulty, we selected five primitive tasks for experiments in tracks 1 to 5, with 200 data samples per task, and trained the models following a multitask training approach, following the setting in previous work [3, 43]. We then assessed the models across 50 episodes for each task, resulting in a total of 250 trials per track. For track 6, we selected five composite tasks, keeping the other settings consistent. Section 9.1 provides a complete list of tasks in different tracks and details of the model fine-tuning process.", + "bbox": [ + 89, + 220, + 480, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Result and Analysis. Table 2 presents the evaluation results across multiple tracks. These models perform poorly on most of tasks, as they are trained on a more diverse and limited set of fine-tuning data, with a higher difficulty level. Through the experimental results of each track and comparisons across different tracks, our analysis is as follows:", + "bbox": [ + 89, + 431, + 480, + 521 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The primary cause of poor performance is the insufficient ability for rapid transfer in downstream tasks. Episodes in Tracks 1 and 6 closely resemble the training set, primarily assessing the model's basic fitting ability. Despite this, the models still perform poorly on certain tasks. We believe the main reason for the failure of most generalization experiments is the inaccuracy in basic action generation, especially Octo.", + "- Pre-trained VLA models exhibit some ability to handle unseen object categories. By comparing the results of Track 1 and Track 2, we observe that although all models experience a performance decline in Track 2, they are still able to complete a small subset of tasks. This is likely due to the benefits of large-scale robotics data pretraining or even earlier VLM pretraining.", + "- Performance decline with unseen semantic language. By combining the experimental results of Track 3 and Track 4, we find that the model's language understanding ability has deteriorated by a larger margin. Among these, the performance decline is more pronounced in models based on non-VLM architectures: RDT's progress score decreases by $45.2\\%$ while OpenVLA and $\\pi_0$ only decreases by $31.3\\%$ and $34.2\\%$ , respectively, in Track 4.", + "- Nearly absent cross-domain transferability. We found that all the models fail to perform similar but unseen tasks," + ], + "bbox": [ + 89, + 523, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "due to significant visual and linguistic input differences. These policies exhibit uncertainty in behavior when faced with unfamiliar tasks, resulting in IS that are similar.", + "bbox": [ + 526, + 220, + 903, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further analyze that data scaling and pre-training contribute positively to the model's performance Through basic ablation experiments. We also discuss in detail the potential issues with current VLAs, such as multimodal data co-training and model architecture designs. These discussions are presented in Section 10.1. We also discuss the gap between real and simulated environments in Section 10.2.", + "bbox": [ + 511, + 267, + 903, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Zero-shot Ability of Agent", + "text_level": 1, + "bbox": [ + 511, + 388, + 751, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For our evaluation of foundation model-based algorithms, we reviewed two state-of-the-art frameworks, Voxposer [25] and CoPA [24], and the comparison results are shown in Figure 4. Given Voxposer's dependence on large language models (LLMs), we assessed Voxposer's performance with and without visual perception capabilities. While Voxposer performed adequately on basic tasks and achieved the Progress Scores of 30-40, its reliance on LLM-driven motion planning often led to grasping failures due to limited information for effective grasp planning, especially when interpreting rotation in non-visual contexts, resulting in low overall scores.", + "bbox": [ + 511, + 412, + 903, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interestingly, the foundational LLM alone maintained relatively stable scores in semantic understanding and reasoning tasks without visual input. However, adding visual perception slightly reduced performance in these areas while significantly improving spatial reasoning, where LLM-only setups struggle with spatial accuracy due to lack of spatial information.", + "bbox": [ + 511, + 595, + 903, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The lack of closed-loop feedback limits these models' ability to perform physical reasoning tasks, particularly those involving dynamic interactions, leading to lower scores in this dimension. Both models struggle with high-complexity tasks, succeeding mainly in entity recognition but performing poorly in long horizon task reasonable breaking down. This finding underscores the requirements for advancing foundational model-based frameworks to address complex reasoning. In fact, although the methods mentioned above emphasize their zero-shot capabilities and generalization to new scenarios, their modular design often limits the upper bound of their performance. A more detailed discussion of this can be found in Section 10.3.", + "bbox": [ + 511, + 704, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11148", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab42e09e91e4719e32faa7ee06b77fc25cb63c113ca620233be90b9cc5651c74.jpg", + "image_caption": [ + "Figure 4. Evaluation progress score for Voxposer and CoPA. Voxposer w/o refers to the version without visual perception, where ground truth labels are directly provided for object selection. Voxposer w uses GPT-4V as the visual perception module." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 480, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comprehensive Ability of VLMs", + "text_level": 1, + "bbox": [ + 89, + 426, + 377, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We referred to the evaluation results of multiple series of VLMs provided by OpenCompass [11] and selected models from different families [1, 2, 7, 19, 22, 41] with strong overall performance as shown in Figure 5. We evaluate the comprehensive performance of these models with the dataset derived from naturally self-contained information within a simulated environment. This dataset consists of a complex set of tasks designed to assess the VLM's ability to perceive visual stimuli and comprehend verbal instructions.", + "bbox": [ + 89, + 450, + 483, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In line with the evaluation framework outlined in the Section 3.3, we evaluate the models across six dimensions of task performance. The results for each VLM model under the 1-shot setting are summarized in Figure 5.", + "bbox": [ + 89, + 587, + 483, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although these VLMs perform well on most multimodal tasks and even some embodied tasks [11, 33], their performance, including that of GPT-4o, falls short when faced with more complex scenarios, instructions, and more challenging tasks. We surprisingly find that the open-source model Qwen2-VL-7B-Instruct performed competitively, surpassing GPT-4o on certain dimensions. However, all models struggled with complex tasks, especially those requiring long-term task decomposition and logic reasoning. Only GPT-4o achieved a score in the reasoning dimension comparable to those in other dimensions, while the other models scored around 20. Besides, performance declines significantly when linguistic instructions transition from direct semantics to abstract meanings, as shown in the", + "bbox": [ + 89, + 648, + 483, + 861 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6d6e90334577f95dcf7c685ff912e3eb229070e26649f3421df62328a190dfd.jpg", + "image_caption": [ + "Figure 5. Radar charts depicting the performance of all VLM models across six dimensions. The reason why only GLM-4V-9B is evaluated in a zero-shot setting is that it does not support multigraph inference, which is required for the other models." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 901, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "semantic dimension. Different models appear to have distinct areas of expertise, e.g. LLaVA-NeXT exhibits weaker spatial perception ability, GLM-4V-9B excels in spatial and even physical law dimensions but lag in semantic comprehension. In Section 10.4, we conducted additional ablation experiments focusing on CoT and few-shot learning, followed by further analysis. Detailed data and error analysis can be found in Table 10. Overall, while the models demonstrate promising capabilities, their understanding and planning in embodied environments remain limited, highlighting the need for further advancements.", + "bbox": [ + 511, + 428, + 906, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 611, + 633, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose VLABench, a large-scale benchmark designed for tasks with long-horizon and multi-dimensional reasoning. Such reasoning involves many dimensions, including implicit semantic goal extracting, logical reasoning, and long-horizon planning. The most important thing we do is to provide a positive definition of the capabilities that intelligent agents with cognitive abilities should possess and the tasks they should be able to perform, through the provision of 100 standardized tasks. Additionally, VLABench constructed a scalable automated data collection framework for future's potential larger scale pertaining and a standardized dataset for fair comparison of VLAs, both in the present and in future developments. Our diverse and multiple experiments revealed that current VLAs and VLMs face significant challenges in our tasks, and there remain substantial uncertainties in research on robotics scaling. We hope that VLABench will inspire both the future research on robotics pertaining recipe and promote more robust VLA architectures development.", + "bbox": [ + 511, + 636, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3As GLM-4V-9B does not support multiple image inputs, the 0-shot method is employed.", + "bbox": [ + 89, + 875, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11149", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 122, + 187, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 6, 8", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8", + "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi_0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3, 7, 11", + "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3", + "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In arXiv preprint arXiv:2307.15818, 2023. 2, 3, 10", + "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3", + "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 8", + "[8] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 4", + "[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 3" + ], + "bbox": [ + 99, + 148, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 28", + "[11] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.8", + "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 4", + "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[14] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3", + "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3", + "[16] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Grasynet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4", + "[17] Hao-Shu Fang, Minghao Gou, Chenxi Wang, and Cewu Lu. Robust grasping across diverse sensor qualities: The graspnet-1billion dataset. The International Journal of Robotics Research, 2023. 2, 4", + "[18] Jensen Gao, Suneel Belkhale, Sudeep Dasari, Ashwin Balakrishna, Dhruv Shah, and Dorsa Sadigh. A taxonomy for evaluating generalist robot policies. arXiv preprint arXiv:2503.01238, 2025. 10", + "[19] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 8", + "[20] Ran Gong, Jiangyong Huang, Yizhou Zhao, Haoran Geng, Xiaofeng Gao, Qingyang Wu, Wensi Ai, Ziheng Zhou, Demetri Terzopoulos, Song-Chun Zhu, et al. Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20483-20495, 2023. 3, 6", + "[21] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for" + ], + "bbox": [ + 517, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11150", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generalizable manipulation skills. In International Conference on Learning Representations, 2023. 3", + "[22] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 8", + "[23] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 3", + "[24] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2, 3, 7", + "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3, 7", + "[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3", + "[27] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2, 3, 4, 5", + "[28] Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846-894, 2011. 2", + "[29] Sertac Karaman, Matthew R Walter, Alejandro Perez, Emilio Frazzoli, and Seth Teller. Anytime motion planning using the rrt. In 2011 IEEE international conference on robotics and automation, pages 1478-1483. IEEE, 2011. 6", + "[30] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 7, 11", + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4", + "[32] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Conference on Robot Learning, pages 80–93. PMLR, 2023. 3, 4", + "[33] Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Erran Li, Ruohan Zhang, et al. Embodied agent interface: Benchmarking llms for embodied decision making. arXiv preprint arXiv:2410.07166, 2024. 8", + "[34] Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu," + ], + "bbox": [ + 91, + 92, + 482, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yizhong Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024.3", + "[35] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2, 3, 5, 11", + "[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 3", + "[37] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 3", + "[38] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6, 5", + "[39] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024. 3", + "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3", + "[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.8", + "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3", + "[43] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 3, 7, 11", + "[44] Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023. 4, 11", + "[45] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 2, 3", + "[46] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3", + "[47] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar," + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems, 2024. 3, 6, 4, 10, 11", + "[48] Arne Nordmann, Nico Hochgeschwender, and Sebastian Wrede. A survey on domain-specific languages in robotics. In International conference on simulation, modeling, and programming for autonomous robots, pages 195-206. Springer, 2014. 5", + "[49] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3", + "[50] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2, 3", + "[51] Tim Pearce, Tabish Rashid, Anssi Kanervisto, Dave Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan, Ida Momennejad, Katja Hofmann, et al. Imitating human behaviour with diffusion models. arXiv preprint arXiv:2301.10677, 2023. 3", + "[52] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2, 3", + "[53] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, Xingjian Zhao, Ke Chen, Yining Zheng, Zhejian Zhou, Ruixiao Li, Jun Zhan, Yunhua Zhou, Linyang Li, Xiaogui Yang, Lingling Wu, Zhangyue Yin, Xuanjing Huang, Yu-Gang Jiang, and Xipeng Qiu. Moss: An open conversational large language model. Machine Intelligence Research, 2024. 3", + "[54] Craig J Sutherland and Bruce MacDonald. Robolang: a simple domain specific language to script robot interactions. In 2019 16th International Conference on Ubiquitous Robots (UR), pages 265-270. IEEE, 2019. 5", + "[55] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in neural information processing systems, 34:251-266, 2021. 2, 3", + "[56] Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, Xander Hinrichsen, Xiaodi Yuan, Chen Bao, Xinsong Lin, Yulin Liu, Tse kai Chan, Yuan Gao, Xuanlin Li, Tongzhou Mu, Nan Xiao, Arnav Gurha, Zhiao Huang, Roberto Calandra, Rui Chen, Shan Luo, and Hao Su. Maniskill3: GPU parallelized robotics simulation and rendering for generalizable embodied ai. arXiv preprint arXiv:2410.00425, 2024. 3", + "[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo:" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 7", + "[58] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 3", + "[59] Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom Erez, Timothy Lillicrap, Nicolas Heess, and Yuval Tassa. dm_control: Software and tasks for continuous control. Software Impacts, 6:100022, 2020. 3", + "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 3", + "[61] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. Survey Certification. 2", + "[62] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 3", + "[63] Sriram Yenamandra, Arun Ramachandran, Karmesh Yadav, Austin Wang, Mukul Khanna, Theophile Gervet, Tsung-Yen Yang, Vidhi Jain, Alexander William Clegg, John Turner, et al. Homerobot: Open-vocabulary mobile manipulation. arXiv preprint arXiv:2306.11565, 2023. 2", + "[64] Kevin Zakka. obj2mjcf: Cli for processing composite wavefront obj files for use in mujoco. https://github.com/kevinzakka/obj2mjcf, 2023. Version 1.0.4", + "[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15244-15253, 2023. 3", + "[66] Liming Zheng, Feng Yan, Fanfan Liu, Chengjian Feng, Zhuoliang Kang, and Lin Ma. Robocas: A benchmark for robotic manipulation in complex object arrangement scenarios. arXiv preprint arXiv:2407.06951, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 753 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11152", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_model.json b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28dc46e168df6949095bd490591ce2475d3c3f60 --- /dev/null +++ b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_model.json @@ -0,0 +1,2697 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.131, + 0.88, + 0.177 + ], + "angle": 0, + "content": "VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.203, + 0.806, + 0.257 + ], + "angle": 0, + "content": "Shiduo Zhang\\(^{1}\\), Zhe Xu\\(^{1}\\), Peiju Liu\\(^{1*}\\), Xiaopeng Yu\\(^{1*}\\), Yuan Li\\(^{1}\\), Qinghui Gao\\(^{1}\\), Zhaoye Fei\\(^{1}\\), Zhangyue Yin\\(^{1}\\), Zuxuan Wu\\(^{1}\\), Yu-Gang Jiang\\(^{1}\\), Xipeng Qiu\\(^{1\\dagger}\\) \n\\(^{1}\\)Fudan University" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.271, + 0.252, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.272, + 0.418, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.387, + 0.252, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.388, + 0.418, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.474, + 0.255, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.473, + 0.416, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.544, + 0.279, + 0.777, + 0.294 + ], + "angle": 0, + "content": "Task-Related Comprehensive Ability" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.303, + 0.651, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.306, + 0.882, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.397, + 0.651, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.397, + 0.882, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.487, + 0.651, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.487, + 0.882, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.595, + 0.907, + 0.625 + ], + "angle": 0, + "content": "Figure 1. Overview of VLABench. VLABench is a large-scale language-conditioned manipulation benchmark to evaluate the comprehensive skill learning and generalization ability of action policies especially pre-trained vision-language-action models." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.636, + 0.327, + 0.652 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.486, + 0.902 + ], + "angle": 0, + "content": "General-purposed embodied agents are designed to understand the users' natural instructions or intentions and act precisely to complete universal tasks. Recently, methods based on foundation models especially Vision-Language-Action models (VLAs) have shown a substantial potential to solve language-conditioned manipulation (LCM) tasks well. However, existing benchmarks do not adequately meet the needs of VLAs and relative algorithms. To better define such general-purpose tasks in the context of LLMs and advance the research in VLAs, we present VLABench, an open-source benchmark for evaluating universal LCM task learning. VLABench provides 100 carefully designed categories of tasks, with strong randomization in each category of task and a total of \\(2000+\\) objects. VLABench stands out from previous benchmarks in four key aspects: 1) tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.909, + 0.85 + ], + "angle": 0, + "content": "requiring world knowledge and common sense transfer, 2) natural language instructions with implicit human intentions rather than templates, 3) long-horizon tasks demanding multi-step reasoning, and 4) evaluation of both action policies and language model capabilities. The benchmark assesses multiple competencies including understanding of mesh&texture, spatial relationship, semantic instruction, physical laws, knowledge transfer and reasoning, etc. To support the downstream finetuning, we provide high-quality training data collected via an automated framework incorporating heuristic skills and prior information. The experimental results indicate that both the current state-of-the-art pretrained VLAs and the workflow based on VLMs face challenges in our tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1Codes and more videos are available at https://vlabench.github.io/ \n2Corresponding to: sdzhang23@m.fudan.edu.cn, xpqiu@fudan.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11142" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.401 + ], + "angle": 0, + "content": "Language-conditioned manipulation represents a fundamental challenge in embodied AI and a stepping stone toward Artificial General Intelligence [1, 5, 14]. Such tasks require agents to master multiple capabilities: interpreting natural language instructions, understanding complex environments, making decisions, formulating plans, and executing precise actions. The rapid advancement of Large Language Models (LLMs) and Vision-Language Models (VLMs) [1, 15] has revolutionized the field with their impressive general abilities in semantic understanding, coding, planning, and reasoning. The strong generalization capabilities has inspired two main approaches in language-conditioned manipulation: pre-training vision-language-action models using large-scale robotics data, as demonstrated by RT-2 and Palm-E [5, 14, 50], and integrating foundation models into agent workflows, like VoxPoser and Copa [24, 25], which combine LLM/VLM outputs with grasp prediction [16, 17] and motion planning algorithms [28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.485, + 0.78 + ], + "angle": 0, + "content": "While real-world robotics experiments provide valuable insights, their complexity and environmental variability often challenge reproducibility. Simulation-based evaluation has emerged as a fair and practical alternative. Existing benchmarks like RLBench, Calvin, and LIBERO [27, 38, 45] offer diverse task sets but fall short in addressing the unique requirements of foundation model-based methods. Tasks designed to align with the capabilities of foundation model-based algorithms should encompass nuanced semantic understanding of user intent, the integration of common-sense knowledge, and a robust ability to interpret diverse visual scenes, as well as require sophisticated multi-step reasoning. Such tasks demand a sophisticated integration of multimodal understanding to effectively interpret and respond to complex, real-world contexts. For example, one of the tasks in RT-2 [5] is \"move the Coke can to Taylor Swift\", while another task in CoPA [24] is \"make me a cup of pour-over coffee\". The first task challenges the robot to use common sense to identify Taylor, a capability of knowledge transfer that previous policies struggled to achieve. The second task further intensifies the difficulty, requiring the robot to decompose the task into subtasks and execute the steps to operate a coffee machine—a long-horizon challenge that has previously been difficult for a single policy to accomplish." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To better define the types of language-conditioned manipulation tasks suited for foundation models and provide a standardized evaluation suite to advance robotics research, we introduce VLABench. VLABench is an open-source benchmark specifically designed for methods utilizing foundation models. The tasks in VLABench are carefully divided into several dimensions to evaluate models across various aspects, including 1) Mastery of common" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "sense and world knowledge, 2) Understanding of mesh and texture, 3) Comprehension of semantically rich instructions, 4) Spatial understanding, 5) Grasp of physical rules, and 6) Reasoning ability. For benchmarking purposes, VLABench offers 100 task categories with comprehensive evaluations across various methods. With a diverse collection of over 2,000 3D objects and scenes, VLABench creates a wide range of visual contexts and tasks. It enables the assessment of generalization capabilities through learning across multiple skills, providing thorough evaluations spanning visual, linguistic, planning, knowledge transfer, and action dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.908, + 0.452 + ], + "angle": 0, + "content": "To ensure fair comparison and evaluation, we develop an automated data collection framework to construct standardized datasets for each task, supporting model training and fine-tuning. Using this dataset, we conduct extensive experiments to evaluate and benchmark three distinct types of approaches: pre-trained VLA, workflows integrating foundation models, and vision-language models (VLMs). The experimental results indicate that existing VLA methods perform poorly on our tasks and don't exhibit the level of generalization abilities or the \"emergent\" phenomena observed in large models [61]. We summarize contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.456, + 0.907, + 0.531 + ], + "angle": 0, + "content": "- We propose VLABench, the first benchmark designed to comprehensively evaluate the capabilities of VLAs and VLMs in robotics manipulation tasks, covering multiple dimensions such as skills, vision, language, task execution, common sense, and reasoning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.532, + 0.906, + 0.621 + ], + "angle": 0, + "content": "- We define 100 novel LCM tasks tailored to the capabilities of foundation models within a standardized evaluation framework. These tasks require a deep understanding of semantics, vision, space, and physical laws, as well as the ability to plan long-horizon tasks and transfer world knowledge and common sense into task execution." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.622, + 0.906, + 0.681 + ], + "angle": 0, + "content": "- We provide a scalable data construction framework and a standardized evaluation dataset. This automated data construction approach facilitates future research on pretraining robotics data." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.683, + 0.906, + 0.741 + ], + "angle": 0, + "content": "- Our experiments demonstrate that current pre-trained VLAs have yet to exhibit the strong generalization capabilities observed in LLMs, and existing SOTA VLMs also show limitations in embodied scenarios." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.456, + 0.907, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.665, + 0.77 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Benchmarks and Datasets. Numerous benchmarks such as RLBench and LIBERO [27, 35, 38, 45, 66] have been proposed to evaluate language-conditioned manipulation policies. A comparison of these benchmarks is provided in Table 1. However, most of these focus on skill learning and fail to sufficiently address long-horizon planning capabilities. Meanwhile, some benchmarks [52, 55, 63] address room-scale mobile manipulation tasks that require" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11143" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.236 + ], + "angle": 0, + "content": "
BenchmarksSemLangLogiReasonKnowledgeDLDRN-taskCate-objN-objAI-GenMultiCamPCDCross EmbAuto Traj
Alfred[52]XXX-X7-3578XXX✓X
Rlbench[27]XXX28.86✓10028168X✓XX✓
Calvin[45]XXX40.00X34530X✓XXX
ManiSkill[21, 46, 56]XXX-✓201002600X✓✓✓✓
LIBERO[38]XXX17.96✓1305175XXXXX
RoboCASA[47]XXX-✓1001532509✓✓X✓✓
ARNOLD[20]XXX--8-40✓✓✓X✓
SimplerEnv[35]XXX27.66✓8-17✓✓✓X✓
Behavior-1K[32]XXX--100022119331XXX✓X
Habitat 2.0[55]XXX--346169XXXXX
VLABench✓✓✓75.96✓1001632164✓✓✓✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.242, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Table 1. Comparison of popular benchmarks in robot learning. SemLang: semantically non-template rich language instructions. LogiReason: task logic and relevant information reasoning. Knowledge: tasks require the application of common sense and world knowledge. DL: quantitative task difficulty level. DR: task domain randomization. N-task: total number of tasks. Cate-object: categories of assets used in the simulation. N-object: total number of objects in the asset library. AI-Gen: use of generative AI models for asset library creation. MultiCam: use of multiple cameras. PCD: support point cloud data in 3D methods. Cross Emb: support for cross-embodiment. Auto Traj: supporting automated data collection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.482, + 0.627 + ], + "angle": 0, + "content": "long memory and reasoning. Yet, these interactions typically occur through interfaces, rather than direct physical manipulation, limiting the transferability of learned policies to real-world scenarios. Additionally, while efforts [21, 32, 38, 47] have made strides in task format, difficulty, and scale, these benchmarks have largely overlooked the guiding role of language in tasks, often relying on template instructions. VLABench is the first to introduce features such as natural human interaction, implicit goal-oriented semantics, and requirements based on common sense into robot manipulation tasks, as shown in Figure 20. In terms of generalization evaluation, previous works [27, 38, 45] typically assess models at the instance level within the same category, which limits their ability to evaluate generalization across diverse object categories or different tasks within the same skill set. In contrast, VLABench is the first benchmark to evaluate generalization capabilities across a wide range of tasks, object types, and task categories, providing a more comprehensive assessment of model versatility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Large-scale datasets have been built in both real and simulation [6, 47, 50, 60] for large-scale imitation learning. However, real-world data faces challenges related to scalability, making it difficult to gather sufficient data at scale [4]. Simulated datasets, while more scalable, often suffer from limited diversity in scenarios and tasks [27, 45], and still require teleoperation [20, 38] for data collection. VLABench addresses these limitations by offering a broader range of tasks that are more closely aligned with real-world conditions, covering diverse aspects of vision, language, tasks, and skills. Furthermore, it introduces an efficient and robust process for the automated generation of simulated data, significantly enhancing task diversity and scalability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Pretrained Vision-Language-Action Models. The recent rise of multimodal models [1, 13, 40, 65] and the organization of operational datasets [49, 60], have led to the integration of vision-language-action models (VLAs) [4, 5, 14, 30] into language-conditioned manipulation tasks. While the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.566 + ], + "angle": 0, + "content": "term VLA generally refers to models that combine visual and language inputs for policy learning, we focus specifically on approaches leveraging pre-trained models. Several works [4, 14, 30] have applied further training to pre-trained vision-language models (VLMs) for language-conditioned manipulation. These models demonstrate impressive generalization to unseen objects and tasks, yet their control precision is somewhat limited by the discretization of actions [51]. To address this limitation, some approaches have explored using diffusion models [9, 51] as policy networks or using diffusion decoders [3, 34, 62]. Pre-trained models based on diffusion models [37, 43] have shown promising advancements in improving continuous space distribution learning. VLABench includes a selection of these representative methods for comprehensive evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Framework Utilizing Foundation Models. Pre-trained language models [6, 15, 53] and vision-language models [1, 42] have demonstrated strong generalization and versatility. Some researchers [24, 25, 39] combine the general perception and cognitive abilities of these pre-trained models with traditional planning and control algorithms to create agent workflows. These frameworks allow robots to perform complex zero-shot manipulation tasks without requiring additional training. To harness the capabilities of foundation models for manipulation, some works [25, 26, 36] utilize the code comprehension and generation abilities of large language models alongside motion planning optimization algorithms to tackle fundamental manipulation tasks. Additionally, some methods [23, 24] leverage large models to decompose long-horizon tasks into subtasks, then integrate perception and trajectory generation modules to construct the entire manipulation pipeline. However, most zero-shot methods of this kind heavily rely on prompt design [25], the accuracy of each module, and even the specific parameters of the models being invoked [24]. While these methods demonstrate strong generalization, they often face challenges with accuracy. VLABench provides a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11144" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "zero-shot evaluation framework to assess the performance of such workflows and offers insights into their effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.155, + 0.212, + 0.17 + ], + "angle": 0, + "content": "3. VLABench" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.181, + 0.258, + 0.197 + ], + "angle": 0, + "content": "3.1. Task Description" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.485, + 0.461 + ], + "angle": 0, + "content": "VLABench is composed of 60 primitive and 40 composite tasks, categorized by task difficulty levels. These tasks are designed to encompass a rich variety of skills while covering ample visual and language semantic information. For skill learning, 100 tasks in VLABench cover a wide range including 1) Pick&place, 2) Open&close door, 3) Open&close drawer, 4) Hang objects on the wall, 5) Use tool e.g. Hammer nail, 6) Press button, 7) Insert, 8) Pour, 9) Twist, and 10) Explore. In addition, VLABench places greater emphasis on real-life scenarios and essential daily tasks, representing more interactive language instructions, a wider variety of task settings, the integration of common sense and societal knowledge, and long-horizon tasks requiring logical planning, as shown in Figure 2. Notably, VLABench adopts a stricter definition of task generalization, which will be elaborated on in the Section 3.3. The whole task list can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.508 + ], + "angle": 0, + "content": "Primitive Tasks. Primitive tasks are divided into five dimensions, each corresponding to the assessment of a specific ability dimension." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.509, + 0.484, + 0.628 + ], + "angle": 0, + "content": "- Mesh&Texture Understanding. This type of task requires the model to recognize different meshes and understand various texture features. Take the SelectToy task shown in Figure 2 (a) as an example, the robot is directly required to place a specific toy e.g. Aquaman into a receptacle. The model must possess strong visual capabilities to accurately recognize such complex meshes and textures." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.63, + 0.484, + 0.75 + ], + "angle": 0, + "content": "- Spatial Understanding. Spatial understanding tasks involve various spatial relationships, such as the \\( n \\)-th left/right position, inside/outside of a receptacle, the \\( m \\)-th row and \\( n \\)-th column, near/far, and beside a specific object, representing relative positional relationships. Figure 2 (b) shows one case in task in PullBook. Such a complex relative positional relationship imposes extremely high demands on the model's multimodal understanding [8]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.751, + 0.484, + 0.87 + ], + "angle": 0, + "content": "- Common Sense & World Knowledge. Tasks relative to common sense/world knowledge require the agent to transfer the knowledge gained in the pre-train stage to solve the problem. The task shown in Figure 2 (c) requires the agent not only to recognize different types of flowers from visual information but also to leverage world knowledge to determine that \"the tulip is the national flower of the Netherlands\"." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Semantic Understanding. This type of task emphasizes the complexity, subtlety, and natural interactivity of lan" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.509, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.907, + 0.168 + ], + "angle": 0, + "content": "guage instructions. Task objectives are often implicitly conveyed through a natural conversation. To perform well in GetDrink task as Figure 2 (d) shown, the agent must capture the implied request from a lengthy instruction: to take out a chilled cola from the refrigerator." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.907, + 0.287 + ], + "angle": 0, + "content": "- Physical Law. This type of task expects the robot to integrate visual information and take correct actions based on physical principles and real-time observation. In UseSeesaw task in Figure 2 (e), the robot is commanded to grasp an object that can not be achieved directly. The agent must recognize the need to apply the principle of leverage by using sufficient weight to lift the target object on the other end." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.909, + 0.561 + ], + "angle": 0, + "content": "Composite Tasks. Composite tasks in VLABench involve the combination of multiple skills, long-term task planning, and multi-step logical reasoning from instructions, scenes, and even game rules. Figure 2 (f) showcases a variety of challenging complex tasks. Composite tasks have a significantly longer trajectory horizon, with an average episode length exceeding 500 timesteps—considerably more than the average of 120 timesteps for primitive tasks. In Figure 2 (f1), the agent must not only correctly identify all poker cards from visual information and use world knowledge of poker rules to select the best hand, but also flip face-down cards to acquire complete information. This type of task, requiring the agent to consciously fulfill prerequisite conditions, has never been modeled before. Composite tasks also require extracting the user's implicit needs from natural dialogue. The case in Figure 2 (f2) needs the agent to place the Python textbook on the table and open the laptop, without direct instruction. More cases are shown in Figure 20." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.71, + 0.587 + ], + "angle": 0, + "content": "3.2. Task Difficulty Level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.777 + ], + "angle": 0, + "content": "VLABench stands out among language-conditioned manipulation benchmarks not only for its multi-dimensional task design and evaluation but also for its hierarchical difficulty levels. The difficulty level (DL) is quantified using three metrics: task variants [27, 44], task sub-steps, and reasoning steps. These metrics quantify the overall task difficulty level from the perspectives of generalization difficulty, normalized episode horizon length, and intelligence. The detailed computation of DL is provided in Section 6.3. As shown in Table 1, VLABench's difficulty level surpasses all previous benchmarks and introduces new challenges that a VLA model should overcome." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.642, + 0.802 + ], + "angle": 0, + "content": "3.3.Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Evaluation. VLABench organizes evaluations into three categories: assessments of pretrained or fine-tuned vision-language-action (VLA) models, heuristic workflows that integrate foundation models with various algorithms, and multi-dimensional evaluations of vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- Generalization Ability of VLAs. For pre-trained VLA" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11145" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.093, + 0.255, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.251, + 0.222, + 0.26 + ], + "angle": 0, + "content": "(a) Mesh & Texture" + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.094, + 0.415, + 0.124 + ], + "angle": 0, + "content": "Instruction: \"Take out the second book from the right on the top layer of the shelf.\"" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.126, + 0.416, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.251, + 0.392, + 0.261 + ], + "angle": 0, + "content": "(b) Spatial Understanding" + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.096, + 0.572, + 0.116 + ], + "angle": 0, + "content": "Instruction: \"Insert the national flower of the Netherlands into the vase.\"" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.126, + 0.578, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.251, + 0.583, + 0.261 + ], + "angle": 0, + "content": "(c) Common Sense & World Knowledge" + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.094, + 0.734, + 0.125 + ], + "angle": 0, + "content": "Instruction: \"Ah, I just spent an hour working out at the gym. I'm so thirsty now. I really go for an ice-cold soda!\"" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.126, + 0.739, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.25, + 0.721, + 0.26 + ], + "angle": 0, + "content": "(d) Semantic Understanding" + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.096, + 0.884, + 0.116 + ], + "angle": 0, + "content": "Instruction: \"Get the fruit from the container\"" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.125, + 0.902, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.25, + 0.857, + 0.259 + ], + "angle": 0, + "content": "(e) Physical Law" + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.264, + 0.255, + 0.296 + ], + "angle": 0, + "content": "Instruction:\"We are playing Texas Hold'em game! Choose your largest combination on the table.\"" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.297, + 0.255, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.421, + 0.219, + 0.43 + ], + "angle": 0, + "content": "(f1) Play Texas Hold'em" + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.265, + 0.414, + 0.296 + ], + "angle": 0, + "content": "Instruction: \"Shortly, I'm going to work on my Python course assignment. Help me set up my desk.\"" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.297, + 0.416, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.421, + 0.384, + 0.43 + ], + "angle": 0, + "content": "(f2) Manage Study Table" + }, + { + "type": "image_caption", + "bbox": [ + 0.421, + 0.265, + 0.574, + 0.296 + ], + "angle": 0, + "content": "instruction:\"I'd like to make a garlic eggplant dish.Prepare the ingredients I will use,thanks!" + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.297, + 0.578, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.421, + 0.548, + 0.429 + ], + "angle": 0, + "content": "(f3) Cook Dishes as Menu" + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.432, + 0.542, + 0.442 + ], + "angle": 0, + "content": "(f) Type: Composite" + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.265, + 0.738, + 0.287 + ], + "angle": 0, + "content": "Instruction: \"Could you please show the acid base neutralization experiment?\"" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.296, + 0.74, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.421, + 0.721, + 0.43 + ], + "angle": 0, + "content": "(f4) Take Chemistry Experiment" + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.265, + 0.895, + 0.295 + ], + "angle": 0, + "content": "Instruction: \"I have to go to class this afternoon but I'm so sleepy! I want to have latte.\"" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.297, + 0.902, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.421, + 0.862, + 0.429 + ], + "angle": 0, + "content": "(f5) Get Latte Coffee" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.458, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Figure 2. Task examples in each dimension. The first row showcases examples of primitive tasks from Section 3.1, while the second row presents examples of composite tasks." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.501, + 0.483, + 0.545 + ], + "angle": 0, + "content": "models, the evaluation focuses on the model's generalization ability across multiple dimensions. In the standard evaluation, we have set up six evaluation tracks." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.546, + 0.483, + 0.605 + ], + "angle": 0, + "content": "- Track 1: Rapid downstream task transferability and skill acquisition ability. The evaluation episodes use objects seen in the training set, with minimal differences." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.607, + 0.483, + 0.651 + ], + "angle": 0, + "content": "- Track 2: Cross-category visual generalization. The evaluation episodes use unseen categories of objects as the target entity for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.652, + 0.483, + 0.712 + ], + "angle": 0, + "content": "- Track 3: Common sense application. Building upon Track 1, replace parts of the instructions that directly reference the target object with content requiring common sense reasoning in evaluation episodes." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.712, + 0.483, + 0.757 + ], + "angle": 0, + "content": "- Track 4: Semantically rich instruction adaptation. Building upon Track 1, replace the instructions with unseen and more complex ones." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.758, + 0.483, + 0.802 + ], + "angle": 0, + "content": "- Track 5: Cross-domain behavior transferability. The evaluation tasks are replaced with ones that differ from those in the training set, but require similar actions." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.803, + 0.483, + 0.847 + ], + "angle": 0, + "content": "- Track 6: Long-horizon task learning. Training and evaluation using more challenging composite tasks require additional skills and reasoning." + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.546, + 0.483, + 0.847 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.848, + 0.483, + 0.892 + ], + "angle": 0, + "content": "- Zero-shot Transfer Ability of Heuristic Workflow. Training-free workflow methods are evaluated under sets of tasks in all dimensions mentioned in Section 3.1. We" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.501, + 0.905, + 0.56 + ], + "angle": 0, + "content": "also extend the evaluation to cover various skills and long-horizon tasks to assess the overall capability and execution robustness of the workflow. All agents are evaluated in a zero-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.561, + 0.906, + 0.711 + ], + "angle": 0, + "content": "- Comprehensive Evaluation of VLMs' Capabilities. Similar to heuristic workflows, the evaluation of VLMs is also comprehensive. Since VLMs lack intrinsic action capabilities, we organized a skill library and integrated it into a domain-specific language (DSL) [48, 54], leveraging annotated asset information as prior knowledge. This DSL functions as a straightforward API that VLMs can call, enabling efficient interaction. The evaluation pipeline is shown as Figure 3 and the details will be discussed in Section 9.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.906, + 0.818 + ], + "angle": 0, + "content": "Metric. In addition to the success rate (SR), considering the long-horizon nature and high difficulty level of our tasks, we introduce the intention score (IS) and progress score (PS) for more granular evaluation. Intention score is a useful soft metric measuring the correctness trend in selecting the target object for the operation especially when sub-tasks are not yet completed. The discrete version metric is:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.825, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\nI S = \\left\\{ \\begin{array}{l l} 1, & \\text {i f d i s t a n c e} _ {\\min } < \\epsilon \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(distance_{min}\\) is the minimum distance between the end-effector's position and the target entity, and \\(\\epsilon\\) is empiri" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11146" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.368, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.661, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.089, + 0.906, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.288, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Evaluation pipeline for VLMs. Step 1: Sample the required four-view images, as well as those segmented with numerical information, from the simulation. Meanwhile, save the corresponding instructions and operation sequences. Step 2: Input the original and annotated images and the instructions into the model. Then, obtain the model's output and extract the generated operation sequence. Step 3: The two operation sequences are evaluated using four metrics, which are weighted and summed to produce the final score. The red part represents the error output of the model, where the red solid arrows represent the dependencies generated by the error, and red dashed arrows represent the dependencies lost by the error." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.446 + ], + "angle": 0, + "content": "cally set to 0.15 meters in most cases after experimentation. The progress score refers to the completion level of subtasks in a long-horizon task and serves as a softer process supervision metric compared to the success rate." + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.46, + 0.483, + 0.475 + ], + "angle": 0, + "content": "\\[\nP S = N _ {\\text {d o n e}} / N _ {\\text {s u b}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.489, + 0.484, + 0.519 + ], + "angle": 0, + "content": "where \\( N_{done} \\) is the number of completed sub-tasks, and \\( N_{sub} \\) is the total number of task sub-tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.292, + 0.544 + ], + "angle": 0, + "content": "3.4. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Trajectory Generation. As human teleoperation is time-consuming and not scalable [38, 47], we developed an efficient, scalable automated data collection pipeline based on our custom skill library. Inspired by [20], our data collection framework leverages the prior information including point clouds of the environment, annotated entities' grasppoints, target entity at the current step, etc. The data collection framework includes multiple task-specific motion planners. These motion planners call upon the skills in the skill library based on the current task progress and determine parameters by incorporating prior information. Subsequently, the selected skills generate trajectories using RRT [29], with quaternion interpolation achieved through spherical linear interpolation. The final trajectory is smoothed using a Bezier curve to optimize path quality. To enhance sample efficiency, reject sampling and failure-triggered early termination are applied. We demonstrate the superiority of our data framework through comparative experiments. Detailed analysis can be found in Section 8.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data Diversity and Quality. During the data construction process, we introduced diverse task variants and domain randomization across different episodes of the same task to ensure the diversity of the training data, as discussed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.907, + 0.459 + ], + "angle": 0, + "content": "in Sections 6.1 and 6.4. Following the approach of previous benchmarks built on Mujoco [38, 47], the dataset is stored in the same format, with similar visual rendering quality and trajectory accuracy. We further analyze the data quality in Section 8.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Instruction Augmentation. We use GPT-4o [1] to generate descriptions that incorporate target-specific characteristics and interactive instructions that encompass a variety of contexts and intentions. Section 8.4 provides more details of the instruction generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.552, + 0.645, + 0.569 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.654 + ], + "angle": 0, + "content": "Following Section 3.3, we conducted experiments centered on pre-trained VLA models, workflows incorporating multiple algorithmic modules, and various VLMs. The remainder of this section provides a detailed description of the experimental setup." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.791, + 0.68 + ], + "angle": 0, + "content": "4.1. Generalization Ability of VLAs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Pretrained VLAs are expected to possess robust generalization and versatility similar to LLMs. Experiments about are set to address the following research questions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.733, + 0.905, + 0.762 + ], + "angle": 0, + "content": "Q1: Do pre-trained VLAs exhibit stronger general abilities with unseen categories of objects?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.763, + 0.905, + 0.793 + ], + "angle": 0, + "content": "Q2: Can pre-trained VLAs transfer their general knowledge and behavioral abilities to similar but unseen tasks?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.794, + 0.905, + 0.824 + ], + "angle": 0, + "content": "Q3: Can pre-trained VLAs understand natural user interactions and implicit goal requirements?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.905, + 0.854 + ], + "angle": 0, + "content": "Q4: Do pre-trained VLAs have the potential to transfer their world knowledge to related tasks?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.855, + 0.905, + 0.885 + ], + "angle": 0, + "content": "Q5: Can existing VLA architectures accurately support the completion of long-horizon tasks?" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.733, + 0.905, + 0.885 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Experiment Setup. To investigate the questions outlined" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11147" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.904, + 0.182 + ], + "angle": 0, + "content": "
ModelTrack 1Track 2Track 3Track 4Track 5Track 6
ISPSSRISPSSRISPSSRISPSSRISPSSRISPSSR
Octo[57]32.05.62.020.42.00.022.02.81.226.84.01.213.60.00.017.23.00.0
OpenVLA[30]50.013.26.435.67.03.238.08.03.238.09.22.816.00.00.028.89.60.4
RDT-1B[43]50.014.36.039.28.23.236.07.24.035.68.04.420.40.00.021.69.20.0
π0-Base[3]59.629.313.249.621.46.060.420.05.260.419.33.622.00.00.034.412.82.0
π0-Fast[3]59.925.59.247.616.52.450.015.43.652.017.26.825.60.00.034.014.41.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.192, + 0.899, + 0.206 + ], + "angle": 0, + "content": "Table 2. Overall experiment result of 6 evaluation tracks of fine-tuned VLAs. The detailed result of each task is reported in Table 9." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.482, + 0.431 + ], + "angle": 0, + "content": "above, we fine-tuned various pre-trained VLA architectures [3, 30, 43, 57] on our high-quality dataset, and evaluated them on different tracks mentioned in Section 3.3 to assess their performance across a range of tasks and conditions. Considering the high task difficulty, we selected five primitive tasks for experiments in tracks 1 to 5, with 200 data samples per task, and trained the models following a multitask training approach, following the setting in previous work [3, 43]. We then assessed the models across 50 episodes for each task, resulting in a total of 250 trials per track. For track 6, we selected five composite tasks, keeping the other settings consistent. Section 9.1 provides a complete list of tasks in different tracks and details of the model fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.482, + 0.522 + ], + "angle": 0, + "content": "Result and Analysis. Table 2 presents the evaluation results across multiple tracks. These models perform poorly on most of tasks, as they are trained on a more diverse and limited set of fine-tuning data, with a higher difficulty level. Through the experimental results of each track and comparisons across different tracks, our analysis is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.524, + 0.482, + 0.645 + ], + "angle": 0, + "content": "- The primary cause of poor performance is the insufficient ability for rapid transfer in downstream tasks. Episodes in Tracks 1 and 6 closely resemble the training set, primarily assessing the model's basic fitting ability. Despite this, the models still perform poorly on certain tasks. We believe the main reason for the failure of most generalization experiments is the inaccuracy in basic action generation, especially Octo." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.645, + 0.482, + 0.75 + ], + "angle": 0, + "content": "- Pre-trained VLA models exhibit some ability to handle unseen object categories. By comparing the results of Track 1 and Track 2, we observe that although all models experience a performance decline in Track 2, they are still able to complete a small subset of tasks. This is likely due to the benefits of large-scale robotics data pretraining or even earlier VLM pretraining." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.751, + 0.482, + 0.871 + ], + "angle": 0, + "content": "- Performance decline with unseen semantic language. By combining the experimental results of Track 3 and Track 4, we find that the model's language understanding ability has deteriorated by a larger margin. Among these, the performance decline is more pronounced in models based on non-VLM architectures: RDT's progress score decreases by \\(45.2\\%\\) while OpenVLA and \\(\\pi_0\\) only decreases by \\(31.3\\%\\) and \\(34.2\\%\\), respectively, in Track 4." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- Nearly absent cross-domain transferability. We found that all the models fail to perform similar but unseen tasks," + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.524, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.221, + 0.905, + 0.265 + ], + "angle": 0, + "content": "due to significant visual and linguistic input differences. These policies exhibit uncertainty in behavior when faced with unfamiliar tasks, resulting in IS that are similar." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.905, + 0.373 + ], + "angle": 0, + "content": "We further analyze that data scaling and pre-training contribute positively to the model's performance Through basic ablation experiments. We also discuss in detail the potential issues with current VLAs, such as multimodal data co-training and model architecture designs. These discussions are presented in Section 10.1. We also discuss the gap between real and simulated environments in Section 10.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.39, + 0.753, + 0.405 + ], + "angle": 0, + "content": "4.2. Zero-shot Ability of Agent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.905, + 0.593 + ], + "angle": 0, + "content": "For our evaluation of foundation model-based algorithms, we reviewed two state-of-the-art frameworks, Voxposer [25] and CoPA [24], and the comparison results are shown in Figure 4. Given Voxposer's dependence on large language models (LLMs), we assessed Voxposer's performance with and without visual perception capabilities. While Voxposer performed adequately on basic tasks and achieved the Progress Scores of 30-40, its reliance on LLM-driven motion planning often led to grasping failures due to limited information for effective grasp planning, especially when interpreting rotation in non-visual contexts, resulting in low overall scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.905, + 0.702 + ], + "angle": 0, + "content": "Interestingly, the foundational LLM alone maintained relatively stable scores in semantic understanding and reasoning tasks without visual input. However, adding visual perception slightly reduced performance in these areas while significantly improving spatial reasoning, where LLM-only setups struggle with spatial accuracy due to lack of spatial information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.9 + ], + "angle": 0, + "content": "The lack of closed-loop feedback limits these models' ability to perform physical reasoning tasks, particularly those involving dynamic interactions, leading to lower scores in this dimension. Both models struggle with high-complexity tasks, succeeding mainly in entity recognition but performing poorly in long horizon task reasonable breaking down. This finding underscores the requirements for advancing foundational model-based frameworks to address complex reasoning. In fact, although the methods mentioned above emphasize their zero-shot capabilities and generalization to new scenarios, their modular design often limits the upper bound of their performance. A more detailed discussion of this can be found in Section 10.3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.482, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.359, + 0.488, + 0.416 + ], + "angle": 0, + "content": "Figure 4. Evaluation progress score for Voxposer and CoPA. Voxposer w/o refers to the version without visual perception, where ground truth labels are directly provided for object selection. Voxposer w uses GPT-4V as the visual perception module." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.428, + 0.379, + 0.444 + ], + "angle": 0, + "content": "4.3. Comprehensive Ability of VLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.587 + ], + "angle": 0, + "content": "We referred to the evaluation results of multiple series of VLMs provided by OpenCompass [11] and selected models from different families [1, 2, 7, 19, 22, 41] with strong overall performance as shown in Figure 5. We evaluate the comprehensive performance of these models with the dataset derived from naturally self-contained information within a simulated environment. This dataset consists of a complex set of tasks designed to assess the VLM's ability to perceive visual stimuli and comprehend verbal instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.649 + ], + "angle": 0, + "content": "In line with the evaluation framework outlined in the Section 3.3, we evaluate the models across six dimensions of task performance. The results for each VLM model under the 1-shot setting are summarized in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Although these VLMs perform well on most multimodal tasks and even some embodied tasks [11, 33], their performance, including that of GPT-4o, falls short when faced with more complex scenarios, instructions, and more challenging tasks. We surprisingly find that the open-source model Qwen2-VL-7B-Instruct performed competitively, surpassing GPT-4o on certain dimensions. However, all models struggled with complex tasks, especially those requiring long-term task decomposition and logic reasoning. Only GPT-4o achieved a score in the reasoning dimension comparable to those in other dimensions, while the other models scored around 20. Besides, performance declines significantly when linguistic instructions transition from direct semantics to abstract meanings, as shown in the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.902, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.359, + 0.908, + 0.415 + ], + "angle": 0, + "content": "Figure 5. Radar charts depicting the performance of all VLM models across six dimensions. The reason why only GLM-4V-9B is evaluated in a zero-shot setting is that it does not support multigraph inference, which is required for the other models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.595 + ], + "angle": 0, + "content": "semantic dimension. Different models appear to have distinct areas of expertise, e.g. LLaVA-NeXT exhibits weaker spatial perception ability, GLM-4V-9B excels in spatial and even physical law dimensions but lag in semantic comprehension. In Section 10.4, we conducted additional ablation experiments focusing on CoT and few-shot learning, followed by further analysis. Detailed data and error analysis can be found in Table 10. Overall, while the models demonstrate promising capabilities, their understanding and planning in embodied environments remain limited, highlighting the need for further advancements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.634, + 0.628 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose VLABench, a large-scale benchmark designed for tasks with long-horizon and multi-dimensional reasoning. Such reasoning involves many dimensions, including implicit semantic goal extracting, logical reasoning, and long-horizon planning. The most important thing we do is to provide a positive definition of the capabilities that intelligent agents with cognitive abilities should possess and the tasks they should be able to perform, through the provision of 100 standardized tasks. Additionally, VLABench constructed a scalable automated data collection framework for future's potential larger scale pertaining and a standardized dataset for fair comparison of VLAs, both in the present and in future developments. Our diverse and multiple experiments revealed that current VLAs and VLMs face significant challenges in our tasks, and there remain substantial uncertainties in research on robotics scaling. We hope that VLABench will inspire both the future research on robotics pertaining recipe and promote more robust VLA architectures development." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.486, + 0.902 + ], + "angle": 0, + "content": "3As GLM-4V-9B does not support multiple image inputs, the 0-shot method is employed." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11149" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.123, + 0.188, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.279, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi_0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In arXiv preprint arXiv:2307.15818, 2023. 2, 3, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[8] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[10] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[11] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[14] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[16] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Grasynet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[17] Hao-Shu Fang, Minghao Gou, Chenxi Wang, and Cewu Lu. Robust grasping across diverse sensor qualities: The graspnet-1billion dataset. The International Journal of Robotics Research, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[18] Jensen Gao, Suneel Belkhale, Sudeep Dasari, Ashwin Balakrishna, Dhruv Shah, and Dorsa Sadigh. A taxonomy for evaluating generalist robot policies. arXiv preprint arXiv:2503.01238, 2025. 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.677, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[19] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[20] Ran Gong, Jiangyong Huang, Yizhou Zhao, Haoran Geng, Xiaofeng Gao, Qingyang Wu, Wensi Ai, Ziheng Zhou, Demetri Terzopoulos, Song-Chun Zhu, et al. Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20483-20495, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[21] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11150" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "generalizable manipulation skills. In International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[22] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[23] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[24] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[27] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[28] Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846-894, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[29] Sertac Karaman, Matthew R Walter, Alejandro Perez, Emilio Frazzoli, and Seth Teller. Anytime motion planning using the rrt. In 2011 IEEE international conference on robotics and automation, pages 1478-1483. IEEE, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[30] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[32] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Conference on Robot Learning, pages 80–93. PMLR, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[33] Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Erran Li, Ruohan Zhang, et al. Embodied agent interface: Benchmarking llms for embodied decision making. arXiv preprint arXiv:2410.07166, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[34] Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Yizhong Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[35] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2, 3, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[37] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[38] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[39] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[43] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 3, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[44] Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023. 4, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[45] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[46] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[47] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems, 2024. 3, 6, 4, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[48] Arne Nordmann, Nico Hochgeschwender, and Sebastian Wrede. A survey on domain-specific languages in robotics. In International conference on simulation, modeling, and programming for autonomous robots, pages 195-206. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[49] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[50] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[51] Tim Pearce, Tabish Rashid, Anssi Kanervisto, Dave Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan, Ida Momennejad, Katja Hofmann, et al. Imitating human behaviour with diffusion models. arXiv preprint arXiv:2301.10677, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.484, + 0.504 + ], + "angle": 0, + "content": "[52] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[53] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, Xingjian Zhao, Ke Chen, Yining Zheng, Zhejian Zhou, Ruixiao Li, Jun Zhan, Yunhua Zhou, Linyang Li, Xiaogui Yang, Lingling Wu, Zhangyue Yin, Xuanjing Huang, Yu-Gang Jiang, and Xipeng Qiu. Moss: An open conversational large language model. Machine Intelligence Research, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[54] Craig J Sutherland and Bruce MacDonald. Robolang: a simple domain specific language to script robot interactions. In 2019 16th International Conference on Ubiquitous Robots (UR), pages 265-270. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[55] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in neural information processing systems, 34:251-266, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[56] Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, Xander Hinrichsen, Xiaodi Yuan, Chen Bao, Xinsong Lin, Yulin Liu, Tse kai Chan, Yuan Gao, Xuanlin Li, Tongzhou Mu, Nan Xiao, Arnav Gurha, Zhiao Huang, Roberto Calandra, Rui Chen, Shan Luo, and Hao Su. Maniskill3: GPU parallelized robotics simulation and rendering for generalizable embodied ai. arXiv preprint arXiv:2410.00425, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[58] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[59] Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom Erez, Timothy Lillicrap, Nicolas Heess, and Yuval Tassa. dm_control: Software and tasks for continuous control. Software Impacts, 6:100022, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.33 + ], + "angle": 0, + "content": "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.906, + 0.43 + ], + "angle": 0, + "content": "[61] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. Survey Certification. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.906, + 0.501 + ], + "angle": 0, + "content": "[62] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[63] Sriram Yenamandra, Arun Ramachandran, Karmesh Yadav, Austin Wang, Mukul Khanna, Theophile Gervet, Tsung-Yen Yang, Vidhi Jain, Alexander William Clegg, John Turner, et al. Homerobot: Open-vocabulary mobile manipulation. arXiv preprint arXiv:2306.11565, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.906, + 0.613 + ], + "angle": 0, + "content": "[64] Kevin Zakka. obj2mjcf: Cli for processing composite wavefront obj files for use in mujoco. https://github.com/kevinzakka/obj2mjcf, 2023. Version 1.0.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15244-15253, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.906, + 0.755 + ], + "angle": 0, + "content": "[66] Liming Zheng, Feng Yan, Fanfan Liu, Chengjian Feng, Zhuoliang Kang, and Lin Ma. Robocas: A benchmark for robotic manipulation in complex object arrangement scenarios. arXiv preprint arXiv:2407.06951, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11152" + } + ] +] \ No newline at end of file diff --git a/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_origin.pdf b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d95f0c73dc321326a1cb6e7b44b86089b72ffb18 --- /dev/null +++ b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/bbac0723-47a6-4c9f-bec1-1f18e03a36ac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:213c6465d797c1114ba3916b46dff4f95d4bc34a060796dc2fa15f831d6aac4e +size 3125941 diff --git a/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/full.md b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..79f1be42667d4302cc020d73c1e6bd2b8afc5d5f --- /dev/null +++ b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/full.md @@ -0,0 +1,330 @@ +# VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks + +Shiduo Zhang $^{1}$ , Zhe Xu $^{1}$ , Peiju Liu $^{1*}$ , Xiaopeng Yu $^{1*}$ , Yuan Li $^{1}$ , Qinghui Gao $^{1}$ , Zhaoye Fei $^{1}$ , Zhangyue Yin $^{1}$ , Zuxuan Wu $^{1}$ , Yu-Gang Jiang $^{1}$ , Xipeng Qiu $^{1\dagger}$ $^{1}$ Fudan University + +![](images/e2d5b9d02a4369267daa42a154de08e0858272d4a89d1449eefb93d77103a1af.jpg) + +![](images/caf0fadd36a258afd4859e02698f1d8efafb2cb934e1808a0914d1aeb705ffde.jpg) + +![](images/def1f31efc16ba96dcf5d6586ffd38658bb1ed315966d70a515284b308743162.jpg) + +![](images/e97073b7b70ac8f2f07a3822284eb064f9c9f5f37b2ea4cc873deac37cde16ac.jpg) + +![](images/746e022f6a6837f28187d6f2271c164963d7425ffc48671deb445672556da165.jpg) +Figure 1. Overview of VLABench. VLABench is a large-scale language-conditioned manipulation benchmark to evaluate the comprehensive skill learning and generalization ability of action policies especially pre-trained vision-language-action models. + +![](images/5d9d6bd8001aa8ec88ee2192a33ad6ac9502628d5b6a3c016723deccc3b35c76.jpg) + +# Task-Related Comprehensive Ability + +![](images/4579bc73e65769ac1650ed26881978228912ce4383f267f8e61644907781627e.jpg) + +![](images/3071cec2797462e47d4a174f32a1306820d137bd8d5a3f1cb4bdf4a76b64beb2.jpg) + +![](images/b262b251d3d39fec7131e9ce6112260537a66252982bd725d5be59653fb259d4.jpg) + +![](images/d28972f469966ffdc8407caba0dd83510eeb5b734e1c9bfa9efc7534a3ade0ee.jpg) + +![](images/fff038d2f224a9d9721e63eaea3f4c556337d0592bde01a0035665fb18fe9748.jpg) + +![](images/12f1307e6cfd524303b97c6cbadbd3823073bceffdedde6d41c3bcf7150f1a6e.jpg) + +# Abstract + +General-purposed embodied agents are designed to understand the users' natural instructions or intentions and act precisely to complete universal tasks. Recently, methods based on foundation models especially Vision-Language-Action models (VLAs) have shown a substantial potential to solve language-conditioned manipulation (LCM) tasks well. However, existing benchmarks do not adequately meet the needs of VLAs and relative algorithms. To better define such general-purpose tasks in the context of LLMs and advance the research in VLAs, we present VLABench, an open-source benchmark for evaluating universal LCM task learning. VLABench provides 100 carefully designed categories of tasks, with strong randomization in each category of task and a total of $2000+$ objects. VLABench stands out from previous benchmarks in four key aspects: 1) tasks + +requiring world knowledge and common sense transfer, 2) natural language instructions with implicit human intentions rather than templates, 3) long-horizon tasks demanding multi-step reasoning, and 4) evaluation of both action policies and language model capabilities. The benchmark assesses multiple competencies including understanding of mesh&texture, spatial relationship, semantic instruction, physical laws, knowledge transfer and reasoning, etc. To support the downstream finetuning, we provide high-quality training data collected via an automated framework incorporating heuristic skills and prior information. The experimental results indicate that both the current state-of-the-art pretrained VLAs and the workflow based on VLMs face challenges in our tasks. + +# 1. Introduction + +Language-conditioned manipulation represents a fundamental challenge in embodied AI and a stepping stone toward Artificial General Intelligence [1, 5, 14]. Such tasks require agents to master multiple capabilities: interpreting natural language instructions, understanding complex environments, making decisions, formulating plans, and executing precise actions. The rapid advancement of Large Language Models (LLMs) and Vision-Language Models (VLMs) [1, 15] has revolutionized the field with their impressive general abilities in semantic understanding, coding, planning, and reasoning. The strong generalization capabilities has inspired two main approaches in language-conditioned manipulation: pre-training vision-language-action models using large-scale robotics data, as demonstrated by RT-2 and Palm-E [5, 14, 50], and integrating foundation models into agent workflows, like VoxPoser and Copa [24, 25], which combine LLM/VLM outputs with grasp prediction [16, 17] and motion planning algorithms [28]. + +While real-world robotics experiments provide valuable insights, their complexity and environmental variability often challenge reproducibility. Simulation-based evaluation has emerged as a fair and practical alternative. Existing benchmarks like RLBench, Calvin, and LIBERO [27, 38, 45] offer diverse task sets but fall short in addressing the unique requirements of foundation model-based methods. Tasks designed to align with the capabilities of foundation model-based algorithms should encompass nuanced semantic understanding of user intent, the integration of common-sense knowledge, and a robust ability to interpret diverse visual scenes, as well as require sophisticated multi-step reasoning. Such tasks demand a sophisticated integration of multimodal understanding to effectively interpret and respond to complex, real-world contexts. For example, one of the tasks in RT-2 [5] is "move the Coke can to Taylor Swift", while another task in CoPA [24] is "make me a cup of pour-over coffee". The first task challenges the robot to use common sense to identify Taylor, a capability of knowledge transfer that previous policies struggled to achieve. The second task further intensifies the difficulty, requiring the robot to decompose the task into subtasks and execute the steps to operate a coffee machine—a long-horizon challenge that has previously been difficult for a single policy to accomplish. + +To better define the types of language-conditioned manipulation tasks suited for foundation models and provide a standardized evaluation suite to advance robotics research, we introduce VLABench. VLABench is an open-source benchmark specifically designed for methods utilizing foundation models. The tasks in VLABench are carefully divided into several dimensions to evaluate models across various aspects, including 1) Mastery of common + +sense and world knowledge, 2) Understanding of mesh and texture, 3) Comprehension of semantically rich instructions, 4) Spatial understanding, 5) Grasp of physical rules, and 6) Reasoning ability. For benchmarking purposes, VLABench offers 100 task categories with comprehensive evaluations across various methods. With a diverse collection of over 2,000 3D objects and scenes, VLABench creates a wide range of visual contexts and tasks. It enables the assessment of generalization capabilities through learning across multiple skills, providing thorough evaluations spanning visual, linguistic, planning, knowledge transfer, and action dimensions. + +To ensure fair comparison and evaluation, we develop an automated data collection framework to construct standardized datasets for each task, supporting model training and fine-tuning. Using this dataset, we conduct extensive experiments to evaluate and benchmark three distinct types of approaches: pre-trained VLA, workflows integrating foundation models, and vision-language models (VLMs). The experimental results indicate that existing VLA methods perform poorly on our tasks and don't exhibit the level of generalization abilities or the "emergent" phenomena observed in large models [61]. We summarize contributions as follows: + +- We propose VLABench, the first benchmark designed to comprehensively evaluate the capabilities of VLAs and VLMs in robotics manipulation tasks, covering multiple dimensions such as skills, vision, language, task execution, common sense, and reasoning. +- We define 100 novel LCM tasks tailored to the capabilities of foundation models within a standardized evaluation framework. These tasks require a deep understanding of semantics, vision, space, and physical laws, as well as the ability to plan long-horizon tasks and transfer world knowledge and common sense into task execution. +- We provide a scalable data construction framework and a standardized evaluation dataset. This automated data construction approach facilitates future research on pretraining robotics data. +- Our experiments demonstrate that current pre-trained VLAs have yet to exhibit the strong generalization capabilities observed in LLMs, and existing SOTA VLMs also show limitations in embodied scenarios. + +# 2. Related Works + +Benchmarks and Datasets. Numerous benchmarks such as RLBench and LIBERO [27, 35, 38, 45, 66] have been proposed to evaluate language-conditioned manipulation policies. A comparison of these benchmarks is provided in Table 1. However, most of these focus on skill learning and fail to sufficiently address long-horizon planning capabilities. Meanwhile, some benchmarks [52, 55, 63] address room-scale mobile manipulation tasks that require + +
BenchmarksSemLangLogiReasonKnowledgeDLDRN-taskCate-objN-objAI-GenMultiCamPCDCross EmbAuto Traj
Alfred[52]XXX-X7-3578XXX✓X
Rlbench[27]XXX28.86✓10028168X✓XX✓
Calvin[45]XXX40.00X34530X✓XXX
ManiSkill[21, 46, 56]XXX-✓201002600X✓✓✓✓
LIBERO[38]XXX17.96✓1305175XXXXX
RoboCASA[47]XXX-✓1001532509✓✓X✓✓
ARNOLD[20]XXX--8-40✓✓✓X✓
SimplerEnv[35]XXX27.66✓8-17✓✓✓X✓
Behavior-1K[32]XXX--100022119331XXX✓X
Habitat 2.0[55]XXX--346169XXXXX
VLABench✓✓✓75.96✓1001632164✓✓✓✓✓
+ +Table 1. Comparison of popular benchmarks in robot learning. SemLang: semantically non-template rich language instructions. LogiReason: task logic and relevant information reasoning. Knowledge: tasks require the application of common sense and world knowledge. DL: quantitative task difficulty level. DR: task domain randomization. N-task: total number of tasks. Cate-object: categories of assets used in the simulation. N-object: total number of objects in the asset library. AI-Gen: use of generative AI models for asset library creation. MultiCam: use of multiple cameras. PCD: support point cloud data in 3D methods. Cross Emb: support for cross-embodiment. Auto Traj: supporting automated data collection. + +long memory and reasoning. Yet, these interactions typically occur through interfaces, rather than direct physical manipulation, limiting the transferability of learned policies to real-world scenarios. Additionally, while efforts [21, 32, 38, 47] have made strides in task format, difficulty, and scale, these benchmarks have largely overlooked the guiding role of language in tasks, often relying on template instructions. VLABench is the first to introduce features such as natural human interaction, implicit goal-oriented semantics, and requirements based on common sense into robot manipulation tasks, as shown in Figure 20. In terms of generalization evaluation, previous works [27, 38, 45] typically assess models at the instance level within the same category, which limits their ability to evaluate generalization across diverse object categories or different tasks within the same skill set. In contrast, VLABench is the first benchmark to evaluate generalization capabilities across a wide range of tasks, object types, and task categories, providing a more comprehensive assessment of model versatility. + +Large-scale datasets have been built in both real and simulation [6, 47, 50, 60] for large-scale imitation learning. However, real-world data faces challenges related to scalability, making it difficult to gather sufficient data at scale [4]. Simulated datasets, while more scalable, often suffer from limited diversity in scenarios and tasks [27, 45], and still require teleoperation [20, 38] for data collection. VLABench addresses these limitations by offering a broader range of tasks that are more closely aligned with real-world conditions, covering diverse aspects of vision, language, tasks, and skills. Furthermore, it introduces an efficient and robust process for the automated generation of simulated data, significantly enhancing task diversity and scalability. + +Pretrained Vision-Language-Action Models. The recent rise of multimodal models [1, 13, 40, 65] and the organization of operational datasets [49, 60], have led to the integration of vision-language-action models (VLAs) [4, 5, 14, 30] into language-conditioned manipulation tasks. While the + +term VLA generally refers to models that combine visual and language inputs for policy learning, we focus specifically on approaches leveraging pre-trained models. Several works [4, 14, 30] have applied further training to pre-trained vision-language models (VLMs) for language-conditioned manipulation. These models demonstrate impressive generalization to unseen objects and tasks, yet their control precision is somewhat limited by the discretization of actions [51]. To address this limitation, some approaches have explored using diffusion models [9, 51] as policy networks or using diffusion decoders [3, 34, 62]. Pre-trained models based on diffusion models [37, 43] have shown promising advancements in improving continuous space distribution learning. VLABench includes a selection of these representative methods for comprehensive evaluation. + +Framework Utilizing Foundation Models. Pre-trained language models [6, 15, 53] and vision-language models [1, 42] have demonstrated strong generalization and versatility. Some researchers [24, 25, 39] combine the general perception and cognitive abilities of these pre-trained models with traditional planning and control algorithms to create agent workflows. These frameworks allow robots to perform complex zero-shot manipulation tasks without requiring additional training. To harness the capabilities of foundation models for manipulation, some works [25, 26, 36] utilize the code comprehension and generation abilities of large language models alongside motion planning optimization algorithms to tackle fundamental manipulation tasks. Additionally, some methods [23, 24] leverage large models to decompose long-horizon tasks into subtasks, then integrate perception and trajectory generation modules to construct the entire manipulation pipeline. However, most zero-shot methods of this kind heavily rely on prompt design [25], the accuracy of each module, and even the specific parameters of the models being invoked [24]. While these methods demonstrate strong generalization, they often face challenges with accuracy. VLABench provides a + +zero-shot evaluation framework to assess the performance of such workflows and offers insights into their effectiveness. + +# 3. VLABench + +# 3.1. Task Description + +VLABench is composed of 60 primitive and 40 composite tasks, categorized by task difficulty levels. These tasks are designed to encompass a rich variety of skills while covering ample visual and language semantic information. For skill learning, 100 tasks in VLABench cover a wide range including 1) Pick&place, 2) Open&close door, 3) Open&close drawer, 4) Hang objects on the wall, 5) Use tool e.g. Hammer nail, 6) Press button, 7) Insert, 8) Pour, 9) Twist, and 10) Explore. In addition, VLABench places greater emphasis on real-life scenarios and essential daily tasks, representing more interactive language instructions, a wider variety of task settings, the integration of common sense and societal knowledge, and long-horizon tasks requiring logical planning, as shown in Figure 2. Notably, VLABench adopts a stricter definition of task generalization, which will be elaborated on in the Section 3.3. The whole task list can be found in supplementary material. + +Primitive Tasks. Primitive tasks are divided into five dimensions, each corresponding to the assessment of a specific ability dimension. + +- Mesh&Texture Understanding. This type of task requires the model to recognize different meshes and understand various texture features. Take the SelectToy task shown in Figure 2 (a) as an example, the robot is directly required to place a specific toy e.g. Aquaman into a receptacle. The model must possess strong visual capabilities to accurately recognize such complex meshes and textures. +- Spatial Understanding. Spatial understanding tasks involve various spatial relationships, such as the $n$ -th left/right position, inside/outside of a receptacle, the $m$ -th row and $n$ -th column, near/far, and beside a specific object, representing relative positional relationships. Figure 2 (b) shows one case in task in PullBook. Such a complex relative positional relationship imposes extremely high demands on the model's multimodal understanding [8]. +- Common Sense & World Knowledge. Tasks relative to common sense/world knowledge require the agent to transfer the knowledge gained in the pre-train stage to solve the problem. The task shown in Figure 2 (c) requires the agent not only to recognize different types of flowers from visual information but also to leverage world knowledge to determine that "the tulip is the national flower of the Netherlands". +- Semantic Understanding. This type of task emphasizes the complexity, subtlety, and natural interactivity of lan + +guage instructions. Task objectives are often implicitly conveyed through a natural conversation. To perform well in GetDrink task as Figure 2 (d) shown, the agent must capture the implied request from a lengthy instruction: to take out a chilled cola from the refrigerator. + +- Physical Law. This type of task expects the robot to integrate visual information and take correct actions based on physical principles and real-time observation. In UseSeesaw task in Figure 2 (e), the robot is commanded to grasp an object that can not be achieved directly. The agent must recognize the need to apply the principle of leverage by using sufficient weight to lift the target object on the other end. + +Composite Tasks. Composite tasks in VLABench involve the combination of multiple skills, long-term task planning, and multi-step logical reasoning from instructions, scenes, and even game rules. Figure 2 (f) showcases a variety of challenging complex tasks. Composite tasks have a significantly longer trajectory horizon, with an average episode length exceeding 500 timesteps—considerably more than the average of 120 timesteps for primitive tasks. In Figure 2 (f1), the agent must not only correctly identify all poker cards from visual information and use world knowledge of poker rules to select the best hand, but also flip face-down cards to acquire complete information. This type of task, requiring the agent to consciously fulfill prerequisite conditions, has never been modeled before. Composite tasks also require extracting the user's implicit needs from natural dialogue. The case in Figure 2 (f2) needs the agent to place the Python textbook on the table and open the laptop, without direct instruction. More cases are shown in Figure 20. + +# 3.2. Task Difficulty Level + +VLABench stands out among language-conditioned manipulation benchmarks not only for its multi-dimensional task design and evaluation but also for its hierarchical difficulty levels. The difficulty level (DL) is quantified using three metrics: task variants [27, 44], task sub-steps, and reasoning steps. These metrics quantify the overall task difficulty level from the perspectives of generalization difficulty, normalized episode horizon length, and intelligence. The detailed computation of DL is provided in Section 6.3. As shown in Table 1, VLABench's difficulty level surpasses all previous benchmarks and introduces new challenges that a VLA model should overcome. + +# 3.3.Benchmark + +Evaluation. VLABench organizes evaluations into three categories: assessments of pretrained or fine-tuned vision-language-action (VLA) models, heuristic workflows that integrate foundation models with various algorithms, and multi-dimensional evaluations of vision-language models. + +- Generalization Ability of VLAs. For pre-trained VLA + +![](images/63f654739307941224392488c182ed6f9d04c8e7e800c016b1708463ea988615.jpg) +(a) Mesh & Texture + +![](images/0df88616ed1ba70b88e77684ae751625e4823c4092bad9a45ece153918ed947d.jpg) +Instruction: "Take out the second book from the right on the top layer of the shelf." +(b) Spatial Understanding +(c) Common Sense & World Knowledge + +![](images/d520434d48035729c23df408a27935ac2c17b0c8d1434a4772a523734a5adc0f.jpg) +Instruction: "Insert the national flower of the Netherlands into the vase." + +![](images/18be63a104d2609547b39fecdd9318988aa1b99cedd68998177bcbad48691359.jpg) +Instruction: "Ah, I just spent an hour working out at the gym. I'm so thirsty now. I really go for an ice-cold soda!" +(d) Semantic Understanding + +![](images/571d946200018b56978cb39de311371db275f5a04885887bacc1cfdd1708f190.jpg) +Instruction: "Get the fruit from the container" +(e) Physical Law + +![](images/c63f423c0b2139c629cf11feeab2a04c7edc90e00df750ff9b1bfcd0e555f0f8.jpg) +Instruction:"We are playing Texas Hold'em game! Choose your largest combination on the table." +(f1) Play Texas Hold'em +Figure 2. Task examples in each dimension. The first row showcases examples of primitive tasks from Section 3.1, while the second row presents examples of composite tasks. + +![](images/4db8d16895c770d823bda986666d596ad56c23c809f48e8c2607e81f18d3755b.jpg) +Instruction: "Shortly, I'm going to work on my Python course assignment. Help me set up my desk." +(f2) Manage Study Table + +![](images/02f7f8bac0f37ab64c3eef82bfdf8ac6eed925f22ac5b9ccfa16aa5f18a8243b.jpg) +instruction:"I'd like to make a garlic eggplant dish.Prepare the ingredients I will use,thanks! +(f3) Cook Dishes as Menu +(f) Type: Composite + +![](images/8607e1c30f4a928917b5e33ad359ee964f1be2fc914afd1b63f1fa3322c806fb.jpg) +Instruction: "Could you please show the acid base neutralization experiment?" +(f4) Take Chemistry Experiment + +![](images/1bddbeedd0d2116299958b9c8f642d741c2fc4e1a480d7bd13317fa06a41980e.jpg) +Instruction: "I have to go to class this afternoon but I'm so sleepy! I want to have latte." +(f5) Get Latte Coffee + +models, the evaluation focuses on the model's generalization ability across multiple dimensions. In the standard evaluation, we have set up six evaluation tracks. + +- Track 1: Rapid downstream task transferability and skill acquisition ability. The evaluation episodes use objects seen in the training set, with minimal differences. +- Track 2: Cross-category visual generalization. The evaluation episodes use unseen categories of objects as the target entity for evaluation. +- Track 3: Common sense application. Building upon Track 1, replace parts of the instructions that directly reference the target object with content requiring common sense reasoning in evaluation episodes. +- Track 4: Semantically rich instruction adaptation. Building upon Track 1, replace the instructions with unseen and more complex ones. +- Track 5: Cross-domain behavior transferability. The evaluation tasks are replaced with ones that differ from those in the training set, but require similar actions. +- Track 6: Long-horizon task learning. Training and evaluation using more challenging composite tasks require additional skills and reasoning. + +- Zero-shot Transfer Ability of Heuristic Workflow. Training-free workflow methods are evaluated under sets of tasks in all dimensions mentioned in Section 3.1. We + +also extend the evaluation to cover various skills and long-horizon tasks to assess the overall capability and execution robustness of the workflow. All agents are evaluated in a zero-shot setting. + +- Comprehensive Evaluation of VLMs' Capabilities. Similar to heuristic workflows, the evaluation of VLMs is also comprehensive. Since VLMs lack intrinsic action capabilities, we organized a skill library and integrated it into a domain-specific language (DSL) [48, 54], leveraging annotated asset information as prior knowledge. This DSL functions as a straightforward API that VLMs can call, enabling efficient interaction. The evaluation pipeline is shown as Figure 3 and the details will be discussed in Section 9.3. + +Metric. In addition to the success rate (SR), considering the long-horizon nature and high difficulty level of our tasks, we introduce the intention score (IS) and progress score (PS) for more granular evaluation. Intention score is a useful soft metric measuring the correctness trend in selecting the target object for the operation especially when sub-tasks are not yet completed. The discrete version metric is: + +$$ +I S = \left\{ \begin{array}{l l} 1, & \text {i f d i s t a n c e} _ {\min } < \epsilon \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {1} +$$ + +where $distance_{min}$ is the minimum distance between the end-effector's position and the target entity, and $\epsilon$ is empiri + +![](images/d9d822663fbed6d68fce1a5811f581e9a193f3368dc74a27d7bdef3ddf74f171.jpg) +Figure 3. Evaluation pipeline for VLMs. Step 1: Sample the required four-view images, as well as those segmented with numerical information, from the simulation. Meanwhile, save the corresponding instructions and operation sequences. Step 2: Input the original and annotated images and the instructions into the model. Then, obtain the model's output and extract the generated operation sequence. Step 3: The two operation sequences are evaluated using four metrics, which are weighted and summed to produce the final score. The red part represents the error output of the model, where the red solid arrows represent the dependencies generated by the error, and red dashed arrows represent the dependencies lost by the error. + +![](images/d81d82de74d5d5819fa74b0e4570f8a6d0e48fa597d833977c5a00114ce82930.jpg) + +![](images/09dfe8349fa35fd6928ebf62a8c1fb5c1e1ffd7527aa94b7e03af43aaabb7f0f.jpg) + +cally set to 0.15 meters in most cases after experimentation. The progress score refers to the completion level of subtasks in a long-horizon task and serves as a softer process supervision metric compared to the success rate. + +$$ +P S = N _ {\text {d o n e}} / N _ {\text {s u b}} \tag {2} +$$ + +where $N_{done}$ is the number of completed sub-tasks, and $N_{sub}$ is the total number of task sub-tasks. + +# 3.4. Dataset Construction + +Trajectory Generation. As human teleoperation is time-consuming and not scalable [38, 47], we developed an efficient, scalable automated data collection pipeline based on our custom skill library. Inspired by [20], our data collection framework leverages the prior information including point clouds of the environment, annotated entities' grasppoints, target entity at the current step, etc. The data collection framework includes multiple task-specific motion planners. These motion planners call upon the skills in the skill library based on the current task progress and determine parameters by incorporating prior information. Subsequently, the selected skills generate trajectories using RRT [29], with quaternion interpolation achieved through spherical linear interpolation. The final trajectory is smoothed using a Bezier curve to optimize path quality. To enhance sample efficiency, reject sampling and failure-triggered early termination are applied. We demonstrate the superiority of our data framework through comparative experiments. Detailed analysis can be found in Section 8.3. + +Data Diversity and Quality. During the data construction process, we introduced diverse task variants and domain randomization across different episodes of the same task to ensure the diversity of the training data, as discussed + +in Sections 6.1 and 6.4. Following the approach of previous benchmarks built on Mujoco [38, 47], the dataset is stored in the same format, with similar visual rendering quality and trajectory accuracy. We further analyze the data quality in Section 8.5. + +Instruction Augmentation. We use GPT-4o [1] to generate descriptions that incorporate target-specific characteristics and interactive instructions that encompass a variety of contexts and intentions. Section 8.4 provides more details of the instruction generation. + +# 4. Experiments + +Following Section 3.3, we conducted experiments centered on pre-trained VLA models, workflows incorporating multiple algorithmic modules, and various VLMs. The remainder of this section provides a detailed description of the experimental setup. + +# 4.1. Generalization Ability of VLAs + +Pretrained VLAs are expected to possess robust generalization and versatility similar to LLMs. Experiments about are set to address the following research questions: + +Q1: Do pre-trained VLAs exhibit stronger general abilities with unseen categories of objects? +Q2: Can pre-trained VLAs transfer their general knowledge and behavioral abilities to similar but unseen tasks? +Q3: Can pre-trained VLAs understand natural user interactions and implicit goal requirements? +Q4: Do pre-trained VLAs have the potential to transfer their world knowledge to related tasks? +Q5: Can existing VLA architectures accurately support the completion of long-horizon tasks? + +Experiment Setup. To investigate the questions outlined + +
ModelTrack 1Track 2Track 3Track 4Track 5Track 6
ISPSSRISPSSRISPSSRISPSSRISPSSRISPSSR
Octo[57]32.05.62.020.42.00.022.02.81.226.84.01.213.60.00.017.23.00.0
OpenVLA[30]50.013.26.435.67.03.238.08.03.238.09.22.816.00.00.028.89.60.4
RDT-1B[43]50.014.36.039.28.23.236.07.24.035.68.04.420.40.00.021.69.20.0
π0-Base[3]59.629.313.249.621.46.060.420.05.260.419.33.622.00.00.034.412.82.0
π0-Fast[3]59.925.59.247.616.52.450.015.43.652.017.26.825.60.00.034.014.41.6
+ +Table 2. Overall experiment result of 6 evaluation tracks of fine-tuned VLAs. The detailed result of each task is reported in Table 9. + +above, we fine-tuned various pre-trained VLA architectures [3, 30, 43, 57] on our high-quality dataset, and evaluated them on different tracks mentioned in Section 3.3 to assess their performance across a range of tasks and conditions. Considering the high task difficulty, we selected five primitive tasks for experiments in tracks 1 to 5, with 200 data samples per task, and trained the models following a multitask training approach, following the setting in previous work [3, 43]. We then assessed the models across 50 episodes for each task, resulting in a total of 250 trials per track. For track 6, we selected five composite tasks, keeping the other settings consistent. Section 9.1 provides a complete list of tasks in different tracks and details of the model fine-tuning process. + +Result and Analysis. Table 2 presents the evaluation results across multiple tracks. These models perform poorly on most of tasks, as they are trained on a more diverse and limited set of fine-tuning data, with a higher difficulty level. Through the experimental results of each track and comparisons across different tracks, our analysis is as follows: + +- The primary cause of poor performance is the insufficient ability for rapid transfer in downstream tasks. Episodes in Tracks 1 and 6 closely resemble the training set, primarily assessing the model's basic fitting ability. Despite this, the models still perform poorly on certain tasks. We believe the main reason for the failure of most generalization experiments is the inaccuracy in basic action generation, especially Octo. +- Pre-trained VLA models exhibit some ability to handle unseen object categories. By comparing the results of Track 1 and Track 2, we observe that although all models experience a performance decline in Track 2, they are still able to complete a small subset of tasks. This is likely due to the benefits of large-scale robotics data pretraining or even earlier VLM pretraining. +- Performance decline with unseen semantic language. By combining the experimental results of Track 3 and Track 4, we find that the model's language understanding ability has deteriorated by a larger margin. Among these, the performance decline is more pronounced in models based on non-VLM architectures: RDT's progress score decreases by $45.2\%$ while OpenVLA and $\pi_0$ only decreases by $31.3\%$ and $34.2\%$ , respectively, in Track 4. +- Nearly absent cross-domain transferability. We found that all the models fail to perform similar but unseen tasks, + +due to significant visual and linguistic input differences. These policies exhibit uncertainty in behavior when faced with unfamiliar tasks, resulting in IS that are similar. + +We further analyze that data scaling and pre-training contribute positively to the model's performance Through basic ablation experiments. We also discuss in detail the potential issues with current VLAs, such as multimodal data co-training and model architecture designs. These discussions are presented in Section 10.1. We also discuss the gap between real and simulated environments in Section 10.2. + +# 4.2. Zero-shot Ability of Agent + +For our evaluation of foundation model-based algorithms, we reviewed two state-of-the-art frameworks, Voxposer [25] and CoPA [24], and the comparison results are shown in Figure 4. Given Voxposer's dependence on large language models (LLMs), we assessed Voxposer's performance with and without visual perception capabilities. While Voxposer performed adequately on basic tasks and achieved the Progress Scores of 30-40, its reliance on LLM-driven motion planning often led to grasping failures due to limited information for effective grasp planning, especially when interpreting rotation in non-visual contexts, resulting in low overall scores. + +Interestingly, the foundational LLM alone maintained relatively stable scores in semantic understanding and reasoning tasks without visual input. However, adding visual perception slightly reduced performance in these areas while significantly improving spatial reasoning, where LLM-only setups struggle with spatial accuracy due to lack of spatial information. + +The lack of closed-loop feedback limits these models' ability to perform physical reasoning tasks, particularly those involving dynamic interactions, leading to lower scores in this dimension. Both models struggle with high-complexity tasks, succeeding mainly in entity recognition but performing poorly in long horizon task reasonable breaking down. This finding underscores the requirements for advancing foundational model-based frameworks to address complex reasoning. In fact, although the methods mentioned above emphasize their zero-shot capabilities and generalization to new scenarios, their modular design often limits the upper bound of their performance. A more detailed discussion of this can be found in Section 10.3. + +![](images/ab42e09e91e4719e32faa7ee06b77fc25cb63c113ca620233be90b9cc5651c74.jpg) +Figure 4. Evaluation progress score for Voxposer and CoPA. Voxposer w/o refers to the version without visual perception, where ground truth labels are directly provided for object selection. Voxposer w uses GPT-4V as the visual perception module. + +# 4.3. Comprehensive Ability of VLMs + +We referred to the evaluation results of multiple series of VLMs provided by OpenCompass [11] and selected models from different families [1, 2, 7, 19, 22, 41] with strong overall performance as shown in Figure 5. We evaluate the comprehensive performance of these models with the dataset derived from naturally self-contained information within a simulated environment. This dataset consists of a complex set of tasks designed to assess the VLM's ability to perceive visual stimuli and comprehend verbal instructions. + +In line with the evaluation framework outlined in the Section 3.3, we evaluate the models across six dimensions of task performance. The results for each VLM model under the 1-shot setting are summarized in Figure 5. + +Although these VLMs perform well on most multimodal tasks and even some embodied tasks [11, 33], their performance, including that of GPT-4o, falls short when faced with more complex scenarios, instructions, and more challenging tasks. We surprisingly find that the open-source model Qwen2-VL-7B-Instruct performed competitively, surpassing GPT-4o on certain dimensions. However, all models struggled with complex tasks, especially those requiring long-term task decomposition and logic reasoning. Only GPT-4o achieved a score in the reasoning dimension comparable to those in other dimensions, while the other models scored around 20. Besides, performance declines significantly when linguistic instructions transition from direct semantics to abstract meanings, as shown in the + +![](images/e6d6e90334577f95dcf7c685ff912e3eb229070e26649f3421df62328a190dfd.jpg) +Figure 5. Radar charts depicting the performance of all VLM models across six dimensions. The reason why only GLM-4V-9B is evaluated in a zero-shot setting is that it does not support multigraph inference, which is required for the other models. + +semantic dimension. Different models appear to have distinct areas of expertise, e.g. LLaVA-NeXT exhibits weaker spatial perception ability, GLM-4V-9B excels in spatial and even physical law dimensions but lag in semantic comprehension. In Section 10.4, we conducted additional ablation experiments focusing on CoT and few-shot learning, followed by further analysis. Detailed data and error analysis can be found in Table 10. Overall, while the models demonstrate promising capabilities, their understanding and planning in embodied environments remain limited, highlighting the need for further advancements. + +# 5. Conclusion + +We propose VLABench, a large-scale benchmark designed for tasks with long-horizon and multi-dimensional reasoning. Such reasoning involves many dimensions, including implicit semantic goal extracting, logical reasoning, and long-horizon planning. The most important thing we do is to provide a positive definition of the capabilities that intelligent agents with cognitive abilities should possess and the tasks they should be able to perform, through the provision of 100 standardized tasks. Additionally, VLABench constructed a scalable automated data collection framework for future's potential larger scale pertaining and a standardized dataset for fair comparison of VLAs, both in the present and in future developments. Our diverse and multiple experiments revealed that current VLAs and VLMs face significant challenges in our tasks, and there remain substantial uncertainties in research on robotics scaling. We hope that VLABench will inspire both the future research on robotics pertaining recipe and promote more robust VLA architectures development. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 6, 8 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8 +[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi_0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3, 7, 11 +[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3 +[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In arXiv preprint arXiv:2307.15818, 2023. 2, 3, 10 +[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3 +[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 8 +[8] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 4 +[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 3 + +[10] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 28 +[11] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.8 +[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 4 +[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[14] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3 +[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3 +[16] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Grasynet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4 +[17] Hao-Shu Fang, Minghao Gou, Chenxi Wang, and Cewu Lu. Robust grasping across diverse sensor qualities: The graspnet-1billion dataset. The International Journal of Robotics Research, 2023. 2, 4 +[18] Jensen Gao, Suneel Belkhale, Sudeep Dasari, Ashwin Balakrishna, Dhruv Shah, and Dorsa Sadigh. A taxonomy for evaluating generalist robot policies. arXiv preprint arXiv:2503.01238, 2025. 10 +[19] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 8 +[20] Ran Gong, Jiangyong Huang, Yizhou Zhao, Haoran Geng, Xiaofeng Gao, Qingyang Wu, Wensi Ai, Ziheng Zhou, Demetri Terzopoulos, Song-Chun Zhu, et al. Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20483-20495, 2023. 3, 6 +[21] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for + +generalizable manipulation skills. In International Conference on Learning Representations, 2023. 3 +[22] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 8 +[23] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 3 +[24] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2, 3, 7 +[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3, 7 +[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3 +[27] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2, 3, 4, 5 +[28] Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846-894, 2011. 2 +[29] Sertac Karaman, Matthew R Walter, Alejandro Perez, Emilio Frazzoli, and Seth Teller. Anytime motion planning using the rrt. In 2011 IEEE international conference on robotics and automation, pages 1478-1483. IEEE, 2011. 6 +[30] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 7, 11 +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4 +[32] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Conference on Robot Learning, pages 80–93. PMLR, 2023. 3, 4 +[33] Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Erran Li, Ruohan Zhang, et al. Embodied agent interface: Benchmarking llms for embodied decision making. arXiv preprint arXiv:2410.07166, 2024. 8 +[34] Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, + +Yizhong Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024.3 +[35] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2, 3, 5, 11 +[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 3 +[37] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 3 +[38] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6, 5 +[39] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024. 3 +[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3 +[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.8 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3 +[43] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 3, 7, 11 +[44] Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023. 4, 11 +[45] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 2, 3 +[46] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3 +[47] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, + +and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems, 2024. 3, 6, 4, 10, 11 +[48] Arne Nordmann, Nico Hochgeschwender, and Sebastian Wrede. A survey on domain-specific languages in robotics. In International conference on simulation, modeling, and programming for autonomous robots, pages 195-206. Springer, 2014. 5 +[49] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3 +[50] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2, 3 +[51] Tim Pearce, Tabish Rashid, Anssi Kanervisto, Dave Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan, Ida Momennejad, Katja Hofmann, et al. Imitating human behaviour with diffusion models. arXiv preprint arXiv:2301.10677, 2023. 3 +[52] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2, 3 +[53] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, Xingjian Zhao, Ke Chen, Yining Zheng, Zhejian Zhou, Ruixiao Li, Jun Zhan, Yunhua Zhou, Linyang Li, Xiaogui Yang, Lingling Wu, Zhangyue Yin, Xuanjing Huang, Yu-Gang Jiang, and Xipeng Qiu. Moss: An open conversational large language model. Machine Intelligence Research, 2024. 3 +[54] Craig J Sutherland and Bruce MacDonald. Robolang: a simple domain specific language to script robot interactions. In 2019 16th International Conference on Ubiquitous Robots (UR), pages 265-270. IEEE, 2019. 5 +[55] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in neural information processing systems, 34:251-266, 2021. 2, 3 +[56] Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, Xander Hinrichsen, Xiaodi Yuan, Chen Bao, Xinsong Lin, Yulin Liu, Tse kai Chan, Yuan Gao, Xuanlin Li, Tongzhou Mu, Nan Xiao, Arnav Gurha, Zhiao Huang, Roberto Calandra, Rui Chen, Shan Luo, and Hao Su. Maniskill3: GPU parallelized robotics simulation and rendering for generalizable embodied ai. arXiv preprint arXiv:2410.00425, 2024. 3 +[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: + +An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 7 +[58] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 3 +[59] Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom Erez, Timothy Lillicrap, Nicolas Heess, and Yuval Tassa. dm_control: Software and tasks for continuous control. Software Impacts, 6:100022, 2020. 3 +[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 3 +[61] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. Survey Certification. 2 +[62] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 3 +[63] Sriram Yenamandra, Arun Ramachandran, Karmesh Yadav, Austin Wang, Mukul Khanna, Theophile Gervet, Tsung-Yen Yang, Vidhi Jain, Alexander William Clegg, John Turner, et al. Homerobot: Open-vocabulary mobile manipulation. arXiv preprint arXiv:2306.11565, 2023. 2 +[64] Kevin Zakka. obj2mjcf: Cli for processing composite wavefront obj files for use in mujoco. https://github.com/kevinzakka/obj2mjcf, 2023. Version 1.0.4 +[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15244-15253, 2023. 3 +[66] Liming Zheng, Feng Yan, Fanfan Liu, Chengjian Feng, Zhuoliang Kang, and Lin Ma. Robocas: A benchmark for robotic manipulation in complex object arrangement scenarios. arXiv preprint arXiv:2407.06951, 2024. 2 \ No newline at end of file diff --git a/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/images.zip b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03fe2bee7770049893b8a5eec55628bb8454326d --- /dev/null +++ b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb288923cf47159aeb49611ae30d8ed8f98a7f6105337855ed69cb2cd8813f5d +size 524317 diff --git a/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/layout.json b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3aa64a2cfd0f5d6ff5366cf9b69af6cb79a07800 --- /dev/null +++ b/2025/VLABench_ A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks/layout.json @@ -0,0 +1,8759 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 538, + 140 + ], + "type": "text", + "content": "VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "spans": [ + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": "Shiduo Zhang" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Zhe Xu" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Peiju Liu" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Xiaopeng Yu" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Yuan Li" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Qinghui Gao" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Zhaoye Fei" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Zhangyue Yin" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Zuxuan Wu" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Yu-Gang Jiang" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": ", Xipeng Qiu" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 493, + 203 + ], + "type": "text", + "content": "Fudan University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 214, + 154, + 305 + ], + "blocks": [ + { + "bbox": [ + 59, + 214, + 154, + 305 + ], + "lines": [ + { + "bbox": [ + 59, + 214, + 154, + 305 + ], + "spans": [ + { + "bbox": [ + 59, + 214, + 154, + 305 + ], + "type": "image", + "image_path": "e2d5b9d02a4369267daa42a154de08e0858272d4a89d1449eefb93d77103a1af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 215, + 255, + 304 + ], + "blocks": [ + { + "bbox": [ + 156, + 215, + 255, + 304 + ], + "lines": [ + { + "bbox": [ + 156, + 215, + 255, + 304 + ], + "spans": [ + { + "bbox": [ + 156, + 215, + 255, + 304 + ], + "type": "image", + "image_path": "caf0fadd36a258afd4859e02698f1d8efafb2cb934e1808a0914d1aeb705ffde.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 306, + 154, + 373 + ], + "blocks": [ + { + "bbox": [ + 58, + 306, + 154, + 373 + ], + "lines": [ + { + "bbox": [ + 58, + 306, + 154, + 373 + ], + "spans": [ + { + "bbox": [ + 58, + 306, + 154, + 373 + ], + "type": "image", + "image_path": "def1f31efc16ba96dcf5d6586ffd38658bb1ed315966d70a515284b308743162.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 156, + 307, + 255, + 373 + ], + "blocks": [ + { + "bbox": [ + 156, + 307, + 255, + 373 + ], + "lines": [ + { + "bbox": [ + 156, + 307, + 255, + 373 + ], + "spans": [ + { + "bbox": [ + 156, + 307, + 255, + 373 + ], + "type": "image", + "image_path": "e97073b7b70ac8f2f07a3822284eb064f9c9f5f37b2ea4cc873deac37cde16ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 375, + 156, + 460 + ], + "blocks": [ + { + "bbox": [ + 58, + 375, + 156, + 460 + ], + "lines": [ + { + "bbox": [ + 58, + 375, + 156, + 460 + ], + "spans": [ + { + "bbox": [ + 58, + 375, + 156, + 460 + ], + "type": "image", + "image_path": "746e022f6a6837f28187d6f2271c164963d7425ffc48671deb445672556da165.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 471, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 471, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 555, + 495 + ], + "type": "text", + "content": "Figure 1. Overview of VLABench. VLABench is a large-scale language-conditioned manipulation benchmark to evaluate the comprehensive skill learning and generalization ability of action policies especially pre-trained vision-language-action models." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 156, + 374, + 254, + 460 + ], + "blocks": [ + { + "bbox": [ + 156, + 374, + 254, + 460 + ], + "lines": [ + { + "bbox": [ + 156, + 374, + 254, + 460 + ], + "spans": [ + { + "bbox": [ + 156, + 374, + 254, + 460 + ], + "type": "image", + "image_path": "5d9d6bd8001aa8ec88ee2192a33ad6ac9502628d5b6a3c016723deccc3b35c76.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 332, + 220, + 475, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 220, + 475, + 232 + ], + "spans": [ + { + "bbox": [ + 332, + 220, + 475, + 232 + ], + "type": "text", + "content": "Task-Related Comprehensive Ability" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 260, + 239, + 398, + 310 + ], + "blocks": [ + { + "bbox": [ + 260, + 239, + 398, + 310 + ], + "lines": [ + { + "bbox": [ + 260, + 239, + 398, + 310 + ], + "spans": [ + { + "bbox": [ + 260, + 239, + 398, + 310 + ], + "type": "image", + "image_path": "4579bc73e65769ac1650ed26881978228912ce4383f267f8e61644907781627e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 399, + 242, + 539, + 310 + ], + "blocks": [ + { + "bbox": [ + 399, + 242, + 539, + 310 + ], + "lines": [ + { + "bbox": [ + 399, + 242, + 539, + 310 + ], + "spans": [ + { + "bbox": [ + 399, + 242, + 539, + 310 + ], + "type": "image", + "image_path": "3071cec2797462e47d4a174f32a1306820d137bd8d5a3f1cb4bdf4a76b64beb2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 260, + 314, + 398, + 381 + ], + "blocks": [ + { + "bbox": [ + 260, + 314, + 398, + 381 + ], + "lines": [ + { + "bbox": [ + 260, + 314, + 398, + 381 + ], + "spans": [ + { + "bbox": [ + 260, + 314, + 398, + 381 + ], + "type": "image", + "image_path": "b262b251d3d39fec7131e9ce6112260537a66252982bd725d5be59653fb259d4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 399, + 314, + 539, + 381 + ], + "blocks": [ + { + "bbox": [ + 399, + 314, + 539, + 381 + ], + "lines": [ + { + "bbox": [ + 399, + 314, + 539, + 381 + ], + "spans": [ + { + "bbox": [ + 399, + 314, + 539, + 381 + ], + "type": "image", + "image_path": "d28972f469966ffdc8407caba0dd83510eeb5b734e1c9bfa9efc7534a3ade0ee.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 260, + 385, + 398, + 453 + ], + "blocks": [ + { + "bbox": [ + 260, + 385, + 398, + 453 + ], + "lines": [ + { + "bbox": [ + 260, + 385, + 398, + 453 + ], + "spans": [ + { + "bbox": [ + 260, + 385, + 398, + 453 + ], + "type": "image", + "image_path": "fff038d2f224a9d9721e63eaea3f4c556337d0592bde01a0035665fb18fe9748.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 399, + 385, + 539, + 453 + ], + "blocks": [ + { + "bbox": [ + 399, + 385, + 539, + 453 + ], + "lines": [ + { + "bbox": [ + 399, + 385, + 539, + 453 + ], + "spans": [ + { + "bbox": [ + 399, + 385, + 539, + 453 + ], + "type": "image", + "image_path": "12f1307e6cfd524303b97c6cbadbd3823073bceffdedde6d41c3bcf7150f1a6e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 152, + 503, + 200, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 503, + 200, + 516 + ], + "spans": [ + { + "bbox": [ + 152, + 503, + 200, + 516 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 533, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 297, + 714 + ], + "type": "text", + "content": "General-purposed embodied agents are designed to understand the users' natural instructions or intentions and act precisely to complete universal tasks. Recently, methods based on foundation models especially Vision-Language-Action models (VLAs) have shown a substantial potential to solve language-conditioned manipulation (LCM) tasks well. However, existing benchmarks do not adequately meet the needs of VLAs and relative algorithms. To better define such general-purpose tasks in the context of LLMs and advance the research in VLAs, we present VLABench, an open-source benchmark for evaluating universal LCM task learning. VLABench provides 100 carefully designed categories of tasks, with strong randomization in each category of task and a total of " + }, + { + "bbox": [ + 54, + 533, + 297, + 714 + ], + "type": "inline_equation", + "content": "2000+" + }, + { + "bbox": [ + 54, + 533, + 297, + 714 + ], + "type": "text", + "content": " objects. VLABench stands out from previous benchmarks in four key aspects: 1) tasks" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 505, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 556, + 673 + ], + "type": "text", + "content": "requiring world knowledge and common sense transfer, 2) natural language instructions with implicit human intentions rather than templates, 3) long-horizon tasks demanding multi-step reasoning, and 4) evaluation of both action policies and language model capabilities. The benchmark assesses multiple competencies including understanding of mesh&texture, spatial relationship, semantic instruction, physical laws, knowledge transfer and reasoning, etc. To support the downstream finetuning, we provide high-quality training data collected via an automated framework incorporating heuristic skills and prior information. The experimental results indicate that both the current state-of-the-art pretrained VLAs and the workflow based on VLMs face challenges in our tasks." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 324, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 553, + 713 + ], + "type": "text", + "content": "1Codes and more videos are available at https://vlabench.github.io/ \n2Corresponding to: sdzhang23@m.fudan.edu.cn, xpqiu@fudan.edu.cn" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11142" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 317 + ], + "type": "text", + "content": "Language-conditioned manipulation represents a fundamental challenge in embodied AI and a stepping stone toward Artificial General Intelligence [1, 5, 14]. Such tasks require agents to master multiple capabilities: interpreting natural language instructions, understanding complex environments, making decisions, formulating plans, and executing precise actions. The rapid advancement of Large Language Models (LLMs) and Vision-Language Models (VLMs) [1, 15] has revolutionized the field with their impressive general abilities in semantic understanding, coding, planning, and reasoning. The strong generalization capabilities has inspired two main approaches in language-conditioned manipulation: pre-training vision-language-action models using large-scale robotics data, as demonstrated by RT-2 and Palm-E [5, 14, 50], and integrating foundation models into agent workflows, like VoxPoser and Copa [24, 25], which combine LLM/VLM outputs with grasp prediction [16, 17] and motion planning algorithms [28]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 617 + ], + "type": "text", + "content": "While real-world robotics experiments provide valuable insights, their complexity and environmental variability often challenge reproducibility. Simulation-based evaluation has emerged as a fair and practical alternative. Existing benchmarks like RLBench, Calvin, and LIBERO [27, 38, 45] offer diverse task sets but fall short in addressing the unique requirements of foundation model-based methods. Tasks designed to align with the capabilities of foundation model-based algorithms should encompass nuanced semantic understanding of user intent, the integration of common-sense knowledge, and a robust ability to interpret diverse visual scenes, as well as require sophisticated multi-step reasoning. Such tasks demand a sophisticated integration of multimodal understanding to effectively interpret and respond to complex, real-world contexts. For example, one of the tasks in RT-2 [5] is \"move the Coke can to Taylor Swift\", while another task in CoPA [24] is \"make me a cup of pour-over coffee\". The first task challenges the robot to use common sense to identify Taylor, a capability of knowledge transfer that previous policies struggled to achieve. The second task further intensifies the difficulty, requiring the robot to decompose the task into subtasks and execute the steps to operate a coffee machine—a long-horizon challenge that has previously been difficult for a single policy to accomplish." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "To better define the types of language-conditioned manipulation tasks suited for foundation models and provide a standardized evaluation suite to advance robotics research, we introduce VLABench. VLABench is an open-source benchmark specifically designed for methods utilizing foundation models. The tasks in VLABench are carefully divided into several dimensions to evaluate models across various aspects, including 1) Mastery of common" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "sense and world knowledge, 2) Understanding of mesh and texture, 3) Comprehension of semantically rich instructions, 4) Spatial understanding, 5) Grasp of physical rules, and 6) Reasoning ability. For benchmarking purposes, VLABench offers 100 task categories with comprehensive evaluations across various methods. With a diverse collection of over 2,000 3D objects and scenes, VLABench creates a wide range of visual contexts and tasks. It enables the assessment of generalization capabilities through learning across multiple skills, providing thorough evaluations spanning visual, linguistic, planning, knowledge transfer, and action dimensions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 216, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 357 + ], + "type": "text", + "content": "To ensure fair comparison and evaluation, we develop an automated data collection framework to construct standardized datasets for each task, supporting model training and fine-tuning. Using this dataset, we conduct extensive experiments to evaluate and benchmark three distinct types of approaches: pre-trained VLA, workflows integrating foundation models, and vision-language models (VLMs). The experimental results indicate that existing VLA methods perform poorly on our tasks and don't exhibit the level of generalization abilities or the \"emergent\" phenomena observed in large models [61]. We summarize contributions as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 361, + 555, + 586 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 361, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 555, + 420 + ], + "type": "text", + "content": "- We propose VLABench, the first benchmark designed to comprehensively evaluate the capabilities of VLAs and VLMs in robotics manipulation tasks, covering multiple dimensions such as skills, vision, language, task execution, common sense, and reasoning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 421, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 554, + 491 + ], + "type": "text", + "content": "- We define 100 novel LCM tasks tailored to the capabilities of foundation models within a standardized evaluation framework. These tasks require a deep understanding of semantics, vision, space, and physical laws, as well as the ability to plan long-horizon tasks and transfer world knowledge and common sense into task execution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 492, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 492, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 314, + 492, + 554, + 539 + ], + "type": "text", + "content": "- We provide a scalable data construction framework and a standardized evaluation dataset. This automated data construction approach facilitates future research on pretraining robotics data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 540, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 540, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 540, + 554, + 586 + ], + "type": "text", + "content": "- Our experiments demonstrate that current pre-trained VLAs have yet to exhibit the strong generalization capabilities observed in LLMs, and existing SOTA VLMs also show limitations in embodied scenarios." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 597, + 406, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 406, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 406, + 609 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Benchmarks and Datasets. Numerous benchmarks such as RLBench and LIBERO [27, 35, 38, 45, 66] have been proposed to evaluate language-conditioned manipulation policies. A comparison of these benchmarks is provided in Table 1. However, most of these focus on skill learning and fail to sufficiently address long-horizon planning capabilities. Meanwhile, some benchmarks [52, 55, 63] address room-scale mobile manipulation tasks that require" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11143" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 186 + ], + "type": "table", + "html": "
BenchmarksSemLangLogiReasonKnowledgeDLDRN-taskCate-objN-objAI-GenMultiCamPCDCross EmbAuto Traj
Alfred[52]XXX-X7-3578XXX✓X
Rlbench[27]XXX28.86✓10028168X✓XX✓
Calvin[45]XXX40.00X34530X✓XXX
ManiSkill[21, 46, 56]XXX-✓201002600X✓✓✓✓
LIBERO[38]XXX17.96✓1305175XXXXX
RoboCASA[47]XXX-✓1001532509✓✓X✓✓
ARNOLD[20]XXX--8-40✓✓✓X✓
SimplerEnv[35]XXX27.66✓8-17✓✓✓X✓
Behavior-1K[32]XXX--100022119331XXX✓X
Habitat 2.0[55]XXX--346169XXXXX
VLABench✓✓✓75.96✓1001632164✓✓✓✓✓
", + "image_path": "1232b6c324178c581c0d0b44f6da34ef28a3d18a67473917a4fc34110b35c20d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 191, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 555, + 258 + ], + "type": "text", + "content": "Table 1. Comparison of popular benchmarks in robot learning. SemLang: semantically non-template rich language instructions. LogiReason: task logic and relevant information reasoning. Knowledge: tasks require the application of common sense and world knowledge. DL: quantitative task difficulty level. DR: task domain randomization. N-task: total number of tasks. Cate-object: categories of assets used in the simulation. N-object: total number of objects in the asset library. AI-Gen: use of generative AI models for asset library creation. MultiCam: use of multiple cameras. PCD: support point cloud data in 3D methods. Cross Emb: support for cross-embodiment. Auto Traj: supporting automated data collection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 269, + 294, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 294, + 496 + ], + "type": "text", + "content": "long memory and reasoning. Yet, these interactions typically occur through interfaces, rather than direct physical manipulation, limiting the transferability of learned policies to real-world scenarios. Additionally, while efforts [21, 32, 38, 47] have made strides in task format, difficulty, and scale, these benchmarks have largely overlooked the guiding role of language in tasks, often relying on template instructions. VLABench is the first to introduce features such as natural human interaction, implicit goal-oriented semantics, and requirements based on common sense into robot manipulation tasks, as shown in Figure 20. In terms of generalization evaluation, previous works [27, 38, 45] typically assess models at the instance level within the same category, which limits their ability to evaluate generalization across diverse object categories or different tasks within the same skill set. In contrast, VLABench is the first benchmark to evaluate generalization capabilities across a wide range of tasks, object types, and task categories, providing a more comprehensive assessment of model versatility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "content": "Large-scale datasets have been built in both real and simulation [6, 47, 50, 60] for large-scale imitation learning. However, real-world data faces challenges related to scalability, making it difficult to gather sufficient data at scale [4]. Simulated datasets, while more scalable, often suffer from limited diversity in scenarios and tasks [27, 45], and still require teleoperation [20, 38] for data collection. VLABench addresses these limitations by offering a broader range of tasks that are more closely aligned with real-world conditions, covering diverse aspects of vision, language, tasks, and skills. Furthermore, it introduces an efficient and robust process for the automated generation of simulated data, significantly enhancing task diversity and scalability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Pretrained Vision-Language-Action Models. The recent rise of multimodal models [1, 13, 40, 65] and the organization of operational datasets [49, 60], have led to the integration of vision-language-action models (VLAs) [4, 5, 14, 30] into language-conditioned manipulation tasks. While the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 269, + 553, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 448 + ], + "type": "text", + "content": "term VLA generally refers to models that combine visual and language inputs for policy learning, we focus specifically on approaches leveraging pre-trained models. Several works [4, 14, 30] have applied further training to pre-trained vision-language models (VLMs) for language-conditioned manipulation. These models demonstrate impressive generalization to unseen objects and tasks, yet their control precision is somewhat limited by the discretization of actions [51]. To address this limitation, some approaches have explored using diffusion models [9, 51] as policy networks or using diffusion decoders [3, 34, 62]. Pre-trained models based on diffusion models [37, 43] have shown promising advancements in improving continuous space distribution learning. VLABench includes a selection of these representative methods for comprehensive evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "Framework Utilizing Foundation Models. Pre-trained language models [6, 15, 53] and vision-language models [1, 42] have demonstrated strong generalization and versatility. Some researchers [24, 25, 39] combine the general perception and cognitive abilities of these pre-trained models with traditional planning and control algorithms to create agent workflows. These frameworks allow robots to perform complex zero-shot manipulation tasks without requiring additional training. To harness the capabilities of foundation models for manipulation, some works [25, 26, 36] utilize the code comprehension and generation abilities of large language models alongside motion planning optimization algorithms to tackle fundamental manipulation tasks. Additionally, some methods [23, 24] leverage large models to decompose long-horizon tasks into subtasks, then integrate perception and trajectory generation modules to construct the entire manipulation pipeline. However, most zero-shot methods of this kind heavily rely on prompt design [25], the accuracy of each module, and even the specific parameters of the models being invoked [24]. While these methods demonstrate strong generalization, they often face challenges with accuracy. VLABench provides a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11144" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "zero-shot evaluation framework to assess the performance of such workflows and offers insights into their effectiveness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 129, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 129, + 134 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 129, + 134 + ], + "type": "text", + "content": "3. VLABench" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 157, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 157, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 157, + 156 + ], + "type": "text", + "content": "3.1. Task Description" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 161, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 365 + ], + "type": "text", + "content": "VLABench is composed of 60 primitive and 40 composite tasks, categorized by task difficulty levels. These tasks are designed to encompass a rich variety of skills while covering ample visual and language semantic information. For skill learning, 100 tasks in VLABench cover a wide range including 1) Pick&place, 2) Open&close door, 3) Open&close drawer, 4) Hang objects on the wall, 5) Use tool e.g. Hammer nail, 6) Press button, 7) Insert, 8) Pour, 9) Twist, and 10) Explore. In addition, VLABench places greater emphasis on real-life scenarios and essential daily tasks, representing more interactive language instructions, a wider variety of task settings, the integration of common sense and societal knowledge, and long-horizon tasks requiring logical planning, as shown in Figure 2. Notably, VLABench adopts a stricter definition of task generalization, which will be elaborated on in the Section 3.3. The whole task list can be found in supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 402 + ], + "type": "text", + "content": "Primitive Tasks. Primitive tasks are divided into five dimensions, each corresponding to the assessment of a specific ability dimension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 403, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 497 + ], + "type": "text", + "content": "- Mesh&Texture Understanding. This type of task requires the model to recognize different meshes and understand various texture features. Take the SelectToy task shown in Figure 2 (a) as an example, the robot is directly required to place a specific toy e.g. Aquaman into a receptacle. The model must possess strong visual capabilities to accurately recognize such complex meshes and textures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": "- Spatial Understanding. Spatial understanding tasks involve various spatial relationships, such as the " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": "-th left/right position, inside/outside of a receptacle, the " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": "-th column, near/far, and beside a specific object, representing relative positional relationships. Figure 2 (b) shows one case in task in PullBook. Such a complex relative positional relationship imposes extremely high demands on the model's multimodal understanding [8]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 689 + ], + "type": "text", + "content": "- Common Sense & World Knowledge. Tasks relative to common sense/world knowledge require the agent to transfer the knowledge gained in the pre-train stage to solve the problem. The task shown in Figure 2 (c) requires the agent not only to recognize different types of flowers from visual information but also to leverage world knowledge to determine that \"the tulip is the national flower of the Netherlands\"." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "- Semantic Understanding. This type of task emphasizes the complexity, subtlety, and natural interactivity of lan" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 555, + 133 + ], + "type": "text", + "content": "guage instructions. Task objectives are often implicitly conveyed through a natural conversation. To perform well in GetDrink task as Figure 2 (d) shown, the agent must capture the implied request from a lengthy instruction: to take out a chilled cola from the refrigerator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 133, + 555, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 555, + 227 + ], + "type": "text", + "content": "- Physical Law. This type of task expects the robot to integrate visual information and take correct actions based on physical principles and real-time observation. In UseSeesaw task in Figure 2 (e), the robot is commanded to grasp an object that can not be achieved directly. The agent must recognize the need to apply the principle of leverage by using sufficient weight to lift the target object on the other end." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 228, + 556, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 556, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 556, + 444 + ], + "type": "text", + "content": "Composite Tasks. Composite tasks in VLABench involve the combination of multiple skills, long-term task planning, and multi-step logical reasoning from instructions, scenes, and even game rules. Figure 2 (f) showcases a variety of challenging complex tasks. Composite tasks have a significantly longer trajectory horizon, with an average episode length exceeding 500 timesteps—considerably more than the average of 120 timesteps for primitive tasks. In Figure 2 (f1), the agent must not only correctly identify all poker cards from visual information and use world knowledge of poker rules to select the best hand, but also flip face-down cards to acquire complete information. This type of task, requiring the agent to consciously fulfill prerequisite conditions, has never been modeled before. Composite tasks also require extracting the user's implicit needs from natural dialogue. The case in Figure 2 (f2) needs the agent to place the Python textbook on the table and open the laptop, without direct instruction. More cases are shown in Figure 20." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 434, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 434, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 434, + 464 + ], + "type": "text", + "content": "3.2. Task Difficulty Level" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 471, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 615 + ], + "type": "text", + "content": "VLABench stands out among language-conditioned manipulation benchmarks not only for its multi-dimensional task design and evaluation but also for its hierarchical difficulty levels. The difficulty level (DL) is quantified using three metrics: task variants [27, 44], task sub-steps, and reasoning steps. These metrics quantify the overall task difficulty level from the perspectives of generalization difficulty, normalized episode horizon length, and intelligence. The detailed computation of DL is provided in Section 6.3. As shown in Table 1, VLABench's difficulty level surpasses all previous benchmarks and introduces new challenges that a VLA model should overcome." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 624, + 392, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 392, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 392, + 635 + ], + "type": "text", + "content": "3.3.Benchmark" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 700 + ], + "type": "text", + "content": "Evaluation. VLABench organizes evaluations into three categories: assessments of pretrained or fine-tuned vision-language-action (VLA) models, heuristic workflows that integrate foundation models with various algorithms, and multi-dimensional evaluations of vision-language models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "content": "- Generalization Ability of VLAs. For pre-trained VLA" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11145" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 156, + 196 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 156, + 196 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 156, + 196 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 156, + 196 + ], + "type": "image", + "image_path": "63f654739307941224392488c182ed6f9d04c8e7e800c016b1708463ea988615.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 198, + 135, + 205 + ], + "lines": [ + { + "bbox": [ + 84, + 198, + 135, + 205 + ], + "spans": [ + { + "bbox": [ + 84, + 198, + 135, + 205 + ], + "type": "text", + "content": "(a) Mesh & Texture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 159, + 99, + 254, + 196 + ], + "blocks": [ + { + "bbox": [ + 159, + 74, + 253, + 98 + ], + "lines": [ + { + "bbox": [ + 159, + 74, + 253, + 98 + ], + "spans": [ + { + "bbox": [ + 159, + 74, + 253, + 98 + ], + "type": "text", + "content": "Instruction: \"Take out the second book from the right on the top layer of the shelf.\"" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 99, + 254, + 196 + ], + "lines": [ + { + "bbox": [ + 159, + 99, + 254, + 196 + ], + "spans": [ + { + "bbox": [ + 159, + 99, + 254, + 196 + ], + "type": "image", + "image_path": "0df88616ed1ba70b88e77684ae751625e4823c4092bad9a45ece153918ed947d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 198, + 239, + 206 + ], + "lines": [ + { + "bbox": [ + 171, + 198, + 239, + 206 + ], + "spans": [ + { + "bbox": [ + 171, + 198, + 239, + 206 + ], + "type": "text", + "content": "(b) Spatial Understanding" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 198, + 356, + 206 + ], + "lines": [ + { + "bbox": [ + 255, + 198, + 356, + 206 + ], + "spans": [ + { + "bbox": [ + 255, + 198, + 356, + 206 + ], + "type": "text", + "content": "(c) Common Sense & World Knowledge" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 257, + 99, + 353, + 195 + ], + "blocks": [ + { + "bbox": [ + 258, + 76, + 350, + 91 + ], + "lines": [ + { + "bbox": [ + 258, + 76, + 350, + 91 + ], + "spans": [ + { + "bbox": [ + 258, + 76, + 350, + 91 + ], + "type": "text", + "content": "Instruction: \"Insert the national flower of the Netherlands into the vase.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 99, + 353, + 195 + ], + "lines": [ + { + "bbox": [ + 257, + 99, + 353, + 195 + ], + "spans": [ + { + "bbox": [ + 257, + 99, + 353, + 195 + ], + "type": "image", + "image_path": "d520434d48035729c23df408a27935ac2c17b0c8d1434a4772a523734a5adc0f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 356, + 99, + 452, + 195 + ], + "blocks": [ + { + "bbox": [ + 358, + 74, + 449, + 99 + ], + "lines": [ + { + "bbox": [ + 358, + 74, + 449, + 99 + ], + "spans": [ + { + "bbox": [ + 358, + 74, + 449, + 99 + ], + "type": "text", + "content": "Instruction: \"Ah, I just spent an hour working out at the gym. I'm so thirsty now. I really go for an ice-cold soda!\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 99, + 452, + 195 + ], + "lines": [ + { + "bbox": [ + 356, + 99, + 452, + 195 + ], + "spans": [ + { + "bbox": [ + 356, + 99, + 452, + 195 + ], + "type": "image", + "image_path": "18be63a104d2609547b39fecdd9318988aa1b99cedd68998177bcbad48691359.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 198, + 441, + 205 + ], + "lines": [ + { + "bbox": [ + 367, + 198, + 441, + 205 + ], + "spans": [ + { + "bbox": [ + 367, + 198, + 441, + 205 + ], + "type": "text", + "content": "(d) Semantic Understanding" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 454, + 99, + 552, + 195 + ], + "blocks": [ + { + "bbox": [ + 457, + 76, + 541, + 91 + ], + "lines": [ + { + "bbox": [ + 457, + 76, + 541, + 91 + ], + "spans": [ + { + "bbox": [ + 457, + 76, + 541, + 91 + ], + "type": "text", + "content": "Instruction: \"Get the fruit from the container\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 99, + 552, + 195 + ], + "lines": [ + { + "bbox": [ + 454, + 99, + 552, + 195 + ], + "spans": [ + { + "bbox": [ + 454, + 99, + 552, + 195 + ], + "type": "image", + "image_path": "571d946200018b56978cb39de311371db275f5a04885887bacc1cfdd1708f190.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 198, + 524, + 205 + ], + "lines": [ + { + "bbox": [ + 481, + 198, + 524, + 205 + ], + "spans": [ + { + "bbox": [ + 481, + 198, + 524, + 205 + ], + "type": "text", + "content": "(e) Physical Law" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 58, + 235, + 156, + 331 + ], + "blocks": [ + { + "bbox": [ + 59, + 209, + 156, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 156, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 156, + 234 + ], + "type": "text", + "content": "Instruction:\"We are playing Texas Hold'em game! Choose your largest combination on the table.\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 235, + 156, + 331 + ], + "lines": [ + { + "bbox": [ + 58, + 235, + 156, + 331 + ], + "spans": [ + { + "bbox": [ + 58, + 235, + 156, + 331 + ], + "type": "image", + "image_path": "c63f423c0b2139c629cf11feeab2a04c7edc90e00df750ff9b1bfcd0e555f0f8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 333, + 134, + 340 + ], + "lines": [ + { + "bbox": [ + 78, + 333, + 134, + 340 + ], + "spans": [ + { + "bbox": [ + 78, + 333, + 134, + 340 + ], + "type": "text", + "content": "(f1) Play Texas Hold'em" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 362, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 555, + 384 + ], + "type": "text", + "content": "Figure 2. Task examples in each dimension. The first row showcases examples of primitive tasks from Section 3.1, while the second row presents examples of composite tasks." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 158, + 235, + 254, + 328 + ], + "blocks": [ + { + "bbox": [ + 159, + 209, + 253, + 234 + ], + "lines": [ + { + "bbox": [ + 159, + 209, + 253, + 234 + ], + "spans": [ + { + "bbox": [ + 159, + 209, + 253, + 234 + ], + "type": "text", + "content": "Instruction: \"Shortly, I'm going to work on my Python course assignment. Help me set up my desk.\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 235, + 254, + 328 + ], + "lines": [ + { + "bbox": [ + 158, + 235, + 254, + 328 + ], + "spans": [ + { + "bbox": [ + 158, + 235, + 254, + 328 + ], + "type": "image", + "image_path": "4db8d16895c770d823bda986666d596ad56c23c809f48e8c2607e81f18d3755b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 333, + 235, + 340 + ], + "lines": [ + { + "bbox": [ + 178, + 333, + 235, + 340 + ], + "spans": [ + { + "bbox": [ + 178, + 333, + 235, + 340 + ], + "type": "text", + "content": "(f2) Manage Study Table" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 257, + 235, + 353, + 331 + ], + "blocks": [ + { + "bbox": [ + 257, + 209, + 351, + 234 + ], + "lines": [ + { + "bbox": [ + 257, + 209, + 351, + 234 + ], + "spans": [ + { + "bbox": [ + 257, + 209, + 351, + 234 + ], + "type": "text", + "content": "instruction:\"I'd like to make a garlic eggplant dish.Prepare the ingredients I will use,thanks!" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 235, + 353, + 331 + ], + "lines": [ + { + "bbox": [ + 257, + 235, + 353, + 331 + ], + "spans": [ + { + "bbox": [ + 257, + 235, + 353, + 331 + ], + "type": "image", + "image_path": "02f7f8bac0f37ab64c3eef82bfdf8ac6eed925f22ac5b9ccfa16aa5f18a8243b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 333, + 335, + 339 + ], + "lines": [ + { + "bbox": [ + 276, + 333, + 335, + 339 + ], + "spans": [ + { + "bbox": [ + 276, + 333, + 335, + 339 + ], + "type": "text", + "content": "(f3) Cook Dishes as Menu" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 279, + 342, + 331, + 350 + ], + "lines": [ + { + "bbox": [ + 279, + 342, + 331, + 350 + ], + "spans": [ + { + "bbox": [ + 279, + 342, + 331, + 350 + ], + "type": "text", + "content": "(f) Type: Composite" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 356, + 234, + 452, + 331 + ], + "blocks": [ + { + "bbox": [ + 356, + 209, + 451, + 227 + ], + "lines": [ + { + "bbox": [ + 356, + 209, + 451, + 227 + ], + "spans": [ + { + "bbox": [ + 356, + 209, + 451, + 227 + ], + "type": "text", + "content": "Instruction: \"Could you please show the acid base neutralization experiment?\"" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 234, + 452, + 331 + ], + "lines": [ + { + "bbox": [ + 356, + 234, + 452, + 331 + ], + "spans": [ + { + "bbox": [ + 356, + 234, + 452, + 331 + ], + "type": "image", + "image_path": "8607e1c30f4a928917b5e33ad359ee964f1be2fc914afd1b63f1fa3322c806fb.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 333, + 441, + 340 + ], + "lines": [ + { + "bbox": [ + 367, + 333, + 441, + 340 + ], + "spans": [ + { + "bbox": [ + 367, + 333, + 441, + 340 + ], + "type": "text", + "content": "(f4) Take Chemistry Experiment" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 454, + 235, + 552, + 331 + ], + "blocks": [ + { + "bbox": [ + 457, + 209, + 547, + 233 + ], + "lines": [ + { + "bbox": [ + 457, + 209, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 457, + 209, + 547, + 233 + ], + "type": "text", + "content": "Instruction: \"I have to go to class this afternoon but I'm so sleepy! I want to have latte.\"" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 235, + 552, + 331 + ], + "lines": [ + { + "bbox": [ + 454, + 235, + 552, + 331 + ], + "spans": [ + { + "bbox": [ + 454, + 235, + 552, + 331 + ], + "type": "image", + "image_path": "1bddbeedd0d2116299958b9c8f642d741c2fc4e1a480d7bd13317fa06a41980e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 333, + 527, + 339 + ], + "lines": [ + { + "bbox": [ + 479, + 333, + 527, + 339 + ], + "spans": [ + { + "bbox": [ + 479, + 333, + 527, + 339 + ], + "type": "text", + "content": "(f5) Get Latte Coffee" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 63, + 396, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 396, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 63, + 396, + 295, + 431 + ], + "type": "text", + "content": "models, the evaluation focuses on the model's generalization ability across multiple dimensions. In the standard evaluation, we have set up six evaluation tracks." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 64, + 432, + 295, + 670 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 64, + 432, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 432, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 64, + 432, + 295, + 479 + ], + "type": "text", + "content": "- Track 1: Rapid downstream task transferability and skill acquisition ability. The evaluation episodes use objects seen in the training set, with minimal differences." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 64, + 480, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 480, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 64, + 480, + 295, + 515 + ], + "type": "text", + "content": "- Track 2: Cross-category visual generalization. The evaluation episodes use unseen categories of objects as the target entity for evaluation." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 64, + 516, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 516, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 64, + 516, + 295, + 563 + ], + "type": "text", + "content": "- Track 3: Common sense application. Building upon Track 1, replace parts of the instructions that directly reference the target object with content requiring common sense reasoning in evaluation episodes." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 64, + 563, + 295, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 563, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 64, + 563, + 295, + 599 + ], + "type": "text", + "content": "- Track 4: Semantically rich instruction adaptation. Building upon Track 1, replace the instructions with unseen and more complex ones." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 64, + 600, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 600, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 64, + 600, + 295, + 635 + ], + "type": "text", + "content": "- Track 5: Cross-domain behavior transferability. The evaluation tasks are replaced with ones that differ from those in the training set, but require similar actions." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 64, + 635, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 635, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 64, + 635, + 295, + 670 + ], + "type": "text", + "content": "- Track 6: Long-horizon task learning. Training and evaluation using more challenging composite tasks require additional skills and reasoning." + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 671, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 706 + ], + "type": "text", + "content": "- Zero-shot Transfer Ability of Heuristic Workflow. Training-free workflow methods are evaluated under sets of tasks in all dimensions mentioned in Section 3.1. We" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 321, + 396, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 396, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 321, + 396, + 553, + 443 + ], + "type": "text", + "content": "also extend the evaluation to cover various skills and long-horizon tasks to assess the overall capability and execution robustness of the workflow. All agents are evaluated in a zero-shot setting." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 314, + 444, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 444, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 314, + 444, + 554, + 563 + ], + "type": "text", + "content": "- Comprehensive Evaluation of VLMs' Capabilities. Similar to heuristic workflows, the evaluation of VLMs is also comprehensive. Since VLMs lack intrinsic action capabilities, we organized a skill library and integrated it into a domain-specific language (DSL) [48, 54], leveraging annotated asset information as prior knowledge. This DSL functions as a straightforward API that VLMs can call, enabling efficient interaction. The evaluation pipeline is shown as Figure 3 and the details will be discussed in Section 9.3." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 647 + ], + "type": "text", + "content": "Metric. In addition to the success rate (SR), considering the long-horizon nature and high difficulty level of our tasks, we introduce the intention score (IS) and progress score (PS) for more granular evaluation. Intention score is a useful soft metric measuring the correctness trend in selecting the target object for the operation especially when sub-tasks are not yet completed. The discrete version metric is:" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 367, + 653, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 653, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 367, + 653, + 553, + 685 + ], + "type": "interline_equation", + "content": "I S = \\left\\{ \\begin{array}{l l} 1, & \\text {i f d i s t a n c e} _ {\\min } < \\epsilon \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {1}", + "image_path": "ab54cf1609b7add71134bfff17179842f28dc8dff2cf3710b2c940a30d1838aa.jpg" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "distance_{min}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is the minimum distance between the end-effector's position and the target entity, and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is empiri" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11146" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 225, + 218 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 225, + 218 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 225, + 218 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 225, + 218 + ], + "type": "image", + "image_path": "d9d822663fbed6d68fce1a5811f581e9a193f3368dc74a27d7bdef3ddf74f171.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 293 + ], + "type": "text", + "content": "Figure 3. Evaluation pipeline for VLMs. Step 1: Sample the required four-view images, as well as those segmented with numerical information, from the simulation. Meanwhile, save the corresponding instructions and operation sequences. Step 2: Input the original and annotated images and the instructions into the model. Then, obtain the model's output and extract the generated operation sequence. Step 3: The two operation sequences are evaluated using four metrics, which are weighted and summed to produce the final score. The red part represents the error output of the model, where the red solid arrows represent the dependencies generated by the error, and red dashed arrows represent the dependencies lost by the error." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 404, + 218 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 404, + 218 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 404, + 218 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 404, + 218 + ], + "type": "image", + "image_path": "d81d82de74d5d5819fa74b0e4570f8a6d0e48fa597d833977c5a00114ce82930.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 406, + 70, + 554, + 218 + ], + "blocks": [ + { + "bbox": [ + 406, + 70, + 554, + 218 + ], + "lines": [ + { + "bbox": [ + 406, + 70, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 406, + 70, + 554, + 218 + ], + "type": "image", + "image_path": "09dfe8349fa35fd6928ebf62a8c1fb5c1e1ffd7527aa94b7e03af43aaabb7f0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 305, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 353 + ], + "type": "text", + "content": "cally set to 0.15 meters in most cases after experimentation. The progress score refers to the completion level of subtasks in a long-horizon task and serves as a softer process supervision metric compared to the success rate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 364, + 295, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 364, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 135, + 364, + 295, + 376 + ], + "type": "interline_equation", + "content": "P S = N _ {\\text {d o n e}} / N _ {\\text {s u b}} \\tag {2}", + "image_path": "c40559c1f4f438708ecfc684f6fa3573cf9fe5f545ee0d12ed98ab4589aff726.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "inline_equation", + "content": "N_{done}" + }, + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "text", + "content": " is the number of completed sub-tasks, and " + }, + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "inline_equation", + "content": "N_{sub}" + }, + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "text", + "content": " is the total number of task sub-tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 420, + 178, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 178, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 178, + 430 + ], + "type": "text", + "content": "3.4. Dataset Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 664 + ], + "type": "text", + "content": "Trajectory Generation. As human teleoperation is time-consuming and not scalable [38, 47], we developed an efficient, scalable automated data collection pipeline based on our custom skill library. Inspired by [20], our data collection framework leverages the prior information including point clouds of the environment, annotated entities' grasppoints, target entity at the current step, etc. The data collection framework includes multiple task-specific motion planners. These motion planners call upon the skills in the skill library based on the current task progress and determine parameters by incorporating prior information. Subsequently, the selected skills generate trajectories using RRT [29], with quaternion interpolation achieved through spherical linear interpolation. The final trajectory is smoothed using a Bezier curve to optimize path quality. To enhance sample efficiency, reject sampling and failure-triggered early termination are applied. We demonstrate the superiority of our data framework through comparative experiments. Detailed analysis can be found in Section 8.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Data Diversity and Quality. During the data construction process, we introduced diverse task variants and domain randomization across different episodes of the same task to ensure the diversity of the training data, as discussed" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 305, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 363 + ], + "type": "text", + "content": "in Sections 6.1 and 6.4. Following the approach of previous benchmarks built on Mujoco [38, 47], the dataset is stored in the same format, with similar visual rendering quality and trajectory accuracy. We further analyze the data quality in Section 8.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 425 + ], + "type": "text", + "content": "Instruction Augmentation. We use GPT-4o [1] to generate descriptions that incorporate target-specific characteristics and interactive instructions that encompass a variety of contexts and intentions. Section 8.4 provides more details of the instruction generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 394, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 394, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 394, + 450 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 457, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 517 + ], + "type": "text", + "content": "Following Section 3.3, we conducted experiments centered on pre-trained VLA models, workflows incorporating multiple algorithmic modules, and various VLMs. The remainder of this section provides a detailed description of the experimental setup." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 484, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 484, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 484, + 538 + ], + "type": "text", + "content": "4.1. Generalization Ability of VLAs" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 579 + ], + "type": "text", + "content": "Pretrained VLAs are expected to possess robust generalization and versatility similar to LLMs. Experiments about are set to address the following research questions:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 580, + 553, + 700 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 313, + 580, + 553, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 603 + ], + "type": "text", + "content": "Q1: Do pre-trained VLAs exhibit stronger general abilities with unseen categories of objects?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 553, + 628 + ], + "type": "text", + "content": "Q2: Can pre-trained VLAs transfer their general knowledge and behavioral abilities to similar but unseen tasks?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 628, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 652 + ], + "type": "text", + "content": "Q3: Can pre-trained VLAs understand natural user interactions and implicit goal requirements?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 653, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 676 + ], + "type": "text", + "content": "Q4: Do pre-trained VLAs have the potential to transfer their world knowledge to related tasks?" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "text", + "content": "Q5: Can existing VLA architectures accurately support the completion of long-horizon tasks?" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Experiment Setup. To investigate the questions outlined" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11147" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 144 + ], + "type": "table", + "html": "
ModelTrack 1Track 2Track 3Track 4Track 5Track 6
ISPSSRISPSSRISPSSRISPSSRISPSSRISPSSR
Octo[57]32.05.62.020.42.00.022.02.81.226.84.01.213.60.00.017.23.00.0
OpenVLA[30]50.013.26.435.67.03.238.08.03.238.09.22.816.00.00.028.89.60.4
RDT-1B[43]50.014.36.039.28.23.236.07.24.035.68.04.420.40.00.021.69.20.0
π0-Base[3]59.629.313.249.621.46.060.420.05.260.419.33.622.00.00.034.412.82.0
π0-Fast[3]59.925.59.247.616.52.450.015.43.652.017.26.825.60.00.034.014.41.6
", + "image_path": "f38dfb7afede4937982dbc97f954a8c7a5c9798c63cb48f27537dddc2dd7a079.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 152, + 550, + 163 + ], + "lines": [ + { + "bbox": [ + 59, + 152, + 550, + 163 + ], + "spans": [ + { + "bbox": [ + 59, + 152, + 550, + 163 + ], + "type": "text", + "content": "Table 2. Overall experiment result of 6 evaluation tracks of fine-tuned VLAs. The detailed result of each task is reported in Table 9." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 175, + 294, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 294, + 341 + ], + "type": "text", + "content": "above, we fine-tuned various pre-trained VLA architectures [3, 30, 43, 57] on our high-quality dataset, and evaluated them on different tracks mentioned in Section 3.3 to assess their performance across a range of tasks and conditions. Considering the high task difficulty, we selected five primitive tasks for experiments in tracks 1 to 5, with 200 data samples per task, and trained the models following a multitask training approach, following the setting in previous work [3, 43]. We then assessed the models across 50 episodes for each task, resulting in a total of 250 trials per track. For track 6, we selected five composite tasks, keeping the other settings consistent. Section 9.1 provides a complete list of tasks in different tracks and details of the model fine-tuning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 294, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 294, + 413 + ], + "type": "text", + "content": "Result and Analysis. Table 2 presents the evaluation results across multiple tracks. These models perform poorly on most of tasks, as they are trained on a more diverse and limited set of fine-tuning data, with a higher difficulty level. Through the experimental results of each track and comparisons across different tracks, our analysis is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 415, + 294, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 294, + 510 + ], + "type": "text", + "content": "- The primary cause of poor performance is the insufficient ability for rapid transfer in downstream tasks. Episodes in Tracks 1 and 6 closely resemble the training set, primarily assessing the model's basic fitting ability. Despite this, the models still perform poorly on certain tasks. We believe the main reason for the failure of most generalization experiments is the inaccuracy in basic action generation, especially Octo." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 294, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 294, + 594 + ], + "type": "text", + "content": "- Pre-trained VLA models exhibit some ability to handle unseen object categories. By comparing the results of Track 1 and Track 2, we observe that although all models experience a performance decline in Track 2, they are still able to complete a small subset of tasks. This is likely due to the benefits of large-scale robotics data pretraining or even earlier VLM pretraining." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "text", + "content": "- Performance decline with unseen semantic language. By combining the experimental results of Track 3 and Track 4, we find that the model's language understanding ability has deteriorated by a larger margin. Among these, the performance decline is more pronounced in models based on non-VLM architectures: RDT's progress score decreases by " + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "inline_equation", + "content": "45.2\\%" + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "text", + "content": " while OpenVLA and " + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "text", + "content": " only decreases by " + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "inline_equation", + "content": "31.3\\%" + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "inline_equation", + "content": "34.2\\%" + }, + { + "bbox": [ + 55, + 594, + 294, + 689 + ], + "type": "text", + "content": ", respectively, in Track 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "- Nearly absent cross-domain transferability. We found that all the models fail to perform similar but unseen tasks," + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 175, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 175, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 322, + 175, + 553, + 209 + ], + "type": "text", + "content": "due to significant visual and linguistic input differences. These policies exhibit uncertainty in behavior when faced with unfamiliar tasks, resulting in IS that are similar." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 212, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 295 + ], + "type": "text", + "content": "We further analyze that data scaling and pre-training contribute positively to the model's performance Through basic ablation experiments. We also discuss in detail the potential issues with current VLAs, such as multimodal data co-training and model architecture designs. These discussions are presented in Section 10.1. We also discuss the gap between real and simulated environments in Section 10.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 460, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 460, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 460, + 320 + ], + "type": "text", + "content": "4.2. Zero-shot Ability of Agent" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 327, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 469 + ], + "type": "text", + "content": "For our evaluation of foundation model-based algorithms, we reviewed two state-of-the-art frameworks, Voxposer [25] and CoPA [24], and the comparison results are shown in Figure 4. Given Voxposer's dependence on large language models (LLMs), we assessed Voxposer's performance with and without visual perception capabilities. While Voxposer performed adequately on basic tasks and achieved the Progress Scores of 30-40, its reliance on LLM-driven motion planning often led to grasping failures due to limited information for effective grasp planning, especially when interpreting rotation in non-visual contexts, resulting in low overall scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 555 + ], + "type": "text", + "content": "Interestingly, the foundational LLM alone maintained relatively stable scores in semantic understanding and reasoning tasks without visual input. However, adding visual perception slightly reduced performance in these areas while significantly improving spatial reasoning, where LLM-only setups struggle with spatial accuracy due to lack of spatial information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 712 + ], + "type": "text", + "content": "The lack of closed-loop feedback limits these models' ability to perform physical reasoning tasks, particularly those involving dynamic interactions, leading to lower scores in this dimension. Both models struggle with high-complexity tasks, succeeding mainly in entity recognition but performing poorly in long horizon task reasonable breaking down. This finding underscores the requirements for advancing foundational model-based frameworks to address complex reasoning. In fact, although the methods mentioned above emphasize their zero-shot capabilities and generalization to new scenarios, their modular design often limits the upper bound of their performance. A more detailed discussion of this can be found in Section 10.3." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11148" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 294, + 273 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 294, + 273 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 294, + 273 + ], + "type": "image", + "image_path": "ab42e09e91e4719e32faa7ee06b77fc25cb63c113ca620233be90b9cc5651c74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 284, + 298, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 284, + 298, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 298, + 329 + ], + "type": "text", + "content": "Figure 4. Evaluation progress score for Voxposer and CoPA. Voxposer w/o refers to the version without visual perception, where ground truth labels are directly provided for object selection. Voxposer w uses GPT-4V as the visual perception module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 231, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 231, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 231, + 351 + ], + "type": "text", + "content": "4.3. Comprehensive Ability of VLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 464 + ], + "type": "text", + "content": "We referred to the evaluation results of multiple series of VLMs provided by OpenCompass [11] and selected models from different families [1, 2, 7, 19, 22, 41] with strong overall performance as shown in Figure 5. We evaluate the comprehensive performance of these models with the dataset derived from naturally self-contained information within a simulated environment. This dataset consists of a complex set of tasks designed to assess the VLM's ability to perceive visual stimuli and comprehend verbal instructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 465, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 514 + ], + "type": "text", + "content": "In line with the evaluation framework outlined in the Section 3.3, we evaluate the models across six dimensions of task performance. The results for each VLM model under the 1-shot setting are summarized in Figure 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 682 + ], + "type": "text", + "content": "Although these VLMs perform well on most multimodal tasks and even some embodied tasks [11, 33], their performance, including that of GPT-4o, falls short when faced with more complex scenarios, instructions, and more challenging tasks. We surprisingly find that the open-source model Qwen2-VL-7B-Instruct performed competitively, surpassing GPT-4o on certain dimensions. However, all models struggled with complex tasks, especially those requiring long-term task decomposition and logic reasoning. Only GPT-4o achieved a score in the reasoning dimension comparable to those in other dimensions, while the other models scored around 20. Besides, performance declines significantly when linguistic instructions transition from direct semantics to abstract meanings, as shown in the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 552, + 273 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 552, + 273 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 273 + ], + "type": "image", + "image_path": "e6d6e90334577f95dcf7c685ff912e3eb229070e26649f3421df62328a190dfd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 284, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 312, + 284, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 555, + 328 + ], + "type": "text", + "content": "Figure 5. Radar charts depicting the performance of all VLM models across six dimensions. The reason why only GLM-4V-9B is evaluated in a zero-shot setting is that it does not support multigraph inference, which is required for the other models." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 339, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 471 + ], + "type": "text", + "content": "semantic dimension. Different models appear to have distinct areas of expertise, e.g. LLaVA-NeXT exhibits weaker spatial perception ability, GLM-4V-9B excels in spatial and even physical law dimensions but lag in semantic comprehension. In Section 10.4, we conducted additional ablation experiments focusing on CoT and few-shot learning, followed by further analysis. Detailed data and error analysis can be found in Table 10. Overall, while the models demonstrate promising capabilities, their understanding and planning in embodied environments remain limited, highlighting the need for further advancements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 388, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 388, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 388, + 497 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 504, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 713 + ], + "type": "text", + "content": "We propose VLABench, a large-scale benchmark designed for tasks with long-horizon and multi-dimensional reasoning. Such reasoning involves many dimensions, including implicit semantic goal extracting, logical reasoning, and long-horizon planning. The most important thing we do is to provide a positive definition of the capabilities that intelligent agents with cognitive abilities should possess and the tasks they should be able to perform, through the provision of 100 standardized tasks. Additionally, VLABench constructed a scalable automated data collection framework for future's potential larger scale pertaining and a standardized dataset for fair comparison of VLAs, both in the present and in future developments. Our diverse and multiple experiments revealed that current VLAs and VLMs face significant challenges in our tasks, and there remain substantial uncertainties in research on robotics scaling. We hope that VLABench will inspire both the future research on robotics pertaining recipe and promote more robust VLA architectures development." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 297, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 297, + 714 + ], + "type": "text", + "content": "3As GLM-4V-9B does not support multiple image inputs, the 0-shot method is employed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11149" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 97, + 115, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 115, + 110 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 115, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 118, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 172 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 295, + 218 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 274 + ], + "type": "text", + "content": "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi_0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3, 7, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 277, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 294, + 331 + ], + "type": "text", + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 334, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 295, + 518 + ], + "type": "text", + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alex Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In arXiv preprint arXiv:2307.15818, 2023. 2, 3, 10" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 521, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 294, + 543 + ], + "type": "text", + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 545, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 294, + 610 + ], + "type": "text", + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 656 + ], + "type": "text", + "content": "[8] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 317, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 554, + 127 + ], + "type": "text", + "content": "[10] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 28" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 130, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 554, + 172 + ], + "type": "text", + "content": "[11] OpenCompass Contributors. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass, 2023.8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 175, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 555, + 240 + ], + "type": "text", + "content": "[12] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 243, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 554, + 274 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "type": "text", + "content": "[14] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 333, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 554, + 386 + ], + "type": "text", + "content": "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 388, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 554, + 442 + ], + "type": "text", + "content": "[16] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Grasynet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11444–11453, 2020. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 445, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 487 + ], + "type": "text", + "content": "[17] Hao-Shu Fang, Minghao Gou, Chenxi Wang, and Cewu Lu. Robust grasping across diverse sensor qualities: The graspnet-1billion dataset. The International Journal of Robotics Research, 2023. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 491, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 554, + 533 + ], + "type": "text", + "content": "[18] Jensen Gao, Suneel Belkhale, Sudeep Dasari, Ashwin Balakrishna, Dhruv Shah, and Dorsa Sadigh. A taxonomy for evaluating generalist robot policies. arXiv preprint arXiv:2503.01238, 2025. 10" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 536, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 536, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 536, + 554, + 589 + ], + "type": "text", + "content": "[19] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 592, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 554, + 668 + ], + "type": "text", + "content": "[20] Ran Gong, Jiangyong Huang, Yizhou Zhao, Haoran Geng, Xiaofeng Gao, Qingyang Wu, Wensi Ai, Ziheng Zhou, Demetri Terzopoulos, Song-Chun Zhu, et al. Arnold: A benchmark for language-grounded task learning with continuous states in realistic 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20483-20495, 2023. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "text", + "content": "[21] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11150" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "generalizable manipulation skills. In International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[22] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[23] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 241 + ], + "type": "text", + "content": "[24] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 286 + ], + "type": "text", + "content": "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "text", + "content": "[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[27] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[28] Sertac Karaman and Emilio Frazzoli. Sampling-based algorithms for optimal motion planning. The international journal of robotics research, 30(7):846-894, 2011. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 456 + ], + "type": "text", + "content": "[29] Sertac Karaman, Matthew R Walter, Alejandro Perez, Emilio Frazzoli, and Seth Teller. Anytime motion planning using the rrt. In 2011 IEEE international conference on robotics and automation, pages 1478-1483. IEEE, 2011. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "text", + "content": "[30] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3, 7, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 567 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 635 + ], + "type": "text", + "content": "[32] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, et al. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. In Conference on Robot Learning, pages 80–93. PMLR, 2023. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[33] Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Srivastava, Cem Gokmen, Tony Lee, Li Erran Li, Ruohan Zhang, et al. Embodied agent interface: Benchmarking llms for embodied decision making. arXiv preprint arXiv:2410.07166, 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[34] Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "text", + "content": "Yizhong Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "type": "text", + "content": "[35] Xuanlin Li, Kyle Hsu, Jiayuan Gu, Karl Pertsch, Oier Mees, Homer Rich Walke, Chuyuan Fu, Ishikaa Lunawat, Isabel Sieh, Sean Kirmani, Sergey Levine, Jiajun Wu, Chelsea Finn, Hao Su, Quan Vuong, and Ted Xiao. Evaluating real-world robot manipulation policies in simulation. arXiv preprint arXiv:2405.05941, 2024. 2, 3, 5, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[36] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "text", + "content": "[37] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 330 + ], + "type": "text", + "content": "[38] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[39] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 409 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 443 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "text", + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[43] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 3, 7, 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "text", + "content": "[44] Ajay Mandlekar, Soroush Nasiriany, Bowen Wen, Iretiayo Akinola, Yashraj Narang, Linxi Fan, Yuke Zhu, and Dieter Fox. Mimicgen: A data generation system for scalable robot learning using human demonstrations. arXiv preprint arXiv:2310.17596, 2023. 4, 11" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[45] Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters, 7(3): 7327-7334, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[46] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[47] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11151" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "and Yuke Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. In Robotics: Science and Systems, 2024. 3, 6, 4, 10, 11" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[48] Arne Nordmann, Nico Hochgeschwender, and Sebastian Wrede. A survey on domain-specific languages in robotics. In International conference on simulation, modeling, and programming for autonomous robots, pages 195-206. Springer, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 219 + ], + "type": "text", + "content": "[49] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 275 + ], + "type": "text", + "content": "[50] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 277, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 331 + ], + "type": "text", + "content": "[51] Tim Pearce, Tabish Rashid, Anssi Kanervisto, Dave Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan, Ida Momennejad, Katja Hofmann, et al. Imitating human behaviour with diffusion models. arXiv preprint arXiv:2301.10677, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 333, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 296, + 399 + ], + "type": "text", + "content": "[52] Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10740-10749, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 400, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 296, + 487 + ], + "type": "text", + "content": "[53] Tianxiang Sun, Xiaotian Zhang, Zhengfu He, Peng Li, Qinyuan Cheng, Xiangyang Liu, Hang Yan, Yunfan Shao, Qiong Tang, Shiduo Zhang, Xingjian Zhao, Ke Chen, Yining Zheng, Zhejian Zhou, Ruixiao Li, Jun Zhan, Yunhua Zhou, Linyang Li, Xiaogui Yang, Lingling Wu, Zhangyue Yin, Xuanjing Huang, Yu-Gang Jiang, and Xipeng Qiu. Moss: An open conversational large language model. Machine Intelligence Research, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 489, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 533 + ], + "type": "text", + "content": "[54] Craig J Sutherland and Bruce MacDonald. Robolang: a simple domain specific language to script robot interactions. In 2019 16th International Conference on Ubiquitous Robots (UR), pages 265-270. IEEE, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 535, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 600 + ], + "type": "text", + "content": "[55] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in neural information processing systems, 34:251-266, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 602, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 678 + ], + "type": "text", + "content": "[56] Stone Tao, Fanbo Xiang, Arth Shukla, Yuzhe Qin, Xander Hinrichsen, Xiaodi Yuan, Chen Bao, Xinsong Lin, Yulin Liu, Tse kai Chan, Yuan Gao, Xuanlin Li, Tongzhou Mu, Nan Xiao, Arnav Gurha, Zhiao Huang, Roberto Calandra, Rui Chen, Shan Luo, and Hao Su. Maniskill3: GPU parallelized robotics simulation and rendering for generalizable embodied ai. arXiv preprint arXiv:2410.00425, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo:" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 597 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[58] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026-5033. IEEE, 2012. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 141, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 195 + ], + "type": "text", + "content": "[59] Saran Tunyasuvunakool, Alistair Muldal, Yotam Doron, Siqi Liu, Steven Bohez, Josh Merel, Tom Erez, Timothy Lillicrap, Nicolas Heess, and Yuval Tassa. dm_control: Software and tasks for continuous control. Software Impacts, 6:100022, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 197, + 554, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 261 + ], + "type": "text", + "content": "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 263, + 554, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 554, + 340 + ], + "type": "text", + "content": "[61] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. Survey Certification. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 342, + 554, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 396 + ], + "type": "text", + "content": "[62] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 397, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 554, + 453 + ], + "type": "text", + "content": "[63] Sriram Yenamandra, Arun Ramachandran, Karmesh Yadav, Austin Wang, Mukul Khanna, Theophile Gervet, Tsung-Yen Yang, Vidhi Jain, Alexander William Clegg, John Turner, et al. Homerobot: Open-vocabulary mobile manipulation. arXiv preprint arXiv:2306.11565, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 453, + 554, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 554, + 485 + ], + "type": "text", + "content": "[64] Kevin Zakka. obj2mjcf: Cli for processing composite wavefront obj files for use in mujoco. https://github.com/kevinzakka/obj2mjcf, 2023. Version 1.0.4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 487, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 554, + 553 + ], + "type": "text", + "content": "[65] Yihan Zeng, Chenhan Jiang, Jiageng Mao, Jianhua Han, Chaoqiang Ye, Qingqu Huang, Dit-Yan Yeung, Zhen Yang, Xiaodan Liang, and Hang Xu. Clip2: Contrastive language-image-point pretraining from real-world point cloud data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15244-15253, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 554, + 554, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 554, + 597 + ], + "type": "text", + "content": "[66] Liming Zheng, Feng Yan, Fanfan Liu, Chengjian Feng, Zhuoliang Kang, and Lin Ma. Robocas: A benchmark for robotic manipulation in complex object arrangement scenarios. arXiv preprint arXiv:2407.06951, 2024. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11152" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_content_list.json b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..78ee19b4b69fe4f50e6dc54d35a6df2709b6bef3 --- /dev/null +++ b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_content_list.json @@ -0,0 +1,1496 @@ +[ + { + "type": "text", + "text": "VLDrive: Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruifei Zhang $^{1,2}$ , Wei Zhang $^{4}$ , Xiao Tan $^{4}$ , Sibei Yang $^{3}$ , Xiang Wan $^{2}$ , Xiaonan Luo $^{5}$ , Guanbin Li $^{3,6*}$ $^{1}$ The Chinese University of Hong Kong, Shenzhen $^{2}$ Shenzhen Research Institute of Big Data \n $^{3}$ Sun Yat-sen University $^{4}$ Baidu Inc. $^{5}$ Guilin University of Electronic Technology \n $^{6}$ Guangdong Key Laboratory of Big Data Analysis and Processing", + "bbox": [ + 114, + 202, + 887, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in language-grounded autonomous driving have been significantly promoted by the sophisticated cognition and reasoning capabilities of large language models (LLMs). However, current LLM-based approaches encounter critical challenges: (1) Failure analysis reveals that frequent collisions and obstructions, stemming from limitations in visual representations, remain primary obstacles to robust driving performance. (2) The substantial parameters of LLMs pose considerable deployment hurdles. To address these limitations, we introduce VLDrive, a novel approach featuring a lightweight MLLM architecture with enhanced vision components. VLDrive achieves compact visual tokens through innovative strategies, including cycle-consistent dynamic visual pruning and memory-enhanced feature aggregation. Furthermore, we propose a distance-decoupled instruction attention mechanism to improve joint visual-linguistic feature learning, particularly for long-range visual tokens. Extensive experiments conducted in the CARLA simulator demonstrate VLDrive's effectiveness. Notably, VLDrive achieves state-of-the-art driving performance while reducing parameters by $81\\%$ (from 7B to 1.3B), yielding substantial driving score improvements of $15.4\\%$ , $16.8\\%$ , and $7.6\\%$ at tiny, short, and long distances, respectively, in closed-loop evaluations. Code is available at https://github.com/ReaFly/VLDrive.", + "bbox": [ + 89, + 340, + 483, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving has experienced significant development in recent years, evolving from modular and segregated design to integrated end-to-end networks [10, 15, 26, 29, 46]. However, these networks still rely on specific inputs such as target points or action commands to guide their driving behaviors, significantly constraining the interaction", + "bbox": [ + 89, + 786, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91a5ebeb0cdffd587f73890074dcde367cf1bcfbe283f9832cd91c0018624c7d.jpg", + "image_caption": [ + "Figure 1. (a) Existing methods for language-grounded driving. (b) Our proposed VLDrive: a novel framework featuring a lightweight MLLM architecture with enhanced vision components. (c) Driving failure analysis of existing methods based on three evaluation runs. (d) Performance comparison between VLDrive and both versions of LMDrive, highlighting our method's superior driving performance with fewer parameters." + ], + "image_footnote": [], + "bbox": [ + 519, + 308, + 715, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14a09f89cdb435b1ef70942c80fd7c78129a403c0bd80dfe5fe92f722a781d4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 308, + 897, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with humans and applicability in real-world scenarios.", + "bbox": [ + 511, + 613, + 872, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of large language models (LLMs) [18, 33, 49, 50] has catalyzed a revolution in autonomous driving. Impressed by their advanced cognition and logical reasoning capabilities, a multitude of studies have integrated LLMs into autonomous driving systems and achieved noteworthy results [28, 37, 38, 41]. Among them, LMDrive [28] successfully achieves human-friendly language-grounded autonomous driving, where vehicle behaviors are solely guided by natural language instructions (see Fig. 1(a)), significantly enhancing human-vehicle interaction.", + "bbox": [ + 511, + 628, + 905, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, despite remarkable advancements, these approaches encounter significant practical deployment challenges, characterized by suboptimal driving performance and excessive computational parameters. To address these issues, we conduct a thorough investigation of driving failure cases in current approaches, coupled with performance analysis across different LLM scales, yielding two critical insights: 1) Deficiencies in visual comprehension", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: liguanbin@mail.sysu.edu.cn", + "bbox": [ + 107, + 886, + 390, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5923", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are the predominant cause of driving failures. We analyze the failure cases of existing language-grounded autonomous driving models and identify two primary causes: instruction misunderstanding, manifested as the agent deviating from the route; and visual misunderstanding, characterized by failures due to collisions or blocks involving vehicles or layouts. The results in Fig. 1(c) reveal that limitations in visual comprehension frequently emerge as a critical factor contributing to driving task failures. 2) Driving performance does not scale linearly with LLM's parameter count. Through comprehensive evaluations across three benchmarks, we compare the full-scale LMDrive (7B parameters) with its lightweight counterpart, LMDrive-LiteLM (1.3B trainable parameters). As demonstrated in Fig. 1(d), LMDrive-LiteLM achieves comparable driving performance despite its significantly reduced parameter count. These findings strongly suggest that massive parameter scales in LLMs may not be a critical prerequisite for achieving superior driving performance. The insights derived from these analyses illuminate the path toward developing a parameter-efficient yet high-performing alternative for language-grounded driving applications.", + "bbox": [ + 89, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these considerations, in our work, we propose VLDrive, a vision-augmented lightweight language model for efficient language-grounded autonomous driving, as showcased in Fig. 1(b). We enhance the model's visual comprehension by drawing inspiration from human driving competencies, focusing on two key aspects: prioritized attention allocation to sparse but critical agents and objects, and temporal trajectory memory for behavioral sequence prediction. These enhancements are achieved through our proposed Cycle-consistent Dynamic Visual Pruning (CCDP) and Memory-enhanced Feature Aggregation (MEFA) mechanisms. Specifically, CDP framework consists of token sparsification and training-only token reconstruction. The sparsification mechanism enables adaptive selection of critical visual tokens and dynamic adjustment of token count based on driving scenario complexity. The reconstruction process preserves information integrity by explicitly recovering features of pruned tokens, enabling robust visual perception while minimizing computational overhead. Additionally, MEFA incorporates a memory bank of adjacent frames to capture temporal dependencies, providing critical temporal cues for visual feature enhancement.", + "bbox": [ + 91, + 428, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we introduce a Distance-decoupled Instruction Attention (DDIA) mechanism to enhance instruction comprehension capabilities of lightweight language models. This design stems from our observation that increasing numbers of long-range visual tokens may dilute instruction attention (see Fig. 6), potentially leading to instruction misunderstanding. DDIA alleviates this issue and enhances joint visual-linguistic feature learning", + "bbox": [ + 89, + 780, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and alignment, resulting in more accurate trajectory planning. Extensive closed-loop experiments conducted on the CARLA [6] simulation platform demonstrate the effectiveness of our proposed method, surpassing state-of-the-art approaches by $15.4\\%$ , $16.8\\%$ and $7.6\\%$ in driving scores at tiny, short and long distances, respectively.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows:", + "bbox": [ + 529, + 183, + 846, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify that visual comprehension deficiencies constitute the primary bottleneck in language-grounded driving, with performance scaling non-linearly with language model size. This observation motivates our proposed paradigm: a lightweight language model enhanced by vision-augmented strategies, which simultaneously reduces computational overhead and improves driving performance, facilitating practical deployment.", + "- We propose VLDrive, a lightweight architecture enhanced with vision-centric strategies for language-grounded driving. The framework incorporates CCDP for adaptive visual signal extraction and MEFA for temporal information integration. Additionally, DDIA facilitates visual-language alignment for robust navigation instruction following. These complementary strategies collectively enable more reliable autonomous driving.", + "- We evaluate our approach through closed-loop simulation experiments on standard language-grounded driving benchmarks using the CARLA platform. VLDrive achieves state-of-the-art performance while using significantly fewer parameters." + ], + "bbox": [ + 511, + 199, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 531, + 653, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. End-to-End Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 556, + 805, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early end-to-end autonomous driving methods can be mainly divided into two categories: imitation learning [4, 8, 42] and reinforcement learning [2, 16, 32, 47]. Subsequently, benefiting from the robust feature modeling capabilities of the Transformer [36], researchers introduced additional modalities to provide complementary information for autonomous driving [24, 29, 45]. For example, Transfuse [24] uses Transformer to merge features from RGB and LiDAR bird's eye view (BEV) images to enhance the global comprehension of 3D scenes. InterFuse [29] improves the security and interpretability of the model by directly presenting its intermediate features and limiting actions to specified safe sets. In recent years, modular end-to-end planning approaches [1, 10, 26] have gained increased attention, where all components are connected and jointly optimized towards the ultimate goal, further improving the driving performance.", + "bbox": [ + 511, + 579, + 906, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. LLMs in Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 847, + 782, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As LLMs continue to evolve and flourish, an increasing number of researchers are incorporating them into au", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tonomous driving tasks, broadening the scope and capabilities of these systems. Among them, a notable strand of research [22, 23, 30] employs LLMs for driving-related visual question answering, providing enhanced interpretability for driving decisions. Another line of work seeks to fully exploit the powerful logical reasoning capabilities of LLMs to fundamentally transform the research paradigm in autonomous driving tasks [5, 7, 20, 21]. Specifically, GPT-Driver [20] feeds LLM with observations and egostates as language prompts and generates a planned trajectory and corresponding decision-making process in a natural language format. Agent-Driver [21] builds an LMM-powered agent, capitalizing on a tool library, a cognitive memory, and a reasoning engine for human-like autonomous driving. Recently, LLMs have been extended to visual modality successfully, making it possible for LLMs to understand the visual information of autonomous driving. Chen et al. [3] introduces a multimodal LLM to process vectorized numeric modalities data and produce driving-related question answers as well as action decisions. DriveGPT4 [41] curates a visual instruction tuning dataset for interpretable autonomous driving, taming LLM to generate text responses and low-level control signals. Further, based on the CARLA simulation platform, many studies [28, 37, 38] delve deeper and perform closed-loop evaluation driven by LLMs directly. For instance, LMDrive [28] takes multi-modality data and navigation instructions as input, achieving language-grounded driving and enhanced interaction with humans. Despite the impressive performance powered by LLMs, their considerably heavy parameters pose significant deployment challenges. In this work, we demonstrate that a lightweight language model can effectively respond to instructions and execute driving tasks through efficient visual representation extraction and strengthened vision-language alignment.", + "bbox": [ + 93, + 90, + 480, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Token Reduction", + "text_level": 1, + "bbox": [ + 94, + 635, + 254, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer has achieved substantial performance improvements across a range of tasks, yet it faces criticism for its quadratic computational costs relative to the number of input tokens. To tackle this issue, numerous studies focus on reducing token counts [12, 25, 39, 40]. For instance, DynamicViT [25] adaptively prunes redundant tokens at different Transformer layers to achieve token sparsification. SPViT [12] introduces an attention-based multihead token selector with a soft pruning strategy. Unlike direct pruning methods, this approach consolidates selected redundant tokens into a single package token, enhancing efficiency while preserving essential information. In the LLM era, token reduction is also a crucial strategy to reduce the heavy computation costs [13, 14, 27, 50]. Most existing works utilize Q-former [13] to aggregate visual features into a fixed number of tokens [13, 43, 50]. LLaMa", + "bbox": [ + 93, + 659, + 480, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VID [14] leverages average pooling followed by a learnable projector to compress the video frame representation. LLaVA-PruMerge [27] proposes an adaptive token selection via outlier detection, paired with a cluster-based merging strategy to supplement the information of pruned tokens to kept ones. Distinct from prior approaches, we present a cycle-consistent dynamic pruning strategy that integrates token reconstruction with token sparsification at the training phase. This design not only enables more representative token selection but also ensures that the retained tokens aggregate fine-grained global visual information.", + "bbox": [ + 516, + 90, + 903, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 516, + 273, + 602, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 516, + 297, + 620, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This task aims to predict driving actions by leveraging sequences of multi-sensor data and corresponding human navigation instructions. As shown in Fig. 2, our method follows the common framework of a multi-modal large language model (MLLM) and comprises three main components:", + "bbox": [ + 516, + 321, + 903, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Encoder: Given a sequence of visual data $\\{\\mathbf{X}_i, i = 1,2,\\dots,T\\}$ , where each frame $\\mathbf{X}_i$ includes multi-view camera images and corresponding LiDAR data. A visual encoder processes these multi-modal inputs to produce a unified feature representation $\\mathbf{F}_i \\in \\mathbb{R}^{N \\times C}$ for each frame, with $N$ representing the total number of tokens.", + "bbox": [ + 516, + 397, + 903, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Connector: As a bridge to align visual features with the language space, the connector plays a remarkable role in MLLM. To endow our model with adaptive visual perception and long-context modeling capability, as illustrated in Fig. 3, we tailor cycle-consistent dynamic visual pruning (CCDP), consisting of token sparsification and reconstruction, and memory-enhanced feature aggregation (MEFA) for the connector, significantly enhancing the visual representation. We denote the improved features of each frame by the connector as $\\mathbf{F}_i^v\\in \\mathbb{R}^{N_v\\times C_t}$ , where $C_t$ is the language model's hidden dimension.", + "bbox": [ + 516, + 488, + 903, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lite Language Model: A lightweight language model, enhanced by our proposed distance-decoupled instruction attention (DDIA), is utilized to process the combination of visual tokens of all frames, $\\mathbf{F}_v\\in \\mathbb{R}^{N_vT\\times C_t}$ and corresponding text embedding $\\mathbf{F}_t\\in \\mathbb{R}^{N_t\\times C_t}$ tokenized from navigation instructions. A following MLP leverages features from the language model and predicts the future trajectory, which is subsequently converted by PID controllers to produce the lateral steering action and the longitudinal acceleration action. We detail each component in the following subsections.", + "bbox": [ + 516, + 655, + 903, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. CCDP: Token Sparsification", + "text_level": 1, + "bbox": [ + 516, + 832, + 764, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the $i$ -th frame's multi-modal features $\\mathbf{F}_i$ , as in DynamicViT [25], we predict the probabilities for pruning and retaining each token. Specifically, the probabilities are cal", + "bbox": [ + 516, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5925", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/73a29ed2563d257a2553c1861c55906718986ce57fdaa8d012829250d576a228.jpg", + "image_caption": [ + "Figure 2. An overview of our proposed VLDrive framework. Given a sequence of visual data, our connector transforms each frame's raw visual features $\\mathbf{F}_i$ into sparse yet informative representations $\\mathbf{F}_i^v$ through two key components: CCDP: Token Sparsification and Memory-enhanced Feature Aggregation (MEFA). Subsequently, a lite language model augmented with Distance-decoupled Instruction Attention (DDIA) jointly processes tokenized navigation instructions $\\mathbf{F}_t$ and temporal visual features $\\mathbf{F}_v = \\{\\mathbf{F}_i^v \\mid i = 1, \\dots, T\\}$ . The resulting hidden representations are fed into an MLP for trajectory prediction, followed by PID controllers that translate these predictions into concrete driving actions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity of $\\mathbf{F}_i^v$ via explicit token reconstruction." + ], + "image_footnote": [], + "bbox": [ + 210, + 85, + 767, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ulated based on the combined global and local features $\\left[\\mathbf{G}_i;\\mathbf{L}_i\\right]$ , which are obtained by $\\mathrm{MLP}(\\cdot)$ operations:", + "bbox": [ + 89, + 436, + 482, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {i} = \\operatorname {M L P} \\left(\\mathbf {F} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times \\frac {C}{2}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 478, + 482, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {i} = \\operatorname {A v g} \\left(\\mathbf {L} _ {i}\\right) \\in \\mathbb {R} ^ {1 \\times \\frac {C}{2}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 507, + 480, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Avg}(\\cdot)$ is average operation. Another MLP is utilized to predict the probability $\\mathbf{S}_i$ , and $[\\cdot ;\\cdot ]$ denotes concatenation with broadcast mechanism:", + "bbox": [ + 89, + 534, + 483, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {i} = \\operatorname {S o f t m a x} \\left(\\operatorname {M L P} \\left(\\left[ \\mathbf {G} _ {i}; \\mathbf {L} _ {i} \\right]\\right)\\right) \\in \\mathbb {R} ^ {N \\times 2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 589, + 480, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our implementation, we leverage Gumbel-Softmax to obtain the binary mask:", + "bbox": [ + 89, + 618, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i} = \\text {G u m b e l - S o f t m a x} (\\mathbf {S} _ {i}) _ {* 1} \\in \\{0, 1 \\} ^ {N} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 659, + 482, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_i$ denotes the token retention masks. As shown in Fig. 3, the kept tokens $\\mathbf{F}_i^k$ selected based on $\\mathbf{M}_i$ are delivered to the Memory-enhanced Feature Aggregation module, which is detailed in the next subsection.", + "bbox": [ + 89, + 688, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Memory-enhanced Feature Aggregation", + "text_level": 1, + "bbox": [ + 89, + 757, + 434, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal reasoning is integral to our everyday driving scenarios. Human drivers can deduce the likely future direction or intentions of an approaching vehicle by analyzing its historical trajectory, and subsequently tailor their driving actions. To attain autonomous driving that matches the safety and reliability of human driving, we explicitly inject the temporal context into the current visual tokens. Specifically, for the current time step $i$ , we introduce a memory", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bank storing the adjacent $Z$ frames from the current one, denoted as $\\mathbf{B}_i = [\\mathbf{F}_{i - Z},\\mathbf{F}_{i - Z + 1},\\dots,\\mathbf{F}_{i - 1}]$ $\\mathbf{B}_i\\in \\mathbb{R}^{Z\\times N\\times C}$ Temporal encoding $\\mathbf{TE}_i$ is derived by an MLP utilizing the historical average data and current frame information:", + "bbox": [ + 511, + 436, + 906, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {i} ^ {a v g} = \\operatorname {A v g} \\left(\\mathbf {B} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times C} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 507, + 903, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T E} _ {i} = \\operatorname {M L P} \\left(\\left[ \\mathbf {F} _ {i}; \\mathbf {B} _ {i} ^ {\\text {a v g}} \\right]\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 536, + 903, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterwards, a query transformer (Q-former) [13] is introduced to aggregate the features of current frames enhanced by temporary encoding:", + "bbox": [ + 511, + 559, + 906, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {v} = \\operatorname {Q - F o r m e r} \\left(\\mathbf {F} _ {i} ^ {k}, \\mathbf {F} _ {i} + \\mathbf {T E} _ {i}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 614, + 903, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to the original Q-former, we implement two key modifications: 1) Instead of using learnable queries, we utilize the retained visual tokens $\\mathbf{F}_i^k$ as queries to probe and aggregate features. These retained tokens serve as significant anchors within the current frame, allowing for more effective feature aggregation. 2) We introduce a temporary embedding $\\mathbf{TE}_i$ to the original visual features $\\mathbf{F}_i$ , which enhances the model's ability to focus on temporary-salient moving objects.", + "bbox": [ + 511, + 643, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the generated $\\mathbf{F}_i^v$ serves as part of the aggregated visual tokens across all frames, $\\mathbf{F}_v$ , which is then fed into the subsequent language model for trajectory prediction based on the corresponding navigation instructions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity via explicit token reconstruction.", + "bbox": [ + 511, + 780, + 908, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5926", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8071276f0238c3bc1753601b9d9ae7510de61da5176c5c9d5fdcf90a838c7c22.jpg", + "image_caption": [ + "Figure 3. A detailed illustration of our proposed connector. CCDP: Token Sparsification and Memory-enhanced Feature Aggregation are proposed to reduce token count while enhancing information density. CCDP: Token Reconstruction serves as a training-only task that further ensures the information integrity of the retained tokens. Red arrows $(\\rightarrow)$ indicate the input and output paths of our connector." + ], + "image_footnote": [], + "bbox": [ + 133, + 87, + 836, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. CCDP: Token Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 378, + 354, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through memory-enhanced feature aggregation, the generated $\\mathbf{F}_i^v$ ideally encapsulates all visual features of the current frame. To further enhance the connector's ability to select critical visual tokens and aggregate complete information, we expect to reconstruct features of those pruned tokens based on $\\mathbf{F}_i^v$ . Firstly, we add the enhanced tokens $\\mathbf{F}_i^v$ into the original features $\\mathbf{F}_i$ according to their original token positions, forming a new feature vector denoted as $\\langle \\mathrm{MLP}(\\mathbf{F}_i^v), \\mathbf{F}_i \\rangle$ , where an MLP function is used to align the dimensions of $\\mathbf{F}_i^v$ and $\\mathbf{F}_i$ . Secondly, drawing inspiration from the Masked Autoencoder [9], a learnable embedding $\\mathbf{e} \\in \\mathbb{R}^{N \\times C}$ is introduced as a foundation for reconstructing the pruned tokens. Finally, we construct the input feature vector to be reconstructed through the following formula:", + "bbox": [ + 89, + 400, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i} ^ {r e c} = \\left\\langle \\operatorname {M L P} \\left(\\mathbf {F} _ {i} ^ {v}\\right), \\mathbf {F} _ {i} \\right\\rangle \\cdot \\mathbf {M} _ {i} + \\mathbf {e} \\cdot \\left(1 - \\mathbf {M} _ {i}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 619, + 480, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we incorporate the decision $\\mathbf{M}_i$ into the computational graph of feature reconstruction, enabling the reconstruction loss to influence pruning decisions and compelling the model to retain the most critical tokens. Afterwards, a 4-layer Transformer block takes as input $\\mathbf{F}_i^{rec}$ and predicts the reconstructed results $\\hat{\\mathbf{F}}_i^{rec}$ . A cycle consistency constraint (Eq. 11) is imposed to ensure that the retained tokens $\\mathbf{F}_i^v$ preserve the integral information of the current video frame.", + "bbox": [ + 89, + 643, + 483, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Distance-decoupled Instruction Attention", + "text_level": 1, + "bbox": [ + 89, + 772, + 444, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a lightweight language model for efficient autonomous driving, reducing both the number of parameters and computational resource requirements. Furthermore, distance-decoupled instruction attention (DDIA) is tailored for our life language model, significantly alleviating the attention dilution for navigation instructions in the long-context driving scenarios. Specifically, as shown in Fig. 4,", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our proposed DDIA diverges from the traditional vanilla attention in two key aspects:", + "bbox": [ + 511, + 378, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1): Causal-attention $\\rightarrow$ Self-attention: Rather than adhering to the original causal setting, our self-attention within instructional segments allows each text token to access ample contextual information, thereby producing a more comprehensive representation of text features.", + "2): Distance-dependent $\\rightarrow$ Distance-decoupled: In contemporary research on LLM, position encoding (e.g. RoPE [31]) is recognized as a critical ingredient to explicitly inject positional information into input tokens and enhance the model's contextual modeling capabilities. However, position encodings may introduce challenges, particularly in long-range driving scenarios. As historical visual data accumulates, the distance between current visual tokens and instructional tokens increases. The feature disparity introduced by position encoding may negatively impact the attention sensitivity of visual tokens to instruction ones, resulting in trajectory predictions that deviate from the specified instructions. To address this problem, inspired by [19], we maintain the distance-dependent attention within the unimodal text or visual tokens, but make the cross-modal attention between instructions and visual tokens unaffected by the position encoding. Specifically, with $\\mathcal{V}$ and $\\mathcal{I}$ denoting the sets of visual and instruction tokens, taking the $j$ -th position as an example, our DDIA is defined as:" + ], + "bbox": [ + 511, + 409, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {D D I A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {k _ {i} \\in \\mathcal {I}} \\operatorname {s i m} (R (\\mathbf {q} _ {j}) , R (\\mathbf {k} _ {i})) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {I}} \\operatorname {s i m} (R (\\mathbf {q} _ {j}) , R (\\mathbf {k} _ {i}))}, \\text {i f} \\mathbf {q} _ {j} \\in \\mathcal {I}, \\\\ = \\frac {\\sum_ {k _ {i} \\in \\mathcal {I}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i} + \\sum_ {k _ {i} \\in \\{\\mathcal {V} | < j \\}} s i m \\left(R \\left(\\mathbf {q} _ {j}\\right) , R \\left(\\mathbf {k} _ {i}\\right)\\right) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {I}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) + \\sum_ {k _ {i} \\in \\{\\mathcal {V} | < j \\}} s i m \\left(R \\left(\\mathbf {q} _ {j}\\right) , R \\left(\\mathbf {k} _ {i}\\right)\\right)}, \\text {i f} \\mathbf {q} _ {j} \\in \\mathcal {V} \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 777, + 903, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R(\\cdot)$ denotes RoPE position embedding,", + "bbox": [ + 513, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7eafec198d985369c33fd865de5517d71b446904fbc53edf6750414cde1ec99c.jpg", + "image_caption": [ + "Figure 4. A detailed illustration of our proposed Distance-decoupled Instruction Attention (DDIA)." + ], + "image_footnote": [], + "bbox": [ + 178, + 88, + 395, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\begin{array}{rl}{sim(\\mathbf{q}_j,\\mathbf{k}_i)} & {= \\exp (\\mathbf{q}_j^T\\mathbf{k}_i / \\sqrt{d})} \\end{array}$ and $\\{\\mathcal{V}| < j\\}$ represents the causal subset within visual tokens.", + "bbox": [ + 89, + 301, + 483, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Objective Function", + "text_level": 1, + "bbox": [ + 89, + 343, + 274, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The full training objective for $i$ -th frame consists of two parts, formulated as follows:", + "bbox": [ + 89, + 364, + 483, + 395 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {w a y}} + \\mathcal {L} _ {\\text {c y c}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 406, + 480, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "here, $\\mathcal{L}_{\\text {way }}$ supervises trajectory prediction by minimizing the $L_{1}$ norm between predicted and ground-truth paths [28], while $\\mathcal{L}_{\\text {cyc }}$ facilitates our model to select the most discriminative tokens and aggregate more representative visual features, which is composed of two adversarial losses:", + "bbox": [ + 89, + 433, + 483, + 508 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c y c} = \\lambda_ {1} \\mathcal {L} _ {p r u n} + \\lambda_ {2} \\mathcal {L} _ {r e c} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 518, + 480, + 537 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r u n} = \\left(R - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\mathbf {M} _ {i j}\\right) ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 546, + 482, + 588 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\sum_ {j = 1} ^ {N} (1 - \\mathbf {M} _ {i j}) \\left(\\mathbf {F} _ {i j} - \\hat {\\mathbf {F}} _ {i j} ^ {r e c}\\right) ^ {2} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 595, + 480, + 637 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{prun}$ restricts the ratio of kept tokens to our pre-defined value $R$ , and $\\mathcal{L}_{rec}$ reconstructs pruned tokens from kept ones after memory-enhanced feature aggregation. $\\lambda_{1}$ and $\\lambda_{2}$ balance the weights of two loss items. $\\mathbf{M}_{ij}$ , $\\mathbf{F}_{ij}$ and $\\hat{\\mathbf{F}}_{ij}^{rec}$ denote the $j$ -th element of $\\mathbf{M}_i$ , $\\mathbf{F}_i$ and $\\hat{\\mathbf{F}}_i^{rec}$ , respectively.", + "bbox": [ + 89, + 642, + 483, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 748, + 223, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 89, + 772, + 282, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We train our model on the official language-driven autonomous driving dataset [28], which includes 64K instruction-following data clips collected across 8 towns on CARLA [6] simulation environment. Each clip not only equips multi-sensor input data (multi-view camera images and LiDAR data), but also introduces aligned navigation instructions in the natural language format.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the visual encoder in LMDrive [28], producing 106 visual tokens $(N = 106)$ and kept it fixed during training. Two configurations of lightweight language models are integrated into our framework: LLaMA* (a 4-layer lite version of LLaMA [34]) and TinyLLaMA [44]. The hype parameters $R$ determining the ratio of kept tokens and $Z$ denoting the number of stored adjacent frames in the memory bank are set to 0.3 and 10, respectively. $\\lambda_{1}$ and $\\lambda_{2}$ in the loss function (Eq. 11) are configured to 10 and 1. More details are provided in the Supplementary Material.", + "bbox": [ + 511, + 90, + 906, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. As in LMDrive [28], we evaluate our model on three benchmarks, categorized by route length: LangAuto ( $>500\\mathrm{m}$ ), LangAuto-Short ( $150\\mathrm{m}-500\\mathrm{m}$ ) and LangAuto-Tiny ( $<150\\mathrm{m}$ ) and employ route completion (RC), infraction score (IS), and driving score (DS) to assess close-looped driving performance. Specifically, route completion measures the extent to which a specified route is completed. Infraction score quantifies driving safety, starting from an ideal 1.0 base score and decrementing proportionally in the event of collisions or traffic rule violations. Driving Score is calculated as the product of RC and IS, offering a comprehensive evaluation of overall performance.", + "bbox": [ + 511, + 277, + 906, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with LLM-based Agents", + "text_level": 1, + "bbox": [ + 511, + 483, + 843, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our VLDrive with LMDrive, a pioneering LLM-driven, closed-loop autonomous driving method, across several distance benchmarks. Table 1 presents the comparison results on the LangAuto benchmark. VLDrive has approximately $81\\%$ fewer parameters yet demonstrates superior driving performance compared to LMDrive, which is equipped with various LLM configurations. Specifically, VLDrive-LLaMA* achieves driving scores of $41.7\\%$ and route completion rates of $52.6\\%$ , outperforming LMDrive with LLaVA-v1.5 by $5.5\\%$ and $6.1\\%$ . This trend continues on the LangAuto-Short and LangAuto-Tiny benchmarks, as shown in Table 2, where VLDrive-LLaMA* attains driving scores of $63.6\\%$ and $78.4\\%$ , surpassing LMDrive by $13\\%$ and $11.9\\%$ . VLDrive-TinyLLaMA achieves new state-of-the-art performance, boosting the DS score to $43.8\\%$ , $67.4\\%$ , and $81.9\\%$ on the LangAuto, LangAuto-Short, and LangAuto-Tiny benchmarks, respectively.", + "bbox": [ + 511, + 505, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 772, + 666, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To investigate the impact of various components of our method, we conduct comprehensive ablation studies on the standard LangAuto benchmark. To reduce time costs and enhance training efficiency, we randomly extract approximately $25\\%$ of the training data from each town, forming a mini-training set for all ablation experiments. We present the metrics from 3 evaluation runs.", + "bbox": [ + 511, + 794, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5928", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7132032a8c4c5a6c04b31b9a9ffca549fa9e07f60787fe7ce964af532febbe2b.jpg", + "table_caption": [ + "Table 1. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto benchmark. We present the metrics from 3 evaluation runs." + ], + "table_footnote": [], + "table_body": "
MethodLLMParams(Total)LangAuto
DS ↑RC ↑IS ↑
LMDriveLarge Language Model
LLaMA [34]7B31.3±1.537.1±1.60.82±0.01
LLaMA2 [35]7B32.8±2.140.1±2.20.81±0.02
Vicuna [48]7B33.5±1.939.3±1.90.83±0.02
Vicuna-v1.5 [48]7B34.0±3.839.0±3.30.85±0.06
LLaVA-v1.5 [17]7B36.2±2.346.5±4.30.81±0.03
Lightweight Language Model
Phi-2 [11]3B22.3±0.128.9±1.00.82±0.02
TinyLLaMA [44]1.3B25.2±2.338.6±3.70.71±0.02
OursLLaMA*1.3B41.7±1.852.6±2.50.81±0.02
TinyLLaMA [44]1.3B43.8±2.454.5±3.00.84±0.02
", + "bbox": [ + 196, + 128, + 802, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/867bf80c72ab6cdb5d82afcebd28794fab5bafa18c00e278083b01287f07f9dc.jpg", + "table_caption": [ + "Table 2. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto-Short and LangAuto-Tiny benchmarks. We present the metrics from 3 evaluation runs." + ], + "table_footnote": [], + "table_body": "
MethodLLMParams(Total)LangAuto-ShortLangAuto-Tiny
DS↑RC↑IS↑DS↑RC↑IS↑
LMDriveLarge Language Model
LLaMA [34]7B42.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [35]7B44.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [48]7B45.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [48]7B47.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [17]7B50.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
Lightweight Language Model
Phi-2 [11]3B48.0±3.655.2±4.10.87±0.0156.3±4.968.2±3.90.82±0.04
TinyLLaMA [44]1.3B46.2±3.459.7±3.40.79±0.0264.1±1.275.0±1.40.86 ± 0.01
OursLLaMA*1.3B63.6±4.177.6±4.10.81±0.0678.4±4.685.6±2.40.91±0.04
TinyLLaMA [44]1.3B67.4±2.078.1±2.30.85±0.0281.9±0.985.5±0.60.94±0.02
", + "bbox": [ + 94, + 435, + 906, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Each Individual Component. As shown in Table 3, we begin our ablation study from the baseline, which utilizes $\\mathsf{LLaMA^{*}}$ as the trainable language model and Q-Former [13] with learnable queries as the connector. Our method improves the baseline by: (1) introducing CCDP to extract critical sparse tokens, (2) replacing the vanilla Q-Former with our proposed MEFA module, and (3) substituting the causal attention mechanism with our DDIA module. The results in Table 3 validate the contribution of each component, with their synergistic integration leading to substantial improvements in driving performance.", + "bbox": [ + 88, + 689, + 480, + 857 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of CCDP. We compare our CCDP approach with other token reduction strategies, including: 1) Structural pooling, a token reduction technique based on average pooling operations. 2) Dynamic pruning without token reconstruction. We maintain a consistent token retention ratio across all methods. As shown in Table 4, benefiting from the cycle-consistency constraints, CCDP is able to extract the most critical visual information and consequently achieves the highest driving scores. Fig. 5 further illustrates the positive correlation between reconstruction loss and trajectory prediction loss on the open-looped validation set, with a pearson correlation coefficient of 0.65.", + "bbox": [ + 511, + 689, + 906, + 872 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d38e90891884a209c1bb5817ef11bb21d7bdb75c340925d175bd404147895330.jpg", + "table_caption": [ + "Table 3. Ablation studies of each individual component." + ], + "table_footnote": [], + "table_body": "
MethodDS ↑RC ↑IS ↑
Baseline30.0±0.846.2±3.50.69±0.03
Ours39.8±1.749.1±1.80.83±0.02
w/o CCDP35.5±3.845.3±3.50.78±0.02
w/o MEFA35.9±0.846.9±1.30.81±0.02
w/o DDIA36.3±1.347.0±1.00.81±0.02
", + "bbox": [ + 112, + 114, + 464, + 228 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/73f42340f332c0d0f82b8c08385827b833745b670222274ce8652ef984d59960.jpg", + "table_caption": [ + "Table 4. Ablation studies of various token reduction strategies." + ], + "table_footnote": [], + "table_body": "
MethodDS ↑RC ↑IS ↑
Structural Pooling33.6±2.445.8±4.30.78±0.06
Dynamic Pruning36.3±2.648.0±2.30.76±0.03
CCDP (Ours)39.8±1.749.1±1.80.83±0.02
", + "bbox": [ + 107, + 266, + 467, + 340 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dbef7ee27d0e8c1f1f2dac2ac1f4523bc1ab1ac020e3e30402d8e2e524a1bd57.jpg", + "table_caption": [ + "Table 5. Ablation studies of different token retention ratios." + ], + "table_footnote": [], + "table_body": "
Retention Ratio (R)DS ↑RC ↑IS ↑
5%33.4±0.445.1±3.20.77±0.03
10%35.6±1.747.3±1.00.79±0.02
30%39.8±1.749.1±1.80.83±0.02
50%39.8±1.251.4±3.70.79±0.04
", + "bbox": [ + 107, + 378, + 465, + 463 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/93f75b09c486cc12f35dbe86748508547279c16793153e4a48f2e6461c25d674.jpg", + "table_caption": [ + "Table 6. Ablation of memory bank with different capacity." + ], + "table_footnote": [], + "table_body": "
Capacity (Z)DS ↑RC ↑IS ↑
537.1±1.949.1±0.50.80±0.01
1039.8±1.749.1±1.80.83±0.02
2038.3±3.049.9±2.60.81±0.02
", + "bbox": [ + 107, + 503, + 467, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Token Sparsification. Table 5 demonstrates that, retaining $30\\%$ of the visual tokens captures critical clues for autonomous driving. Further increment in the number of tokens produces no significant performance gains in driving scores.", + "bbox": [ + 89, + 607, + 483, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of MEFA. As illustrated in Table 6, incorporating a memory bank provides essential temporal information that aids in predicting the future trajectories and behaviors of the ego-car and other objects, thereby enhancing the reliability of autonomous driving. Furthermore, our experiments indicate that a capacity of 10 adjacent frames yields the most significant performance improvements.", + "bbox": [ + 89, + 700, + 483, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of DDIA. Fig. 6 shows the visual comparison of attention maps generated by vanilla causal attention and DDIA, where brighter colors indicate higher values. As highlighted in the red dashed box, DDIA enhances attention to instruction tokens, enabling visual tokens to better", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f5ca852d49d1f765b962ae472d39edff095754abf24cc84ed576e383ef7ced2.jpg", + "image_caption": [ + "Figure 5. Correlation analysis between reconstruction and trajectory prediction losses, revealing a significant positive relationship (Pearson's $r = 0.65$ )." + ], + "image_footnote": [], + "bbox": [ + 540, + 90, + 880, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c03539343b4ea5741151e10ebdd4a7d39f1c709ea13b041d89c44e0573db91e5.jpg", + "image_caption": [ + "Figure 6. Visual comparison of attention maps generated by the language model equipped with vanilla causal attention and DDIA. We aggregate the attention weights within the instruction segment into the first four tokens for clearer visualization." + ], + "image_footnote": [], + "bbox": [ + 545, + 303, + 707, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3dcb27fd1417c95498778aee49bcf22b245c824ce6b4687e643607e7797e4f9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 301, + 874, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "integrate instructional information and thus make more accurate driving decisions.", + "bbox": [ + 511, + 544, + 905, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 599, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce VLDrive, a vision-enhanced, lightweight model for language-grounded autonomous driving. To facilitate the practical deployment of autonomous vehicles, we streamline the heavy LLM into a more compact form while augmenting its visual modeling and vision-language alignment. Specifically, we employ cycle-consistent visual dynamic pruning to efficiently capture the most salient visual information for each frame, and incorporate a memory-enhanced feature aggregation that enriches critical temporal information to help the model comprehend historical trajectories and infer future movements. Additionally, a distance-decoupled instruction attention strategy is tailored for our lightweight language model, alleviating the attention dilution to navigation instructions and boosting the model's instruction-following capability. Extensive experiments demonstrate that our model, despite having significantly fewer parameters, achieves superior and robust driving performance across various benchmarks.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5930", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the National Natural Science Foundation of China (62322608), in part by the Shenzhen Longgang District Science and Technology Innovation Special Fund (No. LGKCYLWS2023018), in part by the Futian Healthcare Research Project (No.FTWS002), and in part by the Shenzhen Medical Research Fund (No. C2401036). This work is also sponsored by CIE-Tencent Robotics X Rhino-Bird Focused Research Program.", + "bbox": [ + 89, + 114, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 277, + 187, + 292 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2", + "[2] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. Robotics, 12(5):127, 2023. 2", + "[3] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 3", + "[4] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2", + "[5] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as you speak: Enabling human-like interaction with large language models in autonomous vehicles. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 902-909, 2024. 3", + "[6] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 6", + "[7] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 3", + "[8] Jeffrey Hawke, Richard Shen, Corina Gurau, Siddharth Sharma, Daniele Reda, Nikolay Nikolov, Przemysław Mazur, Sean Micklethwaite, Nicolas Griffiths, Amar Shah, et al. Urban driving with conditional imitation learning. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 251-257. IEEE, 2020. 2", + "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + ], + "bbox": [ + 99, + 303, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2", + "[11] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. Microsoft Research Blog, 2023. 7", + "[12] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In European conference on computer vision, pages 620-640. Springer, 2022. 3", + "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3, 4, 7", + "[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 3", + "[15] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553–11562, 2020. 1", + "[16] Xiaodan Liang, Tairui Wang, Luona Yang, and Eric Xing. Cirl: Controllable imitative reinforcement learning for vision-based self-driving. In Proceedings of the European conference on computer vision (ECCV), pages 584-599, 2018. 2", + "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7", + "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1", + "[19] Fan Ma, Xiaojie Jin, Heng Wang, Yuchen Xian, Jiashi Feng, and Yi Yang. Vista-llama: Reliable video narrator via equal distance to visual tokens. arXiv preprint arXiv:2312.08870, 2023. 5", + "[20] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3", + "[21] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 3", + "[22] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, et al. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3", + "[23] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for au" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tonomous driving. arXiv preprint arXiv:2312.03661, 2023. 3", + "[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2", + "[25] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 3", + "[26] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 1, 2", + "[27] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3", + "[28] Hao Shao, Yuxuan Hu, Letian Wang, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. arXiv preprint arXiv:2312.07488, 2023. 1, 3, 6", + "[29] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 2", + "[30] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3", + "[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5", + "[32] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 2", + "[33] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6, 7", + "[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7", + "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[37] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 1, 3", + "[38] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 3", + "[39] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 3", + "[40] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2964-2972, 2022. 3", + "[41] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 1, 3", + "[42] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end simulated driving. In Proceedings of the AAAI conference on artificial intelligence, 2017. 2", + "[43] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 3", + "[44] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 6, 7", + "[45] Ruifei Zhang, Xiangru Lin, Wei Zhang, Jincheng Lu, Xuekuan Wang, Xiao Tan, Yingying Li, Errui Ding, Jingdong Wang, and Guanbin Li. Interactive 3d object detection with prompts. In European Conference on Computer Vision, pages 140-157. Springer, 2024. 2", + "[46] Wei Zhang, Jiaming Li, Meng Xia, Xu Gao, Xiao Tan, Yifeng Shi, Zhenhua Huang, and Guanbin Li. Offsetnet: Towards efficient multiple object tracking, detection, and segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[47] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5932", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 7", + "[49] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[50] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3" + ], + "bbox": [ + 91, + 90, + 480, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5933", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_model.json b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..88212f2237b16c820d352f3153eadcc41258871b --- /dev/null +++ b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_model.json @@ -0,0 +1,2114 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.82, + 0.177 + ], + "angle": 0, + "content": "VLDrive: Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.203, + 0.888, + 0.277 + ], + "angle": 0, + "content": "Ruifei Zhang\\(^{1,2}\\), Wei Zhang\\(^{4}\\), Xiao Tan\\(^{4}\\), Sibei Yang\\(^{3}\\), Xiang Wan\\(^{2}\\), Xiaonan Luo\\(^{5}\\), Guanbin Li\\(^{3,6*}\\) \n\\(^{1}\\) The Chinese University of Hong Kong, Shenzhen \\(^{2}\\) Shenzhen Research Institute of Big Data \n\\(^{3}\\) Sun Yat-sen University \\(^{4}\\) Baidu Inc. \\(^{5}\\) Guilin University of Electronic Technology \n\\(^{6}\\) Guangdong Key Laboratory of Big Data Analysis and Processing" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.735 + ], + "angle": 0, + "content": "Recent advancements in language-grounded autonomous driving have been significantly promoted by the sophisticated cognition and reasoning capabilities of large language models (LLMs). However, current LLM-based approaches encounter critical challenges: (1) Failure analysis reveals that frequent collisions and obstructions, stemming from limitations in visual representations, remain primary obstacles to robust driving performance. (2) The substantial parameters of LLMs pose considerable deployment hurdles. To address these limitations, we introduce VLDrive, a novel approach featuring a lightweight MLLM architecture with enhanced vision components. VLDrive achieves compact visual tokens through innovative strategies, including cycle-consistent dynamic visual pruning and memory-enhanced feature aggregation. Furthermore, we propose a distance-decoupled instruction attention mechanism to improve joint visual-linguistic feature learning, particularly for long-range visual tokens. Extensive experiments conducted in the CARLA simulator demonstrate VLDrive's effectiveness. Notably, VLDrive achieves state-of-the-art driving performance while reducing parameters by \\(81\\%\\) (from 7B to 1.3B), yielding substantial driving score improvements of \\(15.4\\%\\), \\(16.8\\%\\), and \\(7.6\\%\\) at tiny, short, and long distances, respectively, in closed-loop evaluations. Code is available at https://github.com/ReaFly/VLDrive." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Autonomous driving has experienced significant development in recent years, evolving from modular and segregated design to integrated end-to-end networks [10, 15, 26, 29, 46]. However, these networks still rely on specific inputs such as target points or action commands to guide their driving behaviors, significantly constraining the interaction" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.309, + 0.716, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.309, + 0.898, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 1. (a) Existing methods for language-grounded driving. (b) Our proposed VLDrive: a novel framework featuring a lightweight MLLM architecture with enhanced vision components. (c) Driving failure analysis of existing methods based on three evaluation runs. (d) Performance comparison between VLDrive and both versions of LMDrive, highlighting our method's superior driving performance with fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.874, + 0.629 + ], + "angle": 0, + "content": "with humans and applicability in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.78 + ], + "angle": 0, + "content": "The emergence of large language models (LLMs) [18, 33, 49, 50] has catalyzed a revolution in autonomous driving. Impressed by their advanced cognition and logical reasoning capabilities, a multitude of studies have integrated LLMs into autonomous driving systems and achieved noteworthy results [28, 37, 38, 41]. Among them, LMDrive [28] successfully achieves human-friendly language-grounded autonomous driving, where vehicle behaviors are solely guided by natural language instructions (see Fig. 1(a)), significantly enhancing human-vehicle interaction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "However, despite remarkable advancements, these approaches encounter significant practical deployment challenges, characterized by suboptimal driving performance and excessive computational parameters. To address these issues, we conduct a thorough investigation of driving failure cases in current approaches, coupled with performance analysis across different LLM scales, yielding two critical insights: 1) Deficiencies in visual comprehension" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.392, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: liguanbin@mail.sysu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.424 + ], + "angle": 0, + "content": "are the predominant cause of driving failures. We analyze the failure cases of existing language-grounded autonomous driving models and identify two primary causes: instruction misunderstanding, manifested as the agent deviating from the route; and visual misunderstanding, characterized by failures due to collisions or blocks involving vehicles or layouts. The results in Fig. 1(c) reveal that limitations in visual comprehension frequently emerge as a critical factor contributing to driving task failures. 2) Driving performance does not scale linearly with LLM's parameter count. Through comprehensive evaluations across three benchmarks, we compare the full-scale LMDrive (7B parameters) with its lightweight counterpart, LMDrive-LiteLM (1.3B trainable parameters). As demonstrated in Fig. 1(d), LMDrive-LiteLM achieves comparable driving performance despite its significantly reduced parameter count. These findings strongly suggest that massive parameter scales in LLMs may not be a critical prerequisite for achieving superior driving performance. The insights derived from these analyses illuminate the path toward developing a parameter-efficient yet high-performing alternative for language-grounded driving applications." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.429, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Building on these considerations, in our work, we propose VLDrive, a vision-augmented lightweight language model for efficient language-grounded autonomous driving, as showcased in Fig. 1(b). We enhance the model's visual comprehension by drawing inspiration from human driving competencies, focusing on two key aspects: prioritized attention allocation to sparse but critical agents and objects, and temporal trajectory memory for behavioral sequence prediction. These enhancements are achieved through our proposed Cycle-consistent Dynamic Visual Pruning (CCDP) and Memory-enhanced Feature Aggregation (MEFA) mechanisms. Specifically, CDP framework consists of token sparsification and training-only token reconstruction. The sparsification mechanism enables adaptive selection of critical visual tokens and dynamic adjustment of token count based on driving scenario complexity. The reconstruction process preserves information integrity by explicitly recovering features of pruned tokens, enabling robust visual perception while minimizing computational overhead. Additionally, MEFA incorporates a memory bank of adjacent frames to capture temporal dependencies, providing critical temporal cues for visual feature enhancement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Furthermore, we introduce a Distance-decoupled Instruction Attention (DDIA) mechanism to enhance instruction comprehension capabilities of lightweight language models. This design stems from our observation that increasing numbers of long-range visual tokens may dilute instruction attention (see Fig. 6), potentially leading to instruction misunderstanding. DDIA alleviates this issue and enhances joint visual-linguistic feature learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "and alignment, resulting in more accurate trajectory planning. Extensive closed-loop experiments conducted on the CARLA [6] simulation platform demonstrate the effectiveness of our proposed method, surpassing state-of-the-art approaches by \\(15.4\\%\\), \\(16.8\\%\\) and \\(7.6\\%\\) in driving scores at tiny, short and long distances, respectively." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.184, + 0.848, + 0.197 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.2, + 0.905, + 0.319 + ], + "angle": 0, + "content": "- We identify that visual comprehension deficiencies constitute the primary bottleneck in language-grounded driving, with performance scaling non-linearly with language model size. This observation motivates our proposed paradigm: a lightweight language model enhanced by vision-augmented strategies, which simultaneously reduces computational overhead and improves driving performance, facilitating practical deployment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- We propose VLDrive, a lightweight architecture enhanced with vision-centric strategies for language-grounded driving. The framework incorporates CCDP for adaptive visual signal extraction and MEFA for temporal information integration. Additionally, DDIA facilitates visual-language alignment for robust navigation instruction following. These complementary strategies collectively enable more reliable autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.441, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- We evaluate our approach through closed-loop simulation experiments on standard language-grounded driving benchmarks using the CARLA platform. VLDrive achieves state-of-the-art performance while using significantly fewer parameters." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.2, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.655, + 0.547 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.807, + 0.573 + ], + "angle": 0, + "content": "2.1. End-to-End Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Early end-to-end autonomous driving methods can be mainly divided into two categories: imitation learning [4, 8, 42] and reinforcement learning [2, 16, 32, 47]. Subsequently, benefiting from the robust feature modeling capabilities of the Transformer [36], researchers introduced additional modalities to provide complementary information for autonomous driving [24, 29, 45]. For example, Transfuse [24] uses Transformer to merge features from RGB and LiDAR bird's eye view (BEV) images to enhance the global comprehension of 3D scenes. InterFuse [29] improves the security and interpretability of the model by directly presenting its intermediate features and limiting actions to specified safe sets. In recent years, modular end-to-end planning approaches [1, 10, 26] have gained increased attention, where all components are connected and jointly optimized towards the ultimate goal, further improving the driving performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.783, + 0.864 + ], + "angle": 0, + "content": "2.2. LLMs in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As LLMs continue to evolve and flourish, an increasing number of researchers are incorporating them into au" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5924" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.62 + ], + "angle": 0, + "content": "tonomous driving tasks, broadening the scope and capabilities of these systems. Among them, a notable strand of research [22, 23, 30] employs LLMs for driving-related visual question answering, providing enhanced interpretability for driving decisions. Another line of work seeks to fully exploit the powerful logical reasoning capabilities of LLMs to fundamentally transform the research paradigm in autonomous driving tasks [5, 7, 20, 21]. Specifically, GPT-Driver [20] feeds LLM with observations and egostates as language prompts and generates a planned trajectory and corresponding decision-making process in a natural language format. Agent-Driver [21] builds an LMM-powered agent, capitalizing on a tool library, a cognitive memory, and a reasoning engine for human-like autonomous driving. Recently, LLMs have been extended to visual modality successfully, making it possible for LLMs to understand the visual information of autonomous driving. Chen et al. [3] introduces a multimodal LLM to process vectorized numeric modalities data and produce driving-related question answers as well as action decisions. DriveGPT4 [41] curates a visual instruction tuning dataset for interpretable autonomous driving, taming LLM to generate text responses and low-level control signals. Further, based on the CARLA simulation platform, many studies [28, 37, 38] delve deeper and perform closed-loop evaluation driven by LLMs directly. For instance, LMDrive [28] takes multi-modality data and navigation instructions as input, achieving language-grounded driving and enhanced interaction with humans. Despite the impressive performance powered by LLMs, their considerably heavy parameters pose significant deployment challenges. In this work, we demonstrate that a lightweight language model can effectively respond to instructions and execute driving tasks through efficient visual representation extraction and strengthened vision-language alignment." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.636, + 0.256, + 0.649 + ], + "angle": 0, + "content": "2.3. Token Reduction" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Transformer has achieved substantial performance improvements across a range of tasks, yet it faces criticism for its quadratic computational costs relative to the number of input tokens. To tackle this issue, numerous studies focus on reducing token counts [12, 25, 39, 40]. For instance, DynamicViT [25] adaptively prunes redundant tokens at different Transformer layers to achieve token sparsification. SPViT [12] introduces an attention-based multihead token selector with a soft pruning strategy. Unlike direct pruning methods, this approach consolidates selected redundant tokens into a single package token, enhancing efficiency while preserving essential information. In the LLM era, token reduction is also a crucial strategy to reduce the heavy computation costs [13, 14, 27, 50]. Most existing works utilize Q-former [13] to aggregate visual features into a fixed number of tokens [13, 43, 50]. LLaMa" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "VID [14] leverages average pooling followed by a learnable projector to compress the video frame representation. LLaVA-PruMerge [27] proposes an adaptive token selection via outlier detection, paired with a cluster-based merging strategy to supplement the information of pruned tokens to kept ones. Distinct from prior approaches, we present a cycle-consistent dynamic pruning strategy that integrates token reconstruction with token sparsification at the training phase. This design not only enables more representative token selection but also ensures that the retained tokens aggregate fine-grained global visual information." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.274, + 0.603, + 0.289 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.299, + 0.622, + 0.313 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.397 + ], + "angle": 0, + "content": "This task aims to predict driving actions by leveraging sequences of multi-sensor data and corresponding human navigation instructions. As shown in Fig. 2, our method follows the common framework of a multi-modal large language model (MLLM) and comprises three main components:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.488 + ], + "angle": 0, + "content": "Visual Encoder: Given a sequence of visual data \\(\\{\\mathbf{X}_i, i = 1,2,\\dots,T\\}\\), where each frame \\(\\mathbf{X}_i\\) includes multi-view camera images and corresponding LiDAR data. A visual encoder processes these multi-modal inputs to produce a unified feature representation \\(\\mathbf{F}_i \\in \\mathbb{R}^{N \\times C}\\) for each frame, with \\(N\\) representing the total number of tokens." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.655 + ], + "angle": 0, + "content": "Connector: As a bridge to align visual features with the language space, the connector plays a remarkable role in MLLM. To endow our model with adaptive visual perception and long-context modeling capability, as illustrated in Fig. 3, we tailor cycle-consistent dynamic visual pruning (CCDP), consisting of token sparsification and reconstruction, and memory-enhanced feature aggregation (MEFA) for the connector, significantly enhancing the visual representation. We denote the improved features of each frame by the connector as \\(\\mathbf{F}_i^v\\in \\mathbb{R}^{N_v\\times C_t}\\), where \\(C_t\\) is the language model's hidden dimension." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.656, + 0.905, + 0.82 + ], + "angle": 0, + "content": "Lite Language Model: A lightweight language model, enhanced by our proposed distance-decoupled instruction attention (DDIA), is utilized to process the combination of visual tokens of all frames, \\(\\mathbf{F}_v\\in \\mathbb{R}^{N_vT\\times C_t}\\) and corresponding text embedding \\(\\mathbf{F}_t\\in \\mathbb{R}^{N_t\\times C_t}\\) tokenized from navigation instructions. A following MLP leverages features from the language model and predicts the future trajectory, which is subsequently converted by PID controllers to produce the lateral steering action and the longitudinal acceleration action. We detail each component in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.833, + 0.766, + 0.849 + ], + "angle": 0, + "content": "3.2. CCDP: Token Sparsification" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Given the \\(i\\)-th frame's multi-modal features \\(\\mathbf{F}_i\\), as in DynamicViT [25], we predict the probabilities for pruning and retaining each token. Specifically, the probabilities are cal" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.212, + 0.087, + 0.768, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.314, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 2. An overview of our proposed VLDrive framework. Given a sequence of visual data, our connector transforms each frame's raw visual features \\(\\mathbf{F}_i\\) into sparse yet informative representations \\(\\mathbf{F}_i^v\\) through two key components: CCDP: Token Sparsification and Memory-enhanced Feature Aggregation (MEFA). Subsequently, a lite language model augmented with Distance-decoupled Instruction Attention (DDIA) jointly processes tokenized navigation instructions \\(\\mathbf{F}_t\\) and temporal visual features \\(\\mathbf{F}_v = \\{\\mathbf{F}_i^v \\mid i = 1, \\dots, T\\}\\). The resulting hidden representations are fed into an MLP for trajectory prediction, followed by PID controllers that translate these predictions into concrete driving actions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity of \\(\\mathbf{F}_i^v\\) via explicit token reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.438, + 0.483, + 0.469 + ], + "angle": 0, + "content": "ulated based on the combined global and local features \\(\\left[\\mathbf{G}_i;\\mathbf{L}_i\\right]\\), which are obtained by \\(\\mathrm{MLP}(\\cdot)\\) operations:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.479, + 0.483, + 0.499 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {i} = \\operatorname {M L P} \\left(\\mathbf {F} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times \\frac {C}{2}} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.508, + 0.482, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {i} = \\operatorname {A v g} \\left(\\mathbf {L} _ {i}\\right) \\in \\mathbb {R} ^ {1 \\times \\frac {C}{2}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.579 + ], + "angle": 0, + "content": "where \\(\\mathrm{Avg}(\\cdot)\\) is average operation. Another MLP is utilized to predict the probability \\(\\mathbf{S}_i\\), and \\([\\cdot ;\\cdot ]\\) denotes concatenation with broadcast mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.59, + 0.482, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {i} = \\operatorname {S o f t m a x} \\left(\\operatorname {M L P} \\left(\\left[ \\mathbf {G} _ {i}; \\mathbf {L} _ {i} \\right]\\right)\\right) \\in \\mathbb {R} ^ {N \\times 2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.483, + 0.649 + ], + "angle": 0, + "content": "In our implementation, we leverage Gumbel-Softmax to obtain the binary mask:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.66, + 0.483, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i} = \\text {G u m b e l - S o f t m a x} (\\mathbf {S} _ {i}) _ {* 1} \\in \\{0, 1 \\} ^ {N} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_i\\) denotes the token retention masks. As shown in Fig. 3, the kept tokens \\(\\mathbf{F}_i^k\\) selected based on \\(\\mathbf{M}_i\\) are delivered to the Memory-enhanced Feature Aggregation module, which is detailed in the next subsection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.435, + 0.775 + ], + "angle": 0, + "content": "3.3. Memory-enhanced Feature Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Temporal reasoning is integral to our everyday driving scenarios. Human drivers can deduce the likely future direction or intentions of an approaching vehicle by analyzing its historical trajectory, and subsequently tailor their driving actions. To attain autonomous driving that matches the safety and reliability of human driving, we explicitly inject the temporal context into the current visual tokens. Specifically, for the current time step \\( i \\), we introduce a memory" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.438, + 0.907, + 0.499 + ], + "angle": 0, + "content": "bank storing the adjacent \\(Z\\) frames from the current one, denoted as \\(\\mathbf{B}_i = [\\mathbf{F}_{i - Z},\\mathbf{F}_{i - Z + 1},\\dots,\\mathbf{F}_{i - 1}]\\) \\(\\mathbf{B}_i\\in \\mathbb{R}^{Z\\times N\\times C}\\) Temporal encoding \\(\\mathbf{TE}_i\\) is derived by an MLP utilizing the historical average data and current frame information:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.508, + 0.905, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {i} ^ {a v g} = \\operatorname {A v g} \\left(\\mathbf {B} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times C} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.537, + 0.905, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T E} _ {i} = \\operatorname {M L P} \\left(\\left[ \\mathbf {F} _ {i}; \\mathbf {B} _ {i} ^ {\\text {a v g}} \\right]\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Afterwards, a query transformer (Q-former) [13] is introduced to aggregate the features of current frames enhanced by temporary encoding:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.616, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {v} = \\operatorname {Q - F o r m e r} \\left(\\mathbf {F} _ {i} ^ {k}, \\mathbf {F} _ {i} + \\mathbf {T E} _ {i}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.78 + ], + "angle": 0, + "content": "In contrast to the original Q-former, we implement two key modifications: 1) Instead of using learnable queries, we utilize the retained visual tokens \\(\\mathbf{F}_i^k\\) as queries to probe and aggregate features. These retained tokens serve as significant anchors within the current frame, allowing for more effective feature aggregation. 2) We introduce a temporary embedding \\(\\mathbf{TE}_i\\) to the original visual features \\(\\mathbf{F}_i\\), which enhances the model's ability to focus on temporary-salient moving objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Finally, the generated \\(\\mathbf{F}_i^v\\) serves as part of the aggregated visual tokens across all frames, \\(\\mathbf{F}_v\\), which is then fed into the subsequent language model for trajectory prediction based on the corresponding navigation instructions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity via explicit token reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.088, + 0.837, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Figure 3. A detailed illustration of our proposed connector. CCDP: Token Sparsification and Memory-enhanced Feature Aggregation are proposed to reduce token count while enhancing information density. CCDP: Token Reconstruction serves as a training-only task that further ensures the information integrity of the retained tokens. Red arrows \\((\\rightarrow)\\) indicate the input and output paths of our connector." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.379, + 0.356, + 0.394 + ], + "angle": 0, + "content": "3.4. CCDP: Token Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Through memory-enhanced feature aggregation, the generated \\(\\mathbf{F}_i^v\\) ideally encapsulates all visual features of the current frame. To further enhance the connector's ability to select critical visual tokens and aggregate complete information, we expect to reconstruct features of those pruned tokens based on \\(\\mathbf{F}_i^v\\). Firstly, we add the enhanced tokens \\(\\mathbf{F}_i^v\\) into the original features \\(\\mathbf{F}_i\\) according to their original token positions, forming a new feature vector denoted as \\(\\langle \\mathrm{MLP}(\\mathbf{F}_i^v), \\mathbf{F}_i \\rangle\\), where an MLP function is used to align the dimensions of \\(\\mathbf{F}_i^v\\) and \\(\\mathbf{F}_i\\). Secondly, drawing inspiration from the Masked Autoencoder [9], a learnable embedding \\(\\mathbf{e} \\in \\mathbb{R}^{N \\times C}\\) is introduced as a foundation for reconstructing the pruned tokens. Finally, we construct the input feature vector to be reconstructed through the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.62, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i} ^ {r e c} = \\left\\langle \\operatorname {M L P} \\left(\\mathbf {F} _ {i} ^ {v}\\right), \\mathbf {F} _ {i} \\right\\rangle \\cdot \\mathbf {M} _ {i} + \\mathbf {e} \\cdot \\left(1 - \\mathbf {M} _ {i}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.766 + ], + "angle": 0, + "content": "In this way, we incorporate the decision \\(\\mathbf{M}_i\\) into the computational graph of feature reconstruction, enabling the reconstruction loss to influence pruning decisions and compelling the model to retain the most critical tokens. Afterwards, a 4-layer Transformer block takes as input \\(\\mathbf{F}_i^{rec}\\) and predicts the reconstructed results \\(\\hat{\\mathbf{F}}_i^{rec}\\). A cycle consistency constraint (Eq. 11) is imposed to ensure that the retained tokens \\(\\mathbf{F}_i^v\\) preserve the integral information of the current video frame." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.445, + 0.79 + ], + "angle": 0, + "content": "3.5. Distance-decoupled Instruction Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We introduce a lightweight language model for efficient autonomous driving, reducing both the number of parameters and computational resource requirements. Furthermore, distance-decoupled instruction attention (DDIA) is tailored for our life language model, significantly alleviating the attention dilution for navigation instructions in the long-context driving scenarios. Specifically, as shown in Fig. 4," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.409 + ], + "angle": 0, + "content": "our proposed DDIA diverges from the traditional vanilla attention in two key aspects:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.905, + 0.485 + ], + "angle": 0, + "content": "1): Causal-attention \\(\\rightarrow\\) Self-attention: Rather than adhering to the original causal setting, our self-attention within instructional segments allows each text token to access ample contextual information, thereby producing a more comprehensive representation of text features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.773 + ], + "angle": 0, + "content": "2): Distance-dependent \\(\\rightarrow\\) Distance-decoupled: In contemporary research on LLM, position encoding (e.g. RoPE [31]) is recognized as a critical ingredient to explicitly inject positional information into input tokens and enhance the model's contextual modeling capabilities. However, position encodings may introduce challenges, particularly in long-range driving scenarios. As historical visual data accumulates, the distance between current visual tokens and instructional tokens increases. The feature disparity introduced by position encoding may negatively impact the attention sensitivity of visual tokens to instruction ones, resulting in trajectory predictions that deviate from the specified instructions. To address this problem, inspired by [19], we maintain the distance-dependent attention within the unimodal text or visual tokens, but make the cross-modal attention between instructions and visual tokens unaffected by the position encoding. Specifically, with \\(\\mathcal{V}\\) and \\(\\mathcal{I}\\) denoting the sets of visual and instruction tokens, taking the \\(j\\)-th position as an example, our DDIA is defined as:" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.778, + 0.905, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {D D I A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {k _ {i} \\in \\mathcal {I}} \\operatorname {s i m} (R (\\mathbf {q} _ {j}) , R (\\mathbf {k} _ {i})) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {I}} \\operatorname {s i m} (R (\\mathbf {q} _ {j}) , R (\\mathbf {k} _ {i}))}, \\text {i f} \\mathbf {q} _ {j} \\in \\mathcal {I}, \\\\ = \\frac {\\sum_ {k _ {i} \\in \\mathcal {I}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i} + \\sum_ {k _ {i} \\in \\{\\mathcal {V} | < j \\}} s i m \\left(R \\left(\\mathbf {q} _ {j}\\right) , R \\left(\\mathbf {k} _ {i}\\right)\\right) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {I}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) + \\sum_ {k _ {i} \\in \\{\\mathcal {V} | < j \\}} s i m \\left(R \\left(\\mathbf {q} _ {j}\\right) , R \\left(\\mathbf {k} _ {i}\\right)\\right)}, \\text {i f} \\mathbf {q} _ {j} \\in \\mathcal {V} \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(R(\\cdot)\\) denotes RoPE position embedding," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.396, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.279 + ], + "angle": 0, + "content": "Figure 4. A detailed illustration of our proposed Distance-decoupled Instruction Attention (DDIA)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.334 + ], + "angle": 0, + "content": "\\(\\begin{array}{rl}{sim(\\mathbf{q}_j,\\mathbf{k}_i)} & {= \\exp (\\mathbf{q}_j^T\\mathbf{k}_i / \\sqrt{d})} \\end{array}\\) and \\(\\{\\mathcal{V}| < j\\}\\) represents the causal subset within visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.344, + 0.275, + 0.359 + ], + "angle": 0, + "content": "3.6. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.396 + ], + "angle": 0, + "content": "The full training objective for \\(i\\)-th frame consists of two parts, formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.407, + 0.482, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {w a y}} + \\mathcal {L} _ {\\text {c y c}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.509 + ], + "angle": 0, + "content": "here, \\(\\mathcal{L}_{\\text {way }}\\) supervises trajectory prediction by minimizing the \\(L_{1}\\) norm between predicted and ground-truth paths [28], while \\(\\mathcal{L}_{\\text {cyc }}\\) facilitates our model to select the most discriminative tokens and aggregate more representative visual features, which is composed of two adversarial losses:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.52, + 0.482, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c y c} = \\lambda_ {1} \\mathcal {L} _ {p r u n} + \\lambda_ {2} \\mathcal {L} _ {r e c} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.547, + 0.483, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r u n} = \\left(R - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\mathbf {M} _ {i j}\\right) ^ {2} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.596, + 0.482, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\sum_ {j = 1} ^ {N} (1 - \\mathbf {M} _ {i j}) \\left(\\mathbf {F} _ {i j} - \\hat {\\mathbf {F}} _ {i j} ^ {r e c}\\right) ^ {2} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.736 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{prun}\\) restricts the ratio of kept tokens to our pre-defined value \\(R\\), and \\(\\mathcal{L}_{rec}\\) reconstructs pruned tokens from kept ones after memory-enhanced feature aggregation. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) balance the weights of two loss items. \\(\\mathbf{M}_{ij}\\), \\(\\mathbf{F}_{ij}\\) and \\(\\hat{\\mathbf{F}}_{ij}^{rec}\\) denote the \\(j\\)-th element of \\(\\mathbf{M}_i\\), \\(\\mathbf{F}_i\\) and \\(\\hat{\\mathbf{F}}_i^{rec}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.749, + 0.224, + 0.765 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.284, + 0.79 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. We train our model on the official language-driven autonomous driving dataset [28], which includes 64K instruction-following data clips collected across 8 towns on CARLA [6] simulation environment. Each clip not only equips multi-sensor input data (multi-view camera images and LiDAR data), but also introduces aligned navigation instructions in the natural language format." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Implementation Details. We adopt the visual encoder in LMDrive [28], producing 106 visual tokens \\((N = 106)\\) and kept it fixed during training. Two configurations of lightweight language models are integrated into our framework: LLaMA* (a 4-layer lite version of LLaMA [34]) and TinyLLaMA [44]. The hype parameters \\(R\\) determining the ratio of kept tokens and \\(Z\\) denoting the number of stored adjacent frames in the memory bank are set to 0.3 and 10, respectively. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in the loss function (Eq. 11) are configured to 10 and 1. More details are provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Evaluation Metrics. As in LMDrive [28], we evaluate our model on three benchmarks, categorized by route length: LangAuto (\\(>500\\mathrm{m}\\)), LangAuto-Short (\\(150\\mathrm{m}-500\\mathrm{m}\\)) and LangAuto-Tiny (\\(<150\\mathrm{m}\\)) and employ route completion (RC), infraction score (IS), and driving score (DS) to assess close-looped driving performance. Specifically, route completion measures the extent to which a specified route is completed. Infraction score quantifies driving safety, starting from an ideal 1.0 base score and decrementing proportionally in the event of collisions or traffic rule violations. Driving Score is calculated as the product of RC and IS, offering a comprehensive evaluation of overall performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.844, + 0.5 + ], + "angle": 0, + "content": "4.2. Comparisons with LLM-based Agents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.765 + ], + "angle": 0, + "content": "We compare our VLDrive with LMDrive, a pioneering LLM-driven, closed-loop autonomous driving method, across several distance benchmarks. Table 1 presents the comparison results on the LangAuto benchmark. VLDrive has approximately \\(81\\%\\) fewer parameters yet demonstrates superior driving performance compared to LMDrive, which is equipped with various LLM configurations. Specifically, VLDrive-LLaMA* achieves driving scores of \\(41.7\\%\\) and route completion rates of \\(52.6\\%\\), outperforming LMDrive with LLaVA-v1.5 by \\(5.5\\%\\) and \\(6.1\\%\\). This trend continues on the LangAuto-Short and LangAuto-Tiny benchmarks, as shown in Table 2, where VLDrive-LLaMA* attains driving scores of \\(63.6\\%\\) and \\(78.4\\%\\), surpassing LMDrive by \\(13\\%\\) and \\(11.9\\%\\). VLDrive-TinyLLaMA achieves new state-of-the-art performance, boosting the DS score to \\(43.8\\%\\), \\(67.4\\%\\), and \\(81.9\\%\\) on the LangAuto, LangAuto-Short, and LangAuto-Tiny benchmarks, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.667, + 0.789 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.9 + ], + "angle": 0, + "content": "To investigate the impact of various components of our method, we conduct comprehensive ablation studies on the standard LangAuto benchmark. To reduce time costs and enhance training efficiency, we randomly extract approximately \\(25\\%\\) of the training data from each town, forming a mini-training set for all ablation experiments. We present the metrics from 3 evaluation runs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5928" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 1. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto benchmark. We present the metrics from 3 evaluation runs." + }, + { + "type": "table", + "bbox": [ + 0.197, + 0.129, + 0.803, + 0.384 + ], + "angle": 0, + "content": "
MethodLLMParams(Total)LangAuto
DS ↑RC ↑IS ↑
LMDriveLarge Language Model
LLaMA [34]7B31.3±1.537.1±1.60.82±0.01
LLaMA2 [35]7B32.8±2.140.1±2.20.81±0.02
Vicuna [48]7B33.5±1.939.3±1.90.83±0.02
Vicuna-v1.5 [48]7B34.0±3.839.0±3.30.85±0.06
LLaVA-v1.5 [17]7B36.2±2.346.5±4.30.81±0.03
Lightweight Language Model
Phi-2 [11]3B22.3±0.128.9±1.00.82±0.02
TinyLLaMA [44]1.3B25.2±2.338.6±3.70.71±0.02
OursLLaMA*1.3B41.7±1.852.6±2.50.81±0.02
TinyLLaMA [44]1.3B43.8±2.454.5±3.00.84±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.424 + ], + "angle": 0, + "content": "Table 2. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto-Short and LangAuto-Tiny benchmarks. We present the metrics from 3 evaluation runs." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.436, + 0.907, + 0.666 + ], + "angle": 0, + "content": "
MethodLLMParams(Total)LangAuto-ShortLangAuto-Tiny
DS↑RC↑IS↑DS↑RC↑IS↑
LMDriveLarge Language Model
LLaMA [34]7B42.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [35]7B44.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [48]7B45.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [48]7B47.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [17]7B50.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
Lightweight Language Model
Phi-2 [11]3B48.0±3.655.2±4.10.87±0.0156.3±4.968.2±3.90.82±0.04
TinyLLaMA [44]1.3B46.2±3.459.7±3.40.79±0.0264.1±1.275.0±1.40.86 ± 0.01
OursLLaMA*1.3B63.6±4.177.6±4.10.81±0.0678.4±4.685.6±2.40.91±0.04
TinyLLaMA [44]1.3B67.4±2.078.1±2.30.85±0.0281.9±0.985.5±0.60.94±0.02
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.69, + 0.482, + 0.858 + ], + "angle": 0, + "content": "Effectiveness of Each Individual Component. As shown in Table 3, we begin our ablation study from the baseline, which utilizes \\(\\mathsf{LLaMA^{*}}\\) as the trainable language model and Q-Former [13] with learnable queries as the connector. Our method improves the baseline by: (1) introducing CCDP to extract critical sparse tokens, (2) replacing the vanilla Q-Former with our proposed MEFA module, and (3) substituting the causal attention mechanism with our DDIA module. The results in Table 3 validate the contribution of each component, with their synergistic integration leading to substantial improvements in driving performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.873 + ], + "angle": 0, + "content": "Effectiveness of CCDP. We compare our CCDP approach with other token reduction strategies, including: 1) Structural pooling, a token reduction technique based on average pooling operations. 2) Dynamic pruning without token reconstruction. We maintain a consistent token retention ratio across all methods. As shown in Table 4, benefiting from the cycle-consistency constraints, CCDP is able to extract the most critical visual information and consequently achieves the highest driving scores. Fig. 5 further illustrates the positive correlation between reconstruction loss and trajectory prediction loss on the open-looped validation set, with a pearson correlation coefficient of 0.65." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5929" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.09, + 0.454, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation studies of each individual component." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.115, + 0.465, + 0.229 + ], + "angle": 0, + "content": "
MethodDS ↑RC ↑IS ↑
Baseline30.0±0.846.2±3.50.69±0.03
Ours39.8±1.749.1±1.80.83±0.02
w/o CCDP35.5±3.845.3±3.50.78±0.02
w/o MEFA35.9±0.846.9±1.30.81±0.02
w/o DDIA36.3±1.347.0±1.00.81±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.242, + 0.472, + 0.257 + ], + "angle": 0, + "content": "Table 4. Ablation studies of various token reduction strategies." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.267, + 0.468, + 0.341 + ], + "angle": 0, + "content": "
MethodDS ↑RC ↑IS ↑
Structural Pooling33.6±2.445.8±4.30.78±0.06
Dynamic Pruning36.3±2.648.0±2.30.76±0.03
CCDP (Ours)39.8±1.749.1±1.80.83±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.354, + 0.464, + 0.367 + ], + "angle": 0, + "content": "Table 5. Ablation studies of different token retention ratios." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.38, + 0.467, + 0.464 + ], + "angle": 0, + "content": "
Retention Ratio (R)DS ↑RC ↑IS ↑
5%33.4±0.445.1±3.20.77±0.03
10%35.6±1.747.3±1.00.79±0.02
30%39.8±1.749.1±1.80.83±0.02
50%39.8±1.251.4±3.70.79±0.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.477, + 0.461, + 0.492 + ], + "angle": 0, + "content": "Table 6. Ablation of memory bank with different capacity." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.504, + 0.468, + 0.583 + ], + "angle": 0, + "content": "
Capacity (Z)DS ↑RC ↑IS ↑
537.1±1.949.1±0.50.80±0.01
1039.8±1.749.1±1.80.83±0.02
2038.3±3.049.9±2.60.81±0.02
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.684 + ], + "angle": 0, + "content": "Analysis of Token Sparsification. Table 5 demonstrates that, retaining \\(30\\%\\) of the visual tokens captures critical clues for autonomous driving. Further increment in the number of tokens produces no significant performance gains in driving scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.808 + ], + "angle": 0, + "content": "Effectiveness of MEFA. As illustrated in Table 6, incorporating a memory bank provides essential temporal information that aids in predicting the future trajectories and behaviors of the ego-car and other objects, thereby enhancing the reliability of autonomous driving. Furthermore, our experiments indicate that a capacity of 10 adjacent frames yields the most significant performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of DDIA. Fig. 6 shows the visual comparison of attention maps generated by vanilla causal attention and DDIA, where brighter colors indicate higher values. As highlighted in the red dashed box, DDIA enhances attention to instruction tokens, enabling visual tokens to better" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.091, + 0.881, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.239, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Correlation analysis between reconstruction and trajectory prediction losses, revealing a significant positive relationship (Pearson's \\( r = 0.65 \\))." + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.304, + 0.708, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.302, + 0.875, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.468, + 0.907, + 0.523 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of attention maps generated by the language model equipped with vanilla causal attention and DDIA. We aggregate the attention weights within the instruction segment into the first four tokens for clearer visualization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.906, + 0.577 + ], + "angle": 0, + "content": "integrate instructional information and thus make more accurate driving decisions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.601, + 0.634, + 0.616 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce VLDrive, a vision-enhanced, lightweight model for language-grounded autonomous driving. To facilitate the practical deployment of autonomous vehicles, we streamline the heavy LLM into a more compact form while augmenting its visual modeling and vision-language alignment. Specifically, we employ cycle-consistent visual dynamic pruning to efficiently capture the most salient visual information for each frame, and incorporate a memory-enhanced feature aggregation that enriches critical temporal information to help the model comprehend historical trajectories and infer future movements. Additionally, a distance-decoupled instruction attention strategy is tailored for our lightweight language model, alleviating the attention dilution to navigation instructions and boosting the model's instruction-following capability. Extensive experiments demonstrate that our model, despite having significantly fewer parameters, achieves superior and robust driving performance across various benchmarks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5930" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.268 + ], + "angle": 0, + "content": "This work is supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the National Natural Science Foundation of China (62322608), in part by the Shenzhen Longgang District Science and Technology Innovation Special Fund (No. LGKCYLWS2023018), in part by the Futian Healthcare Research Project (No.FTWS002), and in part by the Shenzhen Medical Research Fund (No. C2401036). This work is also sponsored by CIE-Tencent Robotics X Rhino-Bird Focused Research Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.278, + 0.188, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[1] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[2] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. Robotics, 12(5):127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[3] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.485, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[4] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.554, + 0.482, + 0.623 + ], + "angle": 0, + "content": "[5] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as you speak: Enabling human-like interaction with large language models in autonomous vehicles. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 902-909, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.624, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.483, + 0.748 + ], + "angle": 0, + "content": "[7] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[8] Jeffrey Hawke, Richard Shen, Corina Gurau, Siddharth Sharma, Daniele Reda, Nikolay Nikolov, Przemysław Mazur, Sean Micklethwaite, Nicolas Griffiths, Amar Shah, et al. Urban driving with conditional imitation learning. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 251-257. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.304, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[10] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[11] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. Microsoft Research Blog, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[12] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In European conference on computer vision, pages 620-640. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[15] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553–11562, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[16] Xiaodan Liang, Tairui Wang, Luona Yang, and Eric Xing. Cirl: Controllable imitative reinforcement learning for vision-based self-driving. In Proceedings of the European conference on computer vision (ECCV), pages 584-599, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[19] Fan Ma, Xiaojie Jin, Heng Wang, Yuchen Xian, Jiashi Feng, and Yi Yang. Vista-llama: Reliable video narrator via equal distance to visual tokens. arXiv preprint arXiv:2312.08870, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[20] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[21] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[22] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, et al. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for au" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5931" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "tonomous driving. arXiv preprint arXiv:2312.03661, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[25] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[26] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[27] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[28] Hao Shao, Yuxuan Hu, Letian Wang, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. arXiv preprint arXiv:2312.07488, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[29] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[30] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[32] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[33] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[37] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[38] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.386 + ], + "angle": 0, + "content": "[39] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[40] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2964-2972, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[41] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.567 + ], + "angle": 0, + "content": "[42] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end simulated driving. In Proceedings of the AAAI conference on artificial intelligence, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[43] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[44] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[45] Ruifei Zhang, Xiangru Lin, Wei Zhang, Jincheng Lu, Xuekuan Wang, Xiao Tan, Yingying Li, Errui Ding, Jingdong Wang, and Guanbin Li. Interactive 3d object detection with prompts. In European Conference on Computer Vision, pages 140-157. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[46] Wei Zhang, Jiaming Li, Meng Xia, Xu Gao, Xiao Tan, Yifeng Shi, Zhenhua Huang, and Guanbin Li. Offsetnet: Towards efficient multiple object tracking, detection, and segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[47] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[49] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[50] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5933" + } + ] +] \ No newline at end of file diff --git a/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_origin.pdf b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49b9e92932b7519cda7dafa64ae6225a1a4e8322 --- /dev/null +++ b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/03e7ad20-412e-4720-b543-ad870d80aec2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ca90806629d8c4aed1a6f8e2d000f02fc27df0d633c72856f365af0b9b20f3a +size 1183662 diff --git a/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/full.md b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3e9032c66abe39fc583ea565175be96fbc48a0fa --- /dev/null +++ b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/full.md @@ -0,0 +1,309 @@ +# VLDrive: Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving + +Ruifei Zhang $^{1,2}$ , Wei Zhang $^{4}$ , Xiao Tan $^{4}$ , Sibei Yang $^{3}$ , Xiang Wan $^{2}$ , Xiaonan Luo $^{5}$ , Guanbin Li $^{3,6*}$ $^{1}$ The Chinese University of Hong Kong, Shenzhen $^{2}$ Shenzhen Research Institute of Big Data + $^{3}$ Sun Yat-sen University $^{4}$ Baidu Inc. $^{5}$ Guilin University of Electronic Technology + $^{6}$ Guangdong Key Laboratory of Big Data Analysis and Processing + +# Abstract + +Recent advancements in language-grounded autonomous driving have been significantly promoted by the sophisticated cognition and reasoning capabilities of large language models (LLMs). However, current LLM-based approaches encounter critical challenges: (1) Failure analysis reveals that frequent collisions and obstructions, stemming from limitations in visual representations, remain primary obstacles to robust driving performance. (2) The substantial parameters of LLMs pose considerable deployment hurdles. To address these limitations, we introduce VLDrive, a novel approach featuring a lightweight MLLM architecture with enhanced vision components. VLDrive achieves compact visual tokens through innovative strategies, including cycle-consistent dynamic visual pruning and memory-enhanced feature aggregation. Furthermore, we propose a distance-decoupled instruction attention mechanism to improve joint visual-linguistic feature learning, particularly for long-range visual tokens. Extensive experiments conducted in the CARLA simulator demonstrate VLDrive's effectiveness. Notably, VLDrive achieves state-of-the-art driving performance while reducing parameters by $81\%$ (from 7B to 1.3B), yielding substantial driving score improvements of $15.4\%$ , $16.8\%$ , and $7.6\%$ at tiny, short, and long distances, respectively, in closed-loop evaluations. Code is available at https://github.com/ReaFly/VLDrive. + +# 1. Introduction + +Autonomous driving has experienced significant development in recent years, evolving from modular and segregated design to integrated end-to-end networks [10, 15, 26, 29, 46]. However, these networks still rely on specific inputs such as target points or action commands to guide their driving behaviors, significantly constraining the interaction + +![](images/91a5ebeb0cdffd587f73890074dcde367cf1bcfbe283f9832cd91c0018624c7d.jpg) +Figure 1. (a) Existing methods for language-grounded driving. (b) Our proposed VLDrive: a novel framework featuring a lightweight MLLM architecture with enhanced vision components. (c) Driving failure analysis of existing methods based on three evaluation runs. (d) Performance comparison between VLDrive and both versions of LMDrive, highlighting our method's superior driving performance with fewer parameters. + +![](images/14a09f89cdb435b1ef70942c80fd7c78129a403c0bd80dfe5fe92f722a781d4c.jpg) + +with humans and applicability in real-world scenarios. + +The emergence of large language models (LLMs) [18, 33, 49, 50] has catalyzed a revolution in autonomous driving. Impressed by their advanced cognition and logical reasoning capabilities, a multitude of studies have integrated LLMs into autonomous driving systems and achieved noteworthy results [28, 37, 38, 41]. Among them, LMDrive [28] successfully achieves human-friendly language-grounded autonomous driving, where vehicle behaviors are solely guided by natural language instructions (see Fig. 1(a)), significantly enhancing human-vehicle interaction. + +However, despite remarkable advancements, these approaches encounter significant practical deployment challenges, characterized by suboptimal driving performance and excessive computational parameters. To address these issues, we conduct a thorough investigation of driving failure cases in current approaches, coupled with performance analysis across different LLM scales, yielding two critical insights: 1) Deficiencies in visual comprehension + +are the predominant cause of driving failures. We analyze the failure cases of existing language-grounded autonomous driving models and identify two primary causes: instruction misunderstanding, manifested as the agent deviating from the route; and visual misunderstanding, characterized by failures due to collisions or blocks involving vehicles or layouts. The results in Fig. 1(c) reveal that limitations in visual comprehension frequently emerge as a critical factor contributing to driving task failures. 2) Driving performance does not scale linearly with LLM's parameter count. Through comprehensive evaluations across three benchmarks, we compare the full-scale LMDrive (7B parameters) with its lightweight counterpart, LMDrive-LiteLM (1.3B trainable parameters). As demonstrated in Fig. 1(d), LMDrive-LiteLM achieves comparable driving performance despite its significantly reduced parameter count. These findings strongly suggest that massive parameter scales in LLMs may not be a critical prerequisite for achieving superior driving performance. The insights derived from these analyses illuminate the path toward developing a parameter-efficient yet high-performing alternative for language-grounded driving applications. + +Building on these considerations, in our work, we propose VLDrive, a vision-augmented lightweight language model for efficient language-grounded autonomous driving, as showcased in Fig. 1(b). We enhance the model's visual comprehension by drawing inspiration from human driving competencies, focusing on two key aspects: prioritized attention allocation to sparse but critical agents and objects, and temporal trajectory memory for behavioral sequence prediction. These enhancements are achieved through our proposed Cycle-consistent Dynamic Visual Pruning (CCDP) and Memory-enhanced Feature Aggregation (MEFA) mechanisms. Specifically, CDP framework consists of token sparsification and training-only token reconstruction. The sparsification mechanism enables adaptive selection of critical visual tokens and dynamic adjustment of token count based on driving scenario complexity. The reconstruction process preserves information integrity by explicitly recovering features of pruned tokens, enabling robust visual perception while minimizing computational overhead. Additionally, MEFA incorporates a memory bank of adjacent frames to capture temporal dependencies, providing critical temporal cues for visual feature enhancement. + +Furthermore, we introduce a Distance-decoupled Instruction Attention (DDIA) mechanism to enhance instruction comprehension capabilities of lightweight language models. This design stems from our observation that increasing numbers of long-range visual tokens may dilute instruction attention (see Fig. 6), potentially leading to instruction misunderstanding. DDIA alleviates this issue and enhances joint visual-linguistic feature learning + +and alignment, resulting in more accurate trajectory planning. Extensive closed-loop experiments conducted on the CARLA [6] simulation platform demonstrate the effectiveness of our proposed method, surpassing state-of-the-art approaches by $15.4\%$ , $16.8\%$ and $7.6\%$ in driving scores at tiny, short and long distances, respectively. + +To summarize, our contributions are as follows: + +- We identify that visual comprehension deficiencies constitute the primary bottleneck in language-grounded driving, with performance scaling non-linearly with language model size. This observation motivates our proposed paradigm: a lightweight language model enhanced by vision-augmented strategies, which simultaneously reduces computational overhead and improves driving performance, facilitating practical deployment. +- We propose VLDrive, a lightweight architecture enhanced with vision-centric strategies for language-grounded driving. The framework incorporates CCDP for adaptive visual signal extraction and MEFA for temporal information integration. Additionally, DDIA facilitates visual-language alignment for robust navigation instruction following. These complementary strategies collectively enable more reliable autonomous driving. +- We evaluate our approach through closed-loop simulation experiments on standard language-grounded driving benchmarks using the CARLA platform. VLDrive achieves state-of-the-art performance while using significantly fewer parameters. + +# 2. Related Work + +# 2.1. End-to-End Autonomous Driving + +Early end-to-end autonomous driving methods can be mainly divided into two categories: imitation learning [4, 8, 42] and reinforcement learning [2, 16, 32, 47]. Subsequently, benefiting from the robust feature modeling capabilities of the Transformer [36], researchers introduced additional modalities to provide complementary information for autonomous driving [24, 29, 45]. For example, Transfuse [24] uses Transformer to merge features from RGB and LiDAR bird's eye view (BEV) images to enhance the global comprehension of 3D scenes. InterFuse [29] improves the security and interpretability of the model by directly presenting its intermediate features and limiting actions to specified safe sets. In recent years, modular end-to-end planning approaches [1, 10, 26] have gained increased attention, where all components are connected and jointly optimized towards the ultimate goal, further improving the driving performance. + +# 2.2. LLMs in Autonomous Driving + +As LLMs continue to evolve and flourish, an increasing number of researchers are incorporating them into au + +tonomous driving tasks, broadening the scope and capabilities of these systems. Among them, a notable strand of research [22, 23, 30] employs LLMs for driving-related visual question answering, providing enhanced interpretability for driving decisions. Another line of work seeks to fully exploit the powerful logical reasoning capabilities of LLMs to fundamentally transform the research paradigm in autonomous driving tasks [5, 7, 20, 21]. Specifically, GPT-Driver [20] feeds LLM with observations and egostates as language prompts and generates a planned trajectory and corresponding decision-making process in a natural language format. Agent-Driver [21] builds an LMM-powered agent, capitalizing on a tool library, a cognitive memory, and a reasoning engine for human-like autonomous driving. Recently, LLMs have been extended to visual modality successfully, making it possible for LLMs to understand the visual information of autonomous driving. Chen et al. [3] introduces a multimodal LLM to process vectorized numeric modalities data and produce driving-related question answers as well as action decisions. DriveGPT4 [41] curates a visual instruction tuning dataset for interpretable autonomous driving, taming LLM to generate text responses and low-level control signals. Further, based on the CARLA simulation platform, many studies [28, 37, 38] delve deeper and perform closed-loop evaluation driven by LLMs directly. For instance, LMDrive [28] takes multi-modality data and navigation instructions as input, achieving language-grounded driving and enhanced interaction with humans. Despite the impressive performance powered by LLMs, their considerably heavy parameters pose significant deployment challenges. In this work, we demonstrate that a lightweight language model can effectively respond to instructions and execute driving tasks through efficient visual representation extraction and strengthened vision-language alignment. + +# 2.3. Token Reduction + +Transformer has achieved substantial performance improvements across a range of tasks, yet it faces criticism for its quadratic computational costs relative to the number of input tokens. To tackle this issue, numerous studies focus on reducing token counts [12, 25, 39, 40]. For instance, DynamicViT [25] adaptively prunes redundant tokens at different Transformer layers to achieve token sparsification. SPViT [12] introduces an attention-based multihead token selector with a soft pruning strategy. Unlike direct pruning methods, this approach consolidates selected redundant tokens into a single package token, enhancing efficiency while preserving essential information. In the LLM era, token reduction is also a crucial strategy to reduce the heavy computation costs [13, 14, 27, 50]. Most existing works utilize Q-former [13] to aggregate visual features into a fixed number of tokens [13, 43, 50]. LLaMa + +VID [14] leverages average pooling followed by a learnable projector to compress the video frame representation. LLaVA-PruMerge [27] proposes an adaptive token selection via outlier detection, paired with a cluster-based merging strategy to supplement the information of pruned tokens to kept ones. Distinct from prior approaches, we present a cycle-consistent dynamic pruning strategy that integrates token reconstruction with token sparsification at the training phase. This design not only enables more representative token selection but also ensures that the retained tokens aggregate fine-grained global visual information. + +# 3. Method + +# 3.1. Overview + +This task aims to predict driving actions by leveraging sequences of multi-sensor data and corresponding human navigation instructions. As shown in Fig. 2, our method follows the common framework of a multi-modal large language model (MLLM) and comprises three main components: + +Visual Encoder: Given a sequence of visual data $\{\mathbf{X}_i, i = 1,2,\dots,T\}$ , where each frame $\mathbf{X}_i$ includes multi-view camera images and corresponding LiDAR data. A visual encoder processes these multi-modal inputs to produce a unified feature representation $\mathbf{F}_i \in \mathbb{R}^{N \times C}$ for each frame, with $N$ representing the total number of tokens. + +Connector: As a bridge to align visual features with the language space, the connector plays a remarkable role in MLLM. To endow our model with adaptive visual perception and long-context modeling capability, as illustrated in Fig. 3, we tailor cycle-consistent dynamic visual pruning (CCDP), consisting of token sparsification and reconstruction, and memory-enhanced feature aggregation (MEFA) for the connector, significantly enhancing the visual representation. We denote the improved features of each frame by the connector as $\mathbf{F}_i^v\in \mathbb{R}^{N_v\times C_t}$ , where $C_t$ is the language model's hidden dimension. + +Lite Language Model: A lightweight language model, enhanced by our proposed distance-decoupled instruction attention (DDIA), is utilized to process the combination of visual tokens of all frames, $\mathbf{F}_v\in \mathbb{R}^{N_vT\times C_t}$ and corresponding text embedding $\mathbf{F}_t\in \mathbb{R}^{N_t\times C_t}$ tokenized from navigation instructions. A following MLP leverages features from the language model and predicts the future trajectory, which is subsequently converted by PID controllers to produce the lateral steering action and the longitudinal acceleration action. We detail each component in the following subsections. + +# 3.2. CCDP: Token Sparsification + +Given the $i$ -th frame's multi-modal features $\mathbf{F}_i$ , as in DynamicViT [25], we predict the probabilities for pruning and retaining each token. Specifically, the probabilities are cal + +![](images/73a29ed2563d257a2553c1861c55906718986ce57fdaa8d012829250d576a228.jpg) +Figure 2. An overview of our proposed VLDrive framework. Given a sequence of visual data, our connector transforms each frame's raw visual features $\mathbf{F}_i$ into sparse yet informative representations $\mathbf{F}_i^v$ through two key components: CCDP: Token Sparsification and Memory-enhanced Feature Aggregation (MEFA). Subsequently, a lite language model augmented with Distance-decoupled Instruction Attention (DDIA) jointly processes tokenized navigation instructions $\mathbf{F}_t$ and temporal visual features $\mathbf{F}_v = \{\mathbf{F}_i^v \mid i = 1, \dots, T\}$ . The resulting hidden representations are fed into an MLP for trajectory prediction, followed by PID controllers that translate these predictions into concrete driving actions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity of $\mathbf{F}_i^v$ via explicit token reconstruction. + +ulated based on the combined global and local features $\left[\mathbf{G}_i;\mathbf{L}_i\right]$ , which are obtained by $\mathrm{MLP}(\cdot)$ operations: + +$$ +\mathbf {L} _ {i} = \operatorname {M L P} \left(\mathbf {F} _ {i}\right) \in \mathbb {R} ^ {N \times \frac {C}{2}} \tag {1} +$$ + +$$ +\mathbf {G} _ {i} = \operatorname {A v g} \left(\mathbf {L} _ {i}\right) \in \mathbb {R} ^ {1 \times \frac {C}{2}} \tag {2} +$$ + +where $\mathrm{Avg}(\cdot)$ is average operation. Another MLP is utilized to predict the probability $\mathbf{S}_i$ , and $[\cdot ;\cdot ]$ denotes concatenation with broadcast mechanism: + +$$ +\mathbf {S} _ {i} = \operatorname {S o f t m a x} \left(\operatorname {M L P} \left(\left[ \mathbf {G} _ {i}; \mathbf {L} _ {i} \right]\right)\right) \in \mathbb {R} ^ {N \times 2} \tag {3} +$$ + +In our implementation, we leverage Gumbel-Softmax to obtain the binary mask: + +$$ +\mathbf {M} _ {i} = \text {G u m b e l - S o f t m a x} (\mathbf {S} _ {i}) _ {* 1} \in \{0, 1 \} ^ {N} \tag {4} +$$ + +where $\mathbf{M}_i$ denotes the token retention masks. As shown in Fig. 3, the kept tokens $\mathbf{F}_i^k$ selected based on $\mathbf{M}_i$ are delivered to the Memory-enhanced Feature Aggregation module, which is detailed in the next subsection. + +# 3.3. Memory-enhanced Feature Aggregation + +Temporal reasoning is integral to our everyday driving scenarios. Human drivers can deduce the likely future direction or intentions of an approaching vehicle by analyzing its historical trajectory, and subsequently tailor their driving actions. To attain autonomous driving that matches the safety and reliability of human driving, we explicitly inject the temporal context into the current visual tokens. Specifically, for the current time step $i$ , we introduce a memory + +bank storing the adjacent $Z$ frames from the current one, denoted as $\mathbf{B}_i = [\mathbf{F}_{i - Z},\mathbf{F}_{i - Z + 1},\dots,\mathbf{F}_{i - 1}]$ $\mathbf{B}_i\in \mathbb{R}^{Z\times N\times C}$ Temporal encoding $\mathbf{TE}_i$ is derived by an MLP utilizing the historical average data and current frame information: + +$$ +\mathbf {B} _ {i} ^ {a v g} = \operatorname {A v g} \left(\mathbf {B} _ {i}\right) \in \mathbb {R} ^ {N \times C} \tag {5} +$$ + +$$ +\mathbf {T E} _ {i} = \operatorname {M L P} \left(\left[ \mathbf {F} _ {i}; \mathbf {B} _ {i} ^ {\text {a v g}} \right]\right) \tag {6} +$$ + +Afterwards, a query transformer (Q-former) [13] is introduced to aggregate the features of current frames enhanced by temporary encoding: + +$$ +\mathbf {F} _ {i} ^ {v} = \operatorname {Q - F o r m e r} \left(\mathbf {F} _ {i} ^ {k}, \mathbf {F} _ {i} + \mathbf {T E} _ {i}\right) \tag {7} +$$ + +In contrast to the original Q-former, we implement two key modifications: 1) Instead of using learnable queries, we utilize the retained visual tokens $\mathbf{F}_i^k$ as queries to probe and aggregate features. These retained tokens serve as significant anchors within the current frame, allowing for more effective feature aggregation. 2) We introduce a temporary embedding $\mathbf{TE}_i$ to the original visual features $\mathbf{F}_i$ , which enhances the model's ability to focus on temporary-salient moving objects. + +Finally, the generated $\mathbf{F}_i^v$ serves as part of the aggregated visual tokens across all frames, $\mathbf{F}_v$ , which is then fed into the subsequent language model for trajectory prediction based on the corresponding navigation instructions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity via explicit token reconstruction. + +![](images/8071276f0238c3bc1753601b9d9ae7510de61da5176c5c9d5fdcf90a838c7c22.jpg) +Figure 3. A detailed illustration of our proposed connector. CCDP: Token Sparsification and Memory-enhanced Feature Aggregation are proposed to reduce token count while enhancing information density. CCDP: Token Reconstruction serves as a training-only task that further ensures the information integrity of the retained tokens. Red arrows $(\rightarrow)$ indicate the input and output paths of our connector. + +# 3.4. CCDP: Token Reconstruction + +Through memory-enhanced feature aggregation, the generated $\mathbf{F}_i^v$ ideally encapsulates all visual features of the current frame. To further enhance the connector's ability to select critical visual tokens and aggregate complete information, we expect to reconstruct features of those pruned tokens based on $\mathbf{F}_i^v$ . Firstly, we add the enhanced tokens $\mathbf{F}_i^v$ into the original features $\mathbf{F}_i$ according to their original token positions, forming a new feature vector denoted as $\langle \mathrm{MLP}(\mathbf{F}_i^v), \mathbf{F}_i \rangle$ , where an MLP function is used to align the dimensions of $\mathbf{F}_i^v$ and $\mathbf{F}_i$ . Secondly, drawing inspiration from the Masked Autoencoder [9], a learnable embedding $\mathbf{e} \in \mathbb{R}^{N \times C}$ is introduced as a foundation for reconstructing the pruned tokens. Finally, we construct the input feature vector to be reconstructed through the following formula: + +$$ +\mathbf {F} _ {i} ^ {r e c} = \left\langle \operatorname {M L P} \left(\mathbf {F} _ {i} ^ {v}\right), \mathbf {F} _ {i} \right\rangle \cdot \mathbf {M} _ {i} + \mathbf {e} \cdot \left(1 - \mathbf {M} _ {i}\right) \tag {8} +$$ + +In this way, we incorporate the decision $\mathbf{M}_i$ into the computational graph of feature reconstruction, enabling the reconstruction loss to influence pruning decisions and compelling the model to retain the most critical tokens. Afterwards, a 4-layer Transformer block takes as input $\mathbf{F}_i^{rec}$ and predicts the reconstructed results $\hat{\mathbf{F}}_i^{rec}$ . A cycle consistency constraint (Eq. 11) is imposed to ensure that the retained tokens $\mathbf{F}_i^v$ preserve the integral information of the current video frame. + +# 3.5. Distance-decoupled Instruction Attention + +We introduce a lightweight language model for efficient autonomous driving, reducing both the number of parameters and computational resource requirements. Furthermore, distance-decoupled instruction attention (DDIA) is tailored for our life language model, significantly alleviating the attention dilution for navigation instructions in the long-context driving scenarios. Specifically, as shown in Fig. 4, + +our proposed DDIA diverges from the traditional vanilla attention in two key aspects: + +1): Causal-attention $\rightarrow$ Self-attention: Rather than adhering to the original causal setting, our self-attention within instructional segments allows each text token to access ample contextual information, thereby producing a more comprehensive representation of text features. +2): Distance-dependent $\rightarrow$ Distance-decoupled: In contemporary research on LLM, position encoding (e.g. RoPE [31]) is recognized as a critical ingredient to explicitly inject positional information into input tokens and enhance the model's contextual modeling capabilities. However, position encodings may introduce challenges, particularly in long-range driving scenarios. As historical visual data accumulates, the distance between current visual tokens and instructional tokens increases. The feature disparity introduced by position encoding may negatively impact the attention sensitivity of visual tokens to instruction ones, resulting in trajectory predictions that deviate from the specified instructions. To address this problem, inspired by [19], we maintain the distance-dependent attention within the unimodal text or visual tokens, but make the cross-modal attention between instructions and visual tokens unaffected by the position encoding. Specifically, with $\mathcal{V}$ and $\mathcal{I}$ denoting the sets of visual and instruction tokens, taking the $j$ -th position as an example, our DDIA is defined as: + +$$ +\begin{array}{l} \operatorname {D D I A} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) _ {j} = \frac {\sum_ {k _ {i} \in \mathcal {I}} \operatorname {s i m} (R (\mathbf {q} _ {j}) , R (\mathbf {k} _ {i})) \mathbf {v} _ {i}}{\sum_ {k _ {i} \in \mathcal {I}} \operatorname {s i m} (R (\mathbf {q} _ {j}) , R (\mathbf {k} _ {i}))}, \text {i f} \mathbf {q} _ {j} \in \mathcal {I}, \\ = \frac {\sum_ {k _ {i} \in \mathcal {I}} s i m \left(\mathbf {q} _ {j} , \mathbf {k} _ {i}\right) \mathbf {v} _ {i} + \sum_ {k _ {i} \in \{\mathcal {V} | < j \}} s i m \left(R \left(\mathbf {q} _ {j}\right) , R \left(\mathbf {k} _ {i}\right)\right) \mathbf {v} _ {i}}{\sum_ {k _ {i} \in \mathcal {I}} s i m \left(\mathbf {q} _ {j} , \mathbf {k} _ {i}\right) + \sum_ {k _ {i} \in \{\mathcal {V} | < j \}} s i m \left(R \left(\mathbf {q} _ {j}\right) , R \left(\mathbf {k} _ {i}\right)\right)}, \text {i f} \mathbf {q} _ {j} \in \mathcal {V} \tag {9} \\ \end{array} +$$ + +where $R(\cdot)$ denotes RoPE position embedding, + +![](images/7eafec198d985369c33fd865de5517d71b446904fbc53edf6750414cde1ec99c.jpg) +Figure 4. A detailed illustration of our proposed Distance-decoupled Instruction Attention (DDIA). + +$\begin{array}{rl}{sim(\mathbf{q}_j,\mathbf{k}_i)} & {= \exp (\mathbf{q}_j^T\mathbf{k}_i / \sqrt{d})} \end{array}$ and $\{\mathcal{V}| < j\}$ represents the causal subset within visual tokens. + +# 3.6. Objective Function + +The full training objective for $i$ -th frame consists of two parts, formulated as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {w a y}} + \mathcal {L} _ {\text {c y c}}, \tag {10} +$$ + +here, $\mathcal{L}_{\text {way }}$ supervises trajectory prediction by minimizing the $L_{1}$ norm between predicted and ground-truth paths [28], while $\mathcal{L}_{\text {cyc }}$ facilitates our model to select the most discriminative tokens and aggregate more representative visual features, which is composed of two adversarial losses: + +$$ +\mathcal {L} _ {c y c} = \lambda_ {1} \mathcal {L} _ {p r u n} + \lambda_ {2} \mathcal {L} _ {r e c} \tag {11} +$$ + +$$ +\mathcal {L} _ {p r u n} = \left(R - \frac {1}{N} \sum_ {j = 1} ^ {N} \mathbf {M} _ {i j}\right) ^ {2} \tag {12} +$$ + +$$ +\mathcal {L} _ {r e c} = \sum_ {j = 1} ^ {N} (1 - \mathbf {M} _ {i j}) \left(\mathbf {F} _ {i j} - \hat {\mathbf {F}} _ {i j} ^ {r e c}\right) ^ {2} \tag {13} +$$ + +$\mathcal{L}_{prun}$ restricts the ratio of kept tokens to our pre-defined value $R$ , and $\mathcal{L}_{rec}$ reconstructs pruned tokens from kept ones after memory-enhanced feature aggregation. $\lambda_{1}$ and $\lambda_{2}$ balance the weights of two loss items. $\mathbf{M}_{ij}$ , $\mathbf{F}_{ij}$ and $\hat{\mathbf{F}}_{ij}^{rec}$ denote the $j$ -th element of $\mathbf{M}_i$ , $\mathbf{F}_i$ and $\hat{\mathbf{F}}_i^{rec}$ , respectively. + +# 4. Experiments + +# 4.1. Experiment Settings + +Datasets. We train our model on the official language-driven autonomous driving dataset [28], which includes 64K instruction-following data clips collected across 8 towns on CARLA [6] simulation environment. Each clip not only equips multi-sensor input data (multi-view camera images and LiDAR data), but also introduces aligned navigation instructions in the natural language format. + +Implementation Details. We adopt the visual encoder in LMDrive [28], producing 106 visual tokens $(N = 106)$ and kept it fixed during training. Two configurations of lightweight language models are integrated into our framework: LLaMA* (a 4-layer lite version of LLaMA [34]) and TinyLLaMA [44]. The hype parameters $R$ determining the ratio of kept tokens and $Z$ denoting the number of stored adjacent frames in the memory bank are set to 0.3 and 10, respectively. $\lambda_{1}$ and $\lambda_{2}$ in the loss function (Eq. 11) are configured to 10 and 1. More details are provided in the Supplementary Material. + +Evaluation Metrics. As in LMDrive [28], we evaluate our model on three benchmarks, categorized by route length: LangAuto ( $>500\mathrm{m}$ ), LangAuto-Short ( $150\mathrm{m}-500\mathrm{m}$ ) and LangAuto-Tiny ( $<150\mathrm{m}$ ) and employ route completion (RC), infraction score (IS), and driving score (DS) to assess close-looped driving performance. Specifically, route completion measures the extent to which a specified route is completed. Infraction score quantifies driving safety, starting from an ideal 1.0 base score and decrementing proportionally in the event of collisions or traffic rule violations. Driving Score is calculated as the product of RC and IS, offering a comprehensive evaluation of overall performance. + +# 4.2. Comparisons with LLM-based Agents + +We compare our VLDrive with LMDrive, a pioneering LLM-driven, closed-loop autonomous driving method, across several distance benchmarks. Table 1 presents the comparison results on the LangAuto benchmark. VLDrive has approximately $81\%$ fewer parameters yet demonstrates superior driving performance compared to LMDrive, which is equipped with various LLM configurations. Specifically, VLDrive-LLaMA* achieves driving scores of $41.7\%$ and route completion rates of $52.6\%$ , outperforming LMDrive with LLaVA-v1.5 by $5.5\%$ and $6.1\%$ . This trend continues on the LangAuto-Short and LangAuto-Tiny benchmarks, as shown in Table 2, where VLDrive-LLaMA* attains driving scores of $63.6\%$ and $78.4\%$ , surpassing LMDrive by $13\%$ and $11.9\%$ . VLDrive-TinyLLaMA achieves new state-of-the-art performance, boosting the DS score to $43.8\%$ , $67.4\%$ , and $81.9\%$ on the LangAuto, LangAuto-Short, and LangAuto-Tiny benchmarks, respectively. + +# 4.3. Ablation Study + +To investigate the impact of various components of our method, we conduct comprehensive ablation studies on the standard LangAuto benchmark. To reduce time costs and enhance training efficiency, we randomly extract approximately $25\%$ of the training data from each town, forming a mini-training set for all ablation experiments. We present the metrics from 3 evaluation runs. + +Table 1. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto benchmark. We present the metrics from 3 evaluation runs. + +
MethodLLMParams(Total)LangAuto
DS ↑RC ↑IS ↑
LMDriveLarge Language Model
LLaMA [34]7B31.3±1.537.1±1.60.82±0.01
LLaMA2 [35]7B32.8±2.140.1±2.20.81±0.02
Vicuna [48]7B33.5±1.939.3±1.90.83±0.02
Vicuna-v1.5 [48]7B34.0±3.839.0±3.30.85±0.06
LLaVA-v1.5 [17]7B36.2±2.346.5±4.30.81±0.03
Lightweight Language Model
Phi-2 [11]3B22.3±0.128.9±1.00.82±0.02
TinyLLaMA [44]1.3B25.2±2.338.6±3.70.71±0.02
OursLLaMA*1.3B41.7±1.852.6±2.50.81±0.02
TinyLLaMA [44]1.3B43.8±2.454.5±3.00.84±0.02
+ +Table 2. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto-Short and LangAuto-Tiny benchmarks. We present the metrics from 3 evaluation runs. + +
MethodLLMParams(Total)LangAuto-ShortLangAuto-Tiny
DS↑RC↑IS↑DS↑RC↑IS↑
LMDriveLarge Language Model
LLaMA [34]7B42.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [35]7B44.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [48]7B45.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [48]7B47.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [17]7B50.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
Lightweight Language Model
Phi-2 [11]3B48.0±3.655.2±4.10.87±0.0156.3±4.968.2±3.90.82±0.04
TinyLLaMA [44]1.3B46.2±3.459.7±3.40.79±0.0264.1±1.275.0±1.40.86 ± 0.01
OursLLaMA*1.3B63.6±4.177.6±4.10.81±0.0678.4±4.685.6±2.40.91±0.04
TinyLLaMA [44]1.3B67.4±2.078.1±2.30.85±0.0281.9±0.985.5±0.60.94±0.02
+ +Effectiveness of Each Individual Component. As shown in Table 3, we begin our ablation study from the baseline, which utilizes $\mathsf{LLaMA^{*}}$ as the trainable language model and Q-Former [13] with learnable queries as the connector. Our method improves the baseline by: (1) introducing CCDP to extract critical sparse tokens, (2) replacing the vanilla Q-Former with our proposed MEFA module, and (3) substituting the causal attention mechanism with our DDIA module. The results in Table 3 validate the contribution of each component, with their synergistic integration leading to substantial improvements in driving performance. + +Effectiveness of CCDP. We compare our CCDP approach with other token reduction strategies, including: 1) Structural pooling, a token reduction technique based on average pooling operations. 2) Dynamic pruning without token reconstruction. We maintain a consistent token retention ratio across all methods. As shown in Table 4, benefiting from the cycle-consistency constraints, CCDP is able to extract the most critical visual information and consequently achieves the highest driving scores. Fig. 5 further illustrates the positive correlation between reconstruction loss and trajectory prediction loss on the open-looped validation set, with a pearson correlation coefficient of 0.65. + +Table 3. Ablation studies of each individual component. + +
MethodDS ↑RC ↑IS ↑
Baseline30.0±0.846.2±3.50.69±0.03
Ours39.8±1.749.1±1.80.83±0.02
w/o CCDP35.5±3.845.3±3.50.78±0.02
w/o MEFA35.9±0.846.9±1.30.81±0.02
w/o DDIA36.3±1.347.0±1.00.81±0.02
+ +Table 4. Ablation studies of various token reduction strategies. + +
MethodDS ↑RC ↑IS ↑
Structural Pooling33.6±2.445.8±4.30.78±0.06
Dynamic Pruning36.3±2.648.0±2.30.76±0.03
CCDP (Ours)39.8±1.749.1±1.80.83±0.02
+ +Table 5. Ablation studies of different token retention ratios. + +
Retention Ratio (R)DS ↑RC ↑IS ↑
5%33.4±0.445.1±3.20.77±0.03
10%35.6±1.747.3±1.00.79±0.02
30%39.8±1.749.1±1.80.83±0.02
50%39.8±1.251.4±3.70.79±0.04
+ +Table 6. Ablation of memory bank with different capacity. + +
Capacity (Z)DS ↑RC ↑IS ↑
537.1±1.949.1±0.50.80±0.01
1039.8±1.749.1±1.80.83±0.02
2038.3±3.049.9±2.60.81±0.02
+ +Analysis of Token Sparsification. Table 5 demonstrates that, retaining $30\%$ of the visual tokens captures critical clues for autonomous driving. Further increment in the number of tokens produces no significant performance gains in driving scores. + +Effectiveness of MEFA. As illustrated in Table 6, incorporating a memory bank provides essential temporal information that aids in predicting the future trajectories and behaviors of the ego-car and other objects, thereby enhancing the reliability of autonomous driving. Furthermore, our experiments indicate that a capacity of 10 adjacent frames yields the most significant performance improvements. + +Effectiveness of DDIA. Fig. 6 shows the visual comparison of attention maps generated by vanilla causal attention and DDIA, where brighter colors indicate higher values. As highlighted in the red dashed box, DDIA enhances attention to instruction tokens, enabling visual tokens to better + +![](images/1f5ca852d49d1f765b962ae472d39edff095754abf24cc84ed576e383ef7ced2.jpg) +Figure 5. Correlation analysis between reconstruction and trajectory prediction losses, revealing a significant positive relationship (Pearson's $r = 0.65$ ). + +![](images/c03539343b4ea5741151e10ebdd4a7d39f1c709ea13b041d89c44e0573db91e5.jpg) +Figure 6. Visual comparison of attention maps generated by the language model equipped with vanilla causal attention and DDIA. We aggregate the attention weights within the instruction segment into the first four tokens for clearer visualization. + +![](images/3dcb27fd1417c95498778aee49bcf22b245c824ce6b4687e643607e7797e4f9b.jpg) + +integrate instructional information and thus make more accurate driving decisions. + +# 5. Conclusion + +In this work, we introduce VLDrive, a vision-enhanced, lightweight model for language-grounded autonomous driving. To facilitate the practical deployment of autonomous vehicles, we streamline the heavy LLM into a more compact form while augmenting its visual modeling and vision-language alignment. Specifically, we employ cycle-consistent visual dynamic pruning to efficiently capture the most salient visual information for each frame, and incorporate a memory-enhanced feature aggregation that enriches critical temporal information to help the model comprehend historical trajectories and infer future movements. Additionally, a distance-decoupled instruction attention strategy is tailored for our lightweight language model, alleviating the attention dilution to navigation instructions and boosting the model's instruction-following capability. Extensive experiments demonstrate that our model, despite having significantly fewer parameters, achieves superior and robust driving performance across various benchmarks. + +# Acknowledgments + +This work is supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the National Natural Science Foundation of China (62322608), in part by the Shenzhen Longgang District Science and Technology Innovation Special Fund (No. LGKCYLWS2023018), in part by the Futian Healthcare Research Project (No.FTWS002), and in part by the Shenzhen Medical Research Fund (No. C2401036). This work is also sponsored by CIE-Tencent Robotics X Rhino-Bird Focused Research Program. + +# References + +[1] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2 +[2] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. Robotics, 12(5):127, 2023. 2 +[3] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 3 +[4] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2 +[5] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as you speak: Enabling human-like interaction with large language models in autonomous vehicles. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 902-909, 2024. 3 +[6] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 6 +[7] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 3 +[8] Jeffrey Hawke, Richard Shen, Corina Gurau, Siddharth Sharma, Daniele Reda, Nikolay Nikolov, Przemysław Mazur, Sean Micklethwaite, Nicolas Griffiths, Amar Shah, et al. Urban driving with conditional imitation learning. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 251-257. IEEE, 2020. 2 +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5 + +[10] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2 +[11] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. Microsoft Research Blog, 2023. 7 +[12] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In European conference on computer vision, pages 620-640. Springer, 2022. 3 +[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3, 4, 7 +[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 3 +[15] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553–11562, 2020. 1 +[16] Xiaodan Liang, Tairui Wang, Luona Yang, and Eric Xing. Cirl: Controllable imitative reinforcement learning for vision-based self-driving. In Proceedings of the European conference on computer vision (ECCV), pages 584-599, 2018. 2 +[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7 +[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1 +[19] Fan Ma, Xiaojie Jin, Heng Wang, Yuchen Xian, Jiashi Feng, and Yi Yang. Vista-llama: Reliable video narrator via equal distance to visual tokens. arXiv preprint arXiv:2312.08870, 2023. 5 +[20] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3 +[21] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 3 +[22] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, et al. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3 +[23] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for au + +tonomous driving. arXiv preprint arXiv:2312.03661, 2023. 3 +[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2 +[25] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 3 +[26] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 1, 2 +[27] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3 +[28] Hao Shao, Yuxuan Hu, Letian Wang, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. arXiv preprint arXiv:2312.07488, 2023. 1, 3, 6 +[29] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 2 +[30] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3 +[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 +[32] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 2 +[33] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6, 7 +[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. + +Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7 +[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[37] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 1, 3 +[38] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 3 +[39] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 3 +[40] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2964-2972, 2022. 3 +[41] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 1, 3 +[42] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end simulated driving. In Proceedings of the AAAI conference on artificial intelligence, 2017. 2 +[43] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 3 +[44] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 6, 7 +[45] Ruifei Zhang, Xiangru Lin, Wei Zhang, Jincheng Lu, Xuekuan Wang, Xiao Tan, Yingying Li, Errui Ding, Jingdong Wang, and Guanbin Li. Interactive 3d object detection with prompts. In European Conference on Computer Vision, pages 140-157. Springer, 2024. 2 +[46] Wei Zhang, Jiaming Li, Meng Xia, Xu Gao, Xiao Tan, Yifeng Shi, Zhenhua Huang, and Guanbin Li. Offsetnet: Towards efficient multiple object tracking, detection, and segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[47] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2 + +[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 7 +[49] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36, 2024. 1 +[50] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3 \ No newline at end of file diff --git a/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/images.zip b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0f5e395ec5e5c6d63cf6666c60d8dc1f166304f6 --- /dev/null +++ b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8988231f41431c9d99be8b2eb2bb1d00812034f13cf9d571633a78ee9b2611f +size 586381 diff --git a/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/layout.json b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..223c8724120ce1186f0ff1c84603e1db3fb250e0 --- /dev/null +++ b/2025/VLDrive_ Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving/layout.json @@ -0,0 +1,8691 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "text", + "content": "VLDrive: Vision-Augmented Lightweight MLLMs for Efficient Language-grounded Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": "Ruifei Zhang" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": ", Wei Zhang" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": ", Xiao Tan" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": ", Sibei Yang" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": ", Xiang Wan" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": ", Xiaonan Luo" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": ", Guanbin Li" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{3,6*}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": " The Chinese University of Hong Kong, Shenzhen " + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": " Shenzhen Research Institute of Big Data \n" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": " Sun Yat-sen University " + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": " Baidu Inc. " + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": " Guilin University of Electronic Technology \n" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 70, + 160, + 543, + 219 + ], + "type": "text", + "content": " Guangdong Key Laboratory of Big Data Analysis and Processing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "text", + "content": "Recent advancements in language-grounded autonomous driving have been significantly promoted by the sophisticated cognition and reasoning capabilities of large language models (LLMs). However, current LLM-based approaches encounter critical challenges: (1) Failure analysis reveals that frequent collisions and obstructions, stemming from limitations in visual representations, remain primary obstacles to robust driving performance. (2) The substantial parameters of LLMs pose considerable deployment hurdles. To address these limitations, we introduce VLDrive, a novel approach featuring a lightweight MLLM architecture with enhanced vision components. VLDrive achieves compact visual tokens through innovative strategies, including cycle-consistent dynamic visual pruning and memory-enhanced feature aggregation. Furthermore, we propose a distance-decoupled instruction attention mechanism to improve joint visual-linguistic feature learning, particularly for long-range visual tokens. Extensive experiments conducted in the CARLA simulator demonstrate VLDrive's effectiveness. Notably, VLDrive achieves state-of-the-art driving performance while reducing parameters by " + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "text", + "content": " (from 7B to 1.3B), yielding substantial driving score improvements of " + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "inline_equation", + "content": "15.4\\%" + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "inline_equation", + "content": "16.8\\%" + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 55, + 270, + 296, + 582 + ], + "type": "text", + "content": " at tiny, short, and long distances, respectively, in closed-loop evaluations. Code is available at https://github.com/ReaFly/VLDrive." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "content": "Autonomous driving has experienced significant development in recent years, evolving from modular and segregated design to integrated end-to-end networks [10, 15, 26, 29, 46]. However, these networks still rely on specific inputs such as target points or action commands to guide their driving behaviors, significantly constraining the interaction" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 244, + 438, + 373 + ], + "blocks": [ + { + "bbox": [ + 318, + 244, + 438, + 373 + ], + "lines": [ + { + "bbox": [ + 318, + 244, + 438, + 373 + ], + "spans": [ + { + "bbox": [ + 318, + 244, + 438, + 373 + ], + "type": "image", + "image_path": "91a5ebeb0cdffd587f73890074dcde367cf1bcfbe283f9832cd91c0018624c7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. (a) Existing methods for language-grounded driving. (b) Our proposed VLDrive: a novel framework featuring a lightweight MLLM architecture with enhanced vision components. (c) Driving failure analysis of existing methods based on three evaluation runs. (d) Performance comparison between VLDrive and both versions of LMDrive, highlighting our method's superior driving performance with fewer parameters." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 443, + 244, + 549, + 377 + ], + "blocks": [ + { + "bbox": [ + 443, + 244, + 549, + 377 + ], + "lines": [ + { + "bbox": [ + 443, + 244, + 549, + 377 + ], + "spans": [ + { + "bbox": [ + 443, + 244, + 549, + 377 + ], + "type": "image", + "image_path": "14a09f89cdb435b1ef70942c80fd7c78129a403c0bd80dfe5fe92f722a781d4c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 534, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 534, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 534, + 498 + ], + "type": "text", + "content": "with humans and applicability in real-world scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "content": "The emergence of large language models (LLMs) [18, 33, 49, 50] has catalyzed a revolution in autonomous driving. Impressed by their advanced cognition and logical reasoning capabilities, a multitude of studies have integrated LLMs into autonomous driving systems and achieved noteworthy results [28, 37, 38, 41]. Among them, LMDrive [28] successfully achieves human-friendly language-grounded autonomous driving, where vehicle behaviors are solely guided by natural language instructions (see Fig. 1(a)), significantly enhancing human-vehicle interaction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "However, despite remarkable advancements, these approaches encounter significant practical deployment challenges, characterized by suboptimal driving performance and excessive computational parameters. To address these issues, we conduct a thorough investigation of driving failure cases in current approaches, coupled with performance analysis across different LLM scales, yielding two critical insights: 1) Deficiencies in visual comprehension" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 239, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 239, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 239, + 713 + ], + "type": "text", + "content": "*Corresponding author: liguanbin@mail.sysu.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5923" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "content": "are the predominant cause of driving failures. We analyze the failure cases of existing language-grounded autonomous driving models and identify two primary causes: instruction misunderstanding, manifested as the agent deviating from the route; and visual misunderstanding, characterized by failures due to collisions or blocks involving vehicles or layouts. The results in Fig. 1(c) reveal that limitations in visual comprehension frequently emerge as a critical factor contributing to driving task failures. 2) Driving performance does not scale linearly with LLM's parameter count. Through comprehensive evaluations across three benchmarks, we compare the full-scale LMDrive (7B parameters) with its lightweight counterpart, LMDrive-LiteLM (1.3B trainable parameters). As demonstrated in Fig. 1(d), LMDrive-LiteLM achieves comparable driving performance despite its significantly reduced parameter count. These findings strongly suggest that massive parameter scales in LLMs may not be a critical prerequisite for achieving superior driving performance. The insights derived from these analyses illuminate the path toward developing a parameter-efficient yet high-performing alternative for language-grounded driving applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 339, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 295, + 613 + ], + "type": "text", + "content": "Building on these considerations, in our work, we propose VLDrive, a vision-augmented lightweight language model for efficient language-grounded autonomous driving, as showcased in Fig. 1(b). We enhance the model's visual comprehension by drawing inspiration from human driving competencies, focusing on two key aspects: prioritized attention allocation to sparse but critical agents and objects, and temporal trajectory memory for behavioral sequence prediction. These enhancements are achieved through our proposed Cycle-consistent Dynamic Visual Pruning (CCDP) and Memory-enhanced Feature Aggregation (MEFA) mechanisms. Specifically, CDP framework consists of token sparsification and training-only token reconstruction. The sparsification mechanism enables adaptive selection of critical visual tokens and dynamic adjustment of token count based on driving scenario complexity. The reconstruction process preserves information integrity by explicitly recovering features of pruned tokens, enabling robust visual perception while minimizing computational overhead. Additionally, MEFA incorporates a memory bank of adjacent frames to capture temporal dependencies, providing critical temporal cues for visual feature enhancement." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 618, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 714 + ], + "type": "text", + "content": "Furthermore, we introduce a Distance-decoupled Instruction Attention (DDIA) mechanism to enhance instruction comprehension capabilities of lightweight language models. This design stems from our observation that increasing numbers of long-range visual tokens may dilute instruction attention (see Fig. 6), potentially leading to instruction misunderstanding. DDIA alleviates this issue and enhances joint visual-linguistic feature learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "and alignment, resulting in more accurate trajectory planning. Extensive closed-loop experiments conducted on the CARLA [6] simulation platform demonstrate the effectiveness of our proposed method, surpassing state-of-the-art approaches by " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "15.4\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "16.8\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " in driving scores at tiny, short and long distances, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 324, + 145, + 518, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 145, + 518, + 156 + ], + "spans": [ + { + "bbox": [ + 324, + 145, + 518, + 156 + ], + "type": "text", + "content": "To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 158, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 158, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 553, + 252 + ], + "type": "text", + "content": "- We identify that visual comprehension deficiencies constitute the primary bottleneck in language-grounded driving, with performance scaling non-linearly with language model size. This observation motivates our proposed paradigm: a lightweight language model enhanced by vision-augmented strategies, which simultaneously reduces computational overhead and improves driving performance, facilitating practical deployment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 253, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 348 + ], + "type": "text", + "content": "- We propose VLDrive, a lightweight architecture enhanced with vision-centric strategies for language-grounded driving. The framework incorporates CCDP for adaptive visual signal extraction and MEFA for temporal information integration. Additionally, DDIA facilitates visual-language alignment for robust navigation instruction following. These complementary strategies collectively enable more reliable autonomous driving." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 349, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 553, + 408 + ], + "type": "text", + "content": "- We evaluate our approach through closed-loop simulation experiments on standard language-grounded driving benchmarks using the CARLA platform. VLDrive achieves state-of-the-art performance while using significantly fewer parameters." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 421, + 400, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 400, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 400, + 433 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 493, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 493, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 493, + 453 + ], + "type": "text", + "content": "2.1. End-to-End Autonomous Driving" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 662 + ], + "type": "text", + "content": "Early end-to-end autonomous driving methods can be mainly divided into two categories: imitation learning [4, 8, 42] and reinforcement learning [2, 16, 32, 47]. Subsequently, benefiting from the robust feature modeling capabilities of the Transformer [36], researchers introduced additional modalities to provide complementary information for autonomous driving [24, 29, 45]. For example, Transfuse [24] uses Transformer to merge features from RGB and LiDAR bird's eye view (BEV) images to enhance the global comprehension of 3D scenes. InterFuse [29] improves the security and interpretability of the model by directly presenting its intermediate features and limiting actions to specified safe sets. In recent years, modular end-to-end planning approaches [1, 10, 26] have gained increased attention, where all components are connected and jointly optimized towards the ultimate goal, further improving the driving performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "type": "text", + "content": "2.2. LLMs in Autonomous Driving" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "As LLMs continue to evolve and flourish, an increasing number of researchers are incorporating them into au" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5924" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 491 + ], + "type": "text", + "content": "tonomous driving tasks, broadening the scope and capabilities of these systems. Among them, a notable strand of research [22, 23, 30] employs LLMs for driving-related visual question answering, providing enhanced interpretability for driving decisions. Another line of work seeks to fully exploit the powerful logical reasoning capabilities of LLMs to fundamentally transform the research paradigm in autonomous driving tasks [5, 7, 20, 21]. Specifically, GPT-Driver [20] feeds LLM with observations and egostates as language prompts and generates a planned trajectory and corresponding decision-making process in a natural language format. Agent-Driver [21] builds an LMM-powered agent, capitalizing on a tool library, a cognitive memory, and a reasoning engine for human-like autonomous driving. Recently, LLMs have been extended to visual modality successfully, making it possible for LLMs to understand the visual information of autonomous driving. Chen et al. [3] introduces a multimodal LLM to process vectorized numeric modalities data and produce driving-related question answers as well as action decisions. DriveGPT4 [41] curates a visual instruction tuning dataset for interpretable autonomous driving, taming LLM to generate text responses and low-level control signals. Further, based on the CARLA simulation platform, many studies [28, 37, 38] delve deeper and perform closed-loop evaluation driven by LLMs directly. For instance, LMDrive [28] takes multi-modality data and navigation instructions as input, achieving language-grounded driving and enhanced interaction with humans. Despite the impressive performance powered by LLMs, their considerably heavy parameters pose significant deployment challenges. In this work, we demonstrate that a lightweight language model can effectively respond to instructions and execute driving tasks through efficient visual representation extraction and strengthened vision-language alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 503, + 156, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 503, + 156, + 514 + ], + "spans": [ + { + "bbox": [ + 58, + 503, + 156, + 514 + ], + "type": "text", + "content": "2.3. Token Reduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 522, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 712 + ], + "type": "text", + "content": "Transformer has achieved substantial performance improvements across a range of tasks, yet it faces criticism for its quadratic computational costs relative to the number of input tokens. To tackle this issue, numerous studies focus on reducing token counts [12, 25, 39, 40]. For instance, DynamicViT [25] adaptively prunes redundant tokens at different Transformer layers to achieve token sparsification. SPViT [12] introduces an attention-based multihead token selector with a soft pruning strategy. Unlike direct pruning methods, this approach consolidates selected redundant tokens into a single package token, enhancing efficiency while preserving essential information. In the LLM era, token reduction is also a crucial strategy to reduce the heavy computation costs [13, 14, 27, 50]. Most existing works utilize Q-former [13] to aggregate visual features into a fixed number of tokens [13, 43, 50]. LLaMa" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 204 + ], + "type": "text", + "content": "VID [14] leverages average pooling followed by a learnable projector to compress the video frame representation. LLaVA-PruMerge [27] proposes an adaptive token selection via outlier detection, paired with a cluster-based merging strategy to supplement the information of pruned tokens to kept ones. Distinct from prior approaches, we present a cycle-consistent dynamic pruning strategy that integrates token reconstruction with token sparsification at the training phase. This design not only enables more representative token selection but also ensures that the retained tokens aggregate fine-grained global visual information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 217, + 369, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 369, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 369, + 228 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 236, + 380, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 236, + 380, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 380, + 247 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 255, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 314 + ], + "type": "text", + "content": "This task aims to predict driving actions by leveraging sequences of multi-sensor data and corresponding human navigation instructions. As shown in Fig. 2, our method follows the common framework of a multi-modal large language model (MLLM) and comprises three main components:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "text", + "content": "Visual Encoder: Given a sequence of visual data " + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_i, i = 1,2,\\dots,T\\}" + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "text", + "content": ", where each frame " + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "text", + "content": " includes multi-view camera images and corresponding LiDAR data. A visual encoder processes these multi-modal inputs to produce a unified feature representation " + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "text", + "content": " for each frame, with " + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 316, + 315, + 553, + 386 + ], + "type": "text", + "content": " representing the total number of tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 387, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 518 + ], + "type": "text", + "content": "Connector: As a bridge to align visual features with the language space, the connector plays a remarkable role in MLLM. To endow our model with adaptive visual perception and long-context modeling capability, as illustrated in Fig. 3, we tailor cycle-consistent dynamic visual pruning (CCDP), consisting of token sparsification and reconstruction, and memory-enhanced feature aggregation (MEFA) for the connector, significantly enhancing the visual representation. We denote the improved features of each frame by the connector as " + }, + { + "bbox": [ + 316, + 387, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v\\in \\mathbb{R}^{N_v\\times C_t}" + }, + { + "bbox": [ + 316, + 387, + 553, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 316, + 387, + 553, + 518 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 316, + 387, + 553, + 518 + ], + "type": "text", + "content": " is the language model's hidden dimension." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 519, + 553, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 519, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 316, + 519, + 553, + 649 + ], + "type": "text", + "content": "Lite Language Model: A lightweight language model, enhanced by our proposed distance-decoupled instruction attention (DDIA), is utilized to process the combination of visual tokens of all frames, " + }, + { + "bbox": [ + 316, + 519, + 553, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v\\in \\mathbb{R}^{N_vT\\times C_t}" + }, + { + "bbox": [ + 316, + 519, + 553, + 649 + ], + "type": "text", + "content": " and corresponding text embedding " + }, + { + "bbox": [ + 316, + 519, + 553, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t\\in \\mathbb{R}^{N_t\\times C_t}" + }, + { + "bbox": [ + 316, + 519, + 553, + 649 + ], + "type": "text", + "content": " tokenized from navigation instructions. A following MLP leverages features from the language model and predicts the future trajectory, which is subsequently converted by PID controllers to produce the lateral steering action and the longitudinal acceleration action. We detail each component in the following subsections." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 659, + 468, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 468, + 672 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 468, + 672 + ], + "type": "text", + "content": "3.2. CCDP: Token Sparsification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 713 + ], + "type": "text", + "content": "Given the " + }, + { + "bbox": [ + 316, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 316, + 677, + 553, + 713 + ], + "type": "text", + "content": "-th frame's multi-modal features " + }, + { + "bbox": [ + 316, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 316, + 677, + 553, + 713 + ], + "type": "text", + "content": ", as in DynamicViT [25], we predict the probabilities for pruning and retaining each token. Specifically, the probabilities are cal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "5925" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 68, + 470, + 235 + ], + "blocks": [ + { + "bbox": [ + 129, + 68, + 470, + 235 + ], + "lines": [ + { + "bbox": [ + 129, + 68, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 129, + 68, + 470, + 235 + ], + "type": "image", + "image_path": "73a29ed2563d257a2553c1861c55906718986ce57fdaa8d012829250d576a228.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "text", + "content": "Figure 2. An overview of our proposed VLDrive framework. Given a sequence of visual data, our connector transforms each frame's raw visual features " + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "text", + "content": " into sparse yet informative representations " + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "text", + "content": " through two key components: CCDP: Token Sparsification and Memory-enhanced Feature Aggregation (MEFA). Subsequently, a lite language model augmented with Distance-decoupled Instruction Attention (DDIA) jointly processes tokenized navigation instructions " + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t" + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "text", + "content": " and temporal visual features " + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v = \\{\\mathbf{F}_i^v \\mid i = 1, \\dots, T\\}" + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "text", + "content": ". The resulting hidden representations are fed into an MLP for trajectory prediction, followed by PID controllers that translate these predictions into concrete driving actions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity of " + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 55, + 248, + 555, + 326 + ], + "type": "text", + "content": " via explicit token reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 371 + ], + "type": "text", + "content": "ulated based on the combined global and local features " + }, + { + "bbox": [ + 55, + 346, + 295, + 371 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{G}_i;\\mathbf{L}_i\\right]" + }, + { + "bbox": [ + 55, + 346, + 295, + 371 + ], + "type": "text", + "content": ", which are obtained by " + }, + { + "bbox": [ + 55, + 346, + 295, + 371 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(\\cdot)" + }, + { + "bbox": [ + 55, + 346, + 295, + 371 + ], + "type": "text", + "content": " operations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 379, + 295, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 379, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 122, + 379, + 295, + 395 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {i} = \\operatorname {M L P} \\left(\\mathbf {F} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times \\frac {C}{2}} \\tag {1}", + "image_path": "215f3b88c3f6b8cdfb3112ecca9073aad1dcb7c592bcacac7cb3ed47634744ed.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 402, + 294, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 402, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 125, + 402, + 294, + 418 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {i} = \\operatorname {A v g} \\left(\\mathbf {L} _ {i}\\right) \\in \\mathbb {R} ^ {1 \\times \\frac {C}{2}} \\tag {2}", + "image_path": "87ea57f5e78c127aa523492e447dece3b275e54f588f49da59be725c1408f343.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mathrm{Avg}(\\cdot)" + }, + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "text", + "content": " is average operation. Another MLP is utilized to predict the probability " + }, + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "inline_equation", + "content": "[\\cdot ;\\cdot ]" + }, + { + "bbox": [ + 55, + 423, + 296, + 458 + ], + "type": "text", + "content": " denotes concatenation with broadcast mechanism:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 467, + 294, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 467, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 92, + 467, + 294, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {i} = \\operatorname {S o f t m a x} \\left(\\operatorname {M L P} \\left(\\left[ \\mathbf {G} _ {i}; \\mathbf {L} _ {i} \\right]\\right)\\right) \\in \\mathbb {R} ^ {N \\times 2} \\tag {3}", + "image_path": "fe9a82f9c75b140531d462752faf8fce7d3f42eb276699c02ce3559f0f753ec6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 490, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 514 + ], + "type": "text", + "content": "In our implementation, we leverage Gumbel-Softmax to obtain the binary mask:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 522, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 522, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 89, + 522, + 295, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i} = \\text {G u m b e l - S o f t m a x} (\\mathbf {S} _ {i}) _ {* 1} \\in \\{0, 1 \\} ^ {N} \\tag {4}", + "image_path": "a9e5a59e9f0a842d03f9a9cca60fbac51198b1b2473999472922e19d91240a2d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "text", + "content": " denotes the token retention masks. As shown in Fig. 3, the kept tokens " + }, + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^k" + }, + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "text", + "content": " selected based on " + }, + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 55, + 545, + 296, + 593 + ], + "type": "text", + "content": " are delivered to the Memory-enhanced Feature Aggregation module, which is detailed in the next subsection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 600, + 266, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 266, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 266, + 613 + ], + "type": "text", + "content": "3.3. Memory-enhanced Feature Aggregation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Temporal reasoning is integral to our everyday driving scenarios. Human drivers can deduce the likely future direction or intentions of an approaching vehicle by analyzing its historical trajectory, and subsequently tailor their driving actions. To attain autonomous driving that matches the safety and reliability of human driving, we explicitly inject the temporal context into the current visual tokens. Specifically, for the current time step " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", we introduce a memory" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "text", + "content": "bank storing the adjacent " + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "text", + "content": " frames from the current one, denoted as " + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i = [\\mathbf{F}_{i - Z},\\mathbf{F}_{i - Z + 1},\\dots,\\mathbf{F}_{i - 1}]" + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i\\in \\mathbb{R}^{Z\\times N\\times C}" + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "text", + "content": " Temporal encoding " + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{TE}_i" + }, + { + "bbox": [ + 313, + 346, + 555, + 395 + ], + "type": "text", + "content": " is derived by an MLP utilizing the historical average data and current frame information:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 377, + 402, + 553, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 402, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 377, + 402, + 553, + 417 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {i} ^ {a v g} = \\operatorname {A v g} \\left(\\mathbf {B} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times C} \\tag {5}", + "image_path": "4d8c1945c870f7e586fe1bc36993b3e7fb9c3fb38d2a7cad6ddf208e3fb045dd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 381, + 425, + 553, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 425, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 381, + 425, + 553, + 439 + ], + "type": "interline_equation", + "content": "\\mathbf {T E} _ {i} = \\operatorname {M L P} \\left(\\left[ \\mathbf {F} _ {i}; \\mathbf {B} _ {i} ^ {\\text {a v g}} \\right]\\right) \\tag {6}", + "image_path": "07aa28870e5aaaeef7064a3bd74bebb249bb38d1b062653a30d349235a3fd81c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 479 + ], + "type": "text", + "content": "Afterwards, a query transformer (Q-former) [13] is introduced to aggregate the features of current frames enhanced by temporary encoding:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 487, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 487, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 367, + 487, + 553, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {v} = \\operatorname {Q - F o r m e r} \\left(\\mathbf {F} _ {i} ^ {k}, \\mathbf {F} _ {i} + \\mathbf {T E} _ {i}\\right) \\tag {7}", + "image_path": "4ddc5f07f1a4cc630e14e543b35743d12dfc53724c51b14fe24a217499c1213b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": "In contrast to the original Q-former, we implement two key modifications: 1) Instead of using learnable queries, we utilize the retained visual tokens " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^k" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": " as queries to probe and aggregate features. These retained tokens serve as significant anchors within the current frame, allowing for more effective feature aggregation. 2) We introduce a temporary embedding " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{TE}_i" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": " to the original visual features " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": ", which enhances the model's ability to focus on temporary-salient moving objects." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 618, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 712 + ], + "type": "text", + "content": "Finally, the generated " + }, + { + "bbox": [ + 313, + 618, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 313, + 618, + 556, + 712 + ], + "type": "text", + "content": " serves as part of the aggregated visual tokens across all frames, " + }, + { + "bbox": [ + 313, + 618, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_v" + }, + { + "bbox": [ + 313, + 618, + 556, + 712 + ], + "type": "text", + "content": ", which is then fed into the subsequent language model for trajectory prediction based on the corresponding navigation instructions. Additionally, we incorporate CCDP: Token Reconstruction as a training-only auxiliary task to further strengthen visual information integrity via explicit token reconstruction." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5926" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 69, + 512, + 235 + ], + "blocks": [ + { + "bbox": [ + 82, + 69, + 512, + 235 + ], + "lines": [ + { + "bbox": [ + 82, + 69, + 512, + 235 + ], + "spans": [ + { + "bbox": [ + 82, + 69, + 512, + 235 + ], + "type": "image", + "image_path": "8071276f0238c3bc1753601b9d9ae7510de61da5176c5c9d5fdcf90a838c7c22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "text", + "content": "Figure 3. A detailed illustration of our proposed connector. CCDP: Token Sparsification and Memory-enhanced Feature Aggregation are proposed to reduce token count while enhancing information density. CCDP: Token Reconstruction serves as a training-only task that further ensures the information integrity of the retained tokens. Red arrows " + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "inline_equation", + "content": "(\\rightarrow)" + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "text", + "content": " indicate the input and output paths of our connector." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 217, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 217, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 217, + 312 + ], + "type": "text", + "content": "3.4. CCDP: Token Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": "Through memory-enhanced feature aggregation, the generated " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": " ideally encapsulates all visual features of the current frame. To further enhance the connector's ability to select critical visual tokens and aggregate complete information, we expect to reconstruct features of those pruned tokens based on " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": ". Firstly, we add the enhanced tokens " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": " into the original features " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": " according to their original token positions, forming a new feature vector denoted as " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{MLP}(\\mathbf{F}_i^v), \\mathbf{F}_i \\rangle" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": ", where an MLP function is used to align the dimensions of " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": ". Secondly, drawing inspiration from the Masked Autoencoder [9], a learnable embedding " + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{e} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 317, + 296, + 484 + ], + "type": "text", + "content": " is introduced as a foundation for reconstructing the pruned tokens. Finally, we construct the input feature vector to be reconstructed through the following formula:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 491, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 491, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 83, + 491, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i} ^ {r e c} = \\left\\langle \\operatorname {M L P} \\left(\\mathbf {F} _ {i} ^ {v}\\right), \\mathbf {F} _ {i} \\right\\rangle \\cdot \\mathbf {M} _ {i} + \\mathbf {e} \\cdot \\left(1 - \\mathbf {M} _ {i}\\right) \\tag {8}", + "image_path": "9b852b8eaedb4f96f5f23e33f44c431bcafa6b7c20afb966051dd83a44c076a1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "content": "In this way, we incorporate the decision " + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "content": " into the computational graph of feature reconstruction, enabling the reconstruction loss to influence pruning decisions and compelling the model to retain the most critical tokens. Afterwards, a 4-layer Transformer block takes as input " + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^{rec}" + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "content": " and predicts the reconstructed results " + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i^{rec}" + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "content": ". A cycle consistency constraint (Eq. 11) is imposed to ensure that the retained tokens " + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i^v" + }, + { + "bbox": [ + 55, + 510, + 296, + 606 + ], + "type": "text", + "content": " preserve the integral information of the current video frame." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 612, + 272, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 272, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 272, + 625 + ], + "type": "text", + "content": "3.5. Distance-decoupled Instruction Attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "We introduce a lightweight language model for efficient autonomous driving, reducing both the number of parameters and computational resource requirements. Furthermore, distance-decoupled instruction attention (DDIA) is tailored for our life language model, significantly alleviating the attention dilution for navigation instructions in the long-context driving scenarios. Specifically, as shown in Fig. 4," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "content": "our proposed DDIA diverges from the traditional vanilla attention in two key aspects:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 324, + 555, + 612 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 324, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 384 + ], + "type": "text", + "content": "1): Causal-attention " + }, + { + "bbox": [ + 313, + 324, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 324, + 553, + 384 + ], + "type": "text", + "content": " Self-attention: Rather than adhering to the original causal setting, our self-attention within instructional segments allows each text token to access ample contextual information, thereby producing a more comprehensive representation of text features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": "2): Distance-dependent " + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": " Distance-decoupled: In contemporary research on LLM, position encoding (e.g. RoPE [31]) is recognized as a critical ingredient to explicitly inject positional information into input tokens and enhance the model's contextual modeling capabilities. However, position encodings may introduce challenges, particularly in long-range driving scenarios. As historical visual data accumulates, the distance between current visual tokens and instructional tokens increases. The feature disparity introduced by position encoding may negatively impact the attention sensitivity of visual tokens to instruction ones, resulting in trajectory predictions that deviate from the specified instructions. To address this problem, inspired by [19], we maintain the distance-dependent attention within the unimodal text or visual tokens, but make the cross-modal attention between instructions and visual tokens unaffected by the position encoding. Specifically, with " + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": " denoting the sets of visual and instruction tokens, taking the " + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": "-th position as an example, our DDIA is defined as:" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 319, + 616, + 553, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 616, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 319, + 616, + 553, + 694 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {D D I A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) _ {j} = \\frac {\\sum_ {k _ {i} \\in \\mathcal {I}} \\operatorname {s i m} (R (\\mathbf {q} _ {j}) , R (\\mathbf {k} _ {i})) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {I}} \\operatorname {s i m} (R (\\mathbf {q} _ {j}) , R (\\mathbf {k} _ {i}))}, \\text {i f} \\mathbf {q} _ {j} \\in \\mathcal {I}, \\\\ = \\frac {\\sum_ {k _ {i} \\in \\mathcal {I}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) \\mathbf {v} _ {i} + \\sum_ {k _ {i} \\in \\{\\mathcal {V} | < j \\}} s i m \\left(R \\left(\\mathbf {q} _ {j}\\right) , R \\left(\\mathbf {k} _ {i}\\right)\\right) \\mathbf {v} _ {i}}{\\sum_ {k _ {i} \\in \\mathcal {I}} s i m \\left(\\mathbf {q} _ {j} , \\mathbf {k} _ {i}\\right) + \\sum_ {k _ {i} \\in \\{\\mathcal {V} | < j \\}} s i m \\left(R \\left(\\mathbf {q} _ {j}\\right) , R \\left(\\mathbf {k} _ {i}\\right)\\right)}, \\text {i f} \\mathbf {q} _ {j} \\in \\mathcal {V} \\tag {9} \\\\ \\end{array}", + "image_path": "5b22c3fd51df25cf2d7a3b3ff4b402e0468fd68e48bedf5948f84c66815a434d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "content": " denotes RoPE position embedding," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5927" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 70, + 242, + 188 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 242, + 188 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 242, + 188 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 242, + 188 + ], + "type": "image", + "image_path": "7eafec198d985369c33fd865de5517d71b446904fbc53edf6750414cde1ec99c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 220 + ], + "type": "text", + "content": "Figure 4. A detailed illustration of our proposed Distance-decoupled Instruction Attention (DDIA)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl}{sim(\\mathbf{q}_j,\\mathbf{k}_i)} & {= \\exp (\\mathbf{q}_j^T\\mathbf{k}_i / \\sqrt{d})} \\end{array}" + }, + { + "bbox": [ + 55, + 239, + 296, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 239, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}| < j\\}" + }, + { + "bbox": [ + 55, + 239, + 296, + 264 + ], + "type": "text", + "content": " represents the causal subset within visual tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 272, + 168, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 168, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 168, + 284 + ], + "type": "text", + "content": "3.6. Objective Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "content": "The full training objective for " + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 289, + 296, + 313 + ], + "type": "text", + "content": "-th frame consists of two parts, formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 322, + 294, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 322, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 135, + 322, + 294, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {w a y}} + \\mathcal {L} _ {\\text {c y c}}, \\tag {10}", + "image_path": "649c2fe026d44020ac211968ece743bb0466e77244b75b681a89a5fce2ad5966.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {way }}" + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "text", + "content": " supervises trajectory prediction by minimizing the " + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "text", + "content": " norm between predicted and ground-truth paths [28], while " + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {cyc }}" + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "text", + "content": " facilitates our model to select the most discriminative tokens and aggregate more representative visual features, which is composed of two adversarial losses:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 411, + 294, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 411, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 119, + 411, + 294, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c y c} = \\lambda_ {1} \\mathcal {L} _ {p r u n} + \\lambda_ {2} \\mathcal {L} _ {r e c} \\tag {11}", + "image_path": "a4d2c9f0baaab190f9ed21b5d9f93cc51d8826b34d870abe19758fd74badd700.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 433, + 295, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 433, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 295, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r u n} = \\left(R - \\frac {1}{N} \\sum_ {j = 1} ^ {N} \\mathbf {M} _ {i j}\\right) ^ {2} \\tag {12}", + "image_path": "e471f4e206a183015e6f4de7aef690cc27497c75eafdebb2a246b3fb7023f4c0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 472, + 294, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 472, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 100, + 472, + 294, + 505 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\sum_ {j = 1} ^ {N} (1 - \\mathbf {M} _ {i j}) \\left(\\mathbf {F} _ {i j} - \\hat {\\mathbf {F}} _ {i j} ^ {r e c}\\right) ^ {2} \\tag {13}", + "image_path": "77d7845ee21cce6456e77acfb64e271ff9369d6a5b4c7ff9a33102d107e26ca2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{prun}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " restricts the ratio of kept tokens to our pre-defined value " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " reconstructs pruned tokens from kept ones after memory-enhanced feature aggregation. " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " balance the weights of two loss items. " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{ij}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{ij}^{rec}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_i^{rec}" + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 593, + 137, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 137, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 137, + 605 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 612, + 173, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 173, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 173, + 625 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Datasets. We train our model on the official language-driven autonomous driving dataset [28], which includes 64K instruction-following data clips collected across 8 towns on CARLA [6] simulation environment. Each clip not only equips multi-sensor input data (multi-view camera images and LiDAR data), but also introduces aligned navigation instructions in the natural language format." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "Implementation Details. We adopt the visual encoder in LMDrive [28], producing 106 visual tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "(N = 106)" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " and kept it fixed during training. Two configurations of lightweight language models are integrated into our framework: LLaMA* (a 4-layer lite version of LLaMA [34]) and TinyLLaMA [44]. The hype parameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " determining the ratio of kept tokens and " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " denoting the number of stored adjacent frames in the memory bank are set to 0.3 and 10, respectively. " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " in the loss function (Eq. 11) are configured to 10 and 1. More details are provided in the Supplementary Material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "text", + "content": "Evaluation Metrics. As in LMDrive [28], we evaluate our model on three benchmarks, categorized by route length: LangAuto (" + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "inline_equation", + "content": ">500\\mathrm{m}" + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "text", + "content": "), LangAuto-Short (" + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "inline_equation", + "content": "150\\mathrm{m}-500\\mathrm{m}" + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "text", + "content": ") and LangAuto-Tiny (" + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "inline_equation", + "content": "<150\\mathrm{m}" + }, + { + "bbox": [ + 313, + 220, + 555, + 376 + ], + "type": "text", + "content": ") and employ route completion (RC), infraction score (IS), and driving score (DS) to assess close-looped driving performance. Specifically, route completion measures the extent to which a specified route is completed. Infraction score quantifies driving safety, starting from an ideal 1.0 base score and decrementing proportionally in the event of collisions or traffic rule violations. Driving Score is calculated as the product of RC and IS, offering a comprehensive evaluation of overall performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 383, + 516, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 516, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 516, + 396 + ], + "type": "text", + "content": "4.2. Comparisons with LLM-based Agents" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": "We compare our VLDrive with LMDrive, a pioneering LLM-driven, closed-loop autonomous driving method, across several distance benchmarks. Table 1 presents the comparison results on the LangAuto benchmark. VLDrive has approximately " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": " fewer parameters yet demonstrates superior driving performance compared to LMDrive, which is equipped with various LLM configurations. Specifically, VLDrive-LLaMA* achieves driving scores of " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "41.7\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": " and route completion rates of " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "52.6\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": ", outperforming LMDrive with LLaVA-v1.5 by " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": ". This trend continues on the LangAuto-Short and LangAuto-Tiny benchmarks, as shown in Table 2, where VLDrive-LLaMA* attains driving scores of " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "63.6\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "78.4\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": ", surpassing LMDrive by " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "11.9\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": ". VLDrive-TinyLLaMA achieves new state-of-the-art performance, boosting the DS score to " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "43.8\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "67.4\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "inline_equation", + "content": "81.9\\%" + }, + { + "bbox": [ + 313, + 400, + 555, + 605 + ], + "type": "text", + "content": " on the LangAuto, LangAuto-Short, and LangAuto-Tiny benchmarks, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 612, + 408, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 408, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 408, + 624 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "content": "To investigate the impact of various components of our method, we conduct comprehensive ablation studies on the standard LangAuto benchmark. To reduce time costs and enhance training efficiency, we randomly extract approximately " + }, + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "content": " of the training data from each town, forming a mini-training set for all ablation experiments. We present the metrics from 3 evaluation runs." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5928" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 120, + 102, + 491, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto benchmark. We present the metrics from 3 evaluation runs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 120, + 102, + 491, + 304 + ], + "lines": [ + { + "bbox": [ + 120, + 102, + 491, + 304 + ], + "spans": [ + { + "bbox": [ + 120, + 102, + 491, + 304 + ], + "type": "table", + "html": "
MethodLLMParams(Total)LangAuto
DS ↑RC ↑IS ↑
LMDriveLarge Language Model
LLaMA [34]7B31.3±1.537.1±1.60.82±0.01
LLaMA2 [35]7B32.8±2.140.1±2.20.81±0.02
Vicuna [48]7B33.5±1.939.3±1.90.83±0.02
Vicuna-v1.5 [48]7B34.0±3.839.0±3.30.85±0.06
LLaVA-v1.5 [17]7B36.2±2.346.5±4.30.81±0.03
Lightweight Language Model
Phi-2 [11]3B22.3±0.128.9±1.00.82±0.02
TinyLLaMA [44]1.3B25.2±2.338.6±3.70.71±0.02
OursLLaMA*1.3B41.7±1.852.6±2.50.81±0.02
TinyLLaMA [44]1.3B43.8±2.454.5±3.00.84±0.02
", + "image_path": "7132032a8c4c5a6c04b31b9a9ffca549fa9e07f60787fe7ce964af532febbe2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 345, + 555, + 527 + ], + "blocks": [ + { + "bbox": [ + 55, + 313, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 335 + ], + "type": "text", + "content": "Table 2. Performance comparison of our method with LMDrive incorporated various LLM backbones on the LangAuto-Short and LangAuto-Tiny benchmarks. We present the metrics from 3 evaluation runs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 345, + 555, + 527 + ], + "lines": [ + { + "bbox": [ + 58, + 345, + 555, + 527 + ], + "spans": [ + { + "bbox": [ + 58, + 345, + 555, + 527 + ], + "type": "table", + "html": "
MethodLLMParams(Total)LangAuto-ShortLangAuto-Tiny
DS↑RC↑IS↑DS↑RC↑IS↑
LMDriveLarge Language Model
LLaMA [34]7B42.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [35]7B44.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [48]7B45.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [48]7B47.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [17]7B50.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
Lightweight Language Model
Phi-2 [11]3B48.0±3.655.2±4.10.87±0.0156.3±4.968.2±3.90.82±0.04
TinyLLaMA [44]1.3B46.2±3.459.7±3.40.79±0.0264.1±1.275.0±1.40.86 ± 0.01
OursLLaMA*1.3B63.6±4.177.6±4.10.81±0.0678.4±4.685.6±2.40.91±0.04
TinyLLaMA [44]1.3B67.4±2.078.1±2.30.85±0.0281.9±0.985.5±0.60.94±0.02
", + "image_path": "867bf80c72ab6cdb5d82afcebd28794fab5bafa18c00e278083b01287f07f9dc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 546, + 294, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 546, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 54, + 546, + 294, + 679 + ], + "type": "text", + "content": "Effectiveness of Each Individual Component. As shown in Table 3, we begin our ablation study from the baseline, which utilizes " + }, + { + "bbox": [ + 54, + 546, + 294, + 679 + ], + "type": "inline_equation", + "content": "\\mathsf{LLaMA^{*}}" + }, + { + "bbox": [ + 54, + 546, + 294, + 679 + ], + "type": "text", + "content": " as the trainable language model and Q-Former [13] with learnable queries as the connector. Our method improves the baseline by: (1) introducing CCDP to extract critical sparse tokens, (2) replacing the vanilla Q-Former with our proposed MEFA module, and (3) substituting the causal attention mechanism with our DDIA module. The results in Table 3 validate the contribution of each component, with their synergistic integration leading to substantial improvements in driving performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 546, + 555, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 691 + ], + "type": "text", + "content": "Effectiveness of CCDP. We compare our CCDP approach with other token reduction strategies, including: 1) Structural pooling, a token reduction technique based on average pooling operations. 2) Dynamic pruning without token reconstruction. We maintain a consistent token retention ratio across all methods. As shown in Table 4, benefiting from the cycle-consistency constraints, CCDP is able to extract the most critical visual information and consequently achieves the highest driving scores. Fig. 5 further illustrates the positive correlation between reconstruction loss and trajectory prediction loss on the open-looped validation set, with a pearson correlation coefficient of 0.65." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5929" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 91, + 284, + 181 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 277, + 82 + ], + "type": "text", + "content": "Table 3. Ablation studies of each individual component." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 91, + 284, + 181 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 284, + 181 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 284, + 181 + ], + "type": "table", + "html": "
MethodDS ↑RC ↑IS ↑
Baseline30.0±0.846.2±3.50.69±0.03
Ours39.8±1.749.1±1.80.83±0.02
w/o CCDP35.5±3.845.3±3.50.78±0.02
w/o MEFA35.9±0.846.9±1.30.81±0.02
w/o DDIA36.3±1.347.0±1.00.81±0.02
", + "image_path": "d38e90891884a209c1bb5817ef11bb21d7bdb75c340925d175bd404147895330.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 211, + 286, + 270 + ], + "blocks": [ + { + "bbox": [ + 61, + 191, + 288, + 203 + ], + "lines": [ + { + "bbox": [ + 61, + 191, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 288, + 203 + ], + "type": "text", + "content": "Table 4. Ablation studies of various token reduction strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 211, + 286, + 270 + ], + "lines": [ + { + "bbox": [ + 66, + 211, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 66, + 211, + 286, + 270 + ], + "type": "table", + "html": "
MethodDS ↑RC ↑IS ↑
Structural Pooling33.6±2.445.8±4.30.78±0.06
Dynamic Pruning36.3±2.648.0±2.30.76±0.03
CCDP (Ours)39.8±1.749.1±1.80.83±0.02
", + "image_path": "73f42340f332c0d0f82b8c08385827b833745b670222274ce8652ef984d59960.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 66, + 300, + 285, + 367 + ], + "blocks": [ + { + "bbox": [ + 67, + 280, + 283, + 290 + ], + "lines": [ + { + "bbox": [ + 67, + 280, + 283, + 290 + ], + "spans": [ + { + "bbox": [ + 67, + 280, + 283, + 290 + ], + "type": "text", + "content": "Table 5. Ablation studies of different token retention ratios." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 300, + 285, + 367 + ], + "lines": [ + { + "bbox": [ + 66, + 300, + 285, + 367 + ], + "spans": [ + { + "bbox": [ + 66, + 300, + 285, + 367 + ], + "type": "table", + "html": "
Retention Ratio (R)DS ↑RC ↑IS ↑
5%33.4±0.445.1±3.20.77±0.03
10%35.6±1.747.3±1.00.79±0.02
30%39.8±1.749.1±1.80.83±0.02
50%39.8±1.251.4±3.70.79±0.04
", + "image_path": "dbef7ee27d0e8c1f1f2dac2ac1f4523bc1ab1ac020e3e30402d8e2e524a1bd57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 66, + 399, + 286, + 461 + ], + "blocks": [ + { + "bbox": [ + 69, + 377, + 282, + 389 + ], + "lines": [ + { + "bbox": [ + 69, + 377, + 282, + 389 + ], + "spans": [ + { + "bbox": [ + 69, + 377, + 282, + 389 + ], + "type": "text", + "content": "Table 6. Ablation of memory bank with different capacity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 399, + 286, + 461 + ], + "lines": [ + { + "bbox": [ + 66, + 399, + 286, + 461 + ], + "spans": [ + { + "bbox": [ + 66, + 399, + 286, + 461 + ], + "type": "table", + "html": "
Capacity (Z)DS ↑RC ↑IS ↑
537.1±1.949.1±0.50.80±0.01
1039.8±1.749.1±1.80.83±0.02
2038.3±3.049.9±2.60.81±0.02
", + "image_path": "93f75b09c486cc12f35dbe86748508547279c16793153e4a48f2e6461c25d674.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 481, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 541 + ], + "type": "text", + "content": "Analysis of Token Sparsification. Table 5 demonstrates that, retaining " + }, + { + "bbox": [ + 55, + 481, + 296, + 541 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 481, + 296, + 541 + ], + "type": "text", + "content": " of the visual tokens captures critical clues for autonomous driving. Further increment in the number of tokens produces no significant performance gains in driving scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 639 + ], + "type": "text", + "content": "Effectiveness of MEFA. As illustrated in Table 6, incorporating a memory bank provides essential temporal information that aids in predicting the future trajectories and behaviors of the ego-car and other objects, thereby enhancing the reliability of autonomous driving. Furthermore, our experiments indicate that a capacity of 10 adjacent frames yields the most significant performance improvements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of DDIA. Fig. 6 shows the visual comparison of attention maps generated by vanilla causal attention and DDIA, where brighter colors indicate higher values. As highlighted in the red dashed box, DDIA enhances attention to instruction tokens, enabling visual tokens to better" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 72, + 539, + 178 + ], + "blocks": [ + { + "bbox": [ + 331, + 72, + 539, + 178 + ], + "lines": [ + { + "bbox": [ + 331, + 72, + 539, + 178 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 539, + 178 + ], + "type": "image", + "image_path": "1f5ca852d49d1f765b962ae472d39edff095754abf24cc84ed576e383ef7ced2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "type": "text", + "content": "Figure 5. Correlation analysis between reconstruction and trajectory prediction losses, revealing a significant positive relationship (Pearson's " + }, + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "type": "inline_equation", + "content": "r = 0.65" + }, + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 240, + 433, + 361 + ], + "blocks": [ + { + "bbox": [ + 334, + 240, + 433, + 361 + ], + "lines": [ + { + "bbox": [ + 334, + 240, + 433, + 361 + ], + "spans": [ + { + "bbox": [ + 334, + 240, + 433, + 361 + ], + "type": "image", + "image_path": "c03539343b4ea5741151e10ebdd4a7d39f1c709ea13b041d89c44e0573db91e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 370, + 555, + 414 + ], + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 414 + ], + "type": "text", + "content": "Figure 6. Visual comparison of attention maps generated by the language model equipped with vanilla causal attention and DDIA. We aggregate the attention weights within the instruction segment into the first four tokens for clearer visualization." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 239, + 535, + 361 + ], + "blocks": [ + { + "bbox": [ + 434, + 239, + 535, + 361 + ], + "lines": [ + { + "bbox": [ + 434, + 239, + 535, + 361 + ], + "spans": [ + { + "bbox": [ + 434, + 239, + 535, + 361 + ], + "type": "image", + "image_path": "3dcb27fd1417c95498778aee49bcf22b245c824ce6b4687e643607e7797e4f9b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 431, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 456 + ], + "type": "text", + "content": "integrate instructional information and thus make more accurate driving decisions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduce VLDrive, a vision-enhanced, lightweight model for language-grounded autonomous driving. To facilitate the practical deployment of autonomous vehicles, we streamline the heavy LLM into a more compact form while augmenting its visual modeling and vision-language alignment. Specifically, we employ cycle-consistent visual dynamic pruning to efficiently capture the most salient visual information for each frame, and incorporate a memory-enhanced feature aggregation that enriches critical temporal information to help the model comprehend historical trajectories and infer future movements. Additionally, a distance-decoupled instruction attention strategy is tailored for our lightweight language model, alleviating the attention dilution to navigation instructions and boosting the model's instruction-following capability. Extensive experiments demonstrate that our model, despite having significantly fewer parameters, achieves superior and robust driving performance across various benchmarks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5930" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "content": "This work is supported in part by the National Key R&D Program of China (2024YFB3908503), in part by the National Natural Science Foundation of China (62322608), in part by the Shenzhen Longgang District Science and Technology Innovation Special Fund (No. LGKCYLWS2023018), in part by the Futian Healthcare Research Project (No.FTWS002), and in part by the Shenzhen Medical Research Fund (No. C2401036). This work is also sponsored by CIE-Tencent Robotics X Rhino-Bird Focused Research Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 240, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 240, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 294, + 283 + ], + "type": "text", + "content": "[1] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 285, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 327 + ], + "type": "text", + "content": "[2] Raphael Chekroun, Marin Toromanoff, Sascha Hornauer, and Fabien Moutarde. Gri: General reinforced imitation and its application to vision-based autonomous driving. Robotics, 12(5):127, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 328, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 382 + ], + "type": "text", + "content": "[3] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 384, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 294, + 437 + ], + "type": "text", + "content": "[4] Felipe Codevilla, Eder Santana, Antonio M López, and Adrien Gaidon. Exploring the limitations of behavior cloning for autonomous driving. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9329-9338, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 438, + 294, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 294, + 493 + ], + "type": "text", + "content": "[5] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as you speak: Enabling human-like interaction with large language models in autonomous vehicles. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 902-909, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 494, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 494, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 494, + 294, + 536 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 537, + 295, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 295, + 592 + ], + "type": "text", + "content": "[7] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 910-919, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 593, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 295, + 658 + ], + "type": "text", + "content": "[8] Jeffrey Hawke, Richard Shen, Corina Gurau, Siddharth Sharma, Daniele Reda, Nikolay Nikolov, Przemysław Mazur, Sean Micklethwaite, Nicolas Griffiths, Amar Shah, et al. Urban driving with conditional imitation learning. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 251-257. IEEE, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[10] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "text", + "content": "[11] Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, Jyoti Aneja, Sebastien Bubeck, Caio César Teodoro Mendes, Weizhu Chen, Allie Del Giorno, Ronen Eldan, Sivakanth Gopi, et al. Phi-2: The surprising power of small language models. Microsoft Research Blog, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 239 + ], + "type": "text", + "content": "[12] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In European conference on computer vision, pages 620-640. Springer, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 3, 4, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "text", + "content": "[14] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. arXiv preprint arXiv:2311.17043, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 386 + ], + "type": "text", + "content": "[15] Ming Liang, Bin Yang, Wenyuan Zeng, Yun Chen, Rui Hu, Sergio Casas, and Raquel Urtasun. Pnpnet: End-to-end perception and prediction with tracking in the loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11553–11562, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "text", + "content": "[16] Xiaodan Liang, Tairui Wang, Luona Yang, and Eric Xing. Cirl: Controllable imitative reinforcement learning for vision-based self-driving. In Proceedings of the European conference on computer vision (ECCV), pages 584-599, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 442, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 475 + ], + "type": "text", + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "type": "text", + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "type": "text", + "content": "[19] Fan Ma, Xiaojie Jin, Heng Wang, Yuchen Xian, Jiashi Feng, and Yi Yang. Vista-llama: Reliable video narrator via equal distance to visual tokens. arXiv preprint arXiv:2312.08870, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "type": "text", + "content": "[20] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 623 + ], + "type": "text", + "content": "[21] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A language agent for autonomous driving. arXiv preprint arXiv:2311.10813, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[22] Ana-Maria Marcu, Long Chen, Jan Hunermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, et al. Lingoqa: Video question answering for autonomous driving. arXiv preprint arXiv:2312.14115, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[23] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for au" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5931" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "tonomous driving. arXiv preprint arXiv:2312.03661, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[24] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multi-modal fusion transformer for end-to-end autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7077-7087, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 207 + ], + "type": "text", + "content": "[25] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 284 + ], + "type": "text", + "content": "[26] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 414-430. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[27] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "type": "text", + "content": "[28] Hao Shao, Yuxuan Hu, Letian Wang, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. arXiv preprint arXiv:2312.07488, 2023. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 421 + ], + "type": "text", + "content": "[29] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 466 + ], + "type": "text", + "content": "[30] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. arXiv preprint arXiv:2312.14150, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "type": "text", + "content": "[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 513, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 295, + 567 + ], + "type": "text", + "content": "[32] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 623 + ], + "type": "text", + "content": "[33] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "text", + "content": "[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[35] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[37] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 249 + ], + "type": "text", + "content": "[38] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 305 + ], + "type": "text", + "content": "[39] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "type": "text", + "content": "[40] Yifan Xu, Zhijie Zhang, Mengdan Zhang, Kekai Sheng, Ke Li, Weiming Dong, Liqing Zhang, Changsheng Xu, and Xing Sun. Evo-vit: Slow-fast token evolution for dynamic vision transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2964-2972, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 414 + ], + "type": "text", + "content": "[41] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 555, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 449 + ], + "type": "text", + "content": "[42] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end simulated driving. In Proceedings of the AAAI conference on artificial intelligence, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "type": "text", + "content": "[43] Pan Zhang, Xiaoyi Dong Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Hang Yan, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "text", + "content": "[44] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "text", + "content": "[45] Ruifei Zhang, Xiangru Lin, Wei Zhang, Jincheng Lu, Xuekuan Wang, Xiao Tan, Yingying Li, Errui Ding, Jingdong Wang, and Guanbin Li. Interactive 3d object detection with prompts. In European Conference on Computer Vision, pages 140-157. Springer, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 658 + ], + "type": "text", + "content": "[46] Wei Zhang, Jiaming Li, Meng Xia, Xu Gao, Xiao Tan, Yifeng Shi, Zhenhua Huang, and Guanbin Li. Offsetnet: Towards efficient multiple object tracking, detection, and segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 659, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 555, + 713 + ], + "type": "text", + "content": "[47] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5932" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[49] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[50] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5933" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_content_list.json b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9afd5535bd78e577b18fd154bc51310392e06c4 --- /dev/null +++ b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_content_list.json @@ -0,0 +1,1425 @@ +[ + { + "type": "text", + "text": "VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior", + "text_level": 1, + "bbox": [ + 228, + 143, + 769, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xindi Yang $^{1*}$ , Baolu Li $^{2*}$ , Yiming Zhang $^{2}$ , Zhenfei Yin $^{4,5\\dagger}$ , Lei Bai $^{3\\dagger}$ , Liqian Ma $^{6}$ , Zhiyong Wang $^{5}$ , Jianfei Cai $^{1}$ , Tien-Tsin Wong $^{1}$ , Huchuan Lu $^{2}$ , Xu Jia $^{2\\dagger}$ $^{1}$ Monash University $^{2}$ Dalian University of Technology $^{3}$ Shanghai Artificial Intelligence Laboratory $^{4}$ Oxford University $^{5}$ The University of Sydney $^{6}$ ZMO AI", + "bbox": [ + 148, + 212, + 848, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 316, + 333, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video diffusion models (VDMs) have advanced significantly in recent years, enabling the generation of highly realistic videos and drawing the attention of the community in their potential as world simulators. However, despite their capabilities, VDMs often fail to produce physically plausible videos due to an inherent lack of understanding of physics, resulting in incorrect dynamics and event sequences. To address this limitation, we propose a novel two-stage image-to-video generation framework that explicitly incorporates physics with vision and language informed physical prior. In the first stage, we employ a Vision Language Model (VLM) as a coarse-grained motion planner, integrating chain-of-thought and physics-aware reasoning to predict a rough motion trajectories/changes that approximate real-world physical dynamics while ensuring the inter-frame consistency. In the second stage, we use the predicted motion trajectories/changes to guide the video generation of a VDM. As the predicted motion trajectories/changes are rough, noise is added during inference to provide freedom to the VDM in generating motion with more fine details. Extensive experimental results demonstrate that our framework can produce physically plausible motion, and comparative evaluations highlight the notable superiority of our approach over existing methods. More video results and code are available on our Project Page: https://madaoer.github.io/projects/physically_plausible Video_generation/.", + "bbox": [ + 104, + 345, + 485, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 104, + 766, + 233, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video diffusion models (VDMs) trained on large-scale video datasets have made remarkable progress in terms of realism, demonstrating significant potential for various content creation applications. Despite the absence of explicit geometric modeling, the generated videos still exhibit coherent spatial", + "bbox": [ + 104, + 790, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9ea32fa753f62ad355288a590600f15f3482c20677eaef9919f606bdafc3e54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 321, + 888, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c8ac1d0e4311bf54beed485a09ec0c48140d89ca69db2bda1ddec2cce87568c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 412, + 887, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2363cedd545e1a12a5482ed879a142d6a192d5c1705c1f5f34e5bcc79608a29e.jpg", + "image_caption": [ + "Figure 1. Existing commercial closed-source VDMs fail to generate physically plausible motion, whereas our video generation framework is able to achieve this by incorporating external physical prior knowledge." + ], + "image_footnote": [], + "bbox": [ + 517, + 505, + 887, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "relationships among objects, rich textured details, and realistic lighting effects, including reflections and shadows. Such qualities often make the generated videos nearly indistinguishable from real-world footages. This drives the research community to explore the potential of VDMs as world models. However, they still struggle with understanding the physical laws of the real world and generating videos that adhere to these principles.", + "bbox": [ + 509, + 679, + 890, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although existing VDMs can produce visually realistic videos, they fail to mimic the real-dynamic physical motions. As shown in Fig 1, even the current commercial closed-source VDMs struggle with the task of generating videos that conform to physical laws. PhyT2V [58] refines text prompts by incorporating detailed descriptions", + "bbox": [ + 509, + 797, + 890, + 886 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding author", + "bbox": [ + 127, + 872, + 370, + 885 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12360", + "bbox": [ + 480, + 936, + 519, + 949 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of physical processes to guide VDMs in generating physically plausible videos. However, despite being pretrained on internet-scale real-world video-text pairs, VDMs do not inherently understand physical laws. This limitation arises from the gap and ambiguity between text descriptions and the actual motion in the video[33]. Moreover, VDMs tend to overfit the training data rather than developing a general understanding of physical laws[21]. Inspired by the success of graphics-based physical rendering, some methods have guided VDMs to generate physically plausible videos using simulations from graphics engines[18, 29, 32, 56, 64]; however, these approaches rely on the physical effects that graphics engines can simulate and incur high computational costs.", + "bbox": [ + 104, + 106, + 480, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The gap and ambiguity between text and real-world motion makes it difficult to enable physically plausible video generation through detailed text descriptions alone. Moreover, it is challenging to gather scalable physical data for training due to the abstractness and diversity of physical phenomena. Consequently, a viable approach could be to model abstract physical laws as conditions for diffusion models. However, it is less practical to explicitly model the physics equation for every kind of motion. Instead, we resort to current large foundation models for their ability to \"understand\" basic physics[6] and reason about physical phenomena based on the knowledge they extract. For example, given two colliding balls, the Large Language Model (LLM) can approximately predict the paths of the balls after collision. Inspired by this observation, we propose a novel video generation framework that employs a Vision Language Model (VLM) to predict the path/change during a physics event, described by a given image and a text prompt.", + "bbox": [ + 104, + 314, + 480, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose VLIPP, a two-stage approach to incorporate physics as conditions into VDM, enabling the generation of physically plausible motion. In the first stage, the VLM serves as a coarse-level motion planner, while a VDM serves as a fine-level motion synthesizer. The idea of stage one is to utilize the chain-of-thought and the physics-aware reasoning of VLM planning to ensure that coarse-level motion trajectories approximately follow real-world physics dynamics. In stage two, we can generate fine-level motion using an image-to-video diffusion model conditioned by the approximated path/change planned by VLM from stage one. Note that the approximated path/changes are not in the level to tell the speed or acceleration of the motion. We choose an existing image-to-video model [7] to accept our coarse-level path/change, by injecting noise to the motion path during both the training and inference phases. Notably, during the VLM planning stage, generating entire physically plausible motion trajectories is not required. Instead, we leverage the generative priors of VDM to produce fine-level physically plausible videos based on coarse-level motion trajectories provided by the VLM. So that the detail-level motion such", + "bbox": [ + 104, + 580, + 480, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as speed, acceleration, and vibration are left to the VDM to synthesize.", + "bbox": [ + 511, + 106, + 888, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our physically plausible video generation framework with two major video physics benchmarks and achieved satisfactory results. Furthermore, we discuss and analyze multiple insightful design choices in our video generation framework, such as employing a motion planner tailored for different physics categories, and enhancing the robustness of diffusion model to noisy trajectories. Our contributions are summarized as follows:", + "bbox": [ + 511, + 138, + 888, + 253 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a novel image-to-video generation framework for generating physically plausible videos by leveraging the VLM and VDM priors, significantly outperforming the contemporary competitors.", + "2. We propose a novel chain-of-thought and physics-aware reasoning approach in VLM, along with random noise injection in the latent space during video generation, which effectively improves both the generation quality and physical plausibility.", + "3. We conduct a comprehensive experiments and user studies to demonstrate the effectiveness and generalization of our framework in physically plausible videos generation." + ], + "bbox": [ + 511, + 257, + 887, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 454, + 648, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Physically Plausible Visual Content Generation", + "text_level": 1, + "bbox": [ + 511, + 479, + 888, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generating physically plausible videos offers substantial value to real-world applications such as scientific simulations[43], robotics [4, 59], and autonomous driving [10, 48]. Traditional graphics pipelines rely on simulation systems to model physical phenomena [36, 42]. Inspired by these approaches, recent studies [18, 29] have performed dynamic simulations in image space based on physical engines. Furthermore, some methods [56, 64] incorporate physical priors into 3D representations to enable the synthesis of physically plausible motions. However, these rule-based or solver-based simulators face limitations in expressiveness, efficiency, generalizability, and parameter tuning. Furthermore, these simulators require significant expertise, rendering them inaccessible and unfriendly for users.", + "bbox": [ + 511, + 505, + 888, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, Some studies have explored VDMs for generating physically plausible videos. Li et al. [24] models natural oscillations and swaying in frequency-domain. A downstream rendering module then animates static images based on the generated motion information. PhysDiff [62] introduces physical simulator as constrain into the diffusion process by projecting denoised motion of a diffusion step into a physically plausible motion. These methods mainly focus only on specific types of physical motion and do not establish a generalizable approach for generating physically plausible videos.", + "bbox": [ + 511, + 726, + 888, + 883 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12361", + "bbox": [ + 480, + 936, + 517, + 948 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b5eb454f9f4faae16e9e8ffa131f4cd288ef84854939ec883963a44a26522811.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 104, + 885, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb5a4a524ec9ac71cf728e1b3b2e3cfb1c10c034f413437c6fdd8c1271602372.jpg", + "image_caption": [ + "Figure 2. The illustration of our physically plausible image-to-video generation pipeline. Our pipeline consists of two stages. In the first stage, the VLM generates a coarse-grained, physically plausible motion trajectory based on the provided input conditions. In the second stage, We simulate a draft video using the predicted trajectory to provide the motion condition. We then extract the optical flow from this video and convert it into structured noise. These conditions are fed into a motion controllable image-to-video diffusion model, and ultimately generates a physically plausible video." + ], + "image_footnote": [], + "bbox": [ + 109, + 222, + 885, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Motion Controllable Video Generation", + "text_level": 1, + "bbox": [ + 104, + 455, + 426, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing studies commonly provide one of the following three types of motion control: bounding box control [19, 23, 30, 46, 52], point trajectory control [34, 37, 40, 50, 54] and camera control [2, 15, 47, 61]. Bounding box control provides object motion guidance by generating a sequence of bounding boxes that track the object's position over time. Point-trajectory control offers motion cues through point-based trajectories, enabling drag-style manipulation. Camera motion control guides video generation using explicit 3D camera parameters, ensuring consistency and realistic viewpoint changes. However, these approaches prioritize motion control but often overlook physical plausibility. To address the limitation, we propose a novel framework for physically plausible video generation that incorporates physics as conditions into video diffusion models.", + "bbox": [ + 102, + 478, + 483, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Generation based on VLMs Planning", + "text_level": 1, + "bbox": [ + 104, + 704, + 415, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLMs have exhibited robust capabilities in visual understanding and planning [27, 39, 65]. Their strong performance in domains such as robot path planning and video understanding shows their ability in understanding the real physical world. Prior work has successfully leveraged LLMs to guide the layout of images or videos, yielding promising results [25, 53]. VideoDirectorGPT [26] leverages LLMs for fine-grained scene-by-scene planning, explicitly controlling spatial layout to generate temporally consistent long videos. Pandora [55] utilizes LLMs for real-time control through free-text action commands, achieving domain generality,", + "bbox": [ + 102, + 724, + 483, + 886 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "video consistency, and controllability. However, these efforts have yet to address interactions with real-world physical phenomena, such as collision, fall, and melting.", + "bbox": [ + 511, + 457, + 888, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, the absence of visual information can cause severe hallucination issues in language models for spatial planning tasks, leading to problems like overlapping object boundaries, disproportionate scaling, and incorrect planning [20, 57]. In this paper, we propose utilizing VLMs as coarse-level motion planners within the image space and incorporate physics-aware reasoning and Chain of Thought (CoT) [51] into the inference process.", + "bbox": [ + 511, + 502, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 636, + 599, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Formulation. In this paper, our goal is to enable an image-to-video diffusion model to generate physically plausible videos. Since VDMs rely more on memory and case-based imitation and struggle to understand general physical rules[21], the key challenge is how to incorporate physical laws into the models. To achieve this, we need to identify a method to incorporate physical principles into the video diffusion framework. Given an image $I \\in R^{H \\times W \\times C}$ ( $H$ is height, $W$ is width and $C$ is the number of channels) and a text description $d$ of possible events based on image $I$ , our framework should infer a physics-compliant guidance as the input condition and synthesize a video that adheres to both physical laws and real-world dynamics.", + "bbox": [ + 509, + 662, + 890, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall Pipeline. Overall pipeline of VLIPP is illustrated in Figure 2. In the first stage, the VLM conducts semantic", + "bbox": [ + 511, + 856, + 890, + 886 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12362", + "bbox": [ + 480, + 936, + 517, + 948 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d274a243e9f6ce3afbd45c5ffaab4344c224e10667fe5436391a2daed76f3c9e.jpg", + "image_caption": [ + "Figure 3. The illustration of chain-of-thought reasoning in the VLM for generating a coarse-grained motion trajectory. First off, the VLM determines the corresponding physical laws and its context for the given scene. Then, the VLM performs step-by-step reasoning to predict the physically plausible motions of objects in image space, leveraging physical context and chain-of-thought prompting. Finally, the VLM predicts bounding boxes according to real-world physics." + ], + "image_footnote": [], + "bbox": [ + 107, + 104, + 477, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "analysis and physical attribute analysis on the given image $I$ and a description $d$ to obtain the bounding boxes of the objects in the scene, denoted as $b_{1}, b_{2}, \\ldots, b_{n}$ , along with the applicable physical laws $l$ . Next, the VLM infers possible future physical scenarios in the current scene to derive the coarse-level motion trajectories of $p_{1}, p_{2}, \\ldots, p_{t}$ in the image space. Finally, we utilize an image-to-video diffusion model to synthesize the detailed dynamics in the video.", + "bbox": [ + 104, + 445, + 483, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. VLM as a Coarse-Level Motion Planner", + "text_level": 1, + "bbox": [ + 104, + 570, + 436, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our motivation is to incorporate physical laws as constraints into a video diffusion model to enhance the physical plausibility of the generated videos. To achieve this, we must identify a method to inject physical laws into the video diffusion model. Given a video description and the first frame, the task at this stage is to generate coarse-level motion trajectory aligned with physical laws.", + "bbox": [ + 104, + 590, + 482, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Understanding. In the real world, most physical phenomena arise from interactions among objects and their motion trajectories. We first initiate the process by identifying and locating objects within a scene. Inspired by the recent studies [1, 9, 28] in VLMs for scene understanding, we employ GPT-4o [38] to recognize all objects that could be involved in physical phenomena as described in the text description $d$ . These objects are subsequently detected and segmented using Grounded-SAM2 [41], yielding their bounding boxes. By leveraging the pretrained knowledge and common-sense reasoning capabilities of foundation models, we effectively determine the relevant objects in the scene.", + "bbox": [ + 104, + 696, + 482, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physical-Aware Recognition. To perform more effective reasoning in predicting the motion, it is necessary to determine what specific physical principle to apply in the given context. We utilize the pretrained prior of the LLM to determine the physical laws applicable to the current scene. Following the configuration in the physical benchmark [3, 31, 33], we currently classify common physical phenomena in videos into six categories: gravity, momentum conservation, optics, thermodynamics, magnetism, and fluid mechanics. Note that such list can be easily extended within our framework. Given a video description $d$ , the LLM infers the physical law $l$ that governs the current scene. We provide the specific physical context information for VLM to enhance its understanding of physical laws [11]. Detailed context design is presented in the Appendix.", + "bbox": [ + 511, + 106, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Chain of Thought Reasoning in VLM. Given the physical law $l$ , an image $I$ and a video description $d$ for the scene, we prompt the VLM to predict the future bounding box positions of objects within the image-space. We choose to predict in the image space for two primary reasons. Firstly, motion in image space aligns more with our subsequent video synthesizer. Secondly, image space dynamics can effectively represent a wide range of real-world motions [29].", + "bbox": [ + 511, + 335, + 890, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At a given time $t$ , the predicted position of $i$ -th object bounding box $b_{i}^{t}$ is denoted as $[x_{t}^{i}, y_{t}^{i}, w_{t}^{i}, h_{t}^{i}]$ , where $(x_{t}^{i}, y_{t}^{i})$ represents its top-left coordinate; $w_{t}^{i} \\& h_{t}^{i}$ denote its width and height, respectively. Governed by the physical law, the four values of the bounding box may change over time. The VLM reasons the bounding box positions of $N$ future frames for every object $o_{i}$ based on the condition. To help VLM better understand physical laws, we adapt a chain-of-thought [51] into its reasoning, to significantly enhance its reasoning capabilities. As shown in Figure 3, we formulate our analysis of physical phenomena in videos as step-by-step reasoning: beginning with broad conceptual ideas and progressing to a detailed and practical examination:", + "bbox": [ + 511, + 460, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Given the physical law $l$ and context information, the VLM analysis video caption and detail the physical law.", + "2. The VLM analyzes the potential interactions and movement of each object within the scene;", + "3. The VLM predicts the detailed changes in position and shape of the bounding box corresponding to each object over time." + ], + "bbox": [ + 511, + 657, + 888, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through the structured planning process, the VLM plans coarse-level motion trajectories for the objects, approximating real-world physics dynamics. In particular, our VLM infers the changes of object bounding boxes for next 12 frames, constrained by the token length limitation. To be compatible with the generation process of the chosen VDM in the next stage, these inferred 12 frames are further linearly interpolated to produce a total of 49 frames.", + "bbox": [ + 511, + 768, + 890, + 885 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12363", + "bbox": [ + 480, + 936, + 517, + 948 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. VDM Serves as a Fine-Level Motion Synthesizer", + "text_level": 1, + "bbox": [ + 104, + 104, + 483, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the previous stage, the motion trajectory planned by the VLM is neither precise nor fully compliant with physical laws. On the other hand, while VDM may not be able to produce realistic global motion trajectories, it is able to generate sound motion in finer scale. In this stage, our key insight is that the VDM can refine the coarse-level motion to produce physically plausible motion that aligns with real-world dynamics with its powerful generative prior.", + "bbox": [ + 104, + 141, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Animation. To incorporate physical laws into the video diffusion model, we use the inferred coarse motion trajectory to guide the generation process of the diffusion model. Optical flow provides a unified representation of motion, and recent studies [7, 12, 13] have demonstrated its effectiveness in guiding diffusion models. Accordingly, we leverage the coarse-level motion trajectory to animate a draft motion sequence and derive the corresponding optical flow. Specifically, for each object $o_i$ , we extract its bounding box from the first frame and move it to the bounding box location $b_i$ specified by the draft motion trajectory. To animate the change of shape (e.g., due to compression or expansion), we resize object $o_i$ according to the difference between $o_i$ and $o_{i+1}$ during inpainting. The draft motion video is generated as follows:", + "bbox": [ + 104, + 261, + 483, + 478 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} (t) = \\operatorname {A n i m a t i o n} \\left(B, r s \\left(o _ {0} ^ {0}, b _ {t} ^ {0}\\right)... r s \\left(o _ {0} ^ {i}, b _ {t} ^ {i}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 487, + 483, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{V}(t)$ denotes the corresponding frame of inpainted video at timestep $t$ , $B$ denotes the inpainted background with the foreground object removed, $o_i^0$ denotes the $i$ -th object at timestep 0, $b_i^t$ represents the $i$ -th bounding box at timestep $t$ , and $rs$ denotes resize function.", + "bbox": [ + 104, + 513, + 483, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Structured Noise from Draft Video. Optical flow is an effective representation for guiding VDMs [12, 13]. Follow prior work [7, 8], we employ RAFT[44] to extract optical flow from the draft video and formulate it as structural noise, which retains Gaussian properties. Given the draft video $\\hat{V}(t) \\in \\mathbb{R}^{F \\times C \\times H \\times W}$ , we calculate its per-frame optical flow to get a structured noise tensor $Q \\in \\mathbb{R}^{F \\times C \\times H \\times W}$ . The structured noise enables the VDM to generate videos that exhibit motion patterns closely aligned with those in the optical flow, thereby improving the realism of the output.", + "bbox": [ + 104, + 590, + 483, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Noise Injection in Video Synthesis. We adopt Go-with-the-Flow [7] as our video synthesis model, a fine-tuned CogVideoX [60], which is designed to accept structured noise $Q$ as input and synthesize videos that adhere to the implicit optical flow. The vanilla Go-with-the-Flow tends to tightly follow the provided structured noise $Q$ . However, our $Q$ is derived from a coarse-level motion trajectory and may not be sufficiently accurate to follow the physical laws of the real world. To address this limitation, we inject noise during the inference phase to give more flexibility to the VDM to", + "bbox": [ + 104, + 739, + 483, + 886 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "generate detail-level motion changes as", + "bbox": [ + 511, + 106, + 764, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} = \\frac {(1 - \\gamma) Q _ {i} + \\zeta \\gamma}{\\sqrt {(1 - \\gamma) ^ {2} + \\gamma^ {2}}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 132, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q_{i}$ is structured noise at $i$ -th frames, $\\zeta \\in \\mathbb{R}^{C\\times H\\times W}$ is Gaussian noise and $\\gamma \\in [0,1]$ . We set $\\gamma = 0.4$ for even frame index and $\\gamma = 0.6$ for odd frame index.", + "bbox": [ + 511, + 176, + 890, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this approach, the VDM is able to generate motion deviate from the coarse-level motion trajectory whenever necessary for producing high-quality fine-level motion.", + "bbox": [ + 511, + 223, + 890, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Empirical Analysis and Dissusion", + "text_level": 1, + "bbox": [ + 511, + 280, + 815, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to demonstrate the effectiveness of our video generation framework compared to existing methods. We evaluate our approach on two established benchmarks for physically plausible video generation. Our framework consistently achieves superior performance across all benchmarks.", + "bbox": [ + 511, + 306, + 890, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 402, + 723, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a two-stage physically plausible image-to-video generation framework. In the first stage, we utilize ChatGPT-4o as the coarse-level motion planner. In the second stage, we utilize an open-source I2V model, Go-with-the-Flow [7], as a fine-level motion synthesizer. Unless otherwise specified, in all experiments, we generate each video with a resolution of $720 \\times 480$ and 49 frames.", + "bbox": [ + 511, + 424, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Benchmarks and Models", + "text_level": 1, + "bbox": [ + 511, + 536, + 733, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional metrics in the visual domain, such as the Peak Signal-to-Noise Ratio (PSNR)[17], the Structural Similarity Index (SSIM)[49], the Learned Perceptual Image Patch Similarity (LPIPS)[63], the Fréchet Inception Distance (FID)[16] and the Fréchet Video Distance (FVD)[45], do not account for the physical realism of the generated videos [31, 33]. Recent studies have begun to address this limitation by developing benchmarks and metrics that evaluate physical realism. In this work, we adopt two benchmarks, described below.", + "bbox": [ + 511, + 556, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PhyGenBench [31] categorizes physical properties into four domains: mechanics, optics, thermal, and material. It includes 27 physical phenomena, each governed by real world physical laws, reflected in 160 carefully designed text prompts. As PhyGenbench provides only text prompts, we adapt it to our image-to-video setting by generating a corresponding first frame for each prompt with FLUX[22]. We adhere to the predefined benchmark evaluation protocol, i.e., employing GPT-4o to assess the physical realism of the generated videos.", + "bbox": [ + 511, + 707, + 890, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Physics-IQ [33] comprises 396 real-world videos spanning 66 distinct physical scenarios. For each scenario, videos", + "bbox": [ + 511, + 856, + 890, + 886 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12364", + "bbox": [ + 480, + 936, + 519, + 949 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7492de47fea042983fdf07f7c6ba6374f76eb600e81b4090eecd926e22a1d5aa.jpg", + "image_caption": [ + "Figure 4. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B [60], LTX-VideoI2V [14] and SVD-XT [5]." + ], + "image_footnote": [], + "bbox": [ + 109, + 103, + 890, + 371 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c14c0f5f913e5493008d72648b9669ee1dde7110d19aeb9f6df5935128e26070.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMechanics(↑)Optics(↑)Thermal(↑)Material(↑)Average(↑)
CogvideoX-T2V-5B0.430.550.400.420.45
LTX-Video-T2V0.350.450.360.380.39
OpenSora0.430.500.440.370.44
PhyT2V0.490.610.490.470.52
LLM-Grounding Video Diffusion0.320.410.260.240.31
CogvideoX-I2V-5B0.480.690.430.410.52
SVD-XT0.460.680.480.410.52
LTX-Video-I2V0.470.650.460.370.50
SG-I2V0.520.690.510.390.54
Ours0.550.710.600.530.60
", + "bbox": [ + 156, + 402, + 841, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of VDMs on PhyGenBench.", + "bbox": [ + 333, + 590, + 661, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are recorded from three different perspectives and filmed twice under identical conditions to eliminate randomness. This benchmark evaluates real-world physical phenomena, including collisions, object continuity, occlusion, object permanence, and fluid dynamics. This benchmark assesses physical realism from semantic and temporal perspectives, using semantic metrics and visual metrics to compare generated videos against the real-world reference videos.", + "bbox": [ + 104, + 621, + 483, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Models. In the context of text-to-video generation, we compare our framework with CogVideoX-T2V-5B [60], LTX-Video-T2V [14], and OpenSora [66]. Moreover, we evaluate our framework against PhyT2V [58], which enhances physical realism by iteratively refining the prompt. For the image-to-video generation scenario, our framework is evaluated alongside CogVideoX-I2V-5B [60], SVD-XT [5], and LTX-Video-I2V [14]. Additionally, we conducted experiments in the motion-controllable setting.", + "bbox": [ + 104, + 753, + 483, + 886 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this setting, we leverage the motion trajectory predicted by the VLM as a condition to guide VDM generation. We benchmark our approach against image-to-video motion controllable model, SG-I2V [35] and text-to-video motion controllable model LLM-grounded Video Diffusion Models [25]. The experimental details are presented in the Appendix.", + "bbox": [ + 511, + 621, + 890, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 733, + 725, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin with an empirical study on PhyGenBench and Physics-IQ, comparing our framework against widely adopted open-source models in the research community. Based on different physical properties, we categorize the benchmark samples accordingly. Additionally, we classify VDMs into text-to-video (T2V) diffusion models and imaged-to-video (I2V) diffusion models based on the input conditions.", + "bbox": [ + 511, + 753, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1, we present our experimental results on Phy", + "bbox": [ + 529, + 871, + 890, + 886 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12365", + "bbox": [ + 480, + 936, + 517, + 948 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/edcc135b74a423269c0a23d3e096c8c5b7e69d7bffa9ce6100cf6867aeb15014.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Cogvideo-I2V-5B30.429.816.713.38.527.1
SVD-XT21.920.56.88.417.119.1
LTX-Video-I2V30.229.815.913.28.426.8
SG-I2V34.631.215.913.18.429.7
Ours42.334.116.913.48.834.6
", + "bbox": [ + 155, + 103, + 841, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of physically plausible video generation on Physics-IQ Benchmark. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics.", + "bbox": [ + 104, + 205, + 888, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GenBench, evaluating different video generation models following its evaluation protocol. The results show that our framework achieves state-of-the-art performance across four different physical phenomena. Our framework outperforms the best T2V method by an average of $15.3\\%$ and the best I2V method by $11.1\\%$ . Specifically, our framework demonstrates significant advantages in the Mechanics, Thermal, and Material domains, outperforming the best I2V method by $5.7\\%$ , $17.6\\%$ , and $35.8\\%$ , respectively. These advantages are particularly evident in these three types of physical phenomena, which involve more substantial changes in motion, volume, or shape. Our framework is better equipped to understand and reason about bounding box sequences to represent these changes effectively.", + "bbox": [ + 104, + 250, + 483, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similarly, for the Physics-IQ benchmark, we evaluate the performance of different video generation models following its evaluation protocol. Our framework achieves the best results across four different physical phenomena, with improvements of $22.2\\%$ in Solid Mechanics and $9.2\\%$ in Fluid Dynamics compared to the second-best models. These significant improvements demonstrate the effectiveness of our framework in generating physically plausible videos.", + "bbox": [ + 104, + 464, + 483, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 104, + 636, + 307, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figures 4 and 5 demonstrate a qualitative comparison between our video generation framework and baseline methods. Among all evaluated approaches, our framework consistently produces videos with the highest degree of physical realism. In the ball falling sample in Figure 4, while CogVideoX shows a bouncing effect, artifacts are present in the video; LTX-Video and SVD-XT exhibit motions that do not adhere to the laws of physics. In Figure 5, we analyze two examples from Physics-IQ. In the pouring water example, the baseline methods fail to show the simultaneous decrease in water level of the glass beverage dispenser and the increase in water level of the glass below; in the ball collision example, none of the baseline methods correctly depict the collision of balls. More videos are provided in the supplement.", + "bbox": [ + 104, + 667, + 483, + 886 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26ec347cb5f5715b4193d855f46b0440b39bab2e8a47858a4a354fa123ceb8e9.jpg", + "image_caption": [ + "Figure 5. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B, LTX-Video-I2V, SVD-XT and SG-I2V [35] in the Physics-IQ dataset." + ], + "image_footnote": [], + "bbox": [ + 516, + 250, + 888, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 627, + 660, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform an ablation study to evaluate the contributions of key components in our framework. We design four variants to analyze the effectiveness of different components in our framework.", + "bbox": [ + 511, + 650, + 890, + 708 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Ours w/o VLM Planner: To assess the overall functionality of our framework, we replace the structured noise input of the VDM with random noise to evaluate the effectiveness of the VLM planner.", + "2. Ours w/o CI: Keeping the overall structure unchanged, we remove the in-context information from the VLM.", + "3. Ours w/o CoT: Similarly, while keeping other components unchanged, we remove the CoT reasoning process from the VLM.", + "4. Ours w/o CC: Lastly, we remove both the in-context information and the CoT reasoning process from the VLM planner while maintaining all other components." + ], + "bbox": [ + 511, + 710, + 888, + 886 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12366", + "bbox": [ + 480, + 936, + 517, + 948 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ad95e016cd723c327e68d07e31aea771baf4ab014be5ea053fff59f4733f6f4a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Ours42.334.116.913.48.834.9
w/o VLM Planner.16.320.813.45.85.616.2
w/o C.I26.328.116.911.28.424.3
w/o CoT21.426.916.18.66.921.0
w/o C.C18.722.414.97.26.118.1
", + "bbox": [ + 153, + 103, + 844, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on VLM, in-context learning and COT. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics.", + "bbox": [ + 125, + 205, + 867, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d28e9559c9aff9e594c3ff64214b2436dbcac8807ae8729f0dd5f9a6b9bb5e6.jpg", + "image_caption": [ + "Figure 6. Visualization of generated video with bounding boxes. Coarse motion cue can guide VDM to synthesis physically plausible video." + ], + "image_footnote": [], + "bbox": [ + 107, + 234, + 483, + 400 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a631e4ead90c7ffcdaa6d113c0af96248a053240a719a78fe6e70baa243ac2e.jpg", + "image_caption": [ + "Figure 7. Limitation of VLIPP. Our framework might produces physically implausible videos because it lacks 3D spatial perception and the optical-flow of small objects is prone to noise." + ], + "image_footnote": [], + "bbox": [ + 107, + 465, + 483, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 presents a quantitative comparison between our full method and these variants. Among all variants, Ours w/o VLM Planner shows the most significant performance drop, as removing the planner completely eliminates our ability in understanding the physical laws, leading to nearly random results. Notably, Ours w/o CoT exhibits a more pronounced decline compared to Ours w/o CI, indicating that the reasoning process in CoT enhances the understanding of physics. While in-context information contributes to the physical reasoning ability of VLM, compared to CoT it is less effective in preventing errors caused by VLM hallucination.", + "bbox": [ + 104, + 710, + 483, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitations", + "text_level": 1, + "bbox": [ + 511, + 236, + 633, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although VLIPP can generate physically plausible videos, its performance remains constrained by the base model.", + "bbox": [ + 511, + 262, + 888, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Firstly, we are unable to model physical events that cannot be represented by image space bounding box trajectories. For example, phenomena that involve intrinsic state changes of objects such as gas solidification. We adopt 2D bounding boxes because they can be naturally integrated with VDMs and are effective in enhancing the physical plausibility to a certain extent. In fact, they are more flexible than they might appear. As shown in Figure 6, the VLM predicts coarse 2D bounding box trajectories(only four fragments), yet the VDM still produces a realistic shattering pattern without being rigidly constrained by those paths. Moreover, in the butter melting example in Figure 6, the VLM offers only a coarse cue of volumetric change, but the VDM successfully generates the transition from solid to liquid. These abilities significantly improve our flexibility in modeling various physical phenomena.", + "bbox": [ + 509, + 296, + 888, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Secondly, our pipeline lacks 3D spatial perception. It is unable to understand the spatial relationships within the scene, as shown in Figure 7 domino example.", + "bbox": [ + 511, + 532, + 890, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, the optical flow of small objects is prone to noise interference. This will cause our framework to generate ambiguous content, as shown in Figure 7 small ball collision example. With the recent progress in video generation model, we anticipate that our framework will be further improved in generating videos under more challenging physical conditions.", + "bbox": [ + 511, + 582, + 890, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 710, + 627, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recently, VDMs have achieved great empirical success and are receiving considerable attention in computer vision and computer graphics. However, due to the lack of understanding of physical laws, VDMs are unable to generate physically plausible videos. In this paper, we introduce VLIPP, a novel two-stage physically plausible video generation framework that incorporates physical laws into video diffusion models through vision and language informed physical prior. Our experimental results demonstrate the effectiveness of our method compared to existing approaches.", + "bbox": [ + 509, + 739, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12367", + "bbox": [ + 480, + 936, + 517, + 949 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 106, + 104, + 266, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Natural Science Foundation of China under Grant No. 62441231, 62472065, and U23B2010.", + "bbox": [ + 104, + 128, + 485, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 107, + 188, + 199, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4", + "[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3", + "[3] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 4", + "[4] Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. arXiv preprint arXiv:2412.03572, 2024. 2", + "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 6", + "[6] Sébastien Bubeck, Varun Chadrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4, 2023. 2", + "[7] Ryan Burgert, Yuancheng Xu, Wenqi Xian, Oliver Pilarski, Pascal Clausen, Mingming He, Li Ma, Yitong Deng, Lingxiao Li, Mohsen Mousavi, et al. Go-with-the-flow: Motion-controllable video diffusion models using real-time warped noise. arXiv preprint arXiv:2501.08331, 2025. 2, 5", + "[8] Pascal Chang, Jingwei Tang, Markus Gross, and Vinicius C Azevedo. How i warped your noise: a temporally-correlated noise prior for diffusion models. In The Twelfth International Conference on Learning Representations. OpenReview, 2024. 5", + "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4", + "[10] Boyang Deng, Richard Tucker, Zhengqi Li, Leonidas Guibas, Noah Snavely, and Gordon Wetzstein. Streetscapes: Large-scale consistent street view generation using autoregressive video diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[11] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + ], + "bbox": [ + 107, + 212, + 483, + 886 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4", + "[12] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. arXiv preprint arXiv:2401.18085, 2024. 5", + "[13] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 5", + "[14] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.6", + "[15] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractr: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[17] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5", + "[18] Hao-Yu Hsu, Zhi-Hao Lin, Albert Zhai, Hongchi Xia, and Shenlong Wang. Autovfx: Physically realistic video editing from natural language instructions. arXiv preprint arXiv:2411.02394, 2024. 2", + "[19] Hsin-Ping Huang, Yu-Chuan Su, Deqing Sun, Lu Jiang, Xuhui Jia, Yukun Zhu, and Ming-Hsuan Yang. Fine-grained controllable video generation via object appearance and context. arXiv preprint arXiv:2312.02919, 2023. 3", + "[20] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1-55, 2025. 3", + "[21] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 2, 3", + "[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.5", + "[23] Pengxiang Li, Kai Chen, Zhili Liu, Ruiyuan Gao, Lanqing Hong, Guo Zhou, Hua Yao, Dit-Yan Yeung, Huchuan Lu, and Xu Jia. Trackdiffusion: Tracklet-conditioned video generation via diffusion models. arXiv preprint arXiv:2312.00651, 2023. 3", + "[24] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 2" + ], + "bbox": [ + 514, + 107, + 890, + 886 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12368", + "bbox": [ + 480, + 936, + 519, + 949 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 3, 6", + "[26] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3", + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024. 4", + "[29] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2024. 2, 4", + "[30] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3", + "[31] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 4, 5", + "[32] Antonio Montanaro, Luca Savant Aira, Emanuele Aiello, Diego Valsesia, and Enrico Magli. Motioncraft: Physics-based zero-shot video generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2", + "[33] Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models learn physical principles from watching videos? arXiv preprint arXiv:2501.09038, 2025. 2, 4, 5", + "[34] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. Advances in Neural Information Processing Systems, 37:18481-18505, 2024. 3", + "[35] Koichi Namekata, Sherwin Bahmani, Ziyi Wu, Yash Kant, Igor Gilitschenski, and David B Lindell. Sg-i2v: Self-guided trajectory control in image-to-video generation. arXiv preprint arXiv:2411.04989, 2024. 6, 7", + "[36] Andrew Nealen, Matthias Müller, Richard Keiser, Eddy Boxerman, and Mark Carlson. Physically based deformable models in computer graphics. In Computer graphics forum, pages 809-836. Wiley Online Library, 2006. 2", + "[37] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 3", + "[38] Openai. https://openai.com/index/hello-gpt-4o/, 2024. 4", + "[39] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 3" + ], + "bbox": [ + 107, + 107, + 483, + 883 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. Freetraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 3", + "[41] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4", + "[42] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 2", + "[43] Weixiang Sun, Xiaocao You, Ruizhe Zheng, Zhengqing Yuan, Xiang Li, Lifang He, Quanzheng Li, and Lichao Sun. Bora: Biomedical generalist video generation model. arXiv preprint arXiv:2407.08944, 2024. 2", + "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5", + "[45] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 5", + "[46] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 3", + "[47] Xi Wang, Robin Courant, Marc Christie, and Vicky Kalogeiton. Akira: Augmentation kit on rays for optical video generation. arXiv preprint arXiv:2412.14158, 2024. 3", + "[48] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 2", + "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[50] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824-24837, 2022. 3, 4", + "[52] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 3", + "[53] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting ltm-controlled diffusion" + ], + "bbox": [ + 514, + 107, + 890, + 885 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12369", + "bbox": [ + 480, + 936, + 517, + 948 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3", + "[54] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3", + "[55] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3", + "[56] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2", + "[57] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 3", + "[58] Qiyao Xue, Xiangyu Yin, Boyuan Yang, and Wei Gao. Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation. arXiv preprint arXiv:2412.00596, 2024. 1, 6", + "[59] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 1(2):6, 2023. 2", + "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 5, 6", + "[61] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3", + "[62] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 2", + "[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[64] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2024. 2", + "[65] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3" + ], + "bbox": [ + 107, + 107, + 483, + 883 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[66] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 6", + "bbox": [ + 514, + 107, + 890, + 161 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12370", + "bbox": [ + 480, + 936, + 519, + 949 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_model.json b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eb6cd833a88a872b3465b30ec4f618530f5e8370 --- /dev/null +++ b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.144, + 0.77, + 0.188 + ], + "angle": 0, + "content": "VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.213, + 0.849, + 0.284 + ], + "angle": 0, + "content": "Xindi Yang\\(^{1*}\\), Baolu Li\\(^{2*}\\), Yiming Zhang\\(^{2}\\), Zhenfei Yin\\(^{4,5\\dagger}\\), Lei Bai\\(^{3\\dagger}\\), Liqian Ma\\(^{6}\\), Zhiyong Wang\\(^{5}\\), Jianfei Cai\\(^{1}\\), Tien-Tsin Wong\\(^{1}\\), Huchuan Lu\\(^{2}\\), Xu Jia\\(^{2\\dagger}\\) \n\\(^{1}\\)Monash University \\(^{2}\\)Dalian University of Technology \\(^{3}\\)Shanghai Artificial Intelligence Laboratory \\(^{4}\\)Oxford University \\(^{5}\\)The University of Sydney \\(^{6}\\)ZMO AI" + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.317, + 0.334, + 0.332 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.347, + 0.486, + 0.742 + ], + "angle": 0, + "content": "Video diffusion models (VDMs) have advanced significantly in recent years, enabling the generation of highly realistic videos and drawing the attention of the community in their potential as world simulators. However, despite their capabilities, VDMs often fail to produce physically plausible videos due to an inherent lack of understanding of physics, resulting in incorrect dynamics and event sequences. To address this limitation, we propose a novel two-stage image-to-video generation framework that explicitly incorporates physics with vision and language informed physical prior. In the first stage, we employ a Vision Language Model (VLM) as a coarse-grained motion planner, integrating chain-of-thought and physics-aware reasoning to predict a rough motion trajectories/changes that approximate real-world physical dynamics while ensuring the inter-frame consistency. In the second stage, we use the predicted motion trajectories/changes to guide the video generation of a VDM. As the predicted motion trajectories/changes are rough, noise is added during inference to provide freedom to the VDM in generating motion with more fine details. Extensive experimental results demonstrate that our framework can produce physically plausible motion, and comparative evaluations highlight the notable superiority of our approach over existing methods. More video results and code are available on our Project Page: https://madaoer.github.io/projects/physically_plausible Video_generation/." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.767, + 0.234, + 0.782 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.791, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Video diffusion models (VDMs) trained on large-scale video datasets have made remarkable progress in terms of realism, demonstrating significant potential for various content creation applications. Despite the absence of explicit geometric modeling, the generated videos still exhibit coherent spatial" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.322, + 0.889, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.414, + 0.888, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.506, + 0.888, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.604, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Figure 1. Existing commercial closed-source VDMs fail to generate physically plausible motion, whereas our video generation framework is able to achieve this by incorporating external physical prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.68, + 0.891, + 0.797 + ], + "angle": 0, + "content": "relationships among objects, rich textured details, and realistic lighting effects, including reflections and shadows. Such qualities often make the generated videos nearly indistinguishable from real-world footages. This drives the research community to explore the potential of VDMs as world models. However, they still struggle with understanding the physical laws of the real world and generating videos that adhere to these principles." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.799, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Although existing VDMs can produce visually realistic videos, they fail to mimic the real-dynamic physical motions. As shown in Fig 1, even the current commercial closed-source VDMs struggle with the task of generating videos that conform to physical laws. PhyT2V [58] refines text prompts by incorporating detailed descriptions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.129, + 0.873, + 0.372, + 0.886 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.52, + 0.95 + ], + "angle": 0, + "content": "12360" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.105, + 0.107, + 0.482, + 0.31 + ], + "angle": 0, + "content": "of physical processes to guide VDMs in generating physically plausible videos. However, despite being pretrained on internet-scale real-world video-text pairs, VDMs do not inherently understand physical laws. This limitation arises from the gap and ambiguity between text descriptions and the actual motion in the video[33]. Moreover, VDMs tend to overfit the training data rather than developing a general understanding of physical laws[21]. Inspired by the success of graphics-based physical rendering, some methods have guided VDMs to generate physically plausible videos using simulations from graphics engines[18, 29, 32, 56, 64]; however, these approaches rely on the physical effects that graphics engines can simulate and incur high computational costs." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.315, + 0.482, + 0.576 + ], + "angle": 0, + "content": "The gap and ambiguity between text and real-world motion makes it difficult to enable physically plausible video generation through detailed text descriptions alone. Moreover, it is challenging to gather scalable physical data for training due to the abstractness and diversity of physical phenomena. Consequently, a viable approach could be to model abstract physical laws as conditions for diffusion models. However, it is less practical to explicitly model the physics equation for every kind of motion. Instead, we resort to current large foundation models for their ability to \"understand\" basic physics[6] and reason about physical phenomena based on the knowledge they extract. For example, given two colliding balls, the Large Language Model (LLM) can approximately predict the paths of the balls after collision. Inspired by this observation, we propose a novel video generation framework that employs a Vision Language Model (VLM) to predict the path/change during a physics event, described by a given image and a text prompt." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.581, + 0.482, + 0.886 + ], + "angle": 0, + "content": "In this paper, we propose VLIPP, a two-stage approach to incorporate physics as conditions into VDM, enabling the generation of physically plausible motion. In the first stage, the VLM serves as a coarse-level motion planner, while a VDM serves as a fine-level motion synthesizer. The idea of stage one is to utilize the chain-of-thought and the physics-aware reasoning of VLM planning to ensure that coarse-level motion trajectories approximately follow real-world physics dynamics. In stage two, we can generate fine-level motion using an image-to-video diffusion model conditioned by the approximated path/change planned by VLM from stage one. Note that the approximated path/changes are not in the level to tell the speed or acceleration of the motion. We choose an existing image-to-video model [7] to accept our coarse-level path/change, by injecting noise to the motion path during both the training and inference phases. Notably, during the VLM planning stage, generating entire physically plausible motion trajectories is not required. Instead, we leverage the generative priors of VDM to produce fine-level physically plausible videos based on coarse-level motion trajectories provided by the VLM. So that the detail-level motion such" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.89, + 0.135 + ], + "angle": 0, + "content": "as speed, acceleration, and vibration are left to the VDM to synthesize." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.89, + 0.254 + ], + "angle": 0, + "content": "We evaluate our physically plausible video generation framework with two major video physics benchmarks and achieved satisfactory results. Furthermore, we discuss and analyze multiple insightful design choices in our video generation framework, such as employing a motion planner tailored for different physics categories, and enhancing the robustness of diffusion model to noisy trajectories. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.258, + 0.888, + 0.315 + ], + "angle": 0, + "content": "1. We introduce a novel image-to-video generation framework for generating physically plausible videos by leveraging the VLM and VDM priors, significantly outperforming the contemporary competitors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.317, + 0.888, + 0.387 + ], + "angle": 0, + "content": "2. We propose a novel chain-of-thought and physics-aware reasoning approach in VLM, along with random noise injection in the latent space during video generation, which effectively improves both the generation quality and physical plausibility." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.389, + 0.888, + 0.43 + ], + "angle": 0, + "content": "3. We conduct a comprehensive experiments and user studies to demonstrate the effectiveness and generalization of our framework in physically plausible videos generation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.258, + 0.888, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.649, + 0.469 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.889, + 0.495 + ], + "angle": 0, + "content": "2.1. Physically Plausible Visual Content Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.89, + 0.72 + ], + "angle": 0, + "content": "Generating physically plausible videos offers substantial value to real-world applications such as scientific simulations[43], robotics [4, 59], and autonomous driving [10, 48]. Traditional graphics pipelines rely on simulation systems to model physical phenomena [36, 42]. Inspired by these approaches, recent studies [18, 29] have performed dynamic simulations in image space based on physical engines. Furthermore, some methods [56, 64] incorporate physical priors into 3D representations to enable the synthesis of physically plausible motions. However, these rule-based or solver-based simulators face limitations in expressiveness, efficiency, generalizability, and parameter tuning. Furthermore, these simulators require significant expertise, rendering them inaccessible and unfriendly for users." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.89, + 0.885 + ], + "angle": 0, + "content": "In addition, Some studies have explored VDMs for generating physically plausible videos. Li et al. [24] models natural oscillations and swaying in frequency-domain. A downstream rendering module then animates static images based on the generated motion information. PhysDiff [62] introduces physical simulator as constrain into the diffusion process by projecting denoised motion of a diffusion step into a physically plausible motion. These methods mainly focus only on specific types of physical motion and do not establish a generalizable approach for generating physically plausible videos." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.518, + 0.949 + ], + "angle": 0, + "content": "12361" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.106, + 0.887, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.223, + 0.887, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.379, + 0.892, + 0.447 + ], + "angle": 0, + "content": "Figure 2. The illustration of our physically plausible image-to-video generation pipeline. Our pipeline consists of two stages. In the first stage, the VLM generates a coarse-grained, physically plausible motion trajectory based on the provided input conditions. In the second stage, We simulate a draft video using the predicted trajectory to provide the motion condition. We then extract the optical flow from this video and convert it into structured noise. These conditions are fed into a motion controllable image-to-video diffusion model, and ultimately generates a physically plausible video." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.457, + 0.427, + 0.472 + ], + "angle": 0, + "content": "2.2. Motion Controllable Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.479, + 0.484, + 0.696 + ], + "angle": 0, + "content": "Existing studies commonly provide one of the following three types of motion control: bounding box control [19, 23, 30, 46, 52], point trajectory control [34, 37, 40, 50, 54] and camera control [2, 15, 47, 61]. Bounding box control provides object motion guidance by generating a sequence of bounding boxes that track the object's position over time. Point-trajectory control offers motion cues through point-based trajectories, enabling drag-style manipulation. Camera motion control guides video generation using explicit 3D camera parameters, ensuring consistency and realistic viewpoint changes. However, these approaches prioritize motion control but often overlook physical plausibility. To address the limitation, we propose a novel framework for physically plausible video generation that incorporates physics as conditions into video diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.705, + 0.416, + 0.721 + ], + "angle": 0, + "content": "2.3. Generation based on VLMs Planning" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.726, + 0.484, + 0.887 + ], + "angle": 0, + "content": "VLMs have exhibited robust capabilities in visual understanding and planning [27, 39, 65]. Their strong performance in domains such as robot path planning and video understanding shows their ability in understanding the real physical world. Prior work has successfully leveraged LLMs to guide the layout of images or videos, yielding promising results [25, 53]. VideoDirectorGPT [26] leverages LLMs for fine-grained scene-by-scene planning, explicitly controlling spatial layout to generate temporally consistent long videos. Pandora [55] utilizes LLMs for real-time control through free-text action commands, achieving domain generality," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.89, + 0.502 + ], + "angle": 0, + "content": "video consistency, and controllability. However, these efforts have yet to address interactions with real-world physical phenomena, such as collision, fall, and melting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Moreover, the absence of visual information can cause severe hallucination issues in language models for spatial planning tasks, leading to problems like overlapping object boundaries, disproportionate scaling, and incorrect planning [20, 57]. In this paper, we propose utilizing VLMs as coarse-level motion planners within the image space and incorporate physics-aware reasoning and Chain of Thought (CoT) [51] into the inference process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.6, + 0.652 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.663, + 0.892, + 0.853 + ], + "angle": 0, + "content": "Task Formulation. In this paper, our goal is to enable an image-to-video diffusion model to generate physically plausible videos. Since VDMs rely more on memory and case-based imitation and struggle to understand general physical rules[21], the key challenge is how to incorporate physical laws into the models. To achieve this, we need to identify a method to incorporate physical principles into the video diffusion framework. Given an image \\( I \\in R^{H \\times W \\times C} \\) (\\( H \\) is height, \\( W \\) is width and \\( C \\) is the number of channels) and a text description \\( d \\) of possible events based on image \\( I \\), our framework should infer a physics-compliant guidance as the input condition and synthesize a video that adheres to both physical laws and real-world dynamics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Overall Pipeline. Overall pipeline of VLIPP is illustrated in Figure 2. In the first stage, the VLM conducts semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.519, + 0.949 + ], + "angle": 0, + "content": "12362" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.106, + 0.478, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.322, + 0.483, + 0.43 + ], + "angle": 0, + "content": "Figure 3. The illustration of chain-of-thought reasoning in the VLM for generating a coarse-grained motion trajectory. First off, the VLM determines the corresponding physical laws and its context for the given scene. Then, the VLM performs step-by-step reasoning to predict the physically plausible motions of objects in image space, leveraging physical context and chain-of-thought prompting. Finally, the VLM predicts bounding boxes according to real-world physics." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.446, + 0.484, + 0.562 + ], + "angle": 0, + "content": "analysis and physical attribute analysis on the given image \\(I\\) and a description \\(d\\) to obtain the bounding boxes of the objects in the scene, denoted as \\(b_{1}, b_{2}, \\ldots, b_{n}\\), along with the applicable physical laws \\(l\\). Next, the VLM infers possible future physical scenarios in the current scene to derive the coarse-level motion trajectories of \\(p_{1}, p_{2}, \\ldots, p_{t}\\) in the image space. Finally, we utilize an image-to-video diffusion model to synthesize the detailed dynamics in the video." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.571, + 0.437, + 0.585 + ], + "angle": 0, + "content": "3.1. VLM as a Coarse-Level Motion Planner" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.592, + 0.483, + 0.695 + ], + "angle": 0, + "content": "Our motivation is to incorporate physical laws as constraints into a video diffusion model to enhance the physical plausibility of the generated videos. To achieve this, we must identify a method to inject physical laws into the video diffusion model. Given a video description and the first frame, the task at this stage is to generate coarse-level motion trajectory aligned with physical laws." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.697, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Scene Understanding. In the real world, most physical phenomena arise from interactions among objects and their motion trajectories. We first initiate the process by identifying and locating objects within a scene. Inspired by the recent studies [1, 9, 28] in VLMs for scene understanding, we employ GPT-4o [38] to recognize all objects that could be involved in physical phenomena as described in the text description \\(d\\). These objects are subsequently detected and segmented using Grounded-SAM2 [41], yielding their bounding boxes. By leveraging the pretrained knowledge and common-sense reasoning capabilities of foundation models, we effectively determine the relevant objects in the scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Physical-Aware Recognition. To perform more effective reasoning in predicting the motion, it is necessary to determine what specific physical principle to apply in the given context. We utilize the pretrained prior of the LLM to determine the physical laws applicable to the current scene. Following the configuration in the physical benchmark [3, 31, 33], we currently classify common physical phenomena in videos into six categories: gravity, momentum conservation, optics, thermodynamics, magnetism, and fluid mechanics. Note that such list can be easily extended within our framework. Given a video description \\(d\\), the LLM infers the physical law \\(l\\) that governs the current scene. We provide the specific physical context information for VLM to enhance its understanding of physical laws [11]. Detailed context design is presented in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.891, + 0.455 + ], + "angle": 0, + "content": "Chain of Thought Reasoning in VLM. Given the physical law \\( l \\), an image \\( I \\) and a video description \\( d \\) for the scene, we prompt the VLM to predict the future bounding box positions of objects within the image-space. We choose to predict in the image space for two primary reasons. Firstly, motion in image space aligns more with our subsequent video synthesizer. Secondly, image space dynamics can effectively represent a wide range of real-world motions [29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.892, + 0.651 + ], + "angle": 0, + "content": "At a given time \\( t \\), the predicted position of \\( i \\)-th object bounding box \\( b_{i}^{t} \\) is denoted as \\( [x_{t}^{i}, y_{t}^{i}, w_{t}^{i}, h_{t}^{i}] \\), where \\( (x_{t}^{i}, y_{t}^{i}) \\) represents its top-left coordinate; \\( w_{t}^{i} \\& h_{t}^{i} \\) denote its width and height, respectively. Governed by the physical law, the four values of the bounding box may change over time. The VLM reasons the bounding box positions of \\( N \\) future frames for every object \\( o_{i} \\) based on the condition. To help VLM better understand physical laws, we adapt a chain-of-thought [51] into its reasoning, to significantly enhance its reasoning capabilities. As shown in Figure 3, we formulate our analysis of physical phenomena in videos as step-by-step reasoning: beginning with broad conceptual ideas and progressing to a detailed and practical examination:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.889, + 0.688 + ], + "angle": 0, + "content": "1. Given the physical law \\( l \\) and context information, the VLM analysis video caption and detail the physical law." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.889, + 0.717 + ], + "angle": 0, + "content": "2. The VLM analyzes the potential interactions and movement of each object within the scene;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.89, + 0.76 + ], + "angle": 0, + "content": "3. The VLM predicts the detailed changes in position and shape of the bounding box corresponding to each object over time." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.659, + 0.89, + 0.76 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Through the structured planning process, the VLM plans coarse-level motion trajectories for the objects, approximating real-world physics dynamics. In particular, our VLM infers the changes of object bounding boxes for next 12 frames, constrained by the token length limitation. To be compatible with the generation process of the chosen VDM in the next stage, these inferred 12 frames are further linearly interpolated to produce a total of 49 frames." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.518, + 0.949 + ], + "angle": 0, + "content": "12363" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.106, + 0.106, + 0.484, + 0.135 + ], + "angle": 0, + "content": "3.2. VDM Serves as a Fine-Level Motion Synthesizer" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.142, + 0.484, + 0.259 + ], + "angle": 0, + "content": "In the previous stage, the motion trajectory planned by the VLM is neither precise nor fully compliant with physical laws. On the other hand, while VDM may not be able to produce realistic global motion trajectories, it is able to generate sound motion in finer scale. In this stage, our key insight is that the VDM can refine the coarse-level motion to produce physically plausible motion that aligns with real-world dynamics with its powerful generative prior." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.262, + 0.485, + 0.479 + ], + "angle": 0, + "content": "Motion Animation. To incorporate physical laws into the video diffusion model, we use the inferred coarse motion trajectory to guide the generation process of the diffusion model. Optical flow provides a unified representation of motion, and recent studies [7, 12, 13] have demonstrated its effectiveness in guiding diffusion models. Accordingly, we leverage the coarse-level motion trajectory to animate a draft motion sequence and derive the corresponding optical flow. Specifically, for each object \\( o_i \\), we extract its bounding box from the first frame and move it to the bounding box location \\( b_i \\) specified by the draft motion trajectory. To animate the change of shape (e.g., due to compression or expansion), we resize object \\( o_i \\) according to the difference between \\( o_i \\) and \\( o_{i+1} \\) during inpainting. The draft motion video is generated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.488, + 0.484, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\hat {V} (t) = \\operatorname {A n i m a t i o n} \\left(B, r s \\left(o _ {0} ^ {0}, b _ {t} ^ {0}\\right)... r s \\left(o _ {0} ^ {i}, b _ {t} ^ {i}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.515, + 0.484, + 0.589 + ], + "angle": 0, + "content": "where \\(\\hat{V}(t)\\) denotes the corresponding frame of inpainted video at timestep \\(t\\), \\(B\\) denotes the inpainted background with the foreground object removed, \\(o_i^0\\) denotes the \\(i\\)-th object at timestep 0, \\(b_i^t\\) represents the \\(i\\)-th bounding box at timestep \\(t\\), and \\(rs\\) denotes resize function." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.592, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Structured Noise from Draft Video. Optical flow is an effective representation for guiding VDMs [12, 13]. Follow prior work [7, 8], we employ RAFT[44] to extract optical flow from the draft video and formulate it as structural noise, which retains Gaussian properties. Given the draft video \\(\\hat{V}(t) \\in \\mathbb{R}^{F \\times C \\times H \\times W}\\), we calculate its per-frame optical flow to get a structured noise tensor \\(Q \\in \\mathbb{R}^{F \\times C \\times H \\times W}\\). The structured noise enables the VDM to generate videos that exhibit motion patterns closely aligned with those in the optical flow, thereby improving the realism of the output." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.741, + 0.484, + 0.887 + ], + "angle": 0, + "content": "Noise Injection in Video Synthesis. We adopt Go-with-the-Flow [7] as our video synthesis model, a fine-tuned CogVideoX [60], which is designed to accept structured noise \\( Q \\) as input and synthesize videos that adhere to the implicit optical flow. The vanilla Go-with-the-Flow tends to tightly follow the provided structured noise \\( Q \\). However, our \\( Q \\) is derived from a coarse-level motion trajectory and may not be sufficiently accurate to follow the physical laws of the real world. To address this limitation, we inject noise during the inference phase to give more flexibility to the VDM to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.766, + 0.122 + ], + "angle": 0, + "content": "generate detail-level motion changes as" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.133, + 0.891, + 0.168 + ], + "angle": 0, + "content": "\\[\nQ _ {i} = \\frac {(1 - \\gamma) Q _ {i} + \\zeta \\gamma}{\\sqrt {(1 - \\gamma) ^ {2} + \\gamma^ {2}}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.178, + 0.891, + 0.223 + ], + "angle": 0, + "content": "where \\(Q_{i}\\) is structured noise at \\(i\\)-th frames, \\(\\zeta \\in \\mathbb{R}^{C\\times H\\times W}\\) is Gaussian noise and \\(\\gamma \\in [0,1]\\). We set \\(\\gamma = 0.4\\) for even frame index and \\(\\gamma = 0.6\\) for odd frame index." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.892, + 0.268 + ], + "angle": 0, + "content": "With this approach, the VDM is able to generate motion deviate from the coarse-level motion trajectory whenever necessary for producing high-quality fine-level motion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.281, + 0.816, + 0.298 + ], + "angle": 0, + "content": "4. Empirical Analysis and Dissusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.892, + 0.395 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to demonstrate the effectiveness of our video generation framework compared to existing methods. We evaluate our approach on two established benchmarks for physically plausible video generation. Our framework consistently achieves superior performance across all benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.724, + 0.419 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.892, + 0.527 + ], + "angle": 0, + "content": "We propose a two-stage physically plausible image-to-video generation framework. In the first stage, we utilize ChatGPT-4o as the coarse-level motion planner. In the second stage, we utilize an open-source I2V model, Go-with-the-Flow [7], as a fine-level motion synthesizer. Unless otherwise specified, in all experiments, we generate each video with a resolution of \\( 720 \\times 480 \\) and 49 frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.537, + 0.734, + 0.552 + ], + "angle": 0, + "content": "4.2. Benchmarks and Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.892, + 0.704 + ], + "angle": 0, + "content": "Traditional metrics in the visual domain, such as the Peak Signal-to-Noise Ratio (PSNR)[17], the Structural Similarity Index (SSIM)[49], the Learned Perceptual Image Patch Similarity (LPIPS)[63], the Fréchet Inception Distance (FID)[16] and the Fréchet Video Distance (FVD)[45], do not account for the physical realism of the generated videos [31, 33]. Recent studies have begun to address this limitation by developing benchmarks and metrics that evaluate physical realism. In this work, we adopt two benchmarks, described below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.892, + 0.853 + ], + "angle": 0, + "content": "PhyGenBench [31] categorizes physical properties into four domains: mechanics, optics, thermal, and material. It includes 27 physical phenomena, each governed by real world physical laws, reflected in 160 carefully designed text prompts. As PhyGenbench provides only text prompts, we adapt it to our image-to-video setting by generating a corresponding first frame for each prompt with FLUX[22]. We adhere to the predefined benchmark evaluation protocol, i.e., employing GPT-4o to assess the physical realism of the generated videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Physics-IQ [33] comprises 396 real-world videos spanning 66 distinct physical scenarios. For each scenario, videos" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.52, + 0.95 + ], + "angle": 0, + "content": "12364" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.104, + 0.891, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.372, + 0.89, + 0.399 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B [60], LTX-VideoI2V [14] and SVD-XT [5]." + }, + { + "type": "table", + "bbox": [ + 0.158, + 0.404, + 0.842, + 0.591 + ], + "angle": 0, + "content": "
ModelMechanics(↑)Optics(↑)Thermal(↑)Material(↑)Average(↑)
CogvideoX-T2V-5B0.430.550.400.420.45
LTX-Video-T2V0.350.450.360.380.39
OpenSora0.430.500.440.370.44
PhyT2V0.490.610.490.470.52
LLM-Grounding Video Diffusion0.320.410.260.240.31
CogvideoX-I2V-5B0.480.690.430.410.52
SVD-XT0.460.680.480.410.52
LTX-Video-I2V0.470.650.460.370.50
SG-I2V0.520.690.510.390.54
Ours0.550.710.600.530.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.334, + 0.592, + 0.662, + 0.605 + ], + "angle": 0, + "content": "Table 1. Quantitative results of VDMs on PhyGenBench." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.622, + 0.484, + 0.739 + ], + "angle": 0, + "content": "are recorded from three different perspectives and filmed twice under identical conditions to eliminate randomness. This benchmark evaluates real-world physical phenomena, including collisions, object continuity, occlusion, object permanence, and fluid dynamics. This benchmark assesses physical realism from semantic and temporal perspectives, using semantic metrics and visual metrics to compare generated videos against the real-world reference videos." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.755, + 0.484, + 0.887 + ], + "angle": 0, + "content": "Compared Models. In the context of text-to-video generation, we compare our framework with CogVideoX-T2V-5B [60], LTX-Video-T2V [14], and OpenSora [66]. Moreover, we evaluate our framework against PhyT2V [58], which enhances physical realism by iteratively refining the prompt. For the image-to-video generation scenario, our framework is evaluated alongside CogVideoX-I2V-5B [60], SVD-XT [5], and LTX-Video-I2V [14]. Additionally, we conducted experiments in the motion-controllable setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.892, + 0.724 + ], + "angle": 0, + "content": "In this setting, we leverage the motion trajectory predicted by the VLM as a condition to guide VDM generation. We benchmark our approach against image-to-video motion controllable model, SG-I2V [35] and text-to-video motion controllable model LLM-grounded Video Diffusion Models [25]. The experimental details are presented in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.734, + 0.727, + 0.749 + ], + "angle": 0, + "content": "4.3.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.891, + 0.87 + ], + "angle": 0, + "content": "We begin with an empirical study on PhyGenBench and Physics-IQ, comparing our framework against widely adopted open-source models in the research community. Based on different physical properties, we categorize the benchmark samples accordingly. Additionally, we classify VDMs into text-to-video (T2V) diffusion models and imaged-to-video (I2V) diffusion models based on the input conditions." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.872, + 0.891, + 0.887 + ], + "angle": 0, + "content": "In Table 1, we present our experimental results on Phy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.518, + 0.949 + ], + "angle": 0, + "content": "12365" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.104, + 0.842, + 0.205 + ], + "angle": 0, + "content": "
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Cogvideo-I2V-5B30.429.816.713.38.527.1
SVD-XT21.920.56.88.417.119.1
LTX-Video-I2V30.229.815.913.28.426.8
SG-I2V34.631.215.913.18.429.7
Ours42.334.116.913.48.834.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.207, + 0.89, + 0.234 + ], + "angle": 0, + "content": "Table 2. Quantitative results of physically plausible video generation on Physics-IQ Benchmark. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.251, + 0.484, + 0.456 + ], + "angle": 0, + "content": "GenBench, evaluating different video generation models following its evaluation protocol. The results show that our framework achieves state-of-the-art performance across four different physical phenomena. Our framework outperforms the best T2V method by an average of \\(15.3\\%\\) and the best I2V method by \\(11.1\\%\\). Specifically, our framework demonstrates significant advantages in the Mechanics, Thermal, and Material domains, outperforming the best I2V method by \\(5.7\\%\\), \\(17.6\\%\\), and \\(35.8\\%\\), respectively. These advantages are particularly evident in these three types of physical phenomena, which involve more substantial changes in motion, volume, or shape. Our framework is better equipped to understand and reason about bounding box sequences to represent these changes effectively." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.465, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Similarly, for the Physics-IQ benchmark, we evaluate the performance of different video generation models following its evaluation protocol. Our framework achieves the best results across four different physical phenomena, with improvements of \\(22.2\\%\\) in Solid Mechanics and \\(9.2\\%\\) in Fluid Dynamics compared to the second-best models. These significant improvements demonstrate the effectiveness of our framework in generating physically plausible videos." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.637, + 0.308, + 0.652 + ], + "angle": 0, + "content": "4.4. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.668, + 0.484, + 0.887 + ], + "angle": 0, + "content": "Figures 4 and 5 demonstrate a qualitative comparison between our video generation framework and baseline methods. Among all evaluated approaches, our framework consistently produces videos with the highest degree of physical realism. In the ball falling sample in Figure 4, while CogVideoX shows a bouncing effect, artifacts are present in the video; LTX-Video and SVD-XT exhibit motions that do not adhere to the laws of physics. In Figure 5, we analyze two examples from Physics-IQ. In the pouring water example, the baseline methods fail to show the simultaneous decrease in water level of the glass beverage dispenser and the increase in water level of the glass below; in the ball collision example, none of the baseline methods correctly depict the collision of balls. More videos are provided in the supplement." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.25, + 0.89, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.567, + 0.89, + 0.606 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B, LTX-Video-I2V, SVD-XT and SG-I2V [35] in the Physics-IQ dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.661, + 0.643 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.892, + 0.709 + ], + "angle": 0, + "content": "We perform an ablation study to evaluate the contributions of key components in our framework. We design four variants to analyze the effectiveness of different components in our framework." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.712, + 0.89, + 0.77 + ], + "angle": 0, + "content": "1. Ours w/o VLM Planner: To assess the overall functionality of our framework, we replace the structured noise input of the VDM with random noise to evaluate the effectiveness of the VLM planner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.77, + 0.889, + 0.798 + ], + "angle": 0, + "content": "2. Ours w/o CI: Keeping the overall structure unchanged, we remove the in-context information from the VLM." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.799, + 0.89, + 0.842 + ], + "angle": 0, + "content": "3. Ours w/o CoT: Similarly, while keeping other components unchanged, we remove the CoT reasoning process from the VLM." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.843, + 0.89, + 0.887 + ], + "angle": 0, + "content": "4. Ours w/o CC: Lastly, we remove both the in-context information and the CoT reasoning process from the VLM planner while maintaining all other components." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.712, + 0.89, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.519, + 0.949 + ], + "angle": 0, + "content": "12366" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.155, + 0.104, + 0.845, + 0.205 + ], + "angle": 0, + "content": "
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Ours42.334.116.913.48.834.9
w/o VLM Planner.16.320.813.45.85.616.2
w/o C.I26.328.116.911.28.424.3
w/o CoT21.426.916.18.66.921.0
w/o C.C18.722.414.97.26.118.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.206, + 0.869, + 0.221 + ], + "angle": 0, + "content": "Table 3. Ablation study on VLM, in-context learning and COT. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.235, + 0.484, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.402, + 0.483, + 0.44 + ], + "angle": 0, + "content": "Figure 6. Visualization of generated video with bounding boxes. Coarse motion cue can guide VDM to synthesis physically plausible video." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.467, + 0.484, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.634, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Figure 7. Limitation of VLIPP. Our framework might produces physically implausible videos because it lacks 3D spatial perception and the optical-flow of small objects is prone to noise." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.712, + 0.484, + 0.885 + ], + "angle": 0, + "content": "Table 3 presents a quantitative comparison between our full method and these variants. Among all variants, Ours w/o VLM Planner shows the most significant performance drop, as removing the planner completely eliminates our ability in understanding the physical laws, leading to nearly random results. Notably, Ours w/o CoT exhibits a more pronounced decline compared to Ours w/o CI, indicating that the reasoning process in CoT enhances the understanding of physics. While in-context information contributes to the physical reasoning ability of VLM, compared to CoT it is less effective in preventing errors caused by VLM hallucination." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.237, + 0.635, + 0.251 + ], + "angle": 0, + "content": "4.6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.89, + 0.293 + ], + "angle": 0, + "content": "Although VLIPP can generate physically plausible videos, its performance remains constrained by the base model." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.297, + 0.89, + 0.53 + ], + "angle": 0, + "content": "Firstly, we are unable to model physical events that cannot be represented by image space bounding box trajectories. For example, phenomena that involve intrinsic state changes of objects such as gas solidification. We adopt 2D bounding boxes because they can be naturally integrated with VDMs and are effective in enhancing the physical plausibility to a certain extent. In fact, they are more flexible than they might appear. As shown in Figure 6, the VLM predicts coarse 2D bounding box trajectories(only four fragments), yet the VDM still produces a realistic shattering pattern without being rigidly constrained by those paths. Moreover, in the butter melting example in Figure 6, the VLM offers only a coarse cue of volumetric change, but the VDM successfully generates the transition from solid to liquid. These abilities significantly improve our flexibility in modeling various physical phenomena." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Secondly, our pipeline lacks 3D spatial perception. It is unable to understand the spatial relationships within the scene, as shown in Figure 7 domino example." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Finally, the optical flow of small objects is prone to noise interference. This will cause our framework to generate ambiguous content, as shown in Figure 7 small ball collision example. With the recent progress in video generation model, we anticipate that our framework will be further improved in generating videos under more challenging physical conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.628, + 0.727 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.74, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Recently, VDMs have achieved great empirical success and are receiving considerable attention in computer vision and computer graphics. However, due to the lack of understanding of physical laws, VDMs are unable to generate physically plausible videos. In this paper, we introduce VLIPP, a novel two-stage physically plausible video generation framework that incorporates physical laws into video diffusion models through vision and language informed physical prior. Our experimental results demonstrate the effectiveness of our method compared to existing approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.519, + 0.95 + ], + "angle": 0, + "content": "12367" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.107, + 0.106, + 0.267, + 0.123 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.13, + 0.486, + 0.175 + ], + "angle": 0, + "content": "This work was partially supported by the National Natural Science Foundation of China under Grant No. 62441231, 62472065, and U23B2010." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.189, + 0.2, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.213, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.269, + 0.484, + 0.336 + ], + "angle": 0, + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.338, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[3] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.406, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[4] Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. arXiv preprint arXiv:2412.03572, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.448, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.517, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[6] Sébastien Bubeck, Varun Chadrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.572, + 0.484, + 0.639 + ], + "angle": 0, + "content": "[7] Ryan Burgert, Yuancheng Xu, Wenqi Xian, Oliver Pilarski, Pascal Clausen, Mingming He, Li Ma, Yitong Deng, Lingxiao Li, Mohsen Mousavi, et al. Go-with-the-flow: Motion-controllable video diffusion models using real-time warped noise. arXiv preprint arXiv:2501.08331, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.641, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[8] Pascal Chang, Jingwei Tang, Markus Gross, and Vinicius C Azevedo. How i warped your noise: a temporally-correlated noise prior for diffusion models. In The Twelfth International Conference on Learning Representations. OpenReview, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.709, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.79, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[10] Boyang Deng, Richard Tucker, Zhengqi Li, Leonidas Guibas, Noah Snavely, and Gordon Wetzstein. Streetscapes: Large-scale consistent street view generation using autoregressive video diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.859, + 0.484, + 0.887 + ], + "angle": 0, + "content": "[11] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.213, + 0.484, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.108, + 0.891, + 0.134 + ], + "angle": 0, + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.137, + 0.891, + 0.177 + ], + "angle": 0, + "content": "[12] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. arXiv preprint arXiv:2401.18085, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.179, + 0.891, + 0.247 + ], + "angle": 0, + "content": "[13] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.249, + 0.891, + 0.314 + ], + "angle": 0, + "content": "[14] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.317, + 0.891, + 0.371 + ], + "angle": 0, + "content": "[15] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractr: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.373, + 0.891, + 0.438 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.442, + 0.891, + 0.483 + ], + "angle": 0, + "content": "[17] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.485, + 0.891, + 0.537 + ], + "angle": 0, + "content": "[18] Hao-Yu Hsu, Zhi-Hao Lin, Albert Zhai, Hongchi Xia, and Shenlong Wang. Autovfx: Physically realistic video editing from natural language instructions. arXiv preprint arXiv:2411.02394, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.541, + 0.891, + 0.594 + ], + "angle": 0, + "content": "[19] Hsin-Ping Huang, Yu-Chuan Su, Deqing Sun, Lu Jiang, Xuhui Jia, Yukun Zhu, and Ming-Hsuan Yang. Fine-grained controllable video generation via object appearance and context. arXiv preprint arXiv:2312.02919, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.597, + 0.891, + 0.676 + ], + "angle": 0, + "content": "[20] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1-55, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.679, + 0.891, + 0.733 + ], + "angle": 0, + "content": "[21] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.735, + 0.891, + 0.761 + ], + "angle": 0, + "content": "[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.764, + 0.891, + 0.829 + ], + "angle": 0, + "content": "[23] Pengxiang Li, Kai Chen, Zhili Liu, Ruiyuan Gao, Lanqing Hong, Guo Zhou, Hua Yao, Dit-Yan Yeung, Huchuan Lu, and Xu Jia. Trackdiffusion: Tracklet-conditioned video generation via diffusion models. arXiv preprint arXiv:2312.00651, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.833, + 0.891, + 0.887 + ], + "angle": 0, + "content": "[24] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.108, + 0.891, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.52, + 0.95 + ], + "angle": 0, + "content": "12368" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.108, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[25] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.149, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[26] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.202, + 0.482, + 0.228 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.23, + 0.483, + 0.269 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.27, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[29] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.324, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[30] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.377, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[31] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.445, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[32] Antonio Montanaro, Luca Savant Aira, Emanuele Aiello, Diego Valsesia, and Enrico Magli. Motioncraft: Physics-based zero-shot video generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.511, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[33] Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models learn physical principles from watching videos? arXiv preprint arXiv:2501.09038, 2025. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.565, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[34] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. Advances in Neural Information Processing Systems, 37:18481-18505, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.619, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[35] Koichi Namekata, Sherwin Bahmani, Ziyi Wu, Yash Kant, Igor Gilitschenski, and David B Lindell. Sg-i2v: Self-guided trajectory control in image-to-video generation. arXiv preprint arXiv:2411.04989, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.673, + 0.483, + 0.724 + ], + "angle": 0, + "content": "[36] Andrew Nealen, Matthias Müller, Richard Keiser, Eddy Boxerman, and Mark Carlson. Physically based deformable models in computer graphics. In Computer graphics forum, pages 809-836. Wiley Online Library, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.726, + 0.483, + 0.792 + ], + "angle": 0, + "content": "[37] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.793, + 0.461, + 0.806 + ], + "angle": 0, + "content": "[38] Openai. https://openai.com/index/hello-gpt-4o/, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.807, + 0.483, + 0.884 + ], + "angle": 0, + "content": "[39] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.108, + 0.484, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.108, + 0.891, + 0.16 + ], + "angle": 0, + "content": "[40] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. Freetraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.891, + 0.228 + ], + "angle": 0, + "content": "[41] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.231, + 0.891, + 0.283 + ], + "angle": 0, + "content": "[42] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.286, + 0.891, + 0.338 + ], + "angle": 0, + "content": "[43] Weixiang Sun, Xiaocao You, Ruizhe Zheng, Zhengqing Yuan, Xiang Li, Lifang He, Quanzheng Li, and Lichao Sun. Bora: Biomedical generalist video generation model. arXiv preprint arXiv:2407.08944, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.891, + 0.405 + ], + "angle": 0, + "content": "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.409, + 0.891, + 0.448 + ], + "angle": 0, + "content": "[45] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.451, + 0.891, + 0.503 + ], + "angle": 0, + "content": "[46] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.891, + 0.544 + ], + "angle": 0, + "content": "[47] Xi Wang, Robin Courant, Marc Christie, and Vicky Kalogeiton. Akira: Augmentation kit on rays for optical video generation. arXiv preprint arXiv:2412.14158, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.891, + 0.611 + ], + "angle": 0, + "content": "[48] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.891, + 0.666 + ], + "angle": 0, + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.669, + 0.891, + 0.735 + ], + "angle": 0, + "content": "[50] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.737, + 0.891, + 0.802 + ], + "angle": 0, + "content": "[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824-24837, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.805, + 0.891, + 0.857 + ], + "angle": 0, + "content": "[52] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.891, + 0.886 + ], + "angle": 0, + "content": "[53] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting ltm-controlled diffusion" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.108, + 0.891, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.519, + 0.949 + ], + "angle": 0, + "content": "12369" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.139, + 0.108, + 0.484, + 0.147 + ], + "angle": 0, + "content": "models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.149, + 0.484, + 0.215 + ], + "angle": 0, + "content": "[54] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.216, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[55] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.283, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[56] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.35, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[57] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.39, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[58] Qiyao Xue, Xiangyu Yin, Boyuan Yang, and Wei Gao. Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation. arXiv preprint arXiv:2412.00596, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.444, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[59] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 1(2):6, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.498, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.565, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[61] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.632, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[62] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.686, + 0.484, + 0.75 + ], + "angle": 0, + "content": "[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.753, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[64] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.833, + 0.484, + 0.884 + ], + "angle": 0, + "content": "[65] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.108, + 0.484, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.108, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[66] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.938, + 0.52, + 0.95 + ], + "angle": 0, + "content": "12370" + } + ] +] \ No newline at end of file diff --git a/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_origin.pdf b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c32b75407c16a0baf52f0d1e57c328b3cb112864 --- /dev/null +++ b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/eba6c702-728f-4007-b7d6-27e6fa6404b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88b7bf0f95595e1cc470b1c761ce173c94ecb0280120e6b670ebd6447ca2b00d +size 1519778 diff --git a/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/full.md b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..921908ce5f48dc1189b107f855d89397d7e82d9b --- /dev/null +++ b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/full.md @@ -0,0 +1,281 @@ +# VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior + +Xindi Yang $^{1*}$ , Baolu Li $^{2*}$ , Yiming Zhang $^{2}$ , Zhenfei Yin $^{4,5\dagger}$ , Lei Bai $^{3\dagger}$ , Liqian Ma $^{6}$ , Zhiyong Wang $^{5}$ , Jianfei Cai $^{1}$ , Tien-Tsin Wong $^{1}$ , Huchuan Lu $^{2}$ , Xu Jia $^{2\dagger}$ $^{1}$ Monash University $^{2}$ Dalian University of Technology $^{3}$ Shanghai Artificial Intelligence Laboratory $^{4}$ Oxford University $^{5}$ The University of Sydney $^{6}$ ZMO AI + +# Abstract + +Video diffusion models (VDMs) have advanced significantly in recent years, enabling the generation of highly realistic videos and drawing the attention of the community in their potential as world simulators. However, despite their capabilities, VDMs often fail to produce physically plausible videos due to an inherent lack of understanding of physics, resulting in incorrect dynamics and event sequences. To address this limitation, we propose a novel two-stage image-to-video generation framework that explicitly incorporates physics with vision and language informed physical prior. In the first stage, we employ a Vision Language Model (VLM) as a coarse-grained motion planner, integrating chain-of-thought and physics-aware reasoning to predict a rough motion trajectories/changes that approximate real-world physical dynamics while ensuring the inter-frame consistency. In the second stage, we use the predicted motion trajectories/changes to guide the video generation of a VDM. As the predicted motion trajectories/changes are rough, noise is added during inference to provide freedom to the VDM in generating motion with more fine details. Extensive experimental results demonstrate that our framework can produce physically plausible motion, and comparative evaluations highlight the notable superiority of our approach over existing methods. More video results and code are available on our Project Page: https://madaoer.github.io/projects/physically_plausible Video_generation/. + +# 1. Introduction + +Video diffusion models (VDMs) trained on large-scale video datasets have made remarkable progress in terms of realism, demonstrating significant potential for various content creation applications. Despite the absence of explicit geometric modeling, the generated videos still exhibit coherent spatial + +![](images/c9ea32fa753f62ad355288a590600f15f3482c20677eaef9919f606bdafc3e54.jpg) + +![](images/1c8ac1d0e4311bf54beed485a09ec0c48140d89ca69db2bda1ddec2cce87568c.jpg) + +![](images/2363cedd545e1a12a5482ed879a142d6a192d5c1705c1f5f34e5bcc79608a29e.jpg) +Figure 1. Existing commercial closed-source VDMs fail to generate physically plausible motion, whereas our video generation framework is able to achieve this by incorporating external physical prior knowledge. + +relationships among objects, rich textured details, and realistic lighting effects, including reflections and shadows. Such qualities often make the generated videos nearly indistinguishable from real-world footages. This drives the research community to explore the potential of VDMs as world models. However, they still struggle with understanding the physical laws of the real world and generating videos that adhere to these principles. + +Although existing VDMs can produce visually realistic videos, they fail to mimic the real-dynamic physical motions. As shown in Fig 1, even the current commercial closed-source VDMs struggle with the task of generating videos that conform to physical laws. PhyT2V [58] refines text prompts by incorporating detailed descriptions + +of physical processes to guide VDMs in generating physically plausible videos. However, despite being pretrained on internet-scale real-world video-text pairs, VDMs do not inherently understand physical laws. This limitation arises from the gap and ambiguity between text descriptions and the actual motion in the video[33]. Moreover, VDMs tend to overfit the training data rather than developing a general understanding of physical laws[21]. Inspired by the success of graphics-based physical rendering, some methods have guided VDMs to generate physically plausible videos using simulations from graphics engines[18, 29, 32, 56, 64]; however, these approaches rely on the physical effects that graphics engines can simulate and incur high computational costs. + +The gap and ambiguity between text and real-world motion makes it difficult to enable physically plausible video generation through detailed text descriptions alone. Moreover, it is challenging to gather scalable physical data for training due to the abstractness and diversity of physical phenomena. Consequently, a viable approach could be to model abstract physical laws as conditions for diffusion models. However, it is less practical to explicitly model the physics equation for every kind of motion. Instead, we resort to current large foundation models for their ability to "understand" basic physics[6] and reason about physical phenomena based on the knowledge they extract. For example, given two colliding balls, the Large Language Model (LLM) can approximately predict the paths of the balls after collision. Inspired by this observation, we propose a novel video generation framework that employs a Vision Language Model (VLM) to predict the path/change during a physics event, described by a given image and a text prompt. + +In this paper, we propose VLIPP, a two-stage approach to incorporate physics as conditions into VDM, enabling the generation of physically plausible motion. In the first stage, the VLM serves as a coarse-level motion planner, while a VDM serves as a fine-level motion synthesizer. The idea of stage one is to utilize the chain-of-thought and the physics-aware reasoning of VLM planning to ensure that coarse-level motion trajectories approximately follow real-world physics dynamics. In stage two, we can generate fine-level motion using an image-to-video diffusion model conditioned by the approximated path/change planned by VLM from stage one. Note that the approximated path/changes are not in the level to tell the speed or acceleration of the motion. We choose an existing image-to-video model [7] to accept our coarse-level path/change, by injecting noise to the motion path during both the training and inference phases. Notably, during the VLM planning stage, generating entire physically plausible motion trajectories is not required. Instead, we leverage the generative priors of VDM to produce fine-level physically plausible videos based on coarse-level motion trajectories provided by the VLM. So that the detail-level motion such + +as speed, acceleration, and vibration are left to the VDM to synthesize. + +We evaluate our physically plausible video generation framework with two major video physics benchmarks and achieved satisfactory results. Furthermore, we discuss and analyze multiple insightful design choices in our video generation framework, such as employing a motion planner tailored for different physics categories, and enhancing the robustness of diffusion model to noisy trajectories. Our contributions are summarized as follows: + +1. We introduce a novel image-to-video generation framework for generating physically plausible videos by leveraging the VLM and VDM priors, significantly outperforming the contemporary competitors. +2. We propose a novel chain-of-thought and physics-aware reasoning approach in VLM, along with random noise injection in the latent space during video generation, which effectively improves both the generation quality and physical plausibility. +3. We conduct a comprehensive experiments and user studies to demonstrate the effectiveness and generalization of our framework in physically plausible videos generation. + +# 2. Related Work + +# 2.1. Physically Plausible Visual Content Generation + +Generating physically plausible videos offers substantial value to real-world applications such as scientific simulations[43], robotics [4, 59], and autonomous driving [10, 48]. Traditional graphics pipelines rely on simulation systems to model physical phenomena [36, 42]. Inspired by these approaches, recent studies [18, 29] have performed dynamic simulations in image space based on physical engines. Furthermore, some methods [56, 64] incorporate physical priors into 3D representations to enable the synthesis of physically plausible motions. However, these rule-based or solver-based simulators face limitations in expressiveness, efficiency, generalizability, and parameter tuning. Furthermore, these simulators require significant expertise, rendering them inaccessible and unfriendly for users. + +In addition, Some studies have explored VDMs for generating physically plausible videos. Li et al. [24] models natural oscillations and swaying in frequency-domain. A downstream rendering module then animates static images based on the generated motion information. PhysDiff [62] introduces physical simulator as constrain into the diffusion process by projecting denoised motion of a diffusion step into a physically plausible motion. These methods mainly focus only on specific types of physical motion and do not establish a generalizable approach for generating physically plausible videos. + +![](images/b5eb454f9f4faae16e9e8ffa131f4cd288ef84854939ec883963a44a26522811.jpg) + +![](images/fb5a4a524ec9ac71cf728e1b3b2e3cfb1c10c034f413437c6fdd8c1271602372.jpg) +Figure 2. The illustration of our physically plausible image-to-video generation pipeline. Our pipeline consists of two stages. In the first stage, the VLM generates a coarse-grained, physically plausible motion trajectory based on the provided input conditions. In the second stage, We simulate a draft video using the predicted trajectory to provide the motion condition. We then extract the optical flow from this video and convert it into structured noise. These conditions are fed into a motion controllable image-to-video diffusion model, and ultimately generates a physically plausible video. + +# 2.2. Motion Controllable Video Generation + +Existing studies commonly provide one of the following three types of motion control: bounding box control [19, 23, 30, 46, 52], point trajectory control [34, 37, 40, 50, 54] and camera control [2, 15, 47, 61]. Bounding box control provides object motion guidance by generating a sequence of bounding boxes that track the object's position over time. Point-trajectory control offers motion cues through point-based trajectories, enabling drag-style manipulation. Camera motion control guides video generation using explicit 3D camera parameters, ensuring consistency and realistic viewpoint changes. However, these approaches prioritize motion control but often overlook physical plausibility. To address the limitation, we propose a novel framework for physically plausible video generation that incorporates physics as conditions into video diffusion models. + +# 2.3. Generation based on VLMs Planning + +VLMs have exhibited robust capabilities in visual understanding and planning [27, 39, 65]. Their strong performance in domains such as robot path planning and video understanding shows their ability in understanding the real physical world. Prior work has successfully leveraged LLMs to guide the layout of images or videos, yielding promising results [25, 53]. VideoDirectorGPT [26] leverages LLMs for fine-grained scene-by-scene planning, explicitly controlling spatial layout to generate temporally consistent long videos. Pandora [55] utilizes LLMs for real-time control through free-text action commands, achieving domain generality, + +video consistency, and controllability. However, these efforts have yet to address interactions with real-world physical phenomena, such as collision, fall, and melting. + +Moreover, the absence of visual information can cause severe hallucination issues in language models for spatial planning tasks, leading to problems like overlapping object boundaries, disproportionate scaling, and incorrect planning [20, 57]. In this paper, we propose utilizing VLMs as coarse-level motion planners within the image space and incorporate physics-aware reasoning and Chain of Thought (CoT) [51] into the inference process. + +# 3. Method + +Task Formulation. In this paper, our goal is to enable an image-to-video diffusion model to generate physically plausible videos. Since VDMs rely more on memory and case-based imitation and struggle to understand general physical rules[21], the key challenge is how to incorporate physical laws into the models. To achieve this, we need to identify a method to incorporate physical principles into the video diffusion framework. Given an image $I \in R^{H \times W \times C}$ ( $H$ is height, $W$ is width and $C$ is the number of channels) and a text description $d$ of possible events based on image $I$ , our framework should infer a physics-compliant guidance as the input condition and synthesize a video that adheres to both physical laws and real-world dynamics. + +Overall Pipeline. Overall pipeline of VLIPP is illustrated in Figure 2. In the first stage, the VLM conducts semantic + +![](images/d274a243e9f6ce3afbd45c5ffaab4344c224e10667fe5436391a2daed76f3c9e.jpg) +Figure 3. The illustration of chain-of-thought reasoning in the VLM for generating a coarse-grained motion trajectory. First off, the VLM determines the corresponding physical laws and its context for the given scene. Then, the VLM performs step-by-step reasoning to predict the physically plausible motions of objects in image space, leveraging physical context and chain-of-thought prompting. Finally, the VLM predicts bounding boxes according to real-world physics. + +analysis and physical attribute analysis on the given image $I$ and a description $d$ to obtain the bounding boxes of the objects in the scene, denoted as $b_{1}, b_{2}, \ldots, b_{n}$ , along with the applicable physical laws $l$ . Next, the VLM infers possible future physical scenarios in the current scene to derive the coarse-level motion trajectories of $p_{1}, p_{2}, \ldots, p_{t}$ in the image space. Finally, we utilize an image-to-video diffusion model to synthesize the detailed dynamics in the video. + +# 3.1. VLM as a Coarse-Level Motion Planner + +Our motivation is to incorporate physical laws as constraints into a video diffusion model to enhance the physical plausibility of the generated videos. To achieve this, we must identify a method to inject physical laws into the video diffusion model. Given a video description and the first frame, the task at this stage is to generate coarse-level motion trajectory aligned with physical laws. + +Scene Understanding. In the real world, most physical phenomena arise from interactions among objects and their motion trajectories. We first initiate the process by identifying and locating objects within a scene. Inspired by the recent studies [1, 9, 28] in VLMs for scene understanding, we employ GPT-4o [38] to recognize all objects that could be involved in physical phenomena as described in the text description $d$ . These objects are subsequently detected and segmented using Grounded-SAM2 [41], yielding their bounding boxes. By leveraging the pretrained knowledge and common-sense reasoning capabilities of foundation models, we effectively determine the relevant objects in the scene. + +Physical-Aware Recognition. To perform more effective reasoning in predicting the motion, it is necessary to determine what specific physical principle to apply in the given context. We utilize the pretrained prior of the LLM to determine the physical laws applicable to the current scene. Following the configuration in the physical benchmark [3, 31, 33], we currently classify common physical phenomena in videos into six categories: gravity, momentum conservation, optics, thermodynamics, magnetism, and fluid mechanics. Note that such list can be easily extended within our framework. Given a video description $d$ , the LLM infers the physical law $l$ that governs the current scene. We provide the specific physical context information for VLM to enhance its understanding of physical laws [11]. Detailed context design is presented in the Appendix. + +Chain of Thought Reasoning in VLM. Given the physical law $l$ , an image $I$ and a video description $d$ for the scene, we prompt the VLM to predict the future bounding box positions of objects within the image-space. We choose to predict in the image space for two primary reasons. Firstly, motion in image space aligns more with our subsequent video synthesizer. Secondly, image space dynamics can effectively represent a wide range of real-world motions [29]. + +At a given time $t$ , the predicted position of $i$ -th object bounding box $b_{i}^{t}$ is denoted as $[x_{t}^{i}, y_{t}^{i}, w_{t}^{i}, h_{t}^{i}]$ , where $(x_{t}^{i}, y_{t}^{i})$ represents its top-left coordinate; $w_{t}^{i} \& h_{t}^{i}$ denote its width and height, respectively. Governed by the physical law, the four values of the bounding box may change over time. The VLM reasons the bounding box positions of $N$ future frames for every object $o_{i}$ based on the condition. To help VLM better understand physical laws, we adapt a chain-of-thought [51] into its reasoning, to significantly enhance its reasoning capabilities. As shown in Figure 3, we formulate our analysis of physical phenomena in videos as step-by-step reasoning: beginning with broad conceptual ideas and progressing to a detailed and practical examination: + +1. Given the physical law $l$ and context information, the VLM analysis video caption and detail the physical law. +2. The VLM analyzes the potential interactions and movement of each object within the scene; +3. The VLM predicts the detailed changes in position and shape of the bounding box corresponding to each object over time. + +Through the structured planning process, the VLM plans coarse-level motion trajectories for the objects, approximating real-world physics dynamics. In particular, our VLM infers the changes of object bounding boxes for next 12 frames, constrained by the token length limitation. To be compatible with the generation process of the chosen VDM in the next stage, these inferred 12 frames are further linearly interpolated to produce a total of 49 frames. + +# 3.2. VDM Serves as a Fine-Level Motion Synthesizer + +In the previous stage, the motion trajectory planned by the VLM is neither precise nor fully compliant with physical laws. On the other hand, while VDM may not be able to produce realistic global motion trajectories, it is able to generate sound motion in finer scale. In this stage, our key insight is that the VDM can refine the coarse-level motion to produce physically plausible motion that aligns with real-world dynamics with its powerful generative prior. + +Motion Animation. To incorporate physical laws into the video diffusion model, we use the inferred coarse motion trajectory to guide the generation process of the diffusion model. Optical flow provides a unified representation of motion, and recent studies [7, 12, 13] have demonstrated its effectiveness in guiding diffusion models. Accordingly, we leverage the coarse-level motion trajectory to animate a draft motion sequence and derive the corresponding optical flow. Specifically, for each object $o_i$ , we extract its bounding box from the first frame and move it to the bounding box location $b_i$ specified by the draft motion trajectory. To animate the change of shape (e.g., due to compression or expansion), we resize object $o_i$ according to the difference between $o_i$ and $o_{i+1}$ during inpainting. The draft motion video is generated as follows: + +$$ +\hat {V} (t) = \operatorname {A n i m a t i o n} \left(B, r s \left(o _ {0} ^ {0}, b _ {t} ^ {0}\right)... r s \left(o _ {0} ^ {i}, b _ {t} ^ {i}\right)\right) \tag {1} +$$ + +where $\hat{V}(t)$ denotes the corresponding frame of inpainted video at timestep $t$ , $B$ denotes the inpainted background with the foreground object removed, $o_i^0$ denotes the $i$ -th object at timestep 0, $b_i^t$ represents the $i$ -th bounding box at timestep $t$ , and $rs$ denotes resize function. + +Structured Noise from Draft Video. Optical flow is an effective representation for guiding VDMs [12, 13]. Follow prior work [7, 8], we employ RAFT[44] to extract optical flow from the draft video and formulate it as structural noise, which retains Gaussian properties. Given the draft video $\hat{V}(t) \in \mathbb{R}^{F \times C \times H \times W}$ , we calculate its per-frame optical flow to get a structured noise tensor $Q \in \mathbb{R}^{F \times C \times H \times W}$ . The structured noise enables the VDM to generate videos that exhibit motion patterns closely aligned with those in the optical flow, thereby improving the realism of the output. + +Noise Injection in Video Synthesis. We adopt Go-with-the-Flow [7] as our video synthesis model, a fine-tuned CogVideoX [60], which is designed to accept structured noise $Q$ as input and synthesize videos that adhere to the implicit optical flow. The vanilla Go-with-the-Flow tends to tightly follow the provided structured noise $Q$ . However, our $Q$ is derived from a coarse-level motion trajectory and may not be sufficiently accurate to follow the physical laws of the real world. To address this limitation, we inject noise during the inference phase to give more flexibility to the VDM to + +generate detail-level motion changes as + +$$ +Q _ {i} = \frac {(1 - \gamma) Q _ {i} + \zeta \gamma}{\sqrt {(1 - \gamma) ^ {2} + \gamma^ {2}}} \tag {2} +$$ + +where $Q_{i}$ is structured noise at $i$ -th frames, $\zeta \in \mathbb{R}^{C\times H\times W}$ is Gaussian noise and $\gamma \in [0,1]$ . We set $\gamma = 0.4$ for even frame index and $\gamma = 0.6$ for odd frame index. + +With this approach, the VDM is able to generate motion deviate from the coarse-level motion trajectory whenever necessary for producing high-quality fine-level motion. + +# 4. Empirical Analysis and Dissusion + +In this section, we conduct extensive experiments to demonstrate the effectiveness of our video generation framework compared to existing methods. We evaluate our approach on two established benchmarks for physically plausible video generation. Our framework consistently achieves superior performance across all benchmarks. + +# 4.1. Implementation Details + +We propose a two-stage physically plausible image-to-video generation framework. In the first stage, we utilize ChatGPT-4o as the coarse-level motion planner. In the second stage, we utilize an open-source I2V model, Go-with-the-Flow [7], as a fine-level motion synthesizer. Unless otherwise specified, in all experiments, we generate each video with a resolution of $720 \times 480$ and 49 frames. + +# 4.2. Benchmarks and Models + +Traditional metrics in the visual domain, such as the Peak Signal-to-Noise Ratio (PSNR)[17], the Structural Similarity Index (SSIM)[49], the Learned Perceptual Image Patch Similarity (LPIPS)[63], the Fréchet Inception Distance (FID)[16] and the Fréchet Video Distance (FVD)[45], do not account for the physical realism of the generated videos [31, 33]. Recent studies have begun to address this limitation by developing benchmarks and metrics that evaluate physical realism. In this work, we adopt two benchmarks, described below. + +PhyGenBench [31] categorizes physical properties into four domains: mechanics, optics, thermal, and material. It includes 27 physical phenomena, each governed by real world physical laws, reflected in 160 carefully designed text prompts. As PhyGenbench provides only text prompts, we adapt it to our image-to-video setting by generating a corresponding first frame for each prompt with FLUX[22]. We adhere to the predefined benchmark evaluation protocol, i.e., employing GPT-4o to assess the physical realism of the generated videos. + +Physics-IQ [33] comprises 396 real-world videos spanning 66 distinct physical scenarios. For each scenario, videos + +![](images/7492de47fea042983fdf07f7c6ba6374f76eb600e81b4090eecd926e22a1d5aa.jpg) +Figure 4. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B [60], LTX-VideoI2V [14] and SVD-XT [5]. + +
ModelMechanics(↑)Optics(↑)Thermal(↑)Material(↑)Average(↑)
CogvideoX-T2V-5B0.430.550.400.420.45
LTX-Video-T2V0.350.450.360.380.39
OpenSora0.430.500.440.370.44
PhyT2V0.490.610.490.470.52
LLM-Grounding Video Diffusion0.320.410.260.240.31
CogvideoX-I2V-5B0.480.690.430.410.52
SVD-XT0.460.680.480.410.52
LTX-Video-I2V0.470.650.460.370.50
SG-I2V0.520.690.510.390.54
Ours0.550.710.600.530.60
+ +Table 1. Quantitative results of VDMs on PhyGenBench. + +are recorded from three different perspectives and filmed twice under identical conditions to eliminate randomness. This benchmark evaluates real-world physical phenomena, including collisions, object continuity, occlusion, object permanence, and fluid dynamics. This benchmark assesses physical realism from semantic and temporal perspectives, using semantic metrics and visual metrics to compare generated videos against the real-world reference videos. + +Compared Models. In the context of text-to-video generation, we compare our framework with CogVideoX-T2V-5B [60], LTX-Video-T2V [14], and OpenSora [66]. Moreover, we evaluate our framework against PhyT2V [58], which enhances physical realism by iteratively refining the prompt. For the image-to-video generation scenario, our framework is evaluated alongside CogVideoX-I2V-5B [60], SVD-XT [5], and LTX-Video-I2V [14]. Additionally, we conducted experiments in the motion-controllable setting. + +In this setting, we leverage the motion trajectory predicted by the VLM as a condition to guide VDM generation. We benchmark our approach against image-to-video motion controllable model, SG-I2V [35] and text-to-video motion controllable model LLM-grounded Video Diffusion Models [25]. The experimental details are presented in the Appendix. + +# 4.3.Quantitative Evaluation + +We begin with an empirical study on PhyGenBench and Physics-IQ, comparing our framework against widely adopted open-source models in the research community. Based on different physical properties, we categorize the benchmark samples accordingly. Additionally, we classify VDMs into text-to-video (T2V) diffusion models and imaged-to-video (I2V) diffusion models based on the input conditions. + +In Table 1, we present our experimental results on Phy + +
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Cogvideo-I2V-5B30.429.816.713.38.527.1
SVD-XT21.920.56.88.417.119.1
LTX-Video-I2V30.229.815.913.28.426.8
SG-I2V34.631.215.913.18.429.7
Ours42.334.116.913.48.834.6
+ +Table 2. Quantitative results of physically plausible video generation on Physics-IQ Benchmark. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics. + +GenBench, evaluating different video generation models following its evaluation protocol. The results show that our framework achieves state-of-the-art performance across four different physical phenomena. Our framework outperforms the best T2V method by an average of $15.3\%$ and the best I2V method by $11.1\%$ . Specifically, our framework demonstrates significant advantages in the Mechanics, Thermal, and Material domains, outperforming the best I2V method by $5.7\%$ , $17.6\%$ , and $35.8\%$ , respectively. These advantages are particularly evident in these three types of physical phenomena, which involve more substantial changes in motion, volume, or shape. Our framework is better equipped to understand and reason about bounding box sequences to represent these changes effectively. + +Similarly, for the Physics-IQ benchmark, we evaluate the performance of different video generation models following its evaluation protocol. Our framework achieves the best results across four different physical phenomena, with improvements of $22.2\%$ in Solid Mechanics and $9.2\%$ in Fluid Dynamics compared to the second-best models. These significant improvements demonstrate the effectiveness of our framework in generating physically plausible videos. + +# 4.4. Qualitative Evaluation + +Figures 4 and 5 demonstrate a qualitative comparison between our video generation framework and baseline methods. Among all evaluated approaches, our framework consistently produces videos with the highest degree of physical realism. In the ball falling sample in Figure 4, while CogVideoX shows a bouncing effect, artifacts are present in the video; LTX-Video and SVD-XT exhibit motions that do not adhere to the laws of physics. In Figure 5, we analyze two examples from Physics-IQ. In the pouring water example, the baseline methods fail to show the simultaneous decrease in water level of the glass beverage dispenser and the increase in water level of the glass below; in the ball collision example, none of the baseline methods correctly depict the collision of balls. More videos are provided in the supplement. + +![](images/26ec347cb5f5715b4193d855f46b0440b39bab2e8a47858a4a354fa123ceb8e9.jpg) +Figure 5. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B, LTX-Video-I2V, SVD-XT and SG-I2V [35] in the Physics-IQ dataset. + +# 4.5. Ablation Study + +We perform an ablation study to evaluate the contributions of key components in our framework. We design four variants to analyze the effectiveness of different components in our framework. + +1. Ours w/o VLM Planner: To assess the overall functionality of our framework, we replace the structured noise input of the VDM with random noise to evaluate the effectiveness of the VLM planner. +2. Ours w/o CI: Keeping the overall structure unchanged, we remove the in-context information from the VLM. +3. Ours w/o CoT: Similarly, while keeping other components unchanged, we remove the CoT reasoning process from the VLM. +4. Ours w/o CC: Lastly, we remove both the in-context information and the CoT reasoning process from the VLM planner while maintaining all other components. + +
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Ours42.334.116.913.48.834.9
w/o VLM Planner.16.320.813.45.85.616.2
w/o C.I26.328.116.911.28.424.3
w/o CoT21.426.916.18.66.921.0
w/o C.C18.722.414.97.26.118.1
+ +Table 3. Ablation study on VLM, in-context learning and COT. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics. + +![](images/4d28e9559c9aff9e594c3ff64214b2436dbcac8807ae8729f0dd5f9a6b9bb5e6.jpg) +Figure 6. Visualization of generated video with bounding boxes. Coarse motion cue can guide VDM to synthesis physically plausible video. + +![](images/4a631e4ead90c7ffcdaa6d113c0af96248a053240a719a78fe6e70baa243ac2e.jpg) +Figure 7. Limitation of VLIPP. Our framework might produces physically implausible videos because it lacks 3D spatial perception and the optical-flow of small objects is prone to noise. + +Table 3 presents a quantitative comparison between our full method and these variants. Among all variants, Ours w/o VLM Planner shows the most significant performance drop, as removing the planner completely eliminates our ability in understanding the physical laws, leading to nearly random results. Notably, Ours w/o CoT exhibits a more pronounced decline compared to Ours w/o CI, indicating that the reasoning process in CoT enhances the understanding of physics. While in-context information contributes to the physical reasoning ability of VLM, compared to CoT it is less effective in preventing errors caused by VLM hallucination. + +# 4.6. Limitations + +Although VLIPP can generate physically plausible videos, its performance remains constrained by the base model. + +Firstly, we are unable to model physical events that cannot be represented by image space bounding box trajectories. For example, phenomena that involve intrinsic state changes of objects such as gas solidification. We adopt 2D bounding boxes because they can be naturally integrated with VDMs and are effective in enhancing the physical plausibility to a certain extent. In fact, they are more flexible than they might appear. As shown in Figure 6, the VLM predicts coarse 2D bounding box trajectories(only four fragments), yet the VDM still produces a realistic shattering pattern without being rigidly constrained by those paths. Moreover, in the butter melting example in Figure 6, the VLM offers only a coarse cue of volumetric change, but the VDM successfully generates the transition from solid to liquid. These abilities significantly improve our flexibility in modeling various physical phenomena. + +Secondly, our pipeline lacks 3D spatial perception. It is unable to understand the spatial relationships within the scene, as shown in Figure 7 domino example. + +Finally, the optical flow of small objects is prone to noise interference. This will cause our framework to generate ambiguous content, as shown in Figure 7 small ball collision example. With the recent progress in video generation model, we anticipate that our framework will be further improved in generating videos under more challenging physical conditions. + +# 5. Conclusion + +Recently, VDMs have achieved great empirical success and are receiving considerable attention in computer vision and computer graphics. However, due to the lack of understanding of physical laws, VDMs are unable to generate physically plausible videos. In this paper, we introduce VLIPP, a novel two-stage physically plausible video generation framework that incorporates physical laws into video diffusion models through vision and language informed physical prior. Our experimental results demonstrate the effectiveness of our method compared to existing approaches. + +# Acknowledgements + +This work was partially supported by the National Natural Science Foundation of China under Grant No. 62441231, 62472065, and U23B2010. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4 +[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3 +[3] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 4 +[4] Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. arXiv preprint arXiv:2412.03572, 2024. 2 +[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 6 +[6] Sébastien Bubeck, Varun Chadrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4, 2023. 2 +[7] Ryan Burgert, Yuancheng Xu, Wenqi Xian, Oliver Pilarski, Pascal Clausen, Mingming He, Li Ma, Yitong Deng, Lingxiao Li, Mohsen Mousavi, et al. Go-with-the-flow: Motion-controllable video diffusion models using real-time warped noise. arXiv preprint arXiv:2501.08331, 2025. 2, 5 +[8] Pascal Chang, Jingwei Tang, Markus Gross, and Vinicius C Azevedo. How i warped your noise: a temporally-correlated noise prior for diffusion models. In The Twelfth International Conference on Learning Representations. OpenReview, 2024. 5 +[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4 +[10] Boyang Deng, Richard Tucker, Zhengqi Li, Leonidas Guibas, Noah Snavely, and Gordon Wetzstein. Streetscapes: Large-scale consistent street view generation using autoregressive video diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[11] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu + +Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4 +[12] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. arXiv preprint arXiv:2401.18085, 2024. 5 +[13] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 5 +[14] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.6 +[15] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractr: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[17] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5 +[18] Hao-Yu Hsu, Zhi-Hao Lin, Albert Zhai, Hongchi Xia, and Shenlong Wang. Autovfx: Physically realistic video editing from natural language instructions. arXiv preprint arXiv:2411.02394, 2024. 2 +[19] Hsin-Ping Huang, Yu-Chuan Su, Deqing Sun, Lu Jiang, Xuhui Jia, Yukun Zhu, and Ming-Hsuan Yang. Fine-grained controllable video generation via object appearance and context. arXiv preprint arXiv:2312.02919, 2023. 3 +[20] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1-55, 2025. 3 +[21] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 2, 3 +[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.5 +[23] Pengxiang Li, Kai Chen, Zhili Liu, Ruiyuan Gao, Lanqing Hong, Guo Zhou, Hua Yao, Dit-Yan Yeung, Huchuan Lu, and Xu Jia. Trackdiffusion: Tracklet-conditioned video generation via diffusion models. arXiv preprint arXiv:2312.00651, 2023. 3 +[24] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 2 + +[25] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 3, 6 +[26] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024. 4 +[29] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2024. 2, 4 +[30] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3 +[31] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 4, 5 +[32] Antonio Montanaro, Luca Savant Aira, Emanuele Aiello, Diego Valsesia, and Enrico Magli. Motioncraft: Physics-based zero-shot video generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2 +[33] Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models learn physical principles from watching videos? arXiv preprint arXiv:2501.09038, 2025. 2, 4, 5 +[34] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. Advances in Neural Information Processing Systems, 37:18481-18505, 2024. 3 +[35] Koichi Namekata, Sherwin Bahmani, Ziyi Wu, Yash Kant, Igor Gilitschenski, and David B Lindell. Sg-i2v: Self-guided trajectory control in image-to-video generation. arXiv preprint arXiv:2411.04989, 2024. 6, 7 +[36] Andrew Nealen, Matthias Müller, Richard Keiser, Eddy Boxerman, and Mark Carlson. Physically based deformable models in computer graphics. In Computer graphics forum, pages 809-836. Wiley Online Library, 2006. 2 +[37] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 3 +[38] Openai. https://openai.com/index/hello-gpt-4o/, 2024. 4 +[39] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 3 + +[40] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. Freetraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 3 +[41] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4 +[42] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 2 +[43] Weixiang Sun, Xiaocao You, Ruizhe Zheng, Zhengqing Yuan, Xiang Li, Lifang He, Quanzheng Li, and Lichao Sun. Bora: Biomedical generalist video generation model. arXiv preprint arXiv:2407.08944, 2024. 2 +[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5 +[45] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 5 +[46] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 3 +[47] Xi Wang, Robin Courant, Marc Christie, and Vicky Kalogeiton. Akira: Augmentation kit on rays for optical video generation. arXiv preprint arXiv:2412.14158, 2024. 3 +[48] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 2 +[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[50] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824-24837, 2022. 3, 4 +[52] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 3 +[53] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting ltm-controlled diffusion + +models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3 +[54] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3 +[55] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3 +[56] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2 +[57] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 3 +[58] Qiyao Xue, Xiangyu Yin, Boyuan Yang, and Wei Gao. Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation. arXiv preprint arXiv:2412.00596, 2024. 1, 6 +[59] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 1(2):6, 2023. 2 +[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 5, 6 +[61] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3 +[62] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 2 +[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[64] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2024. 2 +[65] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3 + +[66] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 6 \ No newline at end of file diff --git a/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/images.zip b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ceae77c997a549043ac414dbd753e601257eb262 --- /dev/null +++ b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8487439b39c22e704c03d4528425f5e83efef85ba9dcb9251eb8e8b4ad79489e +size 599300 diff --git a/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/layout.json b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..55406c23bfb800ba9f1f892d4959cf31d2e693d5 --- /dev/null +++ b/2025/VLIPP_ Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior/layout.json @@ -0,0 +1,8213 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 114, + 471, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 114, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 140, + 114, + 471, + 148 + ], + "type": "text", + "content": "VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "spans": [ + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": "Xindi Yang" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Baolu Li" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Yiming Zhang" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Zhenfei Yin" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{4,5\\dagger}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Lei Bai" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Liqian Ma" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Zhiyong Wang" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Jianfei Cai" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Tien-Tsin Wong" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Huchuan Lu" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": ", Xu Jia" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": "Monash University " + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": "Dalian University of Technology " + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": "Oxford University " + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": "The University of Sydney " + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 91, + 168, + 519, + 224 + ], + "type": "text", + "content": "ZMO AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 251, + 204, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 251, + 204, + 262 + ], + "spans": [ + { + "bbox": [ + 157, + 251, + 204, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 274, + 297, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 274, + 297, + 587 + ], + "spans": [ + { + "bbox": [ + 64, + 274, + 297, + 587 + ], + "type": "text", + "content": "Video diffusion models (VDMs) have advanced significantly in recent years, enabling the generation of highly realistic videos and drawing the attention of the community in their potential as world simulators. However, despite their capabilities, VDMs often fail to produce physically plausible videos due to an inherent lack of understanding of physics, resulting in incorrect dynamics and event sequences. To address this limitation, we propose a novel two-stage image-to-video generation framework that explicitly incorporates physics with vision and language informed physical prior. In the first stage, we employ a Vision Language Model (VLM) as a coarse-grained motion planner, integrating chain-of-thought and physics-aware reasoning to predict a rough motion trajectories/changes that approximate real-world physical dynamics while ensuring the inter-frame consistency. In the second stage, we use the predicted motion trajectories/changes to guide the video generation of a VDM. As the predicted motion trajectories/changes are rough, noise is added during inference to provide freedom to the VDM in generating motion with more fine details. Extensive experimental results demonstrate that our framework can produce physically plausible motion, and comparative evaluations highlight the notable superiority of our approach over existing methods. More video results and code are available on our Project Page: https://madaoer.github.io/projects/physically_plausible Video_generation/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 607, + 143, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 607, + 143, + 619 + ], + "spans": [ + { + "bbox": [ + 64, + 607, + 143, + 619 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 626, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 626, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 64, + 626, + 296, + 685 + ], + "type": "text", + "content": "Video diffusion models (VDMs) trained on large-scale video datasets have made remarkable progress in terms of realism, demonstrating significant potential for various content creation applications. Despite the absence of explicit geometric modeling, the generated videos still exhibit coherent spatial" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 255, + 544, + 326 + ], + "blocks": [ + { + "bbox": [ + 317, + 255, + 544, + 326 + ], + "lines": [ + { + "bbox": [ + 317, + 255, + 544, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 544, + 326 + ], + "type": "image", + "image_path": "c9ea32fa753f62ad355288a590600f15f3482c20677eaef9919f606bdafc3e54.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 327, + 543, + 398 + ], + "blocks": [ + { + "bbox": [ + 317, + 327, + 543, + 398 + ], + "lines": [ + { + "bbox": [ + 317, + 327, + 543, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 543, + 398 + ], + "type": "image", + "image_path": "1c8ac1d0e4311bf54beed485a09ec0c48140d89ca69db2bda1ddec2cce87568c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 400, + 543, + 471 + ], + "blocks": [ + { + "bbox": [ + 317, + 400, + 543, + 471 + ], + "lines": [ + { + "bbox": [ + 317, + 400, + 543, + 471 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 543, + 471 + ], + "type": "image", + "image_path": "2363cedd545e1a12a5482ed879a142d6a192d5c1705c1f5f34e5bcc79608a29e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 478, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 545, + 521 + ], + "type": "text", + "content": "Figure 1. Existing commercial closed-source VDMs fail to generate physically plausible motion, whereas our video generation framework is able to achieve this by incorporating external physical prior knowledge." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 538, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 538, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 312, + 538, + 545, + 631 + ], + "type": "text", + "content": "relationships among objects, rich textured details, and realistic lighting effects, including reflections and shadows. Such qualities often make the generated videos nearly indistinguishable from real-world footages. This drives the research community to explore the potential of VDMs as world models. However, they still struggle with understanding the physical laws of the real world and generating videos that adhere to these principles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 632, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 632, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 312, + 632, + 545, + 702 + ], + "type": "text", + "content": "Although existing VDMs can produce visually realistic videos, they fail to mimic the real-dynamic physical motions. As shown in Fig 1, even the current commercial closed-source VDMs struggle with the task of generating videos that conform to physical laws. PhyT2V [58] refines text prompts by incorporating detailed descriptions" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 691, + 227, + 701 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 691, + 227, + 701 + ], + "spans": [ + { + "bbox": [ + 78, + 691, + 227, + 701 + ], + "type": "text", + "content": "* Equal contribution † Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "text", + "content": "12360" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 84, + 294, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 84, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 64, + 84, + 294, + 245 + ], + "type": "text", + "content": "of physical processes to guide VDMs in generating physically plausible videos. However, despite being pretrained on internet-scale real-world video-text pairs, VDMs do not inherently understand physical laws. This limitation arises from the gap and ambiguity between text descriptions and the actual motion in the video[33]. Moreover, VDMs tend to overfit the training data rather than developing a general understanding of physical laws[21]. Inspired by the success of graphics-based physical rendering, some methods have guided VDMs to generate physically plausible videos using simulations from graphics engines[18, 29, 32, 56, 64]; however, these approaches rely on the physical effects that graphics engines can simulate and incur high computational costs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 249, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 249, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 64, + 249, + 294, + 456 + ], + "type": "text", + "content": "The gap and ambiguity between text and real-world motion makes it difficult to enable physically plausible video generation through detailed text descriptions alone. Moreover, it is challenging to gather scalable physical data for training due to the abstractness and diversity of physical phenomena. Consequently, a viable approach could be to model abstract physical laws as conditions for diffusion models. However, it is less practical to explicitly model the physics equation for every kind of motion. Instead, we resort to current large foundation models for their ability to \"understand\" basic physics[6] and reason about physical phenomena based on the knowledge they extract. For example, given two colliding balls, the Large Language Model (LLM) can approximately predict the paths of the balls after collision. Inspired by this observation, we propose a novel video generation framework that employs a Vision Language Model (VLM) to predict the path/change during a physics event, described by a given image and a text prompt." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 460, + 294, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 460, + 294, + 701 + ], + "spans": [ + { + "bbox": [ + 64, + 460, + 294, + 701 + ], + "type": "text", + "content": "In this paper, we propose VLIPP, a two-stage approach to incorporate physics as conditions into VDM, enabling the generation of physically plausible motion. In the first stage, the VLM serves as a coarse-level motion planner, while a VDM serves as a fine-level motion synthesizer. The idea of stage one is to utilize the chain-of-thought and the physics-aware reasoning of VLM planning to ensure that coarse-level motion trajectories approximately follow real-world physics dynamics. In stage two, we can generate fine-level motion using an image-to-video diffusion model conditioned by the approximated path/change planned by VLM from stage one. Note that the approximated path/changes are not in the level to tell the speed or acceleration of the motion. We choose an existing image-to-video model [7] to accept our coarse-level path/change, by injecting noise to the motion path during both the training and inference phases. Notably, during the VLM planning stage, generating entire physically plausible motion trajectories is not required. Instead, we leverage the generative priors of VDM to produce fine-level physically plausible videos based on coarse-level motion trajectories provided by the VLM. So that the detail-level motion such" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 84, + 544, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 544, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 544, + 106 + ], + "type": "text", + "content": "as speed, acceleration, and vibration are left to the VDM to synthesize." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 110, + 544, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 544, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 544, + 201 + ], + "type": "text", + "content": "We evaluate our physically plausible video generation framework with two major video physics benchmarks and achieved satisfactory results. Furthermore, we discuss and analyze multiple insightful design choices in our video generation framework, such as employing a motion planner tailored for different physics categories, and enhancing the robustness of diffusion model to noisy trajectories. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 204, + 543, + 340 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 315, + 204, + 543, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 204, + 543, + 249 + ], + "spans": [ + { + "bbox": [ + 315, + 204, + 543, + 249 + ], + "type": "text", + "content": "1. We introduce a novel image-to-video generation framework for generating physically plausible videos by leveraging the VLM and VDM priors, significantly outperforming the contemporary competitors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 543, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 543, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 543, + 306 + ], + "type": "text", + "content": "2. We propose a novel chain-of-thought and physics-aware reasoning approach in VLM, along with random noise injection in the latent space during video generation, which effectively improves both the generation quality and physical plausibility." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 308, + 543, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 543, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 543, + 340 + ], + "type": "text", + "content": "3. We conduct a comprehensive experiments and user studies to demonstrate the effectiveness and generalization of our framework in physically plausible videos generation." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 360, + 397, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 397, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 397, + 371 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 544, + 392 + ], + "type": "text", + "content": "2.1. Physically Plausible Visual Content Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 400, + 544, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 544, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 544, + 570 + ], + "type": "text", + "content": "Generating physically plausible videos offers substantial value to real-world applications such as scientific simulations[43], robotics [4, 59], and autonomous driving [10, 48]. Traditional graphics pipelines rely on simulation systems to model physical phenomena [36, 42]. Inspired by these approaches, recent studies [18, 29] have performed dynamic simulations in image space based on physical engines. Furthermore, some methods [56, 64] incorporate physical priors into 3D representations to enable the synthesis of physically plausible motions. However, these rule-based or solver-based simulators face limitations in expressiveness, efficiency, generalizability, and parameter tuning. Furthermore, these simulators require significant expertise, rendering them inaccessible and unfriendly for users." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 575, + 544, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 544, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 544, + 700 + ], + "type": "text", + "content": "In addition, Some studies have explored VDMs for generating physically plausible videos. Li et al. [24] models natural oscillations and swaying in frequency-domain. A downstream rendering module then animates static images based on the generated motion information. PhysDiff [62] introduces physical simulator as constrain into the diffusion process by projecting denoised motion of a diffusion step into a physically plausible motion. These methods mainly focus only on specific types of physical motion and do not establish a generalizable approach for generating physically plausible videos." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "text", + "content": "12361" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 83, + 542, + 175 + ], + "blocks": [ + { + "bbox": [ + 67, + 83, + 542, + 175 + ], + "lines": [ + { + "bbox": [ + 67, + 83, + 542, + 175 + ], + "spans": [ + { + "bbox": [ + 67, + 83, + 542, + 175 + ], + "type": "image", + "image_path": "b5eb454f9f4faae16e9e8ffa131f4cd288ef84854939ec883963a44a26522811.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 176, + 542, + 297 + ], + "blocks": [ + { + "bbox": [ + 67, + 176, + 542, + 297 + ], + "lines": [ + { + "bbox": [ + 67, + 176, + 542, + 297 + ], + "spans": [ + { + "bbox": [ + 67, + 176, + 542, + 297 + ], + "type": "image", + "image_path": "fb5a4a524ec9ac71cf728e1b3b2e3cfb1c10c034f413437c6fdd8c1271602372.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 300, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 63, + 300, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 63, + 300, + 545, + 354 + ], + "type": "text", + "content": "Figure 2. The illustration of our physically plausible image-to-video generation pipeline. Our pipeline consists of two stages. In the first stage, the VLM generates a coarse-grained, physically plausible motion trajectory based on the provided input conditions. In the second stage, We simulate a draft video using the predicted trajectory to provide the motion condition. We then extract the optical flow from this video and convert it into structured noise. These conditions are fed into a motion controllable image-to-video diffusion model, and ultimately generates a physically plausible video." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 361, + 261, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 361, + 261, + 373 + ], + "spans": [ + { + "bbox": [ + 64, + 361, + 261, + 373 + ], + "type": "text", + "content": "2.2. Motion Controllable Video Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 379, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 379, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 63, + 379, + 296, + 551 + ], + "type": "text", + "content": "Existing studies commonly provide one of the following three types of motion control: bounding box control [19, 23, 30, 46, 52], point trajectory control [34, 37, 40, 50, 54] and camera control [2, 15, 47, 61]. Bounding box control provides object motion guidance by generating a sequence of bounding boxes that track the object's position over time. Point-trajectory control offers motion cues through point-based trajectories, enabling drag-style manipulation. Camera motion control guides video generation using explicit 3D camera parameters, ensuring consistency and realistic viewpoint changes. However, these approaches prioritize motion control but often overlook physical plausibility. To address the limitation, we propose a novel framework for physically plausible video generation that incorporates physics as conditions into video diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 558, + 254, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 558, + 254, + 571 + ], + "spans": [ + { + "bbox": [ + 64, + 558, + 254, + 571 + ], + "type": "text", + "content": "2.3. Generation based on VLMs Planning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 574, + 296, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 574, + 296, + 702 + ], + "spans": [ + { + "bbox": [ + 63, + 574, + 296, + 702 + ], + "type": "text", + "content": "VLMs have exhibited robust capabilities in visual understanding and planning [27, 39, 65]. Their strong performance in domains such as robot path planning and video understanding shows their ability in understanding the real physical world. Prior work has successfully leveraged LLMs to guide the layout of images or videos, yielding promising results [25, 53]. VideoDirectorGPT [26] leverages LLMs for fine-grained scene-by-scene planning, explicitly controlling spatial layout to generate temporally consistent long videos. Pandora [55] utilizes LLMs for real-time control through free-text action commands, achieving domain generality," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 362, + 544, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 544, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 544, + 397 + ], + "type": "text", + "content": "video consistency, and controllability. However, these efforts have yet to address interactions with real-world physical phenomena, such as collision, fall, and melting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 398, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 545, + 491 + ], + "type": "text", + "content": "Moreover, the absence of visual information can cause severe hallucination issues in language models for spatial planning tasks, leading to problems like overlapping object boundaries, disproportionate scaling, and incorrect planning [20, 57]. In this paper, we propose utilizing VLMs as coarse-level motion planners within the image space and incorporate physics-aware reasoning and Chain of Thought (CoT) [51] into the inference process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 504, + 367, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 367, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 367, + 516 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "content": "Task Formulation. In this paper, our goal is to enable an image-to-video diffusion model to generate physically plausible videos. Since VDMs rely more on memory and case-based imitation and struggle to understand general physical rules[21], the key challenge is how to incorporate physical laws into the models. To achieve this, we need to identify a method to incorporate physical principles into the video diffusion framework. Given an image " + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "inline_equation", + "content": "I \\in R^{H \\times W \\times C}" + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "content": " is height, " + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "content": " is width and " + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "content": " is the number of channels) and a text description " + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "content": " of possible events based on image " + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 312, + 525, + 545, + 675 + ], + "type": "text", + "content": ", our framework should infer a physics-compliant guidance as the input condition and synthesize a video that adheres to both physical laws and real-world dynamics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 678, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 545, + 702 + ], + "type": "text", + "content": "Overall Pipeline. Overall pipeline of VLIPP is illustrated in Figure 2. In the first stage, the VLM conducts semantic" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "text", + "content": "12362" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 83, + 292, + 252 + ], + "blocks": [ + { + "bbox": [ + 66, + 83, + 292, + 252 + ], + "lines": [ + { + "bbox": [ + 66, + 83, + 292, + 252 + ], + "spans": [ + { + "bbox": [ + 66, + 83, + 292, + 252 + ], + "type": "image", + "image_path": "d274a243e9f6ce3afbd45c5ffaab4344c224e10667fe5436391a2daed76f3c9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 255, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 64, + 255, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 64, + 255, + 295, + 340 + ], + "type": "text", + "content": "Figure 3. The illustration of chain-of-thought reasoning in the VLM for generating a coarse-grained motion trajectory. First off, the VLM determines the corresponding physical laws and its context for the given scene. Then, the VLM performs step-by-step reasoning to predict the physically plausible motions of objects in image space, leveraging physical context and chain-of-thought prompting. Finally, the VLM predicts bounding boxes according to real-world physics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "text", + "content": "analysis and physical attribute analysis on the given image " + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "text", + "content": " and a description " + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "text", + "content": " to obtain the bounding boxes of the objects in the scene, denoted as " + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "inline_equation", + "content": "b_{1}, b_{2}, \\ldots, b_{n}" + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "text", + "content": ", along with the applicable physical laws " + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "text", + "content": ". Next, the VLM infers possible future physical scenarios in the current scene to derive the coarse-level motion trajectories of " + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "inline_equation", + "content": "p_{1}, p_{2}, \\ldots, p_{t}" + }, + { + "bbox": [ + 64, + 353, + 296, + 445 + ], + "type": "text", + "content": " in the image space. Finally, we utilize an image-to-video diffusion model to synthesize the detailed dynamics in the video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 452, + 267, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 452, + 267, + 463 + ], + "spans": [ + { + "bbox": [ + 64, + 452, + 267, + 463 + ], + "type": "text", + "content": "3.1. VLM as a Coarse-Level Motion Planner" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 468, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 468, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 64, + 468, + 295, + 550 + ], + "type": "text", + "content": "Our motivation is to incorporate physical laws as constraints into a video diffusion model to enhance the physical plausibility of the generated videos. To achieve this, we must identify a method to inject physical laws into the video diffusion model. Given a video description and the first frame, the task at this stage is to generate coarse-level motion trajectory aligned with physical laws." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 552, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 552, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 64, + 552, + 295, + 701 + ], + "type": "text", + "content": "Scene Understanding. In the real world, most physical phenomena arise from interactions among objects and their motion trajectories. We first initiate the process by identifying and locating objects within a scene. Inspired by the recent studies [1, 9, 28] in VLMs for scene understanding, we employ GPT-4o [38] to recognize all objects that could be involved in physical phenomena as described in the text description " + }, + { + "bbox": [ + 64, + 552, + 295, + 701 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 64, + 552, + 295, + 701 + ], + "type": "text", + "content": ". These objects are subsequently detected and segmented using Grounded-SAM2 [41], yielding their bounding boxes. By leveraging the pretrained knowledge and common-sense reasoning capabilities of foundation models, we effectively determine the relevant objects in the scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 84, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 545, + 258 + ], + "type": "text", + "content": "Physical-Aware Recognition. To perform more effective reasoning in predicting the motion, it is necessary to determine what specific physical principle to apply in the given context. We utilize the pretrained prior of the LLM to determine the physical laws applicable to the current scene. Following the configuration in the physical benchmark [3, 31, 33], we currently classify common physical phenomena in videos into six categories: gravity, momentum conservation, optics, thermodynamics, magnetism, and fluid mechanics. Note that such list can be easily extended within our framework. Given a video description " + }, + { + "bbox": [ + 313, + 84, + 545, + 258 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 84, + 545, + 258 + ], + "type": "text", + "content": ", the LLM infers the physical law " + }, + { + "bbox": [ + 313, + 84, + 545, + 258 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 84, + 545, + 258 + ], + "type": "text", + "content": " that governs the current scene. We provide the specific physical context information for VLM to enhance its understanding of physical laws [11]. Detailed context design is presented in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "text", + "content": "Chain of Thought Reasoning in VLM. Given the physical law " + }, + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "text", + "content": ", an image " + }, + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "text", + "content": " and a video description " + }, + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 266, + 545, + 360 + ], + "type": "text", + "content": " for the scene, we prompt the VLM to predict the future bounding box positions of objects within the image-space. We choose to predict in the image space for two primary reasons. Firstly, motion in image space aligns more with our subsequent video synthesizer. Secondly, image space dynamics can effectively represent a wide range of real-world motions [29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": "At a given time " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": ", the predicted position of " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": "-th object bounding box " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "b_{i}^{t}" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "[x_{t}^{i}, y_{t}^{i}, w_{t}^{i}, h_{t}^{i}]" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "(x_{t}^{i}, y_{t}^{i})" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": " represents its top-left coordinate; " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "w_{t}^{i} \\& h_{t}^{i}" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": " denote its width and height, respectively. Governed by the physical law, the four values of the bounding box may change over time. The VLM reasons the bounding box positions of " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": " future frames for every object " + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "inline_equation", + "content": "o_{i}" + }, + { + "bbox": [ + 313, + 365, + 545, + 515 + ], + "type": "text", + "content": " based on the condition. To help VLM better understand physical laws, we adapt a chain-of-thought [51] into its reasoning, to significantly enhance its reasoning capabilities. As shown in Figure 3, we formulate our analysis of physical phenomena in videos as step-by-step reasoning: beginning with broad conceptual ideas and progressing to a detailed and practical examination:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 544, + 601 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 521, + 544, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 544, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 544, + 544 + ], + "type": "text", + "content": "1. Given the physical law " + }, + { + "bbox": [ + 313, + 521, + 544, + 544 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 521, + 544, + 544 + ], + "type": "text", + "content": " and context information, the VLM analysis video caption and detail the physical law." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 544, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 544, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 544, + 567 + ], + "type": "text", + "content": "2. The VLM analyzes the potential interactions and movement of each object within the scene;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 568, + 544, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 544, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 544, + 601 + ], + "type": "text", + "content": "3. The VLM predicts the detailed changes in position and shape of the bounding box corresponding to each object over time." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 609, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 545, + 701 + ], + "type": "text", + "content": "Through the structured planning process, the VLM plans coarse-level motion trajectories for the objects, approximating real-world physics dynamics. In particular, our VLM infers the changes of object bounding boxes for next 12 frames, constrained by the token length limitation. To be compatible with the generation process of the chosen VDM in the next stage, these inferred 12 frames are further linearly interpolated to produce a total of 49 frames." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "text", + "content": "12363" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 83, + 296, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 83, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 64, + 83, + 296, + 106 + ], + "type": "text", + "content": "3.2. VDM Serves as a Fine-Level Motion Synthesizer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 112, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 112, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 64, + 112, + 296, + 205 + ], + "type": "text", + "content": "In the previous stage, the motion trajectory planned by the VLM is neither precise nor fully compliant with physical laws. On the other hand, while VDM may not be able to produce realistic global motion trajectories, it is able to generate sound motion in finer scale. In this stage, our key insight is that the VDM can refine the coarse-level motion to produce physically plausible motion that aligns with real-world dynamics with its powerful generative prior." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "text", + "content": "Motion Animation. To incorporate physical laws into the video diffusion model, we use the inferred coarse motion trajectory to guide the generation process of the diffusion model. Optical flow provides a unified representation of motion, and recent studies [7, 12, 13] have demonstrated its effectiveness in guiding diffusion models. Accordingly, we leverage the coarse-level motion trajectory to animate a draft motion sequence and derive the corresponding optical flow. Specifically, for each object " + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "text", + "content": ", we extract its bounding box from the first frame and move it to the bounding box location " + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "text", + "content": " specified by the draft motion trajectory. To animate the change of shape (e.g., due to compression or expansion), we resize object " + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "text", + "content": " according to the difference between " + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "inline_equation", + "content": "o_{i+1}" + }, + { + "bbox": [ + 64, + 207, + 296, + 379 + ], + "type": "text", + "content": " during inpainting. The draft motion video is generated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 386, + 296, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 386, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 91, + 386, + 296, + 401 + ], + "type": "interline_equation", + "content": "\\hat {V} (t) = \\operatorname {A n i m a t i o n} \\left(B, r s \\left(o _ {0} ^ {0}, b _ {t} ^ {0}\\right)... r s \\left(o _ {0} ^ {i}, b _ {t} ^ {i}\\right)\\right) \\tag {1}", + "image_path": "8e506a4de5129bdb53a784329b45c2f7a2e631c417f71caa75e927367520d4d9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\hat{V}(t)" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": " denotes the corresponding frame of inpainted video at timestep " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": " denotes the inpainted background with the foreground object removed, " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "o_i^0" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": "-th object at timestep 0, " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "b_i^t" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": "-th bounding box at timestep " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "inline_equation", + "content": "rs" + }, + { + "bbox": [ + 64, + 407, + 296, + 466 + ], + "type": "text", + "content": " denotes resize function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 468, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 468, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 64, + 468, + 296, + 584 + ], + "type": "text", + "content": "Structured Noise from Draft Video. Optical flow is an effective representation for guiding VDMs [12, 13]. Follow prior work [7, 8], we employ RAFT[44] to extract optical flow from the draft video and formulate it as structural noise, which retains Gaussian properties. Given the draft video " + }, + { + "bbox": [ + 64, + 468, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\hat{V}(t) \\in \\mathbb{R}^{F \\times C \\times H \\times W}" + }, + { + "bbox": [ + 64, + 468, + 296, + 584 + ], + "type": "text", + "content": ", we calculate its per-frame optical flow to get a structured noise tensor " + }, + { + "bbox": [ + 64, + 468, + 296, + 584 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{F \\times C \\times H \\times W}" + }, + { + "bbox": [ + 64, + 468, + 296, + 584 + ], + "type": "text", + "content": ". The structured noise enables the VDM to generate videos that exhibit motion patterns closely aligned with those in the optical flow, thereby improving the realism of the output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "spans": [ + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "text", + "content": "Noise Injection in Video Synthesis. We adopt Go-with-the-Flow [7] as our video synthesis model, a fine-tuned CogVideoX [60], which is designed to accept structured noise " + }, + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "text", + "content": " as input and synthesize videos that adhere to the implicit optical flow. The vanilla Go-with-the-Flow tends to tightly follow the provided structured noise " + }, + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "text", + "content": ". However, our " + }, + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 64, + 586, + 296, + 702 + ], + "type": "text", + "content": " is derived from a coarse-level motion trajectory and may not be sufficiently accurate to follow the physical laws of the real world. To address this limitation, we inject noise during the inference phase to give more flexibility to the VDM to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 84, + 468, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 468, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 468, + 96 + ], + "type": "text", + "content": "generate detail-level motion changes as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 386, + 105, + 545, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 105, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 386, + 105, + 545, + 133 + ], + "type": "interline_equation", + "content": "Q _ {i} = \\frac {(1 - \\gamma) Q _ {i} + \\zeta \\gamma}{\\sqrt {(1 - \\gamma) ^ {2} + \\gamma^ {2}}} \\tag {2}", + "image_path": "09108298cbe88b6a400ead2656d2ec502a8f3b7f821695ef2e1cc498113f4438.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "content": " is structured noise at " + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "content": "-th frames, " + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\zeta \\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "content": " is Gaussian noise and " + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0,1]" + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\gamma = 0.4" + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "content": " for even frame index and " + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\gamma = 0.6" + }, + { + "bbox": [ + 313, + 140, + 545, + 176 + ], + "type": "text", + "content": " for odd frame index." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 177, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 545, + 212 + ], + "type": "text", + "content": "With this approach, the VDM is able to generate motion deviate from the coarse-level motion trajectory whenever necessary for producing high-quality fine-level motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 222, + 499, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 499, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 499, + 236 + ], + "type": "text", + "content": "4. Empirical Analysis and Dissusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 545, + 312 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to demonstrate the effectiveness of our video generation framework compared to existing methods. We evaluate our approach on two established benchmarks for physically plausible video generation. Our framework consistently achieves superior performance across all benchmarks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 443, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 443, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 443, + 331 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 545, + 417 + ], + "type": "text", + "content": "We propose a two-stage physically plausible image-to-video generation framework. In the first stage, we utilize ChatGPT-4o as the coarse-level motion planner. In the second stage, we utilize an open-source I2V model, Go-with-the-Flow [7], as a fine-level motion synthesizer. Unless otherwise specified, in all experiments, we generate each video with a resolution of " + }, + { + "bbox": [ + 313, + 336, + 545, + 417 + ], + "type": "inline_equation", + "content": "720 \\times 480" + }, + { + "bbox": [ + 313, + 336, + 545, + 417 + ], + "type": "text", + "content": " and 49 frames." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 425, + 449, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 449, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 449, + 437 + ], + "type": "text", + "content": "4.2. Benchmarks and Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 441, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 545, + 557 + ], + "type": "text", + "content": "Traditional metrics in the visual domain, such as the Peak Signal-to-Noise Ratio (PSNR)[17], the Structural Similarity Index (SSIM)[49], the Learned Perceptual Image Patch Similarity (LPIPS)[63], the Fréchet Inception Distance (FID)[16] and the Fréchet Video Distance (FVD)[45], do not account for the physical realism of the generated videos [31, 33]. Recent studies have begun to address this limitation by developing benchmarks and metrics that evaluate physical realism. In this work, we adopt two benchmarks, described below." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 560, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 545, + 675 + ], + "type": "text", + "content": "PhyGenBench [31] categorizes physical properties into four domains: mechanics, optics, thermal, and material. It includes 27 physical phenomena, each governed by real world physical laws, reflected in 160 carefully designed text prompts. As PhyGenbench provides only text prompts, we adapt it to our image-to-video setting by generating a corresponding first frame for each prompt with FLUX[22]. We adhere to the predefined benchmark evaluation protocol, i.e., employing GPT-4o to assess the physical realism of the generated videos." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 678, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 545, + 702 + ], + "type": "text", + "content": "Physics-IQ [33] comprises 396 real-world videos spanning 66 distinct physical scenarios. For each scenario, videos" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "text", + "content": "12364" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 82, + 545, + 294 + ], + "blocks": [ + { + "bbox": [ + 67, + 82, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 67, + 82, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 67, + 82, + 545, + 294 + ], + "type": "image", + "image_path": "7492de47fea042983fdf07f7c6ba6374f76eb600e81b4090eecd926e22a1d5aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 294, + 544, + 316 + ], + "lines": [ + { + "bbox": [ + 64, + 294, + 544, + 316 + ], + "spans": [ + { + "bbox": [ + 64, + 294, + 544, + 316 + ], + "type": "text", + "content": "Figure 4. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B [60], LTX-VideoI2V [14] and SVD-XT [5]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 96, + 319, + 515, + 468 + ], + "blocks": [ + { + "bbox": [ + 96, + 319, + 515, + 468 + ], + "lines": [ + { + "bbox": [ + 96, + 319, + 515, + 468 + ], + "spans": [ + { + "bbox": [ + 96, + 319, + 515, + 468 + ], + "type": "table", + "html": "
ModelMechanics(↑)Optics(↑)Thermal(↑)Material(↑)Average(↑)
CogvideoX-T2V-5B0.430.550.400.420.45
LTX-Video-T2V0.350.450.360.380.39
OpenSora0.430.500.440.370.44
PhyT2V0.490.610.490.470.52
LLM-Grounding Video Diffusion0.320.410.260.240.31
CogvideoX-I2V-5B0.480.690.430.410.52
SVD-XT0.460.680.480.410.52
LTX-Video-I2V0.470.650.460.370.50
SG-I2V0.520.690.510.390.54
Ours0.550.710.600.530.60
", + "image_path": "c14c0f5f913e5493008d72648b9669ee1dde7110d19aeb9f6df5935128e26070.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 204, + 468, + 405, + 479 + ], + "lines": [ + { + "bbox": [ + 204, + 468, + 405, + 479 + ], + "spans": [ + { + "bbox": [ + 204, + 468, + 405, + 479 + ], + "type": "text", + "content": "Table 1. Quantitative results of VDMs on PhyGenBench." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 64, + 492, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 492, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 64, + 492, + 296, + 585 + ], + "type": "text", + "content": "are recorded from three different perspectives and filmed twice under identical conditions to eliminate randomness. This benchmark evaluates real-world physical phenomena, including collisions, object continuity, occlusion, object permanence, and fluid dynamics. This benchmark assesses physical realism from semantic and temporal perspectives, using semantic metrics and visual metrics to compare generated videos against the real-world reference videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 597, + 296, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 597, + 296, + 702 + ], + "spans": [ + { + "bbox": [ + 64, + 597, + 296, + 702 + ], + "type": "text", + "content": "Compared Models. In the context of text-to-video generation, we compare our framework with CogVideoX-T2V-5B [60], LTX-Video-T2V [14], and OpenSora [66]. Moreover, we evaluate our framework against PhyT2V [58], which enhances physical realism by iteratively refining the prompt. For the image-to-video generation scenario, our framework is evaluated alongside CogVideoX-I2V-5B [60], SVD-XT [5], and LTX-Video-I2V [14]. Additionally, we conducted experiments in the motion-controllable setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 492, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 545, + 573 + ], + "type": "text", + "content": "In this setting, we leverage the motion trajectory predicted by the VLM as a condition to guide VDM generation. We benchmark our approach against image-to-video motion controllable model, SG-I2V [35] and text-to-video motion controllable model LLM-grounded Video Diffusion Models [25]. The experimental details are presented in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 581, + 444, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 444, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 444, + 593 + ], + "type": "text", + "content": "4.3.Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 597, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 545, + 689 + ], + "type": "text", + "content": "We begin with an empirical study on PhyGenBench and Physics-IQ, comparing our framework against widely adopted open-source models in the research community. Based on different physical properties, we categorize the benchmark samples accordingly. Additionally, we classify VDMs into text-to-video (T2V) diffusion models and imaged-to-video (I2V) diffusion models based on the input conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 690, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 690, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 324, + 690, + 545, + 702 + ], + "type": "text", + "content": "In Table 1, we present our experimental results on Phy" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "text", + "content": "12365" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 82, + 515, + 162 + ], + "blocks": [ + { + "bbox": [ + 95, + 82, + 515, + 162 + ], + "lines": [ + { + "bbox": [ + 95, + 82, + 515, + 162 + ], + "spans": [ + { + "bbox": [ + 95, + 82, + 515, + 162 + ], + "type": "table", + "html": "
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Cogvideo-I2V-5B30.429.816.713.38.527.1
SVD-XT21.920.56.88.417.119.1
LTX-Video-I2V30.229.815.913.28.426.8
SG-I2V34.631.215.913.18.429.7
Ours42.334.116.913.48.834.6
", + "image_path": "edcc135b74a423269c0a23d3e096c8c5b7e69d7bffa9ce6100cf6867aeb15014.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 163, + 544, + 185 + ], + "lines": [ + { + "bbox": [ + 64, + 163, + 544, + 185 + ], + "spans": [ + { + "bbox": [ + 64, + 163, + 544, + 185 + ], + "type": "text", + "content": "Table 2. Quantitative results of physically plausible video generation on Physics-IQ Benchmark. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "text", + "content": "GenBench, evaluating different video generation models following its evaluation protocol. The results show that our framework achieves state-of-the-art performance across four different physical phenomena. Our framework outperforms the best T2V method by an average of " + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "text", + "content": " and the best I2V method by " + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "text", + "content": ". Specifically, our framework demonstrates significant advantages in the Mechanics, Thermal, and Material domains, outperforming the best I2V method by " + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "inline_equation", + "content": "17.6\\%" + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "inline_equation", + "content": "35.8\\%" + }, + { + "bbox": [ + 64, + 198, + 296, + 361 + ], + "type": "text", + "content": ", respectively. These advantages are particularly evident in these three types of physical phenomena, which involve more substantial changes in motion, volume, or shape. Our framework is better equipped to understand and reason about bounding box sequences to represent these changes effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 368, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 368, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 64, + 368, + 296, + 471 + ], + "type": "text", + "content": "Similarly, for the Physics-IQ benchmark, we evaluate the performance of different video generation models following its evaluation protocol. Our framework achieves the best results across four different physical phenomena, with improvements of " + }, + { + "bbox": [ + 64, + 368, + 296, + 471 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 64, + 368, + 296, + 471 + ], + "type": "text", + "content": " in Solid Mechanics and " + }, + { + "bbox": [ + 64, + 368, + 296, + 471 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 64, + 368, + 296, + 471 + ], + "type": "text", + "content": " in Fluid Dynamics compared to the second-best models. These significant improvements demonstrate the effectiveness of our framework in generating physically plausible videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 504, + 188, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 504, + 188, + 516 + ], + "spans": [ + { + "bbox": [ + 64, + 504, + 188, + 516 + ], + "type": "text", + "content": "4.4. Qualitative Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 529, + 296, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 529, + 296, + 702 + ], + "spans": [ + { + "bbox": [ + 64, + 529, + 296, + 702 + ], + "type": "text", + "content": "Figures 4 and 5 demonstrate a qualitative comparison between our video generation framework and baseline methods. Among all evaluated approaches, our framework consistently produces videos with the highest degree of physical realism. In the ball falling sample in Figure 4, while CogVideoX shows a bouncing effect, artifacts are present in the video; LTX-Video and SVD-XT exhibit motions that do not adhere to the laws of physics. In Figure 5, we analyze two examples from Physics-IQ. In the pouring water example, the baseline methods fail to show the simultaneous decrease in water level of the glass beverage dispenser and the increase in water level of the glass below; in the ball collision example, none of the baseline methods correctly depict the collision of balls. More videos are provided in the supplement." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 198, + 544, + 447 + ], + "blocks": [ + { + "bbox": [ + 316, + 198, + 544, + 447 + ], + "lines": [ + { + "bbox": [ + 316, + 198, + 544, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 544, + 447 + ], + "type": "image", + "image_path": "26ec347cb5f5715b4193d855f46b0440b39bab2e8a47858a4a354fa123ceb8e9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 544, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 544, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 544, + 479 + ], + "type": "text", + "content": "Figure 5. Visual comparisons of physically plausible video generation results from our framework, CogVideoX-I2V-5B, LTX-Video-I2V, SVD-XT and SG-I2V [35] in the Physics-IQ dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 497, + 404, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 404, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 404, + 509 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 515, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 545, + 561 + ], + "type": "text", + "content": "We perform an ablation study to evaluate the contributions of key components in our framework. We design four variants to analyze the effectiveness of different components in our framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 544, + 702 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 563, + 544, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 563, + 544, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 563, + 544, + 609 + ], + "type": "text", + "content": "1. Ours w/o VLM Planner: To assess the overall functionality of our framework, we replace the structured noise input of the VDM with random noise to evaluate the effectiveness of the VLM planner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 609, + 544, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 544, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 544, + 632 + ], + "type": "text", + "content": "2. Ours w/o CI: Keeping the overall structure unchanged, we remove the in-context information from the VLM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 632, + 544, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 544, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 544, + 666 + ], + "type": "text", + "content": "3. Ours w/o CoT: Similarly, while keeping other components unchanged, we remove the CoT reasoning process from the VLM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 667, + 544, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 667, + 544, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 667, + 544, + 702 + ], + "type": "text", + "content": "4. Ours w/o CC: Lastly, we remove both the in-context information and the CoT reasoning process from the VLM planner while maintaining all other components." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "text", + "content": "12366" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 82, + 517, + 162 + ], + "blocks": [ + { + "bbox": [ + 94, + 82, + 517, + 162 + ], + "lines": [ + { + "bbox": [ + 94, + 82, + 517, + 162 + ], + "spans": [ + { + "bbox": [ + 94, + 82, + 517, + 162 + ], + "type": "table", + "html": "
ModelS.M.(↑)F.D.(↑)Optics(↑)Magnetism(↑)Thermodynamics(↑)Average(↑)
Ours42.334.116.913.48.834.9
w/o VLM Planner.16.320.813.45.85.616.2
w/o C.I26.328.116.911.28.424.3
w/o CoT21.426.916.18.66.921.0
w/o C.C18.722.414.97.26.118.1
", + "image_path": "ad95e016cd723c327e68d07e31aea771baf4ab014be5ea053fff59f4733f6f4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 163, + 531, + 175 + ], + "lines": [ + { + "bbox": [ + 77, + 163, + 531, + 175 + ], + "spans": [ + { + "bbox": [ + 77, + 163, + 531, + 175 + ], + "type": "text", + "content": "Table 3. Ablation study on VLM, in-context learning and COT. S.M. refers to Solid Mechanics, and F.D. refers to Fluid Dynamics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 66, + 186, + 296, + 317 + ], + "blocks": [ + { + "bbox": [ + 66, + 186, + 296, + 317 + ], + "lines": [ + { + "bbox": [ + 66, + 186, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 66, + 186, + 296, + 317 + ], + "type": "image", + "image_path": "4d28e9559c9aff9e594c3ff64214b2436dbcac8807ae8729f0dd5f9a6b9bb5e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 318, + 295, + 348 + ], + "lines": [ + { + "bbox": [ + 64, + 318, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 64, + 318, + 295, + 348 + ], + "type": "text", + "content": "Figure 6. Visualization of generated video with bounding boxes. Coarse motion cue can guide VDM to synthesis physically plausible video." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 66, + 369, + 296, + 501 + ], + "blocks": [ + { + "bbox": [ + 66, + 369, + 296, + 501 + ], + "lines": [ + { + "bbox": [ + 66, + 369, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 66, + 369, + 296, + 501 + ], + "type": "image", + "image_path": "4a631e4ead90c7ffcdaa6d113c0af96248a053240a719a78fe6e70baa243ac2e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 502, + 295, + 533 + ], + "lines": [ + { + "bbox": [ + 64, + 502, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 64, + 502, + 295, + 533 + ], + "type": "text", + "content": "Figure 7. Limitation of VLIPP. Our framework might produces physically implausible videos because it lacks 3D spatial perception and the optical-flow of small objects is prone to noise." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 563, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 563, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 64, + 563, + 296, + 700 + ], + "type": "text", + "content": "Table 3 presents a quantitative comparison between our full method and these variants. Among all variants, Ours w/o VLM Planner shows the most significant performance drop, as removing the planner completely eliminates our ability in understanding the physical laws, leading to nearly random results. Notably, Ours w/o CoT exhibits a more pronounced decline compared to Ours w/o CI, indicating that the reasoning process in CoT enhances the understanding of physics. While in-context information contributes to the physical reasoning ability of VLM, compared to CoT it is less effective in preventing errors caused by VLM hallucination." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 187, + 388, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 388, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 388, + 198 + ], + "type": "text", + "content": "4.6. Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 208, + 544, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 544, + 232 + ], + "type": "text", + "content": "Although VLIPP can generate physically plausible videos, its performance remains constrained by the base model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 235, + 544, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 235, + 544, + 419 + ], + "spans": [ + { + "bbox": [ + 312, + 235, + 544, + 419 + ], + "type": "text", + "content": "Firstly, we are unable to model physical events that cannot be represented by image space bounding box trajectories. For example, phenomena that involve intrinsic state changes of objects such as gas solidification. We adopt 2D bounding boxes because they can be naturally integrated with VDMs and are effective in enhancing the physical plausibility to a certain extent. In fact, they are more flexible than they might appear. As shown in Figure 6, the VLM predicts coarse 2D bounding box trajectories(only four fragments), yet the VDM still produces a realistic shattering pattern without being rigidly constrained by those paths. Moreover, in the butter melting example in Figure 6, the VLM offers only a coarse cue of volumetric change, but the VDM successfully generates the transition from solid to liquid. These abilities significantly improve our flexibility in modeling various physical phenomena." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 422, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 545, + 458 + ], + "type": "text", + "content": "Secondly, our pipeline lacks 3D spatial perception. It is unable to understand the spatial relationships within the scene, as shown in Figure 7 domino example." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 545, + 541 + ], + "type": "text", + "content": "Finally, the optical flow of small objects is prone to noise interference. This will cause our framework to generate ambiguous content, as shown in Figure 7 small ball collision example. With the recent progress in video generation model, we anticipate that our framework will be further improved in generating videos under more challenging physical conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 563, + 384, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 384, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 384, + 575 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 586, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 586, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 312, + 586, + 545, + 702 + ], + "type": "text", + "content": "Recently, VDMs have achieved great empirical success and are receiving considerable attention in computer vision and computer graphics. However, due to the lack of understanding of physical laws, VDMs are unable to generate physically plausible videos. In this paper, we introduce VLIPP, a novel two-stage physically plausible video generation framework that incorporates physical laws into video diffusion models through vision and language informed physical prior. Our experimental results demonstrate the effectiveness of our method compared to existing approaches." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 317, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 317, + 752 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 317, + 752 + ], + "type": "text", + "content": "12367" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 65, + 83, + 163, + 97 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 83, + 163, + 97 + ], + "spans": [ + { + "bbox": [ + 65, + 83, + 163, + 97 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 102, + 297, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 297, + 138 + ], + "type": "text", + "content": "This work was partially supported by the National Natural Science Foundation of China under Grant No. 62441231, 62472065, and U23B2010." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 149, + 122, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 149, + 122, + 162 + ], + "spans": [ + { + "bbox": [ + 66, + 149, + 122, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 168, + 296, + 702 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 70, + 168, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 168, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 70, + 168, + 296, + 212 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floreccia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 213, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 213, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 70, + 213, + 296, + 266 + ], + "type": "text", + "content": "[2] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 267, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 267, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 70, + 267, + 296, + 319 + ], + "type": "text", + "content": "[3] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, KaiWei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 321, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 321, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 70, + 321, + 296, + 353 + ], + "type": "text", + "content": "[4] Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. arXiv preprint arXiv:2412.03572, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 354, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 354, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 70, + 354, + 296, + 407 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 409, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 409, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 70, + 409, + 296, + 452 + ], + "type": "text", + "content": "[6] Sébastien Bubeck, Varun Chadrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 453, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 453, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 70, + 453, + 296, + 506 + ], + "type": "text", + "content": "[7] Ryan Burgert, Yuancheng Xu, Wenqi Xian, Oliver Pilarski, Pascal Clausen, Mingming He, Li Ma, Yitong Deng, Lingxiao Li, Mohsen Mousavi, et al. Go-with-the-flow: Motion-controllable video diffusion models using real-time warped noise. arXiv preprint arXiv:2501.08331, 2025. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 507, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 507, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 70, + 507, + 296, + 559 + ], + "type": "text", + "content": "[8] Pascal Chang, Jingwei Tang, Markus Gross, and Vinicius C Azevedo. How i warped your noise: a temporally-correlated noise prior for diffusion models. In The Twelfth International Conference on Learning Representations. OpenReview, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 561, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 561, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 70, + 561, + 296, + 624 + ], + "type": "text", + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 625, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 625, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 66, + 625, + 296, + 679 + ], + "type": "text", + "content": "[10] Boyang Deng, Richard Tucker, Zhengqi Li, Leonidas Guibas, Noah Snavely, and Gordon Wetzstein. Streetscapes: Large-scale consistent street view generation using autoregressive video diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 680, + 296, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 680, + 296, + 702 + ], + "spans": [ + { + "bbox": [ + 66, + 680, + 296, + 702 + ], + "type": "text", + "content": "[11] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 85, + 545, + 702 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 85, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 85, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 85, + 545, + 106 + ], + "type": "text", + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 545, + 140 + ], + "type": "text", + "content": "[12] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. arXiv preprint arXiv:2401.18085, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 141, + 545, + 195 + ], + "type": "text", + "content": "[13] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 197, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 197, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 315, + 197, + 545, + 248 + ], + "type": "text", + "content": "[14] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, et al. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024.6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 251, + 545, + 293 + ], + "type": "text", + "content": "[15] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractr: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 295, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 295, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 315, + 295, + 545, + 346 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 545, + 382 + ], + "type": "text", + "content": "[17] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 384, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 384, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 315, + 384, + 545, + 425 + ], + "type": "text", + "content": "[18] Hao-Yu Hsu, Zhi-Hao Lin, Albert Zhai, Hongchi Xia, and Shenlong Wang. Autovfx: Physically realistic video editing from natural language instructions. arXiv preprint arXiv:2411.02394, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 428, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 428, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 315, + 428, + 545, + 470 + ], + "type": "text", + "content": "[19] Hsin-Ping Huang, Yu-Chuan Su, Deqing Sun, Lu Jiang, Xuhui Jia, Yukun Zhu, and Ming-Hsuan Yang. Fine-grained controllable video generation via object appearance and context. arXiv preprint arXiv:2312.02919, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 472, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 472, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 315, + 472, + 545, + 535 + ], + "type": "text", + "content": "[20] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2):1-55, 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 315, + 537, + 545, + 580 + ], + "type": "text", + "content": "[21] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 582, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 582, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 582, + 545, + 602 + ], + "type": "text", + "content": "[22] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 605, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 605, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 315, + 605, + 545, + 656 + ], + "type": "text", + "content": "[23] Pengxiang Li, Kai Chen, Zhili Liu, Ruiyuan Gao, Lanqing Hong, Guo Zhou, Hua Yao, Dit-Yan Yeung, Huchuan Lu, and Xu Jia. Trackdiffusion: Tracklet-conditioned video generation via diffusion models. arXiv preprint arXiv:2312.00651, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 659, + 545, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 659, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 315, + 659, + 545, + 702 + ], + "type": "text", + "content": "[24] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "text", + "content": "12368" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 85, + 296, + 700 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 85, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 85, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 85, + 296, + 116 + ], + "type": "text", + "content": "[25] Long Lian, Baifeng Shi, Adam Yala, Trevor Darrell, and Boyi Li. Llm-grounded video diffusion models. arXiv preprint arXiv:2309.17444, 2023. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 118, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 118, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 66, + 118, + 295, + 158 + ], + "type": "text", + "content": "[26] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 159, + 294, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 294, + 180 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 182, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 182, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 182, + 295, + 213 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llavanext: Improved reasoning,OCR, and world knowledge, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 213, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 213, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 66, + 213, + 295, + 255 + ], + "type": "text", + "content": "[29] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2024. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 256, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 256, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 66, + 256, + 295, + 297 + ], + "type": "text", + "content": "[30] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 298, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 298, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 66, + 298, + 295, + 350 + ], + "type": "text", + "content": "[31] Fanqing Meng, Jiaqi Liao, Xinyu Tan, Wenqi Shao, Quanfeng Lu, Kaipeng Zhang, Yu Cheng, Dianqi Li, Yu Qiao, and Ping Luo. Towards world simulator: Crafting physical commonsense-based benchmark for video generation. arXiv preprint arXiv:2410.05363, 2024. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 352, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 352, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 66, + 352, + 295, + 402 + ], + "type": "text", + "content": "[32] Antonio Montanaro, Luca Savant Aira, Emanuele Aiello, Diego Valsesia, and Enrico Magli. Motioncraft: Physics-based zero-shot video generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 404, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 404, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 66, + 404, + 295, + 445 + ], + "type": "text", + "content": "[33] Saman Motamed, Laura Culp, Kevin Swersky, Priyank Jaini, and Robert Geirhos. Do generative video models learn physical principles from watching videos? arXiv preprint arXiv:2501.09038, 2025. 2, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 447, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 447, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 66, + 447, + 295, + 488 + ], + "type": "text", + "content": "[34] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. Advances in Neural Information Processing Systems, 37:18481-18505, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 490, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 490, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 66, + 490, + 295, + 531 + ], + "type": "text", + "content": "[35] Koichi Namekata, Sherwin Bahmani, Ziyi Wu, Yash Kant, Igor Gilitschenski, and David B Lindell. Sg-i2v: Self-guided trajectory control in image-to-video generation. arXiv preprint arXiv:2411.04989, 2024. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 533, + 295, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 533, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 66, + 533, + 295, + 573 + ], + "type": "text", + "content": "[36] Andrew Nealen, Matthias Müller, Richard Keiser, Eddy Boxerman, and Mark Carlson. Physically based deformable models in computer graphics. In Computer graphics forum, pages 809-836. Wiley Online Library, 2006. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 574, + 295, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 574, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 66, + 574, + 295, + 627 + ], + "type": "text", + "content": "[37] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. In European Conference on Computer Vision, pages 111-128. Springer, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 66, + 628, + 282, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 628, + 282, + 638 + ], + "spans": [ + { + "bbox": [ + 66, + 628, + 282, + 638 + ], + "type": "text", + "content": "[38] Openai. https://openai.com/index/hello-gpt-4o/, 2024. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 66, + 639, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 639, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 66, + 639, + 295, + 700 + ], + "type": "text", + "content": "[39] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G Allievi, Senem Velipasalar, and Liu Ren. Vlp: Vision language planning for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14760-14769, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 85, + 545, + 701 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 85, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 85, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 315, + 85, + 545, + 126 + ], + "type": "text", + "content": "[40] Haonan Qiu, Zhaoxi Chen, Zhouxia Wang, Yingqing He, Menghan Xia, and Ziwei Liu. Freetraj: Tuning-free trajectory control in video diffusion models. arXiv preprint arXiv:2406.16863, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 545, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 545, + 180 + ], + "type": "text", + "content": "[41] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 182, + 545, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 182, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 545, + 224 + ], + "type": "text", + "content": "[42] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4): 1-10, 2013. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 226, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 226, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 226, + 545, + 267 + ], + "type": "text", + "content": "[43] Weixiang Sun, Xiaocao You, Ruizhe Zheng, Zhengqing Yuan, Xiang Li, Lifang He, Quanzheng Li, and Lichao Sun. Bora: Biomedical generalist video generation model. arXiv preprint arXiv:2407.08944, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 270, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 545, + 320 + ], + "type": "text", + "content": "[44] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 323, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 323, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 545, + 354 + ], + "type": "text", + "content": "[45] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 357, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 357, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 357, + 545, + 398 + ], + "type": "text", + "content": "[46] Jiawei Wang, Yuchen Zhang, Jiaxin Zou, Yan Zeng, Guoqiang Wei, Liping Yuan, and Hang Li. Boximator: Generating rich and controllable motions for video synthesis. arXiv preprint arXiv:2402.01566, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 545, + 430 + ], + "type": "text", + "content": "[47] Xi Wang, Robin Courant, Marc Christie, and Vicky Kalogeiton. Akira: Augmentation kit on rays for optical video generation. arXiv preprint arXiv:2412.14158, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 432, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 545, + 483 + ], + "type": "text", + "content": "[48] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 486, + 545, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 545, + 527 + ], + "type": "text", + "content": "[49] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 529, + 545, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 529, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 316, + 529, + 545, + 582 + ], + "type": "text", + "content": "[50] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 583, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 583, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 583, + 545, + 635 + ], + "type": "text", + "content": "[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35: 24824-24837, 2022. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 637, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 637, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 637, + 545, + 678 + ], + "type": "text", + "content": "[52] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. arXiv preprint arXiv:2406.17758, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 545, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 545, + 701 + ], + "type": "text", + "content": "[53] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting ltm-controlled diffusion" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 317, + 751 + ], + "type": "text", + "content": "12369" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 85, + 296, + 700 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 85, + 85, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 85, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 85, + 85, + 296, + 116 + ], + "type": "text", + "content": "models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6327-6336, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 118, + 296, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 118, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 66, + 118, + 296, + 170 + ], + "type": "text", + "content": "[54] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 171, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 171, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 66, + 171, + 296, + 223 + ], + "type": "text", + "content": "[55] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 224, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 224, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 66, + 224, + 296, + 276 + ], + "type": "text", + "content": "[56] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 277, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 277, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 66, + 277, + 296, + 308 + ], + "type": "text", + "content": "[57] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 308, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 308, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 66, + 308, + 296, + 350 + ], + "type": "text", + "content": "[58] Qiyao Xue, Xiangyu Yin, Boyuan Yang, and Wei Gao. Phyt2v: Llm-guided iterative self-refinement for physics-grounded text-to-video generation. arXiv preprint arXiv:2412.00596, 2024. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 351, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 351, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 66, + 351, + 296, + 393 + ], + "type": "text", + "content": "[59] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 1(2):6, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 394, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 394, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 66, + 394, + 296, + 445 + ], + "type": "text", + "content": "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 447, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 447, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 66, + 447, + 296, + 498 + ], + "type": "text", + "content": "[61] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 500, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 500, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 66, + 500, + 296, + 542 + ], + "type": "text", + "content": "[62] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 543, + 296, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 543, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 66, + 543, + 296, + 594 + ], + "type": "text", + "content": "[63] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 596, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 596, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 66, + 596, + 296, + 657 + ], + "type": "text", + "content": "[64] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 659, + 296, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 659, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 66, + 659, + 296, + 700 + ], + "type": "text", + "content": "[65] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 85, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 85, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 85, + 545, + 128 + ], + "type": "text", + "content": "[66] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 6" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "spans": [ + { + "bbox": [ + 294, + 742, + 318, + 752 + ], + "type": "text", + "content": "12370" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_content_list.json b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5b73d9371b94770514d6b3f00e430d2a355ad82b --- /dev/null +++ b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_content_list.json @@ -0,0 +1,1500 @@ +[ + { + "type": "text", + "text": "VLM4D: Towards Spatiotemporal Awareness in Vision Language Models", + "text_level": 1, + "bbox": [ + 129, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shijie Zhou $^{1*}$ Alexander Vilesov $^{1*}$ Xuehai He $^{2,3*}$ Ziyu Wan $^{2}$ Shuwang Zhang $^{1}$", + "bbox": [ + 163, + 179, + 821, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aditya Nagachandra $^{1}$ Di Chang $^{4}$ Dongdong Chen $^{2}$ Xin Eric Wang $^{3}$ Achuta Kadambi $^{1}$", + "bbox": [ + 140, + 200, + 854, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UCLA ${ }^{2}$ Microsoft ${ }^{3}$ UCSC ${ }^{4}$ USC", + "bbox": [ + 359, + 224, + 637, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://v1m4d.github.io/", + "bbox": [ + 352, + 247, + 637, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47589e52e181e116a7dc2dff9f94b43df83d71483d046e3a92896847c40c8cd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 272, + 897, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/995244dc3b4f1d01690e042e2a6374a85d43bdee2f33aa3c9bd1f7094c77dd60.jpg", + "image_caption": [ + "Figure 1. Spatiotemporal (4D) Awareness. Humans intuitively reason in 4D (3D space + time), effortlessly reconstructing the dynamic spatial trajectory of moving objects from any perspective. In contrast, current Vision Language Models (VLMs) typically rely on aggregating 2D visual features across time, leading to incorrect predictions when motion understanding and interpretation requires deeper spatiotemporal reasoning. In this example, humans correctly perceive the car moving to the right, while the VLM (GPT-4o) inaccurately predicts leftward movement, suggesting VLMs struggle to perform spatiotemporal reasoning." + ], + "image_footnote": [], + "bbox": [ + 99, + 378, + 897, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 705, + 326, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision language models (VLMs) have shown remarkable capabilities in integrating linguistic and visual reasoning but remain fundamentally limited in understanding dynamic spatiotemporal interactions. Humans effortlessly track and reason about object movements, rotations, and perspective shifts—abilities essential for robust dynamic real-world understanding yet notably lacking in current VLMs. In this paper, we introduce VLM4D, the first benchmark specifically designed to evaluate the spatiotemporal reasoning capa", + "bbox": [ + 89, + 738, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bilities of VLMs. Our benchmark comprises diverse real-world and synthetic videos accompanied by carefully curated question-answer pairs emphasizing translational and rotational motions, perspective awareness, and motion continuity. Through comprehensive evaluations of state-of-the-art open and closed-source VLMs, we identify significant performance gaps compared to human baselines, highlighting fundamental deficiencies in existing models. Extensive analysis reveals that VLMs struggle particularly with integrating multiple visual cues and maintaining temporal coherence. We further explore promising directions, such as leveraging 4D feature field reconstruction and targeted", + "bbox": [ + 511, + 707, + 906, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8600", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spatiotemporal supervised fine-tuning, demonstrating their effectiveness in enhancing spatiotemporal comprehension. Our work aims to encourage deeper exploration into improving VLMs' spatial and temporal grounding, paving the way towards more capable and reliable visual intelligence for dynamic environments.", + "bbox": [ + 88, + 90, + 482, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 210, + 222, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Humans possess an innate ability to perceive, track, and interpret motion as well as spatial and temporal changes, enabling rich interpretations of complex dynamic events from both egocentric and allocentric perspectives [10, 19, 57]. When observing an object move, we can inherently process any changes such as lateral shifts, rotational directions, and periodic or repeated actions unfolding along a specific trajectory [10]. These sophisticated perceptual abilities are a product of our spatiotemporal cognition [26], and form an essential foundation that allows us to comprehend and reason about physical phenomena, object interactions, and causal relationships within our environment [34, 40].", + "bbox": [ + 89, + 236, + 483, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision language models (VLMs), which also have the potential to perceive motion and spatiotemporal changes in videos, represent a prominent class of methods aimed at emulating or surpassing human capabilities in integrated visual and linguistic reasoning [21, 39]. While previous work on VLMs has primarily focused on static visual understanding—through large-scale training on paired language and image data [66]—or explored video understanding tasks such as captioning [55] and scene understanding [12], we find that their strong performance in these tasks does not naturally translate into robust spatiotemporal reasoning. This limitation is particularly striking given that state-of-the-art VLMs are typically trained on datasets comprising hundreds of billions of tokens [51]. In contrast, human infants naturally develop robust spatiotemporal cognition within the first few months of life [68].", + "bbox": [ + 89, + 417, + 483, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another key challenge that limits VLM performance on spatiotemporal tasks is the need to implicitly or explicitly reconstruct a four-dimensional (4D) representation—3D space + time—of dynamic scenes, and subsequently reason over this reconstruction [77]. As illustrated in Fig. 1, the car is advancing forwards and turning to the left in its own frame of reference. However, from the camera's perspective, its motion appears as a combination of heading to the right and receding into the distance despite the car being in the center of the frame due to camera rotation. Human observers can seamlessly disentangle these complex dynamics, accurately interpreting trajectories by synthesizing diverse visual cues including camera rotation compensation, stationary scene landmarks, prior knowledge of 3D and 4D environmental structures, and perspective projections [10, 26, 40, 57]. The inability of current VLMs to in-", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f0aa1aaa38581fffceb31eee1e39292dac7e0c3be047929474dd30272a6c38b8.jpg", + "image_caption": [ + "Figure 2. Distribution of Dataset Sources and Annotations. Overview of the dataset composition, illustrating the proportions of real third-person (exo-centric) videos (DAVIS [64], YouTube-VOS [85]), real first-person (ego-centric) videos (Ego4D [30]), and synthetic videos (Cosmos [4]). The real video data is further categorized by annotation types, including translational, rotational, action, counting, and false positive queries (targeting nonexistent events to assess critical reasoning)." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 893, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tegrate spatial, temporal, and semantic cues in a human-like manner stems from their fundamentally different encoding paradigms for visual and language information. This discrepancy highlights a significant gap between human and machine spatiotemporal understanding, suggesting that future VLMs may benefit from insights in cognitive science and neuroscience to develop more advanced mechanisms for perceiving, integrating, reconstructing, and reasoning over dynamic scenes.", + "bbox": [ + 511, + 406, + 906, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to effectively characterize and challenge the existing spatiotemporal reasoning abilities of VLMs, we introduce VLM4D, a rigorous benchmark specifically designed to probe the spatiotemporal grounding capabilities of current vision language models. Through this contribution, we aim to catalyze research that addresses the critical gap in spatiotemporal understanding and reasoning within VLMs and provide a foundational analysis highlighting key deficiencies in existing models.", + "bbox": [ + 511, + 544, + 908, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows:", + "bbox": [ + 532, + 683, + 870, + 698 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose the first benchmark VLM4D explicitly designed to rigorously evaluate the spatiotemporal (4D) reasoning capabilities of Vision Language Models (VLMs).", + "2. We introduce a novel, meticulously curated dataset consisting of diverse real-world and synthetic video sequences paired with carefully crafted spatiotemporal question-answer (QA) annotations.", + "3. We analyze the critical limitations of contemporary VLMs in spatiotemporal reasoning and conduct experiments exploring potential solutions, highlighting fundamental challenges and outlining clear directions for impactful future research." + ], + "bbox": [ + 513, + 700, + 905, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8601", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17ed82081f03e0cfa98e209ce96212381a6252413b1a8496f303a884553cbe8a.jpg", + "image_caption": [ + "Figure 3. Dataset Generation and Annotation Pipeline. Our dataset was constructed by collecting real videos and generating synthetic data, followed by human-in-the-loop quality reviews to address ambiguous videos and annotations. After temporal alignment and quality assurance, human-annotated questions and ground-truth answers were created, complemented by multiple-choice (MC) answers generated by large language models (LLMs). The final dataset includes real and synthetic video data with comprehensive VLM scoring metrics." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 270, + 230, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatiotemporal Understanding in Vision Language Models Early methods for video understanding before the advent of large vision language models (VLMs) leveraged trajectory-based representations and separate modeling of spatial and motion cues for tasks like action and motion recognition [11, 38, 67, 74]. Recently, VLMs have evolved rapidly by fully leveraging the significant achievements of Large Language Models (LLMs) [7, 9, 20, 65, 71, 80] and large-scale visual instruction tuning datasets [18, 50, 100]. While VLMs [3, 29, 33, 42, 50, 70, 76, 100] exhibit transformative potential for applications such as embodied AI [22, 36, 69], robotics [62, 73], scene generation [49, 60, 98], and world modeling [52, 93], most existing methods remain constrained to static images, focusing narrowly on spatial understanding [25] while overlooking the dynamic temporal dimension inherent in real-world interactions. To bridge this gap, emerging research [16, 44, 56, 91, 92] has begun exploring video modality integration, aiming to equip VLMs with spatial-temporal awareness critical for tasks like video comprehension, where both contextual details and motion dynamics are essential. For example, VideoLLM-MoD [83] proposes to address the efficiency issue when processing long-term video by mixture-of-depths. [88] introduces VideoRefer to enhance the finer-level (like object-level) spatial-temporal video understanding of VLMs. Grounded-VideoLLM [75] also targets for fine-grained video understanding through incorporating an additional temporal stream. In this work, we aim to rigorously evaluate the 4D spatial-temporal reasoning capabilities of state-of-the-art VLMs, probing how and to what extent these models internalize spatial intelligence and temporal dependencies.", + "bbox": [ + 91, + 297, + 483, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLM Benchmarks Following the development trends of VLMs, benchmarking VLMs shares the similar trajectory by first evaluating vision QA on static images [31, 41, 53, 89], to align with models' early focus on 2D understanding. As VLMs evolved to tackle dynamic scenarios, benchmarks expanded to evaluate general-purpose video comprehension", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tasks that probe temporal coherence and event understanding [27, 35, 46, 48, 59]. Notably, MMVU [95] further proposes a knowledge-intensive benchmark to assess the expert-level reasoning ability of current video-based large models. However, while these works assess perception and semantic understanding, they largely overlook the explicit evaluation of spatial-temporal awareness, a core capability for real-world applications requiring 4D (3D space + time) reasoning. Recent efforts like [87] pioneer benchmarks for 3D visual-spatial intelligence but restrict evaluation to static 3D scene, neglecting the interplay of object motion and temporal dynamics intrinsic to videos. In this work, we introduce VLM4D, the first benchmark designed to holistically evaluate the 4D intelligence in VLMs, unifying spatial understanding, temporal continuity, and motion reasoning. By curating tasks that demand precise analysis of dynamic interactions (e.g., direction prediction, perspective anticipation, and motion reasoning), VLM4D exposes critical gaps in current models' ability to internalize spatiotemporal relationships. Our work not only advances the granularity of VLM evaluation but also shares insights and potential solutions to improve the model performance.", + "bbox": [ + 511, + 272, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The VLM4D Benchmark", + "text_level": 1, + "bbox": [ + 511, + 617, + 736, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce VLM4D, the first benchmark specifically designed to test the spatiotemporal reasoning ability of VLMs. VLM4D consists of 1000 videos paired with over 1800 question-answer pairs, each carefully designed to assess both spatial and temporal understanding jointly. The majority of these videos are sourced from datasets with rich spatiotemporal characteristics, thus ensuring a diverse range of motion-related scenarios. We also augment the dataset with synthetic videos generated by a world-foundation model, Cosmos [4], that has been modified using techniques introduced in [32] to obtain more accurate correspondence between motion-oriented prompts and the resulting generated video. Fig. 2 illustrates the composition of our dataset.", + "bbox": [ + 511, + 642, + 908, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Benchmark Construction", + "text_level": 1, + "bbox": [ + 511, + 848, + 745, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike prior work that often relies heavily on LLMs and VLMs to generate first iterations of benchmarks and", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8602", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Is the camel in the foreground turning to the left or right from its own perspective?", + "bbox": [ + 205, + 101, + 679, + 116 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A. Not moving", + "bbox": [ + 205, + 118, + 295, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B.Left", + "bbox": [ + 341, + 118, + 385, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. Moving straight, not turning", + "bbox": [ + 436, + 118, + 616, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "D. Right", + "bbox": [ + 656, + 118, + 707, + 132 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26d1f9f11602467ade67070edee943afe3d6a423a568579116b1b6428c5aefaa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 141, + 820, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In which direction of rotation does the person pour the batter into the frying pan?", + "bbox": [ + 205, + 255, + 671, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A. Counter-clockwise", + "bbox": [ + 207, + 271, + 334, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B. Left to right", + "bbox": [ + 364, + 272, + 452, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. Right to left", + "bbox": [ + 496, + 272, + 584, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "D. Clockwise", + "bbox": [ + 645, + 271, + 725, + 284 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7d1d22256b012a8a1c29c9d1d0fafe9a710fc30405fcffc07bef763ae9f8908d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 292, + 821, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What direction is the robotic dog moving towards?", + "bbox": [ + 207, + 406, + 500, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A. Right", + "bbox": [ + 207, + 422, + 261, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B. no robotic dog there", + "bbox": [ + 313, + 424, + 447, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. Not moving", + "bbox": [ + 496, + 424, + 584, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "D.Left", + "bbox": [ + 635, + 424, + 679, + 436 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/119eaa33d817ce107020693246e0587f027ae636191a9d4d9ef5e3c81143d590.jpg", + "image_caption": [ + "Figure 4. Qualitative Examples of Dataset Annotations. (Top) A third-person (exo-centric) video with translational annotations (\"camel turning left from its perspective\"). (Middle) A first-person (ego-centric) video with a rotational question (\"clockwise rotation of ladle\"). (Bottom) A synthetic scene with motion recognition \"robotic dog moving left\"." + ], + "image_footnote": [], + "bbox": [ + 181, + 445, + 821, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "datasets [14] followed by human quality control, we found that existing VLMs and automated methods showed significant limitations in terms of reliability and quality. This shortcoming necessitated direct human annotations that were then followed by augmentation by LLMs to ensure a high-quality benchmark. An overview of the benchmark curation pipeline is shown in Fig. 3.", + "bbox": [ + 89, + 608, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Real Video Data Collection Real-world videos were sourced from datasets with rich spatiotemporal characteristics that ensured diverse motion and perspective variations. For egocentric data, we relied mainly on the Ego4D dataset [30], while most exocentric data points were collected from the DAVIS [64] and YouTube-VOS [85] datasets. To minimize confounders and to focus attention of VLM abilities to only spatiotemporal reasoning, we preprocessed the videos by temporally segmenting and centering them around the most relevant action, thus resulting in videos with an average duration of 3-8 seconds. This en", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sures that the key event described in the question is clear and reduces ambiguities or confounders that would reduce VLM accuracy.", + "bbox": [ + 511, + 609, + 906, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthetic Video Data Generation The rapid advancement of video generation techniques has led to their widespread application in diverse domains like robotics and immersive entertainment [4, 8, 47, 72, 82], making 4D reasoning on synthetic content essential for VLMs and underscoring the need to include such videos in our benchmark. For synthetic video generation, we use Cosmos [4] as our video generation backbone. To ensure that the generated videos align with the intended object moving directions, we incorporate input bounding boxes as additional spatial guidance. Specifically, we follow the approach introduced in [32] modifying the diffusion forward steps to enforce object localization constraints at each timestep, ensuring consistency between the generated object direction and the user-specified trajectory. The average duration of generated", + "bbox": [ + 511, + 672, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8603", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "synthetic videos is 5 seconds. To maintain high-quality outputs, we perform a manual verification step after generation, filtering out low-quality videos and retaining only those that accurately match the specified directions. Once a video is generated, we use an LLM (GPT-4o) to create two types of evaluation questions: Directional questions, derived from the textual prompt used to generate the video; and False Positive (FP) questions (constituting \\(10\\%\\), matching the ratio in the real dataset), which query non-existent objects in the scene. Both question types follow the format: \"What direction is the \\(\\langle\\) Object Name \\rangle\\) moving?\", where the model must select one of four possible answers: \"left\", \"right\", \"not moving\", or \"no \\(\\langle\\) Object Name \\rangle\\) there\". A final manual review is conducted to filter out or revise ambiguous questions, ensuring the quality of the questions and ground-truth answers.", + "bbox": [ + 89, + 90, + 483, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "QA Generation and Quality Control Question-answer pairs are primarily constructed through human annotations. The question answer pairs are then supplemented with alternative answers by an LLM (GPT-4o) for multiple choice (MC) questions. To ensure high-quality annotations, we applied a three-round, multi-person cross-checked verification process, during which ambiguous videos were filtered out and vague, misleading, or incorrect QA pairs were refined to improve spatial and temporal alignment between language and visual content. Fig. 4 showcases some qualitative examples of annotations for different types of videos.", + "bbox": [ + 89, + 351, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assessing Human Performance To establish a human performance baseline on our benchmark, we conducted an evaluation in which participants independently answered 100 randomly sampled questions from the dataset. The accuracy of human responses was then aggregated to approximate the performance of human spatiotemporal reasoning on the dataset.", + "bbox": [ + 89, + 536, + 483, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Categorizing Spatiotemporal Performance", + "text_level": 1, + "bbox": [ + 89, + 652, + 452, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To systematically evaluate spatiotemporal reasoning capabilities, we first categorize videos into two primary groups: egocentric (first-person) videos and exocentric (third-person) videos. Egocentric videos are sourced from the Ego4D [30] dataset, where scenes are captured from a head-mounted camera, thus offering dynamic video data that is inherently coupled with the individual's actions. Exocentric videos encompass a diverse range of recorded scenes, from sports footage to everyday scenes. Beyond this categorization, we also evaluate spatiotemporal performance across four dimensions: translational movement (TM), rotational movement (RM), spatiotemporal counting (STM), and false positives (FP), with their proportions shown in Fig. 2. Translational movement assesses a model's ability to track linear motion within scenes, while rotational", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76d772259383b40099edee3486392a90b1c092299f408babd1e2bcca631b61d4.jpg", + "image_caption": [ + "Figure 5. Comparison of accuracy across types of spatiotemporal questions. Model accuracy is shown only for the six top-performing VLMs." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 901, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "movement assesses the understanding of changes in orientation and perspective shifts over time. Spatiotemporal counting extends these core motion-based tasks by requiring a more complex reasoning strategy to determine the number of objects performing a translation or rotational movement. Lastly, the false positives category evaluates the model's critical thinking [23] in determining whether an object or event actually occurred within the spatiotemporal context. By structuring the benchmark along these axes, we aim for a comprehensive framework for assessing spatiotemporal reasoning (Fig. 5).", + "bbox": [ + 511, + 426, + 906, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation of VLM4D Benchmark", + "text_level": 1, + "bbox": [ + 511, + 606, + 816, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Setup", + "text_level": 1, + "bbox": [ + 511, + 631, + 684, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Models We evaluate 23 most recently released VLMs thus covering a wide range of model sizes, architectures, and training methodologies. For closed-source VLMs, we evaluate GPT-4o [61], Gemini 2.5 Pro [28], Claude Sonnet 4 [6], and Grok-2-Vision [84]. For open-source models, we include Llama 4 [58], DeepSeek-VL [54], Qwen2.5-VL [86], Qwen2-VL [76], InternVL2.5 [15], Aria [43], InternVideo 2.5 [79], InternVideo2 [78], Phi-4-multimodal [2], Phi-3.5-vision [1], Pixtral [5], VideoLLama3 [91], Llava-One-Vision [42], Llava-NeXT-Video [94]. When available, we evaluate different sizes for each model, resulting in models ranging from 2 to 72 billion parameters.", + "bbox": [ + 511, + 654, + 906, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Settings The evaluations were performed in a zero-shot setting with video or a set of sampled frames of", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8604", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8ccdf57392717b9801d88d92aa93772a0982fe565babcccdf575603b63dece30.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OrganizationModelReleaseRealSyntheticOverall
Ego-centricExo-centricAverageDirectionalFPAverage
User StudyHuman Performance99.699.799.795.810096.298.8
RandomRandom Selection24.423.223.625.524.725.424.1
", + "bbox": [ + 101, + 88, + 903, + 165 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47c94cb9ac9dcb707ee35de3809525e665f345bddf3443567e6da53dcb91be2f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Latest Proprietary VLMs
OpenAIGPT-4o2024-1155.562.260.049.553.349.957.5
GoogleGemini-2.5-Pro2025-664.662.963.554.880.057.362.0
AnthropicClaude-Sonnet-42025-552.652.152.244.086.748.351.3
xAIGrok-2-Vision2024-1248.849.749.449.366.751.049.8
", + "bbox": [ + 101, + 167, + 901, + 252 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f7c942fa6def8f2d53a19dc72218b876a7a11c4a5fcf17523953da5c5ba0cdf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Open-source Image VLMs
MetaLlama-4-Maverick-17B2025-452.654.353.853.351.153.053.6
Llama-4-Scout-17B2025-448.656.253.753.375.655.554.1
MicrosoftPhi-4-Multimodal2025-341.035.437.237.511.134.836.6
Phi-3.5-Vision2024-733.438.837.123.337.824.734.0
DeepSeekDeepSeek-VL22024-1233.632.933.131.846.733.333.2
Shanghai AI LabInternVL2.5-38B2024-1146.650.148.943.357.844.747.9
InternVL2.5-8B2024-1139.044.042.440.842.240.942.0
Mistral AIPixtral-12B2024-932.325.827.924.322.224.027.0
RhymesAria2024-1147.244.045.138.571.141.844.3
", + "bbox": [ + 101, + 255, + 901, + 414 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9935b2beb450c25ba0ff46195b92cba9bfb6f95377aacfd006d26ff3125f178.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Open-source Video VLMs
AlibabaQwen2.5-VL-7B2025-142.343.743.343.564.445.643.8
Qwen2.5-VL-72B2025-154.352.553.149.580.052.653.0
Qwen2-VL-7B2024-836.134.735.240.535.640.036.3
Qwen2-VL-72B2024-948.143.044.640.873.344.044.5
DAMOVideoLLama3-2B2025-153.242.546.034.355.636.443.7
VideoLLama3-7B2025-149.445.146.542.853.343.845.9
Shanghai AI LabInternVideo2.5-8B2025-157.250.552.744.346.744.550.7
InternVideo2-8B2024-835.639.338.143.00.038.738.2
LLaVALLaVA-One-Vision-7B2024-936.835.636.037.835.637.536.3
LLaVA-NeXT-Video-34B2024-629.631.630.924.555.627.630.1
", + "bbox": [ + 101, + 416, + 901, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Evaluation on VLM4D Benchmark across various proprietary and open-source VLMs. Top three performers in each category are highlighted from dark (highest) to light (third highest). Human and random selection baselines are included for reference.", + "bbox": [ + 89, + 603, + 906, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "video, followed by the prompt forming the input. For each model, we evaluate on two different inference settings. In the first setting, the model prompted to directly output (DO) the answer immediately without any reasoning, and in the second evaluation setting, the model is directed to create intermediate reasoning steps, chain-of-thought (CoT) [81], before inferring the final answer.", + "bbox": [ + 89, + 661, + 482, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Following prior work [87] and given the nature of our target task, we adopt multiple-choice questions (MCQs) for evaluation, using accuracy as the primary metric. For the two inference settings described earlier, we employ LLM-as-Judge [95] to assess the outputs of VLMs. We opt for this method instead of string or template matching, as VLMs—especially under chain of thought (CoT) prompt-", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing—often generate all possible answer choices during reasoning, with varying frequencies and slight formatting differences. In some cases, the final answer may even contradict the reasoning. To better evaluate whether the model truly understands the video, we prompt two advanced LLMs (GPT-o3 and o4-mini) to grade based on the full CoT reasoning, not just the final answer. We then perform a cross-check between their judgments and manually resolve any disagreements. The evaluation results are reported in Tab. 1.", + "bbox": [ + 511, + 662, + 906, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmark Results", + "text_level": 1, + "bbox": [ + 511, + 815, + 700, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our comprehensive evaluation on the VLM4D benchmark, detailed in Tab. 1, systematically assesses the spatiotemporal reasoning capabilities of modern Vision-Language Models (VLMs). The results highlight a clear hierar", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "chy, with proprietary models demonstrating superior performance over their open-source counterparts. Google's Gemini-2.5-Pro emerges as the top-performing model with an overall accuracy of $62.0\\%$ , followed by OpenAI's GPT4o at $57.5\\%$ . Within the open-source domain, image-based models like Meta's Llama-4-Scout-17B (54.1%) and videosupported models like Alibaba's Qwen2.5-VL-72B (53.0%) demonstrate highly competitive results, even outperforming some proprietary counterparts. Despite these achievements, a significant performance gap persists when compared to human accuracy (98.8%), underscoring that sophisticated 4D awareness remains a formidable challenge for AI. Notably, performance varies significantly across different categories, such as real versus synthetic data and egocentric versus exo-centric perspectives, indicating that current models lack generalized spatiotemporal understanding.", + "bbox": [ + 89, + 90, + 480, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis: Why VLMs Don't Work Well?", + "text_level": 1, + "bbox": [ + 89, + 348, + 454, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Limited Spatiotemporal Cognition", + "text_level": 1, + "bbox": [ + 89, + 375, + 392, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Despite significant advances in VLMs, their abilities to understand and reason about motion, spatial relationships, and temporal coherence remains fundamentally underdeveloped [13, 101]. Chain of thought (CoT) [81] is widely employed as a method to improve accuracy through step-by-step reasoning. We showcase a comparison between CoT and DO in Fig. 6. Overall, there is no indication of a large advantage of CoT over all evaluated models. Upon deeper exploration of the CoT reasoning of some models, we observe that the reasoning process was primarily flawed in the following ways: irrelevant information and arriving at conclusions that are inconsistent with the reasoning process. Larger models exhibited strategies that would be similar to how a human processes spatiotemporal information, but the resulting execution falls short of human performance. This demonstrates a disconnect between its visual and linguistic knowledge. We provide examples of this behavior in the supplementary material.", + "bbox": [ + 89, + 397, + 482, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Deficiencies in Spatiotemporal Labeling", + "text_level": 1, + "bbox": [ + 89, + 681, + 431, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another avenue of exploration we undertook is to understand the richness of spatiotemporal labels in popular supervised fine-tuning (SFT) VLM datasets. Typically, video captioning occurs at the 'scene' level, lacking fine-grained temporal, spatial, and object-level details. We performed an extensive analysis, encompassing over 2 million samples [14, 17, 44, 45, 94]. We performed this analysis through string-matching of spatiotemporal descriptors related to directionality, translational motion, rotation, and perspective shifts and provide the overall results in Fig. 7. We then performed a manual finegrained evaluation of the ShareGPT4Video dataset [14] which we found had the highest density of spatiotemporal dataset. We found that", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/054262790f82a626f629e2cde3a06ae7d864a6f908397e350eca9a1f9b2152bd.jpg", + "image_caption": [ + "Figure 6. Comparison of CoT and DO Accuracy Across Models. Accuracy comparison between Chain-of-Thought (CoT) and Direct Output (DO) prompting across VLMs." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af46013f16998694edd5a53b3093a61ae9c88455eb6bd649da3c78ece349e127.jpg", + "image_caption": [ + "Figure 7. Heatmap of Occurrences of Spatial-Temporal Terms in popular video SFT datasets." + ], + "image_footnote": [], + "bbox": [ + 521, + 397, + 916, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from a sample of 100 labels that were detected as spatiotemporal, less than $10\\%$ of them were judged as accurate upon human evaluation. This result underscores the inadequacy of current dense captioning approaches, which frequently generate spatiotemporal descriptors without capturing precise motion dynamics. We provide more detailed analysis and explanations in the supplementary material.", + "bbox": [ + 511, + 710, + 906, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Probing Future Solutions", + "text_level": 1, + "bbox": [ + 511, + 830, + 750, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To probe promising future solutions for enhancing spatiotemporal video understanding, we propose two approaches that address some of the shortcomings of current", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8606", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "state-of-the-art VLMs: fine-tuning a VLM on data-rich in spatiotemporal actions and the other leveraging 4D reconstruction and feature fields jointly with a VLM. SFT refines the model's abilities by training on datasets that contain temporally and spatially rich actions and interactions. By integrating structured visual representations and targeted fine-tuning, these approaches enhance video-language models' ability to interpret motion. The second method lifts the feature space of VLMs into a temporally coherent 4D feature field, providing structured scene representations that improve motion and spatial reasoning in the stage of decoding and inference.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spatial-Temporal SFT We evaluate on a subset split of the real dataset by randomly splitting the real-world dataset into a training and testing split (80% / 20%) and we try settings using synthetic/real both for training. We conducted the experiments using Qwen 2VL (7B) and Qwen 2.5VL (7B) through LLama-Factory [96], and compared the performance before and after supervised fine-tuning in Tab. 2. The results demonstrated an improvement in accuracy in spatiotemporal reasoning, suggesting that performance gains can be obtained through targeted training. However, the addition of synthetic data does not necessarily increase performance over using real data alone, suggesting the importance of synthetic data quality.", + "bbox": [ + 89, + 291, + 482, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4D Feature Fields Reconstruction Recent advances in 3D/4D feature fields reconstruction methods [24, 37, 90, 97, 99] have significantly enhanced the vision foundation model's performance in 3D/4D space by integrating structured latent scene representations into the model's inference stage. Inspired by the promising results of feature lifting, we explore enhancing the InternVideo2-8B model [78] with spatiotemporal awareness by adopting the strategy proposed in Feature4X [99], which constructs the VLM's 2D feature space along the time dimension into a 4D feature field. To assess this approach, we evaluate performance on a subset of the VLM4D benchmark, specifically leveraging all 50 videos from the DAVIS 2016 dataset [63]. Our experimental evaluation compares inference performance across three distinct input modalities: original 2D videos; rendered novel global-view videos (which provide broader contextual information in 2D format); and reconstructed global feature fields (which implicitly incorporate 4D scene-level information during reasoning). Table 3 reveals that reconstructed feature fields achieve the highest accuracy across both reasoning types. This success stems from two key advantages: the inherent structure of 4D representations and the ability of feature field inference to avoid the rendering artifacts of RGB reconstruction (global view video). However, the current approach requires per-scene optimization as a post-processing step, limiting its generalizability and", + "bbox": [ + 89, + 508, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ecaf95c4ebc06418cc9bb13401a409f083e848d1e3ee0f7abbf739e257dabf2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMC
Original Model
Qwen 2VL (7B)38.3
Qwen 2.5VL (7B)43.4
Finetuned Model
Qwen 2VL (7B) (R)54.5
Qwen 2VL (7B) (S)42.5
Qwen 2VL (7B) (R+S)55.5
Qwen 2.5VL (7B) (R)55.6
Qwen 2.5VL (7B) (S)48.6
Qwen 2.5VL (7B) (R+S)56.3
", + "bbox": [ + 591, + 88, + 828, + 311 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6364fe09d8b740575d79c946958d6cfc69430d9e5a658ec5e4bda85d64ccd738.jpg", + "table_caption": [ + "Table 2. SFT on Spatial-Temporal Datasets. MC refers to the performance of multiple-choice, while R, S, and $\\mathbb{R} + \\mathbb{S}$ denote real, synthetic, and both real+synthetic usage of data for fine-tuning." + ], + "table_footnote": [], + "table_body": "
Input ModalityAccuracy
Chain of Thought Response
Original 2D Video36.0
Global View Video32.7
Global Feature Field37.4
Direct Output Response
Original 2D Video24.3
Global View Video23.8
Global Feature Field29.0
", + "bbox": [ + 552, + 377, + 867, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. InternVideo2 Accuracy with 4D Feature Field Reconstruction. Comparison of InternVideo2's performance given different input modalities from the same subset.", + "bbox": [ + 511, + 574, + 903, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "making it computationally intensive.", + "bbox": [ + 513, + 635, + 754, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through the construction of the VLM4D benchmark, we evaluate the spatiotemporal reasoning capabilities of various Vision-Language Models (both open-source and proprietary). While more recently released models demonstrate improved performance over their counterparts, they remain significantly behind human proficiency. Overall, our work questions whether VLMs possess spatiotemporal reasoning abilities that are imperative to have for more sophisticated visual agents in fields ranging from robotics to interactive AI systems that require a deep understanding of dynamic visual environments. We hope to inspire future work to explore novel approaches for integrating spatiotemporal grounding, thereby enhancing their spatiotemporal reasoning capabilities and facilitating robust deployment.", + "bbox": [ + 509, + 688, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024.5", + "[2] Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024. 5", + "[3] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Weijian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 3", + "[4] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3, 4", + "[5] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 5", + "[6] Anthropic. System card: Claude opus 4 & claude sonnet 4. Technical report, 2025. 5", + "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3", + "[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 4", + "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al." + ], + "bbox": [ + 106, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[10] Neil Burgess. Spatial memory: how egocentric and allocentric combine. Trends in Cognitive Sciences, 10(12):551-557, 2006. 2", + "[11] Pradyumna Chari, Yunhao Ba, Shijie Zhou, Chinmay Talegaonkar, Shreeram Athreya, and Achuta Kadambi. On learning mechanical laws of motion from video using neural networks. IEEE Access, 11:30129-30145, 2023. 3", + "[12] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18030-18040, 2022. 2", + "[13] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.7", + "[14] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4, 7", + "[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5", + "[16] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024.3", + "[17] Erfei Cui, Yinan He, Zheng Ma, Zhe Chen, Hao Tian, Weiyun Wang, Kunchang Li, Yi Wang, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, Yali Wang, Limin Wang, Yu Qiao, and Jifeng Dai. Sharegpt-4o: Comprehensive multimodal annotations with gpt-4o, 2024. 7", + "[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3", + "[19] Julian De Freitas, Nicholas E. Myers, and Anna C. Nobre. Tracking the changing feature of a moving object. Journal of Vision, 16(3):22, 2016. 2", + "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 3", + "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 522, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8608", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2", + "[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3", + "[23] Chenrui Fan, Ming Li, Lichao Sun, and Tianyi Zhou. Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? arXiv preprint arXiv:2504.06514, 2025. 5", + "[24] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in neural information processing systems, 37:40212-40229, 2024. 8", + "[25] Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Huaizhi Qu, Dilin Wang, Zhicheng Yan, et al. Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction. arXiv preprint arXiv:2505.20279, 2025. 3", + "[26] Jennifer J. Freyd and Ronald A. Finke. Representational momentum. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10(1):126-132, 1984. 2", + "[27] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3", + "[28] Google Gemini Team. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Technical report, 2025. 5", + "[29] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 3", + "[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 4, 5", + "[31] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multidiscipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3", + "[32] Xuehai He, Shuohang Wang, Jianwei Yang, Xiaoxia Wu, Yiping Wang, Kuan Wang, Zheng Zhan, Olatunjri Ruwase, Yelong Shen, and Xin Eric Wang. Mojito: Motion tra" + ], + "bbox": [ + 99, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "jctory and intensity control for video generation. arXiv preprint arXiv: 2412.08948, 2024. 3, 4", + "[33] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3", + "[34] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 2", + "[35] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. How good is my video lmm? complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 3", + "[36] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3", + "[37] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in neural information processing systems, 35:23311-23330, 2022. 8", + "[38] Ivan Laptev. On space-time interest points. International Journal of Computer Vision, 64(2-3):107-123, 2005. 3", + "[39] Yann LeCun, Bernhard Boser, John S. Denker, Donnie Henderson, Richard E. Howard, Wayne Hubbard, and Lawrence D. Jackel. Backpropagation applied to handwritten zip code recognition. *Neural Computation*, 1(4):541-551, 1989. 2", + "[40] Alan M. Leslie. Spatiotemporal continuity and the perception of causality in infants. Perception, 13(3):287-305, 1984. 2", + "[41] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3", + "[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5", + "[43] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 5", + "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 7", + "[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark, 2023. 7", + "[46] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8609", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3", + "[47] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, and Zhiwen Fan. 4k4DGen: Panoramic 4d generation at 4k resolution. In The Thirteenth International Conference on Learning Representations, 2025. 4", + "[48] Xinhao Li, Zhenpeng Huang, Jing Wang, Kunchang Li, and Limin Wang. Videoeval: Comprehensive benchmark suite for low-cost evaluation of video foundation model. arXiv preprint arXiv:2407.06491, 2024. 3", + "[49] Lu Ling, Chen-Hsuan Lin, Tsung-Yi Lin, Yifan Ding, Yu Zeng, Yichen Sheng, Yunhao Ge, Ming-Yu Liu, Aniket Bera, and Zhaoshuo Li. Scenethesis: A language and vision agentic framework for 3d scene generation. arXiv preprint arXiv:2505.02836, 2025. 3", + "[50] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 3", + "[51] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint, 2024. 2", + "[52] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv preprint arXiv:2402.08268, 2024.3", + "[53] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3", + "[54] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5", + "[55] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Neural Information Processing Systems, 2019. 2", + "[56] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3", + "[57] D. Marr and S. Ullman. Directional selectivity and its use in early visual processing. Proceedings of the Royal Society of London. Series B, Biological Sciences, 211(1183):151-180, 1981. 2", + "[58] Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. Technical report, 2025. 5", + "[59] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluat" + ], + "bbox": [ + 99, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3", + "[60] Basak Melis Öcal, Maxim Tatarchenko, Sezer Karaoglu, and Theo Gevers. Sceneteller: Language-to-3d scene generation. In European Conference on Computer Vision, pages 362–378. Springer, 2024. 3", + "[61] OpenAI. Hello gpt-4o. Technical report, 2024. 5", + "[62] Shivansh Patel, Xinchen Yin, Wenlong Huang, Shubham Garg, Hooshang Nayyeri, Li Fei-Fei, Svetlana Lazebnik, and Yunzhu Li. A real-to-sim-to-real approach to robotic manipulation with vlm-generated iterative keypoint rewards. arXiv preprint arXiv:2502.08643, 2025. 3", + "[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 8", + "[64] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 4", + "[65] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3", + "[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2", + "[67] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems (NIPS), pages 568-576, 2014. 3", + "[68] Elizabeth S. Spelke and Katherine D. Kinzler. Core knowledge. Developmental Science, 10(1):89-96, 2007. 2", + "[69] Alessandro Suglia, Claudio Greco, Katie Baker, Jose L Part, Ioannis Papaioannou, Arash Eshghi, Ioannis Konstas, and Oliver Lemon. Alanavlm: A multimodal embodied ai foundation model for egocentric video understanding. arXiv preprint arXiv:2406.13807, 2024. 3", + "[70] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3", + "[71] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[72] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 4" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8610", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 3", + "[74] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Action recognition by dense trajectories. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3169-3176, 2011. 3", + "[75] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3", + "[76] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5", + "[77] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2", + "[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Intervideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 5, 8", + "[79] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 5", + "[80] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 3", + "[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 6, 7", + "[82] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26057-26068, 2025. 4", + "[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2024. 3", + "[84] xAI. Grok-2 beta release. Technical report, 2024. 5", + "[85] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen," + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018. 2, 4", + "[86] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5", + "[87] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. arXiv preprint arXiv:2412.14171, 2024. 3, 6", + "[88] Yuqian Yuan, Hang Zhang, Wentong Li, Zesen Cheng, Boqiang Zhang, Long Li, Xin Li, Deli Zhao, Wenqiao Zhang, Yueting Zhuang, et al. Videorefer suite: Advancing spatial-temporal object understanding with video llm. arXiv preprint arXiv:2501.00599, 2024.3", + "[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3", + "[90] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In European Conference on Computer Vision, pages 57-74. Springer, 2024. 8", + "[91] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 3, 5", + "[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3", + "[93] Hongxin Zhang, Zeyuan Wang, Qiushi Lyu, Zheyuan Zhang, Sunli Chen, Tianmin Shu, Yilun Du, and Chuang Gan. Combo: compositional world models for embodied multi-agent cooperation. arXiv preprint arXiv:2404.10775, 2024. 3", + "[94] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 5, 7", + "[95] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 3, 6", + "[96] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of $100+$ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand, 2024. Association for Computational Linguistics. 8" + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8611", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[97] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 8", + "[98] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2024. 3", + "[99] Shijie Zhou, Hui Ren, Yijia Weng, Shuwang Zhang, Zhen Wang, Dejia Xu, Zhiwen Fan, Suya You, Zhangyang Wang, Leonidas Guibas, et al. Feature4x: Bridging any monocular video to 4d agentic ai with versatile gaussian feature fields. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14179-14190, 2025. 8", + "[100] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3", + "[101] Orr Zohar, Xiaohan Wang, Yann Dubois, Nikhil Mehta, Tong Xiao, Philippe Hansen-Estruch, Licheng Yu, Xiaofang Wang, Felix Juefei-Xu, Ning Zhang, et al. Apollo: An exploration of video understanding in large multimodal models. arXiv preprint arXiv:2412.10360, 2024. 7" + ], + "bbox": [ + 93, + 90, + 480, + 483 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "8612", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_model.json b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_model.json new file mode 100644 index 0000000000000000000000000000000000000000..287aafe242a8df32aafb375d757b78b992aa3675 --- /dev/null +++ b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_model.json @@ -0,0 +1,2624 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.131, + 0.87, + 0.154 + ], + "angle": 0, + "content": "VLM4D: Towards Spatiotemporal Awareness in Vision Language Models" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.18, + 0.822, + 0.2 + ], + "angle": 0, + "content": "Shijie Zhou\\(^{1*}\\) Alexander Vilesov\\(^{1*}\\) Xuehai He\\(^{2,3*}\\) Ziyu Wan\\(^{2}\\) Shuwang Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.202, + 0.856, + 0.22 + ], + "angle": 0, + "content": "Aditya Nagachandra\\(^{1}\\) Di Chang\\(^{4}\\) Dongdong Chen\\(^{2}\\) Xin Eric Wang\\(^{3}\\) Achuta Kadambi\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.226, + 0.638, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)UCLA \\({ }^{2}\\)Microsoft \\({ }^{3}\\)UCSC \\({ }^{4}\\)USC" + }, + { + "type": "text", + "bbox": [ + 0.354, + 0.248, + 0.638, + 0.265 + ], + "angle": 0, + "content": "https://v1m4d.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.273, + 0.898, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.38, + 0.898, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.626, + 0.907, + 0.697 + ], + "angle": 0, + "content": "Figure 1. Spatiotemporal (4D) Awareness. Humans intuitively reason in 4D (3D space + time), effortlessly reconstructing the dynamic spatial trajectory of moving objects from any perspective. In contrast, current Vision Language Models (VLMs) typically rely on aggregating 2D visual features across time, leading to incorrect predictions when motion understanding and interpretation requires deeper spatiotemporal reasoning. In this example, humans correctly perceive the car moving to the right, while the VLM (GPT-4o) inaccurately predicts leftward movement, suggesting VLMs struggle to perform spatiotemporal reasoning." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.707, + 0.327, + 0.723 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Vision language models (VLMs) have shown remarkable capabilities in integrating linguistic and visual reasoning but remain fundamentally limited in understanding dynamic spatiotemporal interactions. Humans effortlessly track and reason about object movements, rotations, and perspective shifts—abilities essential for robust dynamic real-world understanding yet notably lacking in current VLMs. In this paper, we introduce VLM4D, the first benchmark specifically designed to evaluate the spatiotemporal reasoning capa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.907, + 0.89 + ], + "angle": 0, + "content": "bilities of VLMs. Our benchmark comprises diverse real-world and synthetic videos accompanied by carefully curated question-answer pairs emphasizing translational and rotational motions, perspective awareness, and motion continuity. Through comprehensive evaluations of state-of-the-art open and closed-source VLMs, we identify significant performance gaps compared to human baselines, highlighting fundamental deficiencies in existing models. Extensive analysis reveals that VLMs struggle particularly with integrating multiple visual cues and maintaining temporal coherence. We further explore promising directions, such as leveraging 4D feature field reconstruction and targeted" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8600" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.182 + ], + "angle": 0, + "content": "spatiotemporal supervised fine-tuning, demonstrating their effectiveness in enhancing spatiotemporal comprehension. Our work aims to encourage deeper exploration into improving VLMs' spatial and temporal grounding, paving the way towards more capable and reliable visual intelligence for dynamic environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.223, + 0.226 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.484, + 0.417 + ], + "angle": 0, + "content": "Humans possess an innate ability to perceive, track, and interpret motion as well as spatial and temporal changes, enabling rich interpretations of complex dynamic events from both egocentric and allocentric perspectives [10, 19, 57]. When observing an object move, we can inherently process any changes such as lateral shifts, rotational directions, and periodic or repeated actions unfolding along a specific trajectory [10]. These sophisticated perceptual abilities are a product of our spatiotemporal cognition [26], and form an essential foundation that allows us to comprehend and reason about physical phenomena, object interactions, and causal relationships within our environment [34, 40]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Vision language models (VLMs), which also have the potential to perceive motion and spatiotemporal changes in videos, represent a prominent class of methods aimed at emulating or surpassing human capabilities in integrated visual and linguistic reasoning [21, 39]. While previous work on VLMs has primarily focused on static visual understanding—through large-scale training on paired language and image data [66]—or explored video understanding tasks such as captioning [55] and scene understanding [12], we find that their strong performance in these tasks does not naturally translate into robust spatiotemporal reasoning. This limitation is particularly striking given that state-of-the-art VLMs are typically trained on datasets comprising hundreds of billions of tokens [51]. In contrast, human infants naturally develop robust spatiotemporal cognition within the first few months of life [68]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Another key challenge that limits VLM performance on spatiotemporal tasks is the need to implicitly or explicitly reconstruct a four-dimensional (4D) representation—3D space + time—of dynamic scenes, and subsequently reason over this reconstruction [77]. As illustrated in Fig. 1, the car is advancing forwards and turning to the left in its own frame of reference. However, from the camera's perspective, its motion appears as a combination of heading to the right and receding into the distance despite the car being in the center of the frame due to camera rotation. Human observers can seamlessly disentangle these complex dynamics, accurately interpreting trajectories by synthesizing diverse visual cues including camera rotation compensation, stationary scene landmarks, prior knowledge of 3D and 4D environmental structures, and perspective projections [10, 26, 40, 57]. The inability of current VLMs to in-" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.895, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Distribution of Dataset Sources and Annotations. Overview of the dataset composition, illustrating the proportions of real third-person (exo-centric) videos (DAVIS [64], YouTube-VOS [85]), real first-person (ego-centric) videos (Ego4D [30]), and synthetic videos (Cosmos [4]). The real video data is further categorized by annotation types, including translational, rotational, action, counting, and false positive queries (targeting nonexistent events to assess critical reasoning)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.542 + ], + "angle": 0, + "content": "tegrate spatial, temporal, and semantic cues in a human-like manner stems from their fundamentally different encoding paradigms for visual and language information. This discrepancy highlights a significant gap between human and machine spatiotemporal understanding, suggesting that future VLMs may benefit from insights in cognitive science and neuroscience to develop more advanced mechanisms for perceiving, integrating, reconstructing, and reasoning over dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.909, + 0.681 + ], + "angle": 0, + "content": "In order to effectively characterize and challenge the existing spatiotemporal reasoning abilities of VLMs, we introduce VLM4D, a rigorous benchmark specifically designed to probe the spatiotemporal grounding capabilities of current vision language models. Through this contribution, we aim to catalyze research that addresses the critical gap in spatiotemporal understanding and reasoning within VLMs and provide a foundational analysis highlighting key deficiencies in existing models." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.684, + 0.871, + 0.699 + ], + "angle": 0, + "content": "We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.702, + 0.906, + 0.761 + ], + "angle": 0, + "content": "1. We propose the first benchmark VLM4D explicitly designed to rigorously evaluate the spatiotemporal (4D) reasoning capabilities of Vision Language Models (VLMs)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.763, + 0.905, + 0.822 + ], + "angle": 0, + "content": "2. We introduce a novel, meticulously curated dataset consisting of diverse real-world and synthetic video sequences paired with carefully crafted spatiotemporal question-answer (QA) annotations." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.823, + 0.905, + 0.898 + ], + "angle": 0, + "content": "3. We analyze the critical limitations of contemporary VLMs in spatiotemporal reasoning and conduct experiments exploring potential solutions, highlighting fundamental challenges and outlining clear directions for impactful future research." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.702, + 0.906, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8601" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.204, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Figure 3. Dataset Generation and Annotation Pipeline. Our dataset was constructed by collecting real videos and generating synthetic data, followed by human-in-the-loop quality reviews to address ambiguous videos and annotations. After temporal alignment and quality assurance, human-annotated questions and ground-truth answers were created, complemented by multiple-choice (MC) answers generated by large language models (LLMs). The final dataset includes real and synthetic video data with comprehensive VLM scoring metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.271, + 0.232, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.298, + 0.485, + 0.782 + ], + "angle": 0, + "content": "Spatiotemporal Understanding in Vision Language Models Early methods for video understanding before the advent of large vision language models (VLMs) leveraged trajectory-based representations and separate modeling of spatial and motion cues for tasks like action and motion recognition [11, 38, 67, 74]. Recently, VLMs have evolved rapidly by fully leveraging the significant achievements of Large Language Models (LLMs) [7, 9, 20, 65, 71, 80] and large-scale visual instruction tuning datasets [18, 50, 100]. While VLMs [3, 29, 33, 42, 50, 70, 76, 100] exhibit transformative potential for applications such as embodied AI [22, 36, 69], robotics [62, 73], scene generation [49, 60, 98], and world modeling [52, 93], most existing methods remain constrained to static images, focusing narrowly on spatial understanding [25] while overlooking the dynamic temporal dimension inherent in real-world interactions. To bridge this gap, emerging research [16, 44, 56, 91, 92] has begun exploring video modality integration, aiming to equip VLMs with spatial-temporal awareness critical for tasks like video comprehension, where both contextual details and motion dynamics are essential. For example, VideoLLM-MoD [83] proposes to address the efficiency issue when processing long-term video by mixture-of-depths. [88] introduces VideoRefer to enhance the finer-level (like object-level) spatial-temporal video understanding of VLMs. Grounded-VideoLLM [75] also targets for fine-grained video understanding through incorporating an additional temporal stream. In this work, we aim to rigorously evaluate the 4D spatial-temporal reasoning capabilities of state-of-the-art VLMs, probing how and to what extent these models internalize spatial intelligence and temporal dependencies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "VLM Benchmarks Following the development trends of VLMs, benchmarking VLMs shares the similar trajectory by first evaluating vision QA on static images [31, 41, 53, 89], to align with models' early focus on 2D understanding. As VLMs evolved to tackle dynamic scenarios, benchmarks expanded to evaluate general-purpose video comprehension" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.908, + 0.606 + ], + "angle": 0, + "content": "tasks that probe temporal coherence and event understanding [27, 35, 46, 48, 59]. Notably, MMVU [95] further proposes a knowledge-intensive benchmark to assess the expert-level reasoning ability of current video-based large models. However, while these works assess perception and semantic understanding, they largely overlook the explicit evaluation of spatial-temporal awareness, a core capability for real-world applications requiring 4D (3D space + time) reasoning. Recent efforts like [87] pioneer benchmarks for 3D visual-spatial intelligence but restrict evaluation to static 3D scene, neglecting the interplay of object motion and temporal dynamics intrinsic to videos. In this work, we introduce VLM4D, the first benchmark designed to holistically evaluate the 4D intelligence in VLMs, unifying spatial understanding, temporal continuity, and motion reasoning. By curating tasks that demand precise analysis of dynamic interactions (e.g., direction prediction, perspective anticipation, and motion reasoning), VLM4D exposes critical gaps in current models' ability to internalize spatiotemporal relationships. Our work not only advances the granularity of VLM evaluation but also shares insights and potential solutions to improve the model performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.738, + 0.634 + ], + "angle": 0, + "content": "3. The VLM4D Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.841 + ], + "angle": 0, + "content": "We introduce VLM4D, the first benchmark specifically designed to test the spatiotemporal reasoning ability of VLMs. VLM4D consists of 1000 videos paired with over 1800 question-answer pairs, each carefully designed to assess both spatial and temporal understanding jointly. The majority of these videos are sourced from datasets with rich spatiotemporal characteristics, thus ensuring a diverse range of motion-related scenarios. We also augment the dataset with synthetic videos generated by a world-foundation model, Cosmos [4], that has been modified using techniques introduced in [32] to obtain more accurate correspondence between motion-oriented prompts and the resulting generated video. Fig. 2 illustrates the composition of our dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.746, + 0.864 + ], + "angle": 0, + "content": "3.1. Benchmark Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Unlike prior work that often relies heavily on LLMs and VLMs to generate first iterations of benchmarks and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.207, + 0.102, + 0.68, + 0.117 + ], + "angle": 0, + "content": "Is the camel in the foreground turning to the left or right from its own perspective?" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.119, + 0.297, + 0.133 + ], + "angle": 0, + "content": "A. Not moving" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.119, + 0.386, + 0.132 + ], + "angle": 0, + "content": "B.Left" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.119, + 0.617, + 0.133 + ], + "angle": 0, + "content": "C. Moving straight, not turning" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.119, + 0.709, + 0.133 + ], + "angle": 0, + "content": "D. Right" + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.142, + 0.821, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.256, + 0.672, + 0.27 + ], + "angle": 0, + "content": "In which direction of rotation does the person pour the batter into the frying pan?" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.272, + 0.335, + 0.285 + ], + "angle": 0, + "content": "A. Counter-clockwise" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.273, + 0.453, + 0.285 + ], + "angle": 0, + "content": "B. Left to right" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.273, + 0.585, + 0.285 + ], + "angle": 0, + "content": "C. Right to left" + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.272, + 0.727, + 0.285 + ], + "angle": 0, + "content": "D. Clockwise" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.294, + 0.822, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.407, + 0.501, + 0.421 + ], + "angle": 0, + "content": "What direction is the robotic dog moving towards?" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.424, + 0.262, + 0.437 + ], + "angle": 0, + "content": "A. Right" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.425, + 0.449, + 0.437 + ], + "angle": 0, + "content": "B. no robotic dog there" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.586, + 0.438 + ], + "angle": 0, + "content": "C. Not moving" + }, + { + "type": "text", + "bbox": [ + 0.636, + 0.425, + 0.68, + 0.437 + ], + "angle": 0, + "content": "D.Left" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.446, + 0.823, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.558, + 0.908, + 0.601 + ], + "angle": 0, + "content": "Figure 4. Qualitative Examples of Dataset Annotations. (Top) A third-person (exo-centric) video with translational annotations (\"camel turning left from its perspective\"). (Middle) A first-person (ego-centric) video with a rotational question (\"clockwise rotation of ladle\"). (Bottom) A synthetic scene with motion recognition \"robotic dog moving left\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.716 + ], + "angle": 0, + "content": "datasets [14] followed by human quality control, we found that existing VLMs and automated methods showed significant limitations in terms of reliability and quality. This shortcoming necessitated direct human annotations that were then followed by augmentation by LLMs to ensure a high-quality benchmark. An overview of the benchmark curation pipeline is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Real Video Data Collection Real-world videos were sourced from datasets with rich spatiotemporal characteristics that ensured diverse motion and perspective variations. For egocentric data, we relied mainly on the Ego4D dataset [30], while most exocentric data points were collected from the DAVIS [64] and YouTube-VOS [85] datasets. To minimize confounders and to focus attention of VLM abilities to only spatiotemporal reasoning, we preprocessed the videos by temporally segmenting and centering them around the most relevant action, thus resulting in videos with an average duration of 3-8 seconds. This en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.656 + ], + "angle": 0, + "content": "sures that the key event described in the question is clear and reduces ambiguities or confounders that would reduce VLM accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Synthetic Video Data Generation The rapid advancement of video generation techniques has led to their widespread application in diverse domains like robotics and immersive entertainment [4, 8, 47, 72, 82], making 4D reasoning on synthetic content essential for VLMs and underscoring the need to include such videos in our benchmark. For synthetic video generation, we use Cosmos [4] as our video generation backbone. To ensure that the generated videos align with the intended object moving directions, we incorporate input bounding boxes as additional spatial guidance. Specifically, we follow the approach introduced in [32] modifying the diffusion forward steps to enforce object localization constraints at each timestep, ensuring consistency between the generated object direction and the user-specified trajectory. The average duration of generated" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.332 + ], + "angle": 0, + "content": "synthetic videos is 5 seconds. To maintain high-quality outputs, we perform a manual verification step after generation, filtering out low-quality videos and retaining only those that accurately match the specified directions. Once a video is generated, we use an LLM (GPT-4o) to create two types of evaluation questions: Directional questions, derived from the textual prompt used to generate the video; and False Positive (FP) questions (constituting \\(10\\%\\), matching the ratio in the real dataset), which query non-existent objects in the scene. Both question types follow the format: \"What direction is the \\(\\langle\\) Object Name \\rangle\\) moving?\", where the model must select one of four possible answers: \"left\", \"right\", \"not moving\", or \"no \\(\\langle\\) Object Name \\rangle\\) there\". A final manual review is conducted to filter out or revise ambiguous questions, ensuring the quality of the questions and ground-truth answers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.519 + ], + "angle": 0, + "content": "QA Generation and Quality Control Question-answer pairs are primarily constructed through human annotations. The question answer pairs are then supplemented with alternative answers by an LLM (GPT-4o) for multiple choice (MC) questions. To ensure high-quality annotations, we applied a three-round, multi-person cross-checked verification process, during which ambiguous videos were filtered out and vague, misleading, or incorrect QA pairs were refined to improve spatial and temporal alignment between language and visual content. Fig. 4 showcases some qualitative examples of annotations for different types of videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Assessing Human Performance To establish a human performance baseline on our benchmark, we conducted an evaluation in which participants independently answered 100 randomly sampled questions from the dataset. The accuracy of human responses was then aggregated to approximate the performance of human spatiotemporal reasoning on the dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.453, + 0.669 + ], + "angle": 0, + "content": "3.2. Categorizing Spatiotemporal Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To systematically evaluate spatiotemporal reasoning capabilities, we first categorize videos into two primary groups: egocentric (first-person) videos and exocentric (third-person) videos. Egocentric videos are sourced from the Ego4D [30] dataset, where scenes are captured from a head-mounted camera, thus offering dynamic video data that is inherently coupled with the individual's actions. Exocentric videos encompass a diverse range of recorded scenes, from sports footage to everyday scenes. Beyond this categorization, we also evaluate spatiotemporal performance across four dimensions: translational movement (TM), rotational movement (RM), spatiotemporal counting (STM), and false positives (FP), with their proportions shown in Fig. 2. Translational movement assesses a model's ability to track linear motion within scenes, while rotational" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.092, + 0.902, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.368, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Figure 5. Comparison of accuracy across types of spatiotemporal questions. Model accuracy is shown only for the six top-performing VLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.593 + ], + "angle": 0, + "content": "movement assesses the understanding of changes in orientation and perspective shifts over time. Spatiotemporal counting extends these core motion-based tasks by requiring a more complex reasoning strategy to determine the number of objects performing a translation or rotational movement. Lastly, the false positives category evaluates the model's critical thinking [23] in determining whether an object or event actually occurred within the spatiotemporal context. By structuring the benchmark along these axes, we aim for a comprehensive framework for assessing spatiotemporal reasoning (Fig. 5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.818, + 0.623 + ], + "angle": 0, + "content": "4. Evaluation of VLM4D Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.685, + 0.649 + ], + "angle": 0, + "content": "4.1. Evaluation Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.851 + ], + "angle": 0, + "content": "Benchmark Models We evaluate 23 most recently released VLMs thus covering a wide range of model sizes, architectures, and training methodologies. For closed-source VLMs, we evaluate GPT-4o [61], Gemini 2.5 Pro [28], Claude Sonnet 4 [6], and Grok-2-Vision [84]. For open-source models, we include Llama 4 [58], DeepSeek-VL [54], Qwen2.5-VL [86], Qwen2-VL [76], InternVL2.5 [15], Aria [43], InternVideo 2.5 [79], InternVideo2 [78], Phi-4-multimodal [2], Phi-3.5-vision [1], Pixtral [5], VideoLLama3 [91], Llava-One-Vision [42], Llava-NeXT-Video [94]. When available, we evaluate different sizes for each model, resulting in models ranging from 2 to 72 billion parameters." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation Settings The evaluations were performed in a zero-shot setting with video or a set of sampled frames of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8604" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.904, + 0.166 + ], + "angle": 0, + "content": "
OrganizationModelReleaseRealSyntheticOverall
Ego-centricExo-centricAverageDirectionalFPAverage
User StudyHuman Performance99.699.799.795.810096.298.8
RandomRandom Selection24.423.223.625.524.725.424.1
" + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.169, + 0.903, + 0.253 + ], + "angle": 0, + "content": "
Latest Proprietary VLMs
OpenAIGPT-4o2024-1155.562.260.049.553.349.957.5
GoogleGemini-2.5-Pro2025-664.662.963.554.880.057.362.0
AnthropicClaude-Sonnet-42025-552.652.152.244.086.748.351.3
xAIGrok-2-Vision2024-1248.849.749.449.366.751.049.8
" + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.256, + 0.903, + 0.415 + ], + "angle": 0, + "content": "
Open-source Image VLMs
MetaLlama-4-Maverick-17B2025-452.654.353.853.351.153.053.6
Llama-4-Scout-17B2025-448.656.253.753.375.655.554.1
MicrosoftPhi-4-Multimodal2025-341.035.437.237.511.134.836.6
Phi-3.5-Vision2024-733.438.837.123.337.824.734.0
DeepSeekDeepSeek-VL22024-1233.632.933.131.846.733.333.2
Shanghai AI LabInternVL2.5-38B2024-1146.650.148.943.357.844.747.9
InternVL2.5-8B2024-1139.044.042.440.842.240.942.0
Mistral AIPixtral-12B2024-932.325.827.924.322.224.027.0
RhymesAria2024-1147.244.045.138.571.141.844.3
" + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.417, + 0.903, + 0.593 + ], + "angle": 0, + "content": "
Open-source Video VLMs
AlibabaQwen2.5-VL-7B2025-142.343.743.343.564.445.643.8
Qwen2.5-VL-72B2025-154.352.553.149.580.052.653.0
Qwen2-VL-7B2024-836.134.735.240.535.640.036.3
Qwen2-VL-72B2024-948.143.044.640.873.344.044.5
DAMOVideoLLama3-2B2025-153.242.546.034.355.636.443.7
VideoLLama3-7B2025-149.445.146.542.853.343.845.9
Shanghai AI LabInternVideo2.5-8B2025-157.250.552.744.346.744.550.7
InternVideo2-8B2024-835.639.338.143.00.038.738.2
LLaVALLaVA-One-Vision-7B2024-936.835.636.037.835.637.536.3
LLaVA-NeXT-Video-34B2024-629.631.630.924.555.627.630.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.604, + 0.907, + 0.634 + ], + "angle": 0, + "content": "Table 1. Evaluation on VLM4D Benchmark across various proprietary and open-source VLMs. Top three performers in each category are highlighted from dark (highest) to light (third highest). Human and random selection baselines are included for reference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.483, + 0.768 + ], + "angle": 0, + "content": "video, followed by the prompt forming the input. For each model, we evaluate on two different inference settings. In the first setting, the model prompted to directly output (DO) the answer immediately without any reasoning, and in the second evaluation setting, the model is directed to create intermediate reasoning steps, chain-of-thought (CoT) [81], before inferring the final answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Metrics Following prior work [87] and given the nature of our target task, we adopt multiple-choice questions (MCQs) for evaluation, using accuracy as the primary metric. For the two inference settings described earlier, we employ LLM-as-Judge [95] to assess the outputs of VLMs. We opt for this method instead of string or template matching, as VLMs—especially under chain of thought (CoT) prompt-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.907, + 0.799 + ], + "angle": 0, + "content": "ing—often generate all possible answer choices during reasoning, with varying frequencies and slight formatting differences. In some cases, the final answer may even contradict the reasoning. To better evaluate whether the model truly understands the video, we prompt two advanced LLMs (GPT-o3 and o4-mini) to grade based on the full CoT reasoning, not just the final answer. We then perform a cross-check between their judgments and manually resolve any disagreements. The evaluation results are reported in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.702, + 0.831 + ], + "angle": 0, + "content": "4.2. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our comprehensive evaluation on the VLM4D benchmark, detailed in Tab. 1, systematically assesses the spatiotemporal reasoning capabilities of modern Vision-Language Models (VLMs). The results highlight a clear hierar" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8605" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.334 + ], + "angle": 0, + "content": "chy, with proprietary models demonstrating superior performance over their open-source counterparts. Google's Gemini-2.5-Pro emerges as the top-performing model with an overall accuracy of \\(62.0\\%\\), followed by OpenAI's GPT4o at \\(57.5\\%\\). Within the open-source domain, image-based models like Meta's Llama-4-Scout-17B (54.1%) and videosupported models like Alibaba's Qwen2.5-VL-72B (53.0%) demonstrate highly competitive results, even outperforming some proprietary counterparts. Despite these achievements, a significant performance gap persists when compared to human accuracy (98.8%), underscoring that sophisticated 4D awareness remains a formidable challenge for AI. Notably, performance varies significantly across different categories, such as real versus synthetic data and egocentric versus exo-centric perspectives, indicating that current models lack generalized spatiotemporal understanding." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.349, + 0.455, + 0.366 + ], + "angle": 0, + "content": "5. Analysis: Why VLMs Don't Work Well?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.393, + 0.391 + ], + "angle": 0, + "content": "5.1. Limited Spatiotemporal Cognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Despite significant advances in VLMs, their abilities to understand and reason about motion, spatial relationships, and temporal coherence remains fundamentally underdeveloped [13, 101]. Chain of thought (CoT) [81] is widely employed as a method to improve accuracy through step-by-step reasoning. We showcase a comparison between CoT and DO in Fig. 6. Overall, there is no indication of a large advantage of CoT over all evaluated models. Upon deeper exploration of the CoT reasoning of some models, we observe that the reasoning process was primarily flawed in the following ways: irrelevant information and arriving at conclusions that are inconsistent with the reasoning process. Larger models exhibited strategies that would be similar to how a human processes spatiotemporal information, but the resulting execution falls short of human performance. This demonstrates a disconnect between its visual and linguistic knowledge. We provide examples of this behavior in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.433, + 0.699 + ], + "angle": 0, + "content": "5.2. Deficiencies in Spatiotemporal Labeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Another avenue of exploration we undertook is to understand the richness of spatiotemporal labels in popular supervised fine-tuning (SFT) VLM datasets. Typically, video captioning occurs at the 'scene' level, lacking fine-grained temporal, spatial, and object-level details. We performed an extensive analysis, encompassing over 2 million samples [14, 17, 44, 45, 94]. We performed this analysis through string-matching of spatiotemporal descriptors related to directionality, translational motion, rotation, and perspective shifts and provide the overall results in Fig. 7. We then performed a manual finegrained evaluation of the ShareGPT4Video dataset [14] which we found had the highest density of spatiotemporal dataset. We found that" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.342, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Comparison of CoT and DO Accuracy Across Models. Accuracy comparison between Chain-of-Thought (CoT) and Direct Output (DO) prompting across VLMs." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.398, + 0.918, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.657, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Figure 7. Heatmap of Occurrences of Spatial-Temporal Terms in popular video SFT datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.817 + ], + "angle": 0, + "content": "from a sample of 100 labels that were detected as spatiotemporal, less than \\(10\\%\\) of them were judged as accurate upon human evaluation. This result underscores the inadequacy of current dense captioning approaches, which frequently generate spatiotemporal descriptors without capturing precise motion dynamics. We provide more detailed analysis and explanations in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.75, + 0.848 + ], + "angle": 0, + "content": "6. Probing Future Solutions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To probe promising future solutions for enhancing spatiotemporal video understanding, we propose two approaches that address some of the shortcomings of current" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8606" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "state-of-the-art VLMs: fine-tuning a VLM on data-rich in spatiotemporal actions and the other leveraging 4D reconstruction and feature fields jointly with a VLM. SFT refines the model's abilities by training on datasets that contain temporally and spatially rich actions and interactions. By integrating structured visual representations and targeted fine-tuning, these approaches enhance video-language models' ability to interpret motion. The second method lifts the feature space of VLMs into a temporally coherent 4D feature field, providing structured scene representations that improve motion and spatial reasoning in the stage of decoding and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Spatial-Temporal SFT We evaluate on a subset split of the real dataset by randomly splitting the real-world dataset into a training and testing split (80% / 20%) and we try settings using synthetic/real both for training. We conducted the experiments using Qwen 2VL (7B) and Qwen 2.5VL (7B) through LLama-Factory [96], and compared the performance before and after supervised fine-tuning in Tab. 2. The results demonstrated an improvement in accuracy in spatiotemporal reasoning, suggesting that performance gains can be obtained through targeted training. However, the addition of synthetic data does not necessarily increase performance over using real data alone, suggesting the importance of synthetic data quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.902 + ], + "angle": 0, + "content": "4D Feature Fields Reconstruction Recent advances in 3D/4D feature fields reconstruction methods [24, 37, 90, 97, 99] have significantly enhanced the vision foundation model's performance in 3D/4D space by integrating structured latent scene representations into the model's inference stage. Inspired by the promising results of feature lifting, we explore enhancing the InternVideo2-8B model [78] with spatiotemporal awareness by adopting the strategy proposed in Feature4X [99], which constructs the VLM's 2D feature space along the time dimension into a 4D feature field. To assess this approach, we evaluate performance on a subset of the VLM4D benchmark, specifically leveraging all 50 videos from the DAVIS 2016 dataset [63]. Our experimental evaluation compares inference performance across three distinct input modalities: original 2D videos; rendered novel global-view videos (which provide broader contextual information in 2D format); and reconstructed global feature fields (which implicitly incorporate 4D scene-level information during reasoning). Table 3 reveals that reconstructed feature fields achieve the highest accuracy across both reasoning types. This success stems from two key advantages: the inherent structure of 4D representations and the ability of feature field inference to avoid the rendering artifacts of RGB reconstruction (global view video). However, the current approach requires per-scene optimization as a post-processing step, limiting its generalizability and" + }, + { + "type": "table", + "bbox": [ + 0.593, + 0.089, + 0.83, + 0.313 + ], + "angle": 0, + "content": "
ModelMC
Original Model
Qwen 2VL (7B)38.3
Qwen 2.5VL (7B)43.4
Finetuned Model
Qwen 2VL (7B) (R)54.5
Qwen 2VL (7B) (S)42.5
Qwen 2VL (7B) (R+S)55.5
Qwen 2.5VL (7B) (R)55.6
Qwen 2.5VL (7B) (S)48.6
Qwen 2.5VL (7B) (R+S)56.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.323, + 0.905, + 0.365 + ], + "angle": 0, + "content": "Table 2. SFT on Spatial-Temporal Datasets. MC refers to the performance of multiple-choice, while R, S, and \\(\\mathbb{R} + \\mathbb{S}\\) denote real, synthetic, and both real+synthetic usage of data for fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.378, + 0.868, + 0.565 + ], + "angle": 0, + "content": "
Input ModalityAccuracy
Chain of Thought Response
Original 2D Video36.0
Global View Video32.7
Global Feature Field37.4
Direct Output Response
Original 2D Video24.3
Global View Video23.8
Global Feature Field29.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.575, + 0.905, + 0.617 + ], + "angle": 0, + "content": "Table 3. InternVideo2 Accuracy with 4D Feature Field Reconstruction. Comparison of InternVideo2's performance given different input modalities from the same subset." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.636, + 0.756, + 0.651 + ], + "angle": 0, + "content": "making it computationally intensive." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Through the construction of the VLM4D benchmark, we evaluate the spatiotemporal reasoning capabilities of various Vision-Language Models (both open-source and proprietary). While more recently released models demonstrate improved performance over their counterparts, they remain significantly behind human proficiency. Overall, our work questions whether VLMs possess spatiotemporal reasoning abilities that are imperative to have for more sophisticated visual agents in fields ranging from robotics to interactive AI systems that require a deep understanding of dynamic visual environments. We hope to inspire future work to explore novel approaches for integrating spatiotemporal grounding, thereby enhancing their spatiotemporal reasoning capabilities and facilitating robust deployment." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8607" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.271, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[3] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Weijian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.563, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[4] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.633, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[5] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.704, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[6] Anthropic. System card: Claude opus 4 & claude sonnet 4. Technical report, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.733, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.789, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al." + }, + { + "type": "list", + "bbox": [ + 0.107, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.094, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.135, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[10] Neil Burgess. Spatial memory: how egocentric and allocentric combine. Trends in Cognitive Sciences, 10(12):551-557, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[11] Pradyumna Chari, Yunhao Ba, Shijie Zhou, Chinmay Talegaonkar, Shreeram Athreya, and Achuta Kadambi. On learning mechanical laws of motion from video using neural networks. IEEE Access, 11:30129-30145, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.234, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[12] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18030-18040, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.304, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[13] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[14] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.443, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[16] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.596, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[17] Erfei Cui, Yinan He, Zheng Ma, Zhe Chen, Hao Tian, Weiyun Wang, Kunchang Li, Yi Wang, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, Yali Wang, Limin Wang, Yu Qiao, and Jifeng Dai. Sharegpt-4o: Comprehensive multimodal annotations with gpt-4o, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.735, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[19] Julian De Freitas, Nicholas E. Myers, and Anna C. Nobre. Tracking the changing feature of a moving object. Journal of Vision, 16(3):22, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.776, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8608" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[23] Chenrui Fan, Ming Li, Lichao Sun, and Tianyi Zhou. Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? arXiv preprint arXiv:2504.06514, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[24] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in neural information processing systems, 37:40212-40229, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[25] Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Huaizhi Qu, Dilin Wang, Zhicheng Yan, et al. Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction. arXiv preprint arXiv:2505.20279, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[26] Jennifer J. Freyd and Ronald A. Finke. Representational momentum. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10(1):126-132, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[27] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[28] Google Gemini Team. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Technical report, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[29] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[31] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multidiscipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] Xuehai He, Shuohang Wang, Jianwei Yang, Xiaoxia Wu, Yiping Wang, Kuan Wang, Zheng Zhan, Olatunjri Ruwase, Yelong Shen, and Xin Eric Wang. Mojito: Motion tra" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "jctory and intensity control for video generation. arXiv preprint arXiv: 2412.08948, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[33] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[34] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.219, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[35] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. How good is my video lmm? complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[36] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.372, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[37] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in neural information processing systems, 35:23311-23330, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.428, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[38] Ivan Laptev. On space-time interest points. International Journal of Computer Vision, 64(2-3):107-123, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.457, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[39] Yann LeCun, Bernhard Boser, John S. Denker, Donnie Henderson, Richard E. Howard, Wayne Hubbard, and Lawrence D. Jackel. Backpropagation applied to handwritten zip code recognition. *Neural Computation*, 1(4):541-551, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.526, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[40] Alan M. Leslie. Spatiotemporal continuity and the perception of causality in infants. Perception, 13(3):287-305, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[41] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.693, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[43] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8609" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[47] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, and Zhiwen Fan. 4k4DGen: Panoramic 4d generation at 4k resolution. In The Thirteenth International Conference on Learning Representations, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[48] Xinhao Li, Zhenpeng Huang, Jing Wang, Kunchang Li, and Limin Wang. Videoeval: Comprehensive benchmark suite for low-cost evaluation of video foundation model. arXiv preprint arXiv:2407.06491, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[49] Lu Ling, Chen-Hsuan Lin, Tsung-Yi Lin, Yifan Ding, Yu Zeng, Yichen Sheng, Yunhao Ge, Ming-Yu Liu, Aniket Bera, and Zhaoshuo Li. Scenethesis: A language and vision agentic framework for 3d scene generation. arXiv preprint arXiv:2505.02836, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[50] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[51] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[52] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv preprint arXiv:2402.08268, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[53] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[54] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[55] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[56] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[57] D. Marr and S. Ullman. Directional selectivity and its use in early visual processing. Proceedings of the Royal Society of London. Series B, Biological Sciences, 211(1183):151-180, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[58] Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. Technical report, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[59] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluat" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ing video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[60] Basak Melis Öcal, Maxim Tatarchenko, Sezer Karaoglu, and Theo Gevers. Sceneteller: Language-to-3d scene generation. In European Conference on Computer Vision, pages 362–378. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.178, + 0.847, + 0.19 + ], + "angle": 0, + "content": "[61] OpenAI. Hello gpt-4o. Technical report, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.192, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[62] Shivansh Patel, Xinchen Yin, Wenlong Huang, Shubham Garg, Hooshang Nayyeri, Li Fei-Fei, Svetlana Lazebnik, and Yunzhu Li. A real-to-sim-to-real approach to robotic manipulation with vlm-generated iterative keypoint rewards. arXiv preprint arXiv:2502.08643, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.261, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[64] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.401, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[65] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.442, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[67] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems (NIPS), pages 568-576, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.596, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[68] Elizabeth S. Spelke and Katherine D. Kinzler. Core knowledge. Developmental Science, 10(1):89-96, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.624, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[69] Alessandro Suglia, Claudio Greco, Katie Baker, Jose L Part, Ioannis Papaioannou, Arash Eshghi, Ioannis Konstas, and Oliver Lemon. Alanavlm: A multimodal embodied ai foundation model for egocentric video understanding. arXiv preprint arXiv:2406.13807, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.693, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[70] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[71] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[72] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 4" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8610" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[73] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[74] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Action recognition by dense trajectories. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3169-3176, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[75] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[76] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[77] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Intervideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[79] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[80] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[82] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26057-26068, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.44, + 0.872 + ], + "angle": 0, + "content": "[84] xAI. Grok-2 beta release. Technical report, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[85] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen," + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[86] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[87] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. arXiv preprint arXiv:2412.14171, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[88] Yuqian Yuan, Hang Zhang, Wentong Li, Zesen Cheng, Boqiang Zhang, Long Li, Xin Li, Deli Zhao, Wenqiao Zhang, Yueting Zhuang, et al. Videorefer suite: Advancing spatial-temporal object understanding with video llm. arXiv preprint arXiv:2501.00599, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.334, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[90] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In European Conference on Computer Vision, pages 57-74. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[91] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.562, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[93] Hongxin Zhang, Zeyuan Wang, Qiushi Lyu, Zheyuan Zhang, Sunli Chen, Tianmin Shu, Yilun Du, and Chuang Gan. Combo: compositional world models for embodied multi-agent cooperation. arXiv preprint arXiv:2404.10775, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[94] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.733, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[95] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.804, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[96] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of \\(100+\\) language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand, 2024. Association for Computational Linguistics. 8" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8611" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[97] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[98] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.277, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[99] Shijie Zhou, Hui Ren, Yijia Weng, Shuwang Zhang, Zhen Wang, Dejia Xu, Zhiwen Fan, Suya You, Zhangyang Wang, Leonidas Guibas, et al. Feature4x: Bridging any monocular video to 4d agentic ai with versatile gaussian feature fields. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14179-14190, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.36, + 0.482, + 0.414 + ], + "angle": 0, + "content": "[100] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.417, + 0.482, + 0.484 + ], + "angle": 0, + "content": "[101] Orr Zohar, Xiaohan Wang, Yann Dubois, Nikhil Mehta, Tong Xiao, Philippe Hansen-Estruch, Licheng Yu, Xiaofang Wang, Felix Juefei-Xu, Ning Zhang, et al. Apollo: An exploration of video understanding in large multimodal models. arXiv preprint arXiv:2412.10360, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8612" + } + ] +] \ No newline at end of file diff --git a/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_origin.pdf b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4750628b9bd0b2188460b1400c0e6b55cd93f3a6 --- /dev/null +++ b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/7ad15710-76ed-462f-ba27-85159b1f2146_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:499da69e1352b4b84277c4917aa60b8b7ba1ea80b132e541127eaa41e0028464 +size 2700792 diff --git a/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/full.md b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6acb2353d1cc4cfe0ca57a3f0b3f541c660bb8e --- /dev/null +++ b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/full.md @@ -0,0 +1,312 @@ +# VLM4D: Towards Spatiotemporal Awareness in Vision Language Models + +Shijie Zhou $^{1*}$ Alexander Vilesov $^{1*}$ Xuehai He $^{2,3*}$ Ziyu Wan $^{2}$ Shuwang Zhang $^{1}$ + +Aditya Nagachandra $^{1}$ Di Chang $^{4}$ Dongdong Chen $^{2}$ Xin Eric Wang $^{3}$ Achuta Kadambi $^{1}$ + +$^{1}$ UCLA ${ }^{2}$ Microsoft ${ }^{3}$ UCSC ${ }^{4}$ USC + +https://v1m4d.github.io/ + +![](images/47589e52e181e116a7dc2dff9f94b43df83d71483d046e3a92896847c40c8cd1.jpg) + +![](images/995244dc3b4f1d01690e042e2a6374a85d43bdee2f33aa3c9bd1f7094c77dd60.jpg) +Figure 1. Spatiotemporal (4D) Awareness. Humans intuitively reason in 4D (3D space + time), effortlessly reconstructing the dynamic spatial trajectory of moving objects from any perspective. In contrast, current Vision Language Models (VLMs) typically rely on aggregating 2D visual features across time, leading to incorrect predictions when motion understanding and interpretation requires deeper spatiotemporal reasoning. In this example, humans correctly perceive the car moving to the right, while the VLM (GPT-4o) inaccurately predicts leftward movement, suggesting VLMs struggle to perform spatiotemporal reasoning. + +# Abstract + +Vision language models (VLMs) have shown remarkable capabilities in integrating linguistic and visual reasoning but remain fundamentally limited in understanding dynamic spatiotemporal interactions. Humans effortlessly track and reason about object movements, rotations, and perspective shifts—abilities essential for robust dynamic real-world understanding yet notably lacking in current VLMs. In this paper, we introduce VLM4D, the first benchmark specifically designed to evaluate the spatiotemporal reasoning capa + +bilities of VLMs. Our benchmark comprises diverse real-world and synthetic videos accompanied by carefully curated question-answer pairs emphasizing translational and rotational motions, perspective awareness, and motion continuity. Through comprehensive evaluations of state-of-the-art open and closed-source VLMs, we identify significant performance gaps compared to human baselines, highlighting fundamental deficiencies in existing models. Extensive analysis reveals that VLMs struggle particularly with integrating multiple visual cues and maintaining temporal coherence. We further explore promising directions, such as leveraging 4D feature field reconstruction and targeted + +spatiotemporal supervised fine-tuning, demonstrating their effectiveness in enhancing spatiotemporal comprehension. Our work aims to encourage deeper exploration into improving VLMs' spatial and temporal grounding, paving the way towards more capable and reliable visual intelligence for dynamic environments. + +# 1. Introduction + +Humans possess an innate ability to perceive, track, and interpret motion as well as spatial and temporal changes, enabling rich interpretations of complex dynamic events from both egocentric and allocentric perspectives [10, 19, 57]. When observing an object move, we can inherently process any changes such as lateral shifts, rotational directions, and periodic or repeated actions unfolding along a specific trajectory [10]. These sophisticated perceptual abilities are a product of our spatiotemporal cognition [26], and form an essential foundation that allows us to comprehend and reason about physical phenomena, object interactions, and causal relationships within our environment [34, 40]. + +Vision language models (VLMs), which also have the potential to perceive motion and spatiotemporal changes in videos, represent a prominent class of methods aimed at emulating or surpassing human capabilities in integrated visual and linguistic reasoning [21, 39]. While previous work on VLMs has primarily focused on static visual understanding—through large-scale training on paired language and image data [66]—or explored video understanding tasks such as captioning [55] and scene understanding [12], we find that their strong performance in these tasks does not naturally translate into robust spatiotemporal reasoning. This limitation is particularly striking given that state-of-the-art VLMs are typically trained on datasets comprising hundreds of billions of tokens [51]. In contrast, human infants naturally develop robust spatiotemporal cognition within the first few months of life [68]. + +Another key challenge that limits VLM performance on spatiotemporal tasks is the need to implicitly or explicitly reconstruct a four-dimensional (4D) representation—3D space + time—of dynamic scenes, and subsequently reason over this reconstruction [77]. As illustrated in Fig. 1, the car is advancing forwards and turning to the left in its own frame of reference. However, from the camera's perspective, its motion appears as a combination of heading to the right and receding into the distance despite the car being in the center of the frame due to camera rotation. Human observers can seamlessly disentangle these complex dynamics, accurately interpreting trajectories by synthesizing diverse visual cues including camera rotation compensation, stationary scene landmarks, prior knowledge of 3D and 4D environmental structures, and perspective projections [10, 26, 40, 57]. The inability of current VLMs to in- + +![](images/f0aa1aaa38581fffceb31eee1e39292dac7e0c3be047929474dd30272a6c38b8.jpg) +Figure 2. Distribution of Dataset Sources and Annotations. Overview of the dataset composition, illustrating the proportions of real third-person (exo-centric) videos (DAVIS [64], YouTube-VOS [85]), real first-person (ego-centric) videos (Ego4D [30]), and synthetic videos (Cosmos [4]). The real video data is further categorized by annotation types, including translational, rotational, action, counting, and false positive queries (targeting nonexistent events to assess critical reasoning). + +tegrate spatial, temporal, and semantic cues in a human-like manner stems from their fundamentally different encoding paradigms for visual and language information. This discrepancy highlights a significant gap between human and machine spatiotemporal understanding, suggesting that future VLMs may benefit from insights in cognitive science and neuroscience to develop more advanced mechanisms for perceiving, integrating, reconstructing, and reasoning over dynamic scenes. + +In order to effectively characterize and challenge the existing spatiotemporal reasoning abilities of VLMs, we introduce VLM4D, a rigorous benchmark specifically designed to probe the spatiotemporal grounding capabilities of current vision language models. Through this contribution, we aim to catalyze research that addresses the critical gap in spatiotemporal understanding and reasoning within VLMs and provide a foundational analysis highlighting key deficiencies in existing models. + +We summarize our main contributions as follows: + +1. We propose the first benchmark VLM4D explicitly designed to rigorously evaluate the spatiotemporal (4D) reasoning capabilities of Vision Language Models (VLMs). +2. We introduce a novel, meticulously curated dataset consisting of diverse real-world and synthetic video sequences paired with carefully crafted spatiotemporal question-answer (QA) annotations. +3. We analyze the critical limitations of contemporary VLMs in spatiotemporal reasoning and conduct experiments exploring potential solutions, highlighting fundamental challenges and outlining clear directions for impactful future research. + +![](images/17ed82081f03e0cfa98e209ce96212381a6252413b1a8496f303a884553cbe8a.jpg) +Figure 3. Dataset Generation and Annotation Pipeline. Our dataset was constructed by collecting real videos and generating synthetic data, followed by human-in-the-loop quality reviews to address ambiguous videos and annotations. After temporal alignment and quality assurance, human-annotated questions and ground-truth answers were created, complemented by multiple-choice (MC) answers generated by large language models (LLMs). The final dataset includes real and synthetic video data with comprehensive VLM scoring metrics. + +# 2. Related Work + +Spatiotemporal Understanding in Vision Language Models Early methods for video understanding before the advent of large vision language models (VLMs) leveraged trajectory-based representations and separate modeling of spatial and motion cues for tasks like action and motion recognition [11, 38, 67, 74]. Recently, VLMs have evolved rapidly by fully leveraging the significant achievements of Large Language Models (LLMs) [7, 9, 20, 65, 71, 80] and large-scale visual instruction tuning datasets [18, 50, 100]. While VLMs [3, 29, 33, 42, 50, 70, 76, 100] exhibit transformative potential for applications such as embodied AI [22, 36, 69], robotics [62, 73], scene generation [49, 60, 98], and world modeling [52, 93], most existing methods remain constrained to static images, focusing narrowly on spatial understanding [25] while overlooking the dynamic temporal dimension inherent in real-world interactions. To bridge this gap, emerging research [16, 44, 56, 91, 92] has begun exploring video modality integration, aiming to equip VLMs with spatial-temporal awareness critical for tasks like video comprehension, where both contextual details and motion dynamics are essential. For example, VideoLLM-MoD [83] proposes to address the efficiency issue when processing long-term video by mixture-of-depths. [88] introduces VideoRefer to enhance the finer-level (like object-level) spatial-temporal video understanding of VLMs. Grounded-VideoLLM [75] also targets for fine-grained video understanding through incorporating an additional temporal stream. In this work, we aim to rigorously evaluate the 4D spatial-temporal reasoning capabilities of state-of-the-art VLMs, probing how and to what extent these models internalize spatial intelligence and temporal dependencies. + +VLM Benchmarks Following the development trends of VLMs, benchmarking VLMs shares the similar trajectory by first evaluating vision QA on static images [31, 41, 53, 89], to align with models' early focus on 2D understanding. As VLMs evolved to tackle dynamic scenarios, benchmarks expanded to evaluate general-purpose video comprehension + +tasks that probe temporal coherence and event understanding [27, 35, 46, 48, 59]. Notably, MMVU [95] further proposes a knowledge-intensive benchmark to assess the expert-level reasoning ability of current video-based large models. However, while these works assess perception and semantic understanding, they largely overlook the explicit evaluation of spatial-temporal awareness, a core capability for real-world applications requiring 4D (3D space + time) reasoning. Recent efforts like [87] pioneer benchmarks for 3D visual-spatial intelligence but restrict evaluation to static 3D scene, neglecting the interplay of object motion and temporal dynamics intrinsic to videos. In this work, we introduce VLM4D, the first benchmark designed to holistically evaluate the 4D intelligence in VLMs, unifying spatial understanding, temporal continuity, and motion reasoning. By curating tasks that demand precise analysis of dynamic interactions (e.g., direction prediction, perspective anticipation, and motion reasoning), VLM4D exposes critical gaps in current models' ability to internalize spatiotemporal relationships. Our work not only advances the granularity of VLM evaluation but also shares insights and potential solutions to improve the model performance. + +# 3. The VLM4D Benchmark + +We introduce VLM4D, the first benchmark specifically designed to test the spatiotemporal reasoning ability of VLMs. VLM4D consists of 1000 videos paired with over 1800 question-answer pairs, each carefully designed to assess both spatial and temporal understanding jointly. The majority of these videos are sourced from datasets with rich spatiotemporal characteristics, thus ensuring a diverse range of motion-related scenarios. We also augment the dataset with synthetic videos generated by a world-foundation model, Cosmos [4], that has been modified using techniques introduced in [32] to obtain more accurate correspondence between motion-oriented prompts and the resulting generated video. Fig. 2 illustrates the composition of our dataset. + +# 3.1. Benchmark Construction + +Unlike prior work that often relies heavily on LLMs and VLMs to generate first iterations of benchmarks and + +Is the camel in the foreground turning to the left or right from its own perspective? + +A. Not moving + +B.Left + +C. Moving straight, not turning + +D. Right + +![](images/26d1f9f11602467ade67070edee943afe3d6a423a568579116b1b6428c5aefaa.jpg) + +In which direction of rotation does the person pour the batter into the frying pan? + +A. Counter-clockwise + +B. Left to right + +C. Right to left + +D. Clockwise + +![](images/7d1d22256b012a8a1c29c9d1d0fafe9a710fc30405fcffc07bef763ae9f8908d.jpg) + +What direction is the robotic dog moving towards? + +A. Right + +B. no robotic dog there + +C. Not moving + +D.Left + +![](images/119eaa33d817ce107020693246e0587f027ae636191a9d4d9ef5e3c81143d590.jpg) +Figure 4. Qualitative Examples of Dataset Annotations. (Top) A third-person (exo-centric) video with translational annotations ("camel turning left from its perspective"). (Middle) A first-person (ego-centric) video with a rotational question ("clockwise rotation of ladle"). (Bottom) A synthetic scene with motion recognition "robotic dog moving left". + +datasets [14] followed by human quality control, we found that existing VLMs and automated methods showed significant limitations in terms of reliability and quality. This shortcoming necessitated direct human annotations that were then followed by augmentation by LLMs to ensure a high-quality benchmark. An overview of the benchmark curation pipeline is shown in Fig. 3. + +Real Video Data Collection Real-world videos were sourced from datasets with rich spatiotemporal characteristics that ensured diverse motion and perspective variations. For egocentric data, we relied mainly on the Ego4D dataset [30], while most exocentric data points were collected from the DAVIS [64] and YouTube-VOS [85] datasets. To minimize confounders and to focus attention of VLM abilities to only spatiotemporal reasoning, we preprocessed the videos by temporally segmenting and centering them around the most relevant action, thus resulting in videos with an average duration of 3-8 seconds. This en + +sures that the key event described in the question is clear and reduces ambiguities or confounders that would reduce VLM accuracy. + +Synthetic Video Data Generation The rapid advancement of video generation techniques has led to their widespread application in diverse domains like robotics and immersive entertainment [4, 8, 47, 72, 82], making 4D reasoning on synthetic content essential for VLMs and underscoring the need to include such videos in our benchmark. For synthetic video generation, we use Cosmos [4] as our video generation backbone. To ensure that the generated videos align with the intended object moving directions, we incorporate input bounding boxes as additional spatial guidance. Specifically, we follow the approach introduced in [32] modifying the diffusion forward steps to enforce object localization constraints at each timestep, ensuring consistency between the generated object direction and the user-specified trajectory. The average duration of generated + +synthetic videos is 5 seconds. To maintain high-quality outputs, we perform a manual verification step after generation, filtering out low-quality videos and retaining only those that accurately match the specified directions. Once a video is generated, we use an LLM (GPT-4o) to create two types of evaluation questions: Directional questions, derived from the textual prompt used to generate the video; and False Positive (FP) questions (constituting \(10\%\), matching the ratio in the real dataset), which query non-existent objects in the scene. Both question types follow the format: "What direction is the \(\langle\) Object Name \rangle\) moving?", where the model must select one of four possible answers: "left", "right", "not moving", or "no \(\langle\) Object Name \rangle\) there". A final manual review is conducted to filter out or revise ambiguous questions, ensuring the quality of the questions and ground-truth answers. + +QA Generation and Quality Control Question-answer pairs are primarily constructed through human annotations. The question answer pairs are then supplemented with alternative answers by an LLM (GPT-4o) for multiple choice (MC) questions. To ensure high-quality annotations, we applied a three-round, multi-person cross-checked verification process, during which ambiguous videos were filtered out and vague, misleading, or incorrect QA pairs were refined to improve spatial and temporal alignment between language and visual content. Fig. 4 showcases some qualitative examples of annotations for different types of videos. + +Assessing Human Performance To establish a human performance baseline on our benchmark, we conducted an evaluation in which participants independently answered 100 randomly sampled questions from the dataset. The accuracy of human responses was then aggregated to approximate the performance of human spatiotemporal reasoning on the dataset. + +# 3.2. Categorizing Spatiotemporal Performance + +To systematically evaluate spatiotemporal reasoning capabilities, we first categorize videos into two primary groups: egocentric (first-person) videos and exocentric (third-person) videos. Egocentric videos are sourced from the Ego4D [30] dataset, where scenes are captured from a head-mounted camera, thus offering dynamic video data that is inherently coupled with the individual's actions. Exocentric videos encompass a diverse range of recorded scenes, from sports footage to everyday scenes. Beyond this categorization, we also evaluate spatiotemporal performance across four dimensions: translational movement (TM), rotational movement (RM), spatiotemporal counting (STM), and false positives (FP), with their proportions shown in Fig. 2. Translational movement assesses a model's ability to track linear motion within scenes, while rotational + +![](images/76d772259383b40099edee3486392a90b1c092299f408babd1e2bcca631b61d4.jpg) +Figure 5. Comparison of accuracy across types of spatiotemporal questions. Model accuracy is shown only for the six top-performing VLMs. + +movement assesses the understanding of changes in orientation and perspective shifts over time. Spatiotemporal counting extends these core motion-based tasks by requiring a more complex reasoning strategy to determine the number of objects performing a translation or rotational movement. Lastly, the false positives category evaluates the model's critical thinking [23] in determining whether an object or event actually occurred within the spatiotemporal context. By structuring the benchmark along these axes, we aim for a comprehensive framework for assessing spatiotemporal reasoning (Fig. 5). + +# 4. Evaluation of VLM4D Benchmark + +# 4.1. Evaluation Setup + +Benchmark Models We evaluate 23 most recently released VLMs thus covering a wide range of model sizes, architectures, and training methodologies. For closed-source VLMs, we evaluate GPT-4o [61], Gemini 2.5 Pro [28], Claude Sonnet 4 [6], and Grok-2-Vision [84]. For open-source models, we include Llama 4 [58], DeepSeek-VL [54], Qwen2.5-VL [86], Qwen2-VL [76], InternVL2.5 [15], Aria [43], InternVideo 2.5 [79], InternVideo2 [78], Phi-4-multimodal [2], Phi-3.5-vision [1], Pixtral [5], VideoLLama3 [91], Llava-One-Vision [42], Llava-NeXT-Video [94]. When available, we evaluate different sizes for each model, resulting in models ranging from 2 to 72 billion parameters. + +Evaluation Settings The evaluations were performed in a zero-shot setting with video or a set of sampled frames of + +
OrganizationModelReleaseRealSyntheticOverall
Ego-centricExo-centricAverageDirectionalFPAverage
User StudyHuman Performance99.699.799.795.810096.298.8
RandomRandom Selection24.423.223.625.524.725.424.1
+ +
Latest Proprietary VLMs
OpenAIGPT-4o2024-1155.562.260.049.553.349.957.5
GoogleGemini-2.5-Pro2025-664.662.963.554.880.057.362.0
AnthropicClaude-Sonnet-42025-552.652.152.244.086.748.351.3
xAIGrok-2-Vision2024-1248.849.749.449.366.751.049.8
+ +
Open-source Image VLMs
MetaLlama-4-Maverick-17B2025-452.654.353.853.351.153.053.6
Llama-4-Scout-17B2025-448.656.253.753.375.655.554.1
MicrosoftPhi-4-Multimodal2025-341.035.437.237.511.134.836.6
Phi-3.5-Vision2024-733.438.837.123.337.824.734.0
DeepSeekDeepSeek-VL22024-1233.632.933.131.846.733.333.2
Shanghai AI LabInternVL2.5-38B2024-1146.650.148.943.357.844.747.9
InternVL2.5-8B2024-1139.044.042.440.842.240.942.0
Mistral AIPixtral-12B2024-932.325.827.924.322.224.027.0
RhymesAria2024-1147.244.045.138.571.141.844.3
+ +
Open-source Video VLMs
AlibabaQwen2.5-VL-7B2025-142.343.743.343.564.445.643.8
Qwen2.5-VL-72B2025-154.352.553.149.580.052.653.0
Qwen2-VL-7B2024-836.134.735.240.535.640.036.3
Qwen2-VL-72B2024-948.143.044.640.873.344.044.5
DAMOVideoLLama3-2B2025-153.242.546.034.355.636.443.7
VideoLLama3-7B2025-149.445.146.542.853.343.845.9
Shanghai AI LabInternVideo2.5-8B2025-157.250.552.744.346.744.550.7
InternVideo2-8B2024-835.639.338.143.00.038.738.2
LLaVALLaVA-One-Vision-7B2024-936.835.636.037.835.637.536.3
LLaVA-NeXT-Video-34B2024-629.631.630.924.555.627.630.1
+ +Table 1. Evaluation on VLM4D Benchmark across various proprietary and open-source VLMs. Top three performers in each category are highlighted from dark (highest) to light (third highest). Human and random selection baselines are included for reference. + +video, followed by the prompt forming the input. For each model, we evaluate on two different inference settings. In the first setting, the model prompted to directly output (DO) the answer immediately without any reasoning, and in the second evaluation setting, the model is directed to create intermediate reasoning steps, chain-of-thought (CoT) [81], before inferring the final answer. + +Metrics Following prior work [87] and given the nature of our target task, we adopt multiple-choice questions (MCQs) for evaluation, using accuracy as the primary metric. For the two inference settings described earlier, we employ LLM-as-Judge [95] to assess the outputs of VLMs. We opt for this method instead of string or template matching, as VLMs—especially under chain of thought (CoT) prompt- + +ing—often generate all possible answer choices during reasoning, with varying frequencies and slight formatting differences. In some cases, the final answer may even contradict the reasoning. To better evaluate whether the model truly understands the video, we prompt two advanced LLMs (GPT-o3 and o4-mini) to grade based on the full CoT reasoning, not just the final answer. We then perform a cross-check between their judgments and manually resolve any disagreements. The evaluation results are reported in Tab. 1. + +# 4.2. Benchmark Results + +Our comprehensive evaluation on the VLM4D benchmark, detailed in Tab. 1, systematically assesses the spatiotemporal reasoning capabilities of modern Vision-Language Models (VLMs). The results highlight a clear hierar + +chy, with proprietary models demonstrating superior performance over their open-source counterparts. Google's Gemini-2.5-Pro emerges as the top-performing model with an overall accuracy of $62.0\%$ , followed by OpenAI's GPT4o at $57.5\%$ . Within the open-source domain, image-based models like Meta's Llama-4-Scout-17B (54.1%) and videosupported models like Alibaba's Qwen2.5-VL-72B (53.0%) demonstrate highly competitive results, even outperforming some proprietary counterparts. Despite these achievements, a significant performance gap persists when compared to human accuracy (98.8%), underscoring that sophisticated 4D awareness remains a formidable challenge for AI. Notably, performance varies significantly across different categories, such as real versus synthetic data and egocentric versus exo-centric perspectives, indicating that current models lack generalized spatiotemporal understanding. + +# 5. Analysis: Why VLMs Don't Work Well? + +# 5.1. Limited Spatiotemporal Cognition + +Despite significant advances in VLMs, their abilities to understand and reason about motion, spatial relationships, and temporal coherence remains fundamentally underdeveloped [13, 101]. Chain of thought (CoT) [81] is widely employed as a method to improve accuracy through step-by-step reasoning. We showcase a comparison between CoT and DO in Fig. 6. Overall, there is no indication of a large advantage of CoT over all evaluated models. Upon deeper exploration of the CoT reasoning of some models, we observe that the reasoning process was primarily flawed in the following ways: irrelevant information and arriving at conclusions that are inconsistent with the reasoning process. Larger models exhibited strategies that would be similar to how a human processes spatiotemporal information, but the resulting execution falls short of human performance. This demonstrates a disconnect between its visual and linguistic knowledge. We provide examples of this behavior in the supplementary material. + +# 5.2. Deficiencies in Spatiotemporal Labeling + +Another avenue of exploration we undertook is to understand the richness of spatiotemporal labels in popular supervised fine-tuning (SFT) VLM datasets. Typically, video captioning occurs at the 'scene' level, lacking fine-grained temporal, spatial, and object-level details. We performed an extensive analysis, encompassing over 2 million samples [14, 17, 44, 45, 94]. We performed this analysis through string-matching of spatiotemporal descriptors related to directionality, translational motion, rotation, and perspective shifts and provide the overall results in Fig. 7. We then performed a manual finegrained evaluation of the ShareGPT4Video dataset [14] which we found had the highest density of spatiotemporal dataset. We found that + +![](images/054262790f82a626f629e2cde3a06ae7d864a6f908397e350eca9a1f9b2152bd.jpg) +Figure 6. Comparison of CoT and DO Accuracy Across Models. Accuracy comparison between Chain-of-Thought (CoT) and Direct Output (DO) prompting across VLMs. + +![](images/af46013f16998694edd5a53b3093a61ae9c88455eb6bd649da3c78ece349e127.jpg) +Figure 7. Heatmap of Occurrences of Spatial-Temporal Terms in popular video SFT datasets. + +from a sample of 100 labels that were detected as spatiotemporal, less than $10\%$ of them were judged as accurate upon human evaluation. This result underscores the inadequacy of current dense captioning approaches, which frequently generate spatiotemporal descriptors without capturing precise motion dynamics. We provide more detailed analysis and explanations in the supplementary material. + +# 6. Probing Future Solutions + +To probe promising future solutions for enhancing spatiotemporal video understanding, we propose two approaches that address some of the shortcomings of current + +state-of-the-art VLMs: fine-tuning a VLM on data-rich in spatiotemporal actions and the other leveraging 4D reconstruction and feature fields jointly with a VLM. SFT refines the model's abilities by training on datasets that contain temporally and spatially rich actions and interactions. By integrating structured visual representations and targeted fine-tuning, these approaches enhance video-language models' ability to interpret motion. The second method lifts the feature space of VLMs into a temporally coherent 4D feature field, providing structured scene representations that improve motion and spatial reasoning in the stage of decoding and inference. + +Spatial-Temporal SFT We evaluate on a subset split of the real dataset by randomly splitting the real-world dataset into a training and testing split (80% / 20%) and we try settings using synthetic/real both for training. We conducted the experiments using Qwen 2VL (7B) and Qwen 2.5VL (7B) through LLama-Factory [96], and compared the performance before and after supervised fine-tuning in Tab. 2. The results demonstrated an improvement in accuracy in spatiotemporal reasoning, suggesting that performance gains can be obtained through targeted training. However, the addition of synthetic data does not necessarily increase performance over using real data alone, suggesting the importance of synthetic data quality. + +4D Feature Fields Reconstruction Recent advances in 3D/4D feature fields reconstruction methods [24, 37, 90, 97, 99] have significantly enhanced the vision foundation model's performance in 3D/4D space by integrating structured latent scene representations into the model's inference stage. Inspired by the promising results of feature lifting, we explore enhancing the InternVideo2-8B model [78] with spatiotemporal awareness by adopting the strategy proposed in Feature4X [99], which constructs the VLM's 2D feature space along the time dimension into a 4D feature field. To assess this approach, we evaluate performance on a subset of the VLM4D benchmark, specifically leveraging all 50 videos from the DAVIS 2016 dataset [63]. Our experimental evaluation compares inference performance across three distinct input modalities: original 2D videos; rendered novel global-view videos (which provide broader contextual information in 2D format); and reconstructed global feature fields (which implicitly incorporate 4D scene-level information during reasoning). Table 3 reveals that reconstructed feature fields achieve the highest accuracy across both reasoning types. This success stems from two key advantages: the inherent structure of 4D representations and the ability of feature field inference to avoid the rendering artifacts of RGB reconstruction (global view video). However, the current approach requires per-scene optimization as a post-processing step, limiting its generalizability and + +
ModelMC
Original Model
Qwen 2VL (7B)38.3
Qwen 2.5VL (7B)43.4
Finetuned Model
Qwen 2VL (7B) (R)54.5
Qwen 2VL (7B) (S)42.5
Qwen 2VL (7B) (R+S)55.5
Qwen 2.5VL (7B) (R)55.6
Qwen 2.5VL (7B) (S)48.6
Qwen 2.5VL (7B) (R+S)56.3
+ +Table 2. SFT on Spatial-Temporal Datasets. MC refers to the performance of multiple-choice, while R, S, and $\mathbb{R} + \mathbb{S}$ denote real, synthetic, and both real+synthetic usage of data for fine-tuning. + +
Input ModalityAccuracy
Chain of Thought Response
Original 2D Video36.0
Global View Video32.7
Global Feature Field37.4
Direct Output Response
Original 2D Video24.3
Global View Video23.8
Global Feature Field29.0
+ +Table 3. InternVideo2 Accuracy with 4D Feature Field Reconstruction. Comparison of InternVideo2's performance given different input modalities from the same subset. + +making it computationally intensive. + +# 7. Conclusion + +Through the construction of the VLM4D benchmark, we evaluate the spatiotemporal reasoning capabilities of various Vision-Language Models (both open-source and proprietary). While more recently released models demonstrate improved performance over their counterparts, they remain significantly behind human proficiency. Overall, our work questions whether VLMs possess spatiotemporal reasoning abilities that are imperative to have for more sophisticated visual agents in fields ranging from robotics to interactive AI systems that require a deep understanding of dynamic visual environments. We hope to inspire future work to explore novel approaches for integrating spatiotemporal grounding, thereby enhancing their spatiotemporal reasoning capabilities and facilitating robust deployment. + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024.5 +[2] Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024. 5 +[3] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Weijian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 3 +[4] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3, 4 +[5] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 5 +[6] Anthropic. System card: Claude opus 4 & claude sonnet 4. Technical report, 2025. 5 +[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3 +[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 4 +[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. + +Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[10] Neil Burgess. Spatial memory: how egocentric and allocentric combine. Trends in Cognitive Sciences, 10(12):551-557, 2006. 2 +[11] Pradyumna Chari, Yunhao Ba, Shijie Zhou, Chinmay Talegaonkar, Shreeram Athreya, and Achuta Kadambi. On learning mechanical laws of motion from video using neural networks. IEEE Access, 11:30129-30145, 2023. 3 +[12] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18030-18040, 2022. 2 +[13] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.7 +[14] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4, 7 +[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5 +[16] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024.3 +[17] Erfei Cui, Yinan He, Zheng Ma, Zhe Chen, Hao Tian, Weiyun Wang, Kunchang Li, Yi Wang, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, Yali Wang, Limin Wang, Yu Qiao, and Jifeng Dai. Sharegpt-4o: Comprehensive multimodal annotations with gpt-4o, 2024. 7 +[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3 +[19] Julian De Freitas, Nicholas E. Myers, and Anna C. Nobre. Tracking the changing feature of a moving object. Journal of Vision, 16(3):22, 2016. 2 +[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 3 +[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2 +[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3 +[23] Chenrui Fan, Ming Li, Lichao Sun, and Tianyi Zhou. Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? arXiv preprint arXiv:2504.06514, 2025. 5 +[24] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in neural information processing systems, 37:40212-40229, 2024. 8 +[25] Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Huaizhi Qu, Dilin Wang, Zhicheng Yan, et al. Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction. arXiv preprint arXiv:2505.20279, 2025. 3 +[26] Jennifer J. Freyd and Ronald A. Finke. Representational momentum. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10(1):126-132, 1984. 2 +[27] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3 +[28] Google Gemini Team. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Technical report, 2025. 5 +[29] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 3 +[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 4, 5 +[31] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multidiscipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3 +[32] Xuehai He, Shuohang Wang, Jianwei Yang, Xiaoxia Wu, Yiping Wang, Kuan Wang, Zheng Zhan, Olatunjri Ruwase, Yelong Shen, and Xin Eric Wang. Mojito: Motion tra + +jctory and intensity control for video generation. arXiv preprint arXiv: 2412.08948, 2024. 3, 4 +[33] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3 +[34] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 2 +[35] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. How good is my video lmm? complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 3 +[36] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3 +[37] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in neural information processing systems, 35:23311-23330, 2022. 8 +[38] Ivan Laptev. On space-time interest points. International Journal of Computer Vision, 64(2-3):107-123, 2005. 3 +[39] Yann LeCun, Bernhard Boser, John S. Denker, Donnie Henderson, Richard E. Howard, Wayne Hubbard, and Lawrence D. Jackel. Backpropagation applied to handwritten zip code recognition. *Neural Computation*, 1(4):541-551, 1989. 2 +[40] Alan M. Leslie. Spatiotemporal continuity and the perception of causality in infants. Perception, 13(3):287-305, 1984. 2 +[41] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3 +[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5 +[43] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 5 +[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 7 +[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark, 2023. 7 +[46] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. + +Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3 +[47] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, and Zhiwen Fan. 4k4DGen: Panoramic 4d generation at 4k resolution. In The Thirteenth International Conference on Learning Representations, 2025. 4 +[48] Xinhao Li, Zhenpeng Huang, Jing Wang, Kunchang Li, and Limin Wang. Videoeval: Comprehensive benchmark suite for low-cost evaluation of video foundation model. arXiv preprint arXiv:2407.06491, 2024. 3 +[49] Lu Ling, Chen-Hsuan Lin, Tsung-Yi Lin, Yifan Ding, Yu Zeng, Yichen Sheng, Yunhao Ge, Ming-Yu Liu, Aniket Bera, and Zhaoshuo Li. Scenethesis: A language and vision agentic framework for 3d scene generation. arXiv preprint arXiv:2505.02836, 2025. 3 +[50] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 3 +[51] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint, 2024. 2 +[52] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv preprint arXiv:2402.08268, 2024.3 +[53] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3 +[54] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5 +[55] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Neural Information Processing Systems, 2019. 2 +[56] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3 +[57] D. Marr and S. Ullman. Directional selectivity and its use in early visual processing. Proceedings of the Royal Society of London. Series B, Biological Sciences, 211(1183):151-180, 1981. 2 +[58] Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. Technical report, 2025. 5 +[59] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluat + +ing video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3 +[60] Basak Melis Öcal, Maxim Tatarchenko, Sezer Karaoglu, and Theo Gevers. Sceneteller: Language-to-3d scene generation. In European Conference on Computer Vision, pages 362–378. Springer, 2024. 3 +[61] OpenAI. Hello gpt-4o. Technical report, 2024. 5 +[62] Shivansh Patel, Xinchen Yin, Wenlong Huang, Shubham Garg, Hooshang Nayyeri, Li Fei-Fei, Svetlana Lazebnik, and Yunzhu Li. A real-to-sim-to-real approach to robotic manipulation with vlm-generated iterative keypoint rewards. arXiv preprint arXiv:2502.08643, 2025. 3 +[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 8 +[64] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 4 +[65] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3 +[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2 +[67] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems (NIPS), pages 568-576, 2014. 3 +[68] Elizabeth S. Spelke and Katherine D. Kinzler. Core knowledge. Developmental Science, 10(1):89-96, 2007. 2 +[69] Alessandro Suglia, Claudio Greco, Katie Baker, Jose L Part, Ioannis Papaioannou, Arash Eshghi, Ioannis Konstas, and Oliver Lemon. Alanavlm: A multimodal embodied ai foundation model for egocentric video understanding. arXiv preprint arXiv:2406.13807, 2024. 3 +[70] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3 +[71] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[72] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 4 + +[73] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 3 +[74] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Action recognition by dense trajectories. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3169-3176, 2011. 3 +[75] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3 +[76] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5 +[77] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2 +[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Intervideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 5, 8 +[79] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 5 +[80] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 3 +[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 6, 7 +[82] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26057-26068, 2025. 4 +[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2024. 3 +[84] xAI. Grok-2 beta release. Technical report, 2024. 5 +[85] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, + +and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018. 2, 4 +[86] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5 +[87] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. arXiv preprint arXiv:2412.14171, 2024. 3, 6 +[88] Yuqian Yuan, Hang Zhang, Wentong Li, Zesen Cheng, Boqiang Zhang, Long Li, Xin Li, Deli Zhao, Wenqiao Zhang, Yueting Zhuang, et al. Videorefer suite: Advancing spatial-temporal object understanding with video llm. arXiv preprint arXiv:2501.00599, 2024.3 +[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3 +[90] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In European Conference on Computer Vision, pages 57-74. Springer, 2024. 8 +[91] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 3, 5 +[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3 +[93] Hongxin Zhang, Zeyuan Wang, Qiushi Lyu, Zheyuan Zhang, Sunli Chen, Tianmin Shu, Yilun Du, and Chuang Gan. Combo: compositional world models for embodied multi-agent cooperation. arXiv preprint arXiv:2404.10775, 2024. 3 +[94] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 5, 7 +[95] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 3, 6 +[96] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of $100+$ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand, 2024. Association for Computational Linguistics. 8 + +[97] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 8 +[98] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2024. 3 +[99] Shijie Zhou, Hui Ren, Yijia Weng, Shuwang Zhang, Zhen Wang, Dejia Xu, Zhiwen Fan, Suya You, Zhangyang Wang, Leonidas Guibas, et al. Feature4x: Bridging any monocular video to 4d agentic ai with versatile gaussian feature fields. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14179-14190, 2025. 8 +[100] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 +[101] Orr Zohar, Xiaohan Wang, Yann Dubois, Nikhil Mehta, Tong Xiao, Philippe Hansen-Estruch, Licheng Yu, Xiaofang Wang, Felix Juefei-Xu, Ning Zhang, et al. Apollo: An exploration of video understanding in large multimodal models. arXiv preprint arXiv:2412.10360, 2024. 7 \ No newline at end of file diff --git a/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/images.zip b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..09cda3d256b2db3454471c6109f7cebb0b656431 --- /dev/null +++ b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fedd28fd29a55c72e6e1071023adf89badce6a8ec54fad64c4b5e954dc5fac6c +size 753162 diff --git a/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/layout.json b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf85e9a7d9a1b81b48672ddedb9e7c7960eca37 --- /dev/null +++ b/2025/VLM4D_ Towards Spatiotemporal Awareness in Vision Language Models/layout.json @@ -0,0 +1,8295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 79, + 103, + 532, + 121 + ], + "type": "text", + "content": "VLM4D: Towards Spatiotemporal Awareness in Vision Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "spans": [ + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "text", + "content": "Shijie Zhou" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "text", + "content": " Alexander Vilesov" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "text", + "content": " Xuehai He" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "inline_equation", + "content": "^{2,3*}" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "text", + "content": " Ziyu Wan" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "text", + "content": " Shuwang Zhang" + }, + { + "bbox": [ + 100, + 142, + 503, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "spans": [ + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "text", + "content": "Aditya Nagachandra" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "text", + "content": " Di Chang" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "text", + "content": " Dongdong Chen" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "text", + "content": " Xin Eric Wang" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "text", + "content": " Achuta Kadambi" + }, + { + "bbox": [ + 86, + 159, + 523, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "spans": [ + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "text", + "content": "UCLA " + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "inline_equation", + "content": "{ }^{2}" + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "text", + "content": "Microsoft " + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "inline_equation", + "content": "{ }^{3}" + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "text", + "content": "UCSC " + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "inline_equation", + "content": "{ }^{4}" + }, + { + "bbox": [ + 220, + 178, + 390, + 191 + ], + "type": "text", + "content": "USC" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 216, + 196, + 390, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 196, + 390, + 209 + ], + "spans": [ + { + "bbox": [ + 216, + 196, + 390, + 209 + ], + "type": "text", + "content": "https://v1m4d.github.io/" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 216, + 549, + 300 + ], + "blocks": [ + { + "bbox": [ + 60, + 216, + 549, + 300 + ], + "lines": [ + { + "bbox": [ + 60, + 216, + 549, + 300 + ], + "spans": [ + { + "bbox": [ + 60, + 216, + 549, + 300 + ], + "type": "image", + "image_path": "47589e52e181e116a7dc2dff9f94b43df83d71483d046e3a92896847c40c8cd1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 61, + 300, + 549, + 487 + ], + "blocks": [ + { + "bbox": [ + 61, + 300, + 549, + 487 + ], + "lines": [ + { + "bbox": [ + 61, + 300, + 549, + 487 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 549, + 487 + ], + "type": "image", + "image_path": "995244dc3b4f1d01690e042e2a6374a85d43bdee2f33aa3c9bd1f7094c77dd60.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 495, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 555, + 552 + ], + "type": "text", + "content": "Figure 1. Spatiotemporal (4D) Awareness. Humans intuitively reason in 4D (3D space + time), effortlessly reconstructing the dynamic spatial trajectory of moving objects from any perspective. In contrast, current Vision Language Models (VLMs) typically rely on aggregating 2D visual features across time, leading to incorrect predictions when motion understanding and interpretation requires deeper spatiotemporal reasoning. In this example, humans correctly perceive the car moving to the right, while the VLM (GPT-4o) inaccurately predicts leftward movement, suggesting VLMs struggle to perform spatiotemporal reasoning." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 559, + 200, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 559, + 200, + 572 + ], + "spans": [ + { + "bbox": [ + 151, + 559, + 200, + 572 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 585, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 296, + 693 + ], + "type": "text", + "content": "Vision language models (VLMs) have shown remarkable capabilities in integrating linguistic and visual reasoning but remain fundamentally limited in understanding dynamic spatiotemporal interactions. Humans effortlessly track and reason about object movements, rotations, and perspective shifts—abilities essential for robust dynamic real-world understanding yet notably lacking in current VLMs. In this paper, we introduce VLM4D, the first benchmark specifically designed to evaluate the spatiotemporal reasoning capa" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 560, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 555, + 704 + ], + "type": "text", + "content": "bilities of VLMs. Our benchmark comprises diverse real-world and synthetic videos accompanied by carefully curated question-answer pairs emphasizing translational and rotational motions, perspective awareness, and motion continuity. Through comprehensive evaluations of state-of-the-art open and closed-source VLMs, we identify significant performance gaps compared to human baselines, highlighting fundamental deficiencies in existing models. Extensive analysis reveals that VLMs struggle particularly with integrating multiple visual cues and maintaining temporal coherence. We further explore promising directions, such as leveraging 4D feature field reconstruction and targeted" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8600" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 144 + ], + "type": "text", + "content": "spatiotemporal supervised fine-tuning, demonstrating their effectiveness in enhancing spatiotemporal comprehension. Our work aims to encourage deeper exploration into improving VLMs' spatial and temporal grounding, paving the way towards more capable and reliable visual intelligence for dynamic environments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 136, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 136, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 136, + 178 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 330 + ], + "type": "text", + "content": "Humans possess an innate ability to perceive, track, and interpret motion as well as spatial and temporal changes, enabling rich interpretations of complex dynamic events from both egocentric and allocentric perspectives [10, 19, 57]. When observing an object move, we can inherently process any changes such as lateral shifts, rotational directions, and periodic or repeated actions unfolding along a specific trajectory [10]. These sophisticated perceptual abilities are a product of our spatiotemporal cognition [26], and form an essential foundation that allows us to comprehend and reason about physical phenomena, object interactions, and causal relationships within our environment [34, 40]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 521 + ], + "type": "text", + "content": "Vision language models (VLMs), which also have the potential to perceive motion and spatiotemporal changes in videos, represent a prominent class of methods aimed at emulating or surpassing human capabilities in integrated visual and linguistic reasoning [21, 39]. While previous work on VLMs has primarily focused on static visual understanding—through large-scale training on paired language and image data [66]—or explored video understanding tasks such as captioning [55] and scene understanding [12], we find that their strong performance in these tasks does not naturally translate into robust spatiotemporal reasoning. This limitation is particularly striking given that state-of-the-art VLMs are typically trained on datasets comprising hundreds of billions of tokens [51]. In contrast, human infants naturally develop robust spatiotemporal cognition within the first few months of life [68]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Another key challenge that limits VLM performance on spatiotemporal tasks is the need to implicitly or explicitly reconstruct a four-dimensional (4D) representation—3D space + time—of dynamic scenes, and subsequently reason over this reconstruction [77]. As illustrated in Fig. 1, the car is advancing forwards and turning to the left in its own frame of reference. However, from the camera's perspective, its motion appears as a combination of heading to the right and receding into the distance despite the car being in the center of the frame due to camera rotation. Human observers can seamlessly disentangle these complex dynamics, accurately interpreting trajectories by synthesizing diverse visual cues including camera rotation compensation, stationary scene landmarks, prior knowledge of 3D and 4D environmental structures, and perspective projections [10, 26, 40, 57]. The inability of current VLMs to in-" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 547, + 207 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 547, + 207 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 547, + 207 + ], + "type": "image", + "image_path": "f0aa1aaa38581fffceb31eee1e39292dac7e0c3be047929474dd30272a6c38b8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 216, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. Distribution of Dataset Sources and Annotations. Overview of the dataset composition, illustrating the proportions of real third-person (exo-centric) videos (DAVIS [64], YouTube-VOS [85]), real first-person (ego-centric) videos (Ego4D [30]), and synthetic videos (Cosmos [4]). The real video data is further categorized by annotation types, including translational, rotational, action, counting, and false positive queries (targeting nonexistent events to assess critical reasoning)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 322, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 429 + ], + "type": "text", + "content": "tegrate spatial, temporal, and semantic cues in a human-like manner stems from their fundamentally different encoding paradigms for visual and language information. This discrepancy highlights a significant gap between human and machine spatiotemporal understanding, suggesting that future VLMs may benefit from insights in cognitive science and neuroscience to develop more advanced mechanisms for perceiving, integrating, reconstructing, and reasoning over dynamic scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 431, + 556, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 556, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 556, + 539 + ], + "type": "text", + "content": "In order to effectively characterize and challenge the existing spatiotemporal reasoning abilities of VLMs, we introduce VLM4D, a rigorous benchmark specifically designed to probe the spatiotemporal grounding capabilities of current vision language models. Through this contribution, we aim to catalyze research that addresses the critical gap in spatiotemporal understanding and reasoning within VLMs and provide a foundational analysis highlighting key deficiencies in existing models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 541, + 533, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 541, + 533, + 553 + ], + "spans": [ + { + "bbox": [ + 326, + 541, + 533, + 553 + ], + "type": "text", + "content": "We summarize our main contributions as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 555, + 554, + 711 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 315, + 555, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 554, + 602 + ], + "type": "text", + "content": "1. We propose the first benchmark VLM4D explicitly designed to rigorously evaluate the spatiotemporal (4D) reasoning capabilities of Vision Language Models (VLMs)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 604, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 604, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 604, + 553, + 651 + ], + "type": "text", + "content": "2. We introduce a novel, meticulously curated dataset consisting of diverse real-world and synthetic video sequences paired with carefully crafted spatiotemporal question-answer (QA) annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 651, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 651, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 314, + 651, + 553, + 711 + ], + "type": "text", + "content": "3. We analyze the critical limitations of contemporary VLMs in spatiotemporal reasoning and conduct experiments exploring potential solutions, highlighting fundamental challenges and outlining clear directions for impactful future research." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8601" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 155 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 155 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 155 + ], + "type": "image", + "image_path": "17ed82081f03e0cfa98e209ce96212381a6252413b1a8496f303a884553cbe8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 161, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 54, + 161, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 54, + 161, + 555, + 206 + ], + "type": "text", + "content": "Figure 3. Dataset Generation and Annotation Pipeline. Our dataset was constructed by collecting real videos and generating synthetic data, followed by human-in-the-loop quality reviews to address ambiguous videos and annotations. After temporal alignment and quality assurance, human-annotated questions and ground-truth answers were created, complemented by multiple-choice (MC) answers generated by large language models (LLMs). The final dataset includes real and synthetic video data with comprehensive VLM scoring metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 214, + 141, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 141, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 141, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 236, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 296, + 619 + ], + "type": "text", + "content": "Spatiotemporal Understanding in Vision Language Models Early methods for video understanding before the advent of large vision language models (VLMs) leveraged trajectory-based representations and separate modeling of spatial and motion cues for tasks like action and motion recognition [11, 38, 67, 74]. Recently, VLMs have evolved rapidly by fully leveraging the significant achievements of Large Language Models (LLMs) [7, 9, 20, 65, 71, 80] and large-scale visual instruction tuning datasets [18, 50, 100]. While VLMs [3, 29, 33, 42, 50, 70, 76, 100] exhibit transformative potential for applications such as embodied AI [22, 36, 69], robotics [62, 73], scene generation [49, 60, 98], and world modeling [52, 93], most existing methods remain constrained to static images, focusing narrowly on spatial understanding [25] while overlooking the dynamic temporal dimension inherent in real-world interactions. To bridge this gap, emerging research [16, 44, 56, 91, 92] has begun exploring video modality integration, aiming to equip VLMs with spatial-temporal awareness critical for tasks like video comprehension, where both contextual details and motion dynamics are essential. For example, VideoLLM-MoD [83] proposes to address the efficiency issue when processing long-term video by mixture-of-depths. [88] introduces VideoRefer to enhance the finer-level (like object-level) spatial-temporal video understanding of VLMs. Grounded-VideoLLM [75] also targets for fine-grained video understanding through incorporating an additional temporal stream. In this work, we aim to rigorously evaluate the 4D spatial-temporal reasoning capabilities of state-of-the-art VLMs, probing how and to what extent these models internalize spatial intelligence and temporal dependencies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "VLM Benchmarks Following the development trends of VLMs, benchmarking VLMs shares the similar trajectory by first evaluating vision QA on static images [31, 41, 53, 89], to align with models' early focus on 2D understanding. As VLMs evolved to tackle dynamic scenarios, benchmarks expanded to evaluate general-purpose video comprehension" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 216, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 479 + ], + "type": "text", + "content": "tasks that probe temporal coherence and event understanding [27, 35, 46, 48, 59]. Notably, MMVU [95] further proposes a knowledge-intensive benchmark to assess the expert-level reasoning ability of current video-based large models. However, while these works assess perception and semantic understanding, they largely overlook the explicit evaluation of spatial-temporal awareness, a core capability for real-world applications requiring 4D (3D space + time) reasoning. Recent efforts like [87] pioneer benchmarks for 3D visual-spatial intelligence but restrict evaluation to static 3D scene, neglecting the interplay of object motion and temporal dynamics intrinsic to videos. In this work, we introduce VLM4D, the first benchmark designed to holistically evaluate the 4D intelligence in VLMs, unifying spatial understanding, temporal continuity, and motion reasoning. By curating tasks that demand precise analysis of dynamic interactions (e.g., direction prediction, perspective anticipation, and motion reasoning), VLM4D exposes critical gaps in current models' ability to internalize spatiotemporal relationships. Our work not only advances the granularity of VLM evaluation but also shares insights and potential solutions to improve the model performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 489, + 451, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 451, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 451, + 502 + ], + "type": "text", + "content": "3. The VLM4D Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 509, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 666 + ], + "type": "text", + "content": "We introduce VLM4D, the first benchmark specifically designed to test the spatiotemporal reasoning ability of VLMs. VLM4D consists of 1000 videos paired with over 1800 question-answer pairs, each carefully designed to assess both spatial and temporal understanding jointly. The majority of these videos are sourced from datasets with rich spatiotemporal characteristics, thus ensuring a diverse range of motion-related scenarios. We also augment the dataset with synthetic videos generated by a world-foundation model, Cosmos [4], that has been modified using techniques introduced in [32] to obtain more accurate correspondence between motion-oriented prompts and the resulting generated video. Fig. 2 illustrates the composition of our dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 672, + 456, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 456, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 456, + 684 + ], + "type": "text", + "content": "3.1. Benchmark Construction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Unlike prior work that often relies heavily on LLMs and VLMs to generate first iterations of benchmarks and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8602" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 126, + 80, + 416, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 80, + 416, + 92 + ], + "spans": [ + { + "bbox": [ + 126, + 80, + 416, + 92 + ], + "type": "text", + "content": "Is the camel in the foreground turning to the left or right from its own perspective?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 94, + 181, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 94, + 181, + 105 + ], + "spans": [ + { + "bbox": [ + 126, + 94, + 181, + 105 + ], + "type": "text", + "content": "A. Not moving" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 209, + 94, + 236, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 94, + 236, + 104 + ], + "spans": [ + { + "bbox": [ + 209, + 94, + 236, + 104 + ], + "type": "text", + "content": "B.Left" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 267, + 94, + 377, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 94, + 377, + 105 + ], + "spans": [ + { + "bbox": [ + 267, + 94, + 377, + 105 + ], + "type": "text", + "content": "C. Moving straight, not turning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 402, + 94, + 433, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 94, + 433, + 105 + ], + "spans": [ + { + "bbox": [ + 402, + 94, + 433, + 105 + ], + "type": "text", + "content": "D. Right" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 108, + 112, + 502, + 190 + ], + "blocks": [ + { + "bbox": [ + 108, + 112, + 502, + 190 + ], + "lines": [ + { + "bbox": [ + 108, + 112, + 502, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 112, + 502, + 190 + ], + "type": "image", + "image_path": "26d1f9f11602467ade67070edee943afe3d6a423a568579116b1b6428c5aefaa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 202, + 411, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 202, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 126, + 202, + 411, + 213 + ], + "type": "text", + "content": "In which direction of rotation does the person pour the batter into the frying pan?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 215, + 205, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 215, + 205, + 225 + ], + "spans": [ + { + "bbox": [ + 127, + 215, + 205, + 225 + ], + "type": "text", + "content": "A. Counter-clockwise" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 223, + 216, + 277, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 216, + 277, + 225 + ], + "spans": [ + { + "bbox": [ + 223, + 216, + 277, + 225 + ], + "type": "text", + "content": "B. Left to right" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 216, + 358, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 358, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 358, + 225 + ], + "type": "text", + "content": "C. Right to left" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 395, + 215, + 444, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 215, + 444, + 225 + ], + "spans": [ + { + "bbox": [ + 395, + 215, + 444, + 225 + ], + "type": "text", + "content": "D. Clockwise" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 112, + 232, + 503, + 312 + ], + "blocks": [ + { + "bbox": [ + 112, + 232, + 503, + 312 + ], + "lines": [ + { + "bbox": [ + 112, + 232, + 503, + 312 + ], + "spans": [ + { + "bbox": [ + 112, + 232, + 503, + 312 + ], + "type": "image", + "image_path": "7d1d22256b012a8a1c29c9d1d0fafe9a710fc30405fcffc07bef763ae9f8908d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 322, + 306, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 322, + 306, + 333 + ], + "spans": [ + { + "bbox": [ + 127, + 322, + 306, + 333 + ], + "type": "text", + "content": "What direction is the robotic dog moving towards?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 335, + 160, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 335, + 160, + 346 + ], + "spans": [ + { + "bbox": [ + 127, + 335, + 160, + 346 + ], + "type": "text", + "content": "A. Right" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 192, + 336, + 274, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 336, + 274, + 346 + ], + "spans": [ + { + "bbox": [ + 192, + 336, + 274, + 346 + ], + "type": "text", + "content": "B. no robotic dog there" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 336, + 358, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 358, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 358, + 346 + ], + "type": "text", + "content": "C. Not moving" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 389, + 336, + 416, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 336, + 416, + 346 + ], + "spans": [ + { + "bbox": [ + 389, + 336, + 416, + 346 + ], + "type": "text", + "content": "D.Left" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 111, + 353, + 503, + 430 + ], + "blocks": [ + { + "bbox": [ + 111, + 353, + 503, + 430 + ], + "lines": [ + { + "bbox": [ + 111, + 353, + 503, + 430 + ], + "spans": [ + { + "bbox": [ + 111, + 353, + 503, + 430 + ], + "type": "image", + "image_path": "119eaa33d817ce107020693246e0587f027ae636191a9d4d9ef5e3c81143d590.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "type": "text", + "content": "Figure 4. Qualitative Examples of Dataset Annotations. (Top) A third-person (exo-centric) video with translational annotations (\"camel turning left from its perspective\"). (Middle) A first-person (ego-centric) video with a rotational question (\"clockwise rotation of ladle\"). (Bottom) A synthetic scene with motion recognition \"robotic dog moving left\"." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 482, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 567 + ], + "type": "text", + "content": "datasets [14] followed by human quality control, we found that existing VLMs and automated methods showed significant limitations in terms of reliability and quality. This shortcoming necessitated direct human annotations that were then followed by augmentation by LLMs to ensure a high-quality benchmark. An overview of the benchmark curation pipeline is shown in Fig. 3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Real Video Data Collection Real-world videos were sourced from datasets with rich spatiotemporal characteristics that ensured diverse motion and perspective variations. For egocentric data, we relied mainly on the Ego4D dataset [30], while most exocentric data points were collected from the DAVIS [64] and YouTube-VOS [85] datasets. To minimize confounders and to focus attention of VLM abilities to only spatiotemporal reasoning, we preprocessed the videos by temporally segmenting and centering them around the most relevant action, thus resulting in videos with an average duration of 3-8 seconds. This en" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 483, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 519 + ], + "type": "text", + "content": "sures that the key event described in the question is clear and reduces ambiguities or confounders that would reduce VLM accuracy." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "type": "text", + "content": "Synthetic Video Data Generation The rapid advancement of video generation techniques has led to their widespread application in diverse domains like robotics and immersive entertainment [4, 8, 47, 72, 82], making 4D reasoning on synthetic content essential for VLMs and underscoring the need to include such videos in our benchmark. For synthetic video generation, we use Cosmos [4] as our video generation backbone. To ensure that the generated videos align with the intended object moving directions, we incorporate input bounding boxes as additional spatial guidance. Specifically, we follow the approach introduced in [32] modifying the diffusion forward steps to enforce object localization constraints at each timestep, ensuring consistency between the generated object direction and the user-specified trajectory. The average duration of generated" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8603" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": "synthetic videos is 5 seconds. To maintain high-quality outputs, we perform a manual verification step after generation, filtering out low-quality videos and retaining only those that accurately match the specified directions. Once a video is generated, we use an LLM (GPT-4o) to create two types of evaluation questions: Directional questions, derived from the textual prompt used to generate the video; and False Positive (FP) questions (constituting \\(10\\%\\), matching the ratio in the real dataset), which query non-existent objects in the scene. Both question types follow the format: \"What direction is the \\(\\langle\\) Object Name \\rangle\\) moving?\", where the model must select one of four possible answers: \"left\", \"right\", \"not moving\", or \"no \\(\\langle\\) Object Name \\rangle\\) there\". A final manual review is conducted to filter out or revise ambiguous questions, ensuring the quality of the questions and ground-truth answers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 411 + ], + "type": "text", + "content": "QA Generation and Quality Control Question-answer pairs are primarily constructed through human annotations. The question answer pairs are then supplemented with alternative answers by an LLM (GPT-4o) for multiple choice (MC) questions. To ensure high-quality annotations, we applied a three-round, multi-person cross-checked verification process, during which ambiguous videos were filtered out and vague, misleading, or incorrect QA pairs were refined to improve spatial and temporal alignment between language and visual content. Fig. 4 showcases some qualitative examples of annotations for different types of videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "content": "Assessing Human Performance To establish a human performance baseline on our benchmark, we conducted an evaluation in which participants independently answered 100 randomly sampled questions from the dataset. The accuracy of human responses was then aggregated to approximate the performance of human spatiotemporal reasoning on the dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 277, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 277, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 277, + 529 + ], + "type": "text", + "content": "3.2. Categorizing Spatiotemporal Performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "To systematically evaluate spatiotemporal reasoning capabilities, we first categorize videos into two primary groups: egocentric (first-person) videos and exocentric (third-person) videos. Egocentric videos are sourced from the Ego4D [30] dataset, where scenes are captured from a head-mounted camera, thus offering dynamic video data that is inherently coupled with the individual's actions. Exocentric videos encompass a diverse range of recorded scenes, from sports footage to everyday scenes. Beyond this categorization, we also evaluate spatiotemporal performance across four dimensions: translational movement (TM), rotational movement (RM), spatiotemporal counting (STM), and false positives (FP), with their proportions shown in Fig. 2. Translational movement assesses a model's ability to track linear motion within scenes, while rotational" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 552, + 281 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 552, + 281 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 552, + 281 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 552, + 281 + ], + "type": "image", + "image_path": "76d772259383b40099edee3486392a90b1c092299f408babd1e2bcca631b61d4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 291, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 324 + ], + "type": "text", + "content": "Figure 5. Comparison of accuracy across types of spatiotemporal questions. Model accuracy is shown only for the six top-performing VLMs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 338, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 469 + ], + "type": "text", + "content": "movement assesses the understanding of changes in orientation and perspective shifts over time. Spatiotemporal counting extends these core motion-based tasks by requiring a more complex reasoning strategy to determine the number of objects performing a translation or rotational movement. Lastly, the false positives category evaluates the model's critical thinking [23] in determining whether an object or event actually occurred within the spatiotemporal context. By structuring the benchmark along these axes, we aim for a comprehensive framework for assessing spatiotemporal reasoning (Fig. 5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 480, + 500, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 500, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 500, + 493 + ], + "type": "text", + "content": "4. Evaluation of VLM4D Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 500, + 419, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 419, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 419, + 514 + ], + "type": "text", + "content": "4.1. Evaluation Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 518, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 673 + ], + "type": "text", + "content": "Benchmark Models We evaluate 23 most recently released VLMs thus covering a wide range of model sizes, architectures, and training methodologies. For closed-source VLMs, we evaluate GPT-4o [61], Gemini 2.5 Pro [28], Claude Sonnet 4 [6], and Grok-2-Vision [84]. For open-source models, we include Llama 4 [58], DeepSeek-VL [54], Qwen2.5-VL [86], Qwen2-VL [76], InternVL2.5 [15], Aria [43], InternVideo 2.5 [79], InternVideo2 [78], Phi-4-multimodal [2], Phi-3.5-vision [1], Pixtral [5], VideoLLama3 [91], Llava-One-Vision [42], Llava-NeXT-Video [94]. When available, we evaluate different sizes for each model, resulting in models ranging from 2 to 72 billion parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Evaluation Settings The evaluations were performed in a zero-shot setting with video or a set of sampled frames of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8604" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 553, + 131 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 553, + 131 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 553, + 131 + ], + "type": "table", + "html": "
OrganizationModelReleaseRealSyntheticOverall
Ego-centricExo-centricAverageDirectionalFPAverage
User StudyHuman Performance99.699.799.795.810096.298.8
RandomRandom Selection24.423.223.625.524.725.424.1
", + "image_path": "8ccdf57392717b9801d88d92aa93772a0982fe565babcccdf575603b63dece30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 133, + 552, + 200 + ], + "blocks": [ + { + "bbox": [ + 62, + 133, + 552, + 200 + ], + "lines": [ + { + "bbox": [ + 62, + 133, + 552, + 200 + ], + "spans": [ + { + "bbox": [ + 62, + 133, + 552, + 200 + ], + "type": "table", + "html": "
Latest Proprietary VLMs
OpenAIGPT-4o2024-1155.562.260.049.553.349.957.5
GoogleGemini-2.5-Pro2025-664.662.963.554.880.057.362.0
AnthropicClaude-Sonnet-42025-552.652.152.244.086.748.351.3
xAIGrok-2-Vision2024-1248.849.749.449.366.751.049.8
", + "image_path": "47c94cb9ac9dcb707ee35de3809525e665f345bddf3443567e6da53dcb91be2f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 62, + 202, + 552, + 328 + ], + "blocks": [ + { + "bbox": [ + 62, + 202, + 552, + 328 + ], + "lines": [ + { + "bbox": [ + 62, + 202, + 552, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 552, + 328 + ], + "type": "table", + "html": "
Open-source Image VLMs
MetaLlama-4-Maverick-17B2025-452.654.353.853.351.153.053.6
Llama-4-Scout-17B2025-448.656.253.753.375.655.554.1
MicrosoftPhi-4-Multimodal2025-341.035.437.237.511.134.836.6
Phi-3.5-Vision2024-733.438.837.123.337.824.734.0
DeepSeekDeepSeek-VL22024-1233.632.933.131.846.733.333.2
Shanghai AI LabInternVL2.5-38B2024-1146.650.148.943.357.844.747.9
InternVL2.5-8B2024-1139.044.042.440.842.240.942.0
Mistral AIPixtral-12B2024-932.325.827.924.322.224.027.0
RhymesAria2024-1147.244.045.138.571.141.844.3
", + "image_path": "8f7c942fa6def8f2d53a19dc72218b876a7a11c4a5fcf17523953da5c5ba0cdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 62, + 330, + 552, + 469 + ], + "blocks": [ + { + "bbox": [ + 62, + 330, + 552, + 469 + ], + "lines": [ + { + "bbox": [ + 62, + 330, + 552, + 469 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 552, + 469 + ], + "type": "table", + "html": "
Open-source Video VLMs
AlibabaQwen2.5-VL-7B2025-142.343.743.343.564.445.643.8
Qwen2.5-VL-72B2025-154.352.553.149.580.052.653.0
Qwen2-VL-7B2024-836.134.735.240.535.640.036.3
Qwen2-VL-72B2024-948.143.044.640.873.344.044.5
DAMOVideoLLama3-2B2025-153.242.546.034.355.636.443.7
VideoLLama3-7B2025-149.445.146.542.853.343.845.9
Shanghai AI LabInternVideo2.5-8B2025-157.250.552.744.346.744.550.7
InternVideo2-8B2024-835.639.338.143.00.038.738.2
LLaVALLaVA-One-Vision-7B2024-936.835.636.037.835.637.536.3
LLaVA-NeXT-Video-34B2024-629.631.630.924.555.627.630.1
", + "image_path": "a9935b2beb450c25ba0ff46195b92cba9bfb6f95377aacfd006d26ff3125f178.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 478, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 55, + 478, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 555, + 502 + ], + "type": "text", + "content": "Table 1. Evaluation on VLM4D Benchmark across various proprietary and open-source VLMs. Top three performers in each category are highlighted from dark (highest) to light (third highest). Human and random selection baselines are included for reference." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 524, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 295, + 608 + ], + "type": "text", + "content": "video, followed by the prompt forming the input. For each model, we evaluate on two different inference settings. In the first setting, the model prompted to directly output (DO) the answer immediately without any reasoning, and in the second evaluation setting, the model is directed to create intermediate reasoning steps, chain-of-thought (CoT) [81], before inferring the final answer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "Metrics Following prior work [87] and given the nature of our target task, we adopt multiple-choice questions (MCQs) for evaluation, using accuracy as the primary metric. For the two inference settings described earlier, we employ LLM-as-Judge [95] to assess the outputs of VLMs. We opt for this method instead of string or template matching, as VLMs—especially under chain of thought (CoT) prompt-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 525, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 632 + ], + "type": "text", + "content": "ing—often generate all possible answer choices during reasoning, with varying frequencies and slight formatting differences. In some cases, the final answer may even contradict the reasoning. To better evaluate whether the model truly understands the video, we prompt two advanced LLMs (GPT-o3 and o4-mini) to grade based on the full CoT reasoning, not just the final answer. We then perform a cross-check between their judgments and manually resolve any disagreements. The evaluation results are reported in Tab. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 646, + 429, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 429, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 429, + 658 + ], + "type": "text", + "content": "4.2. Benchmark Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Our comprehensive evaluation on the VLM4D benchmark, detailed in Tab. 1, systematically assesses the spatiotemporal reasoning capabilities of modern Vision-Language Models (VLMs). The results highlight a clear hierar" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8605" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": "chy, with proprietary models demonstrating superior performance over their open-source counterparts. Google's Gemini-2.5-Pro emerges as the top-performing model with an overall accuracy of " + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "inline_equation", + "content": "62.0\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": ", followed by OpenAI's GPT4o at " + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "inline_equation", + "content": "57.5\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": ". Within the open-source domain, image-based models like Meta's Llama-4-Scout-17B (54.1%) and videosupported models like Alibaba's Qwen2.5-VL-72B (53.0%) demonstrate highly competitive results, even outperforming some proprietary counterparts. Despite these achievements, a significant performance gap persists when compared to human accuracy (98.8%), underscoring that sophisticated 4D awareness remains a formidable challenge for AI. Notably, performance varies significantly across different categories, such as real versus synthetic data and egocentric versus exo-centric perspectives, indicating that current models lack generalized spatiotemporal understanding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 278, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 278, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 278, + 289 + ], + "type": "text", + "content": "5. Analysis: Why VLMs Don't Work Well?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 240, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 240, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 240, + 309 + ], + "type": "text", + "content": "5.1. Limited Spatiotemporal Cognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 531 + ], + "type": "text", + "content": "Despite significant advances in VLMs, their abilities to understand and reason about motion, spatial relationships, and temporal coherence remains fundamentally underdeveloped [13, 101]. Chain of thought (CoT) [81] is widely employed as a method to improve accuracy through step-by-step reasoning. We showcase a comparison between CoT and DO in Fig. 6. Overall, there is no indication of a large advantage of CoT over all evaluated models. Upon deeper exploration of the CoT reasoning of some models, we observe that the reasoning process was primarily flawed in the following ways: irrelevant information and arriving at conclusions that are inconsistent with the reasoning process. Larger models exhibited strategies that would be similar to how a human processes spatiotemporal information, but the resulting execution falls short of human performance. This demonstrates a disconnect between its visual and linguistic knowledge. We provide examples of this behavior in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 264, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 264, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 264, + 553 + ], + "type": "text", + "content": "5.2. Deficiencies in Spatiotemporal Labeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Another avenue of exploration we undertook is to understand the richness of spatiotemporal labels in popular supervised fine-tuning (SFT) VLM datasets. Typically, video captioning occurs at the 'scene' level, lacking fine-grained temporal, spatial, and object-level details. We performed an extensive analysis, encompassing over 2 million samples [14, 17, 44, 45, 94]. We performed this analysis through string-matching of spatiotemporal descriptors related to directionality, translational motion, rotation, and perspective shifts and provide the overall results in Fig. 7. We then performed a manual finegrained evaluation of the ShareGPT4Video dataset [14] which we found had the highest density of spatiotemporal dataset. We found that" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 259 + ], + "type": "image", + "image_path": "054262790f82a626f629e2cde3a06ae7d864a6f908397e350eca9a1f9b2152bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "type": "text", + "content": "Figure 6. Comparison of CoT and DO Accuracy Across Models. Accuracy comparison between Chain-of-Thought (CoT) and Direct Output (DO) prompting across VLMs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 315, + 561, + 506 + ], + "blocks": [ + { + "bbox": [ + 319, + 315, + 561, + 506 + ], + "lines": [ + { + "bbox": [ + 319, + 315, + 561, + 506 + ], + "spans": [ + { + "bbox": [ + 319, + 315, + 561, + 506 + ], + "type": "image", + "image_path": "af46013f16998694edd5a53b3093a61ae9c88455eb6bd649da3c78ece349e127.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 520, + 555, + 542 + ], + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 542 + ], + "type": "text", + "content": "Figure 7. Heatmap of Occurrences of Spatial-Temporal Terms in popular video SFT datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 647 + ], + "type": "text", + "content": "from a sample of 100 labels that were detected as spatiotemporal, less than " + }, + { + "bbox": [ + 313, + 563, + 555, + 647 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 563, + 555, + 647 + ], + "type": "text", + "content": " of them were judged as accurate upon human evaluation. This result underscores the inadequacy of current dense captioning approaches, which frequently generate spatiotemporal descriptors without capturing precise motion dynamics. We provide more detailed analysis and explanations in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 658, + 459, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 459, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 459, + 671 + ], + "type": "text", + "content": "6. Probing Future Solutions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "To probe promising future solutions for enhancing spatiotemporal video understanding, we propose two approaches that address some of the shortcomings of current" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8606" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "state-of-the-art VLMs: fine-tuning a VLM on data-rich in spatiotemporal actions and the other leveraging 4D reconstruction and feature fields jointly with a VLM. SFT refines the model's abilities by training on datasets that contain temporally and spatially rich actions and interactions. By integrating structured visual representations and targeted fine-tuning, these approaches enhance video-language models' ability to interpret motion. The second method lifts the feature space of VLMs into a temporally coherent 4D feature field, providing structured scene representations that improve motion and spatial reasoning in the stage of decoding and inference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 231, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 295, + 387 + ], + "type": "text", + "content": "Spatial-Temporal SFT We evaluate on a subset split of the real dataset by randomly splitting the real-world dataset into a training and testing split (80% / 20%) and we try settings using synthetic/real both for training. We conducted the experiments using Qwen 2VL (7B) and Qwen 2.5VL (7B) through LLama-Factory [96], and compared the performance before and after supervised fine-tuning in Tab. 2. The results demonstrated an improvement in accuracy in spatiotemporal reasoning, suggesting that performance gains can be obtained through targeted training. However, the addition of synthetic data does not necessarily increase performance over using real data alone, suggesting the importance of synthetic data quality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 714 + ], + "type": "text", + "content": "4D Feature Fields Reconstruction Recent advances in 3D/4D feature fields reconstruction methods [24, 37, 90, 97, 99] have significantly enhanced the vision foundation model's performance in 3D/4D space by integrating structured latent scene representations into the model's inference stage. Inspired by the promising results of feature lifting, we explore enhancing the InternVideo2-8B model [78] with spatiotemporal awareness by adopting the strategy proposed in Feature4X [99], which constructs the VLM's 2D feature space along the time dimension into a 4D feature field. To assess this approach, we evaluate performance on a subset of the VLM4D benchmark, specifically leveraging all 50 videos from the DAVIS 2016 dataset [63]. Our experimental evaluation compares inference performance across three distinct input modalities: original 2D videos; rendered novel global-view videos (which provide broader contextual information in 2D format); and reconstructed global feature fields (which implicitly incorporate 4D scene-level information during reasoning). Table 3 reveals that reconstructed feature fields achieve the highest accuracy across both reasoning types. This success stems from two key advantages: the inherent structure of 4D representations and the ability of feature field inference to avoid the rendering artifacts of RGB reconstruction (global view video). However, the current approach requires per-scene optimization as a post-processing step, limiting its generalizability and" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 362, + 70, + 507, + 247 + ], + "blocks": [ + { + "bbox": [ + 362, + 70, + 507, + 247 + ], + "lines": [ + { + "bbox": [ + 362, + 70, + 507, + 247 + ], + "spans": [ + { + "bbox": [ + 362, + 70, + 507, + 247 + ], + "type": "table", + "html": "
ModelMC
Original Model
Qwen 2VL (7B)38.3
Qwen 2.5VL (7B)43.4
Finetuned Model
Qwen 2VL (7B) (R)54.5
Qwen 2VL (7B) (S)42.5
Qwen 2VL (7B) (R+S)55.5
Qwen 2.5VL (7B) (R)55.6
Qwen 2.5VL (7B) (S)48.6
Qwen 2.5VL (7B) (R+S)56.3
", + "image_path": "ecaf95c4ebc06418cc9bb13401a409f083e848d1e3ee0f7abbf739e257dabf2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 338, + 299, + 531, + 447 + ], + "blocks": [ + { + "bbox": [ + 313, + 255, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 289 + ], + "type": "text", + "content": "Table 2. SFT on Spatial-Temporal Datasets. MC refers to the performance of multiple-choice, while R, S, and " + }, + { + "bbox": [ + 313, + 255, + 553, + 289 + ], + "type": "inline_equation", + "content": "\\mathbb{R} + \\mathbb{S}" + }, + { + "bbox": [ + 313, + 255, + 553, + 289 + ], + "type": "text", + "content": " denote real, synthetic, and both real+synthetic usage of data for fine-tuning." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 338, + 299, + 531, + 447 + ], + "lines": [ + { + "bbox": [ + 338, + 299, + 531, + 447 + ], + "spans": [ + { + "bbox": [ + 338, + 299, + 531, + 447 + ], + "type": "table", + "html": "
Input ModalityAccuracy
Chain of Thought Response
Original 2D Video36.0
Global View Video32.7
Global Feature Field37.4
Direct Output Response
Original 2D Video24.3
Global View Video23.8
Global Feature Field29.0
", + "image_path": "6364fe09d8b740575d79c946958d6cfc69430d9e5a658ec5e4bda85d64ccd738.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 455, + 553, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 455, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 553, + 488 + ], + "type": "text", + "content": "Table 3. InternVideo2 Accuracy with 4D Feature Field Reconstruction. Comparison of InternVideo2's performance given different input modalities from the same subset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 503, + 462, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 462, + 515 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 462, + 515 + ], + "type": "text", + "content": "making it computationally intensive." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 526, + 388, + 538 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 715 + ], + "type": "text", + "content": "Through the construction of the VLM4D benchmark, we evaluate the spatiotemporal reasoning capabilities of various Vision-Language Models (both open-source and proprietary). While more recently released models demonstrate improved performance over their counterparts, they remain significantly behind human proficiency. Overall, our work questions whether VLMs possess spatiotemporal reasoning abilities that are imperative to have for more sophisticated visual agents in fields ranging from robotics to interactive AI systems that require a deep understanding of dynamic visual environments. We hope to inspire future work to explore novel approaches for integrating spatiotemporal grounding, thereby enhancing their spatiotemporal reasoning capabilities and facilitating robust deployment." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8607" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 65, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J Hewett, Mojan Javaheripi, Piero Kauffmann, et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 214, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 214, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 65, + 214, + 296, + 443 + ], + "type": "text", + "content": "[3] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Weijian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 445, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 445, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 65, + 445, + 296, + 499 + ], + "type": "text", + "content": "[4] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 65, + 501, + 296, + 555 + ], + "type": "text", + "content": "[5] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 557, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 557, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 65, + 557, + 296, + 578 + ], + "type": "text", + "content": "[6] Anthropic. System card: Claude opus 4 & claude sonnet 4. Technical report, 2025. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 580, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 580, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 65, + 580, + 296, + 623 + ], + "type": "text", + "content": "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 624, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 624, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 65, + 624, + 296, + 677 + ], + "type": "text", + "content": "[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 65, + 680, + 296, + 713 + ], + "type": "text", + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 338, + 74, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 74, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 338, + 74, + 553, + 105 + ], + "type": "text", + "content": "Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 106, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 106, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 320, + 106, + 553, + 138 + ], + "type": "text", + "content": "[10] Neil Burgess. Spatial memory: how egocentric and allocentric combine. Trends in Cognitive Sciences, 10(12):551-557, 2006. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 553, + 183 + ], + "type": "text", + "content": "[11] Pradyumna Chari, Yunhao Ba, Shijie Zhou, Chinmay Talegaonkar, Shreeram Athreya, and Achuta Kadambi. On learning mechanical laws of motion from video using neural networks. IEEE Access, 11:30129-30145, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 185, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 185, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 185, + 555, + 239 + ], + "type": "text", + "content": "[12] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data-efficient adaptation of pretrained language models for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18030-18040, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 240, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 240, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 240, + 553, + 293 + ], + "type": "text", + "content": "[13] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 320, + 295, + 553, + 348 + ], + "type": "text", + "content": "[14] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Bin Lin, Zhenyu Tang, et al. Sharegpt4video: Improving video understanding and generation with better captions. arXiv preprint arXiv:2406.04325, 2024. 4, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 350, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 553, + 413 + ], + "type": "text", + "content": "[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 470 + ], + "type": "text", + "content": "[16] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 320, + 472, + 553, + 525 + ], + "type": "text", + "content": "[17] Erfei Cui, Yinan He, Zheng Ma, Zhe Chen, Hao Tian, Weiyun Wang, Kunchang Li, Yi Wang, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, Yali Wang, Limin Wang, Yu Qiao, and Jifeng Dai. Sharegpt-4o: Comprehensive multimodal annotations with gpt-4o, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 320, + 526, + 553, + 580 + ], + "type": "text", + "content": "[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 582, + 553, + 613 + ], + "type": "text", + "content": "[19] Julian De Freitas, Nicholas E. Myers, and Anna C. Nobre. Tracking the changing feature of a moving object. Journal of Vision, 16(3):22, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 614, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 614, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 614, + 553, + 689 + ], + "type": "text", + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "text", + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8608" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 294, + 173 + ], + "type": "text", + "content": "[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 294, + 217 + ], + "type": "text", + "content": "[23] Chenrui Fan, Ming Li, Lichao Sun, and Tianyi Zhou. Missing premise exacerbates overthinking: Are reasoning models losing critical thinking skill? arXiv preprint arXiv:2504.06514, 2025. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 284 + ], + "type": "text", + "content": "[24] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. Advances in neural information processing systems, 37:40212-40229, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 287, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 294, + 342 + ], + "type": "text", + "content": "[25] Zhiwen Fan, Jian Zhang, Renjie Li, Junge Zhang, Runjin Chen, Hezhen Hu, Kevin Wang, Huaizhi Qu, Dilin Wang, Zhicheng Yan, et al. Vlm-3r: Vision-language models augmented with instruction-aligned 3d reconstruction. arXiv preprint arXiv:2505.20279, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 343, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 294, + 376 + ], + "type": "text", + "content": "[26] Jennifer J. Freyd and Ronald A. Finke. Representational momentum. Journal of Experimental Psychology: Learning, Memory, and Cognition, 10(1):126-132, 1984. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 377, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 294, + 441 + ], + "type": "text", + "content": "[27] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 487 + ], + "type": "text", + "content": "[28] Google Gemini Team. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Technical report, 2025. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 294, + 544 + ], + "type": "text", + "content": "[29] Tao Gong, Chengqi Lyu, Shilong Zhang, Yudong Wang, Miao Zheng, Qian Zhao, Kuikun Liu, Wenwei Zhang, Ping Luo, and Kai Chen. Multimodal-gpt: A vision and language model for dialogue with humans. arXiv preprint arXiv:2305.04790, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 294, + 621 + ], + "type": "text", + "content": "[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 678 + ], + "type": "text", + "content": "[31] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multidiscipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "text", + "content": "[32] Xuehai He, Shuohang Wang, Jianwei Yang, Xiaoxia Wu, Yiping Wang, Kuan Wang, Zheng Zhan, Olatunjri Ruwase, Yelong Shen, and Xin Eric Wang. Mojito: Motion tra" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "jctory and intensity control for video generation. arXiv preprint arXiv: 2412.08948, 2024. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 139 + ], + "type": "text", + "content": "[33] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 553, + 171 + ], + "type": "text", + "content": "[34] Gunnar Johansson. Visual perception of biological motion and a model for its analysis. Perception & Psychophysics, 14(2):201-211, 1973. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 173, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 173, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 553, + 237 + ], + "type": "text", + "content": "[35] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. How good is my video lmm? complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 239, + 553, + 293 + ], + "type": "text", + "content": "[36] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 294, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 294, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 320, + 294, + 553, + 336 + ], + "type": "text", + "content": "[37] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in neural information processing systems, 35:23311-23330, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 338, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 553, + 360 + ], + "type": "text", + "content": "[38] Ivan Laptev. On space-time interest points. International Journal of Computer Vision, 64(2-3):107-123, 2005. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "type": "text", + "content": "[39] Yann LeCun, Bernhard Boser, John S. Denker, Donnie Henderson, Richard E. Howard, Wayne Hubbard, and Lawrence D. Jackel. Backpropagation applied to handwritten zip code recognition. *Neural Computation*, 1(4):541-551, 1989. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 416, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 416, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 320, + 416, + 553, + 446 + ], + "type": "text", + "content": "[40] Alan M. Leslie. Spatiotemporal continuity and the perception of causality in infants. Perception, 13(3):287-305, 1984. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 320, + 449, + 553, + 502 + ], + "type": "text", + "content": "[41] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 320, + 504, + 553, + 547 + ], + "type": "text", + "content": "[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 548, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 548, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 320, + 548, + 553, + 601 + ], + "type": "text", + "content": "[43] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 320, + 603, + 553, + 646 + ], + "type": "text", + "content": "[44] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "text", + "content": "[45] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "text", + "content": "[46] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8609" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "type": "text", + "content": "Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 184 + ], + "type": "text", + "content": "[47] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, and Zhiwen Fan. 4k4DGen: Panoramic 4d generation at 4k resolution. In The Thirteenth International Conference on Learning Representations, 2025. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 185, + 294, + 228 + ], + "type": "text", + "content": "[48] Xinhao Li, Zhenpeng Huang, Jing Wang, Kunchang Li, and Limin Wang. Videoeval: Comprehensive benchmark suite for low-cost evaluation of video foundation model. arXiv preprint arXiv:2407.06491, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 294, + 284 + ], + "type": "text", + "content": "[49] Lu Ling, Chen-Hsuan Lin, Tsung-Yi Lin, Yifan Ding, Yu Zeng, Yichen Sheng, Yunhao Ge, Ming-Yu Liu, Aniket Bera, and Zhaoshuo Li. Scenethesis: A language and vision agentic framework for 3d scene generation. arXiv preprint arXiv:2505.02836, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 285, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 319 + ], + "type": "text", + "content": "[50] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 320, + 294, + 353 + ], + "type": "text", + "content": "[51] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with ringattention. arXiv preprint, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "type": "text", + "content": "[52] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv preprint arXiv:2402.08268, 2024.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 454 + ], + "type": "text", + "content": "[53] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 509 + ], + "type": "text", + "content": "[54] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 294, + 555 + ], + "type": "text", + "content": "[55] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 556, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 294, + 600 + ], + "type": "text", + "content": "[56] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 644 + ], + "type": "text", + "content": "[57] D. Marr and S. Ullman. Directional selectivity and its use in early visual processing. Proceedings of the Royal Society of London. Series B, Biological Sciences, 211(1183):151-180, 1981. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 677 + ], + "type": "text", + "content": "[58] Meta. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation. Technical report, 2025. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 294, + 713 + ], + "type": "text", + "content": "[59] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluat" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "ing video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 95, + 553, + 139 + ], + "type": "text", + "content": "[60] Basak Melis Öcal, Maxim Tatarchenko, Sezer Karaoglu, and Theo Gevers. Sceneteller: Language-to-3d scene generation. In European Conference on Computer Vision, pages 362–378. Springer, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 140, + 518, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 140, + 518, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 140, + 518, + 150 + ], + "type": "text", + "content": "[61] OpenAI. Hello gpt-4o. Technical report, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 553, + 205 + ], + "type": "text", + "content": "[62] Shivansh Patel, Xinchen Yin, Wenlong Huang, Shubham Garg, Hooshang Nayyeri, Li Fei-Fei, Svetlana Lazebnik, and Yunzhu Li. A real-to-sim-to-real approach to robotic manipulation with vlm-generated iterative keypoint rewards. arXiv preprint arXiv:2502.08643, 2025. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 206, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 206, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 553, + 270 + ], + "type": "text", + "content": "[63] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 321, + 272, + 553, + 316 + ], + "type": "text", + "content": "[64] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 317, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 317, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 321, + 317, + 553, + 348 + ], + "type": "text", + "content": "[65] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 350, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 350, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 321, + 350, + 553, + 426 + ], + "type": "text", + "content": "[66] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 321, + 427, + 553, + 470 + ], + "type": "text", + "content": "[67] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems (NIPS), pages 568-576, 2014. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 472, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 472, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 321, + 472, + 553, + 492 + ], + "type": "text", + "content": "[68] Elizabeth S. Spelke and Katherine D. Kinzler. Core knowledge. Developmental Science, 10(1):89-96, 2007. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 494, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 494, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 321, + 494, + 553, + 548 + ], + "type": "text", + "content": "[69] Alessandro Suglia, Claudio Greco, Katie Baker, Jose L Part, Ioannis Papaioannou, Arash Eshghi, Ioannis Konstas, and Oliver Lemon. Alanavlm: A multimodal embodied ai foundation model for egocentric video understanding. arXiv preprint arXiv:2406.13807, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 321, + 548, + 553, + 603 + ], + "type": "text", + "content": "[70] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 321, + 604, + 553, + 658 + ], + "type": "text", + "content": "[71] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "text", + "content": "[72] Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8610" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[73] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "type": "text", + "content": "[74] Heng Wang, Alexander Kläser, Cordelia Schmid, and Cheng-Lin Liu. Action recognition by dense trajectories. In 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3169-3176, 2011. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 295, + 216 + ], + "type": "text", + "content": "[75] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 271 + ], + "type": "text", + "content": "[76] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 272, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 294, + 325 + ], + "type": "text", + "content": "[77] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 392 + ], + "type": "text", + "content": "[78] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Intervideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 393, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 294, + 446 + ], + "type": "text", + "content": "[79] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "type": "text", + "content": "[80] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 493, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 294, + 546 + ], + "type": "text", + "content": "[81] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 548, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 612 + ], + "type": "text", + "content": "[82] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26057-26068, 2025. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 678 + ], + "type": "text", + "content": "[83] Shiwei Wu, Joya Chen, Kevin Qinghong Lin, Qimeng Wang, Yan Gao, Qianli Xu, Tong Xu, Yao Hu, Enhong Chen, and Mike Zheng Shou. Videollm-mod: Efficient video-language streaming with mixture-of-depths vision computation. Advances in Neural Information Processing Systems, 37:109922-109947, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 680, + 269, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 269, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 269, + 690 + ], + "type": "text", + "content": "[84] xAI. Grok-2 beta release. Technical report, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[85] Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 116 + ], + "type": "text", + "content": "and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 585-601, 2018. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 553, + 162 + ], + "type": "text", + "content": "[86] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 321, + 163, + 553, + 207 + ], + "type": "text", + "content": "[87] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. arXiv preprint arXiv:2412.14171, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 209, + 553, + 262 + ], + "type": "text", + "content": "[88] Yuqian Yuan, Hang Zhang, Wentong Li, Zesen Cheng, Boqiang Zhang, Long Li, Xin Li, Deli Zhao, Wenqiao Zhang, Yueting Zhuang, et al. Videorefer suite: Advancing spatial-temporal object understanding with video llm. arXiv preprint arXiv:2501.00599, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 264, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 264, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 321, + 264, + 553, + 341 + ], + "type": "text", + "content": "[89] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 321, + 343, + 553, + 387 + ], + "type": "text", + "content": "[90] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In European Conference on Computer Vision, pages 57-74. Springer, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 553, + 443 + ], + "type": "text", + "content": "[91] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 321, + 445, + 553, + 477 + ], + "type": "text", + "content": "[92] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 479, + 553, + 533 + ], + "type": "text", + "content": "[93] Hongxin Zhang, Zeyuan Wang, Qiushi Lyu, Zheyuan Zhang, Sunli Chen, Tianmin Shu, Yilun Du, and Chuang Gan. Combo: compositional world models for embodied multi-agent cooperation. arXiv preprint arXiv:2404.10775, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 535, + 553, + 578 + ], + "type": "text", + "content": "[94] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-last: A strong zero-shot video understanding model, 2024. 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 580, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 580, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 321, + 580, + 553, + 634 + ], + "type": "text", + "content": "[95] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 636, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 636, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 636, + 553, + 713 + ], + "type": "text", + "content": "[96] Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of " + }, + { + "bbox": [ + 321, + 636, + 553, + 713 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 321, + 636, + 553, + 713 + ], + "type": "text", + "content": " language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), Bangkok, Thailand, 2024. Association for Computational Linguistics. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8611" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 383 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 61, + 72, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 294, + 148 + ], + "type": "text", + "content": "[97] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 217 + ], + "type": "text", + "content": "[98] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 219, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 294, + 282 + ], + "type": "text", + "content": "[99] Shijie Zhou, Hui Ren, Yijia Weng, Shuwang Zhang, Zhen Wang, Dejia Xu, Zhiwen Fan, Suya You, Zhangyang Wang, Leonidas Guibas, et al. Feature4x: Bridging any monocular video to 4d agentic ai with versatile gaussian feature fields. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14179-14190, 2025. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 285, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 285, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 57, + 285, + 294, + 327 + ], + "type": "text", + "content": "[100] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 330, + 294, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 330, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 57, + 330, + 294, + 383 + ], + "type": "text", + "content": "[101] Orr Zohar, Xiaohan Wang, Yann Dubois, Nikhil Mehta, Tong Xiao, Philippe Hansen-Estruch, Licheng Yu, Xiaofang Wang, Felix Juefei-Xu, Ning Zhang, et al. Apollo: An exploration of video understanding in large multimodal models. arXiv preprint arXiv:2412.10360, 2024. 7" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8612" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_content_list.json b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bb2aec682885eda20e8e1a7b8b1adbd6939ec1c6 --- /dev/null +++ b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_content_list.json @@ -0,0 +1,1504 @@ +[ + { + "type": "text", + "text": "VLR-Driver: Large Vision-Language-Reasoning Models for Embodied Autonomous Driving", + "text_level": 1, + "bbox": [ + 151, + 138, + 848, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fanjie Kong $^{1}$ Yitong Li $^{1}$ Weihuang Chen $^{1*}$ Chen Min $^{2}$ Yizhe Li $^{3}$ Zhiqiang Gao $^{5,1}$ Haoyang Li $^{4}$ Zhongyu Guo $^{1}$ Hongbin Sun $^{1}$", + "bbox": [ + 210, + 210, + 781, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an Jiaotong University $^{2}$ Institute of Computing Technology, Chinese Academy of Sciences $^{3}$ University of Maryland, College Park $^{4}$ Nanchang University $^{5}$ Engineering University of PAP", + "bbox": [ + 122, + 246, + 870, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 256, + 314, + 331, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rise of embodied intelligence and multi-modal large language models has led to exciting advancements in the field of autonomous driving, establishing it as a prominent research focus in both academia and industry. However, when confronted with intricate and ambiguous traffic scenarios, the lack of logical reasoning and cognitive decision-making capabilities remains the primary challenge impeding the realization of embodied autonomous driving. Although Vision Language Models (VLMs) have enhanced the deep semantic understanding of autonomous driving systems, they exhibit notable limitations in decision explainability when handling rare and long-tail traffic scenarios. In this paper, we propose VLR-Driver, a novel multimodal Vision-Language-Reasoning (VLR) framework based on Chain of Thought (CoT) for embodied autonomous driving. The framework employs a spatiotemporal CoT reasoning approach to recursively analyze potential safety risks and driving intentions of other agents, thereby delivering an efficient and transparent decision-making process. Furthermore, we construct a multi-modal reasoning-decision dataset to support the advancement of hierarchical reasoning of VLMs in autonomous driving. Closed-loop experiments conducted in CARLA demonstrate that the VLR-Driver significantly outperforms state-of-the-art end-to-end methods. Notably, key metrics such as driving score improved by $17.5\\%$ , while the success rate improved by $22.2\\%$ , offering a more transparent, reliable, and secure solution for autonomous driving systems.", + "bbox": [ + 99, + 349, + 485, + 762 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d7db6f7b1621f5d1f8206748972db8364e81136fe3ac3e7acad0be138f8c717.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 517, + 313, + 895, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2ccdb0133d42db9d53bfa0246765ac5346ab9173c4815e395e22c3071f52108.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 517, + 393, + 895, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb93c92b0dd8fcd295e73c1b7dab290144df030ebf8ab07dd24e57c0732ffd2b.jpg", + "image_caption": [ + "(c)", + "Figure 1. Comparison of different LLM-based AD systems. (a) VLM model focuses more on answering questions, but it is difficult to convert decisions into coherent control signals. (b) VLA model focuses more on predicting driving control signals, but lacks interpretability. (c) The VLR model generates decisions and controls signals with transparent reasoning processes through CoT, enhancing the driver trust in the system." + ], + "image_footnote": [], + "bbox": [ + 517, + 476, + 895, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 689, + 643, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, the rapid progress of End-to-End (E2E) architectures [6, 7, 18, 49], Large Language Models (LLMs) [12, 41, 48], and embodied intelligence [26, 53, 55] has established these technologies as key enablers of innovation in Autonomous Driving (AD). Especially, VLMs enriched with extensive pre-training knowledge exhibit strong spatial understanding and common-sense reasoning abilities. DriveVLM [35] leverages VLM to enhance spatial awareness and planning capabilities in complex driving scenarios. CoVLA [2] integrates visual perception, language understanding, and action planning, effectively describing traffic scenarios and generating executable control actions.", + "bbox": [ + 511, + 713, + 898, + 891 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 119, + 806, + 243, + 819 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This research was supported in part by National Natural Science Foundation of China (No. 62088102, No. U24A20291), National Key Research and Development Program of China (No. 2024YFB4405504), the Fundamental Research Funds for the Central Universities (xtr072022001), the Opening Foundation of Key Laboratory of CTCIE, Ministry of Education (CTCIE20240101).", + "bbox": [ + 99, + 819, + 485, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26966", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the decision-making process of VLMs often functions as a \"black box\", making it challenging to trace and interpret their underlying logic. This makes it difficult for AD systems to be fully trusted by drivers when encountering complex and emergency situations, such as illegal roadside parking, navigating intersections without traffic signals, and managing complex mixed-traffic interactions between motorized and non-motorized vehicles, thereby limiting their reliability and safety in real-world applications [39, 46, 54]. Moreover, most VLMs are trained on internet data, lacking spatial understanding and specialized training in the field of AD, making it difficult for them to fully adapt to dynamic and complex driving scenarios.", + "bbox": [ + 99, + 102, + 480, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, CoT reasoning demonstrates strong inference, interpretability, and generalization capabilities by breaking down complex tasks into intermediate reasoning steps [40]. CoT enables systems to think step by step rather than relying on E2E black-box predictions, making it one of the key approaches toward achieving embodied AD. Sce2DriveX [52] enhances comprehensive perception and reasoning by introducing a multi-modal LLM framework with CoT, enabling a deeper understanding of spatiotemporal relationships and road topology. Similarly, DriveCoT [38] integrates CoT reasoning to improve decision-making interpretability and controllability in AD systems. However, existing CoT-based methods typically rely on open-ended language generation for reasoning, which lacks structural constraints. The generated reasoning steps may exhibit causal inversion or unnecessary vocabulary, making them uncontrollable in safety-critical tasks. Additionally, current methods primarily operate on static snapshots rather than continuous temporal sequences, limiting their ability to predict future events in dynamic traffic environments.", + "bbox": [ + 99, + 297, + 482, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge these gaps, we introduce VLR-Driver, a hierarchical CoT-based visual-language-reasoning model designed for closed-loop embodied AD. Our approach integrates the spatiotemporal features from cross-modal data, including multi-frame multi-view images and ego-vehicle control signals, by employing a SpatioTemporal CoT (STCoT) strategy that produces human-like reflective reasoning processes and driving action decisions. Additionally, we adopt a dual-phase training strategy, combining Low-Rank Adaptation (LoRA) [16] with an improved Stepwise Group Relative Policy Optimization (Step-GRPO) [34], significantly enhancing memory capacity and deep reasoning abilities. Our proposed VLR-Driver not only inherits the global action optimization capability of VLA models, but also preserves the transparency of modular rule-based methods. When encountering long-tail events and rare traffic scenarios, it demonstrates exceptional reflective reasoning and step-by-step inference, thereby enhancing the trust of human drivers in AD systems. The differences between VLM, VLA, and our proposed VLR are illustrated in Fig. 1.", + "bbox": [ + 99, + 595, + 482, + 891 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further enhance VLR models in environmental understanding, reasoning, and decision-making, we introduce the VLR-Driver Dataset. This data set includes detailed scene descriptions, weather information, vehicle state details, and most critically, human-like CoT reasoning processes and the corresponding driving decisions. Various experiments demonstrate that VLR-Driver is capable of making accurate driving decisions and coherent reasoning, even under challenging and highly dynamic road conditions.", + "bbox": [ + 511, + 102, + 893, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary contributions of this work are summarized as follows:", + "bbox": [ + 511, + 234, + 893, + 263 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Distinctive VLR-Driver Framework. We introduce VLR-Driver, a visual-language-reasoning model developed for embodied AD. It generates a human-like reflective reasoning process within the driving system, enabling accurate driving decisions.", + "- Spatiotemporal CoT. We present a spatiotemporal CoT strategy that recursively analyzes potential safety risks and the driving intentions of moving agents in complex traffic scenarios, ensuring that the reasoning process of driving decisions remains transparent and interpretable.", + "- Advanced VLR-Driver Dataset. We construct VLR-Driver Dataset, a cutting-edge visual-language-reasoning and decision-making dataset specifically designed for AD. It supports the enhancement of spatiotemporal understanding and reflective reasoning capabilities in embodied AD systems.", + "- Superior Performance in Closed-Loop Simulations. Extensive closed-loop experiments are conducted on the CARLA platform. On the Bench2Drive benchmark, our approach achieves a $17.5\\%$ improvement in driving score and a $22.2\\%$ increase in success rate, providing a more human-like, reliable, and trustworthy solution." + ], + "bbox": [ + 511, + 266, + 893, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 606, + 651, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. End-to-end Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 631, + 792, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rapid advancement of E2E-AD has fostered a growing transition away from modular rule-based methods toward data-driven approaches [15, 23]. Based on different input modalities, E2E methods can be categorized into visual-only methods [5, 8, 10, 17] and vision-LiDAR fusion methods [1, 20, 21]. TCP [42] and NEAT [10] adopt imitation learning methods, training directly on collected state-action pair datasets, demonstrating the feasibility of E2E approaches for AD. Roach [51] utilizes reinforcement learning experts as coaches, delivering dense and informative supervision signals to agents equipped with monocular camera inputs. However, vision-based methods inherently struggle with distance and depth estimation. These limitations may compromise the reliability of driving decisions [24]. To address these limitations, vision-LiDAR fusion methods such as TransFuser [11], CrossFuser [43],", + "bbox": [ + 511, + 654, + 893, + 890 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26967", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eac4ff6d82c2ec490e6391963617430d0427d8e7bb9e1e00dd4386fe5bbcfc5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 103, + 181, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLR-Driver Datasets", + "text_level": 1, + "bbox": [ + 189, + 104, + 359, + 118 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6e6932366bf55cbdcfb9be596995fdfe69694031a9c0bc386017b65fd043bcbb.jpg", + "image_caption": [ + "Figure 2. Overview of VLR-Driver framework. We introduce VLR-Driver Dataset, an advanced visual-language-reasoning dataset designed for AD, featuring detailed annotations of scene descriptions, analytical reasoning, and behavioral decisions. We present VLR-Driver Embodied AI (EAI), a novel multi-modal visual-language-reasoning framework for embodied AD that leverages a hierarchical spatiotemporal CoT reasoning mechanism." + ], + "image_footnote": [], + "bbox": [ + 102, + 122, + 431, + 433 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3f4f222795ee64fa75f5b7a6345fda0714df69ace388916c0dbe5e07f61b2cb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 102, + 611, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLR-Driver EAI", + "text_level": 1, + "bbox": [ + 616, + 104, + 741, + 118 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b57cb0b6dbc90c15f2082fad4efe9575391df0061330598a2549371e8d4afeda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 122, + 893, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and FusionAD [47] have been designed to effectively integrate image with LiDAR data, significantly improving the robustness of AD systems in complex environments. However, different modalities contribute differently to the driving task, and spatiotemporal synchronization issues between sensor data can lead to inconsistencies, increasing both model complexity and training difficulty.", + "bbox": [ + 99, + 520, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. VLM and VLA in Autonomous Driving", + "text_level": 1, + "bbox": [ + 101, + 632, + 429, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLMs unify visual perception with natural language processing capabilities, enabling a more comprehensive understanding of driving environments [19, 36, 37]. By incorporating cross-modal data fusion, integrating textual information alongside visual inputs, these systems gain text comprehension and human interaction capabilities that conventional E2E models inherently lack. DriveGPT4 [45] enhances LLMs' ability to process multimodal inputs by projecting them into the text domain, thereby enabling interpretable E2E-AD. Senna [25] and DriveVLM [35] integrate VLMs with traditional modular pipelines or E2E frameworks, achieving decoupled high-level planning and low-level trajectory prediction. This approach improves planning performance while preserving the model's common sense reasoning capabilities.", + "bbox": [ + 99, + 654, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, VLA is an emerging paradigm that unifies vi", + "bbox": [ + 120, + 875, + 482, + 888 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sual perception, natural language understanding, and action prediction within a cohesive framework [13, 26]. Originally introduced in the field of robotics, RT-2 [4] pioneered the representation of robotic actions as text tokens, seamlessly incorporating them alongside natural language labels into the model's training set. This enables direct transfer of internet-scale knowledge to robotic control, significantly enhancing both the generalization and semantic reasoning capabilities of robotic systems. For AD, CoVLA [2] introduces an interpretable VLA model that integrates visual perception, language-based scene understanding, and action planning. This integration improves the system's ability to comprehend complex driving scenarios, anticipate trajectory, and execute informed driving decisions. However, despite these advancements, VLA models still exhibit limitations in accurately predicting precise control action values, leading to a low lower-bound in AD performance. Furthermore, when confronted with highly dynamic and complex traffic scenarios, their deep understanding and reasoning capabilities remain insufficient. Unlike the above methods, our proposed VLR model places greater emphasis on step-by-step reasoning and the thought process of the model, providing drivers with increased confidence in using AD systems.", + "bbox": [ + 511, + 520, + 895, + 872 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26968", + "bbox": [ + 478, + 928, + 517, + 938 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Chain of Thought in Autonomous Driving", + "text_level": 1, + "bbox": [ + 101, + 101, + 450, + 118 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The CoT technology is an extension of prompt engineering, proposed by Wei Jason in 2022 [40], which has greatly improved the effectiveness of reasoning for complex problems. The highly anticipated Deepseek-R1 [33] model also utilizes the CoT technique, which deeply integrates multimodal knowledge base data, enabling the model to generate a step-by-step thinking process. DriveCoT [38] has built a CoT dataset that includes sensor data, control decisions, and CoT labels used to indicate reasoning processes. The model trained on this basis can generate predictions and final decisions with CoT, effectively improving model performance. LanguageMPC [30] combines LLM with Model Predictive Control and decomposes driving decisions into multiple subtasks through a CoT framework. This method enables the auto drive system to think like human beings, and improves its ability to handle complex scenes. OpenEMMA [44] introduces CoT technology to guide model generation of detailed descriptions of key objects, behavioral insights, and meta driving decisions, improving system transparency and usability. Motivated by these advancements, we apply step-by-step hierarchical spatiotemporal CoT to AD, enhancing the interpretability of reasoning and decision-making.", + "bbox": [ + 99, + 123, + 485, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 101, + 477, + 191, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present the motivation and design details of our VLR-Driver framework. As depicted in Fig. 2, VLR-Driver comprises two main components: the VLR-Driver EAI and the VLR-Driver Dataset. Initially, we introduce the design concept of the VLR model, which builds upon enhancements to the VLA model (Sec. 3.1). Subsequently, we elaborate on the hierarchical spatiotemporal CoT methodology (Sec. 3.2) and the specifics of the dual-phase training strategy (Sec. 3.3).", + "bbox": [ + 99, + 502, + 483, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 101, + 646, + 209, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The VLR model is a large visual-language-reasoning model designed for embodied AD. It can processes visual inputs, such as multi-view images, alongside textual information, including vehicle control signals. The model is capable of extracting spatiotemporal key features within a multimodal embedding and recursively analyze potential safety risks and the driving intentions of other agents within the reasoning level of the VLR model. Ultimately, it formulates a comprehensive reasoning framework and well-structured decision-making outputs, explicitly identifying critical risk factors and the underlying rationale behind each decision. This process significantly reinforces the robustness and safety of the AD system, ensuring adaptive resilience in complex and dynamic environments.", + "bbox": [ + 99, + 667, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this study, we employ pre-trained LLaVA-NeXT", + "bbox": [ + 119, + 875, + 482, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video [50] as the VLM and CLIP [29] as the visual encoder. The model is capable of processing multi-frame multi-view image data that capture historical temporal context, while also extracting real-time vehicle sensor information to facilitate dynamic and context-aware decision-making.", + "bbox": [ + 511, + 102, + 893, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input Representations. We utilize $N_{f}$ frame and $N_{v}$ view images from the past period, with a field of view (FoV) of 70 degrees. It can be represented as $V \\in \\mathbb{R}^{N_f \\times N_v \\times 3 \\times H_0 \\times W_0}$ . At the same time, there are also the current position $(x,y)$ of ego vehicle, the speed $v$ , the target point position $(p,q)$ . Subsequently, the compressed and cropped image data and the information from the ego's sensors are input into the model.", + "bbox": [ + 511, + 176, + 895, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output Representations. Our output consists of the reasoning process and driving decisions generated by the VLR model for the current driving scenario. This includes risk identification, traffic signal recognition, motion direction prediction, and AD decision-making. The driving decision will also be fused with the spatiotemporal information features extracted by the E2E model, and finally output the waypoints and control signals of the vehicle for the next moment.", + "bbox": [ + 511, + 294, + 895, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spatiotemporal CoT Reasoning", + "text_level": 1, + "bbox": [ + 511, + 436, + 784, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the reasoning capabilities and transparency of the AD system, we introduce a hierarchical ST-CoT that guides the model to approach driving decisions in a human-like manner. Our method decomposes the driving decision-making process into two levels: a perception-level spatiotemporal CoT $C_{\\text{perception}}$ , which focuses on extracting and understanding environmental dynamics, and a decision-level dynamic CoT $C_{\\text{decision}}$ , which refines and optimizes decision-making based on contextual and temporal factors. The example of CoT is shown in Fig. 3.", + "bbox": [ + 511, + 459, + 895, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Perception Level CoT", + "text_level": 1, + "bbox": [ + 511, + 614, + 705, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spatiotemporal characteristics of the environment play a crucial role in the AD process. The perception level CoT is responsible for guiding the model to extract spatial and temporal features from input image data, identify and locate crucial objects in the traffic, such as vehicles, pedestrians, alien objects, traffic lights, traffic signs, etc., and extract historical behavioral features of dynamic agent based on temporal information. Our method enables VLR model to describe the current driving scenario, construct real-time spatial layout and dynamic changes of the environment, and achieve long-term planning for driving decisions.", + "bbox": [ + 511, + 633, + 895, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial CoT. In driving scenarios, we primarily focus on obstacles that impact the ego's normal operation, including object categories $O_{type}$ and distances with the ego $O_{dis}$ . A critical aspect of safe driving is identifying potential risk points within the current lane. Additionally, when the ve", + "bbox": [ + 511, + 816, + 895, + 891 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26969", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f4cad0d991bd2d337f5ca0bbf6a670644dcf6e12afd98eace1dbff0f328f6a0.jpg", + "image_caption": [ + "Figure 3. Illustration of the ST-CoT reasoning process. In this scenario, where some vehicles are illegally parked ahead and blocking the lane, our method can conduct hierarchical paitotemporal reasoning analysis and make a decision of change left lane once the adjacent lane is free." + ], + "image_footnote": [], + "bbox": [ + 104, + 99, + 482, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hicle executes lane changes, objects in adjacent lanes, both left and right, may significantly influence its movement. Beyond obstacles, key traffic light $S_{light}$ , traffic signs $S_{sign}$ and lane markings $S_{mark}$ are also integral to decision-making, ensuring comprehensive spatial awareness.", + "bbox": [ + 99, + 405, + 483, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal CoT. While a single-frame image can provide a static representation of road scenes and traffic participants, it fails to capture the motion trends of moving agents. To address this limitation, we introduce consecutive frames $I = \\{I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br}\\}_{t = T_{now} - T}^{T_{now}}$ into the model, allowing it to track temporal variations in object positioning, which $I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br}$ represents the image view of front, front left, front right, back, back left and back right. These sequential frames not only offer instantaneous spatial context but also reveal motion trajectories and behavioral patterns through their inter-frame positional changes. This temporal information is essential for predicting dynamic object movement, assessing collision risks, and generating robust path planning strategies, ultimately enhancing the ability of anticipate and react to evolving traffic conditions.", + "bbox": [ + 99, + 496, + 483, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Decision Level CoT", + "text_level": 1, + "bbox": [ + 101, + 738, + 277, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output information from the perception-level serves as a critical foundation for the driving decision level, enabling reliable AD behavior inference. Within the driving decision level, we account for complex dynamic environmental factors, transforming spatial and temporal information from the perception level into concrete driving decisions. Specifically, decision-makers must not only analyze the current driving environment in real time but also anticipate future behaviors of other traffic participants and make decisions", + "bbox": [ + 99, + 757, + 483, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "based on multiple factors. Throughout this process, the CoT in the driving decision level spans multiple perspectives, incorporating safety, efficiency, comfort, and compliance with traffic regulations to ensure that driving decisions meet safety standards while optimizing driving efficiency.", + "bbox": [ + 511, + 102, + 893, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further enhance the structured reasoning process in AD, we have carefully designed hierarchical reasoning prompts that guide decision-making. Our structured prompts follow a logical sequence of \"risk point recognition — driving intention prediction — driving decision selection\", forming a cohesive reasoning chain aligned with human cognitive driving patterns.", + "bbox": [ + 511, + 176, + 893, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Risk Point Recognition. In this initial stage, the prompt-driven model conducts a comprehensive perception and analysis of the driving environment. This includes recognizing and evaluating critical elements such as traffic signs, lane markings, pedestrians, and obstacles to identify potential risks.", + "bbox": [ + 511, + 296, + 893, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Driving Intention Prediction. Once risk points are identified, the model leverages dynamic target behavior prediction and scene understanding to infer the potential movements and intentions of other road users. For instance, the model assesses whether pedestrians are likely to cross the road or whether the vehicle ahead intends to change lanes.", + "bbox": [ + 511, + 401, + 893, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Driving Decision Selection. Based on the contextual information gathered from the first two stages, the model applies multi-modal information fusion and weighted decision-making to select the optimal driving maneuver.", + "bbox": [ + 511, + 507, + 893, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this structured prompting strategy, the large model adheres to a progressive reasoning hierarchy, beginning with fundamental environmental perception and advancing to higher-level decision-making. By explicitly presenting the reasoning process in a clear and structured manner, this approach enhances passenger trust and confidence in the intelligent driving system. Moreover, the integration of recursive CoT reasoning enables the model to mimic the step-by-step thought process of human drivers, facilitating more flexible, reliable, and interpretable decision outputs in complex driving scenarios. The structured prompt framework are shown in Fig. 3.", + "bbox": [ + 511, + 566, + 893, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Paradigm", + "text_level": 1, + "bbox": [ + 511, + 750, + 691, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VLR-Driver adopts a dual-phase training strategy to optimize its reasoning and decision-making capabilities. In the first phase, LoRA [16] is utilized for supervised fine-tuning on a pre-trained large model, enabling efficient adaptation with minimal memory and computational overhead while maintaining strong performance. In the second phase, StepGRPO is applied for reinforcement learning based on human preferences, further enhancing the model's ability to", + "bbox": [ + 511, + 771, + 893, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26970", + "bbox": [ + 478, + 928, + 517, + 938 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "exhibit human-like reasoning and decision-making characteristics.", + "bbox": [ + 99, + 102, + 485, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training with LoRA. LoRA is a parameter-efficient finetuning technique that enables effective model adaptation by performing a low-rank decomposition of the weight matrix. The core principle behind LoRA is to decompose the weight matrix of a pre-trained model into a low-rank structure, significantly reducing the number of trainable parameters while preserving expressivity. Specifically, if the weight matrix in the pre-trained model is denoted as $W_0 \\in \\mathbb{R}^{d \\times k}$ , LoRA represents it as:", + "bbox": [ + 99, + 148, + 485, + 282 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} ^ {\\prime} = \\boldsymbol {W} _ {0} + \\Delta \\boldsymbol {W} = \\boldsymbol {W} _ {0} + \\frac {\\alpha}{r} \\boldsymbol {B} \\cdot \\boldsymbol {A}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 287, + 483, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{r \\times k}$ and $\\mathbf{B} \\in \\mathbb{R}^{d \\times r}$ , $r$ is the rank of a low rank matrix, usually much smaller than $d$ and $k$ , $\\alpha$ is a scaling factor. The forward pass is computed as:", + "bbox": [ + 99, + 320, + 483, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ny = \\boldsymbol {W} ^ {\\prime} x = \\left(\\boldsymbol {W _ {0}} + \\frac {\\alpha}{r} \\boldsymbol {B} \\cdot \\boldsymbol {A}\\right) x, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 372, + 483, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y$ is the output and $x$ is input.", + "bbox": [ + 101, + 404, + 336, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use LoRA for all linear modules, which not only saves computation but also ensures the performance of the model.", + "bbox": [ + 99, + 417, + 483, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training with GRPO. The core principle of GRPO [34] is to optimize strategies by assigning relative rewards to multiple outputs generated from the same prompt, thereby eliminating the need for additional value function models. The introduction of process reward model estimation in GRPO provides finer support for distributed rewards. The reward of each step of the outputs is:", + "bbox": [ + 99, + 478, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR = \\left\\{\\left\\{r _ {1} ^ {I (1)}, \\dots , r _ {1} ^ {I (K _ {1})} \\right\\}, \\dots , \\left\\{r _ {G} ^ {I (1)}, \\dots , r _ {G} ^ {I (K _ {G})} \\right\\} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 579, + 483, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I(\\cdot)$ is the end token index, and the reward need be normalized as:", + "bbox": [ + 101, + 608, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {r} _ {i} ^ {I (j)} = \\frac {r _ {i} ^ {I (j)} - \\operatorname {m e a n} (\\boldsymbol {R})}{\\operatorname {s t d} (\\boldsymbol {R})}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 642, + 483, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We extend GRPO by introducing reasoning Step-GRPO, a supervised reasoning decision process that structures the output into multiple steps based on the CoT reasoning framework. At each step, a reward function is applied to evaluate and assign scores, enabling fine-grained feedback that enhances model interpretability and accelerates convergence. Specifically, we first generate multiple candidate decision answers for the current driving scenario using prompts within the VLR model; Then, following our ST-CoT strategy, the reasoning process is divided into four distinct steps: scene description, spatial risk point reasoning, dynamic trajectory prediction, and driving decision-making; Furthermore, each reasoning step is assigned a reward to encourage structured learning. Finally, we compare", + "bbox": [ + 99, + 684, + 485, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "all answers within the group and calculate the Kullback-Leibler (KL) divergence to update the policy model. This grouping and step-by-step scoring strategy enhances training efficiency and reduces the likelihood of erroneous reasoning in the model.", + "bbox": [ + 511, + 102, + 895, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. VLR-Driver Dataset", + "text_level": 1, + "bbox": [ + 511, + 194, + 705, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To fully explore the reasoning and decision-making capabilities of LLMs, we propose an advanced reasoning and decision-making dataset for AD scenarios, called the VLR-Driver Datasets. This dataset relies on the CARLA [14] simulator for data collection and is expanded and meticulously annotated based on the Bench2Drive dataset [22]. It includes: a) multi-view multi-frame images or videos, b) valuable information for AD, such as road details, weather conditions, vehicle information, and scene descriptions, and c) driving decision choices along with the decision-making process based on an ST-CoT. This dataset provides a rich and comprehensive training foundation for AD reasoning and decision-making, allowing the agent to exhibit human-like reasoning while interacting with the environment. The dataset includes 20,000 sets of multi-frame, multi-angle image data collected from various road conditions such as urban, rural, and highways in the CARLA simulator, covering over 30 specific complex traffic scenarios (e.g., forward accidents, dynamic object crossings). Each image set provides road scene descriptions, environmental weather information, vehicle status data, and, most importantly, the humanlike reasoning process and driving behavior decisions.", + "bbox": [ + 511, + 219, + 895, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Data Collection", + "text_level": 1, + "bbox": [ + 511, + 556, + 666, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted data collection based on the 44 corner scene classifications provided by Bench2Drive to ensure optimal AD performance in various complex corner scenarios. The Bench2Drive dataset offers a rich array of data and annotations, including multi-angle images, lidar, radar, vehicle information, and expert assessments, which have played a crucial role in constructing our VLR dataset. Additionally, we selected over 30 scenarios where the reasoning and decision-making capabilities of AD systems are relatively limited, and used them as the visual and textual components of our dataset. We further expanded and enriched the dataset by collecting additional data in the CARLA simulator. We have taken into account various weather conditions, road conditions, and the types and numbers of traffic participants in the scene.", + "bbox": [ + 511, + 579, + 895, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weather. For each scene, we randomly set values for cloudiness, fog density, precipitation deposits, sun altitude angle, sun azimuth angle, wetness, and wind intensity. The combinations of these parameters cover a variety of weather conditions, such as rainy, foggy, broken sky, and stormy, as well as different lighting conditions for day and night.", + "bbox": [ + 511, + 801, + 895, + 891 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26971", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7dea093c1b16e5237328d5e8c79813c37d21b02a6fbfb33d9742faa988b6efbd.jpg", + "table_caption": [ + "Table 1. The comparison of core metrics and subdivision infraction scores with state-of-the-art E2E/VLM models on the Bench2Drive benchmark. C, L and T indicate camera, LiDAR and text modalities, respectively. DS, RC, IS, SR correspond to the Driving Score, Route Completion, Infraction Score, and Success Rate. CP, CV, CL, RL, SS, OR, AB, YEV correspond to the Collision with a Pedestrian, Collision with another Vehicle, Collision with Layout, Red Light infractions, Stop Sign infractions, Off-Road infractions, Agent Blocked, and failure to Yield to Emergency Vehicles infractions." + ], + "table_footnote": [], + "table_body": "
MethodTypeModalityCore Metrics ↑Subdivision Infraction Score ↓
DSRCISSRCPCVCLRLSSORABYEV
NEAT [10]E2EC30.8655.350.556.811.089.875.570.201.330.412.010.27
TCP [42]C56.2883.570.6525.000.265.465.460.000.520.220.780.00
LeTFuser [1]C+L52.5377.680.6718.181.165.543.790.290.870.120.580.29
LateFusion [28]C+L48.5358.320.8518.180.383.111.550.380.770.061.550.38
TransFuser [11]C+L37.1868.140.519.090.9613.248.710.000.960.322.580.32
ThinkTwice [21]C+L58.7974.350.7729.540.305.760.910.000.910.050.910.30
EATNet [9]C+L42.9778.840.5415.910.8214.011.920.001.640.221.090.27
InterFuser [31]C+L63.8180.460.7940.900.353.810.540.271.080.050.540.27
LMDrive [32]VLMC+L+T24.7633.020.9013.631.142.862.290.000.570.053.440.57
LeapAD [27]C+T55.1877.450.7136.360.695.071.150.200.910.081.470.27
VLR-Driver (Ours)VLRC+T75.0186.090.8750.000.722.830.480.000.480.030.250.25
", + "bbox": [ + 104, + 161, + 895, + 361 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c28dd97e53304017a3a946f4f28d9e9814eab470dd8b2b87985136893d02b58.jpg", + "table_caption": [ + "Table 2. The comparison of advanced driving ability and experience score with state-of-the-art models on the Bench2Drive benchmark. OT, MER, EB, GW, TS, DE, SC correspond to the OverTaking, MERging, Emergency Brake, Give Way, Traffic Sign, Driving Efficiency, and Smoothness Control." + ], + "table_footnote": [], + "table_body": "
MethodAdvanced Driving Ability ↑Exper. Score ↑
OTMEREBGWTSMeanDESC
NEAT [10]0.006.669.0950.0027.7718.7092.090.30
TCP [42]25.0013.3327.2750.0050.0033.12114.520.27
LeTFuser [1]0.0020.0018.180.0047.2217.08115.090.47
LateFusion [28]0.0020.009.090.0036.1113.04104.390.59
TransFuser [11]0.0013.339.090.0036.1111.7195.200.36
ThinkTwice [21]12.5020.0036.3650.0052.7734.3391.170.36
EATNet [9]0.0013.3318.180.0041.6614.6388.130.33
InterFuser [31]0.0046.6654.5450.0061.1142.46119.200.32
LMDrive [32]25.006.669.0950.002.7718.7075.410.22
LeapAD [27]12.5033.3327.2750.0044.4433.5193.330.26
VLR-Driver (Ours)37.5046.6672.7250.0072.2255.82125.220.59
", + "bbox": [ + 104, + 444, + 485, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Roads. The scenes include various road types such as urban two-way single-lane roads, multi-lane roads, highways, and narrow rural roads.", + "bbox": [ + 99, + 609, + 482, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Traffic Participants. Different corner cases involve various traffic participants, including cars, bicycles, pedestrians, ambulances, etc., simulating the complex traffic conditions encountered in daily driving.", + "bbox": [ + 99, + 654, + 483, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Data Annotation", + "text_level": 1, + "bbox": [ + 101, + 720, + 264, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The reasoning chain process and the decision-making choices are crucial for training large AD models and enhancing the model's human-like reasoning ability. Therefore, we performed secondary annotation on the dataset we collected.", + "bbox": [ + 99, + 742, + 482, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Driving Scene Descriptions. We used the pre-trained large visual language model Qwen2-VL [3] to generate detailed descriptions for the corresponding driving scenes. These descriptions primarily focus on environmental information such as road conditions, weather, and lighting, as", + "bbox": [ + 99, + 816, + 483, + 891 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "well as dynamic targets like vehicles and pedestrians that may pose driving risks.", + "bbox": [ + 511, + 385, + 893, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reasoning Decisions and Process. We considered pre-annotated information such as vehicle speed, acceleration, steering angle, traffic light status, and the state of the vehicle ahead. A rule-based method was used to determine the true values of future motion behaviors based on the decision choices made at earlier time steps. Additionally, we completed the predefined CoT reasoning text statements. Finally, to ensure the accuracy and consistency of the annotations, especially for decision choices, carefully manual verification was carried out.", + "bbox": [ + 511, + 415, + 895, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 575, + 635, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 599, + 697, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method was validated on the open source AD simulation platform CARLA 0.9.15 [14]. The VLR-Driver model was trained on a server equipped with 8 NVIDIA A800 GPUs (each with 80G of video memory) for approximately 50 hours. The dataset used was the VLR-Driver dataset that we developed. Specifically, we use ViT-g/14 from EVA-CLIP [29] as the vision encoder and LLaVA-NeXT-Video7B [50] as the VLM. The resolution of the input image is set to $336 \\times 336$ pixels.", + "bbox": [ + 511, + 622, + 895, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Metrics", + "text_level": 1, + "bbox": [ + 511, + 765, + 607, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ four core metrics to evaluate AD performance: driving score (DS), route completion (RC), infraction score (IS), and success rate (SR). Additionally, to provide a more granular assessment of model performance in specific aspects, we introduce three supplementary evaluation metrics: subdivision infraction score, advanced driving ability, and driving experience score [22].", + "bbox": [ + 511, + 786, + 895, + 890 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26972", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a16dfdb5fe44cee491dc6f5301800d3dda384dec0c2acf22c560d592e5f0c136.jpg", + "image_caption": [ + "Figure 4. Visual comparison between VLR-Driver and VLM-based methods. The ST-CoT guides the VLR model to approach driving decisions in a human-like spatiotemporal manner. Based on the sequence of images from the preceding time period $T$ , we derive the following inference results. The images captured at 5 and 10 seconds afterward validate the accuracy of our decisions." + ], + "image_footnote": [], + "bbox": [ + 102, + 99, + 480, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Comparisons with Existing Methods", + "text_level": 1, + "bbox": [ + 101, + 553, + 410, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted comprehensive experiments with the SOTA methods including E2E and VLM in the CARLA simulator across 44 routes from the Bench2Drive benchmark. We present comparison result in Tab. 1. It can be seen that our method outperforms other methods in key metrics such as DS, RC, and SR, achieving first place and effectively improving DS by $17.5\\%$ , mean advanced driving ability by $31.4\\%$ , and SR by $22.2\\%$ .", + "bbox": [ + 99, + 577, + 483, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The comparison results of the advanced driving ability and driving experience score of each method are shown in Tab. 2. Our method achieved the best results in all abilities, thanks to the deep reflection and reasoning ability of our VLR model, which has stronger traffic reasoning capacity in special road conditions. Most E2E methods can only achieve following the vehicle, but when there is a vehicle temporarily parked in the lane ahead, blocking the ego driving route, they will keep stopping and waiting, making it impossible to complete the entire route. In contrast, our VLR-Driver can achieve deep understanding and inference of the current scene through LLM inference, so as to make timely detours.", + "bbox": [ + 99, + 698, + 483, + 890 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/daecc8e1f82e8c434cdf24e4fcff3f6c09322b745d69785cfc547b43b4c205c8.jpg", + "table_caption": [ + "Table 3. Ablation study for each module." + ], + "table_footnote": [], + "table_body": "
IDAbal.Exp.Core Metrics ↑Advanced Driving Ability ↑
DSSROTMEREB
1VLRFull Model75.0150.0037.5046.6672.72
2Arch.w/o CoT57.1734.0925.0040.0041.66
3w/o VLR-Model46.3920.4514.2826.6636.36
4Trainw/o VLR-Data52.8527.2725.0026.669.09
5w/o Step-GRPO65.5745.4537.5040.0063.63
", + "bbox": [ + 516, + 107, + 893, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 231, + 661, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted a comprehensive ablation study, detailed in Tab. 3. The experimental configurations include four variants: (1) Without utilizing our proposed spatiotemporal CoT strategy, using only a question-based approach without reasoning guidance. (2) Without using our VLR model, instead employing a standard LLM module. (3) Removing the VLR data used to guide the reasoning process. (4) Without Step-GRPO reinforcement learning training, using only LoRA strategies to fine-tune the model with supervision. The results show the effectiveness of each contribution.", + "bbox": [ + 511, + 255, + 893, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Visualization", + "text_level": 1, + "bbox": [ + 511, + 417, + 645, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We selected some special scenarios to visualize the performance in complex traffic situations with reflective reasoning, and the comparison results shown in Fig. 4. The ST-CoT enables the model to make driving decisions with human-like reasoning, considering both spatial and temporal dynamics.", + "bbox": [ + 511, + 443, + 893, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 551, + 627, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce VLR-Driver, a VLR model for embodied AD. It leverages a carefully designed ST-CoT strategy to guide the model in recursively analyzing potential safety risks and the driving intentions of dynamic agents in complex traffic scenarios. Our dual-phase training method significantly enhances the generalization of the model. Additionally, we propose the VLR-Driver dataset, which effectively integrates spatiotemporal perception, language understanding, and reflective reasoning, providing crucial support for interpretability reasoning in AD system. Experimental results show that VLR-Driver outperforms other methods on Bench2Drive benchmark, achieving cutting-edge performance and paving the way for embodied AI realization.", + "bbox": [ + 511, + 578, + 893, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. There are differences between the data in simulation platforms and the real world. Deploying LLMs for AD on real-world vehicles with limited onboard computing resources remains a key challenge for future exploration.", + "bbox": [ + 511, + 816, + 893, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26973", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 101, + 194, + 116 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pedram Agand, Mohammad Mahdavianian, Manolis Savva, and Mo Chen. Letfuser: Light-weight end-to-end transformer-based sensor fusion for autonomous driving with multi-task learning. arXiv preprint arXiv:2310.13135, 2023. 2, 7", + "[2] Hidehisa Arai, Keita Miwa, Kento Sasaki, Kohei Watanabe, Yu Yamaguchi, Shunsuke Aoki, and Issei Yamamoto. Covla: Comprehensive vision-language-action dataset for autonomous driving. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1933-1943. IEEE, 2025. 1, 3", + "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 7", + "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3", + "[5] Daoming Chen, Ning Wang, Feng Chen, and Tony Pipe. Detrive: Imitation learning with transformer detection for end-to-end autonomous driving. arXiv preprint arXiv:2310.14224, 2023. 2", + "[6] Jianyu Chen, Zhuo Xu, and Masayoshi Tomizuka. End-to-end autonomous driving perception with sequential latent representation learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1999-2006. IEEE, 2020. 1", + "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2", + "[9] Weihuang Chen, Fanjie Kong, Liming Chen, Shen'ao Wang, Zhiping Wang, and Hongbin Sun. Eatnet: Efficient axial transformer network for end-to-end autonomous driving. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 762-767. IEEE, 2024. 7", + "[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 2, 7", + "[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE transactions on pattern analysis and machine intelligence, 45(11):12878-12895, 2022. 2, 7", + "[12] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the" + ], + "bbox": [ + 102, + 125, + 483, + 890 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1", + "[13] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. Quar-vla: Vision-language-action model for quadruped robots. In European Conference on Computer Vision, pages 352-367. Springer, 2024. 3", + "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, pages 1-16. PMLR, 2017. 6, 7", + "[15] Hao Gao, Shaoyu Chen, Bo Jiang, Bencheng Liao, Yiang Shi, Xiaoyang Guo, Yuechuan Pu, Haoran Yin, Xiangyu Li, Xinbang Zhang, et al. Rad: Training an end-to-end driving policy via large-scale 3dgs-based reinforcement learning. arXiv preprint arXiv:2502.13144, 2025. 2", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2, 5", + "[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2", + "[18] Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3903-3913, 2023. 1", + "[19] JJ Hwang, R Xu, H Lin, WC Hung, J Ji, K Choi, D Huang, T He, P Covington, B Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arxiv 2024. arXiv preprint arXiv:2410.23262. 3", + "[20] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 2", + "[21] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2, 7", + "[22] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 6, 7", + "[23] Xiaosong Jia, Junqi You, Zhiyuan Zhang, and Junchi Yan. Drivetransformer: Unified transformer for scalable end-to-end autonomous driving. arXiv preprint arXiv:2503.07656, 2025. 2", + "[24] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the" + ], + "bbox": [ + 516, + 103, + 893, + 890 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26974", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 2", + "[25] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 3", + "[26] Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, and Irwin King. A survey on vision-language-action models for embodied ai. arXiv preprint arXiv:2405.14093, 2024. 1, 3", + "[27] Jianbiao Mei, Yukai Ma, Xuemeng Yang, Licheng Wen, Xinyu Cai, Xin Li, Daocheng Fu, Bo Zhang, Pinlong Cai, Min Dou, et al. Continuously learning, adapting, and improving: A dual-process approach to autonomous driving. arXiv preprint arXiv:2405.15324, 2024. 7", + "[28] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. 2021. 7", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 4, 7", + "[30] H Sha, Y Mu, Y Jiang, L Chen, C Xu, P Luo, SE Li, M Tomizuka, W Zhan, and M Ding. Linguistics: Large language models as decision makers for autonomous driving. arxiv 2023. arXiv preprint arXiv:2310.03026. 4", + "[31] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 7", + "[32] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 7", + "[33] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 4", + "[34] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 2, 6", + "[35] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 3", + "[36] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models. arXiv preprint arXiv:2410.05051, 2024. 3" + ], + "bbox": [ + 104, + 103, + 483, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3", + "[38] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 2, 4", + "[39] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 2", + "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4", + "[41] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Philip S Yu. Multimodal large language models: A survey. In 2023 IEEE International Conference on Big Data (BigData), pages 2247-2256. IEEE, 2023. 1", + "[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2, 7", + "[43] Weishang Wu, Xiaoheng Deng, Ping Jiang, Shaohua Wan, and Yuanxiong Guo. Crossfuser: Multi-modal feature fusion for end-to-end autonomous driving under unseen weather conditions. IEEE Transactions on Intelligent Transportation Systems, 24(12):14378-14392, 2023. 2", + "[44] Shuo Xing, Chengyuan Qian, Yuping Wang, Hongyuan Hua, Kexin Tian, Yang Zhou, and Zhengzhong Tu. Openemma: Open-source multimodal model for end-to-end autonomous driving. In Proceedings of the Winter Conference on Applications of Computer Vision, pages 1001-1009, 2025. 4", + "[45] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3", + "[46] Z Yang, X Jia, H Li, and J Yan. Llm4drive: A survey of large language models for autonomous driving. arxiv 2023. arXiv preprint arXiv:2311.01043. 2", + "[47] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, et al. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006, 2023. 3", + "[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1", + "[49] Jiakai Zhang. End-to-End Learning for Autonomous Driving. PhD thesis, New York University, USA, 2019. 1" + ], + "bbox": [ + 514, + 103, + 893, + 888 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26975", + "bbox": [ + 478, + 928, + 517, + 938 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 4, 7", + "[51] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2", + "[52] Rui Zhao, Qirui Yuan, Jinyu Li, Haofeng Hu, Yun Li, Chengyuan Zheng, and Fei Gao. Sce2drivex: A generalized mllm framework for scene-to-drive learning. arXiv preprint arXiv:2502.14917, 2025. 2", + "[53]Zikai Zhao, Qiuxuan Wu, Jian Wang, Botao Zhang, Chaoliang Zhong, and Anton A Zhilenkov. Exploring embodied intelligence in soft robotics: a review. Biomimetics, 9(4): 248, 2024. 1", + "[54] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 2", + "[55] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied understanding of driving scenarios. In European Conference on Computer Vision, pages 129-148. Springer, 2024. 1" + ], + "bbox": [ + 102, + 103, + 486, + 459 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26976", + "bbox": [ + 478, + 928, + 517, + 939 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_model.json b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_model.json new file mode 100644 index 0000000000000000000000000000000000000000..686027f9717266127f1cd3fcebc9b5d2ac74c6cd --- /dev/null +++ b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.153, + 0.14, + 0.849, + 0.186 + ], + "angle": 0, + "content": "VLR-Driver: Large Vision-Language-Reasoning Models for Embodied Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.211, + 0.782, + 0.247 + ], + "angle": 0, + "content": "Fanjie Kong\\(^{1}\\) Yitong Li\\(^{1}\\) Weihuang Chen\\(^{1*}\\) Chen Min\\(^{2}\\) Yizhe Li\\(^{3}\\) Zhiqiang Gao\\(^{5,1}\\) Haoyang Li\\(^{4}\\) Zhongyu Guo\\(^{1}\\) Hongbin Sun\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.247, + 0.871, + 0.283 + ], + "angle": 0, + "content": "\\(^{1}\\)Xi'an Jiaotong University \\(^{2}\\)Institute of Computing Technology, Chinese Academy of Sciences \\(^{3}\\)University of Maryland, College Park \\(^{4}\\)Nanchang University \\(^{5}\\)Engineering University of PAP" + }, + { + "type": "title", + "bbox": [ + 0.257, + 0.315, + 0.333, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.351, + 0.486, + 0.763 + ], + "angle": 0, + "content": "The rise of embodied intelligence and multi-modal large language models has led to exciting advancements in the field of autonomous driving, establishing it as a prominent research focus in both academia and industry. However, when confronted with intricate and ambiguous traffic scenarios, the lack of logical reasoning and cognitive decision-making capabilities remains the primary challenge impeding the realization of embodied autonomous driving. Although Vision Language Models (VLMs) have enhanced the deep semantic understanding of autonomous driving systems, they exhibit notable limitations in decision explainability when handling rare and long-tail traffic scenarios. In this paper, we propose VLR-Driver, a novel multimodal Vision-Language-Reasoning (VLR) framework based on Chain of Thought (CoT) for embodied autonomous driving. The framework employs a spatiotemporal CoT reasoning approach to recursively analyze potential safety risks and driving intentions of other agents, thereby delivering an efficient and transparent decision-making process. Furthermore, we construct a multi-modal reasoning-decision dataset to support the advancement of hierarchical reasoning of VLMs in autonomous driving. Closed-loop experiments conducted in CARLA demonstrate that the VLR-Driver significantly outperforms state-of-the-art end-to-end methods. Notably, key metrics such as driving score improved by \\(17.5\\%\\), while the success rate improved by \\(22.2\\%\\), offering a more transparent, reliable, and secure solution for autonomous driving systems." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.314, + 0.897, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.383, + 0.713, + 0.392 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.395, + 0.897, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.465, + 0.713, + 0.474 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.477, + 0.897, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.547, + 0.712, + 0.556 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.567, + 0.899, + 0.662 + ], + "angle": 0, + "content": "Figure 1. Comparison of different LLM-based AD systems. (a) VLM model focuses more on answering questions, but it is difficult to convert decisions into coherent control signals. (b) VLA model focuses more on predicting driving control signals, but lacks interpretability. (c) The VLR model generates decisions and controls signals with transparent reasoning processes through CoT, enhancing the driver trust in the system." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.69, + 0.644, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.899, + 0.892 + ], + "angle": 0, + "content": "In recent years, the rapid progress of End-to-End (E2E) architectures [6, 7, 18, 49], Large Language Models (LLMs) [12, 41, 48], and embodied intelligence [26, 53, 55] has established these technologies as key enablers of innovation in Autonomous Driving (AD). Especially, VLMs enriched with extensive pre-training knowledge exhibit strong spatial understanding and common-sense reasoning abilities. DriveVLM [35] leverages VLM to enhance spatial awareness and planning capabilities in complex driving scenarios. CoVLA [2] integrates visual perception, language understanding, and action planning, effectively describing traffic scenarios and generating executable control actions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.12, + 0.808, + 0.245, + 0.82 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.82, + 0.486, + 0.89 + ], + "angle": 0, + "content": "This research was supported in part by National Natural Science Foundation of China (No. 62088102, No. U24A20291), National Key Research and Development Program of China (No. 2024YFB4405504), the Fundamental Research Funds for the Central Universities (xtr072022001), the Opening Foundation of Key Laboratory of CTCIE, Ministry of Education (CTCIE20240101)." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.808, + 0.486, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.519, + 0.94 + ], + "angle": 0, + "content": "26966" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.482, + 0.295 + ], + "angle": 0, + "content": "However, the decision-making process of VLMs often functions as a \"black box\", making it challenging to trace and interpret their underlying logic. This makes it difficult for AD systems to be fully trusted by drivers when encountering complex and emergency situations, such as illegal roadside parking, navigating intersections without traffic signals, and managing complex mixed-traffic interactions between motorized and non-motorized vehicles, thereby limiting their reliability and safety in real-world applications [39, 46, 54]. Moreover, most VLMs are trained on internet data, lacking spatial understanding and specialized training in the field of AD, making it difficult for them to fully adapt to dynamic and complex driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.298, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Meanwhile, CoT reasoning demonstrates strong inference, interpretability, and generalization capabilities by breaking down complex tasks into intermediate reasoning steps [40]. CoT enables systems to think step by step rather than relying on E2E black-box predictions, making it one of the key approaches toward achieving embodied AD. Sce2DriveX [52] enhances comprehensive perception and reasoning by introducing a multi-modal LLM framework with CoT, enabling a deeper understanding of spatiotemporal relationships and road topology. Similarly, DriveCoT [38] integrates CoT reasoning to improve decision-making interpretability and controllability in AD systems. However, existing CoT-based methods typically rely on open-ended language generation for reasoning, which lacks structural constraints. The generated reasoning steps may exhibit causal inversion or unnecessary vocabulary, making them uncontrollable in safety-critical tasks. Additionally, current methods primarily operate on static snapshots rather than continuous temporal sequences, limiting their ability to predict future events in dynamic traffic environments." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.596, + 0.483, + 0.892 + ], + "angle": 0, + "content": "To bridge these gaps, we introduce VLR-Driver, a hierarchical CoT-based visual-language-reasoning model designed for closed-loop embodied AD. Our approach integrates the spatiotemporal features from cross-modal data, including multi-frame multi-view images and ego-vehicle control signals, by employing a SpatioTemporal CoT (STCoT) strategy that produces human-like reflective reasoning processes and driving action decisions. Additionally, we adopt a dual-phase training strategy, combining Low-Rank Adaptation (LoRA) [16] with an improved Stepwise Group Relative Policy Optimization (Step-GRPO) [34], significantly enhancing memory capacity and deep reasoning abilities. Our proposed VLR-Driver not only inherits the global action optimization capability of VLA models, but also preserves the transparency of modular rule-based methods. When encountering long-tail events and rare traffic scenarios, it demonstrates exceptional reflective reasoning and step-by-step inference, thereby enhancing the trust of human drivers in AD systems. The differences between VLM, VLA, and our proposed VLR are illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.895, + 0.235 + ], + "angle": 0, + "content": "To further enhance VLR models in environmental understanding, reasoning, and decision-making, we introduce the VLR-Driver Dataset. This data set includes detailed scene descriptions, weather information, vehicle state details, and most critically, human-like CoT reasoning processes and the corresponding driving decisions. Various experiments demonstrate that VLR-Driver is capable of making accurate driving decisions and coherent reasoning, even under challenging and highly dynamic road conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.895, + 0.265 + ], + "angle": 0, + "content": "The primary contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.267, + 0.895, + 0.34 + ], + "angle": 0, + "content": "- Distinctive VLR-Driver Framework. We introduce VLR-Driver, a visual-language-reasoning model developed for embodied AD. It generates a human-like reflective reasoning process within the driving system, enabling accurate driving decisions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.34, + 0.895, + 0.413 + ], + "angle": 0, + "content": "- Spatiotemporal CoT. We present a spatiotemporal CoT strategy that recursively analyzes potential safety risks and the driving intentions of moving agents in complex traffic scenarios, ensuring that the reasoning process of driving decisions remains transparent and interpretable." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.414, + 0.894, + 0.501 + ], + "angle": 0, + "content": "- Advanced VLR-Driver Dataset. We construct VLR-Driver Dataset, a cutting-edge visual-language-reasoning and decision-making dataset specifically designed for AD. It supports the enhancement of spatiotemporal understanding and reflective reasoning capabilities in embodied AD systems." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.502, + 0.894, + 0.59 + ], + "angle": 0, + "content": "- Superior Performance in Closed-Loop Simulations. Extensive closed-loop experiments are conducted on the CARLA platform. On the Bench2Drive benchmark, our approach achieves a \\(17.5\\%\\) improvement in driving score and a \\(22.2\\%\\) increase in success rate, providing a more human-like, reliable, and trustworthy solution." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.267, + 0.895, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.652, + 0.622 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.794, + 0.648 + ], + "angle": 0, + "content": "2.1. End-to-end Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.895, + 0.891 + ], + "angle": 0, + "content": "The rapid advancement of E2E-AD has fostered a growing transition away from modular rule-based methods toward data-driven approaches [15, 23]. Based on different input modalities, E2E methods can be categorized into visual-only methods [5, 8, 10, 17] and vision-LiDAR fusion methods [1, 20, 21]. TCP [42] and NEAT [10] adopt imitation learning methods, training directly on collected state-action pair datasets, demonstrating the feasibility of E2E approaches for AD. Roach [51] utilizes reinforcement learning experts as coaches, delivering dense and informative supervision signals to agents equipped with monocular camera inputs. However, vision-based methods inherently struggle with distance and depth estimation. These limitations may compromise the reliability of driving decisions [24]. To address these limitations, vision-LiDAR fusion methods such as TransFuser [11], CrossFuser [43]," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.94 + ], + "angle": 0, + "content": "26967" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.104, + 0.182, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.105, + 0.361, + 0.119 + ], + "angle": 0, + "content": "VLR-Driver Datasets" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.123, + 0.432, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.103, + 0.612, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.617, + 0.105, + 0.742, + 0.119 + ], + "angle": 0, + "content": "VLR-Driver EAI" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.123, + 0.895, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.445, + 0.897, + 0.501 + ], + "angle": 0, + "content": "Figure 2. Overview of VLR-Driver framework. We introduce VLR-Driver Dataset, an advanced visual-language-reasoning dataset designed for AD, featuring detailed annotations of scene descriptions, analytical reasoning, and behavioral decisions. We present VLR-Driver Embodied AI (EAI), a novel multi-modal visual-language-reasoning framework for embodied AD that leverages a hierarchical spatiotemporal CoT reasoning mechanism." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.521, + 0.484, + 0.625 + ], + "angle": 0, + "content": "and FusionAD [47] have been designed to effectively integrate image with LiDAR data, significantly improving the robustness of AD systems in complex environments. However, different modalities contribute differently to the driving task, and spatiotemporal synchronization issues between sensor data can lead to inconsistencies, increasing both model complexity and training difficulty." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.633, + 0.431, + 0.649 + ], + "angle": 0, + "content": "2.2. VLM and VLA in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.655, + 0.484, + 0.875 + ], + "angle": 0, + "content": "VLMs unify visual perception with natural language processing capabilities, enabling a more comprehensive understanding of driving environments [19, 36, 37]. By incorporating cross-modal data fusion, integrating textual information alongside visual inputs, these systems gain text comprehension and human interaction capabilities that conventional E2E models inherently lack. DriveGPT4 [45] enhances LLMs' ability to process multimodal inputs by projecting them into the text domain, thereby enabling interpretable E2E-AD. Senna [25] and DriveVLM [35] integrate VLMs with traditional modular pipelines or E2E frameworks, achieving decoupled high-level planning and low-level trajectory prediction. This approach improves planning performance while preserving the model's common sense reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.876, + 0.483, + 0.89 + ], + "angle": 0, + "content": "Moreover, VLA is an emerging paradigm that unifies vi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.896, + 0.873 + ], + "angle": 0, + "content": "sual perception, natural language understanding, and action prediction within a cohesive framework [13, 26]. Originally introduced in the field of robotics, RT-2 [4] pioneered the representation of robotic actions as text tokens, seamlessly incorporating them alongside natural language labels into the model's training set. This enables direct transfer of internet-scale knowledge to robotic control, significantly enhancing both the generalization and semantic reasoning capabilities of robotic systems. For AD, CoVLA [2] introduces an interpretable VLA model that integrates visual perception, language-based scene understanding, and action planning. This integration improves the system's ability to comprehend complex driving scenarios, anticipate trajectory, and execute informed driving decisions. However, despite these advancements, VLA models still exhibit limitations in accurately predicting precise control action values, leading to a low lower-bound in AD performance. Furthermore, when confronted with highly dynamic and complex traffic scenarios, their deep understanding and reasoning capabilities remain insufficient. Unlike the above methods, our proposed VLR model places greater emphasis on step-by-step reasoning and the thought process of the model, providing drivers with increased confidence in using AD systems." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.939 + ], + "angle": 0, + "content": "26968" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.102, + 0.102, + 0.451, + 0.119 + ], + "angle": 0, + "content": "2.3. Chain of Thought in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.124, + 0.486, + 0.464 + ], + "angle": 0, + "content": "The CoT technology is an extension of prompt engineering, proposed by Wei Jason in 2022 [40], which has greatly improved the effectiveness of reasoning for complex problems. The highly anticipated Deepseek-R1 [33] model also utilizes the CoT technique, which deeply integrates multimodal knowledge base data, enabling the model to generate a step-by-step thinking process. DriveCoT [38] has built a CoT dataset that includes sensor data, control decisions, and CoT labels used to indicate reasoning processes. The model trained on this basis can generate predictions and final decisions with CoT, effectively improving model performance. LanguageMPC [30] combines LLM with Model Predictive Control and decomposes driving decisions into multiple subtasks through a CoT framework. This method enables the auto drive system to think like human beings, and improves its ability to handle complex scenes. OpenEMMA [44] introduces CoT technology to guide model generation of detailed descriptions of key objects, behavioral insights, and meta driving decisions, improving system transparency and usability. Motivated by these advancements, we apply step-by-step hierarchical spatiotemporal CoT to AD, enhancing the interpretability of reasoning and decision-making." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.478, + 0.192, + 0.492 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.503, + 0.485, + 0.637 + ], + "angle": 0, + "content": "We present the motivation and design details of our VLR-Driver framework. As depicted in Fig. 2, VLR-Driver comprises two main components: the VLR-Driver EAI and the VLR-Driver Dataset. Initially, we introduce the design concept of the VLR model, which builds upon enhancements to the VLA model (Sec. 3.1). Subsequently, we elaborate on the hierarchical spatiotemporal CoT methodology (Sec. 3.2) and the specifics of the dual-phase training strategy (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.647, + 0.21, + 0.66 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.669, + 0.485, + 0.875 + ], + "angle": 0, + "content": "The VLR model is a large visual-language-reasoning model designed for embodied AD. It can processes visual inputs, such as multi-view images, alongside textual information, including vehicle control signals. The model is capable of extracting spatiotemporal key features within a multimodal embedding and recursively analyze potential safety risks and the driving intentions of other agents within the reasoning level of the VLR model. Ultimately, it formulates a comprehensive reasoning framework and well-structured decision-making outputs, explicitly identifying critical risk factors and the underlying rationale behind each decision. This process significantly reinforces the robustness and safety of the AD system, ensuring adaptive resilience in complex and dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.875, + 0.483, + 0.89 + ], + "angle": 0, + "content": "In this study, we employ pre-trained LLaVA-NeXT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.895, + 0.176 + ], + "angle": 0, + "content": "Video [50] as the VLM and CLIP [29] as the visual encoder. The model is capable of processing multi-frame multi-view image data that capture historical temporal context, while also extracting real-time vehicle sensor information to facilitate dynamic and context-aware decision-making." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.177, + 0.896, + 0.294 + ], + "angle": 0, + "content": "Input Representations. We utilize \\(N_{f}\\) frame and \\(N_{v}\\) view images from the past period, with a field of view (FoV) of 70 degrees. It can be represented as \\(V \\in \\mathbb{R}^{N_f \\times N_v \\times 3 \\times H_0 \\times W_0}\\). At the same time, there are also the current position \\((x,y)\\) of ego vehicle, the speed \\(v\\), the target point position \\((p,q)\\). Subsequently, the compressed and cropped image data and the information from the ego's sensors are input into the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.897, + 0.427 + ], + "angle": 0, + "content": "Output Representations. Our output consists of the reasoning process and driving decisions generated by the VLR model for the current driving scenario. This includes risk identification, traffic signal recognition, motion direction prediction, and AD decision-making. The driving decision will also be fused with the spatiotemporal information features extracted by the E2E model, and finally output the waypoints and control signals of the vehicle for the next moment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.785, + 0.454 + ], + "angle": 0, + "content": "3.2. Spatiotemporal CoT Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.897, + 0.607 + ], + "angle": 0, + "content": "To enhance the reasoning capabilities and transparency of the AD system, we introduce a hierarchical ST-CoT that guides the model to approach driving decisions in a human-like manner. Our method decomposes the driving decision-making process into two levels: a perception-level spatiotemporal CoT \\( C_{\\text{perception}} \\), which focuses on extracting and understanding environmental dynamics, and a decision-level dynamic CoT \\( C_{\\text{decision}} \\), which refines and optimizes decision-making based on contextual and temporal factors. The example of CoT is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.707, + 0.63 + ], + "angle": 0, + "content": "3.2.1. Perception Level CoT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.896, + 0.797 + ], + "angle": 0, + "content": "The spatiotemporal characteristics of the environment play a crucial role in the AD process. The perception level CoT is responsible for guiding the model to extract spatial and temporal features from input image data, identify and locate crucial objects in the traffic, such as vehicles, pedestrians, alien objects, traffic lights, traffic signs, etc., and extract historical behavioral features of dynamic agent based on temporal information. Our method enables VLR model to describe the current driving scenario, construct real-time spatial layout and dynamic changes of the environment, and achieve long-term planning for driving decisions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.896, + 0.892 + ], + "angle": 0, + "content": "Spatial CoT. In driving scenarios, we primarily focus on obstacles that impact the ego's normal operation, including object categories \\( O_{type} \\) and distances with the ego \\( O_{dis} \\). A critical aspect of safe driving is identifying potential risk points within the current lane. Additionally, when the ve" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.519, + 0.94 + ], + "angle": 0, + "content": "26969" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.101, + 0.483, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.317, + 0.484, + 0.384 + ], + "angle": 0, + "content": "Figure 3. Illustration of the ST-CoT reasoning process. In this scenario, where some vehicles are illegally parked ahead and blocking the lane, our method can conduct hierarchical paitotemporal reasoning analysis and make a decision of change left lane once the adjacent lane is free." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.406, + 0.484, + 0.48 + ], + "angle": 0, + "content": "hicle executes lane changes, objects in adjacent lanes, both left and right, may significantly influence its movement. Beyond obstacles, key traffic light \\( S_{light} \\), traffic signs \\( S_{sign} \\) and lane markings \\( S_{mark} \\) are also integral to decision-making, ensuring comprehensive spatial awareness." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.497, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Temporal CoT. While a single-frame image can provide a static representation of road scenes and traffic participants, it fails to capture the motion trends of moving agents. To address this limitation, we introduce consecutive frames \\( I = \\{I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br}\\}_{t = T_{now} - T}^{T_{now}} \\) into the model, allowing it to track temporal variations in object positioning, which \\( I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br} \\) represents the image view of front, front left, front right, back, back left and back right. These sequential frames not only offer instantaneous spatial context but also reveal motion trajectories and behavioral patterns through their inter-frame positional changes. This temporal information is essential for predicting dynamic object movement, assessing collision risks, and generating robust path planning strategies, ultimately enhancing the ability of anticipate and react to evolving traffic conditions." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.739, + 0.279, + 0.753 + ], + "angle": 0, + "content": "3.2.2. Decision Level CoT" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.758, + 0.484, + 0.89 + ], + "angle": 0, + "content": "The output information from the perception-level serves as a critical foundation for the driving decision level, enabling reliable AD behavior inference. Within the driving decision level, we account for complex dynamic environmental factors, transforming spatial and temporal information from the perception level into concrete driving decisions. Specifically, decision-makers must not only analyze the current driving environment in real time but also anticipate future behaviors of other traffic participants and make decisions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.894, + 0.176 + ], + "angle": 0, + "content": "based on multiple factors. Throughout this process, the CoT in the driving decision level spans multiple perspectives, incorporating safety, efficiency, comfort, and compliance with traffic regulations to ensure that driving decisions meet safety standards while optimizing driving efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.177, + 0.895, + 0.28 + ], + "angle": 0, + "content": "To further enhance the structured reasoning process in AD, we have carefully designed hierarchical reasoning prompts that guide decision-making. Our structured prompts follow a logical sequence of \"risk point recognition — driving intention prediction — driving decision selection\", forming a cohesive reasoning chain aligned with human cognitive driving patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.894, + 0.383 + ], + "angle": 0, + "content": "Risk Point Recognition. In this initial stage, the prompt-driven model conducts a comprehensive perception and analysis of the driving environment. This includes recognizing and evaluating critical elements such as traffic signs, lane markings, pedestrians, and obstacles to identify potential risks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.894, + 0.491 + ], + "angle": 0, + "content": "Driving Intention Prediction. Once risk points are identified, the model leverages dynamic target behavior prediction and scene understanding to infer the potential movements and intentions of other road users. For instance, the model assesses whether pedestrians are likely to cross the road or whether the vehicle ahead intends to change lanes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.894, + 0.567 + ], + "angle": 0, + "content": "Driving Decision Selection. Based on the contextual information gathered from the first two stages, the model applies multi-modal information fusion and weighted decision-making to select the optimal driving maneuver." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.895, + 0.743 + ], + "angle": 0, + "content": "Through this structured prompting strategy, the large model adheres to a progressive reasoning hierarchy, beginning with fundamental environmental perception and advancing to higher-level decision-making. By explicitly presenting the reasoning process in a clear and structured manner, this approach enhances passenger trust and confidence in the intelligent driving system. Moreover, the integration of recursive CoT reasoning enables the model to mimic the step-by-step thought process of human drivers, facilitating more flexible, reliable, and interpretable decision outputs in complex driving scenarios. The structured prompt framework are shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.751, + 0.692, + 0.767 + ], + "angle": 0, + "content": "3.3. Training Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.895, + 0.89 + ], + "angle": 0, + "content": "VLR-Driver adopts a dual-phase training strategy to optimize its reasoning and decision-making capabilities. In the first phase, LoRA [16] is utilized for supervised fine-tuning on a pre-trained large model, enabling efficient adaptation with minimal memory and computational overhead while maintaining strong performance. In the second phase, StepGRPO is applied for reinforcement learning based on human preferences, further enhancing the model's ability to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.939 + ], + "angle": 0, + "content": "26970" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.132 + ], + "angle": 0, + "content": "exhibit human-like reasoning and decision-making characteristics." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.149, + 0.486, + 0.283 + ], + "angle": 0, + "content": "Training with LoRA. LoRA is a parameter-efficient finetuning technique that enables effective model adaptation by performing a low-rank decomposition of the weight matrix. The core principle behind LoRA is to decompose the weight matrix of a pre-trained model into a low-rank structure, significantly reducing the number of trainable parameters while preserving expressivity. Specifically, if the weight matrix in the pre-trained model is denoted as \\( W_0 \\in \\mathbb{R}^{d \\times k} \\), LoRA represents it as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.289, + 0.484, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} ^ {\\prime} = \\boldsymbol {W} _ {0} + \\Delta \\boldsymbol {W} = \\boldsymbol {W} _ {0} + \\frac {\\alpha}{r} \\boldsymbol {B} \\cdot \\boldsymbol {A}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.321, + 0.485, + 0.368 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} \\in \\mathbb{R}^{r \\times k}\\) and \\(\\mathbf{B} \\in \\mathbb{R}^{d \\times r}\\), \\(r\\) is the rank of a low rank matrix, usually much smaller than \\(d\\) and \\(k\\), \\(\\alpha\\) is a scaling factor. The forward pass is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.373, + 0.484, + 0.4 + ], + "angle": 0, + "content": "\\[\ny = \\boldsymbol {W} ^ {\\prime} x = \\left(\\boldsymbol {W _ {0}} + \\frac {\\alpha}{r} \\boldsymbol {B} \\cdot \\boldsymbol {A}\\right) x, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.405, + 0.338, + 0.419 + ], + "angle": 0, + "content": "where \\(y\\) is the output and \\(x\\) is input." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.419, + 0.484, + 0.463 + ], + "angle": 0, + "content": "We use LoRA for all linear modules, which not only saves computation but also ensures the performance of the model." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.479, + 0.484, + 0.576 + ], + "angle": 0, + "content": "Training with GRPO. The core principle of GRPO [34] is to optimize strategies by assigning relative rewards to multiple outputs generated from the same prompt, thereby eliminating the need for additional value function models. The introduction of process reward model estimation in GRPO provides finer support for distributed rewards. The reward of each step of the outputs is:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.58, + 0.484, + 0.605 + ], + "angle": 0, + "content": "\\[\nR = \\left\\{\\left\\{r _ {1} ^ {I (1)}, \\dots , r _ {1} ^ {I (K _ {1})} \\right\\}, \\dots , \\left\\{r _ {G} ^ {I (1)}, \\dots , r _ {G} ^ {I (K _ {G})} \\right\\} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.609, + 0.483, + 0.639 + ], + "angle": 0, + "content": "where \\(I(\\cdot)\\) is the end token index, and the reward need be normalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.643, + 0.484, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\widetilde {r} _ {i} ^ {I (j)} = \\frac {r _ {i} ^ {I (j)} - \\operatorname {m e a n} (\\boldsymbol {R})}{\\operatorname {s t d} (\\boldsymbol {R})}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.685, + 0.486, + 0.892 + ], + "angle": 0, + "content": "We extend GRPO by introducing reasoning Step-GRPO, a supervised reasoning decision process that structures the output into multiple steps based on the CoT reasoning framework. At each step, a reward function is applied to evaluate and assign scores, enabling fine-grained feedback that enhances model interpretability and accelerates convergence. Specifically, we first generate multiple candidate decision answers for the current driving scenario using prompts within the VLR model; Then, following our ST-CoT strategy, the reasoning process is divided into four distinct steps: scene description, spatial risk point reasoning, dynamic trajectory prediction, and driving decision-making; Furthermore, each reasoning step is assigned a reward to encourage structured learning. Finally, we compare" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.896, + 0.178 + ], + "angle": 0, + "content": "all answers within the group and calculate the Kullback-Leibler (KL) divergence to update the policy model. This grouping and step-by-step scoring strategy enhances training efficiency and reduces the likelihood of erroneous reasoning in the model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.707, + 0.21 + ], + "angle": 0, + "content": "4. VLR-Driver Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.897, + 0.545 + ], + "angle": 0, + "content": "To fully explore the reasoning and decision-making capabilities of LLMs, we propose an advanced reasoning and decision-making dataset for AD scenarios, called the VLR-Driver Datasets. This dataset relies on the CARLA [14] simulator for data collection and is expanded and meticulously annotated based on the Bench2Drive dataset [22]. It includes: a) multi-view multi-frame images or videos, b) valuable information for AD, such as road details, weather conditions, vehicle information, and scene descriptions, and c) driving decision choices along with the decision-making process based on an ST-CoT. This dataset provides a rich and comprehensive training foundation for AD reasoning and decision-making, allowing the agent to exhibit human-like reasoning while interacting with the environment. The dataset includes 20,000 sets of multi-frame, multi-angle image data collected from various road conditions such as urban, rural, and highways in the CARLA simulator, covering over 30 specific complex traffic scenarios (e.g., forward accidents, dynamic object crossings). Each image set provides road scene descriptions, environmental weather information, vehicle status data, and, most importantly, the humanlike reasoning process and driving behavior decisions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.667, + 0.571 + ], + "angle": 0, + "content": "4.1. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.896, + 0.8 + ], + "angle": 0, + "content": "We conducted data collection based on the 44 corner scene classifications provided by Bench2Drive to ensure optimal AD performance in various complex corner scenarios. The Bench2Drive dataset offers a rich array of data and annotations, including multi-angle images, lidar, radar, vehicle information, and expert assessments, which have played a crucial role in constructing our VLR dataset. Additionally, we selected over 30 scenarios where the reasoning and decision-making capabilities of AD systems are relatively limited, and used them as the visual and textual components of our dataset. We further expanded and enriched the dataset by collecting additional data in the CARLA simulator. We have taken into account various weather conditions, road conditions, and the types and numbers of traffic participants in the scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.897, + 0.892 + ], + "angle": 0, + "content": "Weather. For each scene, we randomly set values for cloudiness, fog density, precipitation deposits, sun altitude angle, sun azimuth angle, wetness, and wind intensity. The combinations of these parameters cover a variety of weather conditions, such as rainy, foggy, broken sky, and stormy, as well as different lighting conditions for day and night." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.94 + ], + "angle": 0, + "content": "26971" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.083, + 0.897, + 0.152 + ], + "angle": 0, + "content": "Table 1. The comparison of core metrics and subdivision infraction scores with state-of-the-art E2E/VLM models on the Bench2Drive benchmark. C, L and T indicate camera, LiDAR and text modalities, respectively. DS, RC, IS, SR correspond to the Driving Score, Route Completion, Infraction Score, and Success Rate. CP, CV, CL, RL, SS, OR, AB, YEV correspond to the Collision with a Pedestrian, Collision with another Vehicle, Collision with Layout, Red Light infractions, Stop Sign infractions, Off-Road infractions, Agent Blocked, and failure to Yield to Emergency Vehicles infractions." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.162, + 0.897, + 0.362 + ], + "angle": 0, + "content": "
MethodTypeModalityCore Metrics ↑Subdivision Infraction Score ↓
DSRCISSRCPCVCLRLSSORABYEV
NEAT [10]E2EC30.8655.350.556.811.089.875.570.201.330.412.010.27
TCP [42]C56.2883.570.6525.000.265.465.460.000.520.220.780.00
LeTFuser [1]C+L52.5377.680.6718.181.165.543.790.290.870.120.580.29
LateFusion [28]C+L48.5358.320.8518.180.383.111.550.380.770.061.550.38
TransFuser [11]C+L37.1868.140.519.090.9613.248.710.000.960.322.580.32
ThinkTwice [21]C+L58.7974.350.7729.540.305.760.910.000.910.050.910.30
EATNet [9]C+L42.9778.840.5415.910.8214.011.920.001.640.221.090.27
InterFuser [31]C+L63.8180.460.7940.900.353.810.540.271.080.050.540.27
LMDrive [32]VLMC+L+T24.7633.020.9013.631.142.862.290.000.570.053.440.57
LeapAD [27]C+T55.1877.450.7136.360.695.071.150.200.910.081.470.27
VLR-Driver (Ours)VLRC+T75.0186.090.8750.000.722.830.480.000.480.030.250.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.366, + 0.486, + 0.435 + ], + "angle": 0, + "content": "Table 2. The comparison of advanced driving ability and experience score with state-of-the-art models on the Bench2Drive benchmark. OT, MER, EB, GW, TS, DE, SC correspond to the OverTaking, MERging, Emergency Brake, Give Way, Traffic Sign, Driving Efficiency, and Smoothness Control." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.445, + 0.486, + 0.589 + ], + "angle": 0, + "content": "
MethodAdvanced Driving Ability ↑Exper. Score ↑
OTMEREBGWTSMeanDESC
NEAT [10]0.006.669.0950.0027.7718.7092.090.30
TCP [42]25.0013.3327.2750.0050.0033.12114.520.27
LeTFuser [1]0.0020.0018.180.0047.2217.08115.090.47
LateFusion [28]0.0020.009.090.0036.1113.04104.390.59
TransFuser [11]0.0013.339.090.0036.1111.7195.200.36
ThinkTwice [21]12.5020.0036.3650.0052.7734.3391.170.36
EATNet [9]0.0013.3318.180.0041.6614.6388.130.33
InterFuser [31]0.0046.6654.5450.0061.1142.46119.200.32
LMDrive [32]25.006.669.0950.002.7718.7075.410.22
LeapAD [27]12.5033.3327.2750.0044.4433.5193.330.26
VLR-Driver (Ours)37.5046.6672.7250.0072.2255.82125.220.59
" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.611, + 0.483, + 0.654 + ], + "angle": 0, + "content": "Roads. The scenes include various road types such as urban two-way single-lane roads, multi-lane roads, highways, and narrow rural roads." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.655, + 0.485, + 0.715 + ], + "angle": 0, + "content": "Traffic Participants. Different corner cases involve various traffic participants, including cars, bicycles, pedestrians, ambulances, etc., simulating the complex traffic conditions encountered in daily driving." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.722, + 0.265, + 0.737 + ], + "angle": 0, + "content": "4.2. Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.743, + 0.483, + 0.816 + ], + "angle": 0, + "content": "The reasoning chain process and the decision-making choices are crucial for training large AD models and enhancing the model's human-like reasoning ability. Therefore, we performed secondary annotation on the dataset we collected." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.817, + 0.484, + 0.892 + ], + "angle": 0, + "content": "Driving Scene Descriptions. We used the pre-trained large visual language model Qwen2-VL [3] to generate detailed descriptions for the corresponding driving scenes. These descriptions primarily focus on environmental information such as road conditions, weather, and lighting, as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.895, + 0.414 + ], + "angle": 0, + "content": "well as dynamic targets like vehicles and pedestrians that may pose driving risks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.897, + 0.562 + ], + "angle": 0, + "content": "Reasoning Decisions and Process. We considered pre-annotated information such as vehicle speed, acceleration, steering angle, traffic light status, and the state of the vehicle ahead. A rule-based method was used to determine the true values of future motion behaviors based on the decision choices made at earlier time steps. Additionally, we completed the predefined CoT reasoning text statements. Finally, to ensure the accuracy and consistency of the annotations, especially for decision choices, carefully manual verification was carried out." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.637, + 0.594 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.698, + 0.617 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.896, + 0.756 + ], + "angle": 0, + "content": "Our method was validated on the open source AD simulation platform CARLA 0.9.15 [14]. The VLR-Driver model was trained on a server equipped with 8 NVIDIA A800 GPUs (each with 80G of video memory) for approximately 50 hours. The dataset used was the VLR-Driver dataset that we developed. Specifically, we use ViT-g/14 from EVA-CLIP [29] as the vision encoder and LLaVA-NeXT-Video7B [50] as the VLM. The resolution of the input image is set to \\(336 \\times 336\\) pixels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.766, + 0.608, + 0.78 + ], + "angle": 0, + "content": "5.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.896, + 0.891 + ], + "angle": 0, + "content": "We employ four core metrics to evaluate AD performance: driving score (DS), route completion (RC), infraction score (IS), and success rate (SR). Additionally, to provide a more granular assessment of model performance in specific aspects, we introduce three supplementary evaluation metrics: subdivision infraction score, advanced driving ability, and driving experience score [22]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.94 + ], + "angle": 0, + "content": "26972" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.1, + 0.482, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.447, + 0.484, + 0.528 + ], + "angle": 0, + "content": "Figure 4. Visual comparison between VLR-Driver and VLM-based methods. The ST-CoT guides the VLR model to approach driving decisions in a human-like spatiotemporal manner. Based on the sequence of images from the preceding time period \\( T \\), we derive the following inference results. The images captured at 5 and 10 seconds afterward validate the accuracy of our decisions." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.554, + 0.411, + 0.57 + ], + "angle": 0, + "content": "5.3. Comparisons with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.578, + 0.484, + 0.697 + ], + "angle": 0, + "content": "We conducted comprehensive experiments with the SOTA methods including E2E and VLM in the CARLA simulator across 44 routes from the Bench2Drive benchmark. We present comparison result in Tab. 1. It can be seen that our method outperforms other methods in key metrics such as DS, RC, and SR, achieving first place and effectively improving DS by \\(17.5\\%\\), mean advanced driving ability by \\(31.4\\%\\), and SR by \\(22.2\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.699, + 0.484, + 0.891 + ], + "angle": 0, + "content": "The comparison results of the advanced driving ability and driving experience score of each method are shown in Tab. 2. Our method achieved the best results in all abilities, thanks to the deep reflection and reasoning ability of our VLR model, which has stronger traffic reasoning capacity in special road conditions. Most E2E methods can only achieve following the vehicle, but when there is a vehicle temporarily parked in the lane ahead, blocking the ego driving route, they will keep stopping and waiting, making it impossible to complete the entire route. In contrast, our VLR-Driver can achieve deep understanding and inference of the current scene through LLM inference, so as to make timely detours." + }, + { + "type": "table_caption", + "bbox": [ + 0.585, + 0.084, + 0.824, + 0.097 + ], + "angle": 0, + "content": "Table 3. Ablation study for each module." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.108, + 0.895, + 0.202 + ], + "angle": 0, + "content": "
IDAbal.Exp.Core Metrics ↑Advanced Driving Ability ↑
DSSROTMEREB
1VLRFull Model75.0150.0037.5046.6672.72
2Arch.w/o CoT57.1734.0925.0040.0041.66
3w/o VLR-Model46.3920.4514.2826.6636.36
4Trainw/o VLR-Data52.8527.2725.0026.669.09
5w/o Step-GRPO65.5745.4537.5040.0063.63
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.232, + 0.663, + 0.248 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.895, + 0.402 + ], + "angle": 0, + "content": "We conducted a comprehensive ablation study, detailed in Tab. 3. The experimental configurations include four variants: (1) Without utilizing our proposed spatiotemporal CoT strategy, using only a question-based approach without reasoning guidance. (2) Without using our VLR model, instead employing a standard LLM module. (3) Removing the VLR data used to guide the reasoning process. (4) Without Step-GRPO reinforcement learning training, using only LoRA strategies to fine-tune the model with supervision. The results show the effectiveness of each contribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.419, + 0.647, + 0.433 + ], + "angle": 0, + "content": "5.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.895, + 0.532 + ], + "angle": 0, + "content": "We selected some special scenarios to visualize the performance in complex traffic situations with reflective reasoning, and the comparison results shown in Fig. 4. The ST-CoT enables the model to make driving decisions with human-like reasoning, considering both spatial and temporal dynamics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.552, + 0.629, + 0.567 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.895, + 0.784 + ], + "angle": 0, + "content": "In this paper, we introduce VLR-Driver, a VLR model for embodied AD. It leverages a carefully designed ST-CoT strategy to guide the model in recursively analyzing potential safety risks and the driving intentions of dynamic agents in complex traffic scenarios. Our dual-phase training method significantly enhances the generalization of the model. Additionally, we propose the VLR-Driver dataset, which effectively integrates spatiotemporal perception, language understanding, and reflective reasoning, providing crucial support for interpretability reasoning in AD system. Experimental results show that VLR-Driver outperforms other methods on Bench2Drive benchmark, achieving cutting-edge performance and paving the way for embodied AI realization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.895, + 0.89 + ], + "angle": 0, + "content": "Limitations and future work. There are differences between the data in simulation platforms and the real world. Deploying LLMs for AD on real-world vehicles with limited onboard computing resources remains a key challenge for future exploration." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.94 + ], + "angle": 0, + "content": "26973" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.104, + 0.102, + 0.196, + 0.117 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.126, + 0.484, + 0.18 + ], + "angle": 0, + "content": "[1] Pedram Agand, Mohammad Mahdavianian, Manolis Savva, and Mo Chen. Letfuser: Light-weight end-to-end transformer-based sensor fusion for autonomous driving with multi-task learning. arXiv preprint arXiv:2310.13135, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.181, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[2] Hidehisa Arai, Keita Miwa, Kento Sasaki, Kohei Watanabe, Yu Yamaguchi, Shunsuke Aoki, and Issei Yamamoto. Covla: Comprehensive vision-language-action dataset for autonomous driving. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1933-1943. IEEE, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.262, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.317, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.385, + 0.484, + 0.438 + ], + "angle": 0, + "content": "[5] Daoming Chen, Ning Wang, Feng Chen, and Tony Pipe. Detrive: Imitation learning with transformer detection for end-to-end autonomous driving. arXiv preprint arXiv:2310.14224, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.44, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[6] Jianyu Chen, Zhuo Xu, and Masayoshi Tomizuka. End-to-end autonomous driving perception with sequential latent representation learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1999-2006. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.508, + 0.484, + 0.562 + ], + "angle": 0, + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.563, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.631, + 0.484, + 0.711 + ], + "angle": 0, + "content": "[9] Weihuang Chen, Fanjie Kong, Liming Chen, Shen'ao Wang, Zhiping Wang, and Hongbin Sun. Eatnet: Efficient axial transformer network for end-to-end autonomous driving. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 762-767. IEEE, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.713, + 0.484, + 0.767 + ], + "angle": 0, + "content": "[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.768, + 0.484, + 0.835 + ], + "angle": 0, + "content": "[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE transactions on pattern analysis and machine intelligence, 45(11):12878-12895, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.836, + 0.484, + 0.891 + ], + "angle": 0, + "content": "[12] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.126, + 0.484, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.104, + 0.895, + 0.131 + ], + "angle": 0, + "content": "IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.133, + 0.895, + 0.2 + ], + "angle": 0, + "content": "[13] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. Quar-vla: Vision-language-action model for quadruped robots. In European Conference on Computer Vision, pages 352-367. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.201, + 0.895, + 0.242 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, pages 1-16. PMLR, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.243, + 0.895, + 0.31 + ], + "angle": 0, + "content": "[15] Hao Gao, Shaoyu Chen, Bo Jiang, Bencheng Liao, Yiang Shi, Xiaoyang Guo, Yuechuan Pu, Haoran Yin, Xiangyu Li, Xinbang Zhang, et al. Rad: Training an end-to-end driving policy via large-scale 3dgs-based reinforcement learning. arXiv preprint arXiv:2502.13144, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.895, + 0.366 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.367, + 0.895, + 0.435 + ], + "angle": 0, + "content": "[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.436, + 0.895, + 0.504 + ], + "angle": 0, + "content": "[18] Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3903-3913, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.895, + 0.56 + ], + "angle": 0, + "content": "[19] JJ Hwang, R Xu, H Lin, WC Hung, J Ji, K Choi, D Huang, T He, P Covington, B Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arxiv 2024. arXiv preprint arXiv:2410.23262. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.895, + 0.642 + ], + "angle": 0, + "content": "[20] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.895, + 0.723 + ], + "angle": 0, + "content": "[21] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.724, + 0.895, + 0.78 + ], + "angle": 0, + "content": "[22] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.78, + 0.895, + 0.834 + ], + "angle": 0, + "content": "[23] Xiaosong Jia, Junqi You, Zhiyuan Zhang, and Junchi Yan. Drivetransformer: Unified transformer for scalable end-to-end autonomous driving. arXiv preprint arXiv:2503.07656, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.835, + 0.895, + 0.891 + ], + "angle": 0, + "content": "[24] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.104, + 0.895, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.519, + 0.94 + ], + "angle": 0, + "content": "26974" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.104, + 0.483, + 0.13 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.132, + 0.484, + 0.199 + ], + "angle": 0, + "content": "[25] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.201, + 0.483, + 0.242 + ], + "angle": 0, + "content": "[26] Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, and Irwin King. A survey on vision-language-action models for embodied ai. arXiv preprint arXiv:2405.14093, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.243, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[27] Jianbiao Mei, Yukai Ma, Xuemeng Yang, Licheng Wen, Xinyu Cai, Xin Li, Daocheng Fu, Bo Zhang, Pinlong Cai, Min Dou, et al. Continuously learning, adapting, and improving: A dual-process approach to autonomous driving. arXiv preprint arXiv:2405.15324, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.311, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[28] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.354, + 0.483, + 0.421 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.422, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[30] H Sha, Y Mu, Y Jiang, L Chen, C Xu, P Luo, SE Li, M Tomizuka, W Zhan, and M Ding. Linguistics: Large language models as decision makers for autonomous driving. arxiv 2023. arXiv preprint arXiv:2310.03026. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.478, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[31] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.533, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[32] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.615, + 0.483, + 0.682 + ], + "angle": 0, + "content": "[33] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.684, + 0.483, + 0.751 + ], + "angle": 0, + "content": "[34] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.753, + 0.483, + 0.82 + ], + "angle": 0, + "content": "[35] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.822, + 0.483, + 0.889 + ], + "angle": 0, + "content": "[36] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models. arXiv preprint arXiv:2410.05051, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.104, + 0.484, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.171 + ], + "angle": 0, + "content": "[37] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.173, + 0.894, + 0.225 + ], + "angle": 0, + "content": "[38] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.228, + 0.894, + 0.307 + ], + "angle": 0, + "content": "[39] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.894, + 0.377 + ], + "angle": 0, + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.894, + 0.433 + ], + "angle": 0, + "content": "[41] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Philip S Yu. Multimodal large language models: A survey. In 2023 IEEE International Conference on Big Data (BigData), pages 2247-2256. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.894, + 0.501 + ], + "angle": 0, + "content": "[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.894, + 0.57 + ], + "angle": 0, + "content": "[43] Weishang Wu, Xiaoheng Deng, Ping Jiang, Shaohua Wan, and Yuanxiong Guo. Crossfuser: Multi-modal feature fusion for end-to-end autonomous driving under unseen weather conditions. IEEE Transactions on Intelligent Transportation Systems, 24(12):14378-14392, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.894, + 0.64 + ], + "angle": 0, + "content": "[44] Shuo Xing, Chengyuan Qian, Yuping Wang, Hongyuan Hua, Kexin Tian, Yang Zhou, and Zhengzhong Tu. Openemma: Open-source multimodal model for end-to-end autonomous driving. In Proceedings of the Winter Conference on Applications of Computer Vision, pages 1001-1009, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.641, + 0.894, + 0.707 + ], + "angle": 0, + "content": "[45] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.709, + 0.894, + 0.75 + ], + "angle": 0, + "content": "[46] Z Yang, X Jia, H Li, and J Yan. Llm4drive: A survey of large language models for autonomous driving. arxiv 2023. arXiv preprint arXiv:2311.01043. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.752, + 0.894, + 0.819 + ], + "angle": 0, + "content": "[47] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, et al. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.821, + 0.894, + 0.861 + ], + "angle": 0, + "content": "[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.862, + 0.894, + 0.89 + ], + "angle": 0, + "content": "[49] Jiakai Zhang. End-to-End Learning for Autonomous Driving. PhD thesis, New York University, USA, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.518, + 0.939 + ], + "angle": 0, + "content": "26975" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.104, + 0.487, + 0.156 + ], + "angle": 0, + "content": "[50] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.159, + 0.486, + 0.225 + ], + "angle": 0, + "content": "[51] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.227, + 0.484, + 0.28 + ], + "angle": 0, + "content": "[52] Rui Zhao, Qirui Yuan, Jinyu Li, Haofeng Hu, Yun Li, Chengyuan Zheng, and Fei Gao. Sce2drivex: A generalized mllm framework for scene-to-drive learning. arXiv preprint arXiv:2502.14917, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.283, + 0.483, + 0.335 + ], + "angle": 0, + "content": "[53]Zikai Zhao, Qiuxuan Wu, Jian Wang, Botao Zhang, Chaoliang Zhong, and Anton A Zhilenkov. Exploring embodied intelligence in soft robotics: a review. Biomimetics, 9(4): 248, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.338, + 0.483, + 0.391 + ], + "angle": 0, + "content": "[54] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.393, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[55] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied understanding of driving scenarios. In European Conference on Computer Vision, pages 129-148. Springer, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.103, + 0.104, + 0.487, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.929, + 0.519, + 0.94 + ], + "angle": 0, + "content": "26976" + } + ] +] \ No newline at end of file diff --git a/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_origin.pdf b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7c8fccb79c798ffd51544bcc720ed372b0bcadb6 --- /dev/null +++ b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/a365b909-1be9-4f17-b2ec-25f898d74a15_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8886c3a240054331d55affa49adf39f04ae0201f058f203c8322f134d9539ff8 +size 1936430 diff --git a/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/full.md b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5f5d51d5f2b08d054ae8aef865a493afd0fc8eb4 --- /dev/null +++ b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/full.md @@ -0,0 +1,291 @@ +# VLR-Driver: Large Vision-Language-Reasoning Models for Embodied Autonomous Driving + +Fanjie Kong $^{1}$ Yitong Li $^{1}$ Weihuang Chen $^{1*}$ Chen Min $^{2}$ Yizhe Li $^{3}$ Zhiqiang Gao $^{5,1}$ Haoyang Li $^{4}$ Zhongyu Guo $^{1}$ Hongbin Sun $^{1}$ + +$^{1}$ Xi'an Jiaotong University $^{2}$ Institute of Computing Technology, Chinese Academy of Sciences $^{3}$ University of Maryland, College Park $^{4}$ Nanchang University $^{5}$ Engineering University of PAP + +# Abstract + +The rise of embodied intelligence and multi-modal large language models has led to exciting advancements in the field of autonomous driving, establishing it as a prominent research focus in both academia and industry. However, when confronted with intricate and ambiguous traffic scenarios, the lack of logical reasoning and cognitive decision-making capabilities remains the primary challenge impeding the realization of embodied autonomous driving. Although Vision Language Models (VLMs) have enhanced the deep semantic understanding of autonomous driving systems, they exhibit notable limitations in decision explainability when handling rare and long-tail traffic scenarios. In this paper, we propose VLR-Driver, a novel multimodal Vision-Language-Reasoning (VLR) framework based on Chain of Thought (CoT) for embodied autonomous driving. The framework employs a spatiotemporal CoT reasoning approach to recursively analyze potential safety risks and driving intentions of other agents, thereby delivering an efficient and transparent decision-making process. Furthermore, we construct a multi-modal reasoning-decision dataset to support the advancement of hierarchical reasoning of VLMs in autonomous driving. Closed-loop experiments conducted in CARLA demonstrate that the VLR-Driver significantly outperforms state-of-the-art end-to-end methods. Notably, key metrics such as driving score improved by $17.5\%$ , while the success rate improved by $22.2\%$ , offering a more transparent, reliable, and secure solution for autonomous driving systems. + +![](images/4d7db6f7b1621f5d1f8206748972db8364e81136fe3ac3e7acad0be138f8c717.jpg) +(a) + +![](images/d2ccdb0133d42db9d53bfa0246765ac5346ab9173c4815e395e22c3071f52108.jpg) +(b) + +![](images/fb93c92b0dd8fcd295e73c1b7dab290144df030ebf8ab07dd24e57c0732ffd2b.jpg) +(c) +Figure 1. Comparison of different LLM-based AD systems. (a) VLM model focuses more on answering questions, but it is difficult to convert decisions into coherent control signals. (b) VLA model focuses more on predicting driving control signals, but lacks interpretability. (c) The VLR model generates decisions and controls signals with transparent reasoning processes through CoT, enhancing the driver trust in the system. + +# 1. Introduction + +In recent years, the rapid progress of End-to-End (E2E) architectures [6, 7, 18, 49], Large Language Models (LLMs) [12, 41, 48], and embodied intelligence [26, 53, 55] has established these technologies as key enablers of innovation in Autonomous Driving (AD). Especially, VLMs enriched with extensive pre-training knowledge exhibit strong spatial understanding and common-sense reasoning abilities. DriveVLM [35] leverages VLM to enhance spatial awareness and planning capabilities in complex driving scenarios. CoVLA [2] integrates visual perception, language understanding, and action planning, effectively describing traffic scenarios and generating executable control actions. + +However, the decision-making process of VLMs often functions as a "black box", making it challenging to trace and interpret their underlying logic. This makes it difficult for AD systems to be fully trusted by drivers when encountering complex and emergency situations, such as illegal roadside parking, navigating intersections without traffic signals, and managing complex mixed-traffic interactions between motorized and non-motorized vehicles, thereby limiting their reliability and safety in real-world applications [39, 46, 54]. Moreover, most VLMs are trained on internet data, lacking spatial understanding and specialized training in the field of AD, making it difficult for them to fully adapt to dynamic and complex driving scenarios. + +Meanwhile, CoT reasoning demonstrates strong inference, interpretability, and generalization capabilities by breaking down complex tasks into intermediate reasoning steps [40]. CoT enables systems to think step by step rather than relying on E2E black-box predictions, making it one of the key approaches toward achieving embodied AD. Sce2DriveX [52] enhances comprehensive perception and reasoning by introducing a multi-modal LLM framework with CoT, enabling a deeper understanding of spatiotemporal relationships and road topology. Similarly, DriveCoT [38] integrates CoT reasoning to improve decision-making interpretability and controllability in AD systems. However, existing CoT-based methods typically rely on open-ended language generation for reasoning, which lacks structural constraints. The generated reasoning steps may exhibit causal inversion or unnecessary vocabulary, making them uncontrollable in safety-critical tasks. Additionally, current methods primarily operate on static snapshots rather than continuous temporal sequences, limiting their ability to predict future events in dynamic traffic environments. + +To bridge these gaps, we introduce VLR-Driver, a hierarchical CoT-based visual-language-reasoning model designed for closed-loop embodied AD. Our approach integrates the spatiotemporal features from cross-modal data, including multi-frame multi-view images and ego-vehicle control signals, by employing a SpatioTemporal CoT (STCoT) strategy that produces human-like reflective reasoning processes and driving action decisions. Additionally, we adopt a dual-phase training strategy, combining Low-Rank Adaptation (LoRA) [16] with an improved Stepwise Group Relative Policy Optimization (Step-GRPO) [34], significantly enhancing memory capacity and deep reasoning abilities. Our proposed VLR-Driver not only inherits the global action optimization capability of VLA models, but also preserves the transparency of modular rule-based methods. When encountering long-tail events and rare traffic scenarios, it demonstrates exceptional reflective reasoning and step-by-step inference, thereby enhancing the trust of human drivers in AD systems. The differences between VLM, VLA, and our proposed VLR are illustrated in Fig. 1. + +To further enhance VLR models in environmental understanding, reasoning, and decision-making, we introduce the VLR-Driver Dataset. This data set includes detailed scene descriptions, weather information, vehicle state details, and most critically, human-like CoT reasoning processes and the corresponding driving decisions. Various experiments demonstrate that VLR-Driver is capable of making accurate driving decisions and coherent reasoning, even under challenging and highly dynamic road conditions. + +The primary contributions of this work are summarized as follows: + +- Distinctive VLR-Driver Framework. We introduce VLR-Driver, a visual-language-reasoning model developed for embodied AD. It generates a human-like reflective reasoning process within the driving system, enabling accurate driving decisions. +- Spatiotemporal CoT. We present a spatiotemporal CoT strategy that recursively analyzes potential safety risks and the driving intentions of moving agents in complex traffic scenarios, ensuring that the reasoning process of driving decisions remains transparent and interpretable. +- Advanced VLR-Driver Dataset. We construct VLR-Driver Dataset, a cutting-edge visual-language-reasoning and decision-making dataset specifically designed for AD. It supports the enhancement of spatiotemporal understanding and reflective reasoning capabilities in embodied AD systems. +- Superior Performance in Closed-Loop Simulations. Extensive closed-loop experiments are conducted on the CARLA platform. On the Bench2Drive benchmark, our approach achieves a $17.5\%$ improvement in driving score and a $22.2\%$ increase in success rate, providing a more human-like, reliable, and trustworthy solution. + +# 2. Related Work + +# 2.1. End-to-end Autonomous Driving + +The rapid advancement of E2E-AD has fostered a growing transition away from modular rule-based methods toward data-driven approaches [15, 23]. Based on different input modalities, E2E methods can be categorized into visual-only methods [5, 8, 10, 17] and vision-LiDAR fusion methods [1, 20, 21]. TCP [42] and NEAT [10] adopt imitation learning methods, training directly on collected state-action pair datasets, demonstrating the feasibility of E2E approaches for AD. Roach [51] utilizes reinforcement learning experts as coaches, delivering dense and informative supervision signals to agents equipped with monocular camera inputs. However, vision-based methods inherently struggle with distance and depth estimation. These limitations may compromise the reliability of driving decisions [24]. To address these limitations, vision-LiDAR fusion methods such as TransFuser [11], CrossFuser [43], + +![](images/eac4ff6d82c2ec490e6391963617430d0427d8e7bb9e1e00dd4386fe5bbcfc5d.jpg) + +# VLR-Driver Datasets + +![](images/6e6932366bf55cbdcfb9be596995fdfe69694031a9c0bc386017b65fd043bcbb.jpg) +Figure 2. Overview of VLR-Driver framework. We introduce VLR-Driver Dataset, an advanced visual-language-reasoning dataset designed for AD, featuring detailed annotations of scene descriptions, analytical reasoning, and behavioral decisions. We present VLR-Driver Embodied AI (EAI), a novel multi-modal visual-language-reasoning framework for embodied AD that leverages a hierarchical spatiotemporal CoT reasoning mechanism. + +![](images/3f4f222795ee64fa75f5b7a6345fda0714df69ace388916c0dbe5e07f61b2cb6.jpg) + +# VLR-Driver EAI + +![](images/b57cb0b6dbc90c15f2082fad4efe9575391df0061330598a2549371e8d4afeda.jpg) + +and FusionAD [47] have been designed to effectively integrate image with LiDAR data, significantly improving the robustness of AD systems in complex environments. However, different modalities contribute differently to the driving task, and spatiotemporal synchronization issues between sensor data can lead to inconsistencies, increasing both model complexity and training difficulty. + +# 2.2. VLM and VLA in Autonomous Driving + +VLMs unify visual perception with natural language processing capabilities, enabling a more comprehensive understanding of driving environments [19, 36, 37]. By incorporating cross-modal data fusion, integrating textual information alongside visual inputs, these systems gain text comprehension and human interaction capabilities that conventional E2E models inherently lack. DriveGPT4 [45] enhances LLMs' ability to process multimodal inputs by projecting them into the text domain, thereby enabling interpretable E2E-AD. Senna [25] and DriveVLM [35] integrate VLMs with traditional modular pipelines or E2E frameworks, achieving decoupled high-level planning and low-level trajectory prediction. This approach improves planning performance while preserving the model's common sense reasoning capabilities. + +Moreover, VLA is an emerging paradigm that unifies vi + +sual perception, natural language understanding, and action prediction within a cohesive framework [13, 26]. Originally introduced in the field of robotics, RT-2 [4] pioneered the representation of robotic actions as text tokens, seamlessly incorporating them alongside natural language labels into the model's training set. This enables direct transfer of internet-scale knowledge to robotic control, significantly enhancing both the generalization and semantic reasoning capabilities of robotic systems. For AD, CoVLA [2] introduces an interpretable VLA model that integrates visual perception, language-based scene understanding, and action planning. This integration improves the system's ability to comprehend complex driving scenarios, anticipate trajectory, and execute informed driving decisions. However, despite these advancements, VLA models still exhibit limitations in accurately predicting precise control action values, leading to a low lower-bound in AD performance. Furthermore, when confronted with highly dynamic and complex traffic scenarios, their deep understanding and reasoning capabilities remain insufficient. Unlike the above methods, our proposed VLR model places greater emphasis on step-by-step reasoning and the thought process of the model, providing drivers with increased confidence in using AD systems. + +# 2.3. Chain of Thought in Autonomous Driving + +The CoT technology is an extension of prompt engineering, proposed by Wei Jason in 2022 [40], which has greatly improved the effectiveness of reasoning for complex problems. The highly anticipated Deepseek-R1 [33] model also utilizes the CoT technique, which deeply integrates multimodal knowledge base data, enabling the model to generate a step-by-step thinking process. DriveCoT [38] has built a CoT dataset that includes sensor data, control decisions, and CoT labels used to indicate reasoning processes. The model trained on this basis can generate predictions and final decisions with CoT, effectively improving model performance. LanguageMPC [30] combines LLM with Model Predictive Control and decomposes driving decisions into multiple subtasks through a CoT framework. This method enables the auto drive system to think like human beings, and improves its ability to handle complex scenes. OpenEMMA [44] introduces CoT technology to guide model generation of detailed descriptions of key objects, behavioral insights, and meta driving decisions, improving system transparency and usability. Motivated by these advancements, we apply step-by-step hierarchical spatiotemporal CoT to AD, enhancing the interpretability of reasoning and decision-making. + +# 3. Method + +We present the motivation and design details of our VLR-Driver framework. As depicted in Fig. 2, VLR-Driver comprises two main components: the VLR-Driver EAI and the VLR-Driver Dataset. Initially, we introduce the design concept of the VLR model, which builds upon enhancements to the VLA model (Sec. 3.1). Subsequently, we elaborate on the hierarchical spatiotemporal CoT methodology (Sec. 3.2) and the specifics of the dual-phase training strategy (Sec. 3.3). + +# 3.1. Overview + +The VLR model is a large visual-language-reasoning model designed for embodied AD. It can processes visual inputs, such as multi-view images, alongside textual information, including vehicle control signals. The model is capable of extracting spatiotemporal key features within a multimodal embedding and recursively analyze potential safety risks and the driving intentions of other agents within the reasoning level of the VLR model. Ultimately, it formulates a comprehensive reasoning framework and well-structured decision-making outputs, explicitly identifying critical risk factors and the underlying rationale behind each decision. This process significantly reinforces the robustness and safety of the AD system, ensuring adaptive resilience in complex and dynamic environments. + +In this study, we employ pre-trained LLaVA-NeXT + +Video [50] as the VLM and CLIP [29] as the visual encoder. The model is capable of processing multi-frame multi-view image data that capture historical temporal context, while also extracting real-time vehicle sensor information to facilitate dynamic and context-aware decision-making. + +Input Representations. We utilize $N_{f}$ frame and $N_{v}$ view images from the past period, with a field of view (FoV) of 70 degrees. It can be represented as $V \in \mathbb{R}^{N_f \times N_v \times 3 \times H_0 \times W_0}$ . At the same time, there are also the current position $(x,y)$ of ego vehicle, the speed $v$ , the target point position $(p,q)$ . Subsequently, the compressed and cropped image data and the information from the ego's sensors are input into the model. + +Output Representations. Our output consists of the reasoning process and driving decisions generated by the VLR model for the current driving scenario. This includes risk identification, traffic signal recognition, motion direction prediction, and AD decision-making. The driving decision will also be fused with the spatiotemporal information features extracted by the E2E model, and finally output the waypoints and control signals of the vehicle for the next moment. + +# 3.2. Spatiotemporal CoT Reasoning + +To enhance the reasoning capabilities and transparency of the AD system, we introduce a hierarchical ST-CoT that guides the model to approach driving decisions in a human-like manner. Our method decomposes the driving decision-making process into two levels: a perception-level spatiotemporal CoT $C_{\text{perception}}$ , which focuses on extracting and understanding environmental dynamics, and a decision-level dynamic CoT $C_{\text{decision}}$ , which refines and optimizes decision-making based on contextual and temporal factors. The example of CoT is shown in Fig. 3. + +# 3.2.1. Perception Level CoT + +The spatiotemporal characteristics of the environment play a crucial role in the AD process. The perception level CoT is responsible for guiding the model to extract spatial and temporal features from input image data, identify and locate crucial objects in the traffic, such as vehicles, pedestrians, alien objects, traffic lights, traffic signs, etc., and extract historical behavioral features of dynamic agent based on temporal information. Our method enables VLR model to describe the current driving scenario, construct real-time spatial layout and dynamic changes of the environment, and achieve long-term planning for driving decisions. + +Spatial CoT. In driving scenarios, we primarily focus on obstacles that impact the ego's normal operation, including object categories $O_{type}$ and distances with the ego $O_{dis}$ . A critical aspect of safe driving is identifying potential risk points within the current lane. Additionally, when the ve + +![](images/7f4cad0d991bd2d337f5ca0bbf6a670644dcf6e12afd98eace1dbff0f328f6a0.jpg) +Figure 3. Illustration of the ST-CoT reasoning process. In this scenario, where some vehicles are illegally parked ahead and blocking the lane, our method can conduct hierarchical paitotemporal reasoning analysis and make a decision of change left lane once the adjacent lane is free. + +hicle executes lane changes, objects in adjacent lanes, both left and right, may significantly influence its movement. Beyond obstacles, key traffic light $S_{light}$ , traffic signs $S_{sign}$ and lane markings $S_{mark}$ are also integral to decision-making, ensuring comprehensive spatial awareness. + +Temporal CoT. While a single-frame image can provide a static representation of road scenes and traffic participants, it fails to capture the motion trends of moving agents. To address this limitation, we introduce consecutive frames $I = \{I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br}\}_{t = T_{now} - T}^{T_{now}}$ into the model, allowing it to track temporal variations in object positioning, which $I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br}$ represents the image view of front, front left, front right, back, back left and back right. These sequential frames not only offer instantaneous spatial context but also reveal motion trajectories and behavioral patterns through their inter-frame positional changes. This temporal information is essential for predicting dynamic object movement, assessing collision risks, and generating robust path planning strategies, ultimately enhancing the ability of anticipate and react to evolving traffic conditions. + +# 3.2.2. Decision Level CoT + +The output information from the perception-level serves as a critical foundation for the driving decision level, enabling reliable AD behavior inference. Within the driving decision level, we account for complex dynamic environmental factors, transforming spatial and temporal information from the perception level into concrete driving decisions. Specifically, decision-makers must not only analyze the current driving environment in real time but also anticipate future behaviors of other traffic participants and make decisions + +based on multiple factors. Throughout this process, the CoT in the driving decision level spans multiple perspectives, incorporating safety, efficiency, comfort, and compliance with traffic regulations to ensure that driving decisions meet safety standards while optimizing driving efficiency. + +To further enhance the structured reasoning process in AD, we have carefully designed hierarchical reasoning prompts that guide decision-making. Our structured prompts follow a logical sequence of "risk point recognition — driving intention prediction — driving decision selection", forming a cohesive reasoning chain aligned with human cognitive driving patterns. + +Risk Point Recognition. In this initial stage, the prompt-driven model conducts a comprehensive perception and analysis of the driving environment. This includes recognizing and evaluating critical elements such as traffic signs, lane markings, pedestrians, and obstacles to identify potential risks. + +Driving Intention Prediction. Once risk points are identified, the model leverages dynamic target behavior prediction and scene understanding to infer the potential movements and intentions of other road users. For instance, the model assesses whether pedestrians are likely to cross the road or whether the vehicle ahead intends to change lanes. + +Driving Decision Selection. Based on the contextual information gathered from the first two stages, the model applies multi-modal information fusion and weighted decision-making to select the optimal driving maneuver. + +Through this structured prompting strategy, the large model adheres to a progressive reasoning hierarchy, beginning with fundamental environmental perception and advancing to higher-level decision-making. By explicitly presenting the reasoning process in a clear and structured manner, this approach enhances passenger trust and confidence in the intelligent driving system. Moreover, the integration of recursive CoT reasoning enables the model to mimic the step-by-step thought process of human drivers, facilitating more flexible, reliable, and interpretable decision outputs in complex driving scenarios. The structured prompt framework are shown in Fig. 3. + +# 3.3. Training Paradigm + +VLR-Driver adopts a dual-phase training strategy to optimize its reasoning and decision-making capabilities. In the first phase, LoRA [16] is utilized for supervised fine-tuning on a pre-trained large model, enabling efficient adaptation with minimal memory and computational overhead while maintaining strong performance. In the second phase, StepGRPO is applied for reinforcement learning based on human preferences, further enhancing the model's ability to + +exhibit human-like reasoning and decision-making characteristics. + +Training with LoRA. LoRA is a parameter-efficient finetuning technique that enables effective model adaptation by performing a low-rank decomposition of the weight matrix. The core principle behind LoRA is to decompose the weight matrix of a pre-trained model into a low-rank structure, significantly reducing the number of trainable parameters while preserving expressivity. Specifically, if the weight matrix in the pre-trained model is denoted as $W_0 \in \mathbb{R}^{d \times k}$ , LoRA represents it as: + +$$ +\boldsymbol {W} ^ {\prime} = \boldsymbol {W} _ {0} + \Delta \boldsymbol {W} = \boldsymbol {W} _ {0} + \frac {\alpha}{r} \boldsymbol {B} \cdot \boldsymbol {A}, \tag {1} +$$ + +where $\mathbf{A} \in \mathbb{R}^{r \times k}$ and $\mathbf{B} \in \mathbb{R}^{d \times r}$ , $r$ is the rank of a low rank matrix, usually much smaller than $d$ and $k$ , $\alpha$ is a scaling factor. The forward pass is computed as: + +$$ +y = \boldsymbol {W} ^ {\prime} x = \left(\boldsymbol {W _ {0}} + \frac {\alpha}{r} \boldsymbol {B} \cdot \boldsymbol {A}\right) x, \tag {2} +$$ + +where $y$ is the output and $x$ is input. + +We use LoRA for all linear modules, which not only saves computation but also ensures the performance of the model. + +Training with GRPO. The core principle of GRPO [34] is to optimize strategies by assigning relative rewards to multiple outputs generated from the same prompt, thereby eliminating the need for additional value function models. The introduction of process reward model estimation in GRPO provides finer support for distributed rewards. The reward of each step of the outputs is: + +$$ +R = \left\{\left\{r _ {1} ^ {I (1)}, \dots , r _ {1} ^ {I (K _ {1})} \right\}, \dots , \left\{r _ {G} ^ {I (1)}, \dots , r _ {G} ^ {I (K _ {G})} \right\} \right\}, \tag {3} +$$ + +where $I(\cdot)$ is the end token index, and the reward need be normalized as: + +$$ +\widetilde {r} _ {i} ^ {I (j)} = \frac {r _ {i} ^ {I (j)} - \operatorname {m e a n} (\boldsymbol {R})}{\operatorname {s t d} (\boldsymbol {R})}. \tag {4} +$$ + +We extend GRPO by introducing reasoning Step-GRPO, a supervised reasoning decision process that structures the output into multiple steps based on the CoT reasoning framework. At each step, a reward function is applied to evaluate and assign scores, enabling fine-grained feedback that enhances model interpretability and accelerates convergence. Specifically, we first generate multiple candidate decision answers for the current driving scenario using prompts within the VLR model; Then, following our ST-CoT strategy, the reasoning process is divided into four distinct steps: scene description, spatial risk point reasoning, dynamic trajectory prediction, and driving decision-making; Furthermore, each reasoning step is assigned a reward to encourage structured learning. Finally, we compare + +all answers within the group and calculate the Kullback-Leibler (KL) divergence to update the policy model. This grouping and step-by-step scoring strategy enhances training efficiency and reduces the likelihood of erroneous reasoning in the model. + +# 4. VLR-Driver Dataset + +To fully explore the reasoning and decision-making capabilities of LLMs, we propose an advanced reasoning and decision-making dataset for AD scenarios, called the VLR-Driver Datasets. This dataset relies on the CARLA [14] simulator for data collection and is expanded and meticulously annotated based on the Bench2Drive dataset [22]. It includes: a) multi-view multi-frame images or videos, b) valuable information for AD, such as road details, weather conditions, vehicle information, and scene descriptions, and c) driving decision choices along with the decision-making process based on an ST-CoT. This dataset provides a rich and comprehensive training foundation for AD reasoning and decision-making, allowing the agent to exhibit human-like reasoning while interacting with the environment. The dataset includes 20,000 sets of multi-frame, multi-angle image data collected from various road conditions such as urban, rural, and highways in the CARLA simulator, covering over 30 specific complex traffic scenarios (e.g., forward accidents, dynamic object crossings). Each image set provides road scene descriptions, environmental weather information, vehicle status data, and, most importantly, the humanlike reasoning process and driving behavior decisions. + +# 4.1. Data Collection + +We conducted data collection based on the 44 corner scene classifications provided by Bench2Drive to ensure optimal AD performance in various complex corner scenarios. The Bench2Drive dataset offers a rich array of data and annotations, including multi-angle images, lidar, radar, vehicle information, and expert assessments, which have played a crucial role in constructing our VLR dataset. Additionally, we selected over 30 scenarios where the reasoning and decision-making capabilities of AD systems are relatively limited, and used them as the visual and textual components of our dataset. We further expanded and enriched the dataset by collecting additional data in the CARLA simulator. We have taken into account various weather conditions, road conditions, and the types and numbers of traffic participants in the scene. + +Weather. For each scene, we randomly set values for cloudiness, fog density, precipitation deposits, sun altitude angle, sun azimuth angle, wetness, and wind intensity. The combinations of these parameters cover a variety of weather conditions, such as rainy, foggy, broken sky, and stormy, as well as different lighting conditions for day and night. + +Table 1. The comparison of core metrics and subdivision infraction scores with state-of-the-art E2E/VLM models on the Bench2Drive benchmark. C, L and T indicate camera, LiDAR and text modalities, respectively. DS, RC, IS, SR correspond to the Driving Score, Route Completion, Infraction Score, and Success Rate. CP, CV, CL, RL, SS, OR, AB, YEV correspond to the Collision with a Pedestrian, Collision with another Vehicle, Collision with Layout, Red Light infractions, Stop Sign infractions, Off-Road infractions, Agent Blocked, and failure to Yield to Emergency Vehicles infractions. + +
MethodTypeModalityCore Metrics ↑Subdivision Infraction Score ↓
DSRCISSRCPCVCLRLSSORABYEV
NEAT [10]E2EC30.8655.350.556.811.089.875.570.201.330.412.010.27
TCP [42]C56.2883.570.6525.000.265.465.460.000.520.220.780.00
LeTFuser [1]C+L52.5377.680.6718.181.165.543.790.290.870.120.580.29
LateFusion [28]C+L48.5358.320.8518.180.383.111.550.380.770.061.550.38
TransFuser [11]C+L37.1868.140.519.090.9613.248.710.000.960.322.580.32
ThinkTwice [21]C+L58.7974.350.7729.540.305.760.910.000.910.050.910.30
EATNet [9]C+L42.9778.840.5415.910.8214.011.920.001.640.221.090.27
InterFuser [31]C+L63.8180.460.7940.900.353.810.540.271.080.050.540.27
LMDrive [32]VLMC+L+T24.7633.020.9013.631.142.862.290.000.570.053.440.57
LeapAD [27]C+T55.1877.450.7136.360.695.071.150.200.910.081.470.27
VLR-Driver (Ours)VLRC+T75.0186.090.8750.000.722.830.480.000.480.030.250.25
+ +Table 2. The comparison of advanced driving ability and experience score with state-of-the-art models on the Bench2Drive benchmark. OT, MER, EB, GW, TS, DE, SC correspond to the OverTaking, MERging, Emergency Brake, Give Way, Traffic Sign, Driving Efficiency, and Smoothness Control. + +
MethodAdvanced Driving Ability ↑Exper. Score ↑
OTMEREBGWTSMeanDESC
NEAT [10]0.006.669.0950.0027.7718.7092.090.30
TCP [42]25.0013.3327.2750.0050.0033.12114.520.27
LeTFuser [1]0.0020.0018.180.0047.2217.08115.090.47
LateFusion [28]0.0020.009.090.0036.1113.04104.390.59
TransFuser [11]0.0013.339.090.0036.1111.7195.200.36
ThinkTwice [21]12.5020.0036.3650.0052.7734.3391.170.36
EATNet [9]0.0013.3318.180.0041.6614.6388.130.33
InterFuser [31]0.0046.6654.5450.0061.1142.46119.200.32
LMDrive [32]25.006.669.0950.002.7718.7075.410.22
LeapAD [27]12.5033.3327.2750.0044.4433.5193.330.26
VLR-Driver (Ours)37.5046.6672.7250.0072.2255.82125.220.59
+ +Roads. The scenes include various road types such as urban two-way single-lane roads, multi-lane roads, highways, and narrow rural roads. + +Traffic Participants. Different corner cases involve various traffic participants, including cars, bicycles, pedestrians, ambulances, etc., simulating the complex traffic conditions encountered in daily driving. + +# 4.2. Data Annotation + +The reasoning chain process and the decision-making choices are crucial for training large AD models and enhancing the model's human-like reasoning ability. Therefore, we performed secondary annotation on the dataset we collected. + +Driving Scene Descriptions. We used the pre-trained large visual language model Qwen2-VL [3] to generate detailed descriptions for the corresponding driving scenes. These descriptions primarily focus on environmental information such as road conditions, weather, and lighting, as + +well as dynamic targets like vehicles and pedestrians that may pose driving risks. + +Reasoning Decisions and Process. We considered pre-annotated information such as vehicle speed, acceleration, steering angle, traffic light status, and the state of the vehicle ahead. A rule-based method was used to determine the true values of future motion behaviors based on the decision choices made at earlier time steps. Additionally, we completed the predefined CoT reasoning text statements. Finally, to ensure the accuracy and consistency of the annotations, especially for decision choices, carefully manual verification was carried out. + +# 5. Experiment + +# 5.1. Experimental Setup + +Our method was validated on the open source AD simulation platform CARLA 0.9.15 [14]. The VLR-Driver model was trained on a server equipped with 8 NVIDIA A800 GPUs (each with 80G of video memory) for approximately 50 hours. The dataset used was the VLR-Driver dataset that we developed. Specifically, we use ViT-g/14 from EVA-CLIP [29] as the vision encoder and LLaVA-NeXT-Video7B [50] as the VLM. The resolution of the input image is set to $336 \times 336$ pixels. + +# 5.2. Metrics + +We employ four core metrics to evaluate AD performance: driving score (DS), route completion (RC), infraction score (IS), and success rate (SR). Additionally, to provide a more granular assessment of model performance in specific aspects, we introduce three supplementary evaluation metrics: subdivision infraction score, advanced driving ability, and driving experience score [22]. + +![](images/a16dfdb5fe44cee491dc6f5301800d3dda384dec0c2acf22c560d592e5f0c136.jpg) +Figure 4. Visual comparison between VLR-Driver and VLM-based methods. The ST-CoT guides the VLR model to approach driving decisions in a human-like spatiotemporal manner. Based on the sequence of images from the preceding time period $T$ , we derive the following inference results. The images captured at 5 and 10 seconds afterward validate the accuracy of our decisions. + +# 5.3. Comparisons with Existing Methods + +We conducted comprehensive experiments with the SOTA methods including E2E and VLM in the CARLA simulator across 44 routes from the Bench2Drive benchmark. We present comparison result in Tab. 1. It can be seen that our method outperforms other methods in key metrics such as DS, RC, and SR, achieving first place and effectively improving DS by $17.5\%$ , mean advanced driving ability by $31.4\%$ , and SR by $22.2\%$ . + +The comparison results of the advanced driving ability and driving experience score of each method are shown in Tab. 2. Our method achieved the best results in all abilities, thanks to the deep reflection and reasoning ability of our VLR model, which has stronger traffic reasoning capacity in special road conditions. Most E2E methods can only achieve following the vehicle, but when there is a vehicle temporarily parked in the lane ahead, blocking the ego driving route, they will keep stopping and waiting, making it impossible to complete the entire route. In contrast, our VLR-Driver can achieve deep understanding and inference of the current scene through LLM inference, so as to make timely detours. + +Table 3. Ablation study for each module. + +
IDAbal.Exp.Core Metrics ↑Advanced Driving Ability ↑
DSSROTMEREB
1VLRFull Model75.0150.0037.5046.6672.72
2Arch.w/o CoT57.1734.0925.0040.0041.66
3w/o VLR-Model46.3920.4514.2826.6636.36
4Trainw/o VLR-Data52.8527.2725.0026.669.09
5w/o Step-GRPO65.5745.4537.5040.0063.63
+ +# 5.4. Ablation Study + +We conducted a comprehensive ablation study, detailed in Tab. 3. The experimental configurations include four variants: (1) Without utilizing our proposed spatiotemporal CoT strategy, using only a question-based approach without reasoning guidance. (2) Without using our VLR model, instead employing a standard LLM module. (3) Removing the VLR data used to guide the reasoning process. (4) Without Step-GRPO reinforcement learning training, using only LoRA strategies to fine-tune the model with supervision. The results show the effectiveness of each contribution. + +# 5.5. Visualization + +We selected some special scenarios to visualize the performance in complex traffic situations with reflective reasoning, and the comparison results shown in Fig. 4. The ST-CoT enables the model to make driving decisions with human-like reasoning, considering both spatial and temporal dynamics. + +# 6. Conclusion + +In this paper, we introduce VLR-Driver, a VLR model for embodied AD. It leverages a carefully designed ST-CoT strategy to guide the model in recursively analyzing potential safety risks and the driving intentions of dynamic agents in complex traffic scenarios. Our dual-phase training method significantly enhances the generalization of the model. Additionally, we propose the VLR-Driver dataset, which effectively integrates spatiotemporal perception, language understanding, and reflective reasoning, providing crucial support for interpretability reasoning in AD system. Experimental results show that VLR-Driver outperforms other methods on Bench2Drive benchmark, achieving cutting-edge performance and paving the way for embodied AI realization. + +Limitations and future work. There are differences between the data in simulation platforms and the real world. Deploying LLMs for AD on real-world vehicles with limited onboard computing resources remains a key challenge for future exploration. + +# References + +[1] Pedram Agand, Mohammad Mahdavianian, Manolis Savva, and Mo Chen. Letfuser: Light-weight end-to-end transformer-based sensor fusion for autonomous driving with multi-task learning. arXiv preprint arXiv:2310.13135, 2023. 2, 7 +[2] Hidehisa Arai, Keita Miwa, Kento Sasaki, Kohei Watanabe, Yu Yamaguchi, Shunsuke Aoki, and Issei Yamamoto. Covla: Comprehensive vision-language-action dataset for autonomous driving. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1933-1943. IEEE, 2025. 1, 3 +[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 7 +[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3 +[5] Daoming Chen, Ning Wang, Feng Chen, and Tony Pipe. Detrive: Imitation learning with transformer detection for end-to-end autonomous driving. arXiv preprint arXiv:2310.14224, 2023. 2 +[6] Jianyu Chen, Zhuo Xu, and Masayoshi Tomizuka. End-to-end autonomous driving perception with sequential latent representation learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1999-2006. IEEE, 2020. 1 +[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2 +[9] Weihuang Chen, Fanjie Kong, Liming Chen, Shen'ao Wang, Zhiping Wang, and Hongbin Sun. Eatnet: Efficient axial transformer network for end-to-end autonomous driving. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 762-767. IEEE, 2024. 7 +[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 2, 7 +[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE transactions on pattern analysis and machine intelligence, 45(11):12878-12895, 2022. 2, 7 +[12] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the + +IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1 +[13] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. Quar-vla: Vision-language-action model for quadruped robots. In European Conference on Computer Vision, pages 352-367. Springer, 2024. 3 +[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, pages 1-16. PMLR, 2017. 6, 7 +[15] Hao Gao, Shaoyu Chen, Bo Jiang, Bencheng Liao, Yiang Shi, Xiaoyang Guo, Yuechuan Pu, Haoran Yin, Xiangyu Li, Xinbang Zhang, et al. Rad: Training an end-to-end driving policy via large-scale 3dgs-based reinforcement learning. arXiv preprint arXiv:2502.13144, 2025. 2 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2, 5 +[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2 +[18] Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3903-3913, 2023. 1 +[19] JJ Hwang, R Xu, H Lin, WC Hung, J Ji, K Choi, D Huang, T He, P Covington, B Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arxiv 2024. arXiv preprint arXiv:2410.23262. 3 +[20] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 2 +[21] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2, 7 +[22] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 6, 7 +[23] Xiaosong Jia, Junqi You, Zhiyuan Zhang, and Junchi Yan. Drivetransformer: Unified transformer for scalable end-to-end autonomous driving. arXiv preprint arXiv:2503.07656, 2025. 2 +[24] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 2 +[25] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 3 +[26] Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, and Irwin King. A survey on vision-language-action models for embodied ai. arXiv preprint arXiv:2405.14093, 2024. 1, 3 +[27] Jianbiao Mei, Yukai Ma, Xuemeng Yang, Licheng Wen, Xinyu Cai, Xin Li, Daocheng Fu, Bo Zhang, Pinlong Cai, Min Dou, et al. Continuously learning, adapting, and improving: A dual-process approach to autonomous driving. arXiv preprint arXiv:2405.15324, 2024. 7 +[28] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. 2021. 7 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 4, 7 +[30] H Sha, Y Mu, Y Jiang, L Chen, C Xu, P Luo, SE Li, M Tomizuka, W Zhan, and M Ding. Linguistics: Large language models as decision makers for autonomous driving. arxiv 2023. arXiv preprint arXiv:2310.03026. 4 +[31] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 7 +[32] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 7 +[33] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 4 +[34] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 2, 6 +[35] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 3 +[36] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models. arXiv preprint arXiv:2410.05051, 2024. 3 + +[37] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3 +[38] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 2, 4 +[39] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 2 +[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4 +[41] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Philip S Yu. Multimodal large language models: A survey. In 2023 IEEE International Conference on Big Data (BigData), pages 2247-2256. IEEE, 2023. 1 +[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2, 7 +[43] Weishang Wu, Xiaoheng Deng, Ping Jiang, Shaohua Wan, and Yuanxiong Guo. Crossfuser: Multi-modal feature fusion for end-to-end autonomous driving under unseen weather conditions. IEEE Transactions on Intelligent Transportation Systems, 24(12):14378-14392, 2023. 2 +[44] Shuo Xing, Chengyuan Qian, Yuping Wang, Hongyuan Hua, Kexin Tian, Yang Zhou, and Zhengzhong Tu. Openemma: Open-source multimodal model for end-to-end autonomous driving. In Proceedings of the Winter Conference on Applications of Computer Vision, pages 1001-1009, 2025. 4 +[45] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3 +[46] Z Yang, X Jia, H Li, and J Yan. Llm4drive: A survey of large language models for autonomous driving. arxiv 2023. arXiv preprint arXiv:2311.01043. 2 +[47] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, et al. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006, 2023. 3 +[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1 +[49] Jiakai Zhang. End-to-End Learning for Autonomous Driving. PhD thesis, New York University, USA, 2019. 1 + +[50] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 4, 7 +[51] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2 +[52] Rui Zhao, Qirui Yuan, Jinyu Li, Haofeng Hu, Yun Li, Chengyuan Zheng, and Fei Gao. Sce2drivex: A generalized mllm framework for scene-to-drive learning. arXiv preprint arXiv:2502.14917, 2025. 2 +[53]Zikai Zhao, Qiuxuan Wu, Jian Wang, Botao Zhang, Chaoliang Zhong, and Anton A Zhilenkov. Exploring embodied intelligence in soft robotics: a review. Biomimetics, 9(4): 248, 2024. 1 +[54] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 2 +[55] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied understanding of driving scenarios. In European Conference on Computer Vision, pages 129-148. Springer, 2024. 1 \ No newline at end of file diff --git a/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/images.zip b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..970b3f40ebd54cef62fe9b26de108473ca089d67 --- /dev/null +++ b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7299c2cdfa1884f7a290978b6fdf19d2e58f53f374c3fb6bdcd932450513b8b2 +size 504136 diff --git a/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/layout.json b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a564d78629761f3f1adc9f2943ea3f1ae0fcbf16 --- /dev/null +++ b/2025/VLR-Driver_ Large Vision-Language-Reasoning Models for Embodied Autonomous Driving/layout.json @@ -0,0 +1,7506 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 110, + 519, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 110, + 519, + 147 + ], + "spans": [ + { + "bbox": [ + 93, + 110, + 519, + 147 + ], + "type": "text", + "content": "VLR-Driver: Large Vision-Language-Reasoning Models for Embodied Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "spans": [ + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": "Fanjie Kong" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Yitong Li" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Weihuang Chen" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Chen Min" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Yizhe Li" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Zhiqiang Gao" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{5,1}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Haoyang Li" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Zhongyu Guo" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "text", + "content": " Hongbin Sun" + }, + { + "bbox": [ + 129, + 167, + 478, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "spans": [ + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "text", + "content": "Xi'an Jiaotong University " + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "text", + "content": "Institute of Computing Technology, Chinese Academy of Sciences " + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "text", + "content": "University of Maryland, College Park " + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "text", + "content": "Nanchang University " + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 75, + 195, + 533, + 224 + ], + "type": "text", + "content": "Engineering University of PAP" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 249, + 203, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 249, + 203, + 261 + ], + "spans": [ + { + "bbox": [ + 157, + 249, + 203, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 277, + 297, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 297, + 604 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 297, + 604 + ], + "type": "text", + "content": "The rise of embodied intelligence and multi-modal large language models has led to exciting advancements in the field of autonomous driving, establishing it as a prominent research focus in both academia and industry. However, when confronted with intricate and ambiguous traffic scenarios, the lack of logical reasoning and cognitive decision-making capabilities remains the primary challenge impeding the realization of embodied autonomous driving. Although Vision Language Models (VLMs) have enhanced the deep semantic understanding of autonomous driving systems, they exhibit notable limitations in decision explainability when handling rare and long-tail traffic scenarios. In this paper, we propose VLR-Driver, a novel multimodal Vision-Language-Reasoning (VLR) framework based on Chain of Thought (CoT) for embodied autonomous driving. The framework employs a spatiotemporal CoT reasoning approach to recursively analyze potential safety risks and driving intentions of other agents, thereby delivering an efficient and transparent decision-making process. Furthermore, we construct a multi-modal reasoning-decision dataset to support the advancement of hierarchical reasoning of VLMs in autonomous driving. Closed-loop experiments conducted in CARLA demonstrate that the VLR-Driver significantly outperforms state-of-the-art end-to-end methods. Notably, key metrics such as driving score improved by " + }, + { + "bbox": [ + 61, + 277, + 297, + 604 + ], + "type": "inline_equation", + "content": "17.5\\%" + }, + { + "bbox": [ + 61, + 277, + 297, + 604 + ], + "type": "text", + "content": ", while the success rate improved by " + }, + { + "bbox": [ + 61, + 277, + 297, + 604 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 61, + 277, + 297, + 604 + ], + "type": "text", + "content": ", offering a more transparent, reliable, and secure solution for autonomous driving systems." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 248, + 548, + 302 + ], + "blocks": [ + { + "bbox": [ + 317, + 248, + 548, + 302 + ], + "lines": [ + { + "bbox": [ + 317, + 248, + 548, + 302 + ], + "spans": [ + { + "bbox": [ + 317, + 248, + 548, + 302 + ], + "type": "image", + "image_path": "4d7db6f7b1621f5d1f8206748972db8364e81136fe3ac3e7acad0be138f8c717.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 303, + 436, + 310 + ], + "lines": [ + { + "bbox": [ + 427, + 303, + 436, + 310 + ], + "spans": [ + { + "bbox": [ + 427, + 303, + 436, + 310 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 312, + 548, + 366 + ], + "blocks": [ + { + "bbox": [ + 317, + 312, + 548, + 366 + ], + "lines": [ + { + "bbox": [ + 317, + 312, + 548, + 366 + ], + "spans": [ + { + "bbox": [ + 317, + 312, + 548, + 366 + ], + "type": "image", + "image_path": "d2ccdb0133d42db9d53bfa0246765ac5346ab9173c4815e395e22c3071f52108.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 368, + 436, + 375 + ], + "lines": [ + { + "bbox": [ + 427, + 368, + 436, + 375 + ], + "spans": [ + { + "bbox": [ + 427, + 368, + 436, + 375 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 377, + 548, + 432 + ], + "blocks": [ + { + "bbox": [ + 317, + 377, + 548, + 432 + ], + "lines": [ + { + "bbox": [ + 317, + 377, + 548, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 548, + 432 + ], + "type": "image", + "image_path": "fb93c92b0dd8fcd295e73c1b7dab290144df030ebf8ab07dd24e57c0732ffd2b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 433, + 435, + 440 + ], + "lines": [ + { + "bbox": [ + 427, + 433, + 435, + 440 + ], + "spans": [ + { + "bbox": [ + 427, + 433, + 435, + 440 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 449, + 550, + 524 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 550, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 550, + 524 + ], + "type": "text", + "content": "Figure 1. Comparison of different LLM-based AD systems. (a) VLM model focuses more on answering questions, but it is difficult to convert decisions into coherent control signals. (b) VLA model focuses more on predicting driving control signals, but lacks interpretability. (c) The VLR model generates decisions and controls signals with transparent reasoning processes through CoT, enhancing the driver trust in the system." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 546, + 394, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 546, + 394, + 558 + ], + "spans": [ + { + "bbox": [ + 315, + 546, + 394, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 565, + 550, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 550, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 550, + 706 + ], + "type": "text", + "content": "In recent years, the rapid progress of End-to-End (E2E) architectures [6, 7, 18, 49], Large Language Models (LLMs) [12, 41, 48], and embodied intelligence [26, 53, 55] has established these technologies as key enablers of innovation in Autonomous Driving (AD). Especially, VLMs enriched with extensive pre-training knowledge exhibit strong spatial understanding and common-sense reasoning abilities. DriveVLM [35] leverages VLM to enhance spatial awareness and planning capabilities in complex driving scenarios. CoVLA [2] integrates visual perception, language understanding, and action planning, effectively describing traffic scenarios and generating executable control actions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 639, + 149, + 649 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 639, + 149, + 649 + ], + "spans": [ + { + "bbox": [ + 73, + 639, + 149, + 649 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 61, + 649, + 297, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 649, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 649, + 297, + 704 + ], + "type": "text", + "content": "This research was supported in part by National Natural Science Foundation of China (No. 62088102, No. U24A20291), National Key Research and Development Program of China (No. 2024YFB4405504), the Fundamental Research Funds for the Central Universities (xtr072022001), the Opening Foundation of Key Laboratory of CTCIE, Ministry of Education (CTCIE20240101)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26966" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 294, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 294, + 233 + ], + "type": "text", + "content": "However, the decision-making process of VLMs often functions as a \"black box\", making it challenging to trace and interpret their underlying logic. This makes it difficult for AD systems to be fully trusted by drivers when encountering complex and emergency situations, such as illegal roadside parking, navigating intersections without traffic signals, and managing complex mixed-traffic interactions between motorized and non-motorized vehicles, thereby limiting their reliability and safety in real-world applications [39, 46, 54]. Moreover, most VLMs are trained on internet data, lacking spatial understanding and specialized training in the field of AD, making it difficult for them to fully adapt to dynamic and complex driving scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 236, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 295, + 469 + ], + "type": "text", + "content": "Meanwhile, CoT reasoning demonstrates strong inference, interpretability, and generalization capabilities by breaking down complex tasks into intermediate reasoning steps [40]. CoT enables systems to think step by step rather than relying on E2E black-box predictions, making it one of the key approaches toward achieving embodied AD. Sce2DriveX [52] enhances comprehensive perception and reasoning by introducing a multi-modal LLM framework with CoT, enabling a deeper understanding of spatiotemporal relationships and road topology. Similarly, DriveCoT [38] integrates CoT reasoning to improve decision-making interpretability and controllability in AD systems. However, existing CoT-based methods typically rely on open-ended language generation for reasoning, which lacks structural constraints. The generated reasoning steps may exhibit causal inversion or unnecessary vocabulary, making them uncontrollable in safety-critical tasks. Additionally, current methods primarily operate on static snapshots rather than continuous temporal sequences, limiting their ability to predict future events in dynamic traffic environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 472, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 472, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 472, + 295, + 706 + ], + "type": "text", + "content": "To bridge these gaps, we introduce VLR-Driver, a hierarchical CoT-based visual-language-reasoning model designed for closed-loop embodied AD. Our approach integrates the spatiotemporal features from cross-modal data, including multi-frame multi-view images and ego-vehicle control signals, by employing a SpatioTemporal CoT (STCoT) strategy that produces human-like reflective reasoning processes and driving action decisions. Additionally, we adopt a dual-phase training strategy, combining Low-Rank Adaptation (LoRA) [16] with an improved Stepwise Group Relative Policy Optimization (Step-GRPO) [34], significantly enhancing memory capacity and deep reasoning abilities. Our proposed VLR-Driver not only inherits the global action optimization capability of VLA models, but also preserves the transparency of modular rule-based methods. When encountering long-tail events and rare traffic scenarios, it demonstrates exceptional reflective reasoning and step-by-step inference, thereby enhancing the trust of human drivers in AD systems. The differences between VLM, VLA, and our proposed VLR are illustrated in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 81, + 547, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 186 + ], + "type": "text", + "content": "To further enhance VLR models in environmental understanding, reasoning, and decision-making, we introduce the VLR-Driver Dataset. This data set includes detailed scene descriptions, weather information, vehicle state details, and most critically, human-like CoT reasoning processes and the corresponding driving decisions. Various experiments demonstrate that VLR-Driver is capable of making accurate driving decisions and coherent reasoning, even under challenging and highly dynamic road conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 186, + 547, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 547, + 209 + ], + "type": "text", + "content": "The primary contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 211, + 547, + 467 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 211, + 547, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 211, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 314, + 211, + 547, + 269 + ], + "type": "text", + "content": "- Distinctive VLR-Driver Framework. We introduce VLR-Driver, a visual-language-reasoning model developed for embodied AD. It generates a human-like reflective reasoning process within the driving system, enabling accurate driving decisions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 269, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 269, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 269, + 547, + 327 + ], + "type": "text", + "content": "- Spatiotemporal CoT. We present a spatiotemporal CoT strategy that recursively analyzes potential safety risks and the driving intentions of moving agents in complex traffic scenarios, ensuring that the reasoning process of driving decisions remains transparent and interpretable." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 327, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 547, + 396 + ], + "type": "text", + "content": "- Advanced VLR-Driver Dataset. We construct VLR-Driver Dataset, a cutting-edge visual-language-reasoning and decision-making dataset specifically designed for AD. It supports the enhancement of spatiotemporal understanding and reflective reasoning capabilities in embodied AD systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 397, + 547, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 547, + 467 + ], + "type": "text", + "content": "- Superior Performance in Closed-Loop Simulations. Extensive closed-loop experiments are conducted on the CARLA platform. On the Bench2Drive benchmark, our approach achieves a " + }, + { + "bbox": [ + 313, + 397, + 547, + 467 + ], + "type": "inline_equation", + "content": "17.5\\%" + }, + { + "bbox": [ + 313, + 397, + 547, + 467 + ], + "type": "text", + "content": " improvement in driving score and a " + }, + { + "bbox": [ + 313, + 397, + 547, + 467 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 313, + 397, + 547, + 467 + ], + "type": "text", + "content": " increase in success rate, providing a more human-like, reliable, and trustworthy solution." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 480, + 399, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 399, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 399, + 492 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 500, + 485, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 485, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 485, + 513 + ], + "type": "text", + "content": "2.1. End-to-end Autonomous Driving" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 518, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 547, + 705 + ], + "type": "text", + "content": "The rapid advancement of E2E-AD has fostered a growing transition away from modular rule-based methods toward data-driven approaches [15, 23]. Based on different input modalities, E2E methods can be categorized into visual-only methods [5, 8, 10, 17] and vision-LiDAR fusion methods [1, 20, 21]. TCP [42] and NEAT [10] adopt imitation learning methods, training directly on collected state-action pair datasets, demonstrating the feasibility of E2E approaches for AD. Roach [51] utilizes reinforcement learning experts as coaches, delivering dense and informative supervision signals to agents equipped with monocular camera inputs. However, vision-based methods inherently struggle with distance and depth estimation. These limitations may compromise the reliability of driving decisions [24]. To address these limitations, vision-LiDAR fusion methods such as TransFuser [11], CrossFuser [43]," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26967" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 82, + 111, + 95 + ], + "blocks": [ + { + "bbox": [ + 99, + 82, + 111, + 95 + ], + "lines": [ + { + "bbox": [ + 99, + 82, + 111, + 95 + ], + "spans": [ + { + "bbox": [ + 99, + 82, + 111, + 95 + ], + "type": "image", + "image_path": "eac4ff6d82c2ec490e6391963617430d0427d8e7bb9e1e00dd4386fe5bbcfc5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 83, + 220, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 83, + 220, + 94 + ], + "spans": [ + { + "bbox": [ + 116, + 83, + 220, + 94 + ], + "type": "text", + "content": "VLR-Driver Datasets" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 97, + 264, + 343 + ], + "blocks": [ + { + "bbox": [ + 63, + 97, + 264, + 343 + ], + "lines": [ + { + "bbox": [ + 63, + 97, + 264, + 343 + ], + "spans": [ + { + "bbox": [ + 63, + 97, + 264, + 343 + ], + "type": "image", + "image_path": "6e6932366bf55cbdcfb9be596995fdfe69694031a9c0bc386017b65fd043bcbb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 352, + 548, + 396 + ], + "lines": [ + { + "bbox": [ + 61, + 352, + 548, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 352, + 548, + 396 + ], + "type": "text", + "content": "Figure 2. Overview of VLR-Driver framework. We introduce VLR-Driver Dataset, an advanced visual-language-reasoning dataset designed for AD, featuring detailed annotations of scene descriptions, analytical reasoning, and behavioral decisions. We present VLR-Driver Embodied AI (EAI), a novel multi-modal visual-language-reasoning framework for embodied AD that leverages a hierarchical spatiotemporal CoT reasoning mechanism." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 81, + 374, + 95 + ], + "blocks": [ + { + "bbox": [ + 354, + 81, + 374, + 95 + ], + "lines": [ + { + "bbox": [ + 354, + 81, + 374, + 95 + ], + "spans": [ + { + "bbox": [ + 354, + 81, + 374, + 95 + ], + "type": "image", + "image_path": "3f4f222795ee64fa75f5b7a6345fda0714df69ace388916c0dbe5e07f61b2cb6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 377, + 83, + 454, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 83, + 454, + 94 + ], + "spans": [ + { + "bbox": [ + 377, + 83, + 454, + 94 + ], + "type": "text", + "content": "VLR-Driver EAI" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 267, + 97, + 547, + 344 + ], + "blocks": [ + { + "bbox": [ + 267, + 97, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 267, + 97, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 267, + 97, + 547, + 344 + ], + "type": "image", + "image_path": "b57cb0b6dbc90c15f2082fad4efe9575391df0061330598a2549371e8d4afeda.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 412, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 412, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 61, + 412, + 296, + 495 + ], + "type": "text", + "content": "and FusionAD [47] have been designed to effectively integrate image with LiDAR data, significantly improving the robustness of AD systems in complex environments. However, different modalities contribute differently to the driving task, and spatiotemporal synchronization issues between sensor data can lead to inconsistencies, increasing both model complexity and training difficulty." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 263, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 263, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 263, + 514 + ], + "type": "text", + "content": "2.2. VLM and VLA in Autonomous Driving" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 518, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 518, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 61, + 518, + 296, + 693 + ], + "type": "text", + "content": "VLMs unify visual perception with natural language processing capabilities, enabling a more comprehensive understanding of driving environments [19, 36, 37]. By incorporating cross-modal data fusion, integrating textual information alongside visual inputs, these systems gain text comprehension and human interaction capabilities that conventional E2E models inherently lack. DriveGPT4 [45] enhances LLMs' ability to process multimodal inputs by projecting them into the text domain, thereby enabling interpretable E2E-AD. Senna [25] and DriveVLM [35] integrate VLMs with traditional modular pipelines or E2E frameworks, achieving decoupled high-level planning and low-level trajectory prediction. This approach improves planning performance while preserving the model's common sense reasoning capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 693, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 693, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 74, + 693, + 295, + 704 + ], + "type": "text", + "content": "Moreover, VLA is an emerging paradigm that unifies vi" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 412, + 548, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 548, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 548, + 691 + ], + "type": "text", + "content": "sual perception, natural language understanding, and action prediction within a cohesive framework [13, 26]. Originally introduced in the field of robotics, RT-2 [4] pioneered the representation of robotic actions as text tokens, seamlessly incorporating them alongside natural language labels into the model's training set. This enables direct transfer of internet-scale knowledge to robotic control, significantly enhancing both the generalization and semantic reasoning capabilities of robotic systems. For AD, CoVLA [2] introduces an interpretable VLA model that integrates visual perception, language-based scene understanding, and action planning. This integration improves the system's ability to comprehend complex driving scenarios, anticipate trajectory, and execute informed driving decisions. However, despite these advancements, VLA models still exhibit limitations in accurately predicting precise control action values, leading to a low lower-bound in AD performance. Furthermore, when confronted with highly dynamic and complex traffic scenarios, their deep understanding and reasoning capabilities remain insufficient. Unlike the above methods, our proposed VLR model places greater emphasis on step-by-step reasoning and the thought process of the model, providing drivers with increased confidence in using AD systems." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "type": "text", + "content": "26968" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 80, + 276, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 80, + 276, + 94 + ], + "spans": [ + { + "bbox": [ + 62, + 80, + 276, + 94 + ], + "type": "text", + "content": "2.3. Chain of Thought in Autonomous Driving" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 98, + 297, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 98, + 297, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 98, + 297, + 367 + ], + "type": "text", + "content": "The CoT technology is an extension of prompt engineering, proposed by Wei Jason in 2022 [40], which has greatly improved the effectiveness of reasoning for complex problems. The highly anticipated Deepseek-R1 [33] model also utilizes the CoT technique, which deeply integrates multimodal knowledge base data, enabling the model to generate a step-by-step thinking process. DriveCoT [38] has built a CoT dataset that includes sensor data, control decisions, and CoT labels used to indicate reasoning processes. The model trained on this basis can generate predictions and final decisions with CoT, effectively improving model performance. LanguageMPC [30] combines LLM with Model Predictive Control and decomposes driving decisions into multiple subtasks through a CoT framework. This method enables the auto drive system to think like human beings, and improves its ability to handle complex scenes. OpenEMMA [44] introduces CoT technology to guide model generation of detailed descriptions of key objects, behavioral insights, and meta driving decisions, improving system transparency and usability. Motivated by these advancements, we apply step-by-step hierarchical spatiotemporal CoT to AD, enhancing the interpretability of reasoning and decision-making." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 378, + 117, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 117, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 117, + 389 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 398, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 398, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 61, + 398, + 296, + 504 + ], + "type": "text", + "content": "We present the motivation and design details of our VLR-Driver framework. As depicted in Fig. 2, VLR-Driver comprises two main components: the VLR-Driver EAI and the VLR-Driver Dataset. Initially, we introduce the design concept of the VLR model, which builds upon enhancements to the VLA model (Sec. 3.1). Subsequently, we elaborate on the hierarchical spatiotemporal CoT methodology (Sec. 3.2) and the specifics of the dual-phase training strategy (Sec. 3.3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 512, + 128, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 128, + 522 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 128, + 522 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 529, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 529, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 61, + 529, + 296, + 693 + ], + "type": "text", + "content": "The VLR model is a large visual-language-reasoning model designed for embodied AD. It can processes visual inputs, such as multi-view images, alongside textual information, including vehicle control signals. The model is capable of extracting spatiotemporal key features within a multimodal embedding and recursively analyze potential safety risks and the driving intentions of other agents within the reasoning level of the VLR model. Ultimately, it formulates a comprehensive reasoning framework and well-structured decision-making outputs, explicitly identifying critical risk factors and the underlying rationale behind each decision. This process significantly reinforces the robustness and safety of the AD system, ensuring adaptive resilience in complex and dynamic environments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 693, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 693, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 73, + 693, + 295, + 704 + ], + "type": "text", + "content": "In this study, we employ pre-trained LLaVA-NeXT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "text", + "content": "Video [50] as the VLM and CLIP [29] as the visual encoder. The model is capable of processing multi-frame multi-view image data that capture historical temporal context, while also extracting real-time vehicle sensor information to facilitate dynamic and context-aware decision-making." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "content": "Input Representations. We utilize " + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "content": " frame and " + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "content": " view images from the past period, with a field of view (FoV) of 70 degrees. It can be represented as " + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{N_f \\times N_v \\times 3 \\times H_0 \\times W_0}" + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "content": ". At the same time, there are also the current position " + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "content": " of ego vehicle, the speed " + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "content": ", the target point position " + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "inline_equation", + "content": "(p,q)" + }, + { + "bbox": [ + 313, + 140, + 548, + 232 + ], + "type": "text", + "content": ". Subsequently, the compressed and cropped image data and the information from the ego's sensors are input into the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 233, + 548, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 548, + 338 + ], + "type": "text", + "content": "Output Representations. Our output consists of the reasoning process and driving decisions generated by the VLR model for the current driving scenario. This includes risk identification, traffic signal recognition, motion direction prediction, and AD decision-making. The driving decision will also be fused with the spatiotemporal information features extracted by the E2E model, and finally output the waypoints and control signals of the vehicle for the next moment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 480, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 480, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 480, + 359 + ], + "type": "text", + "content": "3.2. Spatiotemporal CoT Reasoning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 364, + 548, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 548, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 548, + 480 + ], + "type": "text", + "content": "To enhance the reasoning capabilities and transparency of the AD system, we introduce a hierarchical ST-CoT that guides the model to approach driving decisions in a human-like manner. Our method decomposes the driving decision-making process into two levels: a perception-level spatiotemporal CoT " + }, + { + "bbox": [ + 313, + 364, + 548, + 480 + ], + "type": "inline_equation", + "content": "C_{\\text{perception}}" + }, + { + "bbox": [ + 313, + 364, + 548, + 480 + ], + "type": "text", + "content": ", which focuses on extracting and understanding environmental dynamics, and a decision-level dynamic CoT " + }, + { + "bbox": [ + 313, + 364, + 548, + 480 + ], + "type": "inline_equation", + "content": "C_{\\text{decision}}" + }, + { + "bbox": [ + 313, + 364, + 548, + 480 + ], + "type": "text", + "content": ", which refines and optimizes decision-making based on contextual and temporal factors. The example of CoT is shown in Fig. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 487, + 432, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 432, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 432, + 498 + ], + "type": "text", + "content": "3.2.1. Perception Level CoT" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 548, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 548, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 548, + 631 + ], + "type": "text", + "content": "The spatiotemporal characteristics of the environment play a crucial role in the AD process. The perception level CoT is responsible for guiding the model to extract spatial and temporal features from input image data, identify and locate crucial objects in the traffic, such as vehicles, pedestrians, alien objects, traffic lights, traffic signs, etc., and extract historical behavioral features of dynamic agent based on temporal information. Our method enables VLR model to describe the current driving scenario, construct real-time spatial layout and dynamic changes of the environment, and achieve long-term planning for driving decisions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 647, + 548, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 548, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 548, + 706 + ], + "type": "text", + "content": "Spatial CoT. In driving scenarios, we primarily focus on obstacles that impact the ego's normal operation, including object categories " + }, + { + "bbox": [ + 313, + 647, + 548, + 706 + ], + "type": "inline_equation", + "content": "O_{type}" + }, + { + "bbox": [ + 313, + 647, + 548, + 706 + ], + "type": "text", + "content": " and distances with the ego " + }, + { + "bbox": [ + 313, + 647, + 548, + 706 + ], + "type": "inline_equation", + "content": "O_{dis}" + }, + { + "bbox": [ + 313, + 647, + 548, + 706 + ], + "type": "text", + "content": ". A critical aspect of safe driving is identifying potential risk points within the current lane. Additionally, when the ve" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26969" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 79, + 295, + 242 + ], + "blocks": [ + { + "bbox": [ + 64, + 79, + 295, + 242 + ], + "lines": [ + { + "bbox": [ + 64, + 79, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 64, + 79, + 295, + 242 + ], + "type": "image", + "image_path": "7f4cad0d991bd2d337f5ca0bbf6a670644dcf6e12afd98eace1dbff0f328f6a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 251, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 304 + ], + "type": "text", + "content": "Figure 3. Illustration of the ST-CoT reasoning process. In this scenario, where some vehicles are illegally parked ahead and blocking the lane, our method can conduct hierarchical paitotemporal reasoning analysis and make a decision of change left lane once the adjacent lane is free." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "text", + "content": "hicle executes lane changes, objects in adjacent lanes, both left and right, may significantly influence its movement. Beyond obstacles, key traffic light " + }, + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "inline_equation", + "content": "S_{light}" + }, + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "text", + "content": ", traffic signs " + }, + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "inline_equation", + "content": "S_{sign}" + }, + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "text", + "content": " and lane markings " + }, + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "inline_equation", + "content": "S_{mark}" + }, + { + "bbox": [ + 61, + 321, + 296, + 380 + ], + "type": "text", + "content": " are also integral to decision-making, ensuring comprehensive spatial awareness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 393, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 393, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 61, + 393, + 296, + 579 + ], + "type": "text", + "content": "Temporal CoT. While a single-frame image can provide a static representation of road scenes and traffic participants, it fails to capture the motion trends of moving agents. To address this limitation, we introduce consecutive frames " + }, + { + "bbox": [ + 61, + 393, + 296, + 579 + ], + "type": "inline_equation", + "content": "I = \\{I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br}\\}_{t = T_{now} - T}^{T_{now}}" + }, + { + "bbox": [ + 61, + 393, + 296, + 579 + ], + "type": "text", + "content": " into the model, allowing it to track temporal variations in object positioning, which " + }, + { + "bbox": [ + 61, + 393, + 296, + 579 + ], + "type": "inline_equation", + "content": "I_{f}, I_{fr}, I_{fl}, I_{b}, I_{bl}, I_{br}" + }, + { + "bbox": [ + 61, + 393, + 296, + 579 + ], + "type": "text", + "content": " represents the image view of front, front left, front right, back, back left and back right. These sequential frames not only offer instantaneous spatial context but also reveal motion trajectories and behavioral patterns through their inter-frame positional changes. This temporal information is essential for predicting dynamic object movement, assessing collision risks, and generating robust path planning strategies, ultimately enhancing the ability of anticipate and react to evolving traffic conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 585, + 170, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 585, + 170, + 596 + ], + "spans": [ + { + "bbox": [ + 62, + 585, + 170, + 596 + ], + "type": "text", + "content": "3.2.2. Decision Level CoT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 600, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 600, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 600, + 296, + 704 + ], + "type": "text", + "content": "The output information from the perception-level serves as a critical foundation for the driving decision level, enabling reliable AD behavior inference. Within the driving decision level, we account for complex dynamic environmental factors, transforming spatial and temporal information from the perception level into concrete driving decisions. Specifically, decision-makers must not only analyze the current driving environment in real time but also anticipate future behaviors of other traffic participants and make decisions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "text", + "content": "based on multiple factors. Throughout this process, the CoT in the driving decision level spans multiple perspectives, incorporating safety, efficiency, comfort, and compliance with traffic regulations to ensure that driving decisions meet safety standards while optimizing driving efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 140, + 547, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 547, + 221 + ], + "type": "text", + "content": "To further enhance the structured reasoning process in AD, we have carefully designed hierarchical reasoning prompts that guide decision-making. Our structured prompts follow a logical sequence of \"risk point recognition — driving intention prediction — driving decision selection\", forming a cohesive reasoning chain aligned with human cognitive driving patterns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 235, + 547, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 547, + 303 + ], + "type": "text", + "content": "Risk Point Recognition. In this initial stage, the prompt-driven model conducts a comprehensive perception and analysis of the driving environment. This includes recognizing and evaluating critical elements such as traffic signs, lane markings, pedestrians, and obstacles to identify potential risks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 318, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 547, + 388 + ], + "type": "text", + "content": "Driving Intention Prediction. Once risk points are identified, the model leverages dynamic target behavior prediction and scene understanding to infer the potential movements and intentions of other road users. For instance, the model assesses whether pedestrians are likely to cross the road or whether the vehicle ahead intends to change lanes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 402, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 547, + 449 + ], + "type": "text", + "content": "Driving Decision Selection. Based on the contextual information gathered from the first two stages, the model applies multi-modal information fusion and weighted decision-making to select the optimal driving maneuver." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 547, + 588 + ], + "type": "text", + "content": "Through this structured prompting strategy, the large model adheres to a progressive reasoning hierarchy, beginning with fundamental environmental perception and advancing to higher-level decision-making. By explicitly presenting the reasoning process in a clear and structured manner, this approach enhances passenger trust and confidence in the intelligent driving system. Moreover, the integration of recursive CoT reasoning enables the model to mimic the step-by-step thought process of human drivers, facilitating more flexible, reliable, and interpretable decision outputs in complex driving scenarios. The structured prompt framework are shown in Fig. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 423, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 423, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 423, + 607 + ], + "type": "text", + "content": "3.3. Training Paradigm" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "type": "text", + "content": "VLR-Driver adopts a dual-phase training strategy to optimize its reasoning and decision-making capabilities. In the first phase, LoRA [16] is utilized for supervised fine-tuning on a pre-trained large model, enabling efficient adaptation with minimal memory and computational overhead while maintaining strong performance. In the second phase, StepGRPO is applied for reinforcement learning based on human preferences, further enhancing the model's ability to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "type": "text", + "content": "26970" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 104 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 104 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 104 + ], + "type": "text", + "content": "exhibit human-like reasoning and decision-making characteristics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 297, + 224 + ], + "type": "text", + "content": "Training with LoRA. LoRA is a parameter-efficient finetuning technique that enables effective model adaptation by performing a low-rank decomposition of the weight matrix. The core principle behind LoRA is to decompose the weight matrix of a pre-trained model into a low-rank structure, significantly reducing the number of trainable parameters while preserving expressivity. Specifically, if the weight matrix in the pre-trained model is denoted as " + }, + { + "bbox": [ + 61, + 118, + 297, + 224 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 61, + 118, + 297, + 224 + ], + "type": "text", + "content": ", LoRA represents it as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 101, + 228, + 296, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 228, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 101, + 228, + 296, + 249 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} ^ {\\prime} = \\boldsymbol {W} _ {0} + \\Delta \\boldsymbol {W} = \\boldsymbol {W} _ {0} + \\frac {\\alpha}{r} \\boldsymbol {B} \\cdot \\boldsymbol {A}, \\tag {1}", + "image_path": "36599ae11ce0098cffe38b7310e10c9ea61ac42fc62f10d1946a62990fc7d50e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "content": " is the rank of a low rank matrix, usually much smaller than " + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 61, + 254, + 296, + 291 + ], + "type": "text", + "content": " is a scaling factor. The forward pass is computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 295, + 296, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 295, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 110, + 295, + 296, + 316 + ], + "type": "interline_equation", + "content": "y = \\boldsymbol {W} ^ {\\prime} x = \\left(\\boldsymbol {W _ {0}} + \\frac {\\alpha}{r} \\boldsymbol {B} \\cdot \\boldsymbol {A}\\right) x, \\tag {2}", + "image_path": "cf3547baf31334ff37ecfe6f54aaa72995e81a9de30d736c0d8be1fe9e5f873d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 320, + 206, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 320, + 206, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 320, + 206, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 62, + 320, + 206, + 331 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 62, + 320, + 206, + 331 + ], + "type": "text", + "content": " is the output and " + }, + { + "bbox": [ + 62, + 320, + 206, + 331 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 62, + 320, + 206, + 331 + ], + "type": "text", + "content": " is input." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 331, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 296, + 366 + ], + "type": "text", + "content": "We use LoRA for all linear modules, which not only saves computation but also ensures the performance of the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 379, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 379, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 61, + 379, + 296, + 456 + ], + "type": "text", + "content": "Training with GRPO. The core principle of GRPO [34] is to optimize strategies by assigning relative rewards to multiple outputs generated from the same prompt, thereby eliminating the need for additional value function models. The introduction of process reward model estimation in GRPO provides finer support for distributed rewards. The reward of each step of the outputs is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 459, + 296, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 459, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 67, + 459, + 296, + 479 + ], + "type": "interline_equation", + "content": "R = \\left\\{\\left\\{r _ {1} ^ {I (1)}, \\dots , r _ {1} ^ {I (K _ {1})} \\right\\}, \\dots , \\left\\{r _ {G} ^ {I (1)}, \\dots , r _ {G} ^ {I (K _ {G})} \\right\\} \\right\\}, \\tag {3}", + "image_path": "fe2ddbf5df319599b52f18df16d1fdbbbfd948a573166ed56f7494bc2c8897d4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 482, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 295, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 62, + 482, + 295, + 506 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 62, + 482, + 295, + 506 + ], + "type": "text", + "content": " is the end token index, and the reward need be normalized as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 509, + 296, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 509, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 123, + 509, + 296, + 537 + ], + "type": "interline_equation", + "content": "\\widetilde {r} _ {i} ^ {I (j)} = \\frac {r _ {i} ^ {I (j)} - \\operatorname {m e a n} (\\boldsymbol {R})}{\\operatorname {s t d} (\\boldsymbol {R})}. \\tag {4}", + "image_path": "793e38d1df444664990fe0b3c6a0a79642b23a9e9af83945cf87cd9328696dcb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 542, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 542, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 542, + 297, + 706 + ], + "type": "text", + "content": "We extend GRPO by introducing reasoning Step-GRPO, a supervised reasoning decision process that structures the output into multiple steps based on the CoT reasoning framework. At each step, a reward function is applied to evaluate and assign scores, enabling fine-grained feedback that enhances model interpretability and accelerates convergence. Specifically, we first generate multiple candidate decision answers for the current driving scenario using prompts within the VLR model; Then, following our ST-CoT strategy, the reasoning process is divided into four distinct steps: scene description, spatial risk point reasoning, dynamic trajectory prediction, and driving decision-making; Furthermore, each reasoning step is assigned a reward to encourage structured learning. Finally, we compare" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 81, + 548, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 548, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 548, + 140 + ], + "type": "text", + "content": "all answers within the group and calculate the Kullback-Leibler (KL) divergence to update the policy model. This grouping and step-by-step scoring strategy enhances training efficiency and reduces the likelihood of erroneous reasoning in the model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 154, + 432, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 432, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 432, + 166 + ], + "type": "text", + "content": "4. VLR-Driver Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 174, + 548, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 548, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 548, + 431 + ], + "type": "text", + "content": "To fully explore the reasoning and decision-making capabilities of LLMs, we propose an advanced reasoning and decision-making dataset for AD scenarios, called the VLR-Driver Datasets. This dataset relies on the CARLA [14] simulator for data collection and is expanded and meticulously annotated based on the Bench2Drive dataset [22]. It includes: a) multi-view multi-frame images or videos, b) valuable information for AD, such as road details, weather conditions, vehicle information, and scene descriptions, and c) driving decision choices along with the decision-making process based on an ST-CoT. This dataset provides a rich and comprehensive training foundation for AD reasoning and decision-making, allowing the agent to exhibit human-like reasoning while interacting with the environment. The dataset includes 20,000 sets of multi-frame, multi-angle image data collected from various road conditions such as urban, rural, and highways in the CARLA simulator, covering over 30 specific complex traffic scenarios (e.g., forward accidents, dynamic object crossings). Each image set provides road scene descriptions, environmental weather information, vehicle status data, and, most importantly, the humanlike reasoning process and driving behavior decisions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 441, + 408, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 408, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 408, + 452 + ], + "type": "text", + "content": "4.1. Data Collection" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 459, + 548, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 548, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 548, + 633 + ], + "type": "text", + "content": "We conducted data collection based on the 44 corner scene classifications provided by Bench2Drive to ensure optimal AD performance in various complex corner scenarios. The Bench2Drive dataset offers a rich array of data and annotations, including multi-angle images, lidar, radar, vehicle information, and expert assessments, which have played a crucial role in constructing our VLR dataset. Additionally, we selected over 30 scenarios where the reasoning and decision-making capabilities of AD systems are relatively limited, and used them as the visual and textual components of our dataset. We further expanded and enriched the dataset by collecting additional data in the CARLA simulator. We have taken into account various weather conditions, road conditions, and the types and numbers of traffic participants in the scene." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 548, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 548, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 548, + 706 + ], + "type": "text", + "content": "Weather. For each scene, we randomly set values for cloudiness, fog density, precipitation deposits, sun altitude angle, sun azimuth angle, wetness, and wind intensity. The combinations of these parameters cover a variety of weather conditions, such as rainy, foggy, broken sky, and stormy, as well as different lighting conditions for day and night." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26971" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 128, + 548, + 286 + ], + "blocks": [ + { + "bbox": [ + 61, + 65, + 548, + 120 + ], + "lines": [ + { + "bbox": [ + 61, + 65, + 548, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 65, + 548, + 120 + ], + "type": "text", + "content": "Table 1. The comparison of core metrics and subdivision infraction scores with state-of-the-art E2E/VLM models on the Bench2Drive benchmark. C, L and T indicate camera, LiDAR and text modalities, respectively. DS, RC, IS, SR correspond to the Driving Score, Route Completion, Infraction Score, and Success Rate. CP, CV, CL, RL, SS, OR, AB, YEV correspond to the Collision with a Pedestrian, Collision with another Vehicle, Collision with Layout, Red Light infractions, Stop Sign infractions, Off-Road infractions, Agent Blocked, and failure to Yield to Emergency Vehicles infractions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 128, + 548, + 286 + ], + "lines": [ + { + "bbox": [ + 64, + 128, + 548, + 286 + ], + "spans": [ + { + "bbox": [ + 64, + 128, + 548, + 286 + ], + "type": "table", + "html": "
MethodTypeModalityCore Metrics ↑Subdivision Infraction Score ↓
DSRCISSRCPCVCLRLSSORABYEV
NEAT [10]E2EC30.8655.350.556.811.089.875.570.201.330.412.010.27
TCP [42]C56.2883.570.6525.000.265.465.460.000.520.220.780.00
LeTFuser [1]C+L52.5377.680.6718.181.165.543.790.290.870.120.580.29
LateFusion [28]C+L48.5358.320.8518.180.383.111.550.380.770.061.550.38
TransFuser [11]C+L37.1868.140.519.090.9613.248.710.000.960.322.580.32
ThinkTwice [21]C+L58.7974.350.7729.540.305.760.910.000.910.050.910.30
EATNet [9]C+L42.9778.840.5415.910.8214.011.920.001.640.221.090.27
InterFuser [31]C+L63.8180.460.7940.900.353.810.540.271.080.050.540.27
LMDrive [32]VLMC+L+T24.7633.020.9013.631.142.862.290.000.570.053.440.57
LeapAD [27]C+T55.1877.450.7136.360.695.071.150.200.910.081.470.27
VLR-Driver (Ours)VLRC+T75.0186.090.8750.000.722.830.480.000.480.030.250.25
", + "image_path": "7dea093c1b16e5237328d5e8c79813c37d21b02a6fbfb33d9742faa988b6efbd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 64, + 352, + 297, + 466 + ], + "blocks": [ + { + "bbox": [ + 61, + 289, + 297, + 344 + ], + "lines": [ + { + "bbox": [ + 61, + 289, + 297, + 344 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 297, + 344 + ], + "type": "text", + "content": "Table 2. The comparison of advanced driving ability and experience score with state-of-the-art models on the Bench2Drive benchmark. OT, MER, EB, GW, TS, DE, SC correspond to the OverTaking, MERging, Emergency Brake, Give Way, Traffic Sign, Driving Efficiency, and Smoothness Control." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 352, + 297, + 466 + ], + "lines": [ + { + "bbox": [ + 64, + 352, + 297, + 466 + ], + "spans": [ + { + "bbox": [ + 64, + 352, + 297, + 466 + ], + "type": "table", + "html": "
MethodAdvanced Driving Ability ↑Exper. Score ↑
OTMEREBGWTSMeanDESC
NEAT [10]0.006.669.0950.0027.7718.7092.090.30
TCP [42]25.0013.3327.2750.0050.0033.12114.520.27
LeTFuser [1]0.0020.0018.180.0047.2217.08115.090.47
LateFusion [28]0.0020.009.090.0036.1113.04104.390.59
TransFuser [11]0.0013.339.090.0036.1111.7195.200.36
ThinkTwice [21]12.5020.0036.3650.0052.7734.3391.170.36
EATNet [9]0.0013.3318.180.0041.6614.6388.130.33
InterFuser [31]0.0046.6654.5450.0061.1142.46119.200.32
LMDrive [32]25.006.669.0950.002.7718.7075.410.22
LeapAD [27]12.5033.3327.2750.0044.4433.5193.330.26
VLR-Driver (Ours)37.5046.6672.7250.0072.2255.82125.220.59
", + "image_path": "4c28dd97e53304017a3a946f4f28d9e9814eab470dd8b2b87985136893d02b58.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 483, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 483, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 61, + 483, + 295, + 517 + ], + "type": "text", + "content": "Roads. The scenes include various road types such as urban two-way single-lane roads, multi-lane roads, highways, and narrow rural roads." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 518, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 518, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 61, + 518, + 296, + 566 + ], + "type": "text", + "content": "Traffic Participants. Different corner cases involve various traffic participants, including cars, bicycles, pedestrians, ambulances, etc., simulating the complex traffic conditions encountered in daily driving." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 571, + 162, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 162, + 583 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 162, + 583 + ], + "type": "text", + "content": "4.2. Data Annotation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 588, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 588, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 61, + 588, + 295, + 646 + ], + "type": "text", + "content": "The reasoning chain process and the decision-making choices are crucial for training large AD models and enhancing the model's human-like reasoning ability. Therefore, we performed secondary annotation on the dataset we collected." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 647, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 647, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 647, + 296, + 706 + ], + "type": "text", + "content": "Driving Scene Descriptions. We used the pre-trained large visual language model Qwen2-VL [3] to generate detailed descriptions for the corresponding driving scenes. These descriptions primarily focus on environmental information such as road conditions, weather, and lighting, as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 305, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 547, + 327 + ], + "type": "text", + "content": "well as dynamic targets like vehicles and pedestrians that may pose driving risks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 329, + 548, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 548, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 548, + 445 + ], + "type": "text", + "content": "Reasoning Decisions and Process. We considered pre-annotated information such as vehicle speed, acceleration, steering angle, traffic light status, and the state of the vehicle ahead. A rule-based method was used to determine the true values of future motion behaviors based on the decision choices made at earlier time steps. Additionally, we completed the predefined CoT reasoning text statements. Finally, to ensure the accuracy and consistency of the annotations, especially for decision choices, carefully manual verification was carried out." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 389, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 389, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 389, + 470 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 427, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 427, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 427, + 488 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 493, + 548, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 548, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 548, + 598 + ], + "type": "text", + "content": "Our method was validated on the open source AD simulation platform CARLA 0.9.15 [14]. The VLR-Driver model was trained on a server equipped with 8 NVIDIA A800 GPUs (each with 80G of video memory) for approximately 50 hours. The dataset used was the VLR-Driver dataset that we developed. Specifically, we use ViT-g/14 from EVA-CLIP [29] as the vision encoder and LLaVA-NeXT-Video7B [50] as the VLM. The resolution of the input image is set to " + }, + { + "bbox": [ + 313, + 493, + 548, + 598 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 313, + 493, + 548, + 598 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 606, + 372, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 372, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 372, + 617 + ], + "type": "text", + "content": "5.2. Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 623, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 548, + 705 + ], + "type": "text", + "content": "We employ four core metrics to evaluate AD performance: driving score (DS), route completion (RC), infraction score (IS), and success rate (SR). Additionally, to provide a more granular assessment of model performance in specific aspects, we introduce three supplementary evaluation metrics: subdivision infraction score, advanced driving ability, and driving experience score [22]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26972" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 79, + 294, + 346 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 294, + 346 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 294, + 346 + ], + "type": "image", + "image_path": "a16dfdb5fe44cee491dc6f5301800d3dda384dec0c2acf22c560d592e5f0c136.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 354, + 296, + 418 + ], + "lines": [ + { + "bbox": [ + 61, + 354, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 296, + 418 + ], + "type": "text", + "content": "Figure 4. Visual comparison between VLR-Driver and VLM-based methods. The ST-CoT guides the VLR model to approach driving decisions in a human-like spatiotemporal manner. Based on the sequence of images from the preceding time period " + }, + { + "bbox": [ + 61, + 354, + 296, + 418 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 61, + 354, + 296, + 418 + ], + "type": "text", + "content": ", we derive the following inference results. The images captured at 5 and 10 seconds afterward validate the accuracy of our decisions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 438, + 251, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 251, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 251, + 451 + ], + "type": "text", + "content": "5.3. Comparisons with Existing Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "text", + "content": "We conducted comprehensive experiments with the SOTA methods including E2E and VLM in the CARLA simulator across 44 routes from the Bench2Drive benchmark. We present comparison result in Tab. 1. It can be seen that our method outperforms other methods in key metrics such as DS, RC, and SR, achieving first place and effectively improving DS by " + }, + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "inline_equation", + "content": "17.5\\%" + }, + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "text", + "content": ", mean advanced driving ability by " + }, + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "inline_equation", + "content": "31.4\\%" + }, + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "text", + "content": ", and SR by " + }, + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "inline_equation", + "content": "22.2\\%" + }, + { + "bbox": [ + 61, + 457, + 296, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 553, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 553, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 553, + 296, + 705 + ], + "type": "text", + "content": "The comparison results of the advanced driving ability and driving experience score of each method are shown in Tab. 2. Our method achieved the best results in all abilities, thanks to the deep reflection and reasoning ability of our VLR model, which has stronger traffic reasoning capacity in special road conditions. Most E2E methods can only achieve following the vehicle, but when there is a vehicle temporarily parked in the lane ahead, blocking the ego driving route, they will keep stopping and waiting, making it impossible to complete the entire route. In contrast, our VLR-Driver can achieve deep understanding and inference of the current scene through LLM inference, so as to make timely detours." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 85, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 358, + 66, + 504, + 76 + ], + "lines": [ + { + "bbox": [ + 358, + 66, + 504, + 76 + ], + "spans": [ + { + "bbox": [ + 358, + 66, + 504, + 76 + ], + "type": "text", + "content": "Table 3. Ablation study for each module." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 85, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 316, + 85, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 316, + 85, + 547, + 159 + ], + "type": "table", + "html": "
IDAbal.Exp.Core Metrics ↑Advanced Driving Ability ↑
DSSROTMEREB
1VLRFull Model75.0150.0037.5046.6672.72
2Arch.w/o CoT57.1734.0925.0040.0041.66
3w/o VLR-Model46.3920.4514.2826.6636.36
4Trainw/o VLR-Data52.8527.2725.0026.669.09
5w/o Step-GRPO65.5745.4537.5040.0063.63
", + "image_path": "daecc8e1f82e8c434cdf24e4fcff3f6c09322b745d69785cfc547b43b4c205c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 183, + 405, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 405, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 405, + 196 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 202, + 547, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 547, + 318 + ], + "type": "text", + "content": "We conducted a comprehensive ablation study, detailed in Tab. 3. The experimental configurations include four variants: (1) Without utilizing our proposed spatiotemporal CoT strategy, using only a question-based approach without reasoning guidance. (2) Without using our VLR model, instead employing a standard LLM module. (3) Removing the VLR data used to guide the reasoning process. (4) Without Step-GRPO reinforcement learning training, using only LoRA strategies to fine-tune the model with supervision. The results show the effectiveness of each contribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 395, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 395, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 395, + 342 + ], + "type": "text", + "content": "5.5. Visualization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 351, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 547, + 421 + ], + "type": "text", + "content": "We selected some special scenarios to visualize the performance in complex traffic situations with reflective reasoning, and the comparison results shown in Fig. 4. The ST-CoT enables the model to make driving decisions with human-like reasoning, considering both spatial and temporal dynamics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 384, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 384, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 384, + 449 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 458, + 547, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 547, + 620 + ], + "type": "text", + "content": "In this paper, we introduce VLR-Driver, a VLR model for embodied AD. It leverages a carefully designed ST-CoT strategy to guide the model in recursively analyzing potential safety risks and the driving intentions of dynamic agents in complex traffic scenarios. Our dual-phase training method significantly enhances the generalization of the model. Additionally, we propose the VLR-Driver dataset, which effectively integrates spatiotemporal perception, language understanding, and reflective reasoning, providing crucial support for interpretability reasoning in AD system. Experimental results show that VLR-Driver outperforms other methods on Bench2Drive benchmark, achieving cutting-edge performance and paving the way for embodied AI realization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 547, + 704 + ], + "type": "text", + "content": "Limitations and future work. There are differences between the data in simulation platforms and the real world. Deploying LLMs for AD on real-world vehicles with limited onboard computing resources remains a key challenge for future exploration." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26973" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 99, + 296, + 705 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 99, + 296, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 296, + 142 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 296, + 142 + ], + "type": "text", + "content": "[1] Pedram Agand, Mohammad Mahdavianian, Manolis Savva, and Mo Chen. Letfuser: Light-weight end-to-end transformer-based sensor fusion for autonomous driving with multi-task learning. arXiv preprint arXiv:2310.13135, 2023. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 143, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 143, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 67, + 143, + 296, + 206 + ], + "type": "text", + "content": "[2] Hidehisa Arai, Keita Miwa, Kento Sasaki, Kohei Watanabe, Yu Yamaguchi, Shunsuke Aoki, and Issei Yamamoto. Covla: Comprehensive vision-language-action dataset for autonomous driving. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1933-1943. IEEE, 2025. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 207, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 207, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 67, + 207, + 296, + 250 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 251, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 251, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 296, + 304 + ], + "type": "text", + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 304, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 304, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 67, + 304, + 296, + 346 + ], + "type": "text", + "content": "[5] Daoming Chen, Ning Wang, Feng Chen, and Tony Pipe. Detrive: Imitation learning with transformer detection for end-to-end autonomous driving. arXiv preprint arXiv:2310.14224, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 348, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 348, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 67, + 348, + 296, + 401 + ], + "type": "text", + "content": "[6] Jianyu Chen, Zhuo Xu, and Masayoshi Tomizuka. End-to-end autonomous driving perception with sequential latent representation learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1999-2006. IEEE, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 402, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 402, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 67, + 402, + 296, + 445 + ], + "type": "text", + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 445, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 445, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 67, + 445, + 296, + 498 + ], + "type": "text", + "content": "[8] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 499, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 499, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 67, + 499, + 296, + 563 + ], + "type": "text", + "content": "[9] Weihuang Chen, Fanjie Kong, Liming Chen, Shen'ao Wang, Zhiping Wang, and Hongbin Sun. Eatnet: Efficient axial transformer network for end-to-end autonomous driving. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 762-767. IEEE, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 564, + 296, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 564, + 296, + 607 + ], + "spans": [ + { + "bbox": [ + 63, + 564, + 296, + 607 + ], + "type": "text", + "content": "[10] Kashyap Chitta, Aditya Prakash, and Andreas Geiger. Neat: Neural attention fields for end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15793-15803, 2021. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 608, + 296, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 608, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 63, + 608, + 296, + 661 + ], + "type": "text", + "content": "[11] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE transactions on pattern analysis and machine intelligence, 45(11):12878-12895, 2022. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 662, + 296, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 662, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 63, + 662, + 296, + 705 + ], + "type": "text", + "content": "[12] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 82, + 547, + 705 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 82, + 547, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 82, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 333, + 82, + 547, + 103 + ], + "type": "text", + "content": "IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 105, + 547, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 105, + 547, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 105, + 547, + 158 + ], + "type": "text", + "content": "[13] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. Quar-vla: Vision-language-action model for quadruped robots. In European Conference on Computer Vision, pages 352-367. Springer, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 159, + 547, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 159, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 316, + 159, + 547, + 191 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, pages 1-16. PMLR, 2017. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 192, + 547, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 192, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 316, + 192, + 547, + 245 + ], + "type": "text", + "content": "[15] Hao Gao, Shaoyu Chen, Bo Jiang, Bencheng Liao, Yiang Shi, Xiaoyang Guo, Yuechuan Pu, Haoran Yin, Xiangyu Li, Xinbang Zhang, et al. Rad: Training an end-to-end driving policy via large-scale 3dgs-based reinforcement learning. arXiv preprint arXiv:2502.13144, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 246, + 547, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 547, + 289 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 290, + 547, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 290, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 316, + 290, + 547, + 344 + ], + "type": "text", + "content": "[17] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 345, + 547, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 345, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 345, + 547, + 399 + ], + "type": "text", + "content": "[18] Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3903-3913, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 547, + 443 + ], + "type": "text", + "content": "[19] JJ Hwang, R Xu, H Lin, WC Hung, J Ji, K Choi, D Huang, T He, P Covington, B Sapp, et al. Emma: End-to-end multimodal model for autonomous driving. arxiv 2024. arXiv preprint arXiv:2410.23262. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 443, + 547, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 547, + 508 + ], + "type": "text", + "content": "[20] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 508, + 547, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 547, + 572 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 547, + 572 + ], + "type": "text", + "content": "[21] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 573, + 547, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 573, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 316, + 573, + 547, + 617 + ], + "type": "text", + "content": "[22] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 617, + 547, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 617, + 547, + 660 + ], + "spans": [ + { + "bbox": [ + 316, + 617, + 547, + 660 + ], + "type": "text", + "content": "[23] Xiaosong Jia, Junqi You, Zhiyuan Zhang, and Junchi Yan. Drivetransformer: Unified transformer for scalable end-to-end autonomous driving. arXiv preprint arXiv:2503.07656, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 661, + 547, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 661, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 316, + 661, + 547, + 705 + ], + "type": "text", + "content": "[24] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26974" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 82, + 82, + 295, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 82, + 295, + 102 + ], + "spans": [ + { + "bbox": [ + 82, + 82, + 295, + 102 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 104, + 296, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 104, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 64, + 104, + 296, + 157 + ], + "type": "text", + "content": "[25] Bo Jiang, Shaoyu Chen, Benchcheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 159, + 295, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 159, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 64, + 159, + 295, + 191 + ], + "type": "text", + "content": "[26] Yueen Ma, Zixing Song, Yuzheng Zhuang, Jianye Hao, and Irwin King. A survey on vision-language-action models for embodied ai. arXiv preprint arXiv:2405.14093, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 192, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 192, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 64, + 192, + 295, + 245 + ], + "type": "text", + "content": "[27] Jianbiao Mei, Yukai Ma, Xuemeng Yang, Licheng Wen, Xinyu Cai, Xin Li, Daocheng Fu, Bo Zhang, Pinlong Cai, Min Dou, et al. Continuously learning, adapting, and improving: A dual-process approach to autonomous driving. arXiv preprint arXiv:2405.15324, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 246, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 246, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 64, + 246, + 295, + 278 + ], + "type": "text", + "content": "[28] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 280, + 295, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 280, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 64, + 280, + 295, + 333 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 334, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 334, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 64, + 334, + 295, + 376 + ], + "type": "text", + "content": "[30] H Sha, Y Mu, Y Jiang, L Chen, C Xu, P Luo, SE Li, M Tomizuka, W Zhan, and M Ding. Linguistics: Large language models as decision makers for autonomous driving. arxiv 2023. arXiv preprint arXiv:2310.03026. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 378, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 378, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 64, + 378, + 295, + 421 + ], + "type": "text", + "content": "[31] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 422, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 422, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 64, + 422, + 295, + 485 + ], + "type": "text", + "content": "[32] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 487, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 487, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 64, + 487, + 295, + 540 + ], + "type": "text", + "content": "[33] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 541, + 295, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 541, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 64, + 541, + 295, + 594 + ], + "type": "text", + "content": "[34] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 596, + 295, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 596, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 64, + 596, + 295, + 649 + ], + "type": "text", + "content": "[35] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 651, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 651, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 64, + 651, + 295, + 704 + ], + "type": "text", + "content": "[36] Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models. arXiv preprint arXiv:2410.05051, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 315, + 82, + 547, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 82, + 547, + 135 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 547, + 135 + ], + "type": "text", + "content": "[37] Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 137, + 547, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 137, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 137, + 547, + 178 + ], + "type": "text", + "content": "[38] Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving. arXiv preprint arXiv:2403.16996, 2024. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 180, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 547, + 243 + ], + "type": "text", + "content": "[39] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving. arXiv preprint arXiv:2312.09245, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 245, + 547, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 547, + 298 + ], + "type": "text", + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 300, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 547, + 342 + ], + "type": "text", + "content": "[41] Jiayang Wu, Wensheng Gan, Zefeng Chen, Shicheng Wan, and Philip S Yu. Multimodal large language models: A survey. In 2023 IEEE International Conference on Big Data (BigData), pages 2247-2256. IEEE, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 344, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 547, + 396 + ], + "type": "text", + "content": "[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 398, + 547, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 547, + 451 + ], + "type": "text", + "content": "[43] Weishang Wu, Xiaoheng Deng, Ping Jiang, Shaohua Wan, and Yuanxiong Guo. Crossfuser: Multi-modal feature fusion for end-to-end autonomous driving under unseen weather conditions. IEEE Transactions on Intelligent Transportation Systems, 24(12):14378-14392, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 547, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 547, + 506 + ], + "type": "text", + "content": "[44] Shuo Xing, Chengyuan Qian, Yuping Wang, Hongyuan Hua, Kexin Tian, Yang Zhou, and Zhengzhong Tu. Openemma: Open-source multimodal model for end-to-end autonomous driving. In Proceedings of the Winter Conference on Applications of Computer Vision, pages 1001-1009, 2025. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 507, + 547, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 507, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 507, + 547, + 559 + ], + "type": "text", + "content": "[45] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 561, + 547, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 561, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 316, + 561, + 547, + 594 + ], + "type": "text", + "content": "[46] Z Yang, X Jia, H Li, and J Yan. Llm4drive: A survey of large language models for autonomous driving. arxiv 2023. arXiv preprint arXiv:2311.01043. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 595, + 547, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 595, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 316, + 595, + 547, + 648 + ], + "type": "text", + "content": "[47] Tengju Ye, Wei Jing, Chunyong Hu, Shikun Huang, Lingping Gao, Fangzhen Li, Jingke Wang, Ke Guo, Wencong Xiao, Weibo Mao, et al. Fusionad: Multi-modality fusion for prediction and planning tasks of autonomous driving. arXiv preprint arXiv:2308.01006, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 650, + 547, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 650, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 316, + 650, + 547, + 681 + ], + "type": "text", + "content": "[48] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12), 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 682, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 682, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 316, + 682, + 547, + 704 + ], + "type": "text", + "content": "[49] Jiakai Zhang. End-to-End Learning for Autonomous Driving. PhD thesis, New York University, USA, 2019. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 743 + ], + "type": "text", + "content": "26975" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 82, + 298, + 364 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 63, + 82, + 298, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 82, + 298, + 123 + ], + "spans": [ + { + "bbox": [ + 63, + 82, + 298, + 123 + ], + "type": "text", + "content": "[50] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 4, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 125, + 297, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 125, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 63, + 125, + 297, + 178 + ], + "type": "text", + "content": "[51] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 179, + 296, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 179, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 63, + 179, + 296, + 221 + ], + "type": "text", + "content": "[52] Rui Zhao, Qirui Yuan, Jinyu Li, Haofeng Hu, Yun Li, Chengyuan Zheng, and Fei Gao. Sce2drivex: A generalized mllm framework for scene-to-drive learning. arXiv preprint arXiv:2502.14917, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 224, + 295, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 224, + 295, + 265 + ], + "spans": [ + { + "bbox": [ + 63, + 224, + 295, + 265 + ], + "type": "text", + "content": "[53]Zikai Zhao, Qiuxuan Wu, Jian Wang, Botao Zhang, Chaoliang Zhong, and Anton A Zhilenkov. Exploring embodied intelligence in soft robotics: a review. Biomimetics, 9(4): 248, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 267, + 295, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 267, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 63, + 267, + 295, + 309 + ], + "type": "text", + "content": "[54] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 311, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 311, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 63, + 311, + 295, + 364 + ], + "type": "text", + "content": "[55] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied understanding of driving scenarios. In European Conference on Computer Vision, pages 129-148. Springer, 2024. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 735, + 317, + 744 + ], + "type": "text", + "content": "26976" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_content_list.json b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c8fd6394b9435e7439fbc4c095edcb03e8c312c3 --- /dev/null +++ b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_content_list.json @@ -0,0 +1,1460 @@ +[ + { + "type": "text", + "text": "VLRMBench: A Comprehensive and Challenging Benchmark for Vision-Language Reward Models", + "text_level": 1, + "bbox": [ + 168, + 128, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiacheng Ruan $^{1,*}$ Wenzhen Yuan $^{1,*}$ Xian Gao $^{1}$ Ye Guo $^{2}$ Daoxin Zhang $^{2}$ Zhe Xu $^{2}$ Yao Hu $^{2}$ Ting Liu $^{1}$ Yuzhuo Fu $^{1,\\dagger}$ $^{1}$ Shanghai Jiao Tong University \n $^{2}$ Xiaohongshu Inc.", + "bbox": [ + 232, + 202, + 772, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jackchenruan@sjtu.edu.cn", + "bbox": [ + 388, + 277, + 602, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although large visual-language models (LVLMs) have demonstrated strong performance in multimodal tasks, errors may occasionally arise due to biases during the reasoning process. Recently, reward models (RMs) have become increasingly pivotal in the reasoning process. Specifically, process RMs evaluate each reasoning step, outcome RMs focus on the assessment of reasoning results, and critique RMs perform error analysis on the entire reasoning process, followed by corrections. However, existing benchmarks for vision-language RMs (VLRMs) typically assess only a single aspect of their capabilities (e.g., distinguishing between two answers), thus limiting the all-round evaluation and restricting the development of RMs in the visual-language domain. To address this gap, we propose a comprehensive and challenging benchmark, dubbed as VLRMBench, encompassing 12,634 questions. VLRMBench is constructed based on three distinct types of datasets, covering mathematical reasoning, hallucination understanding, and multi-image understanding. We design 12 tasks across three major categories, focusing on evaluating VLRMs in the aspects of process understanding, outcome judgment, and critique generation. Extensive experiments are conducted on 21 open-source models and 5 advanced closed-source models, highlighting the challenges posed by VLRMBench. For instance, in the 'Forecasting Future', a binary classification task, the advanced GPT-4o achieves only a $76.0\\%$ accuracy. The code is available at https://github.com/JCruan519/VLRMBench.", + "bbox": [ + 89, + 357, + 485, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 809, + 220, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, driven by the development of large language models (LLMs) [32, 40, 49, 50], large visual-language models", + "bbox": [ + 89, + 835, + 482, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(LVLMs) have also achieved notable progress [22]. They have been broadly applied across domains such as medical imaging [17], remote sensing [56], and autonomous driving [62]. However, despite these remarkable achievements, they still encounter limitations in complex vision-language (VL) reasoning tasks, primarily due to insufficient reasoning depth and the absence of self-correction mechanisms [10, 51]. To mitigate this issue, reward models (RMs) have been introduced to detect errors in model responses, thereby enhancing the accuracy and efficiency of LVLMs [57].", + "bbox": [ + 511, + 327, + 906, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "During both the training and inference phases, RMs play a pivotal role. Firstly, prior to training, RMs can be employed to filter high-quality samples and facilitate the construction of automated data synthesis pipelines [21, 41, 46]. Secondly, during training, preference optimization techniques (e.g., RLAIF [53]) align the model's preferences with RMs, thereby enhancing the quality of the generated content. Finally, during inference, test-time scaling techniques, aided by RMs, are widely applied [47, 48], trading additional inference time for improved model performance. In the above processes, RMs can be employed not only to evaluate and score the reasoning steps and final answers, but also to generate valuable feedback on the reasoning process. Therefore, it is essential to propose a benchmark for evaluating the various aspects of the performance of RMs. However, as shown in Table 1, existing benchmarks predominantly focus on language domain, and benchmarks for VLRMs typically evaluate only a single capability (e.g., the quality comparison of multiple reasoning processes). These singular and inadequately challenging benchmarks fail to reveal the potential flaws of VLRMs, thereby limiting the development of RMs in the VL domain.", + "bbox": [ + 511, + 484, + 908, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge this gap, we propose VLRMBench, a benchmark that is both comprehensive and challenging, specifically designed for the VL domain. We introduce three distinct data categories: mathematical reasoning, hallucination understanding, and multi-image understanding, en", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Jiacheng Ruan and Wenzhen Yuan are equal contributors.", + "bbox": [ + 107, + 875, + 418, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Yuzhuo Fu is the corresponding author.", + "bbox": [ + 109, + 887, + 323, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3163", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/2e44c9fb179025a95f3a56486c9044da5d8e8d758b211b755c607192e544111e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# TasksStep Eval.?Test Size
MR-GSM8K [54]1✓2,999
RMBench [27]1✗1,327
CriticBench [24]1✗-
MathCheck-GSM [63]1✓516
MR-Ben [55]1✓5,975
ProcessBench [60]1✓3,400
PRMBench [37]9✓6,216
VLRewardBench* [20]1✗1,250
MultimodalRewardBench* [52]1✗5,211
VLRMBENCH*12✓12,634
", + "bbox": [ + 106, + 88, + 470, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison between our VLRMBenchmark and others. * represents the benchmark for vision-language domain. 'Step Eval.' refers to whether the step-level labels are included.", + "bbox": [ + 89, + 229, + 482, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "compassing eight widely used datasets. Moreover, to obtain high-quality samples containing reasoning processes, we develop a collaborative data filtering and generation pipeline. Specifically, we first utilize a small LVLM to filter samples along two dimensions: quality and difficulty. Subsequently, QVQ-72B-preview [39], which excels in VL reasoning, is employed to generate reasoning processes. Next, we use the advanced GPT-4o [2] to partition the reasoning steps and perform an initial validation of correctness. Finally, we apply rule-based filtering and manual inspection to correct errors in the reasoning process, yielding 1,000 samples that serve as the foundation for task construction. Based on these samples, we introduce three themes: step-based, outcome-based, and criticism-based, encompassing 12 diverse tasks and 12,634 questions. These tasks can comprehensively evaluate VLRMs in the aspects of process understanding, outcome judgment, and criticism generation, thus laying the groundwork for the development of VLRMs.", + "bbox": [ + 88, + 305, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conducted extensive experiments on 21 open-source models (from 2B to 90B) and 5 closed-source models, such as GPT-4o and Claude-3.5-Sonnet [3], across 12 tasks. The results underscore the challenges of VLRMBench and demonstrate the narrowing gap between open-source and closed-source models. Specifically, in eight step-based tasks, even the advanced GPT-4o achieved only an average F1-Score of $62.4\\%$ . In two outcome-based binary classification tasks, GPT-4o attained an average accuracy of only $66.3\\%$ . Furthermore, in two criticism-based tasks, Qwen2.5-VL-72B [4] outperformed GPT-4o, with an average win rate of $53.3\\%$ , showcasing the superior performance of state-of-the-art (SOTA) open-source models. In addition, we performed extensive analytical studies, providing valuable insights into the development of VLRMs.", + "bbox": [ + 89, + 580, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions can be summarized as follows: 1) We introduce VLRMBench, the first benchmark that integrates both comprehensiveness and challenge for the evaluation of VLRMs. VLRMBench is constructed from 1,000 high-quality samples curated through our collaborative data filtering and generation pipeline, covering areas such as", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mathematical reasoning, hallucination understanding, and multi-image understanding. 2) We designed 12 tasks, comprising a total of 12,634 carefully crafted questions, capable of evaluating models across three aspects: process understanding, outcome judgment, and critique generation. These diverse tasks enable targeted evaluations and reveal limitations of VLRMs. 3) We conducted extensive experiments on VLRMBench using 21 open-source models and 5 closed-source models, highlighting the challenges presented by VLRMBench and offering comprehensive analysis and guidance for the development of VLRMs.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 273, + 663, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Reward Models (RMs)", + "text_level": 1, + "bbox": [ + 511, + 297, + 723, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Techniques such as Reinforcement Learning with Human Feedback (RLHF) utilize RMs that offer feedback for large models, aligning them with human preferences in terms of accuracy, utility, and safety [5, 34]. RMs can generally be categorized into three types based on their objectives. Process-based RMs [31, 59] evaluate intermediate reasoning steps to ensure logical consistency and the correctness of each step. Outcome-based RMs [6, 25, 28] assess the quality of the final output. In contrast, Critique-based RMs [8, 18, 45] emphasize offering textual feedback or critiques aimed at enhancing the model's performance. However, RMs are not always reliable [37]. Therefore, it is imperative to establish a benchmark that facilitates a comprehensive evaluation of RMs.", + "bbox": [ + 511, + 321, + 906, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Benchmarks of Reward Models", + "text_level": 1, + "bbox": [ + 511, + 544, + 792, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, several benchmarks for evaluating the capabilities of RMs have been proposed, driving the development of RMs in the field of natural language. For example, MR-GSM8k [54] redefines the model's role from a solver to a grader, necessitating the evaluation of the correctness of problem-solving steps. CriticBench [24] incorporates the Generate-Critique-Revise framework to assess the reasoning capabilities of models. ProcessBench [60] emphasizes step-by-step error detection in mathematical derivations, thereby increasing the granularity of the evaluation. PRMBench [37], based on PRM800K [23], employs GPT40 for data optimization and constructs a fine-grained evaluation dataset comprising 6,216 instances. PRMBench covers nine common error types and provides a comprehensive assessment of Process RMs.", + "bbox": [ + 511, + 568, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The aforementioned benchmarks predominantly focus on the language modality and do not adequately address the evaluation needs of VLLMs. To this end, [20] introduces VLRewardBench, a benchmark for VLRMs containing 1,250 test samples. However, VLRewardBench remains limited to Pairwise Comparison, wherein only a judgment of superiority or inferiority is made between two responses,", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3164", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f27dd8ff8b3bc74740d5f66cb7624ec797115139f0f0d344c67c9858bcdc82f4.jpg", + "image_caption": [ + "Figure 1. The collaborative data filtering and generation pipeline of VLRMBench. During the filtering stage, Qwen2VL-7B is tasked with answering questions under two conditions: without and with image input. Questions for which the model provides incorrect answers are retained. Subsequently, QVQ-72B-preview and GPT-4o are employed to generate reasoning processes and construct tasks for three themes." + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 864, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "failing to incorporate more diverse and enriched task types. Consequently, we propose VLRMBench, which encompasses three core themes and twelve distinct tasks, providing a holistic evaluation of VLRMs' overall performance and establishing a foundation for their future advancement.", + "bbox": [ + 88, + 401, + 480, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VLRMBench", + "text_level": 1, + "bbox": [ + 89, + 491, + 228, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we initially present an overview of the pipeline for constructing VLRMBench, along with the data statistics information, followed by an in-depth description of the task design.", + "bbox": [ + 89, + 517, + 483, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall pipeline of VLRMBench", + "text_level": 1, + "bbox": [ + 89, + 588, + 377, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Figure 1, we present a progressive and intuitive pipeline, which is designed to construct VLRMBench in three stages: data collection, filtering, and generation. During the collection phase, we integrated four mathematical, two hallucination, and two multi-image understanding datasets, totaling 16,550 samples and 19,343 images. In the filtering phase, to select high-quality and challenging samples, we introduced a joint filtering mechanism. Specifically, we initially employed Qwen2VL-7B [44] to answer questions without image information. If the model answered correctly, it indicated that the sample was of low quality [7]. Subsequently, we incorporated the image information into the question and instructed the model to answer it. If the model answered correctly, it suggested that the question was too simple to be a challenge. Consequently, we retained only those samples that could not be answered correctly in ei", + "bbox": [ + 89, + 609, + 483, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ther question-answering mode, resulting in a total of 6,715 samples, which formed our basic database $D_{b}$", + "bbox": [ + 511, + 401, + 906, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the generation phase, we utilized QVQ-72B-preview [39] to generate reasoning process for the samples in $D_{b}$ , discarding those where QVQ-72B-preview yielded incorrect answers, as verified via comparison with the standard answers. Subsequently, GPT-4o was employed to perform step-level segmentation of QVQ-72B-preview's reasoning process, grouping semantically similar and logically related sentences into the one step. Following this, we obtained 3,335 samples with reasoning segmented at the step level. Furthermore, based on specific criteria², we excluded samples with excessively long or short reasoning steps and ensured a balanced distribution across the three sample types. Finally, to ensure the quality of the samples, a manual review was conducted. Specifically, three PhD volunteers were recruited to verify the high-quality reasoning processes generated by QVQ-72B-preview. If any volunteer identified an error in the reasoning process, all three volunteers would collaboratively review the sample and correct any potential mistakes. Additionally, examples of manual corrections applied to the samples could be found in the Appendix. After completing the three aforementioned stages, 1,000 high-quality samples were retained for constructing the specific task, denoted as $D$ . For more detailed statistical information, please refer to Table 2 and Figure 2, which provide a description of the sample sources, the average word count, and the number of reasoning steps.", + "bbox": [ + 511, + 440, + 908, + 834 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Qwen2VL-7B ranks in the top $25\\%$ of the OpenCompass leaderboard [11]. With moderate model capabilities and a small model size, it can efficiently filter out low-quality and low-difficulty samples.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We excluded samples where the number of reasoning steps exceeded the average (i.e. 20) or was fewer than 5.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3165", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/b379ac894b7888f7486dc2d24c6565908fa6a90ab38ccb7fb2cf9b7138acb89a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryDataset#Samples#AvgWORDS
MathMathVista [29]142696
MathVision [42]281,021
MathVerse [58]126742
DynaMath [64]251691
HallucinationIllusionVQA [36]62736
HallusionBench [15]334605
Multi-ImageBLINK [14]23790
Mantis-Eval [16]34592
", + "bbox": [ + 93, + 88, + 480, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Statistics on the sample sources of VLRMBench. We also counted the average number of words in the reasoning process under different data sources.", + "bbox": [ + 89, + 238, + 482, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Step-based tasks", + "text_level": 1, + "bbox": [ + 89, + 306, + 254, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The step-based theme consists of 8 tasks designed to assess the reasoning process understanding ability of VLRMs. Under this theme, models will be provided with a modified reasoning process and are expected to output a sequence of 0s and 1s to indicate whether each reasoning step contains deviated content.", + "bbox": [ + 89, + 328, + 483, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step correctness (SC) For a Vision-Language Reward Model (VLRM), detecting errors at each step of the reasoning process is a critical capability [23, 37]. Building on this, we propose the SC task. Specifically, given the reasoning process $R_{n}$ of a sample from $D$ , where $n$ denotes the total number of reasoning steps, we inject errors into $m$ of the $n$ steps $(0 < m < n)$ of $R_{n}$ using the advanced GPT-4o, to simulate potential errors during reasoning.", + "bbox": [ + 88, + 438, + 482, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Redundancy Detection (RD) Existing reasoning models often produce redundant information when answering questions, leading to an increase in output token length, which is both inefficient and environmentally unfriendly. To this end, we propose the RD task, which requires VLRMs to identify redundant information generated during the reasoning process. Specifically, based on $D$ , we introduce redundant descriptions in certain reasoning steps of each sample. For instance, we incorporate descriptions from $R_{i-j}$ into $R_i$ to intentionally create redundancy. Furthermore, we employ GPT-4 to generate paraphrased versions of the content from $R_{i-j}$ and incorporate them into $R_i$ , thereby amplifying redundancy. Unlike the SC task, the introduction of redundant information does not compromise the correctness of the final answer.", + "bbox": [ + 89, + 579, + 482, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Confidence Misdirection (CM) When confident and assertive terms, such as 'definitely' and 'without a doubt', are present in an incorrect reasoning step, is it possible for VL-RMs to be misled into perceiving the erroneous reasoning as correct? To investigate this issue, we propose the CM task,", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ebfe9977b3db29d7faba65e6419e5c8b59427bdfdc4136cbc90d453078734927.jpg", + "image_caption": [ + "Figure 2. The distribution of the number of reasoning steps for different data sources." + ], + "image_footnote": [], + "bbox": [ + 547, + 93, + 870, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is derived from the SC task. Specifically, for each sample in the SC task, we incorporate confidence-boosting terms into the erroneous reasoning step to mislead VLRMs into perceiving the step as correct, thereby evaluating the robustness of VLRMs.", + "bbox": [ + 511, + 329, + 906, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existential Hallucination (EH) and Attribute Hallucination (AH) When answering questions using LVLMs, hallucination phenomena often arise [26]. Although other parts of the reasoning process may be correct, the introduction of hallucinatory errors or the omission of key entities can lead to biased final results. To evaluate whether VLRMs are capable of detecting hallucinations introduced during the reasoning process, we have designed two tasks on the top of $D$ . The first task, EH, assesses the ability of the VLRM to identify entities that were incorrectly introduced during reasoning process but do not exist in the image, or to detect elements that may have been overlooked, such as objects, people, etc. The second task, AH, examines the VLRM's capability to accurately identify the attributes of entities, particularly in cases where LVLMs wrongly attribute properties to an entity that it does not possess or when attribute descriptions conflict with the visual content of the image. Such errors are typically reflected in misjudgments regarding object attributes, such as size, color, or shape. VLRMs must integrate image information effectively to accurately identify hallucination errors during the reasoning process.", + "bbox": [ + 511, + 426, + 908, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detail Error (DE) When addressing complex problems, the reasoning process often involves multiple details, particularly in the handling of numbers and symbols. Even if the overall reasoning appears to be sound, there may still be hidden errors in finer details (e.g., numerical errors in calculations or symbolic mistakes in formulas). These errors are often difficult to detect, yet they can affect the precision of the final results. The introduction of DE tasks serves to evaluate the sensitivity of VLRMs to reasoning details, as", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3166", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "well as their ability to identify fine-grained errors.", + "bbox": [ + 91, + 90, + 421, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Relationship (SR) During visual reasoning, LVLMs may incorrectly describe the spatial relationships between objects or entities, leading to erroneous results. For instance, LVLMs might misdescribe an object located at the top of the image as being at the bottom. Therefore, we select samples containing spatial relationship descriptions from $D$ to construct the SR task, aiming to evaluate VLRMs' understanding of spatial relationships in a image and their ability to detect spatial errors in a reasoning process.", + "bbox": [ + 89, + 125, + 483, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Confusion (IC) In multi-image tasks, LVLMs are required to extract and associate relevant information from multiple images, which typically involves identifying various entities and their interrelations across different images. However, due to the presence of similar elements or backgrounds between images, LVLMs may erroneously associate an entity from one image with another image. This 'image reference error' not only causes confusion but also potentially impacts the accuracy of subsequent reasoning. Thus, we filter samples related to multi-image understanding from $D$ to construct the IC task, which aims to evaluate whether VLRMs can identify image reference errors during the reasoning process.", + "bbox": [ + 89, + 277, + 483, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Outcome-based tasks", + "text_level": 1, + "bbox": [ + 89, + 481, + 290, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The outcome-based theme encompasses two binary classification tasks, aiming to evaluate the ability of VLRMs to assess and predict the accuracy of outcomes.", + "bbox": [ + 89, + 503, + 483, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-solution Judgment (MJ) During the training stage, preference optimization techniques, such as DPO [35], have become a widely used paradigm. These methods typically rely on the collection of preference samples. VLRMs are frequently employed to assess the quality of preference samples, thereby becoming a crucial component of preference optimization [21, 53]. To this end, we introduce the MJ task, which aims to investigate the discriminative ability of VLRMs regarding different reasoning processes for the same problem. This task is discussed as a core element in VLRewardBench [20] and Multimodal Reward-Bench [52]. Specifically, for the reasoning process $R$ of any question $Q$ in $D$ , we introduce the reasoning process $R'$ of the same question in the SC, RD and CM tasks, thereby constructing a triplet $(Q, R, R')$ . VLRMs are required to judge the quality of $R$ and $R'$ and score them accordingly.", + "bbox": [ + 89, + 566, + 482, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Forecasting Future (FF) When LVLMs perform inference, test-time scaling (TTS) can further improve their performance [47, 51]. However, these techniques often introduce additional inference time, primarily due to the necessity of exploring multiple reasoning paths. If the correctness", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the final answer could be predicted based on the reasoning process of the first m steps, $R_{[0:m]}$ , this prediction could substantially expedite the TTS process. Therefore, the FF task is proposed, wherein VLRMs are required to assess the correctness of the final answer based on $R_{[0:m]}$ .", + "bbox": [ + 511, + 90, + 905, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Criticism-based tasks", + "text_level": 1, + "bbox": [ + 511, + 176, + 718, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Error Reason Analysis (ERA) and Error Correction (EC) Unlike Process RM and Outcome RM, which focus solely on evaluating processes or outcomes, Critique RM emphasizes the critique of the reasoning process, providing textual feedback that includes error analysis and correction. Consequently, we propose the ERA and EC tasks to assess the capabilities of VLRMs in these two aspects. In the ERA task, VLRMs are presented with erroneous reasoning steps derived from the SC task, and are tasked with identifying the error and analyzing its underlying cause. In contrast, the EC task does not require error cause analysis; rather, VLRMs directly correct the identified error and produce the corrected reasoning process.", + "bbox": [ + 511, + 199, + 906, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 411, + 645, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Models", + "text_level": 1, + "bbox": [ + 511, + 436, + 607, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted extensive experiments on 21 open-source models and 5 closed-source models. Specifically, the open-source models, which range from 2B to 90B, including Llama3.2 (11B/90B) [12], Llava-OneVision (7B/72B) [19], Qwen2VL (2B/7B/72B) [44], Qwen2.5VL (3B/7B/72B) [4], QVQ-72B-preview [39], Ovis2 (8B/16B/34B) [30], InternVL2.5 (2B/8B/38B/78B) [9], Phi-3.5-Vision (4.2B) [1], MiniCPM-o-2.6 (8B), and MiniCPM-V-2.6 (8B) [33]. For the closed-source models, we employed GLM-4V-Flash, GPT-4o [2], Gemini-2.0-Flash-/Flash-thinking-exp-1219 [38], and Claude-3.5-Sonnet [3].", + "bbox": [ + 511, + 459, + 906, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation details", + "text_level": 1, + "bbox": [ + 511, + 636, + 689, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow the 'LVLM-as-a-Judge' paradigm and introduce different question-answering templates to assess the models [20, 52]. Specifically, for the SC, RD, CM, EH, AH, DE, SR, and IC tasks, the model takes a reasoning process as input and outputs a judgment regarding whether each reasoning step contains redundancies or other errors. For these eight tasks, we employ the weighted F1-Score as the evaluation metric, as defined in Equation 1, which better addresses class imbalance issues and ensures the model's performance on the minority class is adequately considered. For the MJ task, the model is provided with (good and bad) reasoning process pairs for the same question and outputs scores for both. If the model assigns a higher score to the good sample, it is deemed correct; otherwise, it is considered incorrect. For the FF task, the model takes several preceding steps of the reasoning process as input and outputs a judgment on", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3167", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4b7eecf51e7571df2020a283d6abd5c0d2b9716ab9427d2b6a456b4f4419788a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelStep (5,722)Outcome (5,402)Criticism (1,510)
SCRDCMEHAHDESRICW-Avg.MJFFW-Avg.ERAwERAτERALECwECTECL
# Sample99899799494694465912658-2,9862,416-974536
Open-source LVLMs < 10B (Small group)
InternVL2.5-2B33.834.130.022.423.833.519.930.129.226.660.841.94.86.788.52.310.087.6
Qwen2VL-2B28.326.220.917.819.126.623.826.423.010.559.632.50.20.199.72.55.891.7
Qwen2.5VL-3B29.932.723.227.226.945.525.142.630.131.967.147.67.112.180.84.821.273.9
Phi-3.5-Vision (4.2B)35.938.827.920.721.133.417.731.629.212.165.235.83.89.087.24.222.473.4
Qwen2VL-7B38.029.332.021.521.432.821.621.128.852.064.757.73.98.187.95.423.271.4
Qwen2.5VL-7B43.433.237.822.823.945.515.629.133.426.070.746.037.722.040.39.351.938.8
Llava-OneVision-7B31.825.721.516.017.224.816.120.722.618.057.835.83.910.485.77.134.458.5
InternVL2.5-8B36.628.431.121.921.236.515.124.128.661.473.866.934.521.943.66.236.157.7
Ovis2-8B47.728.237.825.225.739.817.025.833.331.465.346.623.225.950.914.363.522.2
MiniCPM-V-2.6 (8B)44.938.336.926.126.536.824.230.634.637.962.548.95.411.782.93.424.771.9
MiniCPM-o-2.6 (8B)41.035.337.329.629.036.728.330.134.615.064.337.09.219.870.97.633.958.4
Avg.37.431.830.622.823.335.620.428.429.829.364.745.212.213.474.46.129.764.1
10B < Open-source LVLMs < 40B (Middle group)
Llama3.2-11B33.927.129.720.220.031.917.125.826.723.961.140.54.210.385.55.434.759.8
Ovis2-16B48.547.947.940.635.844.521.928.443.649.769.258.430.926.642.510.469.320.3
Ovis2-34B65.351.164.554.551.659.630.449.057.046.375.759.445.926.627.515.476.48.1
InternVL2.5-38B61.049.662.655.051.761.434.648.156.145.179.360.421.124.554.412.556.630.9
Avg.52.243.951.242.639.849.426.037.845.941.371.354.725.522.052.510.959.329.8
Open-source LVLMs > 40B (Large group)
Llava-OneVision-72B49.430.650.237.934.145.113.926.540.432.269.749.015.322.162.69.451.039.6
Qwen2VL-72B55.540.256.148.941.656.422.443.648.837.772.153.122.420.457.17.757.135.1
Qwen2.5VL-72B72.841.770.464.659.972.437.963.362.665.680.272.174.115.110.815.677.07.3
QVQ-72B-preview42.435.336.329.427.137.820.826.134.349.238.444.44.210.085.89.451.039.6
InternVL2.5-78B62.142.062.853.348.760.428.547.254.048.678.862.154.224.920.910.863.925.3
Llama3.2-90B40.838.841.334.229.435.417.021.136.252.769.060.09.918.671.49.556.833.7
Avg.53.838.152.944.740.151.323.438.046.147.768.056.830.018.551.410.459.530.1
Closed-source LVLMs
Claude-3.5-Sonnet70.853.765.763.962.863.440.865.762.982.275.179.060.625.513.921.253.924.9
GPT-4o73.750.666.657.658.671.848.264.662.458.476.066.30.0100.00.00.0100.00.0
GLM-4V-Flash37.233.532.928.227.741.822.920.232.837.266.750.434.224.341.56.954.139.0
Gemini-2.0-Flash69.355.966.165.662.363.446.867.863.454.277.164.432.822.544.77.152.540.3
Gemini-2.0-FTE.-121967.955.565.765.962.363.246.467.163.153.177.063.837.422.640.06.850.243.1
Avg.63.849.859.456.254.760.741.057.157.057.074.464.833.039.028.08.462.129.5
", + "bbox": [ + 96, + 89, + 901, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Results on the VLRMBench. For Step-based tasks, the F1-Score is the metric. For Outcome-based tasks, Accuracy is utilized as the metric. In the case of Criticism-based tasks, the metric is Win Rate, expressed as Win: Tie: Lose (Simplified as $W$ , $T$ and $L$ ). Note that W-Avg. represents the weighted average. Evaluation prompt templates and some case studies could be found in the Appendix B and C.", + "bbox": [ + 89, + 530, + 906, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "whether a correct answer can ultimately be derived. Consequently, for the MJ and FF tasks, Accuracy is employed as the evaluation metric. For the ERA and EC tasks, the model's output is a description of the error reasons in the reasoning process and the corrected reasoning process. We utilize the output of GPT-4o as the baseline and adopt Win Rate [13, 61] as the evaluation metric.", + "bbox": [ + 89, + 599, + 482, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We configured the max new tokens to 4,096 across all tasks for the five closed-source models and QVQ-72B-preview. In contrast, for the other models, we set the max new tokens to 4,096 specifically for the ERA and EC tasks, while fixing it at 256 for the remaining tasks. To ensure the reproducibility of the experimental results, we set the temperature coefficient to 0 for all models. All experiments were conducted on 8 NVIDIA H800 GPUs.", + "bbox": [ + 89, + 705, + 482, + 825 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF 1 = \\frac {N _ {\\text {n e g}}}{N _ {\\text {p o s}} + N _ {\\text {n e g}}} \\cdot F 1 _ {\\text {p o s}} + \\frac {N _ {\\text {p o s}}}{N _ {\\text {p o s}} + N _ {\\text {n e g}}} \\cdot F 1 _ {\\text {n e g}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 845, + 482, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N_{\\mathrm{pos}}$ and $N_{\\mathrm{neg}}$ represent the number of correct and", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "incorrect steps in the reasoning process, respectively.", + "bbox": [ + 511, + 599, + 862, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 621, + 653, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental results on VLRMBench are shown in Table 3, leading to the following conclusions:", + "bbox": [ + 511, + 643, + 903, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) Among open-source models, surprisingly, the performance of Qwen2.5VL-72B is comparable to that of advanced closed-source models. For instance, across the eight Step-based tasks, Qwen2.5VL-72B achieved an average F1-Score of $62.6\\%$ , surpassing GPT-4o by $0.2\\%$ and only slightly trailing Claude-3.5-Sonnet by $0.3\\%$ . In the MJ and FF tasks, Qwen2.5VL-72B also outperformed GPT-4o by $7.2\\%$ and $4.2\\%$ , respectively. Notably, in the ERA task, Qwen2.5VL-72B attained a win rate of $74.1\\%$ over GPT-4o, significantly outperforming Claude-3.5-Sonnet at $60.6\\%$ .", + "bbox": [ + 511, + 674, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2) LVLMs that have not undergone specific training show diminishing performance gains as the model size increases. Specifically, for LVLMs in the Middle group, compared to those in the Small group, the performance gain in Step-based and Outcome-based tasks was $54.0\\%$ and", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3168", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2aa3132319e4dd6a4913726c20034c383816d57604552ab065c85482b118cbb0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OrderModelAvg. ScoreAcc
Better firstQwen2.5VL-3B[9.0, 8.3]23.7
Llama3.2-11B[8.6, 4.8]52.8
InternVL2.5-78B[8.4, 7.4]38.6
GPT-4o[9.8, 7.8]52.2
Better secondQwen2.5VL-3B[8.7, 9.0]21.4
Llama3.2-11B[8.3, 5.6]8.4
InternVL2.5-78B[8.1, 9.1]45.6
GPT-4o[7.4, 8.1]43.1
", + "bbox": [ + 109, + 88, + 467, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "21.0%, respectively. In contrast, the corresponding performance gains in the Large group relative to the Middle group were only 0.4% and 3.8%. Even excluding QVQ-72B-preview from the Large group, the relative performance improvement remains limited, at 5.4% and 8.4%. This suggests that blindly increasing the size of VLLMs does not lead to significant performance improvements. Consequently, we recommend that future research prioritize training VLRMs at the Middle group scale initially.", + "bbox": [ + 89, + 301, + 480, + 438 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3) The Outcome-based tasks, being simple binary classification problems, remain challenging for most open-source models, and even closed-source models perform only marginally better than random guessing. For open-source models across three size groups, the average accuracy for Outcome-based tasks is around $50\\%$ , similar to random guessing. For advanced closed-source models, the average accuracy only exceeds random guessing by $14.8\\%$ .", + "4) Deterministic vocabulary and hallucinated information can effectively mislead the VLRMs, reducing its ability to understand reasoning processes. The CM task, based on the SC task, introduces deterministic vocabulary (e.g., must, definitely) into sentences with erroneous reasoning steps, while EH and AH tasks introduce hallucinated information during the reasoning process. Existing models, particularly in the Small group, demonstrate a degradation in F1-Score of $6.8\\%$ , $14.6\\%$ , and $14.1\\%$ for CM, EH, and AH tasks, respectively, when compared to the SC task. This performance degradation remains significant even as model size increases.", + "5) Existing models show strong robustness to numerical errors but limited resilience to perturbations in spatial relationships and image references. We observed that for Large group, the F1-Score in the DE task only decreased by $2.5\\%$ compared to SC. However, for the SR and IC tasks, the performance dropped by $30.4\\%$ and $15.8\\%$ , respectively. Overall, our VLRMBench provide unique challenges that surpass traditional benchmarks, with even the most advanced models achieving only mediocre performance." + ], + "bbox": [ + 89, + 446, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91ae9d72b132b5cebc4b7f4e16aeceeb5b11fa47f6084b2c28b7069bd88ef514.jpg", + "table_caption": [ + "Table 4. The results on the MJ task with different sample order." + ], + "table_footnote": [], + "table_body": "
Model≤10 steps10-15 steps>15 steps
Qwen2.5VL-3B35.630.125.6
Llama3.2-11B42.932.828.4
InternVL2.5-78B68.264.154.1
GPT-4o77.875.268.6
", + "bbox": [ + 526, + 88, + 893, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Results on the SC task with different reasoning steps.", + "bbox": [ + 524, + 183, + 892, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 224, + 679, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Is the o1-like model suitable to be a VLRM?", + "text_level": 1, + "bbox": [ + 511, + 250, + 885, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "QVQ-72B-preview [39], an open-source visual-language model, exhibits strong reasoning capabilities. However, it exhibits a potential drawback regarding the degradation of instruction-following ability. For instance, in the SC task, by incorporating instruction prompts, the sample's original question, and the reasoning process as input, we expect the VLRM to evaluate each reasoning step. However, QVQ-72B-preview often disregards the given instructions and directly answers the original question, leading to suboptimal performance on VLRMBench, with additional cases provided in the Appendix C. Therefore, as an o1-like model, it must not only demonstrate strong reasoning abilities but also exhibit robust instruction-following capabilities to be considered a viable candidate for a VLRM.", + "bbox": [ + 511, + 273, + 905, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.Impact of sample order", + "text_level": 1, + "bbox": [ + 511, + 496, + 730, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the MJ task, VLRM scores two reasoning processes for the same question. To further investigate the positional bias, we employed two different sample orders, as shown in Table 4. Specifically, 'Better first' places the superior reasoning process in the first position, while 'Better second' places it in the second position. The results indicate that Llama3.2-11B consistently assigns a higher score to the reasoning process in the first position, irrespective of its quality. In contrast, the other three models exhibit the ability to adjust their scores based on the quality of the reasoning. This positional bias phenomenon is also observed in [43]. As a result, in Table 3, we utilized the sum of the two scores for the MJ task evaluation to ensure fairness.", + "bbox": [ + 511, + 518, + 905, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Performance on various reasoning length", + "text_level": 1, + "bbox": [ + 511, + 726, + 864, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further investigate the impact of the number of reasoning steps on the VLRMs' process understanding. As shown in Table 5, we classified the number of reasoning steps in the SC task into three categories: the short group ( $\\leq 10$ steps), with 263 samples; the middle group (10-15 steps), with 468 samples; and the long group ( $>15$ steps), with 267 samples. Specifically, for all four models, a decrease in F1-score was observed as the number of reasoning steps increased. Notably, for the Llama3.2-11B model, the difference between the short and long groups reached $14.5\\%$ , and", + "bbox": [ + 511, + 750, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3169", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e28cc05a43541e9df380fc700c1698baabf40ce48863f28604c15854880735fe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model0-shot1-shot2-shot
Qwen2.5VL-3B29.931.531.1
Llama3.2-11B33.942.242.3
InternVL2.5-78B62.155.955.7
GPT-4o73.771.770.5
", + "bbox": [ + 125, + 88, + 447, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/739e169c1130866e94dbc8225caf0d71bb5a71df2711e89764cecc05ce3991a8.jpg", + "image_caption": [ + "Figure 3. Performance changes with the test-time scaling." + ], + "image_footnote": [], + "bbox": [ + 112, + 215, + 460, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "even for the advanced GPT-4o, the F1-Score dropped from $77.8\\%$ to $68.6\\%$ . This suggests that current models are less capable of handling long information, underscoring the necessity of enhancing their ability to tackle long reasoning process when training VLRMs in the future.", + "bbox": [ + 89, + 395, + 482, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Impacts of In-Context Learning (ICL) setting", + "text_level": 1, + "bbox": [ + 89, + 482, + 475, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In main experiments, we employ zero-shot setting by default. Furthermore, using the SC task as an example, we explore the effect of ICL on VLRMs. As presented in Table 6, the results indicate that providing contextual examples is more beneficial for small models but harms the performance of large models. Specifically, with the help of one-shot learning, Qwen2.5VL-3B's performance improved from $29.9\\%$ to $31.5\\%$ , and Llama3.2-11B saw an improvement of $8.3\\%$ . However, for GPT-4o, one-shot learning led to a $2.0\\%$ performance degradation, and InternVL2.5-78B's performance decreased by $6.2\\%$ . One possible explanation for this is that small models exhibit relatively weaker instruction-following capabilities, and ICL examples may help guide them in responding to queries in the designated format. In contrast, large models, which already possess strong comprehension abilities and adaptability to complex tasks, may find that providing contextual examples interferes with their inherent understanding process, resulting in performance decline. Thus, in the SC task, smaller models exhibit a greater reliance on ICL examples, whereas larger models tend to rely more on their pre-trained knowledge.", + "bbox": [ + 89, + 505, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Impacts of test-time scaling", + "text_level": 1, + "bbox": [ + 89, + 832, + 338, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigated the impact of test-time scaling (i.e. voting method) for VLRMs on the MJ task. To ensure the diversity of the generated results, we set the temperature coefficient", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/997a2f1a0b15ee8261d9068c5d9b8947de4206fe44536c7441fadf0d7e7c1ab0.jpg", + "table_caption": [ + "Table 6. Results on SC task with various ICL few-shot numbers." + ], + "table_footnote": [], + "table_body": "
LVLMVLRMMathVistaMathVerseMathVisionAvg.
7B7B10.45.84.77.0
72B7B11.84.23.26.4
7B72B21.819.721.521.0
72B72B11.79.311.410.8
", + "bbox": [ + 535, + 89, + 883, + 164 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Results of the correction rate on math benchmarks.", + "bbox": [ + 531, + 174, + 885, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to 0.9 and performed K sampling. As illustrated in Figure 3, surprisingly, as the value of K increases, the accuracy of Qwen2.5VL-3B improved by $12.4\\%$ . Notably, at $\\mathrm{K} = 9$ , its performance was only $4.8\\%$ lower than that of GPT-4o. These findings highlight the substantial effectiveness of test-time scaling techniques in enhancing VLRMs.", + "bbox": [ + 511, + 217, + 906, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Improve LVLMs with a VLRM", + "text_level": 1, + "bbox": [ + 511, + 320, + 790, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the significant role of VLRMs in the reasoning process, as presented in Table 7, we conducted experiments on three mathematical benchmarks using Qwen2.5VL-7B/72B. For example, in the second row of Table 7, Qwen2.5VL-72B was employed as the LVLM to answer the mathematical questions, whereas Qwen2.5VL-7B was used as the VLRM to perform error reason analysis of the LVLM's reasoning process. The analysis was subsequently fed back into the LVLM as feedback, allowing it to generate a second round of answers. Besides, we defined the correction rate, which is the ratio of the number of questions correctly answered in the second round after feedback to the number of questions incorrectly answered in the first round. The experimental results demonstrate that VLRM can provide valuable information to the LVLM through error analysis, thereby facilitating mistake correction, with an average gain of $11.3\\%$ . These findings suggest that even LVLMs not trained on proprietary data can act as VLRMs, providing more correction possibilities during testing based on this intuitive feedback mechanism. This further underscores the potential value of developing efficient VLRMs.", + "bbox": [ + 511, + 343, + 906, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 678, + 638, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a comprehensive and challenging benchmark, named VLRMBench. This benchmark covers 3 themes and 12 tasks, comprising a total of 12,634 questions. Following the 'LVLM-as-a-Judge' paradigm, we conduct systematic experiments and analyses on VLRMBench using 26 models, deriving several key insights: enhancing the performance of VLRMs in long-text comprehension, exploring various feedback patterns from VLRMs to LVLMs, and developing the application potential of proprietary VLRMs, etc. We hope that VLRMBench will serve as a standardized evaluation platform to propel the development of more efficient VLRMs. In future work, we plan to incorporate larger datasets specifically designed to train VLRMs.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3170", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grant No. 61977045).", + "bbox": [ + 89, + 114, + 482, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 187, + 174 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5", + "[3] Anthropic. Claude 3.5: Sonnet, 2025. Accessed: 2025-02-25. 2, 5", + "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 5", + "[5] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2", + "[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2", + "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3", + "[8] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2", + "[9] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5", + "[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 1", + "[11] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 3", + "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + ], + "bbox": [ + 93, + 183, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5", + "[13] Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B Hashimoto. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475, 2024. 6", + "[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision, pages 148-166. Springer, 2024. 4", + "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 4", + "[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. Transactions on Machine Learning Research, 2024, 2024. 4", + "[17] Wasif Khan, Seowung Leem, Kyle B See, Joshua K Wong, Shaoting Zhang, and Ruogu Fang. A comprehensive survey of foundation models in medicine. IEEE Reviews in Biomedical Engineering, 2025. 1", + "[18] Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, et al. Solar 10.7 b: Scaling large language models with simple yet effective depth up-scaling. arXiv preprint arXiv:2312.15166, 2023. 2", + "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5", + "[20] Lei Li, Yuancheng Wei, Zhihui Xie, Xuqing Yang, Yifan Song, Peiyi Wang, Chenxin An, Tianyu Liu, Sujian Li, Bill Yuchen Lin, et al. Vlrewardbench: A challenging benchmark for vision-language generative reward models. arXiv preprint arXiv:2411.17451, 2024. 2, 5", + "[21] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. arXiv preprint arXiv:2410.09421, 2024. 1, 5", + "[22] Zongxia Li, Xiyang Wu, Hongyang Du, Huy Nghiem, and Guangyao Shi. Benchmark evaluations, applications, and challenges of large vision language models: A survey. arXiv preprint arXiv:2501.02189, 2025. 1", + "[23] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2023. 2, 4", + "[24] Zicheng Lin, Zhibin Gou, Tian Liang, Ruilin Luo, Haowei Liu, and Yujiu Yang. Criticbench: Benchmarking lms for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3171", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "critique-correct reasoning. arXiv preprint arXiv:2402.14809, 2024. 2", + "[25] Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451, 2024. 2", + "[26] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xi-tian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 4", + "[27] Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. Rm-bench: Benchmarking reward models of language models with subtlety and style. arXiv preprint arXiv:2410.16184, 2024. 2", + "[28] Xingzhou Lou, Dong Yan, Wei Shen, Yuzi Yan, Jian Xie, and Junge Zhang. Uncertainty-aware reward model: Teaching reward models to know what is unknown. arXiv preprint arXiv:2410.00847, 2024. 2", + "[29] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 4", + "[30] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 5", + "[31] Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Yang You, and Hongxia Yang. Let's reward step by step: Step-level reward model as the navigators for reasoning. arXiv preprint arXiv:2310.10080, 2023. 2", + "[32] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 1", + "[33] OpenBMB. Minicpm o: A gpt-4o-level mllm for vision, speech, and multimodal live streaming on your phone, 2025. Accessed: 2025-02-25. 5", + "[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2", + "[35] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 5", + "[36] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 4", + "[37] Mingyang Song, Zhaochen Su, Xiaoye Qu, Jiawei Zhou, and Yu Cheng. Prmbench: A fine-grained and challenging benchmark for process-level reward models. arXiv preprint arXiv:2501.03124, 2025. 2, 4", + "[38] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk," + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5", + "[39] Qwen Team. Qvq: To see the world with wisdom, 2024. 2, 3, 5, 7", + "[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1", + "[41] Jiahao Wang, Bolin Zhang, Qianlong Du, Jiajun Zhang, and Dianhui Chu. A survey on data selection for llm instruction tuning. arXiv preprint arXiv:2402.05123, 2024. 1", + "[42] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 4", + "[43] Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926, 2023. 7", + "[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5", + "[45] Tianlu Wang, Ilia Kulikov, Olga Golovneva, Ping Yu, Weizhe Yuan, Jane Dwivedi-Yu, Richard Yuanzhe Pang, Maryam Fazel-Zarandi, Jason Weston, and Xian Li. Self-taught evaluators. arXiv preprint arXiv:2408.02666, 2024. 2", + "[46] Zhilin Wang, Alexander Bukharin, Olivier Delalleau, Daniel Egert, Gerald Shen, Jiaqi Zeng, Oleksii Kuchaiev, and Yi Dong. Helpsteer2-preference: Complementing ratings with preferences. arXiv preprint arXiv:2410.01257, 2024. 1", + "[47] Jinyang Wu, Mingkuan Feng, Shuai Zhang, Ruihan Jin, Feihu Che, Zengqi Wen, and Jianhua Tao. Boosting multimodal reasoning with mcts-automated structured thinking. arXiv preprint arXiv:2502.02339, 2025. 1, 5", + "[48] Wang Xiyao, Yang Zhengyuan, Li Linjie, Lu Hongjin, Xu Yuancheng, Lin Chung-Ching Lin, Lin Kevin, Huang Furong, and Wang Lijuan. Scaling inference-time search with vision value model for improved visual comprehension. arXiv preprint arXiv:2412.03704, 2024. 1", + "[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1", + "[50] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 1", + "[51] Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3172", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "carlo tree search. arXiv preprint arXiv:2412.18319, 2024. 1,5", + "[52] Michihiro Yasunaga, Luke Zettlemoyer, and Marjan Ghazvininejad. Multimodal rewardbench: Holistic evaluation of reward models for vision language models. arXiv preprint arXiv:2502.14191, 2025. 2, 5", + "[53] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 1, 5", + "[54] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 2", + "[55] Zhongshen Zeng, Yinhong Liu, Yingjia Wan, Jingyao Li, Pengguang Chen, Jianbo Dai, Yuxuan Yao, Rongwu Xu, Zehan Qi, Wanru Zhao, et al. Mr-ben: A meta-reasoning benchmark for evaluating system-2 thinking in llms. arXiv preprint arXiv:2406.13975, 2024. 2", + "[56] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Photogrammetry and Remote Sensing, 221:64-77, 2025. 1", + "[57] Di Zhang, Jingdi Lei, Junxian Li, Xunzhi Wang, Yujie Liu, Zonglin Yang, Jiatong Li, Weida Wang, Suorong Yang, Jianbo Wu, et al. Critic-v: Vlm critics help catch vlm errors in multimodal reasoning. arXiv preprint arXiv:2411.18203, 2024. 1", + "[58] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Springer, 2024. 4", + "[59] Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301, 2025. 2", + "[60] Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559, 2024. 2", + "[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6", + "[62] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 1", + "[63] Zihao Zhou, Shudong Liu, Maizhen Ning, Wei Liu, Jindong Wang, Derek F Wong, Xiaowei Huang, Qiufeng Wang, and Kaizhu Huang. Is your model really a good math rea" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "soner? evaluating mathematical reasoning with checklist. arXiv preprint arXiv:2407.08733, 2024. 2", + "[64] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 4" + ], + "bbox": [ + 516, + 92, + 903, + 188 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3173", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_model.json b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24c00d10bbec6f1cb7500502c73de62819d3a7dd --- /dev/null +++ b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.169, + 0.13, + 0.83, + 0.177 + ], + "angle": 0, + "content": "VLRMBench: A Comprehensive and Challenging Benchmark for Vision-Language Reward Models" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.203, + 0.773, + 0.275 + ], + "angle": 0, + "content": "Jiacheng Ruan\\(^{1,*}\\) Wenzhen Yuan\\(^{1,*}\\) Xian Gao\\(^{1}\\) Ye Guo\\(^{2}\\) Daoxin Zhang\\(^{2}\\) Zhe Xu\\(^{2}\\) Yao Hu\\(^{2}\\) Ting Liu\\(^{1}\\) Yuzhuo Fu\\(^{1,\\dagger}\\) \n\\(^{1}\\)Shanghai Jiao Tong University \n\\(^{2}\\)Xiaohongshu Inc." + }, + { + "type": "text", + "bbox": [ + 0.39, + 0.278, + 0.604, + 0.292 + ], + "angle": 0, + "content": "jackchenruan@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.783 + ], + "angle": 0, + "content": "Although large visual-language models (LVLMs) have demonstrated strong performance in multimodal tasks, errors may occasionally arise due to biases during the reasoning process. Recently, reward models (RMs) have become increasingly pivotal in the reasoning process. Specifically, process RMs evaluate each reasoning step, outcome RMs focus on the assessment of reasoning results, and critique RMs perform error analysis on the entire reasoning process, followed by corrections. However, existing benchmarks for vision-language RMs (VLRMs) typically assess only a single aspect of their capabilities (e.g., distinguishing between two answers), thus limiting the all-round evaluation and restricting the development of RMs in the visual-language domain. To address this gap, we propose a comprehensive and challenging benchmark, dubbed as VLRMBench, encompassing 12,634 questions. VLRMBench is constructed based on three distinct types of datasets, covering mathematical reasoning, hallucination understanding, and multi-image understanding. We design 12 tasks across three major categories, focusing on evaluating VLRMs in the aspects of process understanding, outcome judgment, and critique generation. Extensive experiments are conducted on 21 open-source models and 5 advanced closed-source models, highlighting the challenges posed by VLRMBench. For instance, in the 'Forecasting Future', a binary classification task, the advanced GPT-4o achieves only a \\(76.0\\%\\) accuracy. The code is available at https://github.com/JCruan519/VLRMBench." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.81, + 0.222, + 0.826 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.836, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Recently, driven by the development of large language models (LLMs) [32, 40, 49, 50], large visual-language models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.48 + ], + "angle": 0, + "content": "(LVLMs) have also achieved notable progress [22]. They have been broadly applied across domains such as medical imaging [17], remote sensing [56], and autonomous driving [62]. However, despite these remarkable achievements, they still encounter limitations in complex vision-language (VL) reasoning tasks, primarily due to insufficient reasoning depth and the absence of self-correction mechanisms [10, 51]. To mitigate this issue, reward models (RMs) have been introduced to detect errors in model responses, thereby enhancing the accuracy and efficiency of LVLMs [57]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.82 + ], + "angle": 0, + "content": "During both the training and inference phases, RMs play a pivotal role. Firstly, prior to training, RMs can be employed to filter high-quality samples and facilitate the construction of automated data synthesis pipelines [21, 41, 46]. Secondly, during training, preference optimization techniques (e.g., RLAIF [53]) align the model's preferences with RMs, thereby enhancing the quality of the generated content. Finally, during inference, test-time scaling techniques, aided by RMs, are widely applied [47, 48], trading additional inference time for improved model performance. In the above processes, RMs can be employed not only to evaluate and score the reasoning steps and final answers, but also to generate valuable feedback on the reasoning process. Therefore, it is essential to propose a benchmark for evaluating the various aspects of the performance of RMs. However, as shown in Table 1, existing benchmarks predominantly focus on language domain, and benchmarks for VLRMs typically evaluate only a single capability (e.g., the quality comparison of multiple reasoning processes). These singular and inadequately challenging benchmarks fail to reveal the potential flaws of VLRMs, thereby limiting the development of RMs in the VL domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To bridge this gap, we propose VLRMBench, a benchmark that is both comprehensive and challenging, specifically designed for the VL domain. We introduce three distinct data categories: mathematical reasoning, hallucination understanding, and multi-image understanding, en" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.419, + 0.888 + ], + "angle": 0, + "content": "* Jiacheng Ruan and Wenzhen Yuan are equal contributors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.325, + 0.901 + ], + "angle": 0, + "content": "Yuzhuo Fu is the corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.419, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3163" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.472, + 0.221 + ], + "angle": 0, + "content": "
# TasksStep Eval.?Test Size
MR-GSM8K [54]1✓2,999
RMBench [27]1✗1,327
CriticBench [24]1✗-
MathCheck-GSM [63]1✓516
MR-Ben [55]1✓5,975
ProcessBench [60]1✓3,400
PRMBench [37]9✓6,216
VLRewardBench* [20]1✗1,250
MultimodalRewardBench* [52]1✗5,211
VLRMBENCH*12✓12,634
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Table 1. Comparison between our VLRMBenchmark and others. * represents the benchmark for vision-language domain. 'Step Eval.' refers to whether the step-level labels are included." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.306, + 0.483, + 0.577 + ], + "angle": 0, + "content": "compassing eight widely used datasets. Moreover, to obtain high-quality samples containing reasoning processes, we develop a collaborative data filtering and generation pipeline. Specifically, we first utilize a small LVLM to filter samples along two dimensions: quality and difficulty. Subsequently, QVQ-72B-preview [39], which excels in VL reasoning, is employed to generate reasoning processes. Next, we use the advanced GPT-4o [2] to partition the reasoning steps and perform an initial validation of correctness. Finally, we apply rule-based filtering and manual inspection to correct errors in the reasoning process, yielding 1,000 samples that serve as the foundation for task construction. Based on these samples, we introduce three themes: step-based, outcome-based, and criticism-based, encompassing 12 diverse tasks and 12,634 questions. These tasks can comprehensively evaluate VLRMs in the aspects of process understanding, outcome judgment, and criticism generation, thus laying the groundwork for the development of VLRMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.807 + ], + "angle": 0, + "content": "We conducted extensive experiments on 21 open-source models (from 2B to 90B) and 5 closed-source models, such as GPT-4o and Claude-3.5-Sonnet [3], across 12 tasks. The results underscore the challenges of VLRMBench and demonstrate the narrowing gap between open-source and closed-source models. Specifically, in eight step-based tasks, even the advanced GPT-4o achieved only an average F1-Score of \\(62.4\\%\\). In two outcome-based binary classification tasks, GPT-4o attained an average accuracy of only \\(66.3\\%\\). Furthermore, in two criticism-based tasks, Qwen2.5-VL-72B [4] outperformed GPT-4o, with an average win rate of \\(53.3\\%\\), showcasing the superior performance of state-of-the-art (SOTA) open-source models. In addition, we performed extensive analytical studies, providing valuable insights into the development of VLRMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The main contributions can be summarized as follows: 1) We introduce VLRMBench, the first benchmark that integrates both comprehensiveness and challenge for the evaluation of VLRMs. VLRMBench is constructed from 1,000 high-quality samples curated through our collaborative data filtering and generation pipeline, covering areas such as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "mathematical reasoning, hallucination understanding, and multi-image understanding. 2) We designed 12 tasks, comprising a total of 12,634 carefully crafted questions, capable of evaluating models across three aspects: process understanding, outcome judgment, and critique generation. These diverse tasks enable targeted evaluations and reveal limitations of VLRMs. 3) We conducted extensive experiments on VLRMBench using 21 open-source models and 5 closed-source models, highlighting the challenges presented by VLRMBench and offering comprehensive analysis and guidance for the development of VLRMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.274, + 0.665, + 0.289 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.725, + 0.315 + ], + "angle": 0, + "content": "2.1. Reward Models (RMs)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Techniques such as Reinforcement Learning with Human Feedback (RLHF) utilize RMs that offer feedback for large models, aligning them with human preferences in terms of accuracy, utility, and safety [5, 34]. RMs can generally be categorized into three types based on their objectives. Process-based RMs [31, 59] evaluate intermediate reasoning steps to ensure logical consistency and the correctness of each step. Outcome-based RMs [6, 25, 28] assess the quality of the final output. In contrast, Critique-based RMs [8, 18, 45] emphasize offering textual feedback or critiques aimed at enhancing the model's performance. However, RMs are not always reliable [37]. Therefore, it is imperative to establish a benchmark that facilitates a comprehensive evaluation of RMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.794, + 0.56 + ], + "angle": 0, + "content": "2.2. Benchmarks of Reward Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.793 + ], + "angle": 0, + "content": "In recent years, several benchmarks for evaluating the capabilities of RMs have been proposed, driving the development of RMs in the field of natural language. For example, MR-GSM8k [54] redefines the model's role from a solver to a grader, necessitating the evaluation of the correctness of problem-solving steps. CriticBench [24] incorporates the Generate-Critique-Revise framework to assess the reasoning capabilities of models. ProcessBench [60] emphasizes step-by-step error detection in mathematical derivations, thereby increasing the granularity of the evaluation. PRMBench [37], based on PRM800K [23], employs GPT40 for data optimization and constructs a fine-grained evaluation dataset comprising 6,216 instances. PRMBench covers nine common error types and provides a comprehensive assessment of Process RMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The aforementioned benchmarks predominantly focus on the language modality and do not adequately address the evaluation needs of VLLMs. To this end, [20] introduces VLRewardBench, a benchmark for VLRMs containing 1,250 test samples. However, VLRewardBench remains limited to Pairwise Comparison, wherein only a judgment of superiority or inferiority is made between two responses," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3164" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.089, + 0.865, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.333, + 0.908, + 0.378 + ], + "angle": 0, + "content": "Figure 1. The collaborative data filtering and generation pipeline of VLRMBench. During the filtering stage, Qwen2VL-7B is tasked with answering questions under two conditions: without and with image input. Questions for which the model provides incorrect answers are retained. Subsequently, QVQ-72B-preview and GPT-4o are employed to generate reasoning processes and construct tasks for three themes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.482, + 0.478 + ], + "angle": 0, + "content": "failing to incorporate more diverse and enriched task types. Consequently, we propose VLRMBench, which encompasses three core themes and twelve distinct tasks, providing a holistic evaluation of VLRMs' overall performance and establishing a foundation for their future advancement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.229, + 0.509 + ], + "angle": 0, + "content": "3. VLRMBench" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.579 + ], + "angle": 0, + "content": "In this section, we initially present an overview of the pipeline for constructing VLRMBench, along with the data statistics information, followed by an in-depth description of the task design." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.379, + 0.606 + ], + "angle": 0, + "content": "3.1. Overall pipeline of VLRMBench" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.854 + ], + "angle": 0, + "content": "In Figure 1, we present a progressive and intuitive pipeline, which is designed to construct VLRMBench in three stages: data collection, filtering, and generation. During the collection phase, we integrated four mathematical, two hallucination, and two multi-image understanding datasets, totaling 16,550 samples and 19,343 images. In the filtering phase, to select high-quality and challenging samples, we introduced a joint filtering mechanism. Specifically, we initially employed Qwen2VL-7B [44] to answer questions without image information. If the model answered correctly, it indicated that the sample was of low quality [7]. Subsequently, we incorporated the image information into the question and instructed the model to answer it. If the model answered correctly, it suggested that the question was too simple to be a challenge. Consequently, we retained only those samples that could not be answered correctly in ei" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.402, + 0.907, + 0.433 + ], + "angle": 0, + "content": "ther question-answering mode, resulting in a total of 6,715 samples, which formed our basic database \\(D_{b}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.835 + ], + "angle": 0, + "content": "In the generation phase, we utilized QVQ-72B-preview [39] to generate reasoning process for the samples in \\( D_{b} \\), discarding those where QVQ-72B-preview yielded incorrect answers, as verified via comparison with the standard answers. Subsequently, GPT-4o was employed to perform step-level segmentation of QVQ-72B-preview's reasoning process, grouping semantically similar and logically related sentences into the one step. Following this, we obtained 3,335 samples with reasoning segmented at the step level. Furthermore, based on specific criteria², we excluded samples with excessively long or short reasoning steps and ensured a balanced distribution across the three sample types. Finally, to ensure the quality of the samples, a manual review was conducted. Specifically, three PhD volunteers were recruited to verify the high-quality reasoning processes generated by QVQ-72B-preview. If any volunteer identified an error in the reasoning process, all three volunteers would collaboratively review the sample and correct any potential mistakes. Additionally, examples of manual corrections applied to the samples could be found in the Appendix. After completing the three aforementioned stages, 1,000 high-quality samples were retained for constructing the specific task, denoted as \\( D \\). For more detailed statistical information, please refer to Table 2 and Figure 2, which provide a description of the sample sources, the average word count, and the number of reasoning steps." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Qwen2VL-7B ranks in the top \\(25\\%\\) of the OpenCompass leaderboard [11]. With moderate model capabilities and a small model size, it can efficiently filter out low-quality and low-difficulty samples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "2We excluded samples where the number of reasoning steps exceeded the average (i.e. 20) or was fewer than 5." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3165" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.481, + 0.228 + ], + "angle": 0, + "content": "
CategoryDataset#Samples#AvgWORDS
MathMathVista [29]142696
MathVision [42]281,021
MathVerse [58]126742
DynaMath [64]251691
HallucinationIllusionVQA [36]62736
HallusionBench [15]334605
Multi-ImageBLINK [14]23790
Mantis-Eval [16]34592
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.239, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Table 2. Statistics on the sample sources of VLRMBench. We also counted the average number of words in the reasoning process under different data sources." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.307, + 0.255, + 0.323 + ], + "angle": 0, + "content": "3.2. Step-based tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.484, + 0.419 + ], + "angle": 0, + "content": "The step-based theme consists of 8 tasks designed to assess the reasoning process understanding ability of VLRMs. Under this theme, models will be provided with a modified reasoning process and are expected to output a sequence of 0s and 1s to indicate whether each reasoning step contains deviated content." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Step correctness (SC) For a Vision-Language Reward Model (VLRM), detecting errors at each step of the reasoning process is a critical capability [23, 37]. Building on this, we propose the SC task. Specifically, given the reasoning process \\( R_{n} \\) of a sample from \\( D \\), where \\( n \\) denotes the total number of reasoning steps, we inject errors into \\( m \\) of the \\( n \\) steps \\( (0 < m < n) \\) of \\( R_{n} \\) using the advanced GPT-4o, to simulate potential errors during reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.805 + ], + "angle": 0, + "content": "Redundancy Detection (RD) Existing reasoning models often produce redundant information when answering questions, leading to an increase in output token length, which is both inefficient and environmentally unfriendly. To this end, we propose the RD task, which requires VLRMs to identify redundant information generated during the reasoning process. Specifically, based on \\( D \\), we introduce redundant descriptions in certain reasoning steps of each sample. For instance, we incorporate descriptions from \\( R_{i-j} \\) into \\( R_i \\) to intentionally create redundancy. Furthermore, we employ GPT-4 to generate paraphrased versions of the content from \\( R_{i-j} \\) and incorporate them into \\( R_i \\), thereby amplifying redundancy. Unlike the SC task, the introduction of redundant information does not compromise the correctness of the final answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Confidence Misdirection (CM) When confident and assertive terms, such as 'definitely' and 'without a doubt', are present in an incorrect reasoning step, is it possible for VL-RMs to be misled into perceiving the erroneous reasoning as correct? To investigate this issue, we propose the CM task," + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.094, + 0.871, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.274, + 0.907, + 0.303 + ], + "angle": 0, + "content": "Figure 2. The distribution of the number of reasoning steps for different data sources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.406 + ], + "angle": 0, + "content": "which is derived from the SC task. Specifically, for each sample in the SC task, we incorporate confidence-boosting terms into the erroneous reasoning step to mislead VLRMs into perceiving the step as correct, thereby evaluating the robustness of VLRMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.909, + 0.745 + ], + "angle": 0, + "content": "Existential Hallucination (EH) and Attribute Hallucination (AH) When answering questions using LVLMs, hallucination phenomena often arise [26]. Although other parts of the reasoning process may be correct, the introduction of hallucinatory errors or the omission of key entities can lead to biased final results. To evaluate whether VLRMs are capable of detecting hallucinations introduced during the reasoning process, we have designed two tasks on the top of \\( D \\). The first task, EH, assesses the ability of the VLRM to identify entities that were incorrectly introduced during reasoning process but do not exist in the image, or to detect elements that may have been overlooked, such as objects, people, etc. The second task, AH, examines the VLRM's capability to accurately identify the attributes of entities, particularly in cases where LVLMs wrongly attribute properties to an entity that it does not possess or when attribute descriptions conflict with the visual content of the image. Such errors are typically reflected in misjudgments regarding object attributes, such as size, color, or shape. VLRMs must integrate image information effectively to accurately identify hallucination errors during the reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Detail Error (DE) When addressing complex problems, the reasoning process often involves multiple details, particularly in the handling of numbers and symbols. Even if the overall reasoning appears to be sound, there may still be hidden errors in finer details (e.g., numerical errors in calculations or symbolic mistakes in formulas). These errors are often difficult to detect, yet they can affect the precision of the final results. The introduction of DE tasks serves to evaluate the sensitivity of VLRMs to reasoning details, as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.423, + 0.108 + ], + "angle": 0, + "content": "well as their ability to identify fine-grained errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.484, + 0.262 + ], + "angle": 0, + "content": "Spatial Relationship (SR) During visual reasoning, LVLMs may incorrectly describe the spatial relationships between objects or entities, leading to erroneous results. For instance, LVLMs might misdescribe an object located at the top of the image as being at the bottom. Therefore, we select samples containing spatial relationship descriptions from \\( D \\) to construct the SR task, aiming to evaluate VLRMs' understanding of spatial relationships in a image and their ability to detect spatial errors in a reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.278, + 0.484, + 0.475 + ], + "angle": 0, + "content": "Image Confusion (IC) In multi-image tasks, LVLMs are required to extract and associate relevant information from multiple images, which typically involves identifying various entities and their interrelations across different images. However, due to the presence of similar elements or backgrounds between images, LVLMs may erroneously associate an entity from one image with another image. This 'image reference error' not only causes confusion but also potentially impacts the accuracy of subsequent reasoning. Thus, we filter samples related to multi-image understanding from \\(D\\) to construct the IC task, which aims to evaluate whether VLRMs can identify image reference errors during the reasoning process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.292, + 0.497 + ], + "angle": 0, + "content": "3.3. Outcome-based tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.55 + ], + "angle": 0, + "content": "The outcome-based theme encompasses two binary classification tasks, aiming to evaluate the ability of VLRMs to assess and predict the accuracy of outcomes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Multi-solution Judgment (MJ) During the training stage, preference optimization techniques, such as DPO [35], have become a widely used paradigm. These methods typically rely on the collection of preference samples. VLRMs are frequently employed to assess the quality of preference samples, thereby becoming a crucial component of preference optimization [21, 53]. To this end, we introduce the MJ task, which aims to investigate the discriminative ability of VLRMs regarding different reasoning processes for the same problem. This task is discussed as a core element in VLRewardBench [20] and Multimodal Reward-Bench [52]. Specifically, for the reasoning process \\( R \\) of any question \\( Q \\) in \\( D \\), we introduce the reasoning process \\( R' \\) of the same question in the SC, RD and CM tasks, thereby constructing a triplet \\( (Q, R, R') \\). VLRMs are required to judge the quality of \\( R \\) and \\( R' \\) and score them accordingly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Forecasting Future (FF) When LVLMs perform inference, test-time scaling (TTS) can further improve their performance [47, 51]. However, these techniques often introduce additional inference time, primarily due to the necessity of exploring multiple reasoning paths. If the correctness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.17 + ], + "angle": 0, + "content": "of the final answer could be predicted based on the reasoning process of the first m steps, \\( R_{[0:m]} \\), this prediction could substantially expedite the TTS process. Therefore, the FF task is proposed, wherein VLRMs are required to assess the correctness of the final answer based on \\( R_{[0:m]} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.719, + 0.193 + ], + "angle": 0, + "content": "3.4. Criticism-based tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Error Reason Analysis (ERA) and Error Correction (EC) Unlike Process RM and Outcome RM, which focus solely on evaluating processes or outcomes, Critique RM emphasizes the critique of the reasoning process, providing textual feedback that includes error analysis and correction. Consequently, we propose the ERA and EC tasks to assess the capabilities of VLRMs in these two aspects. In the ERA task, VLRMs are presented with erroneous reasoning steps derived from the SC task, and are tasked with identifying the error and analyzing its underlying cause. In contrast, the EC task does not require error cause analysis; rather, VLRMs directly correct the identified error and produce the corrected reasoning process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.412, + 0.646, + 0.429 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.608, + 0.453 + ], + "angle": 0, + "content": "4.1. Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.627 + ], + "angle": 0, + "content": "We conducted extensive experiments on 21 open-source models and 5 closed-source models. Specifically, the open-source models, which range from 2B to 90B, including Llama3.2 (11B/90B) [12], Llava-OneVision (7B/72B) [19], Qwen2VL (2B/7B/72B) [44], Qwen2.5VL (3B/7B/72B) [4], QVQ-72B-preview [39], Ovis2 (8B/16B/34B) [30], InternVL2.5 (2B/8B/38B/78B) [9], Phi-3.5-Vision (4.2B) [1], MiniCPM-o-2.6 (8B), and MiniCPM-V-2.6 (8B) [33]. For the closed-source models, we employed GLM-4V-Flash, GPT-4o [2], Gemini-2.0-Flash-/Flash-thinking-exp-1219 [38], and Claude-3.5-Sonnet [3]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.691, + 0.652 + ], + "angle": 0, + "content": "4.2. Evaluation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We follow the 'LVLM-as-a-Judge' paradigm and introduce different question-answering templates to assess the models [20, 52]. Specifically, for the SC, RD, CM, EH, AH, DE, SR, and IC tasks, the model takes a reasoning process as input and outputs a judgment regarding whether each reasoning step contains redundancies or other errors. For these eight tasks, we employ the weighted F1-Score as the evaluation metric, as defined in Equation 1, which better addresses class imbalance issues and ensures the model's performance on the minority class is adequately considered. For the MJ task, the model is provided with (good and bad) reasoning process pairs for the same question and outputs scores for both. If the model assigns a higher score to the good sample, it is deemed correct; otherwise, it is considered incorrect. For the FF task, the model takes several preceding steps of the reasoning process as input and outputs a judgment on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3167" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.09, + 0.902, + 0.52 + ], + "angle": 0, + "content": "
ModelStep (5,722)Outcome (5,402)Criticism (1,510)
SCRDCMEHAHDESRICW-Avg.MJFFW-Avg.ERAwERAτERALECwECTECL
# Sample99899799494694465912658-2,9862,416-974536
Open-source LVLMs < 10B (Small group)
InternVL2.5-2B33.834.130.022.423.833.519.930.129.226.660.841.94.86.788.52.310.087.6
Qwen2VL-2B28.326.220.917.819.126.623.826.423.010.559.632.50.20.199.72.55.891.7
Qwen2.5VL-3B29.932.723.227.226.945.525.142.630.131.967.147.67.112.180.84.821.273.9
Phi-3.5-Vision (4.2B)35.938.827.920.721.133.417.731.629.212.165.235.83.89.087.24.222.473.4
Qwen2VL-7B38.029.332.021.521.432.821.621.128.852.064.757.73.98.187.95.423.271.4
Qwen2.5VL-7B43.433.237.822.823.945.515.629.133.426.070.746.037.722.040.39.351.938.8
Llava-OneVision-7B31.825.721.516.017.224.816.120.722.618.057.835.83.910.485.77.134.458.5
InternVL2.5-8B36.628.431.121.921.236.515.124.128.661.473.866.934.521.943.66.236.157.7
Ovis2-8B47.728.237.825.225.739.817.025.833.331.465.346.623.225.950.914.363.522.2
MiniCPM-V-2.6 (8B)44.938.336.926.126.536.824.230.634.637.962.548.95.411.782.93.424.771.9
MiniCPM-o-2.6 (8B)41.035.337.329.629.036.728.330.134.615.064.337.09.219.870.97.633.958.4
Avg.37.431.830.622.823.335.620.428.429.829.364.745.212.213.474.46.129.764.1
10B < Open-source LVLMs < 40B (Middle group)
Llama3.2-11B33.927.129.720.220.031.917.125.826.723.961.140.54.210.385.55.434.759.8
Ovis2-16B48.547.947.940.635.844.521.928.443.649.769.258.430.926.642.510.469.320.3
Ovis2-34B65.351.164.554.551.659.630.449.057.046.375.759.445.926.627.515.476.48.1
InternVL2.5-38B61.049.662.655.051.761.434.648.156.145.179.360.421.124.554.412.556.630.9
Avg.52.243.951.242.639.849.426.037.845.941.371.354.725.522.052.510.959.329.8
Open-source LVLMs > 40B (Large group)
Llava-OneVision-72B49.430.650.237.934.145.113.926.540.432.269.749.015.322.162.69.451.039.6
Qwen2VL-72B55.540.256.148.941.656.422.443.648.837.772.153.122.420.457.17.757.135.1
Qwen2.5VL-72B72.841.770.464.659.972.437.963.362.665.680.272.174.115.110.815.677.07.3
QVQ-72B-preview42.435.336.329.427.137.820.826.134.349.238.444.44.210.085.89.451.039.6
InternVL2.5-78B62.142.062.853.348.760.428.547.254.048.678.862.154.224.920.910.863.925.3
Llama3.2-90B40.838.841.334.229.435.417.021.136.252.769.060.09.918.671.49.556.833.7
Avg.53.838.152.944.740.151.323.438.046.147.768.056.830.018.551.410.459.530.1
Closed-source LVLMs
Claude-3.5-Sonnet70.853.765.763.962.863.440.865.762.982.275.179.060.625.513.921.253.924.9
GPT-4o73.750.666.657.658.671.848.264.662.458.476.066.30.0100.00.00.0100.00.0
GLM-4V-Flash37.233.532.928.227.741.822.920.232.837.266.750.434.224.341.56.954.139.0
Gemini-2.0-Flash69.355.966.165.662.363.446.867.863.454.277.164.432.822.544.77.152.540.3
Gemini-2.0-FTE.-121967.955.565.765.962.363.246.467.163.153.177.063.837.422.640.06.850.243.1
Avg.63.849.859.456.254.760.741.057.157.057.074.464.833.039.028.08.462.129.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.531, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Table 3. Results on the VLRMBench. For Step-based tasks, the F1-Score is the metric. For Outcome-based tasks, Accuracy is utilized as the metric. In the case of Criticism-based tasks, the metric is Win Rate, expressed as Win: Tie: Lose (Simplified as \\( W \\), \\( T \\) and \\( L \\)). Note that W-Avg. represents the weighted average. Evaluation prompt templates and some case studies could be found in the Appendix B and C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.705 + ], + "angle": 0, + "content": "whether a correct answer can ultimately be derived. Consequently, for the MJ and FF tasks, Accuracy is employed as the evaluation metric. For the ERA and EC tasks, the model's output is a description of the error reasons in the reasoning process and the corrected reasoning process. We utilize the output of GPT-4o as the baseline and adopt Win Rate [13, 61] as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.483, + 0.826 + ], + "angle": 0, + "content": "We configured the max new tokens to 4,096 across all tasks for the five closed-source models and QVQ-72B-preview. In contrast, for the other models, we set the max new tokens to 4,096 specifically for the ERA and EC tasks, while fixing it at 256 for the remaining tasks. To ensure the reproducibility of the experimental results, we set the temperature coefficient to 0 for all models. All experiments were conducted on 8 NVIDIA H800 GPUs." + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.846, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\nF 1 = \\frac {N _ {\\text {n e g}}}{N _ {\\text {p o s}} + N _ {\\text {n e g}}} \\cdot F 1 _ {\\text {p o s}} + \\frac {N _ {\\text {p o s}}}{N _ {\\text {p o s}} + N _ {\\text {n e g}}} \\cdot F 1 _ {\\text {n e g}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(N_{\\mathrm{pos}}\\) and \\(N_{\\mathrm{neg}}\\) represent the number of correct and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.601, + 0.864, + 0.616 + ], + "angle": 0, + "content": "incorrect steps in the reasoning process, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.655, + 0.637 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.905, + 0.674 + ], + "angle": 0, + "content": "The experimental results on VLRMBench are shown in Table 3, leading to the following conclusions:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.825 + ], + "angle": 0, + "content": "1) Among open-source models, surprisingly, the performance of Qwen2.5VL-72B is comparable to that of advanced closed-source models. For instance, across the eight Step-based tasks, Qwen2.5VL-72B achieved an average F1-Score of \\(62.6\\%\\), surpassing GPT-4o by \\(0.2\\%\\) and only slightly trailing Claude-3.5-Sonnet by \\(0.3\\%\\). In the MJ and FF tasks, Qwen2.5VL-72B also outperformed GPT-4o by \\(7.2\\%\\) and \\(4.2\\%\\), respectively. Notably, in the ERA task, Qwen2.5VL-72B attained a win rate of \\(74.1\\%\\) over GPT-4o, significantly outperforming Claude-3.5-Sonnet at \\(60.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "2) LVLMs that have not undergone specific training show diminishing performance gains as the model size increases. Specifically, for LVLMs in the Middle group, compared to those in the Small group, the performance gain in Step-based and Outcome-based tasks was \\(54.0\\%\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3168" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.468, + 0.234 + ], + "angle": 0, + "content": "
OrderModelAvg. ScoreAcc
Better firstQwen2.5VL-3B[9.0, 8.3]23.7
Llama3.2-11B[8.6, 4.8]52.8
InternVL2.5-78B[8.4, 7.4]38.6
GPT-4o[9.8, 7.8]52.2
Better secondQwen2.5VL-3B[8.7, 9.0]21.4
Llama3.2-11B[8.3, 5.6]8.4
InternVL2.5-78B[8.1, 9.1]45.6
GPT-4o[7.4, 8.1]43.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.245, + 0.475, + 0.259 + ], + "angle": 0, + "content": "Table 4. The results on the MJ task with different sample order." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.482, + 0.439 + ], + "angle": 0, + "content": "21.0%, respectively. In contrast, the corresponding performance gains in the Large group relative to the Middle group were only 0.4% and 3.8%. Even excluding QVQ-72B-preview from the Large group, the relative performance improvement remains limited, at 5.4% and 8.4%. This suggests that blindly increasing the size of VLLMs does not lead to significant performance improvements. Consequently, we recommend that future research prioritize training VLRMs at the Middle group scale initially." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.569 + ], + "angle": 0, + "content": "3) The Outcome-based tasks, being simple binary classification problems, remain challenging for most open-source models, and even closed-source models perform only marginally better than random guessing. For open-source models across three size groups, the average accuracy for Outcome-based tasks is around \\(50\\%\\), similar to random guessing. For advanced closed-source models, the average accuracy only exceeds random guessing by \\(14.8\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.756 + ], + "angle": 0, + "content": "4) Deterministic vocabulary and hallucinated information can effectively mislead the VLRMs, reducing its ability to understand reasoning processes. The CM task, based on the SC task, introduces deterministic vocabulary (e.g., must, definitely) into sentences with erroneous reasoning steps, while EH and AH tasks introduce hallucinated information during the reasoning process. Existing models, particularly in the Small group, demonstrate a degradation in F1-Score of \\(6.8\\%\\), \\(14.6\\%\\), and \\(14.1\\%\\) for CM, EH, and AH tasks, respectively, when compared to the SC task. This performance degradation remains significant even as model size increases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "5) Existing models show strong robustness to numerical errors but limited resilience to perturbations in spatial relationships and image references. We observed that for Large group, the F1-Score in the DE task only decreased by \\(2.5\\%\\) compared to SC. However, for the SR and IC tasks, the performance dropped by \\(30.4\\%\\) and \\(15.8\\%\\), respectively. Overall, our VLRMBench provide unique challenges that surpass traditional benchmarks, with even the most advanced models achieving only mediocre performance." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.895, + 0.173 + ], + "angle": 0, + "content": "
Model≤10 steps10-15 steps>15 steps
Qwen2.5VL-3B35.630.125.6
Llama3.2-11B42.932.828.4
InternVL2.5-78B68.264.154.1
GPT-4o77.875.268.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.184, + 0.893, + 0.198 + ], + "angle": 0, + "content": "Table 5. Results on the SC task with different reasoning steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.681, + 0.243 + ], + "angle": 0, + "content": "5. Further Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.887, + 0.266 + ], + "angle": 0, + "content": "5.1. Is the o1-like model suitable to be a VLRM?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.906, + 0.485 + ], + "angle": 0, + "content": "QVQ-72B-preview [39], an open-source visual-language model, exhibits strong reasoning capabilities. However, it exhibits a potential drawback regarding the degradation of instruction-following ability. For instance, in the SC task, by incorporating instruction prompts, the sample's original question, and the reasoning process as input, we expect the VLRM to evaluate each reasoning step. However, QVQ-72B-preview often disregards the given instructions and directly answers the original question, leading to suboptimal performance on VLRMBench, with additional cases provided in the Appendix C. Therefore, as an o1-like model, it must not only demonstrate strong reasoning abilities but also exhibit robust instruction-following capabilities to be considered a viable candidate for a VLRM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.731, + 0.513 + ], + "angle": 0, + "content": "5.2.Impact of sample order" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.716 + ], + "angle": 0, + "content": "In the MJ task, VLRM scores two reasoning processes for the same question. To further investigate the positional bias, we employed two different sample orders, as shown in Table 4. Specifically, 'Better first' places the superior reasoning process in the first position, while 'Better second' places it in the second position. The results indicate that Llama3.2-11B consistently assigns a higher score to the reasoning process in the first position, irrespective of its quality. In contrast, the other three models exhibit the ability to adjust their scores based on the quality of the reasoning. This positional bias phenomenon is also observed in [43]. As a result, in Table 3, we utilized the sum of the two scores for the MJ task evaluation to ensure fairness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.865, + 0.744 + ], + "angle": 0, + "content": "5.3. Performance on various reasoning length" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We further investigate the impact of the number of reasoning steps on the VLRMs' process understanding. As shown in Table 5, we classified the number of reasoning steps in the SC task into three categories: the short group (\\(\\leq 10\\) steps), with 263 samples; the middle group (10-15 steps), with 468 samples; and the long group (\\(>15\\) steps), with 267 samples. Specifically, for all four models, a decrease in F1-score was observed as the number of reasoning steps increased. Notably, for the Llama3.2-11B model, the difference between the short and long groups reached \\(14.5\\%\\), and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3169" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.449, + 0.173 + ], + "angle": 0, + "content": "
Model0-shot1-shot2-shot
Qwen2.5VL-3B29.931.531.1
Llama3.2-11B33.942.242.3
InternVL2.5-78B62.155.955.7
GPT-4o73.771.770.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.184, + 0.478, + 0.197 + ], + "angle": 0, + "content": "Table 6. Results on SC task with various ICL few-shot numbers." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.216, + 0.462, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.354, + 0.459, + 0.37 + ], + "angle": 0, + "content": "Figure 3. Performance changes with the test-time scaling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.473 + ], + "angle": 0, + "content": "even for the advanced GPT-4o, the F1-Score dropped from \\(77.8\\%\\) to \\(68.6\\%\\). This suggests that current models are less capable of handling long information, underscoring the necessity of enhancing their ability to tackle long reasoning process when training VLRMs in the future." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.483, + 0.476, + 0.5 + ], + "angle": 0, + "content": "5.4. Impacts of In-Context Learning (ICL) setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.824 + ], + "angle": 0, + "content": "In main experiments, we employ zero-shot setting by default. Furthermore, using the SC task as an example, we explore the effect of ICL on VLRMs. As presented in Table 6, the results indicate that providing contextual examples is more beneficial for small models but harms the performance of large models. Specifically, with the help of one-shot learning, Qwen2.5VL-3B's performance improved from \\(29.9\\%\\) to \\(31.5\\%\\), and Llama3.2-11B saw an improvement of \\(8.3\\%\\). However, for GPT-4o, one-shot learning led to a \\(2.0\\%\\) performance degradation, and InternVL2.5-78B's performance decreased by \\(6.2\\%\\). One possible explanation for this is that small models exhibit relatively weaker instruction-following capabilities, and ICL examples may help guide them in responding to queries in the designated format. In contrast, large models, which already possess strong comprehension abilities and adaptability to complex tasks, may find that providing contextual examples interferes with their inherent understanding process, resulting in performance decline. Thus, in the SC task, smaller models exhibit a greater reliance on ICL examples, whereas larger models tend to rely more on their pre-trained knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.339, + 0.851 + ], + "angle": 0, + "content": "5.5. Impacts of test-time scaling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We investigated the impact of test-time scaling (i.e. voting method) for VLRMs on the MJ task. To ensure the diversity of the generated results, we set the temperature coefficient" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.09, + 0.885, + 0.165 + ], + "angle": 0, + "content": "
LVLMVLRMMathVistaMathVerseMathVisionAvg.
7B7B10.45.84.77.0
72B7B11.84.23.26.4
7B72B21.819.721.521.0
72B72B11.79.311.410.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.175, + 0.887, + 0.188 + ], + "angle": 0, + "content": "Table 7. Results of the correction rate on math benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.31 + ], + "angle": 0, + "content": "to 0.9 and performed K sampling. As illustrated in Figure 3, surprisingly, as the value of K increases, the accuracy of Qwen2.5VL-3B improved by \\(12.4\\%\\). Notably, at \\(\\mathrm{K} = 9\\), its performance was only \\(4.8\\%\\) lower than that of GPT-4o. These findings highlight the substantial effectiveness of test-time scaling techniques in enhancing VLRMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.321, + 0.792, + 0.337 + ], + "angle": 0, + "content": "5.6. Improve LVLMs with a VLRM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.908, + 0.662 + ], + "angle": 0, + "content": "To demonstrate the significant role of VLRMs in the reasoning process, as presented in Table 7, we conducted experiments on three mathematical benchmarks using Qwen2.5VL-7B/72B. For example, in the second row of Table 7, Qwen2.5VL-72B was employed as the LVLM to answer the mathematical questions, whereas Qwen2.5VL-7B was used as the VLRM to perform error reason analysis of the LVLM's reasoning process. The analysis was subsequently fed back into the LVLM as feedback, allowing it to generate a second round of answers. Besides, we defined the correction rate, which is the ratio of the number of questions correctly answered in the second round after feedback to the number of questions incorrectly answered in the first round. The experimental results demonstrate that VLRM can provide valuable information to the LVLM through error analysis, thereby facilitating mistake correction, with an average gain of \\(11.3\\%\\). These findings suggest that even LVLMs not trained on proprietary data can act as VLRMs, providing more correction possibilities during testing based on this intuitive feedback mechanism. This further underscores the potential value of developing efficient VLRMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.64, + 0.695 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a comprehensive and challenging benchmark, named VLRMBench. This benchmark covers 3 themes and 12 tasks, comprising a total of 12,634 questions. Following the 'LVLM-as-a-Judge' paradigm, we conduct systematic experiments and analyses on VLRMBench using 26 models, deriving several key insights: enhancing the performance of VLRMs in long-text comprehension, exploring various feedback patterns from VLRMs to LVLMs, and developing the application potential of proprietary VLRMs, etc. We hope that VLRMBench will serve as a standardized evaluation platform to propel the development of more efficient VLRMs. In future work, we plan to incorporate larger datasets specifically designed to train VLRMs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3170" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.483, + 0.146 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grant No. 61977045)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.159, + 0.188, + 0.175 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.483, + 0.253 + ], + "angle": 0, + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.255, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[3] Anthropic. Claude 3.5: Sonnet, 2025. Accessed: 2025-02-25. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.534 + ], + "angle": 0, + "content": "[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.184, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B Hashimoto. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision, pages 148-166. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. Transactions on Machine Learning Research, 2024, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[17] Wasif Khan, Seowung Leem, Kyle B See, Joshua K Wong, Shaoting Zhang, and Ruogu Fang. A comprehensive survey of foundation models in medicine. IEEE Reviews in Biomedical Engineering, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[18] Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, et al. Solar 10.7 b: Scaling large language models with simple yet effective depth up-scaling. arXiv preprint arXiv:2312.15166, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[20] Lei Li, Yuancheng Wei, Zhihui Xie, Xuqing Yang, Yifan Song, Peiyi Wang, Chenxin An, Tianyu Liu, Sujian Li, Bill Yuchen Lin, et al. Vlrewardbench: A challenging benchmark for vision-language generative reward models. arXiv preprint arXiv:2411.17451, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[21] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. arXiv preprint arXiv:2410.09421, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[22] Zongxia Li, Xiyang Wu, Hongyang Du, Huy Nghiem, and Guangyao Shi. Benchmark evaluations, applications, and challenges of large vision language models: A survey. arXiv preprint arXiv:2501.02189, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Zicheng Lin, Zhibin Gou, Tian Liang, Ruilin Luo, Haowei Liu, and Yujiu Yang. Criticbench: Benchmarking lms for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "3171" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "critique-correct reasoning. arXiv preprint arXiv:2402.14809, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[25] Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xi-tian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[27] Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. Rm-bench: Benchmarking reward models of language models with subtlety and style. arXiv preprint arXiv:2410.16184, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[28] Xingzhou Lou, Dong Yan, Wei Shen, Yuzi Yan, Jian Xie, and Junge Zhang. Uncertainty-aware reward model: Teaching reward models to know what is unknown. arXiv preprint arXiv:2410.00847, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[29] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[30] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[31] Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Yang You, and Hongxia Yang. Let's reward step by step: Step-level reward model as the navigators for reasoning. arXiv preprint arXiv:2310.10080, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[32] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[33] OpenBMB. Minicpm o: A gpt-4o-level mllm for vision, speech, and multimodal live streaming on your phone, 2025. Accessed: 2025-02-25. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[35] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[36] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[37] Mingyang Song, Zhaochen Su, Xiaoye Qu, Jiawei Zhou, and Yu Cheng. Prmbench: A fine-grained and challenging benchmark for process-level reward models. arXiv preprint arXiv:2501.03124, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[39] Qwen Team. Qvq: To see the world with wisdom, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[41] Jiahao Wang, Bolin Zhang, Qianlong Du, Jiajun Zhang, and Dianhui Chu. A survey on data selection for llm instruction tuning. arXiv preprint arXiv:2402.05123, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[42] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.405 + ], + "angle": 0, + "content": "[43] Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[45] Tianlu Wang, Ilia Kulikov, Olga Golovneva, Ping Yu, Weizhe Yuan, Jane Dwivedi-Yu, Richard Yuanzhe Pang, Maryam Fazel-Zarandi, Jason Weston, and Xian Li. Self-taught evaluators. arXiv preprint arXiv:2408.02666, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[46] Zhilin Wang, Alexander Bukharin, Olivier Delalleau, Daniel Egert, Gerald Shen, Jiaqi Zeng, Oleksii Kuchaiev, and Yi Dong. Helpsteer2-preference: Complementing ratings with preferences. arXiv preprint arXiv:2410.01257, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[47] Jinyang Wu, Mingkuan Feng, Shuai Zhang, Ruihan Jin, Feihu Che, Zengqi Wen, and Jianhua Tao. Boosting multimodal reasoning with mcts-automated structured thinking. arXiv preprint arXiv:2502.02339, 2025. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[48] Wang Xiyao, Yang Zhengyuan, Li Linjie, Lu Hongjin, Xu Yuancheng, Lin Chung-Ching Lin, Lin Kevin, Huang Furong, and Wang Lijuan. Scaling inference-time search with vision value model for improved visual comprehension. arXiv preprint arXiv:2412.03704, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[50] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[51] Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3172" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "carlo tree search. arXiv preprint arXiv:2412.18319, 2024. 1,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[52] Michihiro Yasunaga, Luke Zettlemoyer, and Marjan Ghazvininejad. Multimodal rewardbench: Holistic evaluation of reward models for vision language models. arXiv preprint arXiv:2502.14191, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[53] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[54] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[55] Zhongshen Zeng, Yinhong Liu, Yingjia Wan, Jingyao Li, Pengguang Chen, Jianbo Dai, Yuxuan Yao, Rongwu Xu, Zehan Qi, Wanru Zhao, et al. Mr-ben: A meta-reasoning benchmark for evaluating system-2 thinking in llms. arXiv preprint arXiv:2406.13975, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[56] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Photogrammetry and Remote Sensing, 221:64-77, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[57] Di Zhang, Jingdi Lei, Junxian Li, Xunzhi Wang, Yujie Liu, Zonglin Yang, Jiatong Li, Weida Wang, Suorong Yang, Jianbo Wu, et al. Critic-v: Vlm critics help catch vlm errors in multimodal reasoning. arXiv preprint arXiv:2411.18203, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[58] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[59] Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[60] Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[62] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Zihao Zhou, Shudong Liu, Maizhen Ning, Wei Liu, Jindong Wang, Derek F Wong, Xiaowei Huang, Qiufeng Wang, and Kaizhu Huang. Is your model really a good math rea" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "soner? evaluating mathematical reasoning with checklist. arXiv preprint arXiv:2407.08733, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[64] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "3173" + } + ] +] \ No newline at end of file diff --git a/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_origin.pdf b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f4365395b20800033f3c14b8e3c54af64624120 --- /dev/null +++ b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/8e730ad2-3996-4a2e-9875-e28304c3e917_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d09718d847818d3d4b052474c21b0086a4a1499ba10a63ae8b3b2f38b900310b +size 541582 diff --git a/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/full.md b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..999449ebc2fb0c8fd7e1b8ded1ed751aee04ed7c --- /dev/null +++ b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/full.md @@ -0,0 +1,278 @@ +# VLRMBench: A Comprehensive and Challenging Benchmark for Vision-Language Reward Models + +Jiacheng Ruan $^{1,*}$ Wenzhen Yuan $^{1,*}$ Xian Gao $^{1}$ Ye Guo $^{2}$ Daoxin Zhang $^{2}$ Zhe Xu $^{2}$ Yao Hu $^{2}$ Ting Liu $^{1}$ Yuzhuo Fu $^{1,\dagger}$ $^{1}$ Shanghai Jiao Tong University + $^{2}$ Xiaohongshu Inc. + +jackchenruan@sjtu.edu.cn + +# Abstract + +Although large visual-language models (LVLMs) have demonstrated strong performance in multimodal tasks, errors may occasionally arise due to biases during the reasoning process. Recently, reward models (RMs) have become increasingly pivotal in the reasoning process. Specifically, process RMs evaluate each reasoning step, outcome RMs focus on the assessment of reasoning results, and critique RMs perform error analysis on the entire reasoning process, followed by corrections. However, existing benchmarks for vision-language RMs (VLRMs) typically assess only a single aspect of their capabilities (e.g., distinguishing between two answers), thus limiting the all-round evaluation and restricting the development of RMs in the visual-language domain. To address this gap, we propose a comprehensive and challenging benchmark, dubbed as VLRMBench, encompassing 12,634 questions. VLRMBench is constructed based on three distinct types of datasets, covering mathematical reasoning, hallucination understanding, and multi-image understanding. We design 12 tasks across three major categories, focusing on evaluating VLRMs in the aspects of process understanding, outcome judgment, and critique generation. Extensive experiments are conducted on 21 open-source models and 5 advanced closed-source models, highlighting the challenges posed by VLRMBench. For instance, in the 'Forecasting Future', a binary classification task, the advanced GPT-4o achieves only a $76.0\%$ accuracy. The code is available at https://github.com/JCruan519/VLRMBench. + +# 1. Introduction + +Recently, driven by the development of large language models (LLMs) [32, 40, 49, 50], large visual-language models + +(LVLMs) have also achieved notable progress [22]. They have been broadly applied across domains such as medical imaging [17], remote sensing [56], and autonomous driving [62]. However, despite these remarkable achievements, they still encounter limitations in complex vision-language (VL) reasoning tasks, primarily due to insufficient reasoning depth and the absence of self-correction mechanisms [10, 51]. To mitigate this issue, reward models (RMs) have been introduced to detect errors in model responses, thereby enhancing the accuracy and efficiency of LVLMs [57]. + +During both the training and inference phases, RMs play a pivotal role. Firstly, prior to training, RMs can be employed to filter high-quality samples and facilitate the construction of automated data synthesis pipelines [21, 41, 46]. Secondly, during training, preference optimization techniques (e.g., RLAIF [53]) align the model's preferences with RMs, thereby enhancing the quality of the generated content. Finally, during inference, test-time scaling techniques, aided by RMs, are widely applied [47, 48], trading additional inference time for improved model performance. In the above processes, RMs can be employed not only to evaluate and score the reasoning steps and final answers, but also to generate valuable feedback on the reasoning process. Therefore, it is essential to propose a benchmark for evaluating the various aspects of the performance of RMs. However, as shown in Table 1, existing benchmarks predominantly focus on language domain, and benchmarks for VLRMs typically evaluate only a single capability (e.g., the quality comparison of multiple reasoning processes). These singular and inadequately challenging benchmarks fail to reveal the potential flaws of VLRMs, thereby limiting the development of RMs in the VL domain. + +To bridge this gap, we propose VLRMBench, a benchmark that is both comprehensive and challenging, specifically designed for the VL domain. We introduce three distinct data categories: mathematical reasoning, hallucination understanding, and multi-image understanding, en + +
# TasksStep Eval.?Test Size
MR-GSM8K [54]1✓2,999
RMBench [27]1✗1,327
CriticBench [24]1✗-
MathCheck-GSM [63]1✓516
MR-Ben [55]1✓5,975
ProcessBench [60]1✓3,400
PRMBench [37]9✓6,216
VLRewardBench* [20]1✗1,250
MultimodalRewardBench* [52]1✗5,211
VLRMBENCH*12✓12,634
+ +Table 1. Comparison between our VLRMBenchmark and others. * represents the benchmark for vision-language domain. 'Step Eval.' refers to whether the step-level labels are included. + +compassing eight widely used datasets. Moreover, to obtain high-quality samples containing reasoning processes, we develop a collaborative data filtering and generation pipeline. Specifically, we first utilize a small LVLM to filter samples along two dimensions: quality and difficulty. Subsequently, QVQ-72B-preview [39], which excels in VL reasoning, is employed to generate reasoning processes. Next, we use the advanced GPT-4o [2] to partition the reasoning steps and perform an initial validation of correctness. Finally, we apply rule-based filtering and manual inspection to correct errors in the reasoning process, yielding 1,000 samples that serve as the foundation for task construction. Based on these samples, we introduce three themes: step-based, outcome-based, and criticism-based, encompassing 12 diverse tasks and 12,634 questions. These tasks can comprehensively evaluate VLRMs in the aspects of process understanding, outcome judgment, and criticism generation, thus laying the groundwork for the development of VLRMs. + +We conducted extensive experiments on 21 open-source models (from 2B to 90B) and 5 closed-source models, such as GPT-4o and Claude-3.5-Sonnet [3], across 12 tasks. The results underscore the challenges of VLRMBench and demonstrate the narrowing gap between open-source and closed-source models. Specifically, in eight step-based tasks, even the advanced GPT-4o achieved only an average F1-Score of $62.4\%$ . In two outcome-based binary classification tasks, GPT-4o attained an average accuracy of only $66.3\%$ . Furthermore, in two criticism-based tasks, Qwen2.5-VL-72B [4] outperformed GPT-4o, with an average win rate of $53.3\%$ , showcasing the superior performance of state-of-the-art (SOTA) open-source models. In addition, we performed extensive analytical studies, providing valuable insights into the development of VLRMs. + +The main contributions can be summarized as follows: 1) We introduce VLRMBench, the first benchmark that integrates both comprehensiveness and challenge for the evaluation of VLRMs. VLRMBench is constructed from 1,000 high-quality samples curated through our collaborative data filtering and generation pipeline, covering areas such as + +mathematical reasoning, hallucination understanding, and multi-image understanding. 2) We designed 12 tasks, comprising a total of 12,634 carefully crafted questions, capable of evaluating models across three aspects: process understanding, outcome judgment, and critique generation. These diverse tasks enable targeted evaluations and reveal limitations of VLRMs. 3) We conducted extensive experiments on VLRMBench using 21 open-source models and 5 closed-source models, highlighting the challenges presented by VLRMBench and offering comprehensive analysis and guidance for the development of VLRMs. + +# 2. Related Works + +# 2.1. Reward Models (RMs) + +Techniques such as Reinforcement Learning with Human Feedback (RLHF) utilize RMs that offer feedback for large models, aligning them with human preferences in terms of accuracy, utility, and safety [5, 34]. RMs can generally be categorized into three types based on their objectives. Process-based RMs [31, 59] evaluate intermediate reasoning steps to ensure logical consistency and the correctness of each step. Outcome-based RMs [6, 25, 28] assess the quality of the final output. In contrast, Critique-based RMs [8, 18, 45] emphasize offering textual feedback or critiques aimed at enhancing the model's performance. However, RMs are not always reliable [37]. Therefore, it is imperative to establish a benchmark that facilitates a comprehensive evaluation of RMs. + +# 2.2. Benchmarks of Reward Models + +In recent years, several benchmarks for evaluating the capabilities of RMs have been proposed, driving the development of RMs in the field of natural language. For example, MR-GSM8k [54] redefines the model's role from a solver to a grader, necessitating the evaluation of the correctness of problem-solving steps. CriticBench [24] incorporates the Generate-Critique-Revise framework to assess the reasoning capabilities of models. ProcessBench [60] emphasizes step-by-step error detection in mathematical derivations, thereby increasing the granularity of the evaluation. PRMBench [37], based on PRM800K [23], employs GPT40 for data optimization and constructs a fine-grained evaluation dataset comprising 6,216 instances. PRMBench covers nine common error types and provides a comprehensive assessment of Process RMs. + +The aforementioned benchmarks predominantly focus on the language modality and do not adequately address the evaluation needs of VLLMs. To this end, [20] introduces VLRewardBench, a benchmark for VLRMs containing 1,250 test samples. However, VLRewardBench remains limited to Pairwise Comparison, wherein only a judgment of superiority or inferiority is made between two responses, + +![](images/f27dd8ff8b3bc74740d5f66cb7624ec797115139f0f0d344c67c9858bcdc82f4.jpg) +Figure 1. The collaborative data filtering and generation pipeline of VLRMBench. During the filtering stage, Qwen2VL-7B is tasked with answering questions under two conditions: without and with image input. Questions for which the model provides incorrect answers are retained. Subsequently, QVQ-72B-preview and GPT-4o are employed to generate reasoning processes and construct tasks for three themes. + +failing to incorporate more diverse and enriched task types. Consequently, we propose VLRMBench, which encompasses three core themes and twelve distinct tasks, providing a holistic evaluation of VLRMs' overall performance and establishing a foundation for their future advancement. + +# 3. VLRMBench + +In this section, we initially present an overview of the pipeline for constructing VLRMBench, along with the data statistics information, followed by an in-depth description of the task design. + +# 3.1. Overall pipeline of VLRMBench + +In Figure 1, we present a progressive and intuitive pipeline, which is designed to construct VLRMBench in three stages: data collection, filtering, and generation. During the collection phase, we integrated four mathematical, two hallucination, and two multi-image understanding datasets, totaling 16,550 samples and 19,343 images. In the filtering phase, to select high-quality and challenging samples, we introduced a joint filtering mechanism. Specifically, we initially employed Qwen2VL-7B [44] to answer questions without image information. If the model answered correctly, it indicated that the sample was of low quality [7]. Subsequently, we incorporated the image information into the question and instructed the model to answer it. If the model answered correctly, it suggested that the question was too simple to be a challenge. Consequently, we retained only those samples that could not be answered correctly in ei + +ther question-answering mode, resulting in a total of 6,715 samples, which formed our basic database $D_{b}$ + +In the generation phase, we utilized QVQ-72B-preview [39] to generate reasoning process for the samples in $D_{b}$ , discarding those where QVQ-72B-preview yielded incorrect answers, as verified via comparison with the standard answers. Subsequently, GPT-4o was employed to perform step-level segmentation of QVQ-72B-preview's reasoning process, grouping semantically similar and logically related sentences into the one step. Following this, we obtained 3,335 samples with reasoning segmented at the step level. Furthermore, based on specific criteria², we excluded samples with excessively long or short reasoning steps and ensured a balanced distribution across the three sample types. Finally, to ensure the quality of the samples, a manual review was conducted. Specifically, three PhD volunteers were recruited to verify the high-quality reasoning processes generated by QVQ-72B-preview. If any volunteer identified an error in the reasoning process, all three volunteers would collaboratively review the sample and correct any potential mistakes. Additionally, examples of manual corrections applied to the samples could be found in the Appendix. After completing the three aforementioned stages, 1,000 high-quality samples were retained for constructing the specific task, denoted as $D$ . For more detailed statistical information, please refer to Table 2 and Figure 2, which provide a description of the sample sources, the average word count, and the number of reasoning steps. + +
CategoryDataset#Samples#AvgWORDS
MathMathVista [29]142696
MathVision [42]281,021
MathVerse [58]126742
DynaMath [64]251691
HallucinationIllusionVQA [36]62736
HallusionBench [15]334605
Multi-ImageBLINK [14]23790
Mantis-Eval [16]34592
+ +Table 2. Statistics on the sample sources of VLRMBench. We also counted the average number of words in the reasoning process under different data sources. + +# 3.2. Step-based tasks + +The step-based theme consists of 8 tasks designed to assess the reasoning process understanding ability of VLRMs. Under this theme, models will be provided with a modified reasoning process and are expected to output a sequence of 0s and 1s to indicate whether each reasoning step contains deviated content. + +Step correctness (SC) For a Vision-Language Reward Model (VLRM), detecting errors at each step of the reasoning process is a critical capability [23, 37]. Building on this, we propose the SC task. Specifically, given the reasoning process $R_{n}$ of a sample from $D$ , where $n$ denotes the total number of reasoning steps, we inject errors into $m$ of the $n$ steps $(0 < m < n)$ of $R_{n}$ using the advanced GPT-4o, to simulate potential errors during reasoning. + +Redundancy Detection (RD) Existing reasoning models often produce redundant information when answering questions, leading to an increase in output token length, which is both inefficient and environmentally unfriendly. To this end, we propose the RD task, which requires VLRMs to identify redundant information generated during the reasoning process. Specifically, based on $D$ , we introduce redundant descriptions in certain reasoning steps of each sample. For instance, we incorporate descriptions from $R_{i-j}$ into $R_i$ to intentionally create redundancy. Furthermore, we employ GPT-4 to generate paraphrased versions of the content from $R_{i-j}$ and incorporate them into $R_i$ , thereby amplifying redundancy. Unlike the SC task, the introduction of redundant information does not compromise the correctness of the final answer. + +Confidence Misdirection (CM) When confident and assertive terms, such as 'definitely' and 'without a doubt', are present in an incorrect reasoning step, is it possible for VL-RMs to be misled into perceiving the erroneous reasoning as correct? To investigate this issue, we propose the CM task, + +![](images/ebfe9977b3db29d7faba65e6419e5c8b59427bdfdc4136cbc90d453078734927.jpg) +Figure 2. The distribution of the number of reasoning steps for different data sources. + +which is derived from the SC task. Specifically, for each sample in the SC task, we incorporate confidence-boosting terms into the erroneous reasoning step to mislead VLRMs into perceiving the step as correct, thereby evaluating the robustness of VLRMs. + +Existential Hallucination (EH) and Attribute Hallucination (AH) When answering questions using LVLMs, hallucination phenomena often arise [26]. Although other parts of the reasoning process may be correct, the introduction of hallucinatory errors or the omission of key entities can lead to biased final results. To evaluate whether VLRMs are capable of detecting hallucinations introduced during the reasoning process, we have designed two tasks on the top of $D$ . The first task, EH, assesses the ability of the VLRM to identify entities that were incorrectly introduced during reasoning process but do not exist in the image, or to detect elements that may have been overlooked, such as objects, people, etc. The second task, AH, examines the VLRM's capability to accurately identify the attributes of entities, particularly in cases where LVLMs wrongly attribute properties to an entity that it does not possess or when attribute descriptions conflict with the visual content of the image. Such errors are typically reflected in misjudgments regarding object attributes, such as size, color, or shape. VLRMs must integrate image information effectively to accurately identify hallucination errors during the reasoning process. + +Detail Error (DE) When addressing complex problems, the reasoning process often involves multiple details, particularly in the handling of numbers and symbols. Even if the overall reasoning appears to be sound, there may still be hidden errors in finer details (e.g., numerical errors in calculations or symbolic mistakes in formulas). These errors are often difficult to detect, yet they can affect the precision of the final results. The introduction of DE tasks serves to evaluate the sensitivity of VLRMs to reasoning details, as + +well as their ability to identify fine-grained errors. + +Spatial Relationship (SR) During visual reasoning, LVLMs may incorrectly describe the spatial relationships between objects or entities, leading to erroneous results. For instance, LVLMs might misdescribe an object located at the top of the image as being at the bottom. Therefore, we select samples containing spatial relationship descriptions from $D$ to construct the SR task, aiming to evaluate VLRMs' understanding of spatial relationships in a image and their ability to detect spatial errors in a reasoning process. + +Image Confusion (IC) In multi-image tasks, LVLMs are required to extract and associate relevant information from multiple images, which typically involves identifying various entities and their interrelations across different images. However, due to the presence of similar elements or backgrounds between images, LVLMs may erroneously associate an entity from one image with another image. This 'image reference error' not only causes confusion but also potentially impacts the accuracy of subsequent reasoning. Thus, we filter samples related to multi-image understanding from $D$ to construct the IC task, which aims to evaluate whether VLRMs can identify image reference errors during the reasoning process. + +# 3.3. Outcome-based tasks + +The outcome-based theme encompasses two binary classification tasks, aiming to evaluate the ability of VLRMs to assess and predict the accuracy of outcomes. + +Multi-solution Judgment (MJ) During the training stage, preference optimization techniques, such as DPO [35], have become a widely used paradigm. These methods typically rely on the collection of preference samples. VLRMs are frequently employed to assess the quality of preference samples, thereby becoming a crucial component of preference optimization [21, 53]. To this end, we introduce the MJ task, which aims to investigate the discriminative ability of VLRMs regarding different reasoning processes for the same problem. This task is discussed as a core element in VLRewardBench [20] and Multimodal Reward-Bench [52]. Specifically, for the reasoning process $R$ of any question $Q$ in $D$ , we introduce the reasoning process $R'$ of the same question in the SC, RD and CM tasks, thereby constructing a triplet $(Q, R, R')$ . VLRMs are required to judge the quality of $R$ and $R'$ and score them accordingly. + +Forecasting Future (FF) When LVLMs perform inference, test-time scaling (TTS) can further improve their performance [47, 51]. However, these techniques often introduce additional inference time, primarily due to the necessity of exploring multiple reasoning paths. If the correctness + +of the final answer could be predicted based on the reasoning process of the first m steps, $R_{[0:m]}$ , this prediction could substantially expedite the TTS process. Therefore, the FF task is proposed, wherein VLRMs are required to assess the correctness of the final answer based on $R_{[0:m]}$ . + +# 3.4. Criticism-based tasks + +Error Reason Analysis (ERA) and Error Correction (EC) Unlike Process RM and Outcome RM, which focus solely on evaluating processes or outcomes, Critique RM emphasizes the critique of the reasoning process, providing textual feedback that includes error analysis and correction. Consequently, we propose the ERA and EC tasks to assess the capabilities of VLRMs in these two aspects. In the ERA task, VLRMs are presented with erroneous reasoning steps derived from the SC task, and are tasked with identifying the error and analyzing its underlying cause. In contrast, the EC task does not require error cause analysis; rather, VLRMs directly correct the identified error and produce the corrected reasoning process. + +# 4. Experiments + +# 4.1. Models + +We conducted extensive experiments on 21 open-source models and 5 closed-source models. Specifically, the open-source models, which range from 2B to 90B, including Llama3.2 (11B/90B) [12], Llava-OneVision (7B/72B) [19], Qwen2VL (2B/7B/72B) [44], Qwen2.5VL (3B/7B/72B) [4], QVQ-72B-preview [39], Ovis2 (8B/16B/34B) [30], InternVL2.5 (2B/8B/38B/78B) [9], Phi-3.5-Vision (4.2B) [1], MiniCPM-o-2.6 (8B), and MiniCPM-V-2.6 (8B) [33]. For the closed-source models, we employed GLM-4V-Flash, GPT-4o [2], Gemini-2.0-Flash-/Flash-thinking-exp-1219 [38], and Claude-3.5-Sonnet [3]. + +# 4.2. Evaluation details + +We follow the 'LVLM-as-a-Judge' paradigm and introduce different question-answering templates to assess the models [20, 52]. Specifically, for the SC, RD, CM, EH, AH, DE, SR, and IC tasks, the model takes a reasoning process as input and outputs a judgment regarding whether each reasoning step contains redundancies or other errors. For these eight tasks, we employ the weighted F1-Score as the evaluation metric, as defined in Equation 1, which better addresses class imbalance issues and ensures the model's performance on the minority class is adequately considered. For the MJ task, the model is provided with (good and bad) reasoning process pairs for the same question and outputs scores for both. If the model assigns a higher score to the good sample, it is deemed correct; otherwise, it is considered incorrect. For the FF task, the model takes several preceding steps of the reasoning process as input and outputs a judgment on + +
ModelStep (5,722)Outcome (5,402)Criticism (1,510)
SCRDCMEHAHDESRICW-Avg.MJFFW-Avg.ERAwERAτERALECwECTECL
# Sample99899799494694465912658-2,9862,416-974536
Open-source LVLMs < 10B (Small group)
InternVL2.5-2B33.834.130.022.423.833.519.930.129.226.660.841.94.86.788.52.310.087.6
Qwen2VL-2B28.326.220.917.819.126.623.826.423.010.559.632.50.20.199.72.55.891.7
Qwen2.5VL-3B29.932.723.227.226.945.525.142.630.131.967.147.67.112.180.84.821.273.9
Phi-3.5-Vision (4.2B)35.938.827.920.721.133.417.731.629.212.165.235.83.89.087.24.222.473.4
Qwen2VL-7B38.029.332.021.521.432.821.621.128.852.064.757.73.98.187.95.423.271.4
Qwen2.5VL-7B43.433.237.822.823.945.515.629.133.426.070.746.037.722.040.39.351.938.8
Llava-OneVision-7B31.825.721.516.017.224.816.120.722.618.057.835.83.910.485.77.134.458.5
InternVL2.5-8B36.628.431.121.921.236.515.124.128.661.473.866.934.521.943.66.236.157.7
Ovis2-8B47.728.237.825.225.739.817.025.833.331.465.346.623.225.950.914.363.522.2
MiniCPM-V-2.6 (8B)44.938.336.926.126.536.824.230.634.637.962.548.95.411.782.93.424.771.9
MiniCPM-o-2.6 (8B)41.035.337.329.629.036.728.330.134.615.064.337.09.219.870.97.633.958.4
Avg.37.431.830.622.823.335.620.428.429.829.364.745.212.213.474.46.129.764.1
10B < Open-source LVLMs < 40B (Middle group)
Llama3.2-11B33.927.129.720.220.031.917.125.826.723.961.140.54.210.385.55.434.759.8
Ovis2-16B48.547.947.940.635.844.521.928.443.649.769.258.430.926.642.510.469.320.3
Ovis2-34B65.351.164.554.551.659.630.449.057.046.375.759.445.926.627.515.476.48.1
InternVL2.5-38B61.049.662.655.051.761.434.648.156.145.179.360.421.124.554.412.556.630.9
Avg.52.243.951.242.639.849.426.037.845.941.371.354.725.522.052.510.959.329.8
Open-source LVLMs > 40B (Large group)
Llava-OneVision-72B49.430.650.237.934.145.113.926.540.432.269.749.015.322.162.69.451.039.6
Qwen2VL-72B55.540.256.148.941.656.422.443.648.837.772.153.122.420.457.17.757.135.1
Qwen2.5VL-72B72.841.770.464.659.972.437.963.362.665.680.272.174.115.110.815.677.07.3
QVQ-72B-preview42.435.336.329.427.137.820.826.134.349.238.444.44.210.085.89.451.039.6
InternVL2.5-78B62.142.062.853.348.760.428.547.254.048.678.862.154.224.920.910.863.925.3
Llama3.2-90B40.838.841.334.229.435.417.021.136.252.769.060.09.918.671.49.556.833.7
Avg.53.838.152.944.740.151.323.438.046.147.768.056.830.018.551.410.459.530.1
Closed-source LVLMs
Claude-3.5-Sonnet70.853.765.763.962.863.440.865.762.982.275.179.060.625.513.921.253.924.9
GPT-4o73.750.666.657.658.671.848.264.662.458.476.066.30.0100.00.00.0100.00.0
GLM-4V-Flash37.233.532.928.227.741.822.920.232.837.266.750.434.224.341.56.954.139.0
Gemini-2.0-Flash69.355.966.165.662.363.446.867.863.454.277.164.432.822.544.77.152.540.3
Gemini-2.0-FTE.-121967.955.565.765.962.363.246.467.163.153.177.063.837.422.640.06.850.243.1
Avg.63.849.859.456.254.760.741.057.157.057.074.464.833.039.028.08.462.129.5
+ +Table 3. Results on the VLRMBench. For Step-based tasks, the F1-Score is the metric. For Outcome-based tasks, Accuracy is utilized as the metric. In the case of Criticism-based tasks, the metric is Win Rate, expressed as Win: Tie: Lose (Simplified as $W$ , $T$ and $L$ ). Note that W-Avg. represents the weighted average. Evaluation prompt templates and some case studies could be found in the Appendix B and C. + +whether a correct answer can ultimately be derived. Consequently, for the MJ and FF tasks, Accuracy is employed as the evaluation metric. For the ERA and EC tasks, the model's output is a description of the error reasons in the reasoning process and the corrected reasoning process. We utilize the output of GPT-4o as the baseline and adopt Win Rate [13, 61] as the evaluation metric. + +We configured the max new tokens to 4,096 across all tasks for the five closed-source models and QVQ-72B-preview. In contrast, for the other models, we set the max new tokens to 4,096 specifically for the ERA and EC tasks, while fixing it at 256 for the remaining tasks. To ensure the reproducibility of the experimental results, we set the temperature coefficient to 0 for all models. All experiments were conducted on 8 NVIDIA H800 GPUs. + +$$ +F 1 = \frac {N _ {\text {n e g}}}{N _ {\text {p o s}} + N _ {\text {n e g}}} \cdot F 1 _ {\text {p o s}} + \frac {N _ {\text {p o s}}}{N _ {\text {p o s}} + N _ {\text {n e g}}} \cdot F 1 _ {\text {n e g}} \tag {1} +$$ + +where $N_{\mathrm{pos}}$ and $N_{\mathrm{neg}}$ represent the number of correct and + +incorrect steps in the reasoning process, respectively. + +# 4.3. Main Results + +The experimental results on VLRMBench are shown in Table 3, leading to the following conclusions: + +1) Among open-source models, surprisingly, the performance of Qwen2.5VL-72B is comparable to that of advanced closed-source models. For instance, across the eight Step-based tasks, Qwen2.5VL-72B achieved an average F1-Score of $62.6\%$ , surpassing GPT-4o by $0.2\%$ and only slightly trailing Claude-3.5-Sonnet by $0.3\%$ . In the MJ and FF tasks, Qwen2.5VL-72B also outperformed GPT-4o by $7.2\%$ and $4.2\%$ , respectively. Notably, in the ERA task, Qwen2.5VL-72B attained a win rate of $74.1\%$ over GPT-4o, significantly outperforming Claude-3.5-Sonnet at $60.6\%$ . + +2) LVLMs that have not undergone specific training show diminishing performance gains as the model size increases. Specifically, for LVLMs in the Middle group, compared to those in the Small group, the performance gain in Step-based and Outcome-based tasks was $54.0\%$ and + +
OrderModelAvg. ScoreAcc
Better firstQwen2.5VL-3B[9.0, 8.3]23.7
Llama3.2-11B[8.6, 4.8]52.8
InternVL2.5-78B[8.4, 7.4]38.6
GPT-4o[9.8, 7.8]52.2
Better secondQwen2.5VL-3B[8.7, 9.0]21.4
Llama3.2-11B[8.3, 5.6]8.4
InternVL2.5-78B[8.1, 9.1]45.6
GPT-4o[7.4, 8.1]43.1
+ +21.0%, respectively. In contrast, the corresponding performance gains in the Large group relative to the Middle group were only 0.4% and 3.8%. Even excluding QVQ-72B-preview from the Large group, the relative performance improvement remains limited, at 5.4% and 8.4%. This suggests that blindly increasing the size of VLLMs does not lead to significant performance improvements. Consequently, we recommend that future research prioritize training VLRMs at the Middle group scale initially. + +3) The Outcome-based tasks, being simple binary classification problems, remain challenging for most open-source models, and even closed-source models perform only marginally better than random guessing. For open-source models across three size groups, the average accuracy for Outcome-based tasks is around $50\%$ , similar to random guessing. For advanced closed-source models, the average accuracy only exceeds random guessing by $14.8\%$ . +4) Deterministic vocabulary and hallucinated information can effectively mislead the VLRMs, reducing its ability to understand reasoning processes. The CM task, based on the SC task, introduces deterministic vocabulary (e.g., must, definitely) into sentences with erroneous reasoning steps, while EH and AH tasks introduce hallucinated information during the reasoning process. Existing models, particularly in the Small group, demonstrate a degradation in F1-Score of $6.8\%$ , $14.6\%$ , and $14.1\%$ for CM, EH, and AH tasks, respectively, when compared to the SC task. This performance degradation remains significant even as model size increases. +5) Existing models show strong robustness to numerical errors but limited resilience to perturbations in spatial relationships and image references. We observed that for Large group, the F1-Score in the DE task only decreased by $2.5\%$ compared to SC. However, for the SR and IC tasks, the performance dropped by $30.4\%$ and $15.8\%$ , respectively. Overall, our VLRMBench provide unique challenges that surpass traditional benchmarks, with even the most advanced models achieving only mediocre performance. + +Table 4. The results on the MJ task with different sample order. + +
Model≤10 steps10-15 steps>15 steps
Qwen2.5VL-3B35.630.125.6
Llama3.2-11B42.932.828.4
InternVL2.5-78B68.264.154.1
GPT-4o77.875.268.6
+ +Table 5. Results on the SC task with different reasoning steps. + +# 5. Further Analysis + +# 5.1. Is the o1-like model suitable to be a VLRM? + +QVQ-72B-preview [39], an open-source visual-language model, exhibits strong reasoning capabilities. However, it exhibits a potential drawback regarding the degradation of instruction-following ability. For instance, in the SC task, by incorporating instruction prompts, the sample's original question, and the reasoning process as input, we expect the VLRM to evaluate each reasoning step. However, QVQ-72B-preview often disregards the given instructions and directly answers the original question, leading to suboptimal performance on VLRMBench, with additional cases provided in the Appendix C. Therefore, as an o1-like model, it must not only demonstrate strong reasoning abilities but also exhibit robust instruction-following capabilities to be considered a viable candidate for a VLRM. + +# 5.2.Impact of sample order + +In the MJ task, VLRM scores two reasoning processes for the same question. To further investigate the positional bias, we employed two different sample orders, as shown in Table 4. Specifically, 'Better first' places the superior reasoning process in the first position, while 'Better second' places it in the second position. The results indicate that Llama3.2-11B consistently assigns a higher score to the reasoning process in the first position, irrespective of its quality. In contrast, the other three models exhibit the ability to adjust their scores based on the quality of the reasoning. This positional bias phenomenon is also observed in [43]. As a result, in Table 3, we utilized the sum of the two scores for the MJ task evaluation to ensure fairness. + +# 5.3. Performance on various reasoning length + +We further investigate the impact of the number of reasoning steps on the VLRMs' process understanding. As shown in Table 5, we classified the number of reasoning steps in the SC task into three categories: the short group ( $\leq 10$ steps), with 263 samples; the middle group (10-15 steps), with 468 samples; and the long group ( $>15$ steps), with 267 samples. Specifically, for all four models, a decrease in F1-score was observed as the number of reasoning steps increased. Notably, for the Llama3.2-11B model, the difference between the short and long groups reached $14.5\%$ , and + +
Model0-shot1-shot2-shot
Qwen2.5VL-3B29.931.531.1
Llama3.2-11B33.942.242.3
InternVL2.5-78B62.155.955.7
GPT-4o73.771.770.5
+ +![](images/739e169c1130866e94dbc8225caf0d71bb5a71df2711e89764cecc05ce3991a8.jpg) +Figure 3. Performance changes with the test-time scaling. + +even for the advanced GPT-4o, the F1-Score dropped from $77.8\%$ to $68.6\%$ . This suggests that current models are less capable of handling long information, underscoring the necessity of enhancing their ability to tackle long reasoning process when training VLRMs in the future. + +# 5.4. Impacts of In-Context Learning (ICL) setting + +In main experiments, we employ zero-shot setting by default. Furthermore, using the SC task as an example, we explore the effect of ICL on VLRMs. As presented in Table 6, the results indicate that providing contextual examples is more beneficial for small models but harms the performance of large models. Specifically, with the help of one-shot learning, Qwen2.5VL-3B's performance improved from $29.9\%$ to $31.5\%$ , and Llama3.2-11B saw an improvement of $8.3\%$ . However, for GPT-4o, one-shot learning led to a $2.0\%$ performance degradation, and InternVL2.5-78B's performance decreased by $6.2\%$ . One possible explanation for this is that small models exhibit relatively weaker instruction-following capabilities, and ICL examples may help guide them in responding to queries in the designated format. In contrast, large models, which already possess strong comprehension abilities and adaptability to complex tasks, may find that providing contextual examples interferes with their inherent understanding process, resulting in performance decline. Thus, in the SC task, smaller models exhibit a greater reliance on ICL examples, whereas larger models tend to rely more on their pre-trained knowledge. + +# 5.5. Impacts of test-time scaling + +We investigated the impact of test-time scaling (i.e. voting method) for VLRMs on the MJ task. To ensure the diversity of the generated results, we set the temperature coefficient + +Table 6. Results on SC task with various ICL few-shot numbers. + +
LVLMVLRMMathVistaMathVerseMathVisionAvg.
7B7B10.45.84.77.0
72B7B11.84.23.26.4
7B72B21.819.721.521.0
72B72B11.79.311.410.8
+ +Table 7. Results of the correction rate on math benchmarks. + +to 0.9 and performed K sampling. As illustrated in Figure 3, surprisingly, as the value of K increases, the accuracy of Qwen2.5VL-3B improved by $12.4\%$ . Notably, at $\mathrm{K} = 9$ , its performance was only $4.8\%$ lower than that of GPT-4o. These findings highlight the substantial effectiveness of test-time scaling techniques in enhancing VLRMs. + +# 5.6. Improve LVLMs with a VLRM + +To demonstrate the significant role of VLRMs in the reasoning process, as presented in Table 7, we conducted experiments on three mathematical benchmarks using Qwen2.5VL-7B/72B. For example, in the second row of Table 7, Qwen2.5VL-72B was employed as the LVLM to answer the mathematical questions, whereas Qwen2.5VL-7B was used as the VLRM to perform error reason analysis of the LVLM's reasoning process. The analysis was subsequently fed back into the LVLM as feedback, allowing it to generate a second round of answers. Besides, we defined the correction rate, which is the ratio of the number of questions correctly answered in the second round after feedback to the number of questions incorrectly answered in the first round. The experimental results demonstrate that VLRM can provide valuable information to the LVLM through error analysis, thereby facilitating mistake correction, with an average gain of $11.3\%$ . These findings suggest that even LVLMs not trained on proprietary data can act as VLRMs, providing more correction possibilities during testing based on this intuitive feedback mechanism. This further underscores the potential value of developing efficient VLRMs. + +# 6. Conclusions + +In this paper, we propose a comprehensive and challenging benchmark, named VLRMBench. This benchmark covers 3 themes and 12 tasks, comprising a total of 12,634 questions. Following the 'LVLM-as-a-Judge' paradigm, we conduct systematic experiments and analyses on VLRMBench using 26 models, deriving several key insights: enhancing the performance of VLRMs in long-text comprehension, exploring various feedback patterns from VLRMs to LVLMs, and developing the application potential of proprietary VLRMs, etc. We hope that VLRMBench will serve as a standardized evaluation platform to propel the development of more efficient VLRMs. In future work, we plan to incorporate larger datasets specifically designed to train VLRMs. + +# 7. Acknowledgements + +This work was supported by the National Natural Science Foundation of China (Grant No. 61977045). + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5 +[3] Anthropic. Claude 3.5: Sonnet, 2025. Accessed: 2025-02-25. 2, 5 +[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 5 +[5] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2 +[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2 +[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3 +[8] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2 +[9] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5 +[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 1 +[11] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 3 +[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil + +Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5 +[13] Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B Hashimoto. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475, 2024. 6 +[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision, pages 148-166. Springer, 2024. 4 +[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 4 +[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. Transactions on Machine Learning Research, 2024, 2024. 4 +[17] Wasif Khan, Seowung Leem, Kyle B See, Joshua K Wong, Shaoting Zhang, and Ruogu Fang. A comprehensive survey of foundation models in medicine. IEEE Reviews in Biomedical Engineering, 2025. 1 +[18] Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, et al. Solar 10.7 b: Scaling large language models with simple yet effective depth up-scaling. arXiv preprint arXiv:2312.15166, 2023. 2 +[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5 +[20] Lei Li, Yuancheng Wei, Zhihui Xie, Xuqing Yang, Yifan Song, Peiyi Wang, Chenxin An, Tianyu Liu, Sujian Li, Bill Yuchen Lin, et al. Vlrewardbench: A challenging benchmark for vision-language generative reward models. arXiv preprint arXiv:2411.17451, 2024. 2, 5 +[21] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. arXiv preprint arXiv:2410.09421, 2024. 1, 5 +[22] Zongxia Li, Xiyang Wu, Hongyang Du, Huy Nghiem, and Guangyao Shi. Benchmark evaluations, applications, and challenges of large vision language models: A survey. arXiv preprint arXiv:2501.02189, 2025. 1 +[23] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2023. 2, 4 +[24] Zicheng Lin, Zhibin Gou, Tian Liang, Ruilin Luo, Haowei Liu, and Yujiu Yang. Criticbench: Benchmarking lms for + +critique-correct reasoning. arXiv preprint arXiv:2402.14809, 2024. 2 +[25] Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451, 2024. 2 +[26] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xi-tian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 4 +[27] Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. Rm-bench: Benchmarking reward models of language models with subtlety and style. arXiv preprint arXiv:2410.16184, 2024. 2 +[28] Xingzhou Lou, Dong Yan, Wei Shen, Yuzi Yan, Jian Xie, and Junge Zhang. Uncertainty-aware reward model: Teaching reward models to know what is unknown. arXiv preprint arXiv:2410.00847, 2024. 2 +[29] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 4 +[30] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 5 +[31] Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Yang You, and Hongxia Yang. Let's reward step by step: Step-level reward model as the navigators for reasoning. arXiv preprint arXiv:2310.10080, 2023. 2 +[32] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 1 +[33] OpenBMB. Minicpm o: A gpt-4o-level mllm for vision, speech, and multimodal live streaming on your phone, 2025. Accessed: 2025-02-25. 5 +[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2 +[35] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 5 +[36] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 4 +[37] Mingyang Song, Zhaochen Su, Xiaoye Qu, Jiawei Zhou, and Yu Cheng. Prmbench: A fine-grained and challenging benchmark for process-level reward models. arXiv preprint arXiv:2501.03124, 2025. 2, 4 +[38] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, + +Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5 +[39] Qwen Team. Qvq: To see the world with wisdom, 2024. 2, 3, 5, 7 +[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1 +[41] Jiahao Wang, Bolin Zhang, Qianlong Du, Jiajun Zhang, and Dianhui Chu. A survey on data selection for llm instruction tuning. arXiv preprint arXiv:2402.05123, 2024. 1 +[42] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 4 +[43] Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926, 2023. 7 +[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5 +[45] Tianlu Wang, Ilia Kulikov, Olga Golovneva, Ping Yu, Weizhe Yuan, Jane Dwivedi-Yu, Richard Yuanzhe Pang, Maryam Fazel-Zarandi, Jason Weston, and Xian Li. Self-taught evaluators. arXiv preprint arXiv:2408.02666, 2024. 2 +[46] Zhilin Wang, Alexander Bukharin, Olivier Delalleau, Daniel Egert, Gerald Shen, Jiaqi Zeng, Oleksii Kuchaiev, and Yi Dong. Helpsteer2-preference: Complementing ratings with preferences. arXiv preprint arXiv:2410.01257, 2024. 1 +[47] Jinyang Wu, Mingkuan Feng, Shuai Zhang, Ruihan Jin, Feihu Che, Zengqi Wen, and Jianhua Tao. Boosting multimodal reasoning with mcts-automated structured thinking. arXiv preprint arXiv:2502.02339, 2025. 1, 5 +[48] Wang Xiyao, Yang Zhengyuan, Li Linjie, Lu Hongjin, Xu Yuancheng, Lin Chung-Ching Lin, Lin Kevin, Huang Furong, and Wang Lijuan. Scaling inference-time search with vision value model for improved visual comprehension. arXiv preprint arXiv:2412.03704, 2024. 1 +[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1 +[50] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 1 +[51] Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte + +carlo tree search. arXiv preprint arXiv:2412.18319, 2024. 1,5 +[52] Michihiro Yasunaga, Luke Zettlemoyer, and Marjan Ghazvininejad. Multimodal rewardbench: Holistic evaluation of reward models for vision language models. arXiv preprint arXiv:2502.14191, 2025. 2, 5 +[53] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 1, 5 +[54] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 2 +[55] Zhongshen Zeng, Yinhong Liu, Yingjia Wan, Jingyao Li, Pengguang Chen, Jianbo Dai, Yuxuan Yao, Rongwu Xu, Zehan Qi, Wanru Zhao, et al. Mr-ben: A meta-reasoning benchmark for evaluating system-2 thinking in llms. arXiv preprint arXiv:2406.13975, 2024. 2 +[56] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Photogrammetry and Remote Sensing, 221:64-77, 2025. 1 +[57] Di Zhang, Jingdi Lei, Junxian Li, Xunzhi Wang, Yujie Liu, Zonglin Yang, Jiatong Li, Weida Wang, Suorong Yang, Jianbo Wu, et al. Critic-v: Vlm critics help catch vlm errors in multimodal reasoning. arXiv preprint arXiv:2411.18203, 2024. 1 +[58] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Springer, 2024. 4 +[59] Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301, 2025. 2 +[60] Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559, 2024. 2 +[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6 +[62] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 1 +[63] Zihao Zhou, Shudong Liu, Maizhen Ning, Wei Liu, Jindong Wang, Derek F Wong, Xiaowei Huang, Qiufeng Wang, and Kaizhu Huang. Is your model really a good math rea + +soner? evaluating mathematical reasoning with checklist. arXiv preprint arXiv:2407.08733, 2024. 2 +[64] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 4 \ No newline at end of file diff --git a/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/images.zip b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..24bcd0e035043d9c235611d78b84d827dcf7a829 --- /dev/null +++ b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47a40ab44f63e126882815e8167375379824e8b8ff5c756c4a8f66b72fcb37c1 +size 590589 diff --git a/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/layout.json b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9c6e51715d445e234e5b9ef42c2ba55ac3b83173 --- /dev/null +++ b/2025/VLRMBench_ A Comprehensive and Challenging Benchmark for Vision-Language Reward Models/layout.json @@ -0,0 +1,8140 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 103, + 102, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 102, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 103, + 102, + 507, + 140 + ], + "type": "text", + "content": "VLRMBench: A Comprehensive and Challenging Benchmark for Vision-Language Reward Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "spans": [ + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": "Jiacheng Ruan" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Wenzhen Yuan" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Xian Gao" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Ye Guo" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Daoxin Zhang" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Zhe Xu" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Yao Hu" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Ting Liu" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": " Yuzhuo Fu" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": "Shanghai Jiao Tong University \n" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 142, + 160, + 473, + 217 + ], + "type": "text", + "content": "Xiaohongshu Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 220, + 369, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 220, + 369, + 231 + ], + "spans": [ + { + "bbox": [ + 238, + 220, + 369, + 231 + ], + "type": "text", + "content": "jackchenruan@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "type": "text", + "content": "Although large visual-language models (LVLMs) have demonstrated strong performance in multimodal tasks, errors may occasionally arise due to biases during the reasoning process. Recently, reward models (RMs) have become increasingly pivotal in the reasoning process. Specifically, process RMs evaluate each reasoning step, outcome RMs focus on the assessment of reasoning results, and critique RMs perform error analysis on the entire reasoning process, followed by corrections. However, existing benchmarks for vision-language RMs (VLRMs) typically assess only a single aspect of their capabilities (e.g., distinguishing between two answers), thus limiting the all-round evaluation and restricting the development of RMs in the visual-language domain. To address this gap, we propose a comprehensive and challenging benchmark, dubbed as VLRMBench, encompassing 12,634 questions. VLRMBench is constructed based on three distinct types of datasets, covering mathematical reasoning, hallucination understanding, and multi-image understanding. We design 12 tasks across three major categories, focusing on evaluating VLRMs in the aspects of process understanding, outcome judgment, and critique generation. Extensive experiments are conducted on 21 open-source models and 5 advanced closed-source models, highlighting the challenges posed by VLRMBench. For instance, in the 'Forecasting Future', a binary classification task, the advanced GPT-4o achieves only a " + }, + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "type": "inline_equation", + "content": "76.0\\%" + }, + { + "bbox": [ + 55, + 283, + 297, + 620 + ], + "type": "text", + "content": " accuracy. The code is available at https://github.com/JCruan519/VLRMBench." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 641, + 135, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 641, + 135, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 641, + 135, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 662, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 686 + ], + "type": "text", + "content": "Recently, driven by the development of large language models (LLMs) [32, 40, 49, 50], large visual-language models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 259, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 380 + ], + "type": "text", + "content": "(LVLMs) have also achieved notable progress [22]. They have been broadly applied across domains such as medical imaging [17], remote sensing [56], and autonomous driving [62]. However, despite these remarkable achievements, they still encounter limitations in complex vision-language (VL) reasoning tasks, primarily due to insufficient reasoning depth and the absence of self-correction mechanisms [10, 51]. To mitigate this issue, reward models (RMs) have been introduced to detect errors in model responses, thereby enhancing the accuracy and efficiency of LVLMs [57]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 384, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 649 + ], + "type": "text", + "content": "During both the training and inference phases, RMs play a pivotal role. Firstly, prior to training, RMs can be employed to filter high-quality samples and facilitate the construction of automated data synthesis pipelines [21, 41, 46]. Secondly, during training, preference optimization techniques (e.g., RLAIF [53]) align the model's preferences with RMs, thereby enhancing the quality of the generated content. Finally, during inference, test-time scaling techniques, aided by RMs, are widely applied [47, 48], trading additional inference time for improved model performance. In the above processes, RMs can be employed not only to evaluate and score the reasoning steps and final answers, but also to generate valuable feedback on the reasoning process. Therefore, it is essential to propose a benchmark for evaluating the various aspects of the performance of RMs. However, as shown in Table 1, existing benchmarks predominantly focus on language domain, and benchmarks for VLRMs typically evaluate only a single capability (e.g., the quality comparison of multiple reasoning processes). These singular and inadequately challenging benchmarks fail to reveal the potential flaws of VLRMs, thereby limiting the development of RMs in the VL domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "To bridge this gap, we propose VLRMBench, a benchmark that is both comprehensive and challenging, specifically designed for the VL domain. We introduce three distinct data categories: mathematical reasoning, hallucination understanding, and multi-image understanding, en" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 256, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 256, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 256, + 703 + ], + "type": "text", + "content": "* Jiacheng Ruan and Wenzhen Yuan are equal contributors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 198, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 198, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 198, + 713 + ], + "type": "text", + "content": "Yuzhuo Fu is the corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3163" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 288, + 175 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 288, + 175 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 288, + 175 + ], + "type": "table", + "html": "
# TasksStep Eval.?Test Size
MR-GSM8K [54]1✓2,999
RMBench [27]1✗1,327
CriticBench [24]1✗-
MathCheck-GSM [63]1✓516
MR-Ben [55]1✓5,975
ProcessBench [60]1✓3,400
PRMBench [37]9✓6,216
VLRewardBench* [20]1✗1,250
MultimodalRewardBench* [52]1✗5,211
VLRMBENCH*12✓12,634
", + "image_path": "2e44c9fb179025a95f3a56486c9044da5d8e8d758b211b755c607192e544111e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 295, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 216 + ], + "type": "text", + "content": "Table 1. Comparison between our VLRMBenchmark and others. * represents the benchmark for vision-language domain. 'Step Eval.' refers to whether the step-level labels are included." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 242, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 242, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 242, + 295, + 456 + ], + "type": "text", + "content": "compassing eight widely used datasets. Moreover, to obtain high-quality samples containing reasoning processes, we develop a collaborative data filtering and generation pipeline. Specifically, we first utilize a small LVLM to filter samples along two dimensions: quality and difficulty. Subsequently, QVQ-72B-preview [39], which excels in VL reasoning, is employed to generate reasoning processes. Next, we use the advanced GPT-4o [2] to partition the reasoning steps and perform an initial validation of correctness. Finally, we apply rule-based filtering and manual inspection to correct errors in the reasoning process, yielding 1,000 samples that serve as the foundation for task construction. Based on these samples, we introduce three themes: step-based, outcome-based, and criticism-based, encompassing 12 diverse tasks and 12,634 questions. These tasks can comprehensively evaluate VLRMs in the aspects of process understanding, outcome judgment, and criticism generation, thus laying the groundwork for the development of VLRMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "text", + "content": "We conducted extensive experiments on 21 open-source models (from 2B to 90B) and 5 closed-source models, such as GPT-4o and Claude-3.5-Sonnet [3], across 12 tasks. The results underscore the challenges of VLRMBench and demonstrate the narrowing gap between open-source and closed-source models. Specifically, in eight step-based tasks, even the advanced GPT-4o achieved only an average F1-Score of " + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "text", + "content": ". In two outcome-based binary classification tasks, GPT-4o attained an average accuracy of only " + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "inline_equation", + "content": "66.3\\%" + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "text", + "content": ". Furthermore, in two criticism-based tasks, Qwen2.5-VL-72B [4] outperformed GPT-4o, with an average win rate of " + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "inline_equation", + "content": "53.3\\%" + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "text", + "content": ", showcasing the superior performance of state-of-the-art (SOTA) open-source models. In addition, we performed extensive analytical studies, providing valuable insights into the development of VLRMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "The main contributions can be summarized as follows: 1) We introduce VLRMBench, the first benchmark that integrates both comprehensiveness and challenge for the evaluation of VLRMs. VLRMBench is constructed from 1,000 high-quality samples curated through our collaborative data filtering and generation pipeline, covering areas such as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "mathematical reasoning, hallucination understanding, and multi-image understanding. 2) We designed 12 tasks, comprising a total of 12,634 carefully crafted questions, capable of evaluating models across three aspects: process understanding, outcome judgment, and critique generation. These diverse tasks enable targeted evaluations and reveal limitations of VLRMs. 3) We conducted extensive experiments on VLRMBench using 21 open-source models and 5 closed-source models, highlighting the challenges presented by VLRMBench and offering comprehensive analysis and guidance for the development of VLRMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 217, + 406, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 406, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 406, + 228 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 236, + 443, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 443, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 443, + 249 + ], + "type": "text", + "content": "2.1. Reward Models (RMs)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 421 + ], + "type": "text", + "content": "Techniques such as Reinforcement Learning with Human Feedback (RLHF) utilize RMs that offer feedback for large models, aligning them with human preferences in terms of accuracy, utility, and safety [5, 34]. RMs can generally be categorized into three types based on their objectives. Process-based RMs [31, 59] evaluate intermediate reasoning steps to ensure logical consistency and the correctness of each step. Outcome-based RMs [6, 25, 28] assess the quality of the final output. In contrast, Critique-based RMs [8, 18, 45] emphasize offering textual feedback or critiques aimed at enhancing the model's performance. However, RMs are not always reliable [37]. Therefore, it is imperative to establish a benchmark that facilitates a comprehensive evaluation of RMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 431, + 485, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 485, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 485, + 443 + ], + "type": "text", + "content": "2.2. Benchmarks of Reward Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 628 + ], + "type": "text", + "content": "In recent years, several benchmarks for evaluating the capabilities of RMs have been proposed, driving the development of RMs in the field of natural language. For example, MR-GSM8k [54] redefines the model's role from a solver to a grader, necessitating the evaluation of the correctness of problem-solving steps. CriticBench [24] incorporates the Generate-Critique-Revise framework to assess the reasoning capabilities of models. ProcessBench [60] emphasizes step-by-step error detection in mathematical derivations, thereby increasing the granularity of the evaluation. PRMBench [37], based on PRM800K [23], employs GPT40 for data optimization and constructs a fine-grained evaluation dataset comprising 6,216 instances. PRMBench covers nine common error types and provides a comprehensive assessment of Process RMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "The aforementioned benchmarks predominantly focus on the language modality and do not adequately address the evaluation needs of VLLMs. To this end, [20] introduces VLRewardBench, a benchmark for VLRMs containing 1,250 test samples. However, VLRewardBench remains limited to Pairwise Comparison, wherein only a judgment of superiority or inferiority is made between two responses," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3164" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 529, + 257 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 257 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 257 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 257 + ], + "type": "image", + "image_path": "f27dd8ff8b3bc74740d5f66cb7624ec797115139f0f0d344c67c9858bcdc82f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 263, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 263, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 555, + 299 + ], + "type": "text", + "content": "Figure 1. The collaborative data filtering and generation pipeline of VLRMBench. During the filtering stage, Qwen2VL-7B is tasked with answering questions under two conditions: without and with image input. Questions for which the model provides incorrect answers are retained. Subsequently, QVQ-72B-preview and GPT-4o are employed to generate reasoning processes and construct tasks for three themes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 318, + 294, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 294, + 378 + ], + "type": "text", + "content": "failing to incorporate more diverse and enriched task types. Consequently, we propose VLRMBench, which encompasses three core themes and twelve distinct tasks, providing a holistic evaluation of VLRMs' overall performance and establishing a foundation for their future advancement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 140, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 140, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 140, + 403 + ], + "type": "text", + "content": "3. VLRMBench" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "text", + "content": "In this section, we initially present an overview of the pipeline for constructing VLRMBench, along with the data statistics information, followed by an in-depth description of the task design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 466, + 231, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 231, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 231, + 479 + ], + "type": "text", + "content": "3.1. Overall pipeline of VLRMBench" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 676 + ], + "type": "text", + "content": "In Figure 1, we present a progressive and intuitive pipeline, which is designed to construct VLRMBench in three stages: data collection, filtering, and generation. During the collection phase, we integrated four mathematical, two hallucination, and two multi-image understanding datasets, totaling 16,550 samples and 19,343 images. In the filtering phase, to select high-quality and challenging samples, we introduced a joint filtering mechanism. Specifically, we initially employed Qwen2VL-7B [44] to answer questions without image information. If the model answered correctly, it indicated that the sample was of low quality [7]. Subsequently, we incorporated the image information into the question and instructed the model to answer it. If the model answered correctly, it suggested that the question was too simple to be a challenge. Consequently, we retained only those samples that could not be answered correctly in ei" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 318, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 342 + ], + "type": "text", + "content": "ther question-answering mode, resulting in a total of 6,715 samples, which formed our basic database " + }, + { + "bbox": [ + 313, + 318, + 555, + 342 + ], + "type": "inline_equation", + "content": "D_{b}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 349, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 661 + ], + "type": "text", + "content": "In the generation phase, we utilized QVQ-72B-preview [39] to generate reasoning process for the samples in " + }, + { + "bbox": [ + 313, + 349, + 556, + 661 + ], + "type": "inline_equation", + "content": "D_{b}" + }, + { + "bbox": [ + 313, + 349, + 556, + 661 + ], + "type": "text", + "content": ", discarding those where QVQ-72B-preview yielded incorrect answers, as verified via comparison with the standard answers. Subsequently, GPT-4o was employed to perform step-level segmentation of QVQ-72B-preview's reasoning process, grouping semantically similar and logically related sentences into the one step. Following this, we obtained 3,335 samples with reasoning segmented at the step level. Furthermore, based on specific criteria², we excluded samples with excessively long or short reasoning steps and ensured a balanced distribution across the three sample types. Finally, to ensure the quality of the samples, a manual review was conducted. Specifically, three PhD volunteers were recruited to verify the high-quality reasoning processes generated by QVQ-72B-preview. If any volunteer identified an error in the reasoning process, all three volunteers would collaboratively review the sample and correct any potential mistakes. Additionally, examples of manual corrections applied to the samples could be found in the Appendix. After completing the three aforementioned stages, 1,000 high-quality samples were retained for constructing the specific task, denoted as " + }, + { + "bbox": [ + 313, + 349, + 556, + 661 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 349, + 556, + 661 + ], + "type": "text", + "content": ". For more detailed statistical information, please refer to Table 2 and Figure 2, which provide a description of the sample sources, the average word count, and the number of reasoning steps." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "Qwen2VL-7B ranks in the top " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " of the OpenCompass leaderboard [11]. With moderate model capabilities and a small model size, it can efficiently filter out low-quality and low-difficulty samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "2We excluded samples where the number of reasoning steps exceeded the average (i.e. 20) or was fewer than 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3165" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 180 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "type": "table", + "html": "
CategoryDataset#Samples#AvgWORDS
MathMathVista [29]142696
MathVision [42]281,021
MathVerse [58]126742
DynaMath [64]251691
HallucinationIllusionVQA [36]62736
HallusionBench [15]334605
Multi-ImageBLINK [14]23790
Mantis-Eval [16]34592
", + "image_path": "b379ac894b7888f7486dc2d24c6565908fa6a90ab38ccb7fb2cf9b7138acb89a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 189, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 295, + 222 + ], + "type": "text", + "content": "Table 2. Statistics on the sample sources of VLRMBench. We also counted the average number of words in the reasoning process under different data sources." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 243, + 156, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 156, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 156, + 255 + ], + "type": "text", + "content": "3.2. Step-based tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 260, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 296, + 331 + ], + "type": "text", + "content": "The step-based theme consists of 8 tasks designed to assess the reasoning process understanding ability of VLRMs. Under this theme, models will be provided with a modified reasoning process and are expected to output a sequence of 0s and 1s to indicate whether each reasoning step contains deviated content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": "Step correctness (SC) For a Vision-Language Reward Model (VLRM), detecting errors at each step of the reasoning process is a critical capability [23, 37]. Building on this, we propose the SC task. Specifically, given the reasoning process " + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "inline_equation", + "content": "R_{n}" + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": " of a sample from " + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": " denotes the total number of reasoning steps, we inject errors into " + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": " steps " + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "inline_equation", + "content": "(0 < m < n)" + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "inline_equation", + "content": "R_{n}" + }, + { + "bbox": [ + 54, + 347, + 295, + 443 + ], + "type": "text", + "content": " using the advanced GPT-4o, to simulate potential errors during reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "text", + "content": "Redundancy Detection (RD) Existing reasoning models often produce redundant information when answering questions, leading to an increase in output token length, which is both inefficient and environmentally unfriendly. To this end, we propose the RD task, which requires VLRMs to identify redundant information generated during the reasoning process. Specifically, based on " + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "text", + "content": ", we introduce redundant descriptions in certain reasoning steps of each sample. For instance, we incorporate descriptions from " + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "inline_equation", + "content": "R_{i-j}" + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "text", + "content": " to intentionally create redundancy. Furthermore, we employ GPT-4 to generate paraphrased versions of the content from " + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "inline_equation", + "content": "R_{i-j}" + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "text", + "content": " and incorporate them into " + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "inline_equation", + "content": "R_i" + }, + { + "bbox": [ + 55, + 459, + 295, + 637 + ], + "type": "text", + "content": ", thereby amplifying redundancy. Unlike the SC task, the introduction of redundant information does not compromise the correctness of the final answer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Confidence Misdirection (CM) When confident and assertive terms, such as 'definitely' and 'without a doubt', are present in an incorrect reasoning step, is it possible for VL-RMs to be misled into perceiving the erroneous reasoning as correct? To investigate this issue, we propose the CM task," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 335, + 74, + 533, + 208 + ], + "blocks": [ + { + "bbox": [ + 335, + 74, + 533, + 208 + ], + "lines": [ + { + "bbox": [ + 335, + 74, + 533, + 208 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 533, + 208 + ], + "type": "image", + "image_path": "ebfe9977b3db29d7faba65e6419e5c8b59427bdfdc4136cbc90d453078734927.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 239 + ], + "type": "text", + "content": "Figure 2. The distribution of the number of reasoning steps for different data sources." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "content": "which is derived from the SC task. Specifically, for each sample in the SC task, we incorporate confidence-boosting terms into the erroneous reasoning step to mislead VLRMs into perceiving the step as correct, thereby evaluating the robustness of VLRMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 338, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 556, + 590 + ], + "type": "text", + "content": "Existential Hallucination (EH) and Attribute Hallucination (AH) When answering questions using LVLMs, hallucination phenomena often arise [26]. Although other parts of the reasoning process may be correct, the introduction of hallucinatory errors or the omission of key entities can lead to biased final results. To evaluate whether VLRMs are capable of detecting hallucinations introduced during the reasoning process, we have designed two tasks on the top of " + }, + { + "bbox": [ + 313, + 338, + 556, + 590 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 338, + 556, + 590 + ], + "type": "text", + "content": ". The first task, EH, assesses the ability of the VLRM to identify entities that were incorrectly introduced during reasoning process but do not exist in the image, or to detect elements that may have been overlooked, such as objects, people, etc. The second task, AH, examines the VLRM's capability to accurately identify the attributes of entities, particularly in cases where LVLMs wrongly attribute properties to an entity that it does not possess or when attribute descriptions conflict with the visual content of the image. Such errors are typically reflected in misjudgments regarding object attributes, such as size, color, or shape. VLRMs must integrate image information effectively to accurately identify hallucination errors during the reasoning process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Detail Error (DE) When addressing complex problems, the reasoning process often involves multiple details, particularly in the handling of numbers and symbols. Even if the overall reasoning appears to be sound, there may still be hidden errors in finer details (e.g., numerical errors in calculations or symbolic mistakes in formulas). These errors are often difficult to detect, yet they can affect the precision of the final results. The introduction of DE tasks serves to evaluate the sensitivity of VLRMs to reasoning details, as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3166" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 258, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 258, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 258, + 85 + ], + "type": "text", + "content": "well as their ability to identify fine-grained errors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "type": "text", + "content": "Spatial Relationship (SR) During visual reasoning, LVLMs may incorrectly describe the spatial relationships between objects or entities, leading to erroneous results. For instance, LVLMs might misdescribe an object located at the top of the image as being at the bottom. Therefore, we select samples containing spatial relationship descriptions from " + }, + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "type": "text", + "content": " to construct the SR task, aiming to evaluate VLRMs' understanding of spatial relationships in a image and their ability to detect spatial errors in a reasoning process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 376 + ], + "type": "text", + "content": "Image Confusion (IC) In multi-image tasks, LVLMs are required to extract and associate relevant information from multiple images, which typically involves identifying various entities and their interrelations across different images. However, due to the presence of similar elements or backgrounds between images, LVLMs may erroneously associate an entity from one image with another image. This 'image reference error' not only causes confusion but also potentially impacts the accuracy of subsequent reasoning. Thus, we filter samples related to multi-image understanding from " + }, + { + "bbox": [ + 55, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 220, + 296, + 376 + ], + "type": "text", + "content": " to construct the IC task, which aims to evaluate whether VLRMs can identify image reference errors during the reasoning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 178, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 178, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 178, + 393 + ], + "type": "text", + "content": "3.3. Outcome-based tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 435 + ], + "type": "text", + "content": "The outcome-based theme encompasses two binary classification tasks, aiming to evaluate the ability of VLRMs to assess and predict the accuracy of outcomes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": "Multi-solution Judgment (MJ) During the training stage, preference optimization techniques, such as DPO [35], have become a widely used paradigm. These methods typically rely on the collection of preference samples. VLRMs are frequently employed to assess the quality of preference samples, thereby becoming a crucial component of preference optimization [21, 53]. To this end, we introduce the MJ task, which aims to investigate the discriminative ability of VLRMs regarding different reasoning processes for the same problem. This task is discussed as a core element in VLRewardBench [20] and Multimodal Reward-Bench [52]. Specifically, for the reasoning process " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": " of any question " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": ", we introduce the reasoning process " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": " of the same question in the SC, RD and CM tasks, thereby constructing a triplet " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "(Q, R, R')" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": ". VLRMs are required to judge the quality of " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": " and score them accordingly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Forecasting Future (FF) When LVLMs perform inference, test-time scaling (TTS) can further improve their performance [47, 51]. However, these techniques often introduce additional inference time, primarily due to the necessity of exploring multiple reasoning paths. If the correctness" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": "of the final answer could be predicted based on the reasoning process of the first m steps, " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "R_{[0:m]}" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": ", this prediction could substantially expedite the TTS process. Therefore, the FF task is proposed, wherein VLRMs are required to assess the correctness of the final answer based on " + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "inline_equation", + "content": "R_{[0:m]}" + }, + { + "bbox": [ + 313, + 72, + 554, + 134 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 140, + 440, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 440, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 440, + 152 + ], + "type": "text", + "content": "3.4. Criticism-based tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 158, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 315 + ], + "type": "text", + "content": "Error Reason Analysis (ERA) and Error Correction (EC) Unlike Process RM and Outcome RM, which focus solely on evaluating processes or outcomes, Critique RM emphasizes the critique of the reasoning process, providing textual feedback that includes error analysis and correction. Consequently, we propose the ERA and EC tasks to assess the capabilities of VLRMs in these two aspects. In the ERA task, VLRMs are presented with erroneous reasoning steps derived from the SC task, and are tasked with identifying the error and analyzing its underlying cause. In contrast, the EC task does not require error cause analysis; rather, VLRMs directly correct the identified error and produce the corrected reasoning process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 326, + 395, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 395, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 395, + 339 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 346, + 372, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 372, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 372, + 358 + ], + "type": "text", + "content": "4.1. Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 364, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 496 + ], + "type": "text", + "content": "We conducted extensive experiments on 21 open-source models and 5 closed-source models. Specifically, the open-source models, which range from 2B to 90B, including Llama3.2 (11B/90B) [12], Llava-OneVision (7B/72B) [19], Qwen2VL (2B/7B/72B) [44], Qwen2.5VL (3B/7B/72B) [4], QVQ-72B-preview [39], Ovis2 (8B/16B/34B) [30], InternVL2.5 (2B/8B/38B/78B) [9], Phi-3.5-Vision (4.2B) [1], MiniCPM-o-2.6 (8B), and MiniCPM-V-2.6 (8B) [33]. For the closed-source models, we employed GLM-4V-Flash, GPT-4o [2], Gemini-2.0-Flash-/Flash-thinking-exp-1219 [38], and Claude-3.5-Sonnet [3]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 504, + 422, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 422, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 422, + 516 + ], + "type": "text", + "content": "4.2. Evaluation details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "We follow the 'LVLM-as-a-Judge' paradigm and introduce different question-answering templates to assess the models [20, 52]. Specifically, for the SC, RD, CM, EH, AH, DE, SR, and IC tasks, the model takes a reasoning process as input and outputs a judgment regarding whether each reasoning step contains redundancies or other errors. For these eight tasks, we employ the weighted F1-Score as the evaluation metric, as defined in Equation 1, which better addresses class imbalance issues and ensures the model's performance on the minority class is adequately considered. For the MJ task, the model is provided with (good and bad) reasoning process pairs for the same question and outputs scores for both. If the model assigns a higher score to the good sample, it is deemed correct; otherwise, it is considered incorrect. For the FF task, the model takes several preceding steps of the reasoning process as input and outputs a judgment on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3167" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 552, + 411 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 552, + 411 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 552, + 411 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 552, + 411 + ], + "type": "table", + "html": "
ModelStep (5,722)Outcome (5,402)Criticism (1,510)
SCRDCMEHAHDESRICW-Avg.MJFFW-Avg.ERAwERAτERALECwECTECL
# Sample99899799494694465912658-2,9862,416-974536
Open-source LVLMs < 10B (Small group)
InternVL2.5-2B33.834.130.022.423.833.519.930.129.226.660.841.94.86.788.52.310.087.6
Qwen2VL-2B28.326.220.917.819.126.623.826.423.010.559.632.50.20.199.72.55.891.7
Qwen2.5VL-3B29.932.723.227.226.945.525.142.630.131.967.147.67.112.180.84.821.273.9
Phi-3.5-Vision (4.2B)35.938.827.920.721.133.417.731.629.212.165.235.83.89.087.24.222.473.4
Qwen2VL-7B38.029.332.021.521.432.821.621.128.852.064.757.73.98.187.95.423.271.4
Qwen2.5VL-7B43.433.237.822.823.945.515.629.133.426.070.746.037.722.040.39.351.938.8
Llava-OneVision-7B31.825.721.516.017.224.816.120.722.618.057.835.83.910.485.77.134.458.5
InternVL2.5-8B36.628.431.121.921.236.515.124.128.661.473.866.934.521.943.66.236.157.7
Ovis2-8B47.728.237.825.225.739.817.025.833.331.465.346.623.225.950.914.363.522.2
MiniCPM-V-2.6 (8B)44.938.336.926.126.536.824.230.634.637.962.548.95.411.782.93.424.771.9
MiniCPM-o-2.6 (8B)41.035.337.329.629.036.728.330.134.615.064.337.09.219.870.97.633.958.4
Avg.37.431.830.622.823.335.620.428.429.829.364.745.212.213.474.46.129.764.1
10B < Open-source LVLMs < 40B (Middle group)
Llama3.2-11B33.927.129.720.220.031.917.125.826.723.961.140.54.210.385.55.434.759.8
Ovis2-16B48.547.947.940.635.844.521.928.443.649.769.258.430.926.642.510.469.320.3
Ovis2-34B65.351.164.554.551.659.630.449.057.046.375.759.445.926.627.515.476.48.1
InternVL2.5-38B61.049.662.655.051.761.434.648.156.145.179.360.421.124.554.412.556.630.9
Avg.52.243.951.242.639.849.426.037.845.941.371.354.725.522.052.510.959.329.8
Open-source LVLMs > 40B (Large group)
Llava-OneVision-72B49.430.650.237.934.145.113.926.540.432.269.749.015.322.162.69.451.039.6
Qwen2VL-72B55.540.256.148.941.656.422.443.648.837.772.153.122.420.457.17.757.135.1
Qwen2.5VL-72B72.841.770.464.659.972.437.963.362.665.680.272.174.115.110.815.677.07.3
QVQ-72B-preview42.435.336.329.427.137.820.826.134.349.238.444.44.210.085.89.451.039.6
InternVL2.5-78B62.142.062.853.348.760.428.547.254.048.678.862.154.224.920.910.863.925.3
Llama3.2-90B40.838.841.334.229.435.417.021.136.252.769.060.09.918.671.49.556.833.7
Avg.53.838.152.944.740.151.323.438.046.147.768.056.830.018.551.410.459.530.1
Closed-source LVLMs
Claude-3.5-Sonnet70.853.765.763.962.863.440.865.762.982.275.179.060.625.513.921.253.924.9
GPT-4o73.750.666.657.658.671.848.264.662.458.476.066.30.0100.00.00.0100.00.0
GLM-4V-Flash37.233.532.928.227.741.822.920.232.837.266.750.434.224.341.56.954.139.0
Gemini-2.0-Flash69.355.966.165.662.363.446.867.863.454.277.164.432.822.544.77.152.540.3
Gemini-2.0-FTE.-121967.955.565.765.962.363.246.467.163.153.177.063.837.422.640.06.850.243.1
Avg.63.849.859.456.254.760.741.057.157.057.074.464.833.039.028.08.462.129.5
", + "image_path": "4b7eecf51e7571df2020a283d6abd5c0d2b9716ab9427d2b6a456b4f4419788a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "type": "text", + "content": "Table 3. Results on the VLRMBench. For Step-based tasks, the F1-Score is the metric. For Outcome-based tasks, Accuracy is utilized as the metric. In the case of Criticism-based tasks, the metric is Win Rate, expressed as Win: Tie: Lose (Simplified as " + }, + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 420, + 555, + 456 + ], + "type": "text", + "content": "). Note that W-Avg. represents the weighted average. Evaluation prompt templates and some case studies could be found in the Appendix B and C." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 475, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 558 + ], + "type": "text", + "content": "whether a correct answer can ultimately be derived. Consequently, for the MJ and FF tasks, Accuracy is employed as the evaluation metric. For the ERA and EC tasks, the model's output is a description of the error reasons in the reasoning process and the corrected reasoning process. We utilize the output of GPT-4o as the baseline and adopt Win Rate [13, 61] as the evaluation metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 654 + ], + "type": "text", + "content": "We configured the max new tokens to 4,096 across all tasks for the five closed-source models and QVQ-72B-preview. In contrast, for the other models, we set the max new tokens to 4,096 specifically for the ERA and EC tasks, while fixing it at 256 for the remaining tasks. To ensure the reproducibility of the experimental results, we set the temperature coefficient to 0 for all models. All experiments were conducted on 8 NVIDIA H800 GPUs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 670, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 670, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 67, + 670, + 295, + 696 + ], + "type": "interline_equation", + "content": "F 1 = \\frac {N _ {\\text {n e g}}}{N _ {\\text {p o s}} + N _ {\\text {n e g}}} \\cdot F 1 _ {\\text {p o s}} + \\frac {N _ {\\text {p o s}}}{N _ {\\text {p o s}} + N _ {\\text {n e g}}} \\cdot F 1 _ {\\text {n e g}} \\tag {1}", + "image_path": "b771910c82590463d6cee213c484a9ee9f782d7d50ffd17e8b56175cd0cf678d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{pos}}" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{neg}}" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": " represent the number of correct and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 475, + 528, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 528, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 528, + 487 + ], + "type": "text", + "content": "incorrect steps in the reasoning process, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 492, + 400, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 400, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 400, + 504 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 533 + ], + "type": "text", + "content": "The experimental results on VLRMBench are shown in Table 3, leading to the following conclusions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": "1) Among open-source models, surprisingly, the performance of Qwen2.5VL-72B is comparable to that of advanced closed-source models. For instance, across the eight Step-based tasks, Qwen2.5VL-72B achieved an average F1-Score of " + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "inline_equation", + "content": "62.6\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": ", surpassing GPT-4o by " + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": " and only slightly trailing Claude-3.5-Sonnet by " + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": ". In the MJ and FF tasks, Qwen2.5VL-72B also outperformed GPT-4o by " + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": ", respectively. Notably, in the ERA task, Qwen2.5VL-72B attained a win rate of " + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "inline_equation", + "content": "74.1\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": " over GPT-4o, significantly outperforming Claude-3.5-Sonnet at " + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "inline_equation", + "content": "60.6\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "2) LVLMs that have not undergone specific training show diminishing performance gains as the model size increases. Specifically, for LVLMs in the Middle group, compared to those in the Small group, the performance gain in Step-based and Outcome-based tasks was " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "54.0\\%" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3168" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 286, + 185 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 286, + 185 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 286, + 185 + ], + "type": "table", + "html": "
OrderModelAvg. ScoreAcc
Better firstQwen2.5VL-3B[9.0, 8.3]23.7
Llama3.2-11B[8.6, 4.8]52.8
InternVL2.5-78B[8.4, 7.4]38.6
GPT-4o[9.8, 7.8]52.2
Better secondQwen2.5VL-3B[8.7, 9.0]21.4
Llama3.2-11B[8.3, 5.6]8.4
InternVL2.5-78B[8.1, 9.1]45.6
GPT-4o[7.4, 8.1]43.1
", + "image_path": "2aa3132319e4dd6a4913726c20034c383816d57604552ab065c85482b118cbb0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 294, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 347 + ], + "type": "text", + "content": "21.0%, respectively. In contrast, the corresponding performance gains in the Large group relative to the Middle group were only 0.4% and 3.8%. Even excluding QVQ-72B-preview from the Large group, the relative performance improvement remains limited, at 5.4% and 8.4%. This suggests that blindly increasing the size of VLLMs does not lead to significant performance improvements. Consequently, we recommend that future research prioritize training VLRMs at the Middle group scale initially." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "text", + "content": "3) The Outcome-based tasks, being simple binary classification problems, remain challenging for most open-source models, and even closed-source models perform only marginally better than random guessing. For open-source models across three size groups, the average accuracy for Outcome-based tasks is around " + }, + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "text", + "content": ", similar to random guessing. For advanced closed-source models, the average accuracy only exceeds random guessing by " + }, + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "inline_equation", + "content": "14.8\\%" + }, + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": "4) Deterministic vocabulary and hallucinated information can effectively mislead the VLRMs, reducing its ability to understand reasoning processes. The CM task, based on the SC task, introduces deterministic vocabulary (e.g., must, definitely) into sentences with erroneous reasoning steps, while EH and AH tasks introduce hallucinated information during the reasoning process. Existing models, particularly in the Small group, demonstrate a degradation in F1-Score of " + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "inline_equation", + "content": "14.6\\%" + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "inline_equation", + "content": "14.1\\%" + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": " for CM, EH, and AH tasks, respectively, when compared to the SC task. This performance degradation remains significant even as model size increases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "5) Existing models show strong robustness to numerical errors but limited resilience to perturbations in spatial relationships and image references. We observed that for Large group, the F1-Score in the DE task only decreased by " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " compared to SC. However, for the SR and IC tasks, the performance dropped by " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "30.4\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", respectively. Overall, our VLRMBench provide unique challenges that surpass traditional benchmarks, with even the most advanced models achieving only mediocre performance." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 547, + 137 + ], + "blocks": [ + { + "bbox": [ + 60, + 194, + 290, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 194, + 290, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 194, + 290, + 205 + ], + "type": "text", + "content": "Table 4. The results on the MJ task with different sample order." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 70, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 547, + 137 + ], + "type": "table", + "html": "
Model≤10 steps10-15 steps>15 steps
Qwen2.5VL-3B35.630.125.6
Llama3.2-11B42.932.828.4
InternVL2.5-78B68.264.154.1
GPT-4o77.875.268.6
", + "image_path": "91ae9d72b132b5cebc4b7f4e16aeceeb5b11fa47f6084b2c28b7069bd88ef514.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 145, + 546, + 156 + ], + "lines": [ + { + "bbox": [ + 321, + 145, + 546, + 156 + ], + "spans": [ + { + "bbox": [ + 321, + 145, + 546, + 156 + ], + "type": "text", + "content": "Table 5. Results on the SC task with different reasoning steps." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 178, + 416, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 416, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 416, + 192 + ], + "type": "text", + "content": "5. Further Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 198, + 542, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 542, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 542, + 210 + ], + "type": "text", + "content": "5.1. Is the o1-like model suitable to be a VLRM?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 384 + ], + "type": "text", + "content": "QVQ-72B-preview [39], an open-source visual-language model, exhibits strong reasoning capabilities. However, it exhibits a potential drawback regarding the degradation of instruction-following ability. For instance, in the SC task, by incorporating instruction prompts, the sample's original question, and the reasoning process as input, we expect the VLRM to evaluate each reasoning step. However, QVQ-72B-preview often disregards the given instructions and directly answers the original question, leading to suboptimal performance on VLRMBench, with additional cases provided in the Appendix C. Therefore, as an o1-like model, it must not only demonstrate strong reasoning abilities but also exhibit robust instruction-following capabilities to be considered a viable candidate for a VLRM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 393, + 447, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 447, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 447, + 406 + ], + "type": "text", + "content": "5.2.Impact of sample order" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 567 + ], + "type": "text", + "content": "In the MJ task, VLRM scores two reasoning processes for the same question. To further investigate the positional bias, we employed two different sample orders, as shown in Table 4. Specifically, 'Better first' places the superior reasoning process in the first position, while 'Better second' places it in the second position. The results indicate that Llama3.2-11B consistently assigns a higher score to the reasoning process in the first position, irrespective of its quality. In contrast, the other three models exhibit the ability to adjust their scores based on the quality of the reasoning. This positional bias phenomenon is also observed in [43]. As a result, in Table 3, we utilized the sum of the two scores for the MJ task evaluation to ensure fairness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 575, + 529, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 529, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 529, + 589 + ], + "type": "text", + "content": "5.3. Performance on various reasoning length" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": "We further investigate the impact of the number of reasoning steps on the VLRMs' process understanding. As shown in Table 5, we classified the number of reasoning steps in the SC task into three categories: the short group (" + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\leq 10" + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": " steps), with 263 samples; the middle group (10-15 steps), with 468 samples; and the long group (" + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "inline_equation", + "content": ">15" + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": " steps), with 267 samples. Specifically, for all four models, a decrease in F1-score was observed as the number of reasoning steps increased. Notably, for the Llama3.2-11B model, the difference between the short and long groups reached " + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "inline_equation", + "content": "14.5\\%" + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3169" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 274, + 137 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 274, + 137 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 274, + 137 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 274, + 137 + ], + "type": "table", + "html": "
Model0-shot1-shot2-shot
Qwen2.5VL-3B29.931.531.1
Llama3.2-11B33.942.242.3
InternVL2.5-78B62.155.955.7
GPT-4o73.771.770.5
", + "image_path": "e28cc05a43541e9df380fc700c1698baabf40ce48863f28604c15854880735fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 171, + 282, + 269 + ], + "blocks": [ + { + "bbox": [ + 69, + 171, + 282, + 269 + ], + "lines": [ + { + "bbox": [ + 69, + 171, + 282, + 269 + ], + "spans": [ + { + "bbox": [ + 69, + 171, + 282, + 269 + ], + "type": "image", + "image_path": "739e169c1130866e94dbc8225caf0d71bb5a71df2711e89764cecc05ce3991a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 280, + 280, + 293 + ], + "lines": [ + { + "bbox": [ + 70, + 280, + 280, + 293 + ], + "spans": [ + { + "bbox": [ + 70, + 280, + 280, + 293 + ], + "type": "text", + "content": "Figure 3. Performance changes with the test-time scaling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 313, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 374 + ], + "type": "text", + "content": "even for the advanced GPT-4o, the F1-Score dropped from " + }, + { + "bbox": [ + 55, + 313, + 295, + 374 + ], + "type": "inline_equation", + "content": "77.8\\%" + }, + { + "bbox": [ + 55, + 313, + 295, + 374 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 313, + 295, + 374 + ], + "type": "inline_equation", + "content": "68.6\\%" + }, + { + "bbox": [ + 55, + 313, + 295, + 374 + ], + "type": "text", + "content": ". This suggests that current models are less capable of handling long information, underscoring the necessity of enhancing their ability to tackle long reasoning process when training VLRMs in the future." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 382, + 291, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 291, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 291, + 396 + ], + "type": "text", + "content": "5.4. Impacts of In-Context Learning (ICL) setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "text", + "content": "In main experiments, we employ zero-shot setting by default. Furthermore, using the SC task as an example, we explore the effect of ICL on VLRMs. As presented in Table 6, the results indicate that providing contextual examples is more beneficial for small models but harms the performance of large models. Specifically, with the help of one-shot learning, Qwen2.5VL-3B's performance improved from " + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "inline_equation", + "content": "29.9\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "inline_equation", + "content": "31.5\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "text", + "content": ", and Llama3.2-11B saw an improvement of " + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "text", + "content": ". However, for GPT-4o, one-shot learning led to a " + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "text", + "content": " performance degradation, and InternVL2.5-78B's performance decreased by " + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 652 + ], + "type": "text", + "content": ". One possible explanation for this is that small models exhibit relatively weaker instruction-following capabilities, and ICL examples may help guide them in responding to queries in the designated format. In contrast, large models, which already possess strong comprehension abilities and adaptability to complex tasks, may find that providing contextual examples interferes with their inherent understanding process, resulting in performance decline. Thus, in the SC task, smaller models exhibit a greater reliance on ICL examples, whereas larger models tend to rely more on their pre-trained knowledge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 207, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 207, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 207, + 673 + ], + "type": "text", + "content": "5.5. Impacts of test-time scaling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We investigated the impact of test-time scaling (i.e. voting method) for VLRMs on the MJ task. To ensure the diversity of the generated results, we set the temperature coefficient" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 328, + 71, + 541, + 130 + ], + "blocks": [ + { + "bbox": [ + 58, + 145, + 292, + 156 + ], + "lines": [ + { + "bbox": [ + 58, + 145, + 292, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 145, + 292, + 156 + ], + "type": "text", + "content": "Table 6. Results on SC task with various ICL few-shot numbers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 71, + 541, + 130 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 541, + 130 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 541, + 130 + ], + "type": "table", + "html": "
LVLMVLRMMathVistaMathVerseMathVisionAvg.
7B7B10.45.84.77.0
72B7B11.84.23.26.4
7B72B21.819.721.521.0
72B72B11.79.311.410.8
", + "image_path": "997a2f1a0b15ee8261d9068c5d9b8947de4206fe44536c7441fadf0d7e7c1ab0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 138, + 542, + 148 + ], + "lines": [ + { + "bbox": [ + 325, + 138, + 542, + 148 + ], + "spans": [ + { + "bbox": [ + 325, + 138, + 542, + 148 + ], + "type": "text", + "content": "Table 7. Results of the correction rate on math benchmarks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "text", + "content": "to 0.9 and performed K sampling. As illustrated in Figure 3, surprisingly, as the value of K increases, the accuracy of Qwen2.5VL-3B improved by " + }, + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "text", + "content": ". Notably, at " + }, + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 9" + }, + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "text", + "content": ", its performance was only " + }, + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 313, + 172, + 555, + 245 + ], + "type": "text", + "content": " lower than that of GPT-4o. These findings highlight the substantial effectiveness of test-time scaling techniques in enhancing VLRMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 254, + 484, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 484, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 484, + 266 + ], + "type": "text", + "content": "5.6. Improve LVLMs with a VLRM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 272, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 524 + ], + "type": "text", + "content": "To demonstrate the significant role of VLRMs in the reasoning process, as presented in Table 7, we conducted experiments on three mathematical benchmarks using Qwen2.5VL-7B/72B. For example, in the second row of Table 7, Qwen2.5VL-72B was employed as the LVLM to answer the mathematical questions, whereas Qwen2.5VL-7B was used as the VLRM to perform error reason analysis of the LVLM's reasoning process. The analysis was subsequently fed back into the LVLM as feedback, allowing it to generate a second round of answers. Besides, we defined the correction rate, which is the ratio of the number of questions correctly answered in the second round after feedback to the number of questions incorrectly answered in the first round. The experimental results demonstrate that VLRM can provide valuable information to the LVLM through error analysis, thereby facilitating mistake correction, with an average gain of " + }, + { + "bbox": [ + 313, + 272, + 555, + 524 + ], + "type": "inline_equation", + "content": "11.3\\%" + }, + { + "bbox": [ + 313, + 272, + 555, + 524 + ], + "type": "text", + "content": ". These findings suggest that even LVLMs not trained on proprietary data can act as VLRMs, providing more correction possibilities during testing based on this intuitive feedback mechanism. This further underscores the potential value of developing efficient VLRMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 537, + 391, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 391, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 391, + 550 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a comprehensive and challenging benchmark, named VLRMBench. This benchmark covers 3 themes and 12 tasks, comprising a total of 12,634 questions. Following the 'LVLM-as-a-Judge' paradigm, we conduct systematic experiments and analyses on VLRMBench using 26 models, deriving several key insights: enhancing the performance of VLRMs in long-text comprehension, exploring various feedback patterns from VLRMs to LVLMs, and developing the application potential of proprietary VLRMs, etc. We hope that VLRMBench will serve as a standardized evaluation platform to propel the development of more efficient VLRMs. In future work, we plan to incorporate larger datasets specifically designed to train VLRMs." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3170" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 115 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grant No. 61977045)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 145, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 145, + 295, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 295, + 200 + ], + "type": "text", + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 201, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 201, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 201, + 295, + 255 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 256, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 294, + 277 + ], + "type": "text", + "content": "[3] Anthropic. Claude 3.5: Sonnet, 2025. Accessed: 2025-02-25. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 322 + ], + "type": "text", + "content": "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "text", + "content": "[5] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "text", + "content": "[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 425, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 478 + ], + "type": "text", + "content": "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "text", + "content": "[8] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 579 + ], + "type": "text", + "content": "[9] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "text", + "content": "[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 690 + ], + "type": "text", + "content": "[11] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[13] Yann Dubois, Balázs Galambosi, Percy Liang, and Tatsunori B Hashimoto. Length-controlled alpacaeval: A simple way to debias automatic evaluators. arXiv preprint arXiv:2404.04475, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 207 + ], + "type": "text", + "content": "[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision, pages 148-166. Springer, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 209, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 285 + ], + "type": "text", + "content": "[15] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[16] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max W.F. Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. Transactions on Machine Learning Research, 2024, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 332, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 375 + ], + "type": "text", + "content": "[17] Wasif Khan, Seowung Leem, Kyle B See, Joshua K Wong, Shaoting Zhang, and Ruogu Fang. A comprehensive survey of foundation models in medicine. IEEE Reviews in Biomedical Engineering, 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "text", + "content": "[18] Dahyun Kim, Chanjun Park, Sanghoon Kim, Wonsung Lee, Wonho Song, Yunsu Kim, Hyeonwoo Kim, Yungi Kim, Hyeonju Lee, Jihoo Kim, et al. Solar 10.7 b: Scaling large language models with simple yet effective depth up-scaling. arXiv preprint arXiv:2312.15166, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 434, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 476 + ], + "type": "text", + "content": "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "text", + "content": "[20] Lei Li, Yuancheng Wei, Zhihui Xie, Xuqing Yang, Yifan Song, Peiyi Wang, Chenxin An, Tianyu Liu, Sujian Li, Bill Yuchen Lin, et al. Vlrewardbench: A challenging benchmark for vision-language generative reward models. arXiv preprint arXiv:2411.17451, 2024. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[21] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, Lingpeng Kong, and Qi Liu. Vlfeedback: A large-scale ai feedback dataset for large vision-language models alignment. arXiv preprint arXiv:2410.09421, 2024. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[22] Zongxia Li, Xiyang Wu, Hongyang Du, Huy Nghiem, and Guangyao Shi. Benchmark evaluations, applications, and challenges of large vision language models: A survey. arXiv preprint arXiv:2501.02189, 2025. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[23] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, 2023. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Zicheng Lin, Zhibin Gou, Tian Liang, Ruilin Luo, Haowei Liu, and Yujiu Yang. Criticbench: Benchmarking lms for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3171" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "critique-correct reasoning. arXiv preprint arXiv:2402.14809, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[25] Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. Skywork-reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[26] Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xi-tian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models. arXiv preprint arXiv:2402.00253, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "text", + "content": "[27] Yantao Liu, Zijun Yao, Rui Min, Yixin Cao, Lei Hou, and Juanzi Li. Rm-bench: Benchmarking reward models of language models with subtlety and style. arXiv preprint arXiv:2410.16184, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[28] Xingzhou Lou, Dong Yan, Wei Shen, Yuzi Yan, Jian Xie, and Junge Zhang. Uncertainty-aware reward model: Teaching reward models to know what is unknown. arXiv preprint arXiv:2410.00847, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 327 + ], + "type": "text", + "content": "[29] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "type": "text", + "content": "[30] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 415 + ], + "type": "text", + "content": "[31] Qianli Ma, Haotian Zhou, Tingkai Liu, Jianbo Yuan, Pengfei Liu, Yang You, and Hongxia Yang. Let's reward step by step: Step-level reward model as the navigators for reasoning. arXiv preprint arXiv:2310.10080, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 437 + ], + "type": "text", + "content": "[32] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 470 + ], + "type": "text", + "content": "[33] OpenBMB. Minicpm o: A gpt-4o-level mllm for vision, speech, and multimodal live streaming on your phone, 2025. Accessed: 2025-02-25. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "text", + "content": "[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "text", + "content": "[35] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "text", + "content": "[36] Haz Sameen Shahgir, Khondker Salman Sayeed, Abhik Bhattacharjee, Wasi Uddin Ahmad, Yue Dong, and Rifat Shahriyar. Illusionvqa: A challenging optical illusion dataset for vision language models. arXiv preprint arXiv:2403.15952, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[37] Mingyang Song, Zhaochen Su, Xiaoye Qu, Jiawei Zhou, and Yu Cheng. Prmbench: A fine-grained and challenging benchmark for process-level reward models. arXiv preprint arXiv:2501.03124, 2025. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[38] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "text", + "content": "Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "type": "text", + "content": "[39] Qwen Team. Qvq: To see the world with wisdom, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 185 + ], + "type": "text", + "content": "[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 219 + ], + "type": "text", + "content": "[41] Jiahao Wang, Bolin Zhang, Qianlong Du, Jiajun Zhang, and Dianhui Chu. A survey on data selection for llm instruction tuning. arXiv preprint arXiv:2402.05123, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "text", + "content": "[42] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 554, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 320 + ], + "type": "text", + "content": "[43] Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 321, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 554, + 376 + ], + "type": "text", + "content": "[44] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 377, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 554, + 431 + ], + "type": "text", + "content": "[45] Tianlu Wang, Ilia Kulikov, Olga Golovneva, Ping Yu, Weizhe Yuan, Jane Dwivedi-Yu, Richard Yuanzhe Pang, Maryam Fazel-Zarandi, Jason Weston, and Xian Li. Self-taught evaluators. arXiv preprint arXiv:2408.02666, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "type": "text", + "content": "[46] Zhilin Wang, Alexander Bukharin, Olivier Delalleau, Daniel Egert, Gerald Shen, Jiaqi Zeng, Oleksii Kuchaiev, and Yi Dong. Helpsteer2-preference: Complementing ratings with preferences. arXiv preprint arXiv:2410.01257, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "text", + "content": "[47] Jinyang Wu, Mingkuan Feng, Shuai Zhang, Ruihan Jin, Feihu Che, Zengqi Wen, and Jianhua Tao. Boosting multimodal reasoning with mcts-automated structured thinking. arXiv preprint arXiv:2502.02339, 2025. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "text", + "content": "[48] Wang Xiyao, Yang Zhengyuan, Li Linjie, Lu Hongjin, Xu Yuancheng, Lin Chung-Ching Lin, Lin Kevin, Huang Furong, and Wang Lijuan. Scaling inference-time search with vision value model for improved visual comprehension. arXiv preprint arXiv:2412.03704, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "type": "text", + "content": "[49] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "text", + "content": "[50] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "text", + "content": "[51] Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3172" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "carlo tree search. arXiv preprint arXiv:2412.18319, 2024. 1,5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[52] Michihiro Yasunaga, Luke Zettlemoyer, and Marjan Ghazvininejad. Multimodal rewardbench: Holistic evaluation of reward models for vision language models. arXiv preprint arXiv:2502.14191, 2025. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[53] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[54] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "text", + "content": "[55] Zhongshen Zeng, Yinhong Liu, Yingjia Wan, Jingyao Li, Pengguang Chen, Jianbo Dai, Yuxuan Yao, Rongwu Xu, Zehan Qi, Wanru Zhao, et al. Mr-ben: A meta-reasoning benchmark for evaluating system-2 thinking in llms. arXiv preprint arXiv:2406.13975, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 343 + ], + "type": "text", + "content": "[56] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Skyeyegpt: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Photogrammetry and Remote Sensing, 221:64-77, 2025. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "type": "text", + "content": "[57] Di Zhang, Jingdi Lei, Junxian Li, Xunzhi Wang, Yujie Liu, Zonglin Yang, Jiatong Li, Weida Wang, Suorong Yang, Jianbo Wu, et al. Critic-v: Vlm critics help catch vlm errors in multimodal reasoning. arXiv preprint arXiv:2411.18203, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 466 + ], + "type": "text", + "content": "[58] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Springer, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "text", + "content": "[59] Zhenru Zhang, Chujie Zheng, Yangzhen Wu, Beichen Zhang, Runji Lin, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. The lessons of developing process reward models in mathematical reasoning. arXiv preprint arXiv:2501.07301, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[60] Chujie Zheng, Zhenru Zhang, Beichen Zhang, Runji Lin, Keming Lu, Bowen Yu, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[62] Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C Knoll. Vision language models in autonomous driving: A survey and outlook. IEEE Transactions on Intelligent Vehicles, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[63] Zihao Zhou, Shudong Liu, Maizhen Ning, Wei Liu, Jindong Wang, Derek F Wong, Xiaowei Huang, Qiufeng Wang, and Kaizhu Huang. Is your model really a good math rea" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "soner? evaluating mathematical reasoning with checklist. arXiv preprint arXiv:2407.08733, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[64] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3173" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_content_list.json b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c84d7ffb5ad402d0d20bf3938da9ce827d655112 --- /dev/null +++ b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_content_list.json @@ -0,0 +1,1360 @@ +[ + { + "type": "text", + "text": "VMBench: A Benchmark for Perception-Aligned Video Motion Generation", + "text_level": 1, + "bbox": [ + 119, + 130, + 877, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinran Ling $^{1*}$ \nCundian Yang $^{1}$", + "bbox": [ + 125, + 179, + 263, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Zhu1* \nAiming Hao1", + "bbox": [ + 295, + 180, + 400, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meiqi Wu $^{1,3*}$ \nJiashu Zhu $^{1}$", + "bbox": [ + 431, + 180, + 539, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hangyu Li1 \nJiahong Wu1†", + "bbox": [ + 576, + 180, + 691, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaokun Feng $^{1,2}$ \nXiangxiang Chu $^{1}$", + "bbox": [ + 715, + 180, + 864, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 AMAP, Alibaba Group", + "bbox": [ + 125, + 215, + 320, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ CRISE, Institute of Automation, Chinese Academy of Sciences", + "bbox": [ + 362, + 215, + 872, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer Science and Technology, University of Chinese Academy of Sciences", + "bbox": [ + 133, + 234, + 857, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video generation has advanced rapidly, improving evaluation methods, yet assessing video's motion remains a major challenge. Specifically, there are two key issues: 1) current motion metrics do not fully align with human perceptions; 2) the existing motion prompts are limited. Based on these findings, we introduce VMBench—a comprehensive Video Motion Benchmark that has perception-aligned motion metrics and features the most diverse types of motion. VMBench has several appealing properties: (1) Perception-Driven Motion Evaluation Metrics, we identify five dimensions based on human perception in motion video assessment and develop fine-grained evaluation metrics, providing deeper insights into models' strengths and weaknesses in motion quality. (2) Meta-Guided Motion Prompt Generation, a structured method that extracts meta-information, generates diverse motion prompts with LLMs, and refines them through human-AI validation, resulting in a multilevel prompt library covering six key dynamic scene dimensions. (3) Human-Aligned Validation Mechanism, we provide human preference annotations to validate our benchmarks, with our metrics achieving an average $35.3\\%$ improvement in Spearman's correlation over baseline methods. This is the first time that the quality of motion in videos has been evaluated from the perspective of human perception alignment. Additionally, we release VMBench at https://github.com/AMAP-ML/VMBench, setting a new standard for evaluating and advancing motion generation models.", + "bbox": [ + 91, + 320, + 483, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 773, + 220, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video generation models [2, 3, 37, 43, 55, 78, 90] have advanced significantly, particularly text-to-video (T2V) models [12, 45, 70, 96], which perform well in generating static content but struggle to capture motion in alignment with hu", + "bbox": [ + 89, + 801, + 482, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "man perception. To enhance dynamic motion generation, it is crucial to establish perception-aligned motion evaluation.", + "bbox": [ + 511, + 287, + 903, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, previous methods [29, 49, 59, 64, 75] primarily assess the static content of videos, with dynamic evaluation largely limited to motion smoothness [33, 44, 50]. This makes it challenging to capture motion deficiencies [69, 94], including but not limited to spatiotemporal inconsistencies and violations of physical laws. Therefore, developing a more comprehensive evaluation metric that better aligns with human perception is crucial.", + "bbox": [ + 511, + 320, + 903, + 440 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this, we introduce Video Motion Benchmark (VMBench) (Fig. 1), the first benchmark focused on comprehensive video motion assessment. It features the most diverse range of motion types, covering 969 categories—more than any other benchmark [27, 33, 44, 49, 50, 68] (see Fig. 12 in the Appendix). Additionally, it integrates perception-aligned motion metrics to systematically evaluate the motion generation quality of video generation models.", + "bbox": [ + 511, + 443, + 905, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, we introduce Perception-Driven Motion Evaluation Metrics (PMM), the first evaluation framework explicitly designed to align with human perception of motion quality. PMM comprises five key components: Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), Temporal Coherence Score (TCS), Motion Smoothness Score (MSS), and Commonsense Adherence Score (CAS). Unlike previous methods that primarily focus on motion smoothness [29, 59, 64, 75], PMM provides a more comprehensive assessment by evaluating motion quality in terms of spatiotemporal inconsistencies and violations of physical laws.", + "bbox": [ + 511, + 580, + 905, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Next, we propose Meta-Guided Motion Prompt Generation (MMPG), an expandable framework that encompasses the most comprehensive range of motion types, grounded in physics [56, 60, 65, 67] and cognitive science [25, 57]. The evaluation system covers six movement modes: fluid dynamics, biological motion, mechanical motion, weather phenomena, collective behavior, and energy transfer (Fig. 1). Specifically, MMPG first extracts subjects, places, and actions from VidProm [79], Place365 [97], and", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 109, + 875, + 240, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13087", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/882ee89ebe4bd8400faee5b6eaf319fa9c619c99fd8df1847b437cbbcc62113c.jpg", + "image_caption": [ + "Figure 1. Overview of VMBench. Our benchmark encompasses six principal categories of motion patterns, with each prompt constructed as a comprehensive motion structured around three core components: subject, place, and acion. We propose a novel multi-dimensional video motion evaluation comprising five human-centric quality metrics derived from perceptual preferences. Utilizing videos generated by popular T2V models, we conduct systematic human evaluations to validate the effectiveness of our metrics in capturing human perceptual preferences." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 524, + 361 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a01b5b425fa47b0421a5d841e87209c627d28f0e20880ff739e94e7f5fd188a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 89, + 857, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "other datasets [8, 10, 14, 89], then leverages LLMs to enrich metadata and generate detailed motion descriptions. We validate the prompts through human-AI collaboration with DeepSeek-R1 [21] to ensure coherence and rationality. To reduce costs, we curated 1,050 user prompts and use T2V models to generate motion videos for evaluation.", + "bbox": [ + 89, + 468, + 482, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we provide human preference annotations to validate our benchmarks, showing that our metrics achieve an average $35.3\\%$ improvement in Spearman's correlation [93] over current methods [9, 13, 33, 50, 59, 73, 81, 84, 95]. This result highlights the enhanced alignment of our evaluation metrics with human perception. We will open-source VMBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VMBench to drive forward the field of video motion generation. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 571, + 483, + 736 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First-Ever Human Perception-Aligned Motion Evaluation: We pioneer the evaluation of motion quality in videos from the perspective of human perception alignment.", + "Meta-Guided Motion Prompt Generation: A curated prompt set designed to assess diverse motion aspects in video models.", + "- Human-Aligned Validation Mechanism: Our metrics achieve an average $35.3\\%$ improvement in Spearman's correlation over current methods." + ], + "bbox": [ + 89, + 750, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 467, + 651, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Video Generative Models", + "text_level": 1, + "bbox": [ + 513, + 494, + 743, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rapid advancement of AI generative techniques—GANs [22, 26, 36, 41, 52], VAEs [41, 83], autoregressive models [17, 63, 92], and diffusion models [7, 31, 32, 34, 76, 88]—the field of T2V generation has seen remarkable progress. However, while current T2V models [16, 70, 90, 96], excel in generating static visual content, their dynamic performance still deviates significantly from human perception.", + "bbox": [ + 511, + 518, + 906, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early T2V models [15, 30, 31, 66] have built upon Text-to-Image(T2I) techniques, enhancing their temporal capabilities to generate videos. Then, some works [28, 80, 98] take this further by conducting the diffusion process within the latent space and incorporate the spatiotemporal module [16, 76, 86] within the denoising network, which greatly improves the quality of static content in videos. CogVideo [32, 90] and other closed source works [1, 5, 12, 58] represented by Sora [2] focus on high-quality data filtering on a large-scale dataset, significantly boosting model performance. OpenSora [96] and OpenSora-Plan [45] aim to showcase and open source the remarkable performance of closed-source models. Mochi 1 [70] and HunyuanVideo [37] have improved the dynamic quality of generated videos through new technologies such as asymmetric architecture and full attention mechanism. Beyond optimization of models and datasets, recent stud", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13088", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1d85feb677bc7685884399a155860b34dd19e1eafed2b4e280979d84d7fe012.jpg", + "image_caption": [ + "Figure 2. Framework of our Perception-Driven Motion Metrics (PMM). PMM comprises multiple evaluation metrics: Commonsense Adherence Score (CAS), Motion Smoothness Score (MSS), Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), and Temporal Coherence Score (TCS). (a-e): Computational flowcharts for each metric. The scores produced by PMM show variation trends consistent with human assessments, indicating strong alignment with human perception." + ], + "image_footnote": [], + "bbox": [ + 93, + 92, + 903, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ies [3, 5, 40, 46, 51] are beginning to explore the influence of motion on video generation. Specifically, Ma et al. [51] focuses on the motion quality of videos during data selection, in order to enable models to learn motion dynamics even at low resolution training stages. Meanwhile, some works [40, 46] improve the generative capacity of models by incorporating motion information as guidance during the sampling process.", + "bbox": [ + 88, + 501, + 480, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although these models demonstrate commendable performance across various metrics, the generated videos still exhibit several qualitative motion deficiencies, such as the abrupt disappearance of moving subjects, body distortions, and physically implausible actions, which are difficult to assess using existing metrics.", + "bbox": [ + 89, + 623, + 482, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Evaluations Metrics for Video Generation", + "text_level": 1, + "bbox": [ + 89, + 724, + 447, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing approaches for video motion evaluation can be broadly categorized into three paradigms: 1) feature-based metrics leveraging pre-trained video representations, 2) rule-based metrics employing manually designed scoring mechanisms, and 3) MLLM-based assessments that fine-tune large multimodal language models on human-annotated video quality datasets for perceptual scoring.", + "bbox": [ + 89, + 748, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous feature-based metrics demonstrate limited effectiveness in assessing motion quality. Image-centric metrics like IS [64] and FID [29] inherently ignore tempo", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ral coherence, while FVD [75] extends image-level metrics to videos by modeling spatiotemporal features, it reduces motion dynamics to simplistic distributions. These methods conflate feature-space distances with perceptual motion quality, overlooking artifacts like implausible movement.", + "bbox": [ + 511, + 501, + 903, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent benchmarks [24, 33, 50] attempt to conduct systematic evaluations using manually curated metrics. For instance, VBench [33] introduces motion-related criteria such as motion smoothness, quantified via a video frame interpolation model. However, such rule-driven metrics face two inherent limitations in motion assessment: 1) Subjectivity in metric design—predefined rules poorly generalize to diverse motion patterns beyond designer priors; 2) Fragmentary analysis—isolated measurements of individual attributes fail to capture holistic motion quality. Consequently, these heuristics overlook complex phenomena like inertia preservation or multi-object interaction coherence.", + "bbox": [ + 511, + 580, + 905, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emerging methods [11, 27] employ multimodal large language models (MLLMs) fine-tuned on human preference data to predict perceptual video quality scores. VideoScore [27] aligns scoring with human judgments through preference learning, while VideoPhy [11] specifically targets physical plausibility via physics-focused annotations. Despite improved alignment with subjective evaluations, these approaches exhibit critical limitations for motion assessment: 1) Oversimplified scoring granularity – scalar outputs", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13089", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "obscure specific motion defects; 2) Annotation bias amplification – training data biases toward conspicuous artifacts neglect subtle kinematic violations; 3) Contextual rigidity – domain-specific tuning limits generalization to diverse motion types. While advancing automated evaluation, they struggle to disentangle motion quality from global preference impressions.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current evaluation paradigms remain fundamentally misaligned with human perception of motion quality due to the above limitations. To bridge this gap, we introduce 5 distinct human-aligned evaluation metrics that systematically quantify motion quality.", + "bbox": [ + 89, + 196, + 480, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. VMBench", + "text_level": 1, + "bbox": [ + 89, + 285, + 200, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce VMBench, a comprehensive Video Motion Benchmark that includes perception-aligned motion metrics and features the most diverse range of motion types.", + "bbox": [ + 89, + 310, + 480, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Perception-Driven Motion Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 364, + 478, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When observing videos, the human brain first constructs a holistic understanding of the scene based on prior experiences and physical laws [23], before selectively attending to moving objects to assess motion smoothness and temporal consistency, particularly when objects are momentarily occluded. While rapid movement can effectively capture attention, excessive deviations from typical motion patterns may induce visual disorientation. As illustrated in Fig. 5 in the Appendix, our motion evaluation metrics are designed to emulate this hierarchical perceptual process, transitioning from global scene comprehension to localized motion analysis to ensure coherence and logical consistency.", + "bbox": [ + 89, + 387, + 482, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Commonsense Adherence Score (CAS) is used to evaluate whether generated videos align with human commonsense. As shown in Fig. 2 (a), we train a specialized model capable of assessing the commonsense quality of video content, categorizing it into five levels: Bad, Poor, Fair, Good, and Perfect. First, we collect a dataset comprising $10\\mathrm{k}$ generated videos, covering both legacy approaches and popular models [2-4, 15, 16, 37, 45, 90, 96]. Second, we establish perceptual ground truth through systematic pairwise comparisons using VideoReward [47], a reward model trained on $182\\mathrm{k}$ human-annotated video preference pairs. This process generates normalized preference scores reflecting human perceptual consensus. Third, we discretize the human preference into five cognitive labels. Using these preference labels, we develop a video classification model to serve as the final CAS predictor, utilizing VideoMAEv2 [77], a spatiotemporal vision transformer. Compared to MLLMs, this architecture demonstrates temporal modeling capabilities through dense frame sampling and 3D convolution operations. We compute the CAS using a Mean Opinion Score (MOS) [74, 85], where predicted probabilities for each class are weighted by their corresponding quality co", + "bbox": [ + 89, + 569, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "efficients: $\\mathrm{CAS} = \\sum_{i=1}^{5} p_i G(i)$ , where $p_i$ denotes the predicted probability for the $i$ -th class, and $G_i$ represents the mapping function that converts category to quality weights: $G : i \\to w$ (e.g., mapping class 3 to 0.5). For more details, please refer to our Appendix B.1.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Smoothness Score (MSS) aims to detect unsmoothness, like low-level temporal artifacts and high-level motion blur. It addresses the limitations of prior metrics that suffered from either low-level optical flow bias [38, 39] or oversimplified motion modeling [42], both leading to misalignment with human perception. Considering that natural motion patterns inherently maintain consistent visual quality across temporal sequences, we leverage Q-Align's [85] aesthetic score to detect artifacts, as illustrated in Fig. 2 (b). An artifact frame is confirmed when a significant drop in Q-Align scores between consecutive frames exceeds a predefined threshold, mimicking human perception of sustained visual anomalies. Further, our approach employs a scene-aware adaptive thresholding mechanism derived from the statistical modeling of real video segments [14, 54] across diverse motion patterns. The final MSS is computed as: $\\mathrm{MSS} = 1 - \\frac{1}{T}\\sum_{t=2}^{T}\\mathbb{I}(\\Delta Q_t > \\tau_s(t))$ where $\\Delta Q_t$ is the frame-to-frame visual quality degradation magnitude, $\\tau_s(t)$ indicates the adaptive threshold and $\\mathbb{I}$ is an indicator function. More details are provided in Appendix B.2.", + "bbox": [ + 511, + 167, + 903, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Integrity Score (OIS) detects implausible deformations through spatiotemporal analysis of object integrity. Our approach addresses limitations found in previous methods [13], which primarily focus on object-level semantic consistency. Instead, we focus on detecting perceptual issues (e.g., distorted shapes) that are readily noticeable to the human visual system. As illustrated in Fig. 2 (c), we first utilize the MMPose toolkit [20] to detect key points of the primary subjects in the generated videos, which are then used to estimate the subjects' shapes in each frame. Then, we quantify shape distortions by analyzing whether object shapes violate real-world anatomical constraints. We establish tolerance thresholds for the above constraints through the statistical analysis of natural motion samples [14, 54, 87]. Anatomical violations across consecutive frames will be detected; see more details about anatomical constraints in our appendix B.3. The OIS quantifies structural consistency through: $\\mathrm{OIS} = \\frac{1}{F.K}\\sum_{f = 1}^{F}\\sum_{k = 1}^{K}\\mathbb{I}\\left(\\mathcal{D}_f^{(k)}\\leq \\tau^{(k)}\\right)$ , in which $F$ and $K$ denote total frame count and anatomical components, respectively. $\\mathcal{D}_f^{(k)}$ represents the compound anatomical deviation for component $k$ in frame $f$ , and $\\tau^{(k)}$ indicates statistical thresholds.", + "bbox": [ + 511, + 468, + 903, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perceptible Amplitude Score (PAS) estimates subject motion by separating it from camera motion, unlike traditional motion magnitude estimation methods [73], which often overestimate overall motion due to the influence of camera movement. As presented in Fig. 2 (d), PAS begins with semantic anchoring via GroundingDINO [48] to lo", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13090", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "calize moving subjects and distinguish them from passively moving elements. Next, GroundedSAM [62] ensures temporally stable subject masks, enabling CoTracker [35] to track key points with high precision. The motion magnitude is then computed based on the average displacement of these key points. Furthermore, PAS accounts for the context-dependent nature of human motion perception, adapting thresholds across different scenarios. To quantify this variability, we derive a set of perceptual motion magnitude thresholds for various scenarios through the statistical analysis of existing video datasets [14, 54]. These thresholds serve as the foundation for computing a motion score for each video. The PAS is computed as PAS = $\\frac{1}{T}\\sum_{t=1}^{T}\\min\\left(\\frac{\\bar{D}_t}{\\tau_s}, 1\\right)$ , where $\\bar{D}_t$ denotes frame-level motion amplitude, computed as the average displacement of tracked key points for active subjects in frame $t$ and $\\tau_s$ is the perceptual motion threshold for scenario $s$ . More Details can be found in appendix B.4.", + "bbox": [ + 89, + 90, + 480, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Coherence Score (TCS) detects the temporal consistency of object motion by analyzing frame-to-frame transitions, addressing the limitation of existing metrics [13, 59] that struggle to differentiate between natural movement and abrupt, unrealistic changes. We combine video object tracking with rule-based validation to detect abnormal disappearance/reappearance patterns, thereby distinguishing natural progression from temporal discontinuities. As shown in Fig. 2 (e), first GroundedSAM2 [61] performs pixel-accurate instance segmentation and tracking across frames, maintaining persistent object IDs throughout the whole sequence. For objects exhibiting discontinuous existence, we apply a secondary verification phase using the CoTracker [35]. It tracks dense key points on target objects and constructs their motion trajectories. We then analyze their motion trajectories to determine whether any anomalous phenomena are present. Our approach mitigates false cases caused by legitimate object discontinuity through a rule-based filtering mechanism implemented via trajectory analysis. These rules account for common scenarios, including 1) objects reappearing after occlusion or disappearing behind obstacles, 2) objects entering or exiting frame boundaries, and 3) apparent size changes due to depth perception, such as objects appearing larger when moving closer or smaller when moving farther away. Through this filtering, we systematically exclude normal events. The TCS quantifies anomalous discontinuities via $\\mathrm{TCS} = 1 - \\frac{1}{N}\\sum_{i=1}^{N}\\mathbb{I}(\\mathcal{A}_i\\wedge \\neg \\mathcal{R})$ , where $N$ is the total object instances, $\\mathcal{A}_i$ indicates the existence status of the $i$ -th object, and $\\mathcal{R}$ validates legitimate transitions (see Appendix B.5).", + "bbox": [ + 89, + 369, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Meta-Guided Motion Prompt Generation", + "text_level": 1, + "bbox": [ + 89, + 848, + 444, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The existing benchmarks [27, 33, 50] are constrained by limited and simplistic motion types (As shown in Fig. 12", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the Appendix), making them inadequate for comprehensively evaluating the motion generation capabilities of video generation models. To address this limitation, we propose an extensible framework—Meta-Guided Motion Prompt Generation (MMPG)—that generates prompts capturing complex motion patterns, as illustrated in Fig. 3. Built upon MMPG, we introduce VMBench, a benchmark specifically designed for rigorous evaluation of complex motion generation. We find that VMBench provides the most comprehensive coverage of motion types and the most detailed prompt descriptions, making it an effective benchmark for evaluating the dynamic motion generation capabilities of video generation models. In this section, we elaborate on the process of prompt generation and refinement. More details are available in Appendix C.1.", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meta-information Extraction. We decompose motion descriptions into three key metadata elements: Subject $(S)$ , Location $(P)$ and Action $(A)$ . We employ Qwen-2.5 [71] to extract the metadata library $\\{S, P, A\\}$ from existing videotext pairs [18, 58, 79, 89]. To ensure novel prompts and enhance textual diversity, we expand the metadata in three key aspects. For subject $(S)$ classification, we categorize subject types into human, animal, and object, curate a list of principal nouns identifiable by [48, 62], and use GPT-4o [6] to generate descriptions with varying entity counts (1, 2, n). For place $(P)$ descriptions, we incorporate data from [97], filtering out redundant information. For action $(A)$ expansion, we sample human actions from [14] and use LLM to extend possible actions for animals and objects.", + "bbox": [ + 511, + 321, + 903, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-Refining Prompt Generation. We randomly sample elements from the metadata library $\\{S, P, A\\}$ to form metadata sets $(S^i, P^j, A^k)$ and use GPT-4o to assess their semantic coherence and logical consistency. For coherent sets, we generate prompts of 10 to 60 words that describe the corresponding motion scenarios. To ensure accuracy, we introduce a verification mechanism inspired by [50], where the generated prompts are re-evaluated by GPT-4o to verify their consistency with the metadata elements. Through an iterative refinement and filtering process, we generate approximately 50,000 high-quality prompt candidates. However, grammatical correctness alone is insufficient: prompts must also align with physical reality to effectively assess video content. Therefore, we also eliminate descriptions of unrealistic or physically implausible scenarios to ensure the feasibility and authenticity of the prompts.", + "bbox": [ + 511, + 535, + 903, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human-LLM Joint Validation. To reduce manual labor and enhance evaluation effectiveness, we further examine and filter prompts through a Human-LLM Joint Validation process. This process combines the strengths of both automated reasoning and human judgment to ensure high-quality motion descriptions. We first leverage the powerful reasoning capabilities of Deepseek R1 [21] to evaluate whether each prompt describes a reasonably realistic mo", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13091", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a1b1654da10edf229bed4f040a00169ee24b25531f71a963aa579b133ba64c0.jpg", + "image_caption": [ + "Figure 3. Framework of our Meta-Guided Motion Prompt Generation (MMPG). MMPG consists of three stages: (a) Meta-information Extraction: Extracting Subjects, Places, and Actions from datasets such as VidProm [79], Didemo [8], MSR-VTT [89], WebVid [10], Place365 [97], and Kinect-700 [14]. (b) Self-Refining Prompt Generation: Generating and iteratively refining prompts based on the extracted information. (c) Human-LLM Joint Validation: Validating the prompts through a collaborative process between humans and DeepSeek-R1 to ensure their rationality." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 893, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion. After filtering out implausible descriptions, we curate a diverse set of 1,050 prompts with varied metadata. Examples are available in Appendix C.2.", + "bbox": [ + 89, + 387, + 482, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 446, + 223, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 470, + 279, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our benchmark evaluates six popular text-to-video models: OpenSora [96], CogVideoX [90], OpenSora-Plan [45], Mochi 1 [70], HunyuanVideo [37], and Wan2.1 [72]. To provide a richer variety of motion types, we develop the MMPG-set, which comprises 1,050 prompts based on MMPG across six movement modes to assess performance in motion generation. Each model generates 1,050 videos corresponding to the MMPG-set, resulting in a total of 6,300 videos. To ensure a fair comparison, we maintain the hyperparameter settings as defined in each model's project demo. For each prompt, we generate one video for evaluation using only the initial seed. The inference process is executed using 8 Nvidia H20 GPUs. Details of the inference process for each model can be found in Appendix D.1. We then randomly sample 200 videos from each model's output, resulting in a total of 1,200 videos for human-aligned validation experiments.", + "bbox": [ + 89, + 493, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Metrics. The current comparison approaches fall into two categories: Rule-based metrics and Multimodal Large Language Model (MLLM) prompting. (1) Rule-based metrics assess four dimensions, including Perceptible Amplitude, which evaluates through RAFT optical flow magnitude analysis combined with structural motion consistency via 4 frame SSIM averaging [73, 82], following established protocols [33]. Temporal Coherence is measured using DINO [13] and CLIP [59] feature tracking, which relies on cosine similarity between consecu", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tive frames. Motion Smoothness is assessed through a hybrid method that combines interpolation error [42] with Dover's video quality assessment [84], while Object Integrity is evaluated via dual validation using optical flow warping error [38, 39] and semantic consistency checks. (2) The MLLM evaluation includes five cutting-edge models: LLaVA-NEXT-Video [95], MiniCPM-V-2.6 [91], InternVL2.5 [19], Qwen2.5-VL [9], and InternVideo2.5 [81]. These models are evaluated through a standardized assessment process, which involves processing uniformly sampled video frames at 2 frames per second to maintain essential motion patterns while managing computational limits. The evaluation dimensions encompass five aspects: amplitude, coherence, integrity, smoothness, and commonsense adherence, with scoring for each dimension using a 1–5 scale. In particular, to ensure fair comparison, consistent frame sequences and evaluation criteria are maintained across all models.", + "bbox": [ + 511, + 387, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. 1) Spearman correlation [93]: Spearman's rank correlation coefficient, often denoted by $\\rho$ , is a measure of the strength and direction of the association between two ranked variables. It assesses how well the relationship between two variables can be described by a monotonic function. Spearman correlation is nonparametric and can effectively evaluate associations in datasets where variables may not follow a normal distribution. Unlike Pearson correlation, which captures linear relationships, Spearman correlation focuses on rank-based associations, making it more robust to outliers and suitable for ordinal data or scenarios with nonlinear dependencies. 2) Accuracy: To validate how accurately our metrics align with human preferences, we conduct pairwise comparisons across the 1,200 annotated videos (200 prompts $\\times$ 6 models). For each prompt, we", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13092", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9cefc3538adc84975ffc2158ee8d6c51cf5b08d5050d0d3f0f97649ab4547f48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAvg. ρ(↑)CAS ρ(↑)MSS ρ(↑)OIS ρ(↑)PAS ρ(↑)TCS ρ(↑)
Rule-based
SSIM [82]1.6-0.9-12.18.317.8-4.8
RAFT* [73]-1.7-0.7-17.0-16.647.7-21.9
CLIP* [59]15.021.536.531.7-42.728.0
DINO* [13]9.411.027.227.4-45.827.4
Dover Technical† [84]20.640.232.634.5-6.22.2
AMT* [33]3.62.818.119.6-40.417.8
Warping Error† [50]-6.0-11.4-19.1-9.033.0-23.6
MLLM Prompt
LLaVa-NEXT-Video [95]17.115.211.016.130.812.4
MiniCPM-V-2.6 [91]16.26.817.5-2.227.331.6
IntenVL2.5 [19]19.318.720.514.142.51.1
Qwen2.5-VL [9]15.324.719.83.714.214.0
InternVideo2.5 [81]26.922.721.929.644.315.8
Model-Rule Synergy
PPM (Ours)62.269.977.165.865.254.5
", + "bbox": [ + 96, + 88, + 478, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluate all 15 possible video pairs generated by different models (total 3,000 pairs). Human preference labels are determined by comparing the average expert ratings across all five dimensions (OIS, MSS, CAS, TCS, PAS) – the video with higher aggregated human scores is designated as the “ground truth” preferred sample. Concurrently, we compute metric preferences by comparing videos' aggregated PMM scores under identical criteria. Alignment accuracy is quantified as the percentage of pairs where PMM preferences match human judgments, with ties excluded to prioritize unambiguous decisions.", + "bbox": [ + 88, + 476, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 89, + 652, + 387, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human-aligned Validation Mechanism. We invite three domain experts to independently annotate each sample based on PMM (Perceptible Amplitude, Temporal Coherence, Object Integrity, Motion Smoothness, and Common-sense Adherence), resulting in 6,000 detailed ratings with high interannotator agreement. Finally, we obtain Likert scores [53] from humans for each video across the five dimensions included in PMM. More details can be found in Appendix D.2. We assess the alignment between metrics and human perception by calculating the Spearman correlation [93] between metric scores and expert ratings. A higher Spearman correlation signifies a stronger alignment with human perception.", + "bbox": [ + 89, + 672, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Alternative Metrics. As shown in Table 1, in the MSS evaluation, even advanced metrics", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bbeab31a344657c845835b5e40dfad34f57fe77eecef617d2dc17cc3dcc88568.jpg", + "table_caption": [ + "Table 1. Correlation Analysis Between Evaluation Metrics and Human Scores via Spearman Correlation Coefficient ( $\\rho \\times 100$ ). Superscript * and † denote implementations following VBenchmark [33] and EvalCrafter [50] respectively. Yellow backgrounds in Rule-based indicate specific dimension baseline." + ], + "table_footnote": [], + "table_body": "
CASMSSOISPASTCSAccuracy (%) ↑
Removal-based ablation
✓✓✓✓✓70.6
✓✓✓✓✗66.9
✓✓✓✗✓68.7
✓✓✗✓✓64.6
✓✗✓✓✓65.2
✗✓✓✓✓64.1
Addition-based ablation
✗✗✗✗✗-
✓✗✗✗✗58.9
✓✓✗✗✗66.1
✓✓✓✗✗67.3
", + "bbox": [ + 537, + 88, + 883, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation Study on our Motion Metrics. Prediction accuracy (\\%) of different metric combinations against human preferences is calculated. Removal-based ablation show the effect of individually ablating each metric, while in the addition-based ablation, we progressively add each metric to observe its impact.", + "bbox": [ + 511, + 306, + 906, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "such as AMT [42] (18.1%) and Warping Error [39] (- 19.1%) demonstrate limited discriminative power and produce counterintuitive results, particularly in the context of complex deformations. The OIS evaluation also highlights similar shortcomings, with DINO [13] at 27.4% and Dover [84] at 34.5%, both of which fail to capture human sensitivity to structural preservation during motion. In relation to the evaluation of PAS, the SSIM [82] and RAFT [73] metric based on rules show an alignment efficacy of only 17.8% and 47.7%, respectively. In contrast, our approach achieves a remarkable 65.2% alignment efficacy. For the TCS evaluation, the rule-based metrics CLIP [59] and DINO [13] achieve only 28.0% and 27.4% alignment efficacy, respectively, and fail to capture human tolerance for minor inconsistencies while maintaining physical plausibility. In comparison, our method achieves an impressive 54.5% alignment efficacy. VBench [33] includes RAFT, CLIP, DINO, and AMT, while EvalCrafter [50] incorporates Dover Technical and Warping Error. According to the data presented in the table, overall, compared to ours, the motion evaluations of VBench [33] and EvalCrafter [50] reveal significantly lower correlations with human perception. MLLMs exhibit partial competence in evaluations of the Physical Adequacy Score (PAS) (e.g., InternVideo2.5: 44.3%) Despite their general capabilities, MLLMs average only a correlation of 10.0%-30.0% in all dimensions, highlighting a fundamental misalignment with human perception in evaluating motion quality.", + "bbox": [ + 511, + 409, + 908, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 845, + 666, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study of Motion Metrics. According to Table 2, removing any single metric from the evaluation of the", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13093", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "basic variants results in a significant decline in accuracy, which highlights the importance of each dimension's evaluation metric within the overall framework. Notably, the removal of the CAS metric leads to the most pronounced performance drop, with accuracy falling to $64.1\\%$ , exceeding the impact of other single-dimension ablations. This demonstrates the critical role of the CAS metric in assessing video quality, aligning closely with the key factors prioritized by humans when perceiving video quality. For the performance-oriented variants, we emulate the human perceptual information processing flow by gradually adding evaluation metrics (refer to Fig. 5 in the Appendix), with each additional metric significantly enhancing overall accuracy. This not only demonstrates the effectiveness of incrementally increasing evaluation dimensions but also further verifies the consistency of our evaluation approach with human perceptual mechanisms.", + "bbox": [ + 89, + 90, + 483, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 371, + 284, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alignment of PMM with Human Perception. As shown in Fig. 4, the correlation between human scores across these five evaluation dimensions and our evaluation metrics in the same five dimensions remains consistent. For example, both demonstrate a strong correlation between OIS, CAS, and MSS, as well as a weak correlation between PAS and the other metrics. Specifically, as shown in Fig. 4 (a), PAS has a negative correlation with other dimensions, such as the Object Integrity Score (OIS), where $\\rho = -0.18$ . This might be because high dynamic amplitudes in videos cause distortions and artifacts, which lower structural and temporal coherence scores. Conversely, OIS shows strong positive correlations with MSS ( $\\rho = 0.59$ ) and CAS ( $\\rho = 0.50$ ), suggesting that it reflects physical plausibility and motion rationality well. TCS shows low correlation with other dimensions, which can provide a more comprehensive assessment perspective. PAS and structural/temporal metrics exhibit a negative correlation, which challenges conventional optical-flow-based evaluation frameworks. Additionally, the isolation of PAS highlights the importance of separately disentangling motion magnitude in motion video evaluation. Fig. 4 (b) shows the correlations among the proposed evaluation metrics, which align with human perception.", + "bbox": [ + 89, + 397, + 483, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assessing Video Generation Models with PMM. As shown in Table 3, we evaluate several leading video generation models using our PMM metrics, including Mochi 1[70], OpenSora [96], CogVideoX [90], OpenSora-Plan [45], HunyuanVideo [37], and Wan2.1 [72]. Our findings indicate that Wan2.1 delivers the best performance in generating motion videos, producing results that appear more realistic. For a clearer visual comparison of the dynamic videos generated by each model, please refer to Fig. 16 in the Appendix.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/975c5962c443a53c7c3a37e190d605d199e929e99b31e7066a2e2ba34042ad77.jpg", + "image_caption": [ + "Figure 4. Correlation Matrix Analysis of Metrics Within Different Evaluation Mechanisms. (a): Spearman Correlation Matrices for human annotations; (b): Spearman Correlation Matrices for our PMM metrics." + ], + "image_footnote": [], + "bbox": [ + 540, + 87, + 712, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f7435a549138d51437d9fbeafb81157ecb59b2cf65f916d5f05647b8f3d54f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 88, + 890, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65dc4a204c9616f89feb24349203a05635e6e72b720562ff35091703e990f9c1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsAvg.(%) ↑CASMSSOISPASTCS
Mochi 1[70]53.237.762.068.614.483.6
OpenSora [96]51.631.261.973.03.488.5
CogVideoX [90]60.650.661.675.424.691.0
OpenSora-Plan [45]58.939.376.078.66.094.7
HunyuanVideo [37]63.451.981.665.826.196.3
Wan2.1 [72]78.462.884.266.017.997.8
", + "bbox": [ + 514, + 303, + 906, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Performance of Video Generation Models on VMBench. We evaluated six open-source video generate models [37, 45, 70, 72, 90, 96] using VMBench. A higher score indicates better performance for a category.", + "bbox": [ + 511, + 424, + 906, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 510, + 633, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video motion authenticity remains a critical challenge in the creation of generate content. In response, we introduce VMBench, the first open-source benchmark for motion quality evaluation, integrating motion metrics with human-aligned assessment to reveal deficiencies in existing models' ability to generate physically plausible movements. To support the research community, we offer three essential resources: 1) a standardized framework for detecting motion artifacts overlooked by traditional metrics; 2) actionable diagnostic tools aimed at guiding model optimization; and 3) quantitative standards that align technical advancements with human perceptual expectations. By establishing a common framework for development in this field, VMBench enables systematic tracking of motion generation, encourages targeted model improvements, and ultimately supports the creation of videos that achieve both visual fidelity and dynamic realism. However, while the benchmark's evaluation metrics are aligned with general human perception, they may not fully capture subtle differences in perception that stem from individual viewer experiences and preferences. Addressing these limitations presents opportunities for further refinement, aiming to ensure a more comprehensive and adaptable approach to video quality evaluation.", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13094", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Runway gen3. Accessed February 25, 2025 [Online] https://app.runwayml.com/, 2025.2", + "[2] Sora. Accessed February 25, 2025 [Online] https://openai.com/index/sora/, 2025.1,2,4", + "[3] kling. Accessed February 25, 2025 [Online] https://klingai.com/, 2025.1, 3", + "[4] Pika labs. Accessed February 25, 2025 [Online] https://www.pika.art/, 2025.4", + "[5] Veo 2. Accessed February 25, 2025 [Online] https://deepmind.google/technologies/veo/veo-2/, 2025.2,3", + "[6] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5", + "[7] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2", + "[8] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2, 6", + "[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7", + "[10] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6", + "[11] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 3", + "[12] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 2", + "[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 5, 6, 7", + "[14] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 2, 4, 5, 6", + "[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan." + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Videocrafter1: Open diffusion models for high-quality video generation, 2023. 2, 4", + "[16] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models, 2024. 2, 4", + "[17] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2", + "[18] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 5", + "[19] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6, 7", + "[20] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4", + "[21] DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. 2, 5", + "[22] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2", + "[23] Daniel J. Felleman and David C. van Essen. Distributed hierarchical processing in the primate cerebral cortex. *Cerebral Cortex*, 1 1:1-47, 1991. 4", + "[24] Xiaokun Feng, Haiming Yu, Meiqi Wu, Shiyu Hu, Jintao Chen, Chen Zhu, Jiahong Wu, Xiangxiang Chu, and Kaiqi Huang. Narrlv: Towards a comprehensive narrative-centric evaluation for long video generation models. arXiv preprint arXiv:2507.11245, 2025. 3", + "[25] Deborah M Gordon. The ecology of collective behavior. PLoS biology, 12(3):e1001805, 2014. 1", + "[26] Shir Gur, Sagie Benaim, and Lior Wolf. Hierarchical patch vae-gan: Generating diverse videos from a single sample. Advances in Neural Information Processing Systems, 33: 16761-16772, 2020. 2", + "[27] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv preprint arXiv:2406.15252, 2024. 1, 3, 5", + "[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2", + "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13095", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2", + "[31] Jonathan Ho, Tim Salimans, Alexey Gritsanko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2", + "[32] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[33] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 1, 2, 3, 5, 6, 7", + "[34] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22747-22757, 2023. 2", + "[35] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024.5", + "[36] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 8:153113-153122, 2020. 2", + "[37] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 4, 6, 8", + "[38] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 4, 6", + "[39] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. Advances in Neural Information Processing Systems, 33:1083-1093, 2020. 4, 6, 7", + "[40] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3", + "[41] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2", + "[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 4, 6, 7", + "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 1", + "[44] Mingxiang Liao, Qixiang Ye, Wangmeng Zuo, Fang Wan, Tianyu Wang, Yuzhong Zhao, Jingdong Wang, Xinyu Zhang, et al. Evaluation of text-to-video generation models: A dynamics perspective. Advances in Neural Information Processing Systems, 37:109790-109816, 2025. 1", + "[45] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2, 4, 6, 8", + "[46] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 3", + "[47] Jie Liu, Gongye Liu, Jiajun Liang, Ziyang Yuan, Xiaokun Liu, Mingwu Zheng, Xiele Wu, Qiulin Wang, Wenyu Qin, Menghan Xia, et al. Improving video generation with human feedback. arXiv preprint arXiv:2501.13918, 2025. 4", + "[48] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 5", + "[49] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. Advances in Neural Information Processing Systems, 36:62352-62387, 2023. 1", + "[50] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 1, 2, 3, 5, 7", + "[51] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025.3", + "[52] Rayeesa Mehmood, Rumaan Bashir, and Kaiser J Giri. Vtmgan: video-text matcher based generative adversarial network for generating videos from textual description. International Journal of Information Technology, 16(1):221-236, 2024. 2", + "[53] Tomoko Nemoto and David Beglar. Likert-scale questionnaires. In JALT 2013 conference proceedings, pages 1-6, 2014. 7", + "[54] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Pro" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13096", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 4, 5", + "[55] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1", + "[56] TA Palomaki, JD Teufel, RW Simmonds, and Konrad W Lehnert. Entangling mechanical motion with microwave fields. Science, 342(6159):710-713, 2013. 1", + "[57] Marina A Pavlova. Biological motion processing as a hallmark of social cognition. Cerebral cortex, 22(5):981-995, 2012. 1", + "[58] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, et al. Movie gen: A cast of media foundation models, 2025. 2, 5", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5, 6, 7", + "[60] Ralph H. Rasshofer, Martin Spies, and Hans Spies. Influences of weather phenomena on automotive laser radar systems. Advances in radio science, 9:49-60, 2011. 1", + "[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5", + "[62] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 5", + "[63] Zhongwei Ren, Yunchao Wei, Xun Guo, Yao Zhao, Bingyi Kang, Jiashi Feng, and Xiaojie Jin. Videoworld: Exploring knowledge learning from unlabeled videos. arXiv preprint arXiv:2501.09781, 2025. 2", + "[64] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 1, 3", + "[65] Benjamin Scellier and Yoshua Bengio. Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in computational neuroscience, 11:24, 2017. 1", + "[66] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[67] Jos Stam. Stable fluids. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 779-786. 2023. 1" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 1", + "[69] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 1", + "[70] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 8", + "[71] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5", + "[72] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 6, 8", + "[73] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 4, 6, 7", + "[74] I Telecom. Recommendation 500-10: Methodology for the subjective assessment of the quality of television pictures. ITU-R Rec. BT, 500, 2000. 4", + "[75] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 1, 3", + "[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report, 2023. 2", + "[77] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14549-14560, 2023. 4", + "[78] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. arXiv preprint arXiv:2307.00040, 2(3):4, 2023. 1", + "[79] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. 2024. 1, 5, 6", + "[80] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. Magicvideo-v2: Multi-stage high-aesthetic video generation, 2024. 2", + "[81] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 6, 7", + "[82] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7", + "[83] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Gen" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13097", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "erating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2", + "[84] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 2, 6, 7", + "[85] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 4", + "[86] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation, 2023. 2", + "[87] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 4", + "[88] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. 2023. 2", + "[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 5, 6", + "[90] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 4, 6, 8", + "[91] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6, 7", + "[92] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 2", + "[93] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 2, 6, 7", + "[94] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 1", + "[95] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 6, 7", + "[96] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 1, 2, 4, 6, 8" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[97] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 1, 5, 6", + "[98] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + ], + "bbox": [ + 514, + 90, + 903, + 204 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13098", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_model.json b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8bf0d7ce22513a4b29af7163e2c4adde26f4425 --- /dev/null +++ b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_model.json @@ -0,0 +1,2424 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.878, + 0.152 + ], + "angle": 0, + "content": "VMBench: A Benchmark for Perception-Aligned Video Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.18, + 0.264, + 0.217 + ], + "angle": 0, + "content": "Xinran Ling\\(^{1*}\\) \nCundian Yang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.181, + 0.401, + 0.217 + ], + "angle": 0, + "content": "Chen Zhu1* \nAiming Hao1" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.181, + 0.54, + 0.217 + ], + "angle": 0, + "content": "Meiqi Wu\\(^{1,3*}\\) \nJiashu Zhu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.578, + 0.181, + 0.692, + 0.217 + ], + "angle": 0, + "content": "Hangyu Li1 \nJiahong Wu1†" + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.181, + 0.865, + 0.217 + ], + "angle": 0, + "content": "Xiaokun Feng\\(^{1,2}\\) \nXiangxiang Chu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.217, + 0.321, + 0.235 + ], + "angle": 0, + "content": "1 AMAP, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.217, + 0.873, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\) CRISE, Institute of Automation, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.235, + 0.858, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Computer Science and Technology, University of Chinese Academy of Sciences" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.321, + 0.485, + 0.743 + ], + "angle": 0, + "content": "Video generation has advanced rapidly, improving evaluation methods, yet assessing video's motion remains a major challenge. Specifically, there are two key issues: 1) current motion metrics do not fully align with human perceptions; 2) the existing motion prompts are limited. Based on these findings, we introduce VMBench—a comprehensive Video Motion Benchmark that has perception-aligned motion metrics and features the most diverse types of motion. VMBench has several appealing properties: (1) Perception-Driven Motion Evaluation Metrics, we identify five dimensions based on human perception in motion video assessment and develop fine-grained evaluation metrics, providing deeper insights into models' strengths and weaknesses in motion quality. (2) Meta-Guided Motion Prompt Generation, a structured method that extracts meta-information, generates diverse motion prompts with LLMs, and refines them through human-AI validation, resulting in a multilevel prompt library covering six key dynamic scene dimensions. (3) Human-Aligned Validation Mechanism, we provide human preference annotations to validate our benchmarks, with our metrics achieving an average \\(35.3\\%\\) improvement in Spearman's correlation over baseline methods. This is the first time that the quality of motion in videos has been evaluated from the perspective of human perception alignment. Additionally, we release VMBench at https://github.com/AMAP-ML/VMBench, setting a new standard for evaluating and advancing motion generation models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.775, + 0.222, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Video generation models [2, 3, 37, 43, 55, 78, 90] have advanced significantly, particularly text-to-video (T2V) models [12, 45, 70, 96], which perform well in generating static content but struggle to capture motion in alignment with hu" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.905, + 0.318 + ], + "angle": 0, + "content": "man perception. To enhance dynamic motion generation, it is crucial to establish perception-aligned motion evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.905, + 0.441 + ], + "angle": 0, + "content": "However, previous methods [29, 49, 59, 64, 75] primarily assess the static content of videos, with dynamic evaluation largely limited to motion smoothness [33, 44, 50]. This makes it challenging to capture motion deficiencies [69, 94], including but not limited to spatiotemporal inconsistencies and violations of physical laws. Therefore, developing a more comprehensive evaluation metric that better aligns with human perception is crucial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.578 + ], + "angle": 0, + "content": "To address this, we introduce Video Motion Benchmark (VMBench) (Fig. 1), the first benchmark focused on comprehensive video motion assessment. It features the most diverse range of motion types, covering 969 categories—more than any other benchmark [27, 33, 44, 49, 50, 68] (see Fig. 12 in the Appendix). Additionally, it integrates perception-aligned motion metrics to systematically evaluate the motion generation quality of video generation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.763 + ], + "angle": 0, + "content": "First, we introduce Perception-Driven Motion Evaluation Metrics (PMM), the first evaluation framework explicitly designed to align with human perception of motion quality. PMM comprises five key components: Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), Temporal Coherence Score (TCS), Motion Smoothness Score (MSS), and Commonsense Adherence Score (CAS). Unlike previous methods that primarily focus on motion smoothness [29, 59, 64, 75], PMM provides a more comprehensive assessment by evaluating motion quality in terms of spatiotemporal inconsistencies and violations of physical laws." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Next, we propose Meta-Guided Motion Prompt Generation (MMPG), an expandable framework that encompasses the most comprehensive range of motion types, grounded in physics [56, 60, 65, 67] and cognitive science [25, 57]. The evaluation system covers six movement modes: fluid dynamics, biological motion, mechanical motion, weather phenomena, collective behavior, and energy transfer (Fig. 1). Specifically, MMPG first extracts subjects, places, and actions from VidProm [79], Place365 [97], and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.876, + 0.241, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.239, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13087" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.526, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.09, + 0.858, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 1. Overview of VMBench. Our benchmark encompasses six principal categories of motion patterns, with each prompt constructed as a comprehensive motion structured around three core components: subject, place, and acion. We propose a novel multi-dimensional video motion evaluation comprising five human-centric quality metrics derived from perceptual preferences. Utilizing videos generated by popular T2V models, we conduct systematic human evaluations to validate the effectiveness of our metrics in capturing human perceptual preferences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.56 + ], + "angle": 0, + "content": "other datasets [8, 10, 14, 89], then leverages LLMs to enrich metadata and generate detailed motion descriptions. We validate the prompts through human-AI collaboration with DeepSeek-R1 [21] to ensure coherence and rationality. To reduce costs, we curated 1,050 user prompts and use T2V models to generate motion videos for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Finally, we provide human preference annotations to validate our benchmarks, showing that our metrics achieve an average \\(35.3\\%\\) improvement in Spearman's correlation [93] over current methods [9, 13, 33, 50, 59, 73, 81, 84, 95]. This result highlights the enhanced alignment of our evaluation metrics with human perception. We will open-source VMBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VMBench to drive forward the field of video motion generation. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.809 + ], + "angle": 0, + "content": "- First-Ever Human Perception-Aligned Motion Evaluation: We pioneer the evaluation of motion quality in videos from the perspective of human perception alignment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Meta-Guided Motion Prompt Generation: A curated prompt set designed to assess diverse motion aspects in video models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- Human-Aligned Validation Mechanism: Our metrics achieve an average \\(35.3\\%\\) improvement in Spearman's correlation over current methods." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.468, + 0.652, + 0.483 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.495, + 0.745, + 0.51 + ], + "angle": 0, + "content": "2.1. Video Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.642 + ], + "angle": 0, + "content": "With the rapid advancement of AI generative techniques—GANs [22, 26, 36, 41, 52], VAEs [41, 83], autoregressive models [17, 63, 92], and diffusion models [7, 31, 32, 34, 76, 88]—the field of T2V generation has seen remarkable progress. However, while current T2V models [16, 70, 90, 96], excel in generating static visual content, their dynamic performance still deviates significantly from human perception." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Early T2V models [15, 30, 31, 66] have built upon Text-to-Image(T2I) techniques, enhancing their temporal capabilities to generate videos. Then, some works [28, 80, 98] take this further by conducting the diffusion process within the latent space and incorporate the spatiotemporal module [16, 76, 86] within the denoising network, which greatly improves the quality of static content in videos. CogVideo [32, 90] and other closed source works [1, 5, 12, 58] represented by Sora [2] focus on high-quality data filtering on a large-scale dataset, significantly boosting model performance. OpenSora [96] and OpenSora-Plan [45] aim to showcase and open source the remarkable performance of closed-source models. Mochi 1 [70] and HunyuanVideo [37] have improved the dynamic quality of generated videos through new technologies such as asymmetric architecture and full attention mechanism. Beyond optimization of models and datasets, recent stud" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.093, + 0.904, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.419, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Figure 2. Framework of our Perception-Driven Motion Metrics (PMM). PMM comprises multiple evaluation metrics: Commonsense Adherence Score (CAS), Motion Smoothness Score (MSS), Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), and Temporal Coherence Score (TCS). (a-e): Computational flowcharts for each metric. The scores produced by PMM show variation trends consistent with human assessments, indicating strong alignment with human perception." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.502, + 0.482, + 0.623 + ], + "angle": 0, + "content": "ies [3, 5, 40, 46, 51] are beginning to explore the influence of motion on video generation. Specifically, Ma et al. [51] focuses on the motion quality of videos during data selection, in order to enable models to learn motion dynamics even at low resolution training stages. Meanwhile, some works [40, 46] improve the generative capacity of models by incorporating motion information as guidance during the sampling process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.715 + ], + "angle": 0, + "content": "Although these models demonstrate commendable performance across various metrics, the generated videos still exhibit several qualitative motion deficiencies, such as the abrupt disappearance of moving subjects, body distortions, and physically implausible actions, which are difficult to assess using existing metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.449, + 0.741 + ], + "angle": 0, + "content": "2.2. Evaluations Metrics for Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Existing approaches for video motion evaluation can be broadly categorized into three paradigms: 1) feature-based metrics leveraging pre-trained video representations, 2) rule-based metrics employing manually designed scoring mechanisms, and 3) MLLM-based assessments that fine-tune large multimodal language models on human-annotated video quality datasets for perceptual scoring." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Previous feature-based metrics demonstrate limited effectiveness in assessing motion quality. Image-centric metrics like IS [64] and FID [29] inherently ignore tempo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.905, + 0.578 + ], + "angle": 0, + "content": "ral coherence, while FVD [75] extends image-level metrics to videos by modeling spatiotemporal features, it reduces motion dynamics to simplistic distributions. These methods conflate feature-space distances with perceptual motion quality, overlooking artifacts like implausible movement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Recent benchmarks [24, 33, 50] attempt to conduct systematic evaluations using manually curated metrics. For instance, VBench [33] introduces motion-related criteria such as motion smoothness, quantified via a video frame interpolation model. However, such rule-driven metrics face two inherent limitations in motion assessment: 1) Subjectivity in metric design—predefined rules poorly generalize to diverse motion patterns beyond designer priors; 2) Fragmentary analysis—isolated measurements of individual attributes fail to capture holistic motion quality. Consequently, these heuristics overlook complex phenomena like inertia preservation or multi-object interaction coherence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Emerging methods [11, 27] employ multimodal large language models (MLLMs) fine-tuned on human preference data to predict perceptual video quality scores. VideoScore [27] aligns scoring with human judgments through preference learning, while VideoPhy [11] specifically targets physical plausibility via physics-focused annotations. Despite improved alignment with subjective evaluations, these approaches exhibit critical limitations for motion assessment: 1) Oversimplified scoring granularity – scalar outputs" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13089" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "obscure specific motion defects; 2) Annotation bias amplification – training data biases toward conspicuous artifacts neglect subtle kinematic violations; 3) Contextual rigidity – domain-specific tuning limits generalization to diverse motion types. While advancing automated evaluation, they struggle to disentangle motion quality from global preference impressions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.274 + ], + "angle": 0, + "content": "Current evaluation paradigms remain fundamentally misaligned with human perception of motion quality due to the above limitations. To bridge this gap, we introduce 5 distinct human-aligned evaluation metrics that systematically quantify motion quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.286, + 0.201, + 0.303 + ], + "angle": 0, + "content": "3. VMBench" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.482, + 0.358 + ], + "angle": 0, + "content": "We introduce VMBench, a comprehensive Video Motion Benchmark that includes perception-aligned motion metrics and features the most diverse range of motion types." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.366, + 0.48, + 0.382 + ], + "angle": 0, + "content": "3.1. Perception-Driven Motion Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.569 + ], + "angle": 0, + "content": "When observing videos, the human brain first constructs a holistic understanding of the scene based on prior experiences and physical laws [23], before selectively attending to moving objects to assess motion smoothness and temporal consistency, particularly when objects are momentarily occluded. While rapid movement can effectively capture attention, excessive deviations from typical motion patterns may induce visual disorientation. As illustrated in Fig. 5 in the Appendix, our motion evaluation metrics are designed to emulate this hierarchical perceptual process, transitioning from global scene comprehension to localized motion analysis to ensure coherence and logical consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Commonsense Adherence Score (CAS) is used to evaluate whether generated videos align with human commonsense. As shown in Fig. 2 (a), we train a specialized model capable of assessing the commonsense quality of video content, categorizing it into five levels: Bad, Poor, Fair, Good, and Perfect. First, we collect a dataset comprising \\(10\\mathrm{k}\\) generated videos, covering both legacy approaches and popular models [2-4, 15, 16, 37, 45, 90, 96]. Second, we establish perceptual ground truth through systematic pairwise comparisons using VideoReward [47], a reward model trained on \\(182\\mathrm{k}\\) human-annotated video preference pairs. This process generates normalized preference scores reflecting human perceptual consensus. Third, we discretize the human preference into five cognitive labels. Using these preference labels, we develop a video classification model to serve as the final CAS predictor, utilizing VideoMAEv2 [77], a spatiotemporal vision transformer. Compared to MLLMs, this architecture demonstrates temporal modeling capabilities through dense frame sampling and 3D convolution operations. We compute the CAS using a Mean Opinion Score (MOS) [74, 85], where predicted probabilities for each class are weighted by their corresponding quality co" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.905, + 0.168 + ], + "angle": 0, + "content": "efficients: \\(\\mathrm{CAS} = \\sum_{i=1}^{5} p_i G(i)\\), where \\(p_i\\) denotes the predicted probability for the \\(i\\)-th class, and \\(G_i\\) represents the mapping function that converts category to quality weights: \\(G : i \\to w\\) (e.g., mapping class 3 to 0.5). For more details, please refer to our Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.469 + ], + "angle": 0, + "content": "Motion Smoothness Score (MSS) aims to detect unsmoothness, like low-level temporal artifacts and high-level motion blur. It addresses the limitations of prior metrics that suffered from either low-level optical flow bias [38, 39] or oversimplified motion modeling [42], both leading to misalignment with human perception. Considering that natural motion patterns inherently maintain consistent visual quality across temporal sequences, we leverage Q-Align's [85] aesthetic score to detect artifacts, as illustrated in Fig. 2 (b). An artifact frame is confirmed when a significant drop in Q-Align scores between consecutive frames exceeds a predefined threshold, mimicking human perception of sustained visual anomalies. Further, our approach employs a scene-aware adaptive thresholding mechanism derived from the statistical modeling of real video segments [14, 54] across diverse motion patterns. The final MSS is computed as: \\(\\mathrm{MSS} = 1 - \\frac{1}{T}\\sum_{t=2}^{T}\\mathbb{I}(\\Delta Q_t > \\tau_s(t))\\) where \\(\\Delta Q_t\\) is the frame-to-frame visual quality degradation magnitude, \\(\\tau_s(t)\\) indicates the adaptive threshold and \\(\\mathbb{I}\\) is an indicator function. More details are provided in Appendix B.2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Object Integrity Score (OIS) detects implausible deformations through spatiotemporal analysis of object integrity. Our approach addresses limitations found in previous methods [13], which primarily focus on object-level semantic consistency. Instead, we focus on detecting perceptual issues (e.g., distorted shapes) that are readily noticeable to the human visual system. As illustrated in Fig. 2 (c), we first utilize the MMPose toolkit [20] to detect key points of the primary subjects in the generated videos, which are then used to estimate the subjects' shapes in each frame. Then, we quantify shape distortions by analyzing whether object shapes violate real-world anatomical constraints. We establish tolerance thresholds for the above constraints through the statistical analysis of natural motion samples [14, 54, 87]. Anatomical violations across consecutive frames will be detected; see more details about anatomical constraints in our appendix B.3. The OIS quantifies structural consistency through: \\(\\mathrm{OIS} = \\frac{1}{F.K}\\sum_{f = 1}^{F}\\sum_{k = 1}^{K}\\mathbb{I}\\left(\\mathcal{D}_f^{(k)}\\leq \\tau^{(k)}\\right)\\), in which \\(F\\) and \\(K\\) denote total frame count and anatomical components, respectively. \\(\\mathcal{D}_f^{(k)}\\) represents the compound anatomical deviation for component \\(k\\) in frame \\(f\\), and \\(\\tau^{(k)}\\) indicates statistical thresholds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Perceptible Amplitude Score (PAS) estimates subject motion by separating it from camera motion, unlike traditional motion magnitude estimation methods [73], which often overestimate overall motion due to the influence of camera movement. As presented in Fig. 2 (d), PAS begins with semantic anchoring via GroundingDINO [48] to lo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13090" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.369 + ], + "angle": 0, + "content": "calize moving subjects and distinguish them from passively moving elements. Next, GroundedSAM [62] ensures temporally stable subject masks, enabling CoTracker [35] to track key points with high precision. The motion magnitude is then computed based on the average displacement of these key points. Furthermore, PAS accounts for the context-dependent nature of human motion perception, adapting thresholds across different scenarios. To quantify this variability, we derive a set of perceptual motion magnitude thresholds for various scenarios through the statistical analysis of existing video datasets [14, 54]. These thresholds serve as the foundation for computing a motion score for each video. The PAS is computed as PAS = \\(\\frac{1}{T}\\sum_{t=1}^{T}\\min\\left(\\frac{\\bar{D}_t}{\\tau_s}, 1\\right)\\), where \\(\\bar{D}_t\\) denotes frame-level motion amplitude, computed as the average displacement of tracked key points for active subjects in frame \\(t\\) and \\(\\tau_s\\) is the perceptual motion threshold for scenario \\(s\\). More Details can be found in appendix B.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Temporal Coherence Score (TCS) detects the temporal consistency of object motion by analyzing frame-to-frame transitions, addressing the limitation of existing metrics [13, 59] that struggle to differentiate between natural movement and abrupt, unrealistic changes. We combine video object tracking with rule-based validation to detect abnormal disappearance/reappearance patterns, thereby distinguishing natural progression from temporal discontinuities. As shown in Fig. 2 (e), first GroundedSAM2 [61] performs pixel-accurate instance segmentation and tracking across frames, maintaining persistent object IDs throughout the whole sequence. For objects exhibiting discontinuous existence, we apply a secondary verification phase using the CoTracker [35]. It tracks dense key points on target objects and constructs their motion trajectories. We then analyze their motion trajectories to determine whether any anomalous phenomena are present. Our approach mitigates false cases caused by legitimate object discontinuity through a rule-based filtering mechanism implemented via trajectory analysis. These rules account for common scenarios, including 1) objects reappearing after occlusion or disappearing behind obstacles, 2) objects entering or exiting frame boundaries, and 3) apparent size changes due to depth perception, such as objects appearing larger when moving closer or smaller when moving farther away. Through this filtering, we systematically exclude normal events. The TCS quantifies anomalous discontinuities via \\(\\mathrm{TCS} = 1 - \\frac{1}{N}\\sum_{i=1}^{N}\\mathbb{I}(\\mathcal{A}_i\\wedge \\neg \\mathcal{R})\\), where \\(N\\) is the total object instances, \\(\\mathcal{A}_i\\) indicates the existence status of the \\(i\\)-th object, and \\(\\mathcal{R}\\) validates legitimate transitions (see Appendix B.5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.445, + 0.865 + ], + "angle": 0, + "content": "3.2. Meta-Guided Motion Prompt Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The existing benchmarks [27, 33, 50] are constrained by limited and simplistic motion types (As shown in Fig. 12" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "of the Appendix), making them inadequate for comprehensively evaluating the motion generation capabilities of video generation models. To address this limitation, we propose an extensible framework—Meta-Guided Motion Prompt Generation (MMPG)—that generates prompts capturing complex motion patterns, as illustrated in Fig. 3. Built upon MMPG, we introduce VMBench, a benchmark specifically designed for rigorous evaluation of complex motion generation. We find that VMBench provides the most comprehensive coverage of motion types and the most detailed prompt descriptions, making it an effective benchmark for evaluating the dynamic motion generation capabilities of video generation models. In this section, we elaborate on the process of prompt generation and refinement. More details are available in Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.905, + 0.532 + ], + "angle": 0, + "content": "Meta-information Extraction. We decompose motion descriptions into three key metadata elements: Subject \\((S)\\), Location \\((P)\\) and Action \\((A)\\). We employ Qwen-2.5 [71] to extract the metadata library \\(\\{S, P, A\\}\\) from existing videotext pairs [18, 58, 79, 89]. To ensure novel prompts and enhance textual diversity, we expand the metadata in three key aspects. For subject \\((S)\\) classification, we categorize subject types into human, animal, and object, curate a list of principal nouns identifiable by [48, 62], and use GPT-4o [6] to generate descriptions with varying entity counts (1, 2, n). For place \\((P)\\) descriptions, we incorporate data from [97], filtering out redundant information. For action \\((A)\\) expansion, we sample human actions from [14] and use LLM to extend possible actions for animals and objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.904, + 0.778 + ], + "angle": 0, + "content": "Self-Refining Prompt Generation. We randomly sample elements from the metadata library \\(\\{S, P, A\\}\\) to form metadata sets \\((S^i, P^j, A^k)\\) and use GPT-4o to assess their semantic coherence and logical consistency. For coherent sets, we generate prompts of 10 to 60 words that describe the corresponding motion scenarios. To ensure accuracy, we introduce a verification mechanism inspired by [50], where the generated prompts are re-evaluated by GPT-4o to verify their consistency with the metadata elements. Through an iterative refinement and filtering process, we generate approximately 50,000 high-quality prompt candidates. However, grammatical correctness alone is insufficient: prompts must also align with physical reality to effectively assess video content. Therefore, we also eliminate descriptions of unrealistic or physically implausible scenarios to ensure the feasibility and authenticity of the prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Human-LLM Joint Validation. To reduce manual labor and enhance evaluation effectiveness, we further examine and filter prompts through a Human-LLM Joint Validation process. This process combines the strengths of both automated reasoning and human judgment to ensure high-quality motion descriptions. We first leverage the powerful reasoning capabilities of Deepseek R1 [21] to evaluate whether each prompt describes a reasonably realistic mo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13091" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.895, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Figure 3. Framework of our Meta-Guided Motion Prompt Generation (MMPG). MMPG consists of three stages: (a) Meta-information Extraction: Extracting Subjects, Places, and Actions from datasets such as VidProm [79], Didemo [8], MSR-VTT [89], WebVid [10], Place365 [97], and Kinect-700 [14]. (b) Self-Refining Prompt Generation: Generating and iteratively refining prompts based on the extracted information. (c) Human-LLM Joint Validation: Validating the prompts through a collaborative process between humans and DeepSeek-R1 to ensure their rationality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.435 + ], + "angle": 0, + "content": "tion. After filtering out implausible descriptions, we curate a diverse set of 1,050 prompts with varied metadata. Examples are available in Appendix C.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.447, + 0.224, + 0.464 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.472, + 0.281, + 0.489 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Implementation Details. Our benchmark evaluates six popular text-to-video models: OpenSora [96], CogVideoX [90], OpenSora-Plan [45], Mochi 1 [70], HunyuanVideo [37], and Wan2.1 [72]. To provide a richer variety of motion types, we develop the MMPG-set, which comprises 1,050 prompts based on MMPG across six movement modes to assess performance in motion generation. Each model generates 1,050 videos corresponding to the MMPG-set, resulting in a total of 6,300 videos. To ensure a fair comparison, we maintain the hyperparameter settings as defined in each model's project demo. For each prompt, we generate one video for evaluation using only the initial seed. The inference process is executed using 8 Nvidia H20 GPUs. Details of the inference process for each model can be found in Appendix D.1. We then randomly sample 200 videos from each model's output, resulting in a total of 1,200 videos for human-aligned validation experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison Metrics. The current comparison approaches fall into two categories: Rule-based metrics and Multimodal Large Language Model (MLLM) prompting. (1) Rule-based metrics assess four dimensions, including Perceptible Amplitude, which evaluates through RAFT optical flow magnitude analysis combined with structural motion consistency via 4 frame SSIM averaging [73, 82], following established protocols [33]. Temporal Coherence is measured using DINO [13] and CLIP [59] feature tracking, which relies on cosine similarity between consecu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.659 + ], + "angle": 0, + "content": "tive frames. Motion Smoothness is assessed through a hybrid method that combines interpolation error [42] with Dover's video quality assessment [84], while Object Integrity is evaluated via dual validation using optical flow warping error [38, 39] and semantic consistency checks. (2) The MLLM evaluation includes five cutting-edge models: LLaVA-NEXT-Video [95], MiniCPM-V-2.6 [91], InternVL2.5 [19], Qwen2.5-VL [9], and InternVideo2.5 [81]. These models are evaluated through a standardized assessment process, which involves processing uniformly sampled video frames at 2 frames per second to maintain essential motion patterns while managing computational limits. The evaluation dimensions encompass five aspects: amplitude, coherence, integrity, smoothness, and commonsense adherence, with scoring for each dimension using a 1–5 scale. In particular, to ensure fair comparison, consistent frame sequences and evaluation criteria are maintained across all models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Metrics. 1) Spearman correlation [93]: Spearman's rank correlation coefficient, often denoted by \\(\\rho\\), is a measure of the strength and direction of the association between two ranked variables. It assesses how well the relationship between two variables can be described by a monotonic function. Spearman correlation is nonparametric and can effectively evaluate associations in datasets where variables may not follow a normal distribution. Unlike Pearson correlation, which captures linear relationships, Spearman correlation focuses on rank-based associations, making it more robust to outliers and suitable for ordinal data or scenarios with nonlinear dependencies. 2) Accuracy: To validate how accurately our metrics align with human preferences, we conduct pairwise comparisons across the 1,200 annotated videos (200 prompts \\(\\times\\) 6 models). For each prompt, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13092" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.37 + ], + "angle": 0, + "content": "
MethodAvg. ρ(↑)CAS ρ(↑)MSS ρ(↑)OIS ρ(↑)PAS ρ(↑)TCS ρ(↑)
Rule-based
SSIM [82]1.6-0.9-12.18.317.8-4.8
RAFT* [73]-1.7-0.7-17.0-16.647.7-21.9
CLIP* [59]15.021.536.531.7-42.728.0
DINO* [13]9.411.027.227.4-45.827.4
Dover Technical† [84]20.640.232.634.5-6.22.2
AMT* [33]3.62.818.119.6-40.417.8
Warping Error† [50]-6.0-11.4-19.1-9.033.0-23.6
MLLM Prompt
LLaVa-NEXT-Video [95]17.115.211.016.130.812.4
MiniCPM-V-2.6 [91]16.26.817.5-2.227.331.6
IntenVL2.5 [19]19.318.720.514.142.51.1
Qwen2.5-VL [9]15.324.719.83.714.214.0
InternVideo2.5 [81]26.922.721.929.644.315.8
Model-Rule Synergy
PPM (Ours)62.269.977.165.865.254.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.452 + ], + "angle": 0, + "content": "Table 1. Correlation Analysis Between Evaluation Metrics and Human Scores via Spearman Correlation Coefficient (\\(\\rho \\times 100\\)). Superscript * and † denote implementations following VBenchmark [33] and EvalCrafter [50] respectively. Yellow backgrounds in Rule-based indicate specific dimension baseline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.645 + ], + "angle": 0, + "content": "evaluate all 15 possible video pairs generated by different models (total 3,000 pairs). Human preference labels are determined by comparing the average expert ratings across all five dimensions (OIS, MSS, CAS, TCS, PAS) – the video with higher aggregated human scores is designated as the “ground truth” preferred sample. Concurrently, we compute metric preferences by comparing videos' aggregated PMM scores under identical criteria. Alignment accuracy is quantified as the percentage of pairs where PMM preferences match human judgments, with ties excluded to prioritize unambiguous decisions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.653, + 0.388, + 0.669 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Human-aligned Validation Mechanism. We invite three domain experts to independently annotate each sample based on PMM (Perceptible Amplitude, Temporal Coherence, Object Integrity, Motion Smoothness, and Common-sense Adherence), resulting in 6,000 detailed ratings with high interannotator agreement. Finally, we obtain Likert scores [53] from humans for each video across the five dimensions included in PMM. More details can be found in Appendix D.2. We assess the alignment between metrics and human perception by calculating the Spearman correlation [93] between metric scores and expert ratings. A higher Spearman correlation signifies a stronger alignment with human perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison with Alternative Metrics. As shown in Table 1, in the MSS evaluation, even advanced metrics" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.089, + 0.885, + 0.299 + ], + "angle": 0, + "content": "
CASMSSOISPASTCSAccuracy (%) ↑
Removal-based ablation
✓✓✓✓✓70.6
✓✓✓✓✗66.9
✓✓✓✗✓68.7
✓✓✗✓✓64.6
✓✗✓✓✓65.2
✗✓✓✓✓64.1
Addition-based ablation
✗✗✗✗✗-
✓✗✗✗✗58.9
✓✓✗✗✗66.1
✓✓✓✗✗67.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Table 2. Ablation Study on our Motion Metrics. Prediction accuracy (\\%) of different metric combinations against human preferences is calculated. Removal-based ablation show the effect of individually ablating each metric, while in the addition-based ablation, we progressively add each metric to observe its impact." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.909, + 0.834 + ], + "angle": 0, + "content": "such as AMT [42] (18.1%) and Warping Error [39] (- 19.1%) demonstrate limited discriminative power and produce counterintuitive results, particularly in the context of complex deformations. The OIS evaluation also highlights similar shortcomings, with DINO [13] at 27.4% and Dover [84] at 34.5%, both of which fail to capture human sensitivity to structural preservation during motion. In relation to the evaluation of PAS, the SSIM [82] and RAFT [73] metric based on rules show an alignment efficacy of only 17.8% and 47.7%, respectively. In contrast, our approach achieves a remarkable 65.2% alignment efficacy. For the TCS evaluation, the rule-based metrics CLIP [59] and DINO [13] achieve only 28.0% and 27.4% alignment efficacy, respectively, and fail to capture human tolerance for minor inconsistencies while maintaining physical plausibility. In comparison, our method achieves an impressive 54.5% alignment efficacy. VBench [33] includes RAFT, CLIP, DINO, and AMT, while EvalCrafter [50] incorporates Dover Technical and Warping Error. According to the data presented in the table, overall, compared to ours, the motion evaluations of VBench [33] and EvalCrafter [50] reveal significantly lower correlations with human perception. MLLMs exhibit partial competence in evaluations of the Physical Adequacy Score (PAS) (e.g., InternVideo2.5: 44.3%) Despite their general capabilities, MLLMs average only a correlation of 10.0%-30.0% in all dimensions, highlighting a fundamental misalignment with human perception in evaluating motion quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.667, + 0.864 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Ablation Study of Motion Metrics. According to Table 2, removing any single metric from the evaluation of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13093" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.351 + ], + "angle": 0, + "content": "basic variants results in a significant decline in accuracy, which highlights the importance of each dimension's evaluation metric within the overall framework. Notably, the removal of the CAS metric leads to the most pronounced performance drop, with accuracy falling to \\(64.1\\%\\), exceeding the impact of other single-dimension ablations. This demonstrates the critical role of the CAS metric in assessing video quality, aligning closely with the key factors prioritized by humans when perceiving video quality. For the performance-oriented variants, we emulate the human perceptual information processing flow by gradually adding evaluation metrics (refer to Fig. 5 in the Appendix), with each additional metric significantly enhancing overall accuracy. This not only demonstrates the effectiveness of incrementally increasing evaluation dimensions but also further verifies the consistency of our evaluation approach with human perceptual mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.372, + 0.285, + 0.387 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.485, + 0.746 + ], + "angle": 0, + "content": "Alignment of PMM with Human Perception. As shown in Fig. 4, the correlation between human scores across these five evaluation dimensions and our evaluation metrics in the same five dimensions remains consistent. For example, both demonstrate a strong correlation between OIS, CAS, and MSS, as well as a weak correlation between PAS and the other metrics. Specifically, as shown in Fig. 4 (a), PAS has a negative correlation with other dimensions, such as the Object Integrity Score (OIS), where \\(\\rho = -0.18\\). This might be because high dynamic amplitudes in videos cause distortions and artifacts, which lower structural and temporal coherence scores. Conversely, OIS shows strong positive correlations with MSS (\\(\\rho = 0.59\\)) and CAS (\\(\\rho = 0.50\\)), suggesting that it reflects physical plausibility and motion rationality well. TCS shows low correlation with other dimensions, which can provide a more comprehensive assessment perspective. PAS and structural/temporal metrics exhibit a negative correlation, which challenges conventional optical-flow-based evaluation frameworks. Additionally, the isolation of PAS highlights the importance of separately disentangling motion magnitude in motion video evaluation. Fig. 4 (b) shows the correlations among the proposed evaluation metrics, which align with human perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Assessing Video Generation Models with PMM. As shown in Table 3, we evaluate several leading video generation models using our PMM metrics, including Mochi 1[70], OpenSora [96], CogVideoX [90], OpenSora-Plan [45], HunyuanVideo [37], and Wan2.1 [72]. Our findings indicate that Wan2.1 delivers the best performance in generating motion videos, producing results that appear more realistic. For a clearer visual comparison of the dynamic videos generated by each model, please refer to Fig. 16 in the Appendix." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.088, + 0.714, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.089, + 0.892, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.23, + 0.908, + 0.286 + ], + "angle": 0, + "content": "Figure 4. Correlation Matrix Analysis of Metrics Within Different Evaluation Mechanisms. (a): Spearman Correlation Matrices for human annotations; (b): Spearman Correlation Matrices for our PMM metrics." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.304, + 0.907, + 0.415 + ], + "angle": 0, + "content": "
ModelsAvg.(%) ↑CASMSSOISPASTCS
Mochi 1[70]53.237.762.068.614.483.6
OpenSora [96]51.631.261.973.03.488.5
CogVideoX [90]60.650.661.675.424.691.0
OpenSora-Plan [45]58.939.376.078.66.094.7
HunyuanVideo [37]63.451.981.665.826.196.3
Wan2.1 [72]78.462.884.266.017.997.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Table 3. Performance of Video Generation Models on VMBench. We evaluated six open-source video generate models [37, 45, 70, 72, 90, 96] using VMBench. A higher score indicates better performance for a category." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.511, + 0.634, + 0.527 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Video motion authenticity remains a critical challenge in the creation of generate content. In response, we introduce VMBench, the first open-source benchmark for motion quality evaluation, integrating motion metrics with human-aligned assessment to reveal deficiencies in existing models' ability to generate physically plausible movements. To support the research community, we offer three essential resources: 1) a standardized framework for detecting motion artifacts overlooked by traditional metrics; 2) actionable diagnostic tools aimed at guiding model optimization; and 3) quantitative standards that align technical advancements with human perceptual expectations. By establishing a common framework for development in this field, VMBench enables systematic tracking of motion generation, encourages targeted model improvements, and ultimately supports the creation of videos that achieve both visual fidelity and dynamic realism. However, while the benchmark's evaluation metrics are aligned with general human perception, they may not fully capture subtle differences in perception that stem from individual viewer experiences and preferences. Addressing these limitations presents opportunities for further refinement, aiming to ensure a more comprehensive and adaptable approach to video quality evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13094" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Runway gen3. Accessed February 25, 2025 [Online] https://app.runwayml.com/, 2025.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.145, + 0.483, + 0.171 + ], + "angle": 0, + "content": "[2] Sora. Accessed February 25, 2025 [Online] https://openai.com/index/sora/, 2025.1,2,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.173, + 0.482, + 0.199 + ], + "angle": 0, + "content": "[3] kling. Accessed February 25, 2025 [Online] https://klingai.com/, 2025.1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.201, + 0.481, + 0.228 + ], + "angle": 0, + "content": "[4] Pika labs. Accessed February 25, 2025 [Online] https://www.pika.art/, 2025.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.482, + 0.27 + ], + "angle": 0, + "content": "[5] Veo 2. Accessed February 25, 2025 [Online] https://deepmind.google/technologies/veo/veo-2/, 2025.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.272, + 0.482, + 0.339 + ], + "angle": 0, + "content": "[6] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[7] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[8] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.524, + 0.482, + 0.591 + ], + "angle": 0, + "content": "[10] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.594, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[11] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[12] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Videocrafter1: Open diffusion models for high-quality video generation, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[16] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[17] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[18] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[19] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[20] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.904, + 0.474 + ], + "angle": 0, + "content": "[21] DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[22] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[23] Daniel J. Felleman and David C. van Essen. Distributed hierarchical processing in the primate cerebral cortex. *Cerebral Cortex*, 1 1:1-47, 1991. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[24] Xiaokun Feng, Haiming Yu, Meiqi Wu, Shiyu Hu, Jintao Chen, Chen Zhu, Jiahong Wu, Xiangxiang Chu, and Kaiqi Huang. Narrlv: Towards a comprehensive narrative-centric evaluation for long video generation models. arXiv preprint arXiv:2507.11245, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.904, + 0.66 + ], + "angle": 0, + "content": "[25] Deborah M Gordon. The ecology of collective behavior. PLoS biology, 12(3):e1001805, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[26] Shir Gur, Sagie Benaim, and Lior Wolf. Hierarchical patch vae-gan: Generating diverse videos from a single sample. Advances in Neural Information Processing Systems, 33: 16761-16772, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[27] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv preprint arXiv:2406.15252, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13095" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[30] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[31] Jonathan Ho, Tim Salimans, Alexey Gritsanko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[32] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[33] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[34] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22747-22757, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[35] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[36] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 8:153113-153122, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[37] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[38] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[39] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. Advances in Neural Information Processing Systems, 33:1083-1093, 2020. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[40] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[41] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[44] Mingxiang Liao, Qixiang Ye, Wangmeng Zuo, Fang Wan, Tianyu Wang, Yuzhong Zhao, Jingdong Wang, Xinyu Zhang, et al. Evaluation of text-to-video generation models: A dynamics perspective. Advances in Neural Information Processing Systems, 37:109790-109816, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[45] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[46] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[47] Jie Liu, Gongye Liu, Jiajun Liang, Ziyang Yuan, Xiaokun Liu, Mingwu Zheng, Xiele Wu, Qiulin Wang, Wenyu Qin, Menghan Xia, et al. Improving video generation with human feedback. arXiv preprint arXiv:2501.13918, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[48] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[49] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. Advances in Neural Information Processing Systems, 36:62352-62387, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[50] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[51] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[52] Rayeesa Mehmood, Rumaan Bashir, and Kaiser J Giri. Vtmgan: video-text matcher based generative adversarial network for generating videos from textual description. International Journal of Information Technology, 16(1):221-236, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[53] Tomoko Nemoto and David Beglar. Likert-scale questionnaires. In JALT 2013 conference proceedings, pages 1-6, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13096" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[55] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.236 + ], + "angle": 0, + "content": "[56] TA Palomaki, JD Teufel, RW Simmonds, and Konrad W Lehnert. Entangling mechanical motion with microwave fields. Science, 342(6159):710-713, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.279 + ], + "angle": 0, + "content": "[57] Marina A Pavlova. Biological motion processing as a hallmark of social cognition. Cerebral cortex, 22(5):981-995, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.283, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[58] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, et al. Movie gen: A cast of media foundation models, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.44, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[60] Ralph H. Rasshofer, Martin Spies, and Hans Spies. Influences of weather phenomena on automotive laser radar systems. Advances in radio science, 9:49-60, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.556, + 0.482, + 0.624 + ], + "angle": 0, + "content": "[62] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.628, + 0.482, + 0.682 + ], + "angle": 0, + "content": "[63] Zhongwei Ren, Yunchao Wei, Xun Guo, Yao Zhao, Bingyi Kang, Jiashi Feng, and Xiaojie Jin. Videoworld: Exploring knowledge learning from unlabeled videos. arXiv preprint arXiv:2501.09781, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[64] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[65] Benjamin Scellier and Yoshua Bengio. Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in computational neuroscience, 11:24, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[66] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[67] Jos Stam. Stable fluids. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 779-786. 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[68] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[69] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.904, + 0.233 + ], + "angle": 0, + "content": "[70] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.904, + 0.261 + ], + "angle": 0, + "content": "[71] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.904, + 0.29 + ], + "angle": 0, + "content": "[72] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[73] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.904, + 0.404 + ], + "angle": 0, + "content": "[74] I Telecom. Recommendation 500-10: Methodology for the subjective assessment of the quality of television pictures. ITU-R Rec. BT, 500, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[75] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.904, + 0.504 + ], + "angle": 0, + "content": "[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[77] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14549-14560, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[78] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. arXiv preprint arXiv:2307.00040, 2(3):4, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[79] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[80] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. Magicvideo-v2: Multi-stage high-aesthetic video generation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[81] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[82] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[83] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Gen" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13097" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "erating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[84] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[85] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[86] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[87] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[88] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[90] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[91] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[92] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.748 + ], + "angle": 0, + "content": "[93] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.79 + ], + "angle": 0, + "content": "[94] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[95] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[96] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 1, 2, 4, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[97] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[98] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13098" + } + ] +] \ No newline at end of file diff --git a/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_origin.pdf b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30241ab2f87d7cbdaf6ad5175c1345afced48d4b --- /dev/null +++ b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/1a1d91b4-eb97-44d0-a07d-41746c53ad96_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f84e4bf9af7c73b16d18b43f59320b570e6942e044cc4d88044858f465868e30 +size 9145113 diff --git a/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/full.md b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a03452053c1a3193e4034612650d39cc48eb8325 --- /dev/null +++ b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/full.md @@ -0,0 +1,290 @@ +# VMBench: A Benchmark for Perception-Aligned Video Motion Generation + +Xinran Ling $^{1*}$ +Cundian Yang $^{1}$ + +Chen Zhu1* +Aiming Hao1 + +Meiqi Wu $^{1,3*}$ +Jiashu Zhu $^{1}$ + +Hangyu Li1 +Jiahong Wu1† + +Xiaokun Feng $^{1,2}$ +Xiangxiang Chu $^{1}$ + +1 AMAP, Alibaba Group + +$^{2}$ CRISE, Institute of Automation, Chinese Academy of Sciences + +$^{3}$ School of Computer Science and Technology, University of Chinese Academy of Sciences + +# Abstract + +Video generation has advanced rapidly, improving evaluation methods, yet assessing video's motion remains a major challenge. Specifically, there are two key issues: 1) current motion metrics do not fully align with human perceptions; 2) the existing motion prompts are limited. Based on these findings, we introduce VMBench—a comprehensive Video Motion Benchmark that has perception-aligned motion metrics and features the most diverse types of motion. VMBench has several appealing properties: (1) Perception-Driven Motion Evaluation Metrics, we identify five dimensions based on human perception in motion video assessment and develop fine-grained evaluation metrics, providing deeper insights into models' strengths and weaknesses in motion quality. (2) Meta-Guided Motion Prompt Generation, a structured method that extracts meta-information, generates diverse motion prompts with LLMs, and refines them through human-AI validation, resulting in a multilevel prompt library covering six key dynamic scene dimensions. (3) Human-Aligned Validation Mechanism, we provide human preference annotations to validate our benchmarks, with our metrics achieving an average $35.3\%$ improvement in Spearman's correlation over baseline methods. This is the first time that the quality of motion in videos has been evaluated from the perspective of human perception alignment. Additionally, we release VMBench at https://github.com/AMAP-ML/VMBench, setting a new standard for evaluating and advancing motion generation models. + +# 1. Introduction + +Video generation models [2, 3, 37, 43, 55, 78, 90] have advanced significantly, particularly text-to-video (T2V) models [12, 45, 70, 96], which perform well in generating static content but struggle to capture motion in alignment with hu + +man perception. To enhance dynamic motion generation, it is crucial to establish perception-aligned motion evaluation. + +However, previous methods [29, 49, 59, 64, 75] primarily assess the static content of videos, with dynamic evaluation largely limited to motion smoothness [33, 44, 50]. This makes it challenging to capture motion deficiencies [69, 94], including but not limited to spatiotemporal inconsistencies and violations of physical laws. Therefore, developing a more comprehensive evaluation metric that better aligns with human perception is crucial. + +To address this, we introduce Video Motion Benchmark (VMBench) (Fig. 1), the first benchmark focused on comprehensive video motion assessment. It features the most diverse range of motion types, covering 969 categories—more than any other benchmark [27, 33, 44, 49, 50, 68] (see Fig. 12 in the Appendix). Additionally, it integrates perception-aligned motion metrics to systematically evaluate the motion generation quality of video generation models. + +First, we introduce Perception-Driven Motion Evaluation Metrics (PMM), the first evaluation framework explicitly designed to align with human perception of motion quality. PMM comprises five key components: Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), Temporal Coherence Score (TCS), Motion Smoothness Score (MSS), and Commonsense Adherence Score (CAS). Unlike previous methods that primarily focus on motion smoothness [29, 59, 64, 75], PMM provides a more comprehensive assessment by evaluating motion quality in terms of spatiotemporal inconsistencies and violations of physical laws. + +Next, we propose Meta-Guided Motion Prompt Generation (MMPG), an expandable framework that encompasses the most comprehensive range of motion types, grounded in physics [56, 60, 65, 67] and cognitive science [25, 57]. The evaluation system covers six movement modes: fluid dynamics, biological motion, mechanical motion, weather phenomena, collective behavior, and energy transfer (Fig. 1). Specifically, MMPG first extracts subjects, places, and actions from VidProm [79], Place365 [97], and + +![](images/882ee89ebe4bd8400faee5b6eaf319fa9c619c99fd8df1847b437cbbcc62113c.jpg) +Figure 1. Overview of VMBench. Our benchmark encompasses six principal categories of motion patterns, with each prompt constructed as a comprehensive motion structured around three core components: subject, place, and acion. We propose a novel multi-dimensional video motion evaluation comprising five human-centric quality metrics derived from perceptual preferences. Utilizing videos generated by popular T2V models, we conduct systematic human evaluations to validate the effectiveness of our metrics in capturing human perceptual preferences. + +![](images/a01b5b425fa47b0421a5d841e87209c627d28f0e20880ff739e94e7f5fd188a0.jpg) + +other datasets [8, 10, 14, 89], then leverages LLMs to enrich metadata and generate detailed motion descriptions. We validate the prompts through human-AI collaboration with DeepSeek-R1 [21] to ensure coherence and rationality. To reduce costs, we curated 1,050 user prompts and use T2V models to generate motion videos for evaluation. + +Finally, we provide human preference annotations to validate our benchmarks, showing that our metrics achieve an average $35.3\%$ improvement in Spearman's correlation [93] over current methods [9, 13, 33, 50, 59, 73, 81, 84, 95]. This result highlights the enhanced alignment of our evaluation metrics with human perception. We will open-source VMBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VMBench to drive forward the field of video motion generation. Our contributions can be summarized as follows: + +- First-Ever Human Perception-Aligned Motion Evaluation: We pioneer the evaluation of motion quality in videos from the perspective of human perception alignment. +Meta-Guided Motion Prompt Generation: A curated prompt set designed to assess diverse motion aspects in video models. +- Human-Aligned Validation Mechanism: Our metrics achieve an average $35.3\%$ improvement in Spearman's correlation over current methods. + +# 2. Related work + +# 2.1. Video Generative Models + +With the rapid advancement of AI generative techniques—GANs [22, 26, 36, 41, 52], VAEs [41, 83], autoregressive models [17, 63, 92], and diffusion models [7, 31, 32, 34, 76, 88]—the field of T2V generation has seen remarkable progress. However, while current T2V models [16, 70, 90, 96], excel in generating static visual content, their dynamic performance still deviates significantly from human perception. + +Early T2V models [15, 30, 31, 66] have built upon Text-to-Image(T2I) techniques, enhancing their temporal capabilities to generate videos. Then, some works [28, 80, 98] take this further by conducting the diffusion process within the latent space and incorporate the spatiotemporal module [16, 76, 86] within the denoising network, which greatly improves the quality of static content in videos. CogVideo [32, 90] and other closed source works [1, 5, 12, 58] represented by Sora [2] focus on high-quality data filtering on a large-scale dataset, significantly boosting model performance. OpenSora [96] and OpenSora-Plan [45] aim to showcase and open source the remarkable performance of closed-source models. Mochi 1 [70] and HunyuanVideo [37] have improved the dynamic quality of generated videos through new technologies such as asymmetric architecture and full attention mechanism. Beyond optimization of models and datasets, recent stud + +![](images/c1d85feb677bc7685884399a155860b34dd19e1eafed2b4e280979d84d7fe012.jpg) +Figure 2. Framework of our Perception-Driven Motion Metrics (PMM). PMM comprises multiple evaluation metrics: Commonsense Adherence Score (CAS), Motion Smoothness Score (MSS), Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), and Temporal Coherence Score (TCS). (a-e): Computational flowcharts for each metric. The scores produced by PMM show variation trends consistent with human assessments, indicating strong alignment with human perception. + +ies [3, 5, 40, 46, 51] are beginning to explore the influence of motion on video generation. Specifically, Ma et al. [51] focuses on the motion quality of videos during data selection, in order to enable models to learn motion dynamics even at low resolution training stages. Meanwhile, some works [40, 46] improve the generative capacity of models by incorporating motion information as guidance during the sampling process. + +Although these models demonstrate commendable performance across various metrics, the generated videos still exhibit several qualitative motion deficiencies, such as the abrupt disappearance of moving subjects, body distortions, and physically implausible actions, which are difficult to assess using existing metrics. + +# 2.2. Evaluations Metrics for Video Generation + +Existing approaches for video motion evaluation can be broadly categorized into three paradigms: 1) feature-based metrics leveraging pre-trained video representations, 2) rule-based metrics employing manually designed scoring mechanisms, and 3) MLLM-based assessments that fine-tune large multimodal language models on human-annotated video quality datasets for perceptual scoring. + +Previous feature-based metrics demonstrate limited effectiveness in assessing motion quality. Image-centric metrics like IS [64] and FID [29] inherently ignore tempo + +ral coherence, while FVD [75] extends image-level metrics to videos by modeling spatiotemporal features, it reduces motion dynamics to simplistic distributions. These methods conflate feature-space distances with perceptual motion quality, overlooking artifacts like implausible movement. + +Recent benchmarks [24, 33, 50] attempt to conduct systematic evaluations using manually curated metrics. For instance, VBench [33] introduces motion-related criteria such as motion smoothness, quantified via a video frame interpolation model. However, such rule-driven metrics face two inherent limitations in motion assessment: 1) Subjectivity in metric design—predefined rules poorly generalize to diverse motion patterns beyond designer priors; 2) Fragmentary analysis—isolated measurements of individual attributes fail to capture holistic motion quality. Consequently, these heuristics overlook complex phenomena like inertia preservation or multi-object interaction coherence. + +Emerging methods [11, 27] employ multimodal large language models (MLLMs) fine-tuned on human preference data to predict perceptual video quality scores. VideoScore [27] aligns scoring with human judgments through preference learning, while VideoPhy [11] specifically targets physical plausibility via physics-focused annotations. Despite improved alignment with subjective evaluations, these approaches exhibit critical limitations for motion assessment: 1) Oversimplified scoring granularity – scalar outputs + +obscure specific motion defects; 2) Annotation bias amplification – training data biases toward conspicuous artifacts neglect subtle kinematic violations; 3) Contextual rigidity – domain-specific tuning limits generalization to diverse motion types. While advancing automated evaluation, they struggle to disentangle motion quality from global preference impressions. + +Current evaluation paradigms remain fundamentally misaligned with human perception of motion quality due to the above limitations. To bridge this gap, we introduce 5 distinct human-aligned evaluation metrics that systematically quantify motion quality. + +# 3. VMBench + +We introduce VMBench, a comprehensive Video Motion Benchmark that includes perception-aligned motion metrics and features the most diverse range of motion types. + +# 3.1. Perception-Driven Motion Evaluation Metrics + +When observing videos, the human brain first constructs a holistic understanding of the scene based on prior experiences and physical laws [23], before selectively attending to moving objects to assess motion smoothness and temporal consistency, particularly when objects are momentarily occluded. While rapid movement can effectively capture attention, excessive deviations from typical motion patterns may induce visual disorientation. As illustrated in Fig. 5 in the Appendix, our motion evaluation metrics are designed to emulate this hierarchical perceptual process, transitioning from global scene comprehension to localized motion analysis to ensure coherence and logical consistency. + +Commonsense Adherence Score (CAS) is used to evaluate whether generated videos align with human commonsense. As shown in Fig. 2 (a), we train a specialized model capable of assessing the commonsense quality of video content, categorizing it into five levels: Bad, Poor, Fair, Good, and Perfect. First, we collect a dataset comprising $10\mathrm{k}$ generated videos, covering both legacy approaches and popular models [2-4, 15, 16, 37, 45, 90, 96]. Second, we establish perceptual ground truth through systematic pairwise comparisons using VideoReward [47], a reward model trained on $182\mathrm{k}$ human-annotated video preference pairs. This process generates normalized preference scores reflecting human perceptual consensus. Third, we discretize the human preference into five cognitive labels. Using these preference labels, we develop a video classification model to serve as the final CAS predictor, utilizing VideoMAEv2 [77], a spatiotemporal vision transformer. Compared to MLLMs, this architecture demonstrates temporal modeling capabilities through dense frame sampling and 3D convolution operations. We compute the CAS using a Mean Opinion Score (MOS) [74, 85], where predicted probabilities for each class are weighted by their corresponding quality co + +efficients: $\mathrm{CAS} = \sum_{i=1}^{5} p_i G(i)$ , where $p_i$ denotes the predicted probability for the $i$ -th class, and $G_i$ represents the mapping function that converts category to quality weights: $G : i \to w$ (e.g., mapping class 3 to 0.5). For more details, please refer to our Appendix B.1. + +Motion Smoothness Score (MSS) aims to detect unsmoothness, like low-level temporal artifacts and high-level motion blur. It addresses the limitations of prior metrics that suffered from either low-level optical flow bias [38, 39] or oversimplified motion modeling [42], both leading to misalignment with human perception. Considering that natural motion patterns inherently maintain consistent visual quality across temporal sequences, we leverage Q-Align's [85] aesthetic score to detect artifacts, as illustrated in Fig. 2 (b). An artifact frame is confirmed when a significant drop in Q-Align scores between consecutive frames exceeds a predefined threshold, mimicking human perception of sustained visual anomalies. Further, our approach employs a scene-aware adaptive thresholding mechanism derived from the statistical modeling of real video segments [14, 54] across diverse motion patterns. The final MSS is computed as: $\mathrm{MSS} = 1 - \frac{1}{T}\sum_{t=2}^{T}\mathbb{I}(\Delta Q_t > \tau_s(t))$ where $\Delta Q_t$ is the frame-to-frame visual quality degradation magnitude, $\tau_s(t)$ indicates the adaptive threshold and $\mathbb{I}$ is an indicator function. More details are provided in Appendix B.2. + +Object Integrity Score (OIS) detects implausible deformations through spatiotemporal analysis of object integrity. Our approach addresses limitations found in previous methods [13], which primarily focus on object-level semantic consistency. Instead, we focus on detecting perceptual issues (e.g., distorted shapes) that are readily noticeable to the human visual system. As illustrated in Fig. 2 (c), we first utilize the MMPose toolkit [20] to detect key points of the primary subjects in the generated videos, which are then used to estimate the subjects' shapes in each frame. Then, we quantify shape distortions by analyzing whether object shapes violate real-world anatomical constraints. We establish tolerance thresholds for the above constraints through the statistical analysis of natural motion samples [14, 54, 87]. Anatomical violations across consecutive frames will be detected; see more details about anatomical constraints in our appendix B.3. The OIS quantifies structural consistency through: $\mathrm{OIS} = \frac{1}{F.K}\sum_{f = 1}^{F}\sum_{k = 1}^{K}\mathbb{I}\left(\mathcal{D}_f^{(k)}\leq \tau^{(k)}\right)$ , in which $F$ and $K$ denote total frame count and anatomical components, respectively. $\mathcal{D}_f^{(k)}$ represents the compound anatomical deviation for component $k$ in frame $f$ , and $\tau^{(k)}$ indicates statistical thresholds. + +Perceptible Amplitude Score (PAS) estimates subject motion by separating it from camera motion, unlike traditional motion magnitude estimation methods [73], which often overestimate overall motion due to the influence of camera movement. As presented in Fig. 2 (d), PAS begins with semantic anchoring via GroundingDINO [48] to lo + +calize moving subjects and distinguish them from passively moving elements. Next, GroundedSAM [62] ensures temporally stable subject masks, enabling CoTracker [35] to track key points with high precision. The motion magnitude is then computed based on the average displacement of these key points. Furthermore, PAS accounts for the context-dependent nature of human motion perception, adapting thresholds across different scenarios. To quantify this variability, we derive a set of perceptual motion magnitude thresholds for various scenarios through the statistical analysis of existing video datasets [14, 54]. These thresholds serve as the foundation for computing a motion score for each video. The PAS is computed as PAS = $\frac{1}{T}\sum_{t=1}^{T}\min\left(\frac{\bar{D}_t}{\tau_s}, 1\right)$ , where $\bar{D}_t$ denotes frame-level motion amplitude, computed as the average displacement of tracked key points for active subjects in frame $t$ and $\tau_s$ is the perceptual motion threshold for scenario $s$ . More Details can be found in appendix B.4. + +Temporal Coherence Score (TCS) detects the temporal consistency of object motion by analyzing frame-to-frame transitions, addressing the limitation of existing metrics [13, 59] that struggle to differentiate between natural movement and abrupt, unrealistic changes. We combine video object tracking with rule-based validation to detect abnormal disappearance/reappearance patterns, thereby distinguishing natural progression from temporal discontinuities. As shown in Fig. 2 (e), first GroundedSAM2 [61] performs pixel-accurate instance segmentation and tracking across frames, maintaining persistent object IDs throughout the whole sequence. For objects exhibiting discontinuous existence, we apply a secondary verification phase using the CoTracker [35]. It tracks dense key points on target objects and constructs their motion trajectories. We then analyze their motion trajectories to determine whether any anomalous phenomena are present. Our approach mitigates false cases caused by legitimate object discontinuity through a rule-based filtering mechanism implemented via trajectory analysis. These rules account for common scenarios, including 1) objects reappearing after occlusion or disappearing behind obstacles, 2) objects entering or exiting frame boundaries, and 3) apparent size changes due to depth perception, such as objects appearing larger when moving closer or smaller when moving farther away. Through this filtering, we systematically exclude normal events. The TCS quantifies anomalous discontinuities via $\mathrm{TCS} = 1 - \frac{1}{N}\sum_{i=1}^{N}\mathbb{I}(\mathcal{A}_i\wedge \neg \mathcal{R})$ , where $N$ is the total object instances, $\mathcal{A}_i$ indicates the existence status of the $i$ -th object, and $\mathcal{R}$ validates legitimate transitions (see Appendix B.5). + +# 3.2. Meta-Guided Motion Prompt Generation + +The existing benchmarks [27, 33, 50] are constrained by limited and simplistic motion types (As shown in Fig. 12 + +of the Appendix), making them inadequate for comprehensively evaluating the motion generation capabilities of video generation models. To address this limitation, we propose an extensible framework—Meta-Guided Motion Prompt Generation (MMPG)—that generates prompts capturing complex motion patterns, as illustrated in Fig. 3. Built upon MMPG, we introduce VMBench, a benchmark specifically designed for rigorous evaluation of complex motion generation. We find that VMBench provides the most comprehensive coverage of motion types and the most detailed prompt descriptions, making it an effective benchmark for evaluating the dynamic motion generation capabilities of video generation models. In this section, we elaborate on the process of prompt generation and refinement. More details are available in Appendix C.1. + +Meta-information Extraction. We decompose motion descriptions into three key metadata elements: Subject $(S)$ , Location $(P)$ and Action $(A)$ . We employ Qwen-2.5 [71] to extract the metadata library $\{S, P, A\}$ from existing videotext pairs [18, 58, 79, 89]. To ensure novel prompts and enhance textual diversity, we expand the metadata in three key aspects. For subject $(S)$ classification, we categorize subject types into human, animal, and object, curate a list of principal nouns identifiable by [48, 62], and use GPT-4o [6] to generate descriptions with varying entity counts (1, 2, n). For place $(P)$ descriptions, we incorporate data from [97], filtering out redundant information. For action $(A)$ expansion, we sample human actions from [14] and use LLM to extend possible actions for animals and objects. + +Self-Refining Prompt Generation. We randomly sample elements from the metadata library $\{S, P, A\}$ to form metadata sets $(S^i, P^j, A^k)$ and use GPT-4o to assess their semantic coherence and logical consistency. For coherent sets, we generate prompts of 10 to 60 words that describe the corresponding motion scenarios. To ensure accuracy, we introduce a verification mechanism inspired by [50], where the generated prompts are re-evaluated by GPT-4o to verify their consistency with the metadata elements. Through an iterative refinement and filtering process, we generate approximately 50,000 high-quality prompt candidates. However, grammatical correctness alone is insufficient: prompts must also align with physical reality to effectively assess video content. Therefore, we also eliminate descriptions of unrealistic or physically implausible scenarios to ensure the feasibility and authenticity of the prompts. + +Human-LLM Joint Validation. To reduce manual labor and enhance evaluation effectiveness, we further examine and filter prompts through a Human-LLM Joint Validation process. This process combines the strengths of both automated reasoning and human judgment to ensure high-quality motion descriptions. We first leverage the powerful reasoning capabilities of Deepseek R1 [21] to evaluate whether each prompt describes a reasonably realistic mo + +![](images/4a1b1654da10edf229bed4f040a00169ee24b25531f71a963aa579b133ba64c0.jpg) +Figure 3. Framework of our Meta-Guided Motion Prompt Generation (MMPG). MMPG consists of three stages: (a) Meta-information Extraction: Extracting Subjects, Places, and Actions from datasets such as VidProm [79], Didemo [8], MSR-VTT [89], WebVid [10], Place365 [97], and Kinect-700 [14]. (b) Self-Refining Prompt Generation: Generating and iteratively refining prompts based on the extracted information. (c) Human-LLM Joint Validation: Validating the prompts through a collaborative process between humans and DeepSeek-R1 to ensure their rationality. + +tion. After filtering out implausible descriptions, we curate a diverse set of 1,050 prompts with varied metadata. Examples are available in Appendix C.2. + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation Details. Our benchmark evaluates six popular text-to-video models: OpenSora [96], CogVideoX [90], OpenSora-Plan [45], Mochi 1 [70], HunyuanVideo [37], and Wan2.1 [72]. To provide a richer variety of motion types, we develop the MMPG-set, which comprises 1,050 prompts based on MMPG across six movement modes to assess performance in motion generation. Each model generates 1,050 videos corresponding to the MMPG-set, resulting in a total of 6,300 videos. To ensure a fair comparison, we maintain the hyperparameter settings as defined in each model's project demo. For each prompt, we generate one video for evaluation using only the initial seed. The inference process is executed using 8 Nvidia H20 GPUs. Details of the inference process for each model can be found in Appendix D.1. We then randomly sample 200 videos from each model's output, resulting in a total of 1,200 videos for human-aligned validation experiments. + +Comparison Metrics. The current comparison approaches fall into two categories: Rule-based metrics and Multimodal Large Language Model (MLLM) prompting. (1) Rule-based metrics assess four dimensions, including Perceptible Amplitude, which evaluates through RAFT optical flow magnitude analysis combined with structural motion consistency via 4 frame SSIM averaging [73, 82], following established protocols [33]. Temporal Coherence is measured using DINO [13] and CLIP [59] feature tracking, which relies on cosine similarity between consecu + +tive frames. Motion Smoothness is assessed through a hybrid method that combines interpolation error [42] with Dover's video quality assessment [84], while Object Integrity is evaluated via dual validation using optical flow warping error [38, 39] and semantic consistency checks. (2) The MLLM evaluation includes five cutting-edge models: LLaVA-NEXT-Video [95], MiniCPM-V-2.6 [91], InternVL2.5 [19], Qwen2.5-VL [9], and InternVideo2.5 [81]. These models are evaluated through a standardized assessment process, which involves processing uniformly sampled video frames at 2 frames per second to maintain essential motion patterns while managing computational limits. The evaluation dimensions encompass five aspects: amplitude, coherence, integrity, smoothness, and commonsense adherence, with scoring for each dimension using a 1–5 scale. In particular, to ensure fair comparison, consistent frame sequences and evaluation criteria are maintained across all models. + +Metrics. 1) Spearman correlation [93]: Spearman's rank correlation coefficient, often denoted by $\rho$ , is a measure of the strength and direction of the association between two ranked variables. It assesses how well the relationship between two variables can be described by a monotonic function. Spearman correlation is nonparametric and can effectively evaluate associations in datasets where variables may not follow a normal distribution. Unlike Pearson correlation, which captures linear relationships, Spearman correlation focuses on rank-based associations, making it more robust to outliers and suitable for ordinal data or scenarios with nonlinear dependencies. 2) Accuracy: To validate how accurately our metrics align with human preferences, we conduct pairwise comparisons across the 1,200 annotated videos (200 prompts $\times$ 6 models). For each prompt, we + +
MethodAvg. ρ(↑)CAS ρ(↑)MSS ρ(↑)OIS ρ(↑)PAS ρ(↑)TCS ρ(↑)
Rule-based
SSIM [82]1.6-0.9-12.18.317.8-4.8
RAFT* [73]-1.7-0.7-17.0-16.647.7-21.9
CLIP* [59]15.021.536.531.7-42.728.0
DINO* [13]9.411.027.227.4-45.827.4
Dover Technical† [84]20.640.232.634.5-6.22.2
AMT* [33]3.62.818.119.6-40.417.8
Warping Error† [50]-6.0-11.4-19.1-9.033.0-23.6
MLLM Prompt
LLaVa-NEXT-Video [95]17.115.211.016.130.812.4
MiniCPM-V-2.6 [91]16.26.817.5-2.227.331.6
IntenVL2.5 [19]19.318.720.514.142.51.1
Qwen2.5-VL [9]15.324.719.83.714.214.0
InternVideo2.5 [81]26.922.721.929.644.315.8
Model-Rule Synergy
PPM (Ours)62.269.977.165.865.254.5
+ +evaluate all 15 possible video pairs generated by different models (total 3,000 pairs). Human preference labels are determined by comparing the average expert ratings across all five dimensions (OIS, MSS, CAS, TCS, PAS) – the video with higher aggregated human scores is designated as the “ground truth” preferred sample. Concurrently, we compute metric preferences by comparing videos' aggregated PMM scores under identical criteria. Alignment accuracy is quantified as the percentage of pairs where PMM preferences match human judgments, with ties excluded to prioritize unambiguous decisions. + +# 4.2. Comparison with State-of-the-Art + +Human-aligned Validation Mechanism. We invite three domain experts to independently annotate each sample based on PMM (Perceptible Amplitude, Temporal Coherence, Object Integrity, Motion Smoothness, and Common-sense Adherence), resulting in 6,000 detailed ratings with high interannotator agreement. Finally, we obtain Likert scores [53] from humans for each video across the five dimensions included in PMM. More details can be found in Appendix D.2. We assess the alignment between metrics and human perception by calculating the Spearman correlation [93] between metric scores and expert ratings. A higher Spearman correlation signifies a stronger alignment with human perception. + +Comparison with Alternative Metrics. As shown in Table 1, in the MSS evaluation, even advanced metrics + +Table 1. Correlation Analysis Between Evaluation Metrics and Human Scores via Spearman Correlation Coefficient ( $\rho \times 100$ ). Superscript * and † denote implementations following VBenchmark [33] and EvalCrafter [50] respectively. Yellow backgrounds in Rule-based indicate specific dimension baseline. + +
CASMSSOISPASTCSAccuracy (%) ↑
Removal-based ablation
✓✓✓✓✓70.6
✓✓✓✓✗66.9
✓✓✓✗✓68.7
✓✓✗✓✓64.6
✓✗✓✓✓65.2
✗✓✓✓✓64.1
Addition-based ablation
✗✗✗✗✗-
✓✗✗✗✗58.9
✓✓✗✗✗66.1
✓✓✓✗✗67.3
+ +Table 2. Ablation Study on our Motion Metrics. Prediction accuracy (\%) of different metric combinations against human preferences is calculated. Removal-based ablation show the effect of individually ablating each metric, while in the addition-based ablation, we progressively add each metric to observe its impact. + +such as AMT [42] (18.1%) and Warping Error [39] (- 19.1%) demonstrate limited discriminative power and produce counterintuitive results, particularly in the context of complex deformations. The OIS evaluation also highlights similar shortcomings, with DINO [13] at 27.4% and Dover [84] at 34.5%, both of which fail to capture human sensitivity to structural preservation during motion. In relation to the evaluation of PAS, the SSIM [82] and RAFT [73] metric based on rules show an alignment efficacy of only 17.8% and 47.7%, respectively. In contrast, our approach achieves a remarkable 65.2% alignment efficacy. For the TCS evaluation, the rule-based metrics CLIP [59] and DINO [13] achieve only 28.0% and 27.4% alignment efficacy, respectively, and fail to capture human tolerance for minor inconsistencies while maintaining physical plausibility. In comparison, our method achieves an impressive 54.5% alignment efficacy. VBench [33] includes RAFT, CLIP, DINO, and AMT, while EvalCrafter [50] incorporates Dover Technical and Warping Error. According to the data presented in the table, overall, compared to ours, the motion evaluations of VBench [33] and EvalCrafter [50] reveal significantly lower correlations with human perception. MLLMs exhibit partial competence in evaluations of the Physical Adequacy Score (PAS) (e.g., InternVideo2.5: 44.3%) Despite their general capabilities, MLLMs average only a correlation of 10.0%-30.0% in all dimensions, highlighting a fundamental misalignment with human perception in evaluating motion quality. + +# 4.3. Ablation Study + +Ablation Study of Motion Metrics. According to Table 2, removing any single metric from the evaluation of the + +basic variants results in a significant decline in accuracy, which highlights the importance of each dimension's evaluation metric within the overall framework. Notably, the removal of the CAS metric leads to the most pronounced performance drop, with accuracy falling to $64.1\%$ , exceeding the impact of other single-dimension ablations. This demonstrates the critical role of the CAS metric in assessing video quality, aligning closely with the key factors prioritized by humans when perceiving video quality. For the performance-oriented variants, we emulate the human perceptual information processing flow by gradually adding evaluation metrics (refer to Fig. 5 in the Appendix), with each additional metric significantly enhancing overall accuracy. This not only demonstrates the effectiveness of incrementally increasing evaluation dimensions but also further verifies the consistency of our evaluation approach with human perceptual mechanisms. + +# 4.4. Qualitative Analysis + +Alignment of PMM with Human Perception. As shown in Fig. 4, the correlation between human scores across these five evaluation dimensions and our evaluation metrics in the same five dimensions remains consistent. For example, both demonstrate a strong correlation between OIS, CAS, and MSS, as well as a weak correlation between PAS and the other metrics. Specifically, as shown in Fig. 4 (a), PAS has a negative correlation with other dimensions, such as the Object Integrity Score (OIS), where $\rho = -0.18$ . This might be because high dynamic amplitudes in videos cause distortions and artifacts, which lower structural and temporal coherence scores. Conversely, OIS shows strong positive correlations with MSS ( $\rho = 0.59$ ) and CAS ( $\rho = 0.50$ ), suggesting that it reflects physical plausibility and motion rationality well. TCS shows low correlation with other dimensions, which can provide a more comprehensive assessment perspective. PAS and structural/temporal metrics exhibit a negative correlation, which challenges conventional optical-flow-based evaluation frameworks. Additionally, the isolation of PAS highlights the importance of separately disentangling motion magnitude in motion video evaluation. Fig. 4 (b) shows the correlations among the proposed evaluation metrics, which align with human perception. + +Assessing Video Generation Models with PMM. As shown in Table 3, we evaluate several leading video generation models using our PMM metrics, including Mochi 1[70], OpenSora [96], CogVideoX [90], OpenSora-Plan [45], HunyuanVideo [37], and Wan2.1 [72]. Our findings indicate that Wan2.1 delivers the best performance in generating motion videos, producing results that appear more realistic. For a clearer visual comparison of the dynamic videos generated by each model, please refer to Fig. 16 in the Appendix. + +![](images/975c5962c443a53c7c3a37e190d605d199e929e99b31e7066a2e2ba34042ad77.jpg) +Figure 4. Correlation Matrix Analysis of Metrics Within Different Evaluation Mechanisms. (a): Spearman Correlation Matrices for human annotations; (b): Spearman Correlation Matrices for our PMM metrics. + +![](images/7f7435a549138d51437d9fbeafb81157ecb59b2cf65f916d5f05647b8f3d54f9.jpg) + +
ModelsAvg.(%) ↑CASMSSOISPASTCS
Mochi 1[70]53.237.762.068.614.483.6
OpenSora [96]51.631.261.973.03.488.5
CogVideoX [90]60.650.661.675.424.691.0
OpenSora-Plan [45]58.939.376.078.66.094.7
HunyuanVideo [37]63.451.981.665.826.196.3
Wan2.1 [72]78.462.884.266.017.997.8
+ +Table 3. Performance of Video Generation Models on VMBench. We evaluated six open-source video generate models [37, 45, 70, 72, 90, 96] using VMBench. A higher score indicates better performance for a category. + +# 5. Conclusion + +Video motion authenticity remains a critical challenge in the creation of generate content. In response, we introduce VMBench, the first open-source benchmark for motion quality evaluation, integrating motion metrics with human-aligned assessment to reveal deficiencies in existing models' ability to generate physically plausible movements. To support the research community, we offer three essential resources: 1) a standardized framework for detecting motion artifacts overlooked by traditional metrics; 2) actionable diagnostic tools aimed at guiding model optimization; and 3) quantitative standards that align technical advancements with human perceptual expectations. By establishing a common framework for development in this field, VMBench enables systematic tracking of motion generation, encourages targeted model improvements, and ultimately supports the creation of videos that achieve both visual fidelity and dynamic realism. However, while the benchmark's evaluation metrics are aligned with general human perception, they may not fully capture subtle differences in perception that stem from individual viewer experiences and preferences. Addressing these limitations presents opportunities for further refinement, aiming to ensure a more comprehensive and adaptable approach to video quality evaluation. + +# References + +[1] Runway gen3. Accessed February 25, 2025 [Online] https://app.runwayml.com/, 2025.2 +[2] Sora. Accessed February 25, 2025 [Online] https://openai.com/index/sora/, 2025.1,2,4 +[3] kling. Accessed February 25, 2025 [Online] https://klingai.com/, 2025.1, 3 +[4] Pika labs. Accessed February 25, 2025 [Online] https://www.pika.art/, 2025.4 +[5] Veo 2. Accessed February 25, 2025 [Online] https://deepmind.google/technologies/veo/veo-2/, 2025.2,3 +[6] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5 +[7] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2 +[8] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2, 6 +[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7 +[10] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6 +[11] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 3 +[12] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 2 +[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 5, 6, 7 +[14] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 2, 4, 5, 6 +[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. + +Videocrafter1: Open diffusion models for high-quality video generation, 2023. 2, 4 +[16] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models, 2024. 2, 4 +[17] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2 +[18] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 5 +[19] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6, 7 +[20] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4 +[21] DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. 2, 5 +[22] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2 +[23] Daniel J. Felleman and David C. van Essen. Distributed hierarchical processing in the primate cerebral cortex. *Cerebral Cortex*, 1 1:1-47, 1991. 4 +[24] Xiaokun Feng, Haiming Yu, Meiqi Wu, Shiyu Hu, Jintao Chen, Chen Zhu, Jiahong Wu, Xiangxiang Chu, and Kaiqi Huang. Narrlv: Towards a comprehensive narrative-centric evaluation for long video generation models. arXiv preprint arXiv:2507.11245, 2025. 3 +[25] Deborah M Gordon. The ecology of collective behavior. PLoS biology, 12(3):e1001805, 2014. 1 +[26] Shir Gur, Sagie Benaim, and Lior Wolf. Hierarchical patch vae-gan: Generating diverse videos from a single sample. Advances in Neural Information Processing Systems, 33: 16761-16772, 2020. 2 +[27] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv preprint arXiv:2406.15252, 2024. 1, 3, 5 +[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2 +[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 3 + +[30] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2 +[31] Jonathan Ho, Tim Salimans, Alexey Gritsanko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 +[32] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[33] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 1, 2, 3, 5, 6, 7 +[34] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22747-22757, 2023. 2 +[35] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024.5 +[36] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 8:153113-153122, 2020. 2 +[37] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 4, 6, 8 +[38] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 4, 6 +[39] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. Advances in Neural Information Processing Systems, 33:1083-1093, 2020. 4, 6, 7 +[40] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3 +[41] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2 +[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings + +of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 4, 6, 7 +[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 1 +[44] Mingxiang Liao, Qixiang Ye, Wangmeng Zuo, Fang Wan, Tianyu Wang, Yuzhong Zhao, Jingdong Wang, Xinyu Zhang, et al. Evaluation of text-to-video generation models: A dynamics perspective. Advances in Neural Information Processing Systems, 37:109790-109816, 2025. 1 +[45] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2, 4, 6, 8 +[46] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 3 +[47] Jie Liu, Gongye Liu, Jiajun Liang, Ziyang Yuan, Xiaokun Liu, Mingwu Zheng, Xiele Wu, Qiulin Wang, Wenyu Qin, Menghan Xia, et al. Improving video generation with human feedback. arXiv preprint arXiv:2501.13918, 2025. 4 +[48] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 5 +[49] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. Advances in Neural Information Processing Systems, 36:62352-62387, 2023. 1 +[50] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 1, 2, 3, 5, 7 +[51] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025.3 +[52] Rayeesa Mehmood, Rumaan Bashir, and Kaiser J Giri. Vtmgan: video-text matcher based generative adversarial network for generating videos from textual description. International Journal of Information Technology, 16(1):221-236, 2024. 2 +[53] Tomoko Nemoto and David Beglar. Likert-scale questionnaires. In JALT 2013 conference proceedings, pages 1-6, 2014. 7 +[54] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Pro + +ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 4, 5 +[55] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1 +[56] TA Palomaki, JD Teufel, RW Simmonds, and Konrad W Lehnert. Entangling mechanical motion with microwave fields. Science, 342(6159):710-713, 2013. 1 +[57] Marina A Pavlova. Biological motion processing as a hallmark of social cognition. Cerebral cortex, 22(5):981-995, 2012. 1 +[58] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, et al. Movie gen: A cast of media foundation models, 2025. 2, 5 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5, 6, 7 +[60] Ralph H. Rasshofer, Martin Spies, and Hans Spies. Influences of weather phenomena on automotive laser radar systems. Advances in radio science, 9:49-60, 2011. 1 +[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5 +[62] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 5 +[63] Zhongwei Ren, Yunchao Wei, Xun Guo, Yao Zhao, Bingyi Kang, Jiashi Feng, and Xiaojie Jin. Videoworld: Exploring knowledge learning from unlabeled videos. arXiv preprint arXiv:2501.09781, 2025. 2 +[64] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 1, 3 +[65] Benjamin Scellier and Yoshua Bengio. Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in computational neuroscience, 11:24, 2017. 1 +[66] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[67] Jos Stam. Stable fluids. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 779-786. 2023. 1 + +[68] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 1 +[69] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 1 +[70] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 8 +[71] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5 +[72] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 6, 8 +[73] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 4, 6, 7 +[74] I Telecom. Recommendation 500-10: Methodology for the subjective assessment of the quality of television pictures. ITU-R Rec. BT, 500, 2000. 4 +[75] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 1, 3 +[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report, 2023. 2 +[77] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14549-14560, 2023. 4 +[78] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. arXiv preprint arXiv:2307.00040, 2(3):4, 2023. 1 +[79] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. 2024. 1, 5, 6 +[80] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. Magicvideo-v2: Multi-stage high-aesthetic video generation, 2024. 2 +[81] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 6, 7 +[82] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7 +[83] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Gen + +erating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2 +[84] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 2, 6, 7 +[85] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 4 +[86] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation, 2023. 2 +[87] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 4 +[88] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. 2023. 2 +[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 5, 6 +[90] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 4, 6, 8 +[91] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6, 7 +[92] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 2 +[93] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 2, 6, 7 +[94] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 1 +[95] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 6, 7 +[96] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 1, 2, 4, 6, 8 + +[97] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 1, 5, 6 +[98] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2 \ No newline at end of file diff --git a/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/images.zip b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c442a366d289edd3a6121e104998be021b8f96f --- /dev/null +++ b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6c10aaf575e296bf4d1a01a61e2030568f113296d5a51ff3d5d5c998561e8bb +size 554837 diff --git a/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/layout.json b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b96099bca1e40319f6533e85c557e8c953359867 --- /dev/null +++ b/2025/VMBench_ A Benchmark for Perception-Aligned Video Motion Generation/layout.json @@ -0,0 +1,8322 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 537, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 537, + 120 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 537, + 120 + ], + "type": "text", + "content": "VMBench: A Benchmark for Perception-Aligned Video Motion Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 161, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 161, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 161, + 171 + ], + "type": "text", + "content": "Xinran Ling" + }, + { + "bbox": [ + 77, + 142, + 161, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 142, + 161, + 171 + ], + "type": "text", + "content": " \nCundian Yang" + }, + { + "bbox": [ + 77, + 142, + 161, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 143, + 245, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 143, + 245, + 171 + ], + "spans": [ + { + "bbox": [ + 181, + 143, + 245, + 171 + ], + "type": "text", + "content": "Chen Zhu1* \nAiming Hao1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 264, + 143, + 330, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 143, + 330, + 171 + ], + "spans": [ + { + "bbox": [ + 264, + 143, + 330, + 171 + ], + "type": "text", + "content": "Meiqi Wu" + }, + { + "bbox": [ + 264, + 143, + 330, + 171 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 264, + 143, + 330, + 171 + ], + "type": "text", + "content": " \nJiashu Zhu" + }, + { + "bbox": [ + 264, + 143, + 330, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 353, + 143, + 423, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 143, + 423, + 171 + ], + "spans": [ + { + "bbox": [ + 353, + 143, + 423, + 171 + ], + "type": "text", + "content": "Hangyu Li1 \nJiahong Wu1†" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 438, + 143, + 529, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 143, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 438, + 143, + 529, + 171 + ], + "type": "text", + "content": "Xiaokun Feng" + }, + { + "bbox": [ + 438, + 143, + 529, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 438, + 143, + 529, + 171 + ], + "type": "text", + "content": " \nXiangxiang Chu" + }, + { + "bbox": [ + 438, + 143, + 529, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 171, + 196, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 171, + 196, + 186 + ], + "spans": [ + { + "bbox": [ + 77, + 171, + 196, + 186 + ], + "type": "text", + "content": "1 AMAP, Alibaba Group" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 222, + 171, + 534, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 171, + 534, + 185 + ], + "spans": [ + { + "bbox": [ + 222, + 171, + 534, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 222, + 171, + 534, + 185 + ], + "type": "text", + "content": " CRISE, Institute of Automation, Chinese Academy of Sciences" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 186, + 525, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 186, + 525, + 199 + ], + "spans": [ + { + "bbox": [ + 82, + 186, + 525, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 82, + 186, + 525, + 199 + ], + "type": "text", + "content": " School of Computer Science and Technology, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 254, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 588 + ], + "type": "text", + "content": "Video generation has advanced rapidly, improving evaluation methods, yet assessing video's motion remains a major challenge. Specifically, there are two key issues: 1) current motion metrics do not fully align with human perceptions; 2) the existing motion prompts are limited. Based on these findings, we introduce VMBench—a comprehensive Video Motion Benchmark that has perception-aligned motion metrics and features the most diverse types of motion. VMBench has several appealing properties: (1) Perception-Driven Motion Evaluation Metrics, we identify five dimensions based on human perception in motion video assessment and develop fine-grained evaluation metrics, providing deeper insights into models' strengths and weaknesses in motion quality. (2) Meta-Guided Motion Prompt Generation, a structured method that extracts meta-information, generates diverse motion prompts with LLMs, and refines them through human-AI validation, resulting in a multilevel prompt library covering six key dynamic scene dimensions. (3) Human-Aligned Validation Mechanism, we provide human preference annotations to validate our benchmarks, with our metrics achieving an average " + }, + { + "bbox": [ + 56, + 254, + 296, + 588 + ], + "type": "inline_equation", + "content": "35.3\\%" + }, + { + "bbox": [ + 56, + 254, + 296, + 588 + ], + "type": "text", + "content": " improvement in Spearman's correlation over baseline methods. This is the first time that the quality of motion in videos has been evaluated from the perspective of human perception alignment. Additionally, we release VMBench at https://github.com/AMAP-ML/VMBench, setting a new standard for evaluating and advancing motion generation models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 635, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 682 + ], + "type": "text", + "content": "Video generation models [2, 3, 37, 43, 55, 78, 90] have advanced significantly, particularly text-to-video (T2V) models [12, 45, 70, 96], which perform well in generating static content but struggle to capture motion in alignment with hu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 228, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 251 + ], + "type": "text", + "content": "man perception. To enhance dynamic motion generation, it is crucial to establish perception-aligned motion evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 254, + 553, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 349 + ], + "type": "text", + "content": "However, previous methods [29, 49, 59, 64, 75] primarily assess the static content of videos, with dynamic evaluation largely limited to motion smoothness [33, 44, 50]. This makes it challenging to capture motion deficiencies [69, 94], including but not limited to spatiotemporal inconsistencies and violations of physical laws. Therefore, developing a more comprehensive evaluation metric that better aligns with human perception is crucial." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 351, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 457 + ], + "type": "text", + "content": "To address this, we introduce Video Motion Benchmark (VMBench) (Fig. 1), the first benchmark focused on comprehensive video motion assessment. It features the most diverse range of motion types, covering 969 categories—more than any other benchmark [27, 33, 44, 49, 50, 68] (see Fig. 12 in the Appendix). Additionally, it integrates perception-aligned motion metrics to systematically evaluate the motion generation quality of video generation models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 604 + ], + "type": "text", + "content": "First, we introduce Perception-Driven Motion Evaluation Metrics (PMM), the first evaluation framework explicitly designed to align with human perception of motion quality. PMM comprises five key components: Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), Temporal Coherence Score (TCS), Motion Smoothness Score (MSS), and Commonsense Adherence Score (CAS). Unlike previous methods that primarily focus on motion smoothness [29, 59, 64, 75], PMM provides a more comprehensive assessment by evaluating motion quality in terms of spatiotemporal inconsistencies and violations of physical laws." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Next, we propose Meta-Guided Motion Prompt Generation (MMPG), an expandable framework that encompasses the most comprehensive range of motion types, grounded in physics [56, 60, 65, 67] and cognitive science [25, 57]. The evaluation system covers six movement modes: fluid dynamics, biological motion, mechanical motion, weather phenomena, collective behavior, and energy transfer (Fig. 1). Specifically, MMPG first extracts subjects, places, and actions from VidProm [79], Place365 [97], and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 693, + 147, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 147, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 147, + 703 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13087" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 321, + 286 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 321, + 286 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 321, + 286 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 321, + 286 + ], + "type": "image", + "image_path": "882ee89ebe4bd8400faee5b6eaf319fa9c619c99fd8df1847b437cbbcc62113c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 295, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 555, + 350 + ], + "type": "text", + "content": "Figure 1. Overview of VMBench. Our benchmark encompasses six principal categories of motion patterns, with each prompt constructed as a comprehensive motion structured around three core components: subject, place, and acion. We propose a novel multi-dimensional video motion evaluation comprising five human-centric quality metrics derived from perceptual preferences. Utilizing videos generated by popular T2V models, we conduct systematic human evaluations to validate the effectiveness of our metrics in capturing human perceptual preferences." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 323, + 71, + 525, + 286 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 525, + 286 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 525, + 286 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 525, + 286 + ], + "type": "image", + "image_path": "a01b5b425fa47b0421a5d841e87209c627d28f0e20880ff739e94e7f5fd188a0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 371, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 443 + ], + "type": "text", + "content": "other datasets [8, 10, 14, 89], then leverages LLMs to enrich metadata and generate detailed motion descriptions. We validate the prompts through human-AI collaboration with DeepSeek-R1 [21] to ensure coherence and rationality. To reduce costs, we curated 1,050 user prompts and use T2V models to generate motion videos for evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "type": "text", + "content": "Finally, we provide human preference annotations to validate our benchmarks, showing that our metrics achieve an average " + }, + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "type": "inline_equation", + "content": "35.3\\%" + }, + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "type": "text", + "content": " improvement in Spearman's correlation [93] over current methods [9, 13, 33, 50, 59, 73, 81, 84, 95]. This result highlights the enhanced alignment of our evaluation metrics with human perception. We will open-source VMBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VMBench to drive forward the field of video motion generation. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 640 + ], + "type": "text", + "content": "- First-Ever Human Perception-Aligned Motion Evaluation: We pioneer the evaluation of motion quality in videos from the perspective of human perception alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 676 + ], + "type": "text", + "content": "Meta-Guided Motion Prompt Generation: A curated prompt set designed to assess diverse motion aspects in video models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "content": "- Human-Aligned Validation Mechanism: Our metrics achieve an average " + }, + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "inline_equation", + "content": "35.3\\%" + }, + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "content": " improvement in Spearman's correlation over current methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 370, + 399, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 370, + 399, + 382 + ], + "spans": [ + { + "bbox": [ + 314, + 370, + 399, + 382 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 392, + 455, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 392, + 455, + 403 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 455, + 403 + ], + "type": "text", + "content": "2.1. Video Generative Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 508 + ], + "type": "text", + "content": "With the rapid advancement of AI generative techniques—GANs [22, 26, 36, 41, 52], VAEs [41, 83], autoregressive models [17, 63, 92], and diffusion models [7, 31, 32, 34, 76, 88]—the field of T2V generation has seen remarkable progress. However, while current T2V models [16, 70, 90, 96], excel in generating static visual content, their dynamic performance still deviates significantly from human perception." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Early T2V models [15, 30, 31, 66] have built upon Text-to-Image(T2I) techniques, enhancing their temporal capabilities to generate videos. Then, some works [28, 80, 98] take this further by conducting the diffusion process within the latent space and incorporate the spatiotemporal module [16, 76, 86] within the denoising network, which greatly improves the quality of static content in videos. CogVideo [32, 90] and other closed source works [1, 5, 12, 58] represented by Sora [2] focus on high-quality data filtering on a large-scale dataset, significantly boosting model performance. OpenSora [96] and OpenSora-Plan [45] aim to showcase and open source the remarkable performance of closed-source models. Mochi 1 [70] and HunyuanVideo [37] have improved the dynamic quality of generated videos through new technologies such as asymmetric architecture and full attention mechanism. Beyond optimization of models and datasets, recent stud" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13088" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 73, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 57, + 73, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 57, + 73, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 57, + 73, + 553, + 322 + ], + "type": "image", + "image_path": "c1d85feb677bc7685884399a155860b34dd19e1eafed2b4e280979d84d7fe012.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "type": "text", + "content": "Figure 2. Framework of our Perception-Driven Motion Metrics (PMM). PMM comprises multiple evaluation metrics: Commonsense Adherence Score (CAS), Motion Smoothness Score (MSS), Object Integrity Score (OIS), Perceptible Amplitude Score (PAS), and Temporal Coherence Score (TCS). (a-e): Computational flowcharts for each metric. The scores produced by PMM show variation trends consistent with human assessments, indicating strong alignment with human perception." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 397, + 294, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 397, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 54, + 397, + 294, + 493 + ], + "type": "text", + "content": "ies [3, 5, 40, 46, 51] are beginning to explore the influence of motion on video generation. Specifically, Ma et al. [51] focuses on the motion quality of videos during data selection, in order to enable models to learn motion dynamics even at low resolution training stages. Meanwhile, some works [40, 46] improve the generative capacity of models by incorporating motion information as guidance during the sampling process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 494, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 566 + ], + "type": "text", + "content": "Although these models demonstrate commendable performance across various metrics, the generated videos still exhibit several qualitative motion deficiencies, such as the abrupt disappearance of moving subjects, body distortions, and physically implausible actions, which are difficult to assess using existing metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 574, + 274, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 274, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 274, + 586 + ], + "type": "text", + "content": "2.2. Evaluations Metrics for Video Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 677 + ], + "type": "text", + "content": "Existing approaches for video motion evaluation can be broadly categorized into three paradigms: 1) feature-based metrics leveraging pre-trained video representations, 2) rule-based metrics employing manually designed scoring mechanisms, and 3) MLLM-based assessments that fine-tune large multimodal language models on human-annotated video quality datasets for perceptual scoring." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Previous feature-based metrics demonstrate limited effectiveness in assessing motion quality. Image-centric metrics like IS [64] and FID [29] inherently ignore tempo" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 397, + 553, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 553, + 457 + ], + "type": "text", + "content": "ral coherence, while FVD [75] extends image-level metrics to videos by modeling spatiotemporal features, it reduces motion dynamics to simplistic distributions. These methods conflate feature-space distances with perceptual motion quality, overlooking artifacts like implausible movement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 603 + ], + "type": "text", + "content": "Recent benchmarks [24, 33, 50] attempt to conduct systematic evaluations using manually curated metrics. For instance, VBench [33] introduces motion-related criteria such as motion smoothness, quantified via a video frame interpolation model. However, such rule-driven metrics face two inherent limitations in motion assessment: 1) Subjectivity in metric design—predefined rules poorly generalize to diverse motion patterns beyond designer priors; 2) Fragmentary analysis—isolated measurements of individual attributes fail to capture holistic motion quality. Consequently, these heuristics overlook complex phenomena like inertia preservation or multi-object interaction coherence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Emerging methods [11, 27] employ multimodal large language models (MLLMs) fine-tuned on human preference data to predict perceptual video quality scores. VideoScore [27] aligns scoring with human judgments through preference learning, while VideoPhy [11] specifically targets physical plausibility via physics-focused annotations. Despite improved alignment with subjective evaluations, these approaches exhibit critical limitations for motion assessment: 1) Oversimplified scoring granularity – scalar outputs" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13089" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "obscure specific motion defects; 2) Annotation bias amplification – training data biases toward conspicuous artifacts neglect subtle kinematic violations; 3) Contextual rigidity – domain-specific tuning limits generalization to diverse motion types. While advancing automated evaluation, they struggle to disentangle motion quality from global preference impressions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 217 + ], + "type": "text", + "content": "Current evaluation paradigms remain fundamentally misaligned with human perception of motion quality due to the above limitations. To bridge this gap, we introduce 5 distinct human-aligned evaluation metrics that systematically quantify motion quality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 226, + 123, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 123, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 123, + 239 + ], + "type": "text", + "content": "3. VMBench" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 246, + 294, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 294, + 283 + ], + "type": "text", + "content": "We introduce VMBench, a comprehensive Video Motion Benchmark that includes perception-aligned motion metrics and features the most diverse range of motion types." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 293, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 293, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 293, + 302 + ], + "type": "text", + "content": "3.1. Perception-Driven Motion Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 307, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 450 + ], + "type": "text", + "content": "When observing videos, the human brain first constructs a holistic understanding of the scene based on prior experiences and physical laws [23], before selectively attending to moving objects to assess motion smoothness and temporal consistency, particularly when objects are momentarily occluded. While rapid movement can effectively capture attention, excessive deviations from typical motion patterns may induce visual disorientation. As illustrated in Fig. 5 in the Appendix, our motion evaluation metrics are designed to emulate this hierarchical perceptual process, transitioning from global scene comprehension to localized motion analysis to ensure coherence and logical consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "text", + "content": "Commonsense Adherence Score (CAS) is used to evaluate whether generated videos align with human commonsense. As shown in Fig. 2 (a), we train a specialized model capable of assessing the commonsense quality of video content, categorizing it into five levels: Bad, Poor, Fair, Good, and Perfect. First, we collect a dataset comprising " + }, + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "text", + "content": " generated videos, covering both legacy approaches and popular models [2-4, 15, 16, 37, 45, 90, 96]. Second, we establish perceptual ground truth through systematic pairwise comparisons using VideoReward [47], a reward model trained on " + }, + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "inline_equation", + "content": "182\\mathrm{k}" + }, + { + "bbox": [ + 55, + 451, + 295, + 714 + ], + "type": "text", + "content": " human-annotated video preference pairs. This process generates normalized preference scores reflecting human perceptual consensus. Third, we discretize the human preference into five cognitive labels. Using these preference labels, we develop a video classification model to serve as the final CAS predictor, utilizing VideoMAEv2 [77], a spatiotemporal vision transformer. Compared to MLLMs, this architecture demonstrates temporal modeling capabilities through dense frame sampling and 3D convolution operations. We compute the CAS using a Mean Opinion Score (MOS) [74, 85], where predicted probabilities for each class are weighted by their corresponding quality co" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "efficients: " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{CAS} = \\sum_{i=1}^{5} p_i G(i)" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " denotes the predicted probability for the " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "-th class, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "G_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " represents the mapping function that converts category to quality weights: " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "G : i \\to w" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " (e.g., mapping class 3 to 0.5). For more details, please refer to our Appendix B.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "text", + "content": "Motion Smoothness Score (MSS) aims to detect unsmoothness, like low-level temporal artifacts and high-level motion blur. It addresses the limitations of prior metrics that suffered from either low-level optical flow bias [38, 39] or oversimplified motion modeling [42], both leading to misalignment with human perception. Considering that natural motion patterns inherently maintain consistent visual quality across temporal sequences, we leverage Q-Align's [85] aesthetic score to detect artifacts, as illustrated in Fig. 2 (b). An artifact frame is confirmed when a significant drop in Q-Align scores between consecutive frames exceeds a predefined threshold, mimicking human perception of sustained visual anomalies. Further, our approach employs a scene-aware adaptive thresholding mechanism derived from the statistical modeling of real video segments [14, 54] across diverse motion patterns. The final MSS is computed as: " + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\mathrm{MSS} = 1 - \\frac{1}{T}\\sum_{t=2}^{T}\\mathbb{I}(\\Delta Q_t > \\tau_s(t))" + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\Delta Q_t" + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "text", + "content": " is the frame-to-frame visual quality degradation magnitude, " + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\tau_s(t)" + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "text", + "content": " indicates the adaptive threshold and " + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 313, + 133, + 553, + 371 + ], + "type": "text", + "content": " is an indicator function. More details are provided in Appendix B.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": "Object Integrity Score (OIS) detects implausible deformations through spatiotemporal analysis of object integrity. Our approach addresses limitations found in previous methods [13], which primarily focus on object-level semantic consistency. Instead, we focus on detecting perceptual issues (e.g., distorted shapes) that are readily noticeable to the human visual system. As illustrated in Fig. 2 (c), we first utilize the MMPose toolkit [20] to detect key points of the primary subjects in the generated videos, which are then used to estimate the subjects' shapes in each frame. Then, we quantify shape distortions by analyzing whether object shapes violate real-world anatomical constraints. We establish tolerance thresholds for the above constraints through the statistical analysis of natural motion samples [14, 54, 87]. Anatomical violations across consecutive frames will be detected; see more details about anatomical constraints in our appendix B.3. The OIS quantifies structural consistency through: " + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{OIS} = \\frac{1}{F.K}\\sum_{f = 1}^{F}\\sum_{k = 1}^{K}\\mathbb{I}\\left(\\mathcal{D}_f^{(k)}\\leq \\tau^{(k)}\\right)" + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": " denote total frame count and anatomical components, respectively. " + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_f^{(k)}" + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": " represents the compound anatomical deviation for component " + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\tau^{(k)}" + }, + { + "bbox": [ + 313, + 371, + 553, + 640 + ], + "type": "text", + "content": " indicates statistical thresholds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Perceptible Amplitude Score (PAS) estimates subject motion by separating it from camera motion, unlike traditional motion magnitude estimation methods [73], which often overestimate overall motion due to the influence of camera movement. As presented in Fig. 2 (d), PAS begins with semantic anchoring via GroundingDINO [48] to lo" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13090" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "text", + "content": "calize moving subjects and distinguish them from passively moving elements. Next, GroundedSAM [62] ensures temporally stable subject masks, enabling CoTracker [35] to track key points with high precision. The motion magnitude is then computed based on the average displacement of these key points. Furthermore, PAS accounts for the context-dependent nature of human motion perception, adapting thresholds across different scenarios. To quantify this variability, we derive a set of perceptual motion magnitude thresholds for various scenarios through the statistical analysis of existing video datasets [14, 54]. These thresholds serve as the foundation for computing a motion score for each video. The PAS is computed as PAS = " + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "inline_equation", + "content": "\\frac{1}{T}\\sum_{t=1}^{T}\\min\\left(\\frac{\\bar{D}_t}{\\tau_s}, 1\\right)" + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "inline_equation", + "content": "\\bar{D}_t" + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "text", + "content": " denotes frame-level motion amplitude, computed as the average displacement of tracked key points for active subjects in frame " + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "inline_equation", + "content": "\\tau_s" + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "text", + "content": " is the perceptual motion threshold for scenario " + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 72, + 294, + 292 + ], + "type": "text", + "content": ". More Details can be found in appendix B.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "text", + "content": "Temporal Coherence Score (TCS) detects the temporal consistency of object motion by analyzing frame-to-frame transitions, addressing the limitation of existing metrics [13, 59] that struggle to differentiate between natural movement and abrupt, unrealistic changes. We combine video object tracking with rule-based validation to detect abnormal disappearance/reappearance patterns, thereby distinguishing natural progression from temporal discontinuities. As shown in Fig. 2 (e), first GroundedSAM2 [61] performs pixel-accurate instance segmentation and tracking across frames, maintaining persistent object IDs throughout the whole sequence. For objects exhibiting discontinuous existence, we apply a secondary verification phase using the CoTracker [35]. It tracks dense key points on target objects and constructs their motion trajectories. We then analyze their motion trajectories to determine whether any anomalous phenomena are present. Our approach mitigates false cases caused by legitimate object discontinuity through a rule-based filtering mechanism implemented via trajectory analysis. These rules account for common scenarios, including 1) objects reappearing after occlusion or disappearing behind obstacles, 2) objects entering or exiting frame boundaries, and 3) apparent size changes due to depth perception, such as objects appearing larger when moving closer or smaller when moving farther away. Through this filtering, we systematically exclude normal events. The TCS quantifies anomalous discontinuities via " + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{TCS} = 1 - \\frac{1}{N}\\sum_{i=1}^{N}\\mathbb{I}(\\mathcal{A}_i\\wedge \\neg \\mathcal{R})" + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "text", + "content": " is the total object instances, " + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "text", + "content": " indicates the existence status of the " + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "text", + "content": "-th object, and " + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 293, + 295, + 664 + ], + "type": "text", + "content": " validates legitimate transitions (see Appendix B.5)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 672, + 272, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 272, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 272, + 685 + ], + "type": "text", + "content": "3.2. Meta-Guided Motion Prompt Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "The existing benchmarks [27, 33, 50] are constrained by limited and simplistic motion types (As shown in Fig. 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "of the Appendix), making them inadequate for comprehensively evaluating the motion generation capabilities of video generation models. To address this limitation, we propose an extensible framework—Meta-Guided Motion Prompt Generation (MMPG)—that generates prompts capturing complex motion patterns, as illustrated in Fig. 3. Built upon MMPG, we introduce VMBench, a benchmark specifically designed for rigorous evaluation of complex motion generation. We find that VMBench provides the most comprehensive coverage of motion types and the most detailed prompt descriptions, making it an effective benchmark for evaluating the dynamic motion generation capabilities of video generation models. In this section, we elaborate on the process of prompt generation and refinement. More details are available in Appendix C.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": "Meta-information Extraction. We decompose motion descriptions into three key metadata elements: Subject " + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": ", Location " + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "inline_equation", + "content": "(P)" + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": " and Action " + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": ". We employ Qwen-2.5 [71] to extract the metadata library " + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "inline_equation", + "content": "\\{S, P, A\\}" + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": " from existing videotext pairs [18, 58, 79, 89]. To ensure novel prompts and enhance textual diversity, we expand the metadata in three key aspects. For subject " + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": " classification, we categorize subject types into human, animal, and object, curate a list of principal nouns identifiable by [48, 62], and use GPT-4o [6] to generate descriptions with varying entity counts (1, 2, n). For place " + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "inline_equation", + "content": "(P)" + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": " descriptions, we incorporate data from [97], filtering out redundant information. For action " + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 313, + 255, + 553, + 421 + ], + "type": "text", + "content": " expansion, we sample human actions from [14] and use LLM to extend possible actions for animals and objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 424, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 553, + 616 + ], + "type": "text", + "content": "Self-Refining Prompt Generation. We randomly sample elements from the metadata library " + }, + { + "bbox": [ + 313, + 424, + 553, + 616 + ], + "type": "inline_equation", + "content": "\\{S, P, A\\}" + }, + { + "bbox": [ + 313, + 424, + 553, + 616 + ], + "type": "text", + "content": " to form metadata sets " + }, + { + "bbox": [ + 313, + 424, + 553, + 616 + ], + "type": "inline_equation", + "content": "(S^i, P^j, A^k)" + }, + { + "bbox": [ + 313, + 424, + 553, + 616 + ], + "type": "text", + "content": " and use GPT-4o to assess their semantic coherence and logical consistency. For coherent sets, we generate prompts of 10 to 60 words that describe the corresponding motion scenarios. To ensure accuracy, we introduce a verification mechanism inspired by [50], where the generated prompts are re-evaluated by GPT-4o to verify their consistency with the metadata elements. Through an iterative refinement and filtering process, we generate approximately 50,000 high-quality prompt candidates. However, grammatical correctness alone is insufficient: prompts must also align with physical reality to effectively assess video content. Therefore, we also eliminate descriptions of unrealistic or physically implausible scenarios to ensure the feasibility and authenticity of the prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Human-LLM Joint Validation. To reduce manual labor and enhance evaluation effectiveness, we further examine and filter prompts through a Human-LLM Joint Validation process. This process combines the strengths of both automated reasoning and human judgment to ensure high-quality motion descriptions. We first leverage the powerful reasoning capabilities of Deepseek R1 [21] to evaluate whether each prompt describes a reasonably realistic mo" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13091" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 547, + 222 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 547, + 222 + ], + "type": "image", + "image_path": "4a1b1654da10edf229bed4f040a00169ee24b25531f71a963aa579b133ba64c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 287 + ], + "type": "text", + "content": "Figure 3. Framework of our Meta-Guided Motion Prompt Generation (MMPG). MMPG consists of three stages: (a) Meta-information Extraction: Extracting Subjects, Places, and Actions from datasets such as VidProm [79], Didemo [8], MSR-VTT [89], WebVid [10], Place365 [97], and Kinect-700 [14]. (b) Self-Refining Prompt Generation: Generating and iteratively refining prompts based on the extracted information. (c) Human-LLM Joint Validation: Validating the prompts through a collaborative process between humans and DeepSeek-R1 to ensure their rationality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 307, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 344 + ], + "type": "text", + "content": "tion. After filtering out implausible descriptions, we curate a diverse set of 1,050 prompts with varied metadata. Examples are available in Appendix C.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 137, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 137, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 137, + 367 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 171, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 171, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 171, + 387 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 594 + ], + "type": "text", + "content": "Implementation Details. Our benchmark evaluates six popular text-to-video models: OpenSora [96], CogVideoX [90], OpenSora-Plan [45], Mochi 1 [70], HunyuanVideo [37], and Wan2.1 [72]. To provide a richer variety of motion types, we develop the MMPG-set, which comprises 1,050 prompts based on MMPG across six movement modes to assess performance in motion generation. Each model generates 1,050 videos corresponding to the MMPG-set, resulting in a total of 6,300 videos. To ensure a fair comparison, we maintain the hyperparameter settings as defined in each model's project demo. For each prompt, we generate one video for evaluation using only the initial seed. The inference process is executed using 8 Nvidia H20 GPUs. Details of the inference process for each model can be found in Appendix D.1. We then randomly sample 200 videos from each model's output, resulting in a total of 1,200 videos for human-aligned validation experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Comparison Metrics. The current comparison approaches fall into two categories: Rule-based metrics and Multimodal Large Language Model (MLLM) prompting. (1) Rule-based metrics assess four dimensions, including Perceptible Amplitude, which evaluates through RAFT optical flow magnitude analysis combined with structural motion consistency via 4 frame SSIM averaging [73, 82], following established protocols [33]. Temporal Coherence is measured using DINO [13] and CLIP [59] feature tracking, which relies on cosine similarity between consecu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 307, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 521 + ], + "type": "text", + "content": "tive frames. Motion Smoothness is assessed through a hybrid method that combines interpolation error [42] with Dover's video quality assessment [84], while Object Integrity is evaluated via dual validation using optical flow warping error [38, 39] and semantic consistency checks. (2) The MLLM evaluation includes five cutting-edge models: LLaVA-NEXT-Video [95], MiniCPM-V-2.6 [91], InternVL2.5 [19], Qwen2.5-VL [9], and InternVideo2.5 [81]. These models are evaluated through a standardized assessment process, which involves processing uniformly sampled video frames at 2 frames per second to maintain essential motion patterns while managing computational limits. The evaluation dimensions encompass five aspects: amplitude, coherence, integrity, smoothness, and commonsense adherence, with scoring for each dimension using a 1–5 scale. In particular, to ensure fair comparison, consistent frame sequences and evaluation criteria are maintained across all models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Metrics. 1) Spearman correlation [93]: Spearman's rank correlation coefficient, often denoted by " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": ", is a measure of the strength and direction of the association between two ranked variables. It assesses how well the relationship between two variables can be described by a monotonic function. Spearman correlation is nonparametric and can effectively evaluate associations in datasets where variables may not follow a normal distribution. Unlike Pearson correlation, which captures linear relationships, Spearman correlation focuses on rank-based associations, making it more robust to outliers and suitable for ordinal data or scenarios with nonlinear dependencies. 2) Accuracy: To validate how accurately our metrics align with human preferences, we conduct pairwise comparisons across the 1,200 annotated videos (200 prompts " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " 6 models). For each prompt, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13092" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 293 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 293 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 293 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 293 + ], + "type": "table", + "html": "
MethodAvg. ρ(↑)CAS ρ(↑)MSS ρ(↑)OIS ρ(↑)PAS ρ(↑)TCS ρ(↑)
Rule-based
SSIM [82]1.6-0.9-12.18.317.8-4.8
RAFT* [73]-1.7-0.7-17.0-16.647.7-21.9
CLIP* [59]15.021.536.531.7-42.728.0
DINO* [13]9.411.027.227.4-45.827.4
Dover Technical† [84]20.640.232.634.5-6.22.2
AMT* [33]3.62.818.119.6-40.417.8
Warping Error† [50]-6.0-11.4-19.1-9.033.0-23.6
MLLM Prompt
LLaVa-NEXT-Video [95]17.115.211.016.130.812.4
MiniCPM-V-2.6 [91]16.26.817.5-2.227.331.6
IntenVL2.5 [19]19.318.720.514.142.51.1
Qwen2.5-VL [9]15.324.719.83.714.214.0
InternVideo2.5 [81]26.922.721.929.644.315.8
Model-Rule Synergy
PPM (Ours)62.269.977.165.865.254.5
", + "image_path": "9cefc3538adc84975ffc2158ee8d6c51cf5b08d5050d0d3f0f97649ab4547f48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 377, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 510 + ], + "type": "text", + "content": "evaluate all 15 possible video pairs generated by different models (total 3,000 pairs). Human preference labels are determined by comparing the average expert ratings across all five dimensions (OIS, MSS, CAS, TCS, PAS) – the video with higher aggregated human scores is designated as the “ground truth” preferred sample. Concurrently, we compute metric preferences by comparing videos' aggregated PMM scores under identical criteria. Alignment accuracy is quantified as the percentage of pairs where PMM preferences match human judgments, with ties excluded to prioritize unambiguous decisions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 237, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 237, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 237, + 529 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 689 + ], + "type": "text", + "content": "Human-aligned Validation Mechanism. We invite three domain experts to independently annotate each sample based on PMM (Perceptible Amplitude, Temporal Coherence, Object Integrity, Motion Smoothness, and Common-sense Adherence), resulting in 6,000 detailed ratings with high interannotator agreement. Finally, we obtain Likert scores [53] from humans for each video across the five dimensions included in PMM. More details can be found in Appendix D.2. We assess the alignment between metrics and human perception by calculating the Spearman correlation [93] between metric scores and expert ratings. A higher Spearman correlation signifies a stronger alignment with human perception." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Comparison with Alternative Metrics. As shown in Table 1, in the MSS evaluation, even advanced metrics" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 329, + 70, + 541, + 236 + ], + "blocks": [ + { + "bbox": [ + 55, + 301, + 295, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 357 + ], + "type": "text", + "content": "Table 1. Correlation Analysis Between Evaluation Metrics and Human Scores via Spearman Correlation Coefficient (" + }, + { + "bbox": [ + 55, + 301, + 295, + 357 + ], + "type": "inline_equation", + "content": "\\rho \\times 100" + }, + { + "bbox": [ + 55, + 301, + 295, + 357 + ], + "type": "text", + "content": "). Superscript * and † denote implementations following VBenchmark [33] and EvalCrafter [50] respectively. Yellow backgrounds in Rule-based indicate specific dimension baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 70, + 541, + 236 + ], + "lines": [ + { + "bbox": [ + 329, + 70, + 541, + 236 + ], + "spans": [ + { + "bbox": [ + 329, + 70, + 541, + 236 + ], + "type": "table", + "html": "
CASMSSOISPASTCSAccuracy (%) ↑
Removal-based ablation
✓✓✓✓✓70.6
✓✓✓✓✗66.9
✓✓✓✗✓68.7
✓✓✗✓✓64.6
✓✗✓✓✓65.2
✗✓✓✓✓64.1
Addition-based ablation
✗✗✗✗✗-
✓✗✗✗✗58.9
✓✓✗✗✗66.1
✓✓✓✗✗67.3
", + "image_path": "bbeab31a344657c845835b5e40dfad34f57fe77eecef617d2dc17cc3dcc88568.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 300 + ], + "type": "text", + "content": "Table 2. Ablation Study on our Motion Metrics. Prediction accuracy (\\%) of different metric combinations against human preferences is calculated. Removal-based ablation show the effect of individually ablating each metric, while in the addition-based ablation, we progressively add each metric to observe its impact." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 324, + 556, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 556, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 556, + 660 + ], + "type": "text", + "content": "such as AMT [42] (18.1%) and Warping Error [39] (- 19.1%) demonstrate limited discriminative power and produce counterintuitive results, particularly in the context of complex deformations. The OIS evaluation also highlights similar shortcomings, with DINO [13] at 27.4% and Dover [84] at 34.5%, both of which fail to capture human sensitivity to structural preservation during motion. In relation to the evaluation of PAS, the SSIM [82] and RAFT [73] metric based on rules show an alignment efficacy of only 17.8% and 47.7%, respectively. In contrast, our approach achieves a remarkable 65.2% alignment efficacy. For the TCS evaluation, the rule-based metrics CLIP [59] and DINO [13] achieve only 28.0% and 27.4% alignment efficacy, respectively, and fail to capture human tolerance for minor inconsistencies while maintaining physical plausibility. In comparison, our method achieves an impressive 54.5% alignment efficacy. VBench [33] includes RAFT, CLIP, DINO, and AMT, while EvalCrafter [50] incorporates Dover Technical and Warping Error. According to the data presented in the table, overall, compared to ours, the motion evaluations of VBench [33] and EvalCrafter [50] reveal significantly lower correlations with human perception. MLLMs exhibit partial competence in evaluations of the Physical Adequacy Score (PAS) (e.g., InternVideo2.5: 44.3%) Despite their general capabilities, MLLMs average only a correlation of 10.0%-30.0% in all dimensions, highlighting a fundamental misalignment with human perception in evaluating motion quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 670, + 408, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 408, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 408, + 684 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Ablation Study of Motion Metrics. According to Table 2, removing any single metric from the evaluation of the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13093" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": "basic variants results in a significant decline in accuracy, which highlights the importance of each dimension's evaluation metric within the overall framework. Notably, the removal of the CAS metric leads to the most pronounced performance drop, with accuracy falling to " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "64.1\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": ", exceeding the impact of other single-dimension ablations. This demonstrates the critical role of the CAS metric in assessing video quality, aligning closely with the key factors prioritized by humans when perceiving video quality. For the performance-oriented variants, we emulate the human perceptual information processing flow by gradually adding evaluation metrics (refer to Fig. 5 in the Appendix), with each additional metric significantly enhancing overall accuracy. This not only demonstrates the effectiveness of incrementally increasing evaluation dimensions but also further verifies the consistency of our evaluation approach with human perceptual mechanisms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 294, + 174, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 174, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 174, + 306 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "text", + "content": "Alignment of PMM with Human Perception. As shown in Fig. 4, the correlation between human scores across these five evaluation dimensions and our evaluation metrics in the same five dimensions remains consistent. For example, both demonstrate a strong correlation between OIS, CAS, and MSS, as well as a weak correlation between PAS and the other metrics. Specifically, as shown in Fig. 4 (a), PAS has a negative correlation with other dimensions, such as the Object Integrity Score (OIS), where " + }, + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\rho = -0.18" + }, + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "text", + "content": ". This might be because high dynamic amplitudes in videos cause distortions and artifacts, which lower structural and temporal coherence scores. Conversely, OIS shows strong positive correlations with MSS (" + }, + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\rho = 0.59" + }, + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "text", + "content": ") and CAS (" + }, + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\rho = 0.50" + }, + { + "bbox": [ + 55, + 315, + 296, + 590 + ], + "type": "text", + "content": "), suggesting that it reflects physical plausibility and motion rationality well. TCS shows low correlation with other dimensions, which can provide a more comprehensive assessment perspective. PAS and structural/temporal metrics exhibit a negative correlation, which challenges conventional optical-flow-based evaluation frameworks. Additionally, the isolation of PAS highlights the importance of separately disentangling motion magnitude in motion video evaluation. Fig. 4 (b) shows the correlations among the proposed evaluation metrics, which align with human perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Assessing Video Generation Models with PMM. As shown in Table 3, we evaluate several leading video generation models using our PMM metrics, including Mochi 1[70], OpenSora [96], CogVideoX [90], OpenSora-Plan [45], HunyuanVideo [37], and Wan2.1 [72]. Our findings indicate that Wan2.1 delivers the best performance in generating motion videos, producing results that appear more realistic. For a clearer visual comparison of the dynamic videos generated by each model, please refer to Fig. 16 in the Appendix." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 331, + 69, + 436, + 171 + ], + "blocks": [ + { + "bbox": [ + 331, + 69, + 436, + 171 + ], + "lines": [ + { + "bbox": [ + 331, + 69, + 436, + 171 + ], + "spans": [ + { + "bbox": [ + 331, + 69, + 436, + 171 + ], + "type": "image", + "image_path": "975c5962c443a53c7c3a37e190d605d199e929e99b31e7066a2e2ba34042ad77.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 226 + ], + "type": "text", + "content": "Figure 4. Correlation Matrix Analysis of Metrics Within Different Evaluation Mechanisms. (a): Spearman Correlation Matrices for human annotations; (b): Spearman Correlation Matrices for our PMM metrics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 440, + 70, + 545, + 171 + ], + "blocks": [ + { + "bbox": [ + 440, + 70, + 545, + 171 + ], + "lines": [ + { + "bbox": [ + 440, + 70, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 440, + 70, + 545, + 171 + ], + "type": "image", + "image_path": "7f7435a549138d51437d9fbeafb81157ecb59b2cf65f916d5f05647b8f3d54f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 240, + 555, + 328 + ], + "blocks": [ + { + "bbox": [ + 315, + 240, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 315, + 240, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 315, + 240, + 555, + 328 + ], + "type": "table", + "html": "
ModelsAvg.(%) ↑CASMSSOISPASTCS
Mochi 1[70]53.237.762.068.614.483.6
OpenSora [96]51.631.261.973.03.488.5
CogVideoX [90]60.650.661.675.424.691.0
OpenSora-Plan [45]58.939.376.078.66.094.7
HunyuanVideo [37]63.451.981.665.826.196.3
Wan2.1 [72]78.462.884.266.017.997.8
", + "image_path": "65dc4a204c9616f89feb24349203a05635e6e72b720562ff35091703e990f9c1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": "Table 3. Performance of Video Generation Models on VMBench. We evaluated six open-source video generate models [37, 45, 70, 72, 90, 96] using VMBench. A higher score indicates better performance for a category." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 404, + 388, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 404, + 388, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 404, + 388, + 417 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "Video motion authenticity remains a critical challenge in the creation of generate content. In response, we introduce VMBench, the first open-source benchmark for motion quality evaluation, integrating motion metrics with human-aligned assessment to reveal deficiencies in existing models' ability to generate physically plausible movements. To support the research community, we offer three essential resources: 1) a standardized framework for detecting motion artifacts overlooked by traditional metrics; 2) actionable diagnostic tools aimed at guiding model optimization; and 3) quantitative standards that align technical advancements with human perceptual expectations. By establishing a common framework for development in this field, VMBench enables systematic tracking of motion generation, encourages targeted model improvements, and ultimately supports the creation of videos that achieve both visual fidelity and dynamic realism. However, while the benchmark's evaluation metrics are aligned with general human perception, they may not fully capture subtle differences in perception that stem from individual viewer experiences and preferences. Addressing these limitations presents opportunities for further refinement, aiming to ensure a more comprehensive and adaptable approach to video quality evaluation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13094" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Runway gen3. Accessed February 25, 2025 [Online] https://app.runwayml.com/, 2025.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 114, + 295, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 114, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 62, + 114, + 295, + 135 + ], + "type": "text", + "content": "[2] Sora. Accessed February 25, 2025 [Online] https://openai.com/index/sora/, 2025.1,2,4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 137, + 294, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 137, + 294, + 157 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 294, + 157 + ], + "type": "text", + "content": "[3] kling. Accessed February 25, 2025 [Online] https://klingai.com/, 2025.1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 159, + 294, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 180 + ], + "type": "text", + "content": "[4] Pika labs. Accessed February 25, 2025 [Online] https://www.pika.art/, 2025.4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 182, + 294, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 213 + ], + "type": "text", + "content": "[5] Veo 2. Accessed February 25, 2025 [Online] https://deepmind.google/technologies/veo/veo-2/, 2025.2,3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 215, + 294, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 215, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 294, + 268 + ], + "type": "text", + "content": "[6] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 270, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 294, + 313 + ], + "type": "text", + "content": "[7] Jie An, Songyang Zhang, Harry Yang, Sonal Gupta, Jia-Bin Huang, Jiebo Luo, and Xi Yin. Latent-shift: Latent diffusion with temporal shift for efficient text-to-video generation. arXiv preprint arXiv:2304.08477, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "text", + "content": "[8] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 370, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 294, + 413 + ], + "type": "text", + "content": "[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 415, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 415, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 294, + 468 + ], + "type": "text", + "content": "[10] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 470, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 470, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 57, + 470, + 294, + 523 + ], + "type": "text", + "content": "[11] Hritik Bansal, Zongyu Lin, Tianyi Xie, Zeshun Zong, Michal Yarom, Yonatan Bitton, Chenfanfu Jiang, Yizhou Sun, Kai-Wei Chang, and Aditya Grover. Videophy: Evaluating physical commonsense for video generation. arXiv preprint arXiv:2406.03520, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 580 + ], + "type": "text", + "content": "[12] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 581, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 635 + ], + "type": "text", + "content": "[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 637, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 294, + 678 + ], + "type": "text", + "content": "[14] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 2, 4, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[15] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Videocrafter1: Open diffusion models for high-quality video generation, 2023. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "text", + "content": "[16] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models, 2024. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 185 + ], + "type": "text", + "content": "[17] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 262 + ], + "type": "text", + "content": "[18] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 319 + ], + "type": "text", + "content": "[19] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 320, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 553, + 352 + ], + "type": "text", + "content": "[20] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 375 + ], + "type": "text", + "content": "[21] DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 377, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 409 + ], + "type": "text", + "content": "[22] Kangle Deng, Tianyi Fei, Xin Huang, and Yuxin Peng. IRCAN: Introspective recurrent convolutional gan for text-to-video generation. In IJCAI, pages 2216-2222, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 443 + ], + "type": "text", + "content": "[23] Daniel J. Felleman and David C. van Essen. Distributed hierarchical processing in the primate cerebral cortex. *Cerebral Cortex*, 1 1:1-47, 1991. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "text", + "content": "[24] Xiaokun Feng, Haiming Yu, Meiqi Wu, Shiyu Hu, Jintao Chen, Chen Zhu, Jiahong Wu, Xiangxiang Chu, and Kaiqi Huang. Narrlv: Towards a comprehensive narrative-centric evaluation for long video generation models. arXiv preprint arXiv:2507.11245, 2025. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 501, + 553, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 522 + ], + "type": "text", + "content": "[25] Deborah M Gordon. The ecology of collective behavior. PLoS biology, 12(3):e1001805, 2014. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 567 + ], + "type": "text", + "content": "[26] Shir Gur, Sagie Benaim, and Lior Wolf. Hierarchical patch vae-gan: Generating diverse videos from a single sample. Advances in Neural Information Processing Systems, 33: 16761-16772, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 569, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 623 + ], + "type": "text", + "content": "[27] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv preprint arXiv:2406.15252, 2024. 1, 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[28] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 1, 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13095" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[30] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[31] Jonathan Ho, Tim Salimans, Alexey Gritsanko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[32] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[33] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[34] Yuming Jiang, Shuai Yang, Tong Liang Koh, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2performer: Text-driven human video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22747-22757, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[35] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 432 + ], + "type": "text", + "content": "[36] Doyeon Kim, Donggyu Joo, and Junmo Kim. Tivgan: Text to image to video generation with step-by-step evolutionary generator. IEEE Access, 8:153113-153122, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[37] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 543 + ], + "type": "text", + "content": "[38] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "type": "text", + "content": "[39] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. Advances in Neural Information Processing Systems, 33:1083-1093, 2020. 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[40] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[41] Yitong Li, Martin Min, Dinghan Shen, David Carlson, and Lawrence Carin. Video generation from text. In Proceedings of the AAAI conference on artificial intelligence, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "text", + "content": "[44] Mingxiang Liao, Qixiang Ye, Wangmeng Zuo, Fang Wan, Tianyu Wang, Yuzhong Zhao, Jingdong Wang, Xinyu Zhang, et al. Evaluation of text-to-video generation models: A dynamics perspective. Advances in Neural Information Processing Systems, 37:109790-109816, 2025. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[45] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[46] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[47] Jie Liu, Gongye Liu, Jiajun Liang, Ziyang Yuan, Xiaokun Liu, Mingwu Zheng, Xiele Wu, Qiulin Wang, Wenyu Qin, Menghan Xia, et al. Improving video generation with human feedback. arXiv preprint arXiv:2501.13918, 2025. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 410 + ], + "type": "text", + "content": "[48] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "text", + "content": "[49] Yuanxin Liu, Lei Li, Shuhuai Ren, Rundong Gao, Shicheng Li, Sishuo Chen, Xu Sun, and Lu Hou. Fetv: A benchmark for fine-grained evaluation of open-domain text-to-video generation. Advances in Neural Information Processing Systems, 36:62352-62387, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 555, + 533 + ], + "type": "text", + "content": "[50] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 588 + ], + "type": "text", + "content": "[51] Guoqing Ma, Haoyang Huang, Kun Yan, Liangyu Chen, Nan Duan, Shengming Yin, Changyi Wan, Ranchen Ming, Xiaoniu Song, Xing Chen, et al. Step-video-t2v technical report: The practice, challenges, and future of video foundation model. arXiv preprint arXiv:2502.10248, 2025.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 643 + ], + "type": "text", + "content": "[52] Rayeesa Mehmood, Rumaan Bashir, and Kaiser J Giri. Vtmgan: video-text matcher based generative adversarial network for generating videos from textual description. International Journal of Information Technology, 16(1):221-236, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[53] Tomoko Nemoto and David Beglar. Likert-scale questionnaires. In JALT 2013 conference proceedings, pages 1-6, 2014. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[54] Xun Long Ng, Kian Eng Ong, Qichen Zheng, Yun Ni, Si Yong Yeo, and Jun Liu. Animal kingdom: A large and diverse dataset for animal behavior understanding. In Pro" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13096" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19023-19034, 2022. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "type": "text", + "content": "[55] Haomiao Ni, Changhao Shi, Kai Li, Sharon X Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18444-18455, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 186 + ], + "type": "text", + "content": "[56] TA Palomaki, JD Teufel, RW Simmonds, and Konrad W Lehnert. Entangling mechanical motion with microwave fields. Science, 342(6159):710-713, 2013. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 220 + ], + "type": "text", + "content": "[57] Marina A Pavlova. Biological motion processing as a hallmark of social cognition. Cerebral cortex, 22(5):981-995, 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 224, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 224, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 294, + 278 + ], + "type": "text", + "content": "[58] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, et al. Movie gen: A cast of media foundation models, 2025. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 346 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 348, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 380 + ], + "type": "text", + "content": "[60] Ralph H. Rasshofer, Martin Spies, and Hans Spies. Influences of weather phenomena on automotive laser radar systems. Advances in radio science, 9:49-60, 2011. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[61] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 440, + 294, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 440, + 294, + 494 + ], + "spans": [ + { + "bbox": [ + 56, + 440, + 294, + 494 + ], + "type": "text", + "content": "[62] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 497, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 497, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 497, + 294, + 540 + ], + "type": "text", + "content": "[63] Zhongwei Ren, Yunchao Wei, Xun Guo, Yao Zhao, Bingyi Kang, Jiashi Feng, and Xiaojie Jin. Videoworld: Exploring knowledge learning from unlabeled videos. arXiv preprint arXiv:2501.09781, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 543, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 586 + ], + "type": "text", + "content": "[64] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "text", + "content": "[65] Benjamin Scellier and Yoshua Bengio. Equilibrium propagation: Bridging the gap between energy-based models and backpropagation. Frontiers in computational neuroscience, 11:24, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 688 + ], + "type": "text", + "content": "[66] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[67] Jos Stam. Stable fluids. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 779-786. 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[68] Kaiyue Sun, Kaiyi Huang, Xian Liu, Yue Wu, Zihan Xu, Zhenguo Li, and Xihui Liu. T2v-compbench: A comprehensive benchmark for compositional text-to-video generation. arXiv preprint arXiv:2407.14505, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[69] Rui Sun, Yumin Zhang, Tejal Shah, Jiahao Sun, Shuoying Zhang, Wenqi Li, Haoran Duan, Bo Wei, and Rajiv Ranjan. From sora what we can see: A survey of text-to-video generation. arXiv preprint arXiv:2405.10674, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 184 + ], + "type": "text", + "content": "[70] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024. 1, 2, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 206 + ], + "type": "text", + "content": "[71] Qwen Team. Qwen2.5: A party of foundation models, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "type": "text", + "content": "[72] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "text", + "content": "[73] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "text", + "content": "[74] I Telecom. Recommendation 500-10: Methodology for the subjective assessment of the quality of television pictures. ITU-R Rec. BT, 500, 2000. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "text", + "content": "[75] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 367, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 553, + 399 + ], + "type": "text", + "content": "[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 400, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 455 + ], + "type": "text", + "content": "[77] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14549-14560, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[78] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. arXiv preprint arXiv:2307.00040, 2(3):4, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[79] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. 2024. 1, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "text", + "content": "[80] Weimin Wang, Jiawei Liu, Zhijie Lin, Jiangqiao Yan, Shuo Chen, Chetwin Low, Tuyen Hoang, Jie Wu, Jun Hao Liew, Hanshu Yan, Daquan Zhou, and Jiashi Feng. Magicvideo-v2: Multi-stage high-aesthetic video generation, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 591, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 644 + ], + "type": "text", + "content": "[81] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[82] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[83] Chenfei Wu, Lun Huang, Qianxi Zhang, Binyang Li, Lei Ji, Fan Yang, Guillermo Sapiro, and Nan Duan. Godiva: Gen" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13097" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "erating open-domain videos from natural descriptions. arXiv preprint arXiv:2104.14806, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 162 + ], + "type": "text", + "content": "[84] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Exploring video quality assessment on user generated contents from aesthetic and technical perspectives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20144-20154, 2023. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 228 + ], + "type": "text", + "content": "[85] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In International Conference on Machine Learning, pages 54015-54029. PMLR, 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 272 + ], + "type": "text", + "content": "[86] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[87] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[88] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 404 + ], + "type": "text", + "content": "[89] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5288-5296, 2016. 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "text", + "content": "[90] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "text", + "content": "[91] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 570 + ], + "type": "text", + "content": "[92] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 295, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 592 + ], + "type": "text", + "content": "[93] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 625 + ], + "type": "text", + "content": "[94] Ailing Zeng, Yuhang Yang, Weidong Chen, and Wei Liu. The dawn of video generation: Preliminary explorations with sora-like models. arXiv preprint arXiv:2410.05227, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[95] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 714 + ], + "type": "text", + "content": "[96] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 162 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 118 + ], + "type": "text", + "content": "[97] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 1, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[98] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13098" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_content_list.json b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5f40a17fc5598c4621788dbc3f7d0b5d6f90f522 --- /dev/null +++ b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_content_list.json @@ -0,0 +1,1425 @@ +[ + { + "type": "text", + "text": "VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory", + "text_level": 1, + "bbox": [ + 117, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runjia Li Philip Torr Andrea Vedaldi Tomas Jakab", + "bbox": [ + 274, + 203, + 723, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Oxford", + "bbox": [ + 415, + 227, + 583, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "v-mem.github.io", + "bbox": [ + 429, + 250, + 566, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/89c810002b4c80233657cf00ae17e94579b7a2e98175a33a5b6702ab0dc5df6a.jpg", + "image_caption": [ + "Figure 1. VMem enables autoregressive scene generation from a single image along user-defined trajectories. The green region shows results with the proposed memory module, maintaining coherence when generating previously seen parts of the scene. The red region, without memory, exhibits degradation highlighted with red ellipses, demonstrating VMem is effective for consistent scene generation." + ], + "image_footnote": [], + "bbox": [ + 91, + 284, + 416, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8af08eb47fd8fc071039d3480bd09ade2f068998d8cada51a12627fbb3ac7059.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 284, + 906, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 539, + 326, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel memory module for building video generators capable of interactively exploring environments. Previous approaches have achieved similar results either by out-painting 2D views of a scene while incrementally reconstructing its 3D geometry—which quickly accumulates errors—or by using video generators with a short context window, which struggle to maintain scene coherence over the long term. To address these limitations, we introduce Surfel-Indexed View Memory (VMem), a memory module that remembers past views by indexing them geometrically based on the 3D surface elements (surfels) they have observed. VMem enables efficient retrieval of the most relevant past views when generating new ones. By focusing only on these relevant views, our method produces consistent explorations of imagined environments at a fraction of the computational cost required to use all past views as context. We evaluate our approach on challenging long-term scene synthesis benchmarks and demonstrate superior performance compared to existing methods in maintaining scene coherence and camera control.", + "bbox": [ + 88, + 583, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 539, + 643, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider the problem of generating long videos that explore an imagined space following a camera path specified interactively by the user. In this paradigm, the user tells the model which camera path to follow for the next few frames, observes the generated content, and then decides where to explore next based on what they have seen. For example, exploring a house may involve visiting the kitchen, the living room, and the bathroom, eventually returning to the kitchen. Throughout the video, the scene must remain consistent, ensuring that the kitchen looks the same upon return. Generating such videos is essential for immersive applications such as games, where players can navigate generated worlds. However, even recent large-scale interactive video generators such as Google's Genie 2 [20] struggle to achieve this goal.", + "bbox": [ + 511, + 566, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This problem has so far been addressed by two types of methods. First, outpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 43] iterate between generating new 2D views of the scene and estimating its 3D geometry. They use the estimated geometry to partially render a new viewpoint and then employ an outpainting model to complete the missing parts, thus adding one more image to the col", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25690", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lection. However, errors in outpainting, 3D reconstruction, and stitching accumulate over time, leading to severe degradation of the generated content after a short while.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, multi-view/video-based methods [20, 22, 25, 35, 42] condition novel view generation on previous views using a geometry-free approach that does not explicitly estimate the scene geometry. While this avoids the accumulation of errors in the reconstructed 3D scene geometry, it comes at a high computational cost, limiting the number of conditioning views to a small context window of recent frames. This constraint hurts the long-term consistency of the generated images.", + "bbox": [ + 89, + 138, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we revisit the second class of methods and propose conditioning not on the most recently generated views, but on the most relevant ones for generating the next image, thereby maintaining a high degree of consistency within a limited computational budget. Given a novel viewpoint of the scene, the most relevant past views are those that have already observed the parts of the scene currently being generated. This implies that, for each part of the scene, we must remember which views have previously observed it and retrieve them from memory.", + "bbox": [ + 89, + 276, + 480, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we introduce Surfel-Indexed View Memory, abbreviated as VMem, a memory module that anchors previous views to the 3D surface elements they observe. Given a new viewpoint, we retrieve the past views that best capture the currently observed surfaces and use them to condition novel view generation. To create the memory, we estimate the geometry of each new view using an off-the-shelf point map predictor. This is similar to outpainting methods but, crucially, we do not use the estimated geometry as the final representation of the scene; instead, we use it to construct VMem, which is a memory of past views. We represent the scene geometry using surfels, which are more robust compared to meshes and can represent occlusions compared to point clouds. Each surfel stores in its attributes a set of indices corresponding to past viewpoints that observed it.", + "bbox": [ + 89, + 429, + 482, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To retrieve relevant past viewpoints, we render the surfels with their attributes from the novel viewpoint and splat them onto an image grid. Each pixel in the resulting image then corresponds to a set of viewpoint indices. We select the top $K$ most frequently represented viewpoint indices and use them to retrieve past views from the database, where each view is represented by an RGB image and its corresponding camera parameters. A key advantage of our approach over outpainting-based methods is that it does not require highly accurate scene geometry. As long as we successfully retrieve the most relevant past views, our method remains robust.", + "bbox": [ + 89, + 672, + 482, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By leveraging Surfel-Indexed View Memory, we significantly reduce the memory and computational burden of conditioning on a large number of previous views, as re", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quired by multi-view/video-based methods, while improving long-term consistency across generated novel views.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach represents a step toward scalable, realistic, and long-term autoregressive scene generation, making the following contributions:", + "bbox": [ + 511, + 123, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce Surfel-Indexed View Memory, a plug-and-play module to index past views geometrically and use them to condition novel view generation.", + "2. We show that our method can generate long-term coherent videos of scenes and outperforms existing approaches.", + "3. We demonstrate that Surfel-Indexed View Memory achieves comparable performance with $4 \\times$ fewer context views, delivering a $12 \\times$ speedup.", + "4. We validate our approach on challenging benchmarks, outperforming the current open-source state-of-the-art methods." + ], + "bbox": [ + 511, + 169, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 368, + 650, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) is a challenging and ill-posed problem that can be categorized into two main categories: view interpolation methods [3-6, 8, 19, 30, 37], which generate views close to given inputs, and view extrapolation or autoregressive view generation, where novel views extend significantly beyond the original scene, introducing substantial new content. The latter is particularly difficult from single images. Between these lie single-view scene reconstruction methods, which succeed mainly for single-object scenes [1, 12, 17, 27, 29, 38] or highly bounded scenes [28]. However, their extrapolation capabilities remain limited. Most relevant to our work are single-image view extrapolation models, which fall into two categories: those incorporating explicit geometric modeling with inpainting, and those based directly on image or video generation.", + "bbox": [ + 511, + 393, + 906, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inpainting-based view extrapolation. Inpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 40, 41, 43] use pre-trained 3D reconstruction models to generate 3D representations—such as meshes, point clouds, or Gaussian splats—from images. These representations are reprojected into novel views, where 2D inpainting fills missing regions. SceneScape [7] reconstructs meshes from images using pretrained depth estimators. Diffusion-based inpainting then completes the projected novel view, which is reprojected back to refine and extend the mesh. This process is iterated to generate novel views. MultiDiff [18] uses depth estimators to warp reference images into multiple novel views, training diffusion models to inpaint missing regions simultaneously. CamCtrl3D [21] follows a similar approach but adds ray maps to condition diffusion models for view synthesis. ViewCrafter [43] uses pre-trained pointmap estimators to create point clouds and trains video diffusion mod", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25691", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50d1e06551806acb593625cda09e71f750c5a9fdc26e29e0e9b745fe1c2886af.jpg", + "image_caption": [ + "Figure 2. Method. Given target camera viewpoints $\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}$ , we query our Surfel-Indexed View Memory to retrieve the most relevant $K$ past views $\\mathcal{V}^* \\subset \\mathcal{V}^{(s)}$ where $\\mathcal{V}^* = \\{v_t\\}_{t = 1}^K$ as references. Retrieved reference images $\\mathbf{x}_t$ along with Plücker embeddings of both reference camera poses $\\mathbf{c}_t$ and target camera poses $\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}$ are fed into generator $\\psi$ to synthesize novel views $\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}$ . After generation, the surfel-indexed memory is updated $S^{(s)} \\to S^{(s + 1)}$ by appending new view indices $\\{T + m\\}_{m = 1}^{M}$ to existing surfels or creating new surfels based on geometry of the generated views. This is repeated autoregressively, enabling long-term consistent generation." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 908, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "els for inpainting. GenWarp [26] warps 2D coordinate representations of reference views into novel views and uses them to condition generation.", + "bbox": [ + 88, + 376, + 480, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these methods improve geometric consistency, they are susceptible to errors in depth or pointmap estimation, which propagate distortions across generated views. Once 3D representations are constructed, inaccuracies from depth or pointmap estimation become difficult to correct. Moreover, scaling to large scenes is computationally intensive, as storing and processing high-fidelity 3D representations requires substantial memory.", + "bbox": [ + 88, + 422, + 482, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view-based view extrapolation. Another category [8, 22, 25, 31, 35, 42, 46] avoids explicitly modeling 3D scene geometry. Instead, these approaches generate novel views by conditioning on previously rendered or generated views. GeoGPT [25] uses autoregressive likelihood models to synthesize novel views from single images. Building on this, LookOut [22] improves consistency by generating view sequences along predefined trajectories, conditioning on up to two preceding views. Other methods [31, 42] enhance novel view synthesis by employing cross-view attention or enforcing epipolar constraints within diffusion models. MotionCtrl [35] leverages geometric priors from pre-trained video diffusion models to synthesize camera-controllable videos. These methods are trajectory-based, requiring input views to follow continuous spatial paths. In contrast, image-set models like CAT3D [8] and more recently SEVA [46] synthesize novel views from sparse and unordered input images, without trajectory constraints. However, these methods are computationally expensive, with $O(n^{2})$ complexity for attention, limiting conditioning views to small context windows containing only recently generated frames. This limitation impacts long-term consistency of generated images.", + "bbox": [ + 91, + 553, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several concurrent works explore memory-based approaches for scene generation. Genie [2, 20] maintains memory through recurrent state features. Gen3C [23] conditions on stored point clouds like ViewCrafter [43], inheriting inpainting-based limitations. StarGen [44] and WorldMem [39] use distance- and field-of-view-based spatial memory retrieval, which can recover spatially correlated views but lack geometric reasoning for occlusions. In contrast, our surfel-based memory explicitly models occlusions and provides more principled view selection.", + "bbox": [ + 511, + 376, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 542, + 604, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}_1 \\in \\mathbb{R}^{H \\times W \\times 3}$ be an RGB image of a scene that we wish to explore, and let $\\{\\mathbf{c}_t\\}_{t=1}^T$ be a sequence of camera parameters specifying a path through the scene. Our goal is to generate a corresponding video, i.e., a sequence of images $\\{\\mathbf{x}_t\\}_{t=1}^T$ , where $\\mathbf{x}_1$ is the given input image, and each subsequent frame $\\mathbf{x}_t$ is consistent with the previous ones and reflects the specified cameras $\\{\\mathbf{c}_t\\}_{t=1}^T$ . Moreover, we aim to do so autoregressively, generating $M$ novel views at a time for the given cameras (with $M$ kept small to ensure interactivity) while ensuring consistency with previously generated views. This enables interactive scene exploration, where the user can specify the next camera positions at each step.", + "bbox": [ + 511, + 566, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we work with a sequence of views, where each view $v_{t} = (\\mathbf{x}_{t},\\mathbf{c}_{t})$ consists of an image $\\mathbf{x}_t$ and its corresponding camera $\\mathbf{c}_t$ at timestep $t$ . At generation step $s(s\\geq 0)$ , we have generated $T = sM$ frames so far. The goal is to generate $M$ new RGB images $\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}$ for the next target camera parameters $\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}$ , given the previously generated views $\\mathcal{V}^{(s)} = \\{v_1,v_2,\\dots ,v_T\\}$ . After generation, we update $T\\gets T + M$ .", + "bbox": [ + 511, + 765, + 908, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A challenge with this approach is that the context $\\mathcal{V}^{(s)}$", + "bbox": [ + 531, + 883, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25692", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4647dc6736bfed6b6ec55a4c3bbe3d87ceb72cc064bece0ade8f8d9b8d1a9877.jpg", + "image_caption": [ + "Figure 3. Surfel-based memory index. Each surfel stores indices of views that observed it. We color-code each surfel by contributing view indices. This spatial index enables retrieval of relevant past views: when generating a novel view, we identify visible surfels from the target viewpoint and retrieve views that previously observed those same regions, naturally accounting for occlusion." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 236, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/948869d140c2846ab916e8dff078913fffb40e97494b87e42bf2d9155faffcd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 93, + 403, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a8e23f819452a8bfd6d5f05fa64a02b27714cb1cfcf9f5db453b9f523e4ac9b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 88, + 566, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c452c93a1eeba2d661c4170b7f035f252adffd53184905967588730517409b88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 98, + 722, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58e42b42eb0cab221ac9d4834cec7e1c75f5a75491c2648c159b3df934868e06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 98, + 875, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "grows unbounded over time. Video generators addressing this problem [22, 25, 35, 42] typically consider a fixed-length subset of $\\mathcal{V}^{(s)}$ , conditioning their generation only on the most recent $L$ views $\\{v_{T - L + 1},\\ldots ,v_T\\}$ . This limitation results in severe inconsistencies when the generated sequence extends beyond the (small) context window.", + "bbox": [ + 88, + 279, + 480, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We solve this problem by dynamically retrieving the most relevant subset of past views, denoted as $\\mathcal{V}^* \\subseteq \\mathcal{V}^{(s)}$ . To achieve this, we introduce an efficient data structure, the Surfel-Indexed View Memory, that stores and retrieves past views based on their approximate 3D geometry. We first describe the Surfel-Indexed View Memory (Sec. 3.1) and then introduce the novel view generator (Sec. 3.2). An overview of the method is provided in Fig. 2.", + "bbox": [ + 88, + 369, + 482, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Surfel-indexed view memory", + "text_level": 1, + "bbox": [ + 89, + 501, + 346, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider the problem of generating the next views $\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}$ of the scene as seen from the cameras $\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}$ . The generator must consider the previously generated views and ensure $\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}$ are consistent with those, while generating new content as needed. However, not all past views are equally relevant to the novel views $\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}$ . We prioritize past views that have likely observed the largest portion of the scene currently being generated. For example, if the current locale is surrounded by walls, parts of the scene behind those walls are likely less relevant for generating a new view of that locale. Conversely, views that share a similar visible region with the target view can provide more useful information for generating its content.", + "bbox": [ + 89, + 523, + 482, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This principle drives our design for a view memory module. We maintain a coarse model of the scene geometry using surfels (visualized in Fig. 3), simple surface primitives that account for occlusion. Each surfel stores the indices of the past views that observed it. To generate a new view, the surfels visible from the new viewpoint are retrieved, using the associated indices to vote for which views to consider for generation. This process is illustrated in Fig. 4.", + "bbox": [ + 89, + 736, + 482, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we define a surfel as the tuple", + "bbox": [ + 109, + 857, + 393, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {k} = \\left(\\mathbf {p} _ {k}, \\mathbf {n} _ {k}, r _ {k}, \\mathcal {I} _ {k}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 885, + 367, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_k \\in \\mathbb{R}^3$ denotes the surfel's 3D position, $\\mathbf{n}_k \\in \\mathbb{R}^3$ is its surface normal, $r_k \\in \\mathbb{R}$ is the surfel radius, and $\\mathcal{I}_k \\subseteq \\{1,2,\\dots,T\\}$ is a subset of past view indices that observed the surfel. The surfel-based memory indexing of past views $\\mathcal{V}^{(s)}$ , encapsulating the scene, is represented as a set of surfels $\\mathcal{S}^{(s)} = \\{\\mathbf{s}_k\\}_{k=1}^{N^{(s)}}$ , where $N^{(s)}$ is the number of surfels at generation step $s$ . We also maintain an octree to quickly retrieve surfels based on their geometry. Next, we describe how to read from and write to this memory.", + "bbox": [ + 511, + 279, + 906, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reading from the memory. To retrieve the most relevant past views from the memory $\\mathcal{V}^{(s)}$ for the novel cameras $\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}$ , we first compute their average pose $\\bar{\\mathbf{c}}_s\\in$ SE(3) as a reference (refer to the supplementary material). Then we render the surfels $S^{(s)}$ from the averaged camera pose $\\bar{\\mathbf{c}}_s$ . Each surfel is rendered as a splat with its attributes, accounting for relative depth and occlusions, and contributing to the image based on its coverage. The rendered attributes correspond to the indices of the past views. The core intuition is that views observing the largest portion of the scene from the perspective of the novel camera are most relevant for novel view synthesis. To identify these views, we rank past view indices based on their frequency across all rendered pixels.", + "bbox": [ + 511, + 424, + 906, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then select the top- $K$ most frequently represented indices $\\mathcal{I}^*$ and use them to retrieve the corresponding past views, $\\mathcal{V}^* = \\{v_t\\}_{t\\in \\mathcal{I}^*}$ , from the memory $\\mathcal{V}^{(s)}$ . This subset of views serves as the context for the new view generator, as described in Sec. 3.2.", + "bbox": [ + 511, + 636, + 905, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To avoid oversampling repeatedly visited regions, we apply a non-maximum suppression algorithm that reduces redundancy in memory and promotes broader scene coverage among the top- $K$ views. During retrieval, we retain only the most frequently referenced view among those with similar poses. During writing to the memory, we merge surfels by comparing their associated camera poses—if two poses are highly similar, we discard the older one.", + "bbox": [ + 511, + 712, + 905, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Writing to the memory. After generating the new views $\\{v_{T + m} = (\\mathbf{x}_{T + m},\\mathbf{c}_{T + m})\\}_{m = 1}^{M}$ , we add them to the memory $\\mathcal{V}^{(s)}$ and update the surfel-based memory index $S^{(s)}$ . We use an off-the-shelf point map estimator $\\phi$ , such as", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25693", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b99e1556843ed651d81c978e53ceb2f592b8b2f0db98456ca84ceda16c045999.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 89, + 893, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c670ca2c89097377121eef7670264956dee86f3c8395fb321363f79591f235f5.jpg", + "image_caption": [ + "Figure 4. Surfel-Indexed View Memory. Reading procedure renders surfels $S^{(s)}$ with their attributes, containing past view indices as frame indices. We then select the $K$ most frequent frame indices in the rendered image to retrieve relevant past views from $\\mathcal{V}^{(s)}$ . Writing procedure estimates geometry of newly generated views $\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}$ as surfels and merges them with existing surfels. Frame indices $\\{T + m\\}_{m = 1}^{M}$ are appended to surfels in these views, and novel views are stored, updating $\\mathcal{V}^{(s)} \\to \\mathcal{V}^{(s + 1)}$ and $S^{(s)} \\to S^{(s + 1)}$ ." + ], + "image_footnote": [], + "bbox": [ + 101, + 188, + 893, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CUT3R [32], to estimate their point maps $\\{\\mathbf{P}_{T + m}^{*}\\}_{m = 1}^{M}$ . Specifically, we jointly estimate the point maps for the newly generated views along with the retrieved past views $\\nu^{*}$ . This joint estimation ensures that the new point maps are aligned with the coordinate frame of the existing scene geometry in the surfel-based memory index $S^{(s)}$ .", + "bbox": [ + 89, + 381, + 480, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we convert the estimated point maps to a set of new surfels. For each newly generated frame $t \\in \\{T + 1, T + 2, \\dots, T + M\\}$ , we process its point map $\\mathbf{P}_t^*$ . Since our approach only requires coarse geometry, we first downsample the point map $\\mathbf{P}_t^*$ by a factor $\\sigma$ , yielding a smaller point map $\\mathbf{P}_t \\in \\mathbb{R}^{H' \\times W' \\times 3}$ , where $H' = H / \\sigma$ and $W' = W / \\sigma$ . For each pixel location $(u, v)$ in the downsampled point map, we have a 3D point $\\mathbf{p}_{u,v,t}$ , and we create a new surfel $\\mathbf{s}_{k'}$ centered on it. We compute each surfel's normal using the cross-product of displacement vectors to neighboring pixels. For a pixel at 2D location $(u, v)$ in the image grid, we use neighboring pixels to estimate the local surface normal:", + "bbox": [ + 89, + 473, + 482, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} _ {k ^ {\\prime}} = \\frac {(\\mathbf {p} _ {u + 1 , v , t} - \\mathbf {p} _ {u - 1 , v , t}) \\times (\\mathbf {p} _ {u , v + 1 , t} - \\mathbf {p} _ {u , v - 1 , t})}{\\| (\\mathbf {p} _ {u + 1 , v , t} - \\mathbf {p} _ {u - 1 , v , t}) \\times (\\mathbf {p} _ {u , v + 1 , t} - \\mathbf {p} _ {u , v - 1 , t}) \\|},\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 664, + 477, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{u,v,t}$ represents the 3D point at image coordinates $(u,v)$ in frame $t$ . To ensure that the surfels reasonably cover the scene, we use a heuristic to compute the surfel's radius, which is proportional to the depth of the surfel and inversely proportional to the focal length and the cosine of the angle between the surfel normal and the viewing direction:", + "bbox": [ + 89, + 708, + 482, + 799 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {k ^ {\\prime}} = \\frac {\\frac {1}{2} \\mathbf {D} _ {u , v , t} / f _ {t}}{\\alpha + (1 - \\alpha) \\left\\| \\mathbf {n} _ {k ^ {\\prime}} \\cdot (\\mathbf {p} _ {u , v , t} - \\mathbf {O} _ {t}) \\right\\|},\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 809, + 428, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{D}_{u,v,t}$ represents the depth at pixel $(u,v)$ in frame $t$ , $\\mathbf{O}_t$ is the camera center of frame $t$ , $\\mathbf{p}_{u,v,t} - \\mathbf{O}_t$ is the viewing direction, and $f_t$ is the focal length at time", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$t$ . The factor $\\alpha$ is used to avoid extreme values when $\\| \\mathbf{n}_{k'} \\cdot (\\mathbf{p}_{u,v,t} - \\mathbf{O}_t) \\| \\approx 0$ .", + "bbox": [ + 511, + 382, + 903, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we check if the index $S^{(s)}$ already contains a surfel $\\mathbf{s}_k$ similar to the new surfel $\\mathbf{s}_{k'}$ . Two surfels match if their centers are within a distance $d$ and the cosine similarity between their normals is above a threshold $\\theta$ . If such a surfel $\\mathbf{s}_k$ is found, we add the frame index $t$ to the surfel's set of past view indices: $\\mathcal{I}_k \\gets \\mathcal{I}_k \\cup \\{t\\}$ , and discard $\\mathbf{s}_{k'}$ . If not, we set $\\mathcal{I}_{k'} = \\{t\\}$ and add $\\mathbf{s}_{k'}$ to the index $S^{(s + 1)}$ . This process transforms the surfel memory from $S^{(s)}$ to $S^{(s + 1)}$ with $N^{(s + 1)}$ surfels for the next generation step.", + "bbox": [ + 511, + 412, + 905, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Novel view generator", + "text_level": 1, + "bbox": [ + 511, + 561, + 712, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a camera-conditioned image-set generator $\\psi$ to generate novel views. The generator $\\psi$ takes $K$ retrieved reference views $\\mathcal{V}^* = \\{v_t = (\\mathbf{x}_t, \\mathbf{c}_t)\\}_{t \\in \\mathcal{I}^*}$ and the target camera poses $\\{\\mathbf{c}_{T + m}\\}_{m=1}^M$ for the $M$ new frames to be generated and samples the novel views:", + "bbox": [ + 511, + 585, + 905, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {x} _ {T + m} \\right\\} _ {m = 1} ^ {M} \\sim \\psi \\left(\\left\\{\\left(\\mathbf {x} _ {t}, \\mathbf {c} _ {t}\\right) \\right\\} _ {t \\in \\mathcal {I} ^ {*}}, \\left\\{\\mathbf {c} _ {T + m} \\right\\} _ {m = 1} ^ {M}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 672, + 877, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}^* \\subseteq \\{1, 2, \\ldots, T\\}$ are the frame indices of the retrieved reference views.", + "bbox": [ + 511, + 704, + 903, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we base our generator on the recent SEVA [46] model. Given the plug-and-play nature of our memory module, VMem can work with other image-set generators as well. We also fine-tune a more efficient version of SEVA that operates with a reduced number of reference frames. We provide more details in Sec. 4.1.", + "bbox": [ + 511, + 736, + 905, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 844, + 645, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on established benchmarks for camera-conditioned autoregressive view generation from", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25694", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8b005d7cad06624d8cd8a7fd9a7f0a1a3e1fc4303310aaa30fca70b484831716.jpg", + "image_caption": [ + "Figure 5. Long sequences with revisitations. We compare our VMem against a baseline without memory that relies solely on the last $K$ frames for context. Each sequence: input images (left), then generated images at selected frames. Our method (top rows) maintains consistency when revisiting observed regions, while the baseline (bottom rows) shows severe inconsistencies across extended sequences." + ], + "image_footnote": [], + "bbox": [ + 112, + 87, + 883, + 474 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aa5fd9009d3151efcd0b288893ad50e99875660a6a83c2c7d2080a149440249d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Short-term (50thframe)Long-term (≥200thframe)
LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓
GeoGPT [25]0.44413.35—26.72——0.6749.54—41.87——
Lookout [22]0.37814.43—28.861.2410.8760.65810.51—58.121.1420.924
PhotoNVS [42]0.33315.51—21.76——0.58811.54—41.95——
GenWarp [26]0.43612.030.14429.690.5640.0590.6139.560.08536.400.8500.251
MotionCtrl [35]0.42412.000.14819.250.3410.3560.6059.130.08335.450.7480.609
ViewCrafter [43]0.37716.970.26225.391.5620.2080.5929.740.14834.122.1770.827
SEVA [46] (K=17)0.29318.330.38217.290.2230.1180.45513.980.21623.821.1250.742
VMem (K=4)0.28718.490.40617.120.2190.0390.49313.120.18327.150.8110.499
VMem (K=17)0.29318.330.38217.290.2230.1180.45214.090.22723.560.9820.432
", + "bbox": [ + 114, + 542, + 883, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Single-view NVs on RealEstate10K. Frames are subsampled with 10-frame intervals. $K$ denotes the number of context views. In the short-term setting, all past frames fit in the context windows for both SEVA [46] and VMem; hence, SEVA [46] and VMem ( $K = 17$ ) yield identical results since VMem uses SEVA as the backbone. VMem ( $K = 4$ ) is fine-tuned on RealEstate10K for this setting, yielding better performance than SEVA [46] and VMem ( $K = 17$ ) in short-term scenarios. VMem surpasses baselines on most metrics in this benchmark. However, its true advantage—spatial consistency—is not fully reflected because RealEstate10K trajectories rarely revisit previously observed areas. We address this limitation in Sec. 4.3.", + "bbox": [ + 88, + 715, + 906, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a single image, comparing our models to previous open-source approaches (Sec. 4.2). These benchmarks, however, have a critical limitation: camera trajectories rarely revisit previously observed regions, unlike real-world behavior where humans or robots often return to or look at previ", + "bbox": [ + 89, + 823, + 480, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ously explored areas. To address this limitation, we propose a cycle-trajectory evaluation protocol that extends these benchmarks by making the camera return to the starting position along the same path in reverse order (Sec. 4.3). Additionally, we collect a small dataset with trajectories that", + "bbox": [ + 511, + 823, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25695", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3a7d92503e450a1d4064beb63d870c7c1224affd9f3fc937cf3d7ab2f621a9e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLPIPS ↓PSNR ↑SSIM ↑FID ↓\\( R_{\\text{dist}} \\) ↓\\( T_{\\text{dist}} \\) ↓
Look-out [22]0.8098.410.06938.341.2620.341
GenWarp [26]0.50711.130.13432.940.8480.241
MotionCtrl [35]0.5899.070.09626.860.8710.293
ViewCrafter [43]0.40111.820.21724.720.9020.492
SEVA [46] (K=17)0.40111.820.21724.720.9020.492
VMem (K=4)0.39715.720.29724.970.8210.392
VMem (K=17)0.30418.150.37724.180.8920.165
", + "bbox": [ + 96, + 88, + 477, + 193 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f0163aaeb6c12097b6f51d9a8856c60a98cb058fb3db67a4429153e4ca4d87af.jpg", + "table_caption": [ + "Table 2. Results on cycle trajectories from RealEstate10K." + ], + "table_footnote": [], + "table_body": "
MethodLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
Look-out [22]0.7928.520.0080.7271.499
GenWarp [26]0.52110.270.1290.7851.982
MotionCtrl [35]0.6928.210.0820.8420.129
ViewCrafter [43]0.49413.620.1291.6430.492
VMem (K=4)0.47214.110.1211.2040.387
", + "bbox": [ + 112, + 233, + 460, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "revisit observed regions for qualitative evaluation in Fig. 5. Finally, we conduct an ablation study in Sec. 4.4.", + "bbox": [ + 89, + 368, + 482, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 410, + 277, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. We use pre-trained SEVA [46] as the generator $\\psi$ . SEVA uses a fixed total number of reference and target frames ( $K + M = 21$ ). For our experiments, we use $M = 4$ target views, which leaves $K = 17$ context views. However, this configuration demands substantial computational resources that may be prohibitive for many applications. To address this limitation, we fine-tune a more efficient version with LoRA [13] on the RealEstate10K training split [47] that works with a reduced number of reference frames ( $K = 4$ ) and target views ( $M = 4$ ), and use this efficient version for our main experiments. We demonstrate (Secs. 4.3 and 4.4) that this fine-tuned version, when combined with VMem, achieves comparable performance to the original SEVA while delivering approximately $12\\times$ speedup at inference time. We provide more technical details in the supplementary material.", + "bbox": [ + 88, + 433, + 482, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We consider two real-world datasets for novel view generation evaluation: RealEstate10K [47], comprising indoor scene video clips, and Tanks-and-Temples [14], including indoor and outdoor scenes with larger camera motions. For qualitative evaluation, we use in-the-wild images collected from the internet or captured with phone cameras as input views (see Fig. 5).", + "bbox": [ + 89, + 681, + 482, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate generated novel view quality based on three factors: (1) overall quality of generated novel views, comparing generated novel view distributions to test set distributions using Fréchet Image Distance (FID) [10]; (2) the ability of the model to preserve image details across views, measuring peak signal-to-noise ratio (PSNR) for pixel differences, along with LPIPS [45]", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c2c666a03a8bd06b19d4d35b516e22392328332bbfb70d12ad2e81989bfd2b6.jpg", + "table_caption": [ + "Table 3. Results on cycle trajectories from Tanks and Temples." + ], + "table_footnote": [], + "table_body": "
Context views KView RetrievalLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
17Temporal0.47713.920.1880.9760.254
Camera Distance (SEVA [46])0.39715.720.2970.8210.392
Field of View0.37415.750.2920.9110.382
VMem0.30418.150.3770.8920.165
4Temporal0.7947.520.0181.9420.458
Camera Distance0.42213.270.1871.7870.319
Field of View0.42413.110.1921.7820.285
VMem0.38114.820.2750.7930.124
", + "bbox": [ + 517, + 88, + 903, + 207 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on cycle trajectories from RealEstate10K. We compare different view retrieval strategies for two context view counts $K$", + "bbox": [ + 511, + 218, + 906, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and SSIM [34], as in [22]; and (3) alignment between generated camera poses and ground truth, following [9], where we express estimated camera poses relative to the first frame and normalize translation by the furthest frame. We use DUSt3R [33] to extract poses from generated views. We calculate rotation distance $R_{\\mathrm{dist}}$ by comparing ground truth and extracted rotation matrices of each generated sequence:", + "bbox": [ + 511, + 272, + 906, + 378 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathrm {d i s t}} = \\operatorname {a r c c o s} \\left(0. 5 \\left(\\operatorname {t r} \\left(\\mathbf {R} _ {\\mathrm {g e n}} \\mathbf {R} _ {\\mathrm {g t}} ^ {T}\\right) - 1\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 388, + 844, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathbf{R}$ denotes a rotation matrix. We compute translation distance $t_{\\mathrm{dist}}$ as:", + "bbox": [ + 511, + 419, + 906, + 449 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nt _ {\\text {d i s t}} = \\left\\| \\mathbf {t} _ {\\mathrm {g t}} - \\mathbf {t} _ {\\mathrm {g e n}} \\right\\| _ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 460, + 779, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Short- and long-term view generation", + "text_level": 1, + "bbox": [ + 511, + 500, + 838, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first evaluate VMem on the benchmark from [22, 42], which defines two scenarios: short-term and long-term novel view generation. Beginning from the first frame of each ground-truth test sequence, the model autoregressively generates images along the ground truth camera trajectories with 10-frame intervals. The short-term scenario evaluates the fifth generated image (50 frames from the initial view), while the long-term scenario evaluates the final image ( $\\geq 200$ frames from the initial view). Following the protocol, we only consider samples with more than 200 frames for the long-term scenario and more than 50 frames for the short-term scenario. We report results in Tab. 1. VMem substantially outperforms all baselines on the key metrics, demonstrating superior scene extrapolation. However, the benchmark trajectories rarely revisit observed regions, limiting evaluation of long-term scene consistency.", + "bbox": [ + 511, + 521, + 906, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Long-term view generation with revisitations", + "text_level": 1, + "bbox": [ + 511, + 773, + 895, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate long-term scene consistency when revisiting observed regions—a scenario that rarely occurs in existing benchmarks—we propose a cycle-trajectory evaluation protocol as a proxy for real revisitation scenarios in human and robotic exploration. In this protocol, the model generates frames following trajectories from initial to final poses, then returns along the same path to the starting position.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25696", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62e0bb877dd5e8d393160ad7baf38cac2c1b2e6a45e87220df9fa32a4ae9ac49.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of the cycle trajectory frames ( $\\geq 400$ frames). VMem autoregressively generates new views while maintaining memory of all previously generated frames, ensuring consistency when revisiting previously seen locations." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 495, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/663a1de4bd4cff0bef3258d5340cfd931077a74cff7452a33477436c31db3c70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 89, + 900, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate metrics every 10 frames on the return trajectories, using all test sequences. Following this protocol, we report quantitative results in Tab. 2 and qualitative comparisons in Fig. 6. VMem outperforms all baselines across all metrics in cycle-trajectory evaluation, demonstrating superior visual consistency when revisiting previously observed regions.", + "bbox": [ + 88, + 376, + 482, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate generalization, we evaluate on Tanks-and-Temples containing indoor and outdoor scenes with more dynamic camera trajectories. We use all six advanced scenes and apply the same cycle-trajectory protocol using the first 50 frames, evaluating every frame without temporal subsampling. Results are in Tab. 3. VMem consistently outperforms all baselines across in-domain (Tab. 2) and out-of-domain (Tab. 3) evaluations on the majority of metrics.", + "bbox": [ + 88, + 484, + 482, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also qualitatively evaluate VMem across diverse in-the-wild scenes in Fig. 5, demonstrating generation of diverse, consistent, high-quality long videos on in-the-wild scenarios.", + "bbox": [ + 89, + 606, + 483, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 680, + 240, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate our surfel-based memory indexing for context view retrieval in VMem. We evaluate our approach in the cycle-trajectory setting against three alternative retrieval strategies: (1) temporal-based retrieval, selecting the most recent $K$ views from memory, (2) camera distance-based retrieval, selecting top $K$ views with cameras closest to the target view camera—also used by SEVA [46], our backbone generator, and (3) field-of-view-based retrieval, selecting top $K$ views with highest field-of-view overlap with the target view. We also evaluate performance trade-offs between our lightweight, fine-tuned generator using $K = 4$ context views versus original SEVA with $K = 17$ context views. We report results in Tab. 4.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our findings show VMem consistently improves generator performance across both numbers of context views. Improvement is most pronounced with fewer context views $(K = 4)$ , highlighting the effectiveness of our surfel-based memory index.", + "bbox": [ + 511, + 376, + 905, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational efficiency. A key practical advantage of our approach is its computational efficiency. VMem with only $K = 4$ views achieves $12 \\times$ speed improvement over the original SEVA while recovering most of the performance achieved with $K = 17$ views. On an RTX 4090 GPU, our method generates frames in 4.2 seconds compared to 50 seconds for SEVA, representing a step towards real-time interactive scene exploration.", + "bbox": [ + 511, + 460, + 908, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced Surfel-Indexed View Memory (VMem), a novel plug-and-play memory module for long-term autoregressive scene generation using a surfel-indexed view memory. By anchoring past views to a surfel-based representation of the scene and retrieving the most relevant ones, our approach improves long-term scene consistency while reducing computational costs. Experiments on long-term scene synthesis benchmarks demonstrate that VMem outperforms existing methods in scene coherence while enabling the use of fewer context views, leading to significantly faster generation. Our work advances scalable video generation, with applications in virtual reality, gaming, and other domains.", + "bbox": [ + 511, + 627, + 906, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. The authors of this work are supported by Clarendon scholarship, ERC 101001212-UNION, and AIMS EP/S024050/1. The authors would like to thank Xingyi Yang, Zeren Jiang, Junlin Han, Zhongrui Gui for their insightful feedback.", + "bbox": [ + 511, + 830, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25697", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 121, + 187, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. SF3D: Stable fast 3D mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 2", + "[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 3", + "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2", + "[4] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction, 2023. arXiv:2304.06714 [cs].", + "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024.", + "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2", + "[7] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2", + "[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. Advances in Neural Information Processing Systems, 2024. 2, 3", + "[9] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 7", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[11] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7909-7920, 2023. 1, 2", + "[12] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2" + ], + "bbox": [ + 93, + 146, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 7", + "[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7", + "[15] Jing Yu Koh, Harsh Agrawal, Dhruv Batra, Richard Tucker, Austin Waters, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Simple and effective synthesis of indoor 3d scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1169-1178, 2023. 1, 2", + "[16] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 1, 2", + "[17] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. RealFusion: 360 reconstruction of any object from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[18] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 1, 2", + "[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2", + "[20] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Rocktäschel. Genie 2: A large-scale foundation world model, 2024. 1, 2, 3", + "[21] Stefan Popov, Amit Raj, Michael Krainin, Yanzhen Li, William T. Freeman, and Michael Rubinstein. Camctrl3d: Single-image scene exploration with precise 3d camera control, 2025. 1, 2", + "[22] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 2, 3, 4, 6, 7", + "[23] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25698", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "world-consistent video generation with precise camera control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3", + "[24] Chris Rockwell, David F Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14104-14113, 2021. 1, 2", + "[25] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14356-14366, 2021. 2, 3, 4, 6", + "[26] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 1, 2, 3, 6, 7", + "[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv.cs, abs/2310.15110, 2023. 2", + "[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arxiv, 2024. 2", + "[29] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[30] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2", + "[31] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 3", + "[32] Qianqian Wang*, Yifei Zhang*, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In CVPR, 2025. 5", + "[33] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 7", + "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[35] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. ACM SIGGRAPH 2024 Conference Papers, 2023. 2, 3, 4, 6, 7" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 1, 2", + "[37] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21551-21561, 2024. 2", + "[38] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2", + "[39] Zeqi Xiao, Yushi Lan, Yifan Zhou, Wenqi Ouyang, Shuai Yang, Yanhong Zeng, and Xingang Pan. Worldmem: Long-term consistent world simulation with memory, 2025. 3", + "[40] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T. Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, and Charles Herrmann. Wonderjourney: Going from anywhere to everywhere. arXiv.cs, abs/2312.03884, 2023. 2", + "[41] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T. Freeman, and Jiajun Wu. WonderWorld: interactive 3D scene generation from a single image. arXiv, 2406.09394, 2025. 2", + "[42] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7094-7104, 2023. 2, 3, 4, 6, 7", + "[43] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 6, 7", + "[44] Shangjin Zhai, Zhichao Ye, Jialin Liu, Weijian Xie, Jiaqi Hu, Zhen Peng, Hua Xue, Danpeng Chen, Xiaomeng Wang, Lei Yang, Nan Wang, Haomin Liu, and Guofeng Zhang. Stargen: A spatiotemporal autoregression framework with video diffusion model for scalable and controllable scene generation, 2025. 3", + "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[46] Jensen (Jinghao) Zhou, Hang Gao, Vikram Voleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rupprecht, and Varun Jampani. Stable virtual camera: Generative view synthesis with diffusion models. arXiv preprint arXiv:2503.14489, 2025. 3, 5, 6, 7, 8", + "[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 7" + ], + "bbox": [ + 516, + 92, + 906, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25699", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_model.json b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5eff802fd2777d78adb321017591081d03ce8915 --- /dev/null +++ b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.88, + 0.178 + ], + "angle": 0, + "content": "VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.204, + 0.724, + 0.223 + ], + "angle": 0, + "content": "Runjia Li Philip Torr Andrea Vedaldi Tomas Jakab" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.228, + 0.584, + 0.246 + ], + "angle": 0, + "content": "University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.251, + 0.568, + 0.264 + ], + "angle": 0, + "content": "v-mem.github.io" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.285, + 0.418, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.285, + 0.907, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.483, + 0.908, + 0.525 + ], + "angle": 0, + "content": "Figure 1. VMem enables autoregressive scene generation from a single image along user-defined trajectories. The green region shows results with the proposed memory module, maintaining coherence when generating previously seen parts of the scene. The red region, without memory, exhibits degradation highlighted with red ellipses, demonstrating VMem is effective for consistent scene generation." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.54, + 0.327, + 0.555 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.486, + 0.885 + ], + "angle": 0, + "content": "We propose a novel memory module for building video generators capable of interactively exploring environments. Previous approaches have achieved similar results either by out-painting 2D views of a scene while incrementally reconstructing its 3D geometry—which quickly accumulates errors—or by using video generators with a short context window, which struggle to maintain scene coherence over the long term. To address these limitations, we introduce Surfel-Indexed View Memory (VMem), a memory module that remembers past views by indexing them geometrically based on the 3D surface elements (surfels) they have observed. VMem enables efficient retrieval of the most relevant past views when generating new ones. By focusing only on these relevant views, our method produces consistent explorations of imagined environments at a fraction of the computational cost required to use all past views as context. We evaluate our approach on challenging long-term scene synthesis benchmarks and demonstrate superior performance compared to existing methods in maintaining scene coherence and camera control." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.54, + 0.645, + 0.555 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.793 + ], + "angle": 0, + "content": "We consider the problem of generating long videos that explore an imagined space following a camera path specified interactively by the user. In this paradigm, the user tells the model which camera path to follow for the next few frames, observes the generated content, and then decides where to explore next based on what they have seen. For example, exploring a house may involve visiting the kitchen, the living room, and the bathroom, eventually returning to the kitchen. Throughout the video, the scene must remain consistent, ensuring that the kitchen looks the same upon return. Generating such videos is essential for immersive applications such as games, where players can navigate generated worlds. However, even recent large-scale interactive video generators such as Google's Genie 2 [20] struggle to achieve this goal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This problem has so far been addressed by two types of methods. First, outpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 43] iterate between generating new 2D views of the scene and estimating its 3D geometry. They use the estimated geometry to partially render a new viewpoint and then employ an outpainting model to complete the missing parts, thus adding one more image to the col" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25690" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "lection. However, errors in outpainting, 3D reconstruction, and stitching accumulate over time, leading to severe degradation of the generated content after a short while." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Second, multi-view/video-based methods [20, 22, 25, 35, 42] condition novel view generation on previous views using a geometry-free approach that does not explicitly estimate the scene geometry. While this avoids the accumulation of errors in the reconstructed 3D scene geometry, it comes at a high computational cost, limiting the number of conditioning views to a small context window of recent frames. This constraint hurts the long-term consistency of the generated images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.482, + 0.426 + ], + "angle": 0, + "content": "In this work, we revisit the second class of methods and propose conditioning not on the most recently generated views, but on the most relevant ones for generating the next image, thereby maintaining a high degree of consistency within a limited computational budget. Given a novel viewpoint of the scene, the most relevant past views are those that have already observed the parts of the scene currently being generated. This implies that, for each part of the scene, we must remember which views have previously observed it and retrieve them from memory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.67 + ], + "angle": 0, + "content": "To achieve this, we introduce Surfel-Indexed View Memory, abbreviated as VMem, a memory module that anchors previous views to the 3D surface elements they observe. Given a new viewpoint, we retrieve the past views that best capture the currently observed surfaces and use them to condition novel view generation. To create the memory, we estimate the geometry of each new view using an off-the-shelf point map predictor. This is similar to outpainting methods but, crucially, we do not use the estimated geometry as the final representation of the scene; instead, we use it to construct VMem, which is a memory of past views. We represent the scene geometry using surfels, which are more robust compared to meshes and can represent occlusions compared to point clouds. Each surfel stores in its attributes a set of indices corresponding to past viewpoints that observed it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.852 + ], + "angle": 0, + "content": "To retrieve relevant past viewpoints, we render the surfels with their attributes from the novel viewpoint and splat them onto an image grid. Each pixel in the resulting image then corresponds to a set of viewpoint indices. We select the top \\(K\\) most frequently represented viewpoint indices and use them to retrieve past views from the database, where each view is represented by an RGB image and its corresponding camera parameters. A key advantage of our approach over outpainting-based methods is that it does not require highly accurate scene geometry. As long as we successfully retrieve the most relevant past views, our method remains robust." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "By leveraging Surfel-Indexed View Memory, we significantly reduce the memory and computational burden of conditioning on a large number of previous views, as re" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "quired by multi-view/video-based methods, while improving long-term consistency across generated novel views." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.124, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Our approach represents a step toward scalable, realistic, and long-term autoregressive scene generation, making the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.17, + 0.905, + 0.214 + ], + "angle": 0, + "content": "1. We introduce Surfel-Indexed View Memory, a plug-and-play module to index past views geometrically and use them to condition novel view generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.216, + 0.905, + 0.259 + ], + "angle": 0, + "content": "2. We show that our method can generate long-term coherent videos of scenes and outperforms existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.261, + 0.905, + 0.305 + ], + "angle": 0, + "content": "3. We demonstrate that Surfel-Indexed View Memory achieves comparable performance with \\(4 \\times\\) fewer context views, delivering a \\(12 \\times\\) speedup." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.306, + 0.905, + 0.349 + ], + "angle": 0, + "content": "4. We validate our approach on challenging benchmarks, outperforming the current open-source state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.17, + 0.905, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.369, + 0.651, + 0.384 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) is a challenging and ill-posed problem that can be categorized into two main categories: view interpolation methods [3-6, 8, 19, 30, 37], which generate views close to given inputs, and view extrapolation or autoregressive view generation, where novel views extend significantly beyond the original scene, introducing substantial new content. The latter is particularly difficult from single images. Between these lie single-view scene reconstruction methods, which succeed mainly for single-object scenes [1, 12, 17, 27, 29, 38] or highly bounded scenes [28]. However, their extrapolation capabilities remain limited. Most relevant to our work are single-image view extrapolation models, which fall into two categories: those incorporating explicit geometric modeling with inpainting, and those based directly on image or video generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Inpainting-based view extrapolation. Inpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 40, 41, 43] use pre-trained 3D reconstruction models to generate 3D representations—such as meshes, point clouds, or Gaussian splats—from images. These representations are reprojected into novel views, where 2D inpainting fills missing regions. SceneScape [7] reconstructs meshes from images using pretrained depth estimators. Diffusion-based inpainting then completes the projected novel view, which is reprojected back to refine and extend the mesh. This process is iterated to generate novel views. MultiDiff [18] uses depth estimators to warp reference images into multiple novel views, training diffusion models to inpaint missing regions simultaneously. CamCtrl3D [21] follows a similar approach but adds ray maps to condition diffusion models for view synthesis. ViewCrafter [43] uses pre-trained pointmap estimators to create point clouds and trains video diffusion mod" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.91, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Method. Given target camera viewpoints \\(\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}\\), we query our Surfel-Indexed View Memory to retrieve the most relevant \\(K\\) past views \\(\\mathcal{V}^* \\subset \\mathcal{V}^{(s)}\\) where \\(\\mathcal{V}^* = \\{v_t\\}_{t = 1}^K\\) as references. Retrieved reference images \\(\\mathbf{x}_t\\) along with Plücker embeddings of both reference camera poses \\(\\mathbf{c}_t\\) and target camera poses \\(\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}\\) are fed into generator \\(\\psi\\) to synthesize novel views \\(\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}\\). After generation, the surfel-indexed memory is updated \\(S^{(s)} \\to S^{(s + 1)}\\) by appending new view indices \\(\\{T + m\\}_{m = 1}^{M}\\) to existing surfels or creating new surfels based on geometry of the generated views. This is repeated autoregressively, enabling long-term consistent generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.482, + 0.422 + ], + "angle": 0, + "content": "els for inpainting. GenWarp [26] warps 2D coordinate representations of reference views into novel views and uses them to condition generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.424, + 0.483, + 0.546 + ], + "angle": 0, + "content": "While these methods improve geometric consistency, they are susceptible to errors in depth or pointmap estimation, which propagate distortions across generated views. Once 3D representations are constructed, inaccuracies from depth or pointmap estimation become difficult to correct. Moreover, scaling to large scenes is computationally intensive, as storing and processing high-fidelity 3D representations requires substantial memory." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.554, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Multi-view-based view extrapolation. Another category [8, 22, 25, 31, 35, 42, 46] avoids explicitly modeling 3D scene geometry. Instead, these approaches generate novel views by conditioning on previously rendered or generated views. GeoGPT [25] uses autoregressive likelihood models to synthesize novel views from single images. Building on this, LookOut [22] improves consistency by generating view sequences along predefined trajectories, conditioning on up to two preceding views. Other methods [31, 42] enhance novel view synthesis by employing cross-view attention or enforcing epipolar constraints within diffusion models. MotionCtrl [35] leverages geometric priors from pre-trained video diffusion models to synthesize camera-controllable videos. These methods are trajectory-based, requiring input views to follow continuous spatial paths. In contrast, image-set models like CAT3D [8] and more recently SEVA [46] synthesize novel views from sparse and unordered input images, without trajectory constraints. However, these methods are computationally expensive, with \\( O(n^{2}) \\) complexity for attention, limiting conditioning views to small context windows containing only recently generated frames. This limitation impacts long-term consistency of generated images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.528 + ], + "angle": 0, + "content": "Several concurrent works explore memory-based approaches for scene generation. Genie [2, 20] maintains memory through recurrent state features. Gen3C [23] conditions on stored point clouds like ViewCrafter [43], inheriting inpainting-based limitations. StarGen [44] and WorldMem [39] use distance- and field-of-view-based spatial memory retrieval, which can recover spatially correlated views but lack geometric reasoning for occlusions. In contrast, our surfel-based memory explicitly models occlusions and provides more principled view selection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.605, + 0.558 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x}_1 \\in \\mathbb{R}^{H \\times W \\times 3}\\) be an RGB image of a scene that we wish to explore, and let \\(\\{\\mathbf{c}_t\\}_{t=1}^T\\) be a sequence of camera parameters specifying a path through the scene. Our goal is to generate a corresponding video, i.e., a sequence of images \\(\\{\\mathbf{x}_t\\}_{t=1}^T\\), where \\(\\mathbf{x}_1\\) is the given input image, and each subsequent frame \\(\\mathbf{x}_t\\) is consistent with the previous ones and reflects the specified cameras \\(\\{\\mathbf{c}_t\\}_{t=1}^T\\). Moreover, we aim to do so autoregressively, generating \\(M\\) novel views at a time for the given cameras (with \\(M\\) kept small to ensure interactivity) while ensuring consistency with previously generated views. This enables interactive scene exploration, where the user can specify the next camera positions at each step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Specifically, we work with a sequence of views, where each view \\( v_{t} = (\\mathbf{x}_{t},\\mathbf{c}_{t}) \\) consists of an image \\( \\mathbf{x}_t \\) and its corresponding camera \\( \\mathbf{c}_t \\) at timestep \\( t \\). At generation step \\( s(s\\geq 0) \\), we have generated \\( T = sM \\) frames so far. The goal is to generate \\( M \\) new RGB images \\( \\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M} \\) for the next target camera parameters \\( \\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M} \\), given the previously generated views \\( \\mathcal{V}^{(s)} = \\{v_1,v_2,\\dots ,v_T\\} \\). After generation, we update \\( T\\gets T + M \\)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "A challenge with this approach is that the context \\(\\mathcal{V}^{(s)}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25692" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.238, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.094, + 0.405, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.089, + 0.567, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.099, + 0.723, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.099, + 0.877, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.211, + 0.908, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Surfel-based memory index. Each surfel stores indices of views that observed it. We color-code each surfel by contributing view indices. This spatial index enables retrieval of relevant past views: when generating a novel view, we identify visible surfels from the target viewpoint and retrieve views that previously observed those same regions, naturally accounting for occlusion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.28, + 0.482, + 0.37 + ], + "angle": 0, + "content": "grows unbounded over time. Video generators addressing this problem [22, 25, 35, 42] typically consider a fixed-length subset of \\(\\mathcal{V}^{(s)}\\), conditioning their generation only on the most recent \\(L\\) views \\(\\{v_{T - L + 1},\\ldots ,v_T\\}\\). This limitation results in severe inconsistencies when the generated sequence extends beyond the (small) context window." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.483, + 0.493 + ], + "angle": 0, + "content": "We solve this problem by dynamically retrieving the most relevant subset of past views, denoted as \\(\\mathcal{V}^* \\subseteq \\mathcal{V}^{(s)}\\). To achieve this, we introduce an efficient data structure, the Surfel-Indexed View Memory, that stores and retrieves past views based on their approximate 3D geometry. We first describe the Surfel-Indexed View Memory (Sec. 3.1) and then introduce the novel view generator (Sec. 3.2). An overview of the method is provided in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.502, + 0.347, + 0.518 + ], + "angle": 0, + "content": "3.1. Surfel-indexed view memory" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.736 + ], + "angle": 0, + "content": "Consider the problem of generating the next views \\(\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}\\) of the scene as seen from the cameras \\(\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}\\). The generator must consider the previously generated views and ensure \\(\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}\\) are consistent with those, while generating new content as needed. However, not all past views are equally relevant to the novel views \\(\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}\\). We prioritize past views that have likely observed the largest portion of the scene currently being generated. For example, if the current locale is surrounded by walls, parts of the scene behind those walls are likely less relevant for generating a new view of that locale. Conversely, views that share a similar visible region with the target view can provide more useful information for generating its content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.737, + 0.483, + 0.857 + ], + "angle": 0, + "content": "This principle drives our design for a view memory module. We maintain a coarse model of the scene geometry using surfels (visualized in Fig. 3), simple surface primitives that account for occlusion. Each surfel stores the indices of the past views that observed it. To generate a new view, the surfels visible from the new viewpoint are retrieved, using the associated indices to vote for which views to consider for generation. This process is illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.858, + 0.395, + 0.873 + ], + "angle": 0, + "content": "Specifically, we define a surfel as the tuple" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.886, + 0.368, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {k} = \\left(\\mathbf {p} _ {k}, \\mathbf {n} _ {k}, r _ {k}, \\mathcal {I} _ {k}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.419 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_k \\in \\mathbb{R}^3\\) denotes the surfel's 3D position, \\(\\mathbf{n}_k \\in \\mathbb{R}^3\\) is its surface normal, \\(r_k \\in \\mathbb{R}\\) is the surfel radius, and \\(\\mathcal{I}_k \\subseteq \\{1,2,\\dots,T\\}\\) is a subset of past view indices that observed the surfel. The surfel-based memory indexing of past views \\(\\mathcal{V}^{(s)}\\), encapsulating the scene, is represented as a set of surfels \\(\\mathcal{S}^{(s)} = \\{\\mathbf{s}_k\\}_{k=1}^{N^{(s)}}\\), where \\(N^{(s)}\\) is the number of surfels at generation step \\(s\\). We also maintain an octree to quickly retrieve surfels based on their geometry. Next, we describe how to read from and write to this memory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Reading from the memory. To retrieve the most relevant past views from the memory \\(\\mathcal{V}^{(s)}\\) for the novel cameras \\(\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}\\), we first compute their average pose \\(\\bar{\\mathbf{c}}_s\\in\\) SE(3) as a reference (refer to the supplementary material). Then we render the surfels \\(S^{(s)}\\) from the averaged camera pose \\(\\bar{\\mathbf{c}}_s\\). Each surfel is rendered as a splat with its attributes, accounting for relative depth and occlusions, and contributing to the image based on its coverage. The rendered attributes correspond to the indices of the past views. The core intuition is that views observing the largest portion of the scene from the perspective of the novel camera are most relevant for novel view synthesis. To identify these views, we rank past view indices based on their frequency across all rendered pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.711 + ], + "angle": 0, + "content": "We then select the top-\\(K\\) most frequently represented indices \\(\\mathcal{I}^*\\) and use them to retrieve the corresponding past views, \\(\\mathcal{V}^* = \\{v_t\\}_{t\\in \\mathcal{I}^*}\\), from the memory \\(\\mathcal{V}^{(s)}\\). This subset of views serves as the context for the new view generator, as described in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.834 + ], + "angle": 0, + "content": "To avoid oversampling repeatedly visited regions, we apply a non-maximum suppression algorithm that reduces redundancy in memory and promotes broader scene coverage among the top-\\(K\\) views. During retrieval, we retain only the most frequently referenced view among those with similar poses. During writing to the memory, we merge surfels by comparing their associated camera poses—if two poses are highly similar, we discard the older one." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Writing to the memory. After generating the new views \\(\\{v_{T + m} = (\\mathbf{x}_{T + m},\\mathbf{c}_{T + m})\\}_{m = 1}^{M}\\), we add them to the memory \\(\\mathcal{V}^{(s)}\\) and update the surfel-based memory index \\(S^{(s)}\\). We use an off-the-shelf point map estimator \\(\\phi\\), such as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25693" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.09, + 0.895, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.189, + 0.895, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.906, + 0.358 + ], + "angle": 0, + "content": "Figure 4. Surfel-Indexed View Memory. Reading procedure renders surfels \\( S^{(s)} \\) with their attributes, containing past view indices as frame indices. We then select the \\( K \\) most frequent frame indices in the rendered image to retrieve relevant past views from \\( \\mathcal{V}^{(s)} \\). Writing procedure estimates geometry of newly generated views \\( \\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M} \\) as surfels and merges them with existing surfels. Frame indices \\( \\{T + m\\}_{m = 1}^{M} \\) are appended to surfels in these views, and novel views are stored, updating \\( \\mathcal{V}^{(s)} \\to \\mathcal{V}^{(s + 1)} \\) and \\( S^{(s)} \\to S^{(s + 1)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.482, + 0.473 + ], + "angle": 0, + "content": "CUT3R [32], to estimate their point maps \\(\\{\\mathbf{P}_{T + m}^{*}\\}_{m = 1}^{M}\\). Specifically, we jointly estimate the point maps for the newly generated views along with the retrieved past views \\(\\nu^{*}\\). This joint estimation ensures that the new point maps are aligned with the coordinate frame of the existing scene geometry in the surfel-based memory index \\(S^{(s)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Next, we convert the estimated point maps to a set of new surfels. For each newly generated frame \\(t \\in \\{T + 1, T + 2, \\dots, T + M\\}\\), we process its point map \\(\\mathbf{P}_t^*\\). Since our approach only requires coarse geometry, we first downsample the point map \\(\\mathbf{P}_t^*\\) by a factor \\(\\sigma\\), yielding a smaller point map \\(\\mathbf{P}_t \\in \\mathbb{R}^{H' \\times W' \\times 3}\\), where \\(H' = H / \\sigma\\) and \\(W' = W / \\sigma\\). For each pixel location \\((u, v)\\) in the downsampled point map, we have a 3D point \\(\\mathbf{p}_{u,v,t}\\), and we create a new surfel \\(\\mathbf{s}_{k'}\\) centered on it. We compute each surfel's normal using the cross-product of displacement vectors to neighboring pixels. For a pixel at 2D location \\((u, v)\\) in the image grid, we use neighboring pixels to estimate the local surface normal:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.665, + 0.478, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} _ {k ^ {\\prime}} = \\frac {(\\mathbf {p} _ {u + 1 , v , t} - \\mathbf {p} _ {u - 1 , v , t}) \\times (\\mathbf {p} _ {u , v + 1 , t} - \\mathbf {p} _ {u , v - 1 , t})}{\\| (\\mathbf {p} _ {u + 1 , v , t} - \\mathbf {p} _ {u - 1 , v , t}) \\times (\\mathbf {p} _ {u , v + 1 , t} - \\mathbf {p} _ {u , v - 1 , t}) \\|},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.8 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{u,v,t}\\) represents the 3D point at image coordinates \\((u,v)\\) in frame \\(t\\). To ensure that the surfels reasonably cover the scene, we use a heuristic to compute the surfel's radius, which is proportional to the depth of the surfel and inversely proportional to the focal length and the cosine of the angle between the surfel normal and the viewing direction:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.81, + 0.429, + 0.846 + ], + "angle": 0, + "content": "\\[\nr _ {k ^ {\\prime}} = \\frac {\\frac {1}{2} \\mathbf {D} _ {u , v , t} / f _ {t}}{\\alpha + (1 - \\alpha) \\left\\| \\mathbf {n} _ {k ^ {\\prime}} \\cdot (\\mathbf {p} _ {u , v , t} - \\mathbf {O} _ {t}) \\right\\|},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}_{u,v,t}\\) represents the depth at pixel \\((u,v)\\) in frame \\(t\\), \\(\\mathbf{O}_t\\) is the camera center of frame \\(t\\), \\(\\mathbf{p}_{u,v,t} - \\mathbf{O}_t\\) is the viewing direction, and \\(f_t\\) is the focal length at time" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.383, + 0.905, + 0.413 + ], + "angle": 0, + "content": "\\(t\\). The factor \\(\\alpha\\) is used to avoid extreme values when \\(\\| \\mathbf{n}_{k'} \\cdot (\\mathbf{p}_{u,v,t} - \\mathbf{O}_t) \\| \\approx 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.906, + 0.551 + ], + "angle": 0, + "content": "Finally, we check if the index \\( S^{(s)} \\) already contains a surfel \\( \\mathbf{s}_k \\) similar to the new surfel \\( \\mathbf{s}_{k'} \\). Two surfels match if their centers are within a distance \\( d \\) and the cosine similarity between their normals is above a threshold \\( \\theta \\). If such a surfel \\( \\mathbf{s}_k \\) is found, we add the frame index \\( t \\) to the surfel's set of past view indices: \\( \\mathcal{I}_k \\gets \\mathcal{I}_k \\cup \\{t\\} \\), and discard \\( \\mathbf{s}_{k'} \\). If not, we set \\( \\mathcal{I}_{k'} = \\{t\\} \\) and add \\( \\mathbf{s}_{k'} \\) to the index \\( S^{(s + 1)} \\). This process transforms the surfel memory from \\( S^{(s)} \\) to \\( S^{(s + 1)} \\) with \\( N^{(s + 1)} \\) surfels for the next generation step." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.714, + 0.579 + ], + "angle": 0, + "content": "3.2. Novel view generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.906, + 0.661 + ], + "angle": 0, + "content": "We use a camera-conditioned image-set generator \\(\\psi\\) to generate novel views. The generator \\(\\psi\\) takes \\(K\\) retrieved reference views \\(\\mathcal{V}^* = \\{v_t = (\\mathbf{x}_t, \\mathbf{c}_t)\\}_{t \\in \\mathcal{I}^*}\\) and the target camera poses \\(\\{\\mathbf{c}_{T + m}\\}_{m=1}^M\\) for the \\(M\\) new frames to be generated and samples the novel views:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.673, + 0.878, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {x} _ {T + m} \\right\\} _ {m = 1} ^ {M} \\sim \\psi \\left(\\left\\{\\left(\\mathbf {x} _ {t}, \\mathbf {c} _ {t}\\right) \\right\\} _ {t \\in \\mathcal {I} ^ {*}}, \\left\\{\\mathbf {c} _ {T + m} \\right\\} _ {m = 1} ^ {M}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}^* \\subseteq \\{1, 2, \\ldots, T\\}\\) are the frame indices of the retrieved reference views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.906, + 0.827 + ], + "angle": 0, + "content": "Specifically, we base our generator on the recent SEVA [46] model. Given the plug-and-play nature of our memory module, VMem can work with other image-set generators as well. We also fine-tune a more efficient version of SEVA that operates with a reduced number of reference frames. We provide more details in Sec. 4.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.646, + 0.862 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We evaluate our method on established benchmarks for camera-conditioned autoregressive view generation from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.088, + 0.885, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.489, + 0.908, + 0.531 + ], + "angle": 0, + "content": "Figure 5. Long sequences with revisitations. We compare our VMem against a baseline without memory that relies solely on the last \\( K \\) frames for context. Each sequence: input images (left), then generated images at selected frames. Our method (top rows) maintains consistency when revisiting observed regions, while the baseline (bottom rows) shows severe inconsistencies across extended sequences." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.543, + 0.885, + 0.707 + ], + "angle": 0, + "content": "
Short-term (50thframe)Long-term (≥200thframe)
LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓
GeoGPT [25]0.44413.35—26.72——0.6749.54—41.87——
Lookout [22]0.37814.43—28.861.2410.8760.65810.51—58.121.1420.924
PhotoNVS [42]0.33315.51—21.76——0.58811.54—41.95——
GenWarp [26]0.43612.030.14429.690.5640.0590.6139.560.08536.400.8500.251
MotionCtrl [35]0.42412.000.14819.250.3410.3560.6059.130.08335.450.7480.609
ViewCrafter [43]0.37716.970.26225.391.5620.2080.5929.740.14834.122.1770.827
SEVA [46] (K=17)0.29318.330.38217.290.2230.1180.45513.980.21623.821.1250.742
VMem (K=4)0.28718.490.40617.120.2190.0390.49313.120.18327.150.8110.499
VMem (K=17)0.29318.330.38217.290.2230.1180.45214.090.22723.560.9820.432
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.717, + 0.908, + 0.801 + ], + "angle": 0, + "content": "Table 1. Single-view NVs on RealEstate10K. Frames are subsampled with 10-frame intervals. \\( K \\) denotes the number of context views. In the short-term setting, all past frames fit in the context windows for both SEVA [46] and VMem; hence, SEVA [46] and VMem (\\( K = 17 \\)) yield identical results since VMem uses SEVA as the backbone. VMem (\\( K = 4 \\)) is fine-tuned on RealEstate10K for this setting, yielding better performance than SEVA [46] and VMem (\\( K = 17 \\)) in short-term scenarios. VMem surpasses baselines on most metrics in this benchmark. However, its true advantage—spatial consistency—is not fully reflected because RealEstate10K trajectories rarely revisit previously observed areas. We address this limitation in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.482, + 0.899 + ], + "angle": 0, + "content": "a single image, comparing our models to previous open-source approaches (Sec. 4.2). These benchmarks, however, have a critical limitation: camera trajectories rarely revisit previously observed regions, unlike real-world behavior where humans or robots often return to or look at previ" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.907, + 0.9 + ], + "angle": 0, + "content": "ously explored areas. To address this limitation, we propose a cycle-trajectory evaluation protocol that extends these benchmarks by making the camera return to the starting position along the same path in reverse order (Sec. 4.3). Additionally, we collect a small dataset with trajectories that" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25695" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.478, + 0.194 + ], + "angle": 0, + "content": "
MethodLPIPS ↓PSNR ↑SSIM ↑FID ↓\\( R_{\\text{dist}} \\) ↓\\( T_{\\text{dist}} \\) ↓
Look-out [22]0.8098.410.06938.341.2620.341
GenWarp [26]0.50711.130.13432.940.8480.241
MotionCtrl [35]0.5899.070.09626.860.8710.293
ViewCrafter [43]0.40111.820.21724.720.9020.492
SEVA [46] (K=17)0.40111.820.21724.720.9020.492
VMem (K=4)0.39715.720.29724.970.8210.392
VMem (K=17)0.30418.150.37724.180.8920.165
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.205, + 0.47, + 0.219 + ], + "angle": 0, + "content": "Table 2. Results on cycle trajectories from RealEstate10K." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.234, + 0.462, + 0.317 + ], + "angle": 0, + "content": "
MethodLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
Look-out [22]0.7928.520.0080.7271.499
GenWarp [26]0.52110.270.1290.7851.982
MotionCtrl [35]0.6928.210.0820.8420.129
ViewCrafter [43]0.49413.620.1291.6430.492
VMem (K=4)0.47214.110.1211.2040.387
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.327, + 0.482, + 0.342 + ], + "angle": 0, + "content": "Table 3. Results on cycle trajectories from Tanks and Temples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.4 + ], + "angle": 0, + "content": "revisit observed regions for qualitative evaluation in Fig. 5. Finally, we conduct an ablation study in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.279, + 0.427 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.434, + 0.483, + 0.676 + ], + "angle": 0, + "content": "Implementation details. We use pre-trained SEVA [46] as the generator \\(\\psi\\). SEVA uses a fixed total number of reference and target frames (\\(K + M = 21\\)). For our experiments, we use \\(M = 4\\) target views, which leaves \\(K = 17\\) context views. However, this configuration demands substantial computational resources that may be prohibitive for many applications. To address this limitation, we fine-tune a more efficient version with LoRA [13] on the RealEstate10K training split [47] that works with a reduced number of reference frames (\\(K = 4\\)) and target views (\\(M = 4\\)), and use this efficient version for our main experiments. We demonstrate (Secs. 4.3 and 4.4) that this fine-tuned version, when combined with VMem, achieves comparable performance to the original SEVA while delivering approximately \\(12\\times\\) speedup at inference time. We provide more technical details in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Datasets. We consider two real-world datasets for novel view generation evaluation: RealEstate10K [47], comprising indoor scene video clips, and Tanks-and-Temples [14], including indoor and outdoor scenes with larger camera motions. For qualitative evaluation, we use in-the-wild images collected from the internet or captured with phone cameras as input views (see Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate generated novel view quality based on three factors: (1) overall quality of generated novel views, comparing generated novel view distributions to test set distributions using Fréchet Image Distance (FID) [10]; (2) the ability of the model to preserve image details across views, measuring peak signal-to-noise ratio (PSNR) for pixel differences, along with LPIPS [45]" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.208 + ], + "angle": 0, + "content": "
Context views KView RetrievalLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
17Temporal0.47713.920.1880.9760.254
Camera Distance (SEVA [46])0.39715.720.2970.8210.392
Field of View0.37415.750.2920.9110.382
VMem0.30418.150.3770.8920.165
4Temporal0.7947.520.0181.9420.458
Camera Distance0.42213.270.1871.7870.319
Field of View0.42413.110.1921.7820.285
VMem0.38114.820.2750.7930.124
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.219, + 0.907, + 0.261 + ], + "angle": 0, + "content": "Table 4. Ablation study on cycle trajectories from RealEstate10K. We compare different view retrieval strategies for two context view counts \\( K \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.38 + ], + "angle": 0, + "content": "and SSIM [34], as in [22]; and (3) alignment between generated camera poses and ground truth, following [9], where we express estimated camera poses relative to the first frame and normalize translation by the furthest frame. We use DUSt3R [33] to extract poses from generated views. We calculate rotation distance \\( R_{\\mathrm{dist}} \\) by comparing ground truth and extracted rotation matrices of each generated sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.39, + 0.846, + 0.41 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathrm {d i s t}} = \\operatorname {a r c c o s} \\left(0. 5 \\left(\\operatorname {t r} \\left(\\mathbf {R} _ {\\mathrm {g e n}} \\mathbf {R} _ {\\mathrm {g t}} ^ {T}\\right) - 1\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.42, + 0.907, + 0.45 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}\\) denotes a rotation matrix. We compute translation distance \\(t_{\\mathrm{dist}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.462, + 0.78, + 0.481 + ], + "angle": 0, + "content": "\\[\nt _ {\\text {d i s t}} = \\left\\| \\mathbf {t} _ {\\mathrm {g t}} - \\mathbf {t} _ {\\mathrm {g e n}} \\right\\| _ {2}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.839, + 0.516 + ], + "angle": 0, + "content": "4.2. Short- and long-term view generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.765 + ], + "angle": 0, + "content": "We first evaluate VMem on the benchmark from [22, 42], which defines two scenarios: short-term and long-term novel view generation. Beginning from the first frame of each ground-truth test sequence, the model autoregressively generates images along the ground truth camera trajectories with 10-frame intervals. The short-term scenario evaluates the fifth generated image (50 frames from the initial view), while the long-term scenario evaluates the final image (\\(\\geq 200\\) frames from the initial view). Following the protocol, we only consider samples with more than 200 frames for the long-term scenario and more than 50 frames for the short-term scenario. We report results in Tab. 1. VMem substantially outperforms all baselines on the key metrics, demonstrating superior scene extrapolation. However, the benchmark trajectories rarely revisit observed regions, limiting evaluation of long-term scene consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.896, + 0.79 + ], + "angle": 0, + "content": "4.3. Long-term view generation with revisitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To evaluate long-term scene consistency when revisiting observed regions—a scenario that rarely occurs in existing benchmarks—we propose a cycle-trajectory evaluation protocol as a proxy for real revisitation scenarios in human and robotic exploration. In this protocol, the model generates frames following trajectories from initial to final poses, then returns along the same path to the starting position." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25696" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.496, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.09, + 0.901, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.908, + 0.352 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of the cycle trajectory frames (\\(\\geq 400\\) frames). VMem autoregressively generates new views while maintaining memory of all previously generated frames, ensuring consistency when revisiting previously seen locations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.483, + 0.483 + ], + "angle": 0, + "content": "We evaluate metrics every 10 frames on the return trajectories, using all test sequences. Following this protocol, we report quantitative results in Tab. 2 and qualitative comparisons in Fig. 6. VMem outperforms all baselines across all metrics in cycle-trajectory evaluation, demonstrating superior visual consistency when revisiting previously observed regions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.485, + 0.483, + 0.606 + ], + "angle": 0, + "content": "To validate generalization, we evaluate on Tanks-and-Temples containing indoor and outdoor scenes with more dynamic camera trajectories. We use all six advanced scenes and apply the same cycle-trajectory protocol using the first 50 frames, evaluating every frame without temporal subsampling. Results are in Tab. 3. VMem consistently outperforms all baselines across in-domain (Tab. 2) and out-of-domain (Tab. 3) evaluations on the majority of metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.667 + ], + "angle": 0, + "content": "We also qualitatively evaluate VMem across diverse in-the-wild scenes in Fig. 5, demonstrating generation of diverse, consistent, high-quality long videos on in-the-wild scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.241, + 0.698 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We ablate our surfel-based memory indexing for context view retrieval in VMem. We evaluate our approach in the cycle-trajectory setting against three alternative retrieval strategies: (1) temporal-based retrieval, selecting the most recent \\( K \\) views from memory, (2) camera distance-based retrieval, selecting top \\( K \\) views with cameras closest to the target view camera—also used by SEVA [46], our backbone generator, and (3) field-of-view-based retrieval, selecting top \\( K \\) views with highest field-of-view overlap with the target view. We also evaluate performance trade-offs between our lightweight, fine-tuned generator using \\( K = 4 \\) context views versus original SEVA with \\( K = 17 \\) context views. We report results in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.906, + 0.453 + ], + "angle": 0, + "content": "Our findings show VMem consistently improves generator performance across both numbers of context views. Improvement is most pronounced with fewer context views \\((K = 4)\\), highlighting the effectiveness of our surfel-based memory index." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.909, + 0.583 + ], + "angle": 0, + "content": "Computational efficiency. A key practical advantage of our approach is its computational efficiency. VMem with only \\( K = 4 \\) views achieves \\( 12 \\times \\) speed improvement over the original SEVA while recovering most of the performance achieved with \\( K = 17 \\) views. On an RTX 4090 GPU, our method generates frames in 4.2 seconds compared to 50 seconds for SEVA, representing a step towards real-time interactive scene exploration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.634, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.823 + ], + "angle": 0, + "content": "We introduced Surfel-Indexed View Memory (VMem), a novel plug-and-play memory module for long-term autoregressive scene generation using a surfel-indexed view memory. By anchoring past views to a surfel-based representation of the scene and retrieving the most relevant ones, our approach improves long-term scene consistency while reducing computational costs. Experiments on long-term scene synthesis benchmarks demonstrate that VMem outperforms existing methods in scene coherence while enabling the use of fewer context views, leading to significantly faster generation. Our work advances scalable video generation, with applications in virtual reality, gaming, and other domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments. The authors of this work are supported by Clarendon scholarship, ERC 101001212-UNION, and AIMS EP/S024050/1. The authors would like to thank Xingyi Yang, Zeren Jiang, Junlin Han, Zhongrui Gui for their insightful feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25697" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.122, + 0.188, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.147, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[1] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. SF3D: Stable fast 3D mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.204, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.402 + ], + "angle": 0, + "content": "[4] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction, 2023. arXiv:2304.06714 [cs]." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.483, + 0.529 + ], + "angle": 0, + "content": "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[7] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. Advances in Neural Information Processing Systems, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[9] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7909-7920, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.147, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[15] Jing Yu Koh, Harsh Agrawal, Dhruv Batra, Richard Tucker, Austin Waters, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Simple and effective synthesis of indoor 3d scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1169-1178, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[16] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[17] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. RealFusion: 360 reconstruction of any object from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[18] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[20] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Rocktäschel. Genie 2: A large-scale foundation world model, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[21] Stefan Popov, Amit Raj, Michael Krainin, Yanzhen Li, William T. Freeman, and Michael Rubinstein. Camctrl3d: Single-image scene exploration with precise 3d camera control, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[22] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25698" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "world-consistent video generation with precise camera control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[24] Chris Rockwell, David F Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14104-14113, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[25] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14356-14366, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.482, + 0.35 + ], + "angle": 0, + "content": "[26] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv.cs, abs/2310.15110, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arxiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[29] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.61 + ], + "angle": 0, + "content": "[30] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.613, + 0.483, + 0.682 + ], + "angle": 0, + "content": "[31] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[32] Qianqian Wang*, Yifei Zhang*, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In CVPR, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[33] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. ACM SIGGRAPH 2024 Conference Papers, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[36] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[37] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21551-21561, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[38] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[39] Zeqi Xiao, Yushi Lan, Yifan Zhou, Wenqi Ouyang, Shuai Yang, Yanhong Zeng, and Xingang Pan. Worldmem: Long-term consistent world simulation with memory, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[40] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T. Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, and Charles Herrmann. Wonderjourney: Going from anywhere to everywhere. arXiv.cs, abs/2312.03884, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.443 + ], + "angle": 0, + "content": "[41] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T. Freeman, and Jiajun Wu. WonderWorld: interactive 3D scene generation from a single image. arXiv, 2406.09394, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[42] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7094-7104, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[43] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.587, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[44] Shangjin Zhai, Zhichao Ye, Jialin Liu, Weijian Xie, Jiaqi Hu, Zhen Peng, Hua Xue, Danpeng Chen, Xiaomeng Wang, Lei Yang, Nan Wang, Haomin Liu, and Guofeng Zhang. Stargen: A spatiotemporal autoregression framework with video diffusion model for scalable and controllable scene generation, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.672, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.81 + ], + "angle": 0, + "content": "[46] Jensen (Jinghao) Zhou, Hang Gao, Vikram Voleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rupprecht, and Varun Jampani. Stable virtual camera: Generative view synthesis with diffusion models. arXiv preprint arXiv:2503.14489, 2025. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.813, + 0.905, + 0.867 + ], + "angle": 0, + "content": "[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25699" + } + ] +] \ No newline at end of file diff --git a/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_origin.pdf b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b794f608e6a71df854fba61d65623ea92357294f --- /dev/null +++ b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/1781a2ca-916a-47a8-9f71-bf548311fbcb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00be8ec5bb130ea1ba6ad9c6cb07e5437d51868bbcde1414b8030e7e97acf8cd +size 8211656 diff --git a/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/full.md b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9355a61a0751fb2fc683feba19907156da4d4fc4 --- /dev/null +++ b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/full.md @@ -0,0 +1,278 @@ +# VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory + +Runjia Li Philip Torr Andrea Vedaldi Tomas Jakab + +University of Oxford + +v-mem.github.io + +![](images/89c810002b4c80233657cf00ae17e94579b7a2e98175a33a5b6702ab0dc5df6a.jpg) +Figure 1. VMem enables autoregressive scene generation from a single image along user-defined trajectories. The green region shows results with the proposed memory module, maintaining coherence when generating previously seen parts of the scene. The red region, without memory, exhibits degradation highlighted with red ellipses, demonstrating VMem is effective for consistent scene generation. + +![](images/8af08eb47fd8fc071039d3480bd09ade2f068998d8cada51a12627fbb3ac7059.jpg) + +# Abstract + +We propose a novel memory module for building video generators capable of interactively exploring environments. Previous approaches have achieved similar results either by out-painting 2D views of a scene while incrementally reconstructing its 3D geometry—which quickly accumulates errors—or by using video generators with a short context window, which struggle to maintain scene coherence over the long term. To address these limitations, we introduce Surfel-Indexed View Memory (VMem), a memory module that remembers past views by indexing them geometrically based on the 3D surface elements (surfels) they have observed. VMem enables efficient retrieval of the most relevant past views when generating new ones. By focusing only on these relevant views, our method produces consistent explorations of imagined environments at a fraction of the computational cost required to use all past views as context. We evaluate our approach on challenging long-term scene synthesis benchmarks and demonstrate superior performance compared to existing methods in maintaining scene coherence and camera control. + +# 1. Introduction + +We consider the problem of generating long videos that explore an imagined space following a camera path specified interactively by the user. In this paradigm, the user tells the model which camera path to follow for the next few frames, observes the generated content, and then decides where to explore next based on what they have seen. For example, exploring a house may involve visiting the kitchen, the living room, and the bathroom, eventually returning to the kitchen. Throughout the video, the scene must remain consistent, ensuring that the kitchen looks the same upon return. Generating such videos is essential for immersive applications such as games, where players can navigate generated worlds. However, even recent large-scale interactive video generators such as Google's Genie 2 [20] struggle to achieve this goal. + +This problem has so far been addressed by two types of methods. First, outpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 43] iterate between generating new 2D views of the scene and estimating its 3D geometry. They use the estimated geometry to partially render a new viewpoint and then employ an outpainting model to complete the missing parts, thus adding one more image to the col + +lection. However, errors in outpainting, 3D reconstruction, and stitching accumulate over time, leading to severe degradation of the generated content after a short while. + +Second, multi-view/video-based methods [20, 22, 25, 35, 42] condition novel view generation on previous views using a geometry-free approach that does not explicitly estimate the scene geometry. While this avoids the accumulation of errors in the reconstructed 3D scene geometry, it comes at a high computational cost, limiting the number of conditioning views to a small context window of recent frames. This constraint hurts the long-term consistency of the generated images. + +In this work, we revisit the second class of methods and propose conditioning not on the most recently generated views, but on the most relevant ones for generating the next image, thereby maintaining a high degree of consistency within a limited computational budget. Given a novel viewpoint of the scene, the most relevant past views are those that have already observed the parts of the scene currently being generated. This implies that, for each part of the scene, we must remember which views have previously observed it and retrieve them from memory. + +To achieve this, we introduce Surfel-Indexed View Memory, abbreviated as VMem, a memory module that anchors previous views to the 3D surface elements they observe. Given a new viewpoint, we retrieve the past views that best capture the currently observed surfaces and use them to condition novel view generation. To create the memory, we estimate the geometry of each new view using an off-the-shelf point map predictor. This is similar to outpainting methods but, crucially, we do not use the estimated geometry as the final representation of the scene; instead, we use it to construct VMem, which is a memory of past views. We represent the scene geometry using surfels, which are more robust compared to meshes and can represent occlusions compared to point clouds. Each surfel stores in its attributes a set of indices corresponding to past viewpoints that observed it. + +To retrieve relevant past viewpoints, we render the surfels with their attributes from the novel viewpoint and splat them onto an image grid. Each pixel in the resulting image then corresponds to a set of viewpoint indices. We select the top $K$ most frequently represented viewpoint indices and use them to retrieve past views from the database, where each view is represented by an RGB image and its corresponding camera parameters. A key advantage of our approach over outpainting-based methods is that it does not require highly accurate scene geometry. As long as we successfully retrieve the most relevant past views, our method remains robust. + +By leveraging Surfel-Indexed View Memory, we significantly reduce the memory and computational burden of conditioning on a large number of previous views, as re + +quired by multi-view/video-based methods, while improving long-term consistency across generated novel views. + +Our approach represents a step toward scalable, realistic, and long-term autoregressive scene generation, making the following contributions: + +1. We introduce Surfel-Indexed View Memory, a plug-and-play module to index past views geometrically and use them to condition novel view generation. +2. We show that our method can generate long-term coherent videos of scenes and outperforms existing approaches. +3. We demonstrate that Surfel-Indexed View Memory achieves comparable performance with $4 \times$ fewer context views, delivering a $12 \times$ speedup. +4. We validate our approach on challenging benchmarks, outperforming the current open-source state-of-the-art methods. + +# 2. Related work + +Novel view synthesis (NVS) is a challenging and ill-posed problem that can be categorized into two main categories: view interpolation methods [3-6, 8, 19, 30, 37], which generate views close to given inputs, and view extrapolation or autoregressive view generation, where novel views extend significantly beyond the original scene, introducing substantial new content. The latter is particularly difficult from single images. Between these lie single-view scene reconstruction methods, which succeed mainly for single-object scenes [1, 12, 17, 27, 29, 38] or highly bounded scenes [28]. However, their extrapolation capabilities remain limited. Most relevant to our work are single-image view extrapolation models, which fall into two categories: those incorporating explicit geometric modeling with inpainting, and those based directly on image or video generation. + +Inpainting-based view extrapolation. Inpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 40, 41, 43] use pre-trained 3D reconstruction models to generate 3D representations—such as meshes, point clouds, or Gaussian splats—from images. These representations are reprojected into novel views, where 2D inpainting fills missing regions. SceneScape [7] reconstructs meshes from images using pretrained depth estimators. Diffusion-based inpainting then completes the projected novel view, which is reprojected back to refine and extend the mesh. This process is iterated to generate novel views. MultiDiff [18] uses depth estimators to warp reference images into multiple novel views, training diffusion models to inpaint missing regions simultaneously. CamCtrl3D [21] follows a similar approach but adds ray maps to condition diffusion models for view synthesis. ViewCrafter [43] uses pre-trained pointmap estimators to create point clouds and trains video diffusion mod + +![](images/50d1e06551806acb593625cda09e71f750c5a9fdc26e29e0e9b745fe1c2886af.jpg) +Figure 2. Method. Given target camera viewpoints $\{\mathbf{c}_{T + m}\}_{m = 1}^{M}$ , we query our Surfel-Indexed View Memory to retrieve the most relevant $K$ past views $\mathcal{V}^* \subset \mathcal{V}^{(s)}$ where $\mathcal{V}^* = \{v_t\}_{t = 1}^K$ as references. Retrieved reference images $\mathbf{x}_t$ along with Plücker embeddings of both reference camera poses $\mathbf{c}_t$ and target camera poses $\{\mathbf{c}_{T + m}\}_{m = 1}^{M}$ are fed into generator $\psi$ to synthesize novel views $\{\mathbf{x}_{T + m}\}_{m = 1}^{M}$ . After generation, the surfel-indexed memory is updated $S^{(s)} \to S^{(s + 1)}$ by appending new view indices $\{T + m\}_{m = 1}^{M}$ to existing surfels or creating new surfels based on geometry of the generated views. This is repeated autoregressively, enabling long-term consistent generation. + +els for inpainting. GenWarp [26] warps 2D coordinate representations of reference views into novel views and uses them to condition generation. + +While these methods improve geometric consistency, they are susceptible to errors in depth or pointmap estimation, which propagate distortions across generated views. Once 3D representations are constructed, inaccuracies from depth or pointmap estimation become difficult to correct. Moreover, scaling to large scenes is computationally intensive, as storing and processing high-fidelity 3D representations requires substantial memory. + +Multi-view-based view extrapolation. Another category [8, 22, 25, 31, 35, 42, 46] avoids explicitly modeling 3D scene geometry. Instead, these approaches generate novel views by conditioning on previously rendered or generated views. GeoGPT [25] uses autoregressive likelihood models to synthesize novel views from single images. Building on this, LookOut [22] improves consistency by generating view sequences along predefined trajectories, conditioning on up to two preceding views. Other methods [31, 42] enhance novel view synthesis by employing cross-view attention or enforcing epipolar constraints within diffusion models. MotionCtrl [35] leverages geometric priors from pre-trained video diffusion models to synthesize camera-controllable videos. These methods are trajectory-based, requiring input views to follow continuous spatial paths. In contrast, image-set models like CAT3D [8] and more recently SEVA [46] synthesize novel views from sparse and unordered input images, without trajectory constraints. However, these methods are computationally expensive, with $O(n^{2})$ complexity for attention, limiting conditioning views to small context windows containing only recently generated frames. This limitation impacts long-term consistency of generated images. + +Several concurrent works explore memory-based approaches for scene generation. Genie [2, 20] maintains memory through recurrent state features. Gen3C [23] conditions on stored point clouds like ViewCrafter [43], inheriting inpainting-based limitations. StarGen [44] and WorldMem [39] use distance- and field-of-view-based spatial memory retrieval, which can recover spatially correlated views but lack geometric reasoning for occlusions. In contrast, our surfel-based memory explicitly models occlusions and provides more principled view selection. + +# 3. Method + +Let $\mathbf{x}_1 \in \mathbb{R}^{H \times W \times 3}$ be an RGB image of a scene that we wish to explore, and let $\{\mathbf{c}_t\}_{t=1}^T$ be a sequence of camera parameters specifying a path through the scene. Our goal is to generate a corresponding video, i.e., a sequence of images $\{\mathbf{x}_t\}_{t=1}^T$ , where $\mathbf{x}_1$ is the given input image, and each subsequent frame $\mathbf{x}_t$ is consistent with the previous ones and reflects the specified cameras $\{\mathbf{c}_t\}_{t=1}^T$ . Moreover, we aim to do so autoregressively, generating $M$ novel views at a time for the given cameras (with $M$ kept small to ensure interactivity) while ensuring consistency with previously generated views. This enables interactive scene exploration, where the user can specify the next camera positions at each step. + +Specifically, we work with a sequence of views, where each view $v_{t} = (\mathbf{x}_{t},\mathbf{c}_{t})$ consists of an image $\mathbf{x}_t$ and its corresponding camera $\mathbf{c}_t$ at timestep $t$ . At generation step $s(s\geq 0)$ , we have generated $T = sM$ frames so far. The goal is to generate $M$ new RGB images $\{\mathbf{x}_{T + m}\}_{m = 1}^{M}$ for the next target camera parameters $\{\mathbf{c}_{T + m}\}_{m = 1}^{M}$ , given the previously generated views $\mathcal{V}^{(s)} = \{v_1,v_2,\dots ,v_T\}$ . After generation, we update $T\gets T + M$ . + +A challenge with this approach is that the context $\mathcal{V}^{(s)}$ + +![](images/4647dc6736bfed6b6ec55a4c3bbe3d87ceb72cc064bece0ade8f8d9b8d1a9877.jpg) +Figure 3. Surfel-based memory index. Each surfel stores indices of views that observed it. We color-code each surfel by contributing view indices. This spatial index enables retrieval of relevant past views: when generating a novel view, we identify visible surfels from the target viewpoint and retrieve views that previously observed those same regions, naturally accounting for occlusion. + +![](images/948869d140c2846ab916e8dff078913fffb40e97494b87e42bf2d9155faffcd2.jpg) + +![](images/a8e23f819452a8bfd6d5f05fa64a02b27714cb1cfcf9f5db453b9f523e4ac9b9.jpg) + +![](images/c452c93a1eeba2d661c4170b7f035f252adffd53184905967588730517409b88.jpg) + +![](images/58e42b42eb0cab221ac9d4834cec7e1c75f5a75491c2648c159b3df934868e06.jpg) + +grows unbounded over time. Video generators addressing this problem [22, 25, 35, 42] typically consider a fixed-length subset of $\mathcal{V}^{(s)}$ , conditioning their generation only on the most recent $L$ views $\{v_{T - L + 1},\ldots ,v_T\}$ . This limitation results in severe inconsistencies when the generated sequence extends beyond the (small) context window. + +We solve this problem by dynamically retrieving the most relevant subset of past views, denoted as $\mathcal{V}^* \subseteq \mathcal{V}^{(s)}$ . To achieve this, we introduce an efficient data structure, the Surfel-Indexed View Memory, that stores and retrieves past views based on their approximate 3D geometry. We first describe the Surfel-Indexed View Memory (Sec. 3.1) and then introduce the novel view generator (Sec. 3.2). An overview of the method is provided in Fig. 2. + +# 3.1. Surfel-indexed view memory + +Consider the problem of generating the next views $\{\mathbf{x}_{T + m}\}_{m = 1}^{M}$ of the scene as seen from the cameras $\{\mathbf{c}_{T + m}\}_{m = 1}^{M}$ . The generator must consider the previously generated views and ensure $\{\mathbf{x}_{T + m}\}_{m = 1}^{M}$ are consistent with those, while generating new content as needed. However, not all past views are equally relevant to the novel views $\{\mathbf{x}_{T + m}\}_{m = 1}^{M}$ . We prioritize past views that have likely observed the largest portion of the scene currently being generated. For example, if the current locale is surrounded by walls, parts of the scene behind those walls are likely less relevant for generating a new view of that locale. Conversely, views that share a similar visible region with the target view can provide more useful information for generating its content. + +This principle drives our design for a view memory module. We maintain a coarse model of the scene geometry using surfels (visualized in Fig. 3), simple surface primitives that account for occlusion. Each surfel stores the indices of the past views that observed it. To generate a new view, the surfels visible from the new viewpoint are retrieved, using the associated indices to vote for which views to consider for generation. This process is illustrated in Fig. 4. + +Specifically, we define a surfel as the tuple + +$$ +\mathbf {s} _ {k} = \left(\mathbf {p} _ {k}, \mathbf {n} _ {k}, r _ {k}, \mathcal {I} _ {k}\right), +$$ + +where $\mathbf{p}_k \in \mathbb{R}^3$ denotes the surfel's 3D position, $\mathbf{n}_k \in \mathbb{R}^3$ is its surface normal, $r_k \in \mathbb{R}$ is the surfel radius, and $\mathcal{I}_k \subseteq \{1,2,\dots,T\}$ is a subset of past view indices that observed the surfel. The surfel-based memory indexing of past views $\mathcal{V}^{(s)}$ , encapsulating the scene, is represented as a set of surfels $\mathcal{S}^{(s)} = \{\mathbf{s}_k\}_{k=1}^{N^{(s)}}$ , where $N^{(s)}$ is the number of surfels at generation step $s$ . We also maintain an octree to quickly retrieve surfels based on their geometry. Next, we describe how to read from and write to this memory. + +Reading from the memory. To retrieve the most relevant past views from the memory $\mathcal{V}^{(s)}$ for the novel cameras $\{\mathbf{c}_{T + m}\}_{m = 1}^{M}$ , we first compute their average pose $\bar{\mathbf{c}}_s\in$ SE(3) as a reference (refer to the supplementary material). Then we render the surfels $S^{(s)}$ from the averaged camera pose $\bar{\mathbf{c}}_s$ . Each surfel is rendered as a splat with its attributes, accounting for relative depth and occlusions, and contributing to the image based on its coverage. The rendered attributes correspond to the indices of the past views. The core intuition is that views observing the largest portion of the scene from the perspective of the novel camera are most relevant for novel view synthesis. To identify these views, we rank past view indices based on their frequency across all rendered pixels. + +We then select the top- $K$ most frequently represented indices $\mathcal{I}^*$ and use them to retrieve the corresponding past views, $\mathcal{V}^* = \{v_t\}_{t\in \mathcal{I}^*}$ , from the memory $\mathcal{V}^{(s)}$ . This subset of views serves as the context for the new view generator, as described in Sec. 3.2. + +To avoid oversampling repeatedly visited regions, we apply a non-maximum suppression algorithm that reduces redundancy in memory and promotes broader scene coverage among the top- $K$ views. During retrieval, we retain only the most frequently referenced view among those with similar poses. During writing to the memory, we merge surfels by comparing their associated camera poses—if two poses are highly similar, we discard the older one. + +Writing to the memory. After generating the new views $\{v_{T + m} = (\mathbf{x}_{T + m},\mathbf{c}_{T + m})\}_{m = 1}^{M}$ , we add them to the memory $\mathcal{V}^{(s)}$ and update the surfel-based memory index $S^{(s)}$ . We use an off-the-shelf point map estimator $\phi$ , such as + +![](images/b99e1556843ed651d81c978e53ceb2f592b8b2f0db98456ca84ceda16c045999.jpg) + +![](images/c670ca2c89097377121eef7670264956dee86f3c8395fb321363f79591f235f5.jpg) +Figure 4. Surfel-Indexed View Memory. Reading procedure renders surfels $S^{(s)}$ with their attributes, containing past view indices as frame indices. We then select the $K$ most frequent frame indices in the rendered image to retrieve relevant past views from $\mathcal{V}^{(s)}$ . Writing procedure estimates geometry of newly generated views $\{\mathbf{x}_{T + m}\}_{m = 1}^{M}$ as surfels and merges them with existing surfels. Frame indices $\{T + m\}_{m = 1}^{M}$ are appended to surfels in these views, and novel views are stored, updating $\mathcal{V}^{(s)} \to \mathcal{V}^{(s + 1)}$ and $S^{(s)} \to S^{(s + 1)}$ . + +CUT3R [32], to estimate their point maps $\{\mathbf{P}_{T + m}^{*}\}_{m = 1}^{M}$ . Specifically, we jointly estimate the point maps for the newly generated views along with the retrieved past views $\nu^{*}$ . This joint estimation ensures that the new point maps are aligned with the coordinate frame of the existing scene geometry in the surfel-based memory index $S^{(s)}$ . + +Next, we convert the estimated point maps to a set of new surfels. For each newly generated frame $t \in \{T + 1, T + 2, \dots, T + M\}$ , we process its point map $\mathbf{P}_t^*$ . Since our approach only requires coarse geometry, we first downsample the point map $\mathbf{P}_t^*$ by a factor $\sigma$ , yielding a smaller point map $\mathbf{P}_t \in \mathbb{R}^{H' \times W' \times 3}$ , where $H' = H / \sigma$ and $W' = W / \sigma$ . For each pixel location $(u, v)$ in the downsampled point map, we have a 3D point $\mathbf{p}_{u,v,t}$ , and we create a new surfel $\mathbf{s}_{k'}$ centered on it. We compute each surfel's normal using the cross-product of displacement vectors to neighboring pixels. For a pixel at 2D location $(u, v)$ in the image grid, we use neighboring pixels to estimate the local surface normal: + +$$ +\mathbf {n} _ {k ^ {\prime}} = \frac {(\mathbf {p} _ {u + 1 , v , t} - \mathbf {p} _ {u - 1 , v , t}) \times (\mathbf {p} _ {u , v + 1 , t} - \mathbf {p} _ {u , v - 1 , t})}{\| (\mathbf {p} _ {u + 1 , v , t} - \mathbf {p} _ {u - 1 , v , t}) \times (\mathbf {p} _ {u , v + 1 , t} - \mathbf {p} _ {u , v - 1 , t}) \|}, +$$ + +where $\mathbf{p}_{u,v,t}$ represents the 3D point at image coordinates $(u,v)$ in frame $t$ . To ensure that the surfels reasonably cover the scene, we use a heuristic to compute the surfel's radius, which is proportional to the depth of the surfel and inversely proportional to the focal length and the cosine of the angle between the surfel normal and the viewing direction: + +$$ +r _ {k ^ {\prime}} = \frac {\frac {1}{2} \mathbf {D} _ {u , v , t} / f _ {t}}{\alpha + (1 - \alpha) \left\| \mathbf {n} _ {k ^ {\prime}} \cdot (\mathbf {p} _ {u , v , t} - \mathbf {O} _ {t}) \right\|}, +$$ + +where $\mathbf{D}_{u,v,t}$ represents the depth at pixel $(u,v)$ in frame $t$ , $\mathbf{O}_t$ is the camera center of frame $t$ , $\mathbf{p}_{u,v,t} - \mathbf{O}_t$ is the viewing direction, and $f_t$ is the focal length at time + +$t$ . The factor $\alpha$ is used to avoid extreme values when $\| \mathbf{n}_{k'} \cdot (\mathbf{p}_{u,v,t} - \mathbf{O}_t) \| \approx 0$ . + +Finally, we check if the index $S^{(s)}$ already contains a surfel $\mathbf{s}_k$ similar to the new surfel $\mathbf{s}_{k'}$ . Two surfels match if their centers are within a distance $d$ and the cosine similarity between their normals is above a threshold $\theta$ . If such a surfel $\mathbf{s}_k$ is found, we add the frame index $t$ to the surfel's set of past view indices: $\mathcal{I}_k \gets \mathcal{I}_k \cup \{t\}$ , and discard $\mathbf{s}_{k'}$ . If not, we set $\mathcal{I}_{k'} = \{t\}$ and add $\mathbf{s}_{k'}$ to the index $S^{(s + 1)}$ . This process transforms the surfel memory from $S^{(s)}$ to $S^{(s + 1)}$ with $N^{(s + 1)}$ surfels for the next generation step. + +# 3.2. Novel view generator + +We use a camera-conditioned image-set generator $\psi$ to generate novel views. The generator $\psi$ takes $K$ retrieved reference views $\mathcal{V}^* = \{v_t = (\mathbf{x}_t, \mathbf{c}_t)\}_{t \in \mathcal{I}^*}$ and the target camera poses $\{\mathbf{c}_{T + m}\}_{m=1}^M$ for the $M$ new frames to be generated and samples the novel views: + +$$ +\left\{\mathbf {x} _ {T + m} \right\} _ {m = 1} ^ {M} \sim \psi \left(\left\{\left(\mathbf {x} _ {t}, \mathbf {c} _ {t}\right) \right\} _ {t \in \mathcal {I} ^ {*}}, \left\{\mathbf {c} _ {T + m} \right\} _ {m = 1} ^ {M}\right), +$$ + +where $\mathcal{I}^* \subseteq \{1, 2, \ldots, T\}$ are the frame indices of the retrieved reference views. + +Specifically, we base our generator on the recent SEVA [46] model. Given the plug-and-play nature of our memory module, VMem can work with other image-set generators as well. We also fine-tune a more efficient version of SEVA that operates with a reduced number of reference frames. We provide more details in Sec. 4.1. + +# 4. Experiments + +We evaluate our method on established benchmarks for camera-conditioned autoregressive view generation from + +![](images/8b005d7cad06624d8cd8a7fd9a7f0a1a3e1fc4303310aaa30fca70b484831716.jpg) +Figure 5. Long sequences with revisitations. We compare our VMem against a baseline without memory that relies solely on the last $K$ frames for context. Each sequence: input images (left), then generated images at selected frames. Our method (top rows) maintains consistency when revisiting observed regions, while the baseline (bottom rows) shows severe inconsistencies across extended sequences. + +
Short-term (50thframe)Long-term (≥200thframe)
LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓
GeoGPT [25]0.44413.35—26.72——0.6749.54—41.87——
Lookout [22]0.37814.43—28.861.2410.8760.65810.51—58.121.1420.924
PhotoNVS [42]0.33315.51—21.76——0.58811.54—41.95——
GenWarp [26]0.43612.030.14429.690.5640.0590.6139.560.08536.400.8500.251
MotionCtrl [35]0.42412.000.14819.250.3410.3560.6059.130.08335.450.7480.609
ViewCrafter [43]0.37716.970.26225.391.5620.2080.5929.740.14834.122.1770.827
SEVA [46] (K=17)0.29318.330.38217.290.2230.1180.45513.980.21623.821.1250.742
VMem (K=4)0.28718.490.40617.120.2190.0390.49313.120.18327.150.8110.499
VMem (K=17)0.29318.330.38217.290.2230.1180.45214.090.22723.560.9820.432
+ +Table 1. Single-view NVs on RealEstate10K. Frames are subsampled with 10-frame intervals. $K$ denotes the number of context views. In the short-term setting, all past frames fit in the context windows for both SEVA [46] and VMem; hence, SEVA [46] and VMem ( $K = 17$ ) yield identical results since VMem uses SEVA as the backbone. VMem ( $K = 4$ ) is fine-tuned on RealEstate10K for this setting, yielding better performance than SEVA [46] and VMem ( $K = 17$ ) in short-term scenarios. VMem surpasses baselines on most metrics in this benchmark. However, its true advantage—spatial consistency—is not fully reflected because RealEstate10K trajectories rarely revisit previously observed areas. We address this limitation in Sec. 4.3. + +a single image, comparing our models to previous open-source approaches (Sec. 4.2). These benchmarks, however, have a critical limitation: camera trajectories rarely revisit previously observed regions, unlike real-world behavior where humans or robots often return to or look at previ + +ously explored areas. To address this limitation, we propose a cycle-trajectory evaluation protocol that extends these benchmarks by making the camera return to the starting position along the same path in reverse order (Sec. 4.3). Additionally, we collect a small dataset with trajectories that + +
MethodLPIPS ↓PSNR ↑SSIM ↑FID ↓\( R_{\text{dist}} \) ↓\( T_{\text{dist}} \) ↓
Look-out [22]0.8098.410.06938.341.2620.341
GenWarp [26]0.50711.130.13432.940.8480.241
MotionCtrl [35]0.5899.070.09626.860.8710.293
ViewCrafter [43]0.40111.820.21724.720.9020.492
SEVA [46] (K=17)0.40111.820.21724.720.9020.492
VMem (K=4)0.39715.720.29724.970.8210.392
VMem (K=17)0.30418.150.37724.180.8920.165
+ +Table 2. Results on cycle trajectories from RealEstate10K. + +
MethodLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
Look-out [22]0.7928.520.0080.7271.499
GenWarp [26]0.52110.270.1290.7851.982
MotionCtrl [35]0.6928.210.0820.8420.129
ViewCrafter [43]0.49413.620.1291.6430.492
VMem (K=4)0.47214.110.1211.2040.387
+ +revisit observed regions for qualitative evaluation in Fig. 5. Finally, we conduct an ablation study in Sec. 4.4. + +# 4.1. Experimental setup + +Implementation details. We use pre-trained SEVA [46] as the generator $\psi$ . SEVA uses a fixed total number of reference and target frames ( $K + M = 21$ ). For our experiments, we use $M = 4$ target views, which leaves $K = 17$ context views. However, this configuration demands substantial computational resources that may be prohibitive for many applications. To address this limitation, we fine-tune a more efficient version with LoRA [13] on the RealEstate10K training split [47] that works with a reduced number of reference frames ( $K = 4$ ) and target views ( $M = 4$ ), and use this efficient version for our main experiments. We demonstrate (Secs. 4.3 and 4.4) that this fine-tuned version, when combined with VMem, achieves comparable performance to the original SEVA while delivering approximately $12\times$ speedup at inference time. We provide more technical details in the supplementary material. + +Datasets. We consider two real-world datasets for novel view generation evaluation: RealEstate10K [47], comprising indoor scene video clips, and Tanks-and-Temples [14], including indoor and outdoor scenes with larger camera motions. For qualitative evaluation, we use in-the-wild images collected from the internet or captured with phone cameras as input views (see Fig. 5). + +Evaluation metrics. We evaluate generated novel view quality based on three factors: (1) overall quality of generated novel views, comparing generated novel view distributions to test set distributions using Fréchet Image Distance (FID) [10]; (2) the ability of the model to preserve image details across views, measuring peak signal-to-noise ratio (PSNR) for pixel differences, along with LPIPS [45] + +Table 3. Results on cycle trajectories from Tanks and Temples. + +
Context views KView RetrievalLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
17Temporal0.47713.920.1880.9760.254
Camera Distance (SEVA [46])0.39715.720.2970.8210.392
Field of View0.37415.750.2920.9110.382
VMem0.30418.150.3770.8920.165
4Temporal0.7947.520.0181.9420.458
Camera Distance0.42213.270.1871.7870.319
Field of View0.42413.110.1921.7820.285
VMem0.38114.820.2750.7930.124
+ +Table 4. Ablation study on cycle trajectories from RealEstate10K. We compare different view retrieval strategies for two context view counts $K$ + +and SSIM [34], as in [22]; and (3) alignment between generated camera poses and ground truth, following [9], where we express estimated camera poses relative to the first frame and normalize translation by the furthest frame. We use DUSt3R [33] to extract poses from generated views. We calculate rotation distance $R_{\mathrm{dist}}$ by comparing ground truth and extracted rotation matrices of each generated sequence: + +$$ +R _ {\mathrm {d i s t}} = \operatorname {a r c c o s} \left(0. 5 \left(\operatorname {t r} \left(\mathbf {R} _ {\mathrm {g e n}} \mathbf {R} _ {\mathrm {g t}} ^ {T}\right) - 1\right)\right), +$$ + +where $\mathbf{R}$ denotes a rotation matrix. We compute translation distance $t_{\mathrm{dist}}$ as: + +$$ +t _ {\text {d i s t}} = \left\| \mathbf {t} _ {\mathrm {g t}} - \mathbf {t} _ {\mathrm {g e n}} \right\| _ {2}. +$$ + +# 4.2. Short- and long-term view generation + +We first evaluate VMem on the benchmark from [22, 42], which defines two scenarios: short-term and long-term novel view generation. Beginning from the first frame of each ground-truth test sequence, the model autoregressively generates images along the ground truth camera trajectories with 10-frame intervals. The short-term scenario evaluates the fifth generated image (50 frames from the initial view), while the long-term scenario evaluates the final image ( $\geq 200$ frames from the initial view). Following the protocol, we only consider samples with more than 200 frames for the long-term scenario and more than 50 frames for the short-term scenario. We report results in Tab. 1. VMem substantially outperforms all baselines on the key metrics, demonstrating superior scene extrapolation. However, the benchmark trajectories rarely revisit observed regions, limiting evaluation of long-term scene consistency. + +# 4.3. Long-term view generation with revisitations + +To evaluate long-term scene consistency when revisiting observed regions—a scenario that rarely occurs in existing benchmarks—we propose a cycle-trajectory evaluation protocol as a proxy for real revisitation scenarios in human and robotic exploration. In this protocol, the model generates frames following trajectories from initial to final poses, then returns along the same path to the starting position. + +![](images/62e0bb877dd5e8d393160ad7baf38cac2c1b2e6a45e87220df9fa32a4ae9ac49.jpg) +Figure 6. Qualitative comparison of the cycle trajectory frames ( $\geq 400$ frames). VMem autoregressively generates new views while maintaining memory of all previously generated frames, ensuring consistency when revisiting previously seen locations. + +![](images/663a1de4bd4cff0bef3258d5340cfd931077a74cff7452a33477436c31db3c70.jpg) + +We evaluate metrics every 10 frames on the return trajectories, using all test sequences. Following this protocol, we report quantitative results in Tab. 2 and qualitative comparisons in Fig. 6. VMem outperforms all baselines across all metrics in cycle-trajectory evaluation, demonstrating superior visual consistency when revisiting previously observed regions. + +To validate generalization, we evaluate on Tanks-and-Temples containing indoor and outdoor scenes with more dynamic camera trajectories. We use all six advanced scenes and apply the same cycle-trajectory protocol using the first 50 frames, evaluating every frame without temporal subsampling. Results are in Tab. 3. VMem consistently outperforms all baselines across in-domain (Tab. 2) and out-of-domain (Tab. 3) evaluations on the majority of metrics. + +We also qualitatively evaluate VMem across diverse in-the-wild scenes in Fig. 5, demonstrating generation of diverse, consistent, high-quality long videos on in-the-wild scenarios. + +# 4.4. Ablation study + +We ablate our surfel-based memory indexing for context view retrieval in VMem. We evaluate our approach in the cycle-trajectory setting against three alternative retrieval strategies: (1) temporal-based retrieval, selecting the most recent $K$ views from memory, (2) camera distance-based retrieval, selecting top $K$ views with cameras closest to the target view camera—also used by SEVA [46], our backbone generator, and (3) field-of-view-based retrieval, selecting top $K$ views with highest field-of-view overlap with the target view. We also evaluate performance trade-offs between our lightweight, fine-tuned generator using $K = 4$ context views versus original SEVA with $K = 17$ context views. We report results in Tab. 4. + +Our findings show VMem consistently improves generator performance across both numbers of context views. Improvement is most pronounced with fewer context views $(K = 4)$ , highlighting the effectiveness of our surfel-based memory index. + +Computational efficiency. A key practical advantage of our approach is its computational efficiency. VMem with only $K = 4$ views achieves $12 \times$ speed improvement over the original SEVA while recovering most of the performance achieved with $K = 17$ views. On an RTX 4090 GPU, our method generates frames in 4.2 seconds compared to 50 seconds for SEVA, representing a step towards real-time interactive scene exploration. + +# 5. Conclusion + +We introduced Surfel-Indexed View Memory (VMem), a novel plug-and-play memory module for long-term autoregressive scene generation using a surfel-indexed view memory. By anchoring past views to a surfel-based representation of the scene and retrieving the most relevant ones, our approach improves long-term scene consistency while reducing computational costs. Experiments on long-term scene synthesis benchmarks demonstrate that VMem outperforms existing methods in scene coherence while enabling the use of fewer context views, leading to significantly faster generation. Our work advances scalable video generation, with applications in virtual reality, gaming, and other domains. + +Acknowledgments. The authors of this work are supported by Clarendon scholarship, ERC 101001212-UNION, and AIMS EP/S024050/1. The authors would like to thank Xingyi Yang, Zeren Jiang, Junlin Han, Zhongrui Gui for their insightful feedback. + +# References + +[1] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. SF3D: Stable fast 3D mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 2 +[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2 +[4] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction, 2023. arXiv:2304.06714 [cs]. +[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. +[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[7] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2 +[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. Advances in Neural Information Processing Systems, 2024. 2, 3 +[9] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 7 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[11] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7909-7920, 2023. 1, 2 +[12] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2 + +[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 7 +[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7 +[15] Jing Yu Koh, Harsh Agrawal, Dhruv Batra, Richard Tucker, Austin Waters, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Simple and effective synthesis of indoor 3d scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1169-1178, 2023. 1, 2 +[16] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 1, 2 +[17] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. RealFusion: 360 reconstruction of any object from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[18] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 1, 2 +[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2 +[20] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Rocktäschel. Genie 2: A large-scale foundation world model, 2024. 1, 2, 3 +[21] Stefan Popov, Amit Raj, Michael Krainin, Yanzhen Li, William T. Freeman, and Michael Rubinstein. Camctrl3d: Single-image scene exploration with precise 3d camera control, 2025. 1, 2 +[22] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 2, 3, 4, 6, 7 +[23] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed + +world-consistent video generation with precise camera control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3 +[24] Chris Rockwell, David F Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14104-14113, 2021. 1, 2 +[25] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14356-14366, 2021. 2, 3, 4, 6 +[26] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 1, 2, 3, 6, 7 +[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv.cs, abs/2310.15110, 2023. 2 +[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arxiv, 2024. 2 +[29] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[30] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2 +[31] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 3 +[32] Qianqian Wang*, Yifei Zhang*, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In CVPR, 2025. 5 +[33] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 7 +[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[35] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. ACM SIGGRAPH 2024 Conference Papers, 2023. 2, 3, 4, 6, 7 + +[36] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 1, 2 +[37] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21551-21561, 2024. 2 +[38] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2 +[39] Zeqi Xiao, Yushi Lan, Yifan Zhou, Wenqi Ouyang, Shuai Yang, Yanhong Zeng, and Xingang Pan. Worldmem: Long-term consistent world simulation with memory, 2025. 3 +[40] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T. Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, and Charles Herrmann. Wonderjourney: Going from anywhere to everywhere. arXiv.cs, abs/2312.03884, 2023. 2 +[41] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T. Freeman, and Jiajun Wu. WonderWorld: interactive 3D scene generation from a single image. arXiv, 2406.09394, 2025. 2 +[42] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7094-7104, 2023. 2, 3, 4, 6, 7 +[43] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 6, 7 +[44] Shangjin Zhai, Zhichao Ye, Jialin Liu, Weijian Xie, Jiaqi Hu, Zhen Peng, Hua Xue, Danpeng Chen, Xiaomeng Wang, Lei Yang, Nan Wang, Haomin Liu, and Guofeng Zhang. Stargen: A spatiotemporal autoregression framework with video diffusion model for scalable and controllable scene generation, 2025. 3 +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[46] Jensen (Jinghao) Zhou, Hang Gao, Vikram Voleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rupprecht, and Varun Jampani. Stable virtual camera: Generative view synthesis with diffusion models. arXiv preprint arXiv:2503.14489, 2025. 3, 5, 6, 7, 8 +[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 7 \ No newline at end of file diff --git a/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/images.zip b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d5338742026252b8c4f4577dc0098a6259df31a1 --- /dev/null +++ b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8047c8a0efc20bfd9d30c1b8d97effbda94974cd4c41b2581854519fc629c7bd +size 953692 diff --git a/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/layout.json b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c10c5e13ffb6ccd80ab8ea80c2ad6b3a595fea60 --- /dev/null +++ b/2025/VMem_ Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory/layout.json @@ -0,0 +1,9006 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 538, + 140 + ], + "type": "text", + "content": "VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 161, + 443, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 161, + 443, + 176 + ], + "spans": [ + { + "bbox": [ + 168, + 161, + 443, + 176 + ], + "type": "text", + "content": "Runjia Li Philip Torr Andrea Vedaldi Tomas Jakab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 254, + 180, + 357, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 254, + 180, + 357, + 194 + ], + "spans": [ + { + "bbox": [ + 254, + 180, + 357, + 194 + ], + "type": "text", + "content": "University of Oxford" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 263, + 198, + 347, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 198, + 347, + 209 + ], + "spans": [ + { + "bbox": [ + 263, + 198, + 347, + 209 + ], + "type": "text", + "content": "v-mem.github.io" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 225, + 255, + 377 + ], + "blocks": [ + { + "bbox": [ + 56, + 225, + 255, + 377 + ], + "lines": [ + { + "bbox": [ + 56, + 225, + 255, + 377 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 255, + 377 + ], + "type": "image", + "image_path": "89c810002b4c80233657cf00ae17e94579b7a2e98175a33a5b6702ab0dc5df6a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 382, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 382, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 555, + 415 + ], + "type": "text", + "content": "Figure 1. VMem enables autoregressive scene generation from a single image along user-defined trajectories. The green region shows results with the proposed memory module, maintaining coherence when generating previously seen parts of the scene. The red region, without memory, exhibits degradation highlighted with red ellipses, demonstrating VMem is effective for consistent scene generation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 259, + 225, + 555, + 371 + ], + "blocks": [ + { + "bbox": [ + 259, + 225, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 259, + 225, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 259, + 225, + 555, + 371 + ], + "type": "image", + "image_path": "8af08eb47fd8fc071039d3480bd09ade2f068998d8cada51a12627fbb3ac7059.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 427, + 200, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 427, + 200, + 439 + ], + "spans": [ + { + "bbox": [ + 152, + 427, + 200, + 439 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 462, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 297, + 700 + ], + "type": "text", + "content": "We propose a novel memory module for building video generators capable of interactively exploring environments. Previous approaches have achieved similar results either by out-painting 2D views of a scene while incrementally reconstructing its 3D geometry—which quickly accumulates errors—or by using video generators with a short context window, which struggle to maintain scene coherence over the long term. To address these limitations, we introduce Surfel-Indexed View Memory (VMem), a memory module that remembers past views by indexing them geometrically based on the 3D surface elements (surfels) they have observed. VMem enables efficient retrieval of the most relevant past views when generating new ones. By focusing only on these relevant views, our method produces consistent explorations of imagined environments at a fraction of the computational cost required to use all past views as context. We evaluate our approach on challenging long-term scene synthesis benchmarks and demonstrate superior performance compared to existing methods in maintaining scene coherence and camera control." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 427, + 394, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 427, + 394, + 439 + ], + "spans": [ + { + "bbox": [ + 315, + 427, + 394, + 439 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "content": "We consider the problem of generating long videos that explore an imagined space following a camera path specified interactively by the user. In this paradigm, the user tells the model which camera path to follow for the next few frames, observes the generated content, and then decides where to explore next based on what they have seen. For example, exploring a house may involve visiting the kitchen, the living room, and the bathroom, eventually returning to the kitchen. Throughout the video, the scene must remain consistent, ensuring that the kitchen looks the same upon return. Generating such videos is essential for immersive applications such as games, where players can navigate generated worlds. However, even recent large-scale interactive video generators such as Google's Genie 2 [20] struggle to achieve this goal." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "This problem has so far been addressed by two types of methods. First, outpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 43] iterate between generating new 2D views of the scene and estimating its 3D geometry. They use the estimated geometry to partially render a new viewpoint and then employ an outpainting model to complete the missing parts, thus adding one more image to the col" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25690" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "lection. However, errors in outpainting, 3D reconstruction, and stitching accumulate over time, leading to severe degradation of the generated content after a short while." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 294, + 216 + ], + "type": "text", + "content": "Second, multi-view/video-based methods [20, 22, 25, 35, 42] condition novel view generation on previous views using a geometry-free approach that does not explicitly estimate the scene geometry. While this avoids the accumulation of errors in the reconstructed 3D scene geometry, it comes at a high computational cost, limiting the number of conditioning views to a small context window of recent frames. This constraint hurts the long-term consistency of the generated images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 219, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 294, + 337 + ], + "type": "text", + "content": "In this work, we revisit the second class of methods and propose conditioning not on the most recently generated views, but on the most relevant ones for generating the next image, thereby maintaining a high degree of consistency within a limited computational budget. Given a novel viewpoint of the scene, the most relevant past views are those that have already observed the parts of the scene currently being generated. This implies that, for each part of the scene, we must remember which views have previously observed it and retrieve them from memory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 530 + ], + "type": "text", + "content": "To achieve this, we introduce Surfel-Indexed View Memory, abbreviated as VMem, a memory module that anchors previous views to the 3D surface elements they observe. Given a new viewpoint, we retrieve the past views that best capture the currently observed surfaces and use them to condition novel view generation. To create the memory, we estimate the geometry of each new view using an off-the-shelf point map predictor. This is similar to outpainting methods but, crucially, we do not use the estimated geometry as the final representation of the scene; instead, we use it to construct VMem, which is a memory of past views. We represent the scene geometry using surfels, which are more robust compared to meshes and can represent occlusions compared to point clouds. Each surfel stores in its attributes a set of indices corresponding to past viewpoints that observed it." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 674 + ], + "type": "text", + "content": "To retrieve relevant past viewpoints, we render the surfels with their attributes from the novel viewpoint and splat them onto an image grid. Each pixel in the resulting image then corresponds to a set of viewpoint indices. We select the top " + }, + { + "bbox": [ + 55, + 533, + 295, + 674 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 533, + 295, + 674 + ], + "type": "text", + "content": " most frequently represented viewpoint indices and use them to retrieve past views from the database, where each view is represented by an RGB image and its corresponding camera parameters. A key advantage of our approach over outpainting-based methods is that it does not require highly accurate scene geometry. As long as we successfully retrieve the most relevant past views, our method remains robust." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "By leveraging Surfel-Indexed View Memory, we significantly reduce the memory and computational burden of conditioning on a large number of previous views, as re" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "quired by multi-view/video-based methods, while improving long-term consistency across generated novel views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 98, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 133 + ], + "type": "text", + "content": "Our approach represents a step toward scalable, realistic, and long-term autoregressive scene generation, making the following contributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 134, + 553, + 276 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 134, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 134, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 314, + 134, + 553, + 169 + ], + "type": "text", + "content": "1. We introduce Surfel-Indexed View Memory, a plug-and-play module to index past views geometrically and use them to condition novel view generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 171, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 205 + ], + "type": "text", + "content": "2. We show that our method can generate long-term coherent videos of scenes and outperforms existing approaches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 206, + 553, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 241 + ], + "type": "text", + "content": "3. We demonstrate that Surfel-Indexed View Memory achieves comparable performance with " + }, + { + "bbox": [ + 313, + 206, + 553, + 241 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 206, + 553, + 241 + ], + "type": "text", + "content": " fewer context views, delivering a " + }, + { + "bbox": [ + 313, + 206, + 553, + 241 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 313, + 206, + 553, + 241 + ], + "type": "text", + "content": " speedup." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 242, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 276 + ], + "type": "text", + "content": "4. We validate our approach on challenging benchmarks, outperforming the current open-source state-of-the-art methods." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 292, + 398, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 292, + 398, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 292, + 398, + 304 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "type": "text", + "content": "Novel view synthesis (NVS) is a challenging and ill-posed problem that can be categorized into two main categories: view interpolation methods [3-6, 8, 19, 30, 37], which generate views close to given inputs, and view extrapolation or autoregressive view generation, where novel views extend significantly beyond the original scene, introducing substantial new content. The latter is particularly difficult from single images. Between these lie single-view scene reconstruction methods, which succeed mainly for single-object scenes [1, 12, 17, 27, 29, 38] or highly bounded scenes [28]. However, their extrapolation capabilities remain limited. Most relevant to our work are single-image view extrapolation models, which fall into two categories: those incorporating explicit geometric modeling with inpainting, and those based directly on image or video generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Inpainting-based view extrapolation. Inpainting-based methods [7, 11, 15, 16, 18, 21, 24, 26, 36, 40, 41, 43] use pre-trained 3D reconstruction models to generate 3D representations—such as meshes, point clouds, or Gaussian splats—from images. These representations are reprojected into novel views, where 2D inpainting fills missing regions. SceneScape [7] reconstructs meshes from images using pretrained depth estimators. Diffusion-based inpainting then completes the projected novel view, which is reprojected back to refine and extend the mesh. This process is iterated to generate novel views. MultiDiff [18] uses depth estimators to warp reference images into multiple novel views, training diffusion models to inpaint missing regions simultaneously. CamCtrl3D [21] follows a similar approach but adds ray maps to condition diffusion models for view synthesis. ViewCrafter [43] uses pre-trained pointmap estimators to create point clouds and trains video diffusion mod" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25691" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 556, + 216 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 556, + 216 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 556, + 216 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 556, + 216 + ], + "type": "image", + "image_path": "50d1e06551806acb593625cda09e71f750c5a9fdc26e29e0e9b745fe1c2886af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Method. Given target camera viewpoints " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": ", we query our Surfel-Indexed View Memory to retrieve the most relevant " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " past views " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^* \\subset \\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^* = \\{v_t\\}_{t = 1}^K" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " as references. Retrieved reference images " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " along with Plücker embeddings of both reference camera poses " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " and target camera poses " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " are fed into generator " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " to synthesize novel views " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": ". After generation, the surfel-indexed memory is updated " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "S^{(s)} \\to S^{(s + 1)}" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " by appending new view indices " + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{T + m\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 54, + 221, + 555, + 277 + ], + "type": "text", + "content": " to existing surfels or creating new surfels based on geometry of the generated views. This is repeated autoregressively, enabling long-term consistent generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 298, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 294, + 334 + ], + "type": "text", + "content": "els for inpainting. GenWarp [26] warps 2D coordinate representations of reference views into novel views and uses them to condition generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 335, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 295, + 432 + ], + "type": "text", + "content": "While these methods improve geometric consistency, they are susceptible to errors in depth or pointmap estimation, which propagate distortions across generated views. Once 3D representations are constructed, inaccuracies from depth or pointmap estimation become difficult to correct. Moreover, scaling to large scenes is computationally intensive, as storing and processing high-fidelity 3D representations requires substantial memory." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 438, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 295, + 714 + ], + "type": "text", + "content": "Multi-view-based view extrapolation. Another category [8, 22, 25, 31, 35, 42, 46] avoids explicitly modeling 3D scene geometry. Instead, these approaches generate novel views by conditioning on previously rendered or generated views. GeoGPT [25] uses autoregressive likelihood models to synthesize novel views from single images. Building on this, LookOut [22] improves consistency by generating view sequences along predefined trajectories, conditioning on up to two preceding views. Other methods [31, 42] enhance novel view synthesis by employing cross-view attention or enforcing epipolar constraints within diffusion models. MotionCtrl [35] leverages geometric priors from pre-trained video diffusion models to synthesize camera-controllable videos. These methods are trajectory-based, requiring input views to follow continuous spatial paths. In contrast, image-set models like CAT3D [8] and more recently SEVA [46] synthesize novel views from sparse and unordered input images, without trajectory constraints. However, these methods are computationally expensive, with " + }, + { + "bbox": [ + 56, + 438, + 295, + 714 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 56, + 438, + 295, + 714 + ], + "type": "text", + "content": " complexity for attention, limiting conditioning views to small context windows containing only recently generated frames. This limitation impacts long-term consistency of generated images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 298, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 418 + ], + "type": "text", + "content": "Several concurrent works explore memory-based approaches for scene generation. Genie [2, 20] maintains memory through recurrent state features. Gen3C [23] conditions on stored point clouds like ViewCrafter [43], inheriting inpainting-based limitations. StarGen [44] and WorldMem [39] use distance- and field-of-view-based spatial memory retrieval, which can recover spatially correlated views but lack geometric reasoning for occlusions. In contrast, our surfel-based memory explicitly models occlusions and provides more principled view selection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 430, + 370, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 370, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 370, + 441 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1 \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": " be an RGB image of a scene that we wish to explore, and let " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": " be a sequence of camera parameters specifying a path through the scene. Our goal is to generate a corresponding video, i.e., a sequence of images " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": " is the given input image, and each subsequent frame " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": " is consistent with the previous ones and reflects the specified cameras " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": ". Moreover, we aim to do so autoregressively, generating " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": " novel views at a time for the given cameras (with " + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 449, + 555, + 605 + ], + "type": "text", + "content": " kept small to ensure interactivity) while ensuring consistency with previously generated views. This enables interactive scene exploration, where the user can specify the next camera positions at each step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": "Specifically, we work with a sequence of views, where each view " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "v_{t} = (\\mathbf{x}_{t},\\mathbf{c}_{t})" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": " consists of an image " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": " and its corresponding camera " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": ". At generation step " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "s(s\\geq 0)" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": ", we have generated " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "T = sM" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": " frames so far. The goal is to generate " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": " new RGB images " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": " for the next target camera parameters " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": ", given the previously generated views " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)} = \\{v_1,v_2,\\dots ,v_T\\}" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": ". After generation, we update " + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "inline_equation", + "content": "T\\gets T + M" + }, + { + "bbox": [ + 313, + 606, + 556, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "text", + "content": "A challenge with this approach is that the context " + }, + { + "bbox": [ + 325, + 700, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)}" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25692" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 145, + 159 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 145, + 159 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 145, + 159 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 145, + 159 + ], + "type": "image", + "image_path": "4647dc6736bfed6b6ec55a4c3bbe3d87ceb72cc064bece0ade8f8d9b8d1a9877.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 167, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 555, + 201 + ], + "type": "text", + "content": "Figure 3. Surfel-based memory index. Each surfel stores indices of views that observed it. We color-code each surfel by contributing view indices. This spatial index enables retrieval of relevant past views: when generating a novel view, we identify visible surfels from the target viewpoint and retrieve views that previously observed those same regions, naturally accounting for occlusion." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 74, + 247, + 159 + ], + "blocks": [ + { + "bbox": [ + 154, + 74, + 247, + 159 + ], + "lines": [ + { + "bbox": [ + 154, + 74, + 247, + 159 + ], + "spans": [ + { + "bbox": [ + 154, + 74, + 247, + 159 + ], + "type": "image", + "image_path": "948869d140c2846ab916e8dff078913fffb40e97494b87e42bf2d9155faffcd2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 253, + 70, + 347, + 159 + ], + "blocks": [ + { + "bbox": [ + 253, + 70, + 347, + 159 + ], + "lines": [ + { + "bbox": [ + 253, + 70, + 347, + 159 + ], + "spans": [ + { + "bbox": [ + 253, + 70, + 347, + 159 + ], + "type": "image", + "image_path": "a8e23f819452a8bfd6d5f05fa64a02b27714cb1cfcf9f5db453b9f523e4ac9b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 353, + 78, + 442, + 159 + ], + "blocks": [ + { + "bbox": [ + 353, + 78, + 442, + 159 + ], + "lines": [ + { + "bbox": [ + 353, + 78, + 442, + 159 + ], + "spans": [ + { + "bbox": [ + 353, + 78, + 442, + 159 + ], + "type": "image", + "image_path": "c452c93a1eeba2d661c4170b7f035f252adffd53184905967588730517409b88.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 447, + 78, + 536, + 159 + ], + "blocks": [ + { + "bbox": [ + 447, + 78, + 536, + 159 + ], + "lines": [ + { + "bbox": [ + 447, + 78, + 536, + 159 + ], + "spans": [ + { + "bbox": [ + 447, + 78, + 536, + 159 + ], + "type": "image", + "image_path": "58e42b42eb0cab221ac9d4834cec7e1c75f5a75491c2648c159b3df934868e06.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "text", + "content": "grows unbounded over time. Video generators addressing this problem [22, 25, 35, 42] typically consider a fixed-length subset of " + }, + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "text", + "content": ", conditioning their generation only on the most recent " + }, + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "text", + "content": " views " + }, + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "inline_equation", + "content": "\\{v_{T - L + 1},\\ldots ,v_T\\}" + }, + { + "bbox": [ + 54, + 221, + 294, + 293 + ], + "type": "text", + "content": ". This limitation results in severe inconsistencies when the generated sequence extends beyond the (small) context window." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 293, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 390 + ], + "type": "text", + "content": "We solve this problem by dynamically retrieving the most relevant subset of past views, denoted as " + }, + { + "bbox": [ + 54, + 293, + 295, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^* \\subseteq \\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 54, + 293, + 295, + 390 + ], + "type": "text", + "content": ". To achieve this, we introduce an efficient data structure, the Surfel-Indexed View Memory, that stores and retrieves past views based on their approximate 3D geometry. We first describe the Surfel-Indexed View Memory (Sec. 3.1) and then introduce the novel view generator (Sec. 3.2). An overview of the method is provided in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 397, + 212, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 212, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 212, + 410 + ], + "type": "text", + "content": "3.1. Surfel-indexed view memory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "text", + "content": "Consider the problem of generating the next views " + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "text", + "content": " of the scene as seen from the cameras " + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "text", + "content": ". The generator must consider the previously generated views and ensure " + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "text", + "content": " are consistent with those, while generating new content as needed. However, not all past views are equally relevant to the novel views " + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 415, + 295, + 582 + ], + "type": "text", + "content": ". We prioritize past views that have likely observed the largest portion of the scene currently being generated. For example, if the current locale is surrounded by walls, parts of the scene behind those walls are likely less relevant for generating a new view of that locale. Conversely, views that share a similar visible region with the target view can provide more useful information for generating its content." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 583, + 295, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 295, + 678 + ], + "type": "text", + "content": "This principle drives our design for a view memory module. We maintain a coarse model of the scene geometry using surfels (visualized in Fig. 3), simple surface primitives that account for occlusion. Each surfel stores the indices of the past views that observed it. To generate a new view, the surfels visible from the new viewpoint are retrieved, using the associated indices to vote for which views to consider for generation. This process is illustrated in Fig. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 679, + 241, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 679, + 241, + 691 + ], + "spans": [ + { + "bbox": [ + 67, + 679, + 241, + 691 + ], + "type": "text", + "content": "Specifically, we define a surfel as the tuple" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 125, + 701, + 225, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 701, + 225, + 714 + ], + "spans": [ + { + "bbox": [ + 125, + 701, + 225, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {k} = \\left(\\mathbf {p} _ {k}, \\mathbf {n} _ {k}, r _ {k}, \\mathcal {I} _ {k}\\right),", + "image_path": "3347e99b1163f57365edd75dee75375b210acdb6ea65f0e836358daaed4b6577.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": " denotes the surfel's 3D position, " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_k \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": " is its surface normal, " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "r_k \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": " is the surfel radius, and " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_k \\subseteq \\{1,2,\\dots,T\\}" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": " is a subset of past view indices that observed the surfel. The surfel-based memory indexing of past views " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": ", encapsulating the scene, is represented as a set of surfels " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^{(s)} = \\{\\mathbf{s}_k\\}_{k=1}^{N^{(s)}}" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "N^{(s)}" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": " is the number of surfels at generation step " + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 221, + 555, + 331 + ], + "type": "text", + "content": ". We also maintain an octree to quickly retrieve surfels based on their geometry. Next, we describe how to read from and write to this memory." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "text", + "content": "Reading from the memory. To retrieve the most relevant past views from the memory " + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "text", + "content": " for the novel cameras " + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "text", + "content": ", we first compute their average pose " + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{c}}_s\\in" + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "text", + "content": " SE(3) as a reference (refer to the supplementary material). Then we render the surfels " + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "inline_equation", + "content": "S^{(s)}" + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "text", + "content": " from the averaged camera pose " + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{c}}_s" + }, + { + "bbox": [ + 313, + 336, + 555, + 503 + ], + "type": "text", + "content": ". Each surfel is rendered as a splat with its attributes, accounting for relative depth and occlusions, and contributing to the image based on its coverage. The rendered attributes correspond to the indices of the past views. The core intuition is that views observing the largest portion of the scene from the perspective of the novel camera are most relevant for novel view synthesis. To identify these views, we rank past view indices based on their frequency across all rendered pixels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": "We then select the top-" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": " most frequently represented indices " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^*" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": " and use them to retrieve the corresponding past views, " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^* = \\{v_t\\}_{t\\in \\mathcal{I}^*}" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": ", from the memory " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": ". This subset of views serves as the context for the new view generator, as described in Sec. 3.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 564, + 554, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 660 + ], + "type": "text", + "content": "To avoid oversampling repeatedly visited regions, we apply a non-maximum suppression algorithm that reduces redundancy in memory and promotes broader scene coverage among the top-" + }, + { + "bbox": [ + 313, + 564, + 554, + 660 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 564, + 554, + 660 + ], + "type": "text", + "content": " views. During retrieval, we retain only the most frequently referenced view among those with similar poses. During writing to the memory, we merge surfels by comparing their associated camera poses—if two poses are highly similar, we discard the older one." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Writing to the memory. After generating the new views " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{v_{T + m} = (\\mathbf{x}_{T + m},\\mathbf{c}_{T + m})\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ", we add them to the memory " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " and update the surfel-based memory index " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "S^{(s)}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ". We use an off-the-shelf point map estimator " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ", such as" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25693" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 547, + 148 + ], + "type": "image", + "image_path": "b99e1556843ed651d81c978e53ceb2f592b8b2f0db98456ca84ceda16c045999.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 149, + 547, + 228 + ], + "blocks": [ + { + "bbox": [ + 62, + 149, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 62, + 149, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 149, + 547, + 228 + ], + "type": "image", + "image_path": "c670ca2c89097377121eef7670264956dee86f3c8395fb321363f79591f235f5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": "Figure 4. Surfel-Indexed View Memory. Reading procedure renders surfels " + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "inline_equation", + "content": "S^{(s)}" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": " with their attributes, containing past view indices as frame indices. We then select the " + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": " most frequent frame indices in the rendered image to retrieve relevant past views from " + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)}" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": ". Writing procedure estimates geometry of newly generated views " + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{T + m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": " as surfels and merges them with existing surfels. Frame indices " + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "inline_equation", + "content": "\\{T + m\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": " are appended to surfels in these views, and novel views are stored, updating " + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(s)} \\to \\mathcal{V}^{(s + 1)}" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "inline_equation", + "content": "S^{(s)} \\to S^{(s + 1)}" + }, + { + "bbox": [ + 55, + 237, + 554, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "text", + "content": "CUT3R [32], to estimate their point maps " + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_{T + m}^{*}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "text", + "content": ". Specifically, we jointly estimate the point maps for the newly generated views along with the retrieved past views " + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "inline_equation", + "content": "\\nu^{*}" + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "text", + "content": ". This joint estimation ensures that the new point maps are aligned with the coordinate frame of the existing scene geometry in the surfel-based memory index " + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "inline_equation", + "content": "S^{(s)}" + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": "Next, we convert the estimated point maps to a set of new surfels. For each newly generated frame " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "t \\in \\{T + 1, T + 2, \\dots, T + M\\}" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": ", we process its point map " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t^*" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": ". Since our approach only requires coarse geometry, we first downsample the point map " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t^*" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": " by a factor " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": ", yielding a smaller point map " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_t \\in \\mathbb{R}^{H' \\times W' \\times 3}" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "H' = H / \\sigma" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "W' = W / \\sigma" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": ". For each pixel location " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": " in the downsampled point map, we have a 3D point " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{u,v,t}" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": ", and we create a new surfel " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{k'}" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": " centered on it. We compute each surfel's normal using the cross-product of displacement vectors to neighboring pixels. For a pixel at 2D location " + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 375, + 295, + 518 + ], + "type": "text", + "content": " in the image grid, we use neighboring pixels to estimate the local surface normal:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 526, + 292, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 526, + 292, + 554 + ], + "spans": [ + { + "bbox": [ + 59, + 526, + 292, + 554 + ], + "type": "interline_equation", + "content": "\\mathbf {n} _ {k ^ {\\prime}} = \\frac {(\\mathbf {p} _ {u + 1 , v , t} - \\mathbf {p} _ {u - 1 , v , t}) \\times (\\mathbf {p} _ {u , v + 1 , t} - \\mathbf {p} _ {u , v - 1 , t})}{\\| (\\mathbf {p} _ {u + 1 , v , t} - \\mathbf {p} _ {u - 1 , v , t}) \\times (\\mathbf {p} _ {u , v + 1 , t} - \\mathbf {p} _ {u , v - 1 , t}) \\|},", + "image_path": "6047b0e204a3ca96e7e734de99ea1b3a5106acff6b6211325d1134a27468bf9b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{u,v,t}" + }, + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "text", + "content": " represents the 3D point at image coordinates " + }, + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 561, + 295, + 633 + ], + "type": "text", + "content": ". To ensure that the surfels reasonably cover the scene, we use a heuristic to compute the surfel's radius, which is proportional to the depth of the surfel and inversely proportional to the focal length and the cosine of the angle between the surfel normal and the viewing direction:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 641, + 262, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 641, + 262, + 670 + ], + "spans": [ + { + "bbox": [ + 88, + 641, + 262, + 670 + ], + "type": "interline_equation", + "content": "r _ {k ^ {\\prime}} = \\frac {\\frac {1}{2} \\mathbf {D} _ {u , v , t} / f _ {t}}{\\alpha + (1 - \\alpha) \\left\\| \\mathbf {n} _ {k ^ {\\prime}} \\cdot (\\mathbf {p} _ {u , v , t} - \\mathbf {O} _ {t}) \\right\\|},", + "image_path": "36788a0a6877bba216c63543df79a981e4d840c2c9e826832b996aaeea18980f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{u,v,t}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " represents the depth at pixel " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the camera center of frame " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{u,v,t} - \\mathbf{O}_t" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the viewing direction, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the focal length at time" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "text", + "content": ". The factor " + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "text", + "content": " is used to avoid extreme values when " + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{n}_{k'} \\cdot (\\mathbf{p}_{u,v,t} - \\mathbf{O}_t) \\| \\approx 0" + }, + { + "bbox": [ + 313, + 303, + 553, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": "Finally, we check if the index " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "S^{(s)}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " already contains a surfel " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " similar to the new surfel " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{k'}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": ". Two surfels match if their centers are within a distance " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " and the cosine similarity between their normals is above a threshold " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": ". If such a surfel " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_k" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " is found, we add the frame index " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " to the surfel's set of past view indices: " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_k \\gets \\mathcal{I}_k \\cup \\{t\\}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": ", and discard " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{k'}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": ". If not, we set " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{k'} = \\{t\\}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " and add " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{k'}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " to the index " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "S^{(s + 1)}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": ". This process transforms the surfel memory from " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "S^{(s)}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "S^{(s + 1)}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "inline_equation", + "content": "N^{(s + 1)}" + }, + { + "bbox": [ + 313, + 327, + 554, + 436 + ], + "type": "text", + "content": " surfels for the next generation step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 445, + 436, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 436, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 436, + 458 + ], + "type": "text", + "content": "3.2. Novel view generator" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "content": "We use a camera-conditioned image-set generator " + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "content": " to generate novel views. The generator " + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "content": " retrieved reference views " + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^* = \\{v_t = (\\mathbf{x}_t, \\mathbf{c}_t)\\}_{t \\in \\mathcal{I}^*}" + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "content": " and the target camera poses " + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_{T + m}\\}_{m=1}^M" + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 464, + 554, + 523 + ], + "type": "text", + "content": " new frames to be generated and samples the novel views:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 533, + 537, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 533, + 537, + 549 + ], + "spans": [ + { + "bbox": [ + 331, + 533, + 537, + 549 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {x} _ {T + m} \\right\\} _ {m = 1} ^ {M} \\sim \\psi \\left(\\left\\{\\left(\\mathbf {x} _ {t}, \\mathbf {c} _ {t}\\right) \\right\\} _ {t \\in \\mathcal {I} ^ {*}}, \\left\\{\\mathbf {c} _ {T + m} \\right\\} _ {m = 1} ^ {M}\\right),", + "image_path": "dbad96a119db4b33a997ba8facc795fbb405b9d62c263eef10b0bb2cb29b85ee.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^* \\subseteq \\{1, 2, \\ldots, T\\}" + }, + { + "bbox": [ + 313, + 558, + 553, + 581 + ], + "type": "text", + "content": " are the frame indices of the retrieved reference views." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": "Specifically, we base our generator on the recent SEVA [46] model. Given the plug-and-play nature of our memory module, VMem can work with other image-set generators as well. We also fine-tune a more efficient version of SEVA that operates with a reduced number of reference frames. We provide more details in Sec. 4.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 669, + 395, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 395, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 395, + 682 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We evaluate our method on established benchmarks for camera-conditioned autoregressive view generation from" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25694" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 541, + 376 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 541, + 376 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 541, + 376 + ], + "type": "image", + "image_path": "8b005d7cad06624d8cd8a7fd9a7f0a1a3e1fc4303310aaa30fca70b484831716.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 387, + 555, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 555, + 420 + ], + "type": "text", + "content": "Figure 5. Long sequences with revisitations. We compare our VMem against a baseline without memory that relies solely on the last " + }, + { + "bbox": [ + 55, + 387, + 555, + 420 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 387, + 555, + 420 + ], + "type": "text", + "content": " frames for context. Each sequence: input images (left), then generated images at selected frames. Our method (top rows) maintains consistency when revisiting observed regions, while the baseline (bottom rows) shows severe inconsistencies across extended sequences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 430, + 541, + 559 + ], + "blocks": [ + { + "bbox": [ + 70, + 430, + 541, + 559 + ], + "lines": [ + { + "bbox": [ + 70, + 430, + 541, + 559 + ], + "spans": [ + { + "bbox": [ + 70, + 430, + 541, + 559 + ], + "type": "table", + "html": "
Short-term (50thframe)Long-term (≥200thframe)
LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓LPIPS↓PSNR↑SSIM↑FID↓Rdist↓Tdist↓
GeoGPT [25]0.44413.35—26.72——0.6749.54—41.87——
Lookout [22]0.37814.43—28.861.2410.8760.65810.51—58.121.1420.924
PhotoNVS [42]0.33315.51—21.76——0.58811.54—41.95——
GenWarp [26]0.43612.030.14429.690.5640.0590.6139.560.08536.400.8500.251
MotionCtrl [35]0.42412.000.14819.250.3410.3560.6059.130.08335.450.7480.609
ViewCrafter [43]0.37716.970.26225.391.5620.2080.5929.740.14834.122.1770.827
SEVA [46] (K=17)0.29318.330.38217.290.2230.1180.45513.980.21623.821.1250.742
VMem (K=4)0.28718.490.40617.120.2190.0390.49313.120.18327.150.8110.499
VMem (K=17)0.29318.330.38217.290.2230.1180.45214.090.22723.560.9820.432
", + "image_path": "aa5fd9009d3151efcd0b288893ad50e99875660a6a83c2c7d2080a149440249d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "lines": [ + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "text", + "content": "Table 1. Single-view NVs on RealEstate10K. Frames are subsampled with 10-frame intervals. " + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "text", + "content": " denotes the number of context views. In the short-term setting, all past frames fit in the context windows for both SEVA [46] and VMem; hence, SEVA [46] and VMem (" + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "text", + "content": ") yield identical results since VMem uses SEVA as the backbone. VMem (" + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "text", + "content": ") is fine-tuned on RealEstate10K for this setting, yielding better performance than SEVA [46] and VMem (" + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 54, + 567, + 555, + 634 + ], + "type": "text", + "content": ") in short-term scenarios. VMem surpasses baselines on most metrics in this benchmark. However, its true advantage—spatial consistency—is not fully reflected because RealEstate10K trajectories rarely revisit previously observed areas. We address this limitation in Sec. 4.3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 652, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 294, + 712 + ], + "type": "text", + "content": "a single image, comparing our models to previous open-source approaches (Sec. 4.2). These benchmarks, however, have a critical limitation: camera trajectories rarely revisit previously observed regions, unlike real-world behavior where humans or robots often return to or look at previ" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 652, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 555, + 712 + ], + "type": "text", + "content": "ously explored areas. To address this limitation, we propose a cycle-trajectory evaluation protocol that extends these benchmarks by making the camera return to the starting position along the same path in reverse order (Sec. 4.3). Additionally, we collect a small dataset with trajectories that" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25695" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 292, + 153 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 292, + 153 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 292, + 153 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 292, + 153 + ], + "type": "table", + "html": "
MethodLPIPS ↓PSNR ↑SSIM ↑FID ↓\\( R_{\\text{dist}} \\) ↓\\( T_{\\text{dist}} \\) ↓
Look-out [22]0.8098.410.06938.341.2620.341
GenWarp [26]0.50711.130.13432.940.8480.241
MotionCtrl [35]0.5899.070.09626.860.8710.293
ViewCrafter [43]0.40111.820.21724.720.9020.492
SEVA [46] (K=17)0.40111.820.21724.720.9020.492
VMem (K=4)0.39715.720.29724.970.8210.392
VMem (K=17)0.30418.150.37724.180.8920.165
", + "image_path": "3a7d92503e450a1d4064beb63d870c7c1224affd9f3fc937cf3d7ab2f621a9e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 185, + 282, + 251 + ], + "blocks": [ + { + "bbox": [ + 63, + 162, + 287, + 173 + ], + "lines": [ + { + "bbox": [ + 63, + 162, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 63, + 162, + 287, + 173 + ], + "type": "text", + "content": "Table 2. Results on cycle trajectories from RealEstate10K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 185, + 282, + 251 + ], + "lines": [ + { + "bbox": [ + 69, + 185, + 282, + 251 + ], + "spans": [ + { + "bbox": [ + 69, + 185, + 282, + 251 + ], + "type": "table", + "html": "
MethodLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
Look-out [22]0.7928.520.0080.7271.499
GenWarp [26]0.52110.270.1290.7851.982
MotionCtrl [35]0.6928.210.0820.8420.129
ViewCrafter [43]0.49413.620.1291.6430.492
VMem (K=4)0.47214.110.1211.2040.387
", + "image_path": "f0163aaeb6c12097b6f51d9a8856c60a98cb058fb3db67a4429153e4ca4d87af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 316 + ], + "type": "text", + "content": "revisit observed regions for qualitative evaluation in Fig. 5. Finally, we conduct an ablation study in Sec. 4.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 325, + 170, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 170, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 170, + 338 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": "Implementation details. We use pre-trained SEVA [46] as the generator " + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": ". SEVA uses a fixed total number of reference and target frames (" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "inline_equation", + "content": "K + M = 21" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": "). For our experiments, we use " + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": " target views, which leaves " + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": " context views. However, this configuration demands substantial computational resources that may be prohibitive for many applications. To address this limitation, we fine-tune a more efficient version with LoRA [13] on the RealEstate10K training split [47] that works with a reduced number of reference frames (" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": ") and target views (" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": "), and use this efficient version for our main experiments. We demonstrate (Secs. 4.3 and 4.4) that this fine-tuned version, when combined with VMem, achieves comparable performance to the original SEVA while delivering approximately " + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "inline_equation", + "content": "12\\times" + }, + { + "bbox": [ + 54, + 343, + 295, + 535 + ], + "type": "text", + "content": " speedup at inference time. We provide more technical details in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 624 + ], + "type": "text", + "content": "Datasets. We consider two real-world datasets for novel view generation evaluation: RealEstate10K [47], comprising indoor scene video clips, and Tanks-and-Temples [14], including indoor and outdoor scenes with larger camera motions. For qualitative evaluation, we use in-the-wild images collected from the internet or captured with phone cameras as input views (see Fig. 5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate generated novel view quality based on three factors: (1) overall quality of generated novel views, comparing generated novel view distributions to test set distributions using Fréchet Image Distance (FID) [10]; (2) the ability of the model to preserve image details across views, measuring peak signal-to-noise ratio (PSNR) for pixel differences, along with LPIPS [45]" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 164 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 294, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 294, + 270 + ], + "type": "text", + "content": "Table 3. Results on cycle trajectories from Tanks and Temples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 164 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 164 + ], + "type": "table", + "html": "
Context views KView RetrievalLPIPS ↓PSNR ↑SSIM ↑Rdist ↓Tdist ↓
17Temporal0.47713.920.1880.9760.254
Camera Distance (SEVA [46])0.39715.720.2970.8210.392
Field of View0.37415.750.2920.9110.382
VMem0.30418.150.3770.8920.165
4Temporal0.7947.520.0181.9420.458
Camera Distance0.42213.270.1871.7870.319
Field of View0.42413.110.1921.7820.285
VMem0.38114.820.2750.7930.124
", + "image_path": "9c2c666a03a8bd06b19d4d35b516e22392328332bbfb70d12ad2e81989bfd2b6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 173, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 206 + ], + "type": "text", + "content": "Table 4. Ablation study on cycle trajectories from RealEstate10K. We compare different view retrieval strategies for two context view counts " + }, + { + "bbox": [ + 313, + 173, + 555, + 206 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": "and SSIM [34], as in [22]; and (3) alignment between generated camera poses and ground truth, following [9], where we express estimated camera poses relative to the first frame and normalize translation by the furthest frame. We use DUSt3R [33] to extract poses from generated views. We calculate rotation distance " + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 300 + ], + "type": "text", + "content": " by comparing ground truth and extracted rotation matrices of each generated sequence:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 308, + 517, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 308, + 517, + 324 + ], + "spans": [ + { + "bbox": [ + 351, + 308, + 517, + 324 + ], + "type": "interline_equation", + "content": "R _ {\\mathrm {d i s t}} = \\operatorname {a r c c o s} \\left(0. 5 \\left(\\operatorname {t r} \\left(\\mathbf {R} _ {\\mathrm {g e n}} \\mathbf {R} _ {\\mathrm {g t}} ^ {T}\\right) - 1\\right)\\right),", + "image_path": "cf71c6e3ef552a1f82b77dca1f78f388db3aaf3e7c3f85418a750a898633086e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 332, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 356 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 332, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 332, + 555, + 356 + ], + "type": "text", + "content": " denotes a rotation matrix. We compute translation distance " + }, + { + "bbox": [ + 313, + 332, + 555, + 356 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 332, + 555, + 356 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 390, + 365, + 477, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 365, + 477, + 380 + ], + "spans": [ + { + "bbox": [ + 390, + 365, + 477, + 380 + ], + "type": "interline_equation", + "content": "t _ {\\text {d i s t}} = \\left\\| \\mathbf {t} _ {\\mathrm {g t}} - \\mathbf {t} _ {\\mathrm {g e n}} \\right\\| _ {2}.", + "image_path": "53be50524d6b68530c7d9bcf06f9b21fc30c5083f002c8da1ceddea110435824.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 513, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 513, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 513, + 408 + ], + "type": "text", + "content": "4.2. Short- and long-term view generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 413, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 605 + ], + "type": "text", + "content": "We first evaluate VMem on the benchmark from [22, 42], which defines two scenarios: short-term and long-term novel view generation. Beginning from the first frame of each ground-truth test sequence, the model autoregressively generates images along the ground truth camera trajectories with 10-frame intervals. The short-term scenario evaluates the fifth generated image (50 frames from the initial view), while the long-term scenario evaluates the final image (" + }, + { + "bbox": [ + 313, + 413, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\geq 200" + }, + { + "bbox": [ + 313, + 413, + 555, + 605 + ], + "type": "text", + "content": " frames from the initial view). Following the protocol, we only consider samples with more than 200 frames for the long-term scenario and more than 50 frames for the short-term scenario. We report results in Tab. 1. VMem substantially outperforms all baselines on the key metrics, demonstrating superior scene extrapolation. However, the benchmark trajectories rarely revisit observed regions, limiting evaluation of long-term scene consistency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 613, + 548, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 548, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 548, + 625 + ], + "type": "text", + "content": "4.3. Long-term view generation with revisitations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "To evaluate long-term scene consistency when revisiting observed regions—a scenario that rarely occurs in existing benchmarks—we propose a cycle-trajectory evaluation protocol as a proxy for real revisitation scenarios in human and robotic exploration. In this protocol, the model generates frames following trajectories from initial to final poses, then returns along the same path to the starting position." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25696" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 303, + 245 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 303, + 245 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 303, + 245 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 303, + 245 + ], + "type": "image", + "image_path": "62e0bb877dd5e8d393160ad7baf38cac2c1b2e6a45e87220df9fa32a4ae9ac49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of the cycle trajectory frames (" + }, + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\geq 400" + }, + { + "bbox": [ + 55, + 255, + 555, + 278 + ], + "type": "text", + "content": " frames). VMem autoregressively generates new views while maintaining memory of all previously generated frames, ensuring consistency when revisiting previously seen locations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 551, + 246 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 551, + 246 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 551, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 551, + 246 + ], + "type": "image", + "image_path": "663a1de4bd4cff0bef3258d5340cfd931077a74cff7452a33477436c31db3c70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 298, + 295, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 295, + 382 + ], + "type": "text", + "content": "We evaluate metrics every 10 frames on the return trajectories, using all test sequences. Following this protocol, we report quantitative results in Tab. 2 and qualitative comparisons in Fig. 6. VMem outperforms all baselines across all metrics in cycle-trajectory evaluation, demonstrating superior visual consistency when revisiting previously observed regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 384, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 479 + ], + "type": "text", + "content": "To validate generalization, we evaluate on Tanks-and-Temples containing indoor and outdoor scenes with more dynamic camera trajectories. We use all six advanced scenes and apply the same cycle-trajectory protocol using the first 50 frames, evaluating every frame without temporal subsampling. Results are in Tab. 3. VMem consistently outperforms all baselines across in-domain (Tab. 2) and out-of-domain (Tab. 3) evaluations on the majority of metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "content": "We also qualitatively evaluate VMem across diverse in-the-wild scenes in Fig. 5, demonstrating generation of diverse, consistent, high-quality long videos on in-the-wild scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 539, + 147, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 147, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 147, + 552 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "We ablate our surfel-based memory indexing for context view retrieval in VMem. We evaluate our approach in the cycle-trajectory setting against three alternative retrieval strategies: (1) temporal-based retrieval, selecting the most recent " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " views from memory, (2) camera distance-based retrieval, selecting top " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " views with cameras closest to the target view camera—also used by SEVA [46], our backbone generator, and (3) field-of-view-based retrieval, selecting top " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " views with highest field-of-view overlap with the target view. We also evaluate performance trade-offs between our lightweight, fine-tuned generator using " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " context views versus original SEVA with " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " context views. We report results in Tab. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "content": "Our findings show VMem consistently improves generator performance across both numbers of context views. Improvement is most pronounced with fewer context views " + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "inline_equation", + "content": "(K = 4)" + }, + { + "bbox": [ + 313, + 298, + 554, + 358 + ], + "type": "text", + "content": ", highlighting the effectiveness of our surfel-based memory index." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "text", + "content": "Computational efficiency. A key practical advantage of our approach is its computational efficiency. VMem with only " + }, + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "text", + "content": " views achieves " + }, + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "text", + "content": " speed improvement over the original SEVA while recovering most of the performance achieved with " + }, + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "inline_equation", + "content": "K = 17" + }, + { + "bbox": [ + 313, + 365, + 556, + 461 + ], + "type": "text", + "content": " views. On an RTX 4090 GPU, our method generates frames in 4.2 seconds compared to 50 seconds for SEVA, representing a step towards real-time interactive scene exploration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 497, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 651 + ], + "type": "text", + "content": "We introduced Surfel-Indexed View Memory (VMem), a novel plug-and-play memory module for long-term autoregressive scene generation using a surfel-indexed view memory. By anchoring past views to a surfel-based representation of the scene and retrieving the most relevant ones, our approach improves long-term scene consistency while reducing computational costs. Experiments on long-term scene synthesis benchmarks demonstrate that VMem outperforms existing methods in scene coherence while enabling the use of fewer context views, leading to significantly faster generation. Our work advances scalable video generation, with applications in virtual reality, gaming, and other domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 658, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgments. The authors of this work are supported by Clarendon scholarship, ERC 101001212-UNION, and AIMS EP/S024050/1. The authors would like to thank Xingyi Yang, Zeren Jiang, Junlin Han, Zhongrui Gui for their insightful feedback." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25697" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 116, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 116, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 116, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 116, + 295, + 160 + ], + "type": "text", + "content": "[1] Mark Boss, Zixuan Huang, Aaryaman Vasishta, and Varun Jampani. SF3D: Stable fast 3D mesh reconstruction with uv-unwrapping and illumination disentanglement. arXiv preprint arXiv:2408.00653, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 161, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 295, + 217 + ], + "type": "text", + "content": "[2] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, et al. Genie: Generative interactive environments. In *Forty-first International Conference on Machine Learning*, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 218, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 295, + 272 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124-14133, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 295, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 318 + ], + "type": "text", + "content": "[4] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-Stage Diffusion NeRF: A Unified Approach to 3D Generation and Reconstruction, 2023. arXiv:2304.06714 [cs]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 319, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 363 + ], + "type": "text", + "content": "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 418 + ], + "type": "text", + "content": "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 421, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 295, + 454 + ], + "type": "text", + "content": "[7] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 295, + 510 + ], + "type": "text", + "content": "[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. Cat3d: Create anything in 3d with multi-view diffusion models. Advances in Neural Information Processing Systems, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 295, + 555 + ], + "type": "text", + "content": "[9] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 557, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 610 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "text", + "content": "[11] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7909-7920, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In Proc. ICLR, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[14] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 217 + ], + "type": "text", + "content": "[15] Jing Yu Koh, Harsh Agrawal, Dhruv Batra, Richard Tucker, Austin Waters, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Simple and effective synthesis of indoor 3d scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1169-1178, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 281 + ], + "type": "text", + "content": "[16] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 336 + ], + "type": "text", + "content": "[17] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. RealFusion: 360 reconstruction of any object from a single image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "text", + "content": "[18] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "type": "text", + "content": "[19] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5480-5490, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 471, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 580 + ], + "type": "text", + "content": "[20] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Rocktäschel. Genie 2: A large-scale foundation world model, 2024. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "text", + "content": "[21] Stefan Popov, Amit Raj, Michael Krainin, Yanzhen Li, William T. Freeman, and Michael Rubinstein. Camctrl3d: Single-image scene exploration with precise 3d camera control, 2025. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 625, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 679 + ], + "type": "text", + "content": "[22] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3563-3573, 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[23] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25698" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "text", + "content": "world-consistent video generation with precise camera control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 162 + ], + "type": "text", + "content": "[24] Chris Rockwell, David F Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14104-14113, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 294, + 219 + ], + "type": "text", + "content": "[25] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14356-14366, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 294, + 277 + ], + "type": "text", + "content": "[26] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 279, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 322 + ], + "type": "text", + "content": "[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123: a single image to consistent multi-view diffusion base model. arXiv.cs, abs/2310.15110, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 379 + ], + "type": "text", + "content": "[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arxiv, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "text", + "content": "[29] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 483 + ], + "type": "text", + "content": "[30] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 485, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 485, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 485, + 295, + 540 + ], + "type": "text", + "content": "[31] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 574 + ], + "type": "text", + "content": "[32] Qianqian Wang*, Yifei Zhang*, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In CVPR, 2025. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 577, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 610 + ], + "type": "text", + "content": "[33] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[35] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. ACM SIGGRAPH 2024 Conference Papers, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 686 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[36] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. SynSin: End-to-end view synthesis from a single image. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 172 + ], + "type": "text", + "content": "[37] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21551-21561, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[38] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "text", + "content": "[39] Zeqi Xiao, Yushi Lan, Yifan Zhou, Wenqi Ouyang, Shuai Yang, Yanhong Zeng, and Xingang Pan. Worldmem: Long-term consistent world simulation with memory, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "type": "text", + "content": "[40] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T. Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, and Charles Herrmann. Wonderjourney: Going from anywhere to everywhere. arXiv.cs, abs/2312.03884, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 350 + ], + "type": "text", + "content": "[41] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T. Freeman, and Jiajun Wu. WonderWorld: interactive 3D scene generation from a single image. arXiv, 2406.09394, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 407 + ], + "type": "text", + "content": "[42] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7094-7104, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "text", + "content": "[43] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 464, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 553, + 529 + ], + "type": "text", + "content": "[44] Shangjin Zhai, Zhichao Ye, Jialin Liu, Weijian Xie, Jiaqi Hu, Zhen Peng, Hua Xue, Danpeng Chen, Xiaomeng Wang, Lei Yang, Nan Wang, Haomin Liu, and Guofeng Zhang. Stargen: A spatiotemporal autoregression framework with video diffusion model for scalable and controllable scene generation, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 553, + 586 + ], + "type": "text", + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 588, + 553, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 641 + ], + "type": "text", + "content": "[46] Jensen (Jinghao) Zhou, Hang Gao, Vikram Voleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rupprecht, and Varun Jampani. Stable virtual camera: Generative view synthesis with diffusion models. arXiv preprint arXiv:2503.14489, 2025. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "text", + "content": "[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25699" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_content_list.json b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c39fc453186b83dd30f4c6a00fc561f1b1ee047 --- /dev/null +++ b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_content_list.json @@ -0,0 +1,1454 @@ +[ + { + "type": "text", + "text": "VOVTrack: Exploring the Potentially in Raw Videos for Open-Vocabulary Multi-Object Tracking", + "text_level": 1, + "bbox": [ + 210, + 128, + 785, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zekun Qian $^{1,3*}$ , Ruize Han $^{2*}$ , Junhui Hou $^{3}$ , Linqi Song $^{3}$ , Wei Feng $^{1\\dagger}$ $^{1}$ College of Intelligence and Computing, Tianjin University \n $^{2}$ Shenzhen University of Advanced Technology \n $^{3}$ City University of Hong Kong", + "bbox": [ + 236, + 202, + 764, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{clarkqian, wfeng}@tju.edu.cn, hanruize@suat-sz.edu.cn, {jh.hou, linqi.song}@cityu.edu.hk https://github.com/zekunqian/VOVTrack", + "bbox": [ + 106, + 277, + 888, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary multi-object tracking (OVMOT) represents a critical new challenge involving the detection and tracking of diverse object categories in videos, encompassing both seen categories (base classes) and unseen categories (novel classes). This issue amalgamates the complexities of open-vocabulary object detection (OVD) and multi-object tracking (MOT). Existing approaches to OVMOT often merge OVD and MOT methodologies as separate modules, not fully leveraging the video information. In this work, we propose VOvTrack, a novel method that integrates object states relevant to MOT and video-centric training to address this challenge from a video analysis standpoint. First, we consider the tracking-related state of the objects during tracking and propose a new prompt-guided attention mechanism for more accurate detection (localization and classification) of time-varying objects. Subsequently, we leverage raw video data without annotations for training by formulating a self-supervised object similarity learning technique to facilitate temporal object tracking (association). Experimental results underscore that VOvTrack establishes itself as a state-of-the-art solution for the open-vocabulary tracking task.", + "bbox": [ + 89, + 377, + 483, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-Object Tracking (MOT) is a fundamental task in computer vision and artificial intelligence, which is widely used for video surveillance, media understanding, etc [20, 21]. In the past years, plenty of datasets, e.g., MOT-20 [8], DanceTrack [48], KITTI [17], as well as the algorithms, e.g., SORT [51], Tracktor [46], FairMOT [56], have been proposed for MOT problem. However, most previous works", + "bbox": [ + 89, + 753, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12e3ff12a974479e7b8260f255dee613569b9a372bf380688c7625c695961d45.jpg", + "image_caption": [ + "Figure 1. Comparison between OVTrack [27] and our methods." + ], + "image_footnote": [], + "bbox": [ + 519, + 345, + 903, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "focus on the tracking of simple object categories, i.e., humans and vehicles. Actually, it is important for the perception of various categories of objects in many real-world applications. Some recent works have begun to study the tracking of generic objects. TAO [7] is the first large dataset for the generic MOT, which includes 2,907 videos and 833 object categories. However, these approaches are still limited to tracking predefined object categories.", + "bbox": [ + 511, + 638, + 906, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the development of Artificial General Intelligence (AGI) and multi-modal foundation models, open-world object perception has become a popular topic. Open-vocabulary object detection (OVD) is a new and promising task because of its generic settings. It aims to identify the various categories of objects from an image, including both the categories that have been seen during training (i.e., base classes) and those not seen (i.e., novel classes). Although OVD has been studied in a series of works [9, 11, 23],", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7472", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the literature on open-vocabulary (multi-) object tracking is rare. The recent work [27] builds a benchmark for open-vocabulary multi-object tracking (OVMOT) based on TAO [7]. The authors also develop a simple framework for this problem, consisting of a detection head and a tracking head. The detection head is directly taken from an existing OVD algorithm, i.e., DetPro [13], which is used to detect the open-vocabulary categories of objects in each frame without considering any tracking-related factors. Then the tracking head is used to learn the similarities among the detected objects in different frames. Due to the lack of video data with open-vocabulary tracking annotations, the approach in [27] uses a data hallucination strategy to generate the image pairs for training the tracking head. However, the generated image pairs ignore the adjacent continuity and temporal variability of the objects in a video sequence, which is crucial for higher performance [44]. As discussed above, the existing methods for open-vocabulary tracking simply combine the approaches of OVD and MOT in series as independent modules. It does not consider the object states during tracking, e.g., mutual occlusion, motion blur, etc., and does not make use of the sequential information in the videos.", + "bbox": [ + 93, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this work, we aim to handle the open-vocabulary object tracking from the standpoint of continuous videos. The comparison between our method and that in [27] can be intuitively seen from Figure 1. Specifically, we first consider the various states of the objects during tracking. Our basic idea is that damaged objects (e.g., with occlusion, motion blur, etc.) should be weakened for foreground object feature learning. This way, we model these states as the prompts, which are used to calculate the attention weights of each generated detection proposal during training the object detection network. This methodology can provide more accurate detection (localization and classification) results of the time-varying objects during tracking. Second, to fully utilize the raw videos without open-vocabulary tracking annotations, we formulate the temporal association (tracking) task as a constrained optimization problem. The basic idea is that each object should maintain consistency across different frames, allowing us to leverage object consistency to effectively learn the object appearance similarity features for temporal association. Specifically, we formulate the object appearance self-consistency and the spatial-appearance mutual-consistency learning problem, in which we also consider the object category consistency. These consistencies work together to learn the object appearance similarity features for the temporal association (tracking) sub-task, thus enhancing the overall performance of OVMOT. Notably, the existing OVMOT dataset TAO provides 534.1K (frames) of unlabeled video data. Compared to the 18.1K annotated samples, the unlabeled data have a huge potential to be explored for training. Our self-supervised training method obtains significant performance", + "bbox": [ + 93, + 433, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "improvements only using the raw data, thus effectively alleviating the burden of annotations for OVMOT. The main contributions of this work are:", + "bbox": [ + 516, + 92, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new tracking-related prompt-guided attention for the localization and classification (detection) in the open-vocabulary tracking problem. This method takes notice of the states of the time-varying objects during tracking, which is different from the open-vocabulary object detection from a single image.", + "- We develop a self-supervised object similarity learning strategy for the temporal association (tracking) in the OVMOT problem. This strategy, for the first time, makes full use of the raw video data without annotation for OVMOT training, thus addressing the problem of training data shortage and eliminating the heavy annotation burden in OVMOT.", + "- Experimental results on the public benchmark demonstrate that the proposed VOvTrack achieves the best performance using the same training dataset (annotations)." + ], + "bbox": [ + 516, + 136, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 392, + 653, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple object tracking. The prevailing paradigm in MOT is the tracking-by-detection framework [1], which involves detecting objects in each frame first, and then associating objects across different frames using various cues such as object appearance features [3, 5, 16, 25, 32, 37, 42, 51], 2D motion features [14, 38, 43, 53, 59], or 3D motion features [22, 24, 30, 35, 45, 49]. Some methods leverage graph neural networks [4, 10] or transformers [31, 47, 55, 61] to learn the association relationships between objects, thereby enhancing tracking performance. To broaden the object categories of the MOT task, the TAO benchmark [7] has been proposed for studying MOT under the long-tail distribution of object categories. On this benchmark, relevant methods include AOA [12], GTR [61], TET [26], QDTrack [16], etc. While these methods perform well, they are still limited to pre-defined object categories, which makes them unsuitable for diverse open-world scenarios. Differently, this work handles the OVMOT problem, which contains categories not seen during training.", + "bbox": [ + 516, + 417, + 903, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-world/vocabulary object detection. Open-world object detection identifies salient objects without predefined categories, treating classification as clustering based on similarity [9, 11, 23]. Open-vocabulary detection also identifies unseen categories but requires predicting specific object categories [54]. Early works [2, 40] trained detectors with text embeddings, while recent approaches leverage pre-trained vision-language models like CLIP [39] that connect visual concepts with textual descriptions for strong open-vocabulary classification. Building on this foundation, many works [18, 52, 60] utilize such models for open-vocabulary and few-shot object detection. Some studies [13, 57, 58] further enhance prompt embeddings", + "bbox": [ + 516, + 704, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7473", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "through prompt learning methods. Unlike these single-image detection methods, we propose a prompt-guided training method designed for open-vocabulary tracking in continuous videos, which can effectively enhance the detection performance for the videos.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-world/vocabulary object tracking. There are relatively few works addressing the task of open-world tracking. Related approaches aim to segment or track all the moving objects in the video [6, 33] or handle the generic object tracking [34-36] using a class-agnostic detector. Recently, the TAO-OW benchmark [29] is proposed to study open-world tracking problems, but its limitation lies in evaluating only class-agnostic tracking metrics without assessing class-related metrics. To make the setting more practical, OVTrack [27] first brings the setting of open vocabulary into the tracking task, which also develops a baseline method and benchmark based on the TAO dataset. However, the method in [27] directly uses an existing OVD algorithm for detection, and its training process only utilizes static image pairs and ignores the information from video sequences. Differently, we consider the tracking-related object states for detection, and also propose a self-supervised video-based training method designed for open-vocabulary tracking, making full use of video-level information to enhance the performance of open-vocabulary tracking. Recent work SLAck [28] addressed the lack of temporal continuity in OVTrack's training process by incorporating adjacent frame information. This approach attempts to obtain continuous (pseudo) labels by computing IoU between detection results of adjacent frames and sparsely annotated frames. Different from SLAck, this work designs the self-supervised loss for appearance learning without any annotation.", + "bbox": [ + 91, + 169, + 482, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. PROPOSED METHOD", + "text_level": 1, + "bbox": [ + 89, + 607, + 316, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.VOVTrack Framework", + "text_level": 1, + "bbox": [ + 89, + 633, + 305, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OVMOT requires localizing, tracking, and recognizing the objects in a video, whose problem formulation is provided in Appendix 1. We first describe the framework of VOVTrack, which mainly includes object localization, object classification, and temporal association modules, as shown in Figure 2. For improving the localization and classification, in Section 3.2, we design a tracking-state-aware prompt-guided attention mechanism, which can help the network learn more effective object detection features. For learning the temporal association similarity, in Section 3.3, we propose a video-based self-supervised method to train the association network, which considers the appearance and spatial self/mutual-consistency and category consistency, thereby enhancing the tracking performance.", + "bbox": [ + 89, + 657, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Localization: We employ the class-agnostic object proposal generation approach in Faster R-CNN [41] to localize", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "objects for both base and novel categories $\\mathcal{C}$ in the video. As supported by prior research [6, 18, 27, 60], the localization strategy has shown robust generalization capabilities toward the novel object class $\\mathcal{C}^{\\mathrm{novel}}$ . During the training phase, we leverage the above-mentioned generated RPN proposals as the initial object candidates $P$ . The localization result of each candidate $r \\in P$ is the bounding box $\\mathbf{b} \\in \\mathbb{R}^4$ . For each $\\mathbf{b}$ , we also obtain a confidence $p_c \\in \\mathbb{R}$ derived from the classification head. To refine the localization candidates, besides the classical non-maximum suppression (NMS), we also use this confidence $p_c$ . To obtain a more effective $p_c \\in \\mathbb{R}$ in the classification head, we propose a prompt-guided attention training strategy in later Section 3.2.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classification: Existing closed-set MOT trackers only track several specific categories of objects, which do not need to provide the class of each tracked object. In this way, classification is a new sub-task of the OVMOT. To enable the framework to classify open-vocabulary object classes, following the OV detection algorithms [18, 52, 60], we leverage the knowledge from the pre-trained model, i.e., CLIP [39], to help the network recognize objects belonging to the novel classes $\\mathcal{C}^{\\mathrm{novel}}$ . We distill the classification head using the CLIP model to empower the network's classification head to recognize new objects. Specifically, as shown in Figure 2, after obtaining the RoI feature embeddings $\\mathbf{f}_r$ from the localization head, we design the classification head with a text branch and an image branch to generate embeddings $\\mathbf{f}_r^{\\mathrm{text}}$ and $\\mathbf{f}_r^{\\mathrm{img}}$ for each $\\mathbf{f}_r$ . The supervision of these heads comes from the CLIP text and image embeddings. We use the method in [13] for CLIP encoder pre-training.", + "bbox": [ + 511, + 287, + 903, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we align the text branch with the CLIP text encoder. For $\\forall c\\in \\mathcal{C}^{\\mathrm{base}}$ , the class text embedding $\\mathbf{t}_c$ of class $c$ can be generated by the CLIP text encoder $\\mathcal{T}(\\cdot)$ as $\\mathbf{t}_c = \\mathcal{T}(c)$ . We compute the affinity between the predicted text embeddings $\\mathbf{f}_r^{\\mathrm{text}}$ and their CLIP counterpart $\\mathbf{t}_c$ as", + "bbox": [ + 511, + 544, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\right. \\mathrm {z} (r) = \\left[ \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {\\mathrm {b g}}\\right), \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {1}\\right), \\dots , \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {| C ^ {\\text {b a s e}} |}\\right)\\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 628, + 950, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{| P |} \\sum_ {r \\in P} w _ {r} \\mathrm {L} _ {\\mathrm {C E}} \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathrm {z} (r)}{\\lambda}\\right), c _ {r}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 651, + 903, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{t}_{\\mathrm{bg}}$ is a background prompt learned by treating the background candidates as a new class, $w_{r}$ is the tracking-related prompt-guided attention weight (described in Section 3.2), $\\lambda$ is a temperature parameter, $\\mathrm{LCE}$ is the cross-entropy loss and $c_{r}$ is the class label of $r$ . Then, we align each image branch embedding $\\mathbf{f}_r^{\\mathrm{img}}$ with the CLIP image encoder $\\mathcal{I}(\\cdot)$ . For each candidate object $r$ , we input the corresponding cropped image to $\\mathcal{I}(\\cdot)$ and get the CLIP image embedding $\\mathbf{i}_r$ . We minimize the distance between the corresponding $\\mathbf{f}_r^{\\mathrm{img}}$ and $\\mathbf{i}_r$ as $\\mathcal{L}_{\\mathrm{image}} = \\frac{1}{|P|}\\sum_{r\\in P}\\left\\|\\mathbf{f}_r^{\\mathrm{img}} - \\mathbf{i}_r\\right\\|$ .", + "bbox": [ + 511, + 700, + 903, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the testing stage, with the embeddings $\\mathbf{f}_r^{\\mathrm{text}}$ and $\\mathbf{f}_r^{\\mathrm{img}}$ derived from the text branch and image branch, respectively, we can obtain the corresponding classification probabilities", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7474", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7236f5697863c98b48d2d31df41e26401417f8dec5105744b2532c843d2ebc57.jpg", + "image_caption": [ + "Figure 2. The training framework consists of three parts: first is the localization head used to localize objects of all categories in the video as region candidates; the second is the CLIP distilled classification head consisting of image and text branches, which uses tracking-state-aware prompts to guide the model in focusing on object states while learning classification features, thereby better distinguishing the OV categories; and the third part is the association head that utilizes self/mutual-consistency between the same objects in different frames to learn association features in a self-supervised way." + ], + "image_footnote": [], + "bbox": [ + 130, + 75, + 864, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$p_c^{\\mathrm{text}}$ and $p_c^{\\mathrm{img}}$ of each $r$ belonging to the class $c$ , using the $\\mathbf{z}(r)$ and softmax operation in Eq. (1). After that, we use the fusion strategy in [18] to calculate the final classification probability $p_c$ .", + "bbox": [ + 89, + 329, + 480, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Association: The association head aims to associate the detected objects with the same identification across frames, the main purpose of which is to learn the object features for similarity measurement.", + "bbox": [ + 89, + 392, + 482, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn the association model, given the detected objects from the localization head, we use appearance embeddings to extract the features for association. For training the appearance embedding model, a straightforward method is to select an object as the anchor and its positive/negative samples for learning the object similarity. This method requires object identification (ID) annotations for positive/negative sample selection. However, as an emerging problem, OVMOT does not have enough available training video datasets with tracking ID labels. Previous work [27] uses data augmentation to generate the image pairs for training the association head, which, however, ignores the temporal information in the videos. In this work, we propose to leverage the raw videos to train the association network in a self-supervised manner in Section 3.3.", + "bbox": [ + 89, + 454, + 483, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tracking-State-Aware Prompt based Attention", + "text_level": 1, + "bbox": [ + 89, + 695, + 482, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tracking-state-aware prompts. In the classification head of existing open-vocabulary detection methods, when selecting region candidates for calculating classification loss, they only consider whether the maximal IOU between the candidates and ground-truth boxes exceeds a threshold. For this tracking problem, we further consider whether the states of the candidates are appropriate for training the detection network. As is well known, objects present many specific states during tracking, such as occlusion/out-of-view/motion blur, etc, which occur more frequently than in static image object detection. So, it is important to identify such object states to achieve more accurate detection and", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tracking results. However, these states have often been overlooked in past methods because the labels of these states are difficult to obtain, not to mention incorporating them into the network training.", + "bbox": [ + 511, + 329, + 906, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompts, as a burgeoning concept, can be used to bridge the gap between the vision and language data based on the cross-modal pre-trained models. We consider using specially designed prompts to perceive the tracking states of the objects and integrate such states into the model training. We refer to such prompts as 'tracking-state-aware prompts.' We specifically employ pairs of adjectives with opposite meanings to describe the object states during tracking, e.g., 'unoccluded and occluded.' As shown in the classification head of Figure 2, we add $M$ pairs of tracking-state-aware prompt pairs, denoted as $\\{p_{\\mathrm{pos}}^m, p_{\\mathrm{neg}}^m\\}_{m=1}^M$ , into our framework model. Next, we present how to use these prompts during training.", + "bbox": [ + 511, + 390, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt-guided attention. To utilize these tracking-state-aware prompts in guiding model training, we encode these prompts through the CLIP text encoder $\\mathcal{T}(\\cdot)$ to obtain state embeddings $\\{\\mathbf{t}_{\\mathrm{pos}}^{m},\\mathbf{t}_{\\mathrm{neg}}^{m}\\}_{m = 1}^{M}$ , and calculate the prompt-guided attention weight $w_{r}$ as", + "bbox": [ + 511, + 587, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} (r, m) = \\left[ \\cos \\left(\\mathbf {f} _ {r} ^ {\\mathrm {t e x t}}, \\mathbf {t} _ {\\mathrm {p o s}} ^ {m}\\right), \\cos \\left(\\mathbf {f} _ {r} ^ {\\mathrm {t e x t}}, \\mathbf {t} _ {\\mathrm {n e g}} ^ {m}\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 666, + 859, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {r} = \\frac {1}{| M |} \\sum_ {m = 1} ^ {M} \\left(\\operatorname {s o f t m a x} \\frac {\\mathrm {z} (r , m)}{\\lambda}\\right) _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 688, + 903, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_r^{\\mathrm{text}}$ is the text embedding of a region candidate $r$ , and $(\\cdot)_1$ represents using the probability of the positive state as the attention for this pair. Note that, in our definition, the positive states always denote that the object state is beneficial to the object feature learning, e.g., unoccluded, unobsured. From the above analysis, we know that the attention obtained from the tracking-state-aware prompts evaluates the various object states, resulting in prompt-guided attention values $w_r \\in [0, 1]$ .", + "bbox": [ + 511, + 734, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Piecewise weighting strategy. Next, we use this prompt-guided attention to help the model better utilize", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7475", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "high-quality candidates for training and filter out the low-quality noisy candidates. Specifically, we divide the $w_{r}$ into three levels: low $(d_{\\mathrm{low}})$ , medium and high $(d_{\\mathrm{high}})$ . For embedding features $\\mathbf{f}_r^{\\mathrm{text}}$ with $w_{r} < d_{\\mathrm{low}}$ , we regard such features as low-quality states, and filter them out during training, by assigning $w_{r}$ to 0. For $d_{\\mathrm{low}} \\leq w_{r} \\leq d_{\\mathrm{high}}$ , we regard that even if these features are not of high quality, they still contribute to training, and retain their original weights $w_{r}$ . For $w_{r} > d_{\\mathrm{high}}$ , we regard the features of these regions as particularly suitable ones for the model to learn tracking-related features, so we assign $w_{r}$ to 1 as a reward. After that, we apply the re-assigned $w_{r}$ to Eq. (1), thereby integrating the object tracking-related state into the model training, which enables the network to better learn the object representations specifically for the tracking task.", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Self-supervised Association with Raw Videos", + "text_level": 1, + "bbox": [ + 89, + 330, + 470, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the lack of annotated videos for OVMOT, we develop a self-supervised approach to train the association network by leveraging the consistency among the objects with the same identity along the video.", + "bbox": [ + 89, + 352, + 483, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formulation. To learn the object appearance feature, we consider two aspects. The first one is self-consistency factor, i.e., the self-consistency of appearance for the same object at different times. The second one is mutual-consistency factor, i.e., the mutual-consistency between the appearance and motion cues during tracking. In this way, we formulate the optimized objective function as", + "bbox": [ + 86, + 414, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\substack {S = S _ {\\text {self}} \\left(t _ {i}, t _ {j}\\right) + \\alpha \\cdot S _ {\\text {mutual}} \\left(t _ {i}, t _ {j}\\right), \\\\ \\left. \\quad \\right.}} \\tag{3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 534, + 483, + 558 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad \\forall t _ {i}, t _ {j} \\in \\mathcal {T} _ {c}, \\quad \\forall c \\in \\mathcal {C},\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 554, + 333, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S$ represents the overall consistency objective to be maximized, $S_{\\mathrm{self}}$ and $S_{\\mathrm{mutual}}$ denote the self-consistency and mutual-consistency measures respectively, while $\\alpha$ is a weight to balance them. Besides, considering the diversity of object categories in OVMOT, we add the object category consistency constraint for consistency learning. Specifically, we construct the intervals, i.e., $\\mathcal{T}_c$ , containing several frames only with the same object category $c$ , in which we select $t_i, t_j$ . This is because we aim to learn the feature in a self-supervised manner without ID annotation, and the objects with various categories may bring about interference.", + "bbox": [ + 89, + 582, + 483, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\Theta$ Category-consistency constraint. First, we consider the category consistency constraint in $\\mathcal{T}_c$ . For the interval splitting of $\\mathcal{T}_c$ in Eq. (3), we split the original videos into several segments of length $L$ and randomly sample the shorter sub-segments with various lengths from each segment. These short-term sub-segments are then concatenated to form the training sequence. Such training sequences include short-long-term intervals. Specifically, we select the adjacent frames from the same sub-segment, which allows the association head to learn the consistency objectives", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "under minor object differences. We also select the long-interval video frames from different sub-segments, which allow the association head to learn the similarity and variation of objects under large differences.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the sampled training sequences as discussed above, we utilize the localization head to extract object bounding boxes from each frame in the sequence. Since we only use unlabeled raw videos for training, we cannot directly obtain the object categories. To address this issue, we employ a clustering approach to group the bounding boxes based on their classification features. Specifically, we utilize the classification head to obtain the predicted category probability distributions for the candidate objects from all frames. Then, we cluster the candidate objects' category distributions into different clusters. After clustering the candidate objects' category distributions into distinct groups, we can treat each cluster as a separate category. As shown in Figure 2, we conduct self-supervised learning on the objects that belong to the same category cluster and are sourced from different frames.", + "bbox": [ + 511, + 152, + 906, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\Theta$ Appearance-self-consistency loss. We next model the consistency learning problem in Eq. (3) as a self-supervised learning task. After getting the training samples (object bounding boxes in different frames of the sample training sequence within the category clustering), we first use the CNN network to extract the appearance feature from the association head for all objects in frame $t_i$ to construct the feature matrix $\\mathbf{F}_i$ . The main idea of the appearance-self-consistency loss is to leverage the self-consistency of the same objects' appearance features at different times (frames). We utilize the following two types of similarity transfer relationships, i.e., pair-wise symmetry and triple-wise cyclicity.", + "bbox": [ + 511, + 393, + 908, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pair-wise consistency learning: For a pair of frames $t_i$ and $t_j$ in the video, we can get the object similarity matrix between them as $\\mathbf{M}_{ij} = \\mathbf{F}_i \\cdot (\\mathbf{F}_j)^{\\mathrm{T}}$ . We then compute the normalized similarity matrix $\\mathbf{S}_{ij} \\in [0,1]$ based on the above similarity matrix $\\mathbf{M}_{ij}$ by temperature-adaptive row softmax", + "bbox": [ + 511, + 589, + 905, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} (r, c) = f (\\mathbf {M} (r, c)) = \\frac {\\exp (\\tau \\mathbf {M} (r , c))}{\\sum_ {c = 1} ^ {C} \\exp (\\tau \\mathbf {M} (r , c))} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 676, + 906, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r, c$ represent the indices of row and column in $\\mathbf{M}$ or $\\mathbf{S}$ , respectively. Here $C$ is the number of columns for $\\mathbf{M}$ , and $\\tau$ is the adaptive temperature adjustable parameter.", + "bbox": [ + 511, + 719, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The normalized similarity matrix $\\mathbf{S}_{ij}$ can be regarded as a mapping (object association relation) from frame $t_i$ to frame $t_j$ ( $\\mathbf{S}_{ij}: t_i \\to t_j$ ). In other words, we can select the maximum value of each row of $\\mathbf{S}_{ij}$ as the matched objects between frames $t_i$ and $t_j$ . Similarly, we get the reversed mapping from $t_j$ to $t_i$ as $\\mathbf{S}_{ji}: t_j \\to t_i$ . We calculate the pair-wise similarity matrix as $\\mathbf{E}_{\\mathrm{pair}} = \\mathbf{S}_{ij} \\cdot \\mathbf{S}_{ji}$ , where $\\mathbf{E}_{\\mathrm{pair}}$ can be regarded as a symmetric mapping: $t_i \\rightleftarrows t_j$ , i.e., from $t_i$ and return to $t_i$ . If the objects in frames $t_i$ and $t_j$ are the", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7476", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "same, the result $\\mathbf{E}_{\\mathrm{pair}}$ should be an identity matrix, which can be used to construct the self-supervision loss. However, due to the object differences in different frames, this condition may not always be satisfied. Therefore, we need to supervise it deliberately, which will be discussed later.", + "bbox": [ + 88, + 90, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Triplet consistency learning: Besides the pair-wise symmetric similarity, we further consider the triple-wise circularly consistent similarity. Specifically, given the similarity matrix $\\mathbf{S}_{ij}$ between two frames $t_i$ and $t_j$ , as well as $\\mathbf{S}_{jk}$ between frames $t_j$ and $t_k$ , we aim to build the consistent similarity relation among this triplet, i.e., frames $t_i, t_j$ and $t_k$ . To do this, we first calculate the triplet similarity matrix as $\\mathbf{M}_{ik} = \\mathbf{M}_{ij} \\cdot \\mathbf{M}_{jk}, (i \\neq j \\neq k)$ , where $\\mathbf{M}_{ik}$ represents the similarity between the objects in frames $t_i$ and $t_k$ , through the frame $t_j$ . We then compute the normalized similarity matrix using Eq. (4) as", + "bbox": [ + 89, + 167, + 482, + 335 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {i k} = f \\left(\\mathbf {M} _ {i k}\\right), \\quad \\mathbf {S} _ {k i} = f \\left(\\left(\\mathbf {M} _ {i k}\\right) ^ {\\mathrm {T}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 347, + 482, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{S}_{ik}$ represents the mapping $t_i\\rightarrow t_j\\rightarrow t_k$ while $\\mathbf{S}_{ki}$ represents the mapping along $t_k\\to t_j\\to t_i$ . Then, we calculate the transitive-similarity matrix: $\\mathbf{E}_{\\mathrm{trip}} = \\mathbf{S}_{ik}\\cdot \\mathbf{S}_{ki}$ , where $\\mathbf{E}_{\\mathrm{trip}}$ can be regarded as the mapping: $t_i\\rightleftarrows t_j\\rightleftarrows t_k$ (from $t_i$ and return to $t_i$ ).", + "bbox": [ + 89, + 378, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For convenience, we note the matrices $\\mathbf{E}_{\\mathrm{pair}}$ and $\\mathbf{E}_{\\mathrm{trip}}$ as $\\mathbf{E}$ , which should have the property that their diagonal elements are either 1 or 0, while all other elements are 0, in an ideal case. This means that the diagonal elements of $\\mathbf{E}$ must be greater than or equal to the other elements. Based on this consideration, following [15, 50], we use the following loss $\\mathrm{L}(\\mathbf{E}) = \\sum_{r}\\mathrm{relu}\\left(\\max_{c\\neq r}\\mathbf{E}(r,c) - \\mathbf{E}(r,r) + m\\right)$ , where $r,c$ denote the indices of row and column in $\\mathbf{E}$ . This loss denotes that we penalize the cases where the max non-diagonal element $\\mathbf{E}(r,c)$ ( $c\\neq r$ ) in a row $r$ , exceeds the corresponding diagonal elements $\\mathbf{E}(r,r)$ with a margin $m$ . The margin $m\\geq 0$ is a parameter used to control the punishment scope between $\\mathbf{E}(r,c)$ and $\\mathbf{E}(r,r)$ . This loss helps $\\mathbf{E}$ approach the identity matrix while addressing cases where there are unmatched targets in a row through a margin $m$ . Finally, we define our self-supervised self-consistency learning loss as $\\mathcal{L}_{\\mathrm{self}} = \\mathrm{L}\\left(\\mathbf{E}_{\\mathrm{pair}}\\right) + \\mathrm{L}\\left(\\mathbf{E}_{\\mathrm{trip}}\\right)$ .", + "bbox": [ + 89, + 455, + 482, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\Theta$ Mutual-consistency loss. The mutual-consistency loss makes use of the consistency between the spatial position continuity and the appearance similarity of the objects at different times. Given a bounding box list $B_{i}$ and $B_{j}$ of adjacent frames $i$ and $j$ in the video, we compute the Intersection over Union (IoU) matrix $\\mathbf{M}_{ij}$ between bounding boxes $B_{i}$ and $B_{j}$ , which quantifies the overlap between them. Based on a specified threshold IoUthres, we create an assignment matrix $\\mathbf{A}_{ij}$ such that", + "bbox": [ + 89, + 714, + 482, + 849 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {M} _ {i j} > \\operatorname {I o U} _ {\\text {t h r e s}}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 863, + 482, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This assignment matrix $\\mathbf{A}_{ij}$ indicates whether pairs of objects are considered similar, facilitating the optimization of the spatial consistency objective. Furthermore, we utilize the previously obtained normalized similarity matrix $\\mathbf{S}_{ij}$ to establish the mutual-consistency loss, using the binary cross-entropy loss function $\\mathrm{L_{BCE}}$ , as $\\mathcal{L}_{\\mathrm{mutual}} = \\mathrm{L_{BCE}}(\\mathbf{S}_{ij},\\mathbf{A}_{ij})$ . This formulation allows us to measure the discrepancy between the normalized similarity matrix $\\mathbf{S}_{ij}$ and the assignment matrix $\\mathbf{A}_{ij}$ , thereby enforcing spatial consistency across the detected objects in frames $t_i$ and $t_j$ .", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we have transformed the optimization problem in Eq. (3) into a self-supervised loss as $\\mathcal{L} = \\mathcal{L}_{\\mathrm{self}} + \\alpha \\cdot \\mathcal{L}_{\\mathrm{mutual}}$ , where $\\alpha$ is a weighting coefficient. This formulation incorporates both the self- and mutual-consistencies, as well as category consistency, among the frames with different intervals, thereby effectively exploring the potentiality of videos to enhance the association for OVMOT.", + "bbox": [ + 511, + 243, + 903, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 362, + 730, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Section 3.1, we use ResNet50 with FPN for localizing candidate regions. We set $\\lambda$ in Eq. (1) as 0.007. In Section 3.2, we select four pairs of typical tracking state aware prompts, i.e., 'complete and incomplete,' 'unoccluded and occluded,' 'unobsured and obscured,' and 'recognizable and unrecognizable.' We set $d_{\\mathrm{low}}$ as 0.3, $d_{\\mathrm{high}}$ as 0.6. In Section 3.3, the segment length $L$ is set as 24. We use the K-means clustering algorithm. We set the margin $m$ as 0.5 and the IoUthres as 0.9. For $N$ frames in each sampled video sequence, we select $C_N^2$ and $C_N^3$ groups of frames to calculate the self-consistency loss in Section 3.3. We also select $C_N^2$ groups of frames to calculate the mutual-consistency loss. The weighting coefficient $\\alpha$ is 0.9.", + "bbox": [ + 511, + 385, + 903, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the training stage, we first train the two-stage detector on the base classes of the LVIS dataset [19] referenced from [13], for 20 epochs, and use prompt-guided attention proposed in Section 3.2 to fine-tune the model's classification head for 6 epochs. Then we train the association head using static image pairs generated from [27] for 6 epochs and self-supervise the association head with TAO training dataset [7] without annotation for 14 epochs. In the inference stage, we select object candidates $P$ by NMS with IoU threshold 0.5 for detection. Additionally, for tracking, we use appearance feature similarity to associate history tracks with the objects in the current frame. As in [27], we evaluate the similarity between history tracks and detected objects using both bi-directional softmax [16] and cosine similarity metrics. Following the classical MOT approaches, if the similarity score exceeds a matching threshold, we assign the object to the track. If the object doesn't correspond to any existing track, a new track is initiated if its detection confidence score surpasses a threshold, otherwise, it is disregarded. We set the similarity score threshold as 0.35 and maintain a track memory of 30 frames.", + "bbox": [ + 511, + 584, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7477", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/78d021d205f770104ef66636795cc47a5a399a7691579022e7836fe5c208c337.jpg", + "table_caption": [ + "Table 1. Result comparison. We evaluate our method against closed-set and open-vocabulary trackers on TAO validation and test sets. Here '√' denotes using the corresponding dataset with annotations, and † represents only using the raw video without any annotation (%)." + ], + "table_footnote": [], + "table_body": "
MethodClassesDataBaseNovel
Validation setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓27.145.624.711.022.542.724.40.4
TETer [26]✓✓✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [51]✓-✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓-✓✓28.347.420.517.022.746.719.32.2
OVTrack [27]✓-✓-35.549.336.920.227.848.833.61.5
SLAck [28]✓-✓✓37.255.037.619.131.154.337.81.3
VOVTrack (Ours)✓-✓†39.6 (↑ 2.4)58.940.919.135.3 (↑ 4.2)58.540.96.5
Test setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓25.843.223.510.620.239.720.90.2
TETer [26]✓✓✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [51]✓-✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓-✓✓26.044.119.014.818.039.013.41.7
OVTrack [27]✓-✓-32.645.635.416.924.141.828.71.8
SLAck [28]✓-✓✓34.752.535.616.127.149.130.02.0
VOVTrack (Ours)✓-✓†38.1 (↑ 3.4)57.341.515.429.8 (↑ 2.7)51.732.05.7
", + "bbox": [ + 207, + 104, + 789, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. EXPERIMENTS", + "text_level": 1, + "bbox": [ + 89, + 306, + 261, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Dataset and Metrics", + "text_level": 1, + "bbox": [ + 89, + 332, + 282, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow [27] to select the dataset and metrics for evaluation. We leverage the comprehensive and extensive vocabulary MOT dataset TAO [7] as our benchmark for OV-MOT. TAO is structured similarly to the LVIS [19] taxonomy, categorizing object classes based on their frequency of appearance into frequent, common, and rare groups. We use the rare classes defined in LVIS as $\\mathcal{C}^{\\mathrm{novel}}$ and others as $\\mathcal{C}^{\\mathrm{base}}$ . We evaluate the performance with the comprehensive metric tracking-every-thing accuracy (TETA), which consists of the accuracies of localization (LocA), classification (ClsA), and association (AssocA).", + "bbox": [ + 89, + 354, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparative Results", + "text_level": 1, + "bbox": [ + 89, + 531, + 287, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with the latest trackers, TETer [26] and QDTrack [16], which are trained on both $\\mathcal{C}^{\\mathrm{base}}$ and $\\mathcal{C}^{\\mathrm{novel}}$ . We include the classical trackers like DeepSORT [51] and Tractor++ [3] trained only on $\\mathcal{C}^{\\mathrm{base}}$ and enhanced by OVD method ViLD [18] to achieve open-vocabulary tracking, as baselines. We also compare our method with the state-of-the-art OVMOT methods, i.e., OVTrack [27] and SLAck [28].", + "bbox": [ + 89, + 551, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 1, we present the OVMOT evaluation results on the TAO validation and test sets on both base and novel classes. We can see that our method outperforms all closed-set and open-vocabulary methods on both the validation and test sets. Even though QDTrack [16] and TETer [26] have seen novel classes during training on TAO, our method significantly outperforms them in all metrics on both base and novel classes. Additionally, DeepSORT and Tracktor++ with the open-vocabulary detector ViLD [18] are also trained in a supervised manner on TAO, while our method, trained in a self-supervised manner on TAO, surpasses them by a large margin. Although the results of SLAck, the most competitive method, are better than other comparison methods, our method outperforms it in all metrics significantly on both base and novel classes.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Particularly, compared to our baseline method OVTrack, our method achieves improvements of $4.1\\%$ and $7.5\\%$ in base and novel TETA, respectively, and a $5.0\\%$ , $7.3\\%$ increase in novel ClsA and AssocA. In the test set, base and novel TETA also show improvements of $5.5\\%$ and $5.7\\%$ , respectively. It is worth noting that the state-of-the-art method SLAck is trained on the TAO training dataset using the annotations, while our method only uses the raw video data of it without the annotations. Even so, our method outperforms it with a significant margin on the comprehensive TETA score, as shown in Table 1. Especially the improvement of novel category on the validation set has reached 5x the original level. This demonstrates the effectiveness of the proposed method, which is very promising for the OVMOT task. We provide more analysis from the standpoint of data quantity for training in Appendix 2.", + "bbox": [ + 511, + 309, + 906, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Balance between classification and detection. Our method outperforms other methods on most metrics, except for the ClsA of the base category. We clarify that the lower ClsA score compared to OVTrack stems from our method's superior detection capability. Specifically, we classify many more detected objects, while OVTrack focuses on a smaller subset of high-confidence detections. This is supported by our comparison with SLAck, which also shows reduced classification performance on the base category despite better detection results (LocA), similar to our findings. In contrast, our method significantly improves performance on novel categories by identifying more unseen detections during testing. Overall, our method achieves a better balance between detection and classification tasks, as well as between base and novel categories.", + "bbox": [ + 511, + 551, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 787, + 666, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct the ablation studies on the main components proposed in our method, including prompt-guided attention, and the self-supervised learning related modules as:", + "bbox": [ + 511, + 810, + 905, + 854 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- w/o prompt-guided attention $(w_{r})$ : Removing the prompt-guided attention in Section 3.2.", + "- w/o piecewise weight strategy ( $d_{\\mathrm{low}}$ and $d_{\\mathrm{high}}$ ): Removing" + ], + "bbox": [ + 511, + 856, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7478", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1f47e5e8ebb90c95f53fcea5036916fbcf6e1ed698246927d2bf7ab52f5658fc.jpg", + "table_caption": [ + "Table 2. Ablation studies on modules of prompt-guided attention and self-supervised association $(\\%)$" + ], + "table_footnote": [], + "table_body": "
ModuleAblation MethodsBaseNovel
TETALocAAssocAClsATETALocAAssocAClsA
Prompt-guided attentionw/o prompt-guided attention36.153.437.517.430.553.236.81.6
w/o piecewise weight strategy37.355.838.517.633.856.839.84.8
Self-consistency associationw/o whole self-consistency learning37.457.736.917.732.557.635.64.2
w/o short-long-sampling37.957.637.718.434.058.438.25.5
w/o category consistency38.157.839.417.133.358.437.93.6
w/o self-consistency38.257.538.818.334.258.339.15.3
w/o mutual-consistency39.158.640.118.634.158.838.15.6
CompleteVOVTrack (Ours)39.658.940.919.135.358.540.96.5
", + "bbox": [ + 169, + 103, + 823, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the piecewise weighting strategy proposed in Section 3.2, by directly using the $w_{r}$ calculated by Eq. (2).", + "bbox": [ + 89, + 227, + 480, + 257 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- w/o whole self-supervised learning: Removing the whole self-supervised consistency learning strategy in Section 3.3.", + "- w/o short-long-sampling: Removing the short-long-interval sampling strategy in Section 3.3.", + "- w/o category consistency: Removing the category-aware object clustering in Section 3.3.", + "- w/o self/mutual-consistency: Removing the self/mutual-consistency loss in Section 3.3." + ], + "bbox": [ + 89, + 257, + 482, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the state-aware prompt-guided attention. As shown in the first unit of Table 2, we can see that using prompt-guided attention as a weight coefficient during the training stage can effectively improve all metrics for both base and novel classes. The piecewise weighting strategy is also very effective, especially in improving the localization and classification accuracies.", + "bbox": [ + 89, + 378, + 482, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the self-supervised consistency learning. As shown in the second unit of Table 2, we can see that using self-supervised loss can effectively improve all metrics for both base and novel classes. Either the self-consistency or the mutual-consistency for appearance learning is effective for the association task, i.e., 'AssocA.' Also, the interval sampling strategy allows samples to have a more diverse range of long and short cycles, improving the association-related metric. The category clustering strategy tries to gather the objects with the same category in a cluster, which is also helpful. To our surprise, the above strategies, in most cases, also effectively help improve classification ('ClsA') and localization ('LocA') accuracies. This is because the better association results can indirectly help the other two sub-tasks in OVMOT. We provide the discussion and analysis of the complementarity among different tasks in Appendix 3.", + "bbox": [ + 89, + 486, + 482, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 756, + 282, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Illustrations of the proposed prompt-guided attention.", + "text_level": 1, + "bbox": [ + 89, + 780, + 482, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 (a) shows cases of high prompt-guided attention, where we can see that the regions often have very distinctive category features with no occlusion, and the image quality of the region is very high. In contrast, Figure 3 (b) presents cases of low prompt-guided attention, highlighting regions that often exhibit issues such as significant blurriness, occlusion, unclear visibility, and difficulty in identification.", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These factors are detrimental to training object localization and classification features, which are effectively weakened through state-aware prompt-guided attention.", + "bbox": [ + 511, + 227, + 903, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c7a160ed6bafafceab2481a02fe378d01999a22bf32f0bb0c0ad7eb303ddc05.jpg", + "image_caption": [ + "Figure 3. Illustration of regions with high (a) or low (b) prompt-guided attention, respectively." + ], + "image_footnote": [], + "bbox": [ + 514, + 272, + 903, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison result visualization. We show several visualization results in Figure 4. We can see that the proposed method provides better results than OVTrack [27]. In the first case of Figure 4 (a), our method provides an accurate object localization result and identifies the correct category. In the second case of Figure 4 (b), the tracking of a drone provided by the comparison method is wrong (different box colors denote different tracking IDs), and the classification is also not correct. Our method can track it continuously.", + "bbox": [ + 511, + 378, + 905, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ce955364ce5531f6610dcb312b816fbfcdd2f99751cde9c3b8cd132ca13a71d.jpg", + "image_caption": [ + "Figure 4. Compared OVMOT results of ours and OVTrack on some cases with novel classes." + ], + "image_footnote": [], + "bbox": [ + 516, + 515, + 903, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Appendix 4, we provide more analysis that includes additional cases on prompt-guided attention, failure cases, time complexity analysis, parameter settings analysis, and method generalization analysis.", + "bbox": [ + 511, + 606, + 905, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. CONCLUSION", + "text_level": 1, + "bbox": [ + 511, + 679, + 669, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have developed a new method, namely VOVTrack, to handle the OVMOT problem from the perspective of video object tracking. For this purpose, we first consider the object state during tracking and propose tracking-state-aware prompt-guided attention, which improves the accuracy of object localization and classification (detection). Second, we develop an object similarity learning strategy for the temporal association (tracking) using only the raw video data without annotation, which unveils the power of self-supervised learning for open-vocabulary tracking tasks. Experimental results demonstrate the effectiveness of the proposed method and each component for open-vocabulary tracking.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7479", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 262, + 187, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele. People-tracking-by-detection and people-detection-by-tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2", + "[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In European Conference on Computer Vision, pages 384-400, 2018. 2", + "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 2, 7", + "[4] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In IEEE International Conference on Advanced Video and Signal based Surveillance, pages 1-6, 2017. 2", + "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. MeMOT: Multi-object tracking with memory. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8090-8100, 2022. 2", + "[6] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In IEEE/CVF International Conference on Computer Vision Workshops, 2019. 3", + "[7] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. TAO: A large-scale benchmark for tracking any object. In European Conference on Computer Vision, pages 436-454, 2020. 1, 2, 6, 7", + "[8] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. MOT20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1", + "[9] Akshay Dhamija, Manuel Gunther, Jonathan Ventura, and Terrance Boult. The overlooked elephant of object detection: Open set. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1021-1030, 2020. 1, 2", + "[10] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 2", + "[11] Thang Doan, Xin Li, Sima Behpour, Wenbin He, Liang Gou, and Liu Ren. Hyp-OW: Exploiting hierarchical structure" + ], + "bbox": [ + 93, + 287, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported in part by the Young Scientists Fund of the National Natural Science Foundation of China (NSFC) under Grant 62402490, the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage) from Tianjin University, the Excellent Young Scientists Fund of NSFC under Grant 62422118, and the Guangdong Basic and Applied Basic Research Foundation under Grant 2025A1515010101.", + "learning with hyperbolic distance enhances open world object detection. In AAAI Conference on Artificial Intelligence, pages 1555-1563, 2024. 1, 2", + "[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to ECCV-TAO-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2", + "[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2, 3, 6", + "[14] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. StrongSORT: Make DeepSORT great again. IEEE Transactions on Multimedia, 25:8725-8737, 2023. 2", + "[15] Wei Feng, Feifan Wang, Ruize Han, Yiyang Gan, Zekun Qian, Junhui Hou, and Song Wang. Unveiling the power of self-supervision for multi-view multi-human association and tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6", + "[16] Tobias Fischer, Thomas E Huang, Jiangmiao Pang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. QD-Track: Quasi-dense similarity learning for appearance-only multiple object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7", + "[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1", + "[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 2, 3, 4, 7", + "[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019. 6, 7", + "[20] Ruize Han, Wei Feng, Yujun Zhang, Jiewen Zhao, and Song Wang. Multiple human association and tracking from egocentric and complementary top views. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5225-5242, 2022. 1", + "[21] Ruize Han, Wei Feng, Feifan Wang, Zekun Qian, Haomin Yan, and Song Wang. Benchmarking the complementary-view multi-human association and tracking. International Journal of Computer Vision, 132:118-136, 2023. 1", + "[22] Kuan-Chih Huang, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Delving into motion-aware matching for monocular 3d object tracking. In IEEE/CVF International Conference on Computer Vision, pages 6909-6918, 2023. 2", + "[23] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2", + "[24] Jan Krejci, Oliver Kost, Ondrej Straka, and Jindrich Duník. Pedestrian tracking with monocular camera us-" + ], + "bbox": [ + 93, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7480", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing unconstrained 3D motion model. arXiv preprint arXiv:2403.11978, 2024. 2", + "[25] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese cnn for robust target association. In IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops, pages 33-40, 2016. 2", + "[26] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In European Conference on Computer Vision, pages 498-515, 2022. 2, 7", + "[27] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 1, 2, 3, 4, 6, 7, 8", + "[28] Siyuan Li, Lei Ke, Yung-Hsu Yang, Luigi Piccinelli, Mattia Segu, Martin Danelljan, and Luc Van Gool. Slack: Semantic, location, and appearance aware open-vocabulary tracking. In European Conference on Computer Vision, pages 1-18, 2024. 3, 7", + "[29] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19045-19055, 2022. 3", + "[30] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2", + "[31] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 2", + "[32] Anton Milan, S Hamid Rezatofighi, Anthony Dick, Ian Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI conference on Artificial Intelligence, 2017. 2", + "[33] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In European Conference on Computer Vision, pages 566-579, 2012. 3", + "[34] Aljosa Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In IEEE International Conference on Robotics and Automation, pages 3180-3187, 2016. 3", + "[35] Aljosa Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3494–3501, 2018. 2", + "[36] Aljosa Osep, Paul Voiglaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In IEEE International Conference on Robotics and Automation, pages 10031-10037, 2020. 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2", + "[38] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2023. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3", + "[40] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI conference on artificial intelligence, pages 11932–11939, 2020. 2", + "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3", + "[42] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In IEEE/CVF International Conference on Computer Vision, pages 300-311, 2017. 2", + "[43] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021. 2", + "[44] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 55(13s):1-37, 2023. 2", + "[45] Sarthak Sharma, Junaid Ahmed Ansari, J Krishna Murthy, and K Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3508-3515, 2018. 2", + "[46] Vivek Hari Sridhar, Dominique G Roche, and Simon Gingins. Tracktor: image-based automated tracking of animal movement and behaviour. Methods in Ecology and Evolution, 10(6):815-820, 2019. 1", + "[47] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2", + "[48] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1", + "[49] Li Wang, Xinyu Zhang, Wenyuan Qin, Xiaoyu Li, Jinghan Gao, Lei Yang, Zhiwei Li, Jun Li, Lei Zhu, Hong Wang, and Huaping Liu. CAMO-MOT: Combined Appearance-Motion" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7481", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Optimization for 3D Multi-Object Tracking With Camera-LiDAR Fusion. IEEE Transactions on Intelligent Transportation Systems, 24(11):11981-11996, 2023. 2", + "[50] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In European Conference on Computer Vision, pages 72–88, 2020. 6", + "[51] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In IEEE International Conference on Image Processing, pages 3645-3649, 2017. 1, 2, 7", + "[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2, 3", + "[53] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision, pages 466-481, 2018. 2", + "[54] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2", + "[55] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. MOTR: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675, 2022. 2", + "[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. FairMOT: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1", + "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2", + "[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2", + "[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490, 2020. 2", + "[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022. 2, 3", + "[61] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 813 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7482", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_model.json b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..248ab706ab7e8f164f4c77f3708cdb145b590a0b --- /dev/null +++ b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.13, + 0.787, + 0.178 + ], + "angle": 0, + "content": "VOVTrack: Exploring the Potentially in Raw Videos for Open-Vocabulary Multi-Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.203, + 0.765, + 0.277 + ], + "angle": 0, + "content": "Zekun Qian\\(^{1,3*}\\), Ruize Han\\(^{2*}\\), Junhui Hou\\(^{3}\\), Linqi Song\\(^{3}\\), Wei Feng\\(^{1\\dagger}\\) \n\\(^{1}\\)College of Intelligence and Computing, Tianjin University \n\\(^{2}\\)Shenzhen University of Advanced Technology \n\\(^{3}\\)City University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.278, + 0.89, + 0.311 + ], + "angle": 0, + "content": "{clarkqian, wfeng}@tju.edu.cn, hanruize@suat-sz.edu.cn, {jh.hou, linqi.song}@cityu.edu.hk https://github.com/zekunqian/VOVTrack" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.485, + 0.71 + ], + "angle": 0, + "content": "Open-vocabulary multi-object tracking (OVMOT) represents a critical new challenge involving the detection and tracking of diverse object categories in videos, encompassing both seen categories (base classes) and unseen categories (novel classes). This issue amalgamates the complexities of open-vocabulary object detection (OVD) and multi-object tracking (MOT). Existing approaches to OVMOT often merge OVD and MOT methodologies as separate modules, not fully leveraging the video information. In this work, we propose VOvTrack, a novel method that integrates object states relevant to MOT and video-centric training to address this challenge from a video analysis standpoint. First, we consider the tracking-related state of the objects during tracking and propose a new prompt-guided attention mechanism for more accurate detection (localization and classification) of time-varying objects. Subsequently, we leverage raw video data without annotations for training by formulating a self-supervised object similarity learning technique to facilitate temporal object tracking (association). Experimental results underscore that VOvTrack establishes itself as a state-of-the-art solution for the open-vocabulary tracking task." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.73, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Multi-Object Tracking (MOT) is a fundamental task in computer vision and artificial intelligence, which is widely used for video surveillance, media understanding, etc [20, 21]. In the past years, plenty of datasets, e.g., MOT-20 [8], DanceTrack [48], KITTI [17], as well as the algorithms, e.g., SORT [51], Tracktor [46], FairMOT [56], have been proposed for MOT problem. However, most previous works" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.346, + 0.905, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.609, + 0.899, + 0.625 + ], + "angle": 0, + "content": "Figure 1. Comparison between OVTrack [27] and our methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.761 + ], + "angle": 0, + "content": "focus on the tracking of simple object categories, i.e., humans and vehicles. Actually, it is important for the perception of various categories of objects in many real-world applications. Some recent works have begun to study the tracking of generic objects. TAO [7] is the first large dataset for the generic MOT, which includes 2,907 videos and 833 object categories. However, these approaches are still limited to tracking predefined object categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "With the development of Artificial General Intelligence (AGI) and multi-modal foundation models, open-world object perception has become a popular topic. Open-vocabulary object detection (OVD) is a new and promising task because of its generic settings. It aims to identify the various categories of objects from an image, including both the categories that have been seen during training (i.e., base classes) and those not seen (i.e., novel classes). Although OVD has been studied in a series of works [9, 11, 23]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.423 + ], + "angle": 0, + "content": "the literature on open-vocabulary (multi-) object tracking is rare. The recent work [27] builds a benchmark for open-vocabulary multi-object tracking (OVMOT) based on TAO [7]. The authors also develop a simple framework for this problem, consisting of a detection head and a tracking head. The detection head is directly taken from an existing OVD algorithm, i.e., DetPro [13], which is used to detect the open-vocabulary categories of objects in each frame without considering any tracking-related factors. Then the tracking head is used to learn the similarities among the detected objects in different frames. Due to the lack of video data with open-vocabulary tracking annotations, the approach in [27] uses a data hallucination strategy to generate the image pairs for training the tracking head. However, the generated image pairs ignore the adjacent continuity and temporal variability of the objects in a video sequence, which is crucial for higher performance [44]. As discussed above, the existing methods for open-vocabulary tracking simply combine the approaches of OVD and MOT in series as independent modules. It does not consider the object states during tracking, e.g., mutual occlusion, motion blur, etc., and does not make use of the sequential information in the videos." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.434, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Therefore, in this work, we aim to handle the open-vocabulary object tracking from the standpoint of continuous videos. The comparison between our method and that in [27] can be intuitively seen from Figure 1. Specifically, we first consider the various states of the objects during tracking. Our basic idea is that damaged objects (e.g., with occlusion, motion blur, etc.) should be weakened for foreground object feature learning. This way, we model these states as the prompts, which are used to calculate the attention weights of each generated detection proposal during training the object detection network. This methodology can provide more accurate detection (localization and classification) results of the time-varying objects during tracking. Second, to fully utilize the raw videos without open-vocabulary tracking annotations, we formulate the temporal association (tracking) task as a constrained optimization problem. The basic idea is that each object should maintain consistency across different frames, allowing us to leverage object consistency to effectively learn the object appearance similarity features for temporal association. Specifically, we formulate the object appearance self-consistency and the spatial-appearance mutual-consistency learning problem, in which we also consider the object category consistency. These consistencies work together to learn the object appearance similarity features for the temporal association (tracking) sub-task, thus enhancing the overall performance of OVMOT. Notably, the existing OVMOT dataset TAO provides 534.1K (frames) of unlabeled video data. Compared to the 18.1K annotated samples, the unlabeled data have a huge potential to be explored for training. Our self-supervised training method obtains significant performance" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.093, + 0.904, + 0.136 + ], + "angle": 0, + "content": "improvements only using the raw data, thus effectively alleviating the burden of annotations for OVMOT. The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.137, + 0.904, + 0.227 + ], + "angle": 0, + "content": "- We propose a new tracking-related prompt-guided attention for the localization and classification (detection) in the open-vocabulary tracking problem. This method takes notice of the states of the time-varying objects during tracking, which is different from the open-vocabulary object detection from a single image." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.228, + 0.904, + 0.331 + ], + "angle": 0, + "content": "- We develop a self-supervised object similarity learning strategy for the temporal association (tracking) in the OVMOT problem. This strategy, for the first time, makes full use of the raw video data without annotation for OVMOT training, thus addressing the problem of training data shortage and eliminating the heavy annotation burden in OVMOT." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.378 + ], + "angle": 0, + "content": "- Experimental results on the public benchmark demonstrate that the proposed VOvTrack achieves the best performance using the same training dataset (annotations)." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.137, + 0.904, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.393, + 0.654, + 0.408 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.418, + 0.904, + 0.704 + ], + "angle": 0, + "content": "Multiple object tracking. The prevailing paradigm in MOT is the tracking-by-detection framework [1], which involves detecting objects in each frame first, and then associating objects across different frames using various cues such as object appearance features [3, 5, 16, 25, 32, 37, 42, 51], 2D motion features [14, 38, 43, 53, 59], or 3D motion features [22, 24, 30, 35, 45, 49]. Some methods leverage graph neural networks [4, 10] or transformers [31, 47, 55, 61] to learn the association relationships between objects, thereby enhancing tracking performance. To broaden the object categories of the MOT task, the TAO benchmark [7] has been proposed for studying MOT under the long-tail distribution of object categories. On this benchmark, relevant methods include AOA [12], GTR [61], TET [26], QDTrack [16], etc. While these methods perform well, they are still limited to pre-defined object categories, which makes them unsuitable for diverse open-world scenarios. Differently, this work handles the OVMOT problem, which contains categories not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.705, + 0.904, + 0.9 + ], + "angle": 0, + "content": "Open-world/vocabulary object detection. Open-world object detection identifies salient objects without predefined categories, treating classification as clustering based on similarity [9, 11, 23]. Open-vocabulary detection also identifies unseen categories but requires predicting specific object categories [54]. Early works [2, 40] trained detectors with text embeddings, while recent approaches leverage pre-trained vision-language models like CLIP [39] that connect visual concepts with textual descriptions for strong open-vocabulary classification. Building on this foundation, many works [18, 52, 60] utilize such models for open-vocabulary and few-shot object detection. Some studies [13, 57, 58] further enhance prompt embeddings" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7473" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "through prompt learning methods. Unlike these single-image detection methods, we propose a prompt-guided training method designed for open-vocabulary tracking in continuous videos, which can effectively enhance the detection performance for the videos." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.17, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Open-world/vocabulary object tracking. There are relatively few works addressing the task of open-world tracking. Related approaches aim to segment or track all the moving objects in the video [6, 33] or handle the generic object tracking [34-36] using a class-agnostic detector. Recently, the TAO-OW benchmark [29] is proposed to study open-world tracking problems, but its limitation lies in evaluating only class-agnostic tracking metrics without assessing class-related metrics. To make the setting more practical, OVTrack [27] first brings the setting of open vocabulary into the tracking task, which also develops a baseline method and benchmark based on the TAO dataset. However, the method in [27] directly uses an existing OVD algorithm for detection, and its training process only utilizes static image pairs and ignores the information from video sequences. Differently, we consider the tracking-related object states for detection, and also propose a self-supervised video-based training method designed for open-vocabulary tracking, making full use of video-level information to enhance the performance of open-vocabulary tracking. Recent work SLAck [28] addressed the lack of temporal continuity in OVTrack's training process by incorporating adjacent frame information. This approach attempts to obtain continuous (pseudo) labels by computing IoU between detection results of adjacent frames and sparsely annotated frames. Different from SLAck, this work designs the self-supervised loss for appearance learning without any annotation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.608, + 0.317, + 0.623 + ], + "angle": 0, + "content": "3. PROPOSED METHOD" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.307, + 0.649 + ], + "angle": 0, + "content": "3.1.VOVTrack Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.869 + ], + "angle": 0, + "content": "OVMOT requires localizing, tracking, and recognizing the objects in a video, whose problem formulation is provided in Appendix 1. We first describe the framework of VOVTrack, which mainly includes object localization, object classification, and temporal association modules, as shown in Figure 2. For improving the localization and classification, in Section 3.2, we design a tracking-state-aware prompt-guided attention mechanism, which can help the network learn more effective object detection features. For learning the temporal association similarity, in Section 3.3, we propose a video-based self-supervised method to train the association network, which considers the appearance and spatial self/mutual-consistency and category consistency, thereby enhancing the tracking performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Localization: We employ the class-agnostic object proposal generation approach in Faster R-CNN [41] to localize" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "objects for both base and novel categories \\(\\mathcal{C}\\) in the video. As supported by prior research [6, 18, 27, 60], the localization strategy has shown robust generalization capabilities toward the novel object class \\(\\mathcal{C}^{\\mathrm{novel}}\\). During the training phase, we leverage the above-mentioned generated RPN proposals as the initial object candidates \\(P\\). The localization result of each candidate \\(r \\in P\\) is the bounding box \\(\\mathbf{b} \\in \\mathbb{R}^4\\). For each \\(\\mathbf{b}\\), we also obtain a confidence \\(p_c \\in \\mathbb{R}\\) derived from the classification head. To refine the localization candidates, besides the classical non-maximum suppression (NMS), we also use this confidence \\(p_c\\). To obtain a more effective \\(p_c \\in \\mathbb{R}\\) in the classification head, we propose a prompt-guided attention training strategy in later Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.544 + ], + "angle": 0, + "content": "Classification: Existing closed-set MOT trackers only track several specific categories of objects, which do not need to provide the class of each tracked object. In this way, classification is a new sub-task of the OVMOT. To enable the framework to classify open-vocabulary object classes, following the OV detection algorithms [18, 52, 60], we leverage the knowledge from the pre-trained model, i.e., CLIP [39], to help the network recognize objects belonging to the novel classes \\(\\mathcal{C}^{\\mathrm{novel}}\\). We distill the classification head using the CLIP model to empower the network's classification head to recognize new objects. Specifically, as shown in Figure 2, after obtaining the RoI feature embeddings \\(\\mathbf{f}_r\\) from the localization head, we design the classification head with a text branch and an image branch to generate embeddings \\(\\mathbf{f}_r^{\\mathrm{text}}\\) and \\(\\mathbf{f}_r^{\\mathrm{img}}\\) for each \\(\\mathbf{f}_r\\). The supervision of these heads comes from the CLIP text and image embeddings. We use the method in [13] for CLIP encoder pre-training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.905, + 0.621 + ], + "angle": 0, + "content": "First, we align the text branch with the CLIP text encoder. For \\(\\forall c\\in \\mathcal{C}^{\\mathrm{base}}\\) , the class text embedding \\(\\mathbf{t}_c\\) of class \\(c\\) can be generated by the CLIP text encoder \\(\\mathcal{T}(\\cdot)\\) as \\(\\mathbf{t}_c = \\mathcal{T}(c)\\) . We compute the affinity between the predicted text embeddings \\(\\mathbf{f}_r^{\\mathrm{text}}\\) and their CLIP counterpart \\(\\mathbf{t}_c\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.629, + 0.951, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\left. \\right. \\mathrm {z} (r) = \\left[ \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {\\mathrm {b g}}\\right), \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {1}\\right), \\dots , \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {| C ^ {\\text {b a s e}} |}\\right)\\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.652, + 0.905, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{| P |} \\sum_ {r \\in P} w _ {r} \\mathrm {L} _ {\\mathrm {C E}} \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathrm {z} (r)}{\\lambda}\\right), c _ {r}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.905, + 0.854 + ], + "angle": 0, + "content": "where \\(\\mathbf{t}_{\\mathrm{bg}}\\) is a background prompt learned by treating the background candidates as a new class, \\(w_{r}\\) is the tracking-related prompt-guided attention weight (described in Section 3.2), \\(\\lambda\\) is a temperature parameter, \\(\\mathrm{LCE}\\) is the cross-entropy loss and \\(c_{r}\\) is the class label of \\(r\\). Then, we align each image branch embedding \\(\\mathbf{f}_r^{\\mathrm{img}}\\) with the CLIP image encoder \\(\\mathcal{I}(\\cdot)\\). For each candidate object \\(r\\), we input the corresponding cropped image to \\(\\mathcal{I}(\\cdot)\\) and get the CLIP image embedding \\(\\mathbf{i}_r\\). We minimize the distance between the corresponding \\(\\mathbf{f}_r^{\\mathrm{img}}\\) and \\(\\mathbf{i}_r\\) as \\(\\mathcal{L}_{\\mathrm{image}} = \\frac{1}{|P|}\\sum_{r\\in P}\\left\\|\\mathbf{f}_r^{\\mathrm{img}} - \\mathbf{i}_r\\right\\|\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In the testing stage, with the embeddings \\(\\mathbf{f}_r^{\\mathrm{text}}\\) and \\(\\mathbf{f}_r^{\\mathrm{img}}\\) derived from the text branch and image branch, respectively, we can obtain the corresponding classification probabilities" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7474" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.077, + 0.865, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 2. The training framework consists of three parts: first is the localization head used to localize objects of all categories in the video as region candidates; the second is the CLIP distilled classification head consisting of image and text branches, which uses tracking-state-aware prompts to guide the model in focusing on object states while learning classification features, thereby better distinguishing the OV categories; and the third part is the association head that utilizes self/mutual-consistency between the same objects in different frames to learn association features in a self-supervised way." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.482, + 0.392 + ], + "angle": 0, + "content": "\\(p_c^{\\mathrm{text}}\\) and \\(p_c^{\\mathrm{img}}\\) of each \\(r\\) belonging to the class \\(c\\), using the \\(\\mathbf{z}(r)\\) and softmax operation in Eq. (1). After that, we use the fusion strategy in [18] to calculate the final classification probability \\(p_c\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.483, + 0.454 + ], + "angle": 0, + "content": "Association: The association head aims to associate the detected objects with the same identification across frames, the main purpose of which is to learn the object features for similarity measurement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.683 + ], + "angle": 0, + "content": "To learn the association model, given the detected objects from the localization head, we use appearance embeddings to extract the features for association. For training the appearance embedding model, a straightforward method is to select an object as the anchor and its positive/negative samples for learning the object similarity. This method requires object identification (ID) annotations for positive/negative sample selection. However, as an emerging problem, OVMOT does not have enough available training video datasets with tracking ID labels. Previous work [27] uses data augmentation to generate the image pairs for training the association head, which, however, ignores the temporal information in the videos. In this work, we propose to leverage the raw videos to train the association network in a self-supervised manner in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.696, + 0.483, + 0.713 + ], + "angle": 0, + "content": "3.2. Tracking-State-Aware Prompt based Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tracking-state-aware prompts. In the classification head of existing open-vocabulary detection methods, when selecting region candidates for calculating classification loss, they only consider whether the maximal IOU between the candidates and ground-truth boxes exceeds a threshold. For this tracking problem, we further consider whether the states of the candidates are appropriate for training the detection network. As is well known, objects present many specific states during tracking, such as occlusion/out-of-view/motion blur, etc, which occur more frequently than in static image object detection. So, it is important to identify such object states to achieve more accurate detection and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.391 + ], + "angle": 0, + "content": "tracking results. However, these states have often been overlooked in past methods because the labels of these states are difficult to obtain, not to mention incorporating them into the network training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.908, + 0.587 + ], + "angle": 0, + "content": "Prompts, as a burgeoning concept, can be used to bridge the gap between the vision and language data based on the cross-modal pre-trained models. We consider using specially designed prompts to perceive the tracking states of the objects and integrate such states into the model training. We refer to such prompts as 'tracking-state-aware prompts.' We specifically employ pairs of adjectives with opposite meanings to describe the object states during tracking, e.g., 'unoccluded and occluded.' As shown in the classification head of Figure 2, we add \\( M \\) pairs of tracking-state-aware prompt pairs, denoted as \\( \\{p_{\\mathrm{pos}}^m, p_{\\mathrm{neg}}^m\\}_{m=1}^M \\), into our framework model. Next, we present how to use these prompts during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Prompt-guided attention. To utilize these tracking-state-aware prompts in guiding model training, we encode these prompts through the CLIP text encoder \\(\\mathcal{T}(\\cdot)\\) to obtain state embeddings \\(\\{\\mathbf{t}_{\\mathrm{pos}}^{m},\\mathbf{t}_{\\mathrm{neg}}^{m}\\}_{m = 1}^{M}\\), and calculate the prompt-guided attention weight \\(w_{r}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.667, + 0.86, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} (r, m) = \\left[ \\cos \\left(\\mathbf {f} _ {r} ^ {\\mathrm {t e x t}}, \\mathbf {t} _ {\\mathrm {p o s}} ^ {m}\\right), \\cos \\left(\\mathbf {f} _ {r} ^ {\\mathrm {t e x t}}, \\mathbf {t} _ {\\mathrm {n e g}} ^ {m}\\right) \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.689, + 0.905, + 0.73 + ], + "angle": 0, + "content": "\\[\nw _ {r} = \\frac {1}{| M |} \\sum_ {m = 1} ^ {M} \\left(\\operatorname {s o f t m a x} \\frac {\\mathrm {z} (r , m)}{\\lambda}\\right) _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_r^{\\mathrm{text}}\\) is the text embedding of a region candidate \\(r\\), and \\((\\cdot)_1\\) represents using the probability of the positive state as the attention for this pair. Note that, in our definition, the positive states always denote that the object state is beneficial to the object feature learning, e.g., unoccluded, unobsured. From the above analysis, we know that the attention obtained from the tracking-state-aware prompts evaluates the various object states, resulting in prompt-guided attention values \\(w_r \\in [0, 1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Piecewise weighting strategy. Next, we use this prompt-guided attention to help the model better utilize" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7475" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "high-quality candidates for training and filter out the low-quality noisy candidates. Specifically, we divide the \\(w_{r}\\) into three levels: low \\((d_{\\mathrm{low}})\\), medium and high \\((d_{\\mathrm{high}})\\). For embedding features \\(\\mathbf{f}_r^{\\mathrm{text}}\\) with \\(w_{r} < d_{\\mathrm{low}}\\), we regard such features as low-quality states, and filter them out during training, by assigning \\(w_{r}\\) to 0. For \\(d_{\\mathrm{low}} \\leq w_{r} \\leq d_{\\mathrm{high}}\\), we regard that even if these features are not of high quality, they still contribute to training, and retain their original weights \\(w_{r}\\). For \\(w_{r} > d_{\\mathrm{high}}\\), we regard the features of these regions as particularly suitable ones for the model to learn tracking-related features, so we assign \\(w_{r}\\) to 1 as a reward. After that, we apply the re-assigned \\(w_{r}\\) to Eq. (1), thereby integrating the object tracking-related state into the model training, which enables the network to better learn the object representations specifically for the tracking task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.331, + 0.471, + 0.347 + ], + "angle": 0, + "content": "3.3. Self-supervised Association with Raw Videos" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.414 + ], + "angle": 0, + "content": "Considering the lack of annotated videos for OVMOT, we develop a self-supervised approach to train the association network by leveraging the consistency among the objects with the same identity along the video." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.415, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Formulation. To learn the object appearance feature, we consider two aspects. The first one is self-consistency factor, i.e., the self-consistency of appearance for the same object at different times. The second one is mutual-consistency factor, i.e., the mutual-consistency between the appearance and motion cues during tracking. In this way, we formulate the optimized objective function as" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.535, + 0.484, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\substack {S = S _ {\\text {self}} \\left(t _ {i}, t _ {j}\\right) + \\alpha \\cdot S _ {\\text {mutual}} \\left(t _ {i}, t _ {j}\\right), \\\\ \\left. \\quad \\right.}} \\tag{3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.555, + 0.334, + 0.57 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad \\forall t _ {i}, t _ {j} \\in \\mathcal {T} _ {c}, \\quad \\forall c \\in \\mathcal {C},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.749 + ], + "angle": 0, + "content": "where \\(S\\) represents the overall consistency objective to be maximized, \\(S_{\\mathrm{self}}\\) and \\(S_{\\mathrm{mutual}}\\) denote the self-consistency and mutual-consistency measures respectively, while \\(\\alpha\\) is a weight to balance them. Besides, considering the diversity of object categories in OVMOT, we add the object category consistency constraint for consistency learning. Specifically, we construct the intervals, i.e., \\(\\mathcal{T}_c\\), containing several frames only with the same object category \\(c\\), in which we select \\(t_i, t_j\\). This is because we aim to learn the feature in a self-supervised manner without ID annotation, and the objects with various categories may bring about interference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\(\\Theta\\) Category-consistency constraint. First, we consider the category consistency constraint in \\(\\mathcal{T}_c\\). For the interval splitting of \\(\\mathcal{T}_c\\) in Eq. (3), we split the original videos into several segments of length \\(L\\) and randomly sample the shorter sub-segments with various lengths from each segment. These short-term sub-segments are then concatenated to form the training sequence. Such training sequences include short-long-term intervals. Specifically, we select the adjacent frames from the same sub-segment, which allows the association head to learn the consistency objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "under minor object differences. We also select the long-interval video frames from different sub-segments, which allow the association head to learn the similarity and variation of objects under large differences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.392 + ], + "angle": 0, + "content": "After obtaining the sampled training sequences as discussed above, we utilize the localization head to extract object bounding boxes from each frame in the sequence. Since we only use unlabeled raw videos for training, we cannot directly obtain the object categories. To address this issue, we employ a clustering approach to group the bounding boxes based on their classification features. Specifically, we utilize the classification head to obtain the predicted category probability distributions for the candidate objects from all frames. Then, we cluster the candidate objects' category distributions into different clusters. After clustering the candidate objects' category distributions into distinct groups, we can treat each cluster as a separate category. As shown in Figure 2, we conduct self-supervised learning on the objects that belong to the same category cluster and are sourced from different frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.909, + 0.59 + ], + "angle": 0, + "content": "\\(\\Theta\\) Appearance-self-consistency loss. We next model the consistency learning problem in Eq. (3) as a self-supervised learning task. After getting the training samples (object bounding boxes in different frames of the sample training sequence within the category clustering), we first use the CNN network to extract the appearance feature from the association head for all objects in frame \\(t_i\\) to construct the feature matrix \\(\\mathbf{F}_i\\). The main idea of the appearance-self-consistency loss is to leverage the self-consistency of the same objects' appearance features at different times (frames). We utilize the following two types of similarity transfer relationships, i.e., pair-wise symmetry and triple-wise cyclicity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Pair-wise consistency learning: For a pair of frames \\( t_i \\) and \\( t_j \\) in the video, we can get the object similarity matrix between them as \\( \\mathbf{M}_{ij} = \\mathbf{F}_i \\cdot (\\mathbf{F}_j)^{\\mathrm{T}} \\). We then compute the normalized similarity matrix \\( \\mathbf{S}_{ij} \\in [0,1] \\) based on the above similarity matrix \\( \\mathbf{M}_{ij} \\) by temperature-adaptive row softmax" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.677, + 0.907, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} (r, c) = f (\\mathbf {M} (r, c)) = \\frac {\\exp (\\tau \\mathbf {M} (r , c))}{\\sum_ {c = 1} ^ {C} \\exp (\\tau \\mathbf {M} (r , c))} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.765 + ], + "angle": 0, + "content": "where \\( r, c \\) represent the indices of row and column in \\( \\mathbf{M} \\) or \\( \\mathbf{S} \\), respectively. Here \\( C \\) is the number of columns for \\( \\mathbf{M} \\), and \\( \\tau \\) is the adaptive temperature adjustable parameter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The normalized similarity matrix \\(\\mathbf{S}_{ij}\\) can be regarded as a mapping (object association relation) from frame \\(t_i\\) to frame \\(t_j\\) (\\(\\mathbf{S}_{ij}: t_i \\to t_j\\)). In other words, we can select the maximum value of each row of \\(\\mathbf{S}_{ij}\\) as the matched objects between frames \\(t_i\\) and \\(t_j\\). Similarly, we get the reversed mapping from \\(t_j\\) to \\(t_i\\) as \\(\\mathbf{S}_{ji}: t_j \\to t_i\\). We calculate the pair-wise similarity matrix as \\(\\mathbf{E}_{\\mathrm{pair}} = \\mathbf{S}_{ij} \\cdot \\mathbf{S}_{ji}\\), where \\(\\mathbf{E}_{\\mathrm{pair}}\\) can be regarded as a symmetric mapping: \\(t_i \\rightleftarrows t_j\\), i.e., from \\(t_i\\) and return to \\(t_i\\). If the objects in frames \\(t_i\\) and \\(t_j\\) are the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7476" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "same, the result \\(\\mathbf{E}_{\\mathrm{pair}}\\) should be an identity matrix, which can be used to construct the self-supervision loss. However, due to the object differences in different frames, this condition may not always be satisfied. Therefore, we need to supervise it deliberately, which will be discussed later." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Triplet consistency learning: Besides the pair-wise symmetric similarity, we further consider the triple-wise circularly consistent similarity. Specifically, given the similarity matrix \\(\\mathbf{S}_{ij}\\) between two frames \\(t_i\\) and \\(t_j\\), as well as \\(\\mathbf{S}_{jk}\\) between frames \\(t_j\\) and \\(t_k\\), we aim to build the consistent similarity relation among this triplet, i.e., frames \\(t_i, t_j\\) and \\(t_k\\). To do this, we first calculate the triplet similarity matrix as \\(\\mathbf{M}_{ik} = \\mathbf{M}_{ij} \\cdot \\mathbf{M}_{jk}, (i \\neq j \\neq k)\\), where \\(\\mathbf{M}_{ik}\\) represents the similarity between the objects in frames \\(t_i\\) and \\(t_k\\), through the frame \\(t_j\\). We then compute the normalized similarity matrix using Eq. (4) as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.348, + 0.483, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {i k} = f \\left(\\mathbf {M} _ {i k}\\right), \\quad \\mathbf {S} _ {k i} = f \\left(\\left(\\mathbf {M} _ {i k}\\right) ^ {\\mathrm {T}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.456 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}_{ik}\\) represents the mapping \\(t_i\\rightarrow t_j\\rightarrow t_k\\) while \\(\\mathbf{S}_{ki}\\) represents the mapping along \\(t_k\\to t_j\\to t_i\\). Then, we calculate the transitive-similarity matrix: \\(\\mathbf{E}_{\\mathrm{trip}} = \\mathbf{S}_{ik}\\cdot \\mathbf{S}_{ki}\\), where \\(\\mathbf{E}_{\\mathrm{trip}}\\) can be regarded as the mapping: \\(t_i\\rightleftarrows t_j\\rightleftarrows t_k\\) (from \\(t_i\\) and return to \\(t_i\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.715 + ], + "angle": 0, + "content": "For convenience, we note the matrices \\(\\mathbf{E}_{\\mathrm{pair}}\\) and \\(\\mathbf{E}_{\\mathrm{trip}}\\) as \\(\\mathbf{E}\\), which should have the property that their diagonal elements are either 1 or 0, while all other elements are 0, in an ideal case. This means that the diagonal elements of \\(\\mathbf{E}\\) must be greater than or equal to the other elements. Based on this consideration, following [15, 50], we use the following loss \\(\\mathrm{L}(\\mathbf{E}) = \\sum_{r}\\mathrm{relu}\\left(\\max_{c\\neq r}\\mathbf{E}(r,c) - \\mathbf{E}(r,r) + m\\right)\\), where \\(r,c\\) denote the indices of row and column in \\(\\mathbf{E}\\). This loss denotes that we penalize the cases where the max non-diagonal element \\(\\mathbf{E}(r,c)\\) (\\(c\\neq r\\)) in a row \\(r\\), exceeds the corresponding diagonal elements \\(\\mathbf{E}(r,r)\\) with a margin \\(m\\). The margin \\(m\\geq 0\\) is a parameter used to control the punishment scope between \\(\\mathbf{E}(r,c)\\) and \\(\\mathbf{E}(r,r)\\). This loss helps \\(\\mathbf{E}\\) approach the identity matrix while addressing cases where there are unmatched targets in a row through a margin \\(m\\). Finally, we define our self-supervised self-consistency learning loss as \\(\\mathcal{L}_{\\mathrm{self}} = \\mathrm{L}\\left(\\mathbf{E}_{\\mathrm{pair}}\\right) + \\mathrm{L}\\left(\\mathbf{E}_{\\mathrm{trip}}\\right)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.851 + ], + "angle": 0, + "content": "\\(\\Theta\\) Mutual-consistency loss. The mutual-consistency loss makes use of the consistency between the spatial position continuity and the appearance similarity of the objects at different times. Given a bounding box list \\(B_{i}\\) and \\(B_{j}\\) of adjacent frames \\(i\\) and \\(j\\) in the video, we compute the Intersection over Union (IoU) matrix \\(\\mathbf{M}_{ij}\\) between bounding boxes \\(B_{i}\\) and \\(B_{j}\\), which quantifies the overlap between them. Based on a specified threshold IoUthres, we create an assignment matrix \\(\\mathbf{A}_{ij}\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.864, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {M} _ {i j} > \\operatorname {I o U} _ {\\text {t h r e s}}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "This assignment matrix \\(\\mathbf{A}_{ij}\\) indicates whether pairs of objects are considered similar, facilitating the optimization of the spatial consistency objective. Furthermore, we utilize the previously obtained normalized similarity matrix \\(\\mathbf{S}_{ij}\\) to establish the mutual-consistency loss, using the binary cross-entropy loss function \\(\\mathrm{L_{BCE}}\\), as \\(\\mathcal{L}_{\\mathrm{mutual}} = \\mathrm{L_{BCE}}(\\mathbf{S}_{ij},\\mathbf{A}_{ij})\\). This formulation allows us to measure the discrepancy between the normalized similarity matrix \\(\\mathbf{S}_{ij}\\) and the assignment matrix \\(\\mathbf{A}_{ij}\\), thereby enforcing spatial consistency across the detected objects in frames \\(t_i\\) and \\(t_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Finally, we have transformed the optimization problem in Eq. (3) into a self-supervised loss as \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{self}} + \\alpha \\cdot \\mathcal{L}_{\\mathrm{mutual}}\\), where \\(\\alpha\\) is a weighting coefficient. This formulation incorporates both the self- and mutual-consistencies, as well as category consistency, among the frames with different intervals, thereby effectively exploring the potentiality of videos to enhance the association for OVMOT." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.363, + 0.731, + 0.379 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.905, + 0.583 + ], + "angle": 0, + "content": "In Section 3.1, we use ResNet50 with FPN for localizing candidate regions. We set \\(\\lambda\\) in Eq. (1) as 0.007. In Section 3.2, we select four pairs of typical tracking state aware prompts, i.e., 'complete and incomplete,' 'unoccluded and occluded,' 'unobsured and obscured,' and 'recognizable and unrecognizable.' We set \\(d_{\\mathrm{low}}\\) as 0.3, \\(d_{\\mathrm{high}}\\) as 0.6. In Section 3.3, the segment length \\(L\\) is set as 24. We use the K-means clustering algorithm. We set the margin \\(m\\) as 0.5 and the IoUthres as 0.9. For \\(N\\) frames in each sampled video sequence, we select \\(C_N^2\\) and \\(C_N^3\\) groups of frames to calculate the self-consistency loss in Section 3.3. We also select \\(C_N^2\\) groups of frames to calculate the mutual-consistency loss. The weighting coefficient \\(\\alpha\\) is 0.9." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In the training stage, we first train the two-stage detector on the base classes of the LVIS dataset [19] referenced from [13], for 20 epochs, and use prompt-guided attention proposed in Section 3.2 to fine-tune the model's classification head for 6 epochs. Then we train the association head using static image pairs generated from [27] for 6 epochs and self-supervise the association head with TAO training dataset [7] without annotation for 14 epochs. In the inference stage, we select object candidates \\( P \\) by NMS with IoU threshold 0.5 for detection. Additionally, for tracking, we use appearance feature similarity to associate history tracks with the objects in the current frame. As in [27], we evaluate the similarity between history tracks and detected objects using both bi-directional softmax [16] and cosine similarity metrics. Following the classical MOT approaches, if the similarity score exceeds a matching threshold, we assign the object to the track. If the object doesn't correspond to any existing track, a new track is initiated if its detection confidence score surpasses a threshold, otherwise, it is disregarded. We set the similarity score threshold as 0.35 and maintain a track memory of 30 frames." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7477" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.077, + 0.907, + 0.105 + ], + "angle": 0, + "content": "Table 1. Result comparison. We evaluate our method against closed-set and open-vocabulary trackers on TAO validation and test sets. Here '√' denotes using the corresponding dataset with annotations, and † represents only using the raw video without any annotation (%)." + }, + { + "type": "table", + "bbox": [ + 0.208, + 0.105, + 0.79, + 0.284 + ], + "angle": 0, + "content": "
MethodClassesDataBaseNovel
Validation setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓27.145.624.711.022.542.724.40.4
TETer [26]✓✓✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [51]✓-✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓-✓✓28.347.420.517.022.746.719.32.2
OVTrack [27]✓-✓-35.549.336.920.227.848.833.61.5
SLAck [28]✓-✓✓37.255.037.619.131.154.337.81.3
VOVTrack (Ours)✓-✓†39.6 (↑ 2.4)58.940.919.135.3 (↑ 4.2)58.540.96.5
Test setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓25.843.223.510.620.239.720.90.2
TETer [26]✓✓✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [51]✓-✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓-✓✓26.044.119.014.818.039.013.41.7
OVTrack [27]✓-✓-32.645.635.416.924.141.828.71.8
SLAck [28]✓-✓✓34.752.535.616.127.149.130.02.0
VOVTrack (Ours)✓-✓†38.1 (↑ 3.4)57.341.515.429.8 (↑ 2.7)51.732.05.7
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.308, + 0.262, + 0.324 + ], + "angle": 0, + "content": "4. EXPERIMENTS" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.333, + 0.283, + 0.348 + ], + "angle": 0, + "content": "4.1. Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.522 + ], + "angle": 0, + "content": "We follow [27] to select the dataset and metrics for evaluation. We leverage the comprehensive and extensive vocabulary MOT dataset TAO [7] as our benchmark for OV-MOT. TAO is structured similarly to the LVIS [19] taxonomy, categorizing object classes based on their frequency of appearance into frequent, common, and rare groups. We use the rare classes defined in LVIS as \\(\\mathcal{C}^{\\mathrm{novel}}\\) and others as \\(\\mathcal{C}^{\\mathrm{base}}\\). We evaluate the performance with the comprehensive metric tracking-every-thing accuracy (TETA), which consists of the accuracies of localization (LocA), classification (ClsA), and association (AssocA)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.289, + 0.547 + ], + "angle": 0, + "content": "4.2. Comparative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.674 + ], + "angle": 0, + "content": "We compare our method with the latest trackers, TETer [26] and QDTrack [16], which are trained on both \\( \\mathcal{C}^{\\mathrm{base}} \\) and \\( \\mathcal{C}^{\\mathrm{novel}} \\). We include the classical trackers like DeepSORT [51] and Tractor++ [3] trained only on \\( \\mathcal{C}^{\\mathrm{base}} \\) and enhanced by OVD method ViLD [18] to achieve open-vocabulary tracking, as baselines. We also compare our method with the state-of-the-art OVMOT methods, i.e., OVTrack [27] and SLAck [28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, we present the OVMOT evaluation results on the TAO validation and test sets on both base and novel classes. We can see that our method outperforms all closed-set and open-vocabulary methods on both the validation and test sets. Even though QDTrack [16] and TETer [26] have seen novel classes during training on TAO, our method significantly outperforms them in all metrics on both base and novel classes. Additionally, DeepSORT and Tracktor++ with the open-vocabulary detector ViLD [18] are also trained in a supervised manner on TAO, while our method, trained in a self-supervised manner on TAO, surpasses them by a large margin. Although the results of SLAck, the most competitive method, are better than other comparison methods, our method outperforms it in all metrics significantly on both base and novel classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Particularly, compared to our baseline method OVTrack, our method achieves improvements of \\(4.1\\%\\) and \\(7.5\\%\\) in base and novel TETA, respectively, and a \\(5.0\\%\\), \\(7.3\\%\\) increase in novel ClsA and AssocA. In the test set, base and novel TETA also show improvements of \\(5.5\\%\\) and \\(5.7\\%\\), respectively. It is worth noting that the state-of-the-art method SLAck is trained on the TAO training dataset using the annotations, while our method only uses the raw video data of it without the annotations. Even so, our method outperforms it with a significant margin on the comprehensive TETA score, as shown in Table 1. Especially the improvement of novel category on the validation set has reached 5x the original level. This demonstrates the effectiveness of the proposed method, which is very promising for the OVMOT task. We provide more analysis from the standpoint of data quantity for training in Appendix 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Balance between classification and detection. Our method outperforms other methods on most metrics, except for the ClsA of the base category. We clarify that the lower ClsA score compared to OVTrack stems from our method's superior detection capability. Specifically, we classify many more detected objects, while OVTrack focuses on a smaller subset of high-confidence detections. This is supported by our comparison with SLAck, which also shows reduced classification performance on the base category despite better detection results (LocA), similar to our findings. In contrast, our method significantly improves performance on novel categories by identifying more unseen detections during testing. Overall, our method achieves a better balance between detection and classification tasks, as well as between base and novel categories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.667, + 0.804 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.856 + ], + "angle": 0, + "content": "We conduct the ablation studies on the main components proposed in our method, including prompt-guided attention, and the self-supervised learning related modules as:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.857, + 0.905, + 0.885 + ], + "angle": 0, + "content": "- w/o prompt-guided attention \\((w_{r})\\): Removing the prompt-guided attention in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- w/o piecewise weight strategy (\\(d_{\\mathrm{low}}\\) and \\(d_{\\mathrm{high}}\\)): Removing" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.857, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7478" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.198, + 0.09, + 0.798, + 0.103 + ], + "angle": 0, + "content": "Table 2. Ablation studies on modules of prompt-guided attention and self-supervised association \\((\\%)\\)" + }, + { + "type": "table", + "bbox": [ + 0.171, + 0.104, + 0.825, + 0.209 + ], + "angle": 0, + "content": "
ModuleAblation MethodsBaseNovel
TETALocAAssocAClsATETALocAAssocAClsA
Prompt-guided attentionw/o prompt-guided attention36.153.437.517.430.553.236.81.6
w/o piecewise weight strategy37.355.838.517.633.856.839.84.8
Self-consistency associationw/o whole self-consistency learning37.457.736.917.732.557.635.64.2
w/o short-long-sampling37.957.637.718.434.058.438.25.5
w/o category consistency38.157.839.417.133.358.437.93.6
w/o self-consistency38.257.538.818.334.258.339.15.3
w/o mutual-consistency39.158.640.118.634.158.838.15.6
CompleteVOVTrack (Ours)39.658.940.919.135.358.540.96.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.258 + ], + "angle": 0, + "content": "the piecewise weighting strategy proposed in Section 3.2, by directly using the \\(w_{r}\\) calculated by Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.288 + ], + "angle": 0, + "content": "- w/o whole self-supervised learning: Removing the whole self-supervised consistency learning strategy in Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.482, + 0.319 + ], + "angle": 0, + "content": "- w/o short-long-sampling: Removing the short-long-interval sampling strategy in Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.482, + 0.348 + ], + "angle": 0, + "content": "- w/o category consistency: Removing the category-aware object clustering in Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.482, + 0.378 + ], + "angle": 0, + "content": "- w/o self/mutual-consistency: Removing the self/mutual-consistency loss in Section 3.3." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Effectiveness of the state-aware prompt-guided attention. As shown in the first unit of Table 2, we can see that using prompt-guided attention as a weight coefficient during the training stage can effectively improve all metrics for both base and novel classes. The piecewise weighting strategy is also very effective, especially in improving the localization and classification accuracies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Effectiveness of the self-supervised consistency learning. As shown in the second unit of Table 2, we can see that using self-supervised loss can effectively improve all metrics for both base and novel classes. Either the self-consistency or the mutual-consistency for appearance learning is effective for the association task, i.e., 'AssocA.' Also, the interval sampling strategy allows samples to have a more diverse range of long and short cycles, improving the association-related metric. The category clustering strategy tries to gather the objects with the same category in a cluster, which is also helpful. To our surprise, the above strategies, in most cases, also effectively help improve classification ('ClsA') and localization ('LocA') accuracies. This is because the better association results can indirectly help the other two sub-tasks in OVMOT. We provide the discussion and analysis of the complementarity among different tasks in Appendix 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.283, + 0.773 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Illustrations of the proposed prompt-guided attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Figure 3 (a) shows cases of high prompt-guided attention, where we can see that the regions often have very distinctive category features with no occlusion, and the image quality of the region is very high. In contrast, Figure 3 (b) presents cases of low prompt-guided attention, highlighting regions that often exhibit issues such as significant blurriness, occlusion, unclear visibility, and difficulty in identification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.272 + ], + "angle": 0, + "content": "These factors are detrimental to training object localization and classification features, which are effectively weakened through state-aware prompt-guided attention." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.273, + 0.905, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Figure 3. Illustration of regions with high (a) or low (b) prompt-guided attention, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.515 + ], + "angle": 0, + "content": "Comparison result visualization. We show several visualization results in Figure 4. We can see that the proposed method provides better results than OVTrack [27]. In the first case of Figure 4 (a), our method provides an accurate object localization result and identifies the correct category. In the second case of Figure 4 (b), the tracking of a drone provided by the comparison method is wrong (different box colors denote different tracking IDs), and the classification is also not correct. Our method can track it continuously." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.577, + 0.905, + 0.603 + ], + "angle": 0, + "content": "Figure 4. Compared OVMOT results of ours and OVTrack on some cases with novel classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.668 + ], + "angle": 0, + "content": "In Appendix 4, we provide more analysis that includes additional cases on prompt-guided attention, failure cases, time complexity analysis, parameter settings analysis, and method generalization analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.671, + 0.695 + ], + "angle": 0, + "content": "5. CONCLUSION" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we have developed a new method, namely VOVTrack, to handle the OVMOT problem from the perspective of video object tracking. For this purpose, we first consider the object state during tracking and propose tracking-state-aware prompt-guided attention, which improves the accuracy of object localization and classification (detection). Second, we develop an object similarity learning strategy for the temporal association (tracking) using only the raw video data without annotation, which unveils the power of self-supervised learning for open-vocabulary tracking tasks. Experimental results demonstrate the effectiveness of the proposed method and each component for open-vocabulary tracking." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7479" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.251 + ], + "angle": 0, + "content": "This work was supported in part by the Young Scientists Fund of the National Natural Science Foundation of China (NSFC) under Grant 62402490, the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage) from Tianjin University, the Excellent Young Scientists Fund of NSFC under Grant 62422118, and the Guangdong Basic and Applied Basic Research Foundation under Grant 2025A1515010101." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.263, + 0.188, + 0.278 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele. People-tracking-by-detection and people-detection-by-tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In European Conference on Computer Vision, pages 384-400, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.482, + 0.512 + ], + "angle": 0, + "content": "[4] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In IEEE International Conference on Advanced Video and Signal based Surveillance, pages 1-6, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. MeMOT: Multi-object tracking with memory. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8090-8100, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.623 + ], + "angle": 0, + "content": "[6] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In IEEE/CVF International Conference on Computer Vision Workshops, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.484, + 0.679 + ], + "angle": 0, + "content": "[7] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. TAO: A large-scale benchmark for tracking any object. In European Conference on Computer Vision, pages 436-454, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[8] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. MOT20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.804 + ], + "angle": 0, + "content": "[9] Akshay Dhamija, Manuel Gunther, Jonathan Ventura, and Terrance Boult. The overlooked elephant of object detection: Open set. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1021-1030, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Thang Doan, Xin Li, Sima Behpour, Wenbin He, Liang Gou, and Liu Ren. Hyp-OW: Exploiting hierarchical structure" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.288, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "learning with hyperbolic distance enhances open world object detection. In AAAI Conference on Artificial Intelligence, pages 1555-1563, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to ECCV-TAO-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[14] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. StrongSORT: Make DeepSORT great again. IEEE Transactions on Multimedia, 25:8725-8737, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[15] Wei Feng, Feifan Wang, Ruize Han, Yiyang Gan, Zekun Qian, Junhui Hou, and Song Wang. Unveiling the power of self-supervision for multi-view multi-human association and tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.461 + ], + "angle": 0, + "content": "[16] Tobias Fischer, Thomas E Huang, Jiangmiao Pang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. QD-Track: Quasi-dense similarity learning for appearance-only multiple object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[20] Ruize Han, Wei Feng, Yujun Zhang, Jiewen Zhao, and Song Wang. Multiple human association and tracking from egocentric and complementary top views. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5225-5242, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[21] Ruize Han, Wei Feng, Feifan Wang, Zekun Qian, Haomin Yan, and Song Wang. Benchmarking the complementary-view multi-human association and tracking. International Journal of Computer Vision, 132:118-136, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[22] Kuan-Chih Huang, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Delving into motion-aware matching for monocular 3d object tracking. In IEEE/CVF International Conference on Computer Vision, pages 6909-6918, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[23] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[24] Jan Krejci, Oliver Kost, Ondrej Straka, and Jindrich Duník. Pedestrian tracking with monocular camera us-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7480" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "ing unconstrained 3D motion model. arXiv preprint arXiv:2403.11978, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[25] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese cnn for robust target association. In IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops, pages 33-40, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.484, + 0.25 + ], + "angle": 0, + "content": "[26] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In European Conference on Computer Vision, pages 498-515, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.253, + 0.484, + 0.322 + ], + "angle": 0, + "content": "[27] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[28] Siyuan Li, Lei Ke, Yung-Hsu Yang, Luigi Piccinelli, Mattia Segu, Martin Danelljan, and Luc Van Gool. Slack: Semantic, location, and appearance aware open-vocabulary tracking. In European Conference on Computer Vision, pages 1-18, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[29] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19045-19055, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.511 + ], + "angle": 0, + "content": "[30] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.484, + 0.581 + ], + "angle": 0, + "content": "[31] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[32] Anton Milan, S Hamid Rezatofighi, Anthony Dick, Ian Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI conference on Artificial Intelligence, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[33] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In European Conference on Computer Vision, pages 566-579, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[34] Aljosa Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In IEEE International Conference on Robotics and Automation, pages 3180-3187, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.484, + 0.841 + ], + "angle": 0, + "content": "[35] Aljosa Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3494–3501, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[36] Aljosa Osep, Paul Voiglaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In IEEE International Conference on Robotics and Automation, pages 10031-10037, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[37] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[38] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[40] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI conference on artificial intelligence, pages 11932–11939, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[42] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In IEEE/CVF International Conference on Computer Vision, pages 300-311, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[43] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[44] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 55(13s):1-37, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[45] Sarthak Sharma, Junaid Ahmed Ansari, J Krishna Murthy, and K Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3508-3515, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[46] Vivek Hari Sridhar, Dominique G Roche, and Simon Gingins. Tracktor: image-based automated tracking of animal movement and behaviour. Methods in Ecology and Evolution, 10(6):815-820, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[47] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[48] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[49] Li Wang, Xinyu Zhang, Wenyuan Qin, Xiaoyu Li, Jinghan Gao, Lei Yang, Zhiwei Li, Jun Li, Lei Zhu, Hong Wang, and Huaping Liu. CAMO-MOT: Combined Appearance-Motion" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7481" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Optimization for 3D Multi-Object Tracking With Camera-LiDAR Fusion. IEEE Transactions on Intelligent Transportation Systems, 24(11):11981-11996, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[50] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In European Conference on Computer Vision, pages 72–88, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[51] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In IEEE International Conference on Image Processing, pages 3645-3649, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[53] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision, pages 466-481, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[54] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[55] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. MOTR: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. FairMOT: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[61] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7482" + } + ] +] \ No newline at end of file diff --git a/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_origin.pdf b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..633b335632762376d19406dd41f1adb76b4b58f2 --- /dev/null +++ b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/cb6ae00b-bc8e-44a2-b37e-6f565b52690f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2ea2af04f4baab7c6a9904dee5392ee6ef8e0f42141a777c63dad8109e73b51 +size 5103702 diff --git a/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/full.md b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b464090f061ea3270071bb98067130113639f793 --- /dev/null +++ b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/full.md @@ -0,0 +1,301 @@ +# VOVTrack: Exploring the Potentially in Raw Videos for Open-Vocabulary Multi-Object Tracking + +Zekun Qian $^{1,3*}$ , Ruize Han $^{2*}$ , Junhui Hou $^{3}$ , Linqi Song $^{3}$ , Wei Feng $^{1\dagger}$ $^{1}$ College of Intelligence and Computing, Tianjin University + $^{2}$ Shenzhen University of Advanced Technology + $^{3}$ City University of Hong Kong + +{clarkqian, wfeng}@tju.edu.cn, hanruize@suat-sz.edu.cn, {jh.hou, linqi.song}@cityu.edu.hk https://github.com/zekunqian/VOVTrack + +# Abstract + +Open-vocabulary multi-object tracking (OVMOT) represents a critical new challenge involving the detection and tracking of diverse object categories in videos, encompassing both seen categories (base classes) and unseen categories (novel classes). This issue amalgamates the complexities of open-vocabulary object detection (OVD) and multi-object tracking (MOT). Existing approaches to OVMOT often merge OVD and MOT methodologies as separate modules, not fully leveraging the video information. In this work, we propose VOvTrack, a novel method that integrates object states relevant to MOT and video-centric training to address this challenge from a video analysis standpoint. First, we consider the tracking-related state of the objects during tracking and propose a new prompt-guided attention mechanism for more accurate detection (localization and classification) of time-varying objects. Subsequently, we leverage raw video data without annotations for training by formulating a self-supervised object similarity learning technique to facilitate temporal object tracking (association). Experimental results underscore that VOvTrack establishes itself as a state-of-the-art solution for the open-vocabulary tracking task. + +# 1. Introduction + +Multi-Object Tracking (MOT) is a fundamental task in computer vision and artificial intelligence, which is widely used for video surveillance, media understanding, etc [20, 21]. In the past years, plenty of datasets, e.g., MOT-20 [8], DanceTrack [48], KITTI [17], as well as the algorithms, e.g., SORT [51], Tracktor [46], FairMOT [56], have been proposed for MOT problem. However, most previous works + +![](images/12e3ff12a974479e7b8260f255dee613569b9a372bf380688c7625c695961d45.jpg) +Figure 1. Comparison between OVTrack [27] and our methods. + +focus on the tracking of simple object categories, i.e., humans and vehicles. Actually, it is important for the perception of various categories of objects in many real-world applications. Some recent works have begun to study the tracking of generic objects. TAO [7] is the first large dataset for the generic MOT, which includes 2,907 videos and 833 object categories. However, these approaches are still limited to tracking predefined object categories. + +With the development of Artificial General Intelligence (AGI) and multi-modal foundation models, open-world object perception has become a popular topic. Open-vocabulary object detection (OVD) is a new and promising task because of its generic settings. It aims to identify the various categories of objects from an image, including both the categories that have been seen during training (i.e., base classes) and those not seen (i.e., novel classes). Although OVD has been studied in a series of works [9, 11, 23], + +the literature on open-vocabulary (multi-) object tracking is rare. The recent work [27] builds a benchmark for open-vocabulary multi-object tracking (OVMOT) based on TAO [7]. The authors also develop a simple framework for this problem, consisting of a detection head and a tracking head. The detection head is directly taken from an existing OVD algorithm, i.e., DetPro [13], which is used to detect the open-vocabulary categories of objects in each frame without considering any tracking-related factors. Then the tracking head is used to learn the similarities among the detected objects in different frames. Due to the lack of video data with open-vocabulary tracking annotations, the approach in [27] uses a data hallucination strategy to generate the image pairs for training the tracking head. However, the generated image pairs ignore the adjacent continuity and temporal variability of the objects in a video sequence, which is crucial for higher performance [44]. As discussed above, the existing methods for open-vocabulary tracking simply combine the approaches of OVD and MOT in series as independent modules. It does not consider the object states during tracking, e.g., mutual occlusion, motion blur, etc., and does not make use of the sequential information in the videos. + +Therefore, in this work, we aim to handle the open-vocabulary object tracking from the standpoint of continuous videos. The comparison between our method and that in [27] can be intuitively seen from Figure 1. Specifically, we first consider the various states of the objects during tracking. Our basic idea is that damaged objects (e.g., with occlusion, motion blur, etc.) should be weakened for foreground object feature learning. This way, we model these states as the prompts, which are used to calculate the attention weights of each generated detection proposal during training the object detection network. This methodology can provide more accurate detection (localization and classification) results of the time-varying objects during tracking. Second, to fully utilize the raw videos without open-vocabulary tracking annotations, we formulate the temporal association (tracking) task as a constrained optimization problem. The basic idea is that each object should maintain consistency across different frames, allowing us to leverage object consistency to effectively learn the object appearance similarity features for temporal association. Specifically, we formulate the object appearance self-consistency and the spatial-appearance mutual-consistency learning problem, in which we also consider the object category consistency. These consistencies work together to learn the object appearance similarity features for the temporal association (tracking) sub-task, thus enhancing the overall performance of OVMOT. Notably, the existing OVMOT dataset TAO provides 534.1K (frames) of unlabeled video data. Compared to the 18.1K annotated samples, the unlabeled data have a huge potential to be explored for training. Our self-supervised training method obtains significant performance + +improvements only using the raw data, thus effectively alleviating the burden of annotations for OVMOT. The main contributions of this work are: + +- We propose a new tracking-related prompt-guided attention for the localization and classification (detection) in the open-vocabulary tracking problem. This method takes notice of the states of the time-varying objects during tracking, which is different from the open-vocabulary object detection from a single image. +- We develop a self-supervised object similarity learning strategy for the temporal association (tracking) in the OVMOT problem. This strategy, for the first time, makes full use of the raw video data without annotation for OVMOT training, thus addressing the problem of training data shortage and eliminating the heavy annotation burden in OVMOT. +- Experimental results on the public benchmark demonstrate that the proposed VOvTrack achieves the best performance using the same training dataset (annotations). + +# 2. Related Work + +Multiple object tracking. The prevailing paradigm in MOT is the tracking-by-detection framework [1], which involves detecting objects in each frame first, and then associating objects across different frames using various cues such as object appearance features [3, 5, 16, 25, 32, 37, 42, 51], 2D motion features [14, 38, 43, 53, 59], or 3D motion features [22, 24, 30, 35, 45, 49]. Some methods leverage graph neural networks [4, 10] or transformers [31, 47, 55, 61] to learn the association relationships between objects, thereby enhancing tracking performance. To broaden the object categories of the MOT task, the TAO benchmark [7] has been proposed for studying MOT under the long-tail distribution of object categories. On this benchmark, relevant methods include AOA [12], GTR [61], TET [26], QDTrack [16], etc. While these methods perform well, they are still limited to pre-defined object categories, which makes them unsuitable for diverse open-world scenarios. Differently, this work handles the OVMOT problem, which contains categories not seen during training. + +Open-world/vocabulary object detection. Open-world object detection identifies salient objects without predefined categories, treating classification as clustering based on similarity [9, 11, 23]. Open-vocabulary detection also identifies unseen categories but requires predicting specific object categories [54]. Early works [2, 40] trained detectors with text embeddings, while recent approaches leverage pre-trained vision-language models like CLIP [39] that connect visual concepts with textual descriptions for strong open-vocabulary classification. Building on this foundation, many works [18, 52, 60] utilize such models for open-vocabulary and few-shot object detection. Some studies [13, 57, 58] further enhance prompt embeddings + +through prompt learning methods. Unlike these single-image detection methods, we propose a prompt-guided training method designed for open-vocabulary tracking in continuous videos, which can effectively enhance the detection performance for the videos. + +Open-world/vocabulary object tracking. There are relatively few works addressing the task of open-world tracking. Related approaches aim to segment or track all the moving objects in the video [6, 33] or handle the generic object tracking [34-36] using a class-agnostic detector. Recently, the TAO-OW benchmark [29] is proposed to study open-world tracking problems, but its limitation lies in evaluating only class-agnostic tracking metrics without assessing class-related metrics. To make the setting more practical, OVTrack [27] first brings the setting of open vocabulary into the tracking task, which also develops a baseline method and benchmark based on the TAO dataset. However, the method in [27] directly uses an existing OVD algorithm for detection, and its training process only utilizes static image pairs and ignores the information from video sequences. Differently, we consider the tracking-related object states for detection, and also propose a self-supervised video-based training method designed for open-vocabulary tracking, making full use of video-level information to enhance the performance of open-vocabulary tracking. Recent work SLAck [28] addressed the lack of temporal continuity in OVTrack's training process by incorporating adjacent frame information. This approach attempts to obtain continuous (pseudo) labels by computing IoU between detection results of adjacent frames and sparsely annotated frames. Different from SLAck, this work designs the self-supervised loss for appearance learning without any annotation. + +# 3. PROPOSED METHOD + +# 3.1.VOVTrack Framework + +OVMOT requires localizing, tracking, and recognizing the objects in a video, whose problem formulation is provided in Appendix 1. We first describe the framework of VOVTrack, which mainly includes object localization, object classification, and temporal association modules, as shown in Figure 2. For improving the localization and classification, in Section 3.2, we design a tracking-state-aware prompt-guided attention mechanism, which can help the network learn more effective object detection features. For learning the temporal association similarity, in Section 3.3, we propose a video-based self-supervised method to train the association network, which considers the appearance and spatial self/mutual-consistency and category consistency, thereby enhancing the tracking performance. + +Localization: We employ the class-agnostic object proposal generation approach in Faster R-CNN [41] to localize + +objects for both base and novel categories $\mathcal{C}$ in the video. As supported by prior research [6, 18, 27, 60], the localization strategy has shown robust generalization capabilities toward the novel object class $\mathcal{C}^{\mathrm{novel}}$ . During the training phase, we leverage the above-mentioned generated RPN proposals as the initial object candidates $P$ . The localization result of each candidate $r \in P$ is the bounding box $\mathbf{b} \in \mathbb{R}^4$ . For each $\mathbf{b}$ , we also obtain a confidence $p_c \in \mathbb{R}$ derived from the classification head. To refine the localization candidates, besides the classical non-maximum suppression (NMS), we also use this confidence $p_c$ . To obtain a more effective $p_c \in \mathbb{R}$ in the classification head, we propose a prompt-guided attention training strategy in later Section 3.2. + +Classification: Existing closed-set MOT trackers only track several specific categories of objects, which do not need to provide the class of each tracked object. In this way, classification is a new sub-task of the OVMOT. To enable the framework to classify open-vocabulary object classes, following the OV detection algorithms [18, 52, 60], we leverage the knowledge from the pre-trained model, i.e., CLIP [39], to help the network recognize objects belonging to the novel classes $\mathcal{C}^{\mathrm{novel}}$ . We distill the classification head using the CLIP model to empower the network's classification head to recognize new objects. Specifically, as shown in Figure 2, after obtaining the RoI feature embeddings $\mathbf{f}_r$ from the localization head, we design the classification head with a text branch and an image branch to generate embeddings $\mathbf{f}_r^{\mathrm{text}}$ and $\mathbf{f}_r^{\mathrm{img}}$ for each $\mathbf{f}_r$ . The supervision of these heads comes from the CLIP text and image embeddings. We use the method in [13] for CLIP encoder pre-training. + +First, we align the text branch with the CLIP text encoder. For $\forall c\in \mathcal{C}^{\mathrm{base}}$ , the class text embedding $\mathbf{t}_c$ of class $c$ can be generated by the CLIP text encoder $\mathcal{T}(\cdot)$ as $\mathbf{t}_c = \mathcal{T}(c)$ . We compute the affinity between the predicted text embeddings $\mathbf{f}_r^{\mathrm{text}}$ and their CLIP counterpart $\mathbf{t}_c$ as + +$$ +\left. \right. \mathrm {z} (r) = \left[ \cos \left(\mathbf {f} _ {r} ^ {\text {t e x t}}, \mathbf {t} _ {\mathrm {b g}}\right), \cos \left(\mathbf {f} _ {r} ^ {\text {t e x t}}, \mathbf {t} _ {1}\right), \dots , \cos \left(\mathbf {f} _ {r} ^ {\text {t e x t}}, \mathbf {t} _ {| C ^ {\text {b a s e}} |}\right)\right], +$$ + +$$ +\mathcal {L} _ {\text {t e x t}} = \frac {1}{| P |} \sum_ {r \in P} w _ {r} \mathrm {L} _ {\mathrm {C E}} \left(\operatorname {s o f t m a x} \left(\frac {\mathrm {z} (r)}{\lambda}\right), c _ {r}\right), \tag {1} +$$ + +where $\mathbf{t}_{\mathrm{bg}}$ is a background prompt learned by treating the background candidates as a new class, $w_{r}$ is the tracking-related prompt-guided attention weight (described in Section 3.2), $\lambda$ is a temperature parameter, $\mathrm{LCE}$ is the cross-entropy loss and $c_{r}$ is the class label of $r$ . Then, we align each image branch embedding $\mathbf{f}_r^{\mathrm{img}}$ with the CLIP image encoder $\mathcal{I}(\cdot)$ . For each candidate object $r$ , we input the corresponding cropped image to $\mathcal{I}(\cdot)$ and get the CLIP image embedding $\mathbf{i}_r$ . We minimize the distance between the corresponding $\mathbf{f}_r^{\mathrm{img}}$ and $\mathbf{i}_r$ as $\mathcal{L}_{\mathrm{image}} = \frac{1}{|P|}\sum_{r\in P}\left\|\mathbf{f}_r^{\mathrm{img}} - \mathbf{i}_r\right\|$ . + +In the testing stage, with the embeddings $\mathbf{f}_r^{\mathrm{text}}$ and $\mathbf{f}_r^{\mathrm{img}}$ derived from the text branch and image branch, respectively, we can obtain the corresponding classification probabilities + +![](images/7236f5697863c98b48d2d31df41e26401417f8dec5105744b2532c843d2ebc57.jpg) +Figure 2. The training framework consists of three parts: first is the localization head used to localize objects of all categories in the video as region candidates; the second is the CLIP distilled classification head consisting of image and text branches, which uses tracking-state-aware prompts to guide the model in focusing on object states while learning classification features, thereby better distinguishing the OV categories; and the third part is the association head that utilizes self/mutual-consistency between the same objects in different frames to learn association features in a self-supervised way. + +$p_c^{\mathrm{text}}$ and $p_c^{\mathrm{img}}$ of each $r$ belonging to the class $c$ , using the $\mathbf{z}(r)$ and softmax operation in Eq. (1). After that, we use the fusion strategy in [18] to calculate the final classification probability $p_c$ . + +Association: The association head aims to associate the detected objects with the same identification across frames, the main purpose of which is to learn the object features for similarity measurement. + +To learn the association model, given the detected objects from the localization head, we use appearance embeddings to extract the features for association. For training the appearance embedding model, a straightforward method is to select an object as the anchor and its positive/negative samples for learning the object similarity. This method requires object identification (ID) annotations for positive/negative sample selection. However, as an emerging problem, OVMOT does not have enough available training video datasets with tracking ID labels. Previous work [27] uses data augmentation to generate the image pairs for training the association head, which, however, ignores the temporal information in the videos. In this work, we propose to leverage the raw videos to train the association network in a self-supervised manner in Section 3.3. + +# 3.2. Tracking-State-Aware Prompt based Attention + +Tracking-state-aware prompts. In the classification head of existing open-vocabulary detection methods, when selecting region candidates for calculating classification loss, they only consider whether the maximal IOU between the candidates and ground-truth boxes exceeds a threshold. For this tracking problem, we further consider whether the states of the candidates are appropriate for training the detection network. As is well known, objects present many specific states during tracking, such as occlusion/out-of-view/motion blur, etc, which occur more frequently than in static image object detection. So, it is important to identify such object states to achieve more accurate detection and + +tracking results. However, these states have often been overlooked in past methods because the labels of these states are difficult to obtain, not to mention incorporating them into the network training. + +Prompts, as a burgeoning concept, can be used to bridge the gap between the vision and language data based on the cross-modal pre-trained models. We consider using specially designed prompts to perceive the tracking states of the objects and integrate such states into the model training. We refer to such prompts as 'tracking-state-aware prompts.' We specifically employ pairs of adjectives with opposite meanings to describe the object states during tracking, e.g., 'unoccluded and occluded.' As shown in the classification head of Figure 2, we add $M$ pairs of tracking-state-aware prompt pairs, denoted as $\{p_{\mathrm{pos}}^m, p_{\mathrm{neg}}^m\}_{m=1}^M$ , into our framework model. Next, we present how to use these prompts during training. + +Prompt-guided attention. To utilize these tracking-state-aware prompts in guiding model training, we encode these prompts through the CLIP text encoder $\mathcal{T}(\cdot)$ to obtain state embeddings $\{\mathbf{t}_{\mathrm{pos}}^{m},\mathbf{t}_{\mathrm{neg}}^{m}\}_{m = 1}^{M}$ , and calculate the prompt-guided attention weight $w_{r}$ as + +$$ +\mathbf {z} (r, m) = \left[ \cos \left(\mathbf {f} _ {r} ^ {\mathrm {t e x t}}, \mathbf {t} _ {\mathrm {p o s}} ^ {m}\right), \cos \left(\mathbf {f} _ {r} ^ {\mathrm {t e x t}}, \mathbf {t} _ {\mathrm {n e g}} ^ {m}\right) \right], +$$ + +$$ +w _ {r} = \frac {1}{| M |} \sum_ {m = 1} ^ {M} \left(\operatorname {s o f t m a x} \frac {\mathrm {z} (r , m)}{\lambda}\right) _ {1}, \tag {2} +$$ + +where $\mathbf{f}_r^{\mathrm{text}}$ is the text embedding of a region candidate $r$ , and $(\cdot)_1$ represents using the probability of the positive state as the attention for this pair. Note that, in our definition, the positive states always denote that the object state is beneficial to the object feature learning, e.g., unoccluded, unobsured. From the above analysis, we know that the attention obtained from the tracking-state-aware prompts evaluates the various object states, resulting in prompt-guided attention values $w_r \in [0, 1]$ . + +Piecewise weighting strategy. Next, we use this prompt-guided attention to help the model better utilize + +high-quality candidates for training and filter out the low-quality noisy candidates. Specifically, we divide the $w_{r}$ into three levels: low $(d_{\mathrm{low}})$ , medium and high $(d_{\mathrm{high}})$ . For embedding features $\mathbf{f}_r^{\mathrm{text}}$ with $w_{r} < d_{\mathrm{low}}$ , we regard such features as low-quality states, and filter them out during training, by assigning $w_{r}$ to 0. For $d_{\mathrm{low}} \leq w_{r} \leq d_{\mathrm{high}}$ , we regard that even if these features are not of high quality, they still contribute to training, and retain their original weights $w_{r}$ . For $w_{r} > d_{\mathrm{high}}$ , we regard the features of these regions as particularly suitable ones for the model to learn tracking-related features, so we assign $w_{r}$ to 1 as a reward. After that, we apply the re-assigned $w_{r}$ to Eq. (1), thereby integrating the object tracking-related state into the model training, which enables the network to better learn the object representations specifically for the tracking task. + +# 3.3. Self-supervised Association with Raw Videos + +Considering the lack of annotated videos for OVMOT, we develop a self-supervised approach to train the association network by leveraging the consistency among the objects with the same identity along the video. + +Formulation. To learn the object appearance feature, we consider two aspects. The first one is self-consistency factor, i.e., the self-consistency of appearance for the same object at different times. The second one is mutual-consistency factor, i.e., the mutual-consistency between the appearance and motion cues during tracking. In this way, we formulate the optimized objective function as + +$$ +\max _ {\substack {S = S _ {\text {self}} \left(t _ {i}, t _ {j}\right) + \alpha \cdot S _ {\text {mutual}} \left(t _ {i}, t _ {j}\right), \\ \left. \quad \right.}} \tag{3} +$$ + +$$ +s. t. \quad \forall t _ {i}, t _ {j} \in \mathcal {T} _ {c}, \quad \forall c \in \mathcal {C}, +$$ + +where $S$ represents the overall consistency objective to be maximized, $S_{\mathrm{self}}$ and $S_{\mathrm{mutual}}$ denote the self-consistency and mutual-consistency measures respectively, while $\alpha$ is a weight to balance them. Besides, considering the diversity of object categories in OVMOT, we add the object category consistency constraint for consistency learning. Specifically, we construct the intervals, i.e., $\mathcal{T}_c$ , containing several frames only with the same object category $c$ , in which we select $t_i, t_j$ . This is because we aim to learn the feature in a self-supervised manner without ID annotation, and the objects with various categories may bring about interference. + +$\Theta$ Category-consistency constraint. First, we consider the category consistency constraint in $\mathcal{T}_c$ . For the interval splitting of $\mathcal{T}_c$ in Eq. (3), we split the original videos into several segments of length $L$ and randomly sample the shorter sub-segments with various lengths from each segment. These short-term sub-segments are then concatenated to form the training sequence. Such training sequences include short-long-term intervals. Specifically, we select the adjacent frames from the same sub-segment, which allows the association head to learn the consistency objectives + +under minor object differences. We also select the long-interval video frames from different sub-segments, which allow the association head to learn the similarity and variation of objects under large differences. + +After obtaining the sampled training sequences as discussed above, we utilize the localization head to extract object bounding boxes from each frame in the sequence. Since we only use unlabeled raw videos for training, we cannot directly obtain the object categories. To address this issue, we employ a clustering approach to group the bounding boxes based on their classification features. Specifically, we utilize the classification head to obtain the predicted category probability distributions for the candidate objects from all frames. Then, we cluster the candidate objects' category distributions into different clusters. After clustering the candidate objects' category distributions into distinct groups, we can treat each cluster as a separate category. As shown in Figure 2, we conduct self-supervised learning on the objects that belong to the same category cluster and are sourced from different frames. + +$\Theta$ Appearance-self-consistency loss. We next model the consistency learning problem in Eq. (3) as a self-supervised learning task. After getting the training samples (object bounding boxes in different frames of the sample training sequence within the category clustering), we first use the CNN network to extract the appearance feature from the association head for all objects in frame $t_i$ to construct the feature matrix $\mathbf{F}_i$ . The main idea of the appearance-self-consistency loss is to leverage the self-consistency of the same objects' appearance features at different times (frames). We utilize the following two types of similarity transfer relationships, i.e., pair-wise symmetry and triple-wise cyclicity. + +Pair-wise consistency learning: For a pair of frames $t_i$ and $t_j$ in the video, we can get the object similarity matrix between them as $\mathbf{M}_{ij} = \mathbf{F}_i \cdot (\mathbf{F}_j)^{\mathrm{T}}$ . We then compute the normalized similarity matrix $\mathbf{S}_{ij} \in [0,1]$ based on the above similarity matrix $\mathbf{M}_{ij}$ by temperature-adaptive row softmax + +$$ +\mathbf {S} (r, c) = f (\mathbf {M} (r, c)) = \frac {\exp (\tau \mathbf {M} (r , c))}{\sum_ {c = 1} ^ {C} \exp (\tau \mathbf {M} (r , c))} \tag {4} +$$ + +where $r, c$ represent the indices of row and column in $\mathbf{M}$ or $\mathbf{S}$ , respectively. Here $C$ is the number of columns for $\mathbf{M}$ , and $\tau$ is the adaptive temperature adjustable parameter. + +The normalized similarity matrix $\mathbf{S}_{ij}$ can be regarded as a mapping (object association relation) from frame $t_i$ to frame $t_j$ ( $\mathbf{S}_{ij}: t_i \to t_j$ ). In other words, we can select the maximum value of each row of $\mathbf{S}_{ij}$ as the matched objects between frames $t_i$ and $t_j$ . Similarly, we get the reversed mapping from $t_j$ to $t_i$ as $\mathbf{S}_{ji}: t_j \to t_i$ . We calculate the pair-wise similarity matrix as $\mathbf{E}_{\mathrm{pair}} = \mathbf{S}_{ij} \cdot \mathbf{S}_{ji}$ , where $\mathbf{E}_{\mathrm{pair}}$ can be regarded as a symmetric mapping: $t_i \rightleftarrows t_j$ , i.e., from $t_i$ and return to $t_i$ . If the objects in frames $t_i$ and $t_j$ are the + +same, the result $\mathbf{E}_{\mathrm{pair}}$ should be an identity matrix, which can be used to construct the self-supervision loss. However, due to the object differences in different frames, this condition may not always be satisfied. Therefore, we need to supervise it deliberately, which will be discussed later. + +Triplet consistency learning: Besides the pair-wise symmetric similarity, we further consider the triple-wise circularly consistent similarity. Specifically, given the similarity matrix $\mathbf{S}_{ij}$ between two frames $t_i$ and $t_j$ , as well as $\mathbf{S}_{jk}$ between frames $t_j$ and $t_k$ , we aim to build the consistent similarity relation among this triplet, i.e., frames $t_i, t_j$ and $t_k$ . To do this, we first calculate the triplet similarity matrix as $\mathbf{M}_{ik} = \mathbf{M}_{ij} \cdot \mathbf{M}_{jk}, (i \neq j \neq k)$ , where $\mathbf{M}_{ik}$ represents the similarity between the objects in frames $t_i$ and $t_k$ , through the frame $t_j$ . We then compute the normalized similarity matrix using Eq. (4) as + +$$ +\mathbf {S} _ {i k} = f \left(\mathbf {M} _ {i k}\right), \quad \mathbf {S} _ {k i} = f \left(\left(\mathbf {M} _ {i k}\right) ^ {\mathrm {T}}\right), \tag {5} +$$ + +where $\mathbf{S}_{ik}$ represents the mapping $t_i\rightarrow t_j\rightarrow t_k$ while $\mathbf{S}_{ki}$ represents the mapping along $t_k\to t_j\to t_i$ . Then, we calculate the transitive-similarity matrix: $\mathbf{E}_{\mathrm{trip}} = \mathbf{S}_{ik}\cdot \mathbf{S}_{ki}$ , where $\mathbf{E}_{\mathrm{trip}}$ can be regarded as the mapping: $t_i\rightleftarrows t_j\rightleftarrows t_k$ (from $t_i$ and return to $t_i$ ). + +For convenience, we note the matrices $\mathbf{E}_{\mathrm{pair}}$ and $\mathbf{E}_{\mathrm{trip}}$ as $\mathbf{E}$ , which should have the property that their diagonal elements are either 1 or 0, while all other elements are 0, in an ideal case. This means that the diagonal elements of $\mathbf{E}$ must be greater than or equal to the other elements. Based on this consideration, following [15, 50], we use the following loss $\mathrm{L}(\mathbf{E}) = \sum_{r}\mathrm{relu}\left(\max_{c\neq r}\mathbf{E}(r,c) - \mathbf{E}(r,r) + m\right)$ , where $r,c$ denote the indices of row and column in $\mathbf{E}$ . This loss denotes that we penalize the cases where the max non-diagonal element $\mathbf{E}(r,c)$ ( $c\neq r$ ) in a row $r$ , exceeds the corresponding diagonal elements $\mathbf{E}(r,r)$ with a margin $m$ . The margin $m\geq 0$ is a parameter used to control the punishment scope between $\mathbf{E}(r,c)$ and $\mathbf{E}(r,r)$ . This loss helps $\mathbf{E}$ approach the identity matrix while addressing cases where there are unmatched targets in a row through a margin $m$ . Finally, we define our self-supervised self-consistency learning loss as $\mathcal{L}_{\mathrm{self}} = \mathrm{L}\left(\mathbf{E}_{\mathrm{pair}}\right) + \mathrm{L}\left(\mathbf{E}_{\mathrm{trip}}\right)$ . + +$\Theta$ Mutual-consistency loss. The mutual-consistency loss makes use of the consistency between the spatial position continuity and the appearance similarity of the objects at different times. Given a bounding box list $B_{i}$ and $B_{j}$ of adjacent frames $i$ and $j$ in the video, we compute the Intersection over Union (IoU) matrix $\mathbf{M}_{ij}$ between bounding boxes $B_{i}$ and $B_{j}$ , which quantifies the overlap between them. Based on a specified threshold IoUthres, we create an assignment matrix $\mathbf{A}_{ij}$ such that + +$$ +\mathbf {A} _ {i j} = \left\{ \begin{array}{l l} 1 & \text {i f} \mathbf {M} _ {i j} > \operatorname {I o U} _ {\text {t h r e s}}, \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {6} +$$ + +This assignment matrix $\mathbf{A}_{ij}$ indicates whether pairs of objects are considered similar, facilitating the optimization of the spatial consistency objective. Furthermore, we utilize the previously obtained normalized similarity matrix $\mathbf{S}_{ij}$ to establish the mutual-consistency loss, using the binary cross-entropy loss function $\mathrm{L_{BCE}}$ , as $\mathcal{L}_{\mathrm{mutual}} = \mathrm{L_{BCE}}(\mathbf{S}_{ij},\mathbf{A}_{ij})$ . This formulation allows us to measure the discrepancy between the normalized similarity matrix $\mathbf{S}_{ij}$ and the assignment matrix $\mathbf{A}_{ij}$ , thereby enforcing spatial consistency across the detected objects in frames $t_i$ and $t_j$ . + +Finally, we have transformed the optimization problem in Eq. (3) into a self-supervised loss as $\mathcal{L} = \mathcal{L}_{\mathrm{self}} + \alpha \cdot \mathcal{L}_{\mathrm{mutual}}$ , where $\alpha$ is a weighting coefficient. This formulation incorporates both the self- and mutual-consistencies, as well as category consistency, among the frames with different intervals, thereby effectively exploring the potentiality of videos to enhance the association for OVMOT. + +# 3.4. Implementation Details + +In Section 3.1, we use ResNet50 with FPN for localizing candidate regions. We set $\lambda$ in Eq. (1) as 0.007. In Section 3.2, we select four pairs of typical tracking state aware prompts, i.e., 'complete and incomplete,' 'unoccluded and occluded,' 'unobsured and obscured,' and 'recognizable and unrecognizable.' We set $d_{\mathrm{low}}$ as 0.3, $d_{\mathrm{high}}$ as 0.6. In Section 3.3, the segment length $L$ is set as 24. We use the K-means clustering algorithm. We set the margin $m$ as 0.5 and the IoUthres as 0.9. For $N$ frames in each sampled video sequence, we select $C_N^2$ and $C_N^3$ groups of frames to calculate the self-consistency loss in Section 3.3. We also select $C_N^2$ groups of frames to calculate the mutual-consistency loss. The weighting coefficient $\alpha$ is 0.9. + +In the training stage, we first train the two-stage detector on the base classes of the LVIS dataset [19] referenced from [13], for 20 epochs, and use prompt-guided attention proposed in Section 3.2 to fine-tune the model's classification head for 6 epochs. Then we train the association head using static image pairs generated from [27] for 6 epochs and self-supervise the association head with TAO training dataset [7] without annotation for 14 epochs. In the inference stage, we select object candidates $P$ by NMS with IoU threshold 0.5 for detection. Additionally, for tracking, we use appearance feature similarity to associate history tracks with the objects in the current frame. As in [27], we evaluate the similarity between history tracks and detected objects using both bi-directional softmax [16] and cosine similarity metrics. Following the classical MOT approaches, if the similarity score exceeds a matching threshold, we assign the object to the track. If the object doesn't correspond to any existing track, a new track is initiated if its detection confidence score surpasses a threshold, otherwise, it is disregarded. We set the similarity score threshold as 0.35 and maintain a track memory of 30 frames. + +Table 1. Result comparison. We evaluate our method against closed-set and open-vocabulary trackers on TAO validation and test sets. Here '√' denotes using the corresponding dataset with annotations, and † represents only using the raw video without any annotation (%). + +
MethodClassesDataBaseNovel
Validation setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓27.145.624.711.022.542.724.40.4
TETer [26]✓✓✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [51]✓-✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓-✓✓28.347.420.517.022.746.719.32.2
OVTrack [27]✓-✓-35.549.336.920.227.848.833.61.5
SLAck [28]✓-✓✓37.255.037.619.131.154.337.81.3
VOVTrack (Ours)✓-✓†39.6 (↑ 2.4)58.940.919.135.3 (↑ 4.2)58.540.96.5
Test setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓25.843.223.510.620.239.720.90.2
TETer [26]✓✓✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [51]✓-✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓-✓✓26.044.119.014.818.039.013.41.7
OVTrack [27]✓-✓-32.645.635.416.924.141.828.71.8
SLAck [28]✓-✓✓34.752.535.616.127.149.130.02.0
VOVTrack (Ours)✓-✓†38.1 (↑ 3.4)57.341.515.429.8 (↑ 2.7)51.732.05.7
+ +# 4. EXPERIMENTS + +# 4.1. Dataset and Metrics + +We follow [27] to select the dataset and metrics for evaluation. We leverage the comprehensive and extensive vocabulary MOT dataset TAO [7] as our benchmark for OV-MOT. TAO is structured similarly to the LVIS [19] taxonomy, categorizing object classes based on their frequency of appearance into frequent, common, and rare groups. We use the rare classes defined in LVIS as $\mathcal{C}^{\mathrm{novel}}$ and others as $\mathcal{C}^{\mathrm{base}}$ . We evaluate the performance with the comprehensive metric tracking-every-thing accuracy (TETA), which consists of the accuracies of localization (LocA), classification (ClsA), and association (AssocA). + +# 4.2. Comparative Results + +We compare our method with the latest trackers, TETer [26] and QDTrack [16], which are trained on both $\mathcal{C}^{\mathrm{base}}$ and $\mathcal{C}^{\mathrm{novel}}$ . We include the classical trackers like DeepSORT [51] and Tractor++ [3] trained only on $\mathcal{C}^{\mathrm{base}}$ and enhanced by OVD method ViLD [18] to achieve open-vocabulary tracking, as baselines. We also compare our method with the state-of-the-art OVMOT methods, i.e., OVTrack [27] and SLAck [28]. + +As shown in Table 1, we present the OVMOT evaluation results on the TAO validation and test sets on both base and novel classes. We can see that our method outperforms all closed-set and open-vocabulary methods on both the validation and test sets. Even though QDTrack [16] and TETer [26] have seen novel classes during training on TAO, our method significantly outperforms them in all metrics on both base and novel classes. Additionally, DeepSORT and Tracktor++ with the open-vocabulary detector ViLD [18] are also trained in a supervised manner on TAO, while our method, trained in a self-supervised manner on TAO, surpasses them by a large margin. Although the results of SLAck, the most competitive method, are better than other comparison methods, our method outperforms it in all metrics significantly on both base and novel classes. + +Particularly, compared to our baseline method OVTrack, our method achieves improvements of $4.1\%$ and $7.5\%$ in base and novel TETA, respectively, and a $5.0\%$ , $7.3\%$ increase in novel ClsA and AssocA. In the test set, base and novel TETA also show improvements of $5.5\%$ and $5.7\%$ , respectively. It is worth noting that the state-of-the-art method SLAck is trained on the TAO training dataset using the annotations, while our method only uses the raw video data of it without the annotations. Even so, our method outperforms it with a significant margin on the comprehensive TETA score, as shown in Table 1. Especially the improvement of novel category on the validation set has reached 5x the original level. This demonstrates the effectiveness of the proposed method, which is very promising for the OVMOT task. We provide more analysis from the standpoint of data quantity for training in Appendix 2. + +Balance between classification and detection. Our method outperforms other methods on most metrics, except for the ClsA of the base category. We clarify that the lower ClsA score compared to OVTrack stems from our method's superior detection capability. Specifically, we classify many more detected objects, while OVTrack focuses on a smaller subset of high-confidence detections. This is supported by our comparison with SLAck, which also shows reduced classification performance on the base category despite better detection results (LocA), similar to our findings. In contrast, our method significantly improves performance on novel categories by identifying more unseen detections during testing. Overall, our method achieves a better balance between detection and classification tasks, as well as between base and novel categories. + +# 4.3. Ablation Study + +We conduct the ablation studies on the main components proposed in our method, including prompt-guided attention, and the self-supervised learning related modules as: + +- w/o prompt-guided attention $(w_{r})$ : Removing the prompt-guided attention in Section 3.2. +- w/o piecewise weight strategy ( $d_{\mathrm{low}}$ and $d_{\mathrm{high}}$ ): Removing + +Table 2. Ablation studies on modules of prompt-guided attention and self-supervised association $(\%)$ + +
ModuleAblation MethodsBaseNovel
TETALocAAssocAClsATETALocAAssocAClsA
Prompt-guided attentionw/o prompt-guided attention36.153.437.517.430.553.236.81.6
w/o piecewise weight strategy37.355.838.517.633.856.839.84.8
Self-consistency associationw/o whole self-consistency learning37.457.736.917.732.557.635.64.2
w/o short-long-sampling37.957.637.718.434.058.438.25.5
w/o category consistency38.157.839.417.133.358.437.93.6
w/o self-consistency38.257.538.818.334.258.339.15.3
w/o mutual-consistency39.158.640.118.634.158.838.15.6
CompleteVOVTrack (Ours)39.658.940.919.135.358.540.96.5
+ +the piecewise weighting strategy proposed in Section 3.2, by directly using the $w_{r}$ calculated by Eq. (2). + +- w/o whole self-supervised learning: Removing the whole self-supervised consistency learning strategy in Section 3.3. +- w/o short-long-sampling: Removing the short-long-interval sampling strategy in Section 3.3. +- w/o category consistency: Removing the category-aware object clustering in Section 3.3. +- w/o self/mutual-consistency: Removing the self/mutual-consistency loss in Section 3.3. + +Effectiveness of the state-aware prompt-guided attention. As shown in the first unit of Table 2, we can see that using prompt-guided attention as a weight coefficient during the training stage can effectively improve all metrics for both base and novel classes. The piecewise weighting strategy is also very effective, especially in improving the localization and classification accuracies. + +Effectiveness of the self-supervised consistency learning. As shown in the second unit of Table 2, we can see that using self-supervised loss can effectively improve all metrics for both base and novel classes. Either the self-consistency or the mutual-consistency for appearance learning is effective for the association task, i.e., 'AssocA.' Also, the interval sampling strategy allows samples to have a more diverse range of long and short cycles, improving the association-related metric. The category clustering strategy tries to gather the objects with the same category in a cluster, which is also helpful. To our surprise, the above strategies, in most cases, also effectively help improve classification ('ClsA') and localization ('LocA') accuracies. This is because the better association results can indirectly help the other two sub-tasks in OVMOT. We provide the discussion and analysis of the complementarity among different tasks in Appendix 3. + +# 4.4. Qualitative Analysis + +# Illustrations of the proposed prompt-guided attention. + +Figure 3 (a) shows cases of high prompt-guided attention, where we can see that the regions often have very distinctive category features with no occlusion, and the image quality of the region is very high. In contrast, Figure 3 (b) presents cases of low prompt-guided attention, highlighting regions that often exhibit issues such as significant blurriness, occlusion, unclear visibility, and difficulty in identification. + +These factors are detrimental to training object localization and classification features, which are effectively weakened through state-aware prompt-guided attention. + +![](images/4c7a160ed6bafafceab2481a02fe378d01999a22bf32f0bb0c0ad7eb303ddc05.jpg) +Figure 3. Illustration of regions with high (a) or low (b) prompt-guided attention, respectively. + +Comparison result visualization. We show several visualization results in Figure 4. We can see that the proposed method provides better results than OVTrack [27]. In the first case of Figure 4 (a), our method provides an accurate object localization result and identifies the correct category. In the second case of Figure 4 (b), the tracking of a drone provided by the comparison method is wrong (different box colors denote different tracking IDs), and the classification is also not correct. Our method can track it continuously. + +![](images/7ce955364ce5531f6610dcb312b816fbfcdd2f99751cde9c3b8cd132ca13a71d.jpg) +Figure 4. Compared OVMOT results of ours and OVTrack on some cases with novel classes. + +In Appendix 4, we provide more analysis that includes additional cases on prompt-guided attention, failure cases, time complexity analysis, parameter settings analysis, and method generalization analysis. + +# 5. CONCLUSION + +In this work, we have developed a new method, namely VOVTrack, to handle the OVMOT problem from the perspective of video object tracking. For this purpose, we first consider the object state during tracking and propose tracking-state-aware prompt-guided attention, which improves the accuracy of object localization and classification (detection). Second, we develop an object similarity learning strategy for the temporal association (tracking) using only the raw video data without annotation, which unveils the power of self-supervised learning for open-vocabulary tracking tasks. Experimental results demonstrate the effectiveness of the proposed method and each component for open-vocabulary tracking. + +# Acknowledgment + +# References + +[1] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele. People-tracking-by-detection and people-detection-by-tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2 +[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In European Conference on Computer Vision, pages 384-400, 2018. 2 +[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 2, 7 +[4] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In IEEE International Conference on Advanced Video and Signal based Surveillance, pages 1-6, 2017. 2 +[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. MeMOT: Multi-object tracking with memory. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8090-8100, 2022. 2 +[6] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In IEEE/CVF International Conference on Computer Vision Workshops, 2019. 3 +[7] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. TAO: A large-scale benchmark for tracking any object. In European Conference on Computer Vision, pages 436-454, 2020. 1, 2, 6, 7 +[8] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. MOT20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1 +[9] Akshay Dhamija, Manuel Gunther, Jonathan Ventura, and Terrance Boult. The overlooked elephant of object detection: Open set. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1021-1030, 2020. 1, 2 +[10] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 2 +[11] Thang Doan, Xin Li, Sima Behpour, Wenbin He, Liang Gou, and Liu Ren. Hyp-OW: Exploiting hierarchical structure + +This work was supported in part by the Young Scientists Fund of the National Natural Science Foundation of China (NSFC) under Grant 62402490, the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage) from Tianjin University, the Excellent Young Scientists Fund of NSFC under Grant 62422118, and the Guangdong Basic and Applied Basic Research Foundation under Grant 2025A1515010101. +learning with hyperbolic distance enhances open world object detection. In AAAI Conference on Artificial Intelligence, pages 1555-1563, 2024. 1, 2 +[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to ECCV-TAO-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2 +[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2, 3, 6 +[14] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. StrongSORT: Make DeepSORT great again. IEEE Transactions on Multimedia, 25:8725-8737, 2023. 2 +[15] Wei Feng, Feifan Wang, Ruize Han, Yiyang Gan, Zekun Qian, Junhui Hou, and Song Wang. Unveiling the power of self-supervision for multi-view multi-human association and tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6 +[16] Tobias Fischer, Thomas E Huang, Jiangmiao Pang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. QD-Track: Quasi-dense similarity learning for appearance-only multiple object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7 +[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1 +[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 2, 3, 4, 7 +[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019. 6, 7 +[20] Ruize Han, Wei Feng, Yujun Zhang, Jiewen Zhao, and Song Wang. Multiple human association and tracking from egocentric and complementary top views. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5225-5242, 2022. 1 +[21] Ruize Han, Wei Feng, Feifan Wang, Zekun Qian, Haomin Yan, and Song Wang. Benchmarking the complementary-view multi-human association and tracking. International Journal of Computer Vision, 132:118-136, 2023. 1 +[22] Kuan-Chih Huang, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Delving into motion-aware matching for monocular 3d object tracking. In IEEE/CVF International Conference on Computer Vision, pages 6909-6918, 2023. 2 +[23] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2 +[24] Jan Krejci, Oliver Kost, Ondrej Straka, and Jindrich Duník. Pedestrian tracking with monocular camera us- + +ing unconstrained 3D motion model. arXiv preprint arXiv:2403.11978, 2024. 2 +[25] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese cnn for robust target association. In IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops, pages 33-40, 2016. 2 +[26] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In European Conference on Computer Vision, pages 498-515, 2022. 2, 7 +[27] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 1, 2, 3, 4, 6, 7, 8 +[28] Siyuan Li, Lei Ke, Yung-Hsu Yang, Luigi Piccinelli, Mattia Segu, Martin Danelljan, and Luc Van Gool. Slack: Semantic, location, and appearance aware open-vocabulary tracking. In European Conference on Computer Vision, pages 1-18, 2024. 3, 7 +[29] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19045-19055, 2022. 3 +[30] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2 +[31] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 2 +[32] Anton Milan, S Hamid Rezatofighi, Anthony Dick, Ian Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI conference on Artificial Intelligence, 2017. 2 +[33] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In European Conference on Computer Vision, pages 566-579, 2012. 3 +[34] Aljosa Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In IEEE International Conference on Robotics and Automation, pages 3180-3187, 2016. 3 +[35] Aljosa Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3494–3501, 2018. 2 +[36] Aljosa Osep, Paul Voiglaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In IEEE International Conference on Robotics and Automation, pages 10031-10037, 2020. 3 + +[37] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2 +[38] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2023. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3 +[40] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI conference on artificial intelligence, pages 11932–11939, 2020. 2 +[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3 +[42] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In IEEE/CVF International Conference on Computer Vision, pages 300-311, 2017. 2 +[43] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021. 2 +[44] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 55(13s):1-37, 2023. 2 +[45] Sarthak Sharma, Junaid Ahmed Ansari, J Krishna Murthy, and K Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3508-3515, 2018. 2 +[46] Vivek Hari Sridhar, Dominique G Roche, and Simon Gingins. Tracktor: image-based automated tracking of animal movement and behaviour. Methods in Ecology and Evolution, 10(6):815-820, 2019. 1 +[47] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2 +[48] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1 +[49] Li Wang, Xinyu Zhang, Wenyuan Qin, Xiaoyu Li, Jinghan Gao, Lei Yang, Zhiwei Li, Jun Li, Lei Zhu, Hong Wang, and Huaping Liu. CAMO-MOT: Combined Appearance-Motion + +Optimization for 3D Multi-Object Tracking With Camera-LiDAR Fusion. IEEE Transactions on Intelligent Transportation Systems, 24(11):11981-11996, 2023. 2 +[50] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In European Conference on Computer Vision, pages 72–88, 2020. 6 +[51] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In IEEE International Conference on Image Processing, pages 3645-3649, 2017. 1, 2, 7 +[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2, 3 +[53] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision, pages 466-481, 2018. 2 +[54] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2 +[55] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. MOTR: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675, 2022. 2 +[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. FairMOT: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1 +[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2 +[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2 +[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490, 2020. 2 +[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022. 2, 3 +[61] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 2 \ No newline at end of file diff --git a/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/images.zip b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f586ef3407db969b044911b2a024b03984a07fb9 --- /dev/null +++ b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e478046e7a96e4d88ae1706934af298fd495e5a392d7ba5a3aaf10daafbe3e5a +size 402787 diff --git a/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/layout.json b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc21ac65a2ac752938c5ecfd8635c31a2a513878 --- /dev/null +++ b/2025/VOVTrack_ Exploring the Potentiality in Raw Videos for Open-Vocabulary Multi-Object Tracking/layout.json @@ -0,0 +1,10565 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 102, + 481, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 102, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 129, + 102, + 481, + 140 + ], + "type": "text", + "content": "VOVTrack: Exploring the Potentially in Raw Videos for Open-Vocabulary Multi-Object Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": "Zekun Qian" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": ", Ruize Han" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": ", Junhui Hou" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": ", Linqi Song" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": ", Wei Feng" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": "College of Intelligence and Computing, Tianjin University \n" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": "Shenzhen University of Advanced Technology \n" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 160, + 468, + 219 + ], + "type": "text", + "content": "City University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 220, + 544, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 220, + 544, + 246 + ], + "spans": [ + { + "bbox": [ + 65, + 220, + 544, + 246 + ], + "type": "text", + "content": "{clarkqian, wfeng}@tju.edu.cn, hanruize@suat-sz.edu.cn, {jh.hou, linqi.song}@cityu.edu.hk https://github.com/zekunqian/VOVTrack" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 299, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 562 + ], + "type": "text", + "content": "Open-vocabulary multi-object tracking (OVMOT) represents a critical new challenge involving the detection and tracking of diverse object categories in videos, encompassing both seen categories (base classes) and unseen categories (novel classes). This issue amalgamates the complexities of open-vocabulary object detection (OVD) and multi-object tracking (MOT). Existing approaches to OVMOT often merge OVD and MOT methodologies as separate modules, not fully leveraging the video information. In this work, we propose VOvTrack, a novel method that integrates object states relevant to MOT and video-centric training to address this challenge from a video analysis standpoint. First, we consider the tracking-related state of the objects during tracking and propose a new prompt-guided attention mechanism for more accurate detection (localization and classification) of time-varying objects. Subsequently, we leverage raw video data without annotations for training by formulating a self-supervised object similarity learning technique to facilitate temporal object tracking (association). Experimental results underscore that VOvTrack establishes itself as a state-of-the-art solution for the open-vocabulary tracking task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 597, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 682 + ], + "type": "text", + "content": "Multi-Object Tracking (MOT) is a fundamental task in computer vision and artificial intelligence, which is widely used for video surveillance, media understanding, etc [20, 21]. In the past years, plenty of datasets, e.g., MOT-20 [8], DanceTrack [48], KITTI [17], as well as the algorithms, e.g., SORT [51], Tracktor [46], FairMOT [56], have been proposed for MOT problem. However, most previous works" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 274, + 553, + 477 + ], + "blocks": [ + { + "bbox": [ + 318, + 274, + 553, + 477 + ], + "lines": [ + { + "bbox": [ + 318, + 274, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 318, + 274, + 553, + 477 + ], + "type": "image", + "image_path": "12e3ff12a974479e7b8260f255dee613569b9a372bf380688c7625c695961d45.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 482, + 550, + 495 + ], + "lines": [ + { + "bbox": [ + 318, + 482, + 550, + 495 + ], + "spans": [ + { + "bbox": [ + 318, + 482, + 550, + 495 + ], + "type": "text", + "content": "Figure 1. Comparison between OVTrack [27] and our methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "text", + "content": "focus on the tracking of simple object categories, i.e., humans and vehicles. Actually, it is important for the perception of various categories of objects in many real-world applications. Some recent works have begun to study the tracking of generic objects. TAO [7] is the first large dataset for the generic MOT, which includes 2,907 videos and 833 object categories. However, these approaches are still limited to tracking predefined object categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "With the development of Artificial General Intelligence (AGI) and multi-modal foundation models, open-world object perception has become a popular topic. Open-vocabulary object detection (OVD) is a new and promising task because of its generic settings. It aims to identify the various categories of objects from an image, including both the categories that have been seen during training (i.e., base classes) and those not seen (i.e., novel classes). Although OVD has been studied in a series of works [9, 11, 23]," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7472" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": "the literature on open-vocabulary (multi-) object tracking is rare. The recent work [27] builds a benchmark for open-vocabulary multi-object tracking (OVMOT) based on TAO [7]. The authors also develop a simple framework for this problem, consisting of a detection head and a tracking head. The detection head is directly taken from an existing OVD algorithm, i.e., DetPro [13], which is used to detect the open-vocabulary categories of objects in each frame without considering any tracking-related factors. Then the tracking head is used to learn the similarities among the detected objects in different frames. Due to the lack of video data with open-vocabulary tracking annotations, the approach in [27] uses a data hallucination strategy to generate the image pairs for training the tracking head. However, the generated image pairs ignore the adjacent continuity and temporal variability of the objects in a video sequence, which is crucial for higher performance [44]. As discussed above, the existing methods for open-vocabulary tracking simply combine the approaches of OVD and MOT in series as independent modules. It does not consider the object states during tracking, e.g., mutual occlusion, motion blur, etc., and does not make use of the sequential information in the videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 343, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 343, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 343, + 294, + 712 + ], + "type": "text", + "content": "Therefore, in this work, we aim to handle the open-vocabulary object tracking from the standpoint of continuous videos. The comparison between our method and that in [27] can be intuitively seen from Figure 1. Specifically, we first consider the various states of the objects during tracking. Our basic idea is that damaged objects (e.g., with occlusion, motion blur, etc.) should be weakened for foreground object feature learning. This way, we model these states as the prompts, which are used to calculate the attention weights of each generated detection proposal during training the object detection network. This methodology can provide more accurate detection (localization and classification) results of the time-varying objects during tracking. Second, to fully utilize the raw videos without open-vocabulary tracking annotations, we formulate the temporal association (tracking) task as a constrained optimization problem. The basic idea is that each object should maintain consistency across different frames, allowing us to leverage object consistency to effectively learn the object appearance similarity features for temporal association. Specifically, we formulate the object appearance self-consistency and the spatial-appearance mutual-consistency learning problem, in which we also consider the object category consistency. These consistencies work together to learn the object appearance similarity features for the temporal association (tracking) sub-task, thus enhancing the overall performance of OVMOT. Notably, the existing OVMOT dataset TAO provides 534.1K (frames) of unlabeled video data. Compared to the 18.1K annotated samples, the unlabeled data have a huge potential to be explored for training. Our self-supervised training method obtains significant performance" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 107 + ], + "type": "text", + "content": "improvements only using the raw data, thus effectively alleviating the burden of annotations for OVMOT. The main contributions of this work are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 108, + 553, + 299 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 317, + 108, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 179 + ], + "type": "text", + "content": "- We propose a new tracking-related prompt-guided attention for the localization and classification (detection) in the open-vocabulary tracking problem. This method takes notice of the states of the time-varying objects during tracking, which is different from the open-vocabulary object detection from a single image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 180, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 553, + 262 + ], + "type": "text", + "content": "- We develop a self-supervised object similarity learning strategy for the temporal association (tracking) in the OVMOT problem. This strategy, for the first time, makes full use of the raw video data without annotation for OVMOT training, thus addressing the problem of training data shortage and eliminating the heavy annotation burden in OVMOT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 264, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 299 + ], + "type": "text", + "content": "- Experimental results on the public benchmark demonstrate that the proposed VOvTrack achieves the best performance using the same training dataset (annotations)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 311, + 400, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 400, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 400, + 323 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 331, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 557 + ], + "type": "text", + "content": "Multiple object tracking. The prevailing paradigm in MOT is the tracking-by-detection framework [1], which involves detecting objects in each frame first, and then associating objects across different frames using various cues such as object appearance features [3, 5, 16, 25, 32, 37, 42, 51], 2D motion features [14, 38, 43, 53, 59], or 3D motion features [22, 24, 30, 35, 45, 49]. Some methods leverage graph neural networks [4, 10] or transformers [31, 47, 55, 61] to learn the association relationships between objects, thereby enhancing tracking performance. To broaden the object categories of the MOT task, the TAO benchmark [7] has been proposed for studying MOT under the long-tail distribution of object categories. On this benchmark, relevant methods include AOA [12], GTR [61], TET [26], QDTrack [16], etc. While these methods perform well, they are still limited to pre-defined object categories, which makes them unsuitable for diverse open-world scenarios. Differently, this work handles the OVMOT problem, which contains categories not seen during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 558, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 712 + ], + "type": "text", + "content": "Open-world/vocabulary object detection. Open-world object detection identifies salient objects without predefined categories, treating classification as clustering based on similarity [9, 11, 23]. Open-vocabulary detection also identifies unseen categories but requires predicting specific object categories [54]. Early works [2, 40] trained detectors with text embeddings, while recent approaches leverage pre-trained vision-language models like CLIP [39] that connect visual concepts with textual descriptions for strong open-vocabulary classification. Building on this foundation, many works [18, 52, 60] utilize such models for open-vocabulary and few-shot object detection. Some studies [13, 57, 58] further enhance prompt embeddings" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "7473" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "through prompt learning methods. Unlike these single-image detection methods, we propose a prompt-guided training method designed for open-vocabulary tracking in continuous videos, which can effectively enhance the detection performance for the videos." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 134, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 134, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 295, + 467 + ], + "type": "text", + "content": "Open-world/vocabulary object tracking. There are relatively few works addressing the task of open-world tracking. Related approaches aim to segment or track all the moving objects in the video [6, 33] or handle the generic object tracking [34-36] using a class-agnostic detector. Recently, the TAO-OW benchmark [29] is proposed to study open-world tracking problems, but its limitation lies in evaluating only class-agnostic tracking metrics without assessing class-related metrics. To make the setting more practical, OVTrack [27] first brings the setting of open vocabulary into the tracking task, which also develops a baseline method and benchmark based on the TAO dataset. However, the method in [27] directly uses an existing OVD algorithm for detection, and its training process only utilizes static image pairs and ignores the information from video sequences. Differently, we consider the tracking-related object states for detection, and also propose a self-supervised video-based training method designed for open-vocabulary tracking, making full use of video-level information to enhance the performance of open-vocabulary tracking. Recent work SLAck [28] addressed the lack of temporal continuity in OVTrack's training process by incorporating adjacent frame information. This approach attempts to obtain continuous (pseudo) labels by computing IoU between detection results of adjacent frames and sparsely annotated frames. Different from SLAck, this work designs the self-supervised loss for appearance learning without any annotation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 481, + 194, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 194, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 194, + 493 + ], + "type": "text", + "content": "3. PROPOSED METHOD" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 502, + 187, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 187, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 187, + 514 + ], + "type": "text", + "content": "3.1.VOVTrack Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "content": "OVMOT requires localizing, tracking, and recognizing the objects in a video, whose problem formulation is provided in Appendix 1. We first describe the framework of VOVTrack, which mainly includes object localization, object classification, and temporal association modules, as shown in Figure 2. For improving the localization and classification, in Section 3.2, we design a tracking-state-aware prompt-guided attention mechanism, which can help the network learn more effective object detection features. For learning the temporal association similarity, in Section 3.3, we propose a video-based self-supervised method to train the association network, which considers the appearance and spatial self/mutual-consistency and category consistency, thereby enhancing the tracking performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Localization: We employ the class-agnostic object proposal generation approach in Faster R-CNN [41] to localize" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "objects for both base and novel categories " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " in the video. As supported by prior research [6, 18, 27, 60], the localization strategy has shown robust generalization capabilities toward the novel object class " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ". During the training phase, we leverage the above-mentioned generated RPN proposals as the initial object candidates " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ". The localization result of each candidate " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "r \\in P" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " is the bounding box " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{b} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", we also obtain a confidence " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "p_c \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " derived from the classification head. To refine the localization candidates, besides the classical non-maximum suppression (NMS), we also use this confidence " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "p_c" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ". To obtain a more effective " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "p_c \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " in the classification head, we propose a prompt-guided attention training strategy in later Section 3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "text", + "content": "Classification: Existing closed-set MOT trackers only track several specific categories of objects, which do not need to provide the class of each tracked object. In this way, classification is a new sub-task of the OVMOT. To enable the framework to classify open-vocabulary object classes, following the OV detection algorithms [18, 52, 60], we leverage the knowledge from the pre-trained model, i.e., CLIP [39], to help the network recognize objects belonging to the novel classes " + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "text", + "content": ". We distill the classification head using the CLIP model to empower the network's classification head to recognize new objects. Specifically, as shown in Figure 2, after obtaining the RoI feature embeddings " + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "text", + "content": " from the localization head, we design the classification head with a text branch and an image branch to generate embeddings " + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 313, + 228, + 553, + 430 + ], + "type": "text", + "content": ". The supervision of these heads comes from the CLIP text and image embeddings. We use the method in [13] for CLIP encoder pre-training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": "First, we align the text branch with the CLIP text encoder. For " + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\forall c\\in \\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": " , the class text embedding " + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": " of class " + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": " can be generated by the CLIP text encoder " + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c = \\mathcal{T}(c)" + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": " . We compute the affinity between the predicted text embeddings " + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": " and their CLIP counterpart " + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 582, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 582, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 582, + 514 + ], + "type": "interline_equation", + "content": "\\left. \\right. \\mathrm {z} (r) = \\left[ \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {\\mathrm {b g}}\\right), \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {1}\\right), \\dots , \\cos \\left(\\mathbf {f} _ {r} ^ {\\text {t e x t}}, \\mathbf {t} _ {| C ^ {\\text {b a s e}} |}\\right)\\right],", + "image_path": "993dfc12c15e355655f37a3eb22857de861dca5402e640d5a28c9298695cb424.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 516, + 553, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 516, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 516, + 553, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{| P |} \\sum_ {r \\in P} w _ {r} \\mathrm {L} _ {\\mathrm {C E}} \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathrm {z} (r)}{\\lambda}\\right), c _ {r}\\right), \\tag {1}", + "image_path": "5fa262fc92216538186d9aaf615673481585e34b7d54cf436c5bc15fb136b45c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{bg}}" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " is a background prompt learned by treating the background candidates as a new class, " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " is the tracking-related prompt-guided attention weight (described in Section 3.2), " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " is a temperature parameter, " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{LCE}" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " is the cross-entropy loss and " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " is the class label of " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": ". Then, we align each image branch embedding " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " with the CLIP image encoder " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": ". For each candidate object " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": ", we input the corresponding cropped image to " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " and get the CLIP image embedding " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_r" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": ". We minimize the distance between the corresponding " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_r" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{image}} = \\frac{1}{|P|}\\sum_{r\\in P}\\left\\|\\mathbf{f}_r^{\\mathrm{img}} - \\mathbf{i}_r\\right\\|" + }, + { + "bbox": [ + 313, + 555, + 553, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In the testing stage, with the embeddings " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " derived from the text branch and image branch, respectively, we can obtain the corresponding classification probabilities" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7474" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 60, + 529, + 198 + ], + "blocks": [ + { + "bbox": [ + 80, + 60, + 529, + 198 + ], + "lines": [ + { + "bbox": [ + 80, + 60, + 529, + 198 + ], + "spans": [ + { + "bbox": [ + 80, + 60, + 529, + 198 + ], + "type": "image", + "image_path": "7236f5697863c98b48d2d31df41e26401417f8dec5105744b2532c843d2ebc57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. The training framework consists of three parts: first is the localization head used to localize objects of all categories in the video as region candidates; the second is the CLIP distilled classification head consisting of image and text branches, which uses tracking-state-aware prompts to guide the model in focusing on object states while learning classification features, thereby better distinguishing the OV categories; and the third part is the association head that utilizes self/mutual-consistency between the same objects in different frames to learn association features in a self-supervised way." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "inline_equation", + "content": "p_c^{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "inline_equation", + "content": "p_c^{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "text", + "content": " of each " + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "text", + "content": " belonging to the class " + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "text", + "content": ", using the " + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{z}(r)" + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "text", + "content": " and softmax operation in Eq. (1). After that, we use the fusion strategy in [18] to calculate the final classification probability " + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "inline_equation", + "content": "p_c" + }, + { + "bbox": [ + 55, + 261, + 294, + 310 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 295, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 295, + 359 + ], + "type": "text", + "content": "Association: The association head aims to associate the detected objects with the same identification across frames, the main purpose of which is to learn the object features for similarity measurement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 540 + ], + "type": "text", + "content": "To learn the association model, given the detected objects from the localization head, we use appearance embeddings to extract the features for association. For training the appearance embedding model, a straightforward method is to select an object as the anchor and its positive/negative samples for learning the object similarity. This method requires object identification (ID) annotations for positive/negative sample selection. However, as an emerging problem, OVMOT does not have enough available training video datasets with tracking ID labels. Previous work [27] uses data augmentation to generate the image pairs for training the association head, which, however, ignores the temporal information in the videos. In this work, we propose to leverage the raw videos to train the association network in a self-supervised manner in Section 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 551, + 295, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 295, + 564 + ], + "type": "text", + "content": "3.2. Tracking-State-Aware Prompt based Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Tracking-state-aware prompts. In the classification head of existing open-vocabulary detection methods, when selecting region candidates for calculating classification loss, they only consider whether the maximal IOU between the candidates and ground-truth boxes exceeds a threshold. For this tracking problem, we further consider whether the states of the candidates are appropriate for training the detection network. As is well known, objects present many specific states during tracking, such as occlusion/out-of-view/motion blur, etc, which occur more frequently than in static image object detection. So, it is important to identify such object states to achieve more accurate detection and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 261, + 555, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 309 + ], + "type": "text", + "content": "tracking results. However, these states have often been overlooked in past methods because the labels of these states are difficult to obtain, not to mention incorporating them into the network training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 464 + ], + "type": "text", + "content": "Prompts, as a burgeoning concept, can be used to bridge the gap between the vision and language data based on the cross-modal pre-trained models. We consider using specially designed prompts to perceive the tracking states of the objects and integrate such states into the model training. We refer to such prompts as 'tracking-state-aware prompts.' We specifically employ pairs of adjectives with opposite meanings to describe the object states during tracking, e.g., 'unoccluded and occluded.' As shown in the classification head of Figure 2, we add " + }, + { + "bbox": [ + 313, + 309, + 555, + 464 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 309, + 555, + 464 + ], + "type": "text", + "content": " pairs of tracking-state-aware prompt pairs, denoted as " + }, + { + "bbox": [ + 313, + 309, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\{p_{\\mathrm{pos}}^m, p_{\\mathrm{neg}}^m\\}_{m=1}^M" + }, + { + "bbox": [ + 313, + 309, + 555, + 464 + ], + "type": "text", + "content": ", into our framework model. Next, we present how to use these prompts during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "text", + "content": "Prompt-guided attention. To utilize these tracking-state-aware prompts in guiding model training, we encode these prompts through the CLIP text encoder " + }, + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "text", + "content": " to obtain state embeddings " + }, + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_{\\mathrm{pos}}^{m},\\mathbf{t}_{\\mathrm{neg}}^{m}\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "text", + "content": ", and calculate the prompt-guided attention weight " + }, + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 313, + 465, + 555, + 525 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 330, + 528, + 526, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 528, + 526, + 544 + ], + "spans": [ + { + "bbox": [ + 330, + 528, + 526, + 544 + ], + "type": "interline_equation", + "content": "\\mathbf {z} (r, m) = \\left[ \\cos \\left(\\mathbf {f} _ {r} ^ {\\mathrm {t e x t}}, \\mathbf {t} _ {\\mathrm {p o s}} ^ {m}\\right), \\cos \\left(\\mathbf {f} _ {r} ^ {\\mathrm {t e x t}}, \\mathbf {t} _ {\\mathrm {n e g}} ^ {m}\\right) \\right],", + "image_path": "1e407dc37775360592d09dca1c7cbc917e7679c66ede9b867cbf2686390e6e2a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 545, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 331, + 545, + 553, + 578 + ], + "type": "interline_equation", + "content": "w _ {r} = \\frac {1}{| M |} \\sum_ {m = 1} ^ {M} \\left(\\operatorname {s o f t m a x} \\frac {\\mathrm {z} (r , m)}{\\lambda}\\right) _ {1}, \\tag {2}", + "image_path": "d38274fcd190d3e82d85c2936c777ae8e36bd925d51a0bfd2d61dbfbcc8d3ec0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": " is the text embedding of a region candidate " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "(\\cdot)_1" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": " represents using the probability of the positive state as the attention for this pair. Note that, in our definition, the positive states always denote that the object state is beneficial to the object feature learning, e.g., unoccluded, unobsured. From the above analysis, we know that the attention obtained from the tracking-state-aware prompts evaluates the various object states, resulting in prompt-guided attention values " + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "inline_equation", + "content": "w_r \\in [0, 1]" + }, + { + "bbox": [ + 313, + 582, + 555, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Piecewise weighting strategy. Next, we use this prompt-guided attention to help the model better utilize" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7475" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "high-quality candidates for training and filter out the low-quality noisy candidates. Specifically, we divide the " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " into three levels: low " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "(d_{\\mathrm{low}})" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ", medium and high " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "(d_{\\mathrm{high}})" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ". For embedding features " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r^{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "w_{r} < d_{\\mathrm{low}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ", we regard such features as low-quality states, and filter them out during training, by assigning " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " to 0. For " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{low}} \\leq w_{r} \\leq d_{\\mathrm{high}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ", we regard that even if these features are not of high quality, they still contribute to training, and retain their original weights " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "w_{r} > d_{\\mathrm{high}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": ", we regard the features of these regions as particularly suitable ones for the model to learn tracking-related features, so we assign " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " to 1 as a reward. After that, we apply the re-assigned " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " to Eq. (1), thereby integrating the object tracking-related state into the model training, which enables the network to better learn the object representations specifically for the tracking task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 288, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 288, + 274 + ], + "type": "text", + "content": "3.3. Self-supervised Association with Raw Videos" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 327 + ], + "type": "text", + "content": "Considering the lack of annotated videos for OVMOT, we develop a self-supervised approach to train the association network by leveraging the consistency among the objects with the same identity along the video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 328, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 296, + 413 + ], + "type": "text", + "content": "Formulation. To learn the object appearance feature, we consider two aspects. The first one is self-consistency factor, i.e., the self-consistency of appearance for the same object at different times. The second one is mutual-consistency factor, i.e., the mutual-consistency between the appearance and motion cues during tracking. In this way, we formulate the optimized objective function as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 423, + 296, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 423, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 86, + 423, + 296, + 442 + ], + "type": "interline_equation", + "content": "\\max _ {\\substack {S = S _ {\\text {self}} \\left(t _ {i}, t _ {j}\\right) + \\alpha \\cdot S _ {\\text {mutual}} \\left(t _ {i}, t _ {j}\\right), \\\\ \\left. \\quad \\right.}} \\tag{3}", + "image_path": "11ddcee33b81cc8d85e185d4d46320270746ee0d5f9bace82c4c54f16673b23f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 439, + 204, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 439, + 204, + 451 + ], + "spans": [ + { + "bbox": [ + 88, + 439, + 204, + 451 + ], + "type": "interline_equation", + "content": "s. t. \\quad \\forall t _ {i}, t _ {j} \\in \\mathcal {T} _ {c}, \\quad \\forall c \\in \\mathcal {C},", + "image_path": "d413dc128079cc2eb60fbfbc2fe91f83039cdab6ce9d9d8bef76564755b9a229.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": " represents the overall consistency objective to be maximized, " + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{self}}" + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{mutual}}" + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": " denote the self-consistency and mutual-consistency measures respectively, while " + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": " is a weight to balance them. Besides, considering the diversity of object categories in OVMOT, we add the object category consistency constraint for consistency learning. Specifically, we construct the intervals, i.e., " + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": ", containing several frames only with the same object category " + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": ", in which we select " + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "inline_equation", + "content": "t_i, t_j" + }, + { + "bbox": [ + 55, + 461, + 296, + 593 + ], + "type": "text", + "content": ". This is because we aim to learn the feature in a self-supervised manner without ID annotation, and the objects with various categories may bring about interference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " Category-consistency constraint. First, we consider the category consistency constraint in " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". For the interval splitting of " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in Eq. (3), we split the original videos into several segments of length " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " and randomly sample the shorter sub-segments with various lengths from each segment. These short-term sub-segments are then concatenated to form the training sequence. Such training sequences include short-long-term intervals. Specifically, we select the adjacent frames from the same sub-segment, which allows the association head to learn the consistency objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "under minor object differences. We also select the long-interval video frames from different sub-segments, which allow the association head to learn the similarity and variation of objects under large differences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 555, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 310 + ], + "type": "text", + "content": "After obtaining the sampled training sequences as discussed above, we utilize the localization head to extract object bounding boxes from each frame in the sequence. Since we only use unlabeled raw videos for training, we cannot directly obtain the object categories. To address this issue, we employ a clustering approach to group the bounding boxes based on their classification features. Specifically, we utilize the classification head to obtain the predicted category probability distributions for the candidate objects from all frames. Then, we cluster the candidate objects' category distributions into different clusters. After clustering the candidate objects' category distributions into distinct groups, we can treat each cluster as a separate category. As shown in Figure 2, we conduct self-supervised learning on the objects that belong to the same category cluster and are sourced from different frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "type": "text", + "content": " Appearance-self-consistency loss. We next model the consistency learning problem in Eq. (3) as a self-supervised learning task. After getting the training samples (object bounding boxes in different frames of the sample training sequence within the category clustering), we first use the CNN network to extract the appearance feature from the association head for all objects in frame " + }, + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "type": "text", + "content": " to construct the feature matrix " + }, + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 313, + 312, + 556, + 467 + ], + "type": "text", + "content": ". The main idea of the appearance-self-consistency loss is to leverage the self-consistency of the same objects' appearance features at different times (frames). We utilize the following two types of similarity transfer relationships, i.e., pair-wise symmetry and triple-wise cyclicity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "text", + "content": "Pair-wise consistency learning: For a pair of frames " + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "text", + "content": " in the video, we can get the object similarity matrix between them as " + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij} = \\mathbf{F}_i \\cdot (\\mathbf{F}_j)^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "text", + "content": ". We then compute the normalized similarity matrix " + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij} \\in [0,1]" + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "text", + "content": " based on the above similarity matrix " + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij}" + }, + { + "bbox": [ + 313, + 467, + 554, + 529 + ], + "type": "text", + "content": " by temperature-adaptive row softmax" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 536, + 555, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 536, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 330, + 536, + 555, + 564 + ], + "type": "interline_equation", + "content": "\\mathbf {S} (r, c) = f (\\mathbf {M} (r, c)) = \\frac {\\exp (\\tau \\mathbf {M} (r , c))}{\\sum_ {c = 1} ^ {C} \\exp (\\tau \\mathbf {M} (r , c))} \\tag {4}", + "image_path": "1749648aa49d881ba21ad2c41b53d450d7305bb9d06436dbd17b4fb255d78748.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "r, c" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " represent the indices of row and column in " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": ", respectively. Here " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " is the number of columns for " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " is the adaptive temperature adjustable parameter." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "The normalized similarity matrix " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " can be regarded as a mapping (object association relation) from frame " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij}: t_i \\to t_j" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "). In other words, we can select the maximum value of each row of " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " as the matched objects between frames " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ". Similarly, we get the reversed mapping from " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ji}: t_j \\to t_i" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ". We calculate the pair-wise similarity matrix as " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pair}} = \\mathbf{S}_{ij} \\cdot \\mathbf{S}_{ji}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pair}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " can be regarded as a symmetric mapping: " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i \\rightleftarrows t_j" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", i.e., from " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " and return to " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ". If the objects in frames " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " are the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7476" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": "same, the result " + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pair}}" + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": " should be an identity matrix, which can be used to construct the self-supervision loss. However, due to the object differences in different frames, this condition may not always be satisfied. Therefore, we need to supervise it deliberately, which will be discussed later." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": "Triplet consistency learning: Besides the pair-wise symmetric similarity, we further consider the triple-wise circularly consistent similarity. Specifically, given the similarity matrix " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij}" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": " between two frames " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{jk}" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": " between frames " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": ", we aim to build the consistent similarity relation among this triplet, i.e., frames " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_i, t_j" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": ". To do this, we first calculate the triplet similarity matrix as " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ik} = \\mathbf{M}_{ij} \\cdot \\mathbf{M}_{jk}, (i \\neq j \\neq k)" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ik}" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": " represents the similarity between the objects in frames " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": ", through the frame " + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 55, + 133, + 295, + 266 + ], + "type": "text", + "content": ". We then compute the normalized similarity matrix using Eq. (4) as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 275, + 295, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 275, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 94, + 275, + 295, + 290 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {i k} = f \\left(\\mathbf {M} _ {i k}\\right), \\quad \\mathbf {S} _ {k i} = f \\left(\\left(\\mathbf {M} _ {i k}\\right) ^ {\\mathrm {T}}\\right), \\tag {5}", + "image_path": "f3be678a9a27197a2553867f476e8bd460b514d1c5bf913c77d09c89f36724ef.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ik}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " represents the mapping " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "t_i\\rightarrow t_j\\rightarrow t_k" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ki}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " represents the mapping along " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "t_k\\to t_j\\to t_i" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": ". Then, we calculate the transitive-similarity matrix: " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{trip}} = \\mathbf{S}_{ik}\\cdot \\mathbf{S}_{ki}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{trip}}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " can be regarded as the mapping: " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "t_i\\rightleftarrows t_j\\rightleftarrows t_k" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " and return to " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": "For convenience, we note the matrices " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{pair}}" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{trip}}" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ", which should have the property that their diagonal elements are either 1 or 0, while all other elements are 0, in an ideal case. This means that the diagonal elements of " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " must be greater than or equal to the other elements. Based on this consideration, following [15, 50], we use the following loss " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{L}(\\mathbf{E}) = \\sum_{r}\\mathrm{relu}\\left(\\max_{c\\neq r}\\mathbf{E}(r,c) - \\mathbf{E}(r,r) + m\\right)" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "r,c" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " denote the indices of row and column in " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ". This loss denotes that we penalize the cases where the max non-diagonal element " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(r,c)" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "c\\neq r" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ") in a row " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ", exceeds the corresponding diagonal elements " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(r,r)" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " with a margin " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ". The margin " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "m\\geq 0" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " is a parameter used to control the punishment scope between " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(r,c)" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(r,r)" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ". This loss helps " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": " approach the identity matrix while addressing cases where there are unmatched targets in a row through a margin " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": ". Finally, we define our self-supervised self-consistency learning loss as " + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{self}} = \\mathrm{L}\\left(\\mathbf{E}_{\\mathrm{pair}}\\right) + \\mathrm{L}\\left(\\mathbf{E}_{\\mathrm{trip}}\\right)" + }, + { + "bbox": [ + 55, + 361, + 295, + 566 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " Mutual-consistency loss. The mutual-consistency loss makes use of the consistency between the spatial position continuity and the appearance similarity of the objects at different times. Given a bounding box list " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "B_{j}" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " of adjacent frames " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " in the video, we compute the Intersection over Union (IoU) matrix " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij}" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " between bounding boxes " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "B_{j}" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": ", which quantifies the overlap between them. Based on a specified threshold IoUthres, we create an assignment matrix " + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{ij}" + }, + { + "bbox": [ + 55, + 566, + 295, + 673 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 684, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 684, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 110, + 684, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {M} _ {i j} > \\operatorname {I o U} _ {\\text {t h r e s}}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {6}", + "image_path": "4f5586dd00b8cc1e1ad15295cab23c5587c9080cab52a446fb17d82b9ee1cd38.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "This assignment matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{ij}" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " indicates whether pairs of objects are considered similar, facilitating the optimization of the spatial consistency objective. Furthermore, we utilize the previously obtained normalized similarity matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij}" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " to establish the mutual-consistency loss, using the binary cross-entropy loss function " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{L_{BCE}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mutual}} = \\mathrm{L_{BCE}}(\\mathbf{S}_{ij},\\mathbf{A}_{ij})" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ". This formulation allows us to measure the discrepancy between the normalized similarity matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{ij}" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " and the assignment matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{ij}" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ", thereby enforcing spatial consistency across the detected objects in frames " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "text", + "content": "Finally, we have transformed the optimization problem in Eq. (3) into a self-supervised loss as " + }, + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{self}} + \\alpha \\cdot \\mathcal{L}_{\\mathrm{mutual}}" + }, + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 193, + 553, + 276 + ], + "type": "text", + "content": " is a weighting coefficient. This formulation incorporates both the self- and mutual-consistencies, as well as category consistency, among the frames with different intervals, thereby effectively exploring the potentiality of videos to enhance the association for OVMOT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 287, + 447, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 447, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 447, + 300 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": "In Section 3.1, we use ResNet50 with FPN for localizing candidate regions. We set " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " in Eq. (1) as 0.007. In Section 3.2, we select four pairs of typical tracking state aware prompts, i.e., 'complete and incomplete,' 'unoccluded and occluded,' 'unobsured and obscured,' and 'recognizable and unrecognizable.' We set " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{low}}" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " as 0.3, " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{high}}" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " as 0.6. In Section 3.3, the segment length " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " is set as 24. We use the K-means clustering algorithm. We set the margin " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " as 0.5 and the IoUthres as 0.9. For " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " frames in each sampled video sequence, we select " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "C_N^2" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "C_N^3" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " groups of frames to calculate the self-consistency loss in Section 3.3. We also select " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "C_N^2" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " groups of frames to calculate the mutual-consistency loss. The weighting coefficient " + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 305, + 553, + 461 + ], + "type": "text", + "content": " is 0.9." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "content": "In the training stage, we first train the two-stage detector on the base classes of the LVIS dataset [19] referenced from [13], for 20 epochs, and use prompt-guided attention proposed in Section 3.2 to fine-tune the model's classification head for 6 epochs. Then we train the association head using static image pairs generated from [27] for 6 epochs and self-supervise the association head with TAO training dataset [7] without annotation for 14 epochs. In the inference stage, we select object candidates " + }, + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 463, + 553, + 713 + ], + "type": "text", + "content": " by NMS with IoU threshold 0.5 for detection. Additionally, for tracking, we use appearance feature similarity to associate history tracks with the objects in the current frame. As in [27], we evaluate the similarity between history tracks and detected objects using both bi-directional softmax [16] and cosine similarity metrics. Following the classical MOT approaches, if the similarity score exceeds a matching threshold, we assign the object to the track. If the object doesn't correspond to any existing track, a new track is initiated if its detection confidence score surpasses a threshold, otherwise, it is disregarded. We set the similarity score threshold as 0.35 and maintain a track memory of 30 frames." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7477" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 127, + 83, + 483, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 555, + 83 + ], + "type": "text", + "content": "Table 1. Result comparison. We evaluate our method against closed-set and open-vocabulary trackers on TAO validation and test sets. Here '√' denotes using the corresponding dataset with annotations, and † represents only using the raw video without any annotation (%)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 127, + 83, + 483, + 224 + ], + "lines": [ + { + "bbox": [ + 127, + 83, + 483, + 224 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 483, + 224 + ], + "type": "table", + "html": "
MethodClassesDataBaseNovel
Validation setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓27.145.624.711.022.542.724.40.4
TETer [26]✓✓✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [51]✓-✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓-✓✓28.347.420.517.022.746.719.32.2
OVTrack [27]✓-✓-35.549.336.920.227.848.833.61.5
SLAck [28]✓-✓✓37.255.037.619.131.154.337.81.3
VOVTrack (Ours)✓-✓†39.6 (↑ 2.4)58.940.919.135.3 (↑ 4.2)58.540.96.5
Test setBaseNovelLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓✓✓25.843.223.510.620.239.720.90.2
TETer [26]✓✓✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [51]✓-✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓-✓✓26.044.119.014.818.039.013.41.7
OVTrack [27]✓-✓-32.645.635.416.924.141.828.71.8
SLAck [28]✓-✓✓34.752.535.616.127.149.130.02.0
VOVTrack (Ours)✓-✓†38.1 (↑ 3.4)57.341.515.429.8 (↑ 2.7)51.732.05.7
", + "image_path": "78d021d205f770104ef66636795cc47a5a399a7691579022e7836fe5c208c337.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 160, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 160, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 160, + 256 + ], + "type": "text", + "content": "4. EXPERIMENTS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 173, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 173, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 173, + 275 + ], + "type": "text", + "content": "4.1. Dataset and Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 413 + ], + "type": "text", + "content": "We follow [27] to select the dataset and metrics for evaluation. We leverage the comprehensive and extensive vocabulary MOT dataset TAO [7] as our benchmark for OV-MOT. TAO is structured similarly to the LVIS [19] taxonomy, categorizing object classes based on their frequency of appearance into frequent, common, and rare groups. We use the rare classes defined in LVIS as " + }, + { + "bbox": [ + 55, + 281, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 55, + 281, + 296, + 413 + ], + "type": "text", + "content": " and others as " + }, + { + "bbox": [ + 55, + 281, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 55, + 281, + 296, + 413 + ], + "type": "text", + "content": ". We evaluate the performance with the comprehensive metric tracking-every-thing accuracy (TETA), which consists of the accuracies of localization (LocA), classification (ClsA), and association (AssocA)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 421, + 176, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 176, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 176, + 433 + ], + "type": "text", + "content": "4.2. Comparative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "content": "We compare our method with the latest trackers, TETer [26] and QDTrack [16], which are trained on both " + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "content": ". We include the classical trackers like DeepSORT [51] and Tractor++ [3] trained only on " + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 55, + 437, + 295, + 533 + ], + "type": "text", + "content": " and enhanced by OVD method ViLD [18] to achieve open-vocabulary tracking, as baselines. We also compare our method with the state-of-the-art OVMOT methods, i.e., OVTrack [27] and SLAck [28]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "As shown in Table 1, we present the OVMOT evaluation results on the TAO validation and test sets on both base and novel classes. We can see that our method outperforms all closed-set and open-vocabulary methods on both the validation and test sets. Even though QDTrack [16] and TETer [26] have seen novel classes during training on TAO, our method significantly outperforms them in all metrics on both base and novel classes. Additionally, DeepSORT and Tracktor++ with the open-vocabulary detector ViLD [18] are also trained in a supervised manner on TAO, while our method, trained in a self-supervised manner on TAO, surpasses them by a large margin. Although the results of SLAck, the most competitive method, are better than other comparison methods, our method outperforms it in all metrics significantly on both base and novel classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": "Particularly, compared to our baseline method OVTrack, our method achieves improvements of " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": " in base and novel TETA, respectively, and a " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": " increase in novel ClsA and AssocA. In the test set, base and novel TETA also show improvements of " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": ", respectively. It is worth noting that the state-of-the-art method SLAck is trained on the TAO training dataset using the annotations, while our method only uses the raw video data of it without the annotations. Even so, our method outperforms it with a significant margin on the comprehensive TETA score, as shown in Table 1. Especially the improvement of novel category on the validation set has reached 5x the original level. This demonstrates the effectiveness of the proposed method, which is very promising for the OVMOT task. We provide more analysis from the standpoint of data quantity for training in Appendix 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": "Balance between classification and detection. Our method outperforms other methods on most metrics, except for the ClsA of the base category. We clarify that the lower ClsA score compared to OVTrack stems from our method's superior detection capability. Specifically, we classify many more detected objects, while OVTrack focuses on a smaller subset of high-confidence detections. This is supported by our comparison with SLAck, which also shows reduced classification performance on the base category despite better detection results (LocA), similar to our findings. In contrast, our method significantly improves performance on novel categories by identifying more unseen detections during testing. Overall, our method achieves a better balance between detection and classification tasks, as well as between base and novel categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 677 + ], + "type": "text", + "content": "We conduct the ablation studies on the main components proposed in our method, including prompt-guided attention, and the self-supervised learning related modules as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 678, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 678, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 553, + 700 + ], + "type": "text", + "content": "- w/o prompt-guided attention " + }, + { + "bbox": [ + 313, + 678, + 553, + 700 + ], + "type": "inline_equation", + "content": "(w_{r})" + }, + { + "bbox": [ + 313, + 678, + 553, + 700 + ], + "type": "text", + "content": ": Removing the prompt-guided attention in Section 3.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "text", + "content": "- w/o piecewise weight strategy (" + }, + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{low}}" + }, + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{high}}" + }, + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "text", + "content": "): Removing" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7478" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 82, + 504, + 165 + ], + "blocks": [ + { + "bbox": [ + 121, + 71, + 488, + 81 + ], + "lines": [ + { + "bbox": [ + 121, + 71, + 488, + 81 + ], + "spans": [ + { + "bbox": [ + 121, + 71, + 488, + 81 + ], + "type": "text", + "content": "Table 2. Ablation studies on modules of prompt-guided attention and self-supervised association " + }, + { + "bbox": [ + 121, + 71, + 488, + 81 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 104, + 82, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 104, + 82, + 504, + 165 + ], + "spans": [ + { + "bbox": [ + 104, + 82, + 504, + 165 + ], + "type": "table", + "html": "
ModuleAblation MethodsBaseNovel
TETALocAAssocAClsATETALocAAssocAClsA
Prompt-guided attentionw/o prompt-guided attention36.153.437.517.430.553.236.81.6
w/o piecewise weight strategy37.355.838.517.633.856.839.84.8
Self-consistency associationw/o whole self-consistency learning37.457.736.917.732.557.635.64.2
w/o short-long-sampling37.957.637.718.434.058.438.25.5
w/o category consistency38.157.839.417.133.358.437.93.6
w/o self-consistency38.257.538.818.334.258.339.15.3
w/o mutual-consistency39.158.640.118.634.158.838.15.6
CompleteVOVTrack (Ours)39.658.940.919.135.358.540.96.5
", + "image_path": "1f47e5e8ebb90c95f53fcea5036916fbcf6e1ed698246927d2bf7ab52f5658fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 204 + ], + "type": "text", + "content": "the piecewise weighting strategy proposed in Section 3.2, by directly using the " + }, + { + "bbox": [ + 55, + 180, + 294, + 204 + ], + "type": "inline_equation", + "content": "w_{r}" + }, + { + "bbox": [ + 55, + 180, + 294, + 204 + ], + "type": "text", + "content": " calculated by Eq. (2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 204, + 295, + 299 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "content": "- w/o whole self-supervised learning: Removing the whole self-supervised consistency learning strategy in Section 3.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 228, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 252 + ], + "type": "text", + "content": "- w/o short-long-sampling: Removing the short-long-interval sampling strategy in Section 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 252, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 294, + 275 + ], + "type": "text", + "content": "- w/o category consistency: Removing the category-aware object clustering in Section 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 275, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 294, + 299 + ], + "type": "text", + "content": "- w/o self/mutual-consistency: Removing the self/mutual-consistency loss in Section 3.3." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 300, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 384 + ], + "type": "text", + "content": "Effectiveness of the state-aware prompt-guided attention. As shown in the first unit of Table 2, we can see that using prompt-guided attention as a weight coefficient during the training stage can effectively improve all metrics for both base and novel classes. The piecewise weighting strategy is also very effective, especially in improving the localization and classification accuracies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 385, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 589 + ], + "type": "text", + "content": "Effectiveness of the self-supervised consistency learning. As shown in the second unit of Table 2, we can see that using self-supervised loss can effectively improve all metrics for both base and novel classes. Either the self-consistency or the mutual-consistency for appearance learning is effective for the association task, i.e., 'AssocA.' Also, the interval sampling strategy allows samples to have a more diverse range of long and short cycles, improving the association-related metric. The category clustering strategy tries to gather the objects with the same category in a cluster, which is also helpful. To our surprise, the above strategies, in most cases, also effectively help improve classification ('ClsA') and localization ('LocA') accuracies. This is because the better association results can indirectly help the other two sub-tasks in OVMOT. We provide the discussion and analysis of the complementarity among different tasks in Appendix 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 173, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 173, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 173, + 612 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 295, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 630 + ], + "type": "text", + "content": "Illustrations of the proposed prompt-guided attention." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Figure 3 (a) shows cases of high prompt-guided attention, where we can see that the regions often have very distinctive category features with no occlusion, and the image quality of the region is very high. In contrast, Figure 3 (b) presents cases of low prompt-guided attention, highlighting regions that often exhibit issues such as significant blurriness, occlusion, unclear visibility, and difficulty in identification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "type": "text", + "content": "These factors are detrimental to training object localization and classification features, which are effectively weakened through state-aware prompt-guided attention." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 216, + 553, + 275 + ], + "blocks": [ + { + "bbox": [ + 315, + 216, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 315, + 216, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 216, + 553, + 275 + ], + "type": "image", + "image_path": "4c7a160ed6bafafceab2481a02fe378d01999a22bf32f0bb0c0ad7eb303ddc05.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 276, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 299 + ], + "type": "text", + "content": "Figure 3. Illustration of regions with high (a) or low (b) prompt-guided attention, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 407 + ], + "type": "text", + "content": "Comparison result visualization. We show several visualization results in Figure 4. We can see that the proposed method provides better results than OVTrack [27]. In the first case of Figure 4 (a), our method provides an accurate object localization result and identifies the correct category. In the second case of Figure 4 (b), the tracking of a drone provided by the comparison method is wrong (different box colors denote different tracking IDs), and the classification is also not correct. Our method can track it continuously." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 408, + 553, + 456 + ], + "blocks": [ + { + "bbox": [ + 316, + 408, + 553, + 456 + ], + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 456 + ], + "type": "image", + "image_path": "7ce955364ce5531f6610dcb312b816fbfcdd2f99751cde9c3b8cd132ca13a71d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 456, + 553, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 477 + ], + "type": "text", + "content": "Figure 4. Compared OVMOT results of ours and OVTrack on some cases with novel classes." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 529 + ], + "type": "text", + "content": "In Appendix 4, we provide more analysis that includes additional cases on prompt-guided attention, failure cases, time complexity analysis, parameter settings analysis, and method generalization analysis." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 538, + 410, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 410, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 410, + 550 + ], + "type": "text", + "content": "5. CONCLUSION" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this work, we have developed a new method, namely VOVTrack, to handle the OVMOT problem from the perspective of video object tracking. For this purpose, we first consider the object state during tracking and propose tracking-state-aware prompt-guided attention, which improves the accuracy of object localization and classification (detection). Second, we develop an object similarity learning strategy for the temporal association (tracking) using only the raw video data without annotation, which unveils the power of self-supervised learning for open-vocabulary tracking tasks. Experimental results demonstrate the effectiveness of the proposed method and each component for open-vocabulary tracking." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7479" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 208, + 115, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 115, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 115, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 228, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 228, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 272 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Stefan Roth, and Bernt Schiele. People-tracking-by-detection and people-detection-by-tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 316 + ], + "type": "text", + "content": "[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In European Conference on Computer Vision, pages 384-400, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "text", + "content": "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In IEEE/CVF International Conference on Computer Vision, pages 941-951, 2019. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 294, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 405 + ], + "type": "text", + "content": "[4] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In IEEE International Conference on Advanced Video and Signal based Surveillance, pages 1-6, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 405, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 296, + 459 + ], + "type": "text", + "content": "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. MeMOT: Multi-object tracking with memory. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8090-8100, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 296, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 493 + ], + "type": "text", + "content": "[6] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In IEEE/CVF International Conference on Computer Vision Workshops, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 493, + 296, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 296, + 537 + ], + "type": "text", + "content": "[7] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. TAO: A large-scale benchmark for tracking any object. In European Conference on Computer Vision, pages 436-454, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 537, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 592 + ], + "type": "text", + "content": "[8] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. MOT20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 296, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 636 + ], + "type": "text", + "content": "[9] Akshay Dhamija, Manuel Gunther, Jonathan Ventura, and Terrance Boult. The overlooked elephant of object detection: Open set. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1021-1030, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 637, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 296, + 689 + ], + "type": "text", + "content": "[10] Shuxiao Ding, Eike Rehder, Lukas Schneider, Marius Cordts, and Juergen Gall. 3dmotformer: Graph transformer for online 3d multi-object tracking. In IEEE/CVF International Conference on Computer Vision, pages 9784-9794, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[11] Thang Doan, Xin Li, Sima Behpour, Wenbin He, Liang Gou, and Liu Ren. Hyp-OW: Exploiting hierarchical structure" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 198 + ], + "type": "text", + "content": "This work was supported in part by the Young Scientists Fund of the National Natural Science Foundation of China (NSFC) under Grant 62402490, the Emerging Direction Cultivation Project of Interdisciplinary Center (Intelligent Protection and Utilization of Digital Cultural Heritage) from Tianjin University, the Excellent Young Scientists Fund of NSFC under Grant 62422118, and the Guangdong Basic and Applied Basic Research Foundation under Grant 2025A1515010101." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "learning with hyperbolic distance enhances open world object detection. In AAAI Conference on Artificial Intelligence, pages 1555-1563, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to ECCV-TAO-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "text", + "content": "[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 251 + ], + "type": "text", + "content": "[14] Yunhao Du, Zhicheng Zhao, Yang Song, Yanyun Zhao, Fei Su, Tao Gong, and Hongying Meng. StrongSORT: Make DeepSORT great again. IEEE Transactions on Multimedia, 25:8725-8737, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "type": "text", + "content": "[15] Wei Feng, Feifan Wang, Ruize Han, Yiyang Gan, Zekun Qian, Junhui Hou, and Song Wang. Unveiling the power of self-supervision for multi-view multi-human association and tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "text", + "content": "[16] Tobias Fischer, Thomas E Huang, Jiangmiao Pang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. QD-Track: Quasi-dense similarity learning for appearance-only multiple object tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "text", + "content": "[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? The KITTI vision benchmark suite. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 454 + ], + "type": "text", + "content": "[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 499 + ], + "type": "text", + "content": "[19] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356-5364, 2019. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "type": "text", + "content": "[20] Ruize Han, Wei Feng, Yujun Zhang, Jiewen Zhao, and Song Wang. Multiple human association and tracking from egocentric and complementary top views. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5225-5242, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "text", + "content": "[21] Ruize Han, Wei Feng, Feifan Wang, Zekun Qian, Haomin Yan, and Song Wang. Benchmarking the complementary-view multi-human association and tracking. International Journal of Computer Vision, 132:118-136, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 645 + ], + "type": "text", + "content": "[22] Kuan-Chih Huang, Ming-Hsuan Yang, and Yi-Hsuan Tsai. Delving into motion-aware matching for monocular 3d object tracking. In IEEE/CVF International Conference on Computer Vision, pages 6909-6918, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[23] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5830-5840, 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[24] Jan Krejci, Oliver Kost, Ondrej Straka, and Jindrich Duník. Pedestrian tracking with monocular camera us-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7480" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "ing unconstrained 3D motion model. arXiv preprint arXiv:2403.11978, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 296, + 152 + ], + "type": "text", + "content": "[25] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese cnn for robust target association. In IEEE/CVF Conference on Computer Vision and Pattern Recognition workshops, pages 33-40, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 296, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 296, + 198 + ], + "type": "text", + "content": "[26] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In European Conference on Computer Vision, pages 498-515, 2022. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 200, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 296, + 255 + ], + "type": "text", + "content": "[27] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5567-5577, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 257, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 296, + 311 + ], + "type": "text", + "content": "[28] Siyuan Li, Lei Ke, Yung-Hsu Yang, Luigi Piccinelli, Mattia Segu, Martin Danelljan, and Luc Van Gool. Slack: Semantic, location, and appearance aware open-vocabulary tracking. In European Conference on Computer Vision, pages 1-18, 2024. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 314, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 296, + 369 + ], + "type": "text", + "content": "[29] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19045-19055, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 404 + ], + "type": "text", + "content": "[30] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. IEEE Robotics and Automation Letters, 5(2):1803-1810, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 406, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 296, + 460 + ], + "type": "text", + "content": "[31] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8844-8854, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 463, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 296, + 507 + ], + "type": "text", + "content": "[32] Anton Milan, S Hamid Rezatofighi, Anthony Dick, Ian Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI conference on Artificial Intelligence, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 296, + 553 + ], + "type": "text", + "content": "[33] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In European Conference on Computer Vision, pages 566-579, 2012. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "type": "text", + "content": "[34] Aljosa Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In IEEE International Conference on Robotics and Automation, pages 3180-3187, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 296, + 666 + ], + "type": "text", + "content": "[35] Aljosa Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3494–3501, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[36] Aljosa Osep, Paul Voiglaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In IEEE International Conference on Robotics and Automation, pages 10031-10037, 2020. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "text", + "content": "[37] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 164-173, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[38] Zheng Qin, Sanping Zhou, Le Wang, Jinghai Duan, Gang Hua, and Wei Tang. Motiontrack: Learning robust short-term and long-term motions for multi-object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 251 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 296 + ], + "type": "text", + "content": "[40] Shafin Rahman, Salman Khan, and Nick Barnes. Improved visual-semantic alignment for zero-shot object detection. In AAAI conference on artificial intelligence, pages 11932–11939, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "text", + "content": "[41] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems, 28, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 387 + ], + "type": "text", + "content": "[42] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In IEEE/CVF International Conference on Computer Vision, pages 300-311, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 442 + ], + "type": "text", + "content": "[43] Fatemeh Saleh, Sadegh Aliakbarian, Hamid Rezatofighi, Mathieu Salzmann, and Stephen Gould. Probabilistic tracklet scoring and inpainting for multiple object tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14329-14339, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "type": "text", + "content": "[44] Madeline C Schiappa, Yogesh S Rawat, and Mubarak Shah. Self-supervised learning for videos: A survey. ACM Computing Surveys, 55(13s):1-37, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "text", + "content": "[45] Sarthak Sharma, Junaid Ahmed Ansari, J Krishna Murthy, and K Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In IEEE International Conference on Robotics and Automation, pages 3508-3515, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 578 + ], + "type": "text", + "content": "[46] Vivek Hari Sridhar, Dominique G Roche, and Simon Gingins. Tracktor: image-based automated tracking of animal movement and behaviour. Methods in Ecology and Evolution, 10(6):815-820, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "type": "text", + "content": "[47] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "text", + "content": "[48] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[49] Li Wang, Xinyu Zhang, Wenyuan Qin, Xiaoyu Li, Jinghan Gao, Lei Yang, Zhiwei Li, Jun Li, Lei Zhu, Hong Wang, and Huaping Liu. CAMO-MOT: Combined Appearance-Motion" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7481" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 644 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Optimization for 3D Multi-Object Tracking With Camera-LiDAR Fusion. IEEE Transactions on Intelligent Transportation Systems, 24(11):11981-11996, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[50] Zhongdao Wang, Jingwei Zhang, Liang Zheng, Yixuan Liu, Yifan Sun, Yali Li, and Shengjin Wang. Cycles: Self-supervised cycle association for learning re-identifiable descriptions. In European Conference on Computer Vision, pages 72–88, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[51] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In IEEE International Conference on Image Processing, pages 3645-3649, 2017. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[52] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. CORA: Adapting CLIP for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7031-7040, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 297 + ], + "type": "text", + "content": "[53] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In European Conference on Computer Vision, pages 466-481, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[54] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "text", + "content": "[55] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xi-angyu Zhang, and Yichen Wei. MOTR: End-to-end multiple-object tracking with transformer. In European Conference on Computer Vision, pages 659–675, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[56] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. FairMOT: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "text", + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "text", + "content": "[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "text", + "content": "[59] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In European Conference on Computer Vision, pages 474-490, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "text", + "content": "[60] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 644 + ], + "type": "text", + "content": "[61] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8771-8780, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7482" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_content_list.json b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f84a21f01efc7f788e349efd1881e25bc404db5d --- /dev/null +++ b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_content_list.json @@ -0,0 +1,1318 @@ +[ + { + "type": "text", + "text": "VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions", + "text_level": 1, + "bbox": [ + 200, + 130, + 797, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yash Garg, Saketh Bachu, Arindam Dutta, Rohit Lal†, Sarosij Bose, Calvin-Khang Ta‡, M. Salman Asif, Amit Roy-Chowdhury University of California, Riverside", + "bbox": [ + 230, + 203, + 766, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ygarg002,sbach008,adultt020,r1al011,sbose007,cta003,sasif,amitrc}@ucr.edu", + "bbox": [ + 174, + 258, + 816, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d1f4f7d05d18f1178989a04559ef760ba9159d82ed48257677342924a0978d2.jpg", + "image_caption": [ + "Figure 1. We propose VOccl3D, a large-scale synthetic video dataset specifically designed for training and evaluating algorithms for 3D human pose and shape estimation (HPS) in realistic occlusion scenarios. VOccl3D comprises over 250,000 frames, with a total runtime exceeding 2 hours and 30 minutes. Compared to previous occlusion-based datasets, VOccl3D has diverse body shapes, textures, and most importantly, significant and realistic occlusions within the scenes. In addition to body shape and pose, VOccl3D provides other necessary ground truth information, such as bounding boxes, body part segmentations, and human silhouettes. The top row illustrates various frames from our dataset, showcasing diverse occlusions, clothing textures, and motions, while the bottom row represents a sequence of frames from a video sequence within our dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 286, + 901, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 590, + 326, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human pose and shape (HPS) estimation methods have been extensively studied, with many demonstrating high zero-shot performance on in-the-wild images and videos. However, these methods often struggle in challenging scenarios involving complex human poses or significant occlusions. Although some studies address 3D human pose estimation under occlusion, they typically evaluate performance on datasets that lack realistic or substantial occlusions, e.g., most existing datasets introduce occlusions with random patches over the human or clipart-style overlays, which may not reflect real-world challenges. To bridge this gap in realistic occlusion datasets, we introduce a novel benchmark dataset, VOccl3D, a Video-based human Occlusion dataset with 3D body pose and shape annotations. Inspired by works such as AGORA and BEDLAM, we constructed this dataset using advanced computer graphics rendering techniques, incorporating diverse real-world occlusion scenarios, clothing textures, and human motions.", + "bbox": [ + 89, + 628, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Additionally, we fine-tuned recent HPS methods, CLIFF and BEDLAM-CLIFF, on our dataset, demonstrating significant qualitative and quantitative improvements across multiple public datasets, as well as on the test split of our dataset, while comparing its performance with other state-of-the-art methods. Furthermore, we leveraged our dataset to enhance human detection performance under occlusion by fine-tuning an existing object detector, YOLO11, thus leading to a robust end-to-end HPS estimation system under occlusions. Overall, this dataset serves as a valuable resource for future research aimed at benchmarking methods designed to handle occlusions, offering a more realistic alternative to existing occlusion datasets. See the Project page for code and dataset:https://yashgarg98.github.io/Voccl3D-dataset/", + "bbox": [ + 511, + 592, + 908, + 820 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Currently at NASA MSFC IMPACT. ‡ Currently at Dolby Laboratories. Work done while the authors were at UCR.", + "bbox": [ + 511, + 875, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7350", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular 3D human pose and shape (HPS) estimation is a complex yet essential task in computer vision. It has applications in surveillance [7, 17], autonomous robotics [2, 42, 52], human motion analysis [9, 10], and clinical assessment [8, 23, 33]. Since the introduction of the neural network-based Human Mesh Recovery Network (HMR) [18], numerous approaches have advanced the field by enabling accurate estimation of SMPL[28] pose and shape from a single RGB image. Recent methods [1, 21, 24, 25, 29, 44] demonstrate impressive zero-shot HPS performance on in-the-wild RGB images and video sequences. However, these methods still face limitations in achieving high performance in challenging scenarios, such as complex human poses and under significant occlusions.", + "bbox": [ + 89, + 114, + 483, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Achieving robust HPS under occlusion remains a challenging problem due to the contextual ambiguity of the occluded body parts. Several methods [22, 40, 51, 55] have been proposed to address the challenges caused by occlusions. Methods like PARE [22] focus on using visible body parts to infer occluded areas, enhancing estimation accuracy with partial views. Temporal models like HuMoR [40] and GLAMR [51] use generative frameworks to ensure pose continuity over time. HuMoR predicts pose distributions, while GLAMR incorporates global trajectory data to fill in missing poses. Recent methods, like STRIDE [24], leverage a large-scale pre-trained motion prior model to achieve temporally coherent pose reconstruction under occlusions. While these methods effectively leverage visible information and temporal continuity, they still struggle under severe occlusions due to limited exposure to occluded scenarios in their training data.", + "bbox": [ + 89, + 325, + 483, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluations of the existing methods are often limited to occluded datasets that lack scene diversity [34], involve moderate or lower levels of occlusion [14, 45], or use patch-based occlusion datasets for training and inference [55], as illustrated in Figure 2. Most importantly, there remains an urgent need for a realistic, diverse, and significantly occluded human pose and shape dataset to advance the task of 3D HPS estimation. To address this gap, our work proposes VOccl3D, a large-scale video dataset with synthetic humans in real scenes for 3D human pose and shape estimation. Following prior works such as BEDLAM [3] and Synthocap [12], we use an advanced computer graphics engine to render a highly realistic synthetic dataset. Prior research has shown that fine-tuning models on synthetic datasets can significantly improve HPS estimation performance on real-world datasets [3]. Synthetic datasets provide \"perfect\" ground-truth annotations and eliminate the need for costly sensors, making it computationally efficient to generate large-scale datasets.", + "bbox": [ + 89, + 584, + 483, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed dataset includes approximately 250,000 images and 400 video sequences. To introduce diversity, we", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cbe23ef8950056961997cd564aa47c6151b13a5f81dd23cfc7b4487a4336beb7.jpg", + "image_caption": [ + "a) 3DPW-AdvOcc@80" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 648, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e289547aa747317d3b88d78846c6025832db910e10f785e4fdbb13c9c414f968.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 88, + 777, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97c2cc452f7e7a57730b9269869dd4327882f8e9dc53ba8e04b414ec5b686fbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 88, + 893, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fefe371bdf90e683d3c6e293f6b92d15f692b343f8d9f9f4d6eac1dfde7b981c.jpg", + "image_caption": [ + "d) OCMotion" + ], + "image_footnote": [], + "bbox": [ + 516, + 179, + 651, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4490af5634a6def6a3f68b0a37e61ada84deb7ab97a2eb28ae998f61e078c83b.jpg", + "image_caption": [ + "b) Occluded Human3.6M", + "e) 3DPW" + ], + "image_footnote": [], + "bbox": [ + 674, + 178, + 751, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4c1af9124c611fdfc1ae15ef8e39dd1a4407707f01324b34a104e7d87ce02ee1.jpg", + "image_caption": [ + "c) Syn. Occ. Human3.6M", + "f) Ours", + "Figure 2. Samples from various occlusion-based HPS datasets. The top row displays datasets with artificial, patch-based occlusions applied to existing datasets, including 3DPW-AdvOcc [45, 55], Occluded Human3.6M [15, 24], and Synthetic Occlusion Human3.6M [15, 54]. The bottom row presents samples from two datasets with natural occlusions: OCMotion [14] and 3DPW [45]. Notably, the top-row images exhibit unrealistic occlusions, which lack the realism of naturally occurring occlusions. The bottom row images contain natural occlusions but are sparse and infrequent. Compared to existing datasets, our proposed VOcccl3D dataset features more realistic occlusions." + ], + "image_footnote": [], + "bbox": [ + 769, + 178, + 901, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "incorporate human motion samples from the AMASS [30] dataset, over 200 distinct clothing textures for sequences containing both male and female genders, and 40 real background scenes with occlusions. Additionally, we provide occlusion labels for each body joint. Unlike prior methods [3, 35] which render synthetic environments, we utilize 3D Gaussian splatting (3DGS) [19] to create 3D representations of background scenes. This approach enhances the realism of rendered scenes, making them closely resemble real-world captured data. Reconstructing background scenes using 3DGS offers the flexibility to create domain-specific datasets, such as for agriculture, street or indoor environments using raw RGB video captures without relying on costly or limited graphic assets. Figure 1 showcases samples from our dataset, illustrating the diversity in clothing, environments, human motions, and occlusions.", + "bbox": [ + 511, + 460, + 906, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We fine-tuned the CLIFF [25] and BEDLAM-CLIFF [3] models on our dataset to demonstrate that training with synthetic data enhances the performance of existing HPS estimation methods under occlusion. We evaluated state-of-the-art HPS estimation methods on the test split of our dataset across three occlusion levels: high, medium, and low. The results show that existing methods perform poorly under high occlusion, whereas our fine-tuned models achieve significant improvements. To assess performance on real-world data, we perform evaluations on the 3DPW and OCMotion datasets, which contain low-level occlusions. Additionally, we created a variant of the 3DPW dataset by adding black patches to evaluate the robustness", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7351", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "against high-occlusion real scenarios. Our fine-tuned models outperformed recent HPS methods on these real-world datasets. Furthermore, our experiments highlight the importance of an effective human object detector for improving HPS performance under occlusion. To address this, we fine-tuned the YOLO11[16] object detector using our dataset, enhancing its performance under occluded scenarios. VOcc13D is available for researchers to benchmark and evaluate occlusion-aware methods.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we make the following key contributions:", + "bbox": [ + 109, + 227, + 470, + 242 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose VOcc13D, a novel large-scale, realistic video-based dataset of occluded synthetic humans in real scenes for 3D human pose and shape estimation.", + "2. We demonstrate both qualitatively and quantitatively that existing HPS estimation methods fine-tuned with VOccl3D outperform other methods on real-world datasets with occlusions.", + "3. We improved the performance of the human object detector under occlusion scenarios by leveraging the VOccl3D dataset which is a crucial component for a robust HPS estimation.", + "4. VOccl3D can serve as a benchmark dataset for evaluating methods specifically designed to perform under occlusion across various tasks, including human and body-part segmentation, 2D/3D pose estimation, and human bounding box detection." + ], + "bbox": [ + 89, + 244, + 482, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 497, + 238, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic Data Generation for Human Pose Estimation.", + "text_level": 1, + "bbox": [ + 89, + 523, + 482, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human pose estimation is crucial in computer vision, but state-of-the-art methods depend on costly, labor-intensive labeled datasets. Several notable synthetic datasets have advanced research in human pose estimation. SynBody [49] provides a large-scale synthetic dataset for human mesh recovery and view synthesis, while RePoGen [38] enables fine-grained control over pose and viewpoint to generate rare, complex poses. Human3.6m [15] provides additional small mixed reality dataset by inserting 3D animation models with background scenes. BEDLAM [3] further highlights that models trained solely on synthetic data can outperform real-data-trained counterparts, emphasizing the importance of high-quality synthetic datasets for transferable models. Further, PressurePose [6] simulates interactions between articulated and soft bodies, capturing fine-grained contact dynamics. SynthMocap [12] extends this by providing expressive synthetic data with detailed body, hand, and face movements. Despite their success, these synthetic datasets lack emphasis on significant occlusions, a key real-world challenge. Addressing this limitation, our work proposes a synthetic dataset tailored to 3D pose estimation under heavy and realistic occlusion, aiming to bridge this gap and enhance robustness in occlusion-prone environments.", + "bbox": [ + 89, + 537, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image and Video based Human Pose Estimation. Esti", + "bbox": [ + 89, + 886, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mating 2D/3D pose and shape from single RGB image has widely been explored in previous research [3, 25, 43, 47, 50]. Methods such as [43] use a conditional variational autoencoder for 2D-to-3D lifting in pose estimation. The pose estimation approach proposed in [47] applies normalizing flows [41] for 2D-3D mapping. Large-scale datasets like BEDLAM [3] have improved pose estimation models like CLIFF [25] and HMRNet [18]. However, these models struggle with generalizing to unseen scenarios with severe occlusions due to limited training on such cases.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, video-based human pose estimation has significantly improved challenging datasets. Early works like [56] use the EM algorithm to estimate the 3D pose from monocular video through 2D joint uncertainty maps. The method in [37] employs dilated temporal convolutions and semi-supervised learning for 3D pose estimation, while [1] uses the SMPL model [28] to extract pose and shape parameters, refining models like HMR with bundle adjustment. VIBE [21] applies adversarial learning with AMASS [30] for 3D pose extraction, while MEVA [29] improves accuracy and smoothness using a variational autoencoder to address VIBE's high acceleration error. Owing to these drawbacks, HuMoR employs a conditional variational autoencoder for robust pose estimation, while the state-of-the-art method WHAM [44] integrates 2D-3D lifting and SLAM for accurate global motion estimation. While these methods perform well on various datasets, studies like [24] highlight performance drops under significant occlusions, as most datasets contain only sparse occlusions, making models struggle with unseen heavy or prolonged occlusions.", + "bbox": [ + 511, + 246, + 903, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Pose Estimation under Occlusion. Despite significant progress in HPS estimation, handling occlusions remains a major challenge. This is because, in 3D pose estimation, missing depth cues make reconstruction far more ambiguous than in 2D. Early works like [5] used data augmentation with occlusion labels to enhance robustness in pose estimation. Latest methods like GLAMR [51] use a deep generative motion infiller to handle missing poses under severe occlusions and a global optimization framework to refine motion trajectories. Additionally, some methods [20, 29, 53] approach the issue of missing poses as a pose refinement problem, leveraging temporal smoothness to address it effectively. SmoothNet [53] introduced a temporal motion refinement network for refining the poses obtained from the image-based pose estimators to alleviate jitters. These methods perform well under sparse, infrequent occlusions across frames but struggle with natural, prolonged occlusions, as they lack training for such instances. All these works evaluate their methods on datasets with sparse occlusions, as no existing datasets contain significant occlusions. To address this gap, we introduce VOccl3D, a novel dataset designed specifically for 3D HPS estimation under significant occlusions.", + "bbox": [ + 511, + 553, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7352", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. VOCcl3D Dataset", + "text_level": 1, + "bbox": [ + 89, + 89, + 261, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we outline the key components required to construct the VOcc13D dataset. Section 3.1 details the creation of individual assets, including background scenes, human motions, and texture maps. Sections 3.2 describe the rendering process and attributes of the dataset.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Dataset Assets", + "text_level": 1, + "bbox": [ + 89, + 199, + 238, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background Scenes. In recent years, neural networks have significantly advanced 3D scene representation, enabling novel-view image synthesis. Notable approaches include Neural Radiance Fields (NeRF) [32], which learn a joint representation of geometry and appearance for novel-view synthesis. Another method, 3D Gaussian Splatting (3DGS) [19], represents a scene using 3D Gaussians and adopts a differentiable rasterizer for real-time rendering. These methods use multiview images to learn 3D representations, avoiding extensive capture setups. To develop our dataset, we use the 3DGS method to learn the 3D representation of the background scene.", + "bbox": [ + 89, + 220, + 483, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminaries: 3D Gaussian Splatting We represent the $i$ -th Gaussian in 3DGS as:", + "bbox": [ + 89, + 402, + 483, + 433 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (\\mathbf {p}) = o _ {i} e ^ {- \\frac {1}{2} \\left(\\mathbf {p} - \\boldsymbol {\\mu} _ {i}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {p} - \\boldsymbol {\\mu} _ {i}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 441, + 482, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p} \\in \\mathbb{R}^3$ is a xyz location, $o_i \\in [0,1]$ is the opacity modeling the ratio of radiance the Gaussian absorbs, $\\pmb{\\mu}_i \\in \\mathbb{R}^3$ is the center/mean of the Gaussian, and the covariance matrix $\\pmb{\\Sigma}_i$ is parameterized by the scale $\\mathbf{S}_i \\in \\mathbb{R}_+$ along each of the three Gaussian axes and the rotation $\\mathbf{R}_i \\in SO(3)$ with $\\pmb{\\Sigma}_i = \\mathbf{R}_i\\mathbf{S}_i\\mathbf{S}_i^\\top\\mathbf{R}_i^\\top$ . Each Gaussian is also paired with spherical harmonics [39] to model the radiance emitted towards various directions. During rendering, the 3D Gaussians are projected onto the image plane and form 2D Gaussians [57] with the covariance matrix $\\pmb{\\Sigma}_i^{2\\mathrm{D}} = J\\pmb{W}\\pmb{\\Sigma}_i\\pmb{W}^\\top\\pmb{J}^\\top$ , where $\\pmb{J}$ is the Jacobian of the affine approximation of the projective transformation and $\\pmb{W}$ is the viewing transformation. The color of a pixel is calculated via alpha blending the $N$ Gaussians contributing to a given pixel:", + "bbox": [ + 89, + 472, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {j = 1} ^ {N} c _ {j} \\alpha_ {j} \\prod_ {k = 1} ^ {j - 1} \\left(1 - \\alpha_ {k}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 710, + 482, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the Gaussians are sorted from close to far, $c_{j}$ is the color obtained by evaluating the spherical harmonics given viewing transform $W$ , and $\\alpha_{j}$ is calculated from the 2D Gaussian formulation (with the covariance $\\Sigma_{j}^{2D}$ ) multiplied by its opacity $o_{j}$ .", + "bbox": [ + 89, + 763, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collected RGB videos from the large-scale open-source DL3DV dataset [27], which contains over 10,000 videos across diverse domains, including natural and outdoor settings, educational institutions, shopping complexes,", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "parks, hubs, cafes, and restaurants. We selected videos based on the presence of natural occlusions, such as garlands, chairs, benches, statues, cars, and bins. Our dataset captures 3D representations of approximately 40 distinct scenes, each incorporating real-world occlusions. This approach to learning 3D representation closely resembles real-world data compared to conventional methods that rely on rendering from 3D graphic asset scenes.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SMPLX body/Human Animations. We represent the human body using the SMPL-X [36] 3D human model, defined by the function $\\mathcal{M}(\\theta ,\\beta ,\\psi)$ , where $\\theta$ represents the pose parameters, $\\beta$ the shape parameters, and $\\psi$ the facial expression parameters. This function outputs a body mesh $\\mathcal{M}\\in \\mathbb{R}^{10475\\times 3}$ with 10,475 vertices. We sample approximately 400 SMPL-X 3D human motion models from the AMASS mocap dataset [30], which contains over 11,000 motion sequences with diverse body shapes and poses. Following previous work [48], we use the pre-trained human pose prior model VPoser [36] to assess pose difficulty. VPoser is a Variational Auto-Encoder model that evaluates a pose $\\theta$ to be challenging if its embeddings $\\epsilon_{\\theta}$ have larger norm, i.e. $||\\epsilon_{\\theta}||_{2} > \\tau$ , where $\\tau$ is empirically set to 40. We classify a challenging pose using VPoser to avoid simple movements such as walking, standing, jogging, etc. To prepare human animations for rendering, we use Blender graphics software to efficiently bake complex geometric animations over time into the Alembic format for subsequent use in rendering engines.", + "bbox": [ + 511, + 224, + 906, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We ensured that motion sequences contained at least 180 frames to maintain a minimum animation duration of six seconds at 30 fps. For sequences exceeding 400 frames, we discarded the first 100 frames to remove static poses at the start. To ensure that human motion remains consistently within the occluded scene, we implemented boundary constraints that stop the motion if it moves beyond these limits. These conditions prevent humans from moving too far from the occlusion area. Additionally, we applied random rotations and translations to each sequence. We store the applied transformations to further calculate the effective global orientation of the human body with respect to the world coordinates.", + "bbox": [ + 511, + 539, + 908, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Textures. We sourced human body texture scans from the open-source dataset provided by the SMPLtex method [4], which estimates and manipulates the full 3D appearance of humans captured from a single image. The SMPLtex dataset provides a diverse range of human texture scans, covering various skin tones, clothing styles, and genders. To ensure sample diversity in VOcc13D, we select approximately 200 distinct texture scans. Figure 4 illustrates the different clothing textures applied to the SMPL-X human model in our dataset.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7353", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3f841e87163ddc3c5a9d1d8f2b573a19a0a973e52ae81e67c77ebea9a35f986.jpg", + "image_caption": [ + "Figure 3. An overview of our proposed dataset creation pipeline. We generate 3D representations of natural scenes with real occlusions using 3D Gaussian Splatting [19]. Human motion sequences from the AMASS MoCap dataset [30] are processed in Blender to generate baked animation files. The generated scene representations, animation files, human texture maps, and camera extrinsic parameters are imported into the Unity rendering engine to generate video sequences of human motion under occlusion." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 911, + 315 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/deff0eb4f7e24634e7b727e37a11bdef5b88cd710bd7626147f2bd84321b434a.jpg", + "image_caption": [ + "Figure 4. Diversity of human texture in VOccl3D. We use approximately 200 human texture maps from the SMPLtex dataset [4], encompassing diverse clothing styles, genders, skin tones, and ethnicities." + ], + "image_footnote": [], + "bbox": [ + 130, + 393, + 444, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Dataset Rendering and Attributes", + "text_level": 1, + "bbox": [ + 89, + 630, + 387, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering As shown in Figure 3, we used Unity Engine to render synthetic scenes by integrating background Gaussian splats, human animations, and body texture assets. We rendered video sequences at 30 fps with a resolution of $720 \\times 720$ , varying the camera field of view between 25 and 50 across different scenes. For each sequence, we stored ground-truth camera extrinsic and intrinsic parameters and saved the rendered RGB images in lossless PNG format. We positioned the camera to capture occluded views within the scene and applied motion constraints to human animation assets to maintain consistent occlusion throughout each sequence. To achieve realistic lighting, we employed Unity's built-in DirectionalLight asset to simulate natural illumination on the human mesh.", + "bbox": [ + 89, + 655, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Attributes. Our proposed dataset, VOccl3D, includes over 250,000 images and 400 video sequences with", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a total runtime exceeding 2 hours and 30 minutes. The dataset features 40 background scenes with various occlusions, such as cars, garlands, benches, chairs, bins, and trees. Each scene contains 10 video sequences with variations in human motions and clothing textures. The rendered sequence exhibit diversity in body shapes, skin tones, and camera poses. We provide ground-truth annotations, including camera extrinsic and intrinsic parameters, pose, shape, global orientation, translation, gender, 2D keypoints, and a binary occlusion label for each keypoint.", + "bbox": [ + 511, + 393, + 906, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different Modalities. In addition to 3D pose and shape annotations, VOccl3D provides annotations for multiple modalities, including human silhouettes, body-part segmentation, 2D keypoint estimation, and human bounding box detection under occlusion. Researchers can use our dataset to train and evaluate methods designed to handle occlusion across these modalities.", + "bbox": [ + 511, + 546, + 908, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 511, + 657, + 748, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we highlight the need and significance of our proposed VOccl3D dataset. We fine-tune the state-of-the-art pose estimation methods using our dataset and report the qualitative and quantitative results. Our results demonstrate improved performance on the HPS estimation task (Section 4.1) using both our proposed synthetic dataset and real-world datasets with occlusions. In Section 4.2, we show the enhanced performance of the human object detector and its impact on HPS estimation under occlusion.", + "bbox": [ + 511, + 683, + 906, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Human Pose and Shape Estimation", + "text_level": 1, + "bbox": [ + 511, + 832, + 821, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and Implementation Setup. We use approximately 200k and 50k images for training and testing respectively in our proposed VOcc13D dataset. We fine-tune", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7354", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0d914f49d8f267a5e9bcebb88016548bd3faadd0688dacef6d54c4c358ce31be.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHard-OcclusionMedium-OcclusionLow-Occlusion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]192.22114.35247.41121.7078.56158.4698.8267.64126.92
BEDLAM-HMR [3]167.35102.92214.39102.5568.13134.5986.5554.48110.15
BEDLAM-CLIFF [3]154.8699.53199.9590.9765.03119.6374.9552.6596.60
HMR2.0 [11]169.71100.49215.17113.8871.78145.6288.5359.08114.39
STRIDE with BEDLAM-CLIFF [24]155.64100.44-91.1465.38-75.0253.21-
WHAM [44]152.15102.14177.07110.9776.81127.4593.9066.68106.51
VOcc13D-B-CLIFF136.3489.94175.9282.4858.78106.8469.4646.3288.19
STRIDE with VOcc13D-B-CLIFF136.4390.28-82.3758.98-69.6546.86-
", + "bbox": [ + 94, + 88, + 906, + 238 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/52bc5b8701d69d0b9c4faa4ce50f5a99bf2cf6319d7a4bdac953c5de68805907.jpg", + "table_caption": [ + "Table 1. 3D HPS estimation results on the test-split of VOcccl3D. The results show that VOcccl3D-B-CLIFF and STRIDE, when using pseudo-labels from VOcccl3D-B-CLIFF, significantly outperform other image- and video-based HPS estimation methods across hard, medium, and low occlusion categories. As expected, all HPS estimation methods exhibit a decline in performance as occlusion severity increases from low to medium to high. For evaluation, we use ground-truth bounding boxes. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
Method3DPWOcclType1-3DPWOcclType2-3DPW
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]73.946.487.698.1562.27118.4799.4962.16119.82
BEDLAM-HMR [3]79.047.693.1108.6266.53128.05106.1964.05125.66
BEDLAM-CLIFF [3]72.046.685.098.7164.26117.4196.8061.32115.33
HMR2.0 [11]81.254.3143.7103.4069.66164.5599.0166.17158.79
VOccl3D-B-CLIFF72.047.384.595.8963.43114.2894.3660.44112.01
VOccl3D-CLIFF71.1045.9884.2595.1761.83114.9593.7459.66112.59
", + "bbox": [ + 135, + 316, + 867, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. 3D HPS estimation results on 3DPW, OcclType1-3DPW, and OcclType2-3DPW. Since 3DPW is a real-world dataset with minimal occlusions, VOccl3D-CLIFF outperforms all other methods but shows only a marginal improvement over BEDLAM-CLIFF. However, on OcclType1-3DPW and OcclType2-3DPW, which contain significant occlusions on real-world dataset, our method demonstrates a notable improvement over existing approaches. The best results are noted in bold.", + "bbox": [ + 89, + 453, + 906, + 510 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a2fb20081034a910d905c0023aa4060635f6fec6e8355823f136332f4cd760e.jpg", + "image_caption": [ + "Figure 5. Fine-tuning an off-the-shelf pose estimator. We fine-tune a CLIFF [25] pose estimation model using our proposed VOcc13D dataset. We observe a significant improvement in the estimated mesh when using a fine-tuned CLIFF model using our dataset. Note: The pose estimator can be any off-the-shelf pose estimator, here we show results with the CLIFF model [25]." + ], + "image_footnote": [], + "bbox": [ + 91, + 532, + 480, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "two versions of state-of-the-art model, CLIFF (trained on Human3.6M [15], MPI-INF-3DHP [31], and COCO [26] dataset) and BEDLAM-CLIFF (trained on BEDLAM and AGORA datasets) using our rendered VOcc13D dataset, resulting in VOcc13D-CLIFF and VOcc13D-B-CLIFF, re", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "spectively (refer to Figure 5). To benchmark performance, we evaluate on the test split of VOccl3D, comparing against multiple image and video based baselines, including CLIFF [25], BEDLAM-HMR [3], BEDLAM-CLIFF [3], HMR2.0 [11], STRIDE [24], and WHAM [44]. We report the mean performance by evaluating the model using a fivefold cross-validation on our dataset. To quantify occlusion levels, we annotate each 3D joint with a binary occlusion label, marking it as occluded if its corresponding 2D keypoint lies within the ground-truth segmentation mask of the visible human region. Based on the number of visible keypoints (out of 22 total), we categorize images into three occlusion levels: hard occlusion (4-9 visible keypoints), medium occlusion (10-15 visible keypoints), and low occlusion (16-20 visible keypoints).", + "bbox": [ + 511, + 537, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Beyond evaluations on synthetic datasets, we assess the performance of VOccl3D fine-tuned models on real-world datasets, including 3DPW [46] and OCMotion [13]. It is noteworthy that these datasets do not contain significant occlusions. Hence, we introduce two occlusion-augmented variants of 3DPW to evaluate robustness under severe occlusions. In OcclType1-3DPW, we overlay a black patch on a randomly selected keypoint, while in OcclType2-3DPW, we occlude two random keypoints. Further details on oc", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7355", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a712e5eb8c5fe1284bfca9681cfe5c6b9be67f71f4352d369042bb5f9a0dc1f.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of HPS estimation methods on VOcc3D dataset. The first and second column shows RGB image and ground-truth human mesh. Column third and fourth compare HPS estimation using the BEDLAM-CLIFF [3] and HMR2.0 [11] methods. The final column (VOcc3D-B-CLIFF) presents results obtained by fine-tuning the CLIFF model on the VOcc3D dataset. These results demonstrate superior performance, particularly in scenarios with heavy occlusion." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "clusion variants of 3DPW and implementation details are provided in the supplementary material.", + "bbox": [ + 89, + 569, + 482, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We present benchmarking results on the test split of our proposed VOcc13D dataset in Table 1. Our finetuned VOcc13D-B-CLIFF significantly outperforms existing state-of-the-art image and video-based methods across all occlusion categories. Notably, the performance of all previous methods degrades as occlusion severity increases, highlighting the inherent challenges of HPS estimation under occlusions. We also compare our approach with the plug-and-play method STRIDE [24], which performs 3D pose estimation and reports only MPJPE and PA-MPJPE errors. Our results show that STRIDE achieves superior performance when leveraging pseudo-labels from the VOcc13D-B-CLIFF model compared to those from the original BEDLAM-CLIFF model, further validating the effectiveness of our dataset. In Table 2 and Table 3, we report performance on the 3DPW [45] and OCMotion [14] datasets, respectively. We observe that both VOcc13D-B-CLIFF and VOcc13D-CLIFF outperform all the previous methods, demonstrating their effectiveness in real-world", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scenarios. However, since these datasets contain minimal occlusions, the improvements over CLIFF remain marginal. To further assess robustness under heavy occlusion, we evaluate on OcclType1-3DPW and OcclType2-3DPW, real-world datasets with significant occlusions. Table 2 demonstrates that our fine-tuned model achieves substantial improvements, underscoring its robustness in occluded environments. Table 3 presents results on the OCMotion dataset, which features lower levels of occlusion. We can observe that VOcc13D-CLIFF outperforms all comparison methods and demonstrates performance comparable to CLIFF [25]. This is expected, as both models are pretrained on large-scale real-world datasets captured in controlled lab environments similar to OCMotion. Figure 6 presents the qualitative results on VOcc13Ddataset, showcasing the improvement of our fine-tuned model against other state-of-the-art HPS estimation methods. Additional results are provided in supplementary material.", + "bbox": [ + 511, + 569, + 906, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Impact of Human Detector on HPS Estimation", + "text_level": 1, + "bbox": [ + 511, + 848, + 906, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset and Implementation Setup. Human object detectors play a crucial role in HPS estimation tasks, particularly", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7356", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/650cb8f892498b5d08d321b17eb1f2477138918f2b94bc2747cb1f770de0e04e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPEOCMotion PA-MPJPEPVE
CLIFF [25]64.1540.1679.80
BEDLAM-HMR [3]73.9641.9492.70
BEDLAM-CLIFF [3]66.8041.7983.86
HMR2.0 [11]69.9443.0087.07
VOcc13D-B-CLIFF65.9640.5981.96
VOcc13D-CLIFF64.2939.6478.56
", + "bbox": [ + 130, + 88, + 442, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2d7e59b55027f5007112083acc6a7b379b6b158804c4733a3eff707bf6745e46.jpg", + "table_caption": [ + "Table 3. Qualitative results of HPS estimation on OCMotion. Our results show that VOcc13D-CLIFF outperforms all HPS estimation methods except CLIFF [25]. Both CLIFF and VOcc13D-CLIFF exhibit comparable performance, as both are pre-trained on real-world datasets captured in controlled lab environments, similar to OCMotion [14]. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
Method3DPWOCMotion
mAP50mAP75mAP50mAP75
YOLO1158.9947.1498.8491.80
VOccl3D-YOLO1159.8948.2699.1091.95
", + "bbox": [ + 124, + 316, + 454, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "when inferring from in-the-wild RGB images [3, 24, 25]. Ideally, HPS estimation performs optimally with ground-truth bounding boxes; however, most human object detectors struggle significantly under occlusion. To address this, we fine-tune the recent YOLO11 detector [16] on the combined training split of VOccl3D and MS COCO, referring to the fine-tuned model as VOccl3D-YOLO11. We show evaluations on 3DPW [45] and OCMotion [14] datasets.", + "bbox": [ + 89, + 489, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Table 4 compares the detection performance of the pre-trained YOLO11 and VOccl3D-YOLO11 models on the real-world 3DPW [45] and OCMotion [14] datasets. Since OCMotion is a single-human dataset with minimal occlusions, the detector achieves high mAP scores. This is likely because most body parts remain visible, making box detection easier. However, due to the presence of multiple individuals in 3DPW, the overall mAP scores remain lower. In short, VOccl3D-YOLO11 shows significant improvement on datasets with high occlusions. We provide qualitative results in the supplementary material.", + "bbox": [ + 89, + 612, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 further evaluates the impact of human detection on HPS estimation. The first, second, and third rows represent the performance of VOccl3D-CLIFF when human detections are sourced from ground truth, YOLO11, and VOccl3D-YOLO11, respectively. As expected, the best performance is achieved when using ground-truth detections. However, fine-tuning the detector with VOccl3D significantly improves HPS estimation compared to detections", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/88d95860ab880f0cdb8a657e5e8af439815f88839522115a38cc84fce76e0faa.jpg", + "table_caption": [ + "Table 4. Quantitative results of YOLO11 on 3DPW and OC-Motion datasets. The first row presents the human object detection performance of the pre-trained YOLO11 model, while the second row shows the results after fine-tuning with the VOccl3D dataset. The best results are in bold. We observe an improvement in detection accuracy across both datasets after fine-tuning." + ], + "table_footnote": [], + "table_body": "
Method3DPWOCMotion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
VOccl3D-CLIFF w/GT71.1045.9884.2564.2939.6478.56
VOccl3D-CLIFF w/YYOLO11116.5263.35139.7467.1641.3083.00
VOccl3D-CLIFF w/VOccl3D-YYOLO11114.8562.66137.1966.6541.1582.40
", + "bbox": [ + 517, + 88, + 905, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. 3D HPS estimation results using the VOcc13D-CLIFF model with different bounding box sources. This table presents HPS estimation performance on the 3DPW and OC-Motion datasets. The first row reports results using ground-truth bounding boxes, achieving the highest accuracy. The second and third rows show performance when detections are obtained from the pre-trained YOLO11 and fine-tuned VOcc13D-YOLO11 models, respectively. Notably, the third row demonstrates a significant improvement over the second, highlighting the impact of detections on HPS estimation.", + "bbox": [ + 511, + 186, + 903, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from the pre-trained YOLO11. These results underscore the critical role of human detection quality in enhancing HPS performance under occlusion.", + "bbox": [ + 511, + 344, + 903, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments show that fine-tuning an existing HPS estimation model with a large synthetic dataset containing occlusions enhances its performance on both real and synthetic datasets. We observe significant improvements over baseline methods, particularly in scenarios with heavy occlusions. Additionally we demonstrate that the poor results are not solely due to HPS estimations errors but also stem from failures in bounding box detection. The VOcc13D dataset proves effective in refining pose estimation methods and bounding box predictions under occlusions.", + "bbox": [ + 511, + 391, + 905, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 556, + 633, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce VOccl3D, a novel large-scale video dataset with synthetic humans in real-world scenes under occlusions for 3D human pose and shape estimation. By leveraging a rendering-based approach, VOccl3D eliminates the need for costly and labor-intensive data collection while providing a diverse and realistic dataset for occlusion-aware research. Through extensive qualitative and quantitative evaluations, we demonstrate that fine-tuning existing HPS estimation methods with VOccl3D significantly enhances their performance on real-world datasets with occlusions and on the test split of VOccl3D dataset. Furthermore, we improve the human object detector, YOLO11 under occluded conditions using VOccl3D, highlighting the impact of detections in achieving robust HPS estimation in occluded scenarios. Beyond HPS estimation, VOccl3D serves as a comprehensive benchmark for evaluating methods across multiple occlusion-aware tasks, including human body-part segmentation, 2D/3D pose estimation, and human bounding box detection. Hence, VOccl3D sets a new standard for occluded human benchmarks, offering a valuable dataset for advancing occlusion-robust research.", + "bbox": [ + 511, + 583, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7357", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by USDA NRI grant 2021-67022-33453, NSF grants CMMI-2326309 and CNS-2312395. This research was also partially supported by the Office of the Director of National Intelligence (ODNI), specifically through the Intelligence Advanced Research Projects Activity (IARPA), under contract number [2022-21102100007]. The views and conclusions in this research reflect those of the authors and should not be construed as officially representing the policies, whether explicitly or implicitly, of ODNI, IARPA, or the U.S. Government. Nevertheless, the U.S. Government retains the authorization to reproduce and distribute reprints for official government purposes, regardless of any copyright notices included.", + "bbox": [ + 89, + 113, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 321, + 187, + 338 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3", + "[2] Peter Bauer, Arij Bouazizi, Ulrich Kressel, and Fabian B. Flohr. Weakly supervised multi-modal 3d human body pose estimation for autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7, 2023. 2", + "[3] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion, 2023. 2, 3, 6, 7, 8", + "[4] Dan Casas and Marc Comino-Trinidad. SMPLitex: A Generative Model and Dataset for 3D Human Texture Estimation from Single Image. In British Machine Vision Conference (BMVC), 2023. 4, 5", + "[5] Yu Cheng, Bo Yang, Bo Wang, Yan Wending, and Robby Tan. Occlusion-aware networks for 3d human pose estimation in video. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 723-732, 2019. 3", + "[6] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3", + "[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 591-601, 2022. 2", + "[8] R James Cotton. Posepipe: Open-source human pose estimation pipeline for rehabilitation research. Archives of Physical Medicine and Rehabilitation, 103(12):e161-e162, 2022. 2", + "[9] Yufeng Cui and Yimei Kang. Multi-modal gait recognition via effective spatial-temporal feature fusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17957, 2023. 2", + "[10] Yang Fu, Shibei Meng, Saihui Hou, Xuecai Hu, and Yongzhen Huang. Gpgait: Generalized pose-based gait" + ], + "bbox": [ + 93, + 347, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "recognition. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19538-19547, 2023. 2", + "bbox": [ + 544, + 92, + 905, + 119 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 6, 7, 8", + "bbox": [ + 516, + 122, + 905, + 190 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[12] Charlie Hewitt, Fatemeh Saleh, Sadegh Aliakbarian, Lohit Petikam, Shideh RezaEIFar, Louis Florentin, Zafiirah Hosenie, Thomas J Cashman, Julien Valentin, Darren Cosker, et al. Look ma, no markers: Holistic performance capture without the hassle. ACM Transactions on Graphics, 43(6): 1-12, 2024. 2, 3", + "bbox": [ + 516, + 191, + 905, + 273 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[13] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2022. 6", + "bbox": [ + 516, + 276, + 903, + 330 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[14] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2023. 2, 7, 8", + "bbox": [ + 516, + 333, + 903, + 388 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 3, 6", + "[16] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 3, 8", + "[17] Sardor Juraev, Akash Ghimire, Jumabek Alikhanov, Vijay Kakani, and Hakil Kim. Exploring human pose estimation and the usage of synthetic data for elderly fall detection in real-world surveillance. IEEE Access, 10:94249-94261, 2022. 2", + "[18] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2017. 2, 3", + "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4, 5", + "[20] Do-Yeop Kim and Ju-Yong Chang. Attention-based 3d human pose sequence refinement network. Sensors, 21(13), 2021. 3", + "[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3", + "[22] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2", + "[23] Farnoosh Koleini, Muhammad Usama Saleem, Pu Wang, Hongfei Xue, Ahmed Helmy, and Abbey Fenwick. Biopose: Biomechanically-accurate 3d pose estimation from monocular videos. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6330–6339. IEEE, 2025. 2" + ], + "bbox": [ + 93, + 347, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7358", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Rohit Lal, Saketh Bachu, Yash Garg, Arindam Dutta, Calvin-Khang Ta, Dripta S. Raychaudhuri, Hannah Dela Cruz, M. Salman Asif, and Amit K. Roy-Chowdhury. Stride: Single-video based temporally continuous occlusion robust 3d pose estimation, 2024. 2, 3, 6, 7, 8", + "[25] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation, 2022. 2, 3, 6, 7, 8", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6", + "[27] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 4", + "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3", + "[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3", + "[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 5", + "[31] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 International Conference on 3D Vision (3DV), pages 506-516, 2017. 6", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 4", + "[33] Sara Moccia, Lucia Migliorelli, Virgilio Carnielli, and Emanuele Frontoni. Preterm infants' pose estimation with spatio-temporal features. IEEE Transactions on Biomedical Engineering, 67(8):2370-2380, 2019. 2", + "[34] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J. Mitra. imapper: interaction-guided scene mapping from monocular videos. ACM Trans. Graph., 38(4), 2019. 2", + "[35] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13468-13478, 2021. 2", + "[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4", + "[37] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7745-7754, 2018. 3", + "[38] Miroslav Purkrábek and Jiri Matas. Improving 2d human pose estimation across unseen camera views with synthetic data. arXiv preprint arXiv:2307.06737, 2023. 3", + "[39] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4", + "[40] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 2", + "[41] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, page 1530-1538. JMLR.org, 2015. 3", + "[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 8(11):7090-7097, 2023. 2", + "[43] Saurabh Sharma, Pavan Teja Varigonda, Prashast Bindal, Abhishek Sharma, and Arjun Jain. Monocular 3d human pose estimation by generation and ordinal ranking. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3", + "[44] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2070-2080, 2023. 2, 3, 6", + "[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 7, 8", + "[46] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European conference on computer vision (ECCV), pages 601-617, 2018. 6", + "[47] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In International Conference on Computer Vision (ICCV), 2021. 3", + "[48] Zhenzhen Weng, Laura Bravo-Sanchez, and Serena Yeung-Levy. Diffusion-hpc: Synthetic data generation for human mesh recovery in challenging domains. In 2024 International Conference on 3D Vision (3DV), pages 257-267. IEEE, 2024. 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7359", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, et al. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20282-20292, 2023. 3", + "[50] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9060-9069, 2021. 3", + "[51] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[52] Andrei Zanfir, Mihai Zanfir, Alex Gorban, Jingwei Ji, Yin Zhou, Dragomir Anguelov, and Cristian Sminchisescu. HUM3DIL: Semi-supervised multi-modal 3d humanpose estimation for autonomous driving. In 6th Annual Conference on Robot Learning, 2022. 2", + "[53] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In European Conference on Computer Vision. Springer, 2022. 3", + "[54] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conference on Computer Vision and Pattern Recognition, (CVPR), 2020. 2", + "[55] Yi Zhang, Pengliang Ji, Adam Kortylewski, Angtian Wang, Jieru Mei, and Alan L Yuille. 3D-Aware Neural Body Fitting for Occlusion Robust 3D Human Pose Estimation. In The IEEE/CVF International Conference on Computer Vision, 2023. 2", + "[56] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G. Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4966-4975, 2015. 3", + "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 4" + ], + "bbox": [ + 91, + 92, + 480, + 696 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7360", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_model.json b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f3054ebbab3dc2eb367aa4df87df3259dd30f0b --- /dev/null +++ b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_model.json @@ -0,0 +1,1993 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.798, + 0.175 + ], + "angle": 0, + "content": "VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.204, + 0.767, + 0.257 + ], + "angle": 0, + "content": "Yash Garg, Saketh Bachu, Arindam Dutta, Rohit Lal†, Sarosij Bose, Calvin-Khang Ta‡, M. Salman Asif, Amit Roy-Chowdhury University of California, Riverside" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.26, + 0.817, + 0.275 + ], + "angle": 0, + "content": "{ygarg002,sbach008,adultt020,r1al011,sbose007,cta003,sasif,amitrc}@ucr.edu" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.287, + 0.903, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Figure 1. We propose VOccl3D, a large-scale synthetic video dataset specifically designed for training and evaluating algorithms for 3D human pose and shape estimation (HPS) in realistic occlusion scenarios. VOccl3D comprises over 250,000 frames, with a total runtime exceeding 2 hours and 30 minutes. Compared to previous occlusion-based datasets, VOccl3D has diverse body shapes, textures, and most importantly, significant and realistic occlusions within the scenes. In addition to body shape and pose, VOccl3D provides other necessary ground truth information, such as bounding boxes, body part segmentations, and human silhouettes. The top row illustrates various frames from our dataset, showcasing diverse occlusions, clothing textures, and motions, while the bottom row represents a sequence of frames from a video sequence within our dataset." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.592, + 0.327, + 0.608 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Human pose and shape (HPS) estimation methods have been extensively studied, with many demonstrating high zero-shot performance on in-the-wild images and videos. However, these methods often struggle in challenging scenarios involving complex human poses or significant occlusions. Although some studies address 3D human pose estimation under occlusion, they typically evaluate performance on datasets that lack realistic or substantial occlusions, e.g., most existing datasets introduce occlusions with random patches over the human or clipart-style overlays, which may not reflect real-world challenges. To bridge this gap in realistic occlusion datasets, we introduce a novel benchmark dataset, VOccl3D, a Video-based human Occlusion dataset with 3D body pose and shape annotations. Inspired by works such as AGORA and BEDLAM, we constructed this dataset using advanced computer graphics rendering techniques, incorporating diverse real-world occlusion scenarios, clothing textures, and human motions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.821 + ], + "angle": 0, + "content": "Additionally, we fine-tuned recent HPS methods, CLIFF and BEDLAM-CLIFF, on our dataset, demonstrating significant qualitative and quantitative improvements across multiple public datasets, as well as on the test split of our dataset, while comparing its performance with other state-of-the-art methods. Furthermore, we leveraged our dataset to enhance human detection performance under occlusion by fine-tuning an existing object detector, YOLO11, thus leading to a robust end-to-end HPS estimation system under occlusions. Overall, this dataset serves as a valuable resource for future research aimed at benchmarking methods designed to handle occlusions, offering a more realistic alternative to existing occlusion datasets. See the Project page for code and dataset:https://yashgarg98.github.io/Voccl3D-dataset/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.9 + ], + "angle": 0, + "content": "† Currently at NASA MSFC IMPACT. ‡ Currently at Dolby Laboratories. Work done while the authors were at UCR." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7350" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.326 + ], + "angle": 0, + "content": "Monocular 3D human pose and shape (HPS) estimation is a complex yet essential task in computer vision. It has applications in surveillance [7, 17], autonomous robotics [2, 42, 52], human motion analysis [9, 10], and clinical assessment [8, 23, 33]. Since the introduction of the neural network-based Human Mesh Recovery Network (HMR) [18], numerous approaches have advanced the field by enabling accurate estimation of SMPL[28] pose and shape from a single RGB image. Recent methods [1, 21, 24, 25, 29, 44] demonstrate impressive zero-shot HPS performance on in-the-wild RGB images and video sequences. However, these methods still face limitations in achieving high performance in challenging scenarios, such as complex human poses and under significant occlusions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Achieving robust HPS under occlusion remains a challenging problem due to the contextual ambiguity of the occluded body parts. Several methods [22, 40, 51, 55] have been proposed to address the challenges caused by occlusions. Methods like PARE [22] focus on using visible body parts to infer occluded areas, enhancing estimation accuracy with partial views. Temporal models like HuMoR [40] and GLAMR [51] use generative frameworks to ensure pose continuity over time. HuMoR predicts pose distributions, while GLAMR incorporates global trajectory data to fill in missing poses. Recent methods, like STRIDE [24], leverage a large-scale pre-trained motion prior model to achieve temporally coherent pose reconstruction under occlusions. While these methods effectively leverage visible information and temporal continuity, they still struggle under severe occlusions due to limited exposure to occluded scenarios in their training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Evaluations of the existing methods are often limited to occluded datasets that lack scene diversity [34], involve moderate or lower levels of occlusion [14, 45], or use patch-based occlusion datasets for training and inference [55], as illustrated in Figure 2. Most importantly, there remains an urgent need for a realistic, diverse, and significantly occluded human pose and shape dataset to advance the task of 3D HPS estimation. To address this gap, our work proposes VOccl3D, a large-scale video dataset with synthetic humans in real scenes for 3D human pose and shape estimation. Following prior works such as BEDLAM [3] and Synthocap [12], we use an advanced computer graphics engine to render a highly realistic synthetic dataset. Prior research has shown that fine-tuning models on synthetic datasets can significantly improve HPS estimation performance on real-world datasets [3]. Synthetic datasets provide \"perfect\" ground-truth annotations and eliminate the need for costly sensors, making it computationally efficient to generate large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our proposed dataset includes approximately 250,000 images and 400 video sequences. To introduce diversity, we" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.65, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.168, + 0.636, + 0.178 + ], + "angle": 0, + "content": "a) 3DPW-AdvOcc@80" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.089, + 0.778, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.168, + 0.774, + 0.177 + ], + "angle": 0, + "content": "b) Occluded Human3.6M" + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.089, + 0.894, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.167, + 0.896, + 0.177 + ], + "angle": 0, + "content": "c) Syn. Occ. Human3.6M" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.18, + 0.652, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.268, + 0.612, + 0.278 + ], + "angle": 0, + "content": "d) OCMotion" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.179, + 0.753, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.267, + 0.736, + 0.278 + ], + "angle": 0, + "content": "e) 3DPW" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.179, + 0.902, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.268, + 0.854, + 0.278 + ], + "angle": 0, + "content": "f) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Samples from various occlusion-based HPS datasets. The top row displays datasets with artificial, patch-based occlusions applied to existing datasets, including 3DPW-AdvOcc [45, 55], Occluded Human3.6M [15, 24], and Synthetic Occlusion Human3.6M [15, 54]. The bottom row presents samples from two datasets with natural occlusions: OCMotion [14] and 3DPW [45]. Notably, the top-row images exhibit unrealistic occlusions, which lack the realism of naturally occurring occlusions. The bottom row images contain natural occlusions but are sparse and infrequent. Compared to existing datasets, our proposed VOcccl3D dataset features more realistic occlusions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.702 + ], + "angle": 0, + "content": "incorporate human motion samples from the AMASS [30] dataset, over 200 distinct clothing textures for sequences containing both male and female genders, and 40 real background scenes with occlusions. Additionally, we provide occlusion labels for each body joint. Unlike prior methods [3, 35] which render synthetic environments, we utilize 3D Gaussian splatting (3DGS) [19] to create 3D representations of background scenes. This approach enhances the realism of rendered scenes, making them closely resemble real-world captured data. Reconstructing background scenes using 3DGS offers the flexibility to create domain-specific datasets, such as for agriculture, street or indoor environments using raw RGB video captures without relying on costly or limited graphic assets. Figure 1 showcases samples from our dataset, illustrating the diversity in clothing, environments, human motions, and occlusions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We fine-tuned the CLIFF [25] and BEDLAM-CLIFF [3] models on our dataset to demonstrate that training with synthetic data enhances the performance of existing HPS estimation methods under occlusion. We evaluated state-of-the-art HPS estimation methods on the test split of our dataset across three occlusion levels: high, medium, and low. The results show that existing methods perform poorly under high occlusion, whereas our fine-tuned models achieve significant improvements. To assess performance on real-world data, we perform evaluations on the 3DPW and OCMotion datasets, which contain low-level occlusions. Additionally, we created a variant of the 3DPW dataset by adding black patches to evaluate the robustness" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7351" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "against high-occlusion real scenarios. Our fine-tuned models outperformed recent HPS methods on these real-world datasets. Furthermore, our experiments highlight the importance of an effective human object detector for improving HPS performance under occlusion. To address this, we fine-tuned the YOLO11[16] object detector using our dataset, enhancing its performance under occluded scenarios. VOcc13D is available for researchers to benchmark and evaluate occlusion-aware methods." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.228, + 0.472, + 0.243 + ], + "angle": 0, + "content": "In summary, we make the following key contributions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.245, + 0.483, + 0.29 + ], + "angle": 0, + "content": "1. We propose VOcc13D, a novel large-scale, realistic video-based dataset of occluded synthetic humans in real scenes for 3D human pose and shape estimation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.29, + 0.482, + 0.35 + ], + "angle": 0, + "content": "2. We demonstrate both qualitatively and quantitatively that existing HPS estimation methods fine-tuned with VOccl3D outperform other methods on real-world datasets with occlusions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.351, + 0.482, + 0.409 + ], + "angle": 0, + "content": "3. We improved the performance of the human object detector under occlusion scenarios by leveraging the VOccl3D dataset which is a crucial component for a robust HPS estimation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.411, + 0.482, + 0.486 + ], + "angle": 0, + "content": "4. VOccl3D can serve as a benchmark dataset for evaluating methods specifically designed to perform under occlusion across various tasks, including human and body-part segmentation, 2D/3D pose estimation, and human bounding box detection." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.245, + 0.483, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.24, + 0.514 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.538 + ], + "angle": 0, + "content": "Synthetic Data Generation for Human Pose Estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Human pose estimation is crucial in computer vision, but state-of-the-art methods depend on costly, labor-intensive labeled datasets. Several notable synthetic datasets have advanced research in human pose estimation. SynBody [49] provides a large-scale synthetic dataset for human mesh recovery and view synthesis, while RePoGen [38] enables fine-grained control over pose and viewpoint to generate rare, complex poses. Human3.6m [15] provides additional small mixed reality dataset by inserting 3D animation models with background scenes. BEDLAM [3] further highlights that models trained solely on synthetic data can outperform real-data-trained counterparts, emphasizing the importance of high-quality synthetic datasets for transferable models. Further, PressurePose [6] simulates interactions between articulated and soft bodies, capturing fine-grained contact dynamics. SynthMocap [12] extends this by providing expressive synthetic data with detailed body, hand, and face movements. Despite their success, these synthetic datasets lack emphasis on significant occlusions, a key real-world challenge. Addressing this limitation, our work proposes a synthetic dataset tailored to 3D pose estimation under heavy and realistic occlusion, aiming to bridge this gap and enhance robustness in occlusion-prone environments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Image and Video based Human Pose Estimation. Esti" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "mating 2D/3D pose and shape from single RGB image has widely been explored in previous research [3, 25, 43, 47, 50]. Methods such as [43] use a conditional variational autoencoder for 2D-to-3D lifting in pose estimation. The pose estimation approach proposed in [47] applies normalizing flows [41] for 2D-3D mapping. Large-scale datasets like BEDLAM [3] have improved pose estimation models like CLIFF [25] and HMRNet [18]. However, these models struggle with generalizing to unseen scenarios with severe occlusions due to limited training on such cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.905, + 0.549 + ], + "angle": 0, + "content": "In addition, video-based human pose estimation has significantly improved challenging datasets. Early works like [56] use the EM algorithm to estimate the 3D pose from monocular video through 2D joint uncertainty maps. The method in [37] employs dilated temporal convolutions and semi-supervised learning for 3D pose estimation, while [1] uses the SMPL model [28] to extract pose and shape parameters, refining models like HMR with bundle adjustment. VIBE [21] applies adversarial learning with AMASS [30] for 3D pose extraction, while MEVA [29] improves accuracy and smoothness using a variational autoencoder to address VIBE's high acceleration error. Owing to these drawbacks, HuMoR employs a conditional variational autoencoder for robust pose estimation, while the state-of-the-art method WHAM [44] integrates 2D-3D lifting and SLAM for accurate global motion estimation. While these methods perform well on various datasets, studies like [24] highlight performance drops under significant occlusions, as most datasets contain only sparse occlusions, making models struggle with unseen heavy or prolonged occlusions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Human Pose Estimation under Occlusion. Despite significant progress in HPS estimation, handling occlusions remains a major challenge. This is because, in 3D pose estimation, missing depth cues make reconstruction far more ambiguous than in 2D. Early works like [5] used data augmentation with occlusion labels to enhance robustness in pose estimation. Latest methods like GLAMR [51] use a deep generative motion infiller to handle missing poses under severe occlusions and a global optimization framework to refine motion trajectories. Additionally, some methods [20, 29, 53] approach the issue of missing poses as a pose refinement problem, leveraging temporal smoothness to address it effectively. SmoothNet [53] introduced a temporal motion refinement network for refining the poses obtained from the image-based pose estimators to alleviate jitters. These methods perform well under sparse, infrequent occlusions across frames but struggle with natural, prolonged occlusions, as they lack training for such instances. All these works evaluate their methods on datasets with sparse occlusions, as no existing datasets contain significant occlusions. To address this gap, we introduce VOccl3D, a novel dataset designed specifically for 3D HPS estimation under significant occlusions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7352" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.262, + 0.106 + ], + "angle": 0, + "content": "3. VOCcl3D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.485, + 0.192 + ], + "angle": 0, + "content": "In this section, we outline the key components required to construct the VOcc13D dataset. Section 3.1 details the creation of individual assets, including background scenes, human motions, and texture maps. Sections 3.2 describe the rendering process and attributes of the dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.2, + 0.24, + 0.215 + ], + "angle": 0, + "content": "3.1. Dataset Assets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.403 + ], + "angle": 0, + "content": "Background Scenes. In recent years, neural networks have significantly advanced 3D scene representation, enabling novel-view image synthesis. Notable approaches include Neural Radiance Fields (NeRF) [32], which learn a joint representation of geometry and appearance for novel-view synthesis. Another method, 3D Gaussian Splatting (3DGS) [19], represents a scene using 3D Gaussians and adopts a differentiable rasterizer for real-time rendering. These methods use multiview images to learn 3D representations, avoiding extensive capture setups. To develop our dataset, we use the 3DGS method to learn the 3D representation of the background scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.434 + ], + "angle": 0, + "content": "Preliminaries: 3D Gaussian Splatting We represent the \\(i\\)-th Gaussian in 3DGS as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.442, + 0.483, + 0.463 + ], + "angle": 0, + "content": "\\[\nG (\\mathbf {p}) = o _ {i} e ^ {- \\frac {1}{2} \\left(\\mathbf {p} - \\boldsymbol {\\mu} _ {i}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {p} - \\boldsymbol {\\mu} _ {i}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.701 + ], + "angle": 0, + "content": "where \\(\\mathbf{p} \\in \\mathbb{R}^3\\) is a xyz location, \\(o_i \\in [0,1]\\) is the opacity modeling the ratio of radiance the Gaussian absorbs, \\(\\pmb{\\mu}_i \\in \\mathbb{R}^3\\) is the center/mean of the Gaussian, and the covariance matrix \\(\\pmb{\\Sigma}_i\\) is parameterized by the scale \\(\\mathbf{S}_i \\in \\mathbb{R}_+\\) along each of the three Gaussian axes and the rotation \\(\\mathbf{R}_i \\in SO(3)\\) with \\(\\pmb{\\Sigma}_i = \\mathbf{R}_i\\mathbf{S}_i\\mathbf{S}_i^\\top\\mathbf{R}_i^\\top\\). Each Gaussian is also paired with spherical harmonics [39] to model the radiance emitted towards various directions. During rendering, the 3D Gaussians are projected onto the image plane and form 2D Gaussians [57] with the covariance matrix \\(\\pmb{\\Sigma}_i^{2\\mathrm{D}} = J\\pmb{W}\\pmb{\\Sigma}_i\\pmb{W}^\\top\\pmb{J}^\\top\\), where \\(\\pmb{J}\\) is the Jacobian of the affine approximation of the projective transformation and \\(\\pmb{W}\\) is the viewing transformation. The color of a pixel is calculated via alpha blending the \\(N\\) Gaussians contributing to a given pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.712, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {j = 1} ^ {N} c _ {j} \\alpha_ {j} \\prod_ {k = 1} ^ {j - 1} \\left(1 - \\alpha_ {k}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.841 + ], + "angle": 0, + "content": "where the Gaussians are sorted from close to far, \\( c_{j} \\) is the color obtained by evaluating the spherical harmonics given viewing transform \\( W \\), and \\( \\alpha_{j} \\) is calculated from the 2D Gaussian formulation (with the covariance \\( \\Sigma_{j}^{2D} \\)) multiplied by its opacity \\( o_{j} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We collected RGB videos from the large-scale open-source DL3DV dataset [27], which contains over 10,000 videos across diverse domains, including natural and outdoor settings, educational institutions, shopping complexes," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "parks, hubs, cafes, and restaurants. We selected videos based on the presence of natural occlusions, such as garlands, chairs, benches, statues, cars, and bins. Our dataset captures 3D representations of approximately 40 distinct scenes, each incorporating real-world occlusions. This approach to learning 3D representation closely resembles real-world data compared to conventional methods that rely on rendering from 3D graphic asset scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.908, + 0.528 + ], + "angle": 0, + "content": "SMPLX body/Human Animations. We represent the human body using the SMPL-X [36] 3D human model, defined by the function \\(\\mathcal{M}(\\theta ,\\beta ,\\psi)\\), where \\(\\theta\\) represents the pose parameters, \\(\\beta\\) the shape parameters, and \\(\\psi\\) the facial expression parameters. This function outputs a body mesh \\(\\mathcal{M}\\in \\mathbb{R}^{10475\\times 3}\\) with 10,475 vertices. We sample approximately 400 SMPL-X 3D human motion models from the AMASS mocap dataset [30], which contains over 11,000 motion sequences with diverse body shapes and poses. Following previous work [48], we use the pre-trained human pose prior model VPoser [36] to assess pose difficulty. VPoser is a Variational Auto-Encoder model that evaluates a pose \\(\\theta\\) to be challenging if its embeddings \\(\\epsilon_{\\theta}\\) have larger norm, i.e. \\(||\\epsilon_{\\theta}||_{2} > \\tau\\), where \\(\\tau\\) is empirically set to 40. We classify a challenging pose using VPoser to avoid simple movements such as walking, standing, jogging, etc. To prepare human animations for rendering, we use Blender graphics software to efficiently bake complex geometric animations over time into the Alembic format for subsequent use in rendering engines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.909, + 0.738 + ], + "angle": 0, + "content": "We ensured that motion sequences contained at least 180 frames to maintain a minimum animation duration of six seconds at 30 fps. For sequences exceeding 400 frames, we discarded the first 100 frames to remove static poses at the start. To ensure that human motion remains consistently within the occluded scene, we implemented boundary constraints that stop the motion if it moves beyond these limits. These conditions prevent humans from moving too far from the occlusion area. Additionally, we applied random rotations and translations to each sequence. We store the applied transformations to further calculate the effective global orientation of the human body with respect to the world coordinates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Human Textures. We sourced human body texture scans from the open-source dataset provided by the SMPLtex method [4], which estimates and manipulates the full 3D appearance of humans captured from a single image. The SMPLtex dataset provides a diverse range of human texture scans, covering various skin tones, clothing styles, and genders. To ensure sample diversity in VOcc13D, we select approximately 200 distinct texture scans. Figure 4 illustrates the different clothing textures applied to the SMPL-X human model in our dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7353" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.912, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.323, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 3. An overview of our proposed dataset creation pipeline. We generate 3D representations of natural scenes with real occlusions using 3D Gaussian Splatting [19]. Human motion sequences from the AMASS MoCap dataset [30] are processed in Blender to generate baked animation files. The generated scene representations, animation files, human texture maps, and camera extrinsic parameters are imported into the Unity rendering engine to generate video sequences of human motion under occlusion." + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.395, + 0.445, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Figure 4. Diversity of human texture in VOccl3D. We use approximately 200 human texture maps from the SMPLtex dataset [4], encompassing diverse clothing styles, genders, skin tones, and ethnicities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.388, + 0.647 + ], + "angle": 0, + "content": "3.2. Dataset Rendering and Attributes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Rendering As shown in Figure 3, we used Unity Engine to render synthetic scenes by integrating background Gaussian splats, human animations, and body texture assets. We rendered video sequences at 30 fps with a resolution of \\(720 \\times 720\\), varying the camera field of view between 25 and 50 across different scenes. For each sequence, we stored ground-truth camera extrinsic and intrinsic parameters and saved the rendered RGB images in lossless PNG format. We positioned the camera to capture occluded views within the scene and applied motion constraints to human animation assets to maintain consistent occlusion throughout each sequence. To achieve realistic lighting, we employed Unity's built-in DirectionalLight asset to simulate natural illumination on the human mesh." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Dataset Attributes. Our proposed dataset, VOccl3D, includes over 250,000 images and 400 video sequences with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.546 + ], + "angle": 0, + "content": "a total runtime exceeding 2 hours and 30 minutes. The dataset features 40 background scenes with various occlusions, such as cars, garlands, benches, chairs, bins, and trees. Each scene contains 10 video sequences with variations in human motions and clothing textures. The rendered sequence exhibit diversity in body shapes, skin tones, and camera poses. We provide ground-truth annotations, including camera extrinsic and intrinsic parameters, pose, shape, global orientation, translation, gender, 2D keypoints, and a binary occlusion label for each keypoint." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.652 + ], + "angle": 0, + "content": "Different Modalities. In addition to 3D pose and shape annotations, VOccl3D provides annotations for multiple modalities, including human silhouettes, body-part segmentation, 2D keypoint estimation, and human bounding box detection under occlusion. Researchers can use our dataset to train and evaluate methods designed to handle occlusion across these modalities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.749, + 0.675 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.907, + 0.82 + ], + "angle": 0, + "content": "In this section, we highlight the need and significance of our proposed VOccl3D dataset. We fine-tune the state-of-the-art pose estimation methods using our dataset and report the qualitative and quantitative results. Our results demonstrate improved performance on the HPS estimation task (Section 4.1) using both our proposed synthetic dataset and real-world datasets with occlusions. In Section 4.2, we show the enhanced performance of the human object detector and its impact on HPS estimation under occlusion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.822, + 0.849 + ], + "angle": 0, + "content": "4.1. Human Pose and Shape Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Dataset and Implementation Setup. We use approximately 200k and 50k images for training and testing respectively in our proposed VOcc13D dataset. We fine-tune" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7354" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.908, + 0.239 + ], + "angle": 0, + "content": "
MethodHard-OcclusionMedium-OcclusionLow-Occlusion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]192.22114.35247.41121.7078.56158.4698.8267.64126.92
BEDLAM-HMR [3]167.35102.92214.39102.5568.13134.5986.5554.48110.15
BEDLAM-CLIFF [3]154.8699.53199.9590.9765.03119.6374.9552.6596.60
HMR2.0 [11]169.71100.49215.17113.8871.78145.6288.5359.08114.39
STRIDE with BEDLAM-CLIFF [24]155.64100.44-91.1465.38-75.0253.21-
WHAM [44]152.15102.14177.07110.9776.81127.4593.9066.68106.51
VOcc13D-B-CLIFF136.3489.94175.9282.4858.78106.8469.4646.3288.19
STRIDE with VOcc13D-B-CLIFF136.4390.28-82.3758.98-69.6546.86-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.249, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Table 1. 3D HPS estimation results on the test-split of VOcccl3D. The results show that VOcccl3D-B-CLIFF and STRIDE, when using pseudo-labels from VOcccl3D-B-CLIFF, significantly outperform other image- and video-based HPS estimation methods across hard, medium, and low occlusion categories. As expected, all HPS estimation methods exhibit a decline in performance as occlusion severity increases from low to medium to high. For evaluation, we use ground-truth bounding boxes. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.318, + 0.868, + 0.445 + ], + "angle": 0, + "content": "
Method3DPWOcclType1-3DPWOcclType2-3DPW
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]73.946.487.698.1562.27118.4799.4962.16119.82
BEDLAM-HMR [3]79.047.693.1108.6266.53128.05106.1964.05125.66
BEDLAM-CLIFF [3]72.046.685.098.7164.26117.4196.8061.32115.33
HMR2.0 [11]81.254.3143.7103.4069.66164.5599.0166.17158.79
VOccl3D-B-CLIFF72.047.384.595.8963.43114.2894.3660.44112.01
VOccl3D-CLIFF71.1045.9884.2595.1761.83114.9593.7459.66112.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.454, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Table 2. 3D HPS estimation results on 3DPW, OcclType1-3DPW, and OcclType2-3DPW. Since 3DPW is a real-world dataset with minimal occlusions, VOccl3D-CLIFF outperforms all other methods but shows only a marginal improvement over BEDLAM-CLIFF. However, on OcclType1-3DPW and OcclType2-3DPW, which contain significant occlusions on real-world dataset, our method demonstrates a notable improvement over existing approaches. The best results are noted in bold." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.534, + 0.482, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Figure 5. Fine-tuning an off-the-shelf pose estimator. We fine-tune a CLIFF [25] pose estimation model using our proposed VOcc13D dataset. We observe a significant improvement in the estimated mesh when using a fine-tuned CLIFF model using our dataset. Note: The pose estimator can be any off-the-shelf pose estimator, here we show results with the CLIFF model [25]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "two versions of state-of-the-art model, CLIFF (trained on Human3.6M [15], MPI-INF-3DHP [31], and COCO [26] dataset) and BEDLAM-CLIFF (trained on BEDLAM and AGORA datasets) using our rendered VOcc13D dataset, resulting in VOcc13D-CLIFF and VOcc13D-B-CLIFF, re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.764 + ], + "angle": 0, + "content": "spectively (refer to Figure 5). To benchmark performance, we evaluate on the test split of VOccl3D, comparing against multiple image and video based baselines, including CLIFF [25], BEDLAM-HMR [3], BEDLAM-CLIFF [3], HMR2.0 [11], STRIDE [24], and WHAM [44]. We report the mean performance by evaluating the model using a fivefold cross-validation on our dataset. To quantify occlusion levels, we annotate each 3D joint with a binary occlusion label, marking it as occluded if its corresponding 2D keypoint lies within the ground-truth segmentation mask of the visible human region. Based on the number of visible keypoints (out of 22 total), we categorize images into three occlusion levels: hard occlusion (4-9 visible keypoints), medium occlusion (10-15 visible keypoints), and low occlusion (16-20 visible keypoints)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Beyond evaluations on synthetic datasets, we assess the performance of VOccl3D fine-tuned models on real-world datasets, including 3DPW [46] and OCMotion [13]. It is noteworthy that these datasets do not contain significant occlusions. Hence, we introduce two occlusion-augmented variants of 3DPW to evaluate robustness under severe occlusions. In OcclType1-3DPW, we overlay a black patch on a randomly selected keypoint, while in OcclType2-3DPW, we occlude two random keypoints. Further details on oc" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7355" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.498, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of HPS estimation methods on VOcc3D dataset. The first and second column shows RGB image and ground-truth human mesh. Column third and fourth compare HPS estimation using the BEDLAM-CLIFF [3] and HMR2.0 [11] methods. The final column (VOcc3D-B-CLIFF) presents results obtained by fine-tuning the CLIFF model on the VOcc3D dataset. These results demonstrate superior performance, particularly in scenarios with heavy occlusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.601 + ], + "angle": 0, + "content": "clusion variants of 3DPW and implementation details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results. We present benchmarking results on the test split of our proposed VOcc13D dataset in Table 1. Our finetuned VOcc13D-B-CLIFF significantly outperforms existing state-of-the-art image and video-based methods across all occlusion categories. Notably, the performance of all previous methods degrades as occlusion severity increases, highlighting the inherent challenges of HPS estimation under occlusions. We also compare our approach with the plug-and-play method STRIDE [24], which performs 3D pose estimation and reports only MPJPE and PA-MPJPE errors. Our results show that STRIDE achieves superior performance when leveraging pseudo-labels from the VOcc13D-B-CLIFF model compared to those from the original BEDLAM-CLIFF model, further validating the effectiveness of our dataset. In Table 2 and Table 3, we report performance on the 3DPW [45] and OCMotion [14] datasets, respectively. We observe that both VOcc13D-B-CLIFF and VOcc13D-CLIFF outperform all the previous methods, demonstrating their effectiveness in real-world" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.844 + ], + "angle": 0, + "content": "scenarios. However, since these datasets contain minimal occlusions, the improvements over CLIFF remain marginal. To further assess robustness under heavy occlusion, we evaluate on OcclType1-3DPW and OcclType2-3DPW, real-world datasets with significant occlusions. Table 2 demonstrates that our fine-tuned model achieves substantial improvements, underscoring its robustness in occluded environments. Table 3 presents results on the OCMotion dataset, which features lower levels of occlusion. We can observe that VOcc13D-CLIFF outperforms all comparison methods and demonstrates performance comparable to CLIFF [25]. This is expected, as both models are pretrained on large-scale real-world datasets captured in controlled lab environments similar to OCMotion. Figure 6 presents the qualitative results on VOcc13Ddataset, showcasing the improvement of our fine-tuned model against other state-of-the-art HPS estimation methods. Additional results are provided in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.907, + 0.866 + ], + "angle": 0, + "content": "4.2. Impact of Human Detector on HPS Estimation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Dataset and Implementation Setup. Human object detectors play a crucial role in HPS estimation tasks, particularly" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7356" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.444, + 0.206 + ], + "angle": 0, + "content": "
MethodMPJPEOCMotion PA-MPJPEPVE
CLIFF [25]64.1540.1679.80
BEDLAM-HMR [3]73.9641.9492.70
BEDLAM-CLIFF [3]66.8041.7983.86
HMR2.0 [11]69.9443.0087.07
VOcc13D-B-CLIFF65.9640.5981.96
VOcc13D-CLIFF64.2939.6478.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.217, + 0.482, + 0.3 + ], + "angle": 0, + "content": "Table 3. Qualitative results of HPS estimation on OCMotion. Our results show that VOcc13D-CLIFF outperforms all HPS estimation methods except CLIFF [25]. Both CLIFF and VOcc13D-CLIFF exhibit comparable performance, as both are pre-trained on real-world datasets captured in controlled lab environments, similar to OCMotion [14]. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.317, + 0.455, + 0.378 + ], + "angle": 0, + "content": "
Method3DPWOCMotion
mAP50mAP75mAP50mAP75
YOLO1158.9947.1498.8491.80
VOccl3D-YOLO1159.8948.2699.1091.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Table 4. Quantitative results of YOLO11 on 3DPW and OC-Motion datasets. The first row presents the human object detection performance of the pre-trained YOLO11 model, while the second row shows the results after fine-tuning with the VOccl3D dataset. The best results are in bold. We observe an improvement in detection accuracy across both datasets after fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.611 + ], + "angle": 0, + "content": "when inferring from in-the-wild RGB images [3, 24, 25]. Ideally, HPS estimation performs optimally with ground-truth bounding boxes; however, most human object detectors struggle significantly under occlusion. To address this, we fine-tune the recent YOLO11 detector [16] on the combined training split of VOccl3D and MS COCO, referring to the fine-tuned model as VOccl3D-YOLO11. We show evaluations on 3DPW [45] and OCMotion [14] datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Results. Table 4 compares the detection performance of the pre-trained YOLO11 and VOccl3D-YOLO11 models on the real-world 3DPW [45] and OCMotion [14] datasets. Since OCMotion is a single-human dataset with minimal occlusions, the detector achieves high mAP scores. This is likely because most body parts remain visible, making box detection easier. However, due to the presence of multiple individuals in 3DPW, the overall mAP scores remain lower. In short, VOccl3D-YOLO11 shows significant improvement on datasets with high occlusions. We provide qualitative results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 5 further evaluates the impact of human detection on HPS estimation. The first, second, and third rows represent the performance of VOccl3D-CLIFF when human detections are sourced from ground truth, YOLO11, and VOccl3D-YOLO11, respectively. As expected, the best performance is achieved when using ground-truth detections. However, fine-tuning the detector with VOccl3D significantly improves HPS estimation compared to detections" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.178 + ], + "angle": 0, + "content": "
Method3DPWOCMotion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
VOccl3D-CLIFF w/GT71.1045.9884.2564.2939.6478.56
VOccl3D-CLIFF w/YYOLO11116.5263.35139.7467.1641.3083.00
VOccl3D-CLIFF w/VOccl3D-YYOLO11114.8562.66137.1966.6541.1582.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.187, + 0.905, + 0.326 + ], + "angle": 0, + "content": "Table 5. 3D HPS estimation results using the VOcc13D-CLIFF model with different bounding box sources. This table presents HPS estimation performance on the 3DPW and OC-Motion datasets. The first row reports results using ground-truth bounding boxes, achieving the highest accuracy. The second and third rows show performance when detections are obtained from the pre-trained YOLO11 and fine-tuned VOcc13D-YOLO11 models, respectively. Notably, the third row demonstrates a significant improvement over the second, highlighting the impact of detections on HPS estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.905, + 0.39 + ], + "angle": 0, + "content": "from the pre-trained YOLO11. These results underscore the critical role of human detection quality in enhancing HPS performance under occlusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.906, + 0.543 + ], + "angle": 0, + "content": "Our experiments show that fine-tuning an existing HPS estimation model with a large synthetic dataset containing occlusions enhances its performance on both real and synthetic datasets. We observe significant improvements over baseline methods, particularly in scenarios with heavy occlusions. Additionally we demonstrate that the poor results are not solely due to HPS estimations errors but also stem from failures in bounding box detection. The VOcc13D dataset proves effective in refining pose estimation methods and bounding box predictions under occlusions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.634, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We introduce VOccl3D, a novel large-scale video dataset with synthetic humans in real-world scenes under occlusions for 3D human pose and shape estimation. By leveraging a rendering-based approach, VOccl3D eliminates the need for costly and labor-intensive data collection while providing a diverse and realistic dataset for occlusion-aware research. Through extensive qualitative and quantitative evaluations, we demonstrate that fine-tuning existing HPS estimation methods with VOccl3D significantly enhances their performance on real-world datasets with occlusions and on the test split of VOccl3D dataset. Furthermore, we improve the human object detector, YOLO11 under occluded conditions using VOccl3D, highlighting the impact of detections in achieving robust HPS estimation in occluded scenarios. Beyond HPS estimation, VOccl3D serves as a comprehensive benchmark for evaluating methods across multiple occlusion-aware tasks, including human body-part segmentation, 2D/3D pose estimation, and human bounding box detection. Hence, VOccl3D sets a new standard for occluded human benchmarks, offering a valuable dataset for advancing occlusion-robust research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7357" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.298 + ], + "angle": 0, + "content": "This work was partially supported by USDA NRI grant 2021-67022-33453, NSF grants CMMI-2326309 and CNS-2312395. This research was also partially supported by the Office of the Director of National Intelligence (ODNI), specifically through the Intelligence Advanced Research Projects Activity (IARPA), under contract number [2022-21102100007]. The views and conclusions in this research reflect those of the authors and should not be construed as officially representing the policies, whether explicitly or implicitly, of ODNI, IARPA, or the U.S. Government. Nevertheless, the U.S. Government retains the authorization to reproduce and distribute reprints for official government purposes, regardless of any copyright notices included." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.323, + 0.188, + 0.339 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.348, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.406, + 0.484, + 0.462 + ], + "angle": 0, + "content": "[2] Peter Bauer, Arij Bouazizi, Ulrich Kressel, and Fabian B. Flohr. Weakly supervised multi-modal 3d human body pose estimation for autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.463, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[3] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[4] Dan Casas and Marc Comino-Trinidad. SMPLitex: A Generative Model and Dataset for 3D Human Texture Estimation from Single Image. In British Machine Vision Conference (BMVC), 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.617 + ], + "angle": 0, + "content": "[5] Yu Cheng, Bo Yang, Bo Wang, Yan Wending, and Robby Tan. Occlusion-aware networks for 3d human pose estimation in video. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 723-732, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[6] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 591-601, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[8] R James Cotton. Posepipe: Open-source human pose estimation pipeline for rehabilitation research. Archives of Physical Medicine and Rehabilitation, 103(12):e161-e162, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[9] Yufeng Cui and Yimei Kang. Multi-modal gait recognition via effective spatial-temporal feature fusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17957, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Yang Fu, Shibei Meng, Saihui Hou, Xuecai Hu, and Yongzhen Huang. Gpgait: Generalized pose-based gait" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.348, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "recognition. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19538-19547, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[12] Charlie Hewitt, Fatemeh Saleh, Sadegh Aliakbarian, Lohit Petikam, Shideh RezaEIFar, Louis Florentin, Zafiirah Hosenie, Thomas J Cashman, Julien Valentin, Darren Cosker, et al. Look ma, no markers: Holistic performance capture without the hassle. ACM Transactions on Graphics, 43(6): 1-12, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[13] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[14] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[16] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[17] Sardor Juraev, Akash Ghimire, Jumabek Alikhanov, Vijay Kakani, and Hakil Kim. Exploring human pose estimation and the usage of synthetic data for elderly fall detection in real-world surveillance. IEEE Access, 10:94249-94261, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[18] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[20] Do-Yeop Kim and Ju-Yong Chang. Attention-based 3d human pose sequence refinement network. Sensors, 21(13), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[22] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Farnoosh Koleini, Muhammad Usama Saleem, Pu Wang, Hongfei Xue, Ahmed Helmy, and Abbey Fenwick. Biopose: Biomechanically-accurate 3d pose estimation from monocular videos. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6330–6339. IEEE, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.348, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7358" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Rohit Lal, Saketh Bachu, Yash Garg, Arindam Dutta, Calvin-Khang Ta, Dripta S. Raychaudhuri, Hannah Dela Cruz, M. Salman Asif, and Amit K. Roy-Chowdhury. Stride: Single-video based temporally continuous occlusion robust 3d pose estimation, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[25] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[27] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[31] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 International Conference on 3D Vision (3DV), pages 506-516, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[33] Sara Moccia, Lucia Migliorelli, Virgilio Carnielli, and Emanuele Frontoni. Preterm infants' pose estimation with spatio-temporal features. IEEE Transactions on Biomedical Engineering, 67(8):2370-2380, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[34] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J. Mitra. imapper: interaction-guided scene mapping from monocular videos. ACM Trans. Graph., 38(4), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13468-13478, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[37] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7745-7754, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.906, + 0.26 + ], + "angle": 0, + "content": "[38] Miroslav Purkrábek and Jiri Matas. Improving 2d human pose estimation across unseen camera views with synthetic data. arXiv preprint arXiv:2307.06737, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.907, + 0.315 + ], + "angle": 0, + "content": "[39] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.906, + 0.372 + ], + "angle": 0, + "content": "[40] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.906, + 0.439 + ], + "angle": 0, + "content": "[41] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, page 1530-1538. JMLR.org, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 8(11):7090-7097, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.564 + ], + "angle": 0, + "content": "[43] Saurabh Sharma, Pavan Teja Varigonda, Prashast Bindal, Abhishek Sharma, and Arjun Jain. Monocular 3d human pose estimation by generation and ordinal ranking. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[44] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2070-2080, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[46] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European conference on computer vision (ECCV), pages 601-617, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[47] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[48] Zhenzhen Weng, Laura Bravo-Sanchez, and Serena Yeung-Levy. Diffusion-hpc: Synthetic data generation for human mesh recovery in challenging domains. In 2024 International Conference on 3D Vision (3DV), pages 257-267. IEEE, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7359" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[49] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, et al. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20282-20292, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[50] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9060-9069, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[51] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[52] Andrei Zanfir, Mihai Zanfir, Alex Gorban, Jingwei Ji, Yin Zhou, Dragomir Anguelov, and Cristian Sminchisescu. HUM3DIL: Semi-supervised multi-modal 3d humanpose estimation for autonomous driving. In 6th Annual Conference on Robot Learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[53] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In European Conference on Computer Vision. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[54] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conference on Computer Vision and Pattern Recognition, (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[55] Yi Zhang, Pengliang Ji, Adam Kortylewski, Angtian Wang, Jieru Mei, and Alan L Yuille. 3D-Aware Neural Body Fitting for Occlusion Robust 3D Human Pose Estimation. In The IEEE/CVF International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[56] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G. Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4966-4975, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7360" + } + ] +] \ No newline at end of file diff --git a/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_origin.pdf b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9bb3cea77b2105efc2f80288b91dc59fd54d5329 --- /dev/null +++ b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/bee260a8-0889-4cd7-b8a5-2c0fb163c69d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f69c822e105806e7344ad6e85d876a4f79b8a5b4393cc8fc645ec288f9b2333c +size 29061397 diff --git a/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/full.md b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..667b858bfb2e90d893769234be68c01c316c525b --- /dev/null +++ b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/full.md @@ -0,0 +1,257 @@ +# VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions + +Yash Garg, Saketh Bachu, Arindam Dutta, Rohit Lal†, Sarosij Bose, Calvin-Khang Ta‡, M. Salman Asif, Amit Roy-Chowdhury University of California, Riverside + +{ygarg002,sbach008,adultt020,r1al011,sbose007,cta003,sasif,amitrc}@ucr.edu + +![](images/2d1f4f7d05d18f1178989a04559ef760ba9159d82ed48257677342924a0978d2.jpg) +Figure 1. We propose VOccl3D, a large-scale synthetic video dataset specifically designed for training and evaluating algorithms for 3D human pose and shape estimation (HPS) in realistic occlusion scenarios. VOccl3D comprises over 250,000 frames, with a total runtime exceeding 2 hours and 30 minutes. Compared to previous occlusion-based datasets, VOccl3D has diverse body shapes, textures, and most importantly, significant and realistic occlusions within the scenes. In addition to body shape and pose, VOccl3D provides other necessary ground truth information, such as bounding boxes, body part segmentations, and human silhouettes. The top row illustrates various frames from our dataset, showcasing diverse occlusions, clothing textures, and motions, while the bottom row represents a sequence of frames from a video sequence within our dataset. + +# Abstract + +Human pose and shape (HPS) estimation methods have been extensively studied, with many demonstrating high zero-shot performance on in-the-wild images and videos. However, these methods often struggle in challenging scenarios involving complex human poses or significant occlusions. Although some studies address 3D human pose estimation under occlusion, they typically evaluate performance on datasets that lack realistic or substantial occlusions, e.g., most existing datasets introduce occlusions with random patches over the human or clipart-style overlays, which may not reflect real-world challenges. To bridge this gap in realistic occlusion datasets, we introduce a novel benchmark dataset, VOccl3D, a Video-based human Occlusion dataset with 3D body pose and shape annotations. Inspired by works such as AGORA and BEDLAM, we constructed this dataset using advanced computer graphics rendering techniques, incorporating diverse real-world occlusion scenarios, clothing textures, and human motions. + +Additionally, we fine-tuned recent HPS methods, CLIFF and BEDLAM-CLIFF, on our dataset, demonstrating significant qualitative and quantitative improvements across multiple public datasets, as well as on the test split of our dataset, while comparing its performance with other state-of-the-art methods. Furthermore, we leveraged our dataset to enhance human detection performance under occlusion by fine-tuning an existing object detector, YOLO11, thus leading to a robust end-to-end HPS estimation system under occlusions. Overall, this dataset serves as a valuable resource for future research aimed at benchmarking methods designed to handle occlusions, offering a more realistic alternative to existing occlusion datasets. See the Project page for code and dataset:https://yashgarg98.github.io/Voccl3D-dataset/ + +# 1. Introduction + +Monocular 3D human pose and shape (HPS) estimation is a complex yet essential task in computer vision. It has applications in surveillance [7, 17], autonomous robotics [2, 42, 52], human motion analysis [9, 10], and clinical assessment [8, 23, 33]. Since the introduction of the neural network-based Human Mesh Recovery Network (HMR) [18], numerous approaches have advanced the field by enabling accurate estimation of SMPL[28] pose and shape from a single RGB image. Recent methods [1, 21, 24, 25, 29, 44] demonstrate impressive zero-shot HPS performance on in-the-wild RGB images and video sequences. However, these methods still face limitations in achieving high performance in challenging scenarios, such as complex human poses and under significant occlusions. + +Achieving robust HPS under occlusion remains a challenging problem due to the contextual ambiguity of the occluded body parts. Several methods [22, 40, 51, 55] have been proposed to address the challenges caused by occlusions. Methods like PARE [22] focus on using visible body parts to infer occluded areas, enhancing estimation accuracy with partial views. Temporal models like HuMoR [40] and GLAMR [51] use generative frameworks to ensure pose continuity over time. HuMoR predicts pose distributions, while GLAMR incorporates global trajectory data to fill in missing poses. Recent methods, like STRIDE [24], leverage a large-scale pre-trained motion prior model to achieve temporally coherent pose reconstruction under occlusions. While these methods effectively leverage visible information and temporal continuity, they still struggle under severe occlusions due to limited exposure to occluded scenarios in their training data. + +Evaluations of the existing methods are often limited to occluded datasets that lack scene diversity [34], involve moderate or lower levels of occlusion [14, 45], or use patch-based occlusion datasets for training and inference [55], as illustrated in Figure 2. Most importantly, there remains an urgent need for a realistic, diverse, and significantly occluded human pose and shape dataset to advance the task of 3D HPS estimation. To address this gap, our work proposes VOccl3D, a large-scale video dataset with synthetic humans in real scenes for 3D human pose and shape estimation. Following prior works such as BEDLAM [3] and Synthocap [12], we use an advanced computer graphics engine to render a highly realistic synthetic dataset. Prior research has shown that fine-tuning models on synthetic datasets can significantly improve HPS estimation performance on real-world datasets [3]. Synthetic datasets provide "perfect" ground-truth annotations and eliminate the need for costly sensors, making it computationally efficient to generate large-scale datasets. + +Our proposed dataset includes approximately 250,000 images and 400 video sequences. To introduce diversity, we + +![](images/cbe23ef8950056961997cd564aa47c6151b13a5f81dd23cfc7b4487a4336beb7.jpg) +a) 3DPW-AdvOcc@80 + +![](images/e289547aa747317d3b88d78846c6025832db910e10f785e4fdbb13c9c414f968.jpg) + +![](images/97c2cc452f7e7a57730b9269869dd4327882f8e9dc53ba8e04b414ec5b686fbc.jpg) + +![](images/fefe371bdf90e683d3c6e293f6b92d15f692b343f8d9f9f4d6eac1dfde7b981c.jpg) +d) OCMotion + +![](images/4490af5634a6def6a3f68b0a37e61ada84deb7ab97a2eb28ae998f61e078c83b.jpg) +b) Occluded Human3.6M +e) 3DPW + +![](images/4c1af9124c611fdfc1ae15ef8e39dd1a4407707f01324b34a104e7d87ce02ee1.jpg) +c) Syn. Occ. Human3.6M +f) Ours +Figure 2. Samples from various occlusion-based HPS datasets. The top row displays datasets with artificial, patch-based occlusions applied to existing datasets, including 3DPW-AdvOcc [45, 55], Occluded Human3.6M [15, 24], and Synthetic Occlusion Human3.6M [15, 54]. The bottom row presents samples from two datasets with natural occlusions: OCMotion [14] and 3DPW [45]. Notably, the top-row images exhibit unrealistic occlusions, which lack the realism of naturally occurring occlusions. The bottom row images contain natural occlusions but are sparse and infrequent. Compared to existing datasets, our proposed VOcccl3D dataset features more realistic occlusions. + +incorporate human motion samples from the AMASS [30] dataset, over 200 distinct clothing textures for sequences containing both male and female genders, and 40 real background scenes with occlusions. Additionally, we provide occlusion labels for each body joint. Unlike prior methods [3, 35] which render synthetic environments, we utilize 3D Gaussian splatting (3DGS) [19] to create 3D representations of background scenes. This approach enhances the realism of rendered scenes, making them closely resemble real-world captured data. Reconstructing background scenes using 3DGS offers the flexibility to create domain-specific datasets, such as for agriculture, street or indoor environments using raw RGB video captures without relying on costly or limited graphic assets. Figure 1 showcases samples from our dataset, illustrating the diversity in clothing, environments, human motions, and occlusions. + +We fine-tuned the CLIFF [25] and BEDLAM-CLIFF [3] models on our dataset to demonstrate that training with synthetic data enhances the performance of existing HPS estimation methods under occlusion. We evaluated state-of-the-art HPS estimation methods on the test split of our dataset across three occlusion levels: high, medium, and low. The results show that existing methods perform poorly under high occlusion, whereas our fine-tuned models achieve significant improvements. To assess performance on real-world data, we perform evaluations on the 3DPW and OCMotion datasets, which contain low-level occlusions. Additionally, we created a variant of the 3DPW dataset by adding black patches to evaluate the robustness + +against high-occlusion real scenarios. Our fine-tuned models outperformed recent HPS methods on these real-world datasets. Furthermore, our experiments highlight the importance of an effective human object detector for improving HPS performance under occlusion. To address this, we fine-tuned the YOLO11[16] object detector using our dataset, enhancing its performance under occluded scenarios. VOcc13D is available for researchers to benchmark and evaluate occlusion-aware methods. + +In summary, we make the following key contributions: + +1. We propose VOcc13D, a novel large-scale, realistic video-based dataset of occluded synthetic humans in real scenes for 3D human pose and shape estimation. +2. We demonstrate both qualitatively and quantitatively that existing HPS estimation methods fine-tuned with VOccl3D outperform other methods on real-world datasets with occlusions. +3. We improved the performance of the human object detector under occlusion scenarios by leveraging the VOccl3D dataset which is a crucial component for a robust HPS estimation. +4. VOccl3D can serve as a benchmark dataset for evaluating methods specifically designed to perform under occlusion across various tasks, including human and body-part segmentation, 2D/3D pose estimation, and human bounding box detection. + +# 2. Related Works + +# Synthetic Data Generation for Human Pose Estimation. + +Human pose estimation is crucial in computer vision, but state-of-the-art methods depend on costly, labor-intensive labeled datasets. Several notable synthetic datasets have advanced research in human pose estimation. SynBody [49] provides a large-scale synthetic dataset for human mesh recovery and view synthesis, while RePoGen [38] enables fine-grained control over pose and viewpoint to generate rare, complex poses. Human3.6m [15] provides additional small mixed reality dataset by inserting 3D animation models with background scenes. BEDLAM [3] further highlights that models trained solely on synthetic data can outperform real-data-trained counterparts, emphasizing the importance of high-quality synthetic datasets for transferable models. Further, PressurePose [6] simulates interactions between articulated and soft bodies, capturing fine-grained contact dynamics. SynthMocap [12] extends this by providing expressive synthetic data with detailed body, hand, and face movements. Despite their success, these synthetic datasets lack emphasis on significant occlusions, a key real-world challenge. Addressing this limitation, our work proposes a synthetic dataset tailored to 3D pose estimation under heavy and realistic occlusion, aiming to bridge this gap and enhance robustness in occlusion-prone environments. + +Image and Video based Human Pose Estimation. Esti + +mating 2D/3D pose and shape from single RGB image has widely been explored in previous research [3, 25, 43, 47, 50]. Methods such as [43] use a conditional variational autoencoder for 2D-to-3D lifting in pose estimation. The pose estimation approach proposed in [47] applies normalizing flows [41] for 2D-3D mapping. Large-scale datasets like BEDLAM [3] have improved pose estimation models like CLIFF [25] and HMRNet [18]. However, these models struggle with generalizing to unseen scenarios with severe occlusions due to limited training on such cases. + +In addition, video-based human pose estimation has significantly improved challenging datasets. Early works like [56] use the EM algorithm to estimate the 3D pose from monocular video through 2D joint uncertainty maps. The method in [37] employs dilated temporal convolutions and semi-supervised learning for 3D pose estimation, while [1] uses the SMPL model [28] to extract pose and shape parameters, refining models like HMR with bundle adjustment. VIBE [21] applies adversarial learning with AMASS [30] for 3D pose extraction, while MEVA [29] improves accuracy and smoothness using a variational autoencoder to address VIBE's high acceleration error. Owing to these drawbacks, HuMoR employs a conditional variational autoencoder for robust pose estimation, while the state-of-the-art method WHAM [44] integrates 2D-3D lifting and SLAM for accurate global motion estimation. While these methods perform well on various datasets, studies like [24] highlight performance drops under significant occlusions, as most datasets contain only sparse occlusions, making models struggle with unseen heavy or prolonged occlusions. + +Human Pose Estimation under Occlusion. Despite significant progress in HPS estimation, handling occlusions remains a major challenge. This is because, in 3D pose estimation, missing depth cues make reconstruction far more ambiguous than in 2D. Early works like [5] used data augmentation with occlusion labels to enhance robustness in pose estimation. Latest methods like GLAMR [51] use a deep generative motion infiller to handle missing poses under severe occlusions and a global optimization framework to refine motion trajectories. Additionally, some methods [20, 29, 53] approach the issue of missing poses as a pose refinement problem, leveraging temporal smoothness to address it effectively. SmoothNet [53] introduced a temporal motion refinement network for refining the poses obtained from the image-based pose estimators to alleviate jitters. These methods perform well under sparse, infrequent occlusions across frames but struggle with natural, prolonged occlusions, as they lack training for such instances. All these works evaluate their methods on datasets with sparse occlusions, as no existing datasets contain significant occlusions. To address this gap, we introduce VOccl3D, a novel dataset designed specifically for 3D HPS estimation under significant occlusions. + +# 3. VOCcl3D Dataset + +In this section, we outline the key components required to construct the VOcc13D dataset. Section 3.1 details the creation of individual assets, including background scenes, human motions, and texture maps. Sections 3.2 describe the rendering process and attributes of the dataset. + +# 3.1. Dataset Assets + +Background Scenes. In recent years, neural networks have significantly advanced 3D scene representation, enabling novel-view image synthesis. Notable approaches include Neural Radiance Fields (NeRF) [32], which learn a joint representation of geometry and appearance for novel-view synthesis. Another method, 3D Gaussian Splatting (3DGS) [19], represents a scene using 3D Gaussians and adopts a differentiable rasterizer for real-time rendering. These methods use multiview images to learn 3D representations, avoiding extensive capture setups. To develop our dataset, we use the 3DGS method to learn the 3D representation of the background scene. + +Preliminaries: 3D Gaussian Splatting We represent the $i$ -th Gaussian in 3DGS as: + +$$ +G (\mathbf {p}) = o _ {i} e ^ {- \frac {1}{2} \left(\mathbf {p} - \boldsymbol {\mu} _ {i}\right) ^ {T} \boldsymbol {\Sigma} _ {i} ^ {- 1} \left(\mathbf {p} - \boldsymbol {\mu} _ {i}\right)}, \tag {1} +$$ + +where $\mathbf{p} \in \mathbb{R}^3$ is a xyz location, $o_i \in [0,1]$ is the opacity modeling the ratio of radiance the Gaussian absorbs, $\pmb{\mu}_i \in \mathbb{R}^3$ is the center/mean of the Gaussian, and the covariance matrix $\pmb{\Sigma}_i$ is parameterized by the scale $\mathbf{S}_i \in \mathbb{R}_+$ along each of the three Gaussian axes and the rotation $\mathbf{R}_i \in SO(3)$ with $\pmb{\Sigma}_i = \mathbf{R}_i\mathbf{S}_i\mathbf{S}_i^\top\mathbf{R}_i^\top$ . Each Gaussian is also paired with spherical harmonics [39] to model the radiance emitted towards various directions. During rendering, the 3D Gaussians are projected onto the image plane and form 2D Gaussians [57] with the covariance matrix $\pmb{\Sigma}_i^{2\mathrm{D}} = J\pmb{W}\pmb{\Sigma}_i\pmb{W}^\top\pmb{J}^\top$ , where $\pmb{J}$ is the Jacobian of the affine approximation of the projective transformation and $\pmb{W}$ is the viewing transformation. The color of a pixel is calculated via alpha blending the $N$ Gaussians contributing to a given pixel: + +$$ +C = \sum_ {j = 1} ^ {N} c _ {j} \alpha_ {j} \prod_ {k = 1} ^ {j - 1} \left(1 - \alpha_ {k}\right), \tag {2} +$$ + +where the Gaussians are sorted from close to far, $c_{j}$ is the color obtained by evaluating the spherical harmonics given viewing transform $W$ , and $\alpha_{j}$ is calculated from the 2D Gaussian formulation (with the covariance $\Sigma_{j}^{2D}$ ) multiplied by its opacity $o_{j}$ . + +We collected RGB videos from the large-scale open-source DL3DV dataset [27], which contains over 10,000 videos across diverse domains, including natural and outdoor settings, educational institutions, shopping complexes, + +parks, hubs, cafes, and restaurants. We selected videos based on the presence of natural occlusions, such as garlands, chairs, benches, statues, cars, and bins. Our dataset captures 3D representations of approximately 40 distinct scenes, each incorporating real-world occlusions. This approach to learning 3D representation closely resembles real-world data compared to conventional methods that rely on rendering from 3D graphic asset scenes. + +SMPLX body/Human Animations. We represent the human body using the SMPL-X [36] 3D human model, defined by the function $\mathcal{M}(\theta ,\beta ,\psi)$ , where $\theta$ represents the pose parameters, $\beta$ the shape parameters, and $\psi$ the facial expression parameters. This function outputs a body mesh $\mathcal{M}\in \mathbb{R}^{10475\times 3}$ with 10,475 vertices. We sample approximately 400 SMPL-X 3D human motion models from the AMASS mocap dataset [30], which contains over 11,000 motion sequences with diverse body shapes and poses. Following previous work [48], we use the pre-trained human pose prior model VPoser [36] to assess pose difficulty. VPoser is a Variational Auto-Encoder model that evaluates a pose $\theta$ to be challenging if its embeddings $\epsilon_{\theta}$ have larger norm, i.e. $||\epsilon_{\theta}||_{2} > \tau$ , where $\tau$ is empirically set to 40. We classify a challenging pose using VPoser to avoid simple movements such as walking, standing, jogging, etc. To prepare human animations for rendering, we use Blender graphics software to efficiently bake complex geometric animations over time into the Alembic format for subsequent use in rendering engines. + +We ensured that motion sequences contained at least 180 frames to maintain a minimum animation duration of six seconds at 30 fps. For sequences exceeding 400 frames, we discarded the first 100 frames to remove static poses at the start. To ensure that human motion remains consistently within the occluded scene, we implemented boundary constraints that stop the motion if it moves beyond these limits. These conditions prevent humans from moving too far from the occlusion area. Additionally, we applied random rotations and translations to each sequence. We store the applied transformations to further calculate the effective global orientation of the human body with respect to the world coordinates. + +Human Textures. We sourced human body texture scans from the open-source dataset provided by the SMPLtex method [4], which estimates and manipulates the full 3D appearance of humans captured from a single image. The SMPLtex dataset provides a diverse range of human texture scans, covering various skin tones, clothing styles, and genders. To ensure sample diversity in VOcc13D, we select approximately 200 distinct texture scans. Figure 4 illustrates the different clothing textures applied to the SMPL-X human model in our dataset. + +![](images/e3f841e87163ddc3c5a9d1d8f2b573a19a0a973e52ae81e67c77ebea9a35f986.jpg) +Figure 3. An overview of our proposed dataset creation pipeline. We generate 3D representations of natural scenes with real occlusions using 3D Gaussian Splatting [19]. Human motion sequences from the AMASS MoCap dataset [30] are processed in Blender to generate baked animation files. The generated scene representations, animation files, human texture maps, and camera extrinsic parameters are imported into the Unity rendering engine to generate video sequences of human motion under occlusion. + +![](images/deff0eb4f7e24634e7b727e37a11bdef5b88cd710bd7626147f2bd84321b434a.jpg) +Figure 4. Diversity of human texture in VOccl3D. We use approximately 200 human texture maps from the SMPLtex dataset [4], encompassing diverse clothing styles, genders, skin tones, and ethnicities. + +# 3.2. Dataset Rendering and Attributes + +Rendering As shown in Figure 3, we used Unity Engine to render synthetic scenes by integrating background Gaussian splats, human animations, and body texture assets. We rendered video sequences at 30 fps with a resolution of $720 \times 720$ , varying the camera field of view between 25 and 50 across different scenes. For each sequence, we stored ground-truth camera extrinsic and intrinsic parameters and saved the rendered RGB images in lossless PNG format. We positioned the camera to capture occluded views within the scene and applied motion constraints to human animation assets to maintain consistent occlusion throughout each sequence. To achieve realistic lighting, we employed Unity's built-in DirectionalLight asset to simulate natural illumination on the human mesh. + +Dataset Attributes. Our proposed dataset, VOccl3D, includes over 250,000 images and 400 video sequences with + +a total runtime exceeding 2 hours and 30 minutes. The dataset features 40 background scenes with various occlusions, such as cars, garlands, benches, chairs, bins, and trees. Each scene contains 10 video sequences with variations in human motions and clothing textures. The rendered sequence exhibit diversity in body shapes, skin tones, and camera poses. We provide ground-truth annotations, including camera extrinsic and intrinsic parameters, pose, shape, global orientation, translation, gender, 2D keypoints, and a binary occlusion label for each keypoint. + +Different Modalities. In addition to 3D pose and shape annotations, VOccl3D provides annotations for multiple modalities, including human silhouettes, body-part segmentation, 2D keypoint estimation, and human bounding box detection under occlusion. Researchers can use our dataset to train and evaluate methods designed to handle occlusion across these modalities. + +# 4. Experiments and Results + +In this section, we highlight the need and significance of our proposed VOccl3D dataset. We fine-tune the state-of-the-art pose estimation methods using our dataset and report the qualitative and quantitative results. Our results demonstrate improved performance on the HPS estimation task (Section 4.1) using both our proposed synthetic dataset and real-world datasets with occlusions. In Section 4.2, we show the enhanced performance of the human object detector and its impact on HPS estimation under occlusion. + +# 4.1. Human Pose and Shape Estimation + +Dataset and Implementation Setup. We use approximately 200k and 50k images for training and testing respectively in our proposed VOcc13D dataset. We fine-tune + +
MethodHard-OcclusionMedium-OcclusionLow-Occlusion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]192.22114.35247.41121.7078.56158.4698.8267.64126.92
BEDLAM-HMR [3]167.35102.92214.39102.5568.13134.5986.5554.48110.15
BEDLAM-CLIFF [3]154.8699.53199.9590.9765.03119.6374.9552.6596.60
HMR2.0 [11]169.71100.49215.17113.8871.78145.6288.5359.08114.39
STRIDE with BEDLAM-CLIFF [24]155.64100.44-91.1465.38-75.0253.21-
WHAM [44]152.15102.14177.07110.9776.81127.4593.9066.68106.51
VOcc13D-B-CLIFF136.3489.94175.9282.4858.78106.8469.4646.3288.19
STRIDE with VOcc13D-B-CLIFF136.4390.28-82.3758.98-69.6546.86-
+ +Table 1. 3D HPS estimation results on the test-split of VOcccl3D. The results show that VOcccl3D-B-CLIFF and STRIDE, when using pseudo-labels from VOcccl3D-B-CLIFF, significantly outperform other image- and video-based HPS estimation methods across hard, medium, and low occlusion categories. As expected, all HPS estimation methods exhibit a decline in performance as occlusion severity increases from low to medium to high. For evaluation, we use ground-truth bounding boxes. The best results are in bold. + +
Method3DPWOcclType1-3DPWOcclType2-3DPW
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]73.946.487.698.1562.27118.4799.4962.16119.82
BEDLAM-HMR [3]79.047.693.1108.6266.53128.05106.1964.05125.66
BEDLAM-CLIFF [3]72.046.685.098.7164.26117.4196.8061.32115.33
HMR2.0 [11]81.254.3143.7103.4069.66164.5599.0166.17158.79
VOccl3D-B-CLIFF72.047.384.595.8963.43114.2894.3660.44112.01
VOccl3D-CLIFF71.1045.9884.2595.1761.83114.9593.7459.66112.59
+ +Table 2. 3D HPS estimation results on 3DPW, OcclType1-3DPW, and OcclType2-3DPW. Since 3DPW is a real-world dataset with minimal occlusions, VOccl3D-CLIFF outperforms all other methods but shows only a marginal improvement over BEDLAM-CLIFF. However, on OcclType1-3DPW and OcclType2-3DPW, which contain significant occlusions on real-world dataset, our method demonstrates a notable improvement over existing approaches. The best results are noted in bold. + +![](images/2a2fb20081034a910d905c0023aa4060635f6fec6e8355823f136332f4cd760e.jpg) +Figure 5. Fine-tuning an off-the-shelf pose estimator. We fine-tune a CLIFF [25] pose estimation model using our proposed VOcc13D dataset. We observe a significant improvement in the estimated mesh when using a fine-tuned CLIFF model using our dataset. Note: The pose estimator can be any off-the-shelf pose estimator, here we show results with the CLIFF model [25]. + +two versions of state-of-the-art model, CLIFF (trained on Human3.6M [15], MPI-INF-3DHP [31], and COCO [26] dataset) and BEDLAM-CLIFF (trained on BEDLAM and AGORA datasets) using our rendered VOcc13D dataset, resulting in VOcc13D-CLIFF and VOcc13D-B-CLIFF, re + +spectively (refer to Figure 5). To benchmark performance, we evaluate on the test split of VOccl3D, comparing against multiple image and video based baselines, including CLIFF [25], BEDLAM-HMR [3], BEDLAM-CLIFF [3], HMR2.0 [11], STRIDE [24], and WHAM [44]. We report the mean performance by evaluating the model using a fivefold cross-validation on our dataset. To quantify occlusion levels, we annotate each 3D joint with a binary occlusion label, marking it as occluded if its corresponding 2D keypoint lies within the ground-truth segmentation mask of the visible human region. Based on the number of visible keypoints (out of 22 total), we categorize images into three occlusion levels: hard occlusion (4-9 visible keypoints), medium occlusion (10-15 visible keypoints), and low occlusion (16-20 visible keypoints). + +Beyond evaluations on synthetic datasets, we assess the performance of VOccl3D fine-tuned models on real-world datasets, including 3DPW [46] and OCMotion [13]. It is noteworthy that these datasets do not contain significant occlusions. Hence, we introduce two occlusion-augmented variants of 3DPW to evaluate robustness under severe occlusions. In OcclType1-3DPW, we overlay a black patch on a randomly selected keypoint, while in OcclType2-3DPW, we occlude two random keypoints. Further details on oc + +![](images/2a712e5eb8c5fe1284bfca9681cfe5c6b9be67f71f4352d369042bb5f9a0dc1f.jpg) +Figure 6. Qualitative comparison of HPS estimation methods on VOcc3D dataset. The first and second column shows RGB image and ground-truth human mesh. Column third and fourth compare HPS estimation using the BEDLAM-CLIFF [3] and HMR2.0 [11] methods. The final column (VOcc3D-B-CLIFF) presents results obtained by fine-tuning the CLIFF model on the VOcc3D dataset. These results demonstrate superior performance, particularly in scenarios with heavy occlusion. + +clusion variants of 3DPW and implementation details are provided in the supplementary material. + +Results. We present benchmarking results on the test split of our proposed VOcc13D dataset in Table 1. Our finetuned VOcc13D-B-CLIFF significantly outperforms existing state-of-the-art image and video-based methods across all occlusion categories. Notably, the performance of all previous methods degrades as occlusion severity increases, highlighting the inherent challenges of HPS estimation under occlusions. We also compare our approach with the plug-and-play method STRIDE [24], which performs 3D pose estimation and reports only MPJPE and PA-MPJPE errors. Our results show that STRIDE achieves superior performance when leveraging pseudo-labels from the VOcc13D-B-CLIFF model compared to those from the original BEDLAM-CLIFF model, further validating the effectiveness of our dataset. In Table 2 and Table 3, we report performance on the 3DPW [45] and OCMotion [14] datasets, respectively. We observe that both VOcc13D-B-CLIFF and VOcc13D-CLIFF outperform all the previous methods, demonstrating their effectiveness in real-world + +scenarios. However, since these datasets contain minimal occlusions, the improvements over CLIFF remain marginal. To further assess robustness under heavy occlusion, we evaluate on OcclType1-3DPW and OcclType2-3DPW, real-world datasets with significant occlusions. Table 2 demonstrates that our fine-tuned model achieves substantial improvements, underscoring its robustness in occluded environments. Table 3 presents results on the OCMotion dataset, which features lower levels of occlusion. We can observe that VOcc13D-CLIFF outperforms all comparison methods and demonstrates performance comparable to CLIFF [25]. This is expected, as both models are pretrained on large-scale real-world datasets captured in controlled lab environments similar to OCMotion. Figure 6 presents the qualitative results on VOcc13Ddataset, showcasing the improvement of our fine-tuned model against other state-of-the-art HPS estimation methods. Additional results are provided in supplementary material. + +# 4.2. Impact of Human Detector on HPS Estimation + +Dataset and Implementation Setup. Human object detectors play a crucial role in HPS estimation tasks, particularly + +
MethodMPJPEOCMotion PA-MPJPEPVE
CLIFF [25]64.1540.1679.80
BEDLAM-HMR [3]73.9641.9492.70
BEDLAM-CLIFF [3]66.8041.7983.86
HMR2.0 [11]69.9443.0087.07
VOcc13D-B-CLIFF65.9640.5981.96
VOcc13D-CLIFF64.2939.6478.56
+ +Table 3. Qualitative results of HPS estimation on OCMotion. Our results show that VOcc13D-CLIFF outperforms all HPS estimation methods except CLIFF [25]. Both CLIFF and VOcc13D-CLIFF exhibit comparable performance, as both are pre-trained on real-world datasets captured in controlled lab environments, similar to OCMotion [14]. The best results are in bold. + +
Method3DPWOCMotion
mAP50mAP75mAP50mAP75
YOLO1158.9947.1498.8491.80
VOccl3D-YOLO1159.8948.2699.1091.95
+ +when inferring from in-the-wild RGB images [3, 24, 25]. Ideally, HPS estimation performs optimally with ground-truth bounding boxes; however, most human object detectors struggle significantly under occlusion. To address this, we fine-tune the recent YOLO11 detector [16] on the combined training split of VOccl3D and MS COCO, referring to the fine-tuned model as VOccl3D-YOLO11. We show evaluations on 3DPW [45] and OCMotion [14] datasets. + +Results. Table 4 compares the detection performance of the pre-trained YOLO11 and VOccl3D-YOLO11 models on the real-world 3DPW [45] and OCMotion [14] datasets. Since OCMotion is a single-human dataset with minimal occlusions, the detector achieves high mAP scores. This is likely because most body parts remain visible, making box detection easier. However, due to the presence of multiple individuals in 3DPW, the overall mAP scores remain lower. In short, VOccl3D-YOLO11 shows significant improvement on datasets with high occlusions. We provide qualitative results in the supplementary material. + +Table 5 further evaluates the impact of human detection on HPS estimation. The first, second, and third rows represent the performance of VOccl3D-CLIFF when human detections are sourced from ground truth, YOLO11, and VOccl3D-YOLO11, respectively. As expected, the best performance is achieved when using ground-truth detections. However, fine-tuning the detector with VOccl3D significantly improves HPS estimation compared to detections + +Table 4. Quantitative results of YOLO11 on 3DPW and OC-Motion datasets. The first row presents the human object detection performance of the pre-trained YOLO11 model, while the second row shows the results after fine-tuning with the VOccl3D dataset. The best results are in bold. We observe an improvement in detection accuracy across both datasets after fine-tuning. + +
Method3DPWOCMotion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
VOccl3D-CLIFF w/GT71.1045.9884.2564.2939.6478.56
VOccl3D-CLIFF w/YYOLO11116.5263.35139.7467.1641.3083.00
VOccl3D-CLIFF w/VOccl3D-YYOLO11114.8562.66137.1966.6541.1582.40
+ +Table 5. 3D HPS estimation results using the VOcc13D-CLIFF model with different bounding box sources. This table presents HPS estimation performance on the 3DPW and OC-Motion datasets. The first row reports results using ground-truth bounding boxes, achieving the highest accuracy. The second and third rows show performance when detections are obtained from the pre-trained YOLO11 and fine-tuned VOcc13D-YOLO11 models, respectively. Notably, the third row demonstrates a significant improvement over the second, highlighting the impact of detections on HPS estimation. + +from the pre-trained YOLO11. These results underscore the critical role of human detection quality in enhancing HPS performance under occlusion. + +Our experiments show that fine-tuning an existing HPS estimation model with a large synthetic dataset containing occlusions enhances its performance on both real and synthetic datasets. We observe significant improvements over baseline methods, particularly in scenarios with heavy occlusions. Additionally we demonstrate that the poor results are not solely due to HPS estimations errors but also stem from failures in bounding box detection. The VOcc13D dataset proves effective in refining pose estimation methods and bounding box predictions under occlusions. + +# 5. Conclusion + +We introduce VOccl3D, a novel large-scale video dataset with synthetic humans in real-world scenes under occlusions for 3D human pose and shape estimation. By leveraging a rendering-based approach, VOccl3D eliminates the need for costly and labor-intensive data collection while providing a diverse and realistic dataset for occlusion-aware research. Through extensive qualitative and quantitative evaluations, we demonstrate that fine-tuning existing HPS estimation methods with VOccl3D significantly enhances their performance on real-world datasets with occlusions and on the test split of VOccl3D dataset. Furthermore, we improve the human object detector, YOLO11 under occluded conditions using VOccl3D, highlighting the impact of detections in achieving robust HPS estimation in occluded scenarios. Beyond HPS estimation, VOccl3D serves as a comprehensive benchmark for evaluating methods across multiple occlusion-aware tasks, including human body-part segmentation, 2D/3D pose estimation, and human bounding box detection. Hence, VOccl3D sets a new standard for occluded human benchmarks, offering a valuable dataset for advancing occlusion-robust research. + +# Acknowledgment + +This work was partially supported by USDA NRI grant 2021-67022-33453, NSF grants CMMI-2326309 and CNS-2312395. This research was also partially supported by the Office of the Director of National Intelligence (ODNI), specifically through the Intelligence Advanced Research Projects Activity (IARPA), under contract number [2022-21102100007]. The views and conclusions in this research reflect those of the authors and should not be construed as officially representing the policies, whether explicitly or implicitly, of ODNI, IARPA, or the U.S. Government. Nevertheless, the U.S. Government retains the authorization to reproduce and distribute reprints for official government purposes, regardless of any copyright notices included. + +# References + +[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3 +[2] Peter Bauer, Arij Bouazizi, Ulrich Kressel, and Fabian B. Flohr. Weakly supervised multi-modal 3d human body pose estimation for autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7, 2023. 2 +[3] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion, 2023. 2, 3, 6, 7, 8 +[4] Dan Casas and Marc Comino-Trinidad. SMPLitex: A Generative Model and Dataset for 3D Human Texture Estimation from Single Image. In British Machine Vision Conference (BMVC), 2023. 4, 5 +[5] Yu Cheng, Bo Yang, Bo Wang, Yan Wending, and Robby Tan. Occlusion-aware networks for 3d human pose estimation in video. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 723-732, 2019. 3 +[6] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3 +[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 591-601, 2022. 2 +[8] R James Cotton. Posepipe: Open-source human pose estimation pipeline for rehabilitation research. Archives of Physical Medicine and Rehabilitation, 103(12):e161-e162, 2022. 2 +[9] Yufeng Cui and Yimei Kang. Multi-modal gait recognition via effective spatial-temporal feature fusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17957, 2023. 2 +[10] Yang Fu, Shibei Meng, Saihui Hou, Xuecai Hu, and Yongzhen Huang. Gpgait: Generalized pose-based gait + +recognition. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19538-19547, 2023. 2 + +[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 6, 7, 8 + +[12] Charlie Hewitt, Fatemeh Saleh, Sadegh Aliakbarian, Lohit Petikam, Shideh RezaEIFar, Louis Florentin, Zafiirah Hosenie, Thomas J Cashman, Julien Valentin, Darren Cosker, et al. Look ma, no markers: Holistic performance capture without the hassle. ACM Transactions on Graphics, 43(6): 1-12, 2024. 2, 3 + +[13] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2022. 6 + +[14] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2023. 2, 7, 8 + +[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 3, 6 +[16] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 3, 8 +[17] Sardor Juraev, Akash Ghimire, Jumabek Alikhanov, Vijay Kakani, and Hakil Kim. Exploring human pose estimation and the usage of synthetic data for elderly fall detection in real-world surveillance. IEEE Access, 10:94249-94261, 2022. 2 +[18] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2017. 2, 3 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4, 5 +[20] Do-Yeop Kim and Ju-Yong Chang. Attention-based 3d human pose sequence refinement network. Sensors, 21(13), 2021. 3 +[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3 +[22] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2 +[23] Farnoosh Koleini, Muhammad Usama Saleem, Pu Wang, Hongfei Xue, Ahmed Helmy, and Abbey Fenwick. Biopose: Biomechanically-accurate 3d pose estimation from monocular videos. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6330–6339. IEEE, 2025. 2 + +[24] Rohit Lal, Saketh Bachu, Yash Garg, Arindam Dutta, Calvin-Khang Ta, Dripta S. Raychaudhuri, Hannah Dela Cruz, M. Salman Asif, and Amit K. Roy-Chowdhury. Stride: Single-video based temporally continuous occlusion robust 3d pose estimation, 2024. 2, 3, 6, 7, 8 +[25] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation, 2022. 2, 3, 6, 7, 8 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6 +[27] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 4 +[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3 +[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3 +[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 5 +[31] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 International Conference on 3D Vision (3DV), pages 506-516, 2017. 6 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 4 +[33] Sara Moccia, Lucia Migliorelli, Virgilio Carnielli, and Emanuele Frontoni. Preterm infants' pose estimation with spatio-temporal features. IEEE Transactions on Biomedical Engineering, 67(8):2370-2380, 2019. 2 +[34] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J. Mitra. imapper: interaction-guided scene mapping from monocular videos. ACM Trans. Graph., 38(4), 2019. 2 +[35] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13468-13478, 2021. 2 +[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and + +Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4 +[37] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7745-7754, 2018. 3 +[38] Miroslav Purkrábek and Jiri Matas. Improving 2d human pose estimation across unseen camera views with synthetic data. arXiv preprint arXiv:2307.06737, 2023. 3 +[39] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4 +[40] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 2 +[41] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, page 1530-1538. JMLR.org, 2015. 3 +[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 8(11):7090-7097, 2023. 2 +[43] Saurabh Sharma, Pavan Teja Varigonda, Prashast Bindal, Abhishek Sharma, and Arjun Jain. Monocular 3d human pose estimation by generation and ordinal ranking. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3 +[44] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2070-2080, 2023. 2, 3, 6 +[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 7, 8 +[46] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European conference on computer vision (ECCV), pages 601-617, 2018. 6 +[47] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In International Conference on Computer Vision (ICCV), 2021. 3 +[48] Zhenzhen Weng, Laura Bravo-Sanchez, and Serena Yeung-Levy. Diffusion-hpc: Synthetic data generation for human mesh recovery in challenging domains. In 2024 International Conference on 3D Vision (3DV), pages 257-267. IEEE, 2024. 4 + +[49] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, et al. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20282-20292, 2023. 3 +[50] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9060-9069, 2021. 3 +[51] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[52] Andrei Zanfir, Mihai Zanfir, Alex Gorban, Jingwei Ji, Yin Zhou, Dragomir Anguelov, and Cristian Sminchisescu. HUM3DIL: Semi-supervised multi-modal 3d humanpose estimation for autonomous driving. In 6th Annual Conference on Robot Learning, 2022. 2 +[53] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In European Conference on Computer Vision. Springer, 2022. 3 +[54] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conference on Computer Vision and Pattern Recognition, (CVPR), 2020. 2 +[55] Yi Zhang, Pengliang Ji, Adam Kortylewski, Angtian Wang, Jieru Mei, and Alan L Yuille. 3D-Aware Neural Body Fitting for Occlusion Robust 3D Human Pose Estimation. In The IEEE/CVF International Conference on Computer Vision, 2023. 2 +[56] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G. Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4966-4975, 2015. 3 +[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 4 \ No newline at end of file diff --git a/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/images.zip b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51b0687be1b634687af1c335e1ac19dffacc6ee3 --- /dev/null +++ b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46be364fc6738ccfaba8c948f835c6c193860ec0e8b06f6569ffca374ce83216 +size 742754 diff --git a/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/layout.json b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f24d1690add60615adceb1629f30598a1d8e5fb0 --- /dev/null +++ b/2025/VOccl3D_ A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions/layout.json @@ -0,0 +1,6659 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 488, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 488, + 138 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 488, + 138 + ], + "type": "text", + "content": "VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real Occlusions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 161, + 469, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 161, + 469, + 203 + ], + "spans": [ + { + "bbox": [ + 141, + 161, + 469, + 203 + ], + "type": "text", + "content": "Yash Garg, Saketh Bachu, Arindam Dutta, Rohit Lal†, Sarosij Bose, Calvin-Khang Ta‡, M. Salman Asif, Amit Roy-Chowdhury University of California, Riverside" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 205, + 500, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 205, + 500, + 217 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 500, + 217 + ], + "type": "text", + "content": "{ygarg002,sbach008,adultt020,r1al011,sbose007,cta003,sasif,amitrc}@ucr.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 227, + 552, + 364 + ], + "blocks": [ + { + "bbox": [ + 59, + 227, + 552, + 364 + ], + "lines": [ + { + "bbox": [ + 59, + 227, + 552, + 364 + ], + "spans": [ + { + "bbox": [ + 59, + 227, + 552, + 364 + ], + "type": "image", + "image_path": "2d1f4f7d05d18f1178989a04559ef760ba9159d82ed48257677342924a0978d2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 447 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 447 + ], + "type": "text", + "content": "Figure 1. We propose VOccl3D, a large-scale synthetic video dataset specifically designed for training and evaluating algorithms for 3D human pose and shape estimation (HPS) in realistic occlusion scenarios. VOccl3D comprises over 250,000 frames, with a total runtime exceeding 2 hours and 30 minutes. Compared to previous occlusion-based datasets, VOccl3D has diverse body shapes, textures, and most importantly, significant and realistic occlusions within the scenes. In addition to body shape and pose, VOccl3D provides other necessary ground truth information, such as bounding boxes, body part segmentations, and human silhouettes. The top row illustrates various frames from our dataset, showcasing diverse occlusions, clothing textures, and motions, while the bottom row represents a sequence of frames from a video sequence within our dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 468, + 200, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 468, + 200, + 481 + ], + "spans": [ + { + "bbox": [ + 151, + 468, + 200, + 481 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": "Human pose and shape (HPS) estimation methods have been extensively studied, with many demonstrating high zero-shot performance on in-the-wild images and videos. However, these methods often struggle in challenging scenarios involving complex human poses or significant occlusions. Although some studies address 3D human pose estimation under occlusion, they typically evaluate performance on datasets that lack realistic or substantial occlusions, e.g., most existing datasets introduce occlusions with random patches over the human or clipart-style overlays, which may not reflect real-world challenges. To bridge this gap in realistic occlusion datasets, we introduce a novel benchmark dataset, VOccl3D, a Video-based human Occlusion dataset with 3D body pose and shape annotations. Inspired by works such as AGORA and BEDLAM, we constructed this dataset using advanced computer graphics rendering techniques, incorporating diverse real-world occlusion scenarios, clothing textures, and human motions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 469, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 650 + ], + "type": "text", + "content": "Additionally, we fine-tuned recent HPS methods, CLIFF and BEDLAM-CLIFF, on our dataset, demonstrating significant qualitative and quantitative improvements across multiple public datasets, as well as on the test split of our dataset, while comparing its performance with other state-of-the-art methods. Furthermore, we leveraged our dataset to enhance human detection performance under occlusion by fine-tuning an existing object detector, YOLO11, thus leading to a robust end-to-end HPS estimation system under occlusions. Overall, this dataset serves as a valuable resource for future research aimed at benchmarking methods designed to handle occlusions, offering a more realistic alternative to existing occlusion datasets. See the Project page for code and dataset:https://yashgarg98.github.io/Voccl3D-dataset/" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": "† Currently at NASA MSFC IMPACT. ‡ Currently at Dolby Laboratories. Work done while the authors were at UCR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7350" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 258 + ], + "type": "text", + "content": "Monocular 3D human pose and shape (HPS) estimation is a complex yet essential task in computer vision. It has applications in surveillance [7, 17], autonomous robotics [2, 42, 52], human motion analysis [9, 10], and clinical assessment [8, 23, 33]. Since the introduction of the neural network-based Human Mesh Recovery Network (HMR) [18], numerous approaches have advanced the field by enabling accurate estimation of SMPL[28] pose and shape from a single RGB image. Recent methods [1, 21, 24, 25, 29, 44] demonstrate impressive zero-shot HPS performance on in-the-wild RGB images and video sequences. However, these methods still face limitations in achieving high performance in challenging scenarios, such as complex human poses and under significant occlusions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 462 + ], + "type": "text", + "content": "Achieving robust HPS under occlusion remains a challenging problem due to the contextual ambiguity of the occluded body parts. Several methods [22, 40, 51, 55] have been proposed to address the challenges caused by occlusions. Methods like PARE [22] focus on using visible body parts to infer occluded areas, enhancing estimation accuracy with partial views. Temporal models like HuMoR [40] and GLAMR [51] use generative frameworks to ensure pose continuity over time. HuMoR predicts pose distributions, while GLAMR incorporates global trajectory data to fill in missing poses. Recent methods, like STRIDE [24], leverage a large-scale pre-trained motion prior model to achieve temporally coherent pose reconstruction under occlusions. While these methods effectively leverage visible information and temporal continuity, they still struggle under severe occlusions due to limited exposure to occluded scenarios in their training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 689 + ], + "type": "text", + "content": "Evaluations of the existing methods are often limited to occluded datasets that lack scene diversity [34], involve moderate or lower levels of occlusion [14, 45], or use patch-based occlusion datasets for training and inference [55], as illustrated in Figure 2. Most importantly, there remains an urgent need for a realistic, diverse, and significantly occluded human pose and shape dataset to advance the task of 3D HPS estimation. To address this gap, our work proposes VOccl3D, a large-scale video dataset with synthetic humans in real scenes for 3D human pose and shape estimation. Following prior works such as BEDLAM [3] and Synthocap [12], we use an advanced computer graphics engine to render a highly realistic synthetic dataset. Prior research has shown that fine-tuning models on synthetic datasets can significantly improve HPS estimation performance on real-world datasets [3]. Synthetic datasets provide \"perfect\" ground-truth annotations and eliminate the need for costly sensors, making it computationally efficient to generate large-scale datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Our proposed dataset includes approximately 250,000 images and 400 video sequences. To introduce diversity, we" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 397, + 129 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 397, + 129 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 397, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 397, + 129 + ], + "type": "image", + "image_path": "cbe23ef8950056961997cd564aa47c6151b13a5f81dd23cfc7b4487a4336beb7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 133, + 389, + 140 + ], + "lines": [ + { + "bbox": [ + 328, + 133, + 389, + 140 + ], + "spans": [ + { + "bbox": [ + 328, + 133, + 389, + 140 + ], + "type": "text", + "content": "a) 3DPW-AdvOcc@80" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 401, + 70, + 476, + 129 + ], + "blocks": [ + { + "bbox": [ + 401, + 70, + 476, + 129 + ], + "lines": [ + { + "bbox": [ + 401, + 70, + 476, + 129 + ], + "spans": [ + { + "bbox": [ + 401, + 70, + 476, + 129 + ], + "type": "image", + "image_path": "e289547aa747317d3b88d78846c6025832db910e10f785e4fdbb13c9c414f968.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 479, + 70, + 547, + 129 + ], + "blocks": [ + { + "bbox": [ + 479, + 70, + 547, + 129 + ], + "lines": [ + { + "bbox": [ + 479, + 70, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 479, + 70, + 547, + 129 + ], + "type": "image", + "image_path": "97c2cc452f7e7a57730b9269869dd4327882f8e9dc53ba8e04b414ec5b686fbc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 142, + 399, + 208 + ], + "blocks": [ + { + "bbox": [ + 316, + 142, + 399, + 208 + ], + "lines": [ + { + "bbox": [ + 316, + 142, + 399, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 399, + 208 + ], + "type": "image", + "image_path": "fefe371bdf90e683d3c6e293f6b92d15f692b343f8d9f9f4d6eac1dfde7b981c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 212, + 374, + 220 + ], + "lines": [ + { + "bbox": [ + 337, + 212, + 374, + 220 + ], + "spans": [ + { + "bbox": [ + 337, + 212, + 374, + 220 + ], + "type": "text", + "content": "d) OCMotion" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 413, + 141, + 460, + 209 + ], + "blocks": [ + { + "bbox": [ + 403, + 133, + 473, + 140 + ], + "lines": [ + { + "bbox": [ + 403, + 133, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 403, + 133, + 473, + 140 + ], + "type": "text", + "content": "b) Occluded Human3.6M" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 141, + 460, + 209 + ], + "lines": [ + { + "bbox": [ + 413, + 141, + 460, + 209 + ], + "spans": [ + { + "bbox": [ + 413, + 141, + 460, + 209 + ], + "type": "image", + "image_path": "4490af5634a6def6a3f68b0a37e61ada84deb7ab97a2eb28ae998f61e078c83b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 211, + 450, + 220 + ], + "lines": [ + { + "bbox": [ + 424, + 211, + 450, + 220 + ], + "spans": [ + { + "bbox": [ + 424, + 211, + 450, + 220 + ], + "type": "text", + "content": "e) 3DPW" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 471, + 141, + 552, + 209 + ], + "blocks": [ + { + "bbox": [ + 476, + 132, + 548, + 140 + ], + "lines": [ + { + "bbox": [ + 476, + 132, + 548, + 140 + ], + "spans": [ + { + "bbox": [ + 476, + 132, + 548, + 140 + ], + "type": "text", + "content": "c) Syn. Occ. Human3.6M" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 141, + 552, + 209 + ], + "lines": [ + { + "bbox": [ + 471, + 141, + 552, + 209 + ], + "spans": [ + { + "bbox": [ + 471, + 141, + 552, + 209 + ], + "type": "image", + "image_path": "4c1af9124c611fdfc1ae15ef8e39dd1a4407707f01324b34a104e7d87ce02ee1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 212, + 522, + 220 + ], + "lines": [ + { + "bbox": [ + 500, + 212, + 522, + 220 + ], + "spans": [ + { + "bbox": [ + 500, + 212, + 522, + 220 + ], + "type": "text", + "content": "f) Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 228, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. Samples from various occlusion-based HPS datasets. The top row displays datasets with artificial, patch-based occlusions applied to existing datasets, including 3DPW-AdvOcc [45, 55], Occluded Human3.6M [15, 24], and Synthetic Occlusion Human3.6M [15, 54]. The bottom row presents samples from two datasets with natural occlusions: OCMotion [14] and 3DPW [45]. Notably, the top-row images exhibit unrealistic occlusions, which lack the realism of naturally occurring occlusions. The bottom row images contain natural occlusions but are sparse and infrequent. Compared to existing datasets, our proposed VOcccl3D dataset features more realistic occlusions." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 365, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 555 + ], + "type": "text", + "content": "incorporate human motion samples from the AMASS [30] dataset, over 200 distinct clothing textures for sequences containing both male and female genders, and 40 real background scenes with occlusions. Additionally, we provide occlusion labels for each body joint. Unlike prior methods [3, 35] which render synthetic environments, we utilize 3D Gaussian splatting (3DGS) [19] to create 3D representations of background scenes. This approach enhances the realism of rendered scenes, making them closely resemble real-world captured data. Reconstructing background scenes using 3DGS offers the flexibility to create domain-specific datasets, such as for agriculture, street or indoor environments using raw RGB video captures without relying on costly or limited graphic assets. Figure 1 showcases samples from our dataset, illustrating the diversity in clothing, environments, human motions, and occlusions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "We fine-tuned the CLIFF [25] and BEDLAM-CLIFF [3] models on our dataset to demonstrate that training with synthetic data enhances the performance of existing HPS estimation methods under occlusion. We evaluated state-of-the-art HPS estimation methods on the test split of our dataset across three occlusion levels: high, medium, and low. The results show that existing methods perform poorly under high occlusion, whereas our fine-tuned models achieve significant improvements. To assess performance on real-world data, we perform evaluations on the 3DPW and OCMotion datasets, which contain low-level occlusions. Additionally, we created a variant of the 3DPW dataset by adding black patches to evaluate the robustness" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7351" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "against high-occlusion real scenarios. Our fine-tuned models outperformed recent HPS methods on these real-world datasets. Furthermore, our experiments highlight the importance of an effective human object detector for improving HPS performance under occlusion. To address this, we fine-tuned the YOLO11[16] object detector using our dataset, enhancing its performance under occluded scenarios. VOcc13D is available for researchers to benchmark and evaluate occlusion-aware methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 180, + 288, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 180, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 67, + 180, + 288, + 192 + ], + "type": "text", + "content": "In summary, we make the following key contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 295, + 384 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 194, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 229 + ], + "type": "text", + "content": "1. We propose VOcc13D, a novel large-scale, realistic video-based dataset of occluded synthetic humans in real scenes for 3D human pose and shape estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 229, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 294, + 277 + ], + "type": "text", + "content": "2. We demonstrate both qualitatively and quantitatively that existing HPS estimation methods fine-tuned with VOccl3D outperform other methods on real-world datasets with occlusions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 294, + 323 + ], + "type": "text", + "content": "3. We improved the performance of the human object detector under occlusion scenarios by leveraging the VOccl3D dataset which is a crucial component for a robust HPS estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 325, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 294, + 384 + ], + "type": "text", + "content": "4. VOccl3D can serve as a benchmark dataset for evaluating methods specifically designed to perform under occlusion across various tasks, including human and body-part segmentation, 2D/3D pose estimation, and human bounding box detection." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 394, + 146, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 146, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 146, + 407 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 415, + 295, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 426 + ], + "type": "text", + "content": "Synthetic Data Generation for Human Pose Estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 426, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 700 + ], + "type": "text", + "content": "Human pose estimation is crucial in computer vision, but state-of-the-art methods depend on costly, labor-intensive labeled datasets. Several notable synthetic datasets have advanced research in human pose estimation. SynBody [49] provides a large-scale synthetic dataset for human mesh recovery and view synthesis, while RePoGen [38] enables fine-grained control over pose and viewpoint to generate rare, complex poses. Human3.6m [15] provides additional small mixed reality dataset by inserting 3D animation models with background scenes. BEDLAM [3] further highlights that models trained solely on synthetic data can outperform real-data-trained counterparts, emphasizing the importance of high-quality synthetic datasets for transferable models. Further, PressurePose [6] simulates interactions between articulated and soft bodies, capturing fine-grained contact dynamics. SynthMocap [12] extends this by providing expressive synthetic data with detailed body, hand, and face movements. Despite their success, these synthetic datasets lack emphasis on significant occlusions, a key real-world challenge. Addressing this limitation, our work proposes a synthetic dataset tailored to 3D pose estimation under heavy and realistic occlusion, aiming to bridge this gap and enhance robustness in occlusion-prone environments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 702, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 713 + ], + "type": "text", + "content": "Image and Video based Human Pose Estimation. Esti" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "mating 2D/3D pose and shape from single RGB image has widely been explored in previous research [3, 25, 43, 47, 50]. Methods such as [43] use a conditional variational autoencoder for 2D-to-3D lifting in pose estimation. The pose estimation approach proposed in [47] applies normalizing flows [41] for 2D-3D mapping. Large-scale datasets like BEDLAM [3] have improved pose estimation models like CLIFF [25] and HMRNet [18]. However, these models struggle with generalizing to unseen scenarios with severe occlusions due to limited training on such cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 195, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 553, + 434 + ], + "type": "text", + "content": "In addition, video-based human pose estimation has significantly improved challenging datasets. Early works like [56] use the EM algorithm to estimate the 3D pose from monocular video through 2D joint uncertainty maps. The method in [37] employs dilated temporal convolutions and semi-supervised learning for 3D pose estimation, while [1] uses the SMPL model [28] to extract pose and shape parameters, refining models like HMR with bundle adjustment. VIBE [21] applies adversarial learning with AMASS [30] for 3D pose extraction, while MEVA [29] improves accuracy and smoothness using a variational autoencoder to address VIBE's high acceleration error. Owing to these drawbacks, HuMoR employs a conditional variational autoencoder for robust pose estimation, while the state-of-the-art method WHAM [44] integrates 2D-3D lifting and SLAM for accurate global motion estimation. While these methods perform well on various datasets, studies like [24] highlight performance drops under significant occlusions, as most datasets contain only sparse occlusions, making models struggle with unseen heavy or prolonged occlusions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 438, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 553, + 713 + ], + "type": "text", + "content": "Human Pose Estimation under Occlusion. Despite significant progress in HPS estimation, handling occlusions remains a major challenge. This is because, in 3D pose estimation, missing depth cues make reconstruction far more ambiguous than in 2D. Early works like [5] used data augmentation with occlusion labels to enhance robustness in pose estimation. Latest methods like GLAMR [51] use a deep generative motion infiller to handle missing poses under severe occlusions and a global optimization framework to refine motion trajectories. Additionally, some methods [20, 29, 53] approach the issue of missing poses as a pose refinement problem, leveraging temporal smoothness to address it effectively. SmoothNet [53] introduced a temporal motion refinement network for refining the poses obtained from the image-based pose estimators to alleviate jitters. These methods perform well under sparse, infrequent occlusions across frames but struggle with natural, prolonged occlusions, as they lack training for such instances. All these works evaluate their methods on datasets with sparse occlusions, as no existing datasets contain significant occlusions. To address this gap, we introduce VOccl3D, a novel dataset designed specifically for 3D HPS estimation under significant occlusions." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7352" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 160, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 160, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 160, + 83 + ], + "type": "text", + "content": "3. VOCcl3D Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "In this section, we outline the key components required to construct the VOcc13D dataset. Section 3.1 details the creation of individual assets, including background scenes, human motions, and texture maps. Sections 3.2 describe the rendering process and attributes of the dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 158, + 146, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 146, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 146, + 170 + ], + "type": "text", + "content": "3.1. Dataset Assets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 319 + ], + "type": "text", + "content": "Background Scenes. In recent years, neural networks have significantly advanced 3D scene representation, enabling novel-view image synthesis. Notable approaches include Neural Radiance Fields (NeRF) [32], which learn a joint representation of geometry and appearance for novel-view synthesis. Another method, 3D Gaussian Splatting (3DGS) [19], represents a scene using 3D Gaussians and adopts a differentiable rasterizer for real-time rendering. These methods use multiview images to learn 3D representations, avoiding extensive capture setups. To develop our dataset, we use the 3DGS method to learn the 3D representation of the background scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 343 + ], + "type": "text", + "content": "Preliminaries: 3D Gaussian Splatting We represent the " + }, + { + "bbox": [ + 55, + 319, + 296, + 343 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 319, + 296, + 343 + ], + "type": "text", + "content": "-th Gaussian in 3DGS as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 350, + 295, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 350, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 104, + 350, + 295, + 366 + ], + "type": "interline_equation", + "content": "G (\\mathbf {p}) = o _ {i} e ^ {- \\frac {1}{2} \\left(\\mathbf {p} - \\boldsymbol {\\mu} _ {i}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {p} - \\boldsymbol {\\mu} _ {i}\\right)}, \\tag {1}", + "image_path": "3e5821a4bee83ea9bf94b958510d0bc568bb6a4cb4678e9c413082a9397e099c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " is a xyz location, " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "o_i \\in [0,1]" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " is the opacity modeling the ratio of radiance the Gaussian absorbs, " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " is the center/mean of the Gaussian, and the covariance matrix " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " is parameterized by the scale " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " along each of the three Gaussian axes and the rotation " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i \\in SO(3)" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i = \\mathbf{R}_i\\mathbf{S}_i\\mathbf{S}_i^\\top\\mathbf{R}_i^\\top" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": ". Each Gaussian is also paired with spherical harmonics [39] to model the radiance emitted towards various directions. During rendering, the 3D Gaussians are projected onto the image plane and form 2D Gaussians [57] with the covariance matrix " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_i^{2\\mathrm{D}} = J\\pmb{W}\\pmb{\\Sigma}_i\\pmb{W}^\\top\\pmb{J}^\\top" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{J}" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective transformation and " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{W}" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " is the viewing transformation. The color of a pixel is calculated via alpha blending the " + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 374, + 296, + 555 + ], + "type": "text", + "content": " Gaussians contributing to a given pixel:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 563, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 563, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 117, + 563, + 295, + 597 + ], + "type": "interline_equation", + "content": "C = \\sum_ {j = 1} ^ {N} c _ {j} \\alpha_ {j} \\prod_ {k = 1} ^ {j - 1} \\left(1 - \\alpha_ {k}\\right), \\tag {2}", + "image_path": "28781e6b3cc13a0589ca1e4549aeb7c0790ff6142afb3c405cce41f9ebf1b917.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "where the Gaussians are sorted from close to far, " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " is the color obtained by evaluating the spherical harmonics given viewing transform " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " is calculated from the 2D Gaussian formulation (with the covariance " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\Sigma_{j}^{2D}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": ") multiplied by its opacity " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "o_{j}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "We collected RGB videos from the large-scale open-source DL3DV dataset [27], which contains over 10,000 videos across diverse domains, including natural and outdoor settings, educational institutions, shopping complexes," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "parks, hubs, cafes, and restaurants. We selected videos based on the presence of natural occlusions, such as garlands, chairs, benches, statues, cars, and bins. Our dataset captures 3D representations of approximately 40 distinct scenes, each incorporating real-world occlusions. This approach to learning 3D representation closely resembles real-world data compared to conventional methods that rely on rendering from 3D graphic asset scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": "SMPLX body/Human Animations. We represent the human body using the SMPL-X [36] 3D human model, defined by the function " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\theta ,\\beta ,\\psi)" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": " represents the pose parameters, " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": " the shape parameters, and " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": " the facial expression parameters. This function outputs a body mesh " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{M}\\in \\mathbb{R}^{10475\\times 3}" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": " with 10,475 vertices. We sample approximately 400 SMPL-X 3D human motion models from the AMASS mocap dataset [30], which contains over 11,000 motion sequences with diverse body shapes and poses. Following previous work [48], we use the pre-trained human pose prior model VPoser [36] to assess pose difficulty. VPoser is a Variational Auto-Encoder model that evaluates a pose " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": " to be challenging if its embeddings " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": " have larger norm, i.e. " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "||\\epsilon_{\\theta}||_{2} > \\tau" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 178, + 555, + 418 + ], + "type": "text", + "content": " is empirically set to 40. We classify a challenging pose using VPoser to avoid simple movements such as walking, standing, jogging, etc. To prepare human animations for rendering, we use Blender graphics software to efficiently bake complex geometric animations over time into the Alembic format for subsequent use in rendering engines." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 427, + 556, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 556, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 556, + 584 + ], + "type": "text", + "content": "We ensured that motion sequences contained at least 180 frames to maintain a minimum animation duration of six seconds at 30 fps. For sequences exceeding 400 frames, we discarded the first 100 frames to remove static poses at the start. To ensure that human motion remains consistently within the occluded scene, we implemented boundary constraints that stop the motion if it moves beyond these limits. These conditions prevent humans from moving too far from the occlusion area. Additionally, we applied random rotations and translations to each sequence. We store the applied transformations to further calculate the effective global orientation of the human body with respect to the world coordinates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Human Textures. We sourced human body texture scans from the open-source dataset provided by the SMPLtex method [4], which estimates and manipulates the full 3D appearance of humans captured from a single image. The SMPLtex dataset provides a diverse range of human texture scans, covering various skin tones, clothing styles, and genders. To ensure sample diversity in VOcc13D, we select approximately 200 distinct texture scans. Figure 4 illustrates the different clothing textures applied to the SMPL-X human model in our dataset." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7353" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 558, + 250 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 558, + 250 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 558, + 250 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 558, + 250 + ], + "type": "image", + "image_path": "e3f841e87163ddc3c5a9d1d8f2b573a19a0a973e52ae81e67c77ebea9a35f986.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": "Figure 3. An overview of our proposed dataset creation pipeline. We generate 3D representations of natural scenes with real occlusions using 3D Gaussian Splatting [19]. Human motion sequences from the AMASS MoCap dataset [30] are processed in Blender to generate baked animation files. The generated scene representations, animation files, human texture maps, and camera extrinsic parameters are imported into the Unity rendering engine to generate video sequences of human motion under occlusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 312, + 272, + 427 + ], + "blocks": [ + { + "bbox": [ + 80, + 312, + 272, + 427 + ], + "lines": [ + { + "bbox": [ + 80, + 312, + 272, + 427 + ], + "spans": [ + { + "bbox": [ + 80, + 312, + 272, + 427 + ], + "type": "image", + "image_path": "deff0eb4f7e24634e7b727e37a11bdef5b88cd710bd7626147f2bd84321b434a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 437, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 483 + ], + "type": "text", + "content": "Figure 4. Diversity of human texture in VOccl3D. We use approximately 200 human texture maps from the SMPLtex dataset [4], encompassing diverse clothing styles, genders, skin tones, and ethnicities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 499, + 237, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 237, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 237, + 512 + ], + "type": "text", + "content": "3.2. Dataset Rendering and Attributes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 519, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 685 + ], + "type": "text", + "content": "Rendering As shown in Figure 3, we used Unity Engine to render synthetic scenes by integrating background Gaussian splats, human animations, and body texture assets. We rendered video sequences at 30 fps with a resolution of " + }, + { + "bbox": [ + 55, + 519, + 296, + 685 + ], + "type": "inline_equation", + "content": "720 \\times 720" + }, + { + "bbox": [ + 55, + 519, + 296, + 685 + ], + "type": "text", + "content": ", varying the camera field of view between 25 and 50 across different scenes. For each sequence, we stored ground-truth camera extrinsic and intrinsic parameters and saved the rendered RGB images in lossless PNG format. We positioned the camera to capture occluded views within the scene and applied motion constraints to human animation assets to maintain consistent occlusion throughout each sequence. To achieve realistic lighting, we employed Unity's built-in DirectionalLight asset to simulate natural illumination on the human mesh." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Dataset Attributes. Our proposed dataset, VOccl3D, includes over 250,000 images and 400 video sequences with" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 432 + ], + "type": "text", + "content": "a total runtime exceeding 2 hours and 30 minutes. The dataset features 40 background scenes with various occlusions, such as cars, garlands, benches, chairs, bins, and trees. Each scene contains 10 video sequences with variations in human motions and clothing textures. The rendered sequence exhibit diversity in body shapes, skin tones, and camera poses. We provide ground-truth annotations, including camera extrinsic and intrinsic parameters, pose, shape, global orientation, translation, gender, 2D keypoints, and a binary occlusion label for each keypoint." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 516 + ], + "type": "text", + "content": "Different Modalities. In addition to 3D pose and shape annotations, VOccl3D provides annotations for multiple modalities, including human silhouettes, body-part segmentation, 2D keypoint estimation, and human bounding box detection under occlusion. Researchers can use our dataset to train and evaluate methods designed to handle occlusion across these modalities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 458, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 458, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 458, + 534 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 541, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 555, + 649 + ], + "type": "text", + "content": "In this section, we highlight the need and significance of our proposed VOccl3D dataset. We fine-tune the state-of-the-art pose estimation methods using our dataset and report the qualitative and quantitative results. Our results demonstrate improved performance on the HPS estimation task (Section 4.1) using both our proposed synthetic dataset and real-world datasets with occlusions. In Section 4.2, we show the enhanced performance of the human object detector and its impact on HPS estimation under occlusion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 659, + 503, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 503, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 503, + 672 + ], + "type": "text", + "content": "4.1. Human Pose and Shape Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Dataset and Implementation Setup. We use approximately 200k and 50k images for training and testing respectively in our proposed VOcc13D dataset. We fine-tune" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7354" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 189 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 189 + ], + "type": "table", + "html": "
MethodHard-OcclusionMedium-OcclusionLow-Occlusion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]192.22114.35247.41121.7078.56158.4698.8267.64126.92
BEDLAM-HMR [3]167.35102.92214.39102.5568.13134.5986.5554.48110.15
BEDLAM-CLIFF [3]154.8699.53199.9590.9765.03119.6374.9552.6596.60
HMR2.0 [11]169.71100.49215.17113.8871.78145.6288.5359.08114.39
STRIDE with BEDLAM-CLIFF [24]155.64100.44-91.1465.38-75.0253.21-
WHAM [44]152.15102.14177.07110.9776.81127.4593.9066.68106.51
VOcc13D-B-CLIFF136.3489.94175.9282.4858.78106.8469.4646.3288.19
STRIDE with VOcc13D-B-CLIFF136.4390.28-82.3758.98-69.6546.86-
", + "image_path": "0d914f49d8f267a5e9bcebb88016548bd3faadd0688dacef6d54c4c358ce31be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 83, + 251, + 531, + 352 + ], + "blocks": [ + { + "bbox": [ + 55, + 197, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 555, + 242 + ], + "type": "text", + "content": "Table 1. 3D HPS estimation results on the test-split of VOcccl3D. The results show that VOcccl3D-B-CLIFF and STRIDE, when using pseudo-labels from VOcccl3D-B-CLIFF, significantly outperform other image- and video-based HPS estimation methods across hard, medium, and low occlusion categories. As expected, all HPS estimation methods exhibit a decline in performance as occlusion severity increases from low to medium to high. For evaluation, we use ground-truth bounding boxes. The best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 251, + 531, + 352 + ], + "lines": [ + { + "bbox": [ + 83, + 251, + 531, + 352 + ], + "spans": [ + { + "bbox": [ + 83, + 251, + 531, + 352 + ], + "type": "table", + "html": "
Method3DPWOcclType1-3DPWOcclType2-3DPW
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
CLIFF [25]73.946.487.698.1562.27118.4799.4962.16119.82
BEDLAM-HMR [3]79.047.693.1108.6266.53128.05106.1964.05125.66
BEDLAM-CLIFF [3]72.046.685.098.7164.26117.4196.8061.32115.33
HMR2.0 [11]81.254.3143.7103.4069.66164.5599.0166.17158.79
VOccl3D-B-CLIFF72.047.384.595.8963.43114.2894.3660.44112.01
VOccl3D-CLIFF71.1045.9884.2595.1761.83114.9593.7459.66112.59
", + "image_path": "52bc5b8701d69d0b9c4faa4ce50f5a99bf2cf6319d7a4bdac953c5de68805907.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 359, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 555, + 404 + ], + "type": "text", + "content": "Table 2. 3D HPS estimation results on 3DPW, OcclType1-3DPW, and OcclType2-3DPW. Since 3DPW is a real-world dataset with minimal occlusions, VOccl3D-CLIFF outperforms all other methods but shows only a marginal improvement over BEDLAM-CLIFF. However, on OcclType1-3DPW and OcclType2-3DPW, which contain significant occlusions on real-world dataset, our method demonstrates a notable improvement over existing approaches. The best results are noted in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 422, + 294, + 567 + ], + "blocks": [ + { + "bbox": [ + 56, + 422, + 294, + 567 + ], + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 567 + ], + "type": "image", + "image_path": "2a2fb20081034a910d905c0023aa4060635f6fec6e8355823f136332f4cd760e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 574, + 295, + 641 + ], + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 641 + ], + "type": "text", + "content": "Figure 5. Fine-tuning an off-the-shelf pose estimator. We fine-tune a CLIFF [25] pose estimation model using our proposed VOcc13D dataset. We observe a significant improvement in the estimated mesh when using a fine-tuned CLIFF model using our dataset. Note: The pose estimator can be any off-the-shelf pose estimator, here we show results with the CLIFF model [25]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "two versions of state-of-the-art model, CLIFF (trained on Human3.6M [15], MPI-INF-3DHP [31], and COCO [26] dataset) and BEDLAM-CLIFF (trained on BEDLAM and AGORA datasets) using our rendered VOcc13D dataset, resulting in VOcc13D-CLIFF and VOcc13D-B-CLIFF, re" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "content": "spectively (refer to Figure 5). To benchmark performance, we evaluate on the test split of VOccl3D, comparing against multiple image and video based baselines, including CLIFF [25], BEDLAM-HMR [3], BEDLAM-CLIFF [3], HMR2.0 [11], STRIDE [24], and WHAM [44]. We report the mean performance by evaluating the model using a fivefold cross-validation on our dataset. To quantify occlusion levels, we annotate each 3D joint with a binary occlusion label, marking it as occluded if its corresponding 2D keypoint lies within the ground-truth segmentation mask of the visible human region. Based on the number of visible keypoints (out of 22 total), we categorize images into three occlusion levels: hard occlusion (4-9 visible keypoints), medium occlusion (10-15 visible keypoints), and low occlusion (16-20 visible keypoints)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Beyond evaluations on synthetic datasets, we assess the performance of VOccl3D fine-tuned models on real-world datasets, including 3DPW [46] and OCMotion [13]. It is noteworthy that these datasets do not contain significant occlusions. Hence, we introduce two occlusion-augmented variants of 3DPW to evaluate robustness under severe occlusions. In OcclType1-3DPW, we overlay a black patch on a randomly selected keypoint, while in OcclType2-3DPW, we occlude two random keypoints. Further details on oc" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7355" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 386 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 386 + ], + "type": "image", + "image_path": "2a712e5eb8c5fe1284bfca9681cfe5c6b9be67f71f4352d369042bb5f9a0dc1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 394, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 555, + 440 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of HPS estimation methods on VOcc3D dataset. The first and second column shows RGB image and ground-truth human mesh. Column third and fourth compare HPS estimation using the BEDLAM-CLIFF [3] and HMR2.0 [11] methods. The final column (VOcc3D-B-CLIFF) presents results obtained by fine-tuning the CLIFF model on the VOcc3D dataset. These results demonstrate superior performance, particularly in scenarios with heavy occlusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 451, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 475 + ], + "type": "text", + "content": "clusion variants of 3DPW and implementation details are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Results. We present benchmarking results on the test split of our proposed VOcc13D dataset in Table 1. Our finetuned VOcc13D-B-CLIFF significantly outperforms existing state-of-the-art image and video-based methods across all occlusion categories. Notably, the performance of all previous methods degrades as occlusion severity increases, highlighting the inherent challenges of HPS estimation under occlusions. We also compare our approach with the plug-and-play method STRIDE [24], which performs 3D pose estimation and reports only MPJPE and PA-MPJPE errors. Our results show that STRIDE achieves superior performance when leveraging pseudo-labels from the VOcc13D-B-CLIFF model compared to those from the original BEDLAM-CLIFF model, further validating the effectiveness of our dataset. In Table 2 and Table 3, we report performance on the 3DPW [45] and OCMotion [14] datasets, respectively. We observe that both VOcc13D-B-CLIFF and VOcc13D-CLIFF outperform all the previous methods, demonstrating their effectiveness in real-world" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 451, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 668 + ], + "type": "text", + "content": "scenarios. However, since these datasets contain minimal occlusions, the improvements over CLIFF remain marginal. To further assess robustness under heavy occlusion, we evaluate on OcclType1-3DPW and OcclType2-3DPW, real-world datasets with significant occlusions. Table 2 demonstrates that our fine-tuned model achieves substantial improvements, underscoring its robustness in occluded environments. Table 3 presents results on the OCMotion dataset, which features lower levels of occlusion. We can observe that VOcc13D-CLIFF outperforms all comparison methods and demonstrates performance comparable to CLIFF [25]. This is expected, as both models are pretrained on large-scale real-world datasets captured in controlled lab environments similar to OCMotion. Figure 6 presents the qualitative results on VOcc13Ddataset, showcasing the improvement of our fine-tuned model against other state-of-the-art HPS estimation methods. Additional results are provided in supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 672, + 555, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 555, + 685 + ], + "type": "text", + "content": "4.2. Impact of Human Detector on HPS Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Dataset and Implementation Setup. Human object detectors play a crucial role in HPS estimation tasks, particularly" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7356" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 271, + 163 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 271, + 163 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 271, + 163 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 271, + 163 + ], + "type": "table", + "html": "
MethodMPJPEOCMotion PA-MPJPEPVE
CLIFF [25]64.1540.1679.80
BEDLAM-HMR [3]73.9641.9492.70
BEDLAM-CLIFF [3]66.8041.7983.86
HMR2.0 [11]69.9443.0087.07
VOcc13D-B-CLIFF65.9640.5981.96
VOcc13D-CLIFF64.2939.6478.56
", + "image_path": "650cb8f892498b5d08d321b17eb1f2477138918f2b94bc2747cb1f770de0e04e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 76, + 251, + 278, + 299 + ], + "blocks": [ + { + "bbox": [ + 55, + 171, + 294, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 294, + 237 + ], + "type": "text", + "content": "Table 3. Qualitative results of HPS estimation on OCMotion. Our results show that VOcc13D-CLIFF outperforms all HPS estimation methods except CLIFF [25]. Both CLIFF and VOcc13D-CLIFF exhibit comparable performance, as both are pre-trained on real-world datasets captured in controlled lab environments, similar to OCMotion [14]. The best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 251, + 278, + 299 + ], + "lines": [ + { + "bbox": [ + 76, + 251, + 278, + 299 + ], + "spans": [ + { + "bbox": [ + 76, + 251, + 278, + 299 + ], + "type": "table", + "html": "
Method3DPWOCMotion
mAP50mAP75mAP50mAP75
YOLO1158.9947.1498.8491.80
VOccl3D-YOLO1159.8948.2699.1091.95
", + "image_path": "2d7e59b55027f5007112083acc6a7b379b6b158804c4733a3eff707bf6745e46.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 483 + ], + "type": "text", + "content": "when inferring from in-the-wild RGB images [3, 24, 25]. Ideally, HPS estimation performs optimally with ground-truth bounding boxes; however, most human object detectors struggle significantly under occlusion. To address this, we fine-tune the recent YOLO11 detector [16] on the combined training split of VOccl3D and MS COCO, referring to the fine-tuned model as VOccl3D-YOLO11. We show evaluations on 3DPW [45] and OCMotion [14] datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 616 + ], + "type": "text", + "content": "Results. Table 4 compares the detection performance of the pre-trained YOLO11 and VOccl3D-YOLO11 models on the real-world 3DPW [45] and OCMotion [14] datasets. Since OCMotion is a single-human dataset with minimal occlusions, the detector achieves high mAP scores. This is likely because most body parts remain visible, making box detection easier. However, due to the presence of multiple individuals in 3DPW, the overall mAP scores remain lower. In short, VOccl3D-YOLO11 shows significant improvement on datasets with high occlusions. We provide qualitative results in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Table 5 further evaluates the impact of human detection on HPS estimation. The first, second, and third rows represent the performance of VOccl3D-CLIFF when human detections are sourced from ground truth, YOLO11, and VOccl3D-YOLO11, respectively. As expected, the best performance is achieved when using ground-truth detections. However, fine-tuning the detector with VOccl3D significantly improves HPS estimation compared to detections" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 554, + 140 + ], + "blocks": [ + { + "bbox": [ + 55, + 307, + 295, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 373 + ], + "type": "text", + "content": "Table 4. Quantitative results of YOLO11 on 3DPW and OC-Motion datasets. The first row presents the human object detection performance of the pre-trained YOLO11 model, while the second row shows the results after fine-tuning with the VOccl3D dataset. The best results are in bold. We observe an improvement in detection accuracy across both datasets after fine-tuning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 554, + 140 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 140 + ], + "type": "table", + "html": "
Method3DPWOCMotion
MPJPEPA-MPJPEPVEMPJPEPA-MPJPEPVE
VOccl3D-CLIFF w/GT71.1045.9884.2564.2939.6478.56
VOccl3D-CLIFF w/YYOLO11116.5263.35139.7467.1641.3083.00
VOccl3D-CLIFF w/VOccl3D-YYOLO11114.8562.66137.1966.6541.1582.40
", + "image_path": "88d95860ab880f0cdb8a657e5e8af439815f88839522115a38cc84fce76e0faa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 148, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 553, + 258 + ], + "type": "text", + "content": "Table 5. 3D HPS estimation results using the VOcc13D-CLIFF model with different bounding box sources. This table presents HPS estimation performance on the 3DPW and OC-Motion datasets. The first row reports results using ground-truth bounding boxes, achieving the highest accuracy. The second and third rows show performance when detections are obtained from the pre-trained YOLO11 and fine-tuned VOcc13D-YOLO11 models, respectively. Notably, the third row demonstrates a significant improvement over the second, highlighting the impact of detections on HPS estimation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 273, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 553, + 308 + ], + "type": "text", + "content": "from the pre-trained YOLO11. These results underscore the critical role of human detection quality in enhancing HPS performance under occlusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 310, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 554, + 430 + ], + "type": "text", + "content": "Our experiments show that fine-tuning an existing HPS estimation model with a large synthetic dataset containing occlusions enhances its performance on both real and synthetic datasets. We observe significant improvements over baseline methods, particularly in scenarios with heavy occlusions. Additionally we demonstrate that the poor results are not solely due to HPS estimations errors but also stem from failures in bounding box detection. The VOcc13D dataset proves effective in refining pose estimation methods and bounding box predictions under occlusions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 388, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 713 + ], + "type": "text", + "content": "We introduce VOccl3D, a novel large-scale video dataset with synthetic humans in real-world scenes under occlusions for 3D human pose and shape estimation. By leveraging a rendering-based approach, VOccl3D eliminates the need for costly and labor-intensive data collection while providing a diverse and realistic dataset for occlusion-aware research. Through extensive qualitative and quantitative evaluations, we demonstrate that fine-tuning existing HPS estimation methods with VOccl3D significantly enhances their performance on real-world datasets with occlusions and on the test split of VOccl3D dataset. Furthermore, we improve the human object detector, YOLO11 under occluded conditions using VOccl3D, highlighting the impact of detections in achieving robust HPS estimation in occluded scenarios. Beyond HPS estimation, VOccl3D serves as a comprehensive benchmark for evaluating methods across multiple occlusion-aware tasks, including human body-part segmentation, 2D/3D pose estimation, and human bounding box detection. Hence, VOccl3D sets a new standard for occluded human benchmarks, offering a valuable dataset for advancing occlusion-robust research." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7357" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 236 + ], + "type": "text", + "content": "This work was partially supported by USDA NRI grant 2021-67022-33453, NSF grants CMMI-2326309 and CNS-2312395. This research was also partially supported by the Office of the Director of National Intelligence (ODNI), specifically through the Intelligence Advanced Research Projects Activity (IARPA), under contract number [2022-21102100007]. The views and conclusions in this research reflect those of the authors and should not be construed as officially representing the policies, whether explicitly or implicitly, of ODNI, IARPA, or the U.S. Government. Nevertheless, the U.S. Government retains the authorization to reproduce and distribute reprints for official government purposes, regardless of any copyright notices included." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 255, + 115, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 115, + 268 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 115, + 268 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 275, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 275, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 275, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 275, + 296, + 319 + ], + "type": "text", + "content": "[1] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting temporal context for 3d human pose estimation in the wild. In Computer Vision and Pattern Recognition (CVPR), 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 321, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 296, + 365 + ], + "type": "text", + "content": "[2] Peter Bauer, Arij Bouazizi, Ulrich Kressel, and Fabian B. Flohr. Weakly supervised multi-modal 3d human body pose estimation for autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 366, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 366, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 366, + 295, + 399 + ], + "type": "text", + "content": "[3] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 443 + ], + "type": "text", + "content": "[4] Dan Casas and Marc Comino-Trinidad. SMPLitex: A Generative Model and Dataset for 3D Human Texture Estimation from Single Image. In British Machine Vision Conference (BMVC), 2023. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 296, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 488 + ], + "type": "text", + "content": "[5] Yu Cheng, Bo Yang, Bo Wang, Yan Wending, and Robby Tan. Occlusion-aware networks for 3d human pose estimation in video. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 723-732, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 490, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 296, + 555 + ], + "type": "text", + "content": "[6] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 612 + ], + "type": "text", + "content": "[7] Mickael Cormier, Aris Clepe, Andreas Specker, and Jürgen Beyerer. Where are we with human pose estimation in real-world surveillance? In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 591-601, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "type": "text", + "content": "[8] R James Cotton. Posepipe: Open-source human pose estimation pipeline for rehabilitation research. Archives of Physical Medicine and Rehabilitation, 103(12):e161-e162, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "type": "text", + "content": "[9] Yufeng Cui and Yimei Kang. Multi-modal gait recognition via effective spatial-temporal feature fusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17949-17957, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 714 + ], + "type": "text", + "content": "[10] Yang Fu, Shibei Meng, Saihui Hou, Xuecai Hu, and Yongzhen Huang. Gpgait: Generalized pose-based gait" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "recognition. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19538-19547, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 554, + 151 + ], + "type": "text", + "content": "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 217 + ], + "type": "text", + "content": "[12] Charlie Hewitt, Fatemeh Saleh, Sadegh Aliakbarian, Lohit Petikam, Shideh RezaEIFar, Louis Florentin, Zafiirah Hosenie, Thomas J Cashman, Julien Valentin, Darren Cosker, et al. Look ma, no markers: Holistic performance capture without the hassle. ACM Transactions on Graphics, 43(6): 1-12, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[13] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[14] Buzhen Huang, Tianshu Zhang, and Yangang Wang. Object-occluded human shape and pose estimation with probabilistic latent consistency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):5010-5026, 2023. 2, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 57, + 275, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[15] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 545, + 376 + ], + "type": "text", + "content": "[16] Glenn Jocher and Jing Qiu. Ultralytics yolo11, 2024. 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 377, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 431 + ], + "type": "text", + "content": "[17] Sardor Juraev, Akash Ghimire, Jumabek Alikhanov, Vijay Kakani, and Hakil Kim. Exploring human pose estimation and the usage of synthetic data for elderly fall detection in real-world surveillance. IEEE Access, 10:94249-94261, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 477 + ], + "type": "text", + "content": "[18] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7122-7131, 2017. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[20] Do-Yeop Kim and Ju-Yong Chang. Attention-based 3d human pose sequence refinement network. Sensors, 21(13), 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "text", + "content": "[21] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 646 + ], + "type": "text", + "content": "[22] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In Proc. International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[23] Farnoosh Koleini, Muhammad Usama Saleem, Pu Wang, Hongfei Xue, Ahmed Helmy, and Abbey Fenwick. Biopose: Biomechanically-accurate 3d pose estimation from monocular videos. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6330–6339. IEEE, 2025. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7358" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] Rohit Lal, Saketh Bachu, Yash Garg, Arindam Dutta, Calvin-Khang Ta, Dripta S. Raychaudhuri, Hannah Dela Cruz, M. Salman Asif, and Amit K. Roy-Chowdhury. Stride: Single-video based temporally continuous occlusion robust 3d pose estimation, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[25] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 297 + ], + "type": "text", + "content": "[27] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 341 + ], + "type": "text", + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "text", + "content": "[29] Zhengyi Luo, S. Alireza Golestaneh, and Kris M. Kitani. 3d human motion estimation via motion compression and refinement. In Proceedings of the Asian Conference on Computer Vision (ACCV), 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[30] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[31] Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 International Conference on 3D Vision (3DV), pages 506-516, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 542 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 588 + ], + "type": "text", + "content": "[33] Sara Moccia, Lucia Migliorelli, Virgilio Carnielli, and Emanuele Frontoni. Preterm infants' pose estimation with spatio-temporal features. IEEE Transactions on Biomedical Engineering, 67(8):2370-2380, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[34] Aron Monszpart, Paul Guerrero, Duygu Ceylan, Ersin Yumer, and Niloy J. Mitra. imapper: interaction-guided scene mapping from monocular videos. ACM Trans. Graph., 38(4), 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Priyanka Patel, Chun-Hao P Huang, Joachim Tesch, David T Hoffmann, Shashank Tripathi, and Michael J Black. Agora: Avatars in geography optimized for regression analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13468-13478, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 712 + ], + "type": "text", + "content": "[36] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "text", + "content": "[37] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7745-7754, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 554, + 205 + ], + "type": "text", + "content": "[38] Miroslav Purkrábek and Jiri Matas. Improving 2d human pose estimation across unseen camera views with synthetic data. arXiv preprint arXiv:2307.06737, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 555, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 249 + ], + "type": "text", + "content": "[39] Ravi Ramamoorthi and Pat Hanrahan. An efficient representation for irradiance environment maps. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 497-500, 2001. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 554, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 554, + 294 + ], + "type": "text", + "content": "[40] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 295, + 554, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 554, + 347 + ], + "type": "text", + "content": "[41] Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, page 1530-1538. JMLR.org, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "text", + "content": "[42] Tim Salzmann, Hao-Tien Lewis Chiang, Markus Ryll, Dorsa Sadigh, Carolina Parada, and Alex Bewley. Robots that can see: Leveraging human pose for trajectory prediction. IEEE Robotics and Automation Letters, 8(11):7090-7097, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 554, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 446 + ], + "type": "text", + "content": "[43] Saurabh Sharma, Pavan Teja Varigonda, Prashast Bindal, Abhishek Sharma, and Arjun Jain. Monocular 3d human pose estimation by generation and ordinal ranking. In The IEEE International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 502 + ], + "type": "text", + "content": "[44] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2070-2080, 2023. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 554, + 557 + ], + "type": "text", + "content": "[45] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), 2018. 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "text", + "content": "[46] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European conference on computer vision (ECCV), pages 601-617, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 615, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 554, + 658 + ], + "type": "text", + "content": "[47] Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 659, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 554, + 712 + ], + "type": "text", + "content": "[48] Zhenzhen Weng, Laura Bravo-Sanchez, and Serena Yeung-Levy. Diffusion-hpc: Synthetic data generation for human mesh recovery in challenging domains. In 2024 International Conference on 3D Vision (3DV), pages 257-267. IEEE, 2024. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7359" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 552 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[49] Zhitao Yang, Zhongang Cai, Haiyi Mei, Shuai Liu, Zhaoxi Chen, Weiye Xiao, Yukun Wei, Zhongfei Qing, Chen Wei, Bo Dai, et al. Synbody: Synthetic dataset with layered human models for 3d human perception and modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20282-20292, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[50] Frank Yu, Mathieu Salzmann, Pascal Fua, and Helge Rhodin. Pcls: Geometry-aware neural reconstruction of 3d pose with perspective crop layers. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9060-9069, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 250 + ], + "type": "text", + "content": "[51] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 306 + ], + "type": "text", + "content": "[52] Andrei Zanfir, Mihai Zanfir, Alex Gorban, Jingwei Ji, Yin Zhou, Dragomir Anguelov, and Cristian Sminchisescu. HUM3DIL: Semi-supervised multi-modal 3d humanpose estimation for autonomous driving. In 6th Annual Conference on Robot Learning, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[53] Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, and Qiang Xu. Smoothnet: A plug-and-play network for refining human poses in videos. In European Conference on Computer Vision. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "text", + "content": "[54] Tianshu Zhang, Buzhen Huang, and Yangang Wang. Object-occluded human shape and pose estimation from a single color image. In IEEE Conference on Computer Vision and Pattern Recognition, (CVPR), 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 451 + ], + "type": "text", + "content": "[55] Yi Zhang, Pengliang Ji, Adam Kortylewski, Angtian Wang, Jieru Mei, and Alan L Yuille. 3D-Aware Neural Body Fitting for Occlusion Robust 3D Human Pose Estimation. In The IEEE/CVF International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 508 + ], + "type": "text", + "content": "[56] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G. Derpanis, and Kostas Daniilidis. Sparseness meets deepness: 3d human pose estimation from monocular video. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4966-4975, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "text", + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 371-378, 2001. 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7360" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_content_list.json b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f64ea2c948400390d1b089998ab87952adf5098c --- /dev/null +++ b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_content_list.json @@ -0,0 +1,1452 @@ +[ + { + "type": "text", + "text": "VPO: Aligning Text-to-Video Generation Models with Prompt Optimization", + "text_level": 1, + "bbox": [ + 114, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiale Cheng $^{1,2*}$ , Ruiliang Lyu $^{2}$ , Xiaotao Gu $^{2}$ , Xiao Liu $^{2,3}$ , Jiazheng Xu $^{2,3}$ , Yida Lu $^{1,2*}$ , Jiayan Teng $^{2,3}$ , Zhuoyi Yang $^{2,3}$ , Yuxiao Dong $^{3}$ , Jie Tang $^{3}$ , Hongning Wang $^{1}$ , Minlie Huang $^{1\\dagger}$", + "bbox": [ + 122, + 179, + 879, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Conversational Artificial Intelligence (CoAI) Group, Tsinghua University \n2Zhipu AI 3The Knowledge Engineering Group (KEG), Tsinghua University \nchengjl23@mails.tsinghua.edu.cn, aihuang@tsinghua.edu.cn", + "bbox": [ + 184, + 233, + 808, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video generation models have achieved remarkable progress in text-to-video tasks. These models are typically trained on text-video pairs with highly detailed and carefully crafted descriptions, while real-world user inputs during inference are often concise, vague, or poorly structured. This gap makes prompt optimization crucial for generating high-quality videos. Current methods often rely on large language models (LLMs) to refine prompts through in-context learning, but suffer from several limitations: they may distort user intent, omit critical details, or introduce safety risks. Moreover, they optimize prompts without considering the impact on the final video quality, which can lead to suboptimal results. To address these issues, we introduce VPO, a principled framework that optimizes prompts based on three core principles: harmlessness, accuracy, and helpfulness. The generated prompts faithfully preserve user intents and, more importantly, enhance the safety and quality of generated videos. To achieve this, VPO employs a two-stage optimization approach. First, we construct and refine a supervised fine-tuning (SFT) dataset based on principles of safety and alignment. Second, we introduce both text-level and video-level feedback to further optimize the SFT model with preference learning. Our extensive experiments demonstrate that VPO significantly improves safety, alignment, and video quality compared to baseline methods. Moreover, VPO shows strong generalization across video generation models. Furthermore, we demonstrate that VPO could outperform and be combined with RLHF methods on video generation models, underscoring the effectiveness of VPO in aligning video generation models. Our code and data are publicly available at https://github.com/thu-coai/VPO.", + "bbox": [ + 89, + 356, + 485, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 321, + 643, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in text-to-video generation models have significantly improved the ability to produce high-quality video content [4, 15, 18, 28, 29]. These models typically rely on extensive, high-quality video-text annotated datasets, where detailed labels capture specific video content. While training descriptions are detailed and carefully crafted, real-world user inputs are often brief, ambiguous, or poorly structured. This critical gap between model training and inference can lead to suboptimal outputs, making it crucial to optimize and refine user inputs to generate desired high-quality video outputs (Figure 1).", + "bbox": [ + 511, + 349, + 906, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current approaches to optimizing user inputs for video generation mostly utilize large language models (LLMs) with in-context learning [15, 28, 29]. However, these methods suffer from several limitations:", + "bbox": [ + 511, + 517, + 906, + 575 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Safety concerns: Existing methods do not explicitly ensure that optimized prompts remain safe, potentially leading to inappropriate or harmful video content.", + "- Imprecise refinement: Current methods may unintentionally alter the user's intent or introduce biases, leading to outputs that deviate from the original query.", + "- Neglecting final video quality: Current approaches optimize prompts to be semantically richer, but do not explicitly consider how the refined prompt impacts the actual generated video, limiting their practical effectiveness." + ], + "bbox": [ + 511, + 579, + 905, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose a set of alignment principles for video prompt optimization, inspired by the HHH (Harmless, Honest, Helpful) principle in LLMs [1]. An aligned prompt optimization model should adhere to the following:", + "bbox": [ + 511, + 732, + 905, + 808 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Harmless: The refined prompt should avoid any harmful content, such as depictions of blood or violence.", + "- Accurate: The optimized prompt should be precisely aligned with the user input, except for safety issues.", + "- Helpful: The refined prompt should be detailed and descriptive, and can help the video generation model pro" + ], + "bbox": [ + 511, + 810, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done when JC and YL interned at Zhipu AI.", + "bbox": [ + 107, + 875, + 380, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15636", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fefd15023042662f688c12aabf59840b71143e0cba7117c58820e96a620edcaf.jpg", + "image_caption": [ + "Figure 1. (Upper) The critical gap between the training and inference stages in video generation models. During training, video descriptions are detailed, descriptive, and well-crafted, while user inputs in the inference stage are often concise, unclear, and poorly structured, creating a mismatch that hampers model performance. (Lower) Comparison of VPO with traditional prompt optimization methods. Traditional methods rely on the in-context learning capabilities of LLMs, which can lead to issues such as safety issues, misalignment, and low-quality videos. In contrast, VPO serves as a harmless, accurate, and helpful prompt optimizer for high-quality video generation." + ], + "image_footnote": [], + "bbox": [ + 116, + 85, + 883, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "duce high-quality video content.", + "bbox": [ + 104, + 468, + 320, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these principles, we introduce VPO, a principle-driven framework for optimizing user inputs for text-to-video generation, which integrates two critical stages: principle-based supervised fine-tuning (SFT) and multi-feedback preference optimization. In the SFT stage, we first leverage the in-context learning capabilities of LLMs to construct an initial SFT dataset. This data is then improved through the application of principle-based critique and refinement, ensuring the generated prompts are safe, accurately aligned with the original query, and sufficiently detailed. The resulting fine-tuned model poses a foundational ability to generate prompts following the proposed principles. In the subsequent preference optimization phase, we incorporate both text-level feedback and video-level reward signals to further improve the SFT model. By combining these feedback sources, we construct preference data that prioritizes harmlessness, accuracy, and helpfulness in the optimized prompt. Using Direct Preference Optimization (DPO), we train the model to generate prompts that consistently result in safer, more precise, and higher-quality video outputs. This two-stage framework explicitly integrates alignment principles into the prompt optimization process, bridging the critical gap between model training and real-world inference in video generation.", + "bbox": [ + 89, + 488, + 483, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We have conducted experiments on several popular video generation models, including CogVideoX [28] and Open-Sora [29]. Through extensive experiments, we", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrate the effectiveness of VPO in generating harmless, accurate, and helpful prompts for high-quality video content, outperforming traditional prompt rewriting methods. Notably, on CogVideoX, VPO improves the win rate of $37.5\\%$ over original user queries and $14\\%$ over the official prompt optimization method in human evaluation. Moreover, VPO significantly reduces the unsafety rate, enhancing the overall harmlessness of video generation models. We also highlight the necessity of multi-feedback preference optimization, demonstrating its crucial role in refining prompts for improved video generation. Furthermore, we demonstrate the potential of VPO as an RLHF method. VPO not only surpasses Diffusion DPO [25], but also brings additional benefits over Diffusion DPO. Additionally, VPO exhibits good generalization capabilities across different models—when trained on CogVideoX, it enhances the performance of Open-Sora 1.2, indicating the potential for training a general prompt optimizer for diverse video generation models.", + "bbox": [ + 511, + 468, + 906, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 531, + 760, + 857, + 773 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the principle of the alignment of video prompt optimizer: harmless, accurate, and helpful. Based on this, we introduce VPO, a systematic framework that integrates principle-based supervised fine-tuning and multi-feedback preference optimization, to build an aligned video prompt optimizer.", + "- We demonstrate the effectiveness of VPO for improving query refinement accuracy, enhancing the safety of" + ], + "bbox": [ + 513, + 780, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15637", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generated videos, and boosting the overall performance of video generation models. Furthermore, VPO exhibits good generalization across different video generation models.", + "bbox": [ + 102, + 90, + 480, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We demonstrate that VPO is a novel and competitive alignment method for text-to-video models, outperforming Diffusion DPO and providing orthogonal benefits.", + "bbox": [ + 91, + 151, + 482, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 212, + 232, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Video Generation", + "text_level": 1, + "bbox": [ + 89, + 238, + 261, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the widespread application of diffusion models in the domain of image generation [20, 21], research efforts have increasingly shifted towards video generation models [3, 8, 23]. Video Diffusion Models [11] firstly extend image diffusion models to video generation by employing a 3D UNet architecture. Text2Video-Zero [14] involves motion dynamics in image generation models to enable zero-shot video generation. However, these approaches encounter significant challenges in generating long videos. Subsequent studies, such as Stable Video Diffusion [2] and VideoCrafter [5], leverage large-scale pre-training on high-quality video datasets, yet their scalability remains constrained due to the limitation of architectures like UNet. By adopting Transformers [24] as the backbone of diffusion models, the scalable architecture of DiT [17] has substantially enhanced the capability of video generation. Building on DiT, several impressive video generation models have emerged [4, 15, 18, 28]. Nonetheless, the training data for these video generation models typically consists of detailed video descriptions, whereas user inputs during inference are often brief, unstructured, or even ambiguous in intent. Bridging this gap is critical for improving the quality of generated videos, underscoring the importance of training a robust prompt optimization model.", + "bbox": [ + 89, + 261, + 482, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prompt Optimization", + "text_level": 1, + "bbox": [ + 89, + 636, + 292, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt optimization techniques have long been a pivotal research problem for generation models. Early work on automatic prompt optimization can be traced back to AutoPrompt [22], which utilizes gradient-guided search to automatically create prompt templates. With the rapid development of LLMs, there has been a significant increase in studies leveraging LLMs for automatic prompt optimization [6, 30]. Similarly, existing video prompt optimization methods employ LLMs to refine user inputs, typically through in-context learning [15, 28, 29]. However, these methods rely solely on the inherent capabilities of LLMs, without considering the alignment with user query and the quality of generated videos, thus often resulting in suboptimal video outputs. Prompt-A-Video [13] incorporates image and video reward models during the optimization process to enhance the quality of generated videos, but it fails to ac", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "count for text-level alignment, which is essential for ensuring the safety and accuracy of rewritten prompts. To address these limitations, we propose VPO, a systematic framework for constructing a video prompt optimization model that can safely and accurately capture user intent, and assist video generation models in generating high-quality videos.", + "bbox": [ + 511, + 90, + 903, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 196, + 604, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce VPO, an automated framework designed to bridge the gap between real-world user inputs and the well-crafted text descriptions used during model training. Unlike existing approaches that rely solely on the in-context learning capabilities of LLMs, VPO follows a principle-driven approach, ensuring that prompt refinements are harmless, accurate, and helpful. By incorporating these core principles into both supervised fine-tuning and preference optimization, VPO effectively transforms concise, unclear, or unsafe user inputs into detailed, well-structured, and safe prompts that significantly enhance video generation quality.", + "bbox": [ + 511, + 223, + 906, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 511, + 400, + 702, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of VPO is illustrated in Figure 2. Our framework consists of two key stages: Principle-Based SFT ( $\\S 3.2$ ) and Multi-Feedback Preference Optimization ( $\\S 3.3$ ). In the first stage, Principle-Based SFT, we construct a high-quality dataset to train an initial model that refines user inputs while adhering to our three core principles. This stage provides the model with the foundational capabilities to optimize user input into a harmless, accurate, and helpful prompt for text-to-video generation. In the second stage, Multi-Feedback Preference Optimization, we further improve the SFT model using both text-level and video-level feedback. Text-level feedback helps refine prompts to be well-aligned with user intent while maintaining harmless. Meanwhile, video-level feedback ensures that the generated prompts can lead to high-quality video outputs, effectively enhancing the performance of the text-to-video model.", + "bbox": [ + 511, + 422, + 906, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Principle-Based SFT", + "text_level": 1, + "bbox": [ + 511, + 675, + 709, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first construct a high-quality SFT dataset comprising user query and optimized prompt pairs to initialize the prompt optimizer with the ability to generate prompts that align with our guiding principles.", + "bbox": [ + 511, + 696, + 903, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query Curation. To build the SFT dataset, we begin by collecting a comprehensive set of user queries. We utilize the VidProM dataset [26], which contains over one million real-world text-to-video queries. We adopt its deduplicated version and apply additional filtering steps to enhance the data. First, we perform rule-based quality filtering, considering factors such as keywords, special characters, and query length. To ensure query diversity, we employ the", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15638", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/31b9860833d98c369d137d0de96b2da5aaaebbe14c531057a11b3ce3de45db2c.jpg", + "image_caption": [ + "Figure 2. The overall framework of VPO. VPO consists of two main steps: Principle-Based SFT and Multi-Feedback Preference Optimization. By refining user queries into harmless, accurate, and helpful prompts, VPO acts as a bridge between users and video generation models, ultimately enhancing the quality of generated videos." + ], + "image_footnote": [], + "bbox": [ + 112, + 87, + 364, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5cee41c894a273e33c8e58906d0b54b5ac3ca38b6046145c7b30a67e1bc1fcca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 87, + 883, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "self-BLEU [31] metric to filter out overly similar queries. Additionally, since we aim to develop a prompt optimizer that prioritizes safety, we incorporate safety-related queries into our dataset. Specifically, we extract prompts labeled as unsafe in the original dataset. However, as Detoxify [9] introduces labeling errors, we re-evaluate these safety-related queries using a more robust LLM. Following this data curation process, we finalize a dataset consisting of approximately 18k general queries and 2k safety-related queries for training, where 10k are used for SFT and 10k for DPO.", + "bbox": [ + 88, + 415, + 485, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimized Prompts Construction. With a diverse and high-quality set of user queries, we employ an LLM, specifically GPT-4o, with in-context learning to generate initial optimized prompts. The demonstrations used in this process are carefully crafted to guide the LLM in producing clear, well-structured prompts. Additionally, we instruct the LLM to account for potential safety concerns, ensuring that the generated prompts avoid harmful content, such as depictions of violence or explicit material. The exact prompt template used in this data construction process is provided in the Appendix.", + "bbox": [ + 88, + 589, + 483, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Principle-Based Refinement. After generating query-prompt pairs $(x,p)$ , where $x$ is the user query and $p$ is the optimized prompt, we further refine the prompts using a principle-driven approach to enhance their quality. We employ an LLM-as-a-judge method to assess prompts based on three key principles: harmlessness, accuracy, and helpfulness. The LLM critiques the generated prompts, identifying issues such as harmful content, missing key details from the", + "bbox": [ + 88, + 779, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "user query, or vague scenario descriptions. We then collect these critiques and represent problematic cases as triplets $(x,p,c)$ , where $c$ is the critique. Based on this critique, we refine the generated prompt $p$ to produce a refined version $p_{\\mathrm{refined}}$ .", + "bbox": [ + 511, + 415, + 908, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Training. With the principle-based prompt refinement, we construct a high-quality SFT dataset, denoted as $(x,s)\\in D_{\\mathrm{SFT}}$ . For the prompt without being pointed out issues within critiques, $s = p$ , retaining the LLM-generated prompt. If issues are identified, we use the refined version, setting $s = p_{\\mathrm{refined}}$ . To initialize the prompt optimizer, we apply standard supervised fine-tuning with the loss function:", + "bbox": [ + 511, + 518, + 908, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log P (s | x, s _ {< i}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 642, + 906, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ represents the length of $s$ .", + "bbox": [ + 511, + 695, + 750, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Feedback Preference Optimization", + "text_level": 1, + "bbox": [ + 511, + 724, + 864, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the SFT model, we further enhance its capability in optimizing text-to-video generation prompts. Following our proposed principles, we incorporate two types of feedback. Text-level feedback assesses whether prompts align with user intent and maintain safety, while video-level feedback ensures that prompts result in high-quality video generation. Combining feedback from both dimensions, we can perform preference optimization on the model, allowing it to generate prompts that are safer, more accurate, and effective.", + "bbox": [ + 511, + 750, + 908, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15639", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Sampling. To support the DPO training, for each input $x$ , we first sample $K$ optimized prompts $(p_1, p_2, \\ldots, p_K)$ from the SFT model. We then construct DPO training pairs based on both text-level and video-level feedback.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-Level Preference Data Construction. A desired prompt optimizer should produce prompts that are harmless, faithfully represent the user's intent, and are clear and descriptive. To achieve this, in the preference optimization stage, we leverage an LLM-as-a-judge approach to provide text-level feedback. Given a user query and an optimized prompt pair $(x,p_{j})$ , the LLM generates a critique to check whether the prompt violates any proposed principles. If a flaw is detected, we refine the prompt based on the critique, producing a revised version $p_{j_{\\mathrm{refined}}}$ . These pairs $(x,p_{j} < p_{j_{\\mathrm{refined}}})$ comprise $D_{\\mathrm{text}}$ .", + "bbox": [ + 89, + 189, + 483, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video-Level Preference Data Construction. Ultimately, the helpfulness of a prompt should be judged by the quality of the video it helps generate. To guide the prompt optimizer in producing high-quality outputs, we utilize VisionReward [27], a state-of-the-art video reward model, to provide automated feedback. For prompts that pass the text-level check, we generate corresponding videos using the target video generation model. VisionReward then evaluates each video and assigns a reward score. This process results in $(x, p_m, r_m)$ , where $r_m$ is the score from VisionReward. By comparing these scores, we determine which prompts lead to better video quality. Assuming $r_m < r_{m+1}$ , we can then construct the video-level DPO pair, $(x, p_m < p_{m+1})$ , resulting in $D_{\\mathrm{video}}$ .", + "bbox": [ + 89, + 378, + 483, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Training. As we get text-level and video-level preference pairs, we can conduct DPO on the SFT model. The training dataset is denoted as $D_{\\mathrm{dpo}} = D_{\\mathrm{text}} \\cup D_{\\mathrm{video}}$ and the DPO loss is described as follows:", + "bbox": [ + 89, + 613, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\pi_ {\\theta}; \\pi_ {\\text {r e f}}\\right) = - \\mathbb {E} _ {\\left(x, p _ {w}, p _ {l}\\right) \\sim D _ {\\text {d p o}}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(p _ {w} \\mid x\\right)}{\\pi_ {\\text {r e f}} \\left(p _ {w} \\mid x\\right)} \\right. \\right. \\\\ \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(p _ {l} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(p _ {l} \\mid x\\right)}\\right) \\Bigg ] \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 684, + 488, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi_{\\theta}$ is the policy model initialized with the SFT model, and $\\pi_{ref}$ is a fixed reference model also initialized with the SFT model. Here, $p_w$ and $p_l$ represent the chosen ( $p_{j_{\\text{refined}}}$ from $D_{\\text{text}}$ and $p_{m+1}$ from $D_{\\text{video}}$ ) and rejected ( $p_j$ from $D_{\\text{text}}$ and $p_m$ from $D_{\\text{video}}$ ) prompts, respectively. This process results in the final DPO model, which is capable of generating prompts that improve video quality while maintaining safety and accuracy.", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 114, + 702, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To thoroughly evaluate the effectiveness of VPO, we conduct a series of comprehensive experiments covering multiple aspects, including the main results on text-to-video benchmarks, text-level alignment, safety evaluation, and comparisons with RLHF methods for video generation models. Additionally, we assess the iterative improvement capability, as well as the generalization performance of VPO across different video generation models. Implementation details are described in the Appendix.", + "bbox": [ + 511, + 138, + 906, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone Models & Baselines. Our experiments are conducted on several popular video generation models, including the CogVideoX series and Open-Sora 1.2. The CogVideoX series excel in producing coherent, motion-rich, long-duration videos, while Open-Sora 1.2 is an efficient open-source alternative. For CogVideoX, we use the official prompt optimization method based on GLM-4 in-context learning, along with GPT-4o in-context learning as baselines. For Open-Sora 1.2, we adopt the official GPT-4o rewriting method as our baseline. Additionally, we introduce an alternative baseline using the image prompt optimizer, Promptist [10]. Moreover, we compare VPO with Prompt-A-Video [13].", + "bbox": [ + 511, + 290, + 906, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Benchmarks. We use VBench [12] and MonetBench [27] for general text-to-video evaluation. VBench assesses multiple aspects, including Human Action, Scene, Multiple Objects, and Appearance Style. MonetBench evaluates diverse scenarios with metrics such as Alignment, Stability, Preservation, and Physics. For query alignment, we curate 500 diverse queries using LLM-based filtering and manual verification. GPT-4o is leveraged to assess whether the optimized prompts violate our predefined principles, such as containing harmful content or omitting key details. For safety, we use a subset of T2VSafetyBench [16]. More details are provided in the Appendix.", + "bbox": [ + 511, + 501, + 908, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Text-to-Video Benchmark Results", + "text_level": 1, + "bbox": [ + 511, + 696, + 808, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1, we implement VPO on CogVideoX-2B and CogVideoX-5B. In both cases, VPO significantly outperforms baseline approaches. The obvious gap between directly using the original query with using prompt optimization techniques highlights the critical role of prompt optimization in video generation. Furthermore, the difference between VPO-SFT and VPO models demonstrates the necessity of incorporating multiple feedback preference optimization. Notably, the superior performance of VPO over VPO without text-level feedback indicates that improving safety and accuracy at the text level enhances general text-to-video tasks.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15640", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/61abdd188914ee82c86868958779a1b967cf06449e558028df17951daecc47ce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMonetBenchVBench
Align-mentStabilityPreservingPhysicsOverallHuman ActionSceneMultiple ObjectsAppear. Style
CogVideoX-2B
Original Query1.110.250.560.313.2780.0028.3440.1722.60
Promptist0.880.250.550.292.8767.4018.3727.4423.12
Prompt-A-Video1.230.270.610.333.5890.6043.8568.2622.33
GLM-4 Few-Shot1.280.270.590.333.5796.2055.5168.4023.47
GPT-4o Few-Shot1.260.270.580.333.5898.2052.5363.6323.73
VPO-SFT1.280.280.600.333.5997.0055.0468.9824.13
VPO w/o TL FDBK1.320.290.620.333.7296.4054.7867.7924.15
VPO1.340.290.630.343.7699.0055.8370.1724.20
CogVideoX-5B
Original Query1.310.290.620.343.7788.0041.3245.6723.37
Promptist1.080.280.620.333.4277.4024.9318.3423.27
Prompt-A-Video1.420.310.660.354.0591.8045.4074.4122.63
GLM-4 Few-Shot1.460.290.640.353.9898.4055.6072.3824.39
GPT-4o Few-Shot1.480.290.640.344.0399.2053.1372.2124.20
VPO-SFT1.470.300.650.354.0197.2058.4073.7024.55
VPO w/o TL FDBK1.520.310.670.354.1297.6054.5972.9923.96
VPO1.520.310.670.364.1599.6055.6875.7324.57
", + "bbox": [ + 114, + 88, + 885, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Main results on MonetBench and VBench (%). 'Appear. Style' stands for Appearance Style. VPO w/o TL FDBK indicates that only video-level rewards are used during the preference optimization stage. The highest results for each video generation model is bolded.", + "bbox": [ + 89, + 458, + 906, + 487 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/315068907e58ac17806c8fcba45a5e6537d6555bf5393dfe7e586890842f55fe.jpg", + "image_caption": [ + "Figure 3. Manual evaluation results comparing VPO with baseline methods. VPO demonstrates a significant advantage in instruction-following, video quality, and overall performance." + ], + "image_footnote": [], + "bbox": [ + 96, + 507, + 352, + 662 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a54233a2dc2740fa5c4d6e6097deaa35ad4849cdef920071b88ca5794b6019db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 507, + 624, + 662 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9204a50c5f60121b1adfab0505dc5a58898fecc1394c5bdebf86ac36fb96378.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 508, + 898, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further validate the effectiveness of VPO, we conduct a pairwise human evaluation on CogVideoX-5B. We ask annotators to focus on the following aspects:", + "bbox": [ + 89, + 720, + 482, + 766 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Instruction-Following: The degree to which the generated video adheres to the user query.", + "- Video Quality: The quality of generated videos, like coherence, stability, and adherence to physics.", + "- Overall Preference: A holistic judgment based on instruction-following and video quality." + ], + "bbox": [ + 89, + 772, + 482, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Figure 3, which are consistent with evaluation on benchmarks. VPO significantly sur", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "passes other methods. The superior performance of VPO over VPO without text-level feedback on the alignment dimension indicates that a more accurate prompt optimizer could help increase the instruction-following abilities of video generation models.", + "bbox": [ + 511, + 720, + 906, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Text-Level Alignment", + "text_level": 1, + "bbox": [ + 511, + 800, + 718, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further evaluate VPO on a more complex test set constructed from real-world user queries, focusing on accuracy and harmlessness. As shown in Table 2, VPO largely enhances text-level alignment, surpassing alternative methods, including much larger LLMs. Importantly, our findings re", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15641", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7b7919fe9a62e08b0e7337dbb8845d6f376a4d7e0e55c1af40ba3e0bbf49724e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAligned ↑Misaligned ↓
UnsafeImpreciseRefusal
GLM-4 Few-Shot83.45.410.01.2
GPT-4o Few-Shot86.42.48.62.6
VPO-SFT93.80.85.40.0
CogVideoX-2B
VPO w/o TL FDBK93.02.05.00.0
VPO94.60.64.80.0
CogVideoX-5B
VPO w/o TL FDBK92.81.26.00.0
VPO94.80.44.80.0
", + "bbox": [ + 102, + 88, + 475, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Evaluation results $(\\%)$ of query alignment. We highlight the best results for each video generation model in bold.", + "bbox": [ + 89, + 266, + 482, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03c9e3106a3104cab9a99ffbed6d25824649cd5025d3291ad4a5ee54ff35ccaa.jpg", + "image_caption": [ + "Figure 4. Manual evaluation results on safety tasks. VPO demonstrates substantially improved harmlessness, particularly in terms of the completely safe rate (Level 1)." + ], + "image_footnote": [], + "bbox": [ + 101, + 314, + 475, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "veal a critical limitation of relying solely on LLM-based in-context learning for prompt optimization: these models may refuse to process queries containing sensitive keywords or abstract contents, such as \"20 - 11 coins = 9 coins.\" This underscores the necessity of training a comprehensive prompt optimizer capable of dealing with diverse user queries.", + "bbox": [ + 89, + 612, + 483, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Safety Evaluation", + "text_level": 1, + "bbox": [ + 89, + 712, + 264, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a manual safety evaluation using a subset of T2VSafetyBench, where annotators assign safety scores to generated videos on a four-point scale:", + "bbox": [ + 89, + 734, + 483, + 779 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Level 1: Completely safe.", + "- Level 2: Safe but contains minor unsafe elements.", + "- Level 3: Unsafe.", + "- Level 4: Extremely unsafe." + ], + "bbox": [ + 91, + 780, + 434, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 4, VPO substantially improves video safety compared to LLM few-shot methods. This enhancement is particularly evident at Level 1, where VPO ensures a significantly higher proportion of completely safe outputs.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bf4016782b666b2d066f2e9a86eb3483dd7201018bccecc60f32d702ae78258.jpg", + "image_caption": [ + "Figure 5. Pairwise evaluation results using VisionReward. Standard refers to the original video generation model with the official prompt rewriting method. VPO outperforms Diffusion DPO and can be combined with it to provide additional benefits." + ], + "image_footnote": [], + "bbox": [ + 519, + 98, + 901, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importantly, incorporating text-level feedback further enhances safety, whereas neglecting this component leads to a lower safety rating than the SFT model, indicating that solely optimizing for video quality could damage safety.", + "bbox": [ + 511, + 359, + 906, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison to RLHF", + "text_level": 1, + "bbox": [ + 511, + 430, + 718, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We reveal that the prompt optimization method is not only important to bridge the training and inference gap of video generation models, but it is also a good way to align the video generation model with human preference. In this experiment, we compare VPO with Diffusion DPO [25]. Diffusion DPO is applied with the official GLM-4 prompt rewriting. Using VisionReward for pairwise evaluation, we find that both methods enhance CogVideoX-5B's performance, with VPO outperforming Diffusion DPO, as shown in Figure 5. Notably, VPO and Diffusion DPO can be combined for additional gains, demonstrating that optimizing both the video generation model and prompt optimizer can lead to superior results.", + "bbox": [ + 511, + 452, + 908, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Generalization across Models", + "text_level": 1, + "bbox": [ + 511, + 657, + 774, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As the preference optimization stage of VPO relies on a specific video generation model to construct DPO pairs, we would like to investigate if the prompt optimizer could generalize to other models. To test this, we apply the prompt optimizer trained on CogVideoX-2B to Open-Sora 1.2. As shown in Table 3 and Table 4, we demonstrate that VPO could generalize across different video generation models, yielding non-trivial performance improvements on both VBenchmark and MonetBench.", + "bbox": [ + 511, + 680, + 908, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Case Study", + "text_level": 1, + "bbox": [ + 511, + 829, + 633, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 presents a comparative analysis of VPO and the GLM-4 Few-Shot approach in handling both harmful and harmless queries. On the left, the original query describes a", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15642", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/785c2e831bb7edf06b9c1c48f2dc68a29f5702b983d91d55501cc3e3a0030a0b.jpg", + "image_caption": [ + "Original Query: In a dimly lit hospital room, patients with exposed wounds and broken bones lie motionless, the floor slick with blood and medical waste." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 495, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aec504165cbf4ef1c8c8d197f6f539757a158fdb10bbe11bdbe440a5a5541c16.jpg", + "image_caption": [ + "GLM-4 Few-Shot: In a somber, dimly lit hospital ward, patients with various injuries—exposed lacerations, bruised limbs, and splinted fractures……, which is wet and shiny with patches of congealed blood and scattered medical debris." + ], + "image_footnote": [], + "bbox": [ + 94, + 165, + 493, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/875e4bf4cedba992373c2b642b373eccaa3453b184375d40e82ca9f78d0e5558.jpg", + "image_caption": [ + "VPO (Ours): In a dimly lit hospital room, the scene focuses on the caring and professional environment. creating a calm and sterile interior. The professionals move with precision and compassion, reflecting a workplace dedicated to healing and care, capturing the resilience and professionalism of healthcare workers.", + "Figure 6. (Left) Case study on a harmful query; (Right) Case study on a harmless query. Some prompts are omitted due to space constraints." + ], + "image_footnote": [], + "bbox": [ + 94, + 253, + 493, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6038611e15a77631c28ddc1688458103af28bb77cefedcb95c59fcb9920bb4ea.jpg", + "image_caption": [ + "Original Query: A person is hugging." + ], + "image_footnote": [], + "bbox": [ + 508, + 88, + 901, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8164d550745491e89cb68674fe4a88f80f30d6c5cfc9e61e06d0da9db2e11017.jpg", + "image_caption": [ + "GLM-4 Few-Shot: A person, enveloped in a cozy knit sweater, stands in the heart of a lush garden where the vibrant colors of autumn leaves cascade around them. They are seen opening their arms wide, a warm and inviting gesture. As they......" + ], + "image_footnote": [], + "bbox": [ + 501, + 165, + 903, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17cf9750610a364db2b0a79e376446c5285cfe02d164422972feec02aeab0c38.jpg", + "image_caption": [ + "VPO (Ours): A young woman with shoulder-length auburn hair is seen in a cozy living room with warm lighting, wearing a light blue sweater and jeans. She stands close to a man with short black hair, dressed in a green flannel shirt and dark trousers. They embrace each other warmly in a friendly and affectionate manner...." + ], + "image_footnote": [], + "bbox": [ + 501, + 253, + 903, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3a070f901d23c1e0c9067d73002a19913787cb543d7abd5cee5f2bad28c41f7a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHuman ActionSceneMultiple ObjectsAppear. Style
Original Query88.8044.0855.9923.87
GPT-4o Few-Shot92.4053.2165.0223.84
VPO-SFT95.8051.4164.2823.86
VPO97.0053.5867.8823.93
", + "bbox": [ + 94, + 401, + 478, + 500 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d6820fb4759afcfafb9e26657c5eba6b0fe30c2d47c7bab519dbd0647a80ed66.jpg", + "table_caption": [ + "Table 3. Evaluation results (%) of Open-Sora 1.2 on VBench." + ], + "table_footnote": [], + "table_body": "
MethodAlign-mentStabilityPreser-vationPhysicsOverall
Original Query1.010.210.540.312.90
GPT-4o Few-Shot1.120.210.550.313.07
VPO-SFT1.120.230.560.313.08
VPO1.130.240.580.333.18
", + "bbox": [ + 94, + 539, + 482, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Evaluation results of Open-Sora 1.2 on MonetBench.", + "bbox": [ + 102, + 637, + 470, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "disturbing hospital scene with exposed wounds, blood, and medical waste. The few-shot method generates a detailed description of these unsettling elements, leading to a harmful video. In contrast, VPO shifts the focus to a professional and sterile hospital environment, emphasizing care and resilience rather than distressing imagery. This demonstrates VPO's ability to refine harmful prompts into ethical and appropriate content while maintaining contextual relevance. On the right, we examine a harmless query about hugging. While the few-shot approach generates a semantically rich scene, the resulting video suffers from poor quality, with unnatural body movements and misplaced limbs. In comparison, VPO produces a more natural and visually coherent video, with smoother and more realistic movements. This highlights the importance of integrating feedback from gen", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "erated videos to build a helpful prompt optimizer. These cases show that VPO not only improves video quality but also ensures that the generated content remains safe and appropriate. More examples are provided in the Appendix.", + "bbox": [ + 511, + 404, + 906, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 482, + 633, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce VPO, a principle-driven framework for optimizing prompts in video generation, ensuring harmlessness, accuracy, and helpfulness. Unlike conventional in-context learning approaches, which rewrite prompts without considering their actual effectiveness, VPO improves video generation quality through principle-based SFT and multi-feedback preference optimization, incorporating both text-level and video-level feedback. Through extensive experiments, we demonstrate the superiority of VPO over the baseline methods, delivering safer, more accurate, and higher-quality video generation, effectively bridging the gap between training and inference in video generation models. Moreover, we reveal that VPO also serves as an effective RLHF method and can be integrated with RLHF methods on video models, bringing orthogonal improvements. Additionally, we demonstrate the generalization of VPO across video models, showing the potential for developing a general text-to-video prompt optimization model.", + "bbox": [ + 511, + 508, + 906, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 813, + 692, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Science Foundation for Distinguished Young Scholars (with No. 62125604). This work was also supported by Tsinghua University Initiative Scientific Research Program. We would", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15643", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also like to thank Zhipu AI for sponsoring GPU computing and API cost consumed in this study.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 133, + 187, + 148 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, et al. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861, 2021. 1", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 3", + "[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3", + "[6] Jiale Cheng, Xiao Liu, Kehan Zheng, Pei Ke, Hongning Wang, Yuxiao Dong, Jie Tang, and Minlie Huang. Blackbox prompt optimization: Aligning large language models without model training. arXiv preprint arXiv:2311.04155, 2023. 3", + "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 11", + "[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3", + "[9] Laura Hanu and Unitary team. Detoxify. Github. https://github.com/unitaryai/detoxify, 2020.4", + "[10] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36:66923-66939, 2023. 5", + "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3", + "[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of" + ], + "bbox": [ + 93, + 159, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5", + "[13] Yatai Ji, Jiacheng Zhang, Jie Wu, Shilong Zhang, Shoufa Chen, Chongjian GE, Peize Sun, Weifeng Chen, Wenqi Shao, Xuefeng Xiao, et al. Prompt-a-video: Prompt your video diffusion model via preference-aligned llm. arXiv preprint arXiv:2412.15156, 2024. 3, 5", + "[14] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 3", + "[15] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 3", + "[16] Yibo Miao, Yifan Zhu, Yinpeng Dong, Lijia Yu, Jun Zhu, and Xiao-Shan Gao. T2vsafetybench: Evaluating the safety of text-to-video generative models. arXiv preprint arXiv:2407.05965, 2024. 5", + "[17] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3", + "[18] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1, 3", + "[19] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 11", + "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15644", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3", + "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[22] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222-4235, 2020. 3", + "[23] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3", + "[24] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[25] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 7", + "[26] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. arXiv preprint arXiv:2403.06098, 2024. 3", + "[27] Jiazheng Xu, Yu Huang, Jiale Cheng, Yuanming Yang, Jiajun Xu, Yuan Wang, Wenbo Duan, Shen Yang, Qunlin Jin, Shurun Li, et al. Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. arXiv preprint arXiv:2412.21059, 2024. 5", + "[28] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3", + "[29] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3", + "[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[31] Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. Texygen: A benchmarking platform for text generation models. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 1097-1100, 2018. 4" + ], + "bbox": [ + 91, + 90, + 485, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15645", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_model.json b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..abd999828f31aa3c779dd43c44301ea9405e33e7 --- /dev/null +++ b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "VPO: Aligning Text-to-Video Generation Models with Prompt Optimization" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.18, + 0.88, + 0.218 + ], + "angle": 0, + "content": "Jiale Cheng\\(^{1,2*}\\), Ruiliang Lyu\\(^{2}\\), Xiaotao Gu\\(^{2}\\), Xiao Liu\\(^{2,3}\\), Jiazheng Xu\\(^{2,3}\\), Yida Lu\\(^{1,2*}\\), Jiayan Teng\\(^{2,3}\\), Zhuoyi Yang\\(^{2,3}\\), Yuxiao Dong\\(^{3}\\), Jie Tang\\(^{3}\\), Hongning Wang\\(^{1}\\), Minlie Huang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.234, + 0.81, + 0.287 + ], + "angle": 0, + "content": "1The Conversational Artificial Intelligence (CoAI) Group, Tsinghua University \n2Zhipu AI 3The Knowledge Engineering Group (KEG), Tsinghua University \nchengjl23@mails.tsinghua.edu.cn, aihuang@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.486, + 0.841 + ], + "angle": 0, + "content": "Video generation models have achieved remarkable progress in text-to-video tasks. These models are typically trained on text-video pairs with highly detailed and carefully crafted descriptions, while real-world user inputs during inference are often concise, vague, or poorly structured. This gap makes prompt optimization crucial for generating high-quality videos. Current methods often rely on large language models (LLMs) to refine prompts through in-context learning, but suffer from several limitations: they may distort user intent, omit critical details, or introduce safety risks. Moreover, they optimize prompts without considering the impact on the final video quality, which can lead to suboptimal results. To address these issues, we introduce VPO, a principled framework that optimizes prompts based on three core principles: harmlessness, accuracy, and helpfulness. The generated prompts faithfully preserve user intents and, more importantly, enhance the safety and quality of generated videos. To achieve this, VPO employs a two-stage optimization approach. First, we construct and refine a supervised fine-tuning (SFT) dataset based on principles of safety and alignment. Second, we introduce both text-level and video-level feedback to further optimize the SFT model with preference learning. Our extensive experiments demonstrate that VPO significantly improves safety, alignment, and video quality compared to baseline methods. Moreover, VPO shows strong generalization across video generation models. Furthermore, we demonstrate that VPO could outperform and be combined with RLHF methods on video generation models, underscoring the effectiveness of VPO in aligning video generation models. Our code and data are publicly available at https://github.com/thu-coai/VPO." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.322, + 0.645, + 0.338 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.516 + ], + "angle": 0, + "content": "Recent advancements in text-to-video generation models have significantly improved the ability to produce high-quality video content [4, 15, 18, 28, 29]. These models typically rely on extensive, high-quality video-text annotated datasets, where detailed labels capture specific video content. While training descriptions are detailed and carefully crafted, real-world user inputs are often brief, ambiguous, or poorly structured. This critical gap between model training and inference can lead to suboptimal outputs, making it crucial to optimize and refine user inputs to generate desired high-quality video outputs (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Current approaches to optimizing user inputs for video generation mostly utilize large language models (LLMs) with in-context learning [15, 28, 29]. However, these methods suffer from several limitations:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.58, + 0.906, + 0.625 + ], + "angle": 0, + "content": "- Safety concerns: Existing methods do not explicitly ensure that optimized prompts remain safe, potentially leading to inappropriate or harmful video content." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.626, + 0.906, + 0.67 + ], + "angle": 0, + "content": "- Imprecise refinement: Current methods may unintentionally alter the user's intent or introduce biases, leading to outputs that deviate from the original query." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.671, + 0.906, + 0.732 + ], + "angle": 0, + "content": "- Neglecting final video quality: Current approaches optimize prompts to be semantically richer, but do not explicitly consider how the refined prompt impacts the actual generated video, limiting their practical effectiveness." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.58, + 0.906, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.906, + 0.809 + ], + "angle": 0, + "content": "To address these challenges, we propose a set of alignment principles for video prompt optimization, inspired by the HHH (Harmless, Honest, Helpful) principle in LLMs [1]. An aligned prompt optimization model should adhere to the following:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.841 + ], + "angle": 0, + "content": "- Harmless: The refined prompt should avoid any harmful content, such as depictions of blood or violence." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.871 + ], + "angle": 0, + "content": "- Accurate: The optimized prompt should be precisely aligned with the user input, except for safety issues." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- Helpful: The refined prompt should be detailed and descriptive, and can help the video generation model pro" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.382, + 0.888 + ], + "angle": 0, + "content": "*Work done when JC and YL interned at Zhipu AI." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15636" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.087, + 0.885, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 1. (Upper) The critical gap between the training and inference stages in video generation models. During training, video descriptions are detailed, descriptive, and well-crafted, while user inputs in the inference stage are often concise, unclear, and poorly structured, creating a mismatch that hampers model performance. (Lower) Comparison of VPO with traditional prompt optimization methods. Traditional methods rely on the in-context learning capabilities of LLMs, which can lead to issues such as safety issues, misalignment, and low-quality videos. In contrast, VPO serves as a harmless, accurate, and helpful prompt optimizer for high-quality video generation." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.469, + 0.321, + 0.484 + ], + "angle": 0, + "content": "duce high-quality video content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Building on these principles, we introduce VPO, a principle-driven framework for optimizing user inputs for text-to-video generation, which integrates two critical stages: principle-based supervised fine-tuning (SFT) and multi-feedback preference optimization. In the SFT stage, we first leverage the in-context learning capabilities of LLMs to construct an initial SFT dataset. This data is then improved through the application of principle-based critique and refinement, ensuring the generated prompts are safe, accurately aligned with the original query, and sufficiently detailed. The resulting fine-tuned model poses a foundational ability to generate prompts following the proposed principles. In the subsequent preference optimization phase, we incorporate both text-level feedback and video-level reward signals to further improve the SFT model. By combining these feedback sources, we construct preference data that prioritizes harmlessness, accuracy, and helpfulness in the optimized prompt. Using Direct Preference Optimization (DPO), we train the model to generate prompts that consistently result in safer, more precise, and higher-quality video outputs. This two-stage framework explicitly integrates alignment principles into the prompt optimization process, bridging the critical gap between model training and real-world inference in video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We have conducted experiments on several popular video generation models, including CogVideoX [28] and Open-Sora [29]. Through extensive experiments, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.756 + ], + "angle": 0, + "content": "demonstrate the effectiveness of VPO in generating harmless, accurate, and helpful prompts for high-quality video content, outperforming traditional prompt rewriting methods. Notably, on CogVideoX, VPO improves the win rate of \\(37.5\\%\\) over original user queries and \\(14\\%\\) over the official prompt optimization method in human evaluation. Moreover, VPO significantly reduces the unsafety rate, enhancing the overall harmlessness of video generation models. We also highlight the necessity of multi-feedback preference optimization, demonstrating its crucial role in refining prompts for improved video generation. Furthermore, we demonstrate the potential of VPO as an RLHF method. VPO not only surpasses Diffusion DPO [25], but also brings additional benefits over Diffusion DPO. Additionally, VPO exhibits good generalization capabilities across different models—when trained on CogVideoX, it enhances the performance of Open-Sora 1.2, indicating the potential for training a general prompt optimizer for diverse video generation models." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.761, + 0.859, + 0.775 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.781, + 0.907, + 0.87 + ], + "angle": 0, + "content": "- We propose the principle of the alignment of video prompt optimizer: harmless, accurate, and helpful. Based on this, we introduce VPO, a systematic framework that integrates principle-based supervised fine-tuning and multi-feedback preference optimization, to build an aligned video prompt optimizer." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of VPO for improving query refinement accuracy, enhancing the safety of" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15637" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "generated videos, and boosting the overall performance of video generation models. Furthermore, VPO exhibits good generalization across different video generation models." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.152, + 0.483, + 0.198 + ], + "angle": 0, + "content": "- We demonstrate that VPO is a novel and competitive alignment method for text-to-video models, outperforming Diffusion DPO and providing orthogonal benefits." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.213, + 0.233, + 0.229 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.263, + 0.254 + ], + "angle": 0, + "content": "2.1. Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.626 + ], + "angle": 0, + "content": "With the widespread application of diffusion models in the domain of image generation [20, 21], research efforts have increasingly shifted towards video generation models [3, 8, 23]. Video Diffusion Models [11] firstly extend image diffusion models to video generation by employing a 3D UNet architecture. Text2Video-Zero [14] involves motion dynamics in image generation models to enable zero-shot video generation. However, these approaches encounter significant challenges in generating long videos. Subsequent studies, such as Stable Video Diffusion [2] and VideoCrafter [5], leverage large-scale pre-training on high-quality video datasets, yet their scalability remains constrained due to the limitation of architectures like UNet. By adopting Transformers [24] as the backbone of diffusion models, the scalable architecture of DiT [17] has substantially enhanced the capability of video generation. Building on DiT, several impressive video generation models have emerged [4, 15, 18, 28]. Nonetheless, the training data for these video generation models typically consists of detailed video descriptions, whereas user inputs during inference are often brief, unstructured, or even ambiguous in intent. Bridging this gap is critical for improving the quality of generated videos, underscoring the importance of training a robust prompt optimization model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.293, + 0.654 + ], + "angle": 0, + "content": "2.2. Prompt Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Prompt optimization techniques have long been a pivotal research problem for generation models. Early work on automatic prompt optimization can be traced back to AutoPrompt [22], which utilizes gradient-guided search to automatically create prompt templates. With the rapid development of LLMs, there has been a significant increase in studies leveraging LLMs for automatic prompt optimization [6, 30]. Similarly, existing video prompt optimization methods employ LLMs to refine user inputs, typically through in-context learning [15, 28, 29]. However, these methods rely solely on the inherent capabilities of LLMs, without considering the alignment with user query and the quality of generated videos, thus often resulting in suboptimal video outputs. Prompt-A-Video [13] incorporates image and video reward models during the optimization process to enhance the quality of generated videos, but it fails to ac" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.184 + ], + "angle": 0, + "content": "count for text-level alignment, which is essential for ensuring the safety and accuracy of rewritten prompts. To address these limitations, we propose VPO, a systematic framework for constructing a video prompt optimization model that can safely and accurately capture user intent, and assist video generation models in generating high-quality videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.198, + 0.605, + 0.213 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.907, + 0.391 + ], + "angle": 0, + "content": "We introduce VPO, an automated framework designed to bridge the gap between real-world user inputs and the well-crafted text descriptions used during model training. Unlike existing approaches that rely solely on the in-context learning capabilities of LLMs, VPO follows a principle-driven approach, ensuring that prompt refinements are harmless, accurate, and helpful. By incorporating these core principles into both supervised fine-tuning and preference optimization, VPO effectively transforms concise, unclear, or unsafe user inputs into detailed, well-structured, and safe prompts that significantly enhance video generation quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.401, + 0.704, + 0.416 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.907, + 0.665 + ], + "angle": 0, + "content": "The overall framework of VPO is illustrated in Figure 2. Our framework consists of two key stages: Principle-Based SFT (\\(\\S 3.2\\)) and Multi-Feedback Preference Optimization (\\(\\S 3.3\\)). In the first stage, Principle-Based SFT, we construct a high-quality dataset to train an initial model that refines user inputs while adhering to our three core principles. This stage provides the model with the foundational capabilities to optimize user input into a harmless, accurate, and helpful prompt for text-to-video generation. In the second stage, Multi-Feedback Preference Optimization, we further improve the SFT model using both text-level and video-level feedback. Text-level feedback helps refine prompts to be well-aligned with user intent while maintaining harmless. Meanwhile, video-level feedback ensures that the generated prompts can lead to high-quality video outputs, effectively enhancing the performance of the text-to-video model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.676, + 0.71, + 0.691 + ], + "angle": 0, + "content": "3.2. Principle-Based SFT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.905, + 0.759 + ], + "angle": 0, + "content": "We first construct a high-quality SFT dataset comprising user query and optimized prompt pairs to initialize the prompt optimizer with the ability to generate prompts that align with our guiding principles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Query Curation. To build the SFT dataset, we begin by collecting a comprehensive set of user queries. We utilize the VidProM dataset [26], which contains over one million real-world text-to-video queries. We adopt its deduplicated version and apply additional filtering steps to enhance the data. First, we perform rule-based quality filtering, considering factors such as keywords, special characters, and query length. To ensure query diversity, we employ the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15638" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.088, + 0.365, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.088, + 0.885, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.347, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 2. The overall framework of VPO. VPO consists of two main steps: Principle-Based SFT and Multi-Feedback Preference Optimization. By refining user queries into harmless, accurate, and helpful prompts, VPO acts as a bridge between users and video generation models, ultimately enhancing the quality of generated videos." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.486, + 0.568 + ], + "angle": 0, + "content": "self-BLEU [31] metric to filter out overly similar queries. Additionally, since we aim to develop a prompt optimizer that prioritizes safety, we incorporate safety-related queries into our dataset. Specifically, we extract prompts labeled as unsafe in the original dataset. However, as Detoxify [9] introduces labeling errors, we re-evaluate these safety-related queries using a more robust LLM. Following this data curation process, we finalize a dataset consisting of approximately 18k general queries and 2k safety-related queries for training, where 10k are used for SFT and 10k for DPO." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.59, + 0.485, + 0.756 + ], + "angle": 0, + "content": "Optimized Prompts Construction. With a diverse and high-quality set of user queries, we employ an LLM, specifically GPT-4o, with in-context learning to generate initial optimized prompts. The demonstrations used in this process are carefully crafted to guide the LLM in producing clear, well-structured prompts. Additionally, we instruct the LLM to account for potential safety concerns, ensuring that the generated prompts avoid harmful content, such as depictions of violence or explicit material. The exact prompt template used in this data construction process is provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Principle-Based Refinement. After generating query-prompt pairs \\((x,p)\\), where \\(x\\) is the user query and \\(p\\) is the optimized prompt, we further refine the prompts using a principle-driven approach to enhance their quality. We employ an LLM-as-a-judge method to assess prompts based on three key principles: harmlessness, accuracy, and helpfulness. The LLM critiques the generated prompts, identifying issues such as harmful content, missing key details from the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.909, + 0.493 + ], + "angle": 0, + "content": "user query, or vague scenario descriptions. We then collect these critiques and represent problematic cases as triplets \\((x,p,c)\\), where \\(c\\) is the critique. Based on this critique, we refine the generated prompt \\(p\\) to produce a refined version \\(p_{\\mathrm{refined}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.64 + ], + "angle": 0, + "content": "Model Training. With the principle-based prompt refinement, we construct a high-quality SFT dataset, denoted as \\((x,s)\\in D_{\\mathrm{SFT}}\\). For the prompt without being pointed out issues within critiques, \\(s = p\\), retaining the LLM-generated prompt. If issues are identified, we use the refined version, setting \\(s = p_{\\mathrm{refined}}\\). To initialize the prompt optimizer, we apply standard supervised fine-tuning with the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.643, + 0.907, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log P (s | x, s _ {< i}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.696, + 0.75, + 0.712 + ], + "angle": 0, + "content": "where \\(N\\) represents the length of \\(s\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.866, + 0.743 + ], + "angle": 0, + "content": "3.3. Multi-Feedback Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.9 + ], + "angle": 0, + "content": "After obtaining the SFT model, we further enhance its capability in optimizing text-to-video generation prompts. Following our proposed principles, we incorporate two types of feedback. Text-level feedback assesses whether prompts align with user intent and maintain safety, while video-level feedback ensures that prompts result in high-quality video generation. Combining feedback from both dimensions, we can perform preference optimization on the model, allowing it to generate prompts that are safer, more accurate, and effective." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15639" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "Data Sampling. To support the DPO training, for each input \\( x \\), we first sample \\( K \\) optimized prompts \\( (p_1, p_2, \\ldots, p_K) \\) from the SFT model. We then construct DPO training pairs based on both text-level and video-level feedback." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.19, + 0.485, + 0.359 + ], + "angle": 0, + "content": "Text-Level Preference Data Construction. A desired prompt optimizer should produce prompts that are harmless, faithfully represent the user's intent, and are clear and descriptive. To achieve this, in the preference optimization stage, we leverage an LLM-as-a-judge approach to provide text-level feedback. Given a user query and an optimized prompt pair \\((x,p_{j})\\), the LLM generates a critique to check whether the prompt violates any proposed principles. If a flaw is detected, we refine the prompt based on the critique, producing a revised version \\(p_{j_{\\mathrm{refined}}}\\). These pairs \\((x,p_{j} < p_{j_{\\mathrm{refined}}})\\) comprise \\(D_{\\mathrm{text}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Video-Level Preference Data Construction. Ultimately, the helpfulness of a prompt should be judged by the quality of the video it helps generate. To guide the prompt optimizer in producing high-quality outputs, we utilize VisionReward [27], a state-of-the-art video reward model, to provide automated feedback. For prompts that pass the text-level check, we generate corresponding videos using the target video generation model. VisionReward then evaluates each video and assigns a reward score. This process results in \\((x, p_m, r_m)\\), where \\(r_m\\) is the score from VisionReward. By comparing these scores, we determine which prompts lead to better video quality. Assuming \\(r_m < r_{m+1}\\), we can then construct the video-level DPO pair, \\((x, p_m < p_{m+1})\\), resulting in \\(D_{\\mathrm{video}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Model Training. As we get text-level and video-level preference pairs, we can conduct DPO on the SFT model. The training dataset is denoted as \\(D_{\\mathrm{dpo}} = D_{\\mathrm{text}} \\cup D_{\\mathrm{video}}\\) and the DPO loss is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.685, + 0.49, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\pi_ {\\theta}; \\pi_ {\\text {r e f}}\\right) = - \\mathbb {E} _ {\\left(x, p _ {w}, p _ {l}\\right) \\sim D _ {\\text {d p o}}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(p _ {w} \\mid x\\right)}{\\pi_ {\\text {r e f}} \\left(p _ {w} \\mid x\\right)} \\right. \\right. \\\\ \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(p _ {l} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(p _ {l} \\mid x\\right)}\\right) \\Bigg ] \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\pi_{\\theta}\\) is the policy model initialized with the SFT model, and \\(\\pi_{ref}\\) is a fixed reference model also initialized with the SFT model. Here, \\(p_w\\) and \\(p_l\\) represent the chosen (\\(p_{j_{\\text{refined}}}\\) from \\(D_{\\text{text}}\\) and \\(p_{m+1}\\) from \\(D_{\\text{video}}\\)) and rejected (\\(p_j\\) from \\(D_{\\text{text}}\\) and \\(p_m\\) from \\(D_{\\text{video}}\\)) prompts, respectively. This process results in the final DPO model, which is capable of generating prompts that improve video quality while maintaining safety and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.704, + 0.133 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.277 + ], + "angle": 0, + "content": "To thoroughly evaluate the effectiveness of VPO, we conduct a series of comprehensive experiments covering multiple aspects, including the main results on text-to-video benchmarks, text-level alignment, safety evaluation, and comparisons with RLHF methods for video generation models. Additionally, we assess the iterative improvement capability, as well as the generalization performance of VPO across different video generation models. Implementation details are described in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Backbone Models & Baselines. Our experiments are conducted on several popular video generation models, including the CogVideoX series and Open-Sora 1.2. The CogVideoX series excel in producing coherent, motion-rich, long-duration videos, while Open-Sora 1.2 is an efficient open-source alternative. For CogVideoX, we use the official prompt optimization method based on GLM-4 in-context learning, along with GPT-4o in-context learning as baselines. For Open-Sora 1.2, we adopt the official GPT-4o rewriting method as our baseline. Additionally, we introduce an alternative baseline using the image prompt optimizer, Promptist [10]. Moreover, we compare VPO with Prompt-A-Video [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Evaluation Benchmarks. We use VBench [12] and MonetBench [27] for general text-to-video evaluation. VBench assesses multiple aspects, including Human Action, Scene, Multiple Objects, and Appearance Style. MonetBench evaluates diverse scenarios with metrics such as Alignment, Stability, Preservation, and Physics. For query alignment, we curate 500 diverse queries using LLM-based filtering and manual verification. GPT-4o is leveraged to assess whether the optimized prompts violate our predefined principles, such as containing harmful content or omitting key details. For safety, we use a subset of T2VSafetyBench [16]. More details are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.81, + 0.712 + ], + "angle": 0, + "content": "4.2. Text-to-Video Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, we implement VPO on CogVideoX-2B and CogVideoX-5B. In both cases, VPO significantly outperforms baseline approaches. The obvious gap between directly using the original query with using prompt optimization techniques highlights the critical role of prompt optimization in video generation. Furthermore, the difference between VPO-SFT and VPO models demonstrates the necessity of incorporating multiple feedback preference optimization. Notably, the superior performance of VPO over VPO without text-level feedback indicates that improving safety and accuracy at the text level enhances general text-to-video tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15640" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.887, + 0.449 + ], + "angle": 0, + "content": "
MethodMonetBenchVBench
Align-mentStabilityPreservingPhysicsOverallHuman ActionSceneMultiple ObjectsAppear. Style
CogVideoX-2B
Original Query1.110.250.560.313.2780.0028.3440.1722.60
Promptist0.880.250.550.292.8767.4018.3727.4423.12
Prompt-A-Video1.230.270.610.333.5890.6043.8568.2622.33
GLM-4 Few-Shot1.280.270.590.333.5796.2055.5168.4023.47
GPT-4o Few-Shot1.260.270.580.333.5898.2052.5363.6323.73
VPO-SFT1.280.280.600.333.5997.0055.0468.9824.13
VPO w/o TL FDBK1.320.290.620.333.7296.4054.7867.7924.15
VPO1.340.290.630.343.7699.0055.8370.1724.20
CogVideoX-5B
Original Query1.310.290.620.343.7788.0041.3245.6723.37
Promptist1.080.280.620.333.4277.4024.9318.3423.27
Prompt-A-Video1.420.310.660.354.0591.8045.4074.4122.63
GLM-4 Few-Shot1.460.290.640.353.9898.4055.6072.3824.39
GPT-4o Few-Shot1.480.290.640.344.0399.2053.1372.2124.20
VPO-SFT1.470.300.650.354.0197.2058.4073.7024.55
VPO w/o TL FDBK1.520.310.670.354.1297.6054.5972.9923.96
VPO1.520.310.670.364.1599.6055.6875.7324.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.459, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Table 1. Main results on MonetBench and VBench (%). 'Appear. Style' stands for Appearance Style. VPO w/o TL FDBK indicates that only video-level rewards are used during the preference optimization stage. The highest results for each video generation model is bolded." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.508, + 0.353, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.508, + 0.625, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.509, + 0.9, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.677, + 0.907, + 0.706 + ], + "angle": 0, + "content": "Figure 3. Manual evaluation results comparing VPO with baseline methods. VPO demonstrates a significant advantage in instruction-following, video quality, and overall performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.767 + ], + "angle": 0, + "content": "To further validate the effectiveness of VPO, we conduct a pairwise human evaluation on CogVideoX-5B. We ask annotators to focus on the following aspects:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.773, + 0.483, + 0.802 + ], + "angle": 0, + "content": "- Instruction-Following: The degree to which the generated video adheres to the user query." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.803, + 0.483, + 0.833 + ], + "angle": 0, + "content": "- Video Quality: The quality of generated videos, like coherence, stability, and adherence to physics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.834, + 0.483, + 0.863 + ], + "angle": 0, + "content": "- Overall Preference: A holistic judgment based on instruction-following and video quality." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.773, + 0.483, + 0.863 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The results are shown in Figure 3, which are consistent with evaluation on benchmarks. VPO significantly sur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.907, + 0.797 + ], + "angle": 0, + "content": "passes other methods. The superior performance of VPO over VPO without text-level feedback on the alignment dimension indicates that a more accurate prompt optimizer could help increase the instruction-following abilities of video generation models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.719, + 0.818 + ], + "angle": 0, + "content": "4.3. Text-Level Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We further evaluate VPO on a more complex test set constructed from real-world user queries, focusing on accuracy and harmlessness. As shown in Table 2, VPO largely enhances text-level alignment, surpassing alternative methods, including much larger LLMs. Importantly, our findings re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15641" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.477, + 0.259 + ], + "angle": 0, + "content": "
MethodAligned ↑Misaligned ↓
UnsafeImpreciseRefusal
GLM-4 Few-Shot83.45.410.01.2
GPT-4o Few-Shot86.42.48.62.6
VPO-SFT93.80.85.40.0
CogVideoX-2B
VPO w/o TL FDBK93.02.05.00.0
VPO94.60.64.80.0
CogVideoX-5B
VPO w/o TL FDBK92.81.26.00.0
VPO94.80.44.80.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.267, + 0.483, + 0.297 + ], + "angle": 0, + "content": "Table 2. Evaluation results \\((\\%)\\) of query alignment. We highlight the best results for each video generation model in bold." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.315, + 0.476, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.589 + ], + "angle": 0, + "content": "Figure 4. Manual evaluation results on safety tasks. VPO demonstrates substantially improved harmlessness, particularly in terms of the completely safe rate (Level 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.705 + ], + "angle": 0, + "content": "veal a critical limitation of relying solely on LLM-based in-context learning for prompt optimization: these models may refuse to process queries containing sensitive keywords or abstract contents, such as \"20 - 11 coins = 9 coins.\" This underscores the necessity of training a comprehensive prompt optimizer capable of dealing with diverse user queries." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.265, + 0.729 + ], + "angle": 0, + "content": "4.4. Safety Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.78 + ], + "angle": 0, + "content": "We conduct a manual safety evaluation using a subset of T2VSafetyBench, where annotators assign safety scores to generated videos on a four-point scale:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.781, + 0.28, + 0.795 + ], + "angle": 0, + "content": "- Level 1: Completely safe." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.796, + 0.435, + 0.81 + ], + "angle": 0, + "content": "- Level 2: Safe but contains minor unsafe elements." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.219, + 0.824 + ], + "angle": 0, + "content": "- Level 3: Unsafe." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.287, + 0.84 + ], + "angle": 0, + "content": "- Level 4: Extremely unsafe." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.781, + 0.435, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 4, VPO substantially improves video safety compared to LLM few-shot methods. This enhancement is particularly evident at Level 1, where VPO ensures a significantly higher proportion of completely safe outputs." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.099, + 0.902, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.29, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Figure 5. Pairwise evaluation results using VisionReward. Standard refers to the original video generation model with the official prompt rewriting method. VPO outperforms Diffusion DPO and can be combined with it to provide additional benefits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Importantly, incorporating text-level feedback further enhances safety, whereas neglecting this component leads to a lower safety rating than the SFT model, indicating that solely optimizing for video quality could damage safety." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.431, + 0.72, + 0.448 + ], + "angle": 0, + "content": "4.5. Comparison to RLHF" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.651 + ], + "angle": 0, + "content": "We reveal that the prompt optimization method is not only important to bridge the training and inference gap of video generation models, but it is also a good way to align the video generation model with human preference. In this experiment, we compare VPO with Diffusion DPO [25]. Diffusion DPO is applied with the official GLM-4 prompt rewriting. Using VisionReward for pairwise evaluation, we find that both methods enhance CogVideoX-5B's performance, with VPO outperforming Diffusion DPO, as shown in Figure 5. Notably, VPO and Diffusion DPO can be combined for additional gains, demonstrating that optimizing both the video generation model and prompt optimizer can lead to superior results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.776, + 0.674 + ], + "angle": 0, + "content": "4.6. Generalization across Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.909, + 0.817 + ], + "angle": 0, + "content": "As the preference optimization stage of VPO relies on a specific video generation model to construct DPO pairs, we would like to investigate if the prompt optimizer could generalize to other models. To test this, we apply the prompt optimizer trained on CogVideoX-2B to Open-Sora 1.2. As shown in Table 3 and Table 4, we demonstrate that VPO could generalize across different video generation models, yielding non-trivial performance improvements on both VBenchmark and MonetBench." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.634, + 0.848 + ], + "angle": 0, + "content": "5. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Figure 6 presents a comparative analysis of VPO and the GLM-4 Few-Shot approach in handling both harmful and harmless queries. On the left, the original query describes a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15642" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.496, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.144, + 0.49, + 0.164 + ], + "angle": 0, + "content": "Original Query: In a dimly lit hospital room, patients with exposed wounds and broken bones lie motionless, the floor slick with blood and medical waste." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.166, + 0.495, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.22, + 0.486, + 0.251 + ], + "angle": 0, + "content": "GLM-4 Few-Shot: In a somber, dimly lit hospital ward, patients with various injuries—exposed lacerations, bruised limbs, and splinted fractures……, which is wet and shiny with patches of congealed blood and scattered medical debris." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.254, + 0.495, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.308, + 0.48, + 0.35 + ], + "angle": 0, + "content": "VPO (Ours): In a dimly lit hospital room, the scene focuses on the caring and professional environment. creating a calm and sterile interior. The professionals move with precision and compassion, reflecting a workplace dedicated to healing and care, capturing the resilience and professionalism of healthcare workers." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.089, + 0.903, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.145, + 0.68, + 0.157 + ], + "angle": 0, + "content": "Original Query: A person is hugging." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.166, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.22, + 0.89, + 0.251 + ], + "angle": 0, + "content": "GLM-4 Few-Shot: A person, enveloped in a cozy knit sweater, stands in the heart of a lush garden where the vibrant colors of autumn leaves cascade around them. They are seen opening their arms wide, a warm and inviting gesture. As they......" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.254, + 0.905, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.308, + 0.898, + 0.35 + ], + "angle": 0, + "content": "VPO (Ours): A young woman with shoulder-length auburn hair is seen in a cozy living room with warm lighting, wearing a light blue sweater and jeans. She stands close to a man with short black hair, dressed in a green flannel shirt and dark trousers. They embrace each other warmly in a friendly and affectionate manner...." + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.365, + 0.905, + 0.38 + ], + "angle": 0, + "content": "Figure 6. (Left) Case study on a harmful query; (Right) Case study on a harmless query. Some prompts are omitted due to space constraints." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.402, + 0.48, + 0.5 + ], + "angle": 0, + "content": "
MethodHuman ActionSceneMultiple ObjectsAppear. Style
Original Query88.8044.0855.9923.87
GPT-4o Few-Shot92.4053.2165.0223.84
VPO-SFT95.8051.4164.2823.86
VPO97.0053.5867.8823.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.51, + 0.47, + 0.525 + ], + "angle": 0, + "content": "Table 3. Evaluation results (%) of Open-Sora 1.2 on VBench." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.54, + 0.483, + 0.628 + ], + "angle": 0, + "content": "
MethodAlign-mentStabilityPreser-vationPhysicsOverall
Original Query1.010.210.540.312.90
GPT-4o Few-Shot1.120.210.550.313.07
VPO-SFT1.120.230.560.313.08
VPO1.130.240.580.333.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.638, + 0.471, + 0.652 + ], + "angle": 0, + "content": "Table 4. Evaluation results of Open-Sora 1.2 on MonetBench." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "disturbing hospital scene with exposed wounds, blood, and medical waste. The few-shot method generates a detailed description of these unsettling elements, leading to a harmful video. In contrast, VPO shifts the focus to a professional and sterile hospital environment, emphasizing care and resilience rather than distressing imagery. This demonstrates VPO's ability to refine harmful prompts into ethical and appropriate content while maintaining contextual relevance. On the right, we examine a harmless query about hugging. While the few-shot approach generates a semantically rich scene, the resulting video suffers from poor quality, with unnatural body movements and misplaced limbs. In comparison, VPO produces a more natural and visually coherent video, with smoother and more realistic movements. This highlights the importance of integrating feedback from gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.468 + ], + "angle": 0, + "content": "erated videos to build a helpful prompt optimizer. These cases show that VPO not only improves video quality but also ensures that the generated content remains safe and appropriate. More examples are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.634, + 0.499 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.796 + ], + "angle": 0, + "content": "In this work, we introduce VPO, a principle-driven framework for optimizing prompts in video generation, ensuring harmlessness, accuracy, and helpfulness. Unlike conventional in-context learning approaches, which rewrite prompts without considering their actual effectiveness, VPO improves video generation quality through principle-based SFT and multi-feedback preference optimization, incorporating both text-level and video-level feedback. Through extensive experiments, we demonstrate the superiority of VPO over the baseline methods, delivering safer, more accurate, and higher-quality video generation, effectively bridging the gap between training and inference in video generation models. Moreover, we reveal that VPO also serves as an effective RLHF method and can be integrated with RLHF methods on video models, bringing orthogonal improvements. Additionally, we demonstrate the generalization of VPO across video models, showing the potential for developing a general text-to-video prompt optimization model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.694, + 0.831 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Science Foundation for Distinguished Young Scholars (with No. 62125604). This work was also supported by Tsinghua University Initiative Scientific Research Program. We would" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15643" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "also like to thank Zhipu AI for sponsoring GPU computing and API cost consumed in this study." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.134, + 0.188, + 0.15 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.16, + 0.483, + 0.228 + ], + "angle": 0, + "content": "[1] Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, et al. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.23, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.3, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[6] Jiale Cheng, Xiao Liu, Kehan Zheng, Pei Ke, Hongning Wang, Yuxiao Dong, Jie Tang, and Minlie Huang. Blackbox prompt optimization: Aligning large language models without model training. arXiv preprint arXiv:2311.04155, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[9] Laura Hanu and Unitary team. Detoxify. Github. https://github.com/unitaryai/detoxify, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36:66923-66939, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.16, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[13] Yatai Ji, Jiacheng Zhang, Jie Wu, Shilong Zhang, Shoufa Chen, Chongjian GE, Peize Sun, Weifeng Chen, Wenqi Shao, Xuefeng Xiao, et al. Prompt-a-video: Prompt your video diffusion model via preference-aligned llm. arXiv preprint arXiv:2412.15156, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.195, + 0.905, + 0.278 + ], + "angle": 0, + "content": "[14] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.28, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[15] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.351, + 0.905, + 0.407 + ], + "angle": 0, + "content": "[16] Yibo Miao, Yifan Zhu, Yinpeng Dong, Lijia Yu, Jun Zhu, and Xiao-Shan Gao. T2vsafetybench: Evaluating the safety of text-to-video generative models. arXiv preprint arXiv:2407.05965, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.409, + 0.905, + 0.464 + ], + "angle": 0, + "content": "[17] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.467, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[18] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[19] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15644" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.486, + 0.191 + ], + "angle": 0, + "content": "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.485, + 0.275 + ], + "angle": 0, + "content": "[22] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222-4235, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[23] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[24] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.485, + 0.459 + ], + "angle": 0, + "content": "[25] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[26] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. arXiv preprint arXiv:2403.06098, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[27] Jiazheng Xu, Yu Huang, Jiale Cheng, Yuanming Yang, Jiajun Xu, Yuan Wang, Wenbo Duan, Shen Yang, Qunlin Jin, Shurun Li, et al. Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. arXiv preprint arXiv:2412.21059, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[28] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[29] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.483, + 0.768 + ], + "angle": 0, + "content": "[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.77, + 0.483, + 0.84 + ], + "angle": 0, + "content": "[31] Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. Texygen: A benchmarking platform for text generation models. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 1097-1100, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15645" + } + ] +] \ No newline at end of file diff --git a/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_origin.pdf b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3326c9e6cf8496928ec35d80f147fa2ea798bc63 --- /dev/null +++ b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/1912f989-0c41-45f7-ba83-c030d1f76cc0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ec0956166502fdbcda5edb84b4b3e736c9f20de3751f45d55c62c7e873c203b +size 8561687 diff --git a/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/full.md b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50c7ac64052a76fd6bc6b8658119121351fb32cc --- /dev/null +++ b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/full.md @@ -0,0 +1,267 @@ +# VPO: Aligning Text-to-Video Generation Models with Prompt Optimization + +Jiale Cheng $^{1,2*}$ , Ruiliang Lyu $^{2}$ , Xiaotao Gu $^{2}$ , Xiao Liu $^{2,3}$ , Jiazheng Xu $^{2,3}$ , Yida Lu $^{1,2*}$ , Jiayan Teng $^{2,3}$ , Zhuoyi Yang $^{2,3}$ , Yuxiao Dong $^{3}$ , Jie Tang $^{3}$ , Hongning Wang $^{1}$ , Minlie Huang $^{1\dagger}$ + +1The Conversational Artificial Intelligence (CoAI) Group, Tsinghua University +2Zhipu AI 3The Knowledge Engineering Group (KEG), Tsinghua University +chengjl23@mails.tsinghua.edu.cn, aihuang@tsinghua.edu.cn + +# Abstract + +Video generation models have achieved remarkable progress in text-to-video tasks. These models are typically trained on text-video pairs with highly detailed and carefully crafted descriptions, while real-world user inputs during inference are often concise, vague, or poorly structured. This gap makes prompt optimization crucial for generating high-quality videos. Current methods often rely on large language models (LLMs) to refine prompts through in-context learning, but suffer from several limitations: they may distort user intent, omit critical details, or introduce safety risks. Moreover, they optimize prompts without considering the impact on the final video quality, which can lead to suboptimal results. To address these issues, we introduce VPO, a principled framework that optimizes prompts based on three core principles: harmlessness, accuracy, and helpfulness. The generated prompts faithfully preserve user intents and, more importantly, enhance the safety and quality of generated videos. To achieve this, VPO employs a two-stage optimization approach. First, we construct and refine a supervised fine-tuning (SFT) dataset based on principles of safety and alignment. Second, we introduce both text-level and video-level feedback to further optimize the SFT model with preference learning. Our extensive experiments demonstrate that VPO significantly improves safety, alignment, and video quality compared to baseline methods. Moreover, VPO shows strong generalization across video generation models. Furthermore, we demonstrate that VPO could outperform and be combined with RLHF methods on video generation models, underscoring the effectiveness of VPO in aligning video generation models. Our code and data are publicly available at https://github.com/thu-coai/VPO. + +# 1. Introduction + +Recent advancements in text-to-video generation models have significantly improved the ability to produce high-quality video content [4, 15, 18, 28, 29]. These models typically rely on extensive, high-quality video-text annotated datasets, where detailed labels capture specific video content. While training descriptions are detailed and carefully crafted, real-world user inputs are often brief, ambiguous, or poorly structured. This critical gap between model training and inference can lead to suboptimal outputs, making it crucial to optimize and refine user inputs to generate desired high-quality video outputs (Figure 1). + +Current approaches to optimizing user inputs for video generation mostly utilize large language models (LLMs) with in-context learning [15, 28, 29]. However, these methods suffer from several limitations: + +- Safety concerns: Existing methods do not explicitly ensure that optimized prompts remain safe, potentially leading to inappropriate or harmful video content. +- Imprecise refinement: Current methods may unintentionally alter the user's intent or introduce biases, leading to outputs that deviate from the original query. +- Neglecting final video quality: Current approaches optimize prompts to be semantically richer, but do not explicitly consider how the refined prompt impacts the actual generated video, limiting their practical effectiveness. + +To address these challenges, we propose a set of alignment principles for video prompt optimization, inspired by the HHH (Harmless, Honest, Helpful) principle in LLMs [1]. An aligned prompt optimization model should adhere to the following: + +- Harmless: The refined prompt should avoid any harmful content, such as depictions of blood or violence. +- Accurate: The optimized prompt should be precisely aligned with the user input, except for safety issues. +- Helpful: The refined prompt should be detailed and descriptive, and can help the video generation model pro + +![](images/fefd15023042662f688c12aabf59840b71143e0cba7117c58820e96a620edcaf.jpg) +Figure 1. (Upper) The critical gap between the training and inference stages in video generation models. During training, video descriptions are detailed, descriptive, and well-crafted, while user inputs in the inference stage are often concise, unclear, and poorly structured, creating a mismatch that hampers model performance. (Lower) Comparison of VPO with traditional prompt optimization methods. Traditional methods rely on the in-context learning capabilities of LLMs, which can lead to issues such as safety issues, misalignment, and low-quality videos. In contrast, VPO serves as a harmless, accurate, and helpful prompt optimizer for high-quality video generation. + +duce high-quality video content. + +Building on these principles, we introduce VPO, a principle-driven framework for optimizing user inputs for text-to-video generation, which integrates two critical stages: principle-based supervised fine-tuning (SFT) and multi-feedback preference optimization. In the SFT stage, we first leverage the in-context learning capabilities of LLMs to construct an initial SFT dataset. This data is then improved through the application of principle-based critique and refinement, ensuring the generated prompts are safe, accurately aligned with the original query, and sufficiently detailed. The resulting fine-tuned model poses a foundational ability to generate prompts following the proposed principles. In the subsequent preference optimization phase, we incorporate both text-level feedback and video-level reward signals to further improve the SFT model. By combining these feedback sources, we construct preference data that prioritizes harmlessness, accuracy, and helpfulness in the optimized prompt. Using Direct Preference Optimization (DPO), we train the model to generate prompts that consistently result in safer, more precise, and higher-quality video outputs. This two-stage framework explicitly integrates alignment principles into the prompt optimization process, bridging the critical gap between model training and real-world inference in video generation. + +We have conducted experiments on several popular video generation models, including CogVideoX [28] and Open-Sora [29]. Through extensive experiments, we + +demonstrate the effectiveness of VPO in generating harmless, accurate, and helpful prompts for high-quality video content, outperforming traditional prompt rewriting methods. Notably, on CogVideoX, VPO improves the win rate of $37.5\%$ over original user queries and $14\%$ over the official prompt optimization method in human evaluation. Moreover, VPO significantly reduces the unsafety rate, enhancing the overall harmlessness of video generation models. We also highlight the necessity of multi-feedback preference optimization, demonstrating its crucial role in refining prompts for improved video generation. Furthermore, we demonstrate the potential of VPO as an RLHF method. VPO not only surpasses Diffusion DPO [25], but also brings additional benefits over Diffusion DPO. Additionally, VPO exhibits good generalization capabilities across different models—when trained on CogVideoX, it enhances the performance of Open-Sora 1.2, indicating the potential for training a general prompt optimizer for diverse video generation models. + +Our contributions can be summarized as follows: + +- We propose the principle of the alignment of video prompt optimizer: harmless, accurate, and helpful. Based on this, we introduce VPO, a systematic framework that integrates principle-based supervised fine-tuning and multi-feedback preference optimization, to build an aligned video prompt optimizer. +- We demonstrate the effectiveness of VPO for improving query refinement accuracy, enhancing the safety of + +generated videos, and boosting the overall performance of video generation models. Furthermore, VPO exhibits good generalization across different video generation models. + +- We demonstrate that VPO is a novel and competitive alignment method for text-to-video models, outperforming Diffusion DPO and providing orthogonal benefits. + +# 2. Related Work + +# 2.1. Video Generation + +With the widespread application of diffusion models in the domain of image generation [20, 21], research efforts have increasingly shifted towards video generation models [3, 8, 23]. Video Diffusion Models [11] firstly extend image diffusion models to video generation by employing a 3D UNet architecture. Text2Video-Zero [14] involves motion dynamics in image generation models to enable zero-shot video generation. However, these approaches encounter significant challenges in generating long videos. Subsequent studies, such as Stable Video Diffusion [2] and VideoCrafter [5], leverage large-scale pre-training on high-quality video datasets, yet their scalability remains constrained due to the limitation of architectures like UNet. By adopting Transformers [24] as the backbone of diffusion models, the scalable architecture of DiT [17] has substantially enhanced the capability of video generation. Building on DiT, several impressive video generation models have emerged [4, 15, 18, 28]. Nonetheless, the training data for these video generation models typically consists of detailed video descriptions, whereas user inputs during inference are often brief, unstructured, or even ambiguous in intent. Bridging this gap is critical for improving the quality of generated videos, underscoring the importance of training a robust prompt optimization model. + +# 2.2. Prompt Optimization + +Prompt optimization techniques have long been a pivotal research problem for generation models. Early work on automatic prompt optimization can be traced back to AutoPrompt [22], which utilizes gradient-guided search to automatically create prompt templates. With the rapid development of LLMs, there has been a significant increase in studies leveraging LLMs for automatic prompt optimization [6, 30]. Similarly, existing video prompt optimization methods employ LLMs to refine user inputs, typically through in-context learning [15, 28, 29]. However, these methods rely solely on the inherent capabilities of LLMs, without considering the alignment with user query and the quality of generated videos, thus often resulting in suboptimal video outputs. Prompt-A-Video [13] incorporates image and video reward models during the optimization process to enhance the quality of generated videos, but it fails to ac + +count for text-level alignment, which is essential for ensuring the safety and accuracy of rewritten prompts. To address these limitations, we propose VPO, a systematic framework for constructing a video prompt optimization model that can safely and accurately capture user intent, and assist video generation models in generating high-quality videos. + +# 3. Method + +We introduce VPO, an automated framework designed to bridge the gap between real-world user inputs and the well-crafted text descriptions used during model training. Unlike existing approaches that rely solely on the in-context learning capabilities of LLMs, VPO follows a principle-driven approach, ensuring that prompt refinements are harmless, accurate, and helpful. By incorporating these core principles into both supervised fine-tuning and preference optimization, VPO effectively transforms concise, unclear, or unsafe user inputs into detailed, well-structured, and safe prompts that significantly enhance video generation quality. + +# 3.1. Overall Framework + +The overall framework of VPO is illustrated in Figure 2. Our framework consists of two key stages: Principle-Based SFT ( $\S 3.2$ ) and Multi-Feedback Preference Optimization ( $\S 3.3$ ). In the first stage, Principle-Based SFT, we construct a high-quality dataset to train an initial model that refines user inputs while adhering to our three core principles. This stage provides the model with the foundational capabilities to optimize user input into a harmless, accurate, and helpful prompt for text-to-video generation. In the second stage, Multi-Feedback Preference Optimization, we further improve the SFT model using both text-level and video-level feedback. Text-level feedback helps refine prompts to be well-aligned with user intent while maintaining harmless. Meanwhile, video-level feedback ensures that the generated prompts can lead to high-quality video outputs, effectively enhancing the performance of the text-to-video model. + +# 3.2. Principle-Based SFT + +We first construct a high-quality SFT dataset comprising user query and optimized prompt pairs to initialize the prompt optimizer with the ability to generate prompts that align with our guiding principles. + +Query Curation. To build the SFT dataset, we begin by collecting a comprehensive set of user queries. We utilize the VidProM dataset [26], which contains over one million real-world text-to-video queries. We adopt its deduplicated version and apply additional filtering steps to enhance the data. First, we perform rule-based quality filtering, considering factors such as keywords, special characters, and query length. To ensure query diversity, we employ the + +![](images/31b9860833d98c369d137d0de96b2da5aaaebbe14c531057a11b3ce3de45db2c.jpg) +Figure 2. The overall framework of VPO. VPO consists of two main steps: Principle-Based SFT and Multi-Feedback Preference Optimization. By refining user queries into harmless, accurate, and helpful prompts, VPO acts as a bridge between users and video generation models, ultimately enhancing the quality of generated videos. + +![](images/5cee41c894a273e33c8e58906d0b54b5ac3ca38b6046145c7b30a67e1bc1fcca.jpg) + +self-BLEU [31] metric to filter out overly similar queries. Additionally, since we aim to develop a prompt optimizer that prioritizes safety, we incorporate safety-related queries into our dataset. Specifically, we extract prompts labeled as unsafe in the original dataset. However, as Detoxify [9] introduces labeling errors, we re-evaluate these safety-related queries using a more robust LLM. Following this data curation process, we finalize a dataset consisting of approximately 18k general queries and 2k safety-related queries for training, where 10k are used for SFT and 10k for DPO. + +Optimized Prompts Construction. With a diverse and high-quality set of user queries, we employ an LLM, specifically GPT-4o, with in-context learning to generate initial optimized prompts. The demonstrations used in this process are carefully crafted to guide the LLM in producing clear, well-structured prompts. Additionally, we instruct the LLM to account for potential safety concerns, ensuring that the generated prompts avoid harmful content, such as depictions of violence or explicit material. The exact prompt template used in this data construction process is provided in the Appendix. + +Principle-Based Refinement. After generating query-prompt pairs $(x,p)$ , where $x$ is the user query and $p$ is the optimized prompt, we further refine the prompts using a principle-driven approach to enhance their quality. We employ an LLM-as-a-judge method to assess prompts based on three key principles: harmlessness, accuracy, and helpfulness. The LLM critiques the generated prompts, identifying issues such as harmful content, missing key details from the + +user query, or vague scenario descriptions. We then collect these critiques and represent problematic cases as triplets $(x,p,c)$ , where $c$ is the critique. Based on this critique, we refine the generated prompt $p$ to produce a refined version $p_{\mathrm{refined}}$ . + +Model Training. With the principle-based prompt refinement, we construct a high-quality SFT dataset, denoted as $(x,s)\in D_{\mathrm{SFT}}$ . For the prompt without being pointed out issues within critiques, $s = p$ , retaining the LLM-generated prompt. If issues are identified, we use the refined version, setting $s = p_{\mathrm{refined}}$ . To initialize the prompt optimizer, we apply standard supervised fine-tuning with the loss function: + +$$ +\mathcal {L} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log P (s | x, s _ {< i}), \tag {1} +$$ + +where $N$ represents the length of $s$ . + +# 3.3. Multi-Feedback Preference Optimization + +After obtaining the SFT model, we further enhance its capability in optimizing text-to-video generation prompts. Following our proposed principles, we incorporate two types of feedback. Text-level feedback assesses whether prompts align with user intent and maintain safety, while video-level feedback ensures that prompts result in high-quality video generation. Combining feedback from both dimensions, we can perform preference optimization on the model, allowing it to generate prompts that are safer, more accurate, and effective. + +Data Sampling. To support the DPO training, for each input $x$ , we first sample $K$ optimized prompts $(p_1, p_2, \ldots, p_K)$ from the SFT model. We then construct DPO training pairs based on both text-level and video-level feedback. + +Text-Level Preference Data Construction. A desired prompt optimizer should produce prompts that are harmless, faithfully represent the user's intent, and are clear and descriptive. To achieve this, in the preference optimization stage, we leverage an LLM-as-a-judge approach to provide text-level feedback. Given a user query and an optimized prompt pair $(x,p_{j})$ , the LLM generates a critique to check whether the prompt violates any proposed principles. If a flaw is detected, we refine the prompt based on the critique, producing a revised version $p_{j_{\mathrm{refined}}}$ . These pairs $(x,p_{j} < p_{j_{\mathrm{refined}}})$ comprise $D_{\mathrm{text}}$ . + +Video-Level Preference Data Construction. Ultimately, the helpfulness of a prompt should be judged by the quality of the video it helps generate. To guide the prompt optimizer in producing high-quality outputs, we utilize VisionReward [27], a state-of-the-art video reward model, to provide automated feedback. For prompts that pass the text-level check, we generate corresponding videos using the target video generation model. VisionReward then evaluates each video and assigns a reward score. This process results in $(x, p_m, r_m)$ , where $r_m$ is the score from VisionReward. By comparing these scores, we determine which prompts lead to better video quality. Assuming $r_m < r_{m+1}$ , we can then construct the video-level DPO pair, $(x, p_m < p_{m+1})$ , resulting in $D_{\mathrm{video}}$ . + +Model Training. As we get text-level and video-level preference pairs, we can conduct DPO on the SFT model. The training dataset is denoted as $D_{\mathrm{dpo}} = D_{\mathrm{text}} \cup D_{\mathrm{video}}$ and the DPO loss is described as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {D P O}} \left(\pi_ {\theta}; \pi_ {\text {r e f}}\right) = - \mathbb {E} _ {\left(x, p _ {w}, p _ {l}\right) \sim D _ {\text {d p o}}} \left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta} \left(p _ {w} \mid x\right)}{\pi_ {\text {r e f}} \left(p _ {w} \mid x\right)} \right. \right. \\ \left. - \beta \log \frac {\pi_ {\theta} \left(p _ {l} \mid x\right)}{\pi_ {\mathrm {r e f}} \left(p _ {l} \mid x\right)}\right) \Bigg ] \tag {2} \\ \end{array} +$$ + +where $\pi_{\theta}$ is the policy model initialized with the SFT model, and $\pi_{ref}$ is a fixed reference model also initialized with the SFT model. Here, $p_w$ and $p_l$ represent the chosen ( $p_{j_{\text{refined}}}$ from $D_{\text{text}}$ and $p_{m+1}$ from $D_{\text{video}}$ ) and rejected ( $p_j$ from $D_{\text{text}}$ and $p_m$ from $D_{\text{video}}$ ) prompts, respectively. This process results in the final DPO model, which is capable of generating prompts that improve video quality while maintaining safety and accuracy. + +# 4. Experiments + +# 4.1. Experimental Setup + +To thoroughly evaluate the effectiveness of VPO, we conduct a series of comprehensive experiments covering multiple aspects, including the main results on text-to-video benchmarks, text-level alignment, safety evaluation, and comparisons with RLHF methods for video generation models. Additionally, we assess the iterative improvement capability, as well as the generalization performance of VPO across different video generation models. Implementation details are described in the Appendix. + +Backbone Models & Baselines. Our experiments are conducted on several popular video generation models, including the CogVideoX series and Open-Sora 1.2. The CogVideoX series excel in producing coherent, motion-rich, long-duration videos, while Open-Sora 1.2 is an efficient open-source alternative. For CogVideoX, we use the official prompt optimization method based on GLM-4 in-context learning, along with GPT-4o in-context learning as baselines. For Open-Sora 1.2, we adopt the official GPT-4o rewriting method as our baseline. Additionally, we introduce an alternative baseline using the image prompt optimizer, Promptist [10]. Moreover, we compare VPO with Prompt-A-Video [13]. + +Evaluation Benchmarks. We use VBench [12] and MonetBench [27] for general text-to-video evaluation. VBench assesses multiple aspects, including Human Action, Scene, Multiple Objects, and Appearance Style. MonetBench evaluates diverse scenarios with metrics such as Alignment, Stability, Preservation, and Physics. For query alignment, we curate 500 diverse queries using LLM-based filtering and manual verification. GPT-4o is leveraged to assess whether the optimized prompts violate our predefined principles, such as containing harmful content or omitting key details. For safety, we use a subset of T2VSafetyBench [16]. More details are provided in the Appendix. + +# 4.2. Text-to-Video Benchmark Results + +As shown in Table 1, we implement VPO on CogVideoX-2B and CogVideoX-5B. In both cases, VPO significantly outperforms baseline approaches. The obvious gap between directly using the original query with using prompt optimization techniques highlights the critical role of prompt optimization in video generation. Furthermore, the difference between VPO-SFT and VPO models demonstrates the necessity of incorporating multiple feedback preference optimization. Notably, the superior performance of VPO over VPO without text-level feedback indicates that improving safety and accuracy at the text level enhances general text-to-video tasks. + +
MethodMonetBenchVBench
Align-mentStabilityPreservingPhysicsOverallHuman ActionSceneMultiple ObjectsAppear. Style
CogVideoX-2B
Original Query1.110.250.560.313.2780.0028.3440.1722.60
Promptist0.880.250.550.292.8767.4018.3727.4423.12
Prompt-A-Video1.230.270.610.333.5890.6043.8568.2622.33
GLM-4 Few-Shot1.280.270.590.333.5796.2055.5168.4023.47
GPT-4o Few-Shot1.260.270.580.333.5898.2052.5363.6323.73
VPO-SFT1.280.280.600.333.5997.0055.0468.9824.13
VPO w/o TL FDBK1.320.290.620.333.7296.4054.7867.7924.15
VPO1.340.290.630.343.7699.0055.8370.1724.20
CogVideoX-5B
Original Query1.310.290.620.343.7788.0041.3245.6723.37
Promptist1.080.280.620.333.4277.4024.9318.3423.27
Prompt-A-Video1.420.310.660.354.0591.8045.4074.4122.63
GLM-4 Few-Shot1.460.290.640.353.9898.4055.6072.3824.39
GPT-4o Few-Shot1.480.290.640.344.0399.2053.1372.2124.20
VPO-SFT1.470.300.650.354.0197.2058.4073.7024.55
VPO w/o TL FDBK1.520.310.670.354.1297.6054.5972.9923.96
VPO1.520.310.670.364.1599.6055.6875.7324.57
+ +Table 1. Main results on MonetBench and VBench (%). 'Appear. Style' stands for Appearance Style. VPO w/o TL FDBK indicates that only video-level rewards are used during the preference optimization stage. The highest results for each video generation model is bolded. + +![](images/315068907e58ac17806c8fcba45a5e6537d6555bf5393dfe7e586890842f55fe.jpg) +Figure 3. Manual evaluation results comparing VPO with baseline methods. VPO demonstrates a significant advantage in instruction-following, video quality, and overall performance. + +![](images/a54233a2dc2740fa5c4d6e6097deaa35ad4849cdef920071b88ca5794b6019db.jpg) + +![](images/b9204a50c5f60121b1adfab0505dc5a58898fecc1394c5bdebf86ac36fb96378.jpg) + +To further validate the effectiveness of VPO, we conduct a pairwise human evaluation on CogVideoX-5B. We ask annotators to focus on the following aspects: + +- Instruction-Following: The degree to which the generated video adheres to the user query. +- Video Quality: The quality of generated videos, like coherence, stability, and adherence to physics. +- Overall Preference: A holistic judgment based on instruction-following and video quality. + +The results are shown in Figure 3, which are consistent with evaluation on benchmarks. VPO significantly sur + +passes other methods. The superior performance of VPO over VPO without text-level feedback on the alignment dimension indicates that a more accurate prompt optimizer could help increase the instruction-following abilities of video generation models. + +# 4.3. Text-Level Alignment + +We further evaluate VPO on a more complex test set constructed from real-world user queries, focusing on accuracy and harmlessness. As shown in Table 2, VPO largely enhances text-level alignment, surpassing alternative methods, including much larger LLMs. Importantly, our findings re + +
MethodAligned ↑Misaligned ↓
UnsafeImpreciseRefusal
GLM-4 Few-Shot83.45.410.01.2
GPT-4o Few-Shot86.42.48.62.6
VPO-SFT93.80.85.40.0
CogVideoX-2B
VPO w/o TL FDBK93.02.05.00.0
VPO94.60.64.80.0
CogVideoX-5B
VPO w/o TL FDBK92.81.26.00.0
VPO94.80.44.80.0
+ +Table 2. Evaluation results $(\%)$ of query alignment. We highlight the best results for each video generation model in bold. + +![](images/03c9e3106a3104cab9a99ffbed6d25824649cd5025d3291ad4a5ee54ff35ccaa.jpg) +Figure 4. Manual evaluation results on safety tasks. VPO demonstrates substantially improved harmlessness, particularly in terms of the completely safe rate (Level 1). + +veal a critical limitation of relying solely on LLM-based in-context learning for prompt optimization: these models may refuse to process queries containing sensitive keywords or abstract contents, such as "20 - 11 coins = 9 coins." This underscores the necessity of training a comprehensive prompt optimizer capable of dealing with diverse user queries. + +# 4.4. Safety Evaluation + +We conduct a manual safety evaluation using a subset of T2VSafetyBench, where annotators assign safety scores to generated videos on a four-point scale: + +- Level 1: Completely safe. +- Level 2: Safe but contains minor unsafe elements. +- Level 3: Unsafe. +- Level 4: Extremely unsafe. + +As shown in Figure 4, VPO substantially improves video safety compared to LLM few-shot methods. This enhancement is particularly evident at Level 1, where VPO ensures a significantly higher proportion of completely safe outputs. + +![](images/0bf4016782b666b2d066f2e9a86eb3483dd7201018bccecc60f32d702ae78258.jpg) +Figure 5. Pairwise evaluation results using VisionReward. Standard refers to the original video generation model with the official prompt rewriting method. VPO outperforms Diffusion DPO and can be combined with it to provide additional benefits. + +Importantly, incorporating text-level feedback further enhances safety, whereas neglecting this component leads to a lower safety rating than the SFT model, indicating that solely optimizing for video quality could damage safety. + +# 4.5. Comparison to RLHF + +We reveal that the prompt optimization method is not only important to bridge the training and inference gap of video generation models, but it is also a good way to align the video generation model with human preference. In this experiment, we compare VPO with Diffusion DPO [25]. Diffusion DPO is applied with the official GLM-4 prompt rewriting. Using VisionReward for pairwise evaluation, we find that both methods enhance CogVideoX-5B's performance, with VPO outperforming Diffusion DPO, as shown in Figure 5. Notably, VPO and Diffusion DPO can be combined for additional gains, demonstrating that optimizing both the video generation model and prompt optimizer can lead to superior results. + +# 4.6. Generalization across Models + +As the preference optimization stage of VPO relies on a specific video generation model to construct DPO pairs, we would like to investigate if the prompt optimizer could generalize to other models. To test this, we apply the prompt optimizer trained on CogVideoX-2B to Open-Sora 1.2. As shown in Table 3 and Table 4, we demonstrate that VPO could generalize across different video generation models, yielding non-trivial performance improvements on both VBenchmark and MonetBench. + +# 5. Case Study + +Figure 6 presents a comparative analysis of VPO and the GLM-4 Few-Shot approach in handling both harmful and harmless queries. On the left, the original query describes a + +![](images/785c2e831bb7edf06b9c1c48f2dc68a29f5702b983d91d55501cc3e3a0030a0b.jpg) +Original Query: In a dimly lit hospital room, patients with exposed wounds and broken bones lie motionless, the floor slick with blood and medical waste. + +![](images/aec504165cbf4ef1c8c8d197f6f539757a158fdb10bbe11bdbe440a5a5541c16.jpg) +GLM-4 Few-Shot: In a somber, dimly lit hospital ward, patients with various injuries—exposed lacerations, bruised limbs, and splinted fractures……, which is wet and shiny with patches of congealed blood and scattered medical debris. + +![](images/875e4bf4cedba992373c2b642b373eccaa3453b184375d40e82ca9f78d0e5558.jpg) +VPO (Ours): In a dimly lit hospital room, the scene focuses on the caring and professional environment. creating a calm and sterile interior. The professionals move with precision and compassion, reflecting a workplace dedicated to healing and care, capturing the resilience and professionalism of healthcare workers. +Figure 6. (Left) Case study on a harmful query; (Right) Case study on a harmless query. Some prompts are omitted due to space constraints. + +![](images/6038611e15a77631c28ddc1688458103af28bb77cefedcb95c59fcb9920bb4ea.jpg) +Original Query: A person is hugging. + +![](images/8164d550745491e89cb68674fe4a88f80f30d6c5cfc9e61e06d0da9db2e11017.jpg) +GLM-4 Few-Shot: A person, enveloped in a cozy knit sweater, stands in the heart of a lush garden where the vibrant colors of autumn leaves cascade around them. They are seen opening their arms wide, a warm and inviting gesture. As they...... + +![](images/17cf9750610a364db2b0a79e376446c5285cfe02d164422972feec02aeab0c38.jpg) +VPO (Ours): A young woman with shoulder-length auburn hair is seen in a cozy living room with warm lighting, wearing a light blue sweater and jeans. She stands close to a man with short black hair, dressed in a green flannel shirt and dark trousers. They embrace each other warmly in a friendly and affectionate manner.... + +
MethodHuman ActionSceneMultiple ObjectsAppear. Style
Original Query88.8044.0855.9923.87
GPT-4o Few-Shot92.4053.2165.0223.84
VPO-SFT95.8051.4164.2823.86
VPO97.0053.5867.8823.93
+ +Table 3. Evaluation results (%) of Open-Sora 1.2 on VBench. + +
MethodAlign-mentStabilityPreser-vationPhysicsOverall
Original Query1.010.210.540.312.90
GPT-4o Few-Shot1.120.210.550.313.07
VPO-SFT1.120.230.560.313.08
VPO1.130.240.580.333.18
+ +Table 4. Evaluation results of Open-Sora 1.2 on MonetBench. + +disturbing hospital scene with exposed wounds, blood, and medical waste. The few-shot method generates a detailed description of these unsettling elements, leading to a harmful video. In contrast, VPO shifts the focus to a professional and sterile hospital environment, emphasizing care and resilience rather than distressing imagery. This demonstrates VPO's ability to refine harmful prompts into ethical and appropriate content while maintaining contextual relevance. On the right, we examine a harmless query about hugging. While the few-shot approach generates a semantically rich scene, the resulting video suffers from poor quality, with unnatural body movements and misplaced limbs. In comparison, VPO produces a more natural and visually coherent video, with smoother and more realistic movements. This highlights the importance of integrating feedback from gen + +erated videos to build a helpful prompt optimizer. These cases show that VPO not only improves video quality but also ensures that the generated content remains safe and appropriate. More examples are provided in the Appendix. + +# 6. Conclusion + +In this work, we introduce VPO, a principle-driven framework for optimizing prompts in video generation, ensuring harmlessness, accuracy, and helpfulness. Unlike conventional in-context learning approaches, which rewrite prompts without considering their actual effectiveness, VPO improves video generation quality through principle-based SFT and multi-feedback preference optimization, incorporating both text-level and video-level feedback. Through extensive experiments, we demonstrate the superiority of VPO over the baseline methods, delivering safer, more accurate, and higher-quality video generation, effectively bridging the gap between training and inference in video generation models. Moreover, we reveal that VPO also serves as an effective RLHF method and can be integrated with RLHF methods on video models, bringing orthogonal improvements. Additionally, we demonstrate the generalization of VPO across video models, showing the potential for developing a general text-to-video prompt optimization model. + +# 7. Acknowledgement + +This work was supported by the National Science Foundation for Distinguished Young Scholars (with No. 62125604). This work was also supported by Tsinghua University Initiative Scientific Research Program. We would + +also like to thank Zhipu AI for sponsoring GPU computing and API cost consumed in this study. + +# References + +[1] Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, et al. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861, 2021. 1 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 3 +[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3 +[6] Jiale Cheng, Xiao Liu, Kehan Zheng, Pei Ke, Hongning Wang, Yuxiao Dong, Jie Tang, and Minlie Huang. Blackbox prompt optimization: Aligning large language models without model training. arXiv preprint arXiv:2311.04155, 2023. 3 +[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 11 +[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3 +[9] Laura Hanu and Unitary team. Detoxify. Github. https://github.com/unitaryai/detoxify, 2020.4 +[10] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36:66923-66939, 2023. 5 +[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3 +[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5 +[13] Yatai Ji, Jiacheng Zhang, Jie Wu, Shilong Zhang, Shoufa Chen, Chongjian GE, Peize Sun, Weifeng Chen, Wenqi Shao, Xuefeng Xiao, et al. Prompt-a-video: Prompt your video diffusion model via preference-aligned llm. arXiv preprint arXiv:2412.15156, 2024. 3, 5 +[14] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 3 +[15] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 3 +[16] Yibo Miao, Yifan Zhu, Yinpeng Dong, Lijia Yu, Jun Zhu, and Xiao-Shan Gao. T2vsafetybench: Evaluating the safety of text-to-video generative models. arXiv preprint arXiv:2407.05965, 2024. 5 +[17] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3 +[18] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1, 3 +[19] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 11 +[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener + +ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3 +[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[22] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222-4235, 2020. 3 +[23] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3 +[24] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[25] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 7 +[26] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. arXiv preprint arXiv:2403.06098, 2024. 3 +[27] Jiazheng Xu, Yu Huang, Jiale Cheng, Yuanming Yang, Jiajun Xu, Yuan Wang, Wenbo Duan, Shen Yang, Qunlin Jin, Shurun Li, et al. Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. arXiv preprint arXiv:2412.21059, 2024. 5 +[28] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3 +[29] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3 +[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations, 2023. 3 +[31] Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. Texygen: A benchmarking platform for text generation models. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 1097-1100, 2018. 4 \ No newline at end of file diff --git a/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/images.zip b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..144d30b93d1eeabf1323e41fe09acf3b4dfe37b5 --- /dev/null +++ b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a90f939010d36068d1ab9b33f6dcd611751de4b8c2c49643de36e490493b857 +size 666758 diff --git a/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/layout.json b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d7718efbf06c1eb51a06e1c594ecfc7cfa1db0d5 --- /dev/null +++ b/2025/VPO_ Aligning Text-to-Video Generation Models with Prompt Optimization/layout.json @@ -0,0 +1,7163 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 539, + 121 + ], + "type": "text", + "content": "VPO: Aligning Text-to-Video Generation Models with Prompt Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": "Jiale Cheng" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Ruiliang Lyu" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Xiaotao Gu" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Xiao Liu" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Jiazheng Xu" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Yida Lu" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Jiayan Teng" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Zhuoyi Yang" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Yuxiao Dong" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Jie Tang" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Hongning Wang" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "text", + "content": ", Minlie Huang" + }, + { + "bbox": [ + 75, + 142, + 538, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 185, + 495, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 185, + 495, + 227 + ], + "spans": [ + { + "bbox": [ + 113, + 185, + 495, + 227 + ], + "type": "text", + "content": "1The Conversational Artificial Intelligence (CoAI) Group, Tsinghua University \n2Zhipu AI 3The Knowledge Engineering Group (KEG), Tsinghua University \nchengjl23@mails.tsinghua.edu.cn, aihuang@tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 282, + 297, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 297, + 666 + ], + "type": "text", + "content": "Video generation models have achieved remarkable progress in text-to-video tasks. These models are typically trained on text-video pairs with highly detailed and carefully crafted descriptions, while real-world user inputs during inference are often concise, vague, or poorly structured. This gap makes prompt optimization crucial for generating high-quality videos. Current methods often rely on large language models (LLMs) to refine prompts through in-context learning, but suffer from several limitations: they may distort user intent, omit critical details, or introduce safety risks. Moreover, they optimize prompts without considering the impact on the final video quality, which can lead to suboptimal results. To address these issues, we introduce VPO, a principled framework that optimizes prompts based on three core principles: harmlessness, accuracy, and helpfulness. The generated prompts faithfully preserve user intents and, more importantly, enhance the safety and quality of generated videos. To achieve this, VPO employs a two-stage optimization approach. First, we construct and refine a supervised fine-tuning (SFT) dataset based on principles of safety and alignment. Second, we introduce both text-level and video-level feedback to further optimize the SFT model with preference learning. Our extensive experiments demonstrate that VPO significantly improves safety, alignment, and video quality compared to baseline methods. Moreover, VPO shows strong generalization across video generation models. Furthermore, we demonstrate that VPO could outperform and be combined with RLHF methods on video generation models, underscoring the effectiveness of VPO in aligning video generation models. Our code and data are publicly available at https://github.com/thu-coai/VPO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 255, + 394, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 255, + 394, + 267 + ], + "spans": [ + { + "bbox": [ + 315, + 255, + 394, + 267 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 277, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 408 + ], + "type": "text", + "content": "Recent advancements in text-to-video generation models have significantly improved the ability to produce high-quality video content [4, 15, 18, 28, 29]. These models typically rely on extensive, high-quality video-text annotated datasets, where detailed labels capture specific video content. While training descriptions are detailed and carefully crafted, real-world user inputs are often brief, ambiguous, or poorly structured. This critical gap between model training and inference can lead to suboptimal outputs, making it crucial to optimize and refine user inputs to generate desired high-quality video outputs (Figure 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 410, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 456 + ], + "type": "text", + "content": "Current approaches to optimizing user inputs for video generation mostly utilize large language models (LLMs) with in-context learning [15, 28, 29]. However, these methods suffer from several limitations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 459, + 554, + 579 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 459, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 459, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 459, + 554, + 495 + ], + "type": "text", + "content": "- Safety concerns: Existing methods do not explicitly ensure that optimized prompts remain safe, potentially leading to inappropriate or harmful video content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 495, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 530 + ], + "type": "text", + "content": "- Imprecise refinement: Current methods may unintentionally alter the user's intent or introduce biases, leading to outputs that deviate from the original query." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 531, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 579 + ], + "type": "text", + "content": "- Neglecting final video quality: Current approaches optimize prompts to be semantically richer, but do not explicitly consider how the refined prompt impacts the actual generated video, limiting their practical effectiveness." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 580, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 554, + 640 + ], + "type": "text", + "content": "To address these challenges, we propose a set of alignment principles for video prompt optimization, inspired by the HHH (Harmless, Honest, Helpful) principle in LLMs [1]. An aligned prompt optimization model should adhere to the following:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": "- Harmless: The refined prompt should avoid any harmful content, such as depictions of blood or violence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": "- Accurate: The optimized prompt should be precisely aligned with the user input, except for safety issues." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "- Helpful: The refined prompt should be detailed and descriptive, and can help the video generation model pro" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 233, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 233, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 233, + 703 + ], + "type": "text", + "content": "*Work done when JC and YL interned at Zhipu AI." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15636" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 68, + 541, + 285 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 541, + 285 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 541, + 285 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 541, + 285 + ], + "type": "image", + "image_path": "fefd15023042662f688c12aabf59840b71143e0cba7117c58820e96a620edcaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 294, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 350 + ], + "type": "text", + "content": "Figure 1. (Upper) The critical gap between the training and inference stages in video generation models. During training, video descriptions are detailed, descriptive, and well-crafted, while user inputs in the inference stage are often concise, unclear, and poorly structured, creating a mismatch that hampers model performance. (Lower) Comparison of VPO with traditional prompt optimization methods. Traditional methods rely on the in-context learning capabilities of LLMs, which can lead to issues such as safety issues, misalignment, and low-quality videos. In contrast, VPO serves as a harmless, accurate, and helpful prompt optimizer for high-quality video generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 371, + 196, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 371, + 196, + 383 + ], + "spans": [ + { + "bbox": [ + 64, + 371, + 196, + 383 + ], + "type": "text", + "content": "duce high-quality video content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 674 + ], + "type": "text", + "content": "Building on these principles, we introduce VPO, a principle-driven framework for optimizing user inputs for text-to-video generation, which integrates two critical stages: principle-based supervised fine-tuning (SFT) and multi-feedback preference optimization. In the SFT stage, we first leverage the in-context learning capabilities of LLMs to construct an initial SFT dataset. This data is then improved through the application of principle-based critique and refinement, ensuring the generated prompts are safe, accurately aligned with the original query, and sufficiently detailed. The resulting fine-tuned model poses a foundational ability to generate prompts following the proposed principles. In the subsequent preference optimization phase, we incorporate both text-level feedback and video-level reward signals to further improve the SFT model. By combining these feedback sources, we construct preference data that prioritizes harmlessness, accuracy, and helpfulness in the optimized prompt. Using Direct Preference Optimization (DPO), we train the model to generate prompts that consistently result in safer, more precise, and higher-quality video outputs. This two-stage framework explicitly integrates alignment principles into the prompt optimization process, bridging the critical gap between model training and real-world inference in video generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We have conducted experiments on several popular video generation models, including CogVideoX [28] and Open-Sora [29]. Through extensive experiments, we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 371, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 598 + ], + "type": "text", + "content": "demonstrate the effectiveness of VPO in generating harmless, accurate, and helpful prompts for high-quality video content, outperforming traditional prompt rewriting methods. Notably, on CogVideoX, VPO improves the win rate of " + }, + { + "bbox": [ + 313, + 371, + 555, + 598 + ], + "type": "inline_equation", + "content": "37.5\\%" + }, + { + "bbox": [ + 313, + 371, + 555, + 598 + ], + "type": "text", + "content": " over original user queries and " + }, + { + "bbox": [ + 313, + 371, + 555, + 598 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 313, + 371, + 555, + 598 + ], + "type": "text", + "content": " over the official prompt optimization method in human evaluation. Moreover, VPO significantly reduces the unsafety rate, enhancing the overall harmlessness of video generation models. We also highlight the necessity of multi-feedback preference optimization, demonstrating its crucial role in refining prompts for improved video generation. Furthermore, we demonstrate the potential of VPO as an RLHF method. VPO not only surpasses Diffusion DPO [25], but also brings additional benefits over Diffusion DPO. Additionally, VPO exhibits good generalization capabilities across different models—when trained on CogVideoX, it enhances the performance of Open-Sora 1.2, indicating the potential for training a general prompt optimizer for diverse video generation models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 602, + 525, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 602, + 525, + 613 + ], + "spans": [ + { + "bbox": [ + 325, + 602, + 525, + 613 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 618, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 618, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 618, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 618, + 555, + 689 + ], + "type": "text", + "content": "- We propose the principle of the alignment of video prompt optimizer: harmless, accurate, and helpful. Based on this, we introduce VPO, a systematic framework that integrates principle-based supervised fine-tuning and multi-feedback preference optimization, to build an aligned video prompt optimizer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 556, + 714 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of VPO for improving query refinement accuracy, enhancing the safety of" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15637" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 294, + 119 + ], + "type": "text", + "content": "generated videos, and boosting the overall performance of video generation models. Furthermore, VPO exhibits good generalization across different video generation models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 120, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 295, + 156 + ], + "type": "text", + "content": "- We demonstrate that VPO is a novel and competitive alignment method for text-to-video models, outperforming Diffusion DPO and providing orthogonal benefits." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 142, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 142, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 142, + 181 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 189, + 160, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 160, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 160, + 201 + ], + "type": "text", + "content": "2.1. Video Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 207, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 495 + ], + "type": "text", + "content": "With the widespread application of diffusion models in the domain of image generation [20, 21], research efforts have increasingly shifted towards video generation models [3, 8, 23]. Video Diffusion Models [11] firstly extend image diffusion models to video generation by employing a 3D UNet architecture. Text2Video-Zero [14] involves motion dynamics in image generation models to enable zero-shot video generation. However, these approaches encounter significant challenges in generating long videos. Subsequent studies, such as Stable Video Diffusion [2] and VideoCrafter [5], leverage large-scale pre-training on high-quality video datasets, yet their scalability remains constrained due to the limitation of architectures like UNet. By adopting Transformers [24] as the backbone of diffusion models, the scalable architecture of DiT [17] has substantially enhanced the capability of video generation. Building on DiT, several impressive video generation models have emerged [4, 15, 18, 28]. Nonetheless, the training data for these video generation models typically consists of detailed video descriptions, whereas user inputs during inference are often brief, unstructured, or even ambiguous in intent. Bridging this gap is critical for improving the quality of generated videos, underscoring the importance of training a robust prompt optimization model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 179, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 179, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 179, + 517 + ], + "type": "text", + "content": "2.2. Prompt Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Prompt optimization techniques have long been a pivotal research problem for generation models. Early work on automatic prompt optimization can be traced back to AutoPrompt [22], which utilizes gradient-guided search to automatically create prompt templates. With the rapid development of LLMs, there has been a significant increase in studies leveraging LLMs for automatic prompt optimization [6, 30]. Similarly, existing video prompt optimization methods employ LLMs to refine user inputs, typically through in-context learning [15, 28, 29]. However, these methods rely solely on the inherent capabilities of LLMs, without considering the alignment with user query and the quality of generated videos, thus often resulting in suboptimal video outputs. Prompt-A-Video [13] incorporates image and video reward models during the optimization process to enhance the quality of generated videos, but it fails to ac" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": "count for text-level alignment, which is essential for ensuring the safety and accuracy of rewritten prompts. To address these limitations, we propose VPO, a systematic framework for constructing a video prompt optimization model that can safely and accurately capture user intent, and assist video generation models in generating high-quality videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 156, + 370, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 370, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 370, + 168 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 177, + 555, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 309 + ], + "type": "text", + "content": "We introduce VPO, an automated framework designed to bridge the gap between real-world user inputs and the well-crafted text descriptions used during model training. Unlike existing approaches that rely solely on the in-context learning capabilities of LLMs, VPO follows a principle-driven approach, ensuring that prompt refinements are harmless, accurate, and helpful. By incorporating these core principles into both supervised fine-tuning and preference optimization, VPO effectively transforms concise, unclear, or unsafe user inputs into detailed, well-structured, and safe prompts that significantly enhance video generation quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 317, + 430, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 430, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 430, + 329 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 526 + ], + "type": "text", + "content": "The overall framework of VPO is illustrated in Figure 2. Our framework consists of two key stages: Principle-Based SFT (" + }, + { + "bbox": [ + 313, + 335, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\S 3.2" + }, + { + "bbox": [ + 313, + 335, + 555, + 526 + ], + "type": "text", + "content": ") and Multi-Feedback Preference Optimization (" + }, + { + "bbox": [ + 313, + 335, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\S 3.3" + }, + { + "bbox": [ + 313, + 335, + 555, + 526 + ], + "type": "text", + "content": "). In the first stage, Principle-Based SFT, we construct a high-quality dataset to train an initial model that refines user inputs while adhering to our three core principles. This stage provides the model with the foundational capabilities to optimize user input into a harmless, accurate, and helpful prompt for text-to-video generation. In the second stage, Multi-Feedback Preference Optimization, we further improve the SFT model using both text-level and video-level feedback. Text-level feedback helps refine prompts to be well-aligned with user intent while maintaining harmless. Meanwhile, video-level feedback ensures that the generated prompts can lead to high-quality video outputs, effectively enhancing the performance of the text-to-video model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 535, + 434, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 434, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 434, + 547 + ], + "type": "text", + "content": "3.2. Principle-Based SFT" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 553, + 601 + ], + "type": "text", + "content": "We first construct a high-quality SFT dataset comprising user query and optimized prompt pairs to initialize the prompt optimizer with the ability to generate prompts that align with our guiding principles." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Query Curation. To build the SFT dataset, we begin by collecting a comprehensive set of user queries. We utilize the VidProM dataset [26], which contains over one million real-world text-to-video queries. We adopt its deduplicated version and apply additional filtering steps to enhance the data. First, we perform rule-based quality filtering, considering factors such as keywords, special characters, and query length. To ensure query diversity, we employ the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15638" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 223, + 266 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 223, + 266 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 223, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 223, + 266 + ], + "type": "image", + "image_path": "31b9860833d98c369d137d0de96b2da5aaaebbe14c531057a11b3ce3de45db2c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 274, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 309 + ], + "type": "text", + "content": "Figure 2. The overall framework of VPO. VPO consists of two main steps: Principle-Based SFT and Multi-Feedback Preference Optimization. By refining user queries into harmless, accurate, and helpful prompts, VPO acts as a bridge between users and video generation models, ultimately enhancing the quality of generated videos." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 226, + 69, + 541, + 266 + ], + "blocks": [ + { + "bbox": [ + 226, + 69, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 226, + 69, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 226, + 69, + 541, + 266 + ], + "type": "image", + "image_path": "5cee41c894a273e33c8e58906d0b54b5ac3ca38b6046145c7b30a67e1bc1fcca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 329, + 297, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 297, + 449 + ], + "type": "text", + "content": "self-BLEU [31] metric to filter out overly similar queries. Additionally, since we aim to develop a prompt optimizer that prioritizes safety, we incorporate safety-related queries into our dataset. Specifically, we extract prompts labeled as unsafe in the original dataset. However, as Detoxify [9] introduces labeling errors, we re-evaluate these safety-related queries using a more robust LLM. Following this data curation process, we finalize a dataset consisting of approximately 18k general queries and 2k safety-related queries for training, where 10k are used for SFT and 10k for DPO." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 467, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 467, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 54, + 467, + 296, + 598 + ], + "type": "text", + "content": "Optimized Prompts Construction. With a diverse and high-quality set of user queries, we employ an LLM, specifically GPT-4o, with in-context learning to generate initial optimized prompts. The demonstrations used in this process are carefully crafted to guide the LLM in producing clear, well-structured prompts. Additionally, we instruct the LLM to account for potential safety concerns, ensuring that the generated prompts avoid harmful content, such as depictions of violence or explicit material. The exact prompt template used in this data construction process is provided in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "text", + "content": "Principle-Based Refinement. After generating query-prompt pairs " + }, + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "inline_equation", + "content": "(x,p)" + }, + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "text", + "content": " is the user query and " + }, + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 617, + 297, + 715 + ], + "type": "text", + "content": " is the optimized prompt, we further refine the prompts using a principle-driven approach to enhance their quality. We employ an LLM-as-a-judge method to assess prompts based on three key principles: harmlessness, accuracy, and helpfulness. The LLM critiques the generated prompts, identifying issues such as harmful content, missing key details from the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "content": "user query, or vague scenario descriptions. We then collect these critiques and represent problematic cases as triplets " + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "inline_equation", + "content": "(x,p,c)" + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "content": " is the critique. Based on this critique, we refine the generated prompt " + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "content": " to produce a refined version " + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{refined}}" + }, + { + "bbox": [ + 313, + 329, + 556, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "text", + "content": "Model Training. With the principle-based prompt refinement, we construct a high-quality SFT dataset, denoted as " + }, + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "inline_equation", + "content": "(x,s)\\in D_{\\mathrm{SFT}}" + }, + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "text", + "content": ". For the prompt without being pointed out issues within critiques, " + }, + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "inline_equation", + "content": "s = p" + }, + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "text", + "content": ", retaining the LLM-generated prompt. If issues are identified, we use the refined version, setting " + }, + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "inline_equation", + "content": "s = p_{\\mathrm{refined}}" + }, + { + "bbox": [ + 313, + 411, + 556, + 506 + ], + "type": "text", + "content": ". To initialize the prompt optimizer, we apply standard supervised fine-tuning with the loss function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 372, + 509, + 555, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 509, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 372, + 509, + 555, + 541 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log P (s | x, s _ {< i}), \\tag {1}", + "image_path": "8b164a768de263f673e776df5fe15c34a7ed438b5bdd9059e6967b0311719c0a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 551, + 459, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 459, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 459, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 551, + 459, + 563 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 551, + 459, + 563 + ], + "type": "text", + "content": " represents the length of " + }, + { + "bbox": [ + 313, + 551, + 459, + 563 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 551, + 459, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 574, + 529, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 529, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 529, + 588 + ], + "type": "text", + "content": "3.3. Multi-Feedback Preference Optimization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 712 + ], + "type": "text", + "content": "After obtaining the SFT model, we further enhance its capability in optimizing text-to-video generation prompts. Following our proposed principles, we incorporate two types of feedback. Text-level feedback assesses whether prompts align with user intent and maintain safety, while video-level feedback ensures that prompts result in high-quality video generation. Combining feedback from both dimensions, we can perform preference optimization on the model, allowing it to generate prompts that are safer, more accurate, and effective." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15639" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "Data Sampling. To support the DPO training, for each input " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", we first sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " optimized prompts " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "(p_1, p_2, \\ldots, p_K)" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " from the SFT model. We then construct DPO training pairs based on both text-level and video-level feedback." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "text", + "content": "Text-Level Preference Data Construction. A desired prompt optimizer should produce prompts that are harmless, faithfully represent the user's intent, and are clear and descriptive. To achieve this, in the preference optimization stage, we leverage an LLM-as-a-judge approach to provide text-level feedback. Given a user query and an optimized prompt pair " + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "inline_equation", + "content": "(x,p_{j})" + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "text", + "content": ", the LLM generates a critique to check whether the prompt violates any proposed principles. If a flaw is detected, we refine the prompt based on the critique, producing a revised version " + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "inline_equation", + "content": "p_{j_{\\mathrm{refined}}}" + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "text", + "content": ". These pairs " + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "inline_equation", + "content": "(x,p_{j} < p_{j_{\\mathrm{refined}}})" + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "text", + "content": " comprise " + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 150, + 296, + 284 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": "Video-Level Preference Data Construction. Ultimately, the helpfulness of a prompt should be judged by the quality of the video it helps generate. To guide the prompt optimizer in producing high-quality outputs, we utilize VisionReward [27], a state-of-the-art video reward model, to provide automated feedback. For prompts that pass the text-level check, we generate corresponding videos using the target video generation model. VisionReward then evaluates each video and assigns a reward score. This process results in " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "(x, p_m, r_m)" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "r_m" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": " is the score from VisionReward. By comparing these scores, we determine which prompts lead to better video quality. Assuming " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "r_m < r_{m+1}" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": ", we can then construct the video-level DPO pair, " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "(x, p_m < p_{m+1})" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{video}}" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "content": "Model Training. As we get text-level and video-level preference pairs, we can conduct DPO on the SFT model. The training dataset is denoted as " + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{dpo}} = D_{\\mathrm{text}} \\cup D_{\\mathrm{video}}" + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "content": " and the DPO loss is described as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 542, + 299, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 542, + 299, + 617 + ], + "spans": [ + { + "bbox": [ + 58, + 542, + 299, + 617 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\pi_ {\\theta}; \\pi_ {\\text {r e f}}\\right) = - \\mathbb {E} _ {\\left(x, p _ {w}, p _ {l}\\right) \\sim D _ {\\text {d p o}}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(p _ {w} \\mid x\\right)}{\\pi_ {\\text {r e f}} \\left(p _ {w} \\mid x\\right)} \\right. \\right. \\\\ \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(p _ {l} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(p _ {l} \\mid x\\right)}\\right) \\Bigg ] \\tag {2} \\\\ \\end{array}", + "image_path": "32c63fb2009b411a56eeee9ebbc8865ad5e5912a77a59fe15ec431662ce757f3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " is the policy model initialized with the SFT model, and " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pi_{ref}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " is a fixed reference model also initialized with the SFT model. Here, " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_w" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " represent the chosen (" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_{j_{\\text{refined}}}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\text{text}}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_{m+1}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\text{video}}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ") and rejected (" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\text{text}}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\text{video}}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ") prompts, respectively. This process results in the final DPO model, which is capable of generating prompts that improve video quality while maintaining safety and accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 430, + 105 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 219 + ], + "type": "text", + "content": "To thoroughly evaluate the effectiveness of VPO, we conduct a series of comprehensive experiments covering multiple aspects, including the main results on text-to-video benchmarks, text-level alignment, safety evaluation, and comparisons with RLHF methods for video generation models. Additionally, we assess the iterative improvement capability, as well as the generalization performance of VPO across different video generation models. Implementation details are described in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 230, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 386 + ], + "type": "text", + "content": "Backbone Models & Baselines. Our experiments are conducted on several popular video generation models, including the CogVideoX series and Open-Sora 1.2. The CogVideoX series excel in producing coherent, motion-rich, long-duration videos, while Open-Sora 1.2 is an efficient open-source alternative. For CogVideoX, we use the official prompt optimization method based on GLM-4 in-context learning, along with GPT-4o in-context learning as baselines. For Open-Sora 1.2, we adopt the official GPT-4o rewriting method as our baseline. Additionally, we introduce an alternative baseline using the image prompt optimizer, Promptist [10]. Moreover, we compare VPO with Prompt-A-Video [13]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 397, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 542 + ], + "type": "text", + "content": "Evaluation Benchmarks. We use VBench [12] and MonetBench [27] for general text-to-video evaluation. VBench assesses multiple aspects, including Human Action, Scene, Multiple Objects, and Appearance Style. MonetBench evaluates diverse scenarios with metrics such as Alignment, Stability, Preservation, and Physics. For query alignment, we curate 500 diverse queries using LLM-based filtering and manual verification. GPT-4o is leveraged to assess whether the optimized prompts violate our predefined principles, such as containing harmful content or omitting key details. For safety, we use a subset of T2VSafetyBench [16]. More details are provided in the Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 552, + 495, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 495, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 495, + 563 + ], + "type": "text", + "content": "4.2. Text-to-Video Benchmark Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "As shown in Table 1, we implement VPO on CogVideoX-2B and CogVideoX-5B. In both cases, VPO significantly outperforms baseline approaches. The obvious gap between directly using the original query with using prompt optimization techniques highlights the critical role of prompt optimization in video generation. Furthermore, the difference between VPO-SFT and VPO models demonstrates the necessity of incorporating multiple feedback preference optimization. Notably, the superior performance of VPO over VPO without text-level feedback indicates that improving safety and accuracy at the text level enhances general text-to-video tasks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15640" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 542, + 355 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 542, + 355 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 542, + 355 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 542, + 355 + ], + "type": "table", + "html": "
MethodMonetBenchVBench
Align-mentStabilityPreservingPhysicsOverallHuman ActionSceneMultiple ObjectsAppear. Style
CogVideoX-2B
Original Query1.110.250.560.313.2780.0028.3440.1722.60
Promptist0.880.250.550.292.8767.4018.3727.4423.12
Prompt-A-Video1.230.270.610.333.5890.6043.8568.2622.33
GLM-4 Few-Shot1.280.270.590.333.5796.2055.5168.4023.47
GPT-4o Few-Shot1.260.270.580.333.5898.2052.5363.6323.73
VPO-SFT1.280.280.600.333.5997.0055.0468.9824.13
VPO w/o TL FDBK1.320.290.620.333.7296.4054.7867.7924.15
VPO1.340.290.630.343.7699.0055.8370.1724.20
CogVideoX-5B
Original Query1.310.290.620.343.7788.0041.3245.6723.37
Promptist1.080.280.620.333.4277.4024.9318.3423.27
Prompt-A-Video1.420.310.660.354.0591.8045.4074.4122.63
GLM-4 Few-Shot1.460.290.640.353.9898.4055.6072.3824.39
GPT-4o Few-Shot1.480.290.640.344.0399.2053.1372.2124.20
VPO-SFT1.470.300.650.354.0197.2058.4073.7024.55
VPO w/o TL FDBK1.520.310.670.354.1297.6054.5972.9923.96
VPO1.520.310.670.364.1599.6055.6875.7324.57
", + "image_path": "61abdd188914ee82c86868958779a1b967cf06449e558028df17951daecc47ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 363, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 55, + 363, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 555, + 386 + ], + "type": "text", + "content": "Table 1. Main results on MonetBench and VBench (%). 'Appear. Style' stands for Appearance Style. VPO w/o TL FDBK indicates that only video-level rewards are used during the preference optimization stage. The highest results for each video generation model is bolded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 402, + 216, + 525 + ], + "blocks": [ + { + "bbox": [ + 59, + 402, + 216, + 525 + ], + "lines": [ + { + "bbox": [ + 59, + 402, + 216, + 525 + ], + "spans": [ + { + "bbox": [ + 59, + 402, + 216, + 525 + ], + "type": "image", + "image_path": "315068907e58ac17806c8fcba45a5e6537d6555bf5393dfe7e586890842f55fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 536, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 555, + 559 + ], + "type": "text", + "content": "Figure 3. Manual evaluation results comparing VPO with baseline methods. VPO demonstrates a significant advantage in instruction-following, video quality, and overall performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 402, + 382, + 525 + ], + "blocks": [ + { + "bbox": [ + 227, + 402, + 382, + 525 + ], + "lines": [ + { + "bbox": [ + 227, + 402, + 382, + 525 + ], + "spans": [ + { + "bbox": [ + 227, + 402, + 382, + 525 + ], + "type": "image", + "image_path": "a54233a2dc2740fa5c4d6e6097deaa35ad4849cdef920071b88ca5794b6019db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 394, + 403, + 550, + 525 + ], + "blocks": [ + { + "bbox": [ + 394, + 403, + 550, + 525 + ], + "lines": [ + { + "bbox": [ + 394, + 403, + 550, + 525 + ], + "spans": [ + { + "bbox": [ + 394, + 403, + 550, + 525 + ], + "type": "image", + "image_path": "b9204a50c5f60121b1adfab0505dc5a58898fecc1394c5bdebf86ac36fb96378.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 571, + 295, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 607 + ], + "type": "text", + "content": "To further validate the effectiveness of VPO, we conduct a pairwise human evaluation on CogVideoX-5B. We ask annotators to focus on the following aspects:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 295, + 683 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "type": "text", + "content": "- Instruction-Following: The degree to which the generated video adheres to the user query." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "text", + "content": "- Video Quality: The quality of generated videos, like coherence, stability, and adherence to physics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 295, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 683 + ], + "type": "text", + "content": "- Overall Preference: A holistic judgment based on instruction-following and video quality." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The results are shown in Figure 3, which are consistent with evaluation on benchmarks. VPO significantly sur" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 555, + 631 + ], + "type": "text", + "content": "passes other methods. The superior performance of VPO over VPO without text-level feedback on the alignment dimension indicates that a more accurate prompt optimizer could help increase the instruction-following abilities of video generation models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 634, + 440, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 440, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 440, + 647 + ], + "type": "text", + "content": "4.3. Text-Level Alignment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "We further evaluate VPO on a more complex test set constructed from real-world user queries, focusing on accuracy and harmlessness. As shown in Table 2, VPO largely enhances text-level alignment, surpassing alternative methods, including much larger LLMs. Importantly, our findings re" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15641" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 291, + 205 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 291, + 205 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 291, + 205 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 291, + 205 + ], + "type": "table", + "html": "
MethodAligned ↑Misaligned ↓
UnsafeImpreciseRefusal
GLM-4 Few-Shot83.45.410.01.2
GPT-4o Few-Shot86.42.48.62.6
VPO-SFT93.80.85.40.0
CogVideoX-2B
VPO w/o TL FDBK93.02.05.00.0
VPO94.60.64.80.0
CogVideoX-5B
VPO w/o TL FDBK92.81.26.00.0
VPO94.80.44.80.0
", + "image_path": "7b7919fe9a62e08b0e7337dbb8845d6f376a4d7e0e55c1af40ba3e0bbf49724e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "type": "text", + "content": "Table 2. Evaluation results " + }, + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "type": "text", + "content": " of query alignment. We highlight the best results for each video generation model in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 249, + 291, + 417 + ], + "blocks": [ + { + "bbox": [ + 62, + 249, + 291, + 417 + ], + "lines": [ + { + "bbox": [ + 62, + 249, + 291, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 249, + 291, + 417 + ], + "type": "image", + "image_path": "03c9e3106a3104cab9a99ffbed6d25824649cd5025d3291ad4a5ee54ff35ccaa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 432, + 295, + 466 + ], + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 466 + ], + "type": "text", + "content": "Figure 4. Manual evaluation results on safety tasks. VPO demonstrates substantially improved harmlessness, particularly in terms of the completely safe rate (Level 1)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 485, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 558 + ], + "type": "text", + "content": "veal a critical limitation of relying solely on LLM-based in-context learning for prompt optimization: these models may refuse to process queries containing sensitive keywords or abstract contents, such as \"20 - 11 coins = 9 coins.\" This underscores the necessity of training a comprehensive prompt optimizer capable of dealing with diverse user queries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 162, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 162, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 162, + 577 + ], + "type": "text", + "content": "4.4. Safety Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 617 + ], + "type": "text", + "content": "We conduct a manual safety evaluation using a subset of T2VSafetyBench, where annotators assign safety scores to generated videos on a four-point scale:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 618, + 266, + 665 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 618, + 171, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 618, + 171, + 629 + ], + "spans": [ + { + "bbox": [ + 56, + 618, + 171, + 629 + ], + "type": "text", + "content": "- Level 1: Completely safe." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 630, + 266, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 266, + 641 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 266, + 641 + ], + "type": "text", + "content": "- Level 2: Safe but contains minor unsafe elements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 642, + 134, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 134, + 652 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 134, + 652 + ], + "type": "text", + "content": "- Level 3: Unsafe." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 654, + 175, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 175, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 175, + 665 + ], + "type": "text", + "content": "- Level 4: Extremely unsafe." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "As shown in Figure 4, VPO substantially improves video safety compared to LLM few-shot methods. This enhancement is particularly evident at Level 1, where VPO ensures a significantly higher proportion of completely safe outputs." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 78, + 552, + 214 + ], + "blocks": [ + { + "bbox": [ + 318, + 78, + 552, + 214 + ], + "lines": [ + { + "bbox": [ + 318, + 78, + 552, + 214 + ], + "spans": [ + { + "bbox": [ + 318, + 78, + 552, + 214 + ], + "type": "image", + "image_path": "0bf4016782b666b2d066f2e9a86eb3483dd7201018bccecc60f32d702ae78258.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 229, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 275 + ], + "type": "text", + "content": "Figure 5. Pairwise evaluation results using VisionReward. Standard refers to the original video generation model with the official prompt rewriting method. VPO outperforms Diffusion DPO and can be combined with it to provide additional benefits." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 285, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 335 + ], + "type": "text", + "content": "Importantly, incorporating text-level feedback further enhances safety, whereas neglecting this component leads to a lower safety rating than the SFT model, indicating that solely optimizing for video quality could damage safety." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 341, + 440, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 440, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 440, + 354 + ], + "type": "text", + "content": "4.5. Comparison to RLHF" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 358, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 515 + ], + "type": "text", + "content": "We reveal that the prompt optimization method is not only important to bridge the training and inference gap of video generation models, but it is also a good way to align the video generation model with human preference. In this experiment, we compare VPO with Diffusion DPO [25]. Diffusion DPO is applied with the official GLM-4 prompt rewriting. Using VisionReward for pairwise evaluation, we find that both methods enhance CogVideoX-5B's performance, with VPO outperforming Diffusion DPO, as shown in Figure 5. Notably, VPO and Diffusion DPO can be combined for additional gains, demonstrating that optimizing both the video generation model and prompt optimizer can lead to superior results." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 474, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 474, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 474, + 533 + ], + "type": "text", + "content": "4.6. Generalization across Models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 539, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 556, + 647 + ], + "type": "text", + "content": "As the preference optimization stage of VPO relies on a specific video generation model to construct DPO pairs, we would like to investigate if the prompt optimizer could generalize to other models. To test this, we apply the prompt optimizer trained on CogVideoX-2B to Open-Sora 1.2. As shown in Table 3 and Table 4, we demonstrate that VPO could generalize across different video generation models, yielding non-trivial performance improvements on both VBenchmark and MonetBench." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 657, + 388, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 388, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 388, + 671 + ], + "type": "text", + "content": "5. Case Study" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Figure 6 presents a comparative analysis of VPO and the GLM-4 Few-Shot approach in handling both harmful and harmless queries. On the left, the original query describes a" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15642" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 303, + 113 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 303, + 113 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 303, + 113 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 303, + 113 + ], + "type": "image", + "image_path": "785c2e831bb7edf06b9c1c48f2dc68a29f5702b983d91d55501cc3e3a0030a0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 114, + 299, + 129 + ], + "lines": [ + { + "bbox": [ + 58, + 114, + 299, + 129 + ], + "spans": [ + { + "bbox": [ + 58, + 114, + 299, + 129 + ], + "type": "text", + "content": "Original Query: In a dimly lit hospital room, patients with exposed wounds and broken bones lie motionless, the floor slick with blood and medical waste." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 131, + 302, + 173 + ], + "blocks": [ + { + "bbox": [ + 58, + 131, + 302, + 173 + ], + "lines": [ + { + "bbox": [ + 58, + 131, + 302, + 173 + ], + "spans": [ + { + "bbox": [ + 58, + 131, + 302, + 173 + ], + "type": "image", + "image_path": "aec504165cbf4ef1c8c8d197f6f539757a158fdb10bbe11bdbe440a5a5541c16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 174, + 297, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 174, + 297, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 174, + 297, + 198 + ], + "type": "text", + "content": "GLM-4 Few-Shot: In a somber, dimly lit hospital ward, patients with various injuries—exposed lacerations, bruised limbs, and splinted fractures……, which is wet and shiny with patches of congealed blood and scattered medical debris." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 201, + 302, + 243 + ], + "blocks": [ + { + "bbox": [ + 58, + 201, + 302, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 302, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 302, + 243 + ], + "type": "image", + "image_path": "875e4bf4cedba992373c2b642b373eccaa3453b184375d40e82ca9f78d0e5558.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 243, + 293, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 243, + 293, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 243, + 293, + 277 + ], + "type": "text", + "content": "VPO (Ours): In a dimly lit hospital room, the scene focuses on the caring and professional environment. creating a calm and sterile interior. The professionals move with precision and compassion, reflecting a workplace dedicated to healing and care, capturing the resilience and professionalism of healthcare workers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 289, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 553, + 300 + ], + "type": "text", + "content": "Figure 6. (Left) Case study on a harmful query; (Right) Case study on a harmless query. Some prompts are omitted due to space constraints." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 311, + 70, + 552, + 114 + ], + "blocks": [ + { + "bbox": [ + 311, + 70, + 552, + 114 + ], + "lines": [ + { + "bbox": [ + 311, + 70, + 552, + 114 + ], + "spans": [ + { + "bbox": [ + 311, + 70, + 552, + 114 + ], + "type": "image", + "image_path": "6038611e15a77631c28ddc1688458103af28bb77cefedcb95c59fcb9920bb4ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 114, + 416, + 124 + ], + "lines": [ + { + "bbox": [ + 310, + 114, + 416, + 124 + ], + "spans": [ + { + "bbox": [ + 310, + 114, + 416, + 124 + ], + "type": "text", + "content": "Original Query: A person is hugging." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 131, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 131, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 131, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 553, + 173 + ], + "type": "image", + "image_path": "8164d550745491e89cb68674fe4a88f80f30d6c5cfc9e61e06d0da9db2e11017.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 174, + 544, + 198 + ], + "lines": [ + { + "bbox": [ + 307, + 174, + 544, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 544, + 198 + ], + "type": "text", + "content": "GLM-4 Few-Shot: A person, enveloped in a cozy knit sweater, stands in the heart of a lush garden where the vibrant colors of autumn leaves cascade around them. They are seen opening their arms wide, a warm and inviting gesture. As they......" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 201, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 307, + 201, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 201, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 553, + 243 + ], + "type": "image", + "image_path": "17cf9750610a364db2b0a79e376446c5285cfe02d164422972feec02aeab0c38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 243, + 549, + 277 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 549, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 549, + 277 + ], + "type": "text", + "content": "VPO (Ours): A young woman with shoulder-length auburn hair is seen in a cozy living room with warm lighting, wearing a light blue sweater and jeans. She stands close to a man with short black hair, dressed in a green flannel shirt and dark trousers. They embrace each other warmly in a friendly and affectionate manner...." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 58, + 318, + 293, + 396 + ], + "blocks": [ + { + "bbox": [ + 58, + 318, + 293, + 396 + ], + "lines": [ + { + "bbox": [ + 58, + 318, + 293, + 396 + ], + "spans": [ + { + "bbox": [ + 58, + 318, + 293, + 396 + ], + "type": "table", + "html": "
MethodHuman ActionSceneMultiple ObjectsAppear. Style
Original Query88.8044.0855.9923.87
GPT-4o Few-Shot92.4053.2165.0223.84
VPO-SFT95.8051.4164.2823.86
VPO97.0053.5867.8823.93
", + "image_path": "3a070f901d23c1e0c9067d73002a19913787cb543d7abd5cee5f2bad28c41f7a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 58, + 427, + 295, + 497 + ], + "blocks": [ + { + "bbox": [ + 63, + 403, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 63, + 403, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 63, + 403, + 287, + 415 + ], + "type": "text", + "content": "Table 3. Evaluation results (%) of Open-Sora 1.2 on VBench." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 427, + 295, + 497 + ], + "lines": [ + { + "bbox": [ + 58, + 427, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 58, + 427, + 295, + 497 + ], + "type": "table", + "html": "
MethodAlign-mentStabilityPreser-vationPhysicsOverall
Original Query1.010.210.540.312.90
GPT-4o Few-Shot1.120.210.550.313.07
VPO-SFT1.120.230.560.313.08
VPO1.130.240.580.333.18
", + "image_path": "d6820fb4759afcfafb9e26657c5eba6b0fe30c2d47c7bab519dbd0647a80ed66.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 505, + 288, + 516 + ], + "lines": [ + { + "bbox": [ + 63, + 505, + 288, + 516 + ], + "spans": [ + { + "bbox": [ + 63, + 505, + 288, + 516 + ], + "type": "text", + "content": "Table 4. Evaluation results of Open-Sora 1.2 on MonetBench." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "disturbing hospital scene with exposed wounds, blood, and medical waste. The few-shot method generates a detailed description of these unsettling elements, leading to a harmful video. In contrast, VPO shifts the focus to a professional and sterile hospital environment, emphasizing care and resilience rather than distressing imagery. This demonstrates VPO's ability to refine harmful prompts into ethical and appropriate content while maintaining contextual relevance. On the right, we examine a harmless query about hugging. While the few-shot approach generates a semantically rich scene, the resulting video suffers from poor quality, with unnatural body movements and misplaced limbs. In comparison, VPO produces a more natural and visually coherent video, with smoother and more realistic movements. This highlights the importance of integrating feedback from gen" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 320, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 370 + ], + "type": "text", + "content": "erated videos to build a helpful prompt optimizer. These cases show that VPO not only improves video quality but also ensures that the generated content remains safe and appropriate. More examples are provided in the Appendix." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 382, + 388, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 388, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 388, + 395 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 403, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 630 + ], + "type": "text", + "content": "In this work, we introduce VPO, a principle-driven framework for optimizing prompts in video generation, ensuring harmlessness, accuracy, and helpfulness. Unlike conventional in-context learning approaches, which rewrite prompts without considering their actual effectiveness, VPO improves video generation quality through principle-based SFT and multi-feedback preference optimization, incorporating both text-level and video-level feedback. Through extensive experiments, we demonstrate the superiority of VPO over the baseline methods, delivering safer, more accurate, and higher-quality video generation, effectively bridging the gap between training and inference in video generation models. Moreover, we reveal that VPO also serves as an effective RLHF method and can be integrated with RLHF methods on video models, bringing orthogonal improvements. Additionally, we demonstrate the generalization of VPO across video models, showing the potential for developing a general text-to-video prompt optimization model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 644, + 424, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 424, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 424, + 658 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "This work was supported by the National Science Foundation for Distinguished Young Scholars (with No. 62125604). This work was also supported by Tsinghua University Initiative Scientific Research Program. We would" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15643" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "also like to thank Zhipu AI for sponsoring GPU computing and API cost consumed in this study." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 115, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 115, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 115, + 118 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 126, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "type": "text", + "content": "[1] Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, et al. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 182, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 295, + 236 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 237, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 237, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 237, + 295, + 302 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 346 + ], + "type": "text", + "content": "[4] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 348, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 295, + 403 + ], + "type": "text", + "content": "[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "text", + "content": "[6] Jiale Cheng, Xiao Liu, Kehan Zheng, Pei Ke, Hongning Wang, Yuxiao Dong, Jie Tang, and Minlie Huang. Blackbox prompt optimization: Aligning large language models without model training. arXiv preprint arXiv:2311.04155, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 512 + ], + "type": "text", + "content": "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 568 + ], + "type": "text", + "content": "[8] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 570, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 295, + 591 + ], + "type": "text", + "content": "[9] Laura Hanu and Unitary team. Detoxify. Github. https://github.com/unitaryai/detoxify, 2020.4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "text", + "content": "[10] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. Advances in Neural Information Processing Systems, 36:66923-66939, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "text", + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "text", + "content": "[12] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 98, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 152 + ], + "type": "text", + "content": "[13] Yatai Ji, Jiacheng Zhang, Jie Wu, Shilong Zhang, Shoufa Chen, Chongjian GE, Peize Sun, Weifeng Chen, Wenqi Shao, Xuefeng Xiao, et al. Prompt-a-video: Prompt your video diffusion model via preference-aligned llm. arXiv preprint arXiv:2412.15156, 2024. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 154, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 154, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 553, + 220 + ], + "type": "text", + "content": "[14] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 221, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 221, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 221, + 553, + 275 + ], + "type": "text", + "content": "[15] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 277, + 553, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 322 + ], + "type": "text", + "content": "[16] Yibo Miao, Yifan Zhu, Yinpeng Dong, Lijia Yu, Jun Zhu, and Xiao-Shan Gao. T2vsafetybench: Evaluating the safety of text-to-video generative models. arXiv preprint arXiv:2407.05965, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 323, + 553, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 323, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 553, + 367 + ], + "type": "text", + "content": "[17] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 369, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 369, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 553, + 632 + ], + "type": "text", + "content": "[18] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[19] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 11" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15644" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 665 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 297, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 297, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 297, + 151 + ], + "type": "text", + "content": "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "text", + "content": "[22] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222-4235, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 272 + ], + "type": "text", + "content": "[23] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 295, + 296 + ], + "type": "text", + "content": "[24] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 363 + ], + "type": "text", + "content": "[25] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 295, + 397 + ], + "type": "text", + "content": "[26] Wenhao Wang and Yi Yang. Vidprom: A million-scale real prompt-gallery dataset for text-to-video diffusion models. arXiv preprint arXiv:2403.06098, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 295, + 453 + ], + "type": "text", + "content": "[27] Jiazheng Xu, Yu Huang, Jiale Cheng, Yuanming Yang, Jiajun Xu, Yuan Wang, Wenbo Duan, Shen Yang, Qunlin Jin, Shurun Li, et al. Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. arXiv preprint arXiv:2412.21059, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 295, + 508 + ], + "type": "text", + "content": "[28] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 295, + 552 + ], + "type": "text", + "content": "[29] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 554, + 295, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 295, + 608 + ], + "type": "text", + "content": "[30] Yongchao Zhou, Andrei Loan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 609, + 295, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 295, + 665 + ], + "type": "text", + "content": "[31] Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. Texygen: A benchmarking platform for text generation models. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 1097-1100, 2018. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15645" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_content_list.json b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..266995525e7d612efec197d77dddbb20753e0d0a --- /dev/null +++ b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_content_list.json @@ -0,0 +1,1642 @@ +[ + { + "type": "text", + "text": "VPR-Cloak: A First Look at Privacy Cloak Against Visual Place Recognition", + "text_level": 1, + "bbox": [ + 111, + 130, + 885, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuting Dong $^{1,2*}$ , Mingzhi Chen $^{3*}$ , Feng Lu $^{1,2}$ , Hao Yu $^{1}$ , Guanghao Li $^{1,3}$ , Zhe Wu $^{2}$ , Ming Tang $^{3\\dagger}$ , Chun Yuan $^{1\\dagger}$", + "bbox": [ + 272, + 179, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University, $^{2}$ Pengcheng Laboratory, $^{3}$ Southern University of Science and Technology", + "bbox": [ + 120, + 215, + 875, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dst21@mails.tsinghua.edu.cn, 12211414@mail.sustech.edu.cn,", + "bbox": [ + 240, + 236, + 751, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lf22@mails.tsinghua.edu.cn, longinyh@gmail.com, ligh24@mails.tsinghua.edu.cn,", + "bbox": [ + 156, + 253, + 836, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wuzh02@pcl.ac.cn, tangm3@sustech.edu.cn, yuanc@sz.tsinghua.edu.cn", + "bbox": [ + 205, + 272, + 785, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid advancement of Visual Place Recognition (VPR) systems, their unauthorized use on social media images enables monitoring of individuals' daily movements, posing serious privacy risks. However, privacy protection for addressing these risks in VPR systems remains an underexplored area. While adversarial perturbations have been widely explored for visual privacy protection, existing methods still fail to simultaneously satisfy the black-box constraint, imperceptibility, and real-time performance required in realistic VPR privacy protection scenarios. In this paper, we present the first look at privacy protection in VPR systems and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify decisive regions for place recognition and propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO) to selectively optimize perturbation generation in these areas. To enhance imperceptibility, we further optimize perturbations in the frequency domain, meticulously refining high-frequency components of perturbations while preserving low-frequency structures essential for human perception. Extensive experiments verify that our method outperforms existing SOTA methods. Additionally, our method achieves a $15 \\times$ speedup in runtime compared to SOTA methods. We also validate on Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios. The code is released at https://github.com/Dilemma-CMZ/VPR-Cloak.", + "bbox": [ + 88, + 354, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 792, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed major advances in visual place recognition (VPR) systems, which have been used in various visual applications. While these systems are designed", + "bbox": [ + 89, + 816, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c357ba79535aa82b63f790630b1646ee2c4607e925ca273671d9b40df20124e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 320, + 906, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a7f86dbde162acf07a30f8fbd355afb3f0a28b2f25ec6185bbf33c9056d180e.jpg", + "image_caption": [ + "Figure 1. Illustrations of (a) privacy leakage risk, and (b) privacy protection in visual place recognition." + ], + "image_footnote": [], + "bbox": [ + 517, + 503, + 903, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with good intentions to enhance localization and navigation, their growing capabilities have intensified privacy concerns. As shown in Fig. 1(a), with the widespread use of social media, images posted by users can be easily crawled and queried without permission, making it easier than ever to locate users' daily movements through place recognition, which brings huge privacy risks. Moreover, location data can often be used to directly or indirectly infer a user's true identity, living environment, or social relationships, further intensifying these privacy risks. Despite these significant", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 107, + 875, + 364, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7197", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "risks, to the best of our knowledge, no existing research has specifically addressed privacy protection in visual place recognition. This highlights the urgent need to develop robust privacy protection mechanisms for place recognition.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While no prior work has specifically addressed privacy protection for visual place recognition, adversarial perturbations [44-46, 53] have been widely explored as a general approach to visual privacy. These methods typically modify images to prevent unauthorized recognition systems from accurately identifying their content, either by introducing dense noise [14, 19, 43, 71] or applying patch-based modifications [5, 22, 48, 52]. However, for these perturbations to be effective, they often need to be highly noticeable, leading to visually abrupt and unnatural distortions. This makes it difficult to maintain image visual fidelity within ideal privacy protection for visual place recognition.", + "bbox": [ + 89, + 152, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, some recent works explore imperceptible perturbations [44, 45, 51, 61, 62], aiming at visual fidelity by targeting a limited number of pixels and yet achieving high success rates for privacy protection. However, a major drawback is their low transferability, meaning these perturbations often fail when transferred to different models, especially in black-box settings where the adversary has no access to the target model's parameters or architecture. To address this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent and leveraging a deep ensemble strategy. However, the long inference time of ANDA and MultiANDA makes them unsuitable for privacy protection of place recognition in real-world applications.", + "bbox": [ + 89, + 335, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the previous analysis, privacy protection for place recognition still remains an underexplored area, as it requires simultaneously addressing three critical real-world challenges: (1) diminishing the success rate of unauthorized VPR models which in real-world applications are often black-box, (2) maintaining high image visual quality to ensure imperceptibility, and (3) achieving near-real-time performance to support practical, real-world applications.", + "bbox": [ + 89, + 564, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge these challenges, we present a first look at privacy protection for place recognition and propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify the decisive regions for place recognition. Based on this, we propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO), which strategically optimizes perturbation generation in images by focusing on these decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way of generating the perturbation protects the most vulnerable areas of the image from unauthorized VPR models.", + "bbox": [ + 89, + 686, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, previous research [63] reveals that deep learning models are highly sensitive to high-frequency tex", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures, while the human visual system primarily perceives low-frequency structural information. Inspired by this, we further elevate perturbation optimization to the frequency domain by isolating the high-frequency components from the perturbation and meticulously refining them using the SAP-PO algorithm. This way ensures that only the high-frequency elements of the perturbation are optimized. As a result, we craft a perturbation that is both precisely targeted and remarkably potent. Through further optimization in the frequency domain, the final output perturbation image (i.e., the protected image) remains imperceptible to the human eye while maintaining a robust defense against VPR models. Our work brings the following contributions:", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a fast and effective network that provides privacy protection for location information in images while ensuring imperceptibility to the human eye. To the best of our knowledge, this is the first attempt to address privacy protection for visual place recognition.", + "- We introduce a SAP-PO algorithm, which strategically refines perturbations by focusing on decisive regions for place recognition.", + "- Extensive experiments on multiple benchmark datasets and on various black-box VPR models verify that our method outperforms existing SOTA methods. Additionally, our method achieves a $15 \\times$ speedup in runtime compared to SOTA methods.", + "- We validate the effectiveness of our privacy protection method based on using commercial APIs, including Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios." + ], + "bbox": [ + 513, + 287, + 903, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 558, + 661, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Place Recognition. With the rapid development of deep learning [4, 7-9, 23, 28, 40, 42, 55-59, 67-70], significant progress has been made in the VPR field [3, 15, 16, 20, 24]. Existing place recognition models have achieved impressive results not only in standard scenarios but also in challenging conditions such as radar scene [39], nighttime [31, 39], rainy [39], transport [21], perceptual aliasing [64], viewpoint changes [2, 41, 50], LiDAR scan [33, 60], occlusion [30, 41, 50], seasonal variation [41, 50], illumination change [32, 41, 50] dynamic objects. While these systems are designed with good intentions to enhance localization and navigation, they can be exploited by adversaries to expose users' location information, leading to potential risks of identity inference and tracking. Although previous work [26] attempts to attack image retrieval systems, it is not specifically designed for the VPR task.", + "bbox": [ + 511, + 583, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Perturbations Generation. While no prior research has specifically focused on privacy protection in visual place recognition, adversarial perturbations [44-46, 53] have been extensively studied as a general approach to visual privacy. Most existing adversarial methods impose norm", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7198", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based constraints when generating perturbations. Specifically, many studies [11, 25, 27, 36, 65, 66] adopt infinity-norm or L2-norm constraints, producing dense perturbations that alter a substantial proportion of pixels. Although these methods are effective, they often introduce visible and unnatural distortions, particularly at higher perturbation levels. To address this, recent studies have explored sparse perturbation techniques [6, 10, 12, 17, 34, 38, 65, 72], which restrict modifications to a limited subset of pixels while maintaining high attack success rates. For instance, $\\mathrm{PGD}_0$ [6] extends the projected gradient descent framework by enforcing a sparsity constraint. SparseFool [34], building on DeepFool [35], generates sparse untargeted perturbations. ANDA [13] and MultiANDA [13] enhance transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent. However, these methods tend to be resource-intensive and cost a long inference time.", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 375, + 181, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 400, + 294, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce real-world location privacy risks. As shown in Fig. 1 (a), the rise of social media makes it easier for adversaries to scrape user-posted photos. Using VPR systems, an adversary can infer where an image was taken. This enables monitoring users' daily movements, uncovering sensitive information about their habits, residences, and frequented places. Moreover, when large-scale user daily movements are monitored, location data can directly or indirectly reveal their identity and social connections, intensifying privacy risks.", + "bbox": [ + 89, + 421, + 483, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate these risks, as shown in Fig. 1 (b), we propose a protection transformation $F$ that converts a user's intended post photo $I$ into a protected representation $I_{p} = F(I)$ . Our goal is for $I_{p}$ to remain visually consistent with the original image $I$ while ensuring that adversaries cannot infer the photo's location using VPR systems. We refer to this process as location privacy protection, which can be formulated as an optimization problem:", + "bbox": [ + 89, + 573, + 483, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {F} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p r i v a c y}} \\left(I _ {p}\\right) + \\lambda \\mathcal {L} _ {\\text {v i s u a l}} \\left(I, I _ {p}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 704, + 482, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the visual loss $\\mathcal{L}_{\\mathrm{visual}}$ ensures that the transformed image remains visually perceptual consistent to the original image, while the privacy loss $\\mathcal{L}_{\\mathrm{privacy}}$ mislead VPR models in place recognition. The hyperparameter $\\lambda$ controls the trade-off between visual fidelity and location privacy. Formal definitions are provided in Sec. 3.3.", + "bbox": [ + 89, + 734, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenges. Privacy protection in visual place recognition within real-world scenarios presents several unique challenges: Firstly, the adversary's VPR models are often black-box in real-world scenarios, meaning their architectures, parameters, and training data are unknown, making it", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "difficult to mislead the black-box VPR models. Secondly, the protection transformation must preserve high visual fidelity and imperceptibility so that the generated protection image $I_{p}$ remains visually consistency from the original image $I$ . Thirdly, the protection method need be computationally efficient, enabling near-real-time processing to support practical use in real-world applications.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To the best of our knowledge, no existing approach effectively addresses all three of these challenges simultaneously for location privacy protection. Our goal is to develop a method that disrupts location inference while preserving image quality and maintaining practical efficiency, making it suitable for real-world deployment.", + "bbox": [ + 511, + 196, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Privacy Cloak Against VPR", + "text_level": 1, + "bbox": [ + 511, + 297, + 764, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous works [44, 45, 51, 61, 62] have explored adding imperceptible perturbations to images to enhance image privacy while aiming to be visually imperceptible. However, these methods often exhibit low transferability, which limits their effectiveness in black-box settings. To handle this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of perturbations. However, the high inference time of ANDA and MultiANDA limits the practicality in real-world applications.", + "bbox": [ + 511, + 319, + 903, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, we propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify decisive regions for place recognition. Based on the saliency-aware prior, we propose a SAP-PO algorithm, which strategically optimizes perturbation generation in images by focusing on decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way protects the vulnerable regions of the image from unauthorized VPR models. Additionally, we further enhance perturbation optimization in the frequency domain. By isolating and refining high-frequency components using SAP-PO, we craft a precisely targeted yet imperceptible perturbation that remains effective against VPR models while staying unnoticeable to the human eye.", + "bbox": [ + 511, + 455, + 905, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Saliency-Aware Prior Generation. Instead of applying perturbations randomly, we first identify critical regions for place recognition and selectively perturb them. To achieve this, we introduce a saliency-aware mechanism that acts as a spatial weight to pinpoint key areas for perturbation. For an input unprotected image $I$ , we extract this information by using the fine-tuned Vision Transformer (fine-tuned ViT), which employs the adapter strategy from [18, 31] to fine-tune the pre-trained DINOv2 backbone [37]. The input image is first resized using $\\mathcal{R}(I)$ to fit the fine-tuned ViT model's input size. Then, saliency-aware prior $P$ , is obtained by:", + "bbox": [ + 511, + 681, + 905, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP = \\mathcal {N} \\left(\\mathcal {A} _ {\\mathrm {V i T}} (\\mathcal {R} (I))\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 862, + 903, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(\\cdot)$ represents normalization to ensure the prior", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7199", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d880ebd1312e8494f071f27831b1a28a09d3677d7dcf4744d63659daa6d2e254.jpg", + "image_caption": [ + "Figure 2. Overview of our VPR-Cloak. We introduce a saliency-aware prior to accurately identify decisive regions for place recognition and propose the SAP-PO algorithm, which strategically generates perturbations focused on these decisive regions while minimizing changes in less relevant areas. To enhance imperceptibility, we further elevate perturbation optimization to the frequency domain by isolating and meticulously optimizing only the high-frequencies using the SAP-PO algorithm." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 890, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proper scaling and sum to 1. $\\mathcal{A}_{\\mathrm{ViT}}$ is the fine-tuned ViT.", + "bbox": [ + 89, + 433, + 460, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on this saliency-aware prior $P$ , we propose SAP-PO to optimize perturbation generation, ensuring perturbations are concentrated in decisive regions for place recognition. We provide two optimization strategies: spatial domain perturbation optimization for defense against VPR and frequency domain perturbation optimization for visual imperceptibility.", + "bbox": [ + 89, + 449, + 483, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Perturbation for Defense against VPR. In the spatial domain, SAP-PO, guided by the saliency prior, focuses on perturbing key regions crucial for place recognition, maximizing its impact on reducing recognition accuracy. The spatially-perturbed image $I_{\\mathrm{spa}}^{\\prime}$ is obtained as:", + "bbox": [ + 89, + 556, + 483, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {s p a}} ^ {\\prime} = \\mathcal {F} _ {\\mathrm {s p a}} (I, P), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 643, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\mathrm{spa}}$ is our proposed SAP-PO in spatial domain.", + "bbox": [ + 89, + 674, + 452, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frequency Perturbation for Imperceptible. Previous study [63] have shown that the human eye is sensitive to low frequencies, while deep learning models are sensitive to high frequencies. Inspired by this, we elevate perturbation optimization to the frequency domain by isolating and meticulously refining only the high-frequency components of the perturbation using the SAP-PO algorithm.", + "bbox": [ + 89, + 690, + 482, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To do this, we apply the Discrete Cosine Transform (DCT) to the perturbed image $I_{\\mathrm{spa}}^{\\prime}$ , yielding its frequency representation $f^{\\prime}$ . We specifically optimize the high-frequency components of the perturbation using SAP-PO and generate the frequency-perturbed image $I_{\\mathrm{fre}}^{\\prime}$ :", + "bbox": [ + 89, + 796, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {\\prime} = \\mathcal {D} \\left(I _ {\\mathrm {s p a}} ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {f r e}} ^ {\\prime} = \\mathcal {F} _ {\\mathrm {f r e}} \\left(f ^ {\\prime}, P\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 448, + 903, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}(\\cdot)$ denotes DCT. $\\mathcal{F}_{\\mathrm{fre}}$ is our proposed SAP-PO in frequency domain.", + "bbox": [ + 511, + 468, + 906, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we apply the inverse DCT to transform the frequency-perturbed image $I_{\\mathrm{fre}}^{\\prime}$ back to the spatial domain, obtaining the final protected image $I_{\\mathrm{p}}$ :", + "bbox": [ + 511, + 498, + 905, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {p}} = \\mathcal {D} ^ {- 1} \\left(I _ {\\mathrm {f r e}} ^ {\\prime}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 547, + 903, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}^{-1}$ is the inverse DCT.", + "bbox": [ + 511, + 571, + 723, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through optimization in the spatial and frequency domains, we strategically refine and enforce the distribution of perturbations to concentrate on decisive regions for place recognition and high-frequency areas.", + "bbox": [ + 511, + 585, + 905, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Saliency-Aware Prior Guided Perturbation Optimization", + "text_level": 1, + "bbox": [ + 511, + 654, + 905, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose SAP-PO to generate effective and visually imperceptible perturbations for defending VPR models. By integrating the saliency-aware prior $P$ , we focus perturbations on the decisive regions for place recognition.", + "bbox": [ + 511, + 691, + 905, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To begin, we initialize the perturbation as: $\\delta = 0$ . For an input image $I$ , we retrieve multiple positive reference images $\\{I_{\\mathrm{pos},i}\\}$ from the database and extract their features $\\{F_{\\mathrm{pos},i}\\}$ using the Surrogate VPR:", + "bbox": [ + 511, + 752, + 906, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{F _ {\\text {p o s}, i} \\right\\} = \\mathcal {M} _ {\\mathrm {V P R}} \\left(\\left\\{I _ {\\text {p o s}, i} \\right\\}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 816, + 903, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_{\\mathrm{VPR}}$ denotes the Surrogate VPR model. To enhance the robustness of our perturbations against unauthorized VPR models, for each input image, we randomly select one from multiple VPR models $\\{\\mathcal{M}_1,\\mathcal{M}_2,\\dots ,\\mathcal{M}_m\\}$", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7200", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO)", + "Ensure: Perturbed Image $I^{\\prime}$" + ], + "code_body": "Require: Image $I$ , Saliency-aware prior $P$ , Surrogate VPR Model $\\mathcal{M}_{\\mathrm{VPR}}$ , Trade-off Coefficient $\\lambda$ , Step Size $\\epsilon$ , Maximum Iterations $K$ , Domain Type $D_{\\mathrm{type}}$", + "guess_lang": "txt", + "bbox": [ + 91, + 125, + 483, + 170 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize perturbation $\\delta \\gets 0$", + "2: Retrieve multiple positive samples from database: $\\{I_{\\mathrm{pos},s}\\} \\gets$ retrieve_positives(I)", + "3: Generate $\\{F_{\\mathrm{pos},i}\\}$ using the Surrogate VPR model: $\\{F_{\\mathrm{pos},i}\\} \\gets \\mathcal{M}_{\\mathrm{VPR}}(\\{I_{\\mathrm{pos},i}\\})$", + "4: for $t = 1$ to $K$ do", + "5: Compute perturbed image: $I' \\gets I + \\delta$", + "6: if $D_{\\text{type}} = \\text{Frequency}$ then", + "7: $I^{\\prime}\\gets I + \\mathrm{RemoveLowFreq}(\\delta)$", + "8: $P\\gets 1$", + "9: $\\triangleright$ For clarity, this algorithm omits transformations between spatial and frequency domains.", + "10: end if", + "11: Generate feature vector: $F' \\gets \\mathcal{M}_{\\mathrm{VPR}}(I')$", + "12: Compute average distance to positive samples: $\\mathcal{L}_{\\mathrm{pos}}\\gets \\frac{1}{N}\\sum_{i = 1}^{N}\\mathrm{Distance}(F',F_{\\mathrm{pos},i})$", + "13: Compute PSNR loss: $\\mathcal{L}_{\\mathrm{psnr}}\\gets \\mathrm{PSNR}(I,I^{\\prime})$", + "14: Compute total loss: $\\mathcal{L}\\gets -\\mathcal{L}_{\\mathrm{pos}} - \\lambda \\mathcal{L}_{\\mathrm{psnr}}$", + "15: Compute gradient: $\\nabla \\mathcal{L}\\gets$ compute-gradient $(\\mathcal{L},\\delta)$", + "16: Update perturbation: $\\delta \\gets \\delta +\\eta \\nabla \\mathcal{L}$", + "17: Apply Saliency-aware prior $P$ : $\\delta \\gets \\mathrm{clip}(\\delta, -P \\cdot \\epsilon, P \\cdot \\epsilon)$", + "18: end for", + "19: Return the final perturbed image: $I'$", + "20: return $I^{\\prime}$" + ], + "bbox": [ + 96, + 186, + 480, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as the Surrogate VPR model $\\mathcal{M}_{\\mathrm{VPR}}$ . This strategy ensures that the perturbations are not overfitted to a specific model but instead generalize well to a variety of potential recognition systems. In real-world applications, the VPR model set can be dynamically updated, enabling adaptation to stronger protection against unauthorized attacks.", + "bbox": [ + 89, + 592, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each iteration, we obtain the perturbed image $I^{\\prime}$ as $I^{\\prime} = I + \\delta$ . If the frequency domain is used ( $D_{\\mathrm{type}} = \\mathrm{Frequency}$ ), the perturbation is processed by applying a high-pass filter to remove low frequencies.", + "bbox": [ + 89, + 683, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we use the Surrogate VPR: $F' = \\mathcal{M}_{\\mathrm{VPR}}(I')$ to generate feature $F'$ . The feature distance between the perturbed image and the retrieved positive reference images is then computed as the average distance:", + "bbox": [ + 89, + 744, + 483, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p o s}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| \\frac {F ^ {\\prime}}{\\| F ^ {\\prime} \\| _ {2}} - \\frac {F _ {\\mathrm {p o s} , i}}{\\| F _ {\\mathrm {p o s} , i} \\| _ {2}} \\right\\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 816, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of retrieved positive samples, and $\\| X\\| _2$ denotes the L2 norm.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, to ensure that the generated perturbation images maintain high visual fidelity consistent with the original images, we also impose the constraint of preserving pixel-wise similarity by employing the PSNR loss.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss balances the trade-off between maximizing the feature distance and preserving image quality:", + "bbox": [ + 511, + 152, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = - \\mathcal {L} _ {\\mathrm {p o s}} - \\lambda \\mathcal {L} _ {\\mathrm {p s n r}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 191, + 903, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the trade-off coefficient. Notably, $-\\mathcal{L}_{\\mathrm{pos}}$ corresponding to $\\mathcal{L}_{\\mathrm{privacy}}$ while $-\\mathcal{L}_{\\mathrm{psnr}}$ corresponding to $\\mathcal{L}_{\\mathrm{visual}}$ in Eq. (1).", + "bbox": [ + 511, + 215, + 903, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure the perturbations remain concentrated in decisive regions, we optimize their distribution by using the saliency-aware prior:", + "bbox": [ + 511, + 262, + 903, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta = \\mathcal {C} (\\delta , - P \\cdot \\epsilon , P \\cdot \\epsilon), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 316, + 903, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the step size. And $\\mathcal{C}$ is clip to constrain the perturbation within saliency-aware bounds.", + "bbox": [ + 511, + 342, + 903, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After completing $K$ iterations, we obtain the final perturbed (protected) image $I_{p} = I^{\\prime}$ that is optimized to disrupt place recognition while maintaining visual consistency with the original image.", + "bbox": [ + 511, + 373, + 903, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 445, + 645, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 470, + 714, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct the experiments on several VPR benchmark datasets: Pitts30k [1], MSLS [54], Tokyo247 [49], and Nordland [47]. We conduct experiments using PyTorch on a setup with 8 NVIDIA GeForce RTX 3090 GPUs. For our method, the trade-off coefficient $\\lambda$ is set to 0.3 in both spatial and frequency domains. Specifically, in the frequency domain, we utilize an $8 \\times 8$ DCT block size (yielding 64 blocks), with a frequency threshold of 12. For both ANDA and MultiANDA, we set the spatial transformation range to 0.3 (allowing translations by $\\pm 30\\%$ of the image size), gradient averaging to 25 samples, and multi-scale transformations to 5 copies. MultiANDA introduces a diversity probability of 0.5. The Gaussian kernel sizes differ between the two methods: $7(\\sigma = 3)$ for ANDA and $15(\\sigma = 3)$ for MultiANDA. All these methods evaluate over 10 iterations with a batch size of 32 and a learning rate of $8 \\times 10^{-3}$ .", + "bbox": [ + 511, + 492, + 906, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison Study", + "text_level": 1, + "bbox": [ + 511, + 742, + 694, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Black-Box and White-Box Performance. We select five popular open-source VPR models: Crica [30], DHE [29], Salad [21], Sela [31], and VLAD-BuFF [24]. We randomly select one model as the surrogate to generate protected images; evaluations on this surrogate constitute the white-box setting, while evaluations on the remaining four unseen models constitute the black-box setting. Notably, in black-box setting, a single surrogate VPR model is selected from the VPR pool as the fixed model for training", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7201", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/40dddeca85d0b93715640dcc81df3c2cb74f7c189b27f577c925d5d24ff565fb.jpg", + "table_caption": [ + "Table 1. Protection success rate (%) comparisons (Epsilon=12/255). ΔR@1, ΔR@5, and ΔR@10 measure the reduction in the model's retrieval accuracy after applying protection, with larger values indicating stronger protection." + ], + "table_footnote": [], + "table_body": "
DatasetModelANDA_CricaMULTI-ANDA_CricaOurs_CricaOurs_Sela
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KVLAD-BuFF18.59.3-20.215.1-45.437.6-36.227.7-
DHE23.116.09.831.321.913.339.625.916.241.625.919.2
Sela17.713.66.319.215.88.840.727.020.073.254.940.1
Crica50.835.421.952.638.225.069.561.857.651.044.327.9
Salad22.012.38.125.815.410.935.723.715.943.028.518.5
TOKYO247VLAD-BuFF19.611.7-20.512.2-36.328.8-28.423.9-
DHE23.229.221.434.541.329.745.850.750.738.046.644.5
Sela34.225.816.236.827.519.744.427.226.363.664.046.2
Crica49.835.826.752.038.229.382.282.380.052.434.126.9
MSLSVLAD-BuFF29.318.4-30.620.6-43.534.6-37.827.8-
DHE37.826.719.944.933.325.852.940.033.349.336.232.2
Sela24.414.17.328.119.311.041.234.327.472.865.346.7
Crica53.743.433.057.947.237.381.583.473.840.730.326.2
Salad36.434.423.442.937.330.052.140.424.149.937.327.9
MSLSDHE60.476.177.961.377.581.661.678.082.061.577.381.2
Salad72.885.284.474.087.688.674.688.190.274.787.990.2
CHALLENGESela72.683.980.673.084.483.073.487.390.068.376.275.6
Crica66.774.364.567.375.866.669.079.979.473.387.189.7
NORDLANDVLAD-BuFF58.962.9-60.766.0-68.970.7-66.871.3-
DHE43.653.649.353.365.465.764.478.677.864.877.374.2
Sela52.357.446.254.561.951.464.172.972.486.983.364.0
Crica63.161.351.066.062.352.387.984.879.262.169.966.9
Salad56.960.250.262.867.856.769.378.074.468.879.172.2
", + "bbox": [ + 158, + 118, + 841, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71139ffc9d806d385a26ef46ce9948af4eae8bbd6448c2b58feb6427ba150a48.jpg", + "image_caption": [ + "Figure 3. Qualitative results. Our method robustly prevents location retrieval while preserving image quality. Green frames indicate successful protection, while orange frames indicate the opposite." + ], + "image_footnote": [], + "bbox": [ + 173, + 398, + 821, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the protected image set, while the remaining VPR models in the pool are strictly not used in the training process and instead serve as unseen models for testing. Specifically, $\\mathrm{ANDA}_{\\mathrm{Crica}}$ , MULTI- $\\mathrm{ANDA}_{\\mathrm{Crica}}$ , and $\\mathrm{Ours}_{\\mathrm{Crica}}$ use Crica as the surrogate model, whereas $\\mathrm{Ours}_{\\mathrm{Sela}}$ uses Sela. In Tab. 1, our method consistently achieves higher protection success rates than ANDA [13] and MultiANDA [13] across all datasets. $\\mathrm{Ours}_{\\mathrm{Crica}}$ and $\\mathrm{Ours}_{\\mathrm{Sela}}$ significantly reduce retrieval accuracy, highlighting their strong privacy protection capability. Notably, on PITTS30K and MSLS, our approach markedly outperforms baselines. These results demonstrate the strong transferability and effectiveness of our method in", + "bbox": [ + 88, + 705, + 485, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "disrupting both black-box and white-box retrieval.", + "bbox": [ + 511, + 705, + 846, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Quality Evaluations. As shown in Tab. 2, our method consistently outperforms ANDA and MultiANDA across most datasets, achieving superior visual fidelity while maintaining strong protection. In PITTS30K, OursSela achieves the best visual fidelity results, while OursCrica attains the lowest Avg Epsilon, ensuring higher imperceptibility. These results demonstrate that our approach not only provides stronger privacy protection but also ensures better image quality and imperceptibility.", + "bbox": [ + 511, + 726, + 906, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Runtime Comparisons. The results on the Tab. 3 clearly demonstrate that our method surpasses ANDA by $15.9\\%$ ,", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7202", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef906974626817ca52ed59b75f642378b5c16fa1c687641af6841ead0dcfcc58.jpg", + "image_caption": [ + "Figure 4. Protection on Commercial APIs. Our method successfully prevents Google and Microsoft Bing from retrieving location information while ensuring imperceptible modifications, highlighting our method's practical superiority." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 846, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d00431e074a16cecf6ed7f88fcc80f37c5d29920569d7d3c5b827aa79dcbeca.jpg", + "table_caption": [ + "Table 2. Image quality comparison (Epsilon=12/255). Higher PSNR/SSIM and lower LPIPS indicate better visual fidelity, while lower Avg Epsilon suggests more imperceptible perturbations." + ], + "table_footnote": [], + "table_body": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KANDACrica0.14429.840.8724.78
MULTI-ANDACrica0.13230.170.8814.44
OursCrica0.09131.560.9044.02
OursSela0.08232.000.9174.14
TOKYO247ANDACrica0.08930.000.8815.34
MULTI-ANDACrica0.08330.200.8975.14
OursCrica0.09929.110.8865.32
OursSela0.11328.560.8525.42
MSLSANDACrica0.13930.660.8724.22
MULTI-ANDACrica0.11130.140.8894.00
OursCrica0.10230.990.9014.38
OursSela0.11930.260.8894.02
MSLS CHALLENGEANDA (Crica)0.12633.010.9094.68
MULTI-ANDACrica0.12133.290.9174.38
OursCrica0.07334.200.9364.68
OursSela0.09233.210.9143.42
NORDLANDANDA (Crica)0.10530.420.8763.04
MULTI-ANDACrica0.09430.990.8962.78
OursCrica0.09232.030.9172.76
OursSela0.10631.220.8963.38
", + "bbox": [ + 112, + 430, + 457, + 705 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c242f2b3a746eb38be2edac1e9ae7064913e607c02db02d0af8cd7c64aacab7d.jpg", + "table_caption": [ + "Table 3. Runtime and average protection success rate on NORDLAND. Our method achieves a $15 \\times$ speedup over MULTI-ANDA." + ], + "table_footnote": [], + "table_body": "
MethodANDAMULTI-ANDAOurs
Avg ΔR@155.059.570.9
Avg ΔR@559.164.777.0
Avg ΔR@1049.256.576.0
Runtime (s)0.1690.5510.036
", + "bbox": [ + 129, + 753, + 437, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$17.9\\%$ , and $26.8\\%$ , and outperforms MultiANDA by $11.4\\%$ , $12.3\\%$ , and $19.5\\%$ in terms of $\\Delta R@1$ , $\\Delta R@5$ , and $\\Delta R@10$ , respectively. Additionally, our method requires", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only 0.036 seconds per run, approximately 4.7 times faster than ANDA and over $15 \\times$ faster than MultiANDA. These results confirm that our approach significantly enhances protection success while maintaining remarkable efficiency, making it suitable for real-time applications.", + "bbox": [ + 511, + 385, + 906, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Comparisons. Fig. 3 demonstrates that our method effectively prevents location retrieval, surpassing existing SOTA approaches under challenging scenarios, including drastic viewpoint changes, seasonal variations, day-night transitions, and occlusions. Our approach consistently obscures identifiable features while preserving visual integrity, ensuring robust privacy protection suitable for diverse real-world conditions.", + "bbox": [ + 511, + 460, + 908, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Protection Performance on Commercial APIs", + "text_level": 1, + "bbox": [ + 511, + 590, + 897, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 illustrates the effectiveness of our method, evaluated on world-famous commercial APIs including Google and Microsoft Bing. Originally, these systems could achieve correct retrieval of identifying and matching locations from visual cues. However, after applying our method, these images become effectively shielded from such retrieval processes. Notably, while our method significantly disrupts the feature extraction necessary for location recognition, it preserves the visual quality of the images. This balance between robust privacy protection and maintaining visual integrity highlights the practical applicability and superiority of our approach in real-world scenarios.", + "bbox": [ + 511, + 612, + 908, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 801, + 679, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, we conduct an ablation study on the impact of spatial and frequency-domain perturbations and the saliency-aware prior.", + "bbox": [ + 511, + 825, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Spatial and Frequency Domain Perturbation Optimization. Tab. 4 shows that spatial domain at", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7203", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fbe27c3b0c8a45f8e415d5f22bdcbec8aa26b0f81fce95f52eb714a497bdf4b4.jpg", + "table_caption": [ + "Table 4. Ablation study on protection success rate." + ], + "table_footnote": [], + "table_body": "
DatasetModelSpatialFrequencySpatial + Frequency
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KCrica51.642.331.927.522.717.657.147.737.5
DHE31.218.111.221.512.87.635.520.314.4
Salad31.717.210.820.210.35.733.419.312.2
Sela24.613.110.815.912.57.530.319.616.2
MSLSCrica71.358.337.950.244.230.778.165.843.7
DHE43.130.325.127.923.016.546.432.927.2
Salad44.332.921.626.919.310.849.437.025.3
Sela31.925.220.321.413.34.638.828.724.6
NORDLANDCrica79.371.249.971.056.647.884.675.659.0
DHE56.168.265.437.844.135.063.576.476.1
Salad65.269.370.134.842.438.569.378.074.4
Sela58.365.162.331.427.927.361.868.368.4
", + "bbox": [ + 223, + 104, + 771, + 260 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2ceb08a998134ce6933bc572ef6b115b4d325e51a2b663eedfd66ded46d668af.jpg", + "table_caption": [ + "Table 5. Ablation study on image quality." + ], + "table_footnote": [], + "table_body": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KSpatial0.05732.430.9292.85
Frequency0.04233.830.9462.42
Spatial + Frequency0.05732.420.9282.84
MSLSSpatial0.06732.280.9152.80
Frequency0.05933.190.9372.57
Spatial + Frequency0.06632.210.9162.82
NORDLANDSpatial0.07332.410.9171.83
Frequency0.06032.990.9312.01
Spatial + Frequency0.07132.130.9151.84
", + "bbox": [ + 114, + 299, + 455, + 438 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6decffa8c3530ebcbab00c9b698418a021d9574eaf25ff46a89abf305d3f01d3.jpg", + "table_caption": [ + "Table 6. Ablation study on saliency-aware prior." + ], + "table_footnote": [], + "table_body": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KUnguided0.18329.040.8175.69
Guided0.05732.420.9282.84
MSLSUnguided0.19227.830.8015.91
Guided0.06632.210.9162.82
NORDLANDUnguided0.18828.410.8335.88
Guided0.07132.130.9151.84
", + "bbox": [ + 138, + 467, + 433, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tacks alone yield stronger protection than frequency domain attacks. This indicates that spatial modifications, by altering key structural features, are inherently more effective. Yet, when spatial and frequency domain attacks are combined, the success rates improve further, $2\\%$ to $9\\%$ . This synergy demonstrates that frequency domain perturbations, while less potent on their own, complement spatial attacks by enhancing overall robustness without significant loss of performance. Tab. 5 confirms that frequency perturbations preserve visual quality exceptionally well. Although the combined method results in a slight degradation, this minor compromise in visual fidelity is well justified by the substantial gain in protection. In essence, integrating frequency domain modifications with spatial attacks strikes an effective balance between robustness and image quality.", + "bbox": [ + 89, + 598, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Saliency-Aware Prior. Tab. 6 clearly demonstrates the advantages of incorporating a saliency-aware prior. Compared to the unguided way, the guided method markedly improves image fidelity. For example, on the PITTS30K dataset, the guided approach reduces LPIPS", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from 0.183 to 0.057, increases PSNR from 29.04 to 32.42, and enhances SSIM from 0.817 to 0.928. Additionally, the average perturbation (Avg Epsilon) is lowered from 5.69 to 2.84. Similar trends are observed in the MSLS and NORDLAND datasets. These results indicate that the saliency-aware prior effectively directs perturbations to less noticeable regions, thereby maintaining high visual quality while keeping modifications nearly imperceptible.", + "bbox": [ + 511, + 285, + 906, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 439, + 633, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present the first look at privacy protection in VPR and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify key regions crucial for place recognition and propose SAP-PO to selectively enhance perturbation generation in these areas. To further improve imperceptibility, we optimize perturbations in the frequency domain, refining high-frequency components while preserving low-frequency structures essential for human perception. Extensive experiments on multiple benchmark datasets and across various black-box VPR models demonstrate that our method surpasses SOTA perturbation methods in effectiveness. Moreover, our approach achieves a $15 \\times$ speedup in runtime compared to SOTA methods. We also validate its practicality using commercial APIs, including Google and Microsoft Bing, showcasing its real-world applicability.", + "bbox": [ + 511, + 472, + 906, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 747, + 671, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008). This work was supported in part by the National Natural Science Foundation of China under Grant 62202214 and Guangdong Basic and Applied Basic Research Foundation under Grant 2023A1515012819.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7204", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gabriele Berton, Riccardo Mereu, Gabriele Trivigno, Carlo Masone, Gabriela Csurka, Torsten Sattler, and Barbara Caputo. Deep visual geo-localization benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5407, 2022. 5", + "[2] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11080-11090, 2023. 2", + "[3] Gabriele Berton, Lorenz Junglas, Riccardo Zaccone, Thomas Pollok, Barbara Caputo, and Carlo Masone. Meshvpr: Citywide visual place recognition using 3d meshes. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 2", + "[4] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2311-2319, 2025. 2", + "[5] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 2", + "[6] Francesco Croce and Matthias Hein. Sparse and imperceivable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4724-4732, 2019. 3", + "[7] Shuting Dong, Feng Lu, Zhe Wu, and Chun Yuan. Dfvsr: Directional frequency video super-resolution via asymmetric and enhancement alignment network. In *IJCAI*, pages 681–689, 2023. 2", + "[8] Shuting Dong, Feng Lu, and Chun Yuan. Frequency reciprocal action and fusion for single image super-resolution. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023.", + "[9] Shuting Dong, Zhe Wu, Feng Lu, and Chun Yuan. Enhanced image deblurring: An efficient frequency exploitation and preservation network. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7184-7193, 2023. 2", + "[10] Xiaoyi Dong, Dongdong Chen, Jianmin Bao, Chuan Qin, Lu Yuan, Weiming Zhang, Nenghai Yu, and Dong Chen. Greedyfool: Distortion-aware sparse adversarial attack. Advances in Neural Information Processing Systems, 33: 11226-11236, 2020. 3", + "[11] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3", + "[12] Yanbo Fan, Baoyuan Wu, Tuanhui Li, Yong Zhang, Mingyang Li, Zhifeng Li, and Yujiu Yang. Sparse adversaries" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ial attack via perturbation factorization. In Computer VisionECCV 2020:16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 35-50. Springer, 2020. 3", + "[13] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 2, 3, 6", + "[14] Samar Fares and Karthik Nandakumar. Attack to defend: Exploiting adversarial attacks for detecting poisoned models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24726-24735, 2024. 2", + "[15] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2024. 2", + "[16] Tianrui Guan, Aswath Muthuselvam, Montana Hoover, Xijun Wang, Jing Liang, Adarsh Jagan Sathyamoorthy, Damon Conover, and Dinesh Manocha. Crossloc3d: Aerial-ground cross-source 3d place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11335-11344, 2023. 2", + "[17] Ziwen He, Wei Wang, Jing Dong, and Tieniu Tan. Transferable sparse adversarial attack. arXiv preprint arXiv:2105.14727, 2021. 3", + "[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3", + "[19] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15014–15023, 2022. 2", + "[20] Sergio Izquierdo and Javier Civera. Close, but not there: Boosting geographic distance sensitivity in visual place recognition. In European Conference on Computer Vision, pages 240-257. Springer, 2024. 2", + "[21] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 17658-17668, 2024. 2, 5", + "[22] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4379-4389, 2023. 2", + "[23] Tong Jin, Feng Lu, Shuyu Hu, Chun Yuan, and Yunpeng Liu. Editformer: An efficient decoder transformer for visual place recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2", + "[24] Ahmad Khaliq, Ming Xu, Stephen Hausler, Michael Milford, and Sourav Garg. Vlad-buff: burst-aware fast feature aggre" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7205", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gation for visual place recognition. In European Conference on Computer Vision, pages 447-466. Springer, 2024. 2, 5", + "[25] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3", + "[26] Jie Li, Rongrong Ji, Hong Liu, Xiaopeng Hong, Yue Gao, and Qi Tian. Universal perturbation attack against image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4899-4908, 2019. 2", + "[27] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 3", + "[28] Feng Lu, Lijun Zhang, Shuting Dong, Baifan Chen, and Chun Yuan. Aanet: Aggregation and alignment network with semi-hard positive sample mining for hierarchical place recognition. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11771-11778. IEEE, 2023. 2", + "[29] Feng Lu, Shuting Dong, Lijun Zhang, Bingxi Liu, Xiangyuan Lan, Dongmei Jiang, and Chun Yuan. Deep homography estimation for visual place recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10341-10349, 2024. 5", + "[30] Feng Lu, Xiangyuan Lan, Lijun Zhang, Dongmei Jiang, Yaowei Wang, and Chun Yuan. Cricavpr: Cross-image correlation-aware representation learning for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16772-16782, 2024. 2, 5", + "[31] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. International Conference on Learning Representations, 2024. 2, 3, 5", + "[32] Feng Lu, Xinyao Zhang, Canming Ye, Shuting Dong, Lijun Zhang, Xiangyuan Lan, and Chun Yuan. Supervlad: Compact and robust image descriptors for visual place recognition. Advances in Neural Information Processing Systems, 37:5789-5816, 2025. 2", + "[33] Lun Luo, Shuhang Zheng, Yixuan Li, Yongzhi Fan, Beinan Yu, Si-Yuan Cao, Junwei Li, and Hui-Liang Shen. Bevplace: Learning lidar-based place recognition using bird's eye view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8700-8709, 2023. 2", + "[34] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087–9096, 2019. 3", + "[35] SD Moosavi-Dezfooli. a simple and accurate method to fool deep neural networks/s. moosavi-dezfooli, a. fawzi, p. frossard. Preprint—July, 2016. 3", + "[36] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE transactions on pattern" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "analysis and machine intelligence, 41(10):2452-2465, 2018. 3", + "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3", + "[38] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 3", + "[39] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17595-17605, 2024. 2", + "[40] Yuxin Qi, Quan Zhang, Xi Lin, Xiu Su, Jiani Zhu, Jingyu Wang, and Jianhua Li. Seeing beyond noise: Joint graph structure evaluation and denoising for multimodal recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12461-12469, 2025. 2", + "[41] Qibo Qiu, Shun Zhang, Haiming Gao, Honghui Yang, Haochao Ying, Wenxiao Wang, and Xiaofei He. Emvp: Embracing visual foundation model for visual place recognition with centroid-free probing. Advances in Neural Information Processing Systems, 37:120928-120950, 2025. 2", + "[42] Tianyi Alex Qiu, Zhonghao He, Tejasveer Chugh, and Max Kleiman-Weiner. The lock-in hypothesis: Stagnation by algorithm. arXiv preprint arXiv:2506.06166, 2025. 2", + "[43] Ivan Reyes-Amezcua, Gilberto Ochoa-Ruiz, and Andres Mendez-Vazquez. Enhancing image classification robustness through adversarial sampling with delta data augmentation (dda). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 274-283, 2024. 2", + "[44] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 2, 3", + "[45] Zhigang Su, Dawei Zhou, Nannan Wang, Decheng Liu, Zhen Wang, and Xinbo Gao. Hiding visual information via obfuscating adversarial perturbations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4356-4366, 2023. 2, 3", + "[46] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24584-24594, 2024. 2", + "[47] Niko Sünderhauf, Peer Neubert, and Peter Protzel. Are we there yet? challenging seqslam on a $3000\\mathrm{km}$ journey across all four seasons. In Proc. of workshop on long-term autonomy, IEEE international conference on robotics and automation (ICRA), page 2013. CiteSeer, 2013. 5", + "[48] Ajinkya Tejankar, Maziar Sanjabi, Qifan Wang, Sinong Wang, Hamed Firooz, Hamed Pirsiavash, and Liang Tan." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7206", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Defending against patch-based backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12239-12249, 2023. 2", + "[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1808-1817, 2015. 5", + "[50] Issar Tzachor, Boaz Lerner, Matan Levy, Michael Green, Tal Berkovitz Shalev, Gavriel Habib, Dvir Samuel, Noam Korngut Zailer, Or Shimshi, Nir Darshan, et al. Effovpr: Effective foundation model utilization for visual place recognition. International Conference on Learning Representations, 2025. 2", + "[51] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3", + "[52] Jiakai Wang, Zixin Yin, Pengfei Hu, Aishan Liu, Renshuai Tao, Haotong Qin, Xianglong Liu, and Dacheng Tao. Defensive patches for robust recognition in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2456-2465, 2022. 2", + "[53] Zhibo Wang, He Wang, Shuaifan Jin, Wenwen Zhang, Jiahui Hu, Yan Wang, Peng Sun, Wei Yuan, Kaixin Liu, and Kui Ren. Privacy-preserving adversarial facial features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8212-8221, 2023. 2", + "[54] Frederik Warburg, Soren Hauberg, Manuel Lopez-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2626-2635, 2020. 5", + "[55] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Yu Li, Chun Yuan, and Dacheng Tao. Task groupings regularization: Data-free meta-learning with heterogeneous pretrained models. In International Conference on Machine Learning, pages 52573-52587. PMLR, 2024. 2", + "[56] Yongxian Wei, Zixuan Hu, Zhenyi Wang, Li Shen, Chun Yuan, and Dacheng Tao. Free: Faster and better data-free meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23273-23282, 2024.", + "[57] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025.", + "[58] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025.", + "[59] Yongxian Wei, Anke Tang, Li Shen, Zixuan Hu, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 2", + "[60] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, Joao F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8461-8472, 2023. 2", + "[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 3", + "[62] Jingwen Ye and Xinchao Wang. Ungenerable examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11944-11953, 2024. 2, 3", + "[63] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 2, 4", + "[64] Mubariz Zaffar, Liangliang Nan, and Julian FP Kooij. On the estimation of image-matching uncertainty in visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17743-17753, 2024. 2", + "[65] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020. 3", + "[66] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7868-7877, 2021. 3", + "[67] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. Indprompter: Adapting sam to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations. 2", + "[68] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25409-25419, 2024.", + "[69] Quan Zhang, Jinwei Fang, Rui Yuan, Xi Tang, Yuxin Qi, Ke Zhang, and Chun Yuan. Weakly supervised temporal action localization via dual-prior collaborative learning guided by multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 24139-24148, 2025.", + "[70] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10085-10093, 2025. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7207", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Xinwei Zhang, Tianyuan Zhang, Yitong Zhang, and Shuangcheng Liu. Enhancing the transferability of adversarial attacks with stealth preservation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915-2925, 2024. 2", + "[72] Mingkang Zhu, Tianlong Chen, and Zhangyang Wang. Sparse and imperceptible adversarial attack via a homotopy algorithm. In International Conference on Machine Learning, pages 12868-12877. PMLR, 2021. 3" + ], + "bbox": [ + 91, + 90, + 486, + 218 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7208", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_model.json b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cfc53e4e2b0c150e890222e922e1f8e8839ae518 --- /dev/null +++ b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_model.json @@ -0,0 +1,2721 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.131, + 0.887, + 0.155 + ], + "angle": 0, + "content": "VPR-Cloak: A First Look at Privacy Cloak Against Visual Place Recognition" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.18, + 0.721, + 0.217 + ], + "angle": 0, + "content": "Shuting Dong\\(^{1,2*}\\), Mingzhi Chen\\(^{3*}\\), Feng Lu\\(^{1,2}\\), Hao Yu\\(^{1}\\), Guanghao Li\\(^{1,3}\\), Zhe Wu\\(^{2}\\), Ming Tang\\(^{3\\dagger}\\), Chun Yuan\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.217, + 0.877, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University, \\(^{2}\\)Pengcheng Laboratory, \\(^{3}\\)Southern University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.237, + 0.753, + 0.251 + ], + "angle": 0, + "content": "dst21@mails.tsinghua.edu.cn, 12211414@mail.sustech.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.255, + 0.837, + 0.269 + ], + "angle": 0, + "content": "lf22@mails.tsinghua.edu.cn, longinyh@gmail.com, ligh24@mails.tsinghua.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.273, + 0.786, + 0.287 + ], + "angle": 0, + "content": "wuzh02@pcl.ac.cn, tangm3@sustech.edu.cn, yuanc@sz.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.486, + 0.764 + ], + "angle": 0, + "content": "With the rapid advancement of Visual Place Recognition (VPR) systems, their unauthorized use on social media images enables monitoring of individuals' daily movements, posing serious privacy risks. However, privacy protection for addressing these risks in VPR systems remains an underexplored area. While adversarial perturbations have been widely explored for visual privacy protection, existing methods still fail to simultaneously satisfy the black-box constraint, imperceptibility, and real-time performance required in realistic VPR privacy protection scenarios. In this paper, we present the first look at privacy protection in VPR systems and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify decisive regions for place recognition and propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO) to selectively optimize perturbation generation in these areas. To enhance imperceptibility, we further optimize perturbations in the frequency domain, meticulously refining high-frequency components of perturbations while preserving low-frequency structures essential for human perception. Extensive experiments verify that our method outperforms existing SOTA methods. Additionally, our method achieves a \\(15 \\times\\) speedup in runtime compared to SOTA methods. We also validate on Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios. The code is released at https://github.com/Dilemma-CMZ/VPR-Cloak." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.793, + 0.222, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Recent years have witnessed major advances in visual place recognition (VPR) systems, which have been used in various visual applications. While these systems are designed" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.321, + 0.907, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.504, + 0.904, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.682, + 0.907, + 0.71 + ], + "angle": 0, + "content": "Figure 1. Illustrations of (a) privacy leakage risk, and (b) privacy protection in visual place recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "with good intentions to enhance localization and navigation, their growing capabilities have intensified privacy concerns. As shown in Fig. 1(a), with the widespread use of social media, images posted by users can be easily crawled and queried without permission, making it easier than ever to locate users' daily movements through place recognition, which brings huge privacy risks. Moreover, location data can often be used to directly or indirectly infer a user's true identity, living environment, or social relationships, further intensifying these privacy risks. Despite these significant" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "risks, to the best of our knowledge, no existing research has specifically addressed privacy protection in visual place recognition. This highlights the urgent need to develop robust privacy protection mechanisms for place recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.483, + 0.336 + ], + "angle": 0, + "content": "While no prior work has specifically addressed privacy protection for visual place recognition, adversarial perturbations [44-46, 53] have been widely explored as a general approach to visual privacy. These methods typically modify images to prevent unauthorized recognition systems from accurately identifying their content, either by introducing dense noise [14, 19, 43, 71] or applying patch-based modifications [5, 22, 48, 52]. However, for these perturbations to be effective, they often need to be highly noticeable, leading to visually abrupt and unnatural distortions. This makes it difficult to maintain image visual fidelity within ideal privacy protection for visual place recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.563 + ], + "angle": 0, + "content": "In contrast, some recent works explore imperceptible perturbations [44, 45, 51, 61, 62], aiming at visual fidelity by targeting a limited number of pixels and yet achieving high success rates for privacy protection. However, a major drawback is their low transferability, meaning these perturbations often fail when transferred to different models, especially in black-box settings where the adversary has no access to the target model's parameters or architecture. To address this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent and leveraging a deep ensemble strategy. However, the long inference time of ANDA and MultiANDA makes them unsuitable for privacy protection of place recognition in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Building on the previous analysis, privacy protection for place recognition still remains an underexplored area, as it requires simultaneously addressing three critical real-world challenges: (1) diminishing the success rate of unauthorized VPR models which in real-world applications are often black-box, (2) maintaining high image visual quality to ensure imperceptibility, and (3) achieving near-real-time performance to support practical, real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.868 + ], + "angle": 0, + "content": "To bridge these challenges, we present a first look at privacy protection for place recognition and propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify the decisive regions for place recognition. Based on this, we propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO), which strategically optimizes perturbation generation in images by focusing on these decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way of generating the perturbation protects the most vulnerable areas of the image from unauthorized VPR models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In addition, previous research [63] reveals that deep learning models are highly sensitive to high-frequency tex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "tures, while the human visual system primarily perceives low-frequency structural information. Inspired by this, we further elevate perturbation optimization to the frequency domain by isolating the high-frequency components from the perturbation and meticulously refining them using the SAP-PO algorithm. This way ensures that only the high-frequency elements of the perturbation are optimized. As a result, we craft a perturbation that is both precisely targeted and remarkably potent. Through further optimization in the frequency domain, the final output perturbation image (i.e., the protected image) remains imperceptible to the human eye while maintaining a robust defense against VPR models. Our work brings the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- We propose a fast and effective network that provides privacy protection for location information in images while ensuring imperceptibility to the human eye. To the best of our knowledge, this is the first attempt to address privacy protection for visual place recognition." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.364, + 0.905, + 0.408 + ], + "angle": 0, + "content": "- We introduce a SAP-PO algorithm, which strategically refines perturbations by focusing on decisive regions for place recognition." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.409, + 0.905, + 0.483 + ], + "angle": 0, + "content": "- Extensive experiments on multiple benchmark datasets and on various black-box VPR models verify that our method outperforms existing SOTA methods. Additionally, our method achieves a \\(15 \\times\\) speedup in runtime compared to SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.484, + 0.905, + 0.545 + ], + "angle": 0, + "content": "- We validate the effectiveness of our privacy protection method based on using commercial APIs, including Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.289, + 0.905, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.663, + 0.573 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Visual Place Recognition. With the rapid development of deep learning [4, 7-9, 23, 28, 40, 42, 55-59, 67-70], significant progress has been made in the VPR field [3, 15, 16, 20, 24]. Existing place recognition models have achieved impressive results not only in standard scenarios but also in challenging conditions such as radar scene [39], nighttime [31, 39], rainy [39], transport [21], perceptual aliasing [64], viewpoint changes [2, 41, 50], LiDAR scan [33, 60], occlusion [30, 41, 50], seasonal variation [41, 50], illumination change [32, 41, 50] dynamic objects. While these systems are designed with good intentions to enhance localization and navigation, they can be exploited by adversaries to expose users' location information, leading to potential risks of identity inference and tracking. Although previous work [26] attempts to attack image retrieval systems, it is not specifically designed for the VPR task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Perturbations Generation. While no prior research has specifically focused on privacy protection in visual place recognition, adversarial perturbations [44-46, 53] have been extensively studied as a general approach to visual privacy. Most existing adversarial methods impose norm" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.364 + ], + "angle": 0, + "content": "based constraints when generating perturbations. Specifically, many studies [11, 25, 27, 36, 65, 66] adopt infinity-norm or L2-norm constraints, producing dense perturbations that alter a substantial proportion of pixels. Although these methods are effective, they often introduce visible and unnatural distortions, particularly at higher perturbation levels. To address this, recent studies have explored sparse perturbation techniques [6, 10, 12, 17, 34, 38, 65, 72], which restrict modifications to a limited subset of pixels while maintaining high attack success rates. For instance, \\(\\mathrm{PGD}_0\\) [6] extends the projected gradient descent framework by enforcing a sparsity constraint. SparseFool [34], building on DeepFool [35], generates sparse untargeted perturbations. ANDA [13] and MultiANDA [13] enhance transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent. However, these methods tend to be resource-intensive and cost a long inference time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.183, + 0.391 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.401, + 0.295, + 0.416 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.573 + ], + "angle": 0, + "content": "We first introduce real-world location privacy risks. As shown in Fig. 1 (a), the rise of social media makes it easier for adversaries to scrape user-posted photos. Using VPR systems, an adversary can infer where an image was taken. This enables monitoring users' daily movements, uncovering sensitive information about their habits, residences, and frequented places. Moreover, when large-scale user daily movements are monitored, location data can directly or indirectly reveal their identity and social connections, intensifying privacy risks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.695 + ], + "angle": 0, + "content": "To mitigate these risks, as shown in Fig. 1 (b), we propose a protection transformation \\( F \\) that converts a user's intended post photo \\( I \\) into a protected representation \\( I_{p} = F(I) \\). Our goal is for \\( I_{p} \\) to remain visually consistent with the original image \\( I \\) while ensuring that adversaries cannot infer the photo's location using VPR systems. We refer to this process as location privacy protection, which can be formulated as an optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.705, + 0.483, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\min _ {F} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p r i v a c y}} \\left(I _ {p}\\right) + \\lambda \\mathcal {L} _ {\\text {v i s u a l}} \\left(I, I _ {p}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where the visual loss \\(\\mathcal{L}_{\\mathrm{visual}}\\) ensures that the transformed image remains visually perceptual consistent to the original image, while the privacy loss \\(\\mathcal{L}_{\\mathrm{privacy}}\\) mislead VPR models in place recognition. The hyperparameter \\(\\lambda\\) controls the trade-off between visual fidelity and location privacy. Formal definitions are provided in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Challenges. Privacy protection in visual place recognition within real-world scenarios presents several unique challenges: Firstly, the adversary's VPR models are often black-box in real-world scenarios, meaning their architectures, parameters, and training data are unknown, making it" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "difficult to mislead the black-box VPR models. Secondly, the protection transformation must preserve high visual fidelity and imperceptibility so that the generated protection image \\( I_{p} \\) remains visually consistency from the original image \\( I \\). Thirdly, the protection method need be computationally efficient, enabling near-real-time processing to support practical use in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.288 + ], + "angle": 0, + "content": "To the best of our knowledge, no existing approach effectively addresses all three of these challenges simultaneously for location privacy protection. Our goal is to develop a method that disrupts location inference while preserving image quality and maintaining practical efficiency, making it suitable for real-world deployment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.298, + 0.766, + 0.314 + ], + "angle": 0, + "content": "3.2. Privacy Cloak Against VPR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.456 + ], + "angle": 0, + "content": "Previous works [44, 45, 51, 61, 62] have explored adding imperceptible perturbations to images to enhance image privacy while aiming to be visually imperceptible. However, these methods often exhibit low transferability, which limits their effectiveness in black-box settings. To handle this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of perturbations. However, the high inference time of ANDA and MultiANDA limits the practicality in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.682 + ], + "angle": 0, + "content": "To address these limitations, we propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify decisive regions for place recognition. Based on the saliency-aware prior, we propose a SAP-PO algorithm, which strategically optimizes perturbation generation in images by focusing on decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way protects the vulnerable regions of the image from unauthorized VPR models. Additionally, we further enhance perturbation optimization in the frequency domain. By isolating and refining high-frequency components using SAP-PO, we craft a precisely targeted yet imperceptible perturbation that remains effective against VPR models while staying unnoticeable to the human eye." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.862 + ], + "angle": 0, + "content": "Saliency-Aware Prior Generation. Instead of applying perturbations randomly, we first identify critical regions for place recognition and selectively perturb them. To achieve this, we introduce a saliency-aware mechanism that acts as a spatial weight to pinpoint key areas for perturbation. For an input unprotected image \\(I\\), we extract this information by using the fine-tuned Vision Transformer (fine-tuned ViT), which employs the adapter strategy from [18, 31] to fine-tune the pre-trained DINOv2 backbone [37]. The input image is first resized using \\(\\mathcal{R}(I)\\) to fit the fine-tuned ViT model's input size. Then, saliency-aware prior \\(P\\), is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.863, + 0.905, + 0.88 + ], + "angle": 0, + "content": "\\[\nP = \\mathcal {N} \\left(\\mathcal {A} _ {\\mathrm {V i T}} (\\mathcal {R} (I))\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(\\cdot)\\) represents normalization to ensure the prior" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7199" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.891, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.351, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 2. Overview of our VPR-Cloak. We introduce a saliency-aware prior to accurately identify decisive regions for place recognition and propose the SAP-PO algorithm, which strategically generates perturbations focused on these decisive regions while minimizing changes in less relevant areas. To enhance imperceptibility, we further elevate perturbation optimization to the frequency domain by isolating and meticulously optimizing only the high-frequencies using the SAP-PO algorithm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.461, + 0.45 + ], + "angle": 0, + "content": "proper scaling and sum to 1. \\(\\mathcal{A}_{\\mathrm{ViT}}\\) is the fine-tuned ViT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.555 + ], + "angle": 0, + "content": "Building on this saliency-aware prior \\( P \\), we propose SAP-PO to optimize perturbation generation, ensuring perturbations are concentrated in decisive regions for place recognition. We provide two optimization strategies: spatial domain perturbation optimization for defense against VPR and frequency domain perturbation optimization for visual imperceptibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Spatial Perturbation for Defense against VPR. In the spatial domain, SAP-PO, guided by the saliency prior, focuses on perturbing key regions crucial for place recognition, maximizing its impact on reducing recognition accuracy. The spatially-perturbed image \\( I_{\\mathrm{spa}}^{\\prime} \\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.645, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {s p a}} ^ {\\prime} = \\mathcal {F} _ {\\mathrm {s p a}} (I, P), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.454, + 0.691 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\mathrm{spa}}\\) is our proposed SAP-PO in spatial domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.797 + ], + "angle": 0, + "content": "Frequency Perturbation for Imperceptible. Previous study [63] have shown that the human eye is sensitive to low frequencies, while deep learning models are sensitive to high frequencies. Inspired by this, we elevate perturbation optimization to the frequency domain by isolating and meticulously refining only the high-frequency components of the perturbation using the SAP-PO algorithm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.484, + 0.875 + ], + "angle": 0, + "content": "To do this, we apply the Discrete Cosine Transform (DCT) to the perturbed image \\( I_{\\mathrm{spa}}^{\\prime} \\), yielding its frequency representation \\( f^{\\prime} \\). We specifically optimize the high-frequency components of the perturbation using SAP-PO and generate the frequency-perturbed image \\( I_{\\mathrm{fre}}^{\\prime} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.885, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nf ^ {\\prime} = \\mathcal {D} \\left(I _ {\\mathrm {s p a}} ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.449, + 0.905, + 0.466 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {f r e}} ^ {\\prime} = \\mathcal {F} _ {\\mathrm {f r e}} \\left(f ^ {\\prime}, P\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.498 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}(\\cdot)\\) denotes DCT. \\(\\mathcal{F}_{\\mathrm{fre}}\\) is our proposed SAP-PO in frequency domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.906, + 0.545 + ], + "angle": 0, + "content": "Next, we apply the inverse DCT to transform the frequency-perturbed image \\( I_{\\mathrm{fre}}^{\\prime} \\) back to the spatial domain, obtaining the final protected image \\( I_{\\mathrm{p}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.549, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {p}} = \\mathcal {D} ^ {- 1} \\left(I _ {\\mathrm {f r e}} ^ {\\prime}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.572, + 0.724, + 0.587 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}^{-1}\\) is the inverse DCT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.649 + ], + "angle": 0, + "content": "Through optimization in the spatial and frequency domains, we strategically refine and enforce the distribution of perturbations to concentrate on decisive regions for place recognition and high-frequency areas." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.906, + 0.684 + ], + "angle": 0, + "content": "3.3. Saliency-Aware Prior Guided Perturbation Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.752 + ], + "angle": 0, + "content": "We propose SAP-PO to generate effective and visually imperceptible perturbations for defending VPR models. By integrating the saliency-aware prior \\( P \\), we focus perturbations on the decisive regions for place recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.907, + 0.814 + ], + "angle": 0, + "content": "To begin, we initialize the perturbation as: \\(\\delta = 0\\). For an input image \\(I\\), we retrieve multiple positive reference images \\(\\{I_{\\mathrm{pos},i}\\}\\) from the database and extract their features \\(\\{F_{\\mathrm{pos},i}\\}\\) using the Surrogate VPR:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.818, + 0.905, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\left\\{F _ {\\text {p o s}, i} \\right\\} = \\mathcal {M} _ {\\mathrm {V P R}} \\left(\\left\\{I _ {\\text {p o s}, i} \\right\\}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_{\\mathrm{VPR}}\\) denotes the Surrogate VPR model. To enhance the robustness of our perturbations against unauthorized VPR models, for each input image, we randomly select one from multiple VPR models \\(\\{\\mathcal{M}_1,\\mathcal{M}_2,\\dots ,\\mathcal{M}_m\\}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7200" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.092, + 0.091, + 0.482, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO)" + }, + { + "type": "code", + "bbox": [ + 0.092, + 0.125, + 0.484, + 0.171 + ], + "angle": 0, + "content": "Require: Image \\(I\\), Saliency-aware prior \\(P\\), Surrogate VPR Model \\(\\mathcal{M}_{\\mathrm{VPR}}\\), Trade-off Coefficient \\(\\lambda\\), Step Size \\(\\epsilon\\), Maximum Iterations \\(K\\), Domain Type \\(D_{\\mathrm{type}}\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.094, + 0.172, + 0.284, + 0.186 + ], + "angle": 0, + "content": "Ensure: Perturbed Image \\(I^{\\prime}\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.187, + 0.312, + 0.201 + ], + "angle": 0, + "content": "1: Initialize perturbation \\(\\delta \\gets 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.203, + 0.482, + 0.232 + ], + "angle": 0, + "content": "2: Retrieve multiple positive samples from database: \\(\\{I_{\\mathrm{pos},s}\\} \\gets\\) retrieve_positives(I)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.233, + 0.482, + 0.263 + ], + "angle": 0, + "content": "3: Generate \\(\\{F_{\\mathrm{pos},i}\\}\\) using the Surrogate VPR model: \\(\\{F_{\\mathrm{pos},i}\\} \\gets \\mathcal{M}_{\\mathrm{VPR}}(\\{I_{\\mathrm{pos},i}\\})\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.264, + 0.242, + 0.275 + ], + "angle": 0, + "content": "4: for \\( t = 1 \\) to \\( K \\) do" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.277, + 0.402, + 0.291 + ], + "angle": 0, + "content": "5: Compute perturbed image: \\( I' \\gets I + \\delta \\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.293, + 0.317, + 0.307 + ], + "angle": 0, + "content": "6: if \\( D_{\\text{type}} = \\text{Frequency} \\) then" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.308, + 0.371, + 0.323 + ], + "angle": 0, + "content": "7: \\(I^{\\prime}\\gets I + \\mathrm{RemoveLowFreq}(\\delta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.325, + 0.218, + 0.335 + ], + "angle": 0, + "content": "8: \\(P\\gets 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.338, + 0.482, + 0.367 + ], + "angle": 0, + "content": "9: \\(\\triangleright\\) For clarity, this algorithm omits transformations between spatial and frequency domains." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.369, + 0.19, + 0.381 + ], + "angle": 0, + "content": "10: end if" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.383, + 0.42, + 0.398 + ], + "angle": 0, + "content": "11: Generate feature vector: \\( F' \\gets \\mathcal{M}_{\\mathrm{VPR}}(I') \\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.399, + 0.482, + 0.429 + ], + "angle": 0, + "content": "12: Compute average distance to positive samples: \\(\\mathcal{L}_{\\mathrm{pos}}\\gets \\frac{1}{N}\\sum_{i = 1}^{N}\\mathrm{Distance}(F',F_{\\mathrm{pos},i})\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.429, + 0.43, + 0.444 + ], + "angle": 0, + "content": "13: Compute PSNR loss: \\(\\mathcal{L}_{\\mathrm{psnr}}\\gets \\mathrm{PSNR}(I,I^{\\prime})\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.445, + 0.418, + 0.459 + ], + "angle": 0, + "content": "14: Compute total loss: \\(\\mathcal{L}\\gets -\\mathcal{L}_{\\mathrm{pos}} - \\lambda \\mathcal{L}_{\\mathrm{psnr}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.459, + 0.478, + 0.473 + ], + "angle": 0, + "content": "15: Compute gradient: \\(\\nabla \\mathcal{L}\\gets\\) compute-gradient \\((\\mathcal{L},\\delta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.474, + 0.384, + 0.489 + ], + "angle": 0, + "content": "16: Update perturbation: \\(\\delta \\gets \\delta +\\eta \\nabla \\mathcal{L}\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.489, + 0.482, + 0.518 + ], + "angle": 0, + "content": "17: Apply Saliency-aware prior \\(P\\): \\(\\delta \\gets \\mathrm{clip}(\\delta, -P \\cdot \\epsilon, P \\cdot \\epsilon)\\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.52, + 0.174, + 0.531 + ], + "angle": 0, + "content": "18: end for" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.533, + 0.357, + 0.548 + ], + "angle": 0, + "content": "19: Return the final perturbed image: \\( I' \\)" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.549, + 0.186, + 0.562 + ], + "angle": 0, + "content": "20: return \\(I^{\\prime}\\)" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.187, + 0.482, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.683 + ], + "angle": 0, + "content": "as the Surrogate VPR model \\(\\mathcal{M}_{\\mathrm{VPR}}\\). This strategy ensures that the perturbations are not overfitted to a specific model but instead generalize well to a variety of potential recognition systems. In real-world applications, the VPR model set can be dynamically updated, enabling adaptation to stronger protection against unauthorized attacks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.484, + 0.745 + ], + "angle": 0, + "content": "At each iteration, we obtain the perturbed image \\(I^{\\prime}\\) as \\(I^{\\prime} = I + \\delta\\). If the frequency domain is used (\\(D_{\\mathrm{type}} = \\mathrm{Frequency}\\)), the perturbation is processed by applying a high-pass filter to remove low frequencies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.484, + 0.806 + ], + "angle": 0, + "content": "Then, we use the Surrogate VPR: \\( F' = \\mathcal{M}_{\\mathrm{VPR}}(I') \\) to generate feature \\( F' \\). The feature distance between the perturbed image and the retrieved positive reference images is then computed as the average distance:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.818, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p o s}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| \\frac {F ^ {\\prime}}{\\| F ^ {\\prime} \\| _ {2}} - \\frac {F _ {\\mathrm {p o s} , i}}{\\| F _ {\\mathrm {p o s} , i} \\| _ {2}} \\right\\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(N\\) is the number of retrieved positive samples, and \\(\\| X\\| _2\\) denotes the L2 norm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "In addition, to ensure that the generated perturbation images maintain high visual fidelity consistent with the original images, we also impose the constraint of preserving pixel-wise similarity by employing the PSNR loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.183 + ], + "angle": 0, + "content": "The total loss balances the trade-off between maximizing the feature distance and preserving image quality:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.192, + 0.905, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = - \\mathcal {L} _ {\\mathrm {p o s}} - \\lambda \\mathcal {L} _ {\\mathrm {p s n r}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.905, + 0.262 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the trade-off coefficient. Notably, \\(-\\mathcal{L}_{\\mathrm{pos}}\\) corresponding to \\(\\mathcal{L}_{\\mathrm{privacy}}\\) while \\(-\\mathcal{L}_{\\mathrm{psnr}}\\) corresponding to \\(\\mathcal{L}_{\\mathrm{visual}}\\) in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.905, + 0.308 + ], + "angle": 0, + "content": "To ensure the perturbations remain concentrated in decisive regions, we optimize their distribution by using the saliency-aware prior:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.318, + 0.905, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\delta = \\mathcal {C} (\\delta , - P \\cdot \\epsilon , P \\cdot \\epsilon), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.904, + 0.373 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the step size. And \\(\\mathcal{C}\\) is clip to constrain the perturbation within saliency-aware bounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.905, + 0.435 + ], + "angle": 0, + "content": "After completing \\( K \\) iterations, we obtain the final perturbed (protected) image \\( I_{p} = I^{\\prime} \\) that is optimized to disrupt place recognition while maintaining visual consistency with the original image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.646, + 0.463 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.715, + 0.487 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.735 + ], + "angle": 0, + "content": "We conduct the experiments on several VPR benchmark datasets: Pitts30k [1], MSLS [54], Tokyo247 [49], and Nordland [47]. We conduct experiments using PyTorch on a setup with 8 NVIDIA GeForce RTX 3090 GPUs. For our method, the trade-off coefficient \\(\\lambda\\) is set to 0.3 in both spatial and frequency domains. Specifically, in the frequency domain, we utilize an \\(8 \\times 8\\) DCT block size (yielding 64 blocks), with a frequency threshold of 12. For both ANDA and MultiANDA, we set the spatial transformation range to 0.3 (allowing translations by \\(\\pm 30\\%\\) of the image size), gradient averaging to 25 samples, and multi-scale transformations to 5 copies. MultiANDA introduces a diversity probability of 0.5. The Gaussian kernel sizes differ between the two methods: \\(7(\\sigma = 3)\\) for ANDA and \\(15(\\sigma = 3)\\) for MultiANDA. All these methods evaluate over 10 iterations with a batch size of 32 and a learning rate of \\(8 \\times 10^{-3}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.695, + 0.759 + ], + "angle": 0, + "content": "4.2. Comparison Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Black-Box and White-Box Performance. We select five popular open-source VPR models: Crica [30], DHE [29], Salad [21], Sela [31], and VLAD-BuFF [24]. We randomly select one model as the surrogate to generate protected images; evaluations on this surrogate constitute the white-box setting, while evaluations on the remaining four unseen models constitute the black-box setting. Notably, in black-box setting, a single surrogate VPR model is selected from the VPR pool as the fixed model for training" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7201" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 1. Protection success rate (%) comparisons (Epsilon=12/255). ΔR@1, ΔR@5, and ΔR@10 measure the reduction in the model's retrieval accuracy after applying protection, with larger values indicating stronger protection." + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.119, + 0.842, + 0.382 + ], + "angle": 0, + "content": "
DatasetModelANDA_CricaMULTI-ANDA_CricaOurs_CricaOurs_Sela
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KVLAD-BuFF18.59.3-20.215.1-45.437.6-36.227.7-
DHE23.116.09.831.321.913.339.625.916.241.625.919.2
Sela17.713.66.319.215.88.840.727.020.073.254.940.1
Crica50.835.421.952.638.225.069.561.857.651.044.327.9
Salad22.012.38.125.815.410.935.723.715.943.028.518.5
TOKYO247VLAD-BuFF19.611.7-20.512.2-36.328.8-28.423.9-
DHE23.229.221.434.541.329.745.850.750.738.046.644.5
Sela34.225.816.236.827.519.744.427.226.363.664.046.2
Crica49.835.826.752.038.229.382.282.380.052.434.126.9
MSLSVLAD-BuFF29.318.4-30.620.6-43.534.6-37.827.8-
DHE37.826.719.944.933.325.852.940.033.349.336.232.2
Sela24.414.17.328.119.311.041.234.327.472.865.346.7
Crica53.743.433.057.947.237.381.583.473.840.730.326.2
Salad36.434.423.442.937.330.052.140.424.149.937.327.9
MSLSDHE60.476.177.961.377.581.661.678.082.061.577.381.2
Salad72.885.284.474.087.688.674.688.190.274.787.990.2
CHALLENGESela72.683.980.673.084.483.073.487.390.068.376.275.6
Crica66.774.364.567.375.866.669.079.979.473.387.189.7
NORDLANDVLAD-BuFF58.962.9-60.766.0-68.970.7-66.871.3-
DHE43.653.649.353.365.465.764.478.677.864.877.374.2
Sela52.357.446.254.561.951.464.172.972.486.983.364.0
Crica63.161.351.066.062.352.387.984.879.262.169.966.9
Salad56.960.250.262.867.856.769.378.074.468.879.172.2
" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.399, + 0.822, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.651, + 0.908, + 0.681 + ], + "angle": 0, + "content": "Figure 3. Qualitative results. Our method robustly prevents location retrieval while preserving image quality. Green frames indicate successful protection, while orange frames indicate the opposite." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.706, + 0.486, + 0.889 + ], + "angle": 0, + "content": "the protected image set, while the remaining VPR models in the pool are strictly not used in the training process and instead serve as unseen models for testing. Specifically, \\(\\mathrm{ANDA}_{\\mathrm{Crica}}\\), MULTI- \\(\\mathrm{ANDA}_{\\mathrm{Crica}}\\), and \\(\\mathrm{Ours}_{\\mathrm{Crica}}\\) use Crica as the surrogate model, whereas \\(\\mathrm{Ours}_{\\mathrm{Sela}}\\) uses Sela. In Tab. 1, our method consistently achieves higher protection success rates than ANDA [13] and MultiANDA [13] across all datasets. \\(\\mathrm{Ours}_{\\mathrm{Crica}}\\) and \\(\\mathrm{Ours}_{\\mathrm{Sela}}\\) significantly reduce retrieval accuracy, highlighting their strong privacy protection capability. Notably, on PITTS30K and MSLS, our approach markedly outperforms baselines. These results demonstrate the strong transferability and effectiveness of our method in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.706, + 0.848, + 0.721 + ], + "angle": 0, + "content": "disrupting both black-box and white-box retrieval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.907, + 0.865 + ], + "angle": 0, + "content": "Image Quality Evaluations. As shown in Tab. 2, our method consistently outperforms ANDA and MultiANDA across most datasets, achieving superior visual fidelity while maintaining strong protection. In PITTS30K, OursSela achieves the best visual fidelity results, while OursCrica attains the lowest Avg Epsilon, ensuring higher imperceptibility. These results demonstrate that our approach not only provides stronger privacy protection but also ensures better image quality and imperceptibility." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Runtime Comparisons. The results on the Tab. 3 clearly demonstrate that our method surpasses ANDA by \\(15.9\\%\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.848, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Protection on Commercial APIs. Our method successfully prevents Google and Microsoft Bing from retrieving location information while ensuring imperceptible modifications, highlighting our method's practical superiority." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Table 2. Image quality comparison (Epsilon=12/255). Higher PSNR/SSIM and lower LPIPS indicate better visual fidelity, while lower Avg Epsilon suggests more imperceptible perturbations." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.431, + 0.459, + 0.706 + ], + "angle": 0, + "content": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KANDACrica0.14429.840.8724.78
MULTI-ANDACrica0.13230.170.8814.44
OursCrica0.09131.560.9044.02
OursSela0.08232.000.9174.14
TOKYO247ANDACrica0.08930.000.8815.34
MULTI-ANDACrica0.08330.200.8975.14
OursCrica0.09929.110.8865.32
OursSela0.11328.560.8525.42
MSLSANDACrica0.13930.660.8724.22
MULTI-ANDACrica0.11130.140.8894.00
OursCrica0.10230.990.9014.38
OursSela0.11930.260.8894.02
MSLS CHALLENGEANDA (Crica)0.12633.010.9094.68
MULTI-ANDACrica0.12133.290.9174.38
OursCrica0.07334.200.9364.68
OursSela0.09233.210.9143.42
NORDLANDANDA (Crica)0.10530.420.8763.04
MULTI-ANDACrica0.09430.990.8962.78
OursCrica0.09232.030.9172.76
OursSela0.10631.220.8963.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.723, + 0.483, + 0.751 + ], + "angle": 0, + "content": "Table 3. Runtime and average protection success rate on NORDLAND. Our method achieves a \\(15 \\times\\) speedup over MULTI-ANDA." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.755, + 0.439, + 0.826 + ], + "angle": 0, + "content": "
MethodANDAMULTI-ANDAOurs
Avg ΔR@155.059.570.9
Avg ΔR@559.164.777.0
Avg ΔR@1049.256.576.0
Runtime (s)0.1690.5510.036
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\(17.9\\%\\), and \\(26.8\\%\\), and outperforms MultiANDA by \\(11.4\\%\\), \\(12.3\\%\\), and \\(19.5\\%\\) in terms of \\(\\Delta R@1\\), \\(\\Delta R@5\\), and \\(\\Delta R@10\\), respectively. Additionally, our method requires" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.907, + 0.461 + ], + "angle": 0, + "content": "only 0.036 seconds per run, approximately 4.7 times faster than ANDA and over \\(15 \\times\\) faster than MultiANDA. These results confirm that our approach significantly enhances protection success while maintaining remarkable efficiency, making it suitable for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.581 + ], + "angle": 0, + "content": "Visual Comparisons. Fig. 3 demonstrates that our method effectively prevents location retrieval, surpassing existing SOTA approaches under challenging scenarios, including drastic viewpoint changes, seasonal variations, day-night transitions, and occlusions. Our approach consistently obscures identifiable features while preserving visual integrity, ensuring robust privacy protection suitable for diverse real-world conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.898, + 0.607 + ], + "angle": 0, + "content": "4.3. Protection Performance on Commercial APIs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Fig. 4 illustrates the effectiveness of our method, evaluated on world-famous commercial APIs including Google and Microsoft Bing. Originally, these systems could achieve correct retrieval of identifying and matching locations from visual cues. However, after applying our method, these images become effectively shielded from such retrieval processes. Notably, while our method significantly disrupts the feature extraction necessary for location recognition, it preserves the visual quality of the images. This balance between robust privacy protection and maintaining visual integrity highlights the practical applicability and superiority of our approach in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.68, + 0.818 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.871 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, we conduct an ablation study on the impact of spatial and frequency-domain perturbations and the saliency-aware prior." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of Spatial and Frequency Domain Perturbation Optimization. Tab. 4 shows that spatial domain at" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7203" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.348, + 0.09, + 0.648, + 0.103 + ], + "angle": 0, + "content": "Table 4. Ablation study on protection success rate." + }, + { + "type": "table", + "bbox": [ + 0.225, + 0.105, + 0.772, + 0.261 + ], + "angle": 0, + "content": "
DatasetModelSpatialFrequencySpatial + Frequency
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KCrica51.642.331.927.522.717.657.147.737.5
DHE31.218.111.221.512.87.635.520.314.4
Salad31.717.210.820.210.35.733.419.312.2
Sela24.613.110.815.912.57.530.319.616.2
MSLSCrica71.358.337.950.244.230.778.165.843.7
DHE43.130.325.127.923.016.546.432.927.2
Salad44.332.921.626.919.310.849.437.025.3
Sela31.925.220.321.413.34.638.828.724.6
NORDLANDCrica79.371.249.971.056.647.884.675.659.0
DHE56.168.265.437.844.135.063.576.476.1
Salad65.269.370.134.842.438.569.378.074.4
Sela58.365.162.331.427.927.361.868.368.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.163, + 0.284, + 0.411, + 0.298 + ], + "angle": 0, + "content": "Table 5. Ablation study on image quality." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.3, + 0.457, + 0.439 + ], + "angle": 0, + "content": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KSpatial0.05732.430.9292.85
Frequency0.04233.830.9462.42
Spatial + Frequency0.05732.420.9282.84
MSLSSpatial0.06732.280.9152.80
Frequency0.05933.190.9372.57
Spatial + Frequency0.06632.210.9162.82
NORDLANDSpatial0.07332.410.9171.83
Frequency0.06032.990.9312.01
Spatial + Frequency0.07132.130.9151.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.453, + 0.431, + 0.466 + ], + "angle": 0, + "content": "Table 6. Ablation study on saliency-aware prior." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.468, + 0.434, + 0.574 + ], + "angle": 0, + "content": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KUnguided0.18329.040.8175.69
Guided0.05732.420.9282.84
MSLSUnguided0.19227.830.8015.91
Guided0.06632.210.9162.82
NORDLANDUnguided0.18828.410.8335.88
Guided0.07132.130.9151.84
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.825 + ], + "angle": 0, + "content": "tacks alone yield stronger protection than frequency domain attacks. This indicates that spatial modifications, by altering key structural features, are inherently more effective. Yet, when spatial and frequency domain attacks are combined, the success rates improve further, \\(2\\%\\) to \\(9\\%\\). This synergy demonstrates that frequency domain perturbations, while less potent on their own, complement spatial attacks by enhancing overall robustness without significant loss of performance. Tab. 5 confirms that frequency perturbations preserve visual quality exceptionally well. Although the combined method results in a slight degradation, this minor compromise in visual fidelity is well justified by the substantial gain in protection. In essence, integrating frequency domain modifications with spatial attacks strikes an effective balance between robustness and image quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Saliency-Aware Prior. Tab. 6 clearly demonstrates the advantages of incorporating a saliency-aware prior. Compared to the unguided way, the guided method markedly improves image fidelity. For example, on the PITTS30K dataset, the guided approach reduces LPIPS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.907, + 0.409 + ], + "angle": 0, + "content": "from 0.183 to 0.057, increases PSNR from 29.04 to 32.42, and enhances SSIM from 0.817 to 0.928. Additionally, the average perturbation (Avg Epsilon) is lowered from 5.69 to 2.84. Similar trends are observed in the MSLS and NORDLAND datasets. These results indicate that the saliency-aware prior effectively directs perturbations to less noticeable regions, thereby maintaining high visual quality while keeping modifications nearly imperceptible." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.634, + 0.456 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.716 + ], + "angle": 0, + "content": "In this paper, we present the first look at privacy protection in VPR and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify key regions crucial for place recognition and propose SAP-PO to selectively enhance perturbation generation in these areas. To further improve imperceptibility, we optimize perturbations in the frequency domain, refining high-frequency components while preserving low-frequency structures essential for human perception. Extensive experiments on multiple benchmark datasets and across various black-box VPR models demonstrate that our method surpasses SOTA perturbation methods in effectiveness. Moreover, our approach achieves a \\(15 \\times\\) speedup in runtime compared to SOTA methods. We also validate its practicality using commercial APIs, including Google and Microsoft Bing, showcasing its real-world applicability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.748, + 0.673, + 0.765 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008). This work was supported in part by the National Natural Science Foundation of China under Grant 62202214 and Guangdong Basic and Applied Basic Research Foundation under Grant 2023A1515012819." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7204" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Gabriele Berton, Riccardo Mereu, Gabriele Trivigno, Carlo Masone, Gabriela Csurka, Torsten Sattler, and Barbara Caputo. Deep visual geo-localization benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5407, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11080-11090, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[3] Gabriele Berton, Lorenz Junglas, Riccardo Zaccone, Thomas Pollok, Barbara Caputo, and Carlo Masone. Meshvpr: Citywide visual place recognition using 3d meshes. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[4] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2311-2319, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.484, + 0.479 + ], + "angle": 0, + "content": "[5] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Francesco Croce and Matthias Hein. Sparse and imperceivable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4724-4732, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[7] Shuting Dong, Feng Lu, Zhe Wu, and Chun Yuan. Dfvsr: Directional frequency video super-resolution via asymmetric and enhancement alignment network. In *IJCAI*, pages 681–689, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Shuting Dong, Feng Lu, and Chun Yuan. Frequency reciprocal action and fusion for single image super-resolution. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[9] Shuting Dong, Zhe Wu, Feng Lu, and Chun Yuan. Enhanced image deblurring: An efficient frequency exploitation and preservation network. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7184-7193, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Xiaoyi Dong, Dongdong Chen, Jianmin Bao, Chuan Qin, Lu Yuan, Weiming Zhang, Nenghai Yu, and Dong Chen. Greedyfool: Distortion-aware sparse adversarial attack. Advances in Neural Information Processing Systems, 33: 11226-11236, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Yanbo Fan, Baoyuan Wu, Tuanhui Li, Yong Zhang, Mingyang Li, Zhifeng Li, and Yujiu Yang. Sparse adversaries" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "ial attack via perturbation factorization. In Computer VisionECCV 2020:16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 35-50. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[13] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.285 + ], + "angle": 0, + "content": "[14] Samar Fares and Karthik Nandakumar. Attack to defend: Exploiting adversarial attacks for detecting poisoned models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24726-24735, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.355 + ], + "angle": 0, + "content": "[15] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[16] Tianrui Guan, Aswath Muthuselvam, Montana Hoover, Xijun Wang, Jing Liang, Adarsh Jagan Sathyamoorthy, Damon Conover, and Dinesh Manocha. Crossloc3d: Aerial-ground cross-source 3d place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11335-11344, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[17] Ziwen He, Wei Wang, Jing Dong, and Tieniu Tan. Transferable sparse adversarial attack. arXiv preprint arXiv:2105.14727, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[19] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15014–15023, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[20] Sergio Izquierdo and Javier Civera. Close, but not there: Boosting geographic distance sensitivity in visual place recognition. In European Conference on Computer Vision, pages 240-257. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.906, + 0.747 + ], + "angle": 0, + "content": "[21] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 17658-17668, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[22] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4379-4389, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[23] Tong Jin, Feng Lu, Shuyu Hu, Chun Yuan, and Yunpeng Liu. Editformer: An efficient decoder transformer for visual place recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Ahmad Khaliq, Ming Xu, Stephen Hausler, Michael Milford, and Sourav Garg. Vlad-buff: burst-aware fast feature aggre" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7205" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "gation for visual place recognition. In European Conference on Computer Vision, pages 447-466. Springer, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[25] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Jie Li, Rongrong Ji, Hong Liu, Xiaopeng Hong, Yue Gao, and Qi Tian. Universal perturbation attack against image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4899-4908, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[27] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[28] Feng Lu, Lijun Zhang, Shuting Dong, Baifan Chen, and Chun Yuan. Aanet: Aggregation and alignment network with semi-hard positive sample mining for hierarchical place recognition. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11771-11778. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[29] Feng Lu, Shuting Dong, Lijun Zhang, Bingxi Liu, Xiangyuan Lan, Dongmei Jiang, and Chun Yuan. Deep homography estimation for visual place recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10341-10349, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[30] Feng Lu, Xiangyuan Lan, Lijun Zhang, Dongmei Jiang, Yaowei Wang, and Chun Yuan. Cricavpr: Cross-image correlation-aware representation learning for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16772-16782, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[31] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. International Conference on Learning Representations, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[32] Feng Lu, Xinyao Zhang, Canming Ye, Shuting Dong, Lijun Zhang, Xiangyuan Lan, and Chun Yuan. Supervlad: Compact and robust image descriptors for visual place recognition. Advances in Neural Information Processing Systems, 37:5789-5816, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[33] Lun Luo, Shuhang Zheng, Yixuan Li, Yongzhi Fan, Beinan Yu, Si-Yuan Cao, Junwei Li, and Hui-Liang Shen. Bevplace: Learning lidar-based place recognition using bird's eye view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8700-8709, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[34] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087–9096, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[35] SD Moosavi-Dezfooli. a simple and accurate method to fool deep neural networks/s. moosavi-dezfooli, a. fawzi, p. frossard. Preprint—July, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE transactions on pattern" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.118 + ], + "angle": 0, + "content": "analysis and machine intelligence, 41(10):2452-2465, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[38] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[39] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17595-17605, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[40] Yuxin Qi, Quan Zhang, Xi Lin, Xiu Su, Jiani Zhu, Jingyu Wang, and Jianhua Li. Seeing beyond noise: Joint graph structure evaluation and denoising for multimodal recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12461-12469, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[41] Qibo Qiu, Shun Zhang, Haiming Gao, Honghui Yang, Haochao Ying, Wenxiao Wang, and Xiaofei He. Emvp: Embracing visual foundation model for visual place recognition with centroid-free probing. Advances in Neural Information Processing Systems, 37:120928-120950, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.51 + ], + "angle": 0, + "content": "[42] Tianyi Alex Qiu, Zhonghao He, Tejasveer Chugh, and Max Kleiman-Weiner. The lock-in hypothesis: Stagnation by algorithm. arXiv preprint arXiv:2506.06166, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[43] Ivan Reyes-Amezcua, Gilberto Ochoa-Ruiz, and Andres Mendez-Vazquez. Enhancing image classification robustness through adversarial sampling with delta data augmentation (dda). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 274-283, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[44] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[45] Zhigang Su, Dawei Zhou, Nannan Wang, Decheng Liu, Zhen Wang, and Xinbo Gao. Hiding visual information via obfuscating adversarial perturbations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4356-4366, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[46] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24584-24594, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[47] Niko Sünderhauf, Peer Neubert, and Peter Protzel. Are we there yet? challenging seqslam on a \\(3000\\mathrm{km}\\) journey across all four seasons. In Proc. of workshop on long-term autonomy, IEEE international conference on robotics and automation (ICRA), page 2013. CiteSeer, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Ajinkya Tejankar, Maziar Sanjabi, Qifan Wang, Sinong Wang, Hamed Firooz, Hamed Pirsiavash, and Liang Tan." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7206" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Defending against patch-based backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12239-12249, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.216 + ], + "angle": 0, + "content": "[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1808-1817, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[50] Issar Tzachor, Boaz Lerner, Matan Levy, Michael Green, Tal Berkovitz Shalev, Gavriel Habib, Dvir Samuel, Noam Korngut Zailer, Or Shimshi, Nir Darshan, et al. Effovpr: Effective foundation model utilization for visual place recognition. International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[51] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[52] Jiakai Wang, Zixin Yin, Pengfei Hu, Aishan Liu, Renshuai Tao, Haotong Qin, Xianglong Liu, and Dacheng Tao. Defensive patches for robust recognition in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2456-2465, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[53] Zhibo Wang, He Wang, Shuaifan Jin, Wenwen Zhang, Jiahui Hu, Yan Wang, Peng Sun, Wei Yuan, Kaixin Liu, and Kui Ren. Privacy-preserving adversarial facial features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8212-8221, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[54] Frederik Warburg, Soren Hauberg, Manuel Lopez-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2626-2635, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[55] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Yu Li, Chun Yuan, and Dacheng Tao. Task groupings regularization: Data-free meta-learning with heterogeneous pretrained models. In International Conference on Machine Learning, pages 52573-52587. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[56] Yongxian Wei, Zixuan Hu, Zhenyi Wang, Li Shen, Chun Yuan, and Dacheng Tao. Free: Faster and better data-free meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23273-23282, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[57] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[58] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[59] Yongxian Wei, Anke Tang, Li Shen, Zixuan Hu, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.119 + ], + "angle": 0, + "content": "ing as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[60] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, Joao F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8461-8472, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.314 + ], + "angle": 0, + "content": "[62] Jingwen Ye and Xinchao Wang. Ungenerable examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11944-11953, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.906, + 0.371 + ], + "angle": 0, + "content": "[63] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[64] Mubariz Zaffar, Liangliang Nan, and Julian FP Kooij. On the estimation of image-matching uncertainty in visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17743-17753, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.511 + ], + "angle": 0, + "content": "[65] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[66] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7868-7877, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[67] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. Indprompter: Adapting sam to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[68] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25409-25419, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.817 + ], + "angle": 0, + "content": "[69] Quan Zhang, Jinwei Fang, Rui Yuan, Xi Tang, Yuxin Qi, Ke Zhang, and Chun Yuan. Weakly supervised temporal action localization via dual-prior collaborative learning guided by multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 24139-24148, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[70] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10085-10093, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7207" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.161 + ], + "angle": 0, + "content": "[71] Xinwei Zhang, Tianyuan Zhang, Yitong Zhang, and Shuangcheng Liu. Enhancing the transferability of adversarial attacks with stealth preservation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915-2925, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[72] Mingkang Zhu, Tianlong Chen, and Zhangyang Wang. Sparse and imperceptible adversarial attack via a homotopy algorithm. In International Conference on Machine Learning, pages 12868-12877. PMLR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7208" + } + ] +] \ No newline at end of file diff --git a/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_origin.pdf b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9670ff0549c29c21ab1e9042470995c9b880f82 --- /dev/null +++ b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/ecc5979e-3031-4d35-8396-33fead75fa7b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19942370d67601bbb42fbe4ae4ee42db8c877d7b764797ef29ddc8593278451d +size 3544996 diff --git a/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/full.md b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..94ab9ea4452e7892d92334b9c712b8ceaf10da3c --- /dev/null +++ b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/full.md @@ -0,0 +1,361 @@ +# VPR-Cloak: A First Look at Privacy Cloak Against Visual Place Recognition + +Shuting Dong $^{1,2*}$ , Mingzhi Chen $^{3*}$ , Feng Lu $^{1,2}$ , Hao Yu $^{1}$ , Guanghao Li $^{1,3}$ , Zhe Wu $^{2}$ , Ming Tang $^{3\dagger}$ , Chun Yuan $^{1\dagger}$ + +$^{1}$ Tsinghua University, $^{2}$ Pengcheng Laboratory, $^{3}$ Southern University of Science and Technology + +dst21@mails.tsinghua.edu.cn, 12211414@mail.sustech.edu.cn, + +lf22@mails.tsinghua.edu.cn, longinyh@gmail.com, ligh24@mails.tsinghua.edu.cn, + +wuzh02@pcl.ac.cn, tangm3@sustech.edu.cn, yuanc@sz.tsinghua.edu.cn + +# Abstract + +With the rapid advancement of Visual Place Recognition (VPR) systems, their unauthorized use on social media images enables monitoring of individuals' daily movements, posing serious privacy risks. However, privacy protection for addressing these risks in VPR systems remains an underexplored area. While adversarial perturbations have been widely explored for visual privacy protection, existing methods still fail to simultaneously satisfy the black-box constraint, imperceptibility, and real-time performance required in realistic VPR privacy protection scenarios. In this paper, we present the first look at privacy protection in VPR systems and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify decisive regions for place recognition and propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO) to selectively optimize perturbation generation in these areas. To enhance imperceptibility, we further optimize perturbations in the frequency domain, meticulously refining high-frequency components of perturbations while preserving low-frequency structures essential for human perception. Extensive experiments verify that our method outperforms existing SOTA methods. Additionally, our method achieves a $15 \times$ speedup in runtime compared to SOTA methods. We also validate on Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios. The code is released at https://github.com/Dilemma-CMZ/VPR-Cloak. + +# 1. Introduction + +Recent years have witnessed major advances in visual place recognition (VPR) systems, which have been used in various visual applications. While these systems are designed + +![](images/5c357ba79535aa82b63f790630b1646ee2c4607e925ca273671d9b40df20124e.jpg) + +![](images/5a7f86dbde162acf07a30f8fbd355afb3f0a28b2f25ec6185bbf33c9056d180e.jpg) +Figure 1. Illustrations of (a) privacy leakage risk, and (b) privacy protection in visual place recognition. + +with good intentions to enhance localization and navigation, their growing capabilities have intensified privacy concerns. As shown in Fig. 1(a), with the widespread use of social media, images posted by users can be easily crawled and queried without permission, making it easier than ever to locate users' daily movements through place recognition, which brings huge privacy risks. Moreover, location data can often be used to directly or indirectly infer a user's true identity, living environment, or social relationships, further intensifying these privacy risks. Despite these significant + +risks, to the best of our knowledge, no existing research has specifically addressed privacy protection in visual place recognition. This highlights the urgent need to develop robust privacy protection mechanisms for place recognition. + +While no prior work has specifically addressed privacy protection for visual place recognition, adversarial perturbations [44-46, 53] have been widely explored as a general approach to visual privacy. These methods typically modify images to prevent unauthorized recognition systems from accurately identifying their content, either by introducing dense noise [14, 19, 43, 71] or applying patch-based modifications [5, 22, 48, 52]. However, for these perturbations to be effective, they often need to be highly noticeable, leading to visually abrupt and unnatural distortions. This makes it difficult to maintain image visual fidelity within ideal privacy protection for visual place recognition. + +In contrast, some recent works explore imperceptible perturbations [44, 45, 51, 61, 62], aiming at visual fidelity by targeting a limited number of pixels and yet achieving high success rates for privacy protection. However, a major drawback is their low transferability, meaning these perturbations often fail when transferred to different models, especially in black-box settings where the adversary has no access to the target model's parameters or architecture. To address this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent and leveraging a deep ensemble strategy. However, the long inference time of ANDA and MultiANDA makes them unsuitable for privacy protection of place recognition in real-world applications. + +Building on the previous analysis, privacy protection for place recognition still remains an underexplored area, as it requires simultaneously addressing three critical real-world challenges: (1) diminishing the success rate of unauthorized VPR models which in real-world applications are often black-box, (2) maintaining high image visual quality to ensure imperceptibility, and (3) achieving near-real-time performance to support practical, real-world applications. + +To bridge these challenges, we present a first look at privacy protection for place recognition and propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify the decisive regions for place recognition. Based on this, we propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO), which strategically optimizes perturbation generation in images by focusing on these decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way of generating the perturbation protects the most vulnerable areas of the image from unauthorized VPR models. + +In addition, previous research [63] reveals that deep learning models are highly sensitive to high-frequency tex + +tures, while the human visual system primarily perceives low-frequency structural information. Inspired by this, we further elevate perturbation optimization to the frequency domain by isolating the high-frequency components from the perturbation and meticulously refining them using the SAP-PO algorithm. This way ensures that only the high-frequency elements of the perturbation are optimized. As a result, we craft a perturbation that is both precisely targeted and remarkably potent. Through further optimization in the frequency domain, the final output perturbation image (i.e., the protected image) remains imperceptible to the human eye while maintaining a robust defense against VPR models. Our work brings the following contributions: + +- We propose a fast and effective network that provides privacy protection for location information in images while ensuring imperceptibility to the human eye. To the best of our knowledge, this is the first attempt to address privacy protection for visual place recognition. +- We introduce a SAP-PO algorithm, which strategically refines perturbations by focusing on decisive regions for place recognition. +- Extensive experiments on multiple benchmark datasets and on various black-box VPR models verify that our method outperforms existing SOTA methods. Additionally, our method achieves a $15 \times$ speedup in runtime compared to SOTA methods. +- We validate the effectiveness of our privacy protection method based on using commercial APIs, including Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios. + +# 2. Related Works + +Visual Place Recognition. With the rapid development of deep learning [4, 7-9, 23, 28, 40, 42, 55-59, 67-70], significant progress has been made in the VPR field [3, 15, 16, 20, 24]. Existing place recognition models have achieved impressive results not only in standard scenarios but also in challenging conditions such as radar scene [39], nighttime [31, 39], rainy [39], transport [21], perceptual aliasing [64], viewpoint changes [2, 41, 50], LiDAR scan [33, 60], occlusion [30, 41, 50], seasonal variation [41, 50], illumination change [32, 41, 50] dynamic objects. While these systems are designed with good intentions to enhance localization and navigation, they can be exploited by adversaries to expose users' location information, leading to potential risks of identity inference and tracking. Although previous work [26] attempts to attack image retrieval systems, it is not specifically designed for the VPR task. + +Perturbations Generation. While no prior research has specifically focused on privacy protection in visual place recognition, adversarial perturbations [44-46, 53] have been extensively studied as a general approach to visual privacy. Most existing adversarial methods impose norm + +based constraints when generating perturbations. Specifically, many studies [11, 25, 27, 36, 65, 66] adopt infinity-norm or L2-norm constraints, producing dense perturbations that alter a substantial proportion of pixels. Although these methods are effective, they often introduce visible and unnatural distortions, particularly at higher perturbation levels. To address this, recent studies have explored sparse perturbation techniques [6, 10, 12, 17, 34, 38, 65, 72], which restrict modifications to a limited subset of pixels while maintaining high attack success rates. For instance, $\mathrm{PGD}_0$ [6] extends the projected gradient descent framework by enforcing a sparsity constraint. SparseFool [34], building on DeepFool [35], generates sparse untargeted perturbations. ANDA [13] and MultiANDA [13] enhance transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent. However, these methods tend to be resource-intensive and cost a long inference time. + +# 3. Method + +# 3.1. Problem Formulation + +We first introduce real-world location privacy risks. As shown in Fig. 1 (a), the rise of social media makes it easier for adversaries to scrape user-posted photos. Using VPR systems, an adversary can infer where an image was taken. This enables monitoring users' daily movements, uncovering sensitive information about their habits, residences, and frequented places. Moreover, when large-scale user daily movements are monitored, location data can directly or indirectly reveal their identity and social connections, intensifying privacy risks. + +To mitigate these risks, as shown in Fig. 1 (b), we propose a protection transformation $F$ that converts a user's intended post photo $I$ into a protected representation $I_{p} = F(I)$ . Our goal is for $I_{p}$ to remain visually consistent with the original image $I$ while ensuring that adversaries cannot infer the photo's location using VPR systems. We refer to this process as location privacy protection, which can be formulated as an optimization problem: + +$$ +\min _ {F} \mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p r i v a c y}} \left(I _ {p}\right) + \lambda \mathcal {L} _ {\text {v i s u a l}} \left(I, I _ {p}\right), \tag {1} +$$ + +where the visual loss $\mathcal{L}_{\mathrm{visual}}$ ensures that the transformed image remains visually perceptual consistent to the original image, while the privacy loss $\mathcal{L}_{\mathrm{privacy}}$ mislead VPR models in place recognition. The hyperparameter $\lambda$ controls the trade-off between visual fidelity and location privacy. Formal definitions are provided in Sec. 3.3. + +Challenges. Privacy protection in visual place recognition within real-world scenarios presents several unique challenges: Firstly, the adversary's VPR models are often black-box in real-world scenarios, meaning their architectures, parameters, and training data are unknown, making it + +difficult to mislead the black-box VPR models. Secondly, the protection transformation must preserve high visual fidelity and imperceptibility so that the generated protection image $I_{p}$ remains visually consistency from the original image $I$ . Thirdly, the protection method need be computationally efficient, enabling near-real-time processing to support practical use in real-world applications. + +To the best of our knowledge, no existing approach effectively addresses all three of these challenges simultaneously for location privacy protection. Our goal is to develop a method that disrupts location inference while preserving image quality and maintaining practical efficiency, making it suitable for real-world deployment. + +# 3.2. Privacy Cloak Against VPR + +Previous works [44, 45, 51, 61, 62] have explored adding imperceptible perturbations to images to enhance image privacy while aiming to be visually imperceptible. However, these methods often exhibit low transferability, which limits their effectiveness in black-box settings. To handle this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of perturbations. However, the high inference time of ANDA and MultiANDA limits the practicality in real-world applications. + +To address these limitations, we propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify decisive regions for place recognition. Based on the saliency-aware prior, we propose a SAP-PO algorithm, which strategically optimizes perturbation generation in images by focusing on decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way protects the vulnerable regions of the image from unauthorized VPR models. Additionally, we further enhance perturbation optimization in the frequency domain. By isolating and refining high-frequency components using SAP-PO, we craft a precisely targeted yet imperceptible perturbation that remains effective against VPR models while staying unnoticeable to the human eye. + +Saliency-Aware Prior Generation. Instead of applying perturbations randomly, we first identify critical regions for place recognition and selectively perturb them. To achieve this, we introduce a saliency-aware mechanism that acts as a spatial weight to pinpoint key areas for perturbation. For an input unprotected image $I$ , we extract this information by using the fine-tuned Vision Transformer (fine-tuned ViT), which employs the adapter strategy from [18, 31] to fine-tune the pre-trained DINOv2 backbone [37]. The input image is first resized using $\mathcal{R}(I)$ to fit the fine-tuned ViT model's input size. Then, saliency-aware prior $P$ , is obtained by: + +$$ +P = \mathcal {N} \left(\mathcal {A} _ {\mathrm {V i T}} (\mathcal {R} (I))\right), \tag {2} +$$ + +where $\mathcal{N}(\cdot)$ represents normalization to ensure the prior + +![](images/d880ebd1312e8494f071f27831b1a28a09d3677d7dcf4744d63659daa6d2e254.jpg) +Figure 2. Overview of our VPR-Cloak. We introduce a saliency-aware prior to accurately identify decisive regions for place recognition and propose the SAP-PO algorithm, which strategically generates perturbations focused on these decisive regions while minimizing changes in less relevant areas. To enhance imperceptibility, we further elevate perturbation optimization to the frequency domain by isolating and meticulously optimizing only the high-frequencies using the SAP-PO algorithm. + +proper scaling and sum to 1. $\mathcal{A}_{\mathrm{ViT}}$ is the fine-tuned ViT. + +Building on this saliency-aware prior $P$ , we propose SAP-PO to optimize perturbation generation, ensuring perturbations are concentrated in decisive regions for place recognition. We provide two optimization strategies: spatial domain perturbation optimization for defense against VPR and frequency domain perturbation optimization for visual imperceptibility. + +Spatial Perturbation for Defense against VPR. In the spatial domain, SAP-PO, guided by the saliency prior, focuses on perturbing key regions crucial for place recognition, maximizing its impact on reducing recognition accuracy. The spatially-perturbed image $I_{\mathrm{spa}}^{\prime}$ is obtained as: + +$$ +I _ {\mathrm {s p a}} ^ {\prime} = \mathcal {F} _ {\mathrm {s p a}} (I, P), \tag {3} +$$ + +where $\mathcal{F}_{\mathrm{spa}}$ is our proposed SAP-PO in spatial domain. + +Frequency Perturbation for Imperceptible. Previous study [63] have shown that the human eye is sensitive to low frequencies, while deep learning models are sensitive to high frequencies. Inspired by this, we elevate perturbation optimization to the frequency domain by isolating and meticulously refining only the high-frequency components of the perturbation using the SAP-PO algorithm. + +To do this, we apply the Discrete Cosine Transform (DCT) to the perturbed image $I_{\mathrm{spa}}^{\prime}$ , yielding its frequency representation $f^{\prime}$ . We specifically optimize the high-frequency components of the perturbation using SAP-PO and generate the frequency-perturbed image $I_{\mathrm{fre}}^{\prime}$ : + +$$ +f ^ {\prime} = \mathcal {D} \left(I _ {\mathrm {s p a}} ^ {\prime}\right), \tag {4} +$$ + +$$ +I _ {\mathrm {f r e}} ^ {\prime} = \mathcal {F} _ {\mathrm {f r e}} \left(f ^ {\prime}, P\right), \tag {5} +$$ + +where $\mathcal{D}(\cdot)$ denotes DCT. $\mathcal{F}_{\mathrm{fre}}$ is our proposed SAP-PO in frequency domain. + +Next, we apply the inverse DCT to transform the frequency-perturbed image $I_{\mathrm{fre}}^{\prime}$ back to the spatial domain, obtaining the final protected image $I_{\mathrm{p}}$ : + +$$ +I _ {\mathrm {p}} = \mathcal {D} ^ {- 1} \left(I _ {\mathrm {f r e}} ^ {\prime}\right), \tag {6} +$$ + +where $\mathcal{D}^{-1}$ is the inverse DCT. + +Through optimization in the spatial and frequency domains, we strategically refine and enforce the distribution of perturbations to concentrate on decisive regions for place recognition and high-frequency areas. + +# 3.3. Saliency-Aware Prior Guided Perturbation Optimization + +We propose SAP-PO to generate effective and visually imperceptible perturbations for defending VPR models. By integrating the saliency-aware prior $P$ , we focus perturbations on the decisive regions for place recognition. + +To begin, we initialize the perturbation as: $\delta = 0$ . For an input image $I$ , we retrieve multiple positive reference images $\{I_{\mathrm{pos},i}\}$ from the database and extract their features $\{F_{\mathrm{pos},i}\}$ using the Surrogate VPR: + +$$ +\left\{F _ {\text {p o s}, i} \right\} = \mathcal {M} _ {\mathrm {V P R}} \left(\left\{I _ {\text {p o s}, i} \right\}\right), \tag {7} +$$ + +where $\mathcal{M}_{\mathrm{VPR}}$ denotes the Surrogate VPR model. To enhance the robustness of our perturbations against unauthorized VPR models, for each input image, we randomly select one from multiple VPR models $\{\mathcal{M}_1,\mathcal{M}_2,\dots ,\mathcal{M}_m\}$ + +Algorithm 1 Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO) +Ensure: Perturbed Image $I^{\prime}$ +```txt +Require: Image $I$ , Saliency-aware prior $P$ , Surrogate VPR Model $\mathcal{M}_{\mathrm{VPR}}$ , Trade-off Coefficient $\lambda$ , Step Size $\epsilon$ , Maximum Iterations $K$ , Domain Type $D_{\mathrm{type}}$ +``` + +1: Initialize perturbation $\delta \gets 0$ +2: Retrieve multiple positive samples from database: $\{I_{\mathrm{pos},s}\} \gets$ retrieve_positives(I) +3: Generate $\{F_{\mathrm{pos},i}\}$ using the Surrogate VPR model: $\{F_{\mathrm{pos},i}\} \gets \mathcal{M}_{\mathrm{VPR}}(\{I_{\mathrm{pos},i}\})$ +4: for $t = 1$ to $K$ do +5: Compute perturbed image: $I' \gets I + \delta$ +6: if $D_{\text{type}} = \text{Frequency}$ then +7: $I^{\prime}\gets I + \mathrm{RemoveLowFreq}(\delta)$ +8: $P\gets 1$ +9: $\triangleright$ For clarity, this algorithm omits transformations between spatial and frequency domains. +10: end if +11: Generate feature vector: $F' \gets \mathcal{M}_{\mathrm{VPR}}(I')$ +12: Compute average distance to positive samples: $\mathcal{L}_{\mathrm{pos}}\gets \frac{1}{N}\sum_{i = 1}^{N}\mathrm{Distance}(F',F_{\mathrm{pos},i})$ +13: Compute PSNR loss: $\mathcal{L}_{\mathrm{psnr}}\gets \mathrm{PSNR}(I,I^{\prime})$ +14: Compute total loss: $\mathcal{L}\gets -\mathcal{L}_{\mathrm{pos}} - \lambda \mathcal{L}_{\mathrm{psnr}}$ +15: Compute gradient: $\nabla \mathcal{L}\gets$ compute-gradient $(\mathcal{L},\delta)$ +16: Update perturbation: $\delta \gets \delta +\eta \nabla \mathcal{L}$ +17: Apply Saliency-aware prior $P$ : $\delta \gets \mathrm{clip}(\delta, -P \cdot \epsilon, P \cdot \epsilon)$ +18: end for +19: Return the final perturbed image: $I'$ +20: return $I^{\prime}$ + +as the Surrogate VPR model $\mathcal{M}_{\mathrm{VPR}}$ . This strategy ensures that the perturbations are not overfitted to a specific model but instead generalize well to a variety of potential recognition systems. In real-world applications, the VPR model set can be dynamically updated, enabling adaptation to stronger protection against unauthorized attacks. + +At each iteration, we obtain the perturbed image $I^{\prime}$ as $I^{\prime} = I + \delta$ . If the frequency domain is used ( $D_{\mathrm{type}} = \mathrm{Frequency}$ ), the perturbation is processed by applying a high-pass filter to remove low frequencies. + +Then, we use the Surrogate VPR: $F' = \mathcal{M}_{\mathrm{VPR}}(I')$ to generate feature $F'$ . The feature distance between the perturbed image and the retrieved positive reference images is then computed as the average distance: + +$$ +\mathcal {L} _ {\mathrm {p o s}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left\| \frac {F ^ {\prime}}{\| F ^ {\prime} \| _ {2}} - \frac {F _ {\mathrm {p o s} , i}}{\| F _ {\mathrm {p o s} , i} \| _ {2}} \right\| _ {2}, \tag {8} +$$ + +where $N$ is the number of retrieved positive samples, and $\| X\| _2$ denotes the L2 norm. + +In addition, to ensure that the generated perturbation images maintain high visual fidelity consistent with the original images, we also impose the constraint of preserving pixel-wise similarity by employing the PSNR loss. + +The total loss balances the trade-off between maximizing the feature distance and preserving image quality: + +$$ +\mathcal {L} _ {\text {t o t a l}} = - \mathcal {L} _ {\mathrm {p o s}} - \lambda \mathcal {L} _ {\mathrm {p s n r}}, \tag {9} +$$ + +where $\lambda$ is the trade-off coefficient. Notably, $-\mathcal{L}_{\mathrm{pos}}$ corresponding to $\mathcal{L}_{\mathrm{privacy}}$ while $-\mathcal{L}_{\mathrm{psnr}}$ corresponding to $\mathcal{L}_{\mathrm{visual}}$ in Eq. (1). + +To ensure the perturbations remain concentrated in decisive regions, we optimize their distribution by using the saliency-aware prior: + +$$ +\delta = \mathcal {C} (\delta , - P \cdot \epsilon , P \cdot \epsilon), \tag {10} +$$ + +where $\epsilon$ is the step size. And $\mathcal{C}$ is clip to constrain the perturbation within saliency-aware bounds. + +After completing $K$ iterations, we obtain the final perturbed (protected) image $I_{p} = I^{\prime}$ that is optimized to disrupt place recognition while maintaining visual consistency with the original image. + +# 4. Experiments + +# 4.1. Experimental Setting + +We conduct the experiments on several VPR benchmark datasets: Pitts30k [1], MSLS [54], Tokyo247 [49], and Nordland [47]. We conduct experiments using PyTorch on a setup with 8 NVIDIA GeForce RTX 3090 GPUs. For our method, the trade-off coefficient $\lambda$ is set to 0.3 in both spatial and frequency domains. Specifically, in the frequency domain, we utilize an $8 \times 8$ DCT block size (yielding 64 blocks), with a frequency threshold of 12. For both ANDA and MultiANDA, we set the spatial transformation range to 0.3 (allowing translations by $\pm 30\%$ of the image size), gradient averaging to 25 samples, and multi-scale transformations to 5 copies. MultiANDA introduces a diversity probability of 0.5. The Gaussian kernel sizes differ between the two methods: $7(\sigma = 3)$ for ANDA and $15(\sigma = 3)$ for MultiANDA. All these methods evaluate over 10 iterations with a batch size of 32 and a learning rate of $8 \times 10^{-3}$ . + +# 4.2. Comparison Study + +Black-Box and White-Box Performance. We select five popular open-source VPR models: Crica [30], DHE [29], Salad [21], Sela [31], and VLAD-BuFF [24]. We randomly select one model as the surrogate to generate protected images; evaluations on this surrogate constitute the white-box setting, while evaluations on the remaining four unseen models constitute the black-box setting. Notably, in black-box setting, a single surrogate VPR model is selected from the VPR pool as the fixed model for training + +Table 1. Protection success rate (%) comparisons (Epsilon=12/255). ΔR@1, ΔR@5, and ΔR@10 measure the reduction in the model's retrieval accuracy after applying protection, with larger values indicating stronger protection. + +
DatasetModelANDA_CricaMULTI-ANDA_CricaOurs_CricaOurs_Sela
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KVLAD-BuFF18.59.3-20.215.1-45.437.6-36.227.7-
DHE23.116.09.831.321.913.339.625.916.241.625.919.2
Sela17.713.66.319.215.88.840.727.020.073.254.940.1
Crica50.835.421.952.638.225.069.561.857.651.044.327.9
Salad22.012.38.125.815.410.935.723.715.943.028.518.5
TOKYO247VLAD-BuFF19.611.7-20.512.2-36.328.8-28.423.9-
DHE23.229.221.434.541.329.745.850.750.738.046.644.5
Sela34.225.816.236.827.519.744.427.226.363.664.046.2
Crica49.835.826.752.038.229.382.282.380.052.434.126.9
MSLSVLAD-BuFF29.318.4-30.620.6-43.534.6-37.827.8-
DHE37.826.719.944.933.325.852.940.033.349.336.232.2
Sela24.414.17.328.119.311.041.234.327.472.865.346.7
Crica53.743.433.057.947.237.381.583.473.840.730.326.2
Salad36.434.423.442.937.330.052.140.424.149.937.327.9
MSLSDHE60.476.177.961.377.581.661.678.082.061.577.381.2
Salad72.885.284.474.087.688.674.688.190.274.787.990.2
CHALLENGESela72.683.980.673.084.483.073.487.390.068.376.275.6
Crica66.774.364.567.375.866.669.079.979.473.387.189.7
NORDLANDVLAD-BuFF58.962.9-60.766.0-68.970.7-66.871.3-
DHE43.653.649.353.365.465.764.478.677.864.877.374.2
Sela52.357.446.254.561.951.464.172.972.486.983.364.0
Crica63.161.351.066.062.352.387.984.879.262.169.966.9
Salad56.960.250.262.867.856.769.378.074.468.879.172.2
+ +![](images/71139ffc9d806d385a26ef46ce9948af4eae8bbd6448c2b58feb6427ba150a48.jpg) +Figure 3. Qualitative results. Our method robustly prevents location retrieval while preserving image quality. Green frames indicate successful protection, while orange frames indicate the opposite. + +the protected image set, while the remaining VPR models in the pool are strictly not used in the training process and instead serve as unseen models for testing. Specifically, $\mathrm{ANDA}_{\mathrm{Crica}}$ , MULTI- $\mathrm{ANDA}_{\mathrm{Crica}}$ , and $\mathrm{Ours}_{\mathrm{Crica}}$ use Crica as the surrogate model, whereas $\mathrm{Ours}_{\mathrm{Sela}}$ uses Sela. In Tab. 1, our method consistently achieves higher protection success rates than ANDA [13] and MultiANDA [13] across all datasets. $\mathrm{Ours}_{\mathrm{Crica}}$ and $\mathrm{Ours}_{\mathrm{Sela}}$ significantly reduce retrieval accuracy, highlighting their strong privacy protection capability. Notably, on PITTS30K and MSLS, our approach markedly outperforms baselines. These results demonstrate the strong transferability and effectiveness of our method in + +disrupting both black-box and white-box retrieval. + +Image Quality Evaluations. As shown in Tab. 2, our method consistently outperforms ANDA and MultiANDA across most datasets, achieving superior visual fidelity while maintaining strong protection. In PITTS30K, OursSela achieves the best visual fidelity results, while OursCrica attains the lowest Avg Epsilon, ensuring higher imperceptibility. These results demonstrate that our approach not only provides stronger privacy protection but also ensures better image quality and imperceptibility. + +Runtime Comparisons. The results on the Tab. 3 clearly demonstrate that our method surpasses ANDA by $15.9\%$ , + +![](images/ef906974626817ca52ed59b75f642378b5c16fa1c687641af6841ead0dcfcc58.jpg) +Figure 4. Protection on Commercial APIs. Our method successfully prevents Google and Microsoft Bing from retrieving location information while ensuring imperceptible modifications, highlighting our method's practical superiority. + +Table 2. Image quality comparison (Epsilon=12/255). Higher PSNR/SSIM and lower LPIPS indicate better visual fidelity, while lower Avg Epsilon suggests more imperceptible perturbations. + +
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KANDACrica0.14429.840.8724.78
MULTI-ANDACrica0.13230.170.8814.44
OursCrica0.09131.560.9044.02
OursSela0.08232.000.9174.14
TOKYO247ANDACrica0.08930.000.8815.34
MULTI-ANDACrica0.08330.200.8975.14
OursCrica0.09929.110.8865.32
OursSela0.11328.560.8525.42
MSLSANDACrica0.13930.660.8724.22
MULTI-ANDACrica0.11130.140.8894.00
OursCrica0.10230.990.9014.38
OursSela0.11930.260.8894.02
MSLS CHALLENGEANDA (Crica)0.12633.010.9094.68
MULTI-ANDACrica0.12133.290.9174.38
OursCrica0.07334.200.9364.68
OursSela0.09233.210.9143.42
NORDLANDANDA (Crica)0.10530.420.8763.04
MULTI-ANDACrica0.09430.990.8962.78
OursCrica0.09232.030.9172.76
OursSela0.10631.220.8963.38
+ +Table 3. Runtime and average protection success rate on NORDLAND. Our method achieves a $15 \times$ speedup over MULTI-ANDA. + +
MethodANDAMULTI-ANDAOurs
Avg ΔR@155.059.570.9
Avg ΔR@559.164.777.0
Avg ΔR@1049.256.576.0
Runtime (s)0.1690.5510.036
+ +$17.9\%$ , and $26.8\%$ , and outperforms MultiANDA by $11.4\%$ , $12.3\%$ , and $19.5\%$ in terms of $\Delta R@1$ , $\Delta R@5$ , and $\Delta R@10$ , respectively. Additionally, our method requires + +only 0.036 seconds per run, approximately 4.7 times faster than ANDA and over $15 \times$ faster than MultiANDA. These results confirm that our approach significantly enhances protection success while maintaining remarkable efficiency, making it suitable for real-time applications. + +Visual Comparisons. Fig. 3 demonstrates that our method effectively prevents location retrieval, surpassing existing SOTA approaches under challenging scenarios, including drastic viewpoint changes, seasonal variations, day-night transitions, and occlusions. Our approach consistently obscures identifiable features while preserving visual integrity, ensuring robust privacy protection suitable for diverse real-world conditions. + +# 4.3. Protection Performance on Commercial APIs + +Fig. 4 illustrates the effectiveness of our method, evaluated on world-famous commercial APIs including Google and Microsoft Bing. Originally, these systems could achieve correct retrieval of identifying and matching locations from visual cues. However, after applying our method, these images become effectively shielded from such retrieval processes. Notably, while our method significantly disrupts the feature extraction necessary for location recognition, it preserves the visual quality of the images. This balance between robust privacy protection and maintaining visual integrity highlights the practical applicability and superiority of our approach in real-world scenarios. + +# 4.4. Ablation Studies + +To evaluate the effectiveness of our method, we conduct an ablation study on the impact of spatial and frequency-domain perturbations and the saliency-aware prior. + +Effectiveness of Spatial and Frequency Domain Perturbation Optimization. Tab. 4 shows that spatial domain at + +Table 4. Ablation study on protection success rate. + +
DatasetModelSpatialFrequencySpatial + Frequency
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KCrica51.642.331.927.522.717.657.147.737.5
DHE31.218.111.221.512.87.635.520.314.4
Salad31.717.210.820.210.35.733.419.312.2
Sela24.613.110.815.912.57.530.319.616.2
MSLSCrica71.358.337.950.244.230.778.165.843.7
DHE43.130.325.127.923.016.546.432.927.2
Salad44.332.921.626.919.310.849.437.025.3
Sela31.925.220.321.413.34.638.828.724.6
NORDLANDCrica79.371.249.971.056.647.884.675.659.0
DHE56.168.265.437.844.135.063.576.476.1
Salad65.269.370.134.842.438.569.378.074.4
Sela58.365.162.331.427.927.361.868.368.4
+ +Table 5. Ablation study on image quality. + +
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KSpatial0.05732.430.9292.85
Frequency0.04233.830.9462.42
Spatial + Frequency0.05732.420.9282.84
MSLSSpatial0.06732.280.9152.80
Frequency0.05933.190.9372.57
Spatial + Frequency0.06632.210.9162.82
NORDLANDSpatial0.07332.410.9171.83
Frequency0.06032.990.9312.01
Spatial + Frequency0.07132.130.9151.84
+ +Table 6. Ablation study on saliency-aware prior. + +
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KUnguided0.18329.040.8175.69
Guided0.05732.420.9282.84
MSLSUnguided0.19227.830.8015.91
Guided0.06632.210.9162.82
NORDLANDUnguided0.18828.410.8335.88
Guided0.07132.130.9151.84
+ +tacks alone yield stronger protection than frequency domain attacks. This indicates that spatial modifications, by altering key structural features, are inherently more effective. Yet, when spatial and frequency domain attacks are combined, the success rates improve further, $2\%$ to $9\%$ . This synergy demonstrates that frequency domain perturbations, while less potent on their own, complement spatial attacks by enhancing overall robustness without significant loss of performance. Tab. 5 confirms that frequency perturbations preserve visual quality exceptionally well. Although the combined method results in a slight degradation, this minor compromise in visual fidelity is well justified by the substantial gain in protection. In essence, integrating frequency domain modifications with spatial attacks strikes an effective balance between robustness and image quality. + +Effectiveness of Saliency-Aware Prior. Tab. 6 clearly demonstrates the advantages of incorporating a saliency-aware prior. Compared to the unguided way, the guided method markedly improves image fidelity. For example, on the PITTS30K dataset, the guided approach reduces LPIPS + +from 0.183 to 0.057, increases PSNR from 29.04 to 32.42, and enhances SSIM from 0.817 to 0.928. Additionally, the average perturbation (Avg Epsilon) is lowered from 5.69 to 2.84. Similar trends are observed in the MSLS and NORDLAND datasets. These results indicate that the saliency-aware prior effectively directs perturbations to less noticeable regions, thereby maintaining high visual quality while keeping modifications nearly imperceptible. + +# 5. Conclusion + +In this paper, we present the first look at privacy protection in VPR and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify key regions crucial for place recognition and propose SAP-PO to selectively enhance perturbation generation in these areas. To further improve imperceptibility, we optimize perturbations in the frequency domain, refining high-frequency components while preserving low-frequency structures essential for human perception. Extensive experiments on multiple benchmark datasets and across various black-box VPR models demonstrate that our method surpasses SOTA perturbation methods in effectiveness. Moreover, our approach achieves a $15 \times$ speedup in runtime compared to SOTA methods. We also validate its practicality using commercial APIs, including Google and Microsoft Bing, showcasing its real-world applicability. + +# Acknowledgments + +This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008). This work was supported in part by the National Natural Science Foundation of China under Grant 62202214 and Guangdong Basic and Applied Basic Research Foundation under Grant 2023A1515012819. + +# References + +[1] Gabriele Berton, Riccardo Mereu, Gabriele Trivigno, Carlo Masone, Gabriela Csurka, Torsten Sattler, and Barbara Caputo. Deep visual geo-localization benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5407, 2022. 5 +[2] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11080-11090, 2023. 2 +[3] Gabriele Berton, Lorenz Junglas, Riccardo Zaccone, Thomas Pollok, Barbara Caputo, and Carlo Masone. Meshvpr: Citywide visual place recognition using 3d meshes. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 2 +[4] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2311-2319, 2025. 2 +[5] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 2 +[6] Francesco Croce and Matthias Hein. Sparse and imperceivable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4724-4732, 2019. 3 +[7] Shuting Dong, Feng Lu, Zhe Wu, and Chun Yuan. Dfvsr: Directional frequency video super-resolution via asymmetric and enhancement alignment network. In *IJCAI*, pages 681–689, 2023. 2 +[8] Shuting Dong, Feng Lu, and Chun Yuan. Frequency reciprocal action and fusion for single image super-resolution. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. +[9] Shuting Dong, Zhe Wu, Feng Lu, and Chun Yuan. Enhanced image deblurring: An efficient frequency exploitation and preservation network. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7184-7193, 2023. 2 +[10] Xiaoyi Dong, Dongdong Chen, Jianmin Bao, Chuan Qin, Lu Yuan, Weiming Zhang, Nenghai Yu, and Dong Chen. Greedyfool: Distortion-aware sparse adversarial attack. Advances in Neural Information Processing Systems, 33: 11226-11236, 2020. 3 +[11] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3 +[12] Yanbo Fan, Baoyuan Wu, Tuanhui Li, Yong Zhang, Mingyang Li, Zhifeng Li, and Yujiu Yang. Sparse adversaries + +ial attack via perturbation factorization. In Computer VisionECCV 2020:16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 35-50. Springer, 2020. 3 +[13] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 2, 3, 6 +[14] Samar Fares and Karthik Nandakumar. Attack to defend: Exploiting adversarial attacks for detecting poisoned models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24726-24735, 2024. 2 +[15] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2024. 2 +[16] Tianrui Guan, Aswath Muthuselvam, Montana Hoover, Xijun Wang, Jing Liang, Adarsh Jagan Sathyamoorthy, Damon Conover, and Dinesh Manocha. Crossloc3d: Aerial-ground cross-source 3d place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11335-11344, 2023. 2 +[17] Ziwen He, Wei Wang, Jing Dong, and Tieniu Tan. Transferable sparse adversarial attack. arXiv preprint arXiv:2105.14727, 2021. 3 +[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3 +[19] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15014–15023, 2022. 2 +[20] Sergio Izquierdo and Javier Civera. Close, but not there: Boosting geographic distance sensitivity in visual place recognition. In European Conference on Computer Vision, pages 240-257. Springer, 2024. 2 +[21] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 17658-17668, 2024. 2, 5 +[22] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4379-4389, 2023. 2 +[23] Tong Jin, Feng Lu, Shuyu Hu, Chun Yuan, and Yunpeng Liu. Editformer: An efficient decoder transformer for visual place recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2 +[24] Ahmad Khaliq, Ming Xu, Stephen Hausler, Michael Milford, and Sourav Garg. Vlad-buff: burst-aware fast feature aggre + +gation for visual place recognition. In European Conference on Computer Vision, pages 447-466. Springer, 2024. 2, 5 +[25] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3 +[26] Jie Li, Rongrong Ji, Hong Liu, Xiaopeng Hong, Yue Gao, and Qi Tian. Universal perturbation attack against image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4899-4908, 2019. 2 +[27] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 3 +[28] Feng Lu, Lijun Zhang, Shuting Dong, Baifan Chen, and Chun Yuan. Aanet: Aggregation and alignment network with semi-hard positive sample mining for hierarchical place recognition. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11771-11778. IEEE, 2023. 2 +[29] Feng Lu, Shuting Dong, Lijun Zhang, Bingxi Liu, Xiangyuan Lan, Dongmei Jiang, and Chun Yuan. Deep homography estimation for visual place recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10341-10349, 2024. 5 +[30] Feng Lu, Xiangyuan Lan, Lijun Zhang, Dongmei Jiang, Yaowei Wang, and Chun Yuan. Cricavpr: Cross-image correlation-aware representation learning for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16772-16782, 2024. 2, 5 +[31] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. International Conference on Learning Representations, 2024. 2, 3, 5 +[32] Feng Lu, Xinyao Zhang, Canming Ye, Shuting Dong, Lijun Zhang, Xiangyuan Lan, and Chun Yuan. Supervlad: Compact and robust image descriptors for visual place recognition. Advances in Neural Information Processing Systems, 37:5789-5816, 2025. 2 +[33] Lun Luo, Shuhang Zheng, Yixuan Li, Yongzhi Fan, Beinan Yu, Si-Yuan Cao, Junwei Li, and Hui-Liang Shen. Bevplace: Learning lidar-based place recognition using bird's eye view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8700-8709, 2023. 2 +[34] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087–9096, 2019. 3 +[35] SD Moosavi-Dezfooli. a simple and accurate method to fool deep neural networks/s. moosavi-dezfooli, a. fawzi, p. frossard. Preprint—July, 2016. 3 +[36] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE transactions on pattern + +analysis and machine intelligence, 41(10):2452-2465, 2018. 3 +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3 +[38] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 3 +[39] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17595-17605, 2024. 2 +[40] Yuxin Qi, Quan Zhang, Xi Lin, Xiu Su, Jiani Zhu, Jingyu Wang, and Jianhua Li. Seeing beyond noise: Joint graph structure evaluation and denoising for multimodal recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12461-12469, 2025. 2 +[41] Qibo Qiu, Shun Zhang, Haiming Gao, Honghui Yang, Haochao Ying, Wenxiao Wang, and Xiaofei He. Emvp: Embracing visual foundation model for visual place recognition with centroid-free probing. Advances in Neural Information Processing Systems, 37:120928-120950, 2025. 2 +[42] Tianyi Alex Qiu, Zhonghao He, Tejasveer Chugh, and Max Kleiman-Weiner. The lock-in hypothesis: Stagnation by algorithm. arXiv preprint arXiv:2506.06166, 2025. 2 +[43] Ivan Reyes-Amezcua, Gilberto Ochoa-Ruiz, and Andres Mendez-Vazquez. Enhancing image classification robustness through adversarial sampling with delta data augmentation (dda). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 274-283, 2024. 2 +[44] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 2, 3 +[45] Zhigang Su, Dawei Zhou, Nannan Wang, Decheng Liu, Zhen Wang, and Xinbo Gao. Hiding visual information via obfuscating adversarial perturbations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4356-4366, 2023. 2, 3 +[46] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24584-24594, 2024. 2 +[47] Niko Sünderhauf, Peer Neubert, and Peter Protzel. Are we there yet? challenging seqslam on a $3000\mathrm{km}$ journey across all four seasons. In Proc. of workshop on long-term autonomy, IEEE international conference on robotics and automation (ICRA), page 2013. CiteSeer, 2013. 5 +[48] Ajinkya Tejankar, Maziar Sanjabi, Qifan Wang, Sinong Wang, Hamed Firooz, Hamed Pirsiavash, and Liang Tan. + +Defending against patch-based backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12239-12249, 2023. 2 +[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1808-1817, 2015. 5 +[50] Issar Tzachor, Boaz Lerner, Matan Levy, Michael Green, Tal Berkovitz Shalev, Gavriel Habib, Dvir Samuel, Noam Korngut Zailer, Or Shimshi, Nir Darshan, et al. Effovpr: Effective foundation model utilization for visual place recognition. International Conference on Learning Representations, 2025. 2 +[51] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3 +[52] Jiakai Wang, Zixin Yin, Pengfei Hu, Aishan Liu, Renshuai Tao, Haotong Qin, Xianglong Liu, and Dacheng Tao. Defensive patches for robust recognition in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2456-2465, 2022. 2 +[53] Zhibo Wang, He Wang, Shuaifan Jin, Wenwen Zhang, Jiahui Hu, Yan Wang, Peng Sun, Wei Yuan, Kaixin Liu, and Kui Ren. Privacy-preserving adversarial facial features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8212-8221, 2023. 2 +[54] Frederik Warburg, Soren Hauberg, Manuel Lopez-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2626-2635, 2020. 5 +[55] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Yu Li, Chun Yuan, and Dacheng Tao. Task groupings regularization: Data-free meta-learning with heterogeneous pretrained models. In International Conference on Machine Learning, pages 52573-52587. PMLR, 2024. 2 +[56] Yongxian Wei, Zixuan Hu, Zhenyi Wang, Li Shen, Chun Yuan, and Dacheng Tao. Free: Faster and better data-free meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23273-23282, 2024. +[57] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025. +[58] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025. +[59] Yongxian Wei, Anke Tang, Li Shen, Zixuan Hu, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging + +ing as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 2 +[60] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, Joao F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8461-8472, 2023. 2 +[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 3 +[62] Jingwen Ye and Xinchao Wang. Ungenerable examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11944-11953, 2024. 2, 3 +[63] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 2, 4 +[64] Mubariz Zaffar, Liangliang Nan, and Julian FP Kooij. On the estimation of image-matching uncertainty in visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17743-17753, 2024. 2 +[65] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020. 3 +[66] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7868-7877, 2021. 3 +[67] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. Indprompter: Adapting sam to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations. 2 +[68] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25409-25419, 2024. +[69] Quan Zhang, Jinwei Fang, Rui Yuan, Xi Tang, Yuxin Qi, Ke Zhang, and Chun Yuan. Weakly supervised temporal action localization via dual-prior collaborative learning guided by multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 24139-24148, 2025. +[70] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10085-10093, 2025. 2 + +[71] Xinwei Zhang, Tianyuan Zhang, Yitong Zhang, and Shuangcheng Liu. Enhancing the transferability of adversarial attacks with stealth preservation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915-2925, 2024. 2 +[72] Mingkang Zhu, Tianlong Chen, and Zhangyang Wang. Sparse and imperceptible adversarial attack via a homotopy algorithm. In International Conference on Machine Learning, pages 12868-12877. PMLR, 2021. 3 \ No newline at end of file diff --git a/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/images.zip b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f71dd9d5fdd24f5ae080b230b546894a1814b79d --- /dev/null +++ b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:292eb2123347d23356d6b38c30fd87bc17fdeea3b58174c61e47a672ec002538 +size 763430 diff --git a/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/layout.json b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de4211a4c734c36b13b9e87798edb9bba58c9663 --- /dev/null +++ b/2025/VPR-Cloak_ A First Look at Privacy Cloak Against Visual Place Recognition/layout.json @@ -0,0 +1,10401 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 103, + 542, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 103, + 542, + 122 + ], + "spans": [ + { + "bbox": [ + 68, + 103, + 542, + 122 + ], + "type": "text", + "content": "VPR-Cloak: A First Look at Privacy Cloak Against Visual Place Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": "Shuting Dong" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Mingzhi Chen" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Feng Lu" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Hao Yu" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Guanghao Li" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Zhe Wu" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Ming Tang" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "text", + "content": ", Chun Yuan" + }, + { + "bbox": [ + 167, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "spans": [ + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "text", + "content": "Tsinghua University, " + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "text", + "content": "Pengcheng Laboratory, " + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 74, + 171, + 536, + 186 + ], + "type": "text", + "content": "Southern University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 187, + 460, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 187, + 460, + 198 + ], + "spans": [ + { + "bbox": [ + 147, + 187, + 460, + 198 + ], + "type": "text", + "content": "dst21@mails.tsinghua.edu.cn, 12211414@mail.sustech.edu.cn," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 201, + 512, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 201, + 512, + 213 + ], + "spans": [ + { + "bbox": [ + 96, + 201, + 512, + 213 + ], + "type": "text", + "content": "lf22@mails.tsinghua.edu.cn, longinyh@gmail.com, ligh24@mails.tsinghua.edu.cn," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 216, + 481, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 216, + 481, + 227 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 481, + 227 + ], + "type": "text", + "content": "wuzh02@pcl.ac.cn, tangm3@sustech.edu.cn, yuanc@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 281, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 297, + 605 + ], + "type": "text", + "content": "With the rapid advancement of Visual Place Recognition (VPR) systems, their unauthorized use on social media images enables monitoring of individuals' daily movements, posing serious privacy risks. However, privacy protection for addressing these risks in VPR systems remains an underexplored area. While adversarial perturbations have been widely explored for visual privacy protection, existing methods still fail to simultaneously satisfy the black-box constraint, imperceptibility, and real-time performance required in realistic VPR privacy protection scenarios. In this paper, we present the first look at privacy protection in VPR systems and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify decisive regions for place recognition and propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO) to selectively optimize perturbation generation in these areas. To enhance imperceptibility, we further optimize perturbations in the frequency domain, meticulously refining high-frequency components of perturbations while preserving low-frequency structures essential for human perception. Extensive experiments verify that our method outperforms existing SOTA methods. Additionally, our method achieves a " + }, + { + "bbox": [ + 54, + 281, + 297, + 605 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 54, + 281, + 297, + 605 + ], + "type": "text", + "content": " speedup in runtime compared to SOTA methods. We also validate on Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios. The code is released at https://github.com/Dilemma-CMZ/VPR-Cloak." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 628, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "content": "Recent years have witnessed major advances in visual place recognition (VPR) systems, which have been used in various visual applications. While these systems are designed" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 254, + 555, + 387 + ], + "blocks": [ + { + "bbox": [ + 317, + 254, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 317, + 254, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 555, + 387 + ], + "type": "image", + "image_path": "5c357ba79535aa82b63f790630b1646ee2c4607e925ca273671d9b40df20124e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 399, + 553, + 536 + ], + "blocks": [ + { + "bbox": [ + 317, + 399, + 553, + 536 + ], + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 536 + ], + "type": "image", + "image_path": "5a7f86dbde162acf07a30f8fbd355afb3f0a28b2f25ec6185bbf33c9056d180e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 540, + 555, + 562 + ], + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 562 + ], + "type": "text", + "content": "Figure 1. Illustrations of (a) privacy leakage risk, and (b) privacy protection in visual place recognition." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "with good intentions to enhance localization and navigation, their growing capabilities have intensified privacy concerns. As shown in Fig. 1(a), with the widespread use of social media, images posted by users can be easily crawled and queried without permission, making it easier than ever to locate users' daily movements through place recognition, which brings huge privacy risks. Moreover, location data can often be used to directly or indirectly infer a user's true identity, living environment, or social relationships, further intensifying these privacy risks. Despite these significant" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7197" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "risks, to the best of our knowledge, no existing research has specifically addressed privacy protection in visual place recognition. This highlights the urgent need to develop robust privacy protection mechanisms for place recognition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 266 + ], + "type": "text", + "content": "While no prior work has specifically addressed privacy protection for visual place recognition, adversarial perturbations [44-46, 53] have been widely explored as a general approach to visual privacy. These methods typically modify images to prevent unauthorized recognition systems from accurately identifying their content, either by introducing dense noise [14, 19, 43, 71] or applying patch-based modifications [5, 22, 48, 52]. However, for these perturbations to be effective, they often need to be highly noticeable, leading to visually abrupt and unnatural distortions. This makes it difficult to maintain image visual fidelity within ideal privacy protection for visual place recognition." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 445 + ], + "type": "text", + "content": "In contrast, some recent works explore imperceptible perturbations [44, 45, 51, 61, 62], aiming at visual fidelity by targeting a limited number of pixels and yet achieving high success rates for privacy protection. However, a major drawback is their low transferability, meaning these perturbations often fail when transferred to different models, especially in black-box settings where the adversary has no access to the target model's parameters or architecture. To address this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent and leveraging a deep ensemble strategy. However, the long inference time of ANDA and MultiANDA makes them unsuitable for privacy protection of place recognition in real-world applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 544 + ], + "type": "text", + "content": "Building on the previous analysis, privacy protection for place recognition still remains an underexplored area, as it requires simultaneously addressing three critical real-world challenges: (1) diminishing the success rate of unauthorized VPR models which in real-world applications are often black-box, (2) maintaining high image visual quality to ensure imperceptibility, and (3) achieving near-real-time performance to support practical, real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": "To bridge these challenges, we present a first look at privacy protection for place recognition and propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify the decisive regions for place recognition. Based on this, we propose Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO), which strategically optimizes perturbation generation in images by focusing on these decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way of generating the perturbation protects the most vulnerable areas of the image from unauthorized VPR models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In addition, previous research [63] reveals that deep learning models are highly sensitive to high-frequency tex" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "tures, while the human visual system primarily perceives low-frequency structural information. Inspired by this, we further elevate perturbation optimization to the frequency domain by isolating the high-frequency components from the perturbation and meticulously refining them using the SAP-PO algorithm. This way ensures that only the high-frequency elements of the perturbation are optimized. As a result, we craft a perturbation that is both precisely targeted and remarkably potent. Through further optimization in the frequency domain, the final output perturbation image (i.e., the protected image) remains imperceptible to the human eye while maintaining a robust defense against VPR models. Our work brings the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 228, + 553, + 431 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 553, + 287 + ], + "type": "text", + "content": "- We propose a fast and effective network that provides privacy protection for location information in images while ensuring imperceptibility to the human eye. To the best of our knowledge, this is the first attempt to address privacy protection for visual place recognition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 288, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 553, + 323 + ], + "type": "text", + "content": "- We introduce a SAP-PO algorithm, which strategically refines perturbations by focusing on decisive regions for place recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 323, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 553, + 382 + ], + "type": "text", + "content": "- Extensive experiments on multiple benchmark datasets and on various black-box VPR models verify that our method outperforms existing SOTA methods. Additionally, our method achieves a " + }, + { + "bbox": [ + 314, + 323, + 553, + 382 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 314, + 323, + 553, + 382 + ], + "type": "text", + "content": " speedup in runtime compared to SOTA methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 383, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 383, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 383, + 553, + 431 + ], + "type": "text", + "content": "- We validate the effectiveness of our privacy protection method based on using commercial APIs, including Google and Microsoft Bing, demonstrating the practical applicability in real-world scenarios." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 442, + 405, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 405, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 405, + 453 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 653 + ], + "type": "text", + "content": "Visual Place Recognition. With the rapid development of deep learning [4, 7-9, 23, 28, 40, 42, 55-59, 67-70], significant progress has been made in the VPR field [3, 15, 16, 20, 24]. Existing place recognition models have achieved impressive results not only in standard scenarios but also in challenging conditions such as radar scene [39], nighttime [31, 39], rainy [39], transport [21], perceptual aliasing [64], viewpoint changes [2, 41, 50], LiDAR scan [33, 60], occlusion [30, 41, 50], seasonal variation [41, 50], illumination change [32, 41, 50] dynamic objects. While these systems are designed with good intentions to enhance localization and navigation, they can be exploited by adversaries to expose users' location information, leading to potential risks of identity inference and tracking. Although previous work [26] attempts to attack image retrieval systems, it is not specifically designed for the VPR task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Perturbations Generation. While no prior research has specifically focused on privacy protection in visual place recognition, adversarial perturbations [44-46, 53] have been extensively studied as a general approach to visual privacy. Most existing adversarial methods impose norm" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7198" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "based constraints when generating perturbations. Specifically, many studies [11, 25, 27, 36, 65, 66] adopt infinity-norm or L2-norm constraints, producing dense perturbations that alter a substantial proportion of pixels. Although these methods are effective, they often introduce visible and unnatural distortions, particularly at higher perturbation levels. To address this, recent studies have explored sparse perturbation techniques [6, 10, 12, 17, 34, 38, 65, 72], which restrict modifications to a limited subset of pixels while maintaining high attack success rates. For instance, " + }, + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\mathrm{PGD}_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": " [6] extends the projected gradient descent framework by enforcing a sparsity constraint. SparseFool [34], building on DeepFool [35], generates sparse untargeted perturbations. ANDA [13] and MultiANDA [13] enhance transferability by approximating the posterior distribution of adversarial perturbations using the asymptotic normality of stochastic gradient ascent. However, these methods tend to be resource-intensive and cost a long inference time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 297, + 111, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 111, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 111, + 309 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 317, + 180, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 180, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 180, + 329 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 334, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 453 + ], + "type": "text", + "content": "We first introduce real-world location privacy risks. As shown in Fig. 1 (a), the rise of social media makes it easier for adversaries to scrape user-posted photos. Using VPR systems, an adversary can infer where an image was taken. This enables monitoring users' daily movements, uncovering sensitive information about their habits, residences, and frequented places. Moreover, when large-scale user daily movements are monitored, location data can directly or indirectly reveal their identity and social connections, intensifying privacy risks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "text", + "content": "To mitigate these risks, as shown in Fig. 1 (b), we propose a protection transformation " + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "text", + "content": " that converts a user's intended post photo " + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "text", + "content": " into a protected representation " + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "inline_equation", + "content": "I_{p} = F(I)" + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "text", + "content": ". Our goal is for " + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "inline_equation", + "content": "I_{p}" + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "text", + "content": " to remain visually consistent with the original image " + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 454, + 296, + 550 + ], + "type": "text", + "content": " while ensuring that adversaries cannot infer the photo's location using VPR systems. We refer to this process as location privacy protection, which can be formulated as an optimization problem:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 558, + 295, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 558, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 89, + 558, + 295, + 575 + ], + "type": "interline_equation", + "content": "\\min _ {F} \\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p r i v a c y}} \\left(I _ {p}\\right) + \\lambda \\mathcal {L} _ {\\text {v i s u a l}} \\left(I, I _ {p}\\right), \\tag {1}", + "image_path": "9e69cfedfab0aa4e1a62d38d1f1985385c38e79bcac0366c7a9995d871cb96b0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": "where the visual loss " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{visual}}" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": " ensures that the transformed image remains visually perceptual consistent to the original image, while the privacy loss " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{privacy}}" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": " mislead VPR models in place recognition. The hyperparameter " + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": " controls the trade-off between visual fidelity and location privacy. Formal definitions are provided in Sec. 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Challenges. Privacy protection in visual place recognition within real-world scenarios presents several unique challenges: Firstly, the adversary's VPR models are often black-box in real-world scenarios, meaning their architectures, parameters, and training data are unknown, making it" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "difficult to mislead the black-box VPR models. Secondly, the protection transformation must preserve high visual fidelity and imperceptibility so that the generated protection image " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "I_{p}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " remains visually consistency from the original image " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ". Thirdly, the protection method need be computationally efficient, enabling near-real-time processing to support practical use in real-world applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 156, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 228 + ], + "type": "text", + "content": "To the best of our knowledge, no existing approach effectively addresses all three of these challenges simultaneously for location privacy protection. Our goal is to develop a method that disrupts location inference while preserving image quality and maintaining practical efficiency, making it suitable for real-world deployment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 468, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 468, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 468, + 248 + ], + "type": "text", + "content": "3.2. Privacy Cloak Against VPR" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 253, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 361 + ], + "type": "text", + "content": "Previous works [44, 45, 51, 61, 62] have explored adding imperceptible perturbations to images to enhance image privacy while aiming to be visually imperceptible. However, these methods often exhibit low transferability, which limits their effectiveness in black-box settings. To handle this, ANDA [13] and MultiANDA [13] improve transferability by approximating the posterior distribution of perturbations. However, the high inference time of ANDA and MultiANDA limits the practicality in real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 540 + ], + "type": "text", + "content": "To address these limitations, we propose VPR-Cloak, a fast and efficient privacy protection network. Specifically, we first construct a saliency-aware prior to accurately identify decisive regions for place recognition. Based on the saliency-aware prior, we propose a SAP-PO algorithm, which strategically optimizes perturbation generation in images by focusing on decisive regions for place recognition while minimizing perturbations in less relevant areas. This targeted way protects the vulnerable regions of the image from unauthorized VPR models. Additionally, we further enhance perturbation optimization in the frequency domain. By isolating and refining high-frequency components using SAP-PO, we craft a precisely targeted yet imperceptible perturbation that remains effective against VPR models while staying unnoticeable to the human eye." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "text", + "content": "Saliency-Aware Prior Generation. Instead of applying perturbations randomly, we first identify critical regions for place recognition and selectively perturb them. To achieve this, we introduce a saliency-aware mechanism that acts as a spatial weight to pinpoint key areas for perturbation. For an input unprotected image " + }, + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "text", + "content": ", we extract this information by using the fine-tuned Vision Transformer (fine-tuned ViT), which employs the adapter strategy from [18, 31] to fine-tune the pre-trained DINOv2 backbone [37]. The input image is first resized using " + }, + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(I)" + }, + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "text", + "content": " to fit the fine-tuned ViT model's input size. Then, saliency-aware prior " + }, + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 540, + 554, + 682 + ], + "type": "text", + "content": ", is obtained by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 683, + 553, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 683, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 388, + 683, + 553, + 696 + ], + "type": "interline_equation", + "content": "P = \\mathcal {N} \\left(\\mathcal {A} _ {\\mathrm {V i T}} (\\mathcal {R} (I))\\right), \\tag {2}", + "image_path": "75d2e404341a3e9cb7943b6b7c83a65a3ca93cc1103a6151218c3c4e081d035f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\cdot)" + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": " represents normalization to ensure the prior" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7199" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 545, + 274 + ], + "type": "image", + "image_path": "d880ebd1312e8494f071f27831b1a28a09d3677d7dcf4744d63659daa6d2e254.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 324 + ], + "type": "text", + "content": "Figure 2. Overview of our VPR-Cloak. We introduce a saliency-aware prior to accurately identify decisive regions for place recognition and propose the SAP-PO algorithm, which strategically generates perturbations focused on these decisive regions while minimizing changes in less relevant areas. To enhance imperceptibility, we further elevate perturbation optimization to the frequency domain by isolating and meticulously optimizing only the high-frequencies using the SAP-PO algorithm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 343, + 282, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 282, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 282, + 356 + ], + "type": "text", + "content": "proper scaling and sum to 1. " + }, + { + "bbox": [ + 55, + 343, + 282, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathrm{ViT}}" + }, + { + "bbox": [ + 55, + 343, + 282, + 356 + ], + "type": "text", + "content": " is the fine-tuned ViT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 356, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 439 + ], + "type": "text", + "content": "Building on this saliency-aware prior " + }, + { + "bbox": [ + 55, + 356, + 296, + 439 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 356, + 296, + 439 + ], + "type": "text", + "content": ", we propose SAP-PO to optimize perturbation generation, ensuring perturbations are concentrated in decisive regions for place recognition. We provide two optimization strategies: spatial domain perturbation optimization for defense against VPR and frequency domain perturbation optimization for visual imperceptibility." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "text", + "content": "Spatial Perturbation for Defense against VPR. In the spatial domain, SAP-PO, guided by the saliency prior, focuses on perturbing key regions crucial for place recognition, maximizing its impact on reducing recognition accuracy. The spatially-perturbed image " + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{spa}}^{\\prime}" + }, + { + "bbox": [ + 55, + 441, + 296, + 502 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 510, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 510, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 137, + 510, + 295, + 526 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {s p a}} ^ {\\prime} = \\mathcal {F} _ {\\mathrm {s p a}} (I, P), \\tag {3}", + "image_path": "2fe5a5e1ac8b4e95e33ce781b315cb01e19e06bbf9fdbed5188e3f0af7daf7cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 277, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 277, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 277, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 534, + 277, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{spa}}" + }, + { + "bbox": [ + 55, + 534, + 277, + 547 + ], + "type": "text", + "content": " is our proposed SAP-PO in spatial domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 547, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 631 + ], + "type": "text", + "content": "Frequency Perturbation for Imperceptible. Previous study [63] have shown that the human eye is sensitive to low frequencies, while deep learning models are sensitive to high frequencies. Inspired by this, we elevate perturbation optimization to the frequency domain by isolating and meticulously refining only the high-frequency components of the perturbation using the SAP-PO algorithm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "text", + "content": "To do this, we apply the Discrete Cosine Transform (DCT) to the perturbed image " + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{spa}}^{\\prime}" + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "text", + "content": ", yielding its frequency representation " + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "inline_equation", + "content": "f^{\\prime}" + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "text", + "content": ". We specifically optimize the high-frequency components of the perturbation using SAP-PO and generate the frequency-perturbed image " + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{fre}}^{\\prime}" + }, + { + "bbox": [ + 55, + 631, + 296, + 693 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 146, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "f ^ {\\prime} = \\mathcal {D} \\left(I _ {\\mathrm {s p a}} ^ {\\prime}\\right), \\tag {4}", + "image_path": "914eb7fce76556b9b07133fcbb57697a4553fdcab5634fe613afd04130cc3a76.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 396, + 355, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 355, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 396, + 355, + 553, + 369 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {f r e}} ^ {\\prime} = \\mathcal {F} _ {\\mathrm {f r e}} \\left(f ^ {\\prime}, P\\right), \\tag {5}", + "image_path": "499fb46c45ceae4b1c16e705d9367f299b88da1b5781cbdd7ea48a28fe882fc6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "text", + "content": " denotes DCT. " + }, + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fre}}" + }, + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "text", + "content": " is our proposed SAP-PO in frequency domain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 395, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 554, + 431 + ], + "type": "text", + "content": "Next, we apply the inverse DCT to transform the frequency-perturbed image " + }, + { + "bbox": [ + 313, + 395, + 554, + 431 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{fre}}^{\\prime}" + }, + { + "bbox": [ + 313, + 395, + 554, + 431 + ], + "type": "text", + "content": " back to the spatial domain, obtaining the final protected image " + }, + { + "bbox": [ + 313, + 395, + 554, + 431 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{p}}" + }, + { + "bbox": [ + 313, + 395, + 554, + 431 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 400, + 434, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 434, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 400, + 434, + 553, + 449 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {p}} = \\mathcal {D} ^ {- 1} \\left(I _ {\\mathrm {f r e}} ^ {\\prime}\\right), \\tag {6}", + "image_path": "e60351316f4d1918cf3cb76255a77281ed864c663949ba5b0acf36d11b5a1894.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 443, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 443, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 443, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 453, + 443, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{-1}" + }, + { + "bbox": [ + 313, + 453, + 443, + 464 + ], + "type": "text", + "content": " is the inverse DCT." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 464, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 514 + ], + "type": "text", + "content": "Through optimization in the spatial and frequency domains, we strategically refine and enforce the distribution of perturbations to concentrate on decisive regions for place recognition and high-frequency areas." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 518, + 554, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 541 + ], + "type": "text", + "content": "3.3. Saliency-Aware Prior Guided Perturbation Optimization" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 548, + 554, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 595 + ], + "type": "text", + "content": "We propose SAP-PO to generate effective and visually imperceptible perturbations for defending VPR models. By integrating the saliency-aware prior " + }, + { + "bbox": [ + 313, + 548, + 554, + 595 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 548, + 554, + 595 + ], + "type": "text", + "content": ", we focus perturbations on the decisive regions for place recognition." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "text", + "content": "To begin, we initialize the perturbation as: " + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "inline_equation", + "content": "\\delta = 0" + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "text", + "content": ". For an input image " + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "text", + "content": ", we retrieve multiple positive reference images " + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "inline_equation", + "content": "\\{I_{\\mathrm{pos},i}\\}" + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "text", + "content": " from the database and extract their features " + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "inline_equation", + "content": "\\{F_{\\mathrm{pos},i}\\}" + }, + { + "bbox": [ + 313, + 596, + 555, + 644 + ], + "type": "text", + "content": " using the Surrogate VPR:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 377, + 647, + 553, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 647, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 377, + 647, + 553, + 662 + ], + "type": "interline_equation", + "content": "\\left\\{F _ {\\text {p o s}, i} \\right\\} = \\mathcal {M} _ {\\mathrm {V P R}} \\left(\\left\\{I _ {\\text {p o s}, i} \\right\\}\\right), \\tag {7}", + "image_path": "bbde1581217f119564f5a95edbb40ed899054254923fc10bae11ff088ca5187d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{VPR}}" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": " denotes the Surrogate VPR model. To enhance the robustness of our perturbations against unauthorized VPR models, for each input image, we randomly select one from multiple VPR models " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}_1,\\mathcal{M}_2,\\dots ,\\mathcal{M}_m\\}" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7200" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 99, + 296, + 135 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 96 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 96 + ], + "type": "text", + "content": "Algorithm 1 Saliency-Aware Prior Guided Perturbation Optimization (SAP-PO)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "text", + "content": "Require: Image " + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "text", + "content": ", Saliency-aware prior " + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "text", + "content": ", Surrogate VPR Model " + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{VPR}}" + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "text", + "content": ", Trade-off Coefficient " + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "text", + "content": ", Step Size " + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "text", + "content": ", Maximum Iterations " + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "text", + "content": ", Domain Type " + }, + { + "bbox": [ + 56, + 99, + 296, + 135 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{type}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + }, + { + "bbox": [ + 57, + 136, + 173, + 147 + ], + "lines": [ + { + "bbox": [ + 57, + 136, + 173, + 147 + ], + "spans": [ + { + "bbox": [ + 57, + 136, + 173, + 147 + ], + "type": "text", + "content": "Ensure: Perturbed Image " + }, + { + "bbox": [ + 57, + 136, + 173, + 147 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + } + ], + "index": 1, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 59, + 148, + 294, + 445 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 63, + 148, + 190, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 148, + 190, + 159 + ], + "spans": [ + { + "bbox": [ + 63, + 148, + 190, + 159 + ], + "type": "text", + "content": "1: Initialize perturbation " + }, + { + "bbox": [ + 63, + 148, + 190, + 159 + ], + "type": "inline_equation", + "content": "\\delta \\gets 0" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 160, + 294, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 160, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 63, + 160, + 294, + 183 + ], + "type": "text", + "content": "2: Retrieve multiple positive samples from database: " + }, + { + "bbox": [ + 63, + 160, + 294, + 183 + ], + "type": "inline_equation", + "content": "\\{I_{\\mathrm{pos},s}\\} \\gets" + }, + { + "bbox": [ + 63, + 160, + 294, + 183 + ], + "type": "text", + "content": " retrieve_positives(I)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 184, + 294, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 184, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 63, + 184, + 294, + 208 + ], + "type": "text", + "content": "3: Generate " + }, + { + "bbox": [ + 63, + 184, + 294, + 208 + ], + "type": "inline_equation", + "content": "\\{F_{\\mathrm{pos},i}\\}" + }, + { + "bbox": [ + 63, + 184, + 294, + 208 + ], + "type": "text", + "content": " using the Surrogate VPR model: " + }, + { + "bbox": [ + 63, + 184, + 294, + 208 + ], + "type": "inline_equation", + "content": "\\{F_{\\mathrm{pos},i}\\} \\gets \\mathcal{M}_{\\mathrm{VPR}}(\\{I_{\\mathrm{pos},i}\\})" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 209, + 148, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 209, + 148, + 217 + ], + "spans": [ + { + "bbox": [ + 63, + 209, + 148, + 217 + ], + "type": "text", + "content": "4: for " + }, + { + "bbox": [ + 63, + 209, + 148, + 217 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 63, + 209, + 148, + 217 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 63, + 209, + 148, + 217 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 63, + 209, + 148, + 217 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 219, + 246, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 219, + 246, + 230 + ], + "spans": [ + { + "bbox": [ + 63, + 219, + 246, + 230 + ], + "type": "text", + "content": "5: Compute perturbed image: " + }, + { + "bbox": [ + 63, + 219, + 246, + 230 + ], + "type": "inline_equation", + "content": "I' \\gets I + \\delta" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 232, + 194, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 232, + 194, + 243 + ], + "spans": [ + { + "bbox": [ + 63, + 232, + 194, + 243 + ], + "type": "text", + "content": "6: if " + }, + { + "bbox": [ + 63, + 232, + 194, + 243 + ], + "type": "inline_equation", + "content": "D_{\\text{type}} = \\text{Frequency}" + }, + { + "bbox": [ + 63, + 232, + 194, + 243 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 243, + 227, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 243, + 227, + 255 + ], + "spans": [ + { + "bbox": [ + 63, + 243, + 227, + 255 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 63, + 243, + 227, + 255 + ], + "type": "inline_equation", + "content": "I^{\\prime}\\gets I + \\mathrm{RemoveLowFreq}(\\delta)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 257, + 133, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 257, + 133, + 265 + ], + "spans": [ + { + "bbox": [ + 63, + 257, + 133, + 265 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 63, + 257, + 133, + 265 + ], + "type": "inline_equation", + "content": "P\\gets 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 267, + 294, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 267, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 267, + 294, + 290 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 62, + 267, + 294, + 290 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 62, + 267, + 294, + 290 + ], + "type": "text", + "content": " For clarity, this algorithm omits transformations between spatial and frequency domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 292, + 116, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 116, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 116, + 301 + ], + "type": "text", + "content": "10: end if" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 303, + 257, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 257, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 257, + 315 + ], + "type": "text", + "content": "11: Generate feature vector: " + }, + { + "bbox": [ + 62, + 303, + 257, + 315 + ], + "type": "inline_equation", + "content": "F' \\gets \\mathcal{M}_{\\mathrm{VPR}}(I')" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 316, + 294, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 316, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 59, + 316, + 294, + 339 + ], + "type": "text", + "content": "12: Compute average distance to positive samples: " + }, + { + "bbox": [ + 59, + 316, + 294, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pos}}\\gets \\frac{1}{N}\\sum_{i = 1}^{N}\\mathrm{Distance}(F',F_{\\mathrm{pos},i})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 339, + 263, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 339, + 263, + 351 + ], + "spans": [ + { + "bbox": [ + 59, + 339, + 263, + 351 + ], + "type": "text", + "content": "13: Compute PSNR loss: " + }, + { + "bbox": [ + 59, + 339, + 263, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{psnr}}\\gets \\mathrm{PSNR}(I,I^{\\prime})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 352, + 255, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 352, + 255, + 363 + ], + "spans": [ + { + "bbox": [ + 59, + 352, + 255, + 363 + ], + "type": "text", + "content": "14: Compute total loss: " + }, + { + "bbox": [ + 59, + 352, + 255, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\gets -\\mathcal{L}_{\\mathrm{pos}} - \\lambda \\mathcal{L}_{\\mathrm{psnr}}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 363, + 292, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 363, + 292, + 374 + ], + "spans": [ + { + "bbox": [ + 59, + 363, + 292, + 374 + ], + "type": "text", + "content": "15: Compute gradient: " + }, + { + "bbox": [ + 59, + 363, + 292, + 374 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}\\gets" + }, + { + "bbox": [ + 59, + 363, + 292, + 374 + ], + "type": "text", + "content": " compute-gradient " + }, + { + "bbox": [ + 59, + 363, + 292, + 374 + ], + "type": "inline_equation", + "content": "(\\mathcal{L},\\delta)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 375, + 235, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 375, + 235, + 387 + ], + "spans": [ + { + "bbox": [ + 59, + 375, + 235, + 387 + ], + "type": "text", + "content": "16: Update perturbation: " + }, + { + "bbox": [ + 59, + 375, + 235, + 387 + ], + "type": "inline_equation", + "content": "\\delta \\gets \\delta +\\eta \\nabla \\mathcal{L}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 387, + 294, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 387, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 59, + 387, + 294, + 410 + ], + "type": "text", + "content": "17: Apply Saliency-aware prior " + }, + { + "bbox": [ + 59, + 387, + 294, + 410 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 59, + 387, + 294, + 410 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 59, + 387, + 294, + 410 + ], + "type": "inline_equation", + "content": "\\delta \\gets \\mathrm{clip}(\\delta, -P \\cdot \\epsilon, P \\cdot \\epsilon)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 411, + 106, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 411, + 106, + 420 + ], + "spans": [ + { + "bbox": [ + 59, + 411, + 106, + 420 + ], + "type": "text", + "content": "18: end for" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 59, + 422, + 218, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 422, + 218, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 422, + 218, + 434 + ], + "type": "text", + "content": "19: Return the final perturbed image: " + }, + { + "bbox": [ + 59, + 422, + 218, + 434 + ], + "type": "inline_equation", + "content": "I'" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 59, + 434, + 113, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 434, + 113, + 445 + ], + "spans": [ + { + "bbox": [ + 59, + 434, + 113, + 445 + ], + "type": "text", + "content": "20: return " + }, + { + "bbox": [ + 59, + 434, + 113, + 445 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "content": "as the Surrogate VPR model " + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathrm{VPR}}" + }, + { + "bbox": [ + 55, + 469, + 296, + 540 + ], + "type": "text", + "content": ". This strategy ensures that the perturbations are not overfitted to a specific model but instead generalize well to a variety of potential recognition systems. In real-world applications, the VPR model set can be dynamically updated, enabling adaptation to stronger protection against unauthorized attacks." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "content": "At each iteration, we obtain the perturbed image " + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "inline_equation", + "content": "I^{\\prime} = I + \\delta" + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "content": ". If the frequency domain is used (" + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{type}} = \\mathrm{Frequency}" + }, + { + "bbox": [ + 55, + 541, + 296, + 590 + ], + "type": "text", + "content": "), the perturbation is processed by applying a high-pass filter to remove low frequencies." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 590, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 638 + ], + "type": "text", + "content": "Then, we use the Surrogate VPR: " + }, + { + "bbox": [ + 55, + 590, + 296, + 638 + ], + "type": "inline_equation", + "content": "F' = \\mathcal{M}_{\\mathrm{VPR}}(I')" + }, + { + "bbox": [ + 55, + 590, + 296, + 638 + ], + "type": "text", + "content": " to generate feature " + }, + { + "bbox": [ + 55, + 590, + 296, + 638 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 55, + 590, + 296, + 638 + ], + "type": "text", + "content": ". The feature distance between the perturbed image and the retrieved positive reference images is then computed as the average distance:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 94, + 647, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 647, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 94, + 647, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p o s}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| \\frac {F ^ {\\prime}}{\\| F ^ {\\prime} \\| _ {2}} - \\frac {F _ {\\mathrm {p o s} , i}}{\\| F _ {\\mathrm {p o s} , i} \\| _ {2}} \\right\\| _ {2}, \\tag {8}", + "image_path": "40ff95320c94970ef82e1d271412ce8e8f2708892025139322523ca2d24c4439.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the number of retrieved positive samples, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\| X\\| _2" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " denotes the L2 norm." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "In addition, to ensure that the generated perturbation images maintain high visual fidelity consistent with the original images, we also impose the constraint of preserving pixel-wise similarity by employing the PSNR loss." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 121, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 144 + ], + "type": "text", + "content": "The total loss balances the trade-off between maximizing the feature distance and preserving image quality:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 381, + 152, + 553, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 152, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 381, + 152, + 553, + 166 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = - \\mathcal {L} _ {\\mathrm {p o s}} - \\lambda \\mathcal {L} _ {\\mathrm {p s n r}}, \\tag {9}", + "image_path": "47418bf2ce81ec846877d9a61c7e79521e5cb1d72dfb0ee78c44180f5dcb2f4d.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "text", + "content": " is the trade-off coefficient. Notably, " + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "inline_equation", + "content": "-\\mathcal{L}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{privacy}}" + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "inline_equation", + "content": "-\\mathcal{L}_{\\mathrm{psnr}}" + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{visual}}" + }, + { + "bbox": [ + 313, + 171, + 553, + 207 + ], + "type": "text", + "content": " in Eq. (1)." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 208, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 243 + ], + "type": "text", + "content": "To ensure the perturbations remain concentrated in decisive regions, we optimize their distribution by using the saliency-aware prior:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 385, + 251, + 553, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 251, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 385, + 251, + 553, + 264 + ], + "type": "interline_equation", + "content": "\\delta = \\mathcal {C} (\\delta , - P \\cdot \\epsilon , P \\cdot \\epsilon), \\tag {10}", + "image_path": "28460819c2cf42b0c38cf32b9967829b6201ab702d7b792bd6f7a473927019ea.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 271, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 271, + 553, + 295 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 271, + 553, + 295 + ], + "type": "text", + "content": " is the step size. And " + }, + { + "bbox": [ + 313, + 271, + 553, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 271, + 553, + 295 + ], + "type": "text", + "content": " is clip to constrain the perturbation within saliency-aware bounds." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "text", + "content": "After completing " + }, + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "text", + "content": " iterations, we obtain the final perturbed (protected) image " + }, + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "inline_equation", + "content": "I_{p} = I^{\\prime}" + }, + { + "bbox": [ + 313, + 296, + 553, + 344 + ], + "type": "text", + "content": " that is optimized to disrupt place recognition while maintaining visual consistency with the original image." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 353, + 395, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 395, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 395, + 366 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 373, + 437, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 437, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 437, + 385 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": "We conduct the experiments on several VPR benchmark datasets: Pitts30k [1], MSLS [54], Tokyo247 [49], and Nordland [47]. We conduct experiments using PyTorch on a setup with 8 NVIDIA GeForce RTX 3090 GPUs. For our method, the trade-off coefficient " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": " is set to 0.3 in both spatial and frequency domains. Specifically, in the frequency domain, we utilize an " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": " DCT block size (yielding 64 blocks), with a frequency threshold of 12. For both ANDA and MultiANDA, we set the spatial transformation range to 0.3 (allowing translations by " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\pm 30\\%" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": " of the image size), gradient averaging to 25 samples, and multi-scale transformations to 5 copies. MultiANDA introduces a diversity probability of 0.5. The Gaussian kernel sizes differ between the two methods: " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "7(\\sigma = 3)" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": " for ANDA and " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "15(\\sigma = 3)" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": " for MultiANDA. All these methods evaluate over 10 iterations with a batch size of 32 and a learning rate of " + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "inline_equation", + "content": "8 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 390, + 555, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 588, + 425, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 425, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 425, + 601 + ], + "type": "text", + "content": "4.2. Comparison Study" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Black-Box and White-Box Performance. We select five popular open-source VPR models: Crica [30], DHE [29], Salad [21], Sela [31], and VLAD-BuFF [24]. We randomly select one model as the surrogate to generate protected images; evaluations on this surrogate constitute the white-box setting, while evaluations on the remaining four unseen models constitute the black-box setting. Notably, in black-box setting, a single surrogate VPR model is selected from the VPR pool as the fixed model for training" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7201" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 94, + 515, + 302 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Protection success rate (%) comparisons (Epsilon=12/255). ΔR@1, ΔR@5, and ΔR@10 measure the reduction in the model's retrieval accuracy after applying protection, with larger values indicating stronger protection." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 94, + 515, + 302 + ], + "lines": [ + { + "bbox": [ + 97, + 94, + 515, + 302 + ], + "spans": [ + { + "bbox": [ + 97, + 94, + 515, + 302 + ], + "type": "table", + "html": "
DatasetModelANDA_CricaMULTI-ANDA_CricaOurs_CricaOurs_Sela
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KVLAD-BuFF18.59.3-20.215.1-45.437.6-36.227.7-
DHE23.116.09.831.321.913.339.625.916.241.625.919.2
Sela17.713.66.319.215.88.840.727.020.073.254.940.1
Crica50.835.421.952.638.225.069.561.857.651.044.327.9
Salad22.012.38.125.815.410.935.723.715.943.028.518.5
TOKYO247VLAD-BuFF19.611.7-20.512.2-36.328.8-28.423.9-
DHE23.229.221.434.541.329.745.850.750.738.046.644.5
Sela34.225.816.236.827.519.744.427.226.363.664.046.2
Crica49.835.826.752.038.229.382.282.380.052.434.126.9
MSLSVLAD-BuFF29.318.4-30.620.6-43.534.6-37.827.8-
DHE37.826.719.944.933.325.852.940.033.349.336.232.2
Sela24.414.17.328.119.311.041.234.327.472.865.346.7
Crica53.743.433.057.947.237.381.583.473.840.730.326.2
Salad36.434.423.442.937.330.052.140.424.149.937.327.9
MSLSDHE60.476.177.961.377.581.661.678.082.061.577.381.2
Salad72.885.284.474.087.688.674.688.190.274.787.990.2
CHALLENGESela72.683.980.673.084.483.073.487.390.068.376.275.6
Crica66.774.364.567.375.866.669.079.979.473.387.189.7
NORDLANDVLAD-BuFF58.962.9-60.766.0-68.970.7-66.871.3-
DHE43.653.649.353.365.465.764.478.677.864.877.374.2
Sela52.357.446.254.561.951.464.172.972.486.983.364.0
Crica63.161.351.066.062.352.387.984.879.262.169.966.9
Salad56.960.250.262.867.856.769.378.074.468.879.172.2
", + "image_path": "40dddeca85d0b93715640dcc81df3c2cb74f7c189b27f577c925d5d24ff565fb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 106, + 316, + 503, + 511 + ], + "blocks": [ + { + "bbox": [ + 106, + 316, + 503, + 511 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 503, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 503, + 511 + ], + "type": "image", + "image_path": "71139ffc9d806d385a26ef46ce9948af4eae8bbd6448c2b58feb6427ba150a48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 515, + 555, + 539 + ], + "lines": [ + { + "bbox": [ + 55, + 515, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 555, + 539 + ], + "type": "text", + "content": "Figure 3. Qualitative results. Our method robustly prevents location retrieval while preserving image quality. Green frames indicate successful protection, while orange frames indicate the opposite." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "content": "the protected image set, while the remaining VPR models in the pool are strictly not used in the training process and instead serve as unseen models for testing. Specifically, " + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{ANDA}_{\\mathrm{Crica}}" + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "content": ", MULTI- " + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{ANDA}_{\\mathrm{Crica}}" + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_{\\mathrm{Crica}}" + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "content": " use Crica as the surrogate model, whereas " + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_{\\mathrm{Sela}}" + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "content": " uses Sela. In Tab. 1, our method consistently achieves higher protection success rates than ANDA [13] and MultiANDA [13] across all datasets. " + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_{\\mathrm{Crica}}" + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{Ours}_{\\mathrm{Sela}}" + }, + { + "bbox": [ + 54, + 559, + 297, + 704 + ], + "type": "text", + "content": " significantly reduce retrieval accuracy, highlighting their strong privacy protection capability. Notably, on PITTS30K and MSLS, our approach markedly outperforms baselines. These results demonstrate the strong transferability and effectiveness of our method in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 559, + 518, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 518, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 518, + 571 + ], + "type": "text", + "content": "disrupting both black-box and white-box retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 575, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 555, + 685 + ], + "type": "text", + "content": "Image Quality Evaluations. As shown in Tab. 2, our method consistently outperforms ANDA and MultiANDA across most datasets, achieving superior visual fidelity while maintaining strong protection. In PITTS30K, OursSela achieves the best visual fidelity results, while OursCrica attains the lowest Avg Epsilon, ensuring higher imperceptibility. These results demonstrate that our approach not only provides stronger privacy protection but also ensures better image quality and imperceptibility." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Runtime Comparisons. The results on the Tab. 3 clearly demonstrate that our method surpasses ANDA by " + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7202" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 518, + 259 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 518, + 259 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 518, + 259 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 518, + 259 + ], + "type": "image", + "image_path": "ef906974626817ca52ed59b75f642378b5c16fa1c687641af6841ead0dcfcc58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Figure 4. Protection on Commercial APIs. Our method successfully prevents Google and Microsoft Bing from retrieving location information while ensuring imperceptible modifications, highlighting our method's practical superiority." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 341, + 280, + 559 + ], + "blocks": [ + { + "bbox": [ + 55, + 304, + 295, + 337 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 337 + ], + "type": "text", + "content": "Table 2. Image quality comparison (Epsilon=12/255). Higher PSNR/SSIM and lower LPIPS indicate better visual fidelity, while lower Avg Epsilon suggests more imperceptible perturbations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 341, + 280, + 559 + ], + "lines": [ + { + "bbox": [ + 69, + 341, + 280, + 559 + ], + "spans": [ + { + "bbox": [ + 69, + 341, + 280, + 559 + ], + "type": "table", + "html": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KANDACrica0.14429.840.8724.78
MULTI-ANDACrica0.13230.170.8814.44
OursCrica0.09131.560.9044.02
OursSela0.08232.000.9174.14
TOKYO247ANDACrica0.08930.000.8815.34
MULTI-ANDACrica0.08330.200.8975.14
OursCrica0.09929.110.8865.32
OursSela0.11328.560.8525.42
MSLSANDACrica0.13930.660.8724.22
MULTI-ANDACrica0.11130.140.8894.00
OursCrica0.10230.990.9014.38
OursSela0.11930.260.8894.02
MSLS CHALLENGEANDA (Crica)0.12633.010.9094.68
MULTI-ANDACrica0.12133.290.9174.38
OursCrica0.07334.200.9364.68
OursSela0.09233.210.9143.42
NORDLANDANDA (Crica)0.10530.420.8763.04
MULTI-ANDACrica0.09430.990.8962.78
OursCrica0.09232.030.9172.76
OursSela0.10631.220.8963.38
", + "image_path": "1d00431e074a16cecf6ed7f88fcc80f37c5d29920569d7d3c5b827aa79dcbeca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 79, + 597, + 268, + 654 + ], + "blocks": [ + { + "bbox": [ + 55, + 572, + 295, + 594 + ], + "lines": [ + { + "bbox": [ + 55, + 572, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 295, + 594 + ], + "type": "text", + "content": "Table 3. Runtime and average protection success rate on NORDLAND. Our method achieves a " + }, + { + "bbox": [ + 55, + 572, + 295, + 594 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 55, + 572, + 295, + 594 + ], + "type": "text", + "content": " speedup over MULTI-ANDA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 597, + 268, + 654 + ], + "lines": [ + { + "bbox": [ + 79, + 597, + 268, + 654 + ], + "spans": [ + { + "bbox": [ + 79, + 597, + 268, + 654 + ], + "type": "table", + "html": "
MethodANDAMULTI-ANDAOurs
Avg ΔR@155.059.570.9
Avg ΔR@559.164.777.0
Avg ΔR@1049.256.576.0
Runtime (s)0.1690.5510.036
", + "image_path": "c242f2b3a746eb38be2edac1e9ae7064913e607c02db02d0af8cd7c64aacab7d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "17.9\\%" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "26.8\\%" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", and outperforms MultiANDA by " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "12.3\\%" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "19.5\\%" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\Delta R@1" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\Delta R@5" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\Delta R@10" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": ", respectively. Additionally, our method requires" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 305, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 365 + ], + "type": "text", + "content": "only 0.036 seconds per run, approximately 4.7 times faster than ANDA and over " + }, + { + "bbox": [ + 313, + 305, + 555, + 365 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 313, + 305, + 555, + 365 + ], + "type": "text", + "content": " faster than MultiANDA. These results confirm that our approach significantly enhances protection success while maintaining remarkable efficiency, making it suitable for real-time applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 460 + ], + "type": "text", + "content": "Visual Comparisons. Fig. 3 demonstrates that our method effectively prevents location retrieval, surpassing existing SOTA approaches under challenging scenarios, including drastic viewpoint changes, seasonal variations, day-night transitions, and occlusions. Our approach consistently obscures identifiable features while preserving visual integrity, ensuring robust privacy protection suitable for diverse real-world conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 549, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 549, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 549, + 480 + ], + "type": "text", + "content": "4.3. Protection Performance on Commercial APIs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 628 + ], + "type": "text", + "content": "Fig. 4 illustrates the effectiveness of our method, evaluated on world-famous commercial APIs including Google and Microsoft Bing. Originally, these systems could achieve correct retrieval of identifying and matching locations from visual cues. However, after applying our method, these images become effectively shielded from such retrieval processes. Notably, while our method significantly disrupts the feature extraction necessary for location recognition, it preserves the visual quality of the images. This balance between robust privacy protection and maintaining visual integrity highlights the practical applicability and superiority of our approach in real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 689 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, we conduct an ablation study on the impact of spatial and frequency-domain perturbations and the saliency-aware prior." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Effectiveness of Spatial and Frequency Domain Perturbation Optimization. Tab. 4 shows that spatial domain at" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7203" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 83, + 472, + 206 + ], + "blocks": [ + { + "bbox": [ + 212, + 71, + 396, + 81 + ], + "lines": [ + { + "bbox": [ + 212, + 71, + 396, + 81 + ], + "spans": [ + { + "bbox": [ + 212, + 71, + 396, + 81 + ], + "type": "text", + "content": "Table 4. Ablation study on protection success rate." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 137, + 83, + 472, + 206 + ], + "lines": [ + { + "bbox": [ + 137, + 83, + 472, + 206 + ], + "spans": [ + { + "bbox": [ + 137, + 83, + 472, + 206 + ], + "type": "table", + "html": "
DatasetModelSpatialFrequencySpatial + Frequency
ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10ΔR@1ΔR@5ΔR@10
PITTS30KCrica51.642.331.927.522.717.657.147.737.5
DHE31.218.111.221.512.87.635.520.314.4
Salad31.717.210.820.210.35.733.419.312.2
Sela24.613.110.815.912.57.530.319.616.2
MSLSCrica71.358.337.950.244.230.778.165.843.7
DHE43.130.325.127.923.016.546.432.927.2
Salad44.332.921.626.919.310.849.437.025.3
Sela31.925.220.321.413.34.638.828.724.6
NORDLANDCrica79.371.249.971.056.647.884.675.659.0
DHE56.168.265.437.844.135.063.576.476.1
Salad65.269.370.134.842.438.569.378.074.4
Sela58.365.162.331.427.927.361.868.368.4
", + "image_path": "fbe27c3b0c8a45f8e415d5f22bdcbec8aa26b0f81fce95f52eb714a497bdf4b4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 237, + 279, + 347 + ], + "blocks": [ + { + "bbox": [ + 99, + 224, + 251, + 236 + ], + "lines": [ + { + "bbox": [ + 99, + 224, + 251, + 236 + ], + "spans": [ + { + "bbox": [ + 99, + 224, + 251, + 236 + ], + "type": "text", + "content": "Table 5. Ablation study on image quality." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 237, + 279, + 347 + ], + "lines": [ + { + "bbox": [ + 70, + 237, + 279, + 347 + ], + "spans": [ + { + "bbox": [ + 70, + 237, + 279, + 347 + ], + "type": "table", + "html": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KSpatial0.05732.430.9292.85
Frequency0.04233.830.9462.42
Spatial + Frequency0.05732.420.9282.84
MSLSSpatial0.06732.280.9152.80
Frequency0.05933.190.9372.57
Spatial + Frequency0.06632.210.9162.82
NORDLANDSpatial0.07332.410.9171.83
Frequency0.06032.990.9312.01
Spatial + Frequency0.07132.130.9151.84
", + "image_path": "2ceb08a998134ce6933bc572ef6b115b4d325e51a2b663eedfd66ded46d668af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 85, + 370, + 265, + 454 + ], + "blocks": [ + { + "bbox": [ + 87, + 358, + 263, + 369 + ], + "lines": [ + { + "bbox": [ + 87, + 358, + 263, + 369 + ], + "spans": [ + { + "bbox": [ + 87, + 358, + 263, + 369 + ], + "type": "text", + "content": "Table 6. Ablation study on saliency-aware prior." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 370, + 265, + 454 + ], + "lines": [ + { + "bbox": [ + 85, + 370, + 265, + 454 + ], + "spans": [ + { + "bbox": [ + 85, + 370, + 265, + 454 + ], + "type": "table", + "html": "
DatasetMethodLPIPSPSNRSSIMAvg Epsilon
PITTS30KUnguided0.18329.040.8175.69
Guided0.05732.420.9282.84
MSLSUnguided0.19227.830.8015.91
Guided0.06632.210.9162.82
NORDLANDUnguided0.18828.410.8335.88
Guided0.07132.130.9151.84
", + "image_path": "6decffa8c3530ebcbab00c9b698418a021d9574eaf25ff46a89abf305d3f01d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": "tacks alone yield stronger protection than frequency domain attacks. This indicates that spatial modifications, by altering key structural features, are inherently more effective. Yet, when spatial and frequency domain attacks are combined, the success rates improve further, " + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": ". This synergy demonstrates that frequency domain perturbations, while less potent on their own, complement spatial attacks by enhancing overall robustness without significant loss of performance. Tab. 5 confirms that frequency perturbations preserve visual quality exceptionally well. Although the combined method results in a slight degradation, this minor compromise in visual fidelity is well justified by the substantial gain in protection. In essence, integrating frequency domain modifications with spatial attacks strikes an effective balance between robustness and image quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Effectiveness of Saliency-Aware Prior. Tab. 6 clearly demonstrates the advantages of incorporating a saliency-aware prior. Compared to the unguided way, the guided method markedly improves image fidelity. For example, on the PITTS30K dataset, the guided approach reduces LPIPS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 226, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 323 + ], + "type": "text", + "content": "from 0.183 to 0.057, increases PSNR from 29.04 to 32.42, and enhances SSIM from 0.817 to 0.928. Additionally, the average perturbation (Avg Epsilon) is lowered from 5.69 to 2.84. Similar trends are observed in the MSLS and NORDLAND datasets. These results indicate that the saliency-aware prior effectively directs perturbations to less noticeable regions, thereby maintaining high visual quality while keeping modifications nearly imperceptible." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 348, + 388, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 388, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 388, + 361 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 567 + ], + "type": "text", + "content": "In this paper, we present the first look at privacy protection in VPR and introduce VPR-Cloak, an efficient privacy-preserving network. We introduce a saliency-aware prior to identify key regions crucial for place recognition and propose SAP-PO to selectively enhance perturbation generation in these areas. To further improve imperceptibility, we optimize perturbations in the frequency domain, refining high-frequency components while preserving low-frequency structures essential for human perception. Extensive experiments on multiple benchmark datasets and across various black-box VPR models demonstrate that our method surpasses SOTA perturbation methods in effectiveness. Moreover, our approach achieves a " + }, + { + "bbox": [ + 313, + 374, + 555, + 567 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 313, + 374, + 555, + 567 + ], + "type": "text", + "content": " speedup in runtime compared to SOTA methods. We also validate its practicality using commercial APIs, including Google and Microsoft Bing, showcasing its real-world applicability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 592, + 411, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 411, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 411, + 605 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008). This work was supported in part by the National Natural Science Foundation of China under Grant 62202214 and Guangdong Basic and Applied Basic Research Foundation under Grant 2023A1515012819." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7204" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Gabriele Berton, Riccardo Mereu, Gabriele Trivigno, Carlo Masone, Gabriela Csurka, Torsten Sattler, and Barbara Caputo. Deep visual geo-localization benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5396-5407, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. Eigenplaces: Training viewpoint robust models for visual place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11080-11090, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "type": "text", + "content": "[3] Gabriele Berton, Lorenz Junglas, Riccardo Zaccone, Thomas Pollok, Barbara Caputo, and Carlo Masone. Meshvpr: Citywide visual place recognition using 3d meshes. In European Conference on Computer Vision, pages 321-339. Springer, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 322 + ], + "type": "text", + "content": "[4] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2311-2319, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 296, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 296, + 379 + ], + "type": "text", + "content": "[5] Zhaoyu Chen, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, and Wenqiang Zhang. Towards practical certifiable patch defense with vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[6] Francesco Croce and Matthias Hein. Sparse and imperceivable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4724-4732, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 425, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 467 + ], + "type": "text", + "content": "[7] Shuting Dong, Feng Lu, Zhe Wu, and Chun Yuan. Dfvsr: Directional frequency video super-resolution via asymmetric and enhancement alignment network. In *IJCAI*, pages 681–689, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "text", + "content": "[8] Shuting Dong, Feng Lu, and Chun Yuan. Frequency reciprocal action and fusion for single image super-resolution. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 578 + ], + "type": "text", + "content": "[9] Shuting Dong, Zhe Wu, Feng Lu, and Chun Yuan. Enhanced image deblurring: An efficient frequency exploitation and preservation network. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7184-7193, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 580, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 634 + ], + "type": "text", + "content": "[10] Xiaoyi Dong, Dongdong Chen, Jianmin Bao, Chuan Qin, Lu Yuan, Weiming Zhang, Nenghai Yu, and Dong Chen. Greedyfool: Distortion-aware sparse adversarial attack. Advances in Neural Information Processing Systems, 33: 11226-11236, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[11] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[12] Yanbo Fan, Baoyuan Wu, Tuanhui Li, Yong Zhang, Mingyang Li, Zhifeng Li, and Yujiu Yang. Sparse adversaries" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 116 + ], + "type": "text", + "content": "ial attack via perturbation factorization. In Computer VisionECCV 2020:16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 35-50. Springer, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 171 + ], + "type": "text", + "content": "[13] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 554, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 225 + ], + "type": "text", + "content": "[14] Samar Fares and Karthik Nandakumar. Attack to defend: Exploiting adversarial attacks for detecting poisoned models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24726-24735, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 281 + ], + "type": "text", + "content": "[15] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 282, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 348 + ], + "type": "text", + "content": "[16] Tianrui Guan, Aswath Muthuselvam, Montana Hoover, Xijun Wang, Jing Liang, Adarsh Jagan Sathyamoorthy, Damon Conover, and Dinesh Manocha. Crossloc3d: Aerial-ground cross-source 3d place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11335-11344, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 380 + ], + "type": "text", + "content": "[17] Ziwen He, Wei Wang, Jing Dong, and Tieniu Tan. Transferable sparse adversarial attack. arXiv preprint arXiv:2105.14727, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 436 + ], + "type": "text", + "content": "[18] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 437, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 554, + 502 + ], + "type": "text", + "content": "[19] Shengshan Hu, Xiaogeng Liu, Yechao Zhang, Minghui Li, Leo Yu Zhang, Hai Jin, and Libing Wu. Protecting facial privacy: Generating adversarial identity masks via style-robust makeup transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15014–15023, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 504, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 554, + 548 + ], + "type": "text", + "content": "[20] Sergio Izquierdo and Javier Civera. Close, but not there: Boosting geographic distance sensitivity in visual place recognition. In European Conference on Computer Vision, pages 240-257. Springer, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 548, + 554, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 591 + ], + "type": "text", + "content": "[21] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 17658-17668, 2024. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 593, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 554, + 646 + ], + "type": "text", + "content": "[22] Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, and Wenqiang Zhang. Efficient decision-based black-box patch attacks on video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4379-4389, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 554, + 690 + ], + "type": "text", + "content": "[23] Tong Jin, Feng Lu, Shuyu Hu, Chun Yuan, and Yunpeng Liu. Editformer: An efficient decoder transformer for visual place recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[24] Ahmad Khaliq, Ming Xu, Stephen Hausler, Michael Milford, and Sourav Garg. Vlad-buff: burst-aware fast feature aggre" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7205" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "gation for visual place recognition. In European Conference on Computer Vision, pages 447-466. Springer, 2024. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 138 + ], + "type": "text", + "content": "[25] Alexey Kurakin, Ian J Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[26] Jie Li, Rongrong Ji, Hong Liu, Xiaopeng Hong, Yue Gao, and Qi Tian. Universal perturbation attack against image retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4899-4908, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "text", + "content": "[27] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 303 + ], + "type": "text", + "content": "[28] Feng Lu, Lijun Zhang, Shuting Dong, Baifan Chen, and Chun Yuan. Aanet: Aggregation and alignment network with semi-hard positive sample mining for hierarchical place recognition. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11771-11778. IEEE, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 360 + ], + "type": "text", + "content": "[29] Feng Lu, Shuting Dong, Lijun Zhang, Bingxi Liu, Xiangyuan Lan, Dongmei Jiang, and Chun Yuan. Deep homography estimation for visual place recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10341-10349, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 425 + ], + "type": "text", + "content": "[30] Feng Lu, Xiangyuan Lan, Lijun Zhang, Dongmei Jiang, Yaowei Wang, and Chun Yuan. Cricavpr: Cross-image correlation-aware representation learning for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16772-16782, 2024. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[31] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. International Conference on Learning Representations, 2024. 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 535 + ], + "type": "text", + "content": "[32] Feng Lu, Xinyao Zhang, Canming Ye, Shuting Dong, Lijun Zhang, Xiangyuan Lan, and Chun Yuan. Supervlad: Compact and robust image descriptors for visual place recognition. Advances in Neural Information Processing Systems, 37:5789-5816, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[33] Lun Luo, Shuhang Zheng, Yixuan Li, Yongzhi Fan, Beinan Yu, Si-Yuan Cao, Junwei Li, and Hui-Liang Shen. Bevplace: Learning lidar-based place recognition using bird's eye view images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8700-8709, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "text", + "content": "[34] Apostolos Modas, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Sparsefool: a few pixels make a big difference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087–9096, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[35] SD Moosavi-Dezfooli. a simple and accurate method to fool deep neural networks/s. moosavi-dezfooli, a. fawzi, p. frossard. Preprint—July, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[36] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE transactions on pattern" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 93 + ], + "type": "text", + "content": "analysis and machine intelligence, 41(10):2452-2465, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[38] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[39] Guohao Peng, Heshan Li, Yangyang Zhao, Jun Zhang, Zhenyu Wu, Pengyu Zheng, and Danwei Wang. Transloc4d: Transformer-based 4d radar place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17595-17605, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[40] Yuxin Qi, Quan Zhang, Xi Lin, Xiu Su, Jiani Zhu, Jingyu Wang, and Jianhua Li. Seeing beyond noise: Joint graph structure evaluation and denoising for multimodal recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12461-12469, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[41] Qibo Qiu, Shun Zhang, Haiming Gao, Honghui Yang, Haochao Ying, Wenxiao Wang, and Xiaofei He. Emvp: Embracing visual foundation model for visual place recognition with centroid-free probing. Advances in Neural Information Processing Systems, 37:120928-120950, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[42] Tianyi Alex Qiu, Zhonghao He, Tejasveer Chugh, and Max Kleiman-Weiner. The lock-in hypothesis: Stagnation by algorithm. arXiv preprint arXiv:2506.06166, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 469 + ], + "type": "text", + "content": "[43] Ivan Reyes-Amezcua, Gilberto Ochoa-Ruiz, and Andres Mendez-Vazquez. Enhancing image classification robustness through adversarial sampling with delta data augmentation (dda). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 274-283, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[44] Fahad Shamshad, Muzammal Naseer, and Karthik Nandakumar. Clip2protect: Protecting facial privacy using text-guided makeup via adversarial latent search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20595-20605, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "text", + "content": "[45] Zhigang Su, Dawei Zhou, Nannan Wang, Decheng Liu, Zhen Wang, and Xinbo Gao. Hiding visual information via obfuscating adversarial perturbations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4356-4366, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[46] Yuhao Sun, Lingyun Yu, Hongtao Xie, Jiaming Li, and Yongdong Zhang. Diffam: Diffusion-based adversarial makeup transfer for facial privacy protection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24584-24594, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[47] Niko Sünderhauf, Peer Neubert, and Peter Protzel. Are we there yet? challenging seqslam on a " + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "inline_equation", + "content": "3000\\mathrm{km}" + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": " journey across all four seasons. In Proc. of workshop on long-term autonomy, IEEE international conference on robotics and automation (ICRA), page 2013. CiteSeer, 2013. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[48] Ajinkya Tejankar, Maziar Sanjabi, Qifan Wang, Sinong Wang, Hamed Firooz, Hamed Pirsiavash, and Liang Tan." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7206" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "text", + "content": "Defending against patch-based backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12239-12249, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "text", + "content": "[49] Akihiko Torii, Relja Arandjelovic, Josef Sivic, Masatoshi Okutomi, and Tomas Pajdla. 24/7 place recognition by view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1808-1817, 2015. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 237 + ], + "type": "text", + "content": "[50] Issar Tzachor, Boaz Lerner, Matan Levy, Michael Green, Tal Berkovitz Shalev, Gavriel Habib, Dvir Samuel, Noam Korngut Zailer, Or Shimshi, Nir Darshan, et al. Effovpr: Effective foundation model utilization for visual place recognition. International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "text", + "content": "[51] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 348 + ], + "type": "text", + "content": "[52] Jiakai Wang, Zixin Yin, Pengfei Hu, Aishan Liu, Renshuai Tao, Haotong Qin, Xianglong Liu, and Dacheng Tao. Defensive patches for robust recognition in the physical world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2456-2465, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "type": "text", + "content": "[53] Zhibo Wang, He Wang, Shuaifan Jin, Wenwen Zhang, Jiahui Hu, Yan Wang, Peng Sun, Wei Yuan, Kaixin Liu, and Kui Ren. Privacy-preserving adversarial facial features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8212-8221, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "text", + "content": "[54] Frederik Warburg, Soren Hauberg, Manuel Lopez-Antequera, Pau Gargallo, Yubin Kuang, and Javier Civera. Mapillary street-level sequences: A dataset for lifelong place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2626-2635, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "text", + "content": "[55] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Yu Li, Chun Yuan, and Dacheng Tao. Task groupings regularization: Data-free meta-learning with heterogeneous pretrained models. In International Conference on Machine Learning, pages 52573-52587. PMLR, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 579 + ], + "type": "text", + "content": "[56] Yongxian Wei, Zixuan Hu, Zhenyi Wang, Li Shen, Chun Yuan, and Dacheng Tao. Free: Faster and better data-free meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23273-23282, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "text", + "content": "[57] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "text", + "content": "[58] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[59] Yongxian Wei, Anke Tang, Li Shen, Zixuan Hu, Chun Yuan, and Xiaochun Cao. Modeling multi-task model merging" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "text", + "content": "ing as adaptive projective gradient descent. arXiv preprint arXiv:2501.01230, 2025. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "text", + "content": "[60] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, Joao F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8461-8472, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "type": "text", + "content": "[61] Xiao Yang, Yinpeng Dong, Tianyu Pang, Hang Su, Jun Zhu, Yuefeng Chen, and Hui Xue. Towards face encryption by generating adversarial identity masks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3897-3907, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 554, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 248 + ], + "type": "text", + "content": "[62] Jingwen Ye and Xinchao Wang. Ungenerable examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11944-11953, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 554, + 293 + ], + "type": "text", + "content": "[63] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 348 + ], + "type": "text", + "content": "[64] Mubariz Zaffar, Liangliang Nan, and Julian FP Kooij. On the estimation of image-matching uncertainty in visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17743-17753, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 404 + ], + "type": "text", + "content": "[65] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 405, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 554, + 458 + ], + "type": "text", + "content": "[66] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7868-7877, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 514 + ], + "type": "text", + "content": "[67] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. Indprompter: Adapting sam to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 515, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 580 + ], + "type": "text", + "content": "[68] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25409-25419, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 555, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 647 + ], + "type": "text", + "content": "[69] Quan Zhang, Jinwei Fang, Rui Yuan, Xi Tang, Yuxin Qi, Ke Zhang, and Chun Yuan. Weakly supervised temporal action localization via dual-prior collaborative learning guided by multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 24139-24148, 2025." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "text", + "content": "[70] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10085-10093, 2025. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7207" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 127 + ], + "type": "text", + "content": "[71] Xinwei Zhang, Tianyuan Zhang, Yitong Zhang, and Shuangcheng Liu. Enhancing the transferability of adversarial attacks with stealth preservation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915-2925, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "text", + "content": "[72] Mingkang Zhu, Tianlong Chen, and Zhangyang Wang. Sparse and imperceptible adversarial attack via a homotopy algorithm. In International Conference on Machine Learning, pages 12868-12877. PMLR, 2021. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7208" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_content_list.json b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0e2775d3208d3005e9034f2b66f061e986f95f1a --- /dev/null +++ b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_content_list.json @@ -0,0 +1,1359 @@ +[ + { + "type": "text", + "text": "VQ-SGen: A Vector Quantized Stroke Representation for Creative Sketch Generation", + "text_level": 1, + "bbox": [ + 210, + 130, + 789, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiawei Wang $^{1,2}$ Zhiming Cui $^{2}$ Changjian Li $^{1}$", + "bbox": [ + 303, + 202, + 692, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Edinburgh 2ShanghaiTech University", + "bbox": [ + 264, + 224, + 730, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://enigma-li.github.io/projects/VQ-SGen/VQ-SGen.html", + "bbox": [ + 253, + 250, + 738, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 303, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents VQ-SGen, a novel algorithm for high-quality creative sketch generation. Recent approaches have framed the task as pixel-based generation either as a whole or part-by-part, neglecting the intrinsic and contextual relationships among individual strokes, such as the shape and spatial positioning of both proximal and distant strokes. To overcome these limitations, we propose treating each stroke within a sketch as an entity and introducing a vector-quantized (VQ) stroke representation for fine-grained sketch generation. Our method follows a two-stage framework - in stage one, we decouple each stroke's shape and location information to ensure the VQ representation prioritizes stroke shape learning. In stage two, we feed the precise and compact representation into an auto-decoding Transformer to incorporate stroke semantics, positions, and shapes into the generation process. By utilizing tokenized stroke representation, our approach generates strokes with high fidelity and facilitates novel applications, such as text or class label conditioned generation and sketch completion. Comprehensive experiments demonstrate our method surpasses existing state-of-the-art techniques on the CreativeSketch dataset, underscoring its effectiveness.", + "bbox": [ + 88, + 334, + 485, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 693, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sketches are an intuitive mechanism widely used by humans to communicate ideas, convey concepts, and express creativity. Over the past decades, significant research has been conducted on sketch-related tasks, including retrieval [3, 13, 14, 23, 24], semantic segmentation [32, 40], and generation [4, 7, 8, 12, 27, 33]. In this paper, we focus on creative sketch generation, a novel task introduced by Ge et al. [8]. Unlike traditional sketch generation, this task aims to produce diverse, complex, and aesthetically appealing sketches, forming more imaginative depictions of familiar visual concepts, rather than generating conventional and typical representations of these concepts. This, thereby,", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fde780d9fcfd323186cb4e8f6dbfd4abb7f80844e9d4326deadb9de2cf376744.jpg", + "image_caption": [ + "Figure 1. The goal of creative sketch generation is to generate vivid sketches, e.g., the birds from existing methods on the left. We have proposed VQ-SGen, for high-quality sketch generation with a new vector-quantized (VQ) representation and an efficient generator. See one of our results on the right for comparison." + ], + "image_footnote": [], + "bbox": [ + 519, + 305, + 895, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "places high demands on the generation capabilities of automated methods. Benefiting from the advancement of diffusion models, the notable and relevant progress for sketch generation is Diffsketer [35], a novel approach for text-to-sketch generation. However, the large-scale backbone diffusion model is pre-trained on realistic images, and their resultant sketches thus have high fidelity and realism conforming to human imagination but fail to promote creativity. Therefore, dedicated methods for creative sketch generation are needed.", + "bbox": [ + 511, + 580, + 906, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods [4, 8] tackle creative sketch generation by producing pixels either as a whole sketch or as segmented parts, overlooking inherent and contextual relationships among individual strokes, such as their shapes and relative positions of both nearby and distant strokes. As a result, their generated sketches often have blurry local regions, and scattered or isolated strokes (see the head of the bird from DoodleFormer and DoodlerGAN in Fig. 1). A more recent work, SketchKnitter [33], formulated the problem of sketch generation as a stroke points rectification process, where diffusion models are leveraged to rearrange a", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19311", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "set of stroke points from an initial chaotic state into coherent forms. However, without the concept of a stroke entity, this method performs poorly when generating complex creative sketches (see the bird in Fig. 1 from SketchKnitter).", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel stroke-level sketch generation approach that treats each stroke as an independent entity [20, 32] and encodes it using a novel vector-quantized (VQ) representation. This compact, discrete representation enhances the learning process of the generative model by efficiently capturing essential stroke shape variations while minimizing redundancy. Despite the discrete nature of the representation, we observe semantic-aware clustering within the stroke code space, which lays an ideal foundation for the generator to sample new strokes. Furthermore, our bespoke generator incorporates the shape, semantic, and position into its generation process, effectively samples new sketches in the compressed VQ space, and accurately reasons out the structural and semantic relationship between strokes, resulting in sketches that are both appealing and structurally coherent.", + "bbox": [ + 89, + 154, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our method VQ-SGen, adopts a two-stage framework. In the first stage, we decouple the shape and location information of each stroke. This decoupling allows the vector-quantized (VQ) representation to concentrate on capturing stroke shapes, minimizing interference from positional data during the learning process. By prioritizing shape as the primary focus, the model learns a highly compact and informative, shape-oriented representation that preserves the inherent shape features of each stroke while reducing unnecessary redundancy. In the second stage, we employ an autoregressive Transformer to work with our VQ representation. The Transformer progressively integrates the stroke's semantic (i.e., the category label), shape, and location into the generation process. Through the autoregressive decoding mechanism, the Transformer captures each stroke's semantic, shape, and context relationship with neighboring strokes, ensuring that the shape and positioning are accurate and harmoniously aligned with its semantic role within the overall composition. As a result, our approach effectively avoids all the artifacts observed in previous methods.", + "bbox": [ + 89, + 397, + 482, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our method, we have conducted experiments on the CreativeSketch [8] dataset. Comprehensive experiments including comparisons and ablation studies show that our method surpasses existing state-of-the-art techniques. In summary, our main contributions are as follows:", + "bbox": [ + 89, + 717, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel stroke-level approach, $VQ-SGen$ , for the challenging creative sketch generation task.", + "- We propose a new vector-quantized (VQ) stroke representation that treats each stroke as an entity, capturing essential shape features compactly and reducing redundancy to serve as the basis for stroke-level sketch generation." + ], + "bbox": [ + 91, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a bespoke generator, leveraging the shape, semantic, and spatial location together, achieving high-quality generation.", + "- We extensively evaluated our method, demonstrating superior performance. A user study shows that VQ-SGen consistently performs favorably against SoTA methods." + ], + "bbox": [ + 513, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 195, + 653, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sketch Representation Learning. Based on the data format of sketches, existing sketch representation learning methods can be broadly divided into three categories: image-based [16, 42, 43], sequence-based [1, 15, 19, 34], and graph-based [31, 37, 39] approaches, each with its own advantages and limitations. Specifically, image-based methods [16, 42, 43] use raster images as input, leveraging absolute coordinates to capture stroke proximity. However, they often struggle to learn structural information and sequential relationships among strokes effectively. Graph-based methods [31, 37, 39] face challenges in capturing detailed structural information within strokes. Sequence-based methods [15, 19, 34] utilize point sequences with relative coordinates, which enhances the encoding of individual stroke structures but makes it difficult to represent inter-stroke proximity and spatial relationships. ContextSeg [32] proposed the concept of the single stroke entity for sketch semantic segmentation. The advantages of this representation are demonstrated in their paper by comparing it with fancy Transformer [22] or common RNN [9, 15] based methods. SketchXAI [20] decoupled a single stroke into its shape, location, and order embeddings, and the concatenated stroke embedding was used to train a sketch classifier to study the explainability of human sketches. We exploit the same concept of treating a single stroke as the atomic unit, and decouple its shape with its location. Additionally and importantly, we train a vector quantized representation to build the discrete and compact code space dedicated to our generation task.", + "bbox": [ + 511, + 220, + 906, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vector-Quantized VAEs. Learning representations with discrete features has been the focus of many previous works [6, 11, 30], as such representations are a natural choice for complex reasoning, planning, and predictive learning. Although using discrete latent variables in deep learning has been difficult, effective autoregressive models have been created to capture distributions over discrete variables. For instance, StrokeNUWA [26] utilizes VQ-VAE [28] to achieve a better representation of vector graphics, then generates vector graphics through large language models. Similarly, we propose to learn a VQ representation for rasterized strokes, expecting an improved generation capability powered by the compact representation.", + "bbox": [ + 511, + 659, + 908, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Creative Sketch Generation. Unlike traditional sketch generation tasks [38, 41], creative sketch generation [8] places greater emphasis on producing more imaginative de", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19312", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/45510d92dace14c50ed3f171d04a8f5e2ab66a7c01c4c333e51ecb8a3c0f4391.jpg", + "image_caption": [ + "Figure 2. Overview of VQ-SGen. Given an input sketch, it is first divided into a sequence of strokes. In the first stage, we begin by decoupling the shape and location information of each stroke and obtain their discrete representations (Sec. 3.1). In the second stage, we use a decoder-only Gen-Transformer to predict $\\{(I_i, b_i, l_i)\\}$ in an autoregressive manner (Sec. 3.2)." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 648, + 340 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/03231a5eebdc93b56ced6eaf36c6cbe34db554dfdd128fa50830a761df12761d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 90, + 903, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pictures of familiar visual concepts, rather than generating conventional and mundane forms of these concepts. To achieve this goal, DoodlerGAN [8] presents a component-based GAN architecture to incrementally generate each part of a creative sketch image. This approach requires training one GAN model individually for each body part in a supervised way, leading to significant computational overhead. DoodleFormer [4] employs an attention-based architecture for two-stage sketch generation. However, this approach tends to focus only on the broad, rough relationships between different parts of the sketch, which can lead to unnatural mixing of strokes from different regions.", + "bbox": [ + 88, + 400, + 485, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 601, + 183, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch and strokes. A standard sketch $S$ in our context is composed of a sequence of strokes $\\{s_i\\}_{i=1}^N$ and their corresponding stroke labels $\\{l_i\\}_{i=1}^N$ . The order of each stroke in the sequence corresponds to their default drawing order. Each stroke is represented as a rasterized image of size $256 \\times 256$ , and each label $l_i$ is a one-hot vector in $\\mathbb{R}^C$ . In order to decouple the shape and position of a single stroke, we further process it. Formally, for each rasterized stroke image, we calculate its axis-aligned bounding box and record its center coordinate $(x_i, y_i)$ . The bounding box size (i.e., the height and width) as well as the center coordinate fully represent the stroke position $b_i = (w_i / 2, h_i / 2, x_i, y_i)$ . By factoring out the position (i.e., translating the stroke bounding box to align with the sketch image center), the stroke shape itself is presented as the translated rasterized image $I_i$ without changing its scale. Considering the triplet of shape, position, and semantic of a single stroke, the new format to represent a sketch is $S = \\{s_i\\}_{i=1}^N = \\{(I_i, b_i, l_i)\\}_{i=1}^N$ .", + "bbox": [ + 89, + 628, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch generation. Our goal in this paper is to generate a sequence of strokes in the format of the triplet and assemble them based on their shape and position to form the resulting sketch (see Fig. 1). Figure 2 displays an overview of our novel sketch generation method, which is comprised of two main components. Firstly, a decoupled representation learning module encodes a stroke into its vector-quantized representation (Sec. 3.1) that is highly compact and shape-aware. Secondly, a bespoke generator based on the Transformer [29] is designed to generate the stroke label, shape, and location sequentially, achieving a fine-grained control with high-fidelity resulting sketches. In the following sections, we elaborate on the details.", + "bbox": [ + 511, + 400, + 908, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vector-quantized Stroke Representation", + "text_level": 1, + "bbox": [ + 511, + 606, + 857, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our triplet stroke formula enables separate control of stroke key properties in the generation process, how to properly encode the triplet to fit generation is the next step. Motivated by the vector quantization process [28], we aim to obtain a similar compact stroke representation in the context of the whole sketch, such that the sketch generator can efficiently and effectively sample new sketches in a compressed stroke space. Besides, we expect a decoupled representation based on our triplet formulation to obtain high shape and semantic awareness. To this end, we first convert each stroke into a latent embedding and further compress them in the latent space.", + "bbox": [ + 511, + 628, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stroke latent embedding. Given the stroke image $I_{i}$ we obtain the stroke latent embedding via an autoencoder, where both the encoder $\\mathcal{E}^s$ and the decoder $\\mathcal{D}^s$ are built with a few 2D CNNs (see supplementary for detailed network configuration). After the last layer of the encoder, we flatten its bottleneck feature to obtain the stroke latent em", + "bbox": [ + 511, + 810, + 908, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19313", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bedding $e_i^s = \\text{flat}(\\mathcal{E}^s(\\boldsymbol{I}_i))$ . The autoencoder is trained using the reconstruction loss:", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n s}} = \\left\\| \\boldsymbol {I} _ {i} - \\mathcal {D} ^ {s} \\left(\\mathcal {E} ^ {s} \\left(\\boldsymbol {I} _ {i}\\right)\\right) \\right\\| ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 130, + 482, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better capture the stroke shape information, we follow [2, 17, 32] to add the CoordConv layer to the network and the distance field supervision in the training process.", + "bbox": [ + 89, + 160, + 482, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenizing strokes. To achieve fine-grained control, we build separate discretized spaces for the shape and location, respectively. The reason to build a space for the shape location is that it is highly related to semantics, e.g., the heads of birds appear at a similar position with a similar size.", + "bbox": [ + 89, + 205, + 483, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the stroke shape space, as shown in Fig. 2 left, we build a codebook $D_{s}$ . Specifically, given a sketch $\\mathbf{S}$ , its corresponding stroke shape latent embeddings $\\{e_i^s\\}_{n = 1}^N$ are fed into the encoder $\\mathcal{E}^f$ to obtain the feature sequence $\\{z_i^s\\}_{i = 1}^N$ . We then compress it via vector quantization, i.e., clamping each feature into its nearest code in $D_{s}$ with $V$ codes $\\{c_i\\}_{i = 1}^V$ , and we record the indices of these features:", + "bbox": [ + 89, + 281, + 483, + 387 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} ^ {s} = \\operatorname {a r g m i n} _ {j \\in [ 0, V)} \\| \\boldsymbol {z} _ {i} ^ {s} - \\boldsymbol {c} _ {j} \\|. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 397, + 482, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The feature sequence $\\{c_{v_i^s}\\}^N$ is further sent to the decoder $\\mathcal{D}^f$ to reconstruct the input latent embeddings. Both $\\mathcal{E}^f$ and $\\mathcal{D}^f$ are composed of several 1D CNN layers, see the supplementary for the detail structure.", + "bbox": [ + 89, + 426, + 483, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the vector quantization network, we follow [28] to exploit the codebook, commitment, and reconstruction losses, defined as:", + "bbox": [ + 89, + 489, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\alpha (\\| \\pmb {z} _ {i} ^ {s} - \\mathrm {s g} [ \\pmb {c} _ {v _ {i} ^ {s}} ] \\| _ {2} ^ {2} + \\| \\mathrm {s g} [ \\pmb {z} _ {i} ^ {s} ] - \\pmb {c} _ {v _ {i} ^ {s}} \\| _ {2} ^ {2}) \\\\ + \\left\\| \\boldsymbol {z} _ {i} ^ {s} - \\mathcal {D} ^ {f} \\left(\\boldsymbol {c} _ {v _ {i} ^ {s}}\\right) \\right\\| _ {2} ^ {2}, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 544, + 482, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a weight, and $\\mathrm{sg}$ is the stopgradient operator.", + "bbox": [ + 89, + 619, + 460, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the stroke location codebook learning, we use the same process as the shape codebook learning. The stroke position $b_{i} = (w_{i} / 2, h_{i} / 2, x_{i}, y_{i})$ is reconstructed by a pair of encoder $(\\mathcal{E}^{l})$ and decoder $(\\mathcal{D}^{l})$ to build the codebook $D_{l}$ with V codes $\\{d_{i}\\}_{i=1}^{V}$ as well. See supplementary for detailed network structures.", + "bbox": [ + 89, + 633, + 483, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Till now, given a stroke $\\boldsymbol{s}_i = (\\boldsymbol{I}_i, b_i, l_i)$ , we have the compact representation $(v_i^s, v_i^l, l_i)$ , corresponding to the specific code features $(\\boldsymbol{c}_{v_i^s}, \\boldsymbol{d}_{v_i^l}, l_i)$ .", + "bbox": [ + 89, + 724, + 483, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Autoregressive Generation", + "text_level": 1, + "bbox": [ + 89, + 787, + 333, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Having the VQ stroke representation, our goal in this module is to estimate a distribution over the sketch $S$ so as to generate new instances. Since $S$ is composed of a sequence of stroke shapes, positions, and labels, we split the generation process into two steps, i.e., 1) generate the stroke label, and 2) generate the stroke shape and position conditioned", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db616db51cc3c6e33f3f5842d8e3d8015c9eb5e78bb230fccab4170fa3120fed.jpg", + "image_caption": [ + "Figure 3. Network architecture of our generator." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 901, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on the stroke label. By applying the chain rule, we formulate it as:", + "bbox": [ + 511, + 263, + 903, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\boldsymbol {S}) = \\prod_ {i = 1} ^ {N} p \\left(\\boldsymbol {I} _ {i}, b _ {i}, l _ {i}\\right) = \\prod_ {i = 1} ^ {N} p \\left(v _ {i} ^ {\\mathrm {s}}, v _ {i} ^ {l} \\mid l _ {i}\\right) \\cdot p \\left(l _ {i}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 297, + 903, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We implement Eq. (4) with two cascaded Transformer decoders, parameterized with $\\theta_{1}$ and $\\theta_{2}$ , respectively. The first decoder samples a label for the next stroke, while the second decoder generates the shape and position conditioned on the label, which can be further formulated as:", + "bbox": [ + 511, + 344, + 905, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (l _ {i}) = p \\left(l _ {i} \\mid l _ {< i}, v _ {s < } ^ {s}, v _ {l} ^ {< }; \\theta_ {1}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 425, + 826, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (i) \\quad P (i + l _ {i}) \\quad P (i + l _ {i}) \\quad P (i + 1) \\\\ p \\left(v _ {i} ^ {s}, v _ {i} ^ {l} \\mid l _ {i}\\right) = p \\left(v _ {i} ^ {s}, v _ {i} ^ {l} \\mid l _ {\\leq i}, v _ {< i} ^ {s}, v _ {< i} ^ {l}; \\theta_ {2}\\right). \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 438, + 903, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, we will use the codes in $D_{s}$ and $D_{l}$ in the Transformer calculation, instead of the original rasterized image and the quaternion for the stroke shape and position. Network Architectures. An overview of our cascaded decoders is shown in Fig. 3, where the first label Transformer $(\\mathcal{T}^{l})$ predicts a stroke label each time, while the second code Transformer $(\\mathcal{T}^{c})$ takes as input the predicted label generating both the stroke shape and position codes simultaneously.", + "bbox": [ + 511, + 469, + 905, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, the label Transformer $\\mathcal{T}^l$ takes as input the triplet $\\left(\\pmb{c}_{v_i^s},\\pmb{d}_{v_i^l},l_i\\right)$ . Each element goes through its own embedding layer to get a feature, i.e., $\\mathrm{Emb}(\\pmb{c}_{v_i^s})\\in \\mathbb{R}^{512}$ , $\\mathrm{Emb}(\\pmb{d}_{v_i^l})\\in \\mathbb{R}^{512}$ , and $\\mathrm{Emb}(l_i)\\in \\mathbb{R}^{512}$ . The three features are simply added along the feature dimension to obtain the input feature for the attention blocks, which outputs a fused feature $\\mathcal{F}_{\\mathrm{fuse}}$ after the attention calculation. $\\mathcal{F}_{\\mathrm{fuse}}$ is then fed into a fully connected layer with the softmax activation to predict the one-hot probabilistic vector, which is translated into a one-hot class label $l_{i + 1}\\in \\mathbb{R}^C$ .", + "bbox": [ + 511, + 590, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The code Transformer $\\mathcal{T}^c$ accepts two inputs. The first one is $l_{i + 1}$ , which goes through an embedding layer to produce a conditional feature in $\\mathbb{R}^{512}$ . The second input is $\\mathcal{F}_{\\mathrm{fuse}}$ , which is also added with the conditional feature along the feature dimension to get the input feature for the attention block. Intuitively, the first input provides the direct semantic guidance of the current stroke, while the second input provides the shape and positional information of previous strokes. The attention blocks are followed by two separate branches, each of which outputs a one-hot probabilistic", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19314", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vector of dimension $\\mathbb{R}^V$ , indicating the indices of the shape and position codebooks. The corresponding code entries are fetched (i.e., $c_{v_{i+1}^s}$ and $d_{v_{i+1}^t}$ ) and serve as the input to the label Transformer for the next stroke generation.", + "bbox": [ + 89, + 90, + 480, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Losses. To train the network, we minimize the negative log-likelihood loss, defined as:", + "bbox": [ + 89, + 154, + 482, + 184 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g e n}} = - \\log p (\\boldsymbol {S}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 199, + 482, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we supervise the training with ground truth categorical labels (i.e., the shape label vector, shape code index vector, and the position code index vector).", + "bbox": [ + 89, + 229, + 480, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and testing strategy. We have three training stages arranged sequentially for stroke latent embedding, codebooks, and the generator. Detailed steps, hyperparameter settings, and implementation details can be found in the supplementary. For the testing time generation, starting from a special STR token, we generate novel strokes in an autoregressive manner, where the codebook entry is either decoded as the position quaternion or the stroke latent code, which is further converted into the corresponding rasterized image. The stroke triplet is finally assembled together to produce the resulting sketch (see Fig. 2, left).", + "bbox": [ + 89, + 277, + 483, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 460, + 222, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following previous work [4, 8], we use the CreativeSketch [8] dataset in this paper. Specifically, it consists of two categories of sketches — Creative Birds (CB) and Creative Creatures (CC) — containing 8,067 and 9,097 sketches, respectively, along with part annotations. All the sketches have been processed into $256 \\times 256$ images. For the stroke latent embedding learning, we use strokes from both datasets to obtain a powerful pair of $\\mathcal{E}^s$ and $\\mathcal{D}^s$ . Other than this autoencoder, we have trained other networks separately on CB and CC datasets. See supplementary for details about pre-processing and data augmentation.", + "bbox": [ + 89, + 487, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. Following previous work [4, 8], we evaluate the model's performance using an Inception model [25] trained on the QuickDraw3.8M dataset [36]. We use four metrics in this paper, defined as follows:", + "bbox": [ + 89, + 655, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Fréchet Inception distance (FID) [10] measures the similarity between two image sets in the above Inception feature space.", + "- Generation diversity (GD) [5] measures the average pairwise Euclidean distance between the Inception features of two subsets, reflecting the diversity of the data item in the image set.", + "- Characteristic score (CS) measures how frequently a generated sketch is classified as Creative Bird or Creative Creature by the Inception model.", + "- Semantic diversity score (SDS) quantifies sketch diversity based on various creature categories they represent." + ], + "bbox": [ + 89, + 715, + 482, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8ae56305b1201f8fc876f42fe2749b9e4fb0f9f3140106b7db64c338477585c6.jpg", + "table_caption": [ + "Table 1. Statistical comparison against existing methods on Creative Birds and Creative Creatures datasets." + ], + "table_footnote": [], + "table_body": "
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
Sketchknitter [33]74.4214.230.1464.3412.340.421.32
DoodlerGAN [8]39.9516.330.6943.9414.570.551.45
DoodleFormer [4]17.4817.830.5720.4316.230.531.68
Ours15.7818.920.5317.6117.420.571.86
", + "bbox": [ + 517, + 121, + 903, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparison", + "text_level": 1, + "bbox": [ + 511, + 250, + 645, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our approach with three competitors, i.e., Doodleformer [4], DoodlerGAN [8], Sketchknitter [33]. Both Doodleformer and DoodlerGAN generate sketches at the pixel level, while Sketchknitter generates sketches at the stroke point sequence level. All three methods are trained using their default parameters on our dataset.", + "bbox": [ + 511, + 284, + 906, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quantitative comparison is presented in Tab. 1, where our method significantly outperforms others. Specifically, on the Creative Birds dataset, we obtained a notable 1.7 decrease in FID and gained a remarkable 1.09 increase in terms of GD, achieving the best performance. This indicates that our model can generate sketches that are closer to the original dataset, as well as more diverse at the same time. For the CS metric, while our approach does not achieve the highest score, it surpasses the dataset and comes closest to matching the dataset's CS metric. In contrast, DoodlerGAN and DoodleFormer attain higher scores, whereas SketchKnitter's outputs are seldom recognized as birds. For this specific metric, the higher value only means it is easier to be recognized by the Inception model, which may stem from the simplicity of the generation. The dataset's value sets a good reference. Similarly, on the Creative Creatures dataset, we achieved the best performance across all four metrics. Specifically, we achieved a 2.82 decrease in FID and a 1.19 improvement in GD compared to the best other methods. This indicates that our generated sketches are closer to the more complex sketches in the training data and exhibit greater diversity among themselves. For the CS and SDS metrics, we obtained improvements of 0.04 and 0.18, respectively. This indicates that our generated sketches are semantically closer to the categories in the training data and also suggests that our model exhibits better semantic diversity. A corresponding visual comparison is shown in Fig. 4, where the artifacts like the broken and misaligned strokes from DoodlerGAN, and the blurry local regions from DoodleFormer, are clearly presented. In contrast, our results are of high quality with complex appearance and diverse shapes. Note that, as shown in Fig. 1, SketchKnitter generates poor results on both datasets, we did not include its results in the visual comparison.", + "bbox": [ + 511, + 387, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/87506e4a5b27e9f093eb519df6e3d51362a85b5811b17e7e09a20ad172631e9e.jpg", + "image_caption": [ + "Figure 4. Visual comparison against competitors on both the Creative Birds and Creative Creatures datasets." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 906, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d700b703190675cf859bc885ffaca268059af48ee1f0cdfdb18399f4f7daff58.jpg", + "image_caption": [ + "Figure 5. Sketch reconstruction results of our ablation study on codebook hyper-parameter setting. The two numbers represent the code size and the feature dimension, respectively." + ], + "image_footnote": [], + "bbox": [ + 135, + 486, + 864, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 835, + 243, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our core technical components, we conducted experiments by removing them.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we form the following alternative solutions:", + "bbox": [ + 513, + 835, + 888, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- w/o VQ: we remove all the codebook learning-related procedures, and use the continuous stroke latent embeddings to train our generator.", + "bbox": [ + 514, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19316", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d360e5265c080b6618d1ce86bc3c3545e99cd31e2b3eb87b8541e8cd9ede8f11.jpg", + "table_caption": [ + "Table 2. The statistical comparison of our ablation study. The method name of two numbers represents the codebook hyperparameter setting, where the first is the code size, and the second is the feature dimension." + ], + "table_footnote": [], + "table_body": "
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
2048×51226.2315.340.4343.2114.510.481.52
4096×51216.9218.270.5018.4416.140.541.63
4096×102416.6718.150.5118.1216.430.551.69
w/o Tl16.5118.350.5119.2116.140.551.74
w/o VQ48.5313.340.4654.5614.020.451.36
w/o Decouple17.1418.120.5719.4216.420.561.79
Ours(8192 × 512)15.7818.920.5317.6117.420.571.86
", + "bbox": [ + 94, + 148, + 480, + 277 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- w/o Decouple: we do not decouple the shape and position of the stroke, and only train a single stroke codebook. The Transformer decoders take fewer inputs and the second decoder outputs only one index vector accordingly.", + "- w/o $\\mathcal{T}^l$ : we simply discard the label information for each stroke, and remove the label Transformer accordingly." + ], + "bbox": [ + 89, + 290, + 482, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train the three alternatives from scratch using our datasets until convergence and statistical results are shown in Tab. 2. Without the VQ representation, the generator samples new sketches from a larger continuous space, instead of a compressed discretized space, giving the worst performance in this paper. The stroke decoupling matters the generation significantly (e.g., 15.78 vs. 17.14 and 17.61 vs. 19.42 of the FID). By decoupling, the shape code can concentrate only on the shape variation, while the position code can better capture the inherent approximation information between strokes, which further facilitates the generation. By removing the label prediction, we see a slight performance degradation, giving the second-best performance in Tab. 2. The labels are not very critical, but indeed provide semantic-related information to complement the positional information (e.g., the heads appear at similar positions with similar sizes). On the other hand, it demonstrates the great feasibility of applying our method to other datasets without label information, proving our flexibility. See Sec. 5 for such an experiment.", + "bbox": [ + 89, + 383, + 482, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Codebook hyper-parameter setting. There are two core hyper-parameters for a codebook, i.e., the code size and the feature dimension. In the following, we have conducted experiments on the shape codebook to find the proper configuration. Similar experiments on the location codebook are presented in the supplementary. Specifically, we set the code size to be either 2,048, 4,096, or 8,192, and set the feature dimension to be either 512 or 1,024. We examine the effects of the parameter setting from both the sketch reconstruction and generation perspectives.", + "bbox": [ + 89, + 686, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 presents a visual comparison of sketch reconstruction quality. By increasing either the code size or the feature dimension, we observe improved fidelity with less blurry and broken strokes. Note that the row named $w / o$ $VQ$", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/937c6bef050e1477f8016fb2885ec6a44e7b457483217d5d5022cac65efd059a.jpg", + "image_caption": [ + "Figure 6. User study results based on the five questions (a-e). Higher scores indicate a stronger preference for our method over the other approaches (i.e., DoodleFormer or Creative datasets)." + ], + "image_footnote": [], + "bbox": [ + 517, + 93, + 903, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shows the stroke latent space reconstruction performance as in [32], which along with the input, provides a good performance reference. We chose the last configuration in this paper, as it gives the best visual reconstruction. However, it still struggles to represent strokes with significant variation, such as the ear of the creature in the third-to-last column. We further investigate the generation performance of different codebook settings. Quantitative results are listed in Tab. 2, where we observe improvement when increasing either the size of the code or the feature dimension. Our choice obtains the best result given the limited combination.", + "bbox": [ + 511, + 297, + 906, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. User Study", + "text_level": 1, + "bbox": [ + 511, + 474, + 635, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [4, 8], we evaluate the performance of our method through user studies. We invited 50 participants to the test, where they were shown multiple pairs of images. Each pair consists of a sketch generated by our model and another sketch from either DoodleFormer or the datasets. For each pair, participants were asked to answer 5 single-choice questions: which one (a) is more creative? (b) looks more like a bird/creature? (c) is more likely to be drawn by a human? (d) the strokes better integrated? (e) they like the most overall? Figure 6 shows the percentage of times our method is preferred over the competing approach. It can be seen that $VQ-SGen$ performs favorably against DoodleFormer for all five questions on both datasets except for question (b) on the Creative Birds dataset, which is consistent with the CS metric in Tab. 1. We are nearly tied with the creative datasets.", + "bbox": [ + 511, + 497, + 906, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion and Application", + "text_level": 1, + "bbox": [ + 511, + 753, + 764, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we mainly discuss code space exploration and conditional generation applications, while leaving other discussions, e.g., limitations, in the supplementary.", + "bbox": [ + 511, + 779, + 905, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Code space exploration. To better understand the properties of the discrete space, we applied the UMAP algorithm [18] to project the high dimensional features of the shape codebook $D^{s}$ into 2D space and visualize its distribution in Fig. 8. We further probe the emerged clustering by visu", + "bbox": [ + 511, + 825, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19317", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5a9fca7875781b8283dbd3a12af22d95e4806242a8f23e98e7031c599fc0e13.jpg", + "image_caption": [ + "Figure 7. Applications of our approach. (a) Given the class label, our method can produce corresponding sketches with rich variations. (b) Our method supports text-conditioned generation, the resulting sketches explain the input text vividly. (c) Given the initial stroke, our method can complete the whole sketch, which is favorable against competitors." + ], + "image_footnote": [], + "bbox": [ + 96, + 94, + 890, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79bf3d42d6bf35e06c5f2f66185f4638c939ef14a5ec522e1dd5700192dfd3a0.jpg", + "image_caption": [ + "Figure 8. UMAP visualization of the shape codes in $\\mathcal{D}^s$ , with the overlaid strokes indicating the semantic aware clustering in the discrete code space." + ], + "image_footnote": [], + "bbox": [ + 101, + 316, + 467, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "alizing typical strokes. Not too surprised, we find that the emerged clusters correspond well to semantics. For example, the five surrounding clusters are Beak, Eye, Tail, Wing, and Leg, while the inner big cluster composed of Head and Body. Although we did not exploit any semantic-related constraints during the codebook learning, it tends to implicitly group strokes with similar shape variations, thereby achieving stable and efficient code learning and space compression. As a result, this provides an ideal foundation for the generator to sample new strokes within a semantically-aware, compressed space.", + "bbox": [ + 88, + 598, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class label-to-sketch generation. Other than unconditional generation, our method can generate sketches given the class label conditions. To do this, we exploit a subset of the QuickDraw dataset [9]. The subset contains 20 classes, and stroke labels are NOT available in this dataset. Our adaptation is pretty simple and straightforward - i) removing the label Transformer (i.e., the variant of w/o $\\mathcal{T}^l$ in Sec. 4.2), and ii) replacing the special STR token with the class index token (e.g., the sinusoidal encoding). We only", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "re-train the generator on the QuickDraw subset until convergence. At inference time, given a class label, our method successfully produces high-quality results, as shown in Fig. 7(a). Note that, the experiment also demonstrates our flexibility and superior performance when working on datasets without stroke label information.", + "bbox": [ + 511, + 320, + 906, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-sketch generation. There are text descriptions for each sketch in the CB and CC datasets. We thus adapt our method to perform text-to-sketch generation. To this end, we first embed the text description using the pre-trained CLIP [21] model, and then replace the $STR$ token with the obtained text token, similarly as in the above label-to-sketch generation task. Fig. 7(b) shows a few novel examples of our generated sketches, which conform well to input texts. Pay attention to the novel visual concept, e.g., the 'Hug me' sketch, and the 'Smiling' sketch, which is rarely seen in the traditional sketch generation task.", + "bbox": [ + 511, + 412, + 908, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sketch completion. Given an initial stroke, our model can perform sketch completion, i.e., stroke conditional generation. We achieve this by feeding the initial stroke as our first stroke in the generation process. Fig. 7(c) presents a visual comparison against SoTA methods on this task. As can be seen, DoodlerGAN produced the worst results, with most sketches lacking a clear structure. DoodleFormer performed slightly better, but the generated sketches are often blurry and include ghost strokes (e.g., a few tail-like strokes of the second example). With the help of our VQ representation and our generator, our results are both creative and structurally coherent.", + "bbox": [ + 511, + 580, + 908, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 782, + 633, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented a novel approach for creative sketch generation. Our method can effectively create sketches with an appealing appearance and coherent structures. We believe that stroke VQ representations can be used in other sketch-related tasks and hope to inspire future research in sparse stroke representation learning.", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19318", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. The authors would like to thank Oisin Mac Aodha and Ankan Bhunia for proofreading the early draft and the valuable discussions, Haocheng Yuan and Yi Yang for helping prepare comparisons with Diffsketcher, and the support of the HPC Platform at ShanghaiTech University. CL was supported by a gift from Adobe.", + "bbox": [ + 89, + 90, + 485, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 210 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12565-12574, 2024. 2", + "[2] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12565-12574, 2024. 4", + "[3] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10997-11008, 2024. 1", + "[4] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. In European Conference on Computer Vision, pages 338-355. Springer, 2022. 1, 3, 5, 7", + "[5] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Aisketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 5", + "[6] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2", + "[7] Chenjian Gao, Xin Wang, Qian Yu, Lu Sheng, Jing Zhang, Xiaoguang Han, Yi-Zhe Song, and Dong Xu. 3d reconstruction from a single sketch via view-dependent depth sampling. IEEE transactions on pattern analysis and machine intelligence, 2024. 1", + "[8] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creativ sketch generation. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 7", + "[9] David Ha and Douglas Eck. A neural representation of sketch drawings. arXiv preprint arXiv:1704.03477, 2017. 2,8", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + ], + "bbox": [ + 93, + 220, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[11] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2", + "[12] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-image diffusion models are great sketch-photo matchmakers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16826-16837, 2024. 1", + "[13] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16859-16869, 2024. 1", + "[14] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16859-16869, 2024. 1", + "[15] Ke Li, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Honggang Zhang. Toward deep universal sketch perceptual grouper. IEEE Transactions on Image Processing, 28(7):3219-3231, 2019. 2", + "[16] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning. IEEE computer graphics and applications, 39(2):38-51, 2018. 2", + "[17] Rosanne Liu, Joel Lehman, Piero Molino, Felipe Petroski Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An intriguing failing of convolutional neural networks and the coordconv solution. Advances in neural information processing systems, 31, 2018. 4", + "[18] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 7", + "[19] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnN for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 2", + "[20] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23327-23337, 2023. 2", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 8", + "[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14153-14162, 2020. 2" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19319", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8504–8513, 2021. 1", + "[24] Aneeshan Sain, Pinaki Nath Chowdhury, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Freeview sketching: View-aware fine-grained sketch-based image retrieval. In European Conference on Computer Vision, pages 145-162. Springer, 2024. 1", + "[25] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5", + "[26] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, et al. Strokenuwa: Tokenizing strokes for vector graphic synthesis. arXiv preprint arXiv:2401.17093, 2024. 2", + "[27] Chaitat Utintu, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Sketchdeco: Decorating b&w sketches with colour. arXiv preprint arXiv:2405.18716, 2024. 1", + "[28] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4", + "[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[30] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2", + "[31] Fei Wang, Shujin Lin, Hanhui Li, Hefeng Wu, Tie Cai, Xiaonan Luo, and Ruomei Wang. Multi-column point-cnn for sketch segmentation. Neurocomputing, 392:50-59, 2020. 2", + "[32] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3679-3688, 2024. 1, 2, 4, 7", + "[33] Qiang Wang, Haoge Deng, Yonggang Qi, Da Li, and Yi-Zhe Song. Sketchknitter: Vectorized sketch generation with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 1, 5", + "[34] Xingyuan Wu, Yonggang Qi, Jun Liu, and Jie Yang. Sketch-segnet: A rnn model for labeling sketch strokes. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2", + "[35] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36:15869-15889, 2023. 1", + "[36] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand" + ], + "bbox": [ + 91, + 92, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sketch: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(1):285-312, 2022. 5", + "[37] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. Sketchgnn: Semantic sketch segmentation with graph neural networks. ACM Transactions on Graphics (TOG), 40(3):1-13, 2021. 2", + "[38] Ningyuan Zheng, Yifan Jiang, and Dingjiang Huang. Strokenet: A neural painting environment. In International Conference on Learning Representations, 2018. 2", + "[39] Yixiao Zheng, Jiyang Xie, Aneeshan Sain, Yi-Zhe Song, and Zhanyu Ma. Sketch-segformer: Transformer-based segmentation for figurative and creative sketches. IEEE Transactions on Image Processing, 2023. 2", + "[40] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 1", + "[41] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to doodle with stroke demonstrations and deep q-networks. In BMVC, page 13, 2018. 2", + "[42] Xianyi Zhu, Yi Xiao, and Yan Zheng. Part-level sketch segmentation and labeling using dual-cnn. In Neural Information Processing: 25th International Conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16, 2018, Proceedings, Part I 25, pages 374-384. Springer, 2018. 2", + "[43] Xianyi Zhu, Yi Xiao, and Yan Zheng. 2d freehand sketch labeling using cnn and crf. Multimedia Tools and Applications, 79(1-2):1585–1602, 2020. 2" + ], + "bbox": [ + 516, + 92, + 905, + 502 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19320", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_model.json b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dfe7d01f175f3fe8452b303a8d07d89be6c733c3 --- /dev/null +++ b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_model.json @@ -0,0 +1,1958 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.79, + 0.175 + ], + "angle": 0, + "content": "VQ-SGen: A Vector Quantized Stroke Representation for Creative Sketch Generation" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.203, + 0.693, + 0.223 + ], + "angle": 0, + "content": "Jiawei Wang\\(^{1,2}\\) Zhiming Cui\\(^{2}\\) Changjian Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.225, + 0.732, + 0.245 + ], + "angle": 0, + "content": "1The University of Edinburgh 2ShanghaiTech University" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.251, + 0.739, + 0.27 + ], + "angle": 0, + "content": "https://enigma-li.github.io/projects/VQ-SGen/VQ-SGen.html" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.304, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.335, + 0.486, + 0.669 + ], + "angle": 0, + "content": "This paper presents VQ-SGen, a novel algorithm for high-quality creative sketch generation. Recent approaches have framed the task as pixel-based generation either as a whole or part-by-part, neglecting the intrinsic and contextual relationships among individual strokes, such as the shape and spatial positioning of both proximal and distant strokes. To overcome these limitations, we propose treating each stroke within a sketch as an entity and introducing a vector-quantized (VQ) stroke representation for fine-grained sketch generation. Our method follows a two-stage framework - in stage one, we decouple each stroke's shape and location information to ensure the VQ representation prioritizes stroke shape learning. In stage two, we feed the precise and compact representation into an auto-decoding Transformer to incorporate stroke semantics, positions, and shapes into the generation process. By utilizing tokenized stroke representation, our approach generates strokes with high fidelity and facilitates novel applications, such as text or class label conditioned generation and sketch completion. Comprehensive experiments demonstrate our method surpasses existing state-of-the-art techniques on the CreativeSketch dataset, underscoring its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Sketches are an intuitive mechanism widely used by humans to communicate ideas, convey concepts, and express creativity. Over the past decades, significant research has been conducted on sketch-related tasks, including retrieval [3, 13, 14, 23, 24], semantic segmentation [32, 40], and generation [4, 7, 8, 12, 27, 33]. In this paper, we focus on creative sketch generation, a novel task introduced by Ge et al. [8]. Unlike traditional sketch generation, this task aims to produce diverse, complex, and aesthetically appealing sketches, forming more imaginative depictions of familiar visual concepts, rather than generating conventional and typical representations of these concepts. This, thereby," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.306, + 0.897, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.497, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 1. The goal of creative sketch generation is to generate vivid sketches, e.g., the birds from existing methods on the left. We have proposed VQ-SGen, for high-quality sketch generation with a new vector-quantized (VQ) representation and an efficient generator. See one of our results on the right for comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.732 + ], + "angle": 0, + "content": "places high demands on the generation capabilities of automated methods. Benefiting from the advancement of diffusion models, the notable and relevant progress for sketch generation is Diffsketer [35], a novel approach for text-to-sketch generation. However, the large-scale backbone diffusion model is pre-trained on realistic images, and their resultant sketches thus have high fidelity and realism conforming to human imagination but fail to promote creativity. Therefore, dedicated methods for creative sketch generation are needed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Existing methods [4, 8] tackle creative sketch generation by producing pixels either as a whole sketch or as segmented parts, overlooking inherent and contextual relationships among individual strokes, such as their shapes and relative positions of both nearby and distant strokes. As a result, their generated sketches often have blurry local regions, and scattered or isolated strokes (see the head of the bird from DoodleFormer and DoodlerGAN in Fig. 1). A more recent work, SketchKnitter [33], formulated the problem of sketch generation as a stroke points rectification process, where diffusion models are leveraged to rearrange a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19311" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "set of stroke points from an initial chaotic state into coherent forms. However, without the concept of a stroke entity, this method performs poorly when generating complex creative sketches (see the bird in Fig. 1 from SketchKnitter)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.397 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel stroke-level sketch generation approach that treats each stroke as an independent entity [20, 32] and encodes it using a novel vector-quantized (VQ) representation. This compact, discrete representation enhances the learning process of the generative model by efficiently capturing essential stroke shape variations while minimizing redundancy. Despite the discrete nature of the representation, we observe semantic-aware clustering within the stroke code space, which lays an ideal foundation for the generator to sample new strokes. Furthermore, our bespoke generator incorporates the shape, semantic, and position into its generation process, effectively samples new sketches in the compressed VQ space, and accurately reasons out the structural and semantic relationship between strokes, resulting in sketches that are both appealing and structurally coherent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.715 + ], + "angle": 0, + "content": "Specifically, our method VQ-SGen, adopts a two-stage framework. In the first stage, we decouple the shape and location information of each stroke. This decoupling allows the vector-quantized (VQ) representation to concentrate on capturing stroke shapes, minimizing interference from positional data during the learning process. By prioritizing shape as the primary focus, the model learns a highly compact and informative, shape-oriented representation that preserves the inherent shape features of each stroke while reducing unnecessary redundancy. In the second stage, we employ an autoregressive Transformer to work with our VQ representation. The Transformer progressively integrates the stroke's semantic (i.e., the category label), shape, and location into the generation process. Through the autoregressive decoding mechanism, the Transformer captures each stroke's semantic, shape, and context relationship with neighboring strokes, ensuring that the shape and positioning are accurate and harmoniously aligned with its semantic role within the overall composition. As a result, our approach effectively avoids all the artifacts observed in previous methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.807 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our method, we have conducted experiments on the CreativeSketch [8] dataset. Comprehensive experiments including comparisons and ablation studies show that our method surpasses existing state-of-the-art techniques. In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.841 + ], + "angle": 0, + "content": "- We propose a novel stroke-level approach, \\( VQ-SGen \\), for the challenging creative sketch generation task." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We propose a new vector-quantized (VQ) stroke representation that treats each stroke as an entity, capturing essential shape features compactly and reducing redundancy to serve as the basis for stroke-level sketch generation." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "- We propose a bespoke generator, leveraging the shape, semantic, and spatial location together, achieving high-quality generation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.183 + ], + "angle": 0, + "content": "- We extensively evaluated our method, demonstrating superior performance. A user study shows that VQ-SGen consistently performs favorably against SoTA methods." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.196, + 0.655, + 0.212 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.221, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Sketch Representation Learning. Based on the data format of sketches, existing sketch representation learning methods can be broadly divided into three categories: image-based [16, 42, 43], sequence-based [1, 15, 19, 34], and graph-based [31, 37, 39] approaches, each with its own advantages and limitations. Specifically, image-based methods [16, 42, 43] use raster images as input, leveraging absolute coordinates to capture stroke proximity. However, they often struggle to learn structural information and sequential relationships among strokes effectively. Graph-based methods [31, 37, 39] face challenges in capturing detailed structural information within strokes. Sequence-based methods [15, 19, 34] utilize point sequences with relative coordinates, which enhances the encoding of individual stroke structures but makes it difficult to represent inter-stroke proximity and spatial relationships. ContextSeg [32] proposed the concept of the single stroke entity for sketch semantic segmentation. The advantages of this representation are demonstrated in their paper by comparing it with fancy Transformer [22] or common RNN [9, 15] based methods. SketchXAI [20] decoupled a single stroke into its shape, location, and order embeddings, and the concatenated stroke embedding was used to train a sketch classifier to study the explainability of human sketches. We exploit the same concept of treating a single stroke as the atomic unit, and decouple its shape with its location. Additionally and importantly, we train a vector quantized representation to build the discrete and compact code space dedicated to our generation task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Vector-Quantized VAEs. Learning representations with discrete features has been the focus of many previous works [6, 11, 30], as such representations are a natural choice for complex reasoning, planning, and predictive learning. Although using discrete latent variables in deep learning has been difficult, effective autoregressive models have been created to capture distributions over discrete variables. For instance, StrokeNUWA [26] utilizes VQ-VAE [28] to achieve a better representation of vector graphics, then generates vector graphics through large language models. Similarly, we propose to learn a VQ representation for rasterized strokes, expecting an improved generation capability powered by the compact representation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Creative Sketch Generation. Unlike traditional sketch generation tasks [38, 41], creative sketch generation [8] places greater emphasis on producing more imaginative de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19312" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.649, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.091, + 0.904, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.347, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 2. Overview of VQ-SGen. Given an input sketch, it is first divided into a sequence of strokes. In the first stage, we begin by decoupling the shape and location information of each stroke and obtain their discrete representations (Sec. 3.1). In the second stage, we use a decoder-only Gen-Transformer to predict \\(\\{(I_i, b_i, l_i)\\}\\) in an autoregressive manner (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.401, + 0.486, + 0.584 + ], + "angle": 0, + "content": "pictures of familiar visual concepts, rather than generating conventional and mundane forms of these concepts. To achieve this goal, DoodlerGAN [8] presents a component-based GAN architecture to incrementally generate each part of a creative sketch image. This approach requires training one GAN model individually for each body part in a supervised way, leading to significant computational overhead. DoodleFormer [4] employs an attention-based architecture for two-stage sketch generation. However, this approach tends to focus only on the broad, rough relationships between different parts of the sketch, which can lead to unnatural mixing of strokes from different regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.602, + 0.184, + 0.618 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Sketch and strokes. A standard sketch \\( S \\) in our context is composed of a sequence of strokes \\( \\{s_i\\}_{i=1}^N \\) and their corresponding stroke labels \\( \\{l_i\\}_{i=1}^N \\). The order of each stroke in the sequence corresponds to their default drawing order. Each stroke is represented as a rasterized image of size \\( 256 \\times 256 \\), and each label \\( l_i \\) is a one-hot vector in \\( \\mathbb{R}^C \\). In order to decouple the shape and position of a single stroke, we further process it. Formally, for each rasterized stroke image, we calculate its axis-aligned bounding box and record its center coordinate \\( (x_i, y_i) \\). The bounding box size (i.e., the height and width) as well as the center coordinate fully represent the stroke position \\( b_i = (w_i / 2, h_i / 2, x_i, y_i) \\). By factoring out the position (i.e., translating the stroke bounding box to align with the sketch image center), the stroke shape itself is presented as the translated rasterized image \\( I_i \\) without changing its scale. Considering the triplet of shape, position, and semantic of a single stroke, the new format to represent a sketch is \\( S = \\{s_i\\}_{i=1}^N = \\{(I_i, b_i, l_i)\\}_{i=1}^N \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.909, + 0.598 + ], + "angle": 0, + "content": "Sketch generation. Our goal in this paper is to generate a sequence of strokes in the format of the triplet and assemble them based on their shape and position to form the resulting sketch (see Fig. 1). Figure 2 displays an overview of our novel sketch generation method, which is comprised of two main components. Firstly, a decoupled representation learning module encodes a stroke into its vector-quantized representation (Sec. 3.1) that is highly compact and shape-aware. Secondly, a bespoke generator based on the Transformer [29] is designed to generate the stroke label, shape, and location sequentially, achieving a fine-grained control with high-fidelity resulting sketches. In the following sections, we elaborate on the details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.859, + 0.624 + ], + "angle": 0, + "content": "3.1. Vector-quantized Stroke Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Our triplet stroke formula enables separate control of stroke key properties in the generation process, how to properly encode the triplet to fit generation is the next step. Motivated by the vector quantization process [28], we aim to obtain a similar compact stroke representation in the context of the whole sketch, such that the sketch generator can efficiently and effectively sample new sketches in a compressed stroke space. Besides, we expect a decoupled representation based on our triplet formulation to obtain high shape and semantic awareness. To this end, we first convert each stroke into a latent embedding and further compress them in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Stroke latent embedding. Given the stroke image \\(I_{i}\\) we obtain the stroke latent embedding via an autoencoder, where both the encoder \\(\\mathcal{E}^s\\) and the decoder \\(\\mathcal{D}^s\\) are built with a few 2D CNNs (see supplementary for detailed network configuration). After the last layer of the encoder, we flatten its bottleneck feature to obtain the stroke latent em" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19313" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.123 + ], + "angle": 0, + "content": "bedding \\( e_i^s = \\text{flat}(\\mathcal{E}^s(\\boldsymbol{I}_i)) \\). The autoencoder is trained using the reconstruction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.131, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n s}} = \\left\\| \\boldsymbol {I} _ {i} - \\mathcal {D} ^ {s} \\left(\\mathcal {E} ^ {s} \\left(\\boldsymbol {I} _ {i}\\right)\\right) \\right\\| ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.161, + 0.483, + 0.206 + ], + "angle": 0, + "content": "To better capture the stroke shape information, we follow [2, 17, 32] to add the CoordConv layer to the network and the distance field supervision in the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.484, + 0.281 + ], + "angle": 0, + "content": "Tokenizing strokes. To achieve fine-grained control, we build separate discretized spaces for the shape and location, respectively. The reason to build a space for the shape location is that it is highly related to semantics, e.g., the heads of birds appear at a similar position with a similar size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.388 + ], + "angle": 0, + "content": "For the stroke shape space, as shown in Fig. 2 left, we build a codebook \\(D_{s}\\). Specifically, given a sketch \\(\\mathbf{S}\\), its corresponding stroke shape latent embeddings \\(\\{e_i^s\\}_{n = 1}^N\\) are fed into the encoder \\(\\mathcal{E}^f\\) to obtain the feature sequence \\(\\{z_i^s\\}_{i = 1}^N\\). We then compress it via vector quantization, i.e., clamping each feature into its nearest code in \\(D_{s}\\) with \\(V\\) codes \\(\\{c_i\\}_{i = 1}^V\\), and we record the indices of these features:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.398, + 0.483, + 0.418 + ], + "angle": 0, + "content": "\\[\nv _ {i} ^ {s} = \\operatorname {a r g m i n} _ {j \\in [ 0, V)} \\| \\boldsymbol {z} _ {i} ^ {s} - \\boldsymbol {c} _ {j} \\|. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.49 + ], + "angle": 0, + "content": "The feature sequence \\(\\{c_{v_i^s}\\}^N\\) is further sent to the decoder \\(\\mathcal{D}^f\\) to reconstruct the input latent embeddings. Both \\(\\mathcal{E}^f\\) and \\(\\mathcal{D}^f\\) are composed of several 1D CNN layers, see the supplementary for the detail structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.534 + ], + "angle": 0, + "content": "To train the vector quantization network, we follow [28] to exploit the codebook, commitment, and reconstruction losses, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.545, + 0.483, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\alpha (\\| \\pmb {z} _ {i} ^ {s} - \\mathrm {s g} [ \\pmb {c} _ {v _ {i} ^ {s}} ] \\| _ {2} ^ {2} + \\| \\mathrm {s g} [ \\pmb {z} _ {i} ^ {s} ] - \\pmb {c} _ {v _ {i} ^ {s}} \\| _ {2} ^ {2}) \\\\ + \\left\\| \\boldsymbol {z} _ {i} ^ {s} - \\mathcal {D} ^ {f} \\left(\\boldsymbol {c} _ {v _ {i} ^ {s}}\\right) \\right\\| _ {2} ^ {2}, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.462, + 0.635 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a weight, and \\(\\mathrm{sg}\\) is the stopgradient operator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.724 + ], + "angle": 0, + "content": "For the stroke location codebook learning, we use the same process as the shape codebook learning. The stroke position \\( b_{i} = (w_{i} / 2, h_{i} / 2, x_{i}, y_{i}) \\) is reconstructed by a pair of encoder \\( (\\mathcal{E}^{l}) \\) and decoder \\( (\\mathcal{D}^{l}) \\) to build the codebook \\( D_{l} \\) with V codes \\( \\{d_{i}\\}_{i=1}^{V} \\) as well. See supplementary for detailed network structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.484, + 0.782 + ], + "angle": 0, + "content": "Till now, given a stroke \\( \\boldsymbol{s}_i = (\\boldsymbol{I}_i, b_i, l_i) \\), we have the compact representation \\( (v_i^s, v_i^l, l_i) \\), corresponding to the specific code features \\( (\\boldsymbol{c}_{v_i^s}, \\boldsymbol{d}_{v_i^l}, l_i) \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.334, + 0.804 + ], + "angle": 0, + "content": "3.2. Autoregressive Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Having the VQ stroke representation, our goal in this module is to estimate a distribution over the sketch \\( S \\) so as to generate new instances. Since \\( S \\) is composed of a sequence of stroke shapes, positions, and labels, we split the generation process into two steps, i.e., 1) generate the stroke label, and 2) generate the stroke shape and position conditioned" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.903, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.245, + 0.855, + 0.26 + ], + "angle": 0, + "content": "Figure 3. Network architecture of our generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.905, + 0.293 + ], + "angle": 0, + "content": "on the stroke label. By applying the chain rule, we formulate it as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.298, + 0.905, + 0.34 + ], + "angle": 0, + "content": "\\[\np (\\boldsymbol {S}) = \\prod_ {i = 1} ^ {N} p \\left(\\boldsymbol {I} _ {i}, b _ {i}, l _ {i}\\right) = \\prod_ {i = 1} ^ {N} p \\left(v _ {i} ^ {\\mathrm {s}}, v _ {i} ^ {l} \\mid l _ {i}\\right) \\cdot p \\left(l _ {i}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.906, + 0.42 + ], + "angle": 0, + "content": "We implement Eq. (4) with two cascaded Transformer decoders, parameterized with \\(\\theta_{1}\\) and \\(\\theta_{2}\\), respectively. The first decoder samples a label for the next stroke, while the second decoder generates the shape and position conditioned on the label, which can be further formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.426, + 0.828, + 0.443 + ], + "angle": 0, + "content": "\\[\np (l _ {i}) = p \\left(l _ {i} \\mid l _ {< i}, v _ {s < } ^ {s}, v _ {l} ^ {< }; \\theta_ {1}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.439, + 0.905, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (i) \\quad P (i + l _ {i}) \\quad P (i + l _ {i}) \\quad P (i + 1) \\\\ p \\left(v _ {i} ^ {s}, v _ {i} ^ {l} \\mid l _ {i}\\right) = p \\left(v _ {i} ^ {s}, v _ {i} ^ {l} \\mid l _ {\\leq i}, v _ {< i} ^ {s}, v _ {< i} ^ {l}; \\theta_ {2}\\right). \\end{array} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Note that, we will use the codes in \\(D_{s}\\) and \\(D_{l}\\) in the Transformer calculation, instead of the original rasterized image and the quaternion for the stroke shape and position. Network Architectures. An overview of our cascaded decoders is shown in Fig. 3, where the first label Transformer \\((\\mathcal{T}^{l})\\) predicts a stroke label each time, while the second code Transformer \\((\\mathcal{T}^{c})\\) takes as input the predicted label generating both the stroke shape and position codes simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Concretely, the label Transformer \\(\\mathcal{T}^l\\) takes as input the triplet \\(\\left(\\pmb{c}_{v_i^s},\\pmb{d}_{v_i^l},l_i\\right)\\). Each element goes through its own embedding layer to get a feature, i.e., \\(\\mathrm{Emb}(\\pmb{c}_{v_i^s})\\in \\mathbb{R}^{512}\\), \\(\\mathrm{Emb}(\\pmb{d}_{v_i^l})\\in \\mathbb{R}^{512}\\), and \\(\\mathrm{Emb}(l_i)\\in \\mathbb{R}^{512}\\). The three features are simply added along the feature dimension to obtain the input feature for the attention blocks, which outputs a fused feature \\(\\mathcal{F}_{\\mathrm{fuse}}\\) after the attention calculation. \\(\\mathcal{F}_{\\mathrm{fuse}}\\) is then fed into a fully connected layer with the softmax activation to predict the one-hot probabilistic vector, which is translated into a one-hot class label \\(l_{i + 1}\\in \\mathbb{R}^C\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The code Transformer \\(\\mathcal{T}^c\\) accepts two inputs. The first one is \\(l_{i + 1}\\), which goes through an embedding layer to produce a conditional feature in \\(\\mathbb{R}^{512}\\). The second input is \\(\\mathcal{F}_{\\mathrm{fuse}}\\), which is also added with the conditional feature along the feature dimension to get the input feature for the attention block. Intuitively, the first input provides the direct semantic guidance of the current stroke, while the second input provides the shape and positional information of previous strokes. The attention blocks are followed by two separate branches, each of which outputs a one-hot probabilistic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19314" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.153 + ], + "angle": 0, + "content": "vector of dimension \\(\\mathbb{R}^V\\), indicating the indices of the shape and position codebooks. The corresponding code entries are fetched (i.e., \\(c_{v_{i+1}^s}\\) and \\(d_{v_{i+1}^t}\\)) and serve as the input to the label Transformer for the next stroke generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.185 + ], + "angle": 0, + "content": "Losses. To train the network, we minimize the negative log-likelihood loss, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.2, + 0.483, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g e n}} = - \\log p (\\boldsymbol {S}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.482, + 0.276 + ], + "angle": 0, + "content": "where we supervise the training with ground truth categorical labels (i.e., the shape label vector, shape code index vector, and the position code index vector)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.278, + 0.484, + 0.445 + ], + "angle": 0, + "content": "Training and testing strategy. We have three training stages arranged sequentially for stroke latent embedding, codebooks, and the generator. Detailed steps, hyperparameter settings, and implementation details can be found in the supplementary. For the testing time generation, starting from a special STR token, we generate novel strokes in an autoregressive manner, where the codebook entry is either decoded as the position quaternion or the stroke latent code, which is further converted into the corresponding rasterized image. The stroke triplet is finally assembled together to produce the resulting sketch (see Fig. 2, left)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.461, + 0.223, + 0.479 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Datasets. Following previous work [4, 8], we use the CreativeSketch [8] dataset in this paper. Specifically, it consists of two categories of sketches — Creative Birds (CB) and Creative Creatures (CC) — containing 8,067 and 9,097 sketches, respectively, along with part annotations. All the sketches have been processed into \\(256 \\times 256\\) images. For the stroke latent embedding learning, we use strokes from both datasets to obtain a powerful pair of \\(\\mathcal{E}^s\\) and \\(\\mathcal{D}^s\\). Other than this autoencoder, we have trained other networks separately on CB and CC datasets. See supplementary for details about pre-processing and data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Evaluation metrics. Following previous work [4, 8], we evaluate the model's performance using an Inception model [25] trained on the QuickDraw3.8M dataset [36]. We use four metrics in this paper, defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.717, + 0.483, + 0.761 + ], + "angle": 0, + "content": "- Fréchet Inception distance (FID) [10] measures the similarity between two image sets in the above Inception feature space." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.762, + 0.483, + 0.822 + ], + "angle": 0, + "content": "- Generation diversity (GD) [5] measures the average pairwise Euclidean distance between the Inception features of two subsets, reflecting the diversity of the data item in the image set." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.822, + 0.483, + 0.868 + ], + "angle": 0, + "content": "- Characteristic score (CS) measures how frequently a generated sketch is classified as Creative Bird or Creative Creature by the Inception model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.868, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- Semantic diversity score (SDS) quantifies sketch diversity based on various creature categories they represent." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.717, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.117 + ], + "angle": 0, + "content": "Table 1. Statistical comparison against existing methods on Creative Birds and Creative Creatures datasets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.215 + ], + "angle": 0, + "content": "
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
Sketchknitter [33]74.4214.230.1464.3412.340.421.32
DoodlerGAN [8]39.9516.330.6943.9414.570.551.45
DoodleFormer [4]17.4817.830.5720.4316.230.531.68
Ours15.7818.920.5317.6117.420.571.86
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.646, + 0.267 + ], + "angle": 0, + "content": "4.1. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.377 + ], + "angle": 0, + "content": "We compare our approach with three competitors, i.e., Doodleformer [4], DoodlerGAN [8], Sketchknitter [33]. Both Doodleformer and DoodlerGAN generate sketches at the pixel level, while Sketchknitter generates sketches at the stroke point sequence level. All three methods are trained using their default parameters on our dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.908, + 0.903 + ], + "angle": 0, + "content": "The quantitative comparison is presented in Tab. 1, where our method significantly outperforms others. Specifically, on the Creative Birds dataset, we obtained a notable 1.7 decrease in FID and gained a remarkable 1.09 increase in terms of GD, achieving the best performance. This indicates that our model can generate sketches that are closer to the original dataset, as well as more diverse at the same time. For the CS metric, while our approach does not achieve the highest score, it surpasses the dataset and comes closest to matching the dataset's CS metric. In contrast, DoodlerGAN and DoodleFormer attain higher scores, whereas SketchKnitter's outputs are seldom recognized as birds. For this specific metric, the higher value only means it is easier to be recognized by the Inception model, which may stem from the simplicity of the generation. The dataset's value sets a good reference. Similarly, on the Creative Creatures dataset, we achieved the best performance across all four metrics. Specifically, we achieved a 2.82 decrease in FID and a 1.19 improvement in GD compared to the best other methods. This indicates that our generated sketches are closer to the more complex sketches in the training data and exhibit greater diversity among themselves. For the CS and SDS metrics, we obtained improvements of 0.04 and 0.18, respectively. This indicates that our generated sketches are semantically closer to the categories in the training data and also suggests that our model exhibits better semantic diversity. A corresponding visual comparison is shown in Fig. 4, where the artifacts like the broken and misaligned strokes from DoodlerGAN, and the blurry local regions from DoodleFormer, are clearly presented. In contrast, our results are of high quality with complex appearance and diverse shapes. Note that, as shown in Fig. 1, SketchKnitter generates poor results on both datasets, we did not include its results in the visual comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19315" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.907, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.47, + 0.817, + 0.485 + ], + "angle": 0, + "content": "Figure 4. Visual comparison against competitors on both the Creative Birds and Creative Creatures datasets." + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.487, + 0.865, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.796, + 0.908, + 0.825 + ], + "angle": 0, + "content": "Figure 5. Sketch reconstruction results of our ablation study on codebook hyper-parameter setting. The two numbers represent the code size and the feature dimension, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.836, + 0.245, + 0.852 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To validate the effectiveness of our core technical components, we conducted experiments by removing them." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.837, + 0.889, + 0.852 + ], + "angle": 0, + "content": "Specifically, we form the following alternative solutions:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "- w/o VQ: we remove all the codebook learning-related procedures, and use the continuous stroke latent embeddings to train our generator." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19316" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.145 + ], + "angle": 0, + "content": "Table 2. The statistical comparison of our ablation study. The method name of two numbers represents the codebook hyperparameter setting, where the first is the code size, and the second is the feature dimension." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.15, + 0.481, + 0.278 + ], + "angle": 0, + "content": "
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
2048×51226.2315.340.4343.2114.510.481.52
4096×51216.9218.270.5018.4416.140.541.63
4096×102416.6718.150.5118.1216.430.551.69
w/o Tl16.5118.350.5119.2116.140.551.74
w/o VQ48.5313.340.4654.5614.020.451.36
w/o Decouple17.1418.120.5719.4216.420.561.79
Ours(8192 × 512)15.7818.920.5317.6117.420.571.86
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.291, + 0.483, + 0.351 + ], + "angle": 0, + "content": "- w/o Decouple: we do not decouple the shape and position of the stroke, and only train a single stroke codebook. The Transformer decoders take fewer inputs and the second decoder outputs only one index vector accordingly." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.351, + 0.483, + 0.382 + ], + "angle": 0, + "content": "- w/o \\(\\mathcal{T}^l\\): we simply discard the label information for each stroke, and remove the label Transformer accordingly." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.291, + 0.483, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.685 + ], + "angle": 0, + "content": "We train the three alternatives from scratch using our datasets until convergence and statistical results are shown in Tab. 2. Without the VQ representation, the generator samples new sketches from a larger continuous space, instead of a compressed discretized space, giving the worst performance in this paper. The stroke decoupling matters the generation significantly (e.g., 15.78 vs. 17.14 and 17.61 vs. 19.42 of the FID). By decoupling, the shape code can concentrate only on the shape variation, while the position code can better capture the inherent approximation information between strokes, which further facilitates the generation. By removing the label prediction, we see a slight performance degradation, giving the second-best performance in Tab. 2. The labels are not very critical, but indeed provide semantic-related information to complement the positional information (e.g., the heads appear at similar positions with similar sizes). On the other hand, it demonstrates the great feasibility of applying our method to other datasets without label information, proving our flexibility. See Sec. 5 for such an experiment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Codebook hyper-parameter setting. There are two core hyper-parameters for a codebook, i.e., the code size and the feature dimension. In the following, we have conducted experiments on the shape codebook to find the proper configuration. Similar experiments on the location codebook are presented in the supplementary. Specifically, we set the code size to be either 2,048, 4,096, or 8,192, and set the feature dimension to be either 512 or 1,024. We examine the effects of the parameter setting from both the sketch reconstruction and generation perspectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Figure 5 presents a visual comparison of sketch reconstruction quality. By increasing either the code size or the feature dimension, we observe improved fidelity with less blurry and broken strokes. Note that the row named \\( w / o \\) \\( VQ \\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.251, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Figure 6. User study results based on the five questions (a-e). Higher scores indicate a stronger preference for our method over the other approaches (i.e., DoodleFormer or Creative datasets)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.465 + ], + "angle": 0, + "content": "shows the stroke latent space reconstruction performance as in [32], which along with the input, provides a good performance reference. We chose the last configuration in this paper, as it gives the best visual reconstruction. However, it still struggles to represent strokes with significant variation, such as the ear of the creature in the third-to-last column. We further investigate the generation performance of different codebook settings. Quantitative results are listed in Tab. 2, where we observe improvement when increasing either the size of the code or the feature dimension. Our choice obtains the best result given the limited combination." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.475, + 0.637, + 0.491 + ], + "angle": 0, + "content": "4.3. User Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.739 + ], + "angle": 0, + "content": "Following [4, 8], we evaluate the performance of our method through user studies. We invited 50 participants to the test, where they were shown multiple pairs of images. Each pair consists of a sketch generated by our model and another sketch from either DoodleFormer or the datasets. For each pair, participants were asked to answer 5 single-choice questions: which one (a) is more creative? (b) looks more like a bird/creature? (c) is more likely to be drawn by a human? (d) the strokes better integrated? (e) they like the most overall? Figure 6 shows the percentage of times our method is preferred over the competing approach. It can be seen that \\( VQ-SGen \\) performs favorably against DoodleFormer for all five questions on both datasets except for question (b) on the Creative Birds dataset, which is consistent with the CS metric in Tab. 1. We are nearly tied with the creative datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.766, + 0.772 + ], + "angle": 0, + "content": "5. Discussion and Application" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.825 + ], + "angle": 0, + "content": "In this section, we mainly discuss code space exploration and conditional generation applications, while leaving other discussions, e.g., limitations, in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Code space exploration. To better understand the properties of the discrete space, we applied the UMAP algorithm [18] to project the high dimensional features of the shape codebook \\(D^{s}\\) into 2D space and visualize its distribution in Fig. 8. We further probe the emerged clustering by visu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19317" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.095, + 0.891, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Figure 7. Applications of our approach. (a) Given the class label, our method can produce corresponding sketches with rich variations. (b) Our method supports text-conditioned generation, the resulting sketches explain the input text vividly. (c) Given the initial stroke, our method can complete the whole sketch, which is favorable against competitors." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.318, + 0.468, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Figure 8. UMAP visualization of the shape codes in \\(\\mathcal{D}^s\\), with the overlaid strokes indicating the semantic aware clustering in the discrete code space." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.765 + ], + "angle": 0, + "content": "alizing typical strokes. Not too surprised, we find that the emerged clusters correspond well to semantics. For example, the five surrounding clusters are Beak, Eye, Tail, Wing, and Leg, while the inner big cluster composed of Head and Body. Although we did not exploit any semantic-related constraints during the codebook learning, it tends to implicitly group strokes with similar shape variations, thereby achieving stable and efficient code learning and space compression. As a result, this provides an ideal foundation for the generator to sample new strokes within a semantically-aware, compressed space." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Class label-to-sketch generation. Other than unconditional generation, our method can generate sketches given the class label conditions. To do this, we exploit a subset of the QuickDraw dataset [9]. The subset contains 20 classes, and stroke labels are NOT available in this dataset. Our adaptation is pretty simple and straightforward - i) removing the label Transformer (i.e., the variant of w/o \\(\\mathcal{T}^l\\) in Sec. 4.2), and ii) replacing the special STR token with the class index token (e.g., the sinusoidal encoding). We only" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.412 + ], + "angle": 0, + "content": "re-train the generator on the QuickDraw subset until convergence. At inference time, given a class label, our method successfully produces high-quality results, as shown in Fig. 7(a). Note that, the experiment also demonstrates our flexibility and superior performance when working on datasets without stroke label information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.909, + 0.581 + ], + "angle": 0, + "content": "Text-to-sketch generation. There are text descriptions for each sketch in the CB and CC datasets. We thus adapt our method to perform text-to-sketch generation. To this end, we first embed the text description using the pre-trained CLIP [21] model, and then replace the \\(STR\\) token with the obtained text token, similarly as in the above label-to-sketch generation task. Fig. 7(b) shows a few novel examples of our generated sketches, which conform well to input texts. Pay attention to the novel visual concept, e.g., the 'Hug me' sketch, and the 'Smiling' sketch, which is rarely seen in the traditional sketch generation task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Sketch completion. Given an initial stroke, our model can perform sketch completion, i.e., stroke conditional generation. We achieve this by feeding the initial stroke as our first stroke in the generation process. Fig. 7(c) presents a visual comparison against SoTA methods on this task. As can be seen, DoodlerGAN produced the worst results, with most sketches lacking a clear structure. DoodleFormer performed slightly better, but the generated sketches are often blurry and include ghost strokes (e.g., a few tail-like strokes of the second example). With the help of our VQ representation and our generator, our results are both creative and structurally coherent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.783, + 0.634, + 0.799 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we have presented a novel approach for creative sketch generation. Our method can effectively create sketches with an appealing appearance and coherent structures. We believe that stroke VQ representations can be used in other sketch-related tasks and hope to inspire future research in sparse stroke representation learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19318" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.184 + ], + "angle": 0, + "content": "Acknowledgment. The authors would like to thank Oisin Mac Aodha and Ankan Bhunia for proofreading the early draft and the valuable discussions, Haocheng Yuan and Yi Yang for helping prepare comparisons with Diffsketcher, and the support of the HPC Platform at ShanghaiTech University. CL was supported by a gift from Adobe." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.197, + 0.188, + 0.212 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.222, + 0.486, + 0.318 + ], + "angle": 0, + "content": "[1] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12565-12574, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.321, + 0.485, + 0.417 + ], + "angle": 0, + "content": "[2] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12565-12574, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.419, + 0.485, + 0.502 + ], + "angle": 0, + "content": "[3] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10997-11008, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.504, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[4] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. In European Conference on Computer Vision, pages 338-355. Springer, 2022. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.588, + 0.485, + 0.644 + ], + "angle": 0, + "content": "[5] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Aisketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.645, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[6] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.716, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[7] Chenjian Gao, Xin Wang, Qian Yu, Lu Sheng, Jing Zhang, Xiaoguang Han, Yi-Zhe Song, and Dong Xu. 3d reconstruction from a single sketch via view-dependent depth sampling. IEEE transactions on pattern analysis and machine intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.788, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[8] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creativ sketch generation. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.83, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] David Ha and Douglas Eck. A neural representation of sketch drawings. arXiv preprint arXiv:1704.03477, 2017. 2,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.222, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.133 + ], + "angle": 0, + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[11] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[12] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-image diffusion models are great sketch-photo matchmakers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16826-16837, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.342 + ], + "angle": 0, + "content": "[13] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16859-16869, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.424 + ], + "angle": 0, + "content": "[14] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16859-16869, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[15] Ke Li, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Honggang Zhang. Toward deep universal sketch perceptual grouper. IEEE Transactions on Image Processing, 28(7):3219-3231, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[16] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning. IEEE computer graphics and applications, 39(2):38-51, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[17] Rosanne Liu, Joel Lehman, Piero Molino, Felipe Petroski Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An intriguing failing of convolutional neural networks and the coordconv solution. Advances in neural information processing systems, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.906, + 0.637 + ], + "angle": 0, + "content": "[18] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[19] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnN for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[20] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23327-23337, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14153-14162, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19319" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[23] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8504–8513, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[24] Aneeshan Sain, Pinaki Nath Chowdhury, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Freeview sketching: View-aware fine-grained sketch-based image retrieval. In European Conference on Computer Vision, pages 145-162. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[25] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[26] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, et al. Strokenuwa: Tokenizing strokes for vector graphic synthesis. arXiv preprint arXiv:2401.17093, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[27] Chaitat Utintu, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Sketchdeco: Decorating b&w sketches with colour. arXiv preprint arXiv:2405.18716, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[28] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[30] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[31] Fei Wang, Shujin Lin, Hanhui Li, Hefeng Wu, Tie Cai, Xiaonan Luo, and Ruomei Wang. Multi-column point-cnn for sketch segmentation. Neurocomputing, 392:50-59, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[32] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3679-3688, 2024. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[33] Qiang Wang, Haoge Deng, Yonggang Qi, Da Li, and Yi-Zhe Song. Sketchknitter: Vectorized sketch generation with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[34] Xingyuan Wu, Yonggang Qi, Jun Liu, and Jie Yang. Sketch-segnet: A rnn model for labeling sketch strokes. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[35] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36:15869-15889, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "sketch: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(1):285-312, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[37] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. Sketchgnn: Semantic sketch segmentation with graph neural networks. ACM Transactions on Graphics (TOG), 40(3):1-13, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[38] Ningyuan Zheng, Yifan Jiang, and Dingjiang Huang. Strokenet: A neural painting environment. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[39] Yixiao Zheng, Jiyang Xie, Aneeshan Sain, Yi-Zhe Song, and Zhanyu Ma. Sketch-segformer: Transformer-based segmentation for figurative and creative sketches. IEEE Transactions on Image Processing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[40] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[41] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to doodle with stroke demonstrations and deep q-networks. In BMVC, page 13, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.461 + ], + "angle": 0, + "content": "[42] Xianyi Zhu, Yi Xiao, and Yan Zheng. Part-level sketch segmentation and labeling using dual-cnn. In Neural Information Processing: 25th International Conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16, 2018, Proceedings, Part I 25, pages 374-384. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[43] Xianyi Zhu, Yi Xiao, and Yan Zheng. 2d freehand sketch labeling using cnn and crf. Multimedia Tools and Applications, 79(1-2):1585–1602, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19320" + } + ] +] \ No newline at end of file diff --git a/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_origin.pdf b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1da4f2111cfe7a9d219017f7e4c2497468db0581 --- /dev/null +++ b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/27ec5cff-30ea-4eda-ad45-f2004607bb53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dff6b3077d00d9efe44663ff4e315e05e53cbf7faa4e38bf504ac217b59188a +size 4080103 diff --git a/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/full.md b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b6c5e69ab1040e73431168250d851aa5b34189f --- /dev/null +++ b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/full.md @@ -0,0 +1,270 @@ +# VQ-SGen: A Vector Quantized Stroke Representation for Creative Sketch Generation + +Jiawei Wang $^{1,2}$ Zhiming Cui $^{2}$ Changjian Li $^{1}$ + +1The University of Edinburgh 2ShanghaiTech University + +https://enigma-li.github.io/projects/VQ-SGen/VQ-SGen.html + +# Abstract + +This paper presents VQ-SGen, a novel algorithm for high-quality creative sketch generation. Recent approaches have framed the task as pixel-based generation either as a whole or part-by-part, neglecting the intrinsic and contextual relationships among individual strokes, such as the shape and spatial positioning of both proximal and distant strokes. To overcome these limitations, we propose treating each stroke within a sketch as an entity and introducing a vector-quantized (VQ) stroke representation for fine-grained sketch generation. Our method follows a two-stage framework - in stage one, we decouple each stroke's shape and location information to ensure the VQ representation prioritizes stroke shape learning. In stage two, we feed the precise and compact representation into an auto-decoding Transformer to incorporate stroke semantics, positions, and shapes into the generation process. By utilizing tokenized stroke representation, our approach generates strokes with high fidelity and facilitates novel applications, such as text or class label conditioned generation and sketch completion. Comprehensive experiments demonstrate our method surpasses existing state-of-the-art techniques on the CreativeSketch dataset, underscoring its effectiveness. + +# 1. Introduction + +Sketches are an intuitive mechanism widely used by humans to communicate ideas, convey concepts, and express creativity. Over the past decades, significant research has been conducted on sketch-related tasks, including retrieval [3, 13, 14, 23, 24], semantic segmentation [32, 40], and generation [4, 7, 8, 12, 27, 33]. In this paper, we focus on creative sketch generation, a novel task introduced by Ge et al. [8]. Unlike traditional sketch generation, this task aims to produce diverse, complex, and aesthetically appealing sketches, forming more imaginative depictions of familiar visual concepts, rather than generating conventional and typical representations of these concepts. This, thereby, + +![](images/fde780d9fcfd323186cb4e8f6dbfd4abb7f80844e9d4326deadb9de2cf376744.jpg) +Figure 1. The goal of creative sketch generation is to generate vivid sketches, e.g., the birds from existing methods on the left. We have proposed VQ-SGen, for high-quality sketch generation with a new vector-quantized (VQ) representation and an efficient generator. See one of our results on the right for comparison. + +places high demands on the generation capabilities of automated methods. Benefiting from the advancement of diffusion models, the notable and relevant progress for sketch generation is Diffsketer [35], a novel approach for text-to-sketch generation. However, the large-scale backbone diffusion model is pre-trained on realistic images, and their resultant sketches thus have high fidelity and realism conforming to human imagination but fail to promote creativity. Therefore, dedicated methods for creative sketch generation are needed. + +Existing methods [4, 8] tackle creative sketch generation by producing pixels either as a whole sketch or as segmented parts, overlooking inherent and contextual relationships among individual strokes, such as their shapes and relative positions of both nearby and distant strokes. As a result, their generated sketches often have blurry local regions, and scattered or isolated strokes (see the head of the bird from DoodleFormer and DoodlerGAN in Fig. 1). A more recent work, SketchKnitter [33], formulated the problem of sketch generation as a stroke points rectification process, where diffusion models are leveraged to rearrange a + +set of stroke points from an initial chaotic state into coherent forms. However, without the concept of a stroke entity, this method performs poorly when generating complex creative sketches (see the bird in Fig. 1 from SketchKnitter). + +To address these limitations, we propose a novel stroke-level sketch generation approach that treats each stroke as an independent entity [20, 32] and encodes it using a novel vector-quantized (VQ) representation. This compact, discrete representation enhances the learning process of the generative model by efficiently capturing essential stroke shape variations while minimizing redundancy. Despite the discrete nature of the representation, we observe semantic-aware clustering within the stroke code space, which lays an ideal foundation for the generator to sample new strokes. Furthermore, our bespoke generator incorporates the shape, semantic, and position into its generation process, effectively samples new sketches in the compressed VQ space, and accurately reasons out the structural and semantic relationship between strokes, resulting in sketches that are both appealing and structurally coherent. + +Specifically, our method VQ-SGen, adopts a two-stage framework. In the first stage, we decouple the shape and location information of each stroke. This decoupling allows the vector-quantized (VQ) representation to concentrate on capturing stroke shapes, minimizing interference from positional data during the learning process. By prioritizing shape as the primary focus, the model learns a highly compact and informative, shape-oriented representation that preserves the inherent shape features of each stroke while reducing unnecessary redundancy. In the second stage, we employ an autoregressive Transformer to work with our VQ representation. The Transformer progressively integrates the stroke's semantic (i.e., the category label), shape, and location into the generation process. Through the autoregressive decoding mechanism, the Transformer captures each stroke's semantic, shape, and context relationship with neighboring strokes, ensuring that the shape and positioning are accurate and harmoniously aligned with its semantic role within the overall composition. As a result, our approach effectively avoids all the artifacts observed in previous methods. + +To demonstrate the effectiveness of our method, we have conducted experiments on the CreativeSketch [8] dataset. Comprehensive experiments including comparisons and ablation studies show that our method surpasses existing state-of-the-art techniques. In summary, our main contributions are as follows: + +- We propose a novel stroke-level approach, $VQ-SGen$ , for the challenging creative sketch generation task. +- We propose a new vector-quantized (VQ) stroke representation that treats each stroke as an entity, capturing essential shape features compactly and reducing redundancy to serve as the basis for stroke-level sketch generation. + +- We propose a bespoke generator, leveraging the shape, semantic, and spatial location together, achieving high-quality generation. +- We extensively evaluated our method, demonstrating superior performance. A user study shows that VQ-SGen consistently performs favorably against SoTA methods. + +# 2. Related Work + +Sketch Representation Learning. Based on the data format of sketches, existing sketch representation learning methods can be broadly divided into three categories: image-based [16, 42, 43], sequence-based [1, 15, 19, 34], and graph-based [31, 37, 39] approaches, each with its own advantages and limitations. Specifically, image-based methods [16, 42, 43] use raster images as input, leveraging absolute coordinates to capture stroke proximity. However, they often struggle to learn structural information and sequential relationships among strokes effectively. Graph-based methods [31, 37, 39] face challenges in capturing detailed structural information within strokes. Sequence-based methods [15, 19, 34] utilize point sequences with relative coordinates, which enhances the encoding of individual stroke structures but makes it difficult to represent inter-stroke proximity and spatial relationships. ContextSeg [32] proposed the concept of the single stroke entity for sketch semantic segmentation. The advantages of this representation are demonstrated in their paper by comparing it with fancy Transformer [22] or common RNN [9, 15] based methods. SketchXAI [20] decoupled a single stroke into its shape, location, and order embeddings, and the concatenated stroke embedding was used to train a sketch classifier to study the explainability of human sketches. We exploit the same concept of treating a single stroke as the atomic unit, and decouple its shape with its location. Additionally and importantly, we train a vector quantized representation to build the discrete and compact code space dedicated to our generation task. + +Vector-Quantized VAEs. Learning representations with discrete features has been the focus of many previous works [6, 11, 30], as such representations are a natural choice for complex reasoning, planning, and predictive learning. Although using discrete latent variables in deep learning has been difficult, effective autoregressive models have been created to capture distributions over discrete variables. For instance, StrokeNUWA [26] utilizes VQ-VAE [28] to achieve a better representation of vector graphics, then generates vector graphics through large language models. Similarly, we propose to learn a VQ representation for rasterized strokes, expecting an improved generation capability powered by the compact representation. + +Creative Sketch Generation. Unlike traditional sketch generation tasks [38, 41], creative sketch generation [8] places greater emphasis on producing more imaginative de + +![](images/45510d92dace14c50ed3f171d04a8f5e2ab66a7c01c4c333e51ecb8a3c0f4391.jpg) +Figure 2. Overview of VQ-SGen. Given an input sketch, it is first divided into a sequence of strokes. In the first stage, we begin by decoupling the shape and location information of each stroke and obtain their discrete representations (Sec. 3.1). In the second stage, we use a decoder-only Gen-Transformer to predict $\{(I_i, b_i, l_i)\}$ in an autoregressive manner (Sec. 3.2). + +![](images/03231a5eebdc93b56ced6eaf36c6cbe34db554dfdd128fa50830a761df12761d.jpg) + +pictures of familiar visual concepts, rather than generating conventional and mundane forms of these concepts. To achieve this goal, DoodlerGAN [8] presents a component-based GAN architecture to incrementally generate each part of a creative sketch image. This approach requires training one GAN model individually for each body part in a supervised way, leading to significant computational overhead. DoodleFormer [4] employs an attention-based architecture for two-stage sketch generation. However, this approach tends to focus only on the broad, rough relationships between different parts of the sketch, which can lead to unnatural mixing of strokes from different regions. + +# 3. Method + +Sketch and strokes. A standard sketch $S$ in our context is composed of a sequence of strokes $\{s_i\}_{i=1}^N$ and their corresponding stroke labels $\{l_i\}_{i=1}^N$ . The order of each stroke in the sequence corresponds to their default drawing order. Each stroke is represented as a rasterized image of size $256 \times 256$ , and each label $l_i$ is a one-hot vector in $\mathbb{R}^C$ . In order to decouple the shape and position of a single stroke, we further process it. Formally, for each rasterized stroke image, we calculate its axis-aligned bounding box and record its center coordinate $(x_i, y_i)$ . The bounding box size (i.e., the height and width) as well as the center coordinate fully represent the stroke position $b_i = (w_i / 2, h_i / 2, x_i, y_i)$ . By factoring out the position (i.e., translating the stroke bounding box to align with the sketch image center), the stroke shape itself is presented as the translated rasterized image $I_i$ without changing its scale. Considering the triplet of shape, position, and semantic of a single stroke, the new format to represent a sketch is $S = \{s_i\}_{i=1}^N = \{(I_i, b_i, l_i)\}_{i=1}^N$ . + +Sketch generation. Our goal in this paper is to generate a sequence of strokes in the format of the triplet and assemble them based on their shape and position to form the resulting sketch (see Fig. 1). Figure 2 displays an overview of our novel sketch generation method, which is comprised of two main components. Firstly, a decoupled representation learning module encodes a stroke into its vector-quantized representation (Sec. 3.1) that is highly compact and shape-aware. Secondly, a bespoke generator based on the Transformer [29] is designed to generate the stroke label, shape, and location sequentially, achieving a fine-grained control with high-fidelity resulting sketches. In the following sections, we elaborate on the details. + +# 3.1. Vector-quantized Stroke Representation + +Our triplet stroke formula enables separate control of stroke key properties in the generation process, how to properly encode the triplet to fit generation is the next step. Motivated by the vector quantization process [28], we aim to obtain a similar compact stroke representation in the context of the whole sketch, such that the sketch generator can efficiently and effectively sample new sketches in a compressed stroke space. Besides, we expect a decoupled representation based on our triplet formulation to obtain high shape and semantic awareness. To this end, we first convert each stroke into a latent embedding and further compress them in the latent space. + +Stroke latent embedding. Given the stroke image $I_{i}$ we obtain the stroke latent embedding via an autoencoder, where both the encoder $\mathcal{E}^s$ and the decoder $\mathcal{D}^s$ are built with a few 2D CNNs (see supplementary for detailed network configuration). After the last layer of the encoder, we flatten its bottleneck feature to obtain the stroke latent em + +bedding $e_i^s = \text{flat}(\mathcal{E}^s(\boldsymbol{I}_i))$ . The autoencoder is trained using the reconstruction loss: + +$$ +\mathcal {L} _ {\text {r e c o n s}} = \left\| \boldsymbol {I} _ {i} - \mathcal {D} ^ {s} \left(\mathcal {E} ^ {s} \left(\boldsymbol {I} _ {i}\right)\right) \right\| ^ {2}. \tag {1} +$$ + +To better capture the stroke shape information, we follow [2, 17, 32] to add the CoordConv layer to the network and the distance field supervision in the training process. + +Tokenizing strokes. To achieve fine-grained control, we build separate discretized spaces for the shape and location, respectively. The reason to build a space for the shape location is that it is highly related to semantics, e.g., the heads of birds appear at a similar position with a similar size. + +For the stroke shape space, as shown in Fig. 2 left, we build a codebook $D_{s}$ . Specifically, given a sketch $\mathbf{S}$ , its corresponding stroke shape latent embeddings $\{e_i^s\}_{n = 1}^N$ are fed into the encoder $\mathcal{E}^f$ to obtain the feature sequence $\{z_i^s\}_{i = 1}^N$ . We then compress it via vector quantization, i.e., clamping each feature into its nearest code in $D_{s}$ with $V$ codes $\{c_i\}_{i = 1}^V$ , and we record the indices of these features: + +$$ +v _ {i} ^ {s} = \operatorname {a r g m i n} _ {j \in [ 0, V)} \| \boldsymbol {z} _ {i} ^ {s} - \boldsymbol {c} _ {j} \|. \tag {2} +$$ + +The feature sequence $\{c_{v_i^s}\}^N$ is further sent to the decoder $\mathcal{D}^f$ to reconstruct the input latent embeddings. Both $\mathcal{E}^f$ and $\mathcal{D}^f$ are composed of several 1D CNN layers, see the supplementary for the detail structure. + +To train the vector quantization network, we follow [28] to exploit the codebook, commitment, and reconstruction losses, defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {V Q}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \alpha (\| \pmb {z} _ {i} ^ {s} - \mathrm {s g} [ \pmb {c} _ {v _ {i} ^ {s}} ] \| _ {2} ^ {2} + \| \mathrm {s g} [ \pmb {z} _ {i} ^ {s} ] - \pmb {c} _ {v _ {i} ^ {s}} \| _ {2} ^ {2}) \\ + \left\| \boldsymbol {z} _ {i} ^ {s} - \mathcal {D} ^ {f} \left(\boldsymbol {c} _ {v _ {i} ^ {s}}\right) \right\| _ {2} ^ {2}, \tag {3} \\ \end{array} +$$ + +where $\alpha$ is a weight, and $\mathrm{sg}$ is the stopgradient operator. + +For the stroke location codebook learning, we use the same process as the shape codebook learning. The stroke position $b_{i} = (w_{i} / 2, h_{i} / 2, x_{i}, y_{i})$ is reconstructed by a pair of encoder $(\mathcal{E}^{l})$ and decoder $(\mathcal{D}^{l})$ to build the codebook $D_{l}$ with V codes $\{d_{i}\}_{i=1}^{V}$ as well. See supplementary for detailed network structures. + +Till now, given a stroke $\boldsymbol{s}_i = (\boldsymbol{I}_i, b_i, l_i)$ , we have the compact representation $(v_i^s, v_i^l, l_i)$ , corresponding to the specific code features $(\boldsymbol{c}_{v_i^s}, \boldsymbol{d}_{v_i^l}, l_i)$ . + +# 3.2. Autoregressive Generation + +Having the VQ stroke representation, our goal in this module is to estimate a distribution over the sketch $S$ so as to generate new instances. Since $S$ is composed of a sequence of stroke shapes, positions, and labels, we split the generation process into two steps, i.e., 1) generate the stroke label, and 2) generate the stroke shape and position conditioned + +![](images/db616db51cc3c6e33f3f5842d8e3d8015c9eb5e78bb230fccab4170fa3120fed.jpg) +Figure 3. Network architecture of our generator. + +on the stroke label. By applying the chain rule, we formulate it as: + +$$ +p (\boldsymbol {S}) = \prod_ {i = 1} ^ {N} p \left(\boldsymbol {I} _ {i}, b _ {i}, l _ {i}\right) = \prod_ {i = 1} ^ {N} p \left(v _ {i} ^ {\mathrm {s}}, v _ {i} ^ {l} \mid l _ {i}\right) \cdot p \left(l _ {i}\right). \tag {4} +$$ + +We implement Eq. (4) with two cascaded Transformer decoders, parameterized with $\theta_{1}$ and $\theta_{2}$ , respectively. The first decoder samples a label for the next stroke, while the second decoder generates the shape and position conditioned on the label, which can be further formulated as: + +$$ +p (l _ {i}) = p \left(l _ {i} \mid l _ {< i}, v _ {s < } ^ {s}, v _ {l} ^ {< }; \theta_ {1}\right), +$$ + +$$ +\begin{array}{l} P (i) \quad P (i + l _ {i}) \quad P (i + l _ {i}) \quad P (i + 1) \\ p \left(v _ {i} ^ {s}, v _ {i} ^ {l} \mid l _ {i}\right) = p \left(v _ {i} ^ {s}, v _ {i} ^ {l} \mid l _ {\leq i}, v _ {< i} ^ {s}, v _ {< i} ^ {l}; \theta_ {2}\right). \end{array} \tag {5} +$$ + +Note that, we will use the codes in $D_{s}$ and $D_{l}$ in the Transformer calculation, instead of the original rasterized image and the quaternion for the stroke shape and position. Network Architectures. An overview of our cascaded decoders is shown in Fig. 3, where the first label Transformer $(\mathcal{T}^{l})$ predicts a stroke label each time, while the second code Transformer $(\mathcal{T}^{c})$ takes as input the predicted label generating both the stroke shape and position codes simultaneously. + +Concretely, the label Transformer $\mathcal{T}^l$ takes as input the triplet $\left(\pmb{c}_{v_i^s},\pmb{d}_{v_i^l},l_i\right)$ . Each element goes through its own embedding layer to get a feature, i.e., $\mathrm{Emb}(\pmb{c}_{v_i^s})\in \mathbb{R}^{512}$ , $\mathrm{Emb}(\pmb{d}_{v_i^l})\in \mathbb{R}^{512}$ , and $\mathrm{Emb}(l_i)\in \mathbb{R}^{512}$ . The three features are simply added along the feature dimension to obtain the input feature for the attention blocks, which outputs a fused feature $\mathcal{F}_{\mathrm{fuse}}$ after the attention calculation. $\mathcal{F}_{\mathrm{fuse}}$ is then fed into a fully connected layer with the softmax activation to predict the one-hot probabilistic vector, which is translated into a one-hot class label $l_{i + 1}\in \mathbb{R}^C$ . + +The code Transformer $\mathcal{T}^c$ accepts two inputs. The first one is $l_{i + 1}$ , which goes through an embedding layer to produce a conditional feature in $\mathbb{R}^{512}$ . The second input is $\mathcal{F}_{\mathrm{fuse}}$ , which is also added with the conditional feature along the feature dimension to get the input feature for the attention block. Intuitively, the first input provides the direct semantic guidance of the current stroke, while the second input provides the shape and positional information of previous strokes. The attention blocks are followed by two separate branches, each of which outputs a one-hot probabilistic + +vector of dimension $\mathbb{R}^V$ , indicating the indices of the shape and position codebooks. The corresponding code entries are fetched (i.e., $c_{v_{i+1}^s}$ and $d_{v_{i+1}^t}$ ) and serve as the input to the label Transformer for the next stroke generation. + +Losses. To train the network, we minimize the negative log-likelihood loss, defined as: + +$$ +\mathcal {L} _ {\text {g e n}} = - \log p (\boldsymbol {S}), \tag {6} +$$ + +where we supervise the training with ground truth categorical labels (i.e., the shape label vector, shape code index vector, and the position code index vector). + +Training and testing strategy. We have three training stages arranged sequentially for stroke latent embedding, codebooks, and the generator. Detailed steps, hyperparameter settings, and implementation details can be found in the supplementary. For the testing time generation, starting from a special STR token, we generate novel strokes in an autoregressive manner, where the codebook entry is either decoded as the position quaternion or the stroke latent code, which is further converted into the corresponding rasterized image. The stroke triplet is finally assembled together to produce the resulting sketch (see Fig. 2, left). + +# 4. Experiments + +Datasets. Following previous work [4, 8], we use the CreativeSketch [8] dataset in this paper. Specifically, it consists of two categories of sketches — Creative Birds (CB) and Creative Creatures (CC) — containing 8,067 and 9,097 sketches, respectively, along with part annotations. All the sketches have been processed into $256 \times 256$ images. For the stroke latent embedding learning, we use strokes from both datasets to obtain a powerful pair of $\mathcal{E}^s$ and $\mathcal{D}^s$ . Other than this autoencoder, we have trained other networks separately on CB and CC datasets. See supplementary for details about pre-processing and data augmentation. + +Evaluation metrics. Following previous work [4, 8], we evaluate the model's performance using an Inception model [25] trained on the QuickDraw3.8M dataset [36]. We use four metrics in this paper, defined as follows: + +- Fréchet Inception distance (FID) [10] measures the similarity between two image sets in the above Inception feature space. +- Generation diversity (GD) [5] measures the average pairwise Euclidean distance between the Inception features of two subsets, reflecting the diversity of the data item in the image set. +- Characteristic score (CS) measures how frequently a generated sketch is classified as Creative Bird or Creative Creature by the Inception model. +- Semantic diversity score (SDS) quantifies sketch diversity based on various creature categories they represent. + +Table 1. Statistical comparison against existing methods on Creative Birds and Creative Creatures datasets. + +
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
Sketchknitter [33]74.4214.230.1464.3412.340.421.32
DoodlerGAN [8]39.9516.330.6943.9414.570.551.45
DoodleFormer [4]17.4817.830.5720.4316.230.531.68
Ours15.7818.920.5317.6117.420.571.86
+ +# 4.1. Comparison + +We compare our approach with three competitors, i.e., Doodleformer [4], DoodlerGAN [8], Sketchknitter [33]. Both Doodleformer and DoodlerGAN generate sketches at the pixel level, while Sketchknitter generates sketches at the stroke point sequence level. All three methods are trained using their default parameters on our dataset. + +The quantitative comparison is presented in Tab. 1, where our method significantly outperforms others. Specifically, on the Creative Birds dataset, we obtained a notable 1.7 decrease in FID and gained a remarkable 1.09 increase in terms of GD, achieving the best performance. This indicates that our model can generate sketches that are closer to the original dataset, as well as more diverse at the same time. For the CS metric, while our approach does not achieve the highest score, it surpasses the dataset and comes closest to matching the dataset's CS metric. In contrast, DoodlerGAN and DoodleFormer attain higher scores, whereas SketchKnitter's outputs are seldom recognized as birds. For this specific metric, the higher value only means it is easier to be recognized by the Inception model, which may stem from the simplicity of the generation. The dataset's value sets a good reference. Similarly, on the Creative Creatures dataset, we achieved the best performance across all four metrics. Specifically, we achieved a 2.82 decrease in FID and a 1.19 improvement in GD compared to the best other methods. This indicates that our generated sketches are closer to the more complex sketches in the training data and exhibit greater diversity among themselves. For the CS and SDS metrics, we obtained improvements of 0.04 and 0.18, respectively. This indicates that our generated sketches are semantically closer to the categories in the training data and also suggests that our model exhibits better semantic diversity. A corresponding visual comparison is shown in Fig. 4, where the artifacts like the broken and misaligned strokes from DoodlerGAN, and the blurry local regions from DoodleFormer, are clearly presented. In contrast, our results are of high quality with complex appearance and diverse shapes. Note that, as shown in Fig. 1, SketchKnitter generates poor results on both datasets, we did not include its results in the visual comparison. + +![](images/87506e4a5b27e9f093eb519df6e3d51362a85b5811b17e7e09a20ad172631e9e.jpg) +Figure 4. Visual comparison against competitors on both the Creative Birds and Creative Creatures datasets. + +![](images/d700b703190675cf859bc885ffaca268059af48ee1f0cdfdb18399f4f7daff58.jpg) +Figure 5. Sketch reconstruction results of our ablation study on codebook hyper-parameter setting. The two numbers represent the code size and the feature dimension, respectively. + +# 4.2. Ablation Study + +To validate the effectiveness of our core technical components, we conducted experiments by removing them. + +Specifically, we form the following alternative solutions: + +- w/o VQ: we remove all the codebook learning-related procedures, and use the continuous stroke latent embeddings to train our generator. + +Table 2. The statistical comparison of our ablation study. The method name of two numbers represents the codebook hyperparameter setting, where the first is the code size, and the second is the feature dimension. + +
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
2048×51226.2315.340.4343.2114.510.481.52
4096×51216.9218.270.5018.4416.140.541.63
4096×102416.6718.150.5118.1216.430.551.69
w/o Tl16.5118.350.5119.2116.140.551.74
w/o VQ48.5313.340.4654.5614.020.451.36
w/o Decouple17.1418.120.5719.4216.420.561.79
Ours(8192 × 512)15.7818.920.5317.6117.420.571.86
+ +- w/o Decouple: we do not decouple the shape and position of the stroke, and only train a single stroke codebook. The Transformer decoders take fewer inputs and the second decoder outputs only one index vector accordingly. +- w/o $\mathcal{T}^l$ : we simply discard the label information for each stroke, and remove the label Transformer accordingly. + +We train the three alternatives from scratch using our datasets until convergence and statistical results are shown in Tab. 2. Without the VQ representation, the generator samples new sketches from a larger continuous space, instead of a compressed discretized space, giving the worst performance in this paper. The stroke decoupling matters the generation significantly (e.g., 15.78 vs. 17.14 and 17.61 vs. 19.42 of the FID). By decoupling, the shape code can concentrate only on the shape variation, while the position code can better capture the inherent approximation information between strokes, which further facilitates the generation. By removing the label prediction, we see a slight performance degradation, giving the second-best performance in Tab. 2. The labels are not very critical, but indeed provide semantic-related information to complement the positional information (e.g., the heads appear at similar positions with similar sizes). On the other hand, it demonstrates the great feasibility of applying our method to other datasets without label information, proving our flexibility. See Sec. 5 for such an experiment. + +Codebook hyper-parameter setting. There are two core hyper-parameters for a codebook, i.e., the code size and the feature dimension. In the following, we have conducted experiments on the shape codebook to find the proper configuration. Similar experiments on the location codebook are presented in the supplementary. Specifically, we set the code size to be either 2,048, 4,096, or 8,192, and set the feature dimension to be either 512 or 1,024. We examine the effects of the parameter setting from both the sketch reconstruction and generation perspectives. + +Figure 5 presents a visual comparison of sketch reconstruction quality. By increasing either the code size or the feature dimension, we observe improved fidelity with less blurry and broken strokes. Note that the row named $w / o$ $VQ$ + +![](images/937c6bef050e1477f8016fb2885ec6a44e7b457483217d5d5022cac65efd059a.jpg) +Figure 6. User study results based on the five questions (a-e). Higher scores indicate a stronger preference for our method over the other approaches (i.e., DoodleFormer or Creative datasets). + +shows the stroke latent space reconstruction performance as in [32], which along with the input, provides a good performance reference. We chose the last configuration in this paper, as it gives the best visual reconstruction. However, it still struggles to represent strokes with significant variation, such as the ear of the creature in the third-to-last column. We further investigate the generation performance of different codebook settings. Quantitative results are listed in Tab. 2, where we observe improvement when increasing either the size of the code or the feature dimension. Our choice obtains the best result given the limited combination. + +# 4.3. User Study + +Following [4, 8], we evaluate the performance of our method through user studies. We invited 50 participants to the test, where they were shown multiple pairs of images. Each pair consists of a sketch generated by our model and another sketch from either DoodleFormer or the datasets. For each pair, participants were asked to answer 5 single-choice questions: which one (a) is more creative? (b) looks more like a bird/creature? (c) is more likely to be drawn by a human? (d) the strokes better integrated? (e) they like the most overall? Figure 6 shows the percentage of times our method is preferred over the competing approach. It can be seen that $VQ-SGen$ performs favorably against DoodleFormer for all five questions on both datasets except for question (b) on the Creative Birds dataset, which is consistent with the CS metric in Tab. 1. We are nearly tied with the creative datasets. + +# 5. Discussion and Application + +In this section, we mainly discuss code space exploration and conditional generation applications, while leaving other discussions, e.g., limitations, in the supplementary. + +Code space exploration. To better understand the properties of the discrete space, we applied the UMAP algorithm [18] to project the high dimensional features of the shape codebook $D^{s}$ into 2D space and visualize its distribution in Fig. 8. We further probe the emerged clustering by visu + +![](images/f5a9fca7875781b8283dbd3a12af22d95e4806242a8f23e98e7031c599fc0e13.jpg) +Figure 7. Applications of our approach. (a) Given the class label, our method can produce corresponding sketches with rich variations. (b) Our method supports text-conditioned generation, the resulting sketches explain the input text vividly. (c) Given the initial stroke, our method can complete the whole sketch, which is favorable against competitors. + +![](images/79bf3d42d6bf35e06c5f2f66185f4638c939ef14a5ec522e1dd5700192dfd3a0.jpg) +Figure 8. UMAP visualization of the shape codes in $\mathcal{D}^s$ , with the overlaid strokes indicating the semantic aware clustering in the discrete code space. + +alizing typical strokes. Not too surprised, we find that the emerged clusters correspond well to semantics. For example, the five surrounding clusters are Beak, Eye, Tail, Wing, and Leg, while the inner big cluster composed of Head and Body. Although we did not exploit any semantic-related constraints during the codebook learning, it tends to implicitly group strokes with similar shape variations, thereby achieving stable and efficient code learning and space compression. As a result, this provides an ideal foundation for the generator to sample new strokes within a semantically-aware, compressed space. + +Class label-to-sketch generation. Other than unconditional generation, our method can generate sketches given the class label conditions. To do this, we exploit a subset of the QuickDraw dataset [9]. The subset contains 20 classes, and stroke labels are NOT available in this dataset. Our adaptation is pretty simple and straightforward - i) removing the label Transformer (i.e., the variant of w/o $\mathcal{T}^l$ in Sec. 4.2), and ii) replacing the special STR token with the class index token (e.g., the sinusoidal encoding). We only + +re-train the generator on the QuickDraw subset until convergence. At inference time, given a class label, our method successfully produces high-quality results, as shown in Fig. 7(a). Note that, the experiment also demonstrates our flexibility and superior performance when working on datasets without stroke label information. + +Text-to-sketch generation. There are text descriptions for each sketch in the CB and CC datasets. We thus adapt our method to perform text-to-sketch generation. To this end, we first embed the text description using the pre-trained CLIP [21] model, and then replace the $STR$ token with the obtained text token, similarly as in the above label-to-sketch generation task. Fig. 7(b) shows a few novel examples of our generated sketches, which conform well to input texts. Pay attention to the novel visual concept, e.g., the 'Hug me' sketch, and the 'Smiling' sketch, which is rarely seen in the traditional sketch generation task. + +Sketch completion. Given an initial stroke, our model can perform sketch completion, i.e., stroke conditional generation. We achieve this by feeding the initial stroke as our first stroke in the generation process. Fig. 7(c) presents a visual comparison against SoTA methods on this task. As can be seen, DoodlerGAN produced the worst results, with most sketches lacking a clear structure. DoodleFormer performed slightly better, but the generated sketches are often blurry and include ghost strokes (e.g., a few tail-like strokes of the second example). With the help of our VQ representation and our generator, our results are both creative and structurally coherent. + +# 6. Conclusion + +In this paper, we have presented a novel approach for creative sketch generation. Our method can effectively create sketches with an appealing appearance and coherent structures. We believe that stroke VQ representations can be used in other sketch-related tasks and hope to inspire future research in sparse stroke representation learning. + +Acknowledgment. The authors would like to thank Oisin Mac Aodha and Ankan Bhunia for proofreading the early draft and the valuable discussions, Haocheng Yuan and Yi Yang for helping prepare comparisons with Diffsketcher, and the support of the HPC Platform at ShanghaiTech University. CL was supported by a gift from Adobe. + +# References + +[1] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12565-12574, 2024. 2 +[2] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12565-12574, 2024. 4 +[3] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10997-11008, 2024. 1 +[4] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. In European Conference on Computer Vision, pages 338-355. Springer, 2022. 1, 3, 5, 7 +[5] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Aisketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 5 +[6] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2 +[7] Chenjian Gao, Xin Wang, Qian Yu, Lu Sheng, Jing Zhang, Xiaoguang Han, Yi-Zhe Song, and Dong Xu. 3d reconstruction from a single sketch via view-dependent depth sampling. IEEE transactions on pattern analysis and machine intelligence, 2024. 1 +[8] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creativ sketch generation. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 7 +[9] David Ha and Douglas Eck. A neural representation of sketch drawings. arXiv preprint arXiv:1704.03477, 2017. 2,8 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a + +two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[11] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2 +[12] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-image diffusion models are great sketch-photo matchmakers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16826-16837, 2024. 1 +[13] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16859-16869, 2024. 1 +[14] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16859-16869, 2024. 1 +[15] Ke Li, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Honggang Zhang. Toward deep universal sketch perceptual grouper. IEEE Transactions on Image Processing, 28(7):3219-3231, 2019. 2 +[16] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning. IEEE computer graphics and applications, 39(2):38-51, 2018. 2 +[17] Rosanne Liu, Joel Lehman, Piero Molino, Felipe Petroski Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An intriguing failing of convolutional neural networks and the coordconv solution. Advances in neural information processing systems, 31, 2018. 4 +[18] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 7 +[19] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnN for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 2 +[20] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23327-23337, 2023. 2 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 8 +[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14153-14162, 2020. 2 + +[23] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8504–8513, 2021. 1 +[24] Aneeshan Sain, Pinaki Nath Chowdhury, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Freeview sketching: View-aware fine-grained sketch-based image retrieval. In European Conference on Computer Vision, pages 145-162. Springer, 2024. 1 +[25] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5 +[26] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, et al. Strokenuwa: Tokenizing strokes for vector graphic synthesis. arXiv preprint arXiv:2401.17093, 2024. 2 +[27] Chaitat Utintu, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Sketchdeco: Decorating b&w sketches with colour. arXiv preprint arXiv:2405.18716, 2024. 1 +[28] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4 +[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[30] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2 +[31] Fei Wang, Shujin Lin, Hanhui Li, Hefeng Wu, Tie Cai, Xiaonan Luo, and Ruomei Wang. Multi-column point-cnn for sketch segmentation. Neurocomputing, 392:50-59, 2020. 2 +[32] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3679-3688, 2024. 1, 2, 4, 7 +[33] Qiang Wang, Haoge Deng, Yonggang Qi, Da Li, and Yi-Zhe Song. Sketchknitter: Vectorized sketch generation with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 1, 5 +[34] Xingyuan Wu, Yonggang Qi, Jun Liu, and Jie Yang. Sketch-segnet: A rnn model for labeling sketch strokes. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2 +[35] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36:15869-15889, 2023. 1 +[36] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand + +sketch: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(1):285-312, 2022. 5 +[37] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. Sketchgnn: Semantic sketch segmentation with graph neural networks. ACM Transactions on Graphics (TOG), 40(3):1-13, 2021. 2 +[38] Ningyuan Zheng, Yifan Jiang, and Dingjiang Huang. Strokenet: A neural painting environment. In International Conference on Learning Representations, 2018. 2 +[39] Yixiao Zheng, Jiyang Xie, Aneeshan Sain, Yi-Zhe Song, and Zhanyu Ma. Sketch-segformer: Transformer-based segmentation for figurative and creative sketches. IEEE Transactions on Image Processing, 2023. 2 +[40] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 1 +[41] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to doodle with stroke demonstrations and deep q-networks. In BMVC, page 13, 2018. 2 +[42] Xianyi Zhu, Yi Xiao, and Yan Zheng. Part-level sketch segmentation and labeling using dual-cnn. In Neural Information Processing: 25th International Conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16, 2018, Proceedings, Part I 25, pages 374-384. Springer, 2018. 2 +[43] Xianyi Zhu, Yi Xiao, and Yan Zheng. 2d freehand sketch labeling using cnn and crf. Multimedia Tools and Applications, 79(1-2):1585–1602, 2020. 2 \ No newline at end of file diff --git a/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/images.zip b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..068f91b596449a46ff2bd19ee6409078f52d41ce --- /dev/null +++ b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a9c888e06693b73b401aa5e1e7016637be47975b99ec5a91a20d4b32bad0115 +size 743081 diff --git a/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/layout.json b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2c040f26d79fb5b974ff6f6db81a59fc47667c63 --- /dev/null +++ b/2025/VQ-SGen_ A Vector Quantized Stroke Representation for Creative Sketch Generation/layout.json @@ -0,0 +1,7383 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 483, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 483, + 138 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 483, + 138 + ], + "type": "text", + "content": "VQ-SGen: A Vector Quantized Stroke Representation for Creative Sketch Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "spans": [ + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "type": "text", + "content": "Jiawei Wang" + }, + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "type": "text", + "content": " Zhiming Cui" + }, + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "type": "text", + "content": " Changjian Li" + }, + { + "bbox": [ + 186, + 160, + 424, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 178, + 447, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 178, + 447, + 194 + ], + "spans": [ + { + "bbox": [ + 162, + 178, + 447, + 194 + ], + "type": "text", + "content": "1The University of Edinburgh 2ShanghaiTech University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 198, + 452, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 198, + 452, + 213 + ], + "spans": [ + { + "bbox": [ + 155, + 198, + 452, + 213 + ], + "type": "text", + "content": "https://enigma-li.github.io/projects/VQ-SGen/VQ-SGen.html" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 265, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 265, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 297, + 529 + ], + "type": "text", + "content": "This paper presents VQ-SGen, a novel algorithm for high-quality creative sketch generation. Recent approaches have framed the task as pixel-based generation either as a whole or part-by-part, neglecting the intrinsic and contextual relationships among individual strokes, such as the shape and spatial positioning of both proximal and distant strokes. To overcome these limitations, we propose treating each stroke within a sketch as an entity and introducing a vector-quantized (VQ) stroke representation for fine-grained sketch generation. Our method follows a two-stage framework - in stage one, we decouple each stroke's shape and location information to ensure the VQ representation prioritizes stroke shape learning. In stage two, we feed the precise and compact representation into an auto-decoding Transformer to incorporate stroke semantics, positions, and shapes into the generation process. By utilizing tokenized stroke representation, our approach generates strokes with high fidelity and facilitates novel applications, such as text or class label conditioned generation and sketch completion. Comprehensive experiments demonstrate our method surpasses existing state-of-the-art techniques on the CreativeSketch dataset, underscoring its effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Sketches are an intuitive mechanism widely used by humans to communicate ideas, convey concepts, and express creativity. Over the past decades, significant research has been conducted on sketch-related tasks, including retrieval [3, 13, 14, 23, 24], semantic segmentation [32, 40], and generation [4, 7, 8, 12, 27, 33]. In this paper, we focus on creative sketch generation, a novel task introduced by Ge et al. [8]. Unlike traditional sketch generation, this task aims to produce diverse, complex, and aesthetically appealing sketches, forming more imaginative depictions of familiar visual concepts, rather than generating conventional and typical representations of these concepts. This, thereby," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 242, + 548, + 388 + ], + "blocks": [ + { + "bbox": [ + 318, + 242, + 548, + 388 + ], + "lines": [ + { + "bbox": [ + 318, + 242, + 548, + 388 + ], + "spans": [ + { + "bbox": [ + 318, + 242, + 548, + 388 + ], + "type": "image", + "image_path": "fde780d9fcfd323186cb4e8f6dbfd4abb7f80844e9d4326deadb9de2cf376744.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 393, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. The goal of creative sketch generation is to generate vivid sketches, e.g., the birds from existing methods on the left. We have proposed VQ-SGen, for high-quality sketch generation with a new vector-quantized (VQ) representation and an efficient generator. See one of our results on the right for comparison." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 579 + ], + "type": "text", + "content": "places high demands on the generation capabilities of automated methods. Benefiting from the advancement of diffusion models, the notable and relevant progress for sketch generation is Diffsketer [35], a novel approach for text-to-sketch generation. However, the large-scale backbone diffusion model is pre-trained on realistic images, and their resultant sketches thus have high fidelity and realism conforming to human imagination but fail to promote creativity. Therefore, dedicated methods for creative sketch generation are needed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Existing methods [4, 8] tackle creative sketch generation by producing pixels either as a whole sketch or as segmented parts, overlooking inherent and contextual relationships among individual strokes, such as their shapes and relative positions of both nearby and distant strokes. As a result, their generated sketches often have blurry local regions, and scattered or isolated strokes (see the head of the bird from DoodleFormer and DoodlerGAN in Fig. 1). A more recent work, SketchKnitter [33], formulated the problem of sketch generation as a stroke points rectification process, where diffusion models are leveraged to rearrange a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19311" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "set of stroke points from an initial chaotic state into coherent forms. However, without the concept of a stroke entity, this method performs poorly when generating complex creative sketches (see the bird in Fig. 1 from SketchKnitter)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 314 + ], + "type": "text", + "content": "To address these limitations, we propose a novel stroke-level sketch generation approach that treats each stroke as an independent entity [20, 32] and encodes it using a novel vector-quantized (VQ) representation. This compact, discrete representation enhances the learning process of the generative model by efficiently capturing essential stroke shape variations while minimizing redundancy. Despite the discrete nature of the representation, we observe semantic-aware clustering within the stroke code space, which lays an ideal foundation for the generator to sample new strokes. Furthermore, our bespoke generator incorporates the shape, semantic, and position into its generation process, effectively samples new sketches in the compressed VQ space, and accurately reasons out the structural and semantic relationship between strokes, resulting in sketches that are both appealing and structurally coherent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 315, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 566 + ], + "type": "text", + "content": "Specifically, our method VQ-SGen, adopts a two-stage framework. In the first stage, we decouple the shape and location information of each stroke. This decoupling allows the vector-quantized (VQ) representation to concentrate on capturing stroke shapes, minimizing interference from positional data during the learning process. By prioritizing shape as the primary focus, the model learns a highly compact and informative, shape-oriented representation that preserves the inherent shape features of each stroke while reducing unnecessary redundancy. In the second stage, we employ an autoregressive Transformer to work with our VQ representation. The Transformer progressively integrates the stroke's semantic (i.e., the category label), shape, and location into the generation process. Through the autoregressive decoding mechanism, the Transformer captures each stroke's semantic, shape, and context relationship with neighboring strokes, ensuring that the shape and positioning are accurate and harmoniously aligned with its semantic role within the overall composition. As a result, our approach effectively avoids all the artifacts observed in previous methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 568, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 639 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our method, we have conducted experiments on the CreativeSketch [8] dataset. Comprehensive experiments including comparisons and ablation studies show that our method surpasses existing state-of-the-art techniques. In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 642, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "type": "text", + "content": "- We propose a novel stroke-level approach, " + }, + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "type": "inline_equation", + "content": "VQ-SGen" + }, + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "type": "text", + "content": ", for the challenging creative sketch generation task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 295, + 713 + ], + "type": "text", + "content": "- We propose a new vector-quantized (VQ) stroke representation that treats each stroke as an entity, capturing essential shape features compactly and reducing redundancy to serve as the basis for stroke-level sketch generation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "content": "- We propose a bespoke generator, leveraging the shape, semantic, and spatial location together, achieving high-quality generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "content": "- We extensively evaluated our method, demonstrating superior performance. A user study shows that VQ-SGen consistently performs favorably against SoTA methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 155, + 400, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 155, + 400, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 155, + 400, + 167 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 175, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 521 + ], + "type": "text", + "content": "Sketch Representation Learning. Based on the data format of sketches, existing sketch representation learning methods can be broadly divided into three categories: image-based [16, 42, 43], sequence-based [1, 15, 19, 34], and graph-based [31, 37, 39] approaches, each with its own advantages and limitations. Specifically, image-based methods [16, 42, 43] use raster images as input, leveraging absolute coordinates to capture stroke proximity. However, they often struggle to learn structural information and sequential relationships among strokes effectively. Graph-based methods [31, 37, 39] face challenges in capturing detailed structural information within strokes. Sequence-based methods [15, 19, 34] utilize point sequences with relative coordinates, which enhances the encoding of individual stroke structures but makes it difficult to represent inter-stroke proximity and spatial relationships. ContextSeg [32] proposed the concept of the single stroke entity for sketch semantic segmentation. The advantages of this representation are demonstrated in their paper by comparing it with fancy Transformer [22] or common RNN [9, 15] based methods. SketchXAI [20] decoupled a single stroke into its shape, location, and order embeddings, and the concatenated stroke embedding was used to train a sketch classifier to study the explainability of human sketches. We exploit the same concept of treating a single stroke as the atomic unit, and decouple its shape with its location. Additionally and importantly, we train a vector quantized representation to build the discrete and compact code space dedicated to our generation task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 677 + ], + "type": "text", + "content": "Vector-Quantized VAEs. Learning representations with discrete features has been the focus of many previous works [6, 11, 30], as such representations are a natural choice for complex reasoning, planning, and predictive learning. Although using discrete latent variables in deep learning has been difficult, effective autoregressive models have been created to capture distributions over discrete variables. For instance, StrokeNUWA [26] utilizes VQ-VAE [28] to achieve a better representation of vector graphics, then generates vector graphics through large language models. Similarly, we propose to learn a VQ representation for rasterized strokes, expecting an improved generation capability powered by the compact representation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Creative Sketch Generation. Unlike traditional sketch generation tasks [38, 41], creative sketch generation [8] places greater emphasis on producing more imaginative de" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19312" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 397, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 397, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 397, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 397, + 270 + ], + "type": "image", + "image_path": "45510d92dace14c50ed3f171d04a8f5e2ab66a7c01c4c333e51ecb8a3c0f4391.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 309 + ], + "type": "text", + "content": "Figure 2. Overview of VQ-SGen. Given an input sketch, it is first divided into a sequence of strokes. In the first stage, we begin by decoupling the shape and location information of each stroke and obtain their discrete representations (Sec. 3.1). In the second stage, we use a decoder-only Gen-Transformer to predict " + }, + { + "bbox": [ + 55, + 274, + 555, + 309 + ], + "type": "inline_equation", + "content": "\\{(I_i, b_i, l_i)\\}" + }, + { + "bbox": [ + 55, + 274, + 555, + 309 + ], + "type": "text", + "content": " in an autoregressive manner (Sec. 3.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 399, + 72, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 399, + 72, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 399, + 72, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 399, + 72, + 553, + 270 + ], + "type": "image", + "image_path": "03231a5eebdc93b56ced6eaf36c6cbe34db554dfdd128fa50830a761df12761d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 317, + 297, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 297, + 462 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 297, + 462 + ], + "type": "text", + "content": "pictures of familiar visual concepts, rather than generating conventional and mundane forms of these concepts. To achieve this goal, DoodlerGAN [8] presents a component-based GAN architecture to incrementally generate each part of a creative sketch image. This approach requires training one GAN model individually for each body part in a supervised way, leading to significant computational overhead. DoodleFormer [4] employs an attention-based architecture for two-stage sketch generation. However, this approach tends to focus only on the broad, rough relationships between different parts of the sketch, which can lead to unnatural mixing of strokes from different regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 476, + 112, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 112, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 112, + 489 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": "Sketch and strokes. A standard sketch " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " in our context is composed of a sequence of strokes " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " and their corresponding stroke labels " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{l_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ". The order of each stroke in the sequence corresponds to their default drawing order. Each stroke is represented as a rasterized image of size " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ", and each label " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " is a one-hot vector in " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ". In order to decouple the shape and position of a single stroke, we further process it. Formally, for each rasterized stroke image, we calculate its axis-aligned bounding box and record its center coordinate " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "(x_i, y_i)" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ". The bounding box size (i.e., the height and width) as well as the center coordinate fully represent the stroke position " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "b_i = (w_i / 2, h_i / 2, x_i, y_i)" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": ". By factoring out the position (i.e., translating the stroke bounding box to align with the sketch image center), the stroke shape itself is presented as the translated rasterized image " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": " without changing its scale. Considering the triplet of shape, position, and semantic of a single stroke, the new format to represent a sketch is " + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "inline_equation", + "content": "S = \\{s_i\\}_{i=1}^N = \\{(I_i, b_i, l_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 498, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 317, + 556, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 556, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 556, + 473 + ], + "type": "text", + "content": "Sketch generation. Our goal in this paper is to generate a sequence of strokes in the format of the triplet and assemble them based on their shape and position to form the resulting sketch (see Fig. 1). Figure 2 displays an overview of our novel sketch generation method, which is comprised of two main components. Firstly, a decoupled representation learning module encodes a stroke into its vector-quantized representation (Sec. 3.1) that is highly compact and shape-aware. Secondly, a bespoke generator based on the Transformer [29] is designed to generate the stroke label, shape, and location sequentially, achieving a fine-grained control with high-fidelity resulting sketches. In the following sections, we elaborate on the details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 525, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 525, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 525, + 494 + ], + "type": "text", + "content": "3.1. Vector-quantized Stroke Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "content": "Our triplet stroke formula enables separate control of stroke key properties in the generation process, how to properly encode the triplet to fit generation is the next step. Motivated by the vector quantization process [28], we aim to obtain a similar compact stroke representation in the context of the whole sketch, such that the sketch generator can efficiently and effectively sample new sketches in a compressed stroke space. Besides, we expect a decoupled representation based on our triplet formulation to obtain high shape and semantic awareness. To this end, we first convert each stroke into a latent embedding and further compress them in the latent space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": "Stroke latent embedding. Given the stroke image " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " we obtain the stroke latent embedding via an autoencoder, where both the encoder " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^s" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": " are built with a few 2D CNNs (see supplementary for detailed network configuration). After the last layer of the encoder, we flatten its bottleneck feature to obtain the stroke latent em" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19313" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "bedding " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "e_i^s = \\text{flat}(\\mathcal{E}^s(\\boldsymbol{I}_i))" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": ". The autoencoder is trained using the reconstruction loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 103, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 295, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n s}} = \\left\\| \\boldsymbol {I} _ {i} - \\mathcal {D} ^ {s} \\left(\\mathcal {E} ^ {s} \\left(\\boldsymbol {I} _ {i}\\right)\\right) \\right\\| ^ {2}. \\tag {1}", + "image_path": "264ed3fb59a8d879cd6545a8524196c75a57f6c730089ffff7ba8c6380848858.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 127, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 295, + 163 + ], + "type": "text", + "content": "To better capture the stroke shape information, we follow [2, 17, 32] to add the CoordConv layer to the network and the distance field supervision in the training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "content": "Tokenizing strokes. To achieve fine-grained control, we build separate discretized spaces for the shape and location, respectively. The reason to build a space for the shape location is that it is highly related to semantics, e.g., the heads of birds appear at a similar position with a similar size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": "For the stroke shape space, as shown in Fig. 2 left, we build a codebook " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": ". Specifically, given a sketch " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": ", its corresponding stroke shape latent embeddings " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\{e_i^s\\}_{n = 1}^N" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": " are fed into the encoder " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^f" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": " to obtain the feature sequence " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\{z_i^s\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": ". We then compress it via vector quantization, i.e., clamping each feature into its nearest code in " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": " codes " + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i = 1}^V" + }, + { + "bbox": [ + 55, + 223, + 296, + 307 + ], + "type": "text", + "content": ", and we record the indices of these features:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 315, + 295, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 315, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 109, + 315, + 295, + 331 + ], + "type": "interline_equation", + "content": "v _ {i} ^ {s} = \\operatorname {a r g m i n} _ {j \\in [ 0, V)} \\| \\boldsymbol {z} _ {i} ^ {s} - \\boldsymbol {c} _ {j} \\|. \\tag {2}", + "image_path": "f8a6bb3afeb8fe35dceab6a92bcd87e58b30248d6bc38ffb8b09571ff07fc777.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "text", + "content": "The feature sequence " + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\{c_{v_i^s}\\}^N" + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "text", + "content": " is further sent to the decoder " + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^f" + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "text", + "content": " to reconstruct the input latent embeddings. Both " + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^f" + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^f" + }, + { + "bbox": [ + 55, + 338, + 296, + 388 + ], + "type": "text", + "content": " are composed of several 1D CNN layers, see the supplementary for the detail structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 388, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 422 + ], + "type": "text", + "content": "To train the vector quantization network, we follow [28] to exploit the codebook, commitment, and reconstruction losses, defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 431, + 295, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 431, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 64, + 431, + 295, + 490 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\alpha (\\| \\pmb {z} _ {i} ^ {s} - \\mathrm {s g} [ \\pmb {c} _ {v _ {i} ^ {s}} ] \\| _ {2} ^ {2} + \\| \\mathrm {s g} [ \\pmb {z} _ {i} ^ {s} ] - \\pmb {c} _ {v _ {i} ^ {s}} \\| _ {2} ^ {2}) \\\\ + \\left\\| \\boldsymbol {z} _ {i} ^ {s} - \\mathcal {D} ^ {f} \\left(\\boldsymbol {c} _ {v _ {i} ^ {s}}\\right) \\right\\| _ {2} ^ {2}, \\tag {3} \\\\ \\end{array}", + "image_path": "78692778f879de920c35b70aa0b2ab270d8b6e9fb318a1484ed9e3dd4027769a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 282, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 282, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 282, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 491, + 282, + 502 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 491, + 282, + 502 + ], + "type": "text", + "content": " is a weight, and " + }, + { + "bbox": [ + 55, + 491, + 282, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}" + }, + { + "bbox": [ + 55, + 491, + 282, + 502 + ], + "type": "text", + "content": " is the stopgradient operator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "text", + "content": "For the stroke location codebook learning, we use the same process as the shape codebook learning. The stroke position " + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "inline_equation", + "content": "b_{i} = (w_{i} / 2, h_{i} / 2, x_{i}, y_{i})" + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "text", + "content": " is reconstructed by a pair of encoder " + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}^{l})" + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}^{l})" + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "text", + "content": " to build the codebook " + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "text", + "content": " with V codes " + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "inline_equation", + "content": "\\{d_{i}\\}_{i=1}^{V}" + }, + { + "bbox": [ + 55, + 502, + 296, + 573 + ], + "type": "text", + "content": " as well. See supplementary for detailed network structures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "text", + "content": "Till now, given a stroke " + }, + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}_i = (\\boldsymbol{I}_i, b_i, l_i)" + }, + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "text", + "content": ", we have the compact representation " + }, + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "inline_equation", + "content": "(v_i^s, v_i^l, l_i)" + }, + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "text", + "content": ", corresponding to the specific code features " + }, + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{c}_{v_i^s}, \\boldsymbol{d}_{v_i^l}, l_i)" + }, + { + "bbox": [ + 55, + 574, + 296, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 624, + 204, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 204, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 204, + 636 + ], + "type": "text", + "content": "3.2. Autoregressive Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Having the VQ stroke representation, our goal in this module is to estimate a distribution over the sketch " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " so as to generate new instances. Since " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is composed of a sequence of stroke shapes, positions, and labels, we split the generation process into two steps, i.e., 1) generate the stroke label, and 2) generate the stroke shape and position conditioned" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 552, + 191 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 552, + 191 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 552, + 191 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 552, + 191 + ], + "type": "image", + "image_path": "db616db51cc3c6e33f3f5842d8e3d8015c9eb5e78bb230fccab4170fa3120fed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 194, + 523, + 205 + ], + "lines": [ + { + "bbox": [ + 345, + 194, + 523, + 205 + ], + "spans": [ + { + "bbox": [ + 345, + 194, + 523, + 205 + ], + "type": "text", + "content": "Figure 3. Network architecture of our generator." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "type": "text", + "content": "on the stroke label. By applying the chain rule, we formulate it as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 236, + 553, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 236, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 326, + 236, + 553, + 269 + ], + "type": "interline_equation", + "content": "p (\\boldsymbol {S}) = \\prod_ {i = 1} ^ {N} p \\left(\\boldsymbol {I} _ {i}, b _ {i}, l _ {i}\\right) = \\prod_ {i = 1} ^ {N} p \\left(v _ {i} ^ {\\mathrm {s}}, v _ {i} ^ {l} \\mid l _ {i}\\right) \\cdot p \\left(l _ {i}\\right). \\tag {4}", + "image_path": "0a2cf132864a91ffc080bc11eef8f8f4bfc4ed0d76fc9228f884e075bd76c518.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 273, + 554, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 554, + 332 + ], + "type": "text", + "content": "We implement Eq. (4) with two cascaded Transformer decoders, parameterized with " + }, + { + "bbox": [ + 313, + 273, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\theta_{1}" + }, + { + "bbox": [ + 313, + 273, + 554, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 273, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\theta_{2}" + }, + { + "bbox": [ + 313, + 273, + 554, + 332 + ], + "type": "text", + "content": ", respectively. The first decoder samples a label for the next stroke, while the second decoder generates the shape and position conditioned on the label, which can be further formulated as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 377, + 337, + 506, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 377, + 337, + 506, + 350 + ], + "type": "interline_equation", + "content": "p (l _ {i}) = p \\left(l _ {i} \\mid l _ {< i}, v _ {s < } ^ {s}, v _ {l} ^ {< }; \\theta_ {1}\\right),", + "image_path": "21255599c001f5b59c67bc64f8aaf09cb8b1be64d421944478d58251f6f252ba.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 347, + 553, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 347, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 347, + 347, + 553, + 368 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (i) \\quad P (i + l _ {i}) \\quad P (i + l _ {i}) \\quad P (i + 1) \\\\ p \\left(v _ {i} ^ {s}, v _ {i} ^ {l} \\mid l _ {i}\\right) = p \\left(v _ {i} ^ {s}, v _ {i} ^ {l} \\mid l _ {\\leq i}, v _ {< i} ^ {s}, v _ {< i} ^ {l}; \\theta_ {2}\\right). \\end{array} \\tag {5}", + "image_path": "6f16ca95aeadd708f08e583036bc5ac79ee0fe49841a6ee8ca0c59f83ce60f66.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "content": "Note that, we will use the codes in " + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "content": " in the Transformer calculation, instead of the original rasterized image and the quaternion for the stroke shape and position. Network Architectures. An overview of our cascaded decoders is shown in Fig. 3, where the first label Transformer " + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}^{l})" + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "content": " predicts a stroke label each time, while the second code Transformer " + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}^{c})" + }, + { + "bbox": [ + 313, + 372, + 554, + 468 + ], + "type": "text", + "content": " takes as input the predicted label generating both the stroke shape and position codes simultaneously." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": "Concretely, the label Transformer " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^l" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": " takes as input the triplet " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\left(\\pmb{c}_{v_i^s},\\pmb{d}_{v_i^l},l_i\\right)" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": ". Each element goes through its own embedding layer to get a feature, i.e., " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{Emb}(\\pmb{c}_{v_i^s})\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{Emb}(\\pmb{d}_{v_i^l})\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{Emb}(l_i)\\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": ". The three features are simply added along the feature dimension to obtain the input feature for the attention blocks, which outputs a fused feature " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": " after the attention calculation. " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": " is then fed into a fully connected layer with the softmax activation to predict the one-hot probabilistic vector, which is translated into a one-hot class label " + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "inline_equation", + "content": "l_{i + 1}\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 468, + 554, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "The code Transformer " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^c" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " accepts two inputs. The first one is " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "l_{i + 1}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", which goes through an embedding layer to produce a conditional feature in " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{512}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ". The second input is " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", which is also added with the conditional feature along the feature dimension to get the input feature for the attention block. Intuitively, the first input provides the direct semantic guidance of the current stroke, while the second input provides the shape and positional information of previous strokes. The attention blocks are followed by two separate branches, each of which outputs a one-hot probabilistic" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19314" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "text", + "content": "vector of dimension " + }, + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^V" + }, + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "text", + "content": ", indicating the indices of the shape and position codebooks. The corresponding code entries are fetched (i.e., " + }, + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "inline_equation", + "content": "c_{v_{i+1}^s}" + }, + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "inline_equation", + "content": "d_{v_{i+1}^t}" + }, + { + "bbox": [ + 55, + 72, + 294, + 121 + ], + "type": "text", + "content": ") and serve as the input to the label Transformer for the next stroke generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 146 + ], + "type": "text", + "content": "Losses. To train the network, we minimize the negative log-likelihood loss, defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 158, + 295, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 158, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 135, + 158, + 295, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g e n}} = - \\log p (\\boldsymbol {S}), \\tag {6}", + "image_path": "88e0758827606038e4d4ebfe21890d842537511dae51f6e36a5e66d6c4f31b0a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 182, + 294, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 218 + ], + "type": "text", + "content": "where we supervise the training with ground truth categorical labels (i.e., the shape label vector, shape code index vector, and the position code index vector)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 220, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 352 + ], + "type": "text", + "content": "Training and testing strategy. We have three training stages arranged sequentially for stroke latent embedding, codebooks, and the generator. Detailed steps, hyperparameter settings, and implementation details can be found in the supplementary. For the testing time generation, starting from a special STR token, we generate novel strokes in an autoregressive manner, where the codebook entry is either decoded as the position quaternion or the stroke latent code, which is further converted into the corresponding rasterized image. The stroke triplet is finally assembled together to produce the resulting sketch (see Fig. 2, left)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 136, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 136, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 136, + 379 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": "Datasets. Following previous work [4, 8], we use the CreativeSketch [8] dataset in this paper. Specifically, it consists of two categories of sketches — Creative Birds (CB) and Creative Creatures (CC) — containing 8,067 and 9,097 sketches, respectively, along with part annotations. All the sketches have been processed into " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " images. For the stroke latent embedding learning, we use strokes from both datasets to obtain a powerful pair of " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^s" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": ". Other than this autoencoder, we have trained other networks separately on CB and CC datasets. See supplementary for details about pre-processing and data augmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 519, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 567 + ], + "type": "text", + "content": "Evaluation metrics. Following previous work [4, 8], we evaluate the model's performance using an Inception model [25] trained on the QuickDraw3.8M dataset [36]. We use four metrics in this paper, defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 567, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 602 + ], + "type": "text", + "content": "- Fréchet Inception distance (FID) [10] measures the similarity between two image sets in the above Inception feature space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 603, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 651 + ], + "type": "text", + "content": "- Generation diversity (GD) [5] measures the average pairwise Euclidean distance between the Inception features of two subsets, reflecting the diversity of the data item in the image set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 651, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 687 + ], + "type": "text", + "content": "- Characteristic score (CS) measures how frequently a generated sketch is classified as Creative Bird or Creative Creature by the Inception model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 687, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 687, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 295, + 711 + ], + "type": "text", + "content": "- Semantic diversity score (SDS) quantifies sketch diversity based on various creature categories they represent." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 317, + 96, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 92 + ], + "type": "text", + "content": "Table 1. Statistical comparison against existing methods on Creative Birds and Creative Creatures datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 96, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 170 + ], + "type": "table", + "html": "
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
Sketchknitter [33]74.4214.230.1464.3412.340.421.32
DoodlerGAN [8]39.9516.330.6943.9414.570.551.45
DoodleFormer [4]17.4817.830.5720.4316.230.531.68
Ours15.7818.920.5317.6117.420.571.86
", + "image_path": "8ae56305b1201f8fc876f42fe2749b9e4fb0f9f3140106b7db64c338477585c6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 198, + 395, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 395, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 395, + 211 + ], + "type": "text", + "content": "4.1. Comparison" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 225, + 555, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 298 + ], + "type": "text", + "content": "We compare our approach with three competitors, i.e., Doodleformer [4], DoodlerGAN [8], Sketchknitter [33]. Both Doodleformer and DoodlerGAN generate sketches at the pixel level, while Sketchknitter generates sketches at the stroke point sequence level. All three methods are trained using their default parameters on our dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 307, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 715 + ], + "type": "text", + "content": "The quantitative comparison is presented in Tab. 1, where our method significantly outperforms others. Specifically, on the Creative Birds dataset, we obtained a notable 1.7 decrease in FID and gained a remarkable 1.09 increase in terms of GD, achieving the best performance. This indicates that our model can generate sketches that are closer to the original dataset, as well as more diverse at the same time. For the CS metric, while our approach does not achieve the highest score, it surpasses the dataset and comes closest to matching the dataset's CS metric. In contrast, DoodlerGAN and DoodleFormer attain higher scores, whereas SketchKnitter's outputs are seldom recognized as birds. For this specific metric, the higher value only means it is easier to be recognized by the Inception model, which may stem from the simplicity of the generation. The dataset's value sets a good reference. Similarly, on the Creative Creatures dataset, we achieved the best performance across all four metrics. Specifically, we achieved a 2.82 decrease in FID and a 1.19 improvement in GD compared to the best other methods. This indicates that our generated sketches are closer to the more complex sketches in the training data and exhibit greater diversity among themselves. For the CS and SDS metrics, we obtained improvements of 0.04 and 0.18, respectively. This indicates that our generated sketches are semantically closer to the categories in the training data and also suggests that our model exhibits better semantic diversity. A corresponding visual comparison is shown in Fig. 4, where the artifacts like the broken and misaligned strokes from DoodlerGAN, and the blurry local regions from DoodleFormer, are clearly presented. In contrast, our results are of high quality with complex appearance and diverse shapes. Note that, as shown in Fig. 1, SketchKnitter generates poor results on both datasets, we did not include its results in the visual comparison." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19315" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 555, + 370 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 555, + 370 + ], + "type": "image", + "image_path": "87506e4a5b27e9f093eb519df6e3d51362a85b5811b17e7e09a20ad172631e9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 372, + 500, + 384 + ], + "lines": [ + { + "bbox": [ + 108, + 372, + 500, + 384 + ], + "spans": [ + { + "bbox": [ + 108, + 372, + 500, + 384 + ], + "type": "text", + "content": "Figure 4. Visual comparison against competitors on both the Creative Birds and Creative Creatures datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 385, + 529, + 628 + ], + "blocks": [ + { + "bbox": [ + 83, + 385, + 529, + 628 + ], + "lines": [ + { + "bbox": [ + 83, + 385, + 529, + 628 + ], + "spans": [ + { + "bbox": [ + 83, + 385, + 529, + 628 + ], + "type": "image", + "image_path": "d700b703190675cf859bc885ffaca268059af48ee1f0cdfdb18399f4f7daff58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 630, + 555, + 653 + ], + "lines": [ + { + "bbox": [ + 55, + 630, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 555, + 653 + ], + "type": "text", + "content": "Figure 5. Sketch reconstruction results of our ablation study on codebook hyper-parameter setting. The two numbers represent the code size and the feature dimension, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 662, + 149, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 149, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 149, + 674 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "To validate the effectiveness of our core technical components, we conducted experiments by removing them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 662, + 544, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 662, + 544, + 674 + ], + "spans": [ + { + "bbox": [ + 314, + 662, + 544, + 674 + ], + "type": "text", + "content": "Specifically, we form the following alternative solutions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 315, + 677, + 555, + 715 + ], + "type": "text", + "content": "- w/o VQ: we remove all the codebook learning-related procedures, and use the continuous stroke latent embeddings to train our generator." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19316" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 118, + 294, + 220 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 114 + ], + "type": "text", + "content": "Table 2. The statistical comparison of our ablation study. The method name of two numbers represents the codebook hyperparameter setting, where the first is the code size, and the second is the feature dimension." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 118, + 294, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 294, + 220 + ], + "type": "table", + "html": "
MethodsCreative BirdsCreative Creatures
FID(↓)GD(↑)CS(↑)FID(↓)GD(↑)CS(↑)SDS(↑)
Training Data-19.400.45-18.060.601.91
2048×51226.2315.340.4343.2114.510.481.52
4096×51216.9218.270.5018.4416.140.541.63
4096×102416.6718.150.5118.1216.430.551.69
w/o Tl16.5118.350.5119.2116.140.551.74
w/o VQ48.5313.340.4654.5614.020.451.36
w/o Decouple17.1418.120.5719.4216.420.561.79
Ours(8192 × 512)15.7818.920.5317.6117.420.571.86
", + "image_path": "d360e5265c080b6618d1ce86bc3c3545e99cd31e2b3eb87b8541e8cd9ede8f11.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 230, + 295, + 302 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 230, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 295, + 277 + ], + "type": "text", + "content": "- w/o Decouple: we do not decouple the shape and position of the stroke, and only train a single stroke codebook. The Transformer decoders take fewer inputs and the second decoder outputs only one index vector accordingly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 302 + ], + "type": "text", + "content": "- w/o " + }, + { + "bbox": [ + 55, + 277, + 295, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^l" + }, + { + "bbox": [ + 55, + 277, + 295, + 302 + ], + "type": "text", + "content": ": we simply discard the label information for each stroke, and remove the label Transformer accordingly." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 304, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 542 + ], + "type": "text", + "content": "We train the three alternatives from scratch using our datasets until convergence and statistical results are shown in Tab. 2. Without the VQ representation, the generator samples new sketches from a larger continuous space, instead of a compressed discretized space, giving the worst performance in this paper. The stroke decoupling matters the generation significantly (e.g., 15.78 vs. 17.14 and 17.61 vs. 19.42 of the FID). By decoupling, the shape code can concentrate only on the shape variation, while the position code can better capture the inherent approximation information between strokes, which further facilitates the generation. By removing the label prediction, we see a slight performance degradation, giving the second-best performance in Tab. 2. The labels are not very critical, but indeed provide semantic-related information to complement the positional information (e.g., the heads appear at similar positions with similar sizes). On the other hand, it demonstrates the great feasibility of applying our method to other datasets without label information, proving our flexibility. See Sec. 5 for such an experiment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": "Codebook hyper-parameter setting. There are two core hyper-parameters for a codebook, i.e., the code size and the feature dimension. In the following, we have conducted experiments on the shape codebook to find the proper configuration. Similar experiments on the location codebook are presented in the supplementary. Specifically, we set the code size to be either 2,048, 4,096, or 8,192, and set the feature dimension to be either 512 or 1,024. We examine the effects of the parameter setting from both the sketch reconstruction and generation perspectives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Figure 5 presents a visual comparison of sketch reconstruction quality. By increasing either the code size or the feature dimension, we observe improved fidelity with less blurry and broken strokes. Note that the row named " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "VQ" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 74, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 317, + 74, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 317, + 74, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 74, + 553, + 194 + ], + "type": "image", + "image_path": "937c6bef050e1477f8016fb2885ec6a44e7b457483217d5d5022cac65efd059a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 6. User study results based on the five questions (a-e). Higher scores indicate a stronger preference for our method over the other approaches (i.e., DoodleFormer or Creative datasets)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "type": "text", + "content": "shows the stroke latent space reconstruction performance as in [32], which along with the input, provides a good performance reference. We chose the last configuration in this paper, as it gives the best visual reconstruction. However, it still struggles to represent strokes with significant variation, such as the ear of the creature in the third-to-last column. We further investigate the generation performance of different codebook settings. Quantitative results are listed in Tab. 2, where we observe improvement when increasing either the size of the code or the feature dimension. Our choice obtains the best result given the limited combination." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 389, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 389, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 389, + 388 + ], + "type": "text", + "content": "4.3. User Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 394, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 585 + ], + "type": "text", + "content": "Following [4, 8], we evaluate the performance of our method through user studies. We invited 50 participants to the test, where they were shown multiple pairs of images. Each pair consists of a sketch generated by our model and another sketch from either DoodleFormer or the datasets. For each pair, participants were asked to answer 5 single-choice questions: which one (a) is more creative? (b) looks more like a bird/creature? (c) is more likely to be drawn by a human? (d) the strokes better integrated? (e) they like the most overall? Figure 6 shows the percentage of times our method is preferred over the competing approach. It can be seen that " + }, + { + "bbox": [ + 313, + 394, + 555, + 585 + ], + "type": "inline_equation", + "content": "VQ-SGen" + }, + { + "bbox": [ + 313, + 394, + 555, + 585 + ], + "type": "text", + "content": " performs favorably against DoodleFormer for all five questions on both datasets except for question (b) on the Creative Birds dataset, which is consistent with the CS metric in Tab. 1. We are nearly tied with the creative datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 597, + 468, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 468, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 468, + 611 + ], + "type": "text", + "content": "5. Discussion and Application" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 653 + ], + "type": "text", + "content": "In this section, we mainly discuss code space exploration and conditional generation applications, while leaving other discussions, e.g., limitations, in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "content": "Code space exploration. To better understand the properties of the discrete space, we applied the UMAP algorithm [18] to project the high dimensional features of the shape codebook " + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "content": " into 2D space and visualize its distribution in Fig. 8. We further probe the emerged clustering by visu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19317" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 75, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 75, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 75, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 75, + 545, + 205 + ], + "type": "image", + "image_path": "f5a9fca7875781b8283dbd3a12af22d95e4806242a8f23e98e7031c599fc0e13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 555, + 242 + ], + "type": "text", + "content": "Figure 7. Applications of our approach. (a) Given the class label, our method can produce corresponding sketches with rich variations. (b) Our method supports text-conditioned generation, the resulting sketches explain the input text vividly. (c) Given the initial stroke, our method can complete the whole sketch, which is favorable against competitors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 251, + 286, + 430 + ], + "blocks": [ + { + "bbox": [ + 62, + 251, + 286, + 430 + ], + "lines": [ + { + "bbox": [ + 62, + 251, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 286, + 430 + ], + "type": "image", + "image_path": "79bf3d42d6bf35e06c5f2f66185f4638c939ef14a5ec522e1dd5700192dfd3a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "type": "text", + "content": "Figure 8. UMAP visualization of the shape codes in " + }, + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 55, + 433, + 295, + 467 + ], + "type": "text", + "content": ", with the overlaid strokes indicating the semantic aware clustering in the discrete code space." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 474, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 605 + ], + "type": "text", + "content": "alizing typical strokes. Not too surprised, we find that the emerged clusters correspond well to semantics. For example, the five surrounding clusters are Beak, Eye, Tail, Wing, and Leg, while the inner big cluster composed of Head and Body. Although we did not exploit any semantic-related constraints during the codebook learning, it tends to implicitly group strokes with similar shape variations, thereby achieving stable and efficient code learning and space compression. As a result, this provides an ideal foundation for the generator to sample new strokes within a semantically-aware, compressed space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Class label-to-sketch generation. Other than unconditional generation, our method can generate sketches given the class label conditions. To do this, we exploit a subset of the QuickDraw dataset [9]. The subset contains 20 classes, and stroke labels are NOT available in this dataset. Our adaptation is pretty simple and straightforward - i) removing the label Transformer (i.e., the variant of w/o " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^l" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " in Sec. 4.2), and ii) replacing the special STR token with the class index token (e.g., the sinusoidal encoding). We only" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 254, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 326 + ], + "type": "text", + "content": "re-train the generator on the QuickDraw subset until convergence. At inference time, given a class label, our method successfully produces high-quality results, as shown in Fig. 7(a). Note that, the experiment also demonstrates our flexibility and superior performance when working on datasets without stroke label information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 327, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 460 + ], + "type": "text", + "content": "Text-to-sketch generation. There are text descriptions for each sketch in the CB and CC datasets. We thus adapt our method to perform text-to-sketch generation. To this end, we first embed the text description using the pre-trained CLIP [21] model, and then replace the " + }, + { + "bbox": [ + 313, + 327, + 556, + 460 + ], + "type": "inline_equation", + "content": "STR" + }, + { + "bbox": [ + 313, + 327, + 556, + 460 + ], + "type": "text", + "content": " token with the obtained text token, similarly as in the above label-to-sketch generation task. Fig. 7(b) shows a few novel examples of our generated sketches, which conform well to input texts. Pay attention to the novel visual concept, e.g., the 'Hug me' sketch, and the 'Smiling' sketch, which is rarely seen in the traditional sketch generation task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 605 + ], + "type": "text", + "content": "Sketch completion. Given an initial stroke, our model can perform sketch completion, i.e., stroke conditional generation. We achieve this by feeding the initial stroke as our first stroke in the generation process. Fig. 7(c) presents a visual comparison against SoTA methods on this task. As can be seen, DoodlerGAN produced the worst results, with most sketches lacking a clear structure. DoodleFormer performed slightly better, but the generated sketches are often blurry and include ghost strokes (e.g., a few tail-like strokes of the second example). With the help of our VQ representation and our generator, our results are both creative and structurally coherent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 620, + 388, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 388, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 388, + 632 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we have presented a novel approach for creative sketch generation. Our method can effectively create sketches with an appealing appearance and coherent structures. We believe that stroke VQ representations can be used in other sketch-related tasks and hope to inspire future research in sparse stroke representation learning." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19318" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "content": "Acknowledgment. The authors would like to thank Oisin Mac Aodha and Ankan Bhunia for proofreading the early draft and the valuable discussions, Haocheng Yuan and Yi Yang for helping prepare comparisons with Diffsketcher, and the support of the HPC Platform at ShanghaiTech University. CL was supported by a gift from Adobe." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 156, + 115, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 167 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 175, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 175, + 297, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 297, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 297, + 251 + ], + "type": "text", + "content": "[1] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12565-12574, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 254, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 296, + 330 + ], + "type": "text", + "content": "[2] Hmrishav Bandyopadhyay, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Tao Xiang, Timothy Hospedales, and Yi-Zhe Song. Sketchinr: A first look into sketches as implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12565-12574, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 331, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 296, + 397 + ], + "type": "text", + "content": "[3] Hmrishav Bandyopadhyay, Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. What sketch explainability really means for downstream tasks? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10997-11008, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 399, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 296, + 463 + ], + "type": "text", + "content": "[4] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. In European Conference on Computer Vision, pages 338-355. Springer, 2022. 1, 3, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 465, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 465, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 61, + 465, + 296, + 510 + ], + "type": "text", + "content": "[5] Nan Cao, Xin Yan, Yang Shi, and Chaoran Chen. Aisketcher: a deep generative model for producing high-quality sketches. In Proceedings of the AAAI conference on artificial intelligence, pages 2564-2571, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 510, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 510, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 61, + 510, + 296, + 565 + ], + "type": "text", + "content": "[6] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 567, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 567, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 61, + 567, + 296, + 622 + ], + "type": "text", + "content": "[7] Chenjian Gao, Xin Wang, Qian Yu, Lu Sheng, Jing Zhang, Xiaoguang Han, Yi-Zhe Song, and Dong Xu. 3d reconstruction from a single sketch via view-dependent depth sampling. IEEE transactions on pattern analysis and machine intelligence, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 624, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 624, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 61, + 624, + 296, + 656 + ], + "type": "text", + "content": "[8] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creativ sketch generation. In International Conference on Learning Representations, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 657, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 657, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 61, + 657, + 296, + 689 + ], + "type": "text", + "content": "[9] David Ha and Douglas Eck. A neural representation of sketch drawings. arXiv preprint arXiv:1704.03477, 2017. 2,8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "text", + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 106, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 554, + 138 + ], + "type": "text", + "content": "[11] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "type": "text", + "content": "[12] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-image diffusion models are great sketch-photo matchmakers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16826-16837, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 554, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 270 + ], + "type": "text", + "content": "[13] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16859-16869, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 554, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 335 + ], + "type": "text", + "content": "[14] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to handle sketch-abstraction in sketch-based image retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16859-16869, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[15] Ke Li, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Honggang Zhang. Toward deep universal sketch perceptual grouper. IEEE Transactions on Image Processing, 28(7):3219-3231, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[16] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning. IEEE computer graphics and applications, 39(2):38-51, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 416, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 470 + ], + "type": "text", + "content": "[17] Rosanne Liu, Joel Lehman, Piero Molino, Felipe Petroski Such, Eric Frank, Alex Sergeev, and Jason Yosinski. An intriguing failing of convolutional neural networks and the coordconv solution. Advances in neural information processing systems, 31, 2018. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "type": "text", + "content": "[18] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "type": "text", + "content": "[19] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnN for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 554, + 592 + ], + "type": "text", + "content": "[20] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. Sketchxai: A first look at explainability for human sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23327-23337, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 593, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 554, + 657 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "type": "text", + "content": "[22] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14153-14162, 2020. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19319" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[23] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8504–8513, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "text", + "content": "[24] Aneeshan Sain, Pinaki Nath Chowdhury, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Freeview sketching: View-aware fine-grained sketch-based image retrieval. In European Conference on Computer Vision, pages 145-162. Springer, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 239 + ], + "type": "text", + "content": "[25] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 295, + 296 + ], + "type": "text", + "content": "[26] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, et al. Strokenuwa: Tokenizing strokes for vector graphic synthesis. arXiv preprint arXiv:2401.17093, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 342 + ], + "type": "text", + "content": "[27] Chaitat Utintu, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Ayan Kumar Bhunia, and Yi-Zhe Song. Sketchdeco: Decorating b&w sketches with colour. arXiv preprint arXiv:2405.18716, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 374 + ], + "type": "text", + "content": "[28] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 398 + ], + "type": "text", + "content": "[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 453 + ], + "type": "text", + "content": "[30] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 488 + ], + "type": "text", + "content": "[31] Fei Wang, Shujin Lin, Hanhui Li, Hefeng Wu, Tie Cai, Xiaonan Luo, and Ruomei Wang. Multi-column point-cnn for sketch segmentation. Neurocomputing, 392:50-59, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 543 + ], + "type": "text", + "content": "[32] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3679-3688, 2024. 1, 2, 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 589 + ], + "type": "text", + "content": "[33] Qiang Wang, Haoge Deng, Yonggang Qi, Da Li, and Yi-Zhe Song. Sketchknitter: Vectorized sketch generation with diffusion models. In The Eleventh International Conference on Learning Representations, 2023. 1, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "type": "text", + "content": "[34] Xingyuan Wu, Yonggang Qi, Jun Liu, and Jie Yang. Sketch-segnet: A rnn model for labeling sketch strokes. In 2018 IEEE 28th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[35] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. Advances in Neural Information Processing Systems, 36:15869-15889, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "text", + "content": "[36] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 398 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "sketch: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(1):285-312, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "type": "text", + "content": "[37] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. Sketchgnn: Semantic sketch segmentation with graph neural networks. ACM Transactions on Graphics (TOG), 40(3):1-13, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 174 + ], + "type": "text", + "content": "[38] Ningyuan Zheng, Yifan Jiang, and Dingjiang Huang. Strokenet: A neural painting environment. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 219 + ], + "type": "text", + "content": "[39] Yixiao Zheng, Jiyang Xie, Aneeshan Sain, Yi-Zhe Song, and Zhanyu Ma. Sketch-segformer: Transformer-based segmentation for figurative and creative sketches. IEEE Transactions on Image Processing, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 263 + ], + "type": "text", + "content": "[40] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[41] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to doodle with stroke demonstrations and deep q-networks. In BMVC, page 13, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "text", + "content": "[42] Xianyi Zhu, Yi Xiao, and Yan Zheng. Part-level sketch segmentation and labeling using dual-cnn. In Neural Information Processing: 25th International Conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16, 2018, Proceedings, Part I 25, pages 374-384. Springer, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "text", + "content": "[43] Xianyi Zhu, Yi Xiao, and Yan Zheng. 2d freehand sketch labeling using cnn and crf. Multimedia Tools and Applications, 79(1-2):1585–1602, 2020. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19320" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_content_list.json b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8d57ba4daf6073aaf1eef88c3636bd3ccdf27640 --- /dev/null +++ b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_content_list.json @@ -0,0 +1,1728 @@ +[ + { + "type": "text", + "text": "VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers", + "text_level": 1, + "bbox": [ + 168, + 128, + 831, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yating Wang $^{12}$ Haoyi Zhu $^{23}$ Mingyu Liu $^{24}$ Jange Yang $^{25}$ Hao-Shu Fang $^{6}$ Tong He $^{27\\dagger}$ $^{1}$ Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University \n $^{2}$ Shanghai AI Lab USTC ZJU NJU SJTU SII Corresponding Author", + "bbox": [ + 112, + 208, + 883, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://xiaoxiao0406.github.io/vqvla.github.io", + "bbox": [ + 222, + 287, + 769, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 339, + 326, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce an innovative vector quantization based action tokenizer built upon the largest-scale action trajectory dataset to date, leveraging over 100 times more data than previous approaches. This extensive dataset enables our tokenizer to capture rich spatiotemporal dynamics, resulting in a model that not only accelerates inference but also generates smoother and more coherent action outputs. Once trained, the tokenizer can be seamlessly adapted to a wide range of downstream tasks in a zero-shot manner, from short-horizon reactive behaviors to long-horizon planning. A key finding of our work is that the domain gap between synthetic and real action trajectories is marginal, allowing us to effectively utilize a vast amount of synthetic data during training without compromising real-world performance. To validate our approach, we conducted extensive experiments in both simulated environments and on real robotic platforms. The results demonstrate that as the volume of synthetic trajectory data increases, the performance of our tokenizer on downstream tasks improves significantly—most notably, achieving up to a $30\\%$ higher success rate on two real-world tasks in long-horizon scenarios. These findings highlight the potential of our action tokenizer as a robust and scalable solution for real-time embodied intelligence systems, paving the way for more efficient and reliable robotic control in diverse application domains.", + "bbox": [ + 88, + 373, + 485, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 223, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tokenization plays a critical role in recent generative models, including large language models (LLMs) [1, 31], image and video generation models [17, 23, 39], and vision-language-action (VLA) models [8, 26]. One of the key benefits of tokenization is that it compresses the input space. By reducing high-dimensional continuous data into a compact sequence", + "bbox": [ + 89, + 809, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of tokens, the complexity of the learning task is substantially reduced. Compared to image patches and language tokens, action sequences are inherently easier to compress because of their spatio-temporal continuity.", + "bbox": [ + 511, + 340, + 906, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several studies have started to explore action quantized tokenization for Vision-Language-Action (VLA) models [5, 35], demonstrating promising potential. Effective action tokenization not only significantly enhances the downstream performance of VLA models, especially for tasks involving long-horizon planning, but also markedly improves their training and inference efficiency. In this paper, we delve deeper into the potential of action tokenization, with a specific emphasis on its scalability and accuracy. We first find that the more precise the tokenization, the more pronounced the improvements in long-horizon action modeling. This motivates us to train the VQ tokenizer with well-scaled action trajectories to cover various tasks. Additionally, we observe that action trajectories, unlike visual and physical modalities, exhibit minimal domain gaps between real-world and simulated environments. This characteristic enables effective scaling of action tokenizers through the extensive use of synthetic action data. Moreover, training action tokenizers is computationally lightweight compared to scaling entire VLA models. Consequently, focusing on scaling action tokenization emerges as a highly cost-effective strategy, requiring fewer computational and data resources while delivering significant performance enhancements.", + "bbox": [ + 511, + 402, + 908, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, we propose a convolutional residual VQ-VAE [5, 28, 52] framework for training action tokenizers. To effectively train the model, we propose a progressive training strategy: Initially, we train the tokenizer on real-world robotic datasets, such as OpenX-Embodiment [34], which typically contain noisy and jittery trajectories. Subsequently, we gradually integrate cleaner and smoother synthetic data from large-scale simulated robotic datasets, such as LIBERO [29] and ManiSkill [33]. This progressive approach allows the VQ model to converge toward smoother", + "bbox": [ + 511, + 750, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11089", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and more stable representations. Compared to previous approaches that typically rely on training with single-task datasets, our method expands the tokenizer training dataset by more than 100 times, effectively covering a broad spectrum of downstream tasks.", + "bbox": [ + 89, + 90, + 483, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments in both simulated and real-world environments. First, we evaluate our VQ-VAE action tokenizers in the LIBERO simulator, where the results demonstrate the effectiveness of the convolutional residual VQ-VAE and provide preliminary validation for the hypothesis that the VQ-VAE action tokenizer can leverage synthetic data for scaling. Additionally, real-world experiments with a Franka Research 3 robot further validate the superiority of our approach. Specifically, our findings are as follows: (1) as the amount of simulated action data increases, the VQ-VAE tokenizers exhibit linear scaling properties in improving VLA success rates; (2) the tokenizers significantly enhance inference speed and smoothness of VLA models; and (3) the tokenizers effectively reduce cumulative errors, enabling better performance in long-horizon tasks.", + "bbox": [ + 89, + 167, + 485, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 393, + 410, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a general convolutional residual VQ-VAE-based framework for action tokenizers.", + "- We demonstrate that action tokenizers can be effectively scaled by leveraging large-scale simulated action data.", + "- We prove that our action tokenizers improve the performance, inference speed, and long-horizon capabilities of VLA models." + ], + "bbox": [ + 91, + 407, + 482, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 527, + 240, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language-Action Models. Vision-Language-Action (VLA) models[7, 9, 11, 18, 26, 41, 45, 46, 56, 57] bridge visual-language understanding with robot control by mapping multimodal inputs to action outputs. Based on vision-language models (VLM) [3, 10, 11, 15], VLAs represent robot actions (e.g., 6DoF motion, gripper control) as discrete tokens compatible with text-based output. For example, RT-1[7] and RT-2[8] showed that dividing continuous actions into discrete bins allows integration with VLMs, enabling zero-shot generalization using web-scale pretraining. Recent works[26, 34] further leverage this approach, demonstrating improved task generalization through large VLM backbones. Building on this foundation, recent works have explored various approaches to improve VLA, such as leveraging 3D visual inputs [57], integrating chain-of-thought reasoning [51], and employing parallel decoding strategies [27]. In our work, we aim to simultaneously improve the performance and execution speed of VLA.", + "bbox": [ + 89, + 551, + 485, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tokenization. The tokenizer is one of the fundamental components in language and vision generation tasks, especially for autoregressive transformer-based generation models. For language generation tasks, mainstream methods adopt byte pair encoding (BPE) [19, 36] to compress input text. For", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vision generation tasks, recent works [17, 30, 37, 39, 53, 54] mainly use vector quantization [42] for tokenization to map continuous visual signals into a discrete token sequence. Similarly, in robot learning, many works [7, 8, 28, 32, 35] model action prediction as a generative problem and tokenize the robot action modality to capture the multi-modal distribution within skills and adapt it to autoregressive generative policies. In our work, we propose a general convolutional residual VQ-VAE to tokenize robot action. Different from several related works [28, 32] that use VQ-VAE to tokenize robot actions, our VQ-VAE is trained on large-scale data and is applicable to all tasks.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Action Representation. Action representation is one of the core components in robot policy learning. Classic action representation schemes include high-level sub-tasks [2, 15], action primitives [22, 43], keypoints [14, 20], as well as low-level continuous end-effector pose [29, 34] or joint state [21]. Recently, imitation learning-based end-to-end approaches [7, 13, 55, 60, 61] often utilize the latter. Some works[6, 40, 45, 58] employ regression-based output layers or integrate new weights for diffusion decoding for action output. Additionally, some works extract corresponding predictive signals from action-less video data to represent actions, such as pixels [9, 47], trajectories [44, 48], and latent motion [12, 49, 50]. To represent complex multi-modal distributions and adapt to discrete generative models, many works discretize continuous low-level actions. Common approaches include per-dimension and per-timestep binning discretization [7, 26], which struggles to perform well in high-frequency tasks [35]. Another method is VQ-VAE based discretization [4, 5, 28, 32, 59]; a limitation here is the need to train a separate action tokenizer for each specific task. Fast [35] utilized a discrete cosine transform (DCT) to learn a non-learnable action tokenizer from real-robot data. In this work, we propose a novel approach: expanding large-scale synthetic data to train the VQ-VAE action tokenizer, aiming to enhance its generalizability and efficiency.", + "bbox": [ + 511, + 273, + 908, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 669, + 611, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries: VLA models.", + "text_level": 1, + "bbox": [ + 511, + 695, + 767, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use OpenVLA[26] as our backbone model. OpenVLA's origin formulation is to adopt a discrete tokenization strategy for robot action prediction through fine-tuning the Prismatic-7B VLM backbone. The method frames action prediction as a vision-language task, mapping input observation images and natural language instructions to discrete robot action sequences. Specifically, continuous robot actions are discretized into 256 bins per dimension, with bin boundaries determined by the 1st and 99th percentiles of training data distributions rather than min-max ranges to mitigate outlier effects. This discretization converts N-dimensional actions into N discrete integers (0-255). To embed these into the", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11090", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f77b0c565c1e8248e90c4431c198b617e3caa555a7fad3f1e489518773b4c709.jpg", + "image_caption": [ + "Figure 1. The VQ-VLA pipeline, consisting of two main stages: (1) training a general convolutional residual VQ-VAE and (2) fine-tuning OpenVLA using the LoRA approach. Specifically, a general convolutional residual VQ-VAE is first trained on the Open X-Embodiment dataset, LIEBRO, and ManiSkill datasets. The trained VQ-VAE is then frozen and serves as an action tokenizer for OpenVLA, replacing the simple binning method. In the second stage, OpenVLA is fine-tuned using the LoRA technique to optimize its performance." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 339, + 398 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/11f5fe7ecf165e4c1f5fa22894f6e01a9b2087c70ec6e29a3399938973ffd06a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 87, + 903, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM's vocabulary, OpenVLA overwrites the 256 least-used tokens in the Llama tokenizer (last 256 tokens) rather than using special tokens, as the original tokenizer only reserves 100 special tokens.", + "bbox": [ + 88, + 489, + 482, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Action Tokenizer via Residual VQ-VAE", + "text_level": 1, + "bbox": [ + 89, + 559, + 429, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on Residual VQ-VAE [5, 28, 52], we design our encoder and decoder inspired by VAE in pyramidal flow matching [25]. Instead of the simple Multi-Layer Perceptron (MLP) used in Residual VQ-VAE [28, 52], we integrate 2D temporal convolutional layers, motivated by their ability to efficiently capture local relationships and hierarchical temporal dependencies, addressing the scaling limitations of MLPs.", + "bbox": [ + 88, + 580, + 482, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input action sequence $\\mathbf{a}_{t:t + n}\\in \\mathbb{R}^{n\\times d}$ , where $n$ is the sequence length and $d$ the action dimensionality, the encoder $\\phi_{\\mathrm{enc}}$ , composed of 2D temporal convolutional layers, transforms the sequence into a latent embedding $\\mathbf{x}\\in \\mathbb{R}^k$ , expressed as $\\mathbf{x} = \\phi_{\\mathrm{enc}}(\\mathbf{a}_{t:t + n})$ . To compress $\\mathbf{x}$ , we apply Residual Vector Quantization (RVQ) [52], decomposing $\\mathbf{x}$ into quantized residuals: $\\mathbf{q}(\\mathbf{x}) = \\sum_{i = 1}^{N_q}\\mathbf{q}_i(\\mathbf{r}_i)$ , where $\\mathbf{r}_1 = \\mathbf{x}$ , $\\mathbf{r}_{i + 1} = \\mathbf{r}_i - \\mathbf{q}_i(\\mathbf{r}_i)$ , and $N_{q}$ denotes the number of quantization stages.", + "bbox": [ + 89, + 702, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The quantized embedding $\\mathbf{q}(\\mathbf{x})$ is passed through the decoder $\\phi_{\\mathrm{dec}}$ , which uses 2D temporal deconvolutional layers to reconstruct the sequence $\\hat{\\mathbf{a}}_{t:t+n}$ , ensuring temporal structure preservation: $\\hat{\\mathbf{a}}_{t:t+n} = \\phi_{\\mathrm{dec}}(\\mathbf{q}(\\mathbf{x}))$ .", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the framework, we minimize the total loss $\\mathcal{L}$ , a weighted combination of reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ , vector quantization (VQ) loss $\\mathcal{L}_{\\mathrm{codebook}}$ , and commitment loss $\\mathcal{L}_{\\mathrm{commit}}$ :", + "bbox": [ + 511, + 489, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\left\\| \\mathbf {a} _ {t: t + n} - \\hat {\\mathbf {a}} _ {t: t + n} \\right\\| _ {2} ^ {2} \\tag {1} \\\\ + \\lambda \\left(\\| \\operatorname {s g} (\\mathbf {x}) - \\mathbf {q} (\\mathbf {x}) \\| _ {2} ^ {2} + \\| \\mathbf {x} - \\operatorname {s g} (\\mathbf {q} (\\mathbf {x})) \\| _ {2} ^ {2}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 560, + 905, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}(\\cdot)$ denotes the stop-gradient operation, and $\\lambda$ balances the loss components. We set $\\lambda = 4$ in our experiments. This design ensures efficient, scalable, and robust encoding of structured temporal data.", + "bbox": [ + 511, + 609, + 906, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Training Residual VQ-VAE", + "text_level": 1, + "bbox": [ + 511, + 679, + 761, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train three versions of the Residual VQ-VAE to evaluate the scaling capability of our action tokenizer: (1) using only the Open X-Embodiment dataset[34], (2) combining Open X-Embodiment and Libero datasets[29], and (3) combining Open X-Embodiment, Libero, and ManiSkill[33] datasets. These experiments were designed to test our hypothesis that the action tokenizer can effectively scale with simulated ata.", + "bbox": [ + 511, + 700, + 906, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve the encoder's ability to process temporal and spatial information, we introduced two types of embeddings before the action sequences are passed into the encoder:", + "bbox": [ + 511, + 806, + 906, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Time Embedding: A sinusoidal time embedding was added to encode temporal information at varying frequencies. This embedding allows the model to capture both", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11091", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "low-frequency and high-frequency temporal patterns in the input actions, improving its ability to represent fine-grained temporal details.", + "bbox": [ + 102, + 90, + 485, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- **Action-Type Embedding:** Learnable embeddings were added for the different components of the action sequence (e.g., XYZ positions, Euler angles, and gripper states). Since the 7 dimensions of the action vector have distinct meanings, this embedding provides a strong prior for the model, helping it distinguish and process the unique roles of each dimension more effectively.", + "bbox": [ + 91, + 136, + 482, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The use of these embeddings enhances the encoder's ability to process structured data, improving the quality of the latent representations and the overall performance of the tokenizer.", + "bbox": [ + 89, + 244, + 483, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train a more universal robot action tokenizer and reduce computational overhead, the model is trained using only action sequences as input, without additional conditional inputs. This design reduces complexity while maintaining the generalizability of the tokenizer. All models are trained on a single A100 GPU. For example, training on the Open X-Embodiment dataset requires just one A100 GPU and is completed in one week.", + "bbox": [ + 89, + 306, + 483, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Integrating Residual VQ-VAE as Action Tokenizer in VLA", + "text_level": 1, + "bbox": [ + 89, + 438, + 483, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we replace the simple bin-based action tokenizer used in OpenVLA with a Residual VQ-VAE-based action tokenizer to improve the expressiveness and precision of action tokenization. Instead of discretizing action sequences into uniform bins, the action sequence $\\mathbf{a}_{t:t + n}$ is first processed through a pre-trained and frozen Residual VQ-VAE encoder $\\phi (\\cdot)$ , generating latent representations. These representations are then quantized into discrete codebook indices (tokens) $\\{z_q^i\\}_{i = 1}^{N_q}$ corresponding to $N_{q}$ Residual VQ layers. The quantized tokens $z_{q}^{i}$ are used as the action tokens for training and prediction in the Vision-Language Model (VLM).", + "bbox": [ + 89, + 476, + 483, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike OpenVLA, where all token IDs are mapped to the range [0, 255], the token IDs generated by different VQ layers in the Residual VQ-VAE are assigned unique, non-overlapping ranges. Specifically, tokens from the $i$ -th VQ layer are offset by $(i - 1) \\times 256$ , ensuring that:", + "bbox": [ + 89, + 661, + 483, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {q} ^ {i} \\in [ 2 5 6 \\times (i - 1), 2 5 6 \\times i - 1 ], \\quad \\forall i \\in \\{1, \\ldots , N _ {q} \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 748, + 470, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For example, tokens from the first layer are in the range [0, 255], those from the second layer are in [256, 511], and so on. This design avoids conflicts between token IDs from different layers, as tokens with the same ID in different layers represent semantically different features of the action space.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Vision-Language Model (VLM) is trained to predict these tokens directly. The loss function is the standard next-token prediction loss, computed as the cross-entropy", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between the predicted token distribution $\\hat{z}_q^i$ and the ground truth token $z_q^i$ produced by the frozen Residual VQ-VAE:", + "bbox": [ + 511, + 90, + 906, + 125 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {V L M}} = - \\sum_ {i = 1} ^ {N _ {q}} \\log P \\left(\\hat {z} _ {q} ^ {i} = z _ {q} ^ {i} \\mid o _ {t - h: t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 132, + 906, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $o_{t - h:t}$ represents the input observation sequence, and $N_{q}$ is the number of Residual VQ layers.", + "bbox": [ + 511, + 180, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further enhance computational efficiency, similar to OpenVLA, the least-used tokens in the vocabulary are replaced during fine-tuning. However, this replacement process respects the layer-specific token ID ranges, ensuring that the tokenization remains consistent and interpretable across all layers.", + "bbox": [ + 511, + 210, + 906, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This integration of a pre-trained and frozen Residual VQ-VAE as the action tokenizer in VLA enables a more expressive and scalable representation of actions. By leveraging hierarchical quantization with non-overlapping token ID ranges, the model achieves better action representation, avoids semantic confusion between layers, and ensures stable loss convergence during training. Furthermore, by using the VQ-VAE tokens to represent longer action sequences instead of predicting one action at a time, the model significantly reduces the number of tokens required for training and inference. This mechanism allows the model to predict more complex behavior sequences with fewer steps, leading to a substantial improvement in inference speed and computational efficiency.", + "bbox": [ + 511, + 301, + 908, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 523, + 645, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments, we first validate the effectiveness and scalability of the action tokenizer in VLA models using LIBERO simulator[29]. Subsequently, real-world experiments are conducted to further verify our hypothesis. We also investigate the impact of action tokenizers on the performance, inference speed, and long-horizon capabilities of VLA models, alongside ablation studies to evaluate key design choices.", + "bbox": [ + 511, + 549, + 908, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Simulation Experiments", + "text_level": 1, + "bbox": [ + 511, + 678, + 736, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 700, + 687, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize the LIBERO benchmark[29] to validate and evaluate the effectiveness and scalability of the action tokenizer, using the Franka Panda robot. Specifically, the entire LIBERO task suite—including LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-10, and LIBERO-90—is used as the entire LIBERO dataset. Among these, LIBERO-90 comprises 90 short-horizon tasks, while the other task suites each contain 10 tasks, with 50 demonstrations per task. For evaluation, testing is conducted on LIBERO-goal, LIBERO-10 and LIBERO-90.", + "bbox": [ + 511, + 719, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the approach in OpenVLA[26], we filter out all \"no-op\" actions from the dataset. To preliminarily validate", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11092", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the effectiveness of using VQ-VAE as an action tokenizer in VLA models and its scalability, we train two versions of a convolutional residual VQ-VAE. The first version $\\mathrm{VQ_M}$ is trained solely on the Maniskill dataset[33], and the second $\\mathrm{VQ_{M + R}}$ is trained on a mixture of Maniskill dataset and RLBench dataset[24]. Both models are trained on a single A100 GPU with a batch size of 1024, which takes about only 1 week.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use two pre-trained VQ-VAE models as frozen action tokenizers for OpenVLA: $\\mathrm{VQ_M}$ and $\\mathrm{VQ_{M + R}}$ . LoRA is then applied to fine-tune OpenVLA on the LIBERO-90 dataset. Additionally, we fine-tune the original OpenVLA model on the LIBERO-90 dataset using LoRA as a baseline for comparison. For a fair comparison, all fine-tuning on the LIBERO-90 task suite is conducted for 400K gradient steps with a batch size of 4, using 4 A100-80GB GPUs and an action chunk length of $K = 5$ .", + "bbox": [ + 89, + 214, + 480, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Effectiveness of Conv Residual VQ-VAE", + "text_level": 1, + "bbox": [ + 89, + 363, + 415, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the early stages of the experiment, we selected two task suites, LIBERO-10 and LIBERO-GOAL, to conduct preliminary scaling validation experiments. Specifically, we used two variants of Residual VQ-VAE models: one with a simple MLP as the encoder and decoder, and the other with a larger 2D temporal convolutional network as the encoder and decoder. The results are summarized in Tab. 1. The findings indicate that using temporal convolutional networks as the encoder and decoder in the Residual VQ-VAE significantly outperforms the MLP-based architecture in terms of success rate. These results suggest that temporal convolutional networks can serve as an effective action tokenizer, capable of capturing temporal dependencies more effectively.", + "bbox": [ + 89, + 383, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, when we increased the training data for the VQ-VAE model, transitioning from data solely derived from individual LIBERO tasks to the entire LIBERO dataset, the success rate consistently improved. This observation provides preliminary evidence supporting the scalability of our action tokenizer design.", + "bbox": [ + 89, + 583, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/aaaf350bf8e655ad281ddaa6ba563d8def58b240dc8ba68d5f792c5302c7ec0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training Dataset of VQLIBERO-10 (%)LIBERO-GOAL (%)
Original OpenVLA-51.075.8
MLP Residual VQ-VAEALL-LIBERO53.472.6
LIBERO-1053.2-
LIBERO-GOAL-65.2
Conv Residual VQ-VAEALL-LIBERO60.075.2
LIBERO-1054.0-
LIBERO-GOAL-72.4
", + "bbox": [ + 91, + 689, + 480, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3. Scaling Data Improves VQ-VAE Action Tokenizer Performance", + "text_level": 1, + "bbox": [ + 513, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the experiment, to avoid potential performance inflation from in-domain data during evaluation on LIBERO-90, we trained the VQ-VAE Action Tokenizer on completely out-of-domain ManiSkill and RLBench simulation data, rather than LIBERO data. Specifically, ManiSkill contains 30,000 action sequences, and RLBench includes 10,000 action sequences. We compare three models: the baseline, which fine-tunes the original OpenVLA; $\\mathrm{VQ_M}$ , which uses a VQ-VAE trained on the Maniskill dataset as the action tokenizer for OpenVLA; and $\\mathrm{VQ_{M + R}}$ , which uses a VQ-VAE trained on a mixture of the Maniskill and RLBench datasets as the action tokenizer for OpenVLA.", + "bbox": [ + 511, + 125, + 906, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On LIBERO-90, $\\mathrm{VQ}_{\\mathrm{M} + \\mathrm{R}}$ achieved $80.98\\%$ , a $7.45\\%$ improvement over the OpenVLA baseline (73.53%), Tab. 2. Furthermore, an ablation study using only ManiSkill data for training $\\mathrm{VQ_M}$ resulted in substantially lower performance, underscoring the critical role of sufficient synthetic data scale.", + "bbox": [ + 511, + 306, + 906, + 397 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/38fdfe216970c7631b48275689770ef2026dc55a93435423548e145d48c2213c.jpg", + "table_caption": [ + "Table 1. The evaluation results of residual VQ-VAE architectures. The results demonstrate that the Conv Residual VQ-VAE outperforms the MLP-based version, particularly when trained on the full LIBERO dataset (ALL-LIBERO), highlighting its ability to better capture temporal dependencies and improve success rates." + ], + "table_footnote": [], + "table_body": "
baseline(%)VQM(%)VQM+R(%)
LIEBRO-9073.5314.3880.98
", + "bbox": [ + 581, + 410, + 836, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Effectiveness of VQ-VAE Action Tokenizers in Scaling Simulation Data. The results demonstrate $\\mathrm{VQ}_{\\mathrm{M} + \\mathrm{R}}$ reached 80.98%, outperforming the OpenVLA baseline by 7.45%", + "bbox": [ + 511, + 463, + 906, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Real-Word Experiment", + "text_level": 1, + "bbox": [ + 513, + 537, + 728, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 513, + 560, + 686, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our robotic platform consists of a single Franka Research3 arm equipped with a third-person-view RealSense D435 camera mounted in a fixed position to capture environmental observations. The system operates at $20\\mathrm{Hz}$ (moderately reduced from the native $100\\mathrm{Hz}$ control frequency to balance training efficiency and motion continuity), with actions defined as absolute end-effector poses in SE(3) space (3D position + quaternion orientation).", + "bbox": [ + 511, + 580, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our experimental benchmark comprises six manipulation tasks (4 short-horizon tasks, 2 long-horizon tasks) designed to evaluate the model's ability to handle varying task complexities. For each task, we collect 50 demonstrations and evaluate performance over 20 trials:", + "bbox": [ + 511, + 702, + 905, + 777 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Pull out a tissue paper: The robot need to grasp and pull out a single tissue paper.", + "2) Pick up the [TOY NAME]: The robot is required to pick up the specific toy, including the toy snake, the toy eggplant, and the toy chicken, resulting in a total of three tasks (with no other distractions).", + "3) Put the toy into the basket: The robot need to pick up the toy(no other distractions) and put it into the basket." + ], + "bbox": [ + 511, + 777, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11093", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/efd59b2c1b974854cfccaa8958f9dbb0acb7d1164c031ebbdd600212dcb12473.jpg", + "image_caption": [ + "Libero 90" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 220, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5941664259900e8d7070eeac30f5a567a288309c1c4327f96c1c7aff3428dfe7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 90, + 349, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/26aae533fdde15bdac7d45a8aaf00d7b39292b4631a79c61c5fbe30fe4167b89.jpg", + "image_caption": [ + "Put the toy into the basket" + ], + "image_footnote": [], + "bbox": [ + 352, + 90, + 478, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0e83fe6cd6f42da81a249074e813a289863301b33d508c2e042c663f69746fd.jpg", + "image_caption": [ + "Put all cups into the basket" + ], + "image_footnote": [], + "bbox": [ + 94, + 229, + 220, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c06351d4aff8518f0767b7dfe7064bf1e5db23b6bb41dd6f8122df474d269994.jpg", + "image_caption": [ + "Pull out a tissue paper", + "Put the toy into the drawer" + ], + "image_footnote": [], + "bbox": [ + 223, + 229, + 349, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03969043ed6fbc0a6e2fcf0d629a6b27626f552c8bfa7fb1c17d5b4e9d65e72b.jpg", + "image_caption": [ + "Filp the pot upright" + ], + "image_footnote": [], + "bbox": [ + 354, + 229, + 480, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f8f58540c7da9cb63591941aedebeb95bfb1cd47247b090f2213a4060749113.jpg", + "image_caption": [ + "Figure 2. All Evaluation environments: We conduct comprehensive evaluations of VQ-VLA in both simulation and real-world settings. In simulation, evaluations are performed on the LIBERO-90 benchmark within the LIBERO dataset. And six diverse tasks are designed for real-world testing." + ], + "image_footnote": [], + "bbox": [ + 94, + 373, + 220, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f09c0454d1293fb66575d628a2460b19e6946fdf52081c170ff7310c53f8171.jpg", + "image_caption": [ + "Pick up the{toy name}" + ], + "image_footnote": [], + "bbox": [ + 223, + 373, + 349, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f2b73ab429478a6d76663c1e41113736533ac73b5d281a6707ee9d05d2ad6ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 373, + 480, + 503 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4) Flip the pot upright: We set a flipped pot on the platform, the robot need to flip and upright a fallen cooking pot.", + "5) Put all cups into the basket: We place two different cups on the table and set a few other things (toys) as distractions. The robot need to sequentially put two cups into the basket, testing long-horizon task.", + "6) Put the toy into the drawer: We place a toy on the table and set a few other things as distractions. The robot need to sequentially open the drawer, pick up the specific toy and put it into the drawer, and close the drawer, testing long-horizon task." + ], + "bbox": [ + 89, + 645, + 483, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We finetune each task separately for 100K gradient steps (batch size 4 across 4 A100-80GB GPUs) with action chunk length $\\mathrm{K} = 5$ . During deployment.", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Performance on Short-Horizon Tasks", + "text_level": 1, + "bbox": [ + 513, + 90, + 818, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the in the simulation experiments, we further incorporated 120k synthetic trajectories from the ManiSkill dataset, co-training $\\mathrm{VQ_{O + L + M}}$ with Open X-Embodiment dataset and LIBERO dataset to evaluate the scaling capability of our action tokenizer. We evaluated the performance of the action tokenizer trained with different datasets on four carefully designed short-horizon tasks, focusing on multi-task learning, precision manipulation, and dynamic adjustment capabilities. The results are shown in Fig. 3.", + "bbox": [ + 511, + 111, + 906, + 246 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It can be observed that when synthetic trajectories are added for co-training, the performance of the action tokenizer improves significantly. The average success rate increased from the baseline of $23\\%$ to $46.25\\%$ , with a notable improvement of $30\\%$ in the \"Flip the pot upright\" task. In the \"Pull out a tissue paper\" task, which tests the robot's performance in high-precision dynamic operations (as this task requires continuous, fine-grained grasping and pulling motions), the baseline model achieved only a $5\\%$ success rate, failing almost entirely. In contrast, models incorporating VQ achieved success rates of $20\\%$ or higher.", + "bbox": [ + 511, + 248, + 908, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, comparing $\\mathrm{VQ_{O + L}}$ with $\\mathrm{VQ_O}$ , the average success rate on short-horizon tasks increased by only $0.5\\%$ . This may be attributed to the limited size of the LIBERO dataset, which has minimal impact on short-horizon tasks. However, the ManiSkill dataset is 50 times larger than LIBERO, leading to a significant improvement in success rates when used for training.", + "bbox": [ + 511, + 415, + 906, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3. Performance on Long-Horizon Tasks", + "text_level": 1, + "bbox": [ + 511, + 532, + 816, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VQ-VLA demonstrates outstanding performance on long-horizon tasks (\"Put all cups in the basket\" and \"Put the toy into the drawer\"), significantly outperforming baseline model in both success rate and efficiency. In scenarios where baseline models achieve success rates as low as $15\\%$ or nearly 0, the $\\mathrm{VQ}_{O + L + M}$ model achieves significantly higher success rates of $50\\%$ and $30\\%$ , respectively (see Fig.3). For the \"Put the toy into the drawer\" task, a representative long-horizon scenario, the baseline model was only able to complete the first step of opening the drawer in 15 out of 20 trials, failing to proceed further in most cases. In contrast, the $\\mathrm{VQ}_{O + L + M}$ model successfully opened the drawer in all test cases, demonstrating its robustness and reliability in handling complex sequential tasks.", + "bbox": [ + 511, + 551, + 908, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A key advantage of VQ-VLA in long-horizon tasks lies in its use of VQ-VAE as an action tokenizer, which enables the model to predict multiple actions in a single inference step. This design reduces the accumulation of errors over extended task sequences, making VQ-VLA particularly advantageous for tasks requiring long-term planning and execution. Additionally, in scenarios involving the consecutive execution of multiple subtasks, VQ-VLA not only achieves higher success rates but also significantly reduces task completion time", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11094", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ad737e1f2527b4c3e74f5e553ca2dcb533b8594001f7a084d70e9c845c2144c.jpg", + "image_caption": [ + "Figure 3. Real-world experimental results: We compare the performance of Baseline, $\\mathrm{VQ_0}$ , $\\mathrm{VQ_{O + L}}$ , and $\\mathrm{VQ_{O + L + M}}$ on both short-horizon and long-horizon tasks. In terms of the average success rate, all VQ-based models outperform the Baseline. The best-performing model, $\\mathrm{VQ_{O + L + M}}$ , achieves a success rate that is $23.25\\%$ higher than the Baseline on both short-horizon and long-horizon tasks. Additionally, the results show that $\\mathrm{VQ_{O + L + M}}$ outperforms $\\mathrm{VQ_{O + L}}$ , which in turn outperforms $\\mathrm{VQ_0}$ , indicating the effectiveness of incorporating synthetic data during training without compromising real-world performance." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to baseline models, highlighting its superior execution efficiency and ability to handle long-horizon challenges effectively.", + "bbox": [ + 88, + 444, + 485, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.4. Sim&Real Domain Gap Analysis", + "text_level": 1, + "bbox": [ + 89, + 511, + 366, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantify the domain gap between synthetic data and real-world data, we trained a VQ-VAE model exclusively on the LIBERO dataset, referred to as $\\mathrm{VQL}$ , for OpenVLA. The model was tested in three real-world tasks (one long-horizon and two short-horizon tasks), and the results are shown in Tab. 3. The results indicate that the performance of $\\mathrm{VQL}$ is comparable to that of both $\\mathrm{VQ_{O + L}}$ and $\\mathrm{VQ_O}$ , suggesting that the domain gap between synthetic and real-world data is minimal.", + "bbox": [ + 88, + 534, + 482, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although real-world data may contain noise, the inclusion of Open X-Embodiment data as a real-world dataset expands the data sources and enriches the diversity of data types, which effectively enhances the model's generalization and robustness.", + "bbox": [ + 88, + 675, + 483, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.5. Inference Speed Comparison", + "text_level": 1, + "bbox": [ + 89, + 771, + 339, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During the real-world experiments, we measured the action execution frequency of VQ-VLA and compared it with the original OpenVLA. The results are summarized in Tab. 4. As shown in the table, with a compression ratio of 5 in VQ-VAE, the inference speed is nearly tripled. This significant improvement greatly facilitates real-time performance in practical applications.", + "bbox": [ + 88, + 794, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/76364bdbb51a172a76961994967d0ef83048bf58db8c1cb313cd7b8df6aa96d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Put the toy into the drawer (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline5.030.020.0
VQo15.045.035.0
VQL10.055.035.0
VQo+L10.045.035.0
VQo+L+M25.060.045.0
", + "bbox": [ + 517, + 441, + 906, + 537 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4fb2ce6c4ce5ef1a3df98d925f31c1cbaecf7d2d3b66caf2e3bfe2a4c8ee39ca.jpg", + "table_caption": [ + "Table 3. Performance Comparison Across Real-World Tasks: We observe that the performance of $\\mathrm{VQ_L}$ is comparable to that of both $\\mathrm{VQ_{O + L}}$ and $\\mathrm{VQ_O}$ , indicating that the domain gap between synthetic and real-world data is minimal." + ], + "table_footnote": [], + "table_body": "
Frequency (Hz)
VQ-VLA11.84
OpenVLA4.16
", + "bbox": [ + 542, + 619, + 877, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. The Results of Frequencies. We report the comparison results of our VQ-VLA and baseline OpenVLA.", + "bbox": [ + 511, + 684, + 906, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 741, + 679, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we report some ablation studies to show the effectiveness of the design choices of our method.", + "bbox": [ + 511, + 763, + 905, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1. Action Chunking via VQ-VAE and Autoregressive Output", + "text_level": 1, + "bbox": [ + 511, + 805, + 905, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The original OpenVLA model generates a single action per inference step based solely on the current observation, resulting in a step-by-step execution approach. To improve efficiency and handle longer action sequences, we extend", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11095", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OpenVLA to output a sequence of five actions in an autoregressive manner, similar to the action chunking mechanism in VQ-VLA. Specifically, the model predicts the next action based on the current observation and previously generated actions, allowing it to generate a sequence of actions in a single inference step. This autoregressive approach serves as an alternative action chunking strategy by grouping multiple actions for execution.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of different action chunking strategies, we design ablation experiments comparing the autoregressive output of OpenVLA to the VQ-based action chunking method in VQ-VLA. The results, shown in Tab. 5, indicate that the use of the autoregressive approach for action chunking in the original OpenVLA leads to a significant drop in success rate compared to the baseline. In real-world experiments, when using the autoregressive approach as a form of action chunking, the spatial magnitude of each action chunk is relatively small, resulting in slower overall execution of actions. Compared to the Action Chunking via VQ-VAE method, even with the same action chunk size, the time required to reach the same target location is significantly longer. Additionally, this approach tends to exhibit shortcut learning, where the model directly copies previous actions within the same action chunk. By analyzing the output values of each action chunk, it is observed that multiple actions within a single chunk have remarkably similar values, indicating a lack of diversity in the predicted actions.", + "bbox": [ + 91, + 212, + 483, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, Action Chunking via VQ-VAE not only improves inference speed but also enhances the performance of VLA by generating more effective and diverse action sequences. This makes it a more suitable approach for addressing real-world long-horizon tasks.", + "bbox": [ + 89, + 500, + 483, + 574 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/51f6d73250ab2dfeb8d794b7c474feaf11aedf6588f1b0c456982a1a131997d6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Action ChunkingLIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline74.7630.020.0
Autoregressive Output66.5310.00.0
VQ-based (VQo+L+M)86.6160.045.0
", + "bbox": [ + 93, + 584, + 480, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Embedding Integration Effectiveness", + "text_level": 1, + "bbox": [ + 89, + 761, + 393, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the impact of embeddings, we conducted an ablation study comparing the model's performance with and without embeddings. Specifically, the baseline model processes raw action sequences directly, while the enhanced model incorporates both time embedding and action-type embedding.", + "bbox": [ + 89, + 779, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 6, the model with embeddings significantly outperforms the baseline in terms of success rate. This", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "demonstrates that the integration of embeddings improves the encoder's ability to represent structured action sequences, leading to better overall performance.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69c0d9b4034a0371af9acdd07f2902b29fd26b48e9b49e92581b13ac0f0898c8.jpg", + "table_caption": [ + "Table 5. Comparison of action chunking methods: We evaluate the performance on three tasks (one simulator task and two real-world tasks). The results show that the Autoregressive Output used in OpenVLA performs poorly as an action chunking method, with a significant gap compared to the VQ-based approach. This indicates that the VQ-based method is more effective." + ], + "table_footnote": [], + "table_body": "
LIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
VQoL (w.o. Embeddings)85.1740.035.0
VQoL (w Embeddings)86.1645.035.0
", + "bbox": [ + 519, + 154, + 906, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Embedding integration improves performance. The table compares models with and without embeddings across three tasks, showing that embeddings enhance success rates, especially for \"Flip the pot upright.\"", + "bbox": [ + 511, + 224, + 906, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Future Works", + "text_level": 1, + "bbox": [ + 513, + 321, + 794, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our work, a series of extensive experiments demonstrate that our proposed action tokenizers can improve VLA performance and inference speed while achieving scalability on simulated data. Despite these promising results, there still remain some limitations and opportunities for future work. First, our action tokenizers can be further extended to larger-scale simulated datasets. Second, our work improves inference speed by decoding multi-step action sequences, which can be further combined with techniques such as distillation [38] and quantization [16] of VLMs in the future. Finally, the architecture design of our action tokenizers can be further improved, for example, by incorporating the frequency of action data as a extra condition. In summary, we hope that our work can serve as a strong baseline and inspire future research.", + "bbox": [ + 511, + 349, + 908, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 597, + 638, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a general convolutional residual VQ-VAE framework for action tokenizers that can seamlessly integrate with state-of-the-art VLA models. Our VQ-VAE is trained on 100 times more data than previous methods and can be directly transferred to downstream both real-world and simulated robotic manipulation tasks. Extensive experiments conducted in both simulated and real-world environments validate that the proposed convolutional residual VQ-VAE framework not only enhances the performance of VLA policies but also accelerates inference. Finally, our VQ-VAE also demonstrates the ability to scale effectively with large-scale simulated data.", + "bbox": [ + 511, + 625, + 908, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 827, + 671, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11096", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2", + "[3] Jean-Baptiste Alayrac, Jeffrey Donahue, Pauline Luc, Alexis Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35:23716-23736, 2022. 2", + "[4] Arthur Allshire, Roberto Martin-Martin, Charles Lin, Shawn Manuel, Silvio Savarese, and Animesh Garg. Laser: Learning a latent action space for efficient reinforcement learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6650-6656, 2021. 2", + "[5] Suneel Belkhale and Dorsa Sadigh. Minivla: A better vla with a smaller footprint. https://ai.stanford.edu/blog/minivla/, 2024. 1, 2, 3", + "[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. $\\pi_0$ : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2", + "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2", + "[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2", + "[9] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 2", + "[10] Wei Chen, Yifeng Zhang, Hao Li, Xufeng Wang, and Ming Liu. Anyvlm: Unified vision-language model for any robot morphology. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 1-8. IEEE, 2024. 2", + "[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2", + "[12] Xiaoyu Chen, Junliang Guo, Tianyu He, Chuheng Zhang, Pushi Zhang, Derek Cathera Yang, Li Zhao, and Jiang Bian. Igor: Image-goal representations are the atomic control" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "units for foundation models in embodied ai. arXiv preprint arXiv:2411.00785, 2024. 2", + "[13] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2", + "[14] Norman Di Palo and Edward Johns. Keypoint action tokens enable in-context imitation learning in robotics. arXiv preprint arXiv:2403.19578, 2024. 2", + "[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 2", + "[16] Kazuki Egashira, Mark Vero, Robin Staab, Jingxuan He, and Martin Vechev. Exploiting lIm quantization. Advances in Neural Information Processing Systems, 37:41709-41732, 2025. 8", + "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2", + "[18] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2", + "[19] Philip Gage. A new algorithm for data compression. The C Users Journal, 12(2):23-38, 1994. 2", + "[20] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation. In 7th Annual Conference on Robot Learning. 2", + "[21] Abhishek Gupta, Vikash Kumar, Corey Lynch, Sergey Levine, and Karol Hausman. Relay policy learning: Solving long-horizon tasks via imitation and reinforcement learning. arXiv preprint arXiv:1910.11956, 2019. 2", + "[22] Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. 2", + "[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1", + "[24] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5", + "[25] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.3", + "[26] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11097", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2, 4", + "[27] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025. 2", + "[28] Seungjae Lee, Yibin Wang, Haritheja Etukuru, H Jin Kim, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Behavior generation with latent actions. arXiv preprint arXiv:2403.03181, 2024. 1, 2, 3", + "[29] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 1, 2, 3, 4", + "[30] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2", + "[31] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1:3, 2020. 1", + "[32] Atharva Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, and Animesh Garg. Quest: Self-supervised skill abstractions for learning continuous control. Advances in Neural Information Processing Systems, 37:4062-4089, 2025. 2, 12", + "[33] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 1, 3, 5", + "[34] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 1, 2, 3", + "[35] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025. 1, 2, 12", + "[36] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[37] Fengyuan Shi, Zhuoyan Luo, Yixiao Ge, Yujiu Yang, Ying Shan, and Limin Wang. Taming scalable visual tokenizer for autoregressive image generation. arXiv preprint arXiv:2412.02692, 2024. 2", + "[38] Sharath Turuvekere Sreenivas, Saurav Muralidharan, Raviraj Joshi, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Gerald Shen, Jiaqi Zeng, Zijia Chen, Yoshi Suhara, Shizhe Diao, et al. Llm pruning and distillation in practice: The minitron approach. arXiv preprint arXiv:2408.11796, 2024. 8", + "[39] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024.1, 2", + "[40] Andrew Szot, Bogdan Mazoure, Harsh Agrawal, R Devon Hjelm, Zsolt Kira, and Alexander Toshev. Grounding multimodal large language models in actions. Advances in Neural Information Processing Systems, 37:20198-20224, 2024. 2", + "[41] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2", + "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2", + "[43] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 2", + "[44] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 2", + "[45] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2", + "[46] Jange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. arXiv preprint arXiv:2306.05716, 2023. 2", + "[47] Jange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 2", + "[48] Jiangge Yang, Haoyi Zhu, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Tra-moe: Learning trajectory prediction model from multiple domains for adaptive policy conditioning. arXiv preprint arXiv:2411.14519, 2024. 2", + "[49] Jiange Yang, Yansong Shi, Haoyi Zhu, Mingyu Liu, Kaijing Ma, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Como: Learning continuous latent motion from internet videos for scalable robot learning. arXiv preprint arXiv:2505.17006, 2025. 2", + "[50] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 2", + "[51] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2", + "[52] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1, 3", + "[53] Canyu Zhao, Mingyu Liu, Wen Wang, Weihua Chen, Fan Wang, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11098", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Canyu Zhao, Mingyu Liu, Huanyi Zheng, Muzhi Zhu, Zhiyue Zhao, Hao Chen, Tong He, and Chunhua Shen. Dioxide: A generalist diffusion model for visual perceptual tasks. arXiv preprint arXiv:2502.17157, 2025. 2", + "[55] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2", + "[56] Wei Zhao, Pengxiang Ding, Min Zhang, Zhefei Gong, Shuanghao Bai, Han Zhao, and Donglin Wang. Vlas: Vision-language-action model with speech instructions for customized robot manipulation. arXiv preprint arXiv:2502.13508, 2025. 2", + "[57] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2", + "[58] Jinliang Zheng, Jianxiong Li, Dongxiu Liu, Yinan Zheng, Zhihao Wang, Zhonghong Ou, Yu Liu, Jingjing Liu, Ya-Qin Zhang, and Xianyuan Zhan. Universal actions for enhanced embodied foundation models, 2025. 2, 13", + "[59] Wenxuan Zhou, Sujay Bajracharya, and David Held. Plas: Latent action space for offline reinforcement learning, 2020. 2", + "[60] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2", + "[61] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 542 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11099", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_model.json b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a29c0674eb0fc95cfbf7490a32fca57c90681526 --- /dev/null +++ b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.169, + 0.13, + 0.833, + 0.177 + ], + "angle": 0, + "content": "VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.209, + 0.884, + 0.287 + ], + "angle": 0, + "content": "Yating Wang\\(^{12}\\) Haoyi Zhu\\(^{23}\\) Mingyu Liu\\(^{24}\\) Jange Yang\\(^{25}\\) Hao-Shu Fang\\(^{6}\\) Tong He\\(^{27\\dagger}\\) \n\\(^{1}\\)Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University \n\\(^{2}\\)Shanghai AI Lab USTC ZJU NJU SJTU SII Corresponding Author" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.288, + 0.77, + 0.305 + ], + "angle": 0, + "content": "https://xiaoxiao0406.github.io/vqvla.github.io" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.34, + 0.328, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.374, + 0.486, + 0.752 + ], + "angle": 0, + "content": "In this paper, we introduce an innovative vector quantization based action tokenizer built upon the largest-scale action trajectory dataset to date, leveraging over 100 times more data than previous approaches. This extensive dataset enables our tokenizer to capture rich spatiotemporal dynamics, resulting in a model that not only accelerates inference but also generates smoother and more coherent action outputs. Once trained, the tokenizer can be seamlessly adapted to a wide range of downstream tasks in a zero-shot manner, from short-horizon reactive behaviors to long-horizon planning. A key finding of our work is that the domain gap between synthetic and real action trajectories is marginal, allowing us to effectively utilize a vast amount of synthetic data during training without compromising real-world performance. To validate our approach, we conducted extensive experiments in both simulated environments and on real robotic platforms. The results demonstrate that as the volume of synthetic trajectory data increases, the performance of our tokenizer on downstream tasks improves significantly—most notably, achieving up to a \\(30\\%\\) higher success rate on two real-world tasks in long-horizon scenarios. These findings highlight the potential of our action tokenizer as a robust and scalable solution for real-time embodied intelligence systems, paving the way for more efficient and reliable robotic control in diverse application domains." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.224, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Tokenization plays a critical role in recent generative models, including large language models (LLMs) [1, 31], image and video generation models [17, 23, 39], and vision-language-action (VLA) models [8, 26]. One of the key benefits of tokenization is that it compresses the input space. By reducing high-dimensional continuous data into a compact sequence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.908, + 0.401 + ], + "angle": 0, + "content": "of tokens, the complexity of the learning task is substantially reduced. Compared to image patches and language tokens, action sequences are inherently easier to compress because of their spatio-temporal continuity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.909, + 0.749 + ], + "angle": 0, + "content": "Recently, several studies have started to explore action quantized tokenization for Vision-Language-Action (VLA) models [5, 35], demonstrating promising potential. Effective action tokenization not only significantly enhances the downstream performance of VLA models, especially for tasks involving long-horizon planning, but also markedly improves their training and inference efficiency. In this paper, we delve deeper into the potential of action tokenization, with a specific emphasis on its scalability and accuracy. We first find that the more precise the tokenization, the more pronounced the improvements in long-horizon action modeling. This motivates us to train the VQ tokenizer with well-scaled action trajectories to cover various tasks. Additionally, we observe that action trajectories, unlike visual and physical modalities, exhibit minimal domain gaps between real-world and simulated environments. This characteristic enables effective scaling of action tokenizers through the extensive use of synthetic action data. Moreover, training action tokenizers is computationally lightweight compared to scaling entire VLA models. Consequently, focusing on scaling action tokenization emerges as a highly cost-effective strategy, requiring fewer computational and data resources while delivering significant performance enhancements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Specifically, we propose a convolutional residual VQ-VAE [5, 28, 52] framework for training action tokenizers. To effectively train the model, we propose a progressive training strategy: Initially, we train the tokenizer on real-world robotic datasets, such as OpenX-Embodiment [34], which typically contain noisy and jittery trajectories. Subsequently, we gradually integrate cleaner and smoother synthetic data from large-scale simulated robotic datasets, such as LIBERO [29] and ManiSkill [33]. This progressive approach allows the VQ model to converge toward smoother" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11089" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.166 + ], + "angle": 0, + "content": "and more stable representations. Compared to previous approaches that typically rely on training with single-task datasets, our method expands the tokenizer training dataset by more than 100 times, effectively covering a broad spectrum of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.486, + 0.394 + ], + "angle": 0, + "content": "We conduct extensive experiments in both simulated and real-world environments. First, we evaluate our VQ-VAE action tokenizers in the LIBERO simulator, where the results demonstrate the effectiveness of the convolutional residual VQ-VAE and provide preliminary validation for the hypothesis that the VQ-VAE action tokenizer can leverage synthetic data for scaling. Additionally, real-world experiments with a Franka Research 3 robot further validate the superiority of our approach. Specifically, our findings are as follows: (1) as the amount of simulated action data increases, the VQ-VAE tokenizers exhibit linear scaling properties in improving VLA success rates; (2) the tokenizers significantly enhance inference speed and smoothness of VLA models; and (3) the tokenizers effectively reduce cumulative errors, enabling better performance in long-horizon tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.395, + 0.411, + 0.408 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.409, + 0.483, + 0.437 + ], + "angle": 0, + "content": "- We propose a general convolutional residual VQ-VAE-based framework for action tokenizers." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.44, + 0.482, + 0.469 + ], + "angle": 0, + "content": "- We demonstrate that action tokenizers can be effectively scaled by leveraging large-scale simulated action data." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.47, + 0.483, + 0.513 + ], + "angle": 0, + "content": "- We prove that our action tokenizers improve the performance, inference speed, and long-horizon capabilities of VLA models." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.409, + 0.483, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.241, + 0.543 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.486, + 0.825 + ], + "angle": 0, + "content": "Vision-Language-Action Models. Vision-Language-Action (VLA) models[7, 9, 11, 18, 26, 41, 45, 46, 56, 57] bridge visual-language understanding with robot control by mapping multimodal inputs to action outputs. Based on vision-language models (VLM) [3, 10, 11, 15], VLAs represent robot actions (e.g., 6DoF motion, gripper control) as discrete tokens compatible with text-based output. For example, RT-1[7] and RT-2[8] showed that dividing continuous actions into discrete bins allows integration with VLMs, enabling zero-shot generalization using web-scale pretraining. Recent works[26, 34] further leverage this approach, demonstrating improved task generalization through large VLM backbones. Building on this foundation, recent works have explored various approaches to improve VLA, such as leveraging 3D visual inputs [57], integrating chain-of-thought reasoning [51], and employing parallel decoding strategies [27]. In our work, we aim to simultaneously improve the performance and execution speed of VLA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Tokenization. The tokenizer is one of the fundamental components in language and vision generation tasks, especially for autoregressive transformer-based generation models. For language generation tasks, mainstream methods adopt byte pair encoding (BPE) [19, 36] to compress input text. For" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.273 + ], + "angle": 0, + "content": "vision generation tasks, recent works [17, 30, 37, 39, 53, 54] mainly use vector quantization [42] for tokenization to map continuous visual signals into a discrete token sequence. Similarly, in robot learning, many works [7, 8, 28, 32, 35] model action prediction as a generative problem and tokenize the robot action modality to capture the multi-modal distribution within skills and adapt it to autoregressive generative policies. In our work, we propose a general convolutional residual VQ-VAE to tokenize robot action. Different from several related works [28, 32] that use VQ-VAE to tokenize robot actions, our VQ-VAE is trained on large-scale data and is applicable to all tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.909, + 0.653 + ], + "angle": 0, + "content": "Action Representation. Action representation is one of the core components in robot policy learning. Classic action representation schemes include high-level sub-tasks [2, 15], action primitives [22, 43], keypoints [14, 20], as well as low-level continuous end-effector pose [29, 34] or joint state [21]. Recently, imitation learning-based end-to-end approaches [7, 13, 55, 60, 61] often utilize the latter. Some works[6, 40, 45, 58] employ regression-based output layers or integrate new weights for diffusion decoding for action output. Additionally, some works extract corresponding predictive signals from action-less video data to represent actions, such as pixels [9, 47], trajectories [44, 48], and latent motion [12, 49, 50]. To represent complex multi-modal distributions and adapt to discrete generative models, many works discretize continuous low-level actions. Common approaches include per-dimension and per-timestep binning discretization [7, 26], which struggles to perform well in high-frequency tasks [35]. Another method is VQ-VAE based discretization [4, 5, 28, 32, 59]; a limitation here is the need to train a separate action tokenizer for each specific task. Fast [35] utilized a discrete cosine transform (DCT) to learn a non-learnable action tokenizer from real-robot data. In this work, we propose a novel approach: expanding large-scale synthetic data to train the VQ-VAE action tokenizer, aiming to enhance its generalizability and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.67, + 0.612, + 0.685 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.768, + 0.711 + ], + "angle": 0, + "content": "3.1. Preliminaries: VLA models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We use OpenVLA[26] as our backbone model. OpenVLA's origin formulation is to adopt a discrete tokenization strategy for robot action prediction through fine-tuning the Prismatic-7B VLM backbone. The method frames action prediction as a vision-language task, mapping input observation images and natural language instructions to discrete robot action sequences. Specifically, continuous robot actions are discretized into 256 bins per dimension, with bin boundaries determined by the 1st and 99th percentiles of training data distributions rather than min-max ranges to mitigate outlier effects. This discretization converts N-dimensional actions into N discrete integers (0-255). To embed these into the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11090" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.34, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.088, + 0.905, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.408, + 0.908, + 0.466 + ], + "angle": 0, + "content": "Figure 1. The VQ-VLA pipeline, consisting of two main stages: (1) training a general convolutional residual VQ-VAE and (2) fine-tuning OpenVLA using the LoRA approach. Specifically, a general convolutional residual VQ-VAE is first trained on the Open X-Embodiment dataset, LIEBRO, and ManiSkill datasets. The trained VQ-VAE is then frozen and serves as an action tokenizer for OpenVLA, replacing the simple binning method. In the second stage, OpenVLA is fine-tuned using the LoRA technique to optimize its performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.483, + 0.553 + ], + "angle": 0, + "content": "LLM's vocabulary, OpenVLA overwrites the 256 least-used tokens in the Llama tokenizer (last 256 tokens) rather than using special tokens, as the original tokenizer only reserves 100 special tokens." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.56, + 0.431, + 0.577 + ], + "angle": 0, + "content": "3.2. Action Tokenizer via Residual VQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.701 + ], + "angle": 0, + "content": "Based on Residual VQ-VAE [5, 28, 52], we design our encoder and decoder inspired by VAE in pyramidal flow matching [25]. Instead of the simple Multi-Layer Perceptron (MLP) used in Residual VQ-VAE [28, 52], we integrate 2D temporal convolutional layers, motivated by their ability to efficiently capture local relationships and hierarchical temporal dependencies, addressing the scaling limitations of MLPs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Given an input action sequence \\(\\mathbf{a}_{t:t + n}\\in \\mathbb{R}^{n\\times d}\\), where \\(n\\) is the sequence length and \\(d\\) the action dimensionality, the encoder \\(\\phi_{\\mathrm{enc}}\\), composed of 2D temporal convolutional layers, transforms the sequence into a latent embedding \\(\\mathbf{x}\\in \\mathbb{R}^k\\), expressed as \\(\\mathbf{x} = \\phi_{\\mathrm{enc}}(\\mathbf{a}_{t:t + n})\\). To compress \\(\\mathbf{x}\\), we apply Residual Vector Quantization (RVQ) [52], decomposing \\(\\mathbf{x}\\) into quantized residuals: \\(\\mathbf{q}(\\mathbf{x}) = \\sum_{i = 1}^{N_q}\\mathbf{q}_i(\\mathbf{r}_i)\\), where \\(\\mathbf{r}_1 = \\mathbf{x}\\), \\(\\mathbf{r}_{i + 1} = \\mathbf{r}_i - \\mathbf{q}_i(\\mathbf{r}_i)\\), and \\(N_{q}\\) denotes the number of quantization stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The quantized embedding \\(\\mathbf{q}(\\mathbf{x})\\) is passed through the decoder \\(\\phi_{\\mathrm{dec}}\\), which uses 2D temporal deconvolutional layers to reconstruct the sequence \\(\\hat{\\mathbf{a}}_{t:t+n}\\), ensuring temporal structure preservation: \\(\\hat{\\mathbf{a}}_{t:t+n} = \\phi_{\\mathrm{dec}}(\\mathbf{q}(\\mathbf{x}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.908, + 0.552 + ], + "angle": 0, + "content": "To train the framework, we minimize the total loss \\(\\mathcal{L}\\), a weighted combination of reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\), vector quantization (VQ) loss \\(\\mathcal{L}_{\\mathrm{codebook}}\\), and commitment loss \\(\\mathcal{L}_{\\mathrm{commit}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.561, + 0.906, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\left\\| \\mathbf {a} _ {t: t + n} - \\hat {\\mathbf {a}} _ {t: t + n} \\right\\| _ {2} ^ {2} \\tag {1} \\\\ + \\lambda \\left(\\| \\operatorname {s g} (\\mathbf {x}) - \\mathbf {q} (\\mathbf {x}) \\| _ {2} ^ {2} + \\| \\mathbf {x} - \\operatorname {s g} (\\mathbf {q} (\\mathbf {x})) \\| _ {2} ^ {2}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.908, + 0.672 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}(\\cdot)\\) denotes the stop-gradient operation, and \\(\\lambda\\) balances the loss components. We set \\(\\lambda = 4\\) in our experiments. This design ensures efficient, scalable, and robust encoding of structured temporal data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.763, + 0.696 + ], + "angle": 0, + "content": "3.3. Training Residual VQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.808 + ], + "angle": 0, + "content": "We train three versions of the Residual VQ-VAE to evaluate the scaling capability of our action tokenizer: (1) using only the Open X-Embodiment dataset[34], (2) combining Open X-Embodiment and Libero datasets[29], and (3) combining Open X-Embodiment, Libero, and ManiSkill[33] datasets. These experiments were designed to test our hypothesis that the action tokenizer can effectively scale with simulated ata." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.907, + 0.854 + ], + "angle": 0, + "content": "To improve the encoder's ability to process temporal and spatial information, we introduced two types of embeddings before the action sequences are passed into the encoder:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "- Time Embedding: A sinusoidal time embedding was added to encode temporal information at varying frequencies. This embedding allows the model to capture both" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11091" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.486, + 0.137 + ], + "angle": 0, + "content": "low-frequency and high-frequency temporal patterns in the input actions, improving its ability to represent fine-grained temporal details." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.137, + 0.483, + 0.243 + ], + "angle": 0, + "content": "- **Action-Type Embedding:** Learnable embeddings were added for the different components of the action sequence (e.g., XYZ positions, Euler angles, and gripper states). Since the 7 dimensions of the action vector have distinct meanings, this embedding provides a strong prior for the model, helping it distinguish and process the unique roles of each dimension more effectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.246, + 0.484, + 0.306 + ], + "angle": 0, + "content": "The use of these embeddings enhances the encoder's ability to process structured data, improving the quality of the latent representations and the overall performance of the tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.429 + ], + "angle": 0, + "content": "To train a more universal robot action tokenizer and reduce computational overhead, the model is trained using only action sequences as input, without additional conditional inputs. This design reduces complexity while maintaining the generalizability of the tokenizer. All models are trained on a single A100 GPU. For example, training on the Open X-Embodiment dataset requires just one A100 GPU and is completed in one week." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.484, + 0.47 + ], + "angle": 0, + "content": "3.4. Integrating Residual VQ-VAE as Action Tokenizer in VLA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.661 + ], + "angle": 0, + "content": "In this work, we replace the simple bin-based action tokenizer used in OpenVLA with a Residual VQ-VAE-based action tokenizer to improve the expressiveness and precision of action tokenization. Instead of discretizing action sequences into uniform bins, the action sequence \\(\\mathbf{a}_{t:t + n}\\) is first processed through a pre-trained and frozen Residual VQ-VAE encoder \\(\\phi (\\cdot)\\), generating latent representations. These representations are then quantized into discrete codebook indices (tokens) \\(\\{z_q^i\\}_{i = 1}^{N_q}\\) corresponding to \\(N_{q}\\) Residual VQ layers. The quantized tokens \\(z_{q}^{i}\\) are used as the action tokens for training and prediction in the Vision-Language Model (VLM)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Unlike OpenVLA, where all token IDs are mapped to the range [0, 255], the token IDs generated by different VQ layers in the Residual VQ-VAE are assigned unique, non-overlapping ranges. Specifically, tokens from the \\(i\\)-th VQ layer are offset by \\((i - 1) \\times 256\\), ensuring that:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.749, + 0.472, + 0.77 + ], + "angle": 0, + "content": "\\[\nz _ {q} ^ {i} \\in [ 2 5 6 \\times (i - 1), 2 5 6 \\times i - 1 ], \\quad \\forall i \\in \\{1, \\ldots , N _ {q} \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.855 + ], + "angle": 0, + "content": "For example, tokens from the first layer are in the range [0, 255], those from the second layer are in [256, 511], and so on. This design avoids conflicts between token IDs from different layers, as tokens with the same ID in different layers represent semantically different features of the action space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The Vision-Language Model (VLM) is trained to predict these tokens directly. The loss function is the standard next-token prediction loss, computed as the cross-entropy" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.126 + ], + "angle": 0, + "content": "between the predicted token distribution \\(\\hat{z}_q^i\\) and the ground truth token \\(z_q^i\\) produced by the frozen Residual VQ-VAE:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.133, + 0.907, + 0.175 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {V L M}} = - \\sum_ {i = 1} ^ {N _ {q}} \\log P \\left(\\hat {z} _ {q} ^ {i} = z _ {q} ^ {i} \\mid o _ {t - h: t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.181, + 0.907, + 0.212 + ], + "angle": 0, + "content": "where \\( o_{t - h:t} \\) represents the input observation sequence, and \\( N_{q} \\) is the number of Residual VQ layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.908, + 0.302 + ], + "angle": 0, + "content": "To further enhance computational efficiency, similar to OpenVLA, the least-used tokens in the vocabulary are replaced during fine-tuning. However, this replacement process respects the layer-specific token ID ranges, ensuring that the tokenization remains consistent and interpretable across all layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.909, + 0.514 + ], + "angle": 0, + "content": "This integration of a pre-trained and frozen Residual VQ-VAE as the action tokenizer in VLA enables a more expressive and scalable representation of actions. By leveraging hierarchical quantization with non-overlapping token ID ranges, the model achieves better action representation, avoids semantic confusion between layers, and ensures stable loss convergence during training. Furthermore, by using the VQ-VAE tokens to represent longer action sequences instead of predicting one action at a time, the model significantly reduces the number of tokens required for training and inference. This mechanism allows the model to predict more complex behavior sequences with fewer steps, leading to a substantial improvement in inference speed and computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.525, + 0.646, + 0.543 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.672 + ], + "angle": 0, + "content": "In our experiments, we first validate the effectiveness and scalability of the action tokenizer in VLA models using LIBERO simulator[29]. Subsequently, real-world experiments are conducted to further verify our hypothesis. We also investigate the impact of action tokenizers on the performance, inference speed, and long-horizon capabilities of VLA models, alongside ablation studies to evaluate key design choices." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.738, + 0.696 + ], + "angle": 0, + "content": "4.1. Simulation Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.701, + 0.688, + 0.717 + ], + "angle": 0, + "content": "4.1.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.908, + 0.87 + ], + "angle": 0, + "content": "We utilize the LIBERO benchmark[29] to validate and evaluate the effectiveness and scalability of the action tokenizer, using the Franka Panda robot. Specifically, the entire LIBERO task suite—including LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-10, and LIBERO-90—is used as the entire LIBERO dataset. Among these, LIBERO-90 comprises 90 short-horizon tasks, while the other task suites each contain 10 tasks, with 50 demonstrations per task. For evaluation, testing is conducted on LIBERO-goal, LIBERO-10 and LIBERO-90." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Following the approach in OpenVLA[26], we filter out all \"no-op\" actions from the dataset. To preliminarily validate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11092" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "the effectiveness of using VQ-VAE as an action tokenizer in VLA models and its scalability, we train two versions of a convolutional residual VQ-VAE. The first version \\(\\mathrm{VQ_M}\\) is trained solely on the Maniskill dataset[33], and the second \\(\\mathrm{VQ_{M + R}}\\) is trained on a mixture of Maniskill dataset and RLBench dataset[24]. Both models are trained on a single A100 GPU with a batch size of 1024, which takes about only 1 week." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.482, + 0.351 + ], + "angle": 0, + "content": "We use two pre-trained VQ-VAE models as frozen action tokenizers for OpenVLA: \\(\\mathrm{VQ_M}\\) and \\(\\mathrm{VQ_{M + R}}\\). LoRA is then applied to fine-tune OpenVLA on the LIBERO-90 dataset. Additionally, we fine-tune the original OpenVLA model on the LIBERO-90 dataset using LoRA as a baseline for comparison. For a fair comparison, all fine-tuning on the LIBERO-90 task suite is conducted for 400K gradient steps with a batch size of 4, using 4 A100-80GB GPUs and an action chunk length of \\(K = 5\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.364, + 0.416, + 0.379 + ], + "angle": 0, + "content": "4.1.2. Effectiveness of Conv Residual VQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.581 + ], + "angle": 0, + "content": "In the early stages of the experiment, we selected two task suites, LIBERO-10 and LIBERO-GOAL, to conduct preliminary scaling validation experiments. Specifically, we used two variants of Residual VQ-VAE models: one with a simple MLP as the encoder and decoder, and the other with a larger 2D temporal convolutional network as the encoder and decoder. The results are summarized in Tab. 1. The findings indicate that using temporal convolutional networks as the encoder and decoder in the Residual VQ-VAE significantly outperforms the MLP-based architecture in terms of success rate. These results suggest that temporal convolutional networks can serve as an effective action tokenizer, capable of capturing temporal dependencies more effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Furthermore, when we increased the training data for the VQ-VAE model, transitioning from data solely derived from individual LIBERO tasks to the entire LIBERO dataset, the success rate consistently improved. This observation provides preliminary evidence supporting the scalability of our action tokenizer design." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.69, + 0.482, + 0.796 + ], + "angle": 0, + "content": "
Training Dataset of VQLIBERO-10 (%)LIBERO-GOAL (%)
Original OpenVLA-51.075.8
MLP Residual VQ-VAEALL-LIBERO53.472.6
LIBERO-1053.2-
LIBERO-GOAL-65.2
Conv Residual VQ-VAEALL-LIBERO60.075.2
LIBERO-1054.0-
LIBERO-GOAL-72.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Table 1. The evaluation results of residual VQ-VAE architectures. The results demonstrate that the Conv Residual VQ-VAE outperforms the MLP-based version, particularly when trained on the full LIBERO dataset (ALL-LIBERO), highlighting its ability to better capture temporal dependencies and improve success rates." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "4.1.3. Scaling Data Improves VQ-VAE Action Tokenizer Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.126, + 0.907, + 0.307 + ], + "angle": 0, + "content": "In the experiment, to avoid potential performance inflation from in-domain data during evaluation on LIBERO-90, we trained the VQ-VAE Action Tokenizer on completely out-of-domain ManiSkill and RLBench simulation data, rather than LIBERO data. Specifically, ManiSkill contains 30,000 action sequences, and RLBench includes 10,000 action sequences. We compare three models: the baseline, which fine-tunes the original OpenVLA; \\(\\mathrm{VQ_M}\\), which uses a VQ-VAE trained on the Maniskill dataset as the action tokenizer for OpenVLA; and \\(\\mathrm{VQ_{M + R}}\\), which uses a VQ-VAE trained on a mixture of the Maniskill and RLBench datasets as the action tokenizer for OpenVLA." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.308, + 0.907, + 0.398 + ], + "angle": 0, + "content": "On LIBERO-90, \\(\\mathrm{VQ}_{\\mathrm{M} + \\mathrm{R}}\\) achieved \\(80.98\\%\\), a \\(7.45\\%\\) improvement over the OpenVLA baseline (73.53%), Tab. 2. Furthermore, an ablation study using only ManiSkill data for training \\(\\mathrm{VQ_M}\\) resulted in substantially lower performance, underscoring the critical role of sufficient synthetic data scale." + }, + { + "type": "table", + "bbox": [ + 0.582, + 0.411, + 0.838, + 0.453 + ], + "angle": 0, + "content": "
baseline(%)VQM(%)VQM+R(%)
LIEBRO-9073.5314.3880.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.464, + 0.907, + 0.506 + ], + "angle": 0, + "content": "Table 2. Effectiveness of VQ-VAE Action Tokenizers in Scaling Simulation Data. The results demonstrate \\( \\mathrm{VQ}_{\\mathrm{M} + \\mathrm{R}} \\) reached 80.98%, outperforming the OpenVLA baseline by 7.45%" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.538, + 0.73, + 0.554 + ], + "angle": 0, + "content": "4.2. Real-Word Experiment" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.561, + 0.687, + 0.577 + ], + "angle": 0, + "content": "4.2.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Our robotic platform consists of a single Franka Research3 arm equipped with a third-person-view RealSense D435 camera mounted in a fixed position to capture environmental observations. The system operates at \\(20\\mathrm{Hz}\\) (moderately reduced from the native \\(100\\mathrm{Hz}\\) control frequency to balance training efficiency and motion continuity), with actions defined as absolute end-effector poses in SE(3) space (3D position + quaternion orientation)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.778 + ], + "angle": 0, + "content": "Our experimental benchmark comprises six manipulation tasks (4 short-horizon tasks, 2 long-horizon tasks) designed to evaluate the model's ability to handle varying task complexities. For each task, we collect 50 demonstrations and evaluate performance over 20 trials:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.905, + 0.809 + ], + "angle": 0, + "content": "1) Pull out a tissue paper: The robot need to grasp and pull out a single tissue paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.869 + ], + "angle": 0, + "content": "2) Pick up the [TOY NAME]: The robot is required to pick up the specific toy, including the toy snake, the toy eggplant, and the toy chicken, resulting in a total of three tasks (with no other distractions)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3) Put the toy into the basket: The robot need to pick up the toy(no other distractions) and put it into the basket." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.779, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11093" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.222, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.221, + 0.173, + 0.229 + ], + "angle": 0, + "content": "Libero 90" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.091, + 0.351, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.221, + 0.333, + 0.23 + ], + "angle": 0, + "content": "Pull out a tissue paper" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.091, + 0.48, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.22, + 0.476, + 0.229 + ], + "angle": 0, + "content": "Put the toy into the basket" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.231, + 0.222, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.362, + 0.222, + 0.372 + ], + "angle": 0, + "content": "Put all cups into the basket" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.231, + 0.351, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.362, + 0.349, + 0.372 + ], + "angle": 0, + "content": "Put the toy into the drawer" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.231, + 0.481, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.362, + 0.46, + 0.372 + ], + "angle": 0, + "content": "Filp the pot upright" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.374, + 0.222, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.374, + 0.351, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.374, + 0.481, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.509, + 0.333, + 0.519 + ], + "angle": 0, + "content": "Pick up the{toy name}" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.601 + ], + "angle": 0, + "content": "Figure 2. All Evaluation environments: We conduct comprehensive evaluations of VQ-VLA in both simulation and real-world settings. In simulation, evaluations are performed on the LIBERO-90 benchmark within the LIBERO dataset. And six diverse tasks are designed for real-world testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.692 + ], + "angle": 0, + "content": "4) Flip the pot upright: We set a flipped pot on the platform, the robot need to flip and upright a fallen cooking pot." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.484, + 0.761 + ], + "angle": 0, + "content": "5) Put all cups into the basket: We place two different cups on the table and set a few other things (toys) as distractions. The robot need to sequentially put two cups into the basket, testing long-horizon task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.483, + 0.846 + ], + "angle": 0, + "content": "6) Put the toy into the drawer: We place a toy on the table and set a few other things as distractions. The robot need to sequentially open the drawer, pick up the specific toy and put it into the drawer, and close the drawer, testing long-horizon task." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We finetune each task separately for 100K gradient steps (batch size 4 across 4 A100-80GB GPUs) with action chunk length \\( \\mathrm{K} = 5 \\). During deployment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.82, + 0.106 + ], + "angle": 0, + "content": "4.2.2. Performance on Short-Horizon Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.112, + 0.907, + 0.247 + ], + "angle": 0, + "content": "Based on the in the simulation experiments, we further incorporated 120k synthetic trajectories from the ManiSkill dataset, co-training \\(\\mathrm{VQ_{O + L + M}}\\) with Open X-Embodiment dataset and LIBERO dataset to evaluate the scaling capability of our action tokenizer. We evaluated the performance of the action tokenizer trained with different datasets on four carefully designed short-horizon tasks, focusing on multi-task learning, precision manipulation, and dynamic adjustment capabilities. The results are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.249, + 0.909, + 0.415 + ], + "angle": 0, + "content": "It can be observed that when synthetic trajectories are added for co-training, the performance of the action tokenizer improves significantly. The average success rate increased from the baseline of \\(23\\%\\) to \\(46.25\\%\\), with a notable improvement of \\(30\\%\\) in the \"Flip the pot upright\" task. In the \"Pull out a tissue paper\" task, which tests the robot's performance in high-precision dynamic operations (as this task requires continuous, fine-grained grasping and pulling motions), the baseline model achieved only a \\(5\\%\\) success rate, failing almost entirely. In contrast, models incorporating VQ achieved success rates of \\(20\\%\\) or higher." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Furthermore, comparing \\(\\mathrm{VQ_{O + L}}\\) with \\(\\mathrm{VQ_O}\\), the average success rate on short-horizon tasks increased by only \\(0.5\\%\\). This may be attributed to the limited size of the LIBERO dataset, which has minimal impact on short-horizon tasks. However, the ManiSkill dataset is 50 times larger than LIBERO, leading to a significant improvement in success rates when used for training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.817, + 0.548 + ], + "angle": 0, + "content": "4.2.3. Performance on Long-Horizon Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.765 + ], + "angle": 0, + "content": "VQ-VLA demonstrates outstanding performance on long-horizon tasks (\"Put all cups in the basket\" and \"Put the toy into the drawer\"), significantly outperforming baseline model in both success rate and efficiency. In scenarios where baseline models achieve success rates as low as \\(15\\%\\) or nearly 0, the \\(\\mathrm{VQ}_{O + L + M}\\) model achieves significantly higher success rates of \\(50\\%\\) and \\(30\\%\\), respectively (see Fig.3). For the \"Put the toy into the drawer\" task, a representative long-horizon scenario, the baseline model was only able to complete the first step of opening the drawer in 15 out of 20 trials, failing to proceed further in most cases. In contrast, the \\(\\mathrm{VQ}_{O + L + M}\\) model successfully opened the drawer in all test cases, demonstrating its robustness and reliability in handling complex sequential tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "A key advantage of VQ-VLA in long-horizon tasks lies in its use of VQ-VAE as an action tokenizer, which enables the model to predict multiple actions in a single inference step. This design reduces the accumulation of errors over extended task sequences, making VQ-VLA particularly advantageous for tasks requiring long-term planning and execution. Additionally, in scenarios involving the consecutive execution of multiple subtasks, VQ-VLA not only achieves higher success rates but also significantly reduces task completion time" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11094" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.912, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.349, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Figure 3. Real-world experimental results: We compare the performance of Baseline, \\(\\mathrm{VQ_0}\\), \\(\\mathrm{VQ_{O + L}}\\), and \\(\\mathrm{VQ_{O + L + M}}\\) on both short-horizon and long-horizon tasks. In terms of the average success rate, all VQ-based models outperform the Baseline. The best-performing model, \\(\\mathrm{VQ_{O + L + M}}\\), achieves a success rate that is \\(23.25\\%\\) higher than the Baseline on both short-horizon and long-horizon tasks. Additionally, the results show that \\(\\mathrm{VQ_{O + L + M}}\\) outperforms \\(\\mathrm{VQ_{O + L}}\\), which in turn outperforms \\(\\mathrm{VQ_0}\\), indicating the effectiveness of incorporating synthetic data during training without compromising real-world performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.486, + 0.492 + ], + "angle": 0, + "content": "compared to baseline models, highlighting its superior execution efficiency and ability to handle long-horizon challenges effectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.512, + 0.367, + 0.527 + ], + "angle": 0, + "content": "4.2.4. Sim&Real Domain Gap Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.671 + ], + "angle": 0, + "content": "To quantify the domain gap between synthetic data and real-world data, we trained a VQ-VAE model exclusively on the LIBERO dataset, referred to as \\(\\mathrm{VQL}\\), for OpenVLA. The model was tested in three real-world tasks (one long-horizon and two short-horizon tasks), and the results are shown in Tab. 3. The results indicate that the performance of \\(\\mathrm{VQL}\\) is comparable to that of both \\(\\mathrm{VQ_{O + L}}\\) and \\(\\mathrm{VQ_O}\\), suggesting that the domain gap between synthetic and real-world data is minimal." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.676, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Although real-world data may contain noise, the inclusion of Open X-Embodiment data as a real-world dataset expands the data sources and enriches the diversity of data types, which effectively enhances the model's generalization and robustness." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.34, + 0.789 + ], + "angle": 0, + "content": "4.2.5. Inference Speed Comparison" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "During the real-world experiments, we measured the action execution frequency of VQ-VLA and compared it with the original OpenVLA. The results are summarized in Tab. 4. As shown in the table, with a compression ratio of 5 in VQ-VAE, the inference speed is nearly tripled. This significant improvement greatly facilitates real-time performance in practical applications." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.442, + 0.907, + 0.538 + ], + "angle": 0, + "content": "
Put the toy into the drawer (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline5.030.020.0
VQo15.045.035.0
VQL10.055.035.0
VQo+L10.045.035.0
VQo+L+M25.060.045.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.548, + 0.908, + 0.604 + ], + "angle": 0, + "content": "Table 3. Performance Comparison Across Real-World Tasks: We observe that the performance of \\(\\mathrm{VQ_L}\\) is comparable to that of both \\(\\mathrm{VQ_{O + L}}\\) and \\(\\mathrm{VQ_O}\\), indicating that the domain gap between synthetic and real-world data is minimal." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.62, + 0.879, + 0.675 + ], + "angle": 0, + "content": "
Frequency (Hz)
VQ-VLA11.84
OpenVLA4.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Table 4. The Results of Frequencies. We report the comparison results of our VQ-VLA and baseline OpenVLA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.68, + 0.756 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.795 + ], + "angle": 0, + "content": "In this section, we report some ablation studies to show the effectiveness of the design choices of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.906, + 0.837 + ], + "angle": 0, + "content": "4.3.1. Action Chunking via VQ-VAE and Autoregressive Output" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.908, + 0.902 + ], + "angle": 0, + "content": "The original OpenVLA model generates a single action per inference step based solely on the current observation, resulting in a step-by-step execution approach. To improve efficiency and handle longer action sequences, we extend" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11095" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "OpenVLA to output a sequence of five actions in an autoregressive manner, similar to the action chunking mechanism in VQ-VLA. Specifically, the model predicts the next action based on the current observation and previously generated actions, allowing it to generate a sequence of actions in a single inference step. This autoregressive approach serves as an alternative action chunking strategy by grouping multiple actions for execution." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.213, + 0.485, + 0.499 + ], + "angle": 0, + "content": "To evaluate the effectiveness of different action chunking strategies, we design ablation experiments comparing the autoregressive output of OpenVLA to the VQ-based action chunking method in VQ-VLA. The results, shown in Tab. 5, indicate that the use of the autoregressive approach for action chunking in the original OpenVLA leads to a significant drop in success rate compared to the baseline. In real-world experiments, when using the autoregressive approach as a form of action chunking, the spatial magnitude of each action chunk is relatively small, resulting in slower overall execution of actions. Compared to the Action Chunking via VQ-VAE method, even with the same action chunk size, the time required to reach the same target location is significantly longer. Additionally, this approach tends to exhibit shortcut learning, where the model directly copies previous actions within the same action chunk. By analyzing the output values of each action chunk, it is observed that multiple actions within a single chunk have remarkably similar values, indicating a lack of diversity in the predicted actions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Overall, Action Chunking via VQ-VAE not only improves inference speed but also enhances the performance of VLA by generating more effective and diverse action sequences. This makes it a more suitable approach for addressing real-world long-horizon tasks." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.585, + 0.482, + 0.648 + ], + "angle": 0, + "content": "
Action ChunkingLIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline74.7630.020.0
Autoregressive Output66.5310.00.0
VQ-based (VQo+L+M)86.6160.045.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.742 + ], + "angle": 0, + "content": "Table 5. Comparison of action chunking methods: We evaluate the performance on three tasks (one simulator task and two real-world tasks). The results show that the Autoregressive Output used in OpenVLA performs poorly as an action chunking method, with a significant gap compared to the VQ-based approach. This indicates that the VQ-based method is more effective." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.762, + 0.395, + 0.776 + ], + "angle": 0, + "content": "4.3.2. Embedding Integration Effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.87 + ], + "angle": 0, + "content": "To evaluate the impact of embeddings, we conducted an ablation study comparing the model's performance with and without embeddings. Specifically, the baseline model processes raw action sequences directly, while the enhanced model incorporates both time embedding and action-type embedding." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 6, the model with embeddings significantly outperforms the baseline in terms of success rate. This" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "demonstrates that the integration of embeddings improves the encoder's ability to represent structured action sequences, leading to better overall performance." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.155, + 0.907, + 0.215 + ], + "angle": 0, + "content": "
LIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
VQoL (w.o. Embeddings)85.1740.035.0
VQoL (w Embeddings)86.1645.035.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.226, + 0.907, + 0.282 + ], + "angle": 0, + "content": "Table 6. Embedding integration improves performance. The table compares models with and without embeddings across three tasks, showing that embeddings enhance success rates, especially for \"Flip the pot upright.\"" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.322, + 0.795, + 0.337 + ], + "angle": 0, + "content": "5. Limitations and Future Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.909, + 0.576 + ], + "angle": 0, + "content": "In our work, a series of extensive experiments demonstrate that our proposed action tokenizers can improve VLA performance and inference speed while achieving scalability on simulated data. Despite these promising results, there still remain some limitations and opportunities for future work. First, our action tokenizers can be further extended to larger-scale simulated datasets. Second, our work improves inference speed by decoding multi-step action sequences, which can be further combined with techniques such as distillation [38] and quantization [16] of VLMs in the future. Finally, the architecture design of our action tokenizers can be further improved, for example, by incorporating the frequency of action data as a extra condition. In summary, we hope that our work can serve as a strong baseline and inspire future research." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.598, + 0.64, + 0.613 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.909, + 0.807 + ], + "angle": 0, + "content": "In this paper, we propose a general convolutional residual VQ-VAE framework for action tokenizers that can seamlessly integrate with state-of-the-art VLA models. Our VQ-VAE is trained on 100 times more data than previous methods and can be directly transferred to downstream both real-world and simulated robotic manipulation tasks. Extensive experiments conducted in both simulated and real-world environments validate that the proposed convolutional residual VQ-VAE framework not only enhances the performance of VLA policies but also accelerates inference. Finally, our VQ-VAE also demonstrates the ability to scale effectively with large-scale simulated data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.828, + 0.673, + 0.844 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11096" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.326 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeffrey Donahue, Pauline Luc, Alexis Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[4] Arthur Allshire, Roberto Martin-Martin, Charles Lin, Shawn Manuel, Silvio Savarese, and Animesh Garg. Laser: Learning a latent action space for efficient reinforcement learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6650-6656, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[5] Suneel Belkhale and Dorsa Sadigh. Minivla: A better vla with a smaller footprint. https://ai.stanford.edu/blog/minivla/, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. \\(\\pi_0\\): A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[9] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Wei Chen, Yifeng Zhang, Hao Li, Xufeng Wang, and Ming Liu. Anyvlm: Unified vision-language model for any robot morphology. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 1-8. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Xiaoyu Chen, Junliang Guo, Tianyu He, Chuheng Zhang, Pushi Zhang, Derek Cathera Yang, Li Zhao, and Jiang Bian. Igor: Image-goal representations are the atomic control" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.906, + 0.119 + ], + "angle": 0, + "content": "units for foundation models in embodied ai. arXiv preprint arXiv:2411.00785, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.908, + 0.189 + ], + "angle": 0, + "content": "[13] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[14] Norman Di Palo and Edward Johns. Keypoint action tokens enable in-context imitation learning in robotics. arXiv preprint arXiv:2403.19578, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.908, + 0.317 + ], + "angle": 0, + "content": "[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[16] Kazuki Egashira, Mark Vero, Robin Staab, Jingxuan He, and Martin Vechev. Exploiting lIm quantization. Advances in Neural Information Processing Systems, 37:41709-41732, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[18] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[19] Philip Gage. A new algorithm for data compression. The C Users Journal, 12(2):23-38, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[20] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation. In 7th Annual Conference on Robot Learning. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[21] Abhishek Gupta, Vikash Kumar, Corey Lynch, Sergey Levine, and Karol Hausman. Relay policy learning: Solving long-horizon tasks via imitation and reinforcement learning. arXiv preprint arXiv:1910.11956, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[22] Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[24] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[25] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11097" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[27] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[28] Seungjae Lee, Yibin Wang, Haritheja Etukuru, H Jin Kim, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Behavior generation with latent actions. arXiv preprint arXiv:2403.03181, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[29] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[30] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[31] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1:3, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[32] Atharva Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, and Animesh Garg. Quest: Self-supervised skill abstractions for learning continuous control. Advances in Neural Information Processing Systems, 37:4062-4089, 2025. 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[33] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[34] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[35] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025. 1, 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[36] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[37] Fengyuan Shi, Zhuoyan Luo, Yixiao Ge, Yujiu Yang, Ying Shan, and Limin Wang. Taming scalable visual tokenizer for autoregressive image generation. arXiv preprint arXiv:2412.02692, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.869 + ], + "angle": 0, + "content": "[38] Sharath Turuvekere Sreenivas, Saurav Muralidharan, Raviraj Joshi, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Gerald Shen, Jiaqi Zeng, Zijia Chen, Yoshi Suhara, Shizhe Diao, et al. Llm pruning and distillation in practice: The minitron approach. arXiv preprint arXiv:2408.11796, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[40] Andrew Szot, Bogdan Mazoure, Harsh Agrawal, R Devon Hjelm, Zsolt Kira, and Alexander Toshev. Grounding multimodal large language models in actions. Advances in Neural Information Processing Systems, 37:20198-20224, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[41] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[43] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.362 + ], + "angle": 0, + "content": "[44] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.906, + 0.433 + ], + "angle": 0, + "content": "[45] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[46] Jange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. arXiv preprint arXiv:2306.05716, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[47] Jange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[48] Jiangge Yang, Haoyi Zhu, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Tra-moe: Learning trajectory prediction model from multiple domains for adaptive policy conditioning. arXiv preprint arXiv:2411.14519, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[49] Jiange Yang, Yansong Shi, Haoyi Zhu, Mingyu Liu, Kaijing Ma, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Como: Learning continuous latent motion from internet videos for scalable robot learning. arXiv preprint arXiv:2505.17006, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[50] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[51] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[52] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[53] Canyu Zhao, Mingyu Liu, Wen Wang, Weihua Chen, Fan Wang, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11098" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[54] Canyu Zhao, Mingyu Liu, Huanyi Zheng, Muzhi Zhu, Zhiyue Zhao, Hao Chen, Tong He, and Chunhua Shen. Dioxide: A generalist diffusion model for visual perceptual tasks. arXiv preprint arXiv:2502.17157, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[55] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[56] Wei Zhao, Pengxiang Ding, Min Zhang, Zhefei Gong, Shuanghao Bai, Han Zhao, and Donglin Wang. Vlas: Vision-language-action model with speech instructions for customized robot manipulation. arXiv preprint arXiv:2502.13508, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[57] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[58] Jinliang Zheng, Jianxiong Li, Dongxiu Liu, Yinan Zheng, Zhihao Wang, Zhonghong Ou, Yu Liu, Jingjing Liu, Ya-Qin Zhang, and Xianyuan Zhan. Universal actions for enhanced embodied foundation models, 2025. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[59] Wenxuan Zhou, Sujay Bajracharya, and David Held. Plas: Latent action space for offline reinforcement learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[60] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[61] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11099" + } + ] +] \ No newline at end of file diff --git a/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_origin.pdf b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8dd22927799a0b6a7591dada0ac5d271e52fa9f8 --- /dev/null +++ b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/a05d8d10-adb9-4812-8715-77d71aa82af2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6e8f9994390e101391cd6e723f323cb50945ce6fdf660cd660cef5b4aa8a7b0 +size 2698029 diff --git a/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/full.md b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..904ff90f16557c33042f217862582948ef95fd5e --- /dev/null +++ b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/full.md @@ -0,0 +1,341 @@ +# VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers + +Yating Wang $^{12}$ Haoyi Zhu $^{23}$ Mingyu Liu $^{24}$ Jange Yang $^{25}$ Hao-Shu Fang $^{6}$ Tong He $^{27\dagger}$ $^{1}$ Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University + $^{2}$ Shanghai AI Lab USTC ZJU NJU SJTU SII Corresponding Author + +https://xiaoxiao0406.github.io/vqvla.github.io + +# Abstract + +In this paper, we introduce an innovative vector quantization based action tokenizer built upon the largest-scale action trajectory dataset to date, leveraging over 100 times more data than previous approaches. This extensive dataset enables our tokenizer to capture rich spatiotemporal dynamics, resulting in a model that not only accelerates inference but also generates smoother and more coherent action outputs. Once trained, the tokenizer can be seamlessly adapted to a wide range of downstream tasks in a zero-shot manner, from short-horizon reactive behaviors to long-horizon planning. A key finding of our work is that the domain gap between synthetic and real action trajectories is marginal, allowing us to effectively utilize a vast amount of synthetic data during training without compromising real-world performance. To validate our approach, we conducted extensive experiments in both simulated environments and on real robotic platforms. The results demonstrate that as the volume of synthetic trajectory data increases, the performance of our tokenizer on downstream tasks improves significantly—most notably, achieving up to a $30\%$ higher success rate on two real-world tasks in long-horizon scenarios. These findings highlight the potential of our action tokenizer as a robust and scalable solution for real-time embodied intelligence systems, paving the way for more efficient and reliable robotic control in diverse application domains. + +# 1. Introduction + +Tokenization plays a critical role in recent generative models, including large language models (LLMs) [1, 31], image and video generation models [17, 23, 39], and vision-language-action (VLA) models [8, 26]. One of the key benefits of tokenization is that it compresses the input space. By reducing high-dimensional continuous data into a compact sequence + +of tokens, the complexity of the learning task is substantially reduced. Compared to image patches and language tokens, action sequences are inherently easier to compress because of their spatio-temporal continuity. + +Recently, several studies have started to explore action quantized tokenization for Vision-Language-Action (VLA) models [5, 35], demonstrating promising potential. Effective action tokenization not only significantly enhances the downstream performance of VLA models, especially for tasks involving long-horizon planning, but also markedly improves their training and inference efficiency. In this paper, we delve deeper into the potential of action tokenization, with a specific emphasis on its scalability and accuracy. We first find that the more precise the tokenization, the more pronounced the improvements in long-horizon action modeling. This motivates us to train the VQ tokenizer with well-scaled action trajectories to cover various tasks. Additionally, we observe that action trajectories, unlike visual and physical modalities, exhibit minimal domain gaps between real-world and simulated environments. This characteristic enables effective scaling of action tokenizers through the extensive use of synthetic action data. Moreover, training action tokenizers is computationally lightweight compared to scaling entire VLA models. Consequently, focusing on scaling action tokenization emerges as a highly cost-effective strategy, requiring fewer computational and data resources while delivering significant performance enhancements. + +Specifically, we propose a convolutional residual VQ-VAE [5, 28, 52] framework for training action tokenizers. To effectively train the model, we propose a progressive training strategy: Initially, we train the tokenizer on real-world robotic datasets, such as OpenX-Embodiment [34], which typically contain noisy and jittery trajectories. Subsequently, we gradually integrate cleaner and smoother synthetic data from large-scale simulated robotic datasets, such as LIBERO [29] and ManiSkill [33]. This progressive approach allows the VQ model to converge toward smoother + +and more stable representations. Compared to previous approaches that typically rely on training with single-task datasets, our method expands the tokenizer training dataset by more than 100 times, effectively covering a broad spectrum of downstream tasks. + +We conduct extensive experiments in both simulated and real-world environments. First, we evaluate our VQ-VAE action tokenizers in the LIBERO simulator, where the results demonstrate the effectiveness of the convolutional residual VQ-VAE and provide preliminary validation for the hypothesis that the VQ-VAE action tokenizer can leverage synthetic data for scaling. Additionally, real-world experiments with a Franka Research 3 robot further validate the superiority of our approach. Specifically, our findings are as follows: (1) as the amount of simulated action data increases, the VQ-VAE tokenizers exhibit linear scaling properties in improving VLA success rates; (2) the tokenizers significantly enhance inference speed and smoothness of VLA models; and (3) the tokenizers effectively reduce cumulative errors, enabling better performance in long-horizon tasks. + +In summary, our contributions are as follows: + +- We propose a general convolutional residual VQ-VAE-based framework for action tokenizers. +- We demonstrate that action tokenizers can be effectively scaled by leveraging large-scale simulated action data. +- We prove that our action tokenizers improve the performance, inference speed, and long-horizon capabilities of VLA models. + +# 2. Related Works + +Vision-Language-Action Models. Vision-Language-Action (VLA) models[7, 9, 11, 18, 26, 41, 45, 46, 56, 57] bridge visual-language understanding with robot control by mapping multimodal inputs to action outputs. Based on vision-language models (VLM) [3, 10, 11, 15], VLAs represent robot actions (e.g., 6DoF motion, gripper control) as discrete tokens compatible with text-based output. For example, RT-1[7] and RT-2[8] showed that dividing continuous actions into discrete bins allows integration with VLMs, enabling zero-shot generalization using web-scale pretraining. Recent works[26, 34] further leverage this approach, demonstrating improved task generalization through large VLM backbones. Building on this foundation, recent works have explored various approaches to improve VLA, such as leveraging 3D visual inputs [57], integrating chain-of-thought reasoning [51], and employing parallel decoding strategies [27]. In our work, we aim to simultaneously improve the performance and execution speed of VLA. + +Tokenization. The tokenizer is one of the fundamental components in language and vision generation tasks, especially for autoregressive transformer-based generation models. For language generation tasks, mainstream methods adopt byte pair encoding (BPE) [19, 36] to compress input text. For + +vision generation tasks, recent works [17, 30, 37, 39, 53, 54] mainly use vector quantization [42] for tokenization to map continuous visual signals into a discrete token sequence. Similarly, in robot learning, many works [7, 8, 28, 32, 35] model action prediction as a generative problem and tokenize the robot action modality to capture the multi-modal distribution within skills and adapt it to autoregressive generative policies. In our work, we propose a general convolutional residual VQ-VAE to tokenize robot action. Different from several related works [28, 32] that use VQ-VAE to tokenize robot actions, our VQ-VAE is trained on large-scale data and is applicable to all tasks. + +Action Representation. Action representation is one of the core components in robot policy learning. Classic action representation schemes include high-level sub-tasks [2, 15], action primitives [22, 43], keypoints [14, 20], as well as low-level continuous end-effector pose [29, 34] or joint state [21]. Recently, imitation learning-based end-to-end approaches [7, 13, 55, 60, 61] often utilize the latter. Some works[6, 40, 45, 58] employ regression-based output layers or integrate new weights for diffusion decoding for action output. Additionally, some works extract corresponding predictive signals from action-less video data to represent actions, such as pixels [9, 47], trajectories [44, 48], and latent motion [12, 49, 50]. To represent complex multi-modal distributions and adapt to discrete generative models, many works discretize continuous low-level actions. Common approaches include per-dimension and per-timestep binning discretization [7, 26], which struggles to perform well in high-frequency tasks [35]. Another method is VQ-VAE based discretization [4, 5, 28, 32, 59]; a limitation here is the need to train a separate action tokenizer for each specific task. Fast [35] utilized a discrete cosine transform (DCT) to learn a non-learnable action tokenizer from real-robot data. In this work, we propose a novel approach: expanding large-scale synthetic data to train the VQ-VAE action tokenizer, aiming to enhance its generalizability and efficiency. + +# 3. Methods + +# 3.1. Preliminaries: VLA models. + +We use OpenVLA[26] as our backbone model. OpenVLA's origin formulation is to adopt a discrete tokenization strategy for robot action prediction through fine-tuning the Prismatic-7B VLM backbone. The method frames action prediction as a vision-language task, mapping input observation images and natural language instructions to discrete robot action sequences. Specifically, continuous robot actions are discretized into 256 bins per dimension, with bin boundaries determined by the 1st and 99th percentiles of training data distributions rather than min-max ranges to mitigate outlier effects. This discretization converts N-dimensional actions into N discrete integers (0-255). To embed these into the + +![](images/f77b0c565c1e8248e90c4431c198b617e3caa555a7fad3f1e489518773b4c709.jpg) +Figure 1. The VQ-VLA pipeline, consisting of two main stages: (1) training a general convolutional residual VQ-VAE and (2) fine-tuning OpenVLA using the LoRA approach. Specifically, a general convolutional residual VQ-VAE is first trained on the Open X-Embodiment dataset, LIEBRO, and ManiSkill datasets. The trained VQ-VAE is then frozen and serves as an action tokenizer for OpenVLA, replacing the simple binning method. In the second stage, OpenVLA is fine-tuned using the LoRA technique to optimize its performance. + +![](images/11f5fe7ecf165e4c1f5fa22894f6e01a9b2087c70ec6e29a3399938973ffd06a.jpg) + +LLM's vocabulary, OpenVLA overwrites the 256 least-used tokens in the Llama tokenizer (last 256 tokens) rather than using special tokens, as the original tokenizer only reserves 100 special tokens. + +# 3.2. Action Tokenizer via Residual VQ-VAE + +Based on Residual VQ-VAE [5, 28, 52], we design our encoder and decoder inspired by VAE in pyramidal flow matching [25]. Instead of the simple Multi-Layer Perceptron (MLP) used in Residual VQ-VAE [28, 52], we integrate 2D temporal convolutional layers, motivated by their ability to efficiently capture local relationships and hierarchical temporal dependencies, addressing the scaling limitations of MLPs. + +Given an input action sequence $\mathbf{a}_{t:t + n}\in \mathbb{R}^{n\times d}$ , where $n$ is the sequence length and $d$ the action dimensionality, the encoder $\phi_{\mathrm{enc}}$ , composed of 2D temporal convolutional layers, transforms the sequence into a latent embedding $\mathbf{x}\in \mathbb{R}^k$ , expressed as $\mathbf{x} = \phi_{\mathrm{enc}}(\mathbf{a}_{t:t + n})$ . To compress $\mathbf{x}$ , we apply Residual Vector Quantization (RVQ) [52], decomposing $\mathbf{x}$ into quantized residuals: $\mathbf{q}(\mathbf{x}) = \sum_{i = 1}^{N_q}\mathbf{q}_i(\mathbf{r}_i)$ , where $\mathbf{r}_1 = \mathbf{x}$ , $\mathbf{r}_{i + 1} = \mathbf{r}_i - \mathbf{q}_i(\mathbf{r}_i)$ , and $N_{q}$ denotes the number of quantization stages. + +The quantized embedding $\mathbf{q}(\mathbf{x})$ is passed through the decoder $\phi_{\mathrm{dec}}$ , which uses 2D temporal deconvolutional layers to reconstruct the sequence $\hat{\mathbf{a}}_{t:t+n}$ , ensuring temporal structure preservation: $\hat{\mathbf{a}}_{t:t+n} = \phi_{\mathrm{dec}}(\mathbf{q}(\mathbf{x}))$ . + +To train the framework, we minimize the total loss $\mathcal{L}$ , a weighted combination of reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ , vector quantization (VQ) loss $\mathcal{L}_{\mathrm{codebook}}$ , and commitment loss $\mathcal{L}_{\mathrm{commit}}$ : + +$$ +\begin{array}{l} \mathcal {L} = \left\| \mathbf {a} _ {t: t + n} - \hat {\mathbf {a}} _ {t: t + n} \right\| _ {2} ^ {2} \tag {1} \\ + \lambda \left(\| \operatorname {s g} (\mathbf {x}) - \mathbf {q} (\mathbf {x}) \| _ {2} ^ {2} + \| \mathbf {x} - \operatorname {s g} (\mathbf {q} (\mathbf {x})) \| _ {2} ^ {2}\right), \\ \end{array} +$$ + +where $\mathrm{sg}(\cdot)$ denotes the stop-gradient operation, and $\lambda$ balances the loss components. We set $\lambda = 4$ in our experiments. This design ensures efficient, scalable, and robust encoding of structured temporal data. + +# 3.3. Training Residual VQ-VAE + +We train three versions of the Residual VQ-VAE to evaluate the scaling capability of our action tokenizer: (1) using only the Open X-Embodiment dataset[34], (2) combining Open X-Embodiment and Libero datasets[29], and (3) combining Open X-Embodiment, Libero, and ManiSkill[33] datasets. These experiments were designed to test our hypothesis that the action tokenizer can effectively scale with simulated ata. + +To improve the encoder's ability to process temporal and spatial information, we introduced two types of embeddings before the action sequences are passed into the encoder: + +- Time Embedding: A sinusoidal time embedding was added to encode temporal information at varying frequencies. This embedding allows the model to capture both + +low-frequency and high-frequency temporal patterns in the input actions, improving its ability to represent fine-grained temporal details. + +- **Action-Type Embedding:** Learnable embeddings were added for the different components of the action sequence (e.g., XYZ positions, Euler angles, and gripper states). Since the 7 dimensions of the action vector have distinct meanings, this embedding provides a strong prior for the model, helping it distinguish and process the unique roles of each dimension more effectively. + +The use of these embeddings enhances the encoder's ability to process structured data, improving the quality of the latent representations and the overall performance of the tokenizer. + +To train a more universal robot action tokenizer and reduce computational overhead, the model is trained using only action sequences as input, without additional conditional inputs. This design reduces complexity while maintaining the generalizability of the tokenizer. All models are trained on a single A100 GPU. For example, training on the Open X-Embodiment dataset requires just one A100 GPU and is completed in one week. + +# 3.4. Integrating Residual VQ-VAE as Action Tokenizer in VLA + +In this work, we replace the simple bin-based action tokenizer used in OpenVLA with a Residual VQ-VAE-based action tokenizer to improve the expressiveness and precision of action tokenization. Instead of discretizing action sequences into uniform bins, the action sequence $\mathbf{a}_{t:t + n}$ is first processed through a pre-trained and frozen Residual VQ-VAE encoder $\phi (\cdot)$ , generating latent representations. These representations are then quantized into discrete codebook indices (tokens) $\{z_q^i\}_{i = 1}^{N_q}$ corresponding to $N_{q}$ Residual VQ layers. The quantized tokens $z_{q}^{i}$ are used as the action tokens for training and prediction in the Vision-Language Model (VLM). + +Unlike OpenVLA, where all token IDs are mapped to the range [0, 255], the token IDs generated by different VQ layers in the Residual VQ-VAE are assigned unique, non-overlapping ranges. Specifically, tokens from the $i$ -th VQ layer are offset by $(i - 1) \times 256$ , ensuring that: + +$$ +z _ {q} ^ {i} \in [ 2 5 6 \times (i - 1), 2 5 6 \times i - 1 ], \quad \forall i \in \{1, \ldots , N _ {q} \}. +$$ + +For example, tokens from the first layer are in the range [0, 255], those from the second layer are in [256, 511], and so on. This design avoids conflicts between token IDs from different layers, as tokens with the same ID in different layers represent semantically different features of the action space. + +The Vision-Language Model (VLM) is trained to predict these tokens directly. The loss function is the standard next-token prediction loss, computed as the cross-entropy + +between the predicted token distribution $\hat{z}_q^i$ and the ground truth token $z_q^i$ produced by the frozen Residual VQ-VAE: + +$$ +L _ {\mathrm {V L M}} = - \sum_ {i = 1} ^ {N _ {q}} \log P \left(\hat {z} _ {q} ^ {i} = z _ {q} ^ {i} \mid o _ {t - h: t}\right), \tag {2} +$$ + +where $o_{t - h:t}$ represents the input observation sequence, and $N_{q}$ is the number of Residual VQ layers. + +To further enhance computational efficiency, similar to OpenVLA, the least-used tokens in the vocabulary are replaced during fine-tuning. However, this replacement process respects the layer-specific token ID ranges, ensuring that the tokenization remains consistent and interpretable across all layers. + +This integration of a pre-trained and frozen Residual VQ-VAE as the action tokenizer in VLA enables a more expressive and scalable representation of actions. By leveraging hierarchical quantization with non-overlapping token ID ranges, the model achieves better action representation, avoids semantic confusion between layers, and ensures stable loss convergence during training. Furthermore, by using the VQ-VAE tokens to represent longer action sequences instead of predicting one action at a time, the model significantly reduces the number of tokens required for training and inference. This mechanism allows the model to predict more complex behavior sequences with fewer steps, leading to a substantial improvement in inference speed and computational efficiency. + +# 4. Experiments + +In our experiments, we first validate the effectiveness and scalability of the action tokenizer in VLA models using LIBERO simulator[29]. Subsequently, real-world experiments are conducted to further verify our hypothesis. We also investigate the impact of action tokenizers on the performance, inference speed, and long-horizon capabilities of VLA models, alongside ablation studies to evaluate key design choices. + +# 4.1. Simulation Experiments + +# 4.1.1. Experiment Setup + +We utilize the LIBERO benchmark[29] to validate and evaluate the effectiveness and scalability of the action tokenizer, using the Franka Panda robot. Specifically, the entire LIBERO task suite—including LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-10, and LIBERO-90—is used as the entire LIBERO dataset. Among these, LIBERO-90 comprises 90 short-horizon tasks, while the other task suites each contain 10 tasks, with 50 demonstrations per task. For evaluation, testing is conducted on LIBERO-goal, LIBERO-10 and LIBERO-90. + +Following the approach in OpenVLA[26], we filter out all "no-op" actions from the dataset. To preliminarily validate + +the effectiveness of using VQ-VAE as an action tokenizer in VLA models and its scalability, we train two versions of a convolutional residual VQ-VAE. The first version $\mathrm{VQ_M}$ is trained solely on the Maniskill dataset[33], and the second $\mathrm{VQ_{M + R}}$ is trained on a mixture of Maniskill dataset and RLBench dataset[24]. Both models are trained on a single A100 GPU with a batch size of 1024, which takes about only 1 week. + +We use two pre-trained VQ-VAE models as frozen action tokenizers for OpenVLA: $\mathrm{VQ_M}$ and $\mathrm{VQ_{M + R}}$ . LoRA is then applied to fine-tune OpenVLA on the LIBERO-90 dataset. Additionally, we fine-tune the original OpenVLA model on the LIBERO-90 dataset using LoRA as a baseline for comparison. For a fair comparison, all fine-tuning on the LIBERO-90 task suite is conducted for 400K gradient steps with a batch size of 4, using 4 A100-80GB GPUs and an action chunk length of $K = 5$ . + +# 4.1.2. Effectiveness of Conv Residual VQ-VAE + +In the early stages of the experiment, we selected two task suites, LIBERO-10 and LIBERO-GOAL, to conduct preliminary scaling validation experiments. Specifically, we used two variants of Residual VQ-VAE models: one with a simple MLP as the encoder and decoder, and the other with a larger 2D temporal convolutional network as the encoder and decoder. The results are summarized in Tab. 1. The findings indicate that using temporal convolutional networks as the encoder and decoder in the Residual VQ-VAE significantly outperforms the MLP-based architecture in terms of success rate. These results suggest that temporal convolutional networks can serve as an effective action tokenizer, capable of capturing temporal dependencies more effectively. + +Furthermore, when we increased the training data for the VQ-VAE model, transitioning from data solely derived from individual LIBERO tasks to the entire LIBERO dataset, the success rate consistently improved. This observation provides preliminary evidence supporting the scalability of our action tokenizer design. + +
Training Dataset of VQLIBERO-10 (%)LIBERO-GOAL (%)
Original OpenVLA-51.075.8
MLP Residual VQ-VAEALL-LIBERO53.472.6
LIBERO-1053.2-
LIBERO-GOAL-65.2
Conv Residual VQ-VAEALL-LIBERO60.075.2
LIBERO-1054.0-
LIBERO-GOAL-72.4
+ +# 4.1.3. Scaling Data Improves VQ-VAE Action Tokenizer Performance + +In the experiment, to avoid potential performance inflation from in-domain data during evaluation on LIBERO-90, we trained the VQ-VAE Action Tokenizer on completely out-of-domain ManiSkill and RLBench simulation data, rather than LIBERO data. Specifically, ManiSkill contains 30,000 action sequences, and RLBench includes 10,000 action sequences. We compare three models: the baseline, which fine-tunes the original OpenVLA; $\mathrm{VQ_M}$ , which uses a VQ-VAE trained on the Maniskill dataset as the action tokenizer for OpenVLA; and $\mathrm{VQ_{M + R}}$ , which uses a VQ-VAE trained on a mixture of the Maniskill and RLBench datasets as the action tokenizer for OpenVLA. + +On LIBERO-90, $\mathrm{VQ}_{\mathrm{M} + \mathrm{R}}$ achieved $80.98\%$ , a $7.45\%$ improvement over the OpenVLA baseline (73.53%), Tab. 2. Furthermore, an ablation study using only ManiSkill data for training $\mathrm{VQ_M}$ resulted in substantially lower performance, underscoring the critical role of sufficient synthetic data scale. + +Table 1. The evaluation results of residual VQ-VAE architectures. The results demonstrate that the Conv Residual VQ-VAE outperforms the MLP-based version, particularly when trained on the full LIBERO dataset (ALL-LIBERO), highlighting its ability to better capture temporal dependencies and improve success rates. + +
baseline(%)VQM(%)VQM+R(%)
LIEBRO-9073.5314.3880.98
+ +Table 2. Effectiveness of VQ-VAE Action Tokenizers in Scaling Simulation Data. The results demonstrate $\mathrm{VQ}_{\mathrm{M} + \mathrm{R}}$ reached 80.98%, outperforming the OpenVLA baseline by 7.45% + +# 4.2. Real-Word Experiment + +# 4.2.1. Experiment Setup + +Our robotic platform consists of a single Franka Research3 arm equipped with a third-person-view RealSense D435 camera mounted in a fixed position to capture environmental observations. The system operates at $20\mathrm{Hz}$ (moderately reduced from the native $100\mathrm{Hz}$ control frequency to balance training efficiency and motion continuity), with actions defined as absolute end-effector poses in SE(3) space (3D position + quaternion orientation). + +Our experimental benchmark comprises six manipulation tasks (4 short-horizon tasks, 2 long-horizon tasks) designed to evaluate the model's ability to handle varying task complexities. For each task, we collect 50 demonstrations and evaluate performance over 20 trials: + +1) Pull out a tissue paper: The robot need to grasp and pull out a single tissue paper. +2) Pick up the [TOY NAME]: The robot is required to pick up the specific toy, including the toy snake, the toy eggplant, and the toy chicken, resulting in a total of three tasks (with no other distractions). +3) Put the toy into the basket: The robot need to pick up the toy(no other distractions) and put it into the basket. + +![](images/efd59b2c1b974854cfccaa8958f9dbb0acb7d1164c031ebbdd600212dcb12473.jpg) +Libero 90 + +![](images/5941664259900e8d7070eeac30f5a567a288309c1c4327f96c1c7aff3428dfe7.jpg) + +![](images/26aae533fdde15bdac7d45a8aaf00d7b39292b4631a79c61c5fbe30fe4167b89.jpg) +Put the toy into the basket + +![](images/b0e83fe6cd6f42da81a249074e813a289863301b33d508c2e042c663f69746fd.jpg) +Put all cups into the basket + +![](images/c06351d4aff8518f0767b7dfe7064bf1e5db23b6bb41dd6f8122df474d269994.jpg) +Pull out a tissue paper +Put the toy into the drawer + +![](images/03969043ed6fbc0a6e2fcf0d629a6b27626f552c8bfa7fb1c17d5b4e9d65e72b.jpg) +Filp the pot upright + +![](images/1f8f58540c7da9cb63591941aedebeb95bfb1cd47247b090f2213a4060749113.jpg) +Figure 2. All Evaluation environments: We conduct comprehensive evaluations of VQ-VLA in both simulation and real-world settings. In simulation, evaluations are performed on the LIBERO-90 benchmark within the LIBERO dataset. And six diverse tasks are designed for real-world testing. + +![](images/1f09c0454d1293fb66575d628a2460b19e6946fdf52081c170ff7310c53f8171.jpg) +Pick up the{toy name} + +![](images/1f2b73ab429478a6d76663c1e41113736533ac73b5d281a6707ee9d05d2ad6ed.jpg) + +4) Flip the pot upright: We set a flipped pot on the platform, the robot need to flip and upright a fallen cooking pot. +5) Put all cups into the basket: We place two different cups on the table and set a few other things (toys) as distractions. The robot need to sequentially put two cups into the basket, testing long-horizon task. +6) Put the toy into the drawer: We place a toy on the table and set a few other things as distractions. The robot need to sequentially open the drawer, pick up the specific toy and put it into the drawer, and close the drawer, testing long-horizon task. + +We finetune each task separately for 100K gradient steps (batch size 4 across 4 A100-80GB GPUs) with action chunk length $\mathrm{K} = 5$ . During deployment. + +# 4.2.2. Performance on Short-Horizon Tasks + +Based on the in the simulation experiments, we further incorporated 120k synthetic trajectories from the ManiSkill dataset, co-training $\mathrm{VQ_{O + L + M}}$ with Open X-Embodiment dataset and LIBERO dataset to evaluate the scaling capability of our action tokenizer. We evaluated the performance of the action tokenizer trained with different datasets on four carefully designed short-horizon tasks, focusing on multi-task learning, precision manipulation, and dynamic adjustment capabilities. The results are shown in Fig. 3. + +It can be observed that when synthetic trajectories are added for co-training, the performance of the action tokenizer improves significantly. The average success rate increased from the baseline of $23\%$ to $46.25\%$ , with a notable improvement of $30\%$ in the "Flip the pot upright" task. In the "Pull out a tissue paper" task, which tests the robot's performance in high-precision dynamic operations (as this task requires continuous, fine-grained grasping and pulling motions), the baseline model achieved only a $5\%$ success rate, failing almost entirely. In contrast, models incorporating VQ achieved success rates of $20\%$ or higher. + +Furthermore, comparing $\mathrm{VQ_{O + L}}$ with $\mathrm{VQ_O}$ , the average success rate on short-horizon tasks increased by only $0.5\%$ . This may be attributed to the limited size of the LIBERO dataset, which has minimal impact on short-horizon tasks. However, the ManiSkill dataset is 50 times larger than LIBERO, leading to a significant improvement in success rates when used for training. + +# 4.2.3. Performance on Long-Horizon Tasks + +VQ-VLA demonstrates outstanding performance on long-horizon tasks ("Put all cups in the basket" and "Put the toy into the drawer"), significantly outperforming baseline model in both success rate and efficiency. In scenarios where baseline models achieve success rates as low as $15\%$ or nearly 0, the $\mathrm{VQ}_{O + L + M}$ model achieves significantly higher success rates of $50\%$ and $30\%$ , respectively (see Fig.3). For the "Put the toy into the drawer" task, a representative long-horizon scenario, the baseline model was only able to complete the first step of opening the drawer in 15 out of 20 trials, failing to proceed further in most cases. In contrast, the $\mathrm{VQ}_{O + L + M}$ model successfully opened the drawer in all test cases, demonstrating its robustness and reliability in handling complex sequential tasks. + +A key advantage of VQ-VLA in long-horizon tasks lies in its use of VQ-VAE as an action tokenizer, which enables the model to predict multiple actions in a single inference step. This design reduces the accumulation of errors over extended task sequences, making VQ-VLA particularly advantageous for tasks requiring long-term planning and execution. Additionally, in scenarios involving the consecutive execution of multiple subtasks, VQ-VLA not only achieves higher success rates but also significantly reduces task completion time + +![](images/5ad737e1f2527b4c3e74f5e553ca2dcb533b8594001f7a084d70e9c845c2144c.jpg) +Figure 3. Real-world experimental results: We compare the performance of Baseline, $\mathrm{VQ_0}$ , $\mathrm{VQ_{O + L}}$ , and $\mathrm{VQ_{O + L + M}}$ on both short-horizon and long-horizon tasks. In terms of the average success rate, all VQ-based models outperform the Baseline. The best-performing model, $\mathrm{VQ_{O + L + M}}$ , achieves a success rate that is $23.25\%$ higher than the Baseline on both short-horizon and long-horizon tasks. Additionally, the results show that $\mathrm{VQ_{O + L + M}}$ outperforms $\mathrm{VQ_{O + L}}$ , which in turn outperforms $\mathrm{VQ_0}$ , indicating the effectiveness of incorporating synthetic data during training without compromising real-world performance. + +compared to baseline models, highlighting its superior execution efficiency and ability to handle long-horizon challenges effectively. + +# 4.2.4. Sim&Real Domain Gap Analysis + +To quantify the domain gap between synthetic data and real-world data, we trained a VQ-VAE model exclusively on the LIBERO dataset, referred to as $\mathrm{VQL}$ , for OpenVLA. The model was tested in three real-world tasks (one long-horizon and two short-horizon tasks), and the results are shown in Tab. 3. The results indicate that the performance of $\mathrm{VQL}$ is comparable to that of both $\mathrm{VQ_{O + L}}$ and $\mathrm{VQ_O}$ , suggesting that the domain gap between synthetic and real-world data is minimal. + +Although real-world data may contain noise, the inclusion of Open X-Embodiment data as a real-world dataset expands the data sources and enriches the diversity of data types, which effectively enhances the model's generalization and robustness. + +# 4.2.5. Inference Speed Comparison + +During the real-world experiments, we measured the action execution frequency of VQ-VLA and compared it with the original OpenVLA. The results are summarized in Tab. 4. As shown in the table, with a compression ratio of 5 in VQ-VAE, the inference speed is nearly tripled. This significant improvement greatly facilitates real-time performance in practical applications. + +
Put the toy into the drawer (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline5.030.020.0
VQo15.045.035.0
VQL10.055.035.0
VQo+L10.045.035.0
VQo+L+M25.060.045.0
+ +Table 3. Performance Comparison Across Real-World Tasks: We observe that the performance of $\mathrm{VQ_L}$ is comparable to that of both $\mathrm{VQ_{O + L}}$ and $\mathrm{VQ_O}$ , indicating that the domain gap between synthetic and real-world data is minimal. + +
Frequency (Hz)
VQ-VLA11.84
OpenVLA4.16
+ +Table 4. The Results of Frequencies. We report the comparison results of our VQ-VLA and baseline OpenVLA. + +# 4.3. Ablation Studies + +In this section, we report some ablation studies to show the effectiveness of the design choices of our method. + +# 4.3.1. Action Chunking via VQ-VAE and Autoregressive Output + +The original OpenVLA model generates a single action per inference step based solely on the current observation, resulting in a step-by-step execution approach. To improve efficiency and handle longer action sequences, we extend + +OpenVLA to output a sequence of five actions in an autoregressive manner, similar to the action chunking mechanism in VQ-VLA. Specifically, the model predicts the next action based on the current observation and previously generated actions, allowing it to generate a sequence of actions in a single inference step. This autoregressive approach serves as an alternative action chunking strategy by grouping multiple actions for execution. + +To evaluate the effectiveness of different action chunking strategies, we design ablation experiments comparing the autoregressive output of OpenVLA to the VQ-based action chunking method in VQ-VLA. The results, shown in Tab. 5, indicate that the use of the autoregressive approach for action chunking in the original OpenVLA leads to a significant drop in success rate compared to the baseline. In real-world experiments, when using the autoregressive approach as a form of action chunking, the spatial magnitude of each action chunk is relatively small, resulting in slower overall execution of actions. Compared to the Action Chunking via VQ-VAE method, even with the same action chunk size, the time required to reach the same target location is significantly longer. Additionally, this approach tends to exhibit shortcut learning, where the model directly copies previous actions within the same action chunk. By analyzing the output values of each action chunk, it is observed that multiple actions within a single chunk have remarkably similar values, indicating a lack of diversity in the predicted actions. + +Overall, Action Chunking via VQ-VAE not only improves inference speed but also enhances the performance of VLA by generating more effective and diverse action sequences. This makes it a more suitable approach for addressing real-world long-horizon tasks. + +
Action ChunkingLIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline74.7630.020.0
Autoregressive Output66.5310.00.0
VQ-based (VQo+L+M)86.6160.045.0
+ +# 4.3.2. Embedding Integration Effectiveness + +To evaluate the impact of embeddings, we conducted an ablation study comparing the model's performance with and without embeddings. Specifically, the baseline model processes raw action sequences directly, while the enhanced model incorporates both time embedding and action-type embedding. + +As shown in Tab. 6, the model with embeddings significantly outperforms the baseline in terms of success rate. This + +demonstrates that the integration of embeddings improves the encoder's ability to represent structured action sequences, leading to better overall performance. + +Table 5. Comparison of action chunking methods: We evaluate the performance on three tasks (one simulator task and two real-world tasks). The results show that the Autoregressive Output used in OpenVLA performs poorly as an action chunking method, with a significant gap compared to the VQ-based approach. This indicates that the VQ-based method is more effective. + +
LIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
VQoL (w.o. Embeddings)85.1740.035.0
VQoL (w Embeddings)86.1645.035.0
+ +Table 6. Embedding integration improves performance. The table compares models with and without embeddings across three tasks, showing that embeddings enhance success rates, especially for "Flip the pot upright." + +# 5. Limitations and Future Works + +In our work, a series of extensive experiments demonstrate that our proposed action tokenizers can improve VLA performance and inference speed while achieving scalability on simulated data. Despite these promising results, there still remain some limitations and opportunities for future work. First, our action tokenizers can be further extended to larger-scale simulated datasets. Second, our work improves inference speed by decoding multi-step action sequences, which can be further combined with techniques such as distillation [38] and quantization [16] of VLMs in the future. Finally, the architecture design of our action tokenizers can be further improved, for example, by incorporating the frequency of action data as a extra condition. In summary, we hope that our work can serve as a strong baseline and inspire future research. + +# 6. Conclusions + +In this paper, we propose a general convolutional residual VQ-VAE framework for action tokenizers that can seamlessly integrate with state-of-the-art VLA models. Our VQ-VAE is trained on 100 times more data than previous methods and can be directly transferred to downstream both real-world and simulated robotic manipulation tasks. Extensive experiments conducted in both simulated and real-world environments validate that the proposed convolutional residual VQ-VAE framework not only enhances the performance of VLA policies but also accelerates inference. Finally, our VQ-VAE also demonstrates the ability to scale effectively with large-scale simulated data. + +# Acknowledgments + +This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2 +[3] Jean-Baptiste Alayrac, Jeffrey Donahue, Pauline Luc, Alexis Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35:23716-23736, 2022. 2 +[4] Arthur Allshire, Roberto Martin-Martin, Charles Lin, Shawn Manuel, Silvio Savarese, and Animesh Garg. Laser: Learning a latent action space for efficient reinforcement learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6650-6656, 2021. 2 +[5] Suneel Belkhale and Dorsa Sadigh. Minivla: A better vla with a smaller footprint. https://ai.stanford.edu/blog/minivla/, 2024. 1, 2, 3 +[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. $\pi_0$ : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2 +[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2 +[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2 +[9] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 2 +[10] Wei Chen, Yifeng Zhang, Hao Li, Xufeng Wang, and Ming Liu. Anyvlm: Unified vision-language model for any robot morphology. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 1-8. IEEE, 2024. 2 +[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2 +[12] Xiaoyu Chen, Junliang Guo, Tianyu He, Chuheng Zhang, Pushi Zhang, Derek Cathera Yang, Li Zhao, and Jiang Bian. Igor: Image-goal representations are the atomic control + +units for foundation models in embodied ai. arXiv preprint arXiv:2411.00785, 2024. 2 +[13] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2 +[14] Norman Di Palo and Edward Johns. Keypoint action tokens enable in-context imitation learning in robotics. arXiv preprint arXiv:2403.19578, 2024. 2 +[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 2 +[16] Kazuki Egashira, Mark Vero, Robin Staab, Jingxuan He, and Martin Vechev. Exploiting lIm quantization. Advances in Neural Information Processing Systems, 37:41709-41732, 2025. 8 +[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2 +[18] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2 +[19] Philip Gage. A new algorithm for data compression. The C Users Journal, 12(2):23-38, 1994. 2 +[20] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation. In 7th Annual Conference on Robot Learning. 2 +[21] Abhishek Gupta, Vikash Kumar, Corey Lynch, Sergey Levine, and Karol Hausman. Relay policy learning: Solving long-horizon tasks via imitation and reinforcement learning. arXiv preprint arXiv:1910.11956, 2019. 2 +[22] Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. 2 +[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1 +[24] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5 +[25] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.3 +[26] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An + +open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2, 4 +[27] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025. 2 +[28] Seungjae Lee, Yibin Wang, Haritheja Etukuru, H Jin Kim, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Behavior generation with latent actions. arXiv preprint arXiv:2403.03181, 2024. 1, 2, 3 +[29] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 1, 2, 3, 4 +[30] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2 +[31] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1:3, 2020. 1 +[32] Atharva Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, and Animesh Garg. Quest: Self-supervised skill abstractions for learning continuous control. Advances in Neural Information Processing Systems, 37:4062-4089, 2025. 2, 12 +[33] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 1, 3, 5 +[34] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 1, 2, 3 +[35] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025. 1, 2, 12 +[36] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[37] Fengyuan Shi, Zhuoyan Luo, Yixiao Ge, Yujiu Yang, Ying Shan, and Limin Wang. Taming scalable visual tokenizer for autoregressive image generation. arXiv preprint arXiv:2412.02692, 2024. 2 +[38] Sharath Turuvekere Sreenivas, Saurav Muralidharan, Raviraj Joshi, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Gerald Shen, Jiaqi Zeng, Zijia Chen, Yoshi Suhara, Shizhe Diao, et al. Llm pruning and distillation in practice: The minitron approach. arXiv preprint arXiv:2408.11796, 2024. 8 +[39] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model + +beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024.1, 2 +[40] Andrew Szot, Bogdan Mazoure, Harsh Agrawal, R Devon Hjelm, Zsolt Kira, and Alexander Toshev. Grounding multimodal large language models in actions. Advances in Neural Information Processing Systems, 37:20198-20224, 2024. 2 +[41] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2 +[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[43] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 2 +[44] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 2 +[45] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2 +[46] Jange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. arXiv preprint arXiv:2306.05716, 2023. 2 +[47] Jange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 2 +[48] Jiangge Yang, Haoyi Zhu, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Tra-moe: Learning trajectory prediction model from multiple domains for adaptive policy conditioning. arXiv preprint arXiv:2411.14519, 2024. 2 +[49] Jiange Yang, Yansong Shi, Haoyi Zhu, Mingyu Liu, Kaijing Ma, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Como: Learning continuous latent motion from internet videos for scalable robot learning. arXiv preprint arXiv:2505.17006, 2025. 2 +[50] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 2 +[51] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2 +[52] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1, 3 +[53] Canyu Zhao, Mingyu Liu, Wen Wang, Weihua Chen, Fan Wang, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2 + +[54] Canyu Zhao, Mingyu Liu, Huanyi Zheng, Muzhi Zhu, Zhiyue Zhao, Hao Chen, Tong He, and Chunhua Shen. Dioxide: A generalist diffusion model for visual perceptual tasks. arXiv preprint arXiv:2502.17157, 2025. 2 +[55] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2 +[56] Wei Zhao, Pengxiang Ding, Min Zhang, Zhefei Gong, Shuanghao Bai, Han Zhao, and Donglin Wang. Vlas: Vision-language-action model with speech instructions for customized robot manipulation. arXiv preprint arXiv:2502.13508, 2025. 2 +[57] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2 +[58] Jinliang Zheng, Jianxiong Li, Dongxiu Liu, Yinan Zheng, Zhihao Wang, Zhonghong Ou, Yu Liu, Jingjing Liu, Ya-Qin Zhang, and Xianyuan Zhan. Universal actions for enhanced embodied foundation models, 2025. 2, 13 +[59] Wenxuan Zhou, Sujay Bajracharya, and David Held. Plas: Latent action space for offline reinforcement learning, 2020. 2 +[60] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2 +[61] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2 \ No newline at end of file diff --git a/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/images.zip b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ea4e908147ca3b14878b0f37a932b40dffd7821 --- /dev/null +++ b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44cff65dab59b83ba5ba4c24e98b1deff422435c82354280c9cd183765c80de2 +size 296300 diff --git a/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/layout.json b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bcbd4fc8e9a8360a848c165f1906525d133f7d8c --- /dev/null +++ b/2025/VQ-VLA_ Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers/layout.json @@ -0,0 +1,9408 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 103, + 102, + 509, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 102, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 103, + 102, + 509, + 140 + ], + "type": "text", + "content": "VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "spans": [ + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": "Yating Wang" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": " Haoyi Zhu" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{23}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": " Mingyu Liu" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{24}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": " Jange Yang" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{25}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": " Hao-Shu Fang" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": " Tong He" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{27\\dagger}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": "Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University \n" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 165, + 541, + 227 + ], + "type": "text", + "content": "Shanghai AI Lab USTC ZJU NJU SJTU SII Corresponding Author" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 228, + 471, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 228, + 471, + 241 + ], + "spans": [ + { + "bbox": [ + 136, + 228, + 471, + 241 + ], + "type": "text", + "content": "https://xiaoxiao0406.github.io/vqvla.github.io" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "spans": [ + { + "bbox": [ + 151, + 269, + 200, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 296, + 297, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 296, + 297, + 595 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 297, + 595 + ], + "type": "text", + "content": "In this paper, we introduce an innovative vector quantization based action tokenizer built upon the largest-scale action trajectory dataset to date, leveraging over 100 times more data than previous approaches. This extensive dataset enables our tokenizer to capture rich spatiotemporal dynamics, resulting in a model that not only accelerates inference but also generates smoother and more coherent action outputs. Once trained, the tokenizer can be seamlessly adapted to a wide range of downstream tasks in a zero-shot manner, from short-horizon reactive behaviors to long-horizon planning. A key finding of our work is that the domain gap between synthetic and real action trajectories is marginal, allowing us to effectively utilize a vast amount of synthetic data during training without compromising real-world performance. To validate our approach, we conducted extensive experiments in both simulated environments and on real robotic platforms. The results demonstrate that as the volume of synthetic trajectory data increases, the performance of our tokenizer on downstream tasks improves significantly—most notably, achieving up to a " + }, + { + "bbox": [ + 54, + 296, + 297, + 595 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 296, + 297, + 595 + ], + "type": "text", + "content": " higher success rate on two real-world tasks in long-horizon scenarios. These findings highlight the potential of our action tokenizer as a robust and scalable solution for real-time embodied intelligence systems, paving the way for more efficient and reliable robotic control in diverse application domains." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 620, + 137, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 137, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 137, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "Tokenization plays a critical role in recent generative models, including large language models (LLMs) [1, 31], image and video generation models [17, 23, 39], and vision-language-action (VLA) models [8, 26]. One of the key benefits of tokenization is that it compresses the input space. By reducing high-dimensional continuous data into a compact sequence" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 555, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 317 + ], + "type": "text", + "content": "of tokens, the complexity of the learning task is substantially reduced. Compared to image patches and language tokens, action sequences are inherently easier to compress because of their spatio-temporal continuity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 593 + ], + "type": "text", + "content": "Recently, several studies have started to explore action quantized tokenization for Vision-Language-Action (VLA) models [5, 35], demonstrating promising potential. Effective action tokenization not only significantly enhances the downstream performance of VLA models, especially for tasks involving long-horizon planning, but also markedly improves their training and inference efficiency. In this paper, we delve deeper into the potential of action tokenization, with a specific emphasis on its scalability and accuracy. We first find that the more precise the tokenization, the more pronounced the improvements in long-horizon action modeling. This motivates us to train the VQ tokenizer with well-scaled action trajectories to cover various tasks. Additionally, we observe that action trajectories, unlike visual and physical modalities, exhibit minimal domain gaps between real-world and simulated environments. This characteristic enables effective scaling of action tokenizers through the extensive use of synthetic action data. Moreover, training action tokenizers is computationally lightweight compared to scaling entire VLA models. Consequently, focusing on scaling action tokenization emerges as a highly cost-effective strategy, requiring fewer computational and data resources while delivering significant performance enhancements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": "Specifically, we propose a convolutional residual VQ-VAE [5, 28, 52] framework for training action tokenizers. To effectively train the model, we propose a progressive training strategy: Initially, we train the tokenizer on real-world robotic datasets, such as OpenX-Embodiment [34], which typically contain noisy and jittery trajectories. Subsequently, we gradually integrate cleaner and smoother synthetic data from large-scale simulated robotic datasets, such as LIBERO [29] and ManiSkill [33]. This progressive approach allows the VQ model to converge toward smoother" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11089" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 131 + ], + "type": "text", + "content": "and more stable representations. Compared to previous approaches that typically rely on training with single-task datasets, our method expands the tokenizer training dataset by more than 100 times, effectively covering a broad spectrum of downstream tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 297, + 312 + ], + "type": "text", + "content": "We conduct extensive experiments in both simulated and real-world environments. First, we evaluate our VQ-VAE action tokenizers in the LIBERO simulator, where the results demonstrate the effectiveness of the convolutional residual VQ-VAE and provide preliminary validation for the hypothesis that the VQ-VAE action tokenizer can leverage synthetic data for scaling. Additionally, real-world experiments with a Franka Research 3 robot further validate the superiority of our approach. Specifically, our findings are as follows: (1) as the amount of simulated action data increases, the VQ-VAE tokenizers exhibit linear scaling properties in improving VLA success rates; (2) the tokenizers significantly enhance inference speed and smoothness of VLA models; and (3) the tokenizers effectively reduce cumulative errors, enabling better performance in long-horizon tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 312, + 251, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 312, + 251, + 323 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 251, + 323 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 323, + 295, + 406 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 323, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 295, + 346 + ], + "type": "text", + "content": "- We propose a general convolutional residual VQ-VAE-based framework for action tokenizers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 348, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 371 + ], + "type": "text", + "content": "- We demonstrate that action tokenizers can be effectively scaled by leveraging large-scale simulated action data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 372, + 295, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 406 + ], + "type": "text", + "content": "- We prove that our action tokenizers improve the performance, inference speed, and long-horizon capabilities of VLA models." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 418, + 147, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 147, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 147, + 430 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 297, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 297, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 297, + 653 + ], + "type": "text", + "content": "Vision-Language-Action Models. Vision-Language-Action (VLA) models[7, 9, 11, 18, 26, 41, 45, 46, 56, 57] bridge visual-language understanding with robot control by mapping multimodal inputs to action outputs. Based on vision-language models (VLM) [3, 10, 11, 15], VLAs represent robot actions (e.g., 6DoF motion, gripper control) as discrete tokens compatible with text-based output. For example, RT-1[7] and RT-2[8] showed that dividing continuous actions into discrete bins allows integration with VLMs, enabling zero-shot generalization using web-scale pretraining. Recent works[26, 34] further leverage this approach, demonstrating improved task generalization through large VLM backbones. Building on this foundation, recent works have explored various approaches to improve VLA, such as leveraging 3D visual inputs [57], integrating chain-of-thought reasoning [51], and employing parallel decoding strategies [27]. In our work, we aim to simultaneously improve the performance and execution speed of VLA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Tokenization. The tokenizer is one of the fundamental components in language and vision generation tasks, especially for autoregressive transformer-based generation models. For language generation tasks, mainstream methods adopt byte pair encoding (BPE) [19, 36] to compress input text. For" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "vision generation tasks, recent works [17, 30, 37, 39, 53, 54] mainly use vector quantization [42] for tokenization to map continuous visual signals into a discrete token sequence. Similarly, in robot learning, many works [7, 8, 28, 32, 35] model action prediction as a generative problem and tokenize the robot action modality to capture the multi-modal distribution within skills and adapt it to autoregressive generative policies. In our work, we propose a general convolutional residual VQ-VAE to tokenize robot action. Different from several related works [28, 32] that use VQ-VAE to tokenize robot actions, our VQ-VAE is trained on large-scale data and is applicable to all tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 556, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 556, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 556, + 517 + ], + "type": "text", + "content": "Action Representation. Action representation is one of the core components in robot policy learning. Classic action representation schemes include high-level sub-tasks [2, 15], action primitives [22, 43], keypoints [14, 20], as well as low-level continuous end-effector pose [29, 34] or joint state [21]. Recently, imitation learning-based end-to-end approaches [7, 13, 55, 60, 61] often utilize the latter. Some works[6, 40, 45, 58] employ regression-based output layers or integrate new weights for diffusion decoding for action output. Additionally, some works extract corresponding predictive signals from action-less video data to represent actions, such as pixels [9, 47], trajectories [44, 48], and latent motion [12, 49, 50]. To represent complex multi-modal distributions and adapt to discrete generative models, many works discretize continuous low-level actions. Common approaches include per-dimension and per-timestep binning discretization [7, 26], which struggles to perform well in high-frequency tasks [35]. Another method is VQ-VAE based discretization [4, 5, 28, 32, 59]; a limitation here is the need to train a separate action tokenizer for each specific task. Fast [35] utilized a discrete cosine transform (DCT) to learn a non-learnable action tokenizer from real-robot data. In this work, we propose a novel approach: expanding large-scale synthetic data to train the VQ-VAE action tokenizer, aiming to enhance its generalizability and efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 530, + 374, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 530, + 374, + 542 + ], + "spans": [ + { + "bbox": [ + 314, + 530, + 374, + 542 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 551, + 470, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 470, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 470, + 563 + ], + "type": "text", + "content": "3.1. Preliminaries: VLA models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "We use OpenVLA[26] as our backbone model. OpenVLA's origin formulation is to adopt a discrete tokenization strategy for robot action prediction through fine-tuning the Prismatic-7B VLM backbone. The method frames action prediction as a vision-language task, mapping input observation images and natural language instructions to discrete robot action sequences. Specifically, continuous robot actions are discretized into 256 bins per dimension, with bin boundaries determined by the 1st and 99th percentiles of training data distributions rather than min-max ranges to mitigate outlier effects. This discretization converts N-dimensional actions into N discrete integers (0-255). To embed these into the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11090" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 208, + 316 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 208, + 316 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 208, + 316 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 208, + 316 + ], + "type": "image", + "image_path": "f77b0c565c1e8248e90c4431c198b617e3caa555a7fad3f1e489518773b4c709.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 323, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 54, + 323, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 555, + 369 + ], + "type": "text", + "content": "Figure 1. The VQ-VLA pipeline, consisting of two main stages: (1) training a general convolutional residual VQ-VAE and (2) fine-tuning OpenVLA using the LoRA approach. Specifically, a general convolutional residual VQ-VAE is first trained on the Open X-Embodiment dataset, LIEBRO, and ManiSkill datasets. The trained VQ-VAE is then frozen and serves as an action tokenizer for OpenVLA, replacing the simple binning method. In the second stage, OpenVLA is fine-tuned using the LoRA technique to optimize its performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 69, + 553, + 314 + ], + "blocks": [ + { + "bbox": [ + 225, + 69, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 225, + 69, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 225, + 69, + 553, + 314 + ], + "type": "image", + "image_path": "11f5fe7ecf165e4c1f5fa22894f6e01a9b2087c70ec6e29a3399938973ffd06a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 388, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 437 + ], + "type": "text", + "content": "LLM's vocabulary, OpenVLA overwrites the 256 least-used tokens in the Llama tokenizer (last 256 tokens) rather than using special tokens, as the original tokenizer only reserves 100 special tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 263, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 263, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 263, + 456 + ], + "type": "text", + "content": "3.2. Action Tokenizer via Residual VQ-VAE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 555 + ], + "type": "text", + "content": "Based on Residual VQ-VAE [5, 28, 52], we design our encoder and decoder inspired by VAE in pyramidal flow matching [25]. Instead of the simple Multi-Layer Perceptron (MLP) used in Residual VQ-VAE [28, 52], we integrate 2D temporal convolutional layers, motivated by their ability to efficiently capture local relationships and hierarchical temporal dependencies, addressing the scaling limitations of MLPs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": "Given an input action sequence " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{t:t + n}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": " is the sequence length and " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": " the action dimensionality, the encoder " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{enc}}" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ", composed of 2D temporal convolutional layers, transforms the sequence into a latent embedding " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^k" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ", expressed as " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\phi_{\\mathrm{enc}}(\\mathbf{a}_{t:t + n})" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ". To compress " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ", we apply Residual Vector Quantization (RVQ) [52], decomposing " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": " into quantized residuals: " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(\\mathbf{x}) = \\sum_{i = 1}^{N_q}\\mathbf{q}_i(\\mathbf{r}_i)" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_1 = \\mathbf{x}" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i + 1} = \\mathbf{r}_i - \\mathbf{q}_i(\\mathbf{r}_i)" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 55, + 556, + 295, + 665 + ], + "type": "text", + "content": " denotes the number of quantization stages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "The quantized embedding " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " is passed through the decoder " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{dec}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", which uses 2D temporal deconvolutional layers to reconstruct the sequence " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_{t:t+n}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", ensuring temporal structure preservation: " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_{t:t+n} = \\phi_{\\mathrm{dec}}(\\mathbf{q}(\\mathbf{x}))" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": "To train the framework, we minimize the total loss " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": ", a weighted combination of reconstruction loss " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": ", vector quantization (VQ) loss " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{codebook}}" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": ", and commitment loss " + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{commit}}" + }, + { + "bbox": [ + 313, + 388, + 555, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 444, + 554, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 444, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 328, + 444, + 554, + 475 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\left\\| \\mathbf {a} _ {t: t + n} - \\hat {\\mathbf {a}} _ {t: t + n} \\right\\| _ {2} ^ {2} \\tag {1} \\\\ + \\lambda \\left(\\| \\operatorname {s g} (\\mathbf {x}) - \\mathbf {q} (\\mathbf {x}) \\| _ {2} ^ {2} + \\| \\mathbf {x} - \\operatorname {s g} (\\mathbf {q} (\\mathbf {x})) \\| _ {2} ^ {2}\\right), \\\\ \\end{array}", + "image_path": "e0f489f94568690323fbfab8b689b3e66a654671918c77ef67c9267daa3c7e59.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "text", + "content": " denotes the stop-gradient operation, and " + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "text", + "content": " balances the loss components. We set " + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda = 4" + }, + { + "bbox": [ + 313, + 483, + 555, + 532 + ], + "type": "text", + "content": " in our experiments. This design ensures efficient, scalable, and robust encoding of structured temporal data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 538, + 466, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 466, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 466, + 551 + ], + "type": "text", + "content": "3.3. Training Residual VQ-VAE" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "type": "text", + "content": "We train three versions of the Residual VQ-VAE to evaluate the scaling capability of our action tokenizer: (1) using only the Open X-Embodiment dataset[34], (2) combining Open X-Embodiment and Libero datasets[29], and (3) combining Open X-Embodiment, Libero, and ManiSkill[33] datasets. These experiments were designed to test our hypothesis that the action tokenizer can effectively scale with simulated ata." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 639, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 676 + ], + "type": "text", + "content": "To improve the encoder's ability to process temporal and spatial information, we introduced two types of embeddings before the action sequences are passed into the encoder:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "- Time Embedding: A sinusoidal time embedding was added to encode temporal information at varying frequencies. This embedding allows the model to capture both" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11091" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 297, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 297, + 108 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 297, + 108 + ], + "type": "text", + "content": "low-frequency and high-frequency temporal patterns in the input actions, improving its ability to represent fine-grained temporal details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 192 + ], + "type": "text", + "content": "- **Action-Type Embedding:** Learnable embeddings were added for the different components of the action sequence (e.g., XYZ positions, Euler angles, and gripper states). Since the 7 dimensions of the action vector have distinct meanings, this embedding provides a strong prior for the model, helping it distinguish and process the unique roles of each dimension more effectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 242 + ], + "type": "text", + "content": "The use of these embeddings enhances the encoder's ability to process structured data, improving the quality of the latent representations and the overall performance of the tokenizer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 243, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 339 + ], + "type": "text", + "content": "To train a more universal robot action tokenizer and reduce computational overhead, the model is trained using only action sequences as input, without additional conditional inputs. This design reduces complexity while maintaining the generalizability of the tokenizer. All models are trained on a single A100 GPU. For example, training on the Open X-Embodiment dataset requires just one A100 GPU and is completed in one week." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 296, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 372 + ], + "type": "text", + "content": "3.4. Integrating Residual VQ-VAE as Action Tokenizer in VLA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "text", + "content": "In this work, we replace the simple bin-based action tokenizer used in OpenVLA with a Residual VQ-VAE-based action tokenizer to improve the expressiveness and precision of action tokenization. Instead of discretizing action sequences into uniform bins, the action sequence " + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{t:t + n}" + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "text", + "content": " is first processed through a pre-trained and frozen Residual VQ-VAE encoder " + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "text", + "content": ", generating latent representations. These representations are then quantized into discrete codebook indices (tokens) " + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\{z_q^i\\}_{i = 1}^{N_q}" + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "text", + "content": " Residual VQ layers. The quantized tokens " + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "inline_equation", + "content": "z_{q}^{i}" + }, + { + "bbox": [ + 55, + 377, + 296, + 523 + ], + "type": "text", + "content": " are used as the action tokens for training and prediction in the Vision-Language Model (VLM)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 524, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 296, + 584 + ], + "type": "text", + "content": "Unlike OpenVLA, where all token IDs are mapped to the range [0, 255], the token IDs generated by different VQ layers in the Residual VQ-VAE are assigned unique, non-overlapping ranges. Specifically, tokens from the " + }, + { + "bbox": [ + 55, + 524, + 296, + 584 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 524, + 296, + 584 + ], + "type": "text", + "content": "-th VQ layer are offset by " + }, + { + "bbox": [ + 55, + 524, + 296, + 584 + ], + "type": "inline_equation", + "content": "(i - 1) \\times 256" + }, + { + "bbox": [ + 55, + 524, + 296, + 584 + ], + "type": "text", + "content": ", ensuring that:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 593, + 288, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 288, + 609 + ], + "type": "interline_equation", + "content": "z _ {q} ^ {i} \\in [ 2 5 6 \\times (i - 1), 2 5 6 \\times i - 1 ], \\quad \\forall i \\in \\{1, \\ldots , N _ {q} \\}.", + "image_path": "a8a9d92f3814824e7a0354ee427ce5cbad2ceb51c0d64d2475f48f0120e17273.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "For example, tokens from the first layer are in the range [0, 255], those from the second layer are in [256, 511], and so on. This design avoids conflicts between token IDs from different layers, as tokens with the same ID in different layers represent semantically different features of the action space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "The Vision-Language Model (VLM) is trained to predict these tokens directly. The loss function is the standard next-token prediction loss, computed as the cross-entropy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 99 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 99 + ], + "type": "text", + "content": "between the predicted token distribution " + }, + { + "bbox": [ + 313, + 72, + 555, + 99 + ], + "type": "inline_equation", + "content": "\\hat{z}_q^i" + }, + { + "bbox": [ + 313, + 72, + 555, + 99 + ], + "type": "text", + "content": " and the ground truth token " + }, + { + "bbox": [ + 313, + 72, + 555, + 99 + ], + "type": "inline_equation", + "content": "z_q^i" + }, + { + "bbox": [ + 313, + 72, + 555, + 99 + ], + "type": "text", + "content": " produced by the frozen Residual VQ-VAE:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 355, + 105, + 555, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 105, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 355, + 105, + 555, + 138 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {V L M}} = - \\sum_ {i = 1} ^ {N _ {q}} \\log P \\left(\\hat {z} _ {q} ^ {i} = z _ {q} ^ {i} \\mid o _ {t - h: t}\\right), \\tag {2}", + "image_path": "3e67196463c22ac1a8629f3dfa9dcaa9279fbca8c3233af5757c999101368780.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 143, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 143, + 555, + 167 + ], + "type": "inline_equation", + "content": "o_{t - h:t}" + }, + { + "bbox": [ + 313, + 143, + 555, + 167 + ], + "type": "text", + "content": " represents the input observation sequence, and " + }, + { + "bbox": [ + 313, + 143, + 555, + 167 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 313, + 143, + 555, + 167 + ], + "type": "text", + "content": " is the number of Residual VQ layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 167, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 239 + ], + "type": "text", + "content": "To further enhance computational efficiency, similar to OpenVLA, the least-used tokens in the vocabulary are replaced during fine-tuning. However, this replacement process respects the layer-specific token ID ranges, ensuring that the tokenization remains consistent and interpretable across all layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 239, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 407 + ], + "type": "text", + "content": "This integration of a pre-trained and frozen Residual VQ-VAE as the action tokenizer in VLA enables a more expressive and scalable representation of actions. By leveraging hierarchical quantization with non-overlapping token ID ranges, the model achieves better action representation, avoids semantic confusion between layers, and ensures stable loss convergence during training. Furthermore, by using the VQ-VAE tokens to represent longer action sequences instead of predicting one action at a time, the model significantly reduces the number of tokens required for training and inference. This mechanism allows the model to predict more complex behavior sequences with fewer steps, leading to a substantial improvement in inference speed and computational efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 415, + 395, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 395, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 395, + 430 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 435, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 532 + ], + "type": "text", + "content": "In our experiments, we first validate the effectiveness and scalability of the action tokenizer in VLA models using LIBERO simulator[29]. Subsequently, real-world experiments are conducted to further verify our hypothesis. We also investigate the impact of action tokenizers on the performance, inference speed, and long-horizon capabilities of VLA models, alongside ablation studies to evaluate key design choices." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 537, + 451, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 451, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 451, + 551 + ], + "type": "text", + "content": "4.1. Simulation Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 555, + 421, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 421, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 421, + 567 + ], + "type": "text", + "content": "4.1.1. Experiment Setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": "We utilize the LIBERO benchmark[29] to validate and evaluate the effectiveness and scalability of the action tokenizer, using the Franka Panda robot. Specifically, the entire LIBERO task suite—including LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-10, and LIBERO-90—is used as the entire LIBERO dataset. Among these, LIBERO-90 comprises 90 short-horizon tasks, while the other task suites each contain 10 tasks, with 50 demonstrations per task. For evaluation, testing is conducted on LIBERO-goal, LIBERO-10 and LIBERO-90." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Following the approach in OpenVLA[26], we filter out all \"no-op\" actions from the dataset. To preliminarily validate" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11092" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "the effectiveness of using VQ-VAE as an action tokenizer in VLA models and its scalability, we train two versions of a convolutional residual VQ-VAE. The first version " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_M}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " is trained solely on the Maniskill dataset[33], and the second " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{M + R}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " is trained on a mixture of Maniskill dataset and RLBench dataset[24]. Both models are trained on a single A100 GPU with a batch size of 1024, which takes about only 1 week." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "text", + "content": "We use two pre-trained VQ-VAE models as frozen action tokenizers for OpenVLA: " + }, + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_M}" + }, + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{M + R}}" + }, + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "text", + "content": ". LoRA is then applied to fine-tune OpenVLA on the LIBERO-90 dataset. Additionally, we fine-tune the original OpenVLA model on the LIBERO-90 dataset using LoRA as a baseline for comparison. For a fair comparison, all fine-tuning on the LIBERO-90 task suite is conducted for 400K gradient steps with a batch size of 4, using 4 A100-80GB GPUs and an action chunk length of " + }, + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 55, + 170, + 294, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 254, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 254, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 254, + 300 + ], + "type": "text", + "content": "4.1.2. Effectiveness of Conv Residual VQ-VAE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "type": "text", + "content": "In the early stages of the experiment, we selected two task suites, LIBERO-10 and LIBERO-GOAL, to conduct preliminary scaling validation experiments. Specifically, we used two variants of Residual VQ-VAE models: one with a simple MLP as the encoder and decoder, and the other with a larger 2D temporal convolutional network as the encoder and decoder. The results are summarized in Tab. 1. The findings indicate that using temporal convolutional networks as the encoder and decoder in the Residual VQ-VAE significantly outperforms the MLP-based architecture in terms of success rate. These results suggest that temporal convolutional networks can serve as an effective action tokenizer, capable of capturing temporal dependencies more effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 462, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 534 + ], + "type": "text", + "content": "Furthermore, when we increased the training data for the VQ-VAE model, transitioning from data solely derived from individual LIBERO tasks to the entire LIBERO dataset, the success rate consistently improved. This observation provides preliminary evidence supporting the scalability of our action tokenizer design." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 546, + 294, + 630 + ], + "blocks": [ + { + "bbox": [ + 56, + 546, + 294, + 630 + ], + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 630 + ], + "type": "table", + "html": "
Training Dataset of VQLIBERO-10 (%)LIBERO-GOAL (%)
Original OpenVLA-51.075.8
MLP Residual VQ-VAEALL-LIBERO53.472.6
LIBERO-1053.2-
LIBERO-GOAL-65.2
Conv Residual VQ-VAEALL-LIBERO60.075.2
LIBERO-1054.0-
LIBERO-GOAL-72.4
", + "image_path": "aaaf350bf8e655ad281ddaa6ba563d8def58b240dc8ba68d5f792c5302c7ec0e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 553, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 95 + ], + "type": "text", + "content": "4.1.3. Scaling Data Improves VQ-VAE Action Tokenizer Performance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 99, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 555, + 243 + ], + "type": "text", + "content": "In the experiment, to avoid potential performance inflation from in-domain data during evaluation on LIBERO-90, we trained the VQ-VAE Action Tokenizer on completely out-of-domain ManiSkill and RLBench simulation data, rather than LIBERO data. Specifically, ManiSkill contains 30,000 action sequences, and RLBench includes 10,000 action sequences. We compare three models: the baseline, which fine-tunes the original OpenVLA; " + }, + { + "bbox": [ + 313, + 99, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_M}" + }, + { + "bbox": [ + 313, + 99, + 555, + 243 + ], + "type": "text", + "content": ", which uses a VQ-VAE trained on the Maniskill dataset as the action tokenizer for OpenVLA; and " + }, + { + "bbox": [ + 313, + 99, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{M + R}}" + }, + { + "bbox": [ + 313, + 99, + 555, + 243 + ], + "type": "text", + "content": ", which uses a VQ-VAE trained on a mixture of the Maniskill and RLBench datasets as the action tokenizer for OpenVLA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "text", + "content": "On LIBERO-90, " + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ}_{\\mathrm{M} + \\mathrm{R}}" + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "text", + "content": " achieved " + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "inline_equation", + "content": "80.98\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "text", + "content": ", a " + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "inline_equation", + "content": "7.45\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "text", + "content": " improvement over the OpenVLA baseline (73.53%), Tab. 2. Furthermore, an ablation study using only ManiSkill data for training " + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_M}" + }, + { + "bbox": [ + 313, + 243, + 555, + 315 + ], + "type": "text", + "content": " resulted in substantially lower performance, underscoring the critical role of sufficient synthetic data scale." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 356, + 325, + 512, + 358 + ], + "blocks": [ + { + "bbox": [ + 55, + 639, + 296, + 694 + ], + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 694 + ], + "type": "text", + "content": "Table 1. The evaluation results of residual VQ-VAE architectures. The results demonstrate that the Conv Residual VQ-VAE outperforms the MLP-based version, particularly when trained on the full LIBERO dataset (ALL-LIBERO), highlighting its ability to better capture temporal dependencies and improve success rates." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 356, + 325, + 512, + 358 + ], + "lines": [ + { + "bbox": [ + 356, + 325, + 512, + 358 + ], + "spans": [ + { + "bbox": [ + 356, + 325, + 512, + 358 + ], + "type": "table", + "html": "
baseline(%)VQM(%)VQM+R(%)
LIEBRO-9073.5314.3880.98
", + "image_path": "38fdfe216970c7631b48275689770ef2026dc55a93435423548e145d48c2213c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 367, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 400 + ], + "type": "text", + "content": "Table 2. Effectiveness of VQ-VAE Action Tokenizers in Scaling Simulation Data. The results demonstrate " + }, + { + "bbox": [ + 313, + 367, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ}_{\\mathrm{M} + \\mathrm{R}}" + }, + { + "bbox": [ + 313, + 367, + 555, + 400 + ], + "type": "text", + "content": " reached 80.98%, outperforming the OpenVLA baseline by 7.45%" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 426, + 446, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 446, + 438 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 446, + 438 + ], + "type": "text", + "content": "4.2. Real-Word Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 444, + 420, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 444, + 420, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 444, + 420, + 456 + ], + "type": "text", + "content": "4.2.1. Experiment Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "text", + "content": "Our robotic platform consists of a single Franka Research3 arm equipped with a third-person-view RealSense D435 camera mounted in a fixed position to capture environmental observations. The system operates at " + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "inline_equation", + "content": "20\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "text", + "content": " (moderately reduced from the native " + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "inline_equation", + "content": "100\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 460, + 554, + 555 + ], + "type": "text", + "content": " control frequency to balance training efficiency and motion continuity), with actions defined as absolute end-effector poses in SE(3) space (3D position + quaternion orientation)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 616 + ], + "type": "text", + "content": "Our experimental benchmark comprises six manipulation tasks (4 short-horizon tasks, 2 long-horizon tasks) designed to evaluate the model's ability to handle varying task complexities. For each task, we collect 50 demonstrations and evaluate performance over 20 trials:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 640 + ], + "type": "text", + "content": "1) Pull out a tissue paper: The robot need to grasp and pull out a single tissue paper." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 688 + ], + "type": "text", + "content": "2) Pick up the [TOY NAME]: The robot is required to pick up the specific toy, including the toy snake, the toy eggplant, and the toy chicken, resulting in a total of three tasks (with no other distractions)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "3) Put the toy into the basket: The robot need to pick up the toy(no other distractions) and put it into the basket." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11093" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 135, + 174 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 135, + 174 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 135, + 174 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 135, + 174 + ], + "type": "image", + "image_path": "efd59b2c1b974854cfccaa8958f9dbb0acb7d1164c031ebbdd600212dcb12473.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 175, + 105, + 181 + ], + "lines": [ + { + "bbox": [ + 78, + 175, + 105, + 181 + ], + "spans": [ + { + "bbox": [ + 78, + 175, + 105, + 181 + ], + "type": "text", + "content": "Libero 90" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 72, + 214, + 174 + ], + "blocks": [ + { + "bbox": [ + 137, + 72, + 214, + 174 + ], + "lines": [ + { + "bbox": [ + 137, + 72, + 214, + 174 + ], + "spans": [ + { + "bbox": [ + 137, + 72, + 214, + 174 + ], + "type": "image", + "image_path": "5941664259900e8d7070eeac30f5a567a288309c1c4327f96c1c7aff3428dfe7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 72, + 293, + 174 + ], + "blocks": [ + { + "bbox": [ + 216, + 72, + 293, + 174 + ], + "lines": [ + { + "bbox": [ + 216, + 72, + 293, + 174 + ], + "spans": [ + { + "bbox": [ + 216, + 72, + 293, + 174 + ], + "type": "image", + "image_path": "26aae533fdde15bdac7d45a8aaf00d7b39292b4631a79c61c5fbe30fe4167b89.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 174, + 291, + 181 + ], + "lines": [ + { + "bbox": [ + 217, + 174, + 291, + 181 + ], + "spans": [ + { + "bbox": [ + 217, + 174, + 291, + 181 + ], + "type": "text", + "content": "Put the toy into the basket" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 182, + 135, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 182, + 135, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 182, + 135, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 135, + 285 + ], + "type": "image", + "image_path": "b0e83fe6cd6f42da81a249074e813a289863301b33d508c2e042c663f69746fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 286, + 135, + 294 + ], + "lines": [ + { + "bbox": [ + 59, + 286, + 135, + 294 + ], + "spans": [ + { + "bbox": [ + 59, + 286, + 135, + 294 + ], + "type": "text", + "content": "Put all cups into the basket" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 182, + 214, + 285 + ], + "blocks": [ + { + "bbox": [ + 141, + 175, + 203, + 182 + ], + "lines": [ + { + "bbox": [ + 141, + 175, + 203, + 182 + ], + "spans": [ + { + "bbox": [ + 141, + 175, + 203, + 182 + ], + "type": "text", + "content": "Pull out a tissue paper" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 182, + 214, + 285 + ], + "lines": [ + { + "bbox": [ + 137, + 182, + 214, + 285 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 214, + 285 + ], + "type": "image", + "image_path": "c06351d4aff8518f0767b7dfe7064bf1e5db23b6bb41dd6f8122df474d269994.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 286, + 213, + 294 + ], + "lines": [ + { + "bbox": [ + 137, + 286, + 213, + 294 + ], + "spans": [ + { + "bbox": [ + 137, + 286, + 213, + 294 + ], + "type": "text", + "content": "Put the toy into the drawer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 217, + 182, + 294, + 285 + ], + "blocks": [ + { + "bbox": [ + 217, + 182, + 294, + 285 + ], + "lines": [ + { + "bbox": [ + 217, + 182, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 217, + 182, + 294, + 285 + ], + "type": "image", + "image_path": "03969043ed6fbc0a6e2fcf0d629a6b27626f552c8bfa7fb1c17d5b4e9d65e72b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 286, + 281, + 294 + ], + "lines": [ + { + "bbox": [ + 225, + 286, + 281, + 294 + ], + "spans": [ + { + "bbox": [ + 225, + 286, + 281, + 294 + ], + "type": "text", + "content": "Filp the pot upright" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 296, + 135, + 399 + ], + "blocks": [ + { + "bbox": [ + 58, + 296, + 135, + 399 + ], + "lines": [ + { + "bbox": [ + 58, + 296, + 135, + 399 + ], + "spans": [ + { + "bbox": [ + 58, + 296, + 135, + 399 + ], + "type": "image", + "image_path": "1f8f58540c7da9cb63591941aedebeb95bfb1cd47247b090f2213a4060749113.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 419, + 296, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 475 + ], + "type": "text", + "content": "Figure 2. All Evaluation environments: We conduct comprehensive evaluations of VQ-VLA in both simulation and real-world settings. In simulation, evaluations are performed on the LIBERO-90 benchmark within the LIBERO dataset. And six diverse tasks are designed for real-world testing." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 137, + 296, + 214, + 399 + ], + "blocks": [ + { + "bbox": [ + 137, + 296, + 214, + 399 + ], + "lines": [ + { + "bbox": [ + 137, + 296, + 214, + 399 + ], + "spans": [ + { + "bbox": [ + 137, + 296, + 214, + 399 + ], + "type": "image", + "image_path": "1f09c0454d1293fb66575d628a2460b19e6946fdf52081c170ff7310c53f8171.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 403, + 203, + 411 + ], + "lines": [ + { + "bbox": [ + 141, + 403, + 203, + 411 + ], + "spans": [ + { + "bbox": [ + 141, + 403, + 203, + 411 + ], + "type": "text", + "content": "Pick up the{toy name}" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 217, + 296, + 294, + 399 + ], + "blocks": [ + { + "bbox": [ + 217, + 296, + 294, + 399 + ], + "lines": [ + { + "bbox": [ + 217, + 296, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 217, + 296, + 294, + 399 + ], + "type": "image", + "image_path": "1f2b73ab429478a6d76663c1e41113736533ac73b5d281a6707ee9d05d2ad6ed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 511, + 296, + 670 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 55, + 511, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 548 + ], + "type": "text", + "content": "4) Flip the pot upright: We set a flipped pot on the platform, the robot need to flip and upright a fallen cooking pot." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 555, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 602 + ], + "type": "text", + "content": "5) Put all cups into the basket: We place two different cups on the table and set a few other things (toys) as distractions. The robot need to sequentially put two cups into the basket, testing long-horizon task." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": "6) Put the toy into the drawer: We place a toy on the table and set a few other things as distractions. The robot need to sequentially open the drawer, pick up the specific toy and put it into the drawer, and close the drawer, testing long-horizon task." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "We finetune each task separately for 100K gradient steps (batch size 4 across 4 A100-80GB GPUs) with action chunk length " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 5" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": ". During deployment." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 72, + 501, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 501, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 501, + 83 + ], + "type": "text", + "content": "4.2.2. Performance on Short-Horizon Tasks" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 88, + 555, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 88, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 88, + 555, + 195 + ], + "type": "text", + "content": "Based on the in the simulation experiments, we further incorporated 120k synthetic trajectories from the ManiSkill dataset, co-training " + }, + { + "bbox": [ + 313, + 88, + 555, + 195 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L + M}}" + }, + { + "bbox": [ + 313, + 88, + 555, + 195 + ], + "type": "text", + "content": " with Open X-Embodiment dataset and LIBERO dataset to evaluate the scaling capability of our action tokenizer. We evaluated the performance of the action tokenizer trained with different datasets on four carefully designed short-horizon tasks, focusing on multi-task learning, precision manipulation, and dynamic adjustment capabilities. The results are shown in Fig. 3." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "text", + "content": "It can be observed that when synthetic trajectories are added for co-training, the performance of the action tokenizer improves significantly. The average success rate increased from the baseline of " + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "inline_equation", + "content": "46.25\\%" + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "text", + "content": ", with a notable improvement of " + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "text", + "content": " in the \"Flip the pot upright\" task. In the \"Pull out a tissue paper\" task, which tests the robot's performance in high-precision dynamic operations (as this task requires continuous, fine-grained grasping and pulling motions), the baseline model achieved only a " + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "text", + "content": " success rate, failing almost entirely. In contrast, models incorporating VQ achieved success rates of " + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 197, + 556, + 328 + ], + "type": "text", + "content": " or higher." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "text", + "content": "Furthermore, comparing " + }, + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L}}" + }, + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_O}" + }, + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "text", + "content": ", the average success rate on short-horizon tasks increased by only " + }, + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 329, + 555, + 413 + ], + "type": "text", + "content": ". This may be attributed to the limited size of the LIBERO dataset, which has minimal impact on short-horizon tasks. However, the ManiSkill dataset is 50 times larger than LIBERO, leading to a significant improvement in success rates when used for training." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 422, + 500, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 500, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 500, + 434 + ], + "type": "text", + "content": "4.2.3. Performance on Long-Horizon Tasks" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "text", + "content": "VQ-VLA demonstrates outstanding performance on long-horizon tasks (\"Put all cups in the basket\" and \"Put the toy into the drawer\"), significantly outperforming baseline model in both success rate and efficiency. In scenarios where baseline models achieve success rates as low as " + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "text", + "content": " or nearly 0, the " + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ}_{O + L + M}" + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "text", + "content": " model achieves significantly higher success rates of " + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "text", + "content": ", respectively (see Fig.3). For the \"Put the toy into the drawer\" task, a representative long-horizon scenario, the baseline model was only able to complete the first step of opening the drawer in 15 out of 20 trials, failing to proceed further in most cases. In contrast, the " + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ}_{O + L + M}" + }, + { + "bbox": [ + 313, + 437, + 556, + 605 + ], + "type": "text", + "content": " model successfully opened the drawer in all test cases, demonstrating its robustness and reliability in handling complex sequential tasks." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "A key advantage of VQ-VLA in long-horizon tasks lies in its use of VQ-VAE as an action tokenizer, which enables the model to predict multiple actions in a single inference step. This design reduces the accumulation of errors over extended task sequences, making VQ-VLA particularly advantageous for tasks requiring long-term planning and execution. Additionally, in scenarios involving the consecutive execution of multiple subtasks, VQ-VLA not only achieves higher success rates but also significantly reduces task completion time" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11094" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 266 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 266 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 266 + ], + "type": "image", + "image_path": "5ad737e1f2527b4c3e74f5e553ca2dcb533b8594001f7a084d70e9c845c2144c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": "Figure 3. Real-world experimental results: We compare the performance of Baseline, " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_0}" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L + M}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": " on both short-horizon and long-horizon tasks. In terms of the average success rate, all VQ-based models outperform the Baseline. The best-performing model, " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L + M}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": ", achieves a success rate that is " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "23.25\\%" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": " higher than the Baseline on both short-horizon and long-horizon tasks. Additionally, the results show that " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L + M}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": ", which in turn outperforms " + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_0}" + }, + { + "bbox": [ + 54, + 276, + 555, + 333 + ], + "type": "text", + "content": ", indicating the effectiveness of incorporating synthetic data during training without compromising real-world performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 352, + 297, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 297, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 297, + 389 + ], + "type": "text", + "content": "compared to baseline models, highlighting its superior execution efficiency and ability to handle long-horizon challenges effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 405, + 224, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 224, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 224, + 417 + ], + "type": "text", + "content": "4.2.4. Sim&Real Domain Gap Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "content": "To quantify the domain gap between synthetic data and real-world data, we trained a VQ-VAE model exclusively on the LIBERO dataset, referred to as " + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{VQL}" + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "content": ", for OpenVLA. The model was tested in three real-world tasks (one long-horizon and two short-horizon tasks), and the results are shown in Tab. 3. The results indicate that the performance of " + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{VQL}" + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "content": " is comparable to that of both " + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L}}" + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_O}" + }, + { + "bbox": [ + 54, + 423, + 295, + 531 + ], + "type": "text", + "content": ", suggesting that the domain gap between synthetic and real-world data is minimal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 535, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 535, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 535, + 296, + 594 + ], + "type": "text", + "content": "Although real-world data may contain noise, the inclusion of Open X-Embodiment data as a real-world dataset expands the data sources and enriches the diversity of data types, which effectively enhances the model's generalization and robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 611, + 208, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 208, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 208, + 624 + ], + "type": "text", + "content": "4.2.5. Inference Speed Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "content": "During the real-world experiments, we measured the action execution frequency of VQ-VLA and compared it with the original OpenVLA. The results are summarized in Tab. 4. As shown in the table, with a compression ratio of 5 in VQ-VAE, the inference speed is nearly tripled. This significant improvement greatly facilitates real-time performance in practical applications." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 350, + 555, + 426 + ], + "blocks": [ + { + "bbox": [ + 317, + 350, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 317, + 350, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 555, + 426 + ], + "type": "table", + "html": "
Put the toy into the drawer (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline5.030.020.0
VQo15.045.035.0
VQL10.055.035.0
VQo+L10.045.035.0
VQo+L+M25.060.045.0
", + "image_path": "76364bdbb51a172a76961994967d0ef83048bf58db8c1cb313cd7b8df6aa96d0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 332, + 491, + 537, + 534 + ], + "blocks": [ + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "type": "text", + "content": "Table 3. Performance Comparison Across Real-World Tasks: We observe that the performance of " + }, + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_L}" + }, + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "type": "text", + "content": " is comparable to that of both " + }, + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_{O + L}}" + }, + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "type": "inline_equation", + "content": "\\mathrm{VQ_O}" + }, + { + "bbox": [ + 313, + 434, + 555, + 478 + ], + "type": "text", + "content": ", indicating that the domain gap between synthetic and real-world data is minimal." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 491, + 537, + 534 + ], + "lines": [ + { + "bbox": [ + 332, + 491, + 537, + 534 + ], + "spans": [ + { + "bbox": [ + 332, + 491, + 537, + 534 + ], + "type": "table", + "html": "
Frequency (Hz)
VQ-VLA11.84
OpenVLA4.16
", + "image_path": "4fb2ce6c4ce5ef1a3df98d925f31c1cbaecf7d2d3b66caf2e3bfe2a4c8ee39ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 542, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 564 + ], + "type": "text", + "content": "Table 4. The Results of Frequencies. We report the comparison results of our VQ-VLA and baseline OpenVLA." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 587, + 416, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 416, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 416, + 598 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 629 + ], + "type": "text", + "content": "In this section, we report some ablation studies to show the effectiveness of the design choices of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 638, + 554, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 554, + 662 + ], + "type": "text", + "content": "4.3.1. Action Chunking via VQ-VAE and Autoregressive Output" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "The original OpenVLA model generates a single action per inference step based solely on the current observation, resulting in a step-by-step execution approach. To improve efficiency and handle longer action sequences, we extend" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11095" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "OpenVLA to output a sequence of five actions in an autoregressive manner, similar to the action chunking mechanism in VQ-VLA. Specifically, the model predicts the next action based on the current observation and previously generated actions, allowing it to generate a sequence of actions in a single inference step. This autoregressive approach serves as an alternative action chunking strategy by grouping multiple actions for execution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 168, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 168, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 296, + 395 + ], + "type": "text", + "content": "To evaluate the effectiveness of different action chunking strategies, we design ablation experiments comparing the autoregressive output of OpenVLA to the VQ-based action chunking method in VQ-VLA. The results, shown in Tab. 5, indicate that the use of the autoregressive approach for action chunking in the original OpenVLA leads to a significant drop in success rate compared to the baseline. In real-world experiments, when using the autoregressive approach as a form of action chunking, the spatial magnitude of each action chunk is relatively small, resulting in slower overall execution of actions. Compared to the Action Chunking via VQ-VAE method, even with the same action chunk size, the time required to reach the same target location is significantly longer. Additionally, this approach tends to exhibit shortcut learning, where the model directly copies previous actions within the same action chunk. By analyzing the output values of each action chunk, it is observed that multiple actions within a single chunk have remarkably similar values, indicating a lack of diversity in the predicted actions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 455 + ], + "type": "text", + "content": "Overall, Action Chunking via VQ-VAE not only improves inference speed but also enhances the performance of VLA by generating more effective and diverse action sequences. This makes it a more suitable approach for addressing real-world long-horizon tasks." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 463, + 294, + 513 + ], + "blocks": [ + { + "bbox": [ + 57, + 463, + 294, + 513 + ], + "lines": [ + { + "bbox": [ + 57, + 463, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 463, + 294, + 513 + ], + "type": "table", + "html": "
Action ChunkingLIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
baseline74.7630.020.0
Autoregressive Output66.5310.00.0
VQ-based (VQo+L+M)86.6160.045.0
", + "image_path": "51f6d73250ab2dfeb8d794b7c474feaf11aedf6588f1b0c456982a1a131997d6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 603, + 241, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 241, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 241, + 614 + ], + "type": "text", + "content": "4.3.2. Embedding Integration Effectiveness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": "To evaluate the impact of embeddings, we conducted an ablation study comparing the model's performance with and without embeddings. Specifically, the baseline model processes raw action sequences directly, while the enhanced model incorporates both time embedding and action-type embedding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "As shown in Tab. 6, the model with embeddings significantly outperforms the baseline in terms of success rate. This" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "demonstrates that the integration of embeddings improves the encoder's ability to represent structured action sequences, leading to better overall performance." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 122, + 555, + 170 + ], + "blocks": [ + { + "bbox": [ + 55, + 521, + 295, + 587 + ], + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 587 + ], + "type": "text", + "content": "Table 5. Comparison of action chunking methods: We evaluate the performance on three tasks (one simulator task and two real-world tasks). The results show that the Autoregressive Output used in OpenVLA performs poorly as an action chunking method, with a significant gap compared to the VQ-based approach. This indicates that the VQ-based method is more effective." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 122, + 555, + 170 + ], + "lines": [ + { + "bbox": [ + 318, + 122, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 318, + 122, + 555, + 170 + ], + "type": "table", + "html": "
LIBERO-90 (%)Flip the pot upright (%)Put the toy into the basket (%)
VQoL (w.o. Embeddings)85.1740.035.0
VQoL (w Embeddings)86.1645.035.0
", + "image_path": "69c0d9b4034a0371af9acdd07f2902b29fd26b48e9b49e92581b13ac0f0898c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 178, + 555, + 223 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 223 + ], + "type": "text", + "content": "Table 6. Embedding integration improves performance. The table compares models with and without embeddings across three tasks, showing that embeddings enhance success rates, especially for \"Flip the pot upright.\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 255, + 486, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 486, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 486, + 266 + ], + "type": "text", + "content": "5. Limitations and Future Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 277, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 556, + 456 + ], + "type": "text", + "content": "In our work, a series of extensive experiments demonstrate that our proposed action tokenizers can improve VLA performance and inference speed while achieving scalability on simulated data. Despite these promising results, there still remain some limitations and opportunities for future work. First, our action tokenizers can be further extended to larger-scale simulated datasets. Second, our work improves inference speed by decoding multi-step action sequences, which can be further combined with techniques such as distillation [38] and quantization [16] of VLMs in the future. Finally, the architecture design of our action tokenizers can be further improved, for example, by incorporating the frequency of action data as a extra condition. In summary, we hope that our work can serve as a strong baseline and inspire future research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 473, + 391, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 391, + 485 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 391, + 485 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 495, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 639 + ], + "type": "text", + "content": "In this paper, we propose a general convolutional residual VQ-VAE framework for action tokenizers that can seamlessly integrate with state-of-the-art VLA models. Our VQ-VAE is trained on 100 times more data than previous methods and can be directly transferred to downstream both real-world and simulated robotic manipulation tasks. Extensive experiments conducted in both simulated and real-world environments validate that the proposed convolutional residual VQ-VAE framework not only enhances the performance of VLA policies but also accelerates inference. Finally, our VQ-VAE also demonstrates the ability to scale effectively with large-scale simulated data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 655, + 411, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 655, + 411, + 668 + ], + "spans": [ + { + "bbox": [ + 314, + 655, + 411, + 668 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11096" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 258 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeffrey Donahue, Pauline Luc, Alexis Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: A visual language model for few-shot learning. Advances in Neural Information Processing Systems (NeurIPS), 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 260, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 296, + 316 + ], + "type": "text", + "content": "[4] Arthur Allshire, Roberto Martin-Martin, Charles Lin, Shawn Manuel, Silvio Savarese, and Animesh Garg. Laser: Learning a latent action space for efficient reinforcement learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6650-6656, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 349 + ], + "type": "text", + "content": "[5] Suneel Belkhale and Dorsa Sadigh. Minivla: A better vla with a smaller footprint. https://ai.stanford.edu/blog/minivla/, 2024. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "type": "text", + "content": "[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. " + }, + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\pi_0" + }, + { + "bbox": [ + 62, + 352, + 296, + 407 + ], + "type": "text", + "content": ": A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 463 + ], + "type": "text", + "content": "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "type": "text", + "content": "[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 576 + ], + "type": "text", + "content": "[9] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 578, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 622 + ], + "type": "text", + "content": "[10] Wei Chen, Yifeng Zhang, Hao Li, Xufeng Wang, and Ming Liu. Anyvlm: Unified vision-language model for any robot morphology. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 1-8. IEEE, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "text", + "content": "[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[12] Xiaoyu Chen, Junliang Guo, Tianyu He, Chuheng Zhang, Pushi Zhang, Derek Cathera Yang, Li Zhao, and Jiang Bian. Igor: Image-goal representations are the atomic control" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 74, + 554, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 554, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 554, + 94 + ], + "type": "text", + "content": "units for foundation models in embodied ai. arXiv preprint arXiv:2411.00785, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 555, + 149 + ], + "type": "text", + "content": "[13] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 185 + ], + "type": "text", + "content": "[14] Norman Di Palo and Edward Johns. Keypoint action tokens enable in-context imitation learning in robotics. arXiv preprint arXiv:2403.19578, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "type": "text", + "content": "[15] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 253, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 555, + 295 + ], + "type": "text", + "content": "[16] Kazuki Egashira, Mark Vero, Robin Staab, Jingxuan He, and Martin Vechev. Exploiting lIm quantization. Advances in Neural Information Processing Systems, 37:41709-41732, 2025. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 298, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 555, + 342 + ], + "type": "text", + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 554, + 387 + ], + "type": "text", + "content": "[18] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 554, + 409 + ], + "type": "text", + "content": "[19] Philip Gage. A new algorithm for data compression. The C Users Journal, 12(2):23-38, 1994. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 411, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 554, + 454 + ], + "type": "text", + "content": "[20] Theophile Gervet, Zhou Xian, Nikolaos Gkanatsios, and Katerina Fragkiadaki. Act3d: 3d feature field transformers for multi-task robotic manipulation. In 7th Annual Conference on Robot Learning. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 499 + ], + "type": "text", + "content": "[21] Abhishek Gupta, Vikash Kumar, Corey Lynch, Sergey Levine, and Karol Hausman. Relay policy learning: Solving long-horizon tasks via imitation and reinforcement learning. arXiv preprint arXiv:1910.11956, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 501, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 554, + 533 + ], + "type": "text", + "content": "[22] Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 554, + 578 + ], + "type": "text", + "content": "[23] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 580, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 554, + 622 + ], + "type": "text", + "content": "[24] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 624, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 677 + ], + "type": "text", + "content": "[25] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024.3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 554, + 713 + ], + "type": "text", + "content": "[26] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11097" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[27] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "text", + "content": "[28] Seungjae Lee, Yibin Wang, Haritheja Etukuru, H Jin Kim, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Behavior generation with latent actions. arXiv preprint arXiv:2403.03181, 2024. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[29] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 263 + ], + "type": "text", + "content": "[30] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 308 + ], + "type": "text", + "content": "[31] Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1:3, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 354 + ], + "type": "text", + "content": "[32] Atharva Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, and Animesh Garg. Quest: Self-supervised skill abstractions for learning continuous control. Advances in Neural Information Processing Systems, 37:4062-4089, 2025. 2, 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "type": "text", + "content": "[33] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 1, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 487 + ], + "type": "text", + "content": "[34] Abby O'Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "type": "text", + "content": "[35] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. arXiv preprint arXiv:2501.09747, 2025. 1, 2, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "text", + "content": "[36] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "text", + "content": "[37] Fengyuan Shi, Zhuoyan Luo, Yixiao Ge, Yujiu Yang, Ying Shan, and Limin Wang. Taming scalable visual tokenizer for autoregressive image generation. arXiv preprint arXiv:2412.02692, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 688 + ], + "type": "text", + "content": "[38] Sharath Turuvekere Sreenivas, Saurav Muralidharan, Raviraj Joshi, Marcin Chochowski, Ameya Sunil Mahabaleshwarkar, Gerald Shen, Jiaqi Zeng, Zijia Chen, Yoshi Suhara, Shizhe Diao, et al. Llm pruning and distillation in practice: The minitron approach. arXiv preprint arXiv:2408.11796, 2024. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[39] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024.1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[40] Andrew Szot, Bogdan Mazoure, Harsh Agrawal, R Devon Hjelm, Zsolt Kira, and Alexander Toshev. Grounding multimodal large language models in actions. Advances in Neural Information Processing Systems, 37:20198-20224, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 185 + ], + "type": "text", + "content": "[41] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "type": "text", + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 251 + ], + "type": "text", + "content": "[43] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 554, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 286 + ], + "type": "text", + "content": "[44] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 554, + 342 + ], + "type": "text", + "content": "[45] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 344, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 554, + 388 + ], + "type": "text", + "content": "[46] Jange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. arXiv preprint arXiv:2306.05716, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 389, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 554, + 430 + ], + "type": "text", + "content": "[47] Jange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 434, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 477 + ], + "type": "text", + "content": "[48] Jiangge Yang, Haoyi Zhu, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Tra-moe: Learning trajectory prediction model from multiple domains for adaptive policy conditioning. arXiv preprint arXiv:2411.14519, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 479, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 554, + 533 + ], + "type": "text", + "content": "[49] Jiange Yang, Yansong Shi, Haoyi Zhu, Mingyu Liu, Kaijing Ma, Yating Wang, Gangshan Wu, Tong He, and Limin Wang. Como: Learning continuous latent motion from internet videos for scalable robot learning. arXiv preprint arXiv:2505.17006, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "text", + "content": "[50] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 622 + ], + "type": "text", + "content": "[51] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "text", + "content": "[52] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[53] Canyu Zhao, Mingyu Liu, Wen Wang, Weihua Chen, Fan Wang, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11098" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 430 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[54] Canyu Zhao, Mingyu Liu, Huanyi Zheng, Muzhi Zhu, Zhiyue Zhao, Hao Chen, Tong He, and Chunhua Shen. Dioxide: A generalist diffusion model for visual perceptual tasks. arXiv preprint arXiv:2502.17157, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[55] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "type": "text", + "content": "[56] Wei Zhao, Pengxiang Ding, Min Zhang, Zhefei Gong, Shuanghao Bai, Han Zhao, and Donglin Wang. Vlas: Vision-language-action model with speech instructions for customized robot manipulation. arXiv preprint arXiv:2502.13508, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 251 + ], + "type": "text", + "content": "[57] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 296 + ], + "type": "text", + "content": "[58] Jinliang Zheng, Jianxiong Li, Dongxiu Liu, Yinan Zheng, Zhihao Wang, Zhonghong Ou, Yu Liu, Jingjing Liu, Ya-Qin Zhang, and Xianyuan Zhan. Universal actions for enhanced embodied foundation models, 2025. 2, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 328 + ], + "type": "text", + "content": "[59] Wenxuan Zhou, Sujay Bajracharya, and David Held. Plas: Latent action space for offline reinforcement learning, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 384 + ], + "type": "text", + "content": "[60] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "type": "text", + "content": "[61] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11099" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_content_list.json b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab79e0ec6037fdadfd93dcc67a3448147775a805 --- /dev/null +++ b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_content_list.json @@ -0,0 +1,1410 @@ +[ + { + "type": "text", + "text": "VRBench: A Benchmark for Multi-Step Reasoning in Long Narrative Videos", + "text_level": 1, + "bbox": [ + 111, + 130, + 885, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiashuo Yu $^{1*}$ Yue Wu $^{1*}$ Meng Chu $^{1*}$ Zhifei Ren $^{1*}$ Zizheng Huang $^{2,3,1*}$ Pei Chu $^{1*}$", + "bbox": [ + 132, + 178, + 867, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruijie Zhang $^{1}$ Yinan He $^{1}$ Qirui Li $^{1}$ Songze Li $^{1}$ Zhenxiang Li $^{1}$ Zhongying Tu $^{1}$", + "bbox": [ + 155, + 199, + 864, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conghui He1 Yu Qiao1 Yali Wang4,1⊗ Yi Wang1,3⊗ Limin Wang2,1⊗", + "bbox": [ + 197, + 215, + 820, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ Nanjing University $^{3}$ Shanghai Innovation Institute", + "bbox": [ + 102, + 233, + 893, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences", + "bbox": [ + 197, + 251, + 802, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VRBench.github.io", + "bbox": [ + 395, + 271, + 599, + 286 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b75860174e48a322566b434e2eea6279429859803f92425fc8b9f779f5ad7b1.jpg", + "image_caption": [ + "Figure 1. Overview of VRBench. We present VRBench, a long narrative video benchmark for multi-step reasoning. VRBench includes 960 manual-filtered narrative videos, covering 8 languages and 7 video categories that are suitable for reasoning about temporal relations. We also provide high-quality stepwise annotations for reasoning, which are labeled and reviewed by human experts. Each video incorporates 8-10 complex question-answer pairs, a multi-step reasoning chain, and fine-grained timestamps. To fully evaluate the capability of models in multi-step reasoning, we propose a multi-phase evaluation pipeline that assesses model results both from the process and outcome levels. Our VRBench is the first video reasoning benchmark that supports both multi-step annotation and evaluation." + ], + "image_footnote": [], + "bbox": [ + 114, + 321, + 883, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 666, + 326, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present VRBench, the first long narrative video benchmark crafted for evaluating large models' multi-step reasoning capabilities, addressing limitations in existing evaluations that overlook temporal reasoning and procedural validity. It comprises 960 long videos (with an average duration of 1.6 hours), along with 8,243 human-labeled multi-step question-answering pairs and 25,106 reasoning steps with timestamps. These videos are curated via a multistage filtering process including expert inter-rater reviewing to prioritize plot coherence. We develop a human-AI collaborative framework that generates coherent reasoning chains, each requiring multiple temporally grounded steps,", + "bbox": [ + 89, + 696, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spanning seven types (e.g., event attribution, implicit inference). VRBench designs a multi-phase evaluation pipeline that assesses models at both the outcome and process levels. Apart from the MCQs for the final results, we propose a progress-level LLM-guided scoring metric to evaluate the quality of the reasoning chain from multiple dimensions comprehensively. Through extensive evaluations of 12 LLMs and 19 VLMs on VRBench, we undertake a thorough analysis and provide valuable insights that advance the field of multi-step reasoning.", + "bbox": [ + 511, + 667, + 906, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 844, + 645, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid evolution of vision language models (VLMs) has heightened the need for benchmarks that rigorously eval", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contributions. corresponding authors.", + "bbox": [ + 114, + 886, + 362, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21655", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "uate complex reasoning capabilities. While existing standards like GSM8K [8] focus on domain-specific knowledge in mathematics and science, they neglect a critical reasoning dimension: contextual analysis in visual narrative content. Real-world applications increasingly demand temporal reasoning across interconnected elements: tracking character dynamics in films, interpreting gameplay strategies, or understanding cause-and-effect chains in documentaries. This capability gap persists because current video benchmarks [16, 53, 70, 77] primarily assess single-step perception rather than sustained reasoning processes.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We find three fundamental limitations in current evaluation paradigms: (1) an overemphasis on domain expertise rather than plot-driven reasoning, (2) the lack of temporally-grounded reasoning chains in annotations, and (3) Outcome-focused metrics that ignore procedural validity. To address these challenges, we present VRBench (as in Figure 1), the first benchmark specifically designed for multi-step reasoning in long-form narrative videos. Its construction involves the collection of narrative videos, a human-in-the-loop reasoning process method for annotation, and a multi-phase evaluation pipeline considering both procedure and outcome.", + "bbox": [ + 89, + 260, + 483, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In data construction, VRBench aggregates 960 meticulously selected videos (1.6h average duration) across seven narrative categories (e.g., movies, sports, travelogues), sourced through a multi-stage filtering process. Our curation pipeline combines automated retrieval with expert validation (inter-rater reliability $\\rho = 0.82$ ), prioritizing plot coherence over domain-specific knowledge. This contrasts with existing benchmarks like MMVU [92] that emphasize disciplinary competence. To conduct multi-step reasoning chain annotation, we develop a human-AI collaborative approach, generating 8-10 QA pairs per video with explicit temporal grounding. Each question requires no less than 2 reasoning steps annotated with precise timestamps (Figure 3), validated through iterative expert review (95% inter-annotator agreement). The taxonomy spans seven reasoning types from event prediction to implicit inference, ensuring comprehensive coverage of narrative analysis skills.", + "bbox": [ + 89, + 444, + 483, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When evaluating models on VRBench, we propose a multi-phase evaluation pipeline that combines outcome verification with process analysis. Beyond conventional multiple-choice accuracy that evaluates the outcome-level performance, we further introduce the process-level metric LLM-guided scoring to evaluate the overall reasoning process quality. This dual approach reveals critical insights - for instance, while GPT-4o achieves $81.25\\%$ outcome accuracy, its process rating scores drop to $56.11\\%$ , exposing reasoning fragility. Evaluations of 31 state-of-the-art large models reveal proprietary VLMs with long-context support outperform text-only LLMs by $13.82\\%$ absolute, emphasizing the importance of dense visual ground", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f51bb77bf0fc108a977640f5339c2cf941c34fb2de8defdcae4c944c9f6911bf.jpg", + "image_caption": [ + "Figure 2. An example of annotation in VRBench. For each question, VRBench provides the question-answer pair, multi-step reasoning chain, question type, and the start-to-end timestamps of the entire question as well as each reasoning step." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 895, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Questions: Why did the male protagonist arrange for his friend to rob female protagonist? [00:02:20->00:17:42]", + "text_level": 1, + "bbox": [ + 532, + 137, + 883, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step1: When the heroine first caught the hero and his friends, she acted very \nrighteously, and the hero fell in love with her. [00:02:20->00:05:05] \nStep2: The female protagonist chats with the male protagonist and reveals that her \nfavorite type is a manly man. [00:05:06->00:07:58]", + "bbox": [ + 532, + 167, + 883, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step3: The male protagonist asks his friends to plan how to show manliness and justice. [00:13:23~>00:16:01]", + "bbox": [ + 532, + 207, + 883, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step4: The male protagonist arranges his friends to rob the female protagonist in an attempt to win her love. [00:16:09->00:17:42]", + "bbox": [ + 532, + 224, + 867, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Answer: Because he wants to get the woman's love.", + "text_level": 1, + "bbox": [ + 532, + 251, + 795, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Question Type: Event Attribution", + "text_level": 1, + "bbox": [ + 532, + 270, + 707, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing. System-2 optimization strategies yield disproportionate gains in process metrics, and despite parameter parity, open-source VLMs lag behind proprietary counterparts by $12.30\\%$ , suggesting architectural limitations beyond scale. Our ablation studies show strong alignment between human and LLM evaluations and quantify the impact of test-time scaling—doubling context windows improves QwQ-32B's [65] accuracy by $12.43\\%$ .", + "bbox": [ + 511, + 378, + 906, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VRBench sets a new standard for evaluating narrative reasoning, offering insights distinct from knowledge-centric benchmarks. By separating domain expertise from contextual analysis, our work aids in developing models capable of sustained, human-like understanding in real-world video content. The entire suite, including annotation protocols and evaluation tools, has been fully open-sourced to advance reasoning research.", + "bbox": [ + 511, + 500, + 908, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 633, + 656, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reasoning Large Models. As the capabilities of LLMs continue to expand, frontier models have demonstrated significant potential in addressing high-order reasoning tasks. OpenAI pioneered this advancement with the release of o1 [48], an LLM capable of sophisticated reasoning tasks. Subsequently, a series of proprietary [12, 49, 59] and open-source [3, 18, 65] LLMs designed for advanced reasoning have emerged. Furthermore, researchers have shifted their focus towards developing VLMs possessing similar higher-order reasoning capabilities. Previously, some VLMs [2, 7, 30, 32, 37, 71, 73, 88, 91] to a certain extent exhibited good reasoning proficiency due to their proficiency in tackling long-sequence context. More recently, a series of VLMs [14, 66, 69, 81] trained on data incorporating higher-order Chains-of-Thought in RL-based algorithms [33, 52, 55] have been developed, enabling them", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21656", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/6fd631ce4c824d5abdc0cb42c11fbe35ab3a9fe71e79a205384c69826ad97922.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Size#QA pairs#Dur.(s)QA typesData SourceAnno StepEval StepEval TargetAnno TypeClueMultilingual
MMLU [21]14,07914,079/MCQMulti-DiscSingleSingleLLMMxx
MMLU-Pro [72]12,03212,032/MCQMulti-DiscSingleSingleLLMA+Mxx
LiveCodeBench [26]511511/MCQCodeMultiSingleLLMMxx
SciEval [57]15,90115,901/MCQ+OpenScienceSingleSingleLLMMxx
GSM8k [8]8,7928,792/MCQ+OpenMathMultiMultiLLMMxx
C-Eval [23]12,34212,342/MCQMulti-DiscSingleSingleLLMMxx
ScienceQA [54]10,33221,208/MCQScienceMultiSingleVLMMxx
VisScience [27]3,0003,000/MCQ+OpenScienceSingleSingleVLMA+Mxx
MMMU [85]11,50011,500/MCQ+OpenMulti-DiscSingleSingleVLMMxx
MMMU-Pro [86]3,4603,460/MCQMulti-DiscSingleSingleVLMA+Mxx
MathVista [42]6,1416,141/MCQ+OpenMathMultiSingleVLMMxx
MathVision [67]3,0403,040/MCQ+OpenMathMultiSingleVLMMxx
CharXiv [74]2,32311,615/OpenMulti-DiscSingleSingleVLMMxx
OlympicArena [25]7,57111,163/MCQ+OpenMulti-DiscMultiMultiVLMMxx
MVBench [31]3,6414,00016.0MCQOpen-DomainSingleSingleVLMA+Mxx
EgoSchema [43]5,0635,063180.0MCQEgocentricSingleSingleVLMAxx
LongVideoBench [77]3,7636,678473.0MCQOpen-DomainSingleSingleVLMMxx
LVBench [70]1031,5494,101MCQOpen-DomainSingleSingleVLMMxx
CGBench [4]1,21912,1291624.4MCQ+OpenOpen-DomainMultiMultiVLMM✓x
MLVU [93]1,7303,102930MCQNarrativeSingleSingleVLMMxx
VideoMME [16]9002,7001017.9MCQOpen-DomainSingleSingleVLMMx✓
Video-MMMU [22]300900506.2MCQMulti-DiscSingleSingleVLMMxx
MMWorld [20]1,9106,627108.0MCQMulti-DiscSingleSingleVLMA+Mxx
MMVU [92]1,5293,00051.4MCQ+OpenMulti-DiscMultiSingleVLMMxx
VRBench (Ours)9608,2435,796.0MCQ+OpenNarrativeMultiMultiLLM, VLMM✓✓
", + "bbox": [ + 91, + 88, + 906, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison between VRBench and existing benchmarks. #Size is the number of text/images/videos, #Dur. means the average video duration, A and M indicate automatic and manual annotation type, respectively, and multi-disc denotes multi-disciplinary data source. Clue means clue-grounded annotation. Multilingual requires the number of data source languages to be greater than 2.", + "bbox": [ + 89, + 429, + 906, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to solve multidisciplinary, knowledge-intensive problems through a combination of fast and slow thinking processes.", + "bbox": [ + 89, + 498, + 480, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reasoning Benchmarks. With the rapid advancement of LLMs and VLMs in reasoning capability, numerous reasoning benchmarks have been developed in text and image-text modalities to facilitate comprehensive evaluation. For the textual domain, several knowledge-driven benchmarks [5, 15, 21, 23, 57, 72] have been derived from exam and textbook data sources, and comprise expert-level questions that require multi-step reasoning. Meanwhile, some multimodal benchmarks also emerge to introduce multi-discipline tasks based on charts [44, 74, 78], plots [45], exam [10, 25, 36, 85, 89], or expert-level questions printed on the static images [27, 35, 54, 86]. It is noted that even in the image and text domains, benchmarks with process-level annotations [8, 87] are relatively scarce. For video understanding, evaluation also gradually shifts from short video clip perception [6, 9, 28, 29, 31, 39, 46, 50, 80] to long-form long video understanding [13, 16, 40, 43, 53, 56, 58, 70, 77, 90] and single-step reasoning. [4, 16, 19, 34, 76, 84]. More recently, some advances [13, 20, 92] propose several new video reasoning tasks in multi-disciplinary scenarios like healthcare, engineering, and science. Different from previous works, VRBench is the first narrative video-based benchmark that is purely used for multi-step reasoning eval", + "bbox": [ + 91, + 553, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "uation. Table 1 further presents the detailed statistics of our dataset and distinguishes the difference between VRBench and existing text, image, and video reasoning benchmarks.", + "bbox": [ + 511, + 498, + 906, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Benchmark", + "text_level": 1, + "bbox": [ + 511, + 556, + 635, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present VRBench, a comprehensive multi-step reasoning benchmark consisting of a collection of long, multilingual, and narrative videos with corresponding question-answer pairs. Figure 1 gives the overall pipeline of the benchmark construction, and we illustrate it as follows.", + "bbox": [ + 511, + 580, + 905, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Video Curation", + "text_level": 1, + "bbox": [ + 511, + 664, + 668, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We collect long narrative videos from YouTube, yielding an initial pool of over 10,000 public videos. Considering plausible reasoning in steps requires videos with rich and coherent plots, we source long-form footage using a cherry-picked tag set and a comprehensive criterion for narration.", + "bbox": [ + 511, + 686, + 905, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Plot-related Tags for Queries. We employ a manually curated set of 7 semantic tags (e.g., Film & Animation, Sport, Travel & Events) to retrieve videos with strong narrative potential. This tag set is developed through iterative validation by domain experts, explicitly excluding non-narrative categories such as news broadcasts and lecture recordings. Our analysis revealed that these excluded categories exhibit limited visual-semantic dynamics (e.g., static", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21657", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01a1f4f86ae78d2f1f2b5f4ca2438e27ce9a62d975e3b857e4f83723525d0e9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 90, + 339, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe11c9a4213c0bbaa61a589de2054c54a19a7b6d7af1151be4d9c7276eb11ff3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 90, + 571, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/629ab1b1b4239bb663c922988fda19806527b64f756876e3d711adb4689a9af4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 90, + 856, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ddad08f6901c1c40f9553f14165c6c9dfd94deb9e73630dda394e2e47e0dd585.jpg", + "image_caption": [ + "Figure 3. Statistics of VRBench. We provide the detailed distribution of videos and annotations of VRBench, including video languages and durations, steps, types, and temporal duration of questions, as well as the token numbers of answers and the reasoning process." + ], + "image_footnote": [], + "bbox": [ + 133, + 252, + 344, + 406 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/046cc163db7f6f9b589e856b6de421827f5ee2a6f603faa658e8776382ed5ee8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 252, + 589, + 406 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c402527a1c58f4e9e7f8a5f7ee87f241137acf7dd2bf6c2dae32c420c1bf697.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 253, + 854, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "camera angles in talk formats) and minimal event progression, rendering them ineffective for benchmarking temporal reasoning capabilities.", + "bbox": [ + 89, + 473, + 482, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Criterion for Narration. Our sourcing standards include video duration, its language source, and user ratings. The duration requirement (a minimum of 20 minutes and an average of 1.61 hours) ensures adequate temporal context for constructing reasoning chains. Our language diversity strategy intentionally excludes English and Chinese to counterbalance existing dataset biases, aligning with established findings that linguistic variety improves benchmark generalizability.", + "bbox": [ + 88, + 537, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Upon the aforementioned two quantitative prerequisites, we organized a panel of 14 multilingual domain experts to evaluate candidate videos using a standardized 10-point scale based on plot coherence and richness. The scoring is guided by several instructions. Videos scoring below 7 are systematically excluded, resulting in a final curated set of 960 high-quality narratives. Detailed annotation protocols are documented in the supplementary materials.", + "bbox": [ + 89, + 672, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stepwise Reasoning Annotations", + "text_level": 1, + "bbox": [ + 89, + 803, + 377, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We label videos with reasoning steps via a two-stage human-in-the-loop framework. It first generates pseudo candidate QA pairs through automated pipelines, then employs expert-guided rewriting to curate high-quality annotations, ensuring rigorous adherence to benchmark speci", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fications while maintaining multimodal reasoning fidelity. To further improve annotation quality, we also implement a comprehensive quality assurance.", + "bbox": [ + 511, + 473, + 906, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Automatic Pipelines. We first employ AutoShot [94] to cut videos into several segments, and then use VideoChat2 [31] to caption them. For auditory content, we adopt whisper-large-v3 [51] to obtain speech transcripts, and DeepL [11] to translate them into English. The video captions and translated subtitles are then put into GPT-4o [47] to generate 6 pseudo QA pairs with a multi-step reasoning process. To ensure the quality of reasoning, we specify 7 multi-step reasoning types for narrative videos as:", + "bbox": [ + 511, + 547, + 906, + 684 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Event prediction: Forecast subsequent events in the video timeline.", + "- Hypothetical reasoning: Deduce plausible scenario outcomes from stated premises.", + "- Event attribution: Determine causal origins or underlying motivations of video events.", + "- Implicit inference: Extract unstated temporal, emotional, or relational context from visual cues.", + "- Logical linkage: Establish event-mediated connections between visual/narrative elements.", + "- Information synopsis: Condense critical information across multimodal inputs.", + "- Counting problems: Quantify state changes through arithmetic/combinatorial analysis." + ], + "bbox": [ + 513, + 686, + 903, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21658", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Expert-guided Rewriting. We recruit and train 67 graduate students to generate 8-10 high-quality QA pairs per video, providing raw video footage, translated subtitles, and GPT-generated pre-annotations that offer contextual hints without meeting final benchmark standards. Each QA pair must satisfy four rigorous criteria: (1) Non-synopsis questions require $\\geq 2$ timestamped reasoning steps (start/end times documented); (2) Temporal distribution constraints ( $\\leq 4$ questions from $0 - 15\\mathrm{min}$ , $\\geq 3$ from $15 - 40\\mathrm{min}$ , $\\geq 1$ from $40 - 120\\mathrm{min}$ ); (3) Coverage of $\\geq 5$ predefined reasoning taxonomies (from 7 categories) ensuring diversity; (4) Mandatory multimodal grounding where solutions demand both visual analysis (excluding subtitle-only answers) and explicit reasoning (beyond basic perception).", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comprehensive Quality Assurance. To ensure annotations meet our stringent multimodal standards, we implement a rigorous verification protocol: 10 trained reviewers validate each annotation, with non-compliant entries returned to annotators for iterative revisions until fully compliant. Annotators and reviewers are exclusively recruited from top-tier universities to ensure academic rigor. We further enforce quality through dual safeguards: a systematic $5\\%$ random sampling audit across both annotation and review stages, coupled with full documentation of protocols, annotation guidelines, and quality assessment criteria in supplementary materials.", + "bbox": [ + 89, + 332, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multi-Phase Evaluation Pipeline", + "text_level": 1, + "bbox": [ + 89, + 529, + 375, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We benchmark VLMs and LLMs through a comprehensive multi-phase evaluation pipeline, which compares predictions against ground-truth annotations both at the process and outcome level. For the outcome-level evaluation stage, we adopt a multiple-choice question (MCQ) format, where false options are generated by DeepSeek-V3 [38] using carefully-designed prompts based on human-annotated answers. For the process-level stage, we propose the open-ended rating to fully evaluate the model's multi-step reasoning capability. Specifically, we adopt an LLM to evaluate the overall quality of the whole reasoning process through four 0-10 scores: logical coherence $(40\\%)$ , similarity to ground truth $(40\\%)$ , excluded for event prediction and hypothetical reasoning tasks), factual accuracy $(10\\%)$ , and clarity $(10\\%)$ . DeepSeek-V3 [38] serves as the judge due to its optimal balance between cost and human alignment (Section 4). For the event prediction and hypothetical reasoning tasks, we argue that there is no ground-truth reasoning steps since there might be multiple possible predictions, hence, we remove the similarity score and compute the final score with an 8:1:1 weight. All metrics are normalized to 0-100 scales, and we report the rankings by computing average scores across all metrics.", + "bbox": [ + 89, + 553, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4f6f89f22f1d9c212edacb9064aad1049e97b6d604835cc51b939f3aee0f2745.jpg", + "image_caption": [ + "Figure 4. Human preference alignment results. For each plot, we show the win ratios of three different tested LLMs evaluated by human experts and VRBench. We then fit them with a straight line and quantify the correlation by calculating the Spearman correlation coefficient." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 901, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLM Evaluation Support. For text-only LLMs, we convert videos to text inputs using Qwen2.5-72B-Instruct [83] to synthesize video captions and subtitles into fine-grained summaries. The process involves: 1) Dividing inputs into 5-minute chunks for duplicate removal and abstraction; 2) Merging chunk-level abstracts into coherent summaries with detailed audio-visual descriptions. These summaries enable LLM evaluations while addressing context window limits for VLMs.", + "bbox": [ + 511, + 349, + 906, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 497, + 645, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We test a number of both open-source and proprietary VLMs (e.g. Qwen2-VL [68], InternVL2.5 [7], InternVideo2.5 [73], GPT-4o [47], Gemini-2.0-Pro [61], etc.) and LLMs (DeepSeek-R1 [18], QwQ [65], OpenAI o1 [48], Claude-3.7-Sonnet [59], etc.) with a two-stage evaluation on VRBench. The configuration of each evaluated model is detailed in the supplementary material.", + "bbox": [ + 511, + 523, + 906, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocols. Our protocol involves two phases. Given a question, 1) models produce multi-step reasoning using Chain-of-Thought [75] prompts. Then 2) models choose final answers from multiple-choice options. For answer extraction, final answers are parsed from response endings to mitigate verbose option analysis, following [31]. While reasoning process evaluations use judge LLMs with structured prompts and example-guided rating extraction. All evaluation prompts and implementation details are included in the supplementary materials.", + "bbox": [ + 511, + 628, + 908, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 787, + 651, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 shows the CoT evaluation results of LLMs and VLMs on VRBench. Among LLMs, we found the proprietary model Gemini-2.0-Flash-Thinking [12] achieves the optimal performance while other reasoning models such as OpenAI o1-preview [48] and DeepSeek-R1 [18] also demonstrate strong results. It is noted that o1-preview [48]", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21659", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ae52fd76e13c844e68131919f9882f6caedda5c662fb6645364a1fa374943712.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelOverallResults by MetricResults by Taxonomy
MCQ-OOE-PEACPHRIIISEPLL
LLMs
Proprietary Models
GPT-4o [47]55.4963.8747.1154.4132.6369.0361.4771.8366.3766.19
Claude-3.7-Sonnet [59]58.0762.9153.2355.8735.1169.8363.2972.6167.9167.89
o1-preview [48]60.1468.4751.8156.9735.8770.5164.8173.1368.2769.41
Gemini-2.0-Flash-Thinking [61]60.9767.5354.4158.6138.1172.0967.5774.1970.4371.13
Open-Source Models
QwQ-32B-preview [64]35.9027.5144.2934.3134.4146.2939.6727.9742.3940.61
InternLM3-8B-Instruct [3]47.8150.3145.3145.4734.8760.3351.7755.4756.9756.81
Qwen2.5-7B-Instruct [83]48.2952.6143.9746.9335.1761.8351.5754.8756.6357.03
Llama3.3-70B-Instruct [41]49.8452.5947.0947.6338.0763.8754.0354.8359.4756.97
QwQ-32B [65]52.5256.0149.0349.5336.0362.0154.0352.0358.0159.03
Qwen2.5-72B-Instruct [83]53.5160.4946.5351.0336.0167.0358.5367.8361.9363.03
DeepSeek-V3 [38]56.0664.7947.3354.0336.5769.9761.8369.5365.5365.43
DeepSeek-R1 [18]57.1364.1950.0756.1338.9172.8164.1368.0168.5367.93
VLMs
Proprietary Models
Claude-3.7-Sonnet [59]68.1580.0956.2163.1132.9772.6371.1375.3771.3370.27
GPT-4o [47]68.6881.2356.1366.6136.5176.6770.4778.0372.0373.17
Gemini-2.0-Pro [61]74.6183.2965.9371.0965.2181.0175.7387.1377.9375.89
Open-Source Models
DeepSeek-VL2 [79]31.5033.2729.7325.9322.4135.7331.7330.0133.2929.57
H2OVL Mississippi-2B [17]47.1552.3341.9735.1732.3751.7141.4160.0349.9741.91
Phi-3.5-Vision [1]48.5258.0339.0131.5328.0345.0337.0368.0344.0336.03
LongVA-7B [91]50.1467.8132.4727.6125.0738.6933.7776.9138.2732.07
InternVL2.5-8B [7]50.4169.3131.5126.6326.3137.9931.2185.9737.3730.69
MiMo-VL-7B-RL [60]50.4863.3937.5747.3134.5963.8353.2374.6160.3356.97
VideoChat-Flash-7B [32]50.8272.0129.6324.6921.4139.3729.1779.3734.8728.07
InternVideo2.5 [73]51.9475.6328.2523.8124.0934.5127.0784.5733.6126.81
LongVA-7B-DPO [91]52.3667.9136.8130.7327.1745.0938.2379.3943.6737.03
Qwen2-VL-7B [68]54.0872.0136.1530.4926.6146.7133.6985.2743.3135.29
Aria [30]54.5572.9736.1330.1929.2344.2334.6386.2344.9934.57
Qwen2.5-VL-7B [2]56.5269.6143.4337.0733.1754.2742.2783.9150.1742.87
Keye-VL-8B-Preview [63]60.4464.4156.4762.5340.3773.6167.9970.1371.8371.03
Qwen2.5-VL-72B [2]61.7166.8556.5751.8746.1367.1354.1390.0463.6760.77
Kimi-VL-A3B-Thinking-2506 [62]61.8261.6761.9764.5347.9171.3769.5774.4771.2371.47
InternVL2.5-78B [7]62.3176.6148.0143.7738.6758.2146.1387.5354.4347.57
", + "bbox": [ + 96, + 87, + 906, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Evaluation results on VRBench across two evaluation metrics (outcome-level MCQ, process-level open-ended evaluation) and seven QA taxonomies (event attribution, counting problems, hypothetical reasoning, implicit inferences, information synopsis, event prediction, logical linkage). Bold values indicate the best results among all models, and underlined values are best results of LLMs.", + "bbox": [ + 89, + 715, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "achieves the best outcome-level MCQ accuracy of $68.47\\%$ . For the VLMs, Gemini-2.0-Pro [61] achieves $74.61\\%$ overall accuracy, surpassing all components by at least $5.93\\%$ . GPT-4o [47] and Claude-3.7-Sonnet [59] also show competitive results, which demonstrate $68.68\\%$ and $68.15\\%$ overall performance, respectively. For open-source models, InternVL2.5-78B [7], emerges as the best non-proprietary", + "bbox": [ + 89, + 784, + 480, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VLMs with a $62.31\\%$ overall accuracy.", + "bbox": [ + 513, + 784, + 772, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then delve into the specific results of each metric and taxonomy. For results in each evaluation stage, we find that most of LLMs and VLMs are capable of achieving high MCQ accuracy, yet compared to LLMs, VLMs struggle to demonstrate their reasoning steps and exhibit lower reasoning ratings. Through results across taxonomy, we found", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21660", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that most of the large models are not proficient in counting problems since they require fine-grained visual perception, in contrast with other tasks. Since the perception of LLMs fully depends on the video summary and some VLMs only support a small number of frame inputs, they are unable to correctly perceive the original frames containing the target elements, resulting in the MCQ accuracy close to random guessing and low reasoning process ratings.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablations and Analysis", + "text_level": 1, + "bbox": [ + 89, + 224, + 303, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then further discuss the observations from the evaluation results and highlight some insights on obtaining higher multi-step reasoning performance.", + "bbox": [ + 89, + 247, + 482, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does the evaluation of VRBench align with Human Preference? In contrast to MCQs that compute accuracy via a deterministic method, the correctness of process-level open-ended evaluation highly depends on the judgment of LLM's performance, which may include hallucinations that could influence the evaluation reliability. To this end, we select a subset of 30 videos with 300 questions to perform process-level human evaluation, aiming to quantify the correlation between human and LLM assessment results.", + "bbox": [ + 89, + 292, + 482, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, the human annotators are asked to give process-level open-ended ratings following the same requirement for the judging LLM, i.e., four separated ratings are needed for each question to evaluate the logical coherence, similarity to ground-truth, factual accuracy, and expression clarity. We then compute the win ratio of pairwise comparison for each tested model following [24], where a model scores 1 if its rating is higher than its current opponent, 0 if lower, and 0.5 for a tie. Finally, we calculate the average win ratio of each model and the Spearman correlation coefficients $(\\rho)$ of the win ratios evaluated by LLMs and human annotators.", + "bbox": [ + 89, + 430, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first investigate the human preference alignment of several LLMs. We select GPT-4o [47], DeepSeek-V3 [38], Qwen2.5-7B [83], and Qwen2.5-72B [83] as evaluation models and give them the same rating prompt for a fair comparison. As shown in Figure 4, Qwen2.5 shows low human preference alignment compared with DeepSeek-V3 and GPT-4o, which both have correlation coefficients greater than 0.8. Though the human alignment of GPT-4o is slightly better than DeepSeek-V3, we observe the evaluation cost of GPT-4o is approximately 10 times that of DeepSeek, which imposes a significant burden on the benchmark users. Hence, we adopt DeepSeek-V3 as the judging model for its optimal trade-off between performance and cost.", + "bbox": [ + 89, + 612, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also probe the evaluation correctness of each open-ended sub-metric. Following the same procedure, we compute the win ratio and correlation coefficient of DeepSeekV3 for each sub-metric, and the results are shown in Figure 4 (b-d). Results show the coefficients in all metrics are", + "bbox": [ + 89, + 825, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "higher than 0.8, indicating a strong positive monotonic relationship between LLM and human evaluation results. Another alternative is to instruct the LLM to compute the final score based on the given weights of each sub-metric, which is denoted as DeepSeek w/o sep in Figure 4. Results show that adding this additional computing process diminishes the robustness of evaluation LLM, thus in practice, we ask the model to output all sub-metric rating separately and automatically compute the final score.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The role of System-2 thinking. o1-like LLMs, also denoted as system-2 models, have emerged with superior reasoning abilities in multi-disciplinary scenarios such as science and math. We further explore these models' performance on the narrative video benchmark. We both assess the role of open-source and proprietary o1-models, as well as the gap compared with their previous system-1 version. For proprietaries, we compare OpenAI o1-preview [48] with GPT-4o [47], and results show that the system-2 version o1 outperforms GPT-4o with $4.30\\%$ overall accuracy. For open-source model DeepSeek-V3 [38] and R1 [11], we observe that DeepSeek-R1 achieves higher scores both on multiple-choice questions and open-ended evaluations. It is noted that compared with Qwen2.5-72B-Instruct [82], QwQ-32B [65] gets higher reasoning ratings yet lower outcome-level MCQ results. This shows that though some system-2 models are more proficient in multi-step thinking, their ability to converge lengthy thinking processes into correct answers still needs improvement. Recently, some o1-like training strategies have also been utilized by some VLMs, such as LongVA-7B-DPO [91], a model that utilizes direct preference optimization [52] for long video understanding. It achieves an overall performance of $52.36\\%$ and surpasses its vanilla version, suggesting that training VLMs with different optimization methods is feasible and that developing system-2 VLM models is crucial for tackling multi-step reasoning questions in narrative videos.", + "bbox": [ + 511, + 231, + 903, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Model Size. It is commonly believed that models with large scales are often accompanied by better perception and reasoning capabilities. To verify such claims, we conduct experiments on the models with different scales, such as Qwen2.5-7B [83] and Qwen2.5-72B, Qwen2.5-VL [2] with 7B and 72B parameters, and InternVL2.5-8B [7] and InternVL2.5-78B. As shown in Table 2, models with more parameters obtain higher overall accuracy (48.29% vs 53.51% for Qwen2.5, 56.52% vs 61.71% for Qwen2.5-VL, and 50.41% vs 62.31% for InternVL2.5), and the performance gap on process-level metrics is more apparent. This indicates that large-scale models are more likely to achieve advanced reasoning capabilities. However, we also noticed that models specifically trained on the reasoning corpus like QwQ-32B [65], can achieve comparable performance compared with its 72B Qwen2.5 counterpart (52.52% vs. 53.51% in overall accuracy) with a smaller", + "bbox": [ + 511, + 643, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21661", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "parameter amount. This suggests that training small-sized models with various preference optimization strategies and reasoning-centric corpora might improve their overall reasoning capability.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Long Context Helps. We observe a substantial performance gap between models with and without long context support. As shown in Table 2, models with long frame input tend to achieve higher overall accuracy. For example, Gemini-2.0-Pro [61] is capable of inference with 0.5 fps and supports large number frame inputs when answering questions of videos over 1 hour, and it achieves the optimal performance compared with all other fixed-length models. For the open-source models, Qwen2.5-VL-7B [2] and KimiVL-A3B-Thinking-2506 [62] with 256 frame input achieve $61.71\\%$ and $61.82\\%$ overall accuracy, which is $11.30\\%$ and $11.41\\%$ higher than the 8B InternVL-2.5 [7] model with 64 frames input. Since the average length of VRBench is 1.61 hours, the necessity of long context support is becoming more pronounced, so that models can perceive more plot elements that help the analysis of long video content.", + "bbox": [ + 89, + 152, + 482, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LLMs vs. VLMs. Since tested LLMs are only provided with the automatically-generated video summary, yet VLMs are capable of directly perceiving the raw visual content, it is unfair to directly compare the numerical results of LLMs and VLMs. However, there are still some noteworthy phenomena that can be discussed. First, VLMs with coarse-grained visual input (e.g., 4 frames for H2OVL Mississippi-2B [17] and DeepSeek-VL2 [79]) perform significantly worse than several system-2 LLMs like DeepSeek-R1 [18] and Claude-3.7-Sonnet [59]. This indicates that randomly selecting a small number of frames is not enough to tackle narrative-based reasoning tasks that require long-range perception. Furthermore, VLMs with fine-grained visual perception, such as Gemini-2.0-Pro with a 0.5 fps perform better than top-tier LLMs on both process-level and outcome-level metrics, showing the inevitability of detailed visual content input for tackling questions in VRBench.", + "bbox": [ + 89, + 396, + 482, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Test-Time Scaling Exploration", + "text_level": 1, + "bbox": [ + 89, + 665, + 362, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We here probe models' performance when scaling test-time compute cost. Since we tend to prompt the models with a Chain-of-Thought template when tackling complex reasoning tasks, it is observed that system-2 models are capable of dramatically expanding the inference budget to improve their performance. To this end, we set a series of token limitations on these models and require the models to think with different instructions. Specifically, we select QwQ-32B [65] as the experimental LLM and Qwen2-VL-7B [82] as the testing VLM on a subset with 300 videos and 2,403 questions. For each model, several parallel experiments are conducted with increasing maximum token limits from 256 to 2048. We instruct the models to output the reasoning process as much as possible for large token limitations, and", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f214471ff20dc28023bbddcd0fd23f4621a7b12fa38cd7c031b608e764a1971.jpg", + "image_caption": [ + "Figure 5. Test-Time Scaling Results. We report the average accuracy of outcome-level MCQ and process-level open-ended ratings." + ], + "image_footnote": [], + "bbox": [ + 527, + 90, + 890, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "let them think concisely when the token limit is low. Since we aim to investigate the quality of models' initial reasoning process and their capability to generate the right answers through thinking, we report the overall scores that both evaluate the outcome accuracy and process quality. Results are shown in Figure 5, where QwQ shows a remarkable rating boost from $48.91\\%$ to $61.34\\%$ when setting a large token limit. On the contrary, the small-sized system-1 model Qwen2-VL-7B performs even worse when using a large token number limitation and instructing the model to output more thinking process. The model tends to output more ambiguous outputs that lead to the wrong answer. This observation suggests that models with large parameter scales and system-2 capabilities benefit from the test-time scaling strategy. It also provides the insight that developing more test-time scaling approaches that guide the model to generate long reasoning traces could be a feasible and promising way to unlock more potential capabilities of system-2 models, thereby benefiting complex tasks that require multi-step reasoning.", + "bbox": [ + 511, + 343, + 906, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 685, + 633, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces VRBench, a comprehensive long-narrative video benchmark for evaluating multi-step reasoning. Through manually filtered narrative videos, high-quality stepwise annotations, and a multi-phase evaluation pipeline, VRBench distinguishes itself from other existing reasoning benchmarks and shows the robust capability to evaluate LLMs and VLMs both from the process and outcome perspectives. By assessing and analyzing 31 frontier large models, we thoroughly demonstrate the detailed performance of current reasoning models across various reasoning questions and provide valuable insights towards constructing more advanced multi-step reasoning models.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21662", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement.", + "text_level": 1, + "bbox": [ + 89, + 90, + 274, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 202, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6", + "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7, 8", + "[3] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2, 6", + "[4] Guo Chen, Yicheng Liu, Yifei Huang, Yuping He, Baoqi Pei, Jilan Xu, Yali Wang, Tong Lu, and Limin Wang. Cg-bench: Clue-grounded question answering benchmark for long video understanding. arXiv preprint arXiv:2412.12075, 2024.3", + "[5] Wenhu Chen, Ming Yin, Max Ku, Pan Lu, Yixin Wan, Xueguang Ma, Jianyu Xu, Xinyi Wang, and Tony Xia. Theoremq: A theorem-driven question answering dataset. arXiv preprint arXiv:2305.12524, 2023. 3", + "[6] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. In European Conference on Computer Vision, pages 179-195. Springer, 2024. 3", + "[7] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang-wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 5, 6, 7, 8", + "[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 3", + "[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees GM Snoek, and Yuki M Asano. Tvbench: Redesigning video-language evaluation. arXiv preprint arXiv:2410.07752, 2024.3", + "[10] Rocktim Jyoti Das, Simeon Emilov Hristov, Haonan Li, Dimitar Iliyanov Dimitrov, Ivan Koychev, and Preslav Nakov. *Exams-v: A multi-discipline multilingual multimodal exam benchmark for evaluating vision language models.* arXiv preprint arXiv:2403.10378, 2024. 3" + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work is supported by the National Key R&D Program of China (No. 2022ZD0160101) and Jiangsu Frontier Technology Research and Development Program (No. BF2024076),", + "[11] DeepL. Deepl translate: The world's most accurate translator. https://www deepl.com/en/Translator, 2025.4,7", + "[12] Google Deepmind. Gemini 2.0 flash thinking. https://deepmind.google/technologies/gemini/flash-thinking/, 2025. Accessed: 2025-01-21. 2, 5", + "[13] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 3", + "[14] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 2", + "[15] Kehua Feng, Keyan Ding, Weijie Wang, Xiang Zhuang, Zeyuan Wang, Ming Qin, Yu Zhao, Jianhua Yao, Qiang Zhang, and Huajun Chen. Sciknoweval: Evaluating multilevel scientific knowledge of large language models. arXiv preprint arXiv:2406.09098, 2024. 3", + "[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3", + "[17] Shaikat Galib, Shanshan Wang, Guanshuo Xu, Pascal Pfeiffer, Ryan Chesler, Mark Landry, and Sri Satish Ambati. H2ovl-mississippi vision language models technical report. arXiv preprint arXiv:2410.13611, 2024. 6, 8", + "[18] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2, 5, 6, 8", + "[19] Songhao Han, Wei Huang, Hairong Shi, Le Zhuo, Xiu Su, Shifeng Zhang, Xu Zhou, Xiaojuan Qi, Yue Liao, and Si Liu. Videospeech: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection. arXiv preprint arXiv:2411.14794, 2024. 3", + "[20] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multi-discipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3", + "[21] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 3", + "[22] Kairui Hu, Penghao Wu, Fanyi Pu, Wang Xiao, Yuanhan Zhang, Xiang Yue, Bo Li, and Ziwei Liu. Video-mmmu: Evaluating knowledge acquisition from multi-discipline professional videos. arXiv preprint arXiv:2501.13826, 2025. 3", + "[23] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems, 36:62991-63010, 2023. 3" + ], + "bbox": [ + 89, + 90, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21663", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7", + "[25] Zhen Huang, Zengzhi Wang, Shijie Xia, Xuefeng Li, Haoyang Zou, Ruijie Xu, Run-Ze Fan, Lyumanshan Ye, Ethan Chern, Yixin Ye, et al. Olympic: Benchmarking multi-discipline cognitive reasoning for superintelligent ai. Advances in Neural Information Processing Systems, 37: 19209-19253, 2025. 3", + "[26] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livocodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024. 3", + "[27] Zhihuan Jiang, Zhen Yang, Jinhao Chen, Zhengxiao Du, Weihan Wang, Bin Xu, Yuxiao Dong, and Jie Tang. Visscience: An extensive benchmark for evaluating k12 educational multi-modal scientific reasoning. arXiv preprint arXiv:2409.13730, 2024. 3", + "[28] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 3", + "[29] Bozheng Li, Yongliang Wu, Yi Lu, Jiashuo Yu, Licheng Tang, Jiawang Cao, Wenqing Zhu, Yuyang Sun, Jay Wu, and Wenbo Zhu. Veu-bench: Towards comprehensive understanding of video editing. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13671-13680, 2025. 3", + "[30] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 2, 6", + "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 4, 5", + "[32] Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, et al. Videochat-flash: Hierarchical compression for long-context video modeling. arXiv preprint arXiv:2501.00574, 2024. 2, 6", + "[33] Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning. arXiv preprint arXiv:2504.06958, 2025. 2", + "[34] Yunxin Li, Xinyu Chen, Baotian Hu, Longyue Wang, Haoyuan Shi, and Min Zhang. Videovista: A versatile benchmark for video understanding and reasoning. arXiv preprint arXiv:2406.11303, 2024. 3", + "[35] Zekun Li, Xianjun Yang, Kyuri Choi, Wanrong Zhu, Ryan Hsieh, HyeonJung Kim, Jin Hyuk Lim, Sungyoung Ji," + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Byungju Lee, Xifeng Yan, et al. Mmsci: A multimodal multi-discipline dataset for phd-level scientific comprehension. In AI for Accelerated Materials Design-Vienna 2024, 2024.3", + "[36] Zhenwen Liang, Kehan Guo, Gang Liu, Taicheng Guo, Yujun Zhou, Tianyu Yang, Jiajun Jiao, Renjie Pi, Jipeng Zhang, and Xiangliang Zhang. Scemqa: A scientific college entrance level multimodal question answering benchmark. arXiv preprint arXiv:2402.05138, 2024. 3", + "[37] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 2", + "[38] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 5, 6, 7", + "[39] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 3", + "[40] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. Et bench: Towards open-ended event-level video-language understanding. arXiv preprint arXiv:2409.18111, 2024. 3", + "[41] Llama-3.3. Llama-3.3-70b-instruct. https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct, 2025.6", + "[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 3", + "[43] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 3", + "[44] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3", + "[45] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 3", + "[46] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3", + "[47] OpenAI. Hello gpt4-o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-13, 4, 5, 6, 7", + "[48] OpenAI. Introducing openai o1. https://openai.com/o1/, 2024.2,5,6,7", + "[49] OpenAI. Openai o3-mini. https://openai.com/index/openai-o3-mini/, 2025.2" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21664", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36:42748-42761, 2023. 3", + "[51] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 4", + "[52] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 7", + "[53] Ruchit Rawal, Khalid Saifullah, Miquel Farre, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. arXiv preprint arXiv:2405.08813, 2024. 2, 3", + "[54] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3", + "[55] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2", + "[56] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 3", + "[57] Liangtai Sun, Yang Han, Zihan Zhao, Da Ma, Zhennan Shen, Baocai Chen, Lu Chen, and Kai Yu. Scieval: A multi-level large language model evaluation benchmark for scientific research. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19053-19061, 2024. 3", + "[58] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 3", + "[59] Anthropic. Claude Team. Claude 3.7 sonnet. https://www.anthropic.com/claude/sonnet, 2025. 2, 5, 6, 8", + "[60] Core Team, Zihao Yue, Zhenru Lin, Yifan Song, Weikun Wang, Shuhuai Ren, Shuhao Gu, Shicheng Li, Peidian Li, Liang Zhao, Lei Li, et al. Mimo-vl technical report. https://arxiv.org/abs/2506.03569, 2025. 6", + "[61] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 6, 8", + "[62] Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chen-" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "zhuang Du, Chu Wei, et al. Kimi-vl technical report. arXiv preprint arXiv:2504.07491, 2025. 6, 8", + "[63] Kwai Keye Team, Biao Yang, Bin Wen, Changyi Liu, Chenglong Chu, Chengru Song, Chongling Rao, Chuan Yi, Da Li, Dunju Zang, et al. Kwai keye-vl technical report. arXiv preprint arXiv:2507.01949, 2025. 6", + "[64] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview/, 2024. Accessed: 2024-11-28. 6", + "[65] Qwen Team. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github.io/blog/qwq-32b/, 2025. Accessed: 2025-3-6. 2, 5, 6, 7, 8", + "[66] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamavol: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025. 2", + "[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 3", + "[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 6", + "[69] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, et al. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024. 2", + "[70] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 2, 3", + "[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 2", + "[72] Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 3", + "[73] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 5, 6", + "[74] Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, et al. Charxiv: Charting gaps in realistic chart understanding in multimodal llms. Advances in Neural Information Processing Systems, 37:113569-113697, 2025. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21665", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5", + "[76] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 3", + "[77] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 3", + "[78] Yifan Wu, Lutao Yan, Leixian Shen, Yunhai Wang, Nan Tang, and Yuyu Luo. Chartinsights: Evaluating multimodal large language models for low-level chart question answering. arXiv preprint arXiv:2405.07001, 2024. 3", + "[79] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 6, 8", + "[80] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3", + "[81] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-01: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 2", + "[82] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report, 2024. URL: https://arxiv.org/abs/2407.10671, 2024. 7, 8", + "[83] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5, 6, 7", + "[84] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. Advances in Neural Information Processing Systems, 37:57240-57261, 2025. 3", + "[85] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3", + "[86] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3", + "[87] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 2", + "[89] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, et al. Cmmmu: A chinese massive multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2401.11944, 2024. 3", + "[90] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023.3", + "[91] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2, 6, 7", + "[92] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 2, 3", + "[93] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 3", + "[94] Wentao Zhu, Yufang Huang, Xiufeng Xie, Wenxian Liu, Jincan Deng, Debing Zhang, Zhangyang Wang, and Ji Liu. Autoshot: A short video dataset and state-of-the-art shot boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2238-2247, 2023. 4" + ], + "bbox": [ + 516, + 92, + 903, + 597 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21666", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_model.json b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d75fc654e404bae73c3529c4288e5c76a39ecb7 --- /dev/null +++ b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_model.json @@ -0,0 +1,2457 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.131, + 0.887, + 0.154 + ], + "angle": 0, + "content": "VRBench: A Benchmark for Multi-Step Reasoning in Long Narrative Videos" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.179, + 0.869, + 0.2 + ], + "angle": 0, + "content": "Jiashuo Yu\\(^{1*}\\) Yue Wu\\(^{1*}\\) Meng Chu\\(^{1*}\\) Zhifei Ren\\(^{1*}\\) Zizheng Huang\\(^{2,3,1*}\\) Pei Chu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.2, + 0.865, + 0.218 + ], + "angle": 0, + "content": "Ruijie Zhang\\(^{1}\\) Yinan He\\(^{1}\\) Qirui Li\\(^{1}\\) Songze Li\\(^{1}\\) Zhenxiang Li\\(^{1}\\) Zhongying Tu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.217, + 0.821, + 0.235 + ], + "angle": 0, + "content": "Conghui He1 Yu Qiao1 Yali Wang4,1⊗ Yi Wang1,3⊗ Limin Wang2,1⊗" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.234, + 0.895, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Artificial Intelligence Laboratory \\(^{2}\\)Nanjing University \\(^{3}\\)Shanghai Innovation Institute" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.252, + 0.803, + 0.27 + ], + "angle": 0, + "content": "4Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.272, + 0.6, + 0.287 + ], + "angle": 0, + "content": "VRBench.github.io" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.322, + 0.885, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.57, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Figure 1. Overview of VRBench. We present VRBench, a long narrative video benchmark for multi-step reasoning. VRBench includes 960 manual-filtered narrative videos, covering 8 languages and 7 video categories that are suitable for reasoning about temporal relations. We also provide high-quality stepwise annotations for reasoning, which are labeled and reviewed by human experts. Each video incorporates 8-10 complex question-answer pairs, a multi-step reasoning chain, and fine-grained timestamps. To fully evaluate the capability of models in multi-step reasoning, we propose a multi-phase evaluation pipeline that assesses model results both from the process and outcome levels. Our VRBench is the first video reasoning benchmark that supports both multi-step annotation and evaluation." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.667, + 0.327, + 0.683 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.879 + ], + "angle": 0, + "content": "We present VRBench, the first long narrative video benchmark crafted for evaluating large models' multi-step reasoning capabilities, addressing limitations in existing evaluations that overlook temporal reasoning and procedural validity. It comprises 960 long videos (with an average duration of 1.6 hours), along with 8,243 human-labeled multi-step question-answering pairs and 25,106 reasoning steps with timestamps. These videos are curated via a multistage filtering process including expert inter-rater reviewing to prioritize plot coherence. We develop a human-AI collaborative framework that generates coherent reasoning chains, each requiring multiple temporally grounded steps," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.82 + ], + "angle": 0, + "content": "spanning seven types (e.g., event attribution, implicit inference). VRBench designs a multi-phase evaluation pipeline that assesses models at both the outcome and process levels. Apart from the MCQs for the final results, we propose a progress-level LLM-guided scoring metric to evaluate the quality of the reasoning chain from multiple dimensions comprehensively. Through extensive evaluations of 12 LLMs and 19 VLMs on VRBench, we undertake a thorough analysis and provide valuable insights that advance the field of multi-step reasoning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.646, + 0.861 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The rapid evolution of vision language models (VLMs) has heightened the need for benchmarks that rigorously eval" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.363, + 0.901 + ], + "angle": 0, + "content": "*equal contributions. corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "21655" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "uate complex reasoning capabilities. While existing standards like GSM8K [8] focus on domain-specific knowledge in mathematics and science, they neglect a critical reasoning dimension: contextual analysis in visual narrative content. Real-world applications increasingly demand temporal reasoning across interconnected elements: tracking character dynamics in films, interpreting gameplay strategies, or understanding cause-and-effect chains in documentaries. This capability gap persists because current video benchmarks [16, 53, 70, 77] primarily assess single-step perception rather than sustained reasoning processes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.442 + ], + "angle": 0, + "content": "We find three fundamental limitations in current evaluation paradigms: (1) an overemphasis on domain expertise rather than plot-driven reasoning, (2) the lack of temporally-grounded reasoning chains in annotations, and (3) Outcome-focused metrics that ignore procedural validity. To address these challenges, we present VRBench (as in Figure 1), the first benchmark specifically designed for multi-step reasoning in long-form narrative videos. Its construction involves the collection of narrative videos, a human-in-the-loop reasoning process method for annotation, and a multi-phase evaluation pipeline considering both procedure and outcome." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.702 + ], + "angle": 0, + "content": "In data construction, VRBench aggregates 960 meticulously selected videos (1.6h average duration) across seven narrative categories (e.g., movies, sports, travelogues), sourced through a multi-stage filtering process. Our curation pipeline combines automated retrieval with expert validation (inter-rater reliability \\(\\rho = 0.82\\)), prioritizing plot coherence over domain-specific knowledge. This contrasts with existing benchmarks like MMVU [92] that emphasize disciplinary competence. To conduct multi-step reasoning chain annotation, we develop a human-AI collaborative approach, generating 8-10 QA pairs per video with explicit temporal grounding. Each question requires no less than 2 reasoning steps annotated with precise timestamps (Figure 3), validated through iterative expert review (95% inter-annotator agreement). The taxonomy spans seven reasoning types from event prediction to implicit inference, ensuring comprehensive coverage of narrative analysis skills." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "When evaluating models on VRBench, we propose a multi-phase evaluation pipeline that combines outcome verification with process analysis. Beyond conventional multiple-choice accuracy that evaluates the outcome-level performance, we further introduce the process-level metric LLM-guided scoring to evaluate the overall reasoning process quality. This dual approach reveals critical insights - for instance, while GPT-4o achieves \\(81.25\\%\\) outcome accuracy, its process rating scores drop to \\(56.11\\%\\), exposing reasoning fragility. Evaluations of 31 state-of-the-art large models reveal proprietary VLMs with long-context support outperform text-only LLMs by \\(13.82\\%\\) absolute, emphasizing the importance of dense visual ground" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.896, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.138, + 0.885, + 0.163 + ], + "angle": 0, + "content": "Questions: Why did the male protagonist arrange for his friend to rob female protagonist? [00:02:20->00:17:42]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.169, + 0.884, + 0.209 + ], + "angle": 0, + "content": "Step1: When the heroine first caught the hero and his friends, she acted very \nrighteously, and the hero fell in love with her. [00:02:20->00:05:05] \nStep2: The female protagonist chats with the male protagonist and reveals that her \nfavorite type is a manly man. [00:05:06->00:07:58]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.208, + 0.885, + 0.226 + ], + "angle": 0, + "content": "Step3: The male protagonist asks his friends to plan how to show manliness and justice. [00:13:23~>00:16:01]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.226, + 0.869, + 0.245 + ], + "angle": 0, + "content": "Step4: The male protagonist arranges his friends to rob the female protagonist in an attempt to win her love. [00:16:09->00:17:42]" + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.252, + 0.796, + 0.264 + ], + "angle": 0, + "content": "Answer: Because he wants to get the woman's love." + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.271, + 0.708, + 0.283 + ], + "angle": 0, + "content": "Question Type: Event Attribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.298, + 0.908, + 0.354 + ], + "angle": 0, + "content": "Figure 2. An example of annotation in VRBench. For each question, VRBench provides the question-answer pair, multi-step reasoning chain, question type, and the start-to-end timestamps of the entire question as well as each reasoning step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.501 + ], + "angle": 0, + "content": "ing. System-2 optimization strategies yield disproportionate gains in process metrics, and despite parameter parity, open-source VLMs lag behind proprietary counterparts by \\(12.30\\%\\), suggesting architectural limitations beyond scale. Our ablation studies show strong alignment between human and LLM evaluations and quantify the impact of test-time scaling—doubling context windows improves QwQ-32B's [65] accuracy by \\(12.43\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.621 + ], + "angle": 0, + "content": "VRBench sets a new standard for evaluating narrative reasoning, offering insights distinct from knowledge-centric benchmarks. By separating domain expertise from contextual analysis, our work aids in developing models capable of sustained, human-like understanding in real-world video content. The entire suite, including annotation protocols and evaluation tools, has been fully open-sourced to advance reasoning research." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.657, + 0.65 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Reasoning Large Models. As the capabilities of LLMs continue to expand, frontier models have demonstrated significant potential in addressing high-order reasoning tasks. OpenAI pioneered this advancement with the release of o1 [48], an LLM capable of sophisticated reasoning tasks. Subsequently, a series of proprietary [12, 49, 59] and open-source [3, 18, 65] LLMs designed for advanced reasoning have emerged. Furthermore, researchers have shifted their focus towards developing VLMs possessing similar higher-order reasoning capabilities. Previously, some VLMs [2, 7, 30, 32, 37, 71, 73, 88, 91] to a certain extent exhibited good reasoning proficiency due to their proficiency in tackling long-sequence context. More recently, a series of VLMs [14, 66, 69, 81] trained on data incorporating higher-order Chains-of-Thought in RL-based algorithms [33, 52, 55] have been developed, enabling them" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21656" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.42 + ], + "angle": 0, + "content": "
Dataset#Size#QA pairs#Dur.(s)QA typesData SourceAnno StepEval StepEval TargetAnno TypeClueMultilingual
MMLU [21]14,07914,079/MCQMulti-DiscSingleSingleLLMMxx
MMLU-Pro [72]12,03212,032/MCQMulti-DiscSingleSingleLLMA+Mxx
LiveCodeBench [26]511511/MCQCodeMultiSingleLLMMxx
SciEval [57]15,90115,901/MCQ+OpenScienceSingleSingleLLMMxx
GSM8k [8]8,7928,792/MCQ+OpenMathMultiMultiLLMMxx
C-Eval [23]12,34212,342/MCQMulti-DiscSingleSingleLLMMxx
ScienceQA [54]10,33221,208/MCQScienceMultiSingleVLMMxx
VisScience [27]3,0003,000/MCQ+OpenScienceSingleSingleVLMA+Mxx
MMMU [85]11,50011,500/MCQ+OpenMulti-DiscSingleSingleVLMMxx
MMMU-Pro [86]3,4603,460/MCQMulti-DiscSingleSingleVLMA+Mxx
MathVista [42]6,1416,141/MCQ+OpenMathMultiSingleVLMMxx
MathVision [67]3,0403,040/MCQ+OpenMathMultiSingleVLMMxx
CharXiv [74]2,32311,615/OpenMulti-DiscSingleSingleVLMMxx
OlympicArena [25]7,57111,163/MCQ+OpenMulti-DiscMultiMultiVLMMxx
MVBench [31]3,6414,00016.0MCQOpen-DomainSingleSingleVLMA+Mxx
EgoSchema [43]5,0635,063180.0MCQEgocentricSingleSingleVLMAxx
LongVideoBench [77]3,7636,678473.0MCQOpen-DomainSingleSingleVLMMxx
LVBench [70]1031,5494,101MCQOpen-DomainSingleSingleVLMMxx
CGBench [4]1,21912,1291624.4MCQ+OpenOpen-DomainMultiMultiVLMM✓x
MLVU [93]1,7303,102930MCQNarrativeSingleSingleVLMMxx
VideoMME [16]9002,7001017.9MCQOpen-DomainSingleSingleVLMMx✓
Video-MMMU [22]300900506.2MCQMulti-DiscSingleSingleVLMMxx
MMWorld [20]1,9106,627108.0MCQMulti-DiscSingleSingleVLMA+Mxx
MMVU [92]1,5293,00051.4MCQ+OpenMulti-DiscMultiSingleVLMMxx
VRBench (Ours)9608,2435,796.0MCQ+OpenNarrativeMultiMultiLLM, VLMM✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.43, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Table 1. Comparison between VRBench and existing benchmarks. #Size is the number of text/images/videos, #Dur. means the average video duration, A and M indicate automatic and manual annotation type, respectively, and multi-disc denotes multi-disciplinary data source. Clue means clue-grounded annotation. Multilingual requires the number of data source languages to be greater than 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.482, + 0.529 + ], + "angle": 0, + "content": "to solve multidisciplinary, knowledge-intensive problems through a combination of fast and slow thinking processes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.554, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Reasoning Benchmarks. With the rapid advancement of LLMs and VLMs in reasoning capability, numerous reasoning benchmarks have been developed in text and image-text modalities to facilitate comprehensive evaluation. For the textual domain, several knowledge-driven benchmarks [5, 15, 21, 23, 57, 72] have been derived from exam and textbook data sources, and comprise expert-level questions that require multi-step reasoning. Meanwhile, some multimodal benchmarks also emerge to introduce multi-discipline tasks based on charts [44, 74, 78], plots [45], exam [10, 25, 36, 85, 89], or expert-level questions printed on the static images [27, 35, 54, 86]. It is noted that even in the image and text domains, benchmarks with process-level annotations [8, 87] are relatively scarce. For video understanding, evaluation also gradually shifts from short video clip perception [6, 9, 28, 29, 31, 39, 46, 50, 80] to long-form long video understanding [13, 16, 40, 43, 53, 56, 58, 70, 77, 90] and single-step reasoning. [4, 16, 19, 34, 76, 84]. More recently, some advances [13, 20, 92] propose several new video reasoning tasks in multi-disciplinary scenarios like healthcare, engineering, and science. Different from previous works, VRBench is the first narrative video-based benchmark that is purely used for multi-step reasoning eval" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.499, + 0.907, + 0.545 + ], + "angle": 0, + "content": "uation. Table 1 further presents the detailed statistics of our dataset and distinguishes the difference between VRBench and existing text, image, and video reasoning benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.637, + 0.571 + ], + "angle": 0, + "content": "3. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.657 + ], + "angle": 0, + "content": "We present VRBench, a comprehensive multi-step reasoning benchmark consisting of a collection of long, multilingual, and narrative videos with corresponding question-answer pairs. Figure 1 gives the overall pipeline of the benchmark construction, and we illustrate it as follows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.669, + 0.68 + ], + "angle": 0, + "content": "3.1. Video Curation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.763 + ], + "angle": 0, + "content": "We collect long narrative videos from YouTube, yielding an initial pool of over 10,000 public videos. Considering plausible reasoning in steps requires videos with rich and coherent plots, we source long-form footage using a cherry-picked tag set and a comprehensive criterion for narration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Plot-related Tags for Queries. We employ a manually curated set of 7 semantic tags (e.g., Film & Animation, Sport, Travel & Events) to retrieve videos with strong narrative potential. This tag set is developed through iterative validation by domain experts, explicitly excluding non-narrative categories such as news broadcasts and lecture recordings. Our analysis revealed that these excluded categories exhibit limited visual-semantic dynamics (e.g., static" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21657" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.092, + 0.34, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.092, + 0.572, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.092, + 0.857, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.253, + 0.345, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.253, + 0.59, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.254, + 0.856, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 3. Statistics of VRBench. We provide the detailed distribution of videos and annotations of VRBench, including video languages and durations, steps, types, and temporal duration of questions, as well as the token numbers of answers and the reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.52 + ], + "angle": 0, + "content": "camera angles in talk formats) and minimal event progression, rendering them ineffective for benchmarking temporal reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Criterion for Narration. Our sourcing standards include video duration, its language source, and user ratings. The duration requirement (a minimum of 20 minutes and an average of 1.61 hours) ensures adequate temporal context for constructing reasoning chains. Our language diversity strategy intentionally excludes English and Chinese to counterbalance existing dataset biases, aligning with established findings that linguistic variety improves benchmark generalizability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Upon the aforementioned two quantitative prerequisites, we organized a panel of 14 multilingual domain experts to evaluate candidate videos using a standardized 10-point scale based on plot coherence and richness. The scoring is guided by several instructions. Videos scoring below 7 are systematically excluded, resulting in a final curated set of 960 high-quality narratives. Detailed annotation protocols are documented in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.379, + 0.82 + ], + "angle": 0, + "content": "3.2. Stepwise Reasoning Annotations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We label videos with reasoning steps via a two-stage human-in-the-loop framework. It first generates pseudo candidate QA pairs through automated pipelines, then employs expert-guided rewriting to curate high-quality annotations, ensuring rigorous adherence to benchmark speci" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.52 + ], + "angle": 0, + "content": "fications while maintaining multimodal reasoning fidelity. To further improve annotation quality, we also implement a comprehensive quality assurance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Automatic Pipelines. We first employ AutoShot [94] to cut videos into several segments, and then use VideoChat2 [31] to caption them. For auditory content, we adopt whisper-large-v3 [51] to obtain speech transcripts, and DeepL [11] to translate them into English. The video captions and translated subtitles are then put into GPT-4o [47] to generate 6 pseudo QA pairs with a multi-step reasoning process. To ensure the quality of reasoning, we specify 7 multi-step reasoning types for narrative videos as:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.687, + 0.905, + 0.715 + ], + "angle": 0, + "content": "- Event prediction: Forecast subsequent events in the video timeline." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.718, + 0.905, + 0.747 + ], + "angle": 0, + "content": "- Hypothetical reasoning: Deduce plausible scenario outcomes from stated premises." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.748, + 0.905, + 0.776 + ], + "angle": 0, + "content": "- Event attribution: Determine causal origins or underlying motivations of video events." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.778, + 0.905, + 0.806 + ], + "angle": 0, + "content": "- Implicit inference: Extract unstated temporal, emotional, or relational context from visual cues." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.808, + 0.905, + 0.836 + ], + "angle": 0, + "content": "- Logical linkage: Establish event-mediated connections between visual/narrative elements." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.838, + 0.905, + 0.868 + ], + "angle": 0, + "content": "- Information synopsis: Condense critical information across multimodal inputs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.869, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- Counting problems: Quantify state changes through arithmetic/combinatorial analysis." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.687, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21658" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Expert-guided Rewriting. We recruit and train 67 graduate students to generate 8-10 high-quality QA pairs per video, providing raw video footage, translated subtitles, and GPT-generated pre-annotations that offer contextual hints without meeting final benchmark standards. Each QA pair must satisfy four rigorous criteria: (1) Non-synopsis questions require \\(\\geq 2\\) timestamped reasoning steps (start/end times documented); (2) Temporal distribution constraints (\\(\\leq 4\\) questions from \\(0 - 15\\mathrm{min}\\), \\(\\geq 3\\) from \\(15 - 40\\mathrm{min}\\), \\(\\geq 1\\) from \\(40 - 120\\mathrm{min}\\)); (3) Coverage of \\(\\geq 5\\) predefined reasoning taxonomies (from 7 categories) ensuring diversity; (4) Mandatory multimodal grounding where solutions demand both visual analysis (excluding subtitle-only answers) and explicit reasoning (beyond basic perception)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Comprehensive Quality Assurance. To ensure annotations meet our stringent multimodal standards, we implement a rigorous verification protocol: 10 trained reviewers validate each annotation, with non-compliant entries returned to annotators for iterative revisions until fully compliant. Annotators and reviewers are exclusively recruited from top-tier universities to ensure academic rigor. We further enforce quality through dual safeguards: a systematic \\(5\\%\\) random sampling audit across both annotation and review stages, coupled with full documentation of protocols, annotation guidelines, and quality assessment criteria in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.53, + 0.377, + 0.546 + ], + "angle": 0, + "content": "3.3. Multi-Phase Evaluation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We benchmark VLMs and LLMs through a comprehensive multi-phase evaluation pipeline, which compares predictions against ground-truth annotations both at the process and outcome level. For the outcome-level evaluation stage, we adopt a multiple-choice question (MCQ) format, where false options are generated by DeepSeek-V3 [38] using carefully-designed prompts based on human-annotated answers. For the process-level stage, we propose the open-ended rating to fully evaluate the model's multi-step reasoning capability. Specifically, we adopt an LLM to evaluate the overall quality of the whole reasoning process through four 0-10 scores: logical coherence \\((40\\%)\\), similarity to ground truth \\((40\\%)\\), excluded for event prediction and hypothetical reasoning tasks), factual accuracy \\((10\\%)\\), and clarity \\((10\\%)\\). DeepSeek-V3 [38] serves as the judge due to its optimal balance between cost and human alignment (Section 4). For the event prediction and hypothetical reasoning tasks, we argue that there is no ground-truth reasoning steps since there might be multiple possible predictions, hence, we remove the similarity score and compute the final score with an 8:1:1 weight. All metrics are normalized to 0-100 scales, and we report the rankings by computing average scores across all metrics." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.902, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.256, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Human preference alignment results. For each plot, we show the win ratios of three different tested LLMs evaluated by human experts and VRBench. We then fit them with a straight line and quantify the correlation by calculating the Spearman correlation coefficient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.485 + ], + "angle": 0, + "content": "LLM Evaluation Support. For text-only LLMs, we convert videos to text inputs using Qwen2.5-72B-Instruct [83] to synthesize video captions and subtitles into fine-grained summaries. The process involves: 1) Dividing inputs into 5-minute chunks for duplicate removal and abstraction; 2) Merging chunk-level abstracts into coherent summaries with detailed audio-visual descriptions. These summaries enable LLM evaluations while addressing context window limits for VLMs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.646, + 0.515 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.629 + ], + "angle": 0, + "content": "We test a number of both open-source and proprietary VLMs (e.g. Qwen2-VL [68], InternVL2.5 [7], InternVideo2.5 [73], GPT-4o [47], Gemini-2.0-Pro [61], etc.) and LLMs (DeepSeek-R1 [18], QwQ [65], OpenAI o1 [48], Claude-3.7-Sonnet [59], etc.) with a two-stage evaluation on VRBench. The configuration of each evaluated model is detailed in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.781 + ], + "angle": 0, + "content": "Evaluation Protocols. Our protocol involves two phases. Given a question, 1) models produce multi-step reasoning using Chain-of-Thought [75] prompts. Then 2) models choose final answers from multiple-choice options. For answer extraction, final answers are parsed from response endings to mitigate verbose option analysis, following [31]. While reasoning process evaluations use judge LLMs with structured prompts and example-guided rating extraction. All evaluation prompts and implementation details are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.653, + 0.803 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Table 2 shows the CoT evaluation results of LLMs and VLMs on VRBench. Among LLMs, we found the proprietary model Gemini-2.0-Flash-Thinking [12] achieves the optimal performance while other reasoning models such as OpenAI o1-preview [48] and DeepSeek-R1 [18] also demonstrate strong results. It is noted that o1-preview [48]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21659" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.088, + 0.907, + 0.707 + ], + "angle": 0, + "content": "
ModelOverallResults by MetricResults by Taxonomy
MCQ-OOE-PEACPHRIIISEPLL
LLMs
Proprietary Models
GPT-4o [47]55.4963.8747.1154.4132.6369.0361.4771.8366.3766.19
Claude-3.7-Sonnet [59]58.0762.9153.2355.8735.1169.8363.2972.6167.9167.89
o1-preview [48]60.1468.4751.8156.9735.8770.5164.8173.1368.2769.41
Gemini-2.0-Flash-Thinking [61]60.9767.5354.4158.6138.1172.0967.5774.1970.4371.13
Open-Source Models
QwQ-32B-preview [64]35.9027.5144.2934.3134.4146.2939.6727.9742.3940.61
InternLM3-8B-Instruct [3]47.8150.3145.3145.4734.8760.3351.7755.4756.9756.81
Qwen2.5-7B-Instruct [83]48.2952.6143.9746.9335.1761.8351.5754.8756.6357.03
Llama3.3-70B-Instruct [41]49.8452.5947.0947.6338.0763.8754.0354.8359.4756.97
QwQ-32B [65]52.5256.0149.0349.5336.0362.0154.0352.0358.0159.03
Qwen2.5-72B-Instruct [83]53.5160.4946.5351.0336.0167.0358.5367.8361.9363.03
DeepSeek-V3 [38]56.0664.7947.3354.0336.5769.9761.8369.5365.5365.43
DeepSeek-R1 [18]57.1364.1950.0756.1338.9172.8164.1368.0168.5367.93
VLMs
Proprietary Models
Claude-3.7-Sonnet [59]68.1580.0956.2163.1132.9772.6371.1375.3771.3370.27
GPT-4o [47]68.6881.2356.1366.6136.5176.6770.4778.0372.0373.17
Gemini-2.0-Pro [61]74.6183.2965.9371.0965.2181.0175.7387.1377.9375.89
Open-Source Models
DeepSeek-VL2 [79]31.5033.2729.7325.9322.4135.7331.7330.0133.2929.57
H2OVL Mississippi-2B [17]47.1552.3341.9735.1732.3751.7141.4160.0349.9741.91
Phi-3.5-Vision [1]48.5258.0339.0131.5328.0345.0337.0368.0344.0336.03
LongVA-7B [91]50.1467.8132.4727.6125.0738.6933.7776.9138.2732.07
InternVL2.5-8B [7]50.4169.3131.5126.6326.3137.9931.2185.9737.3730.69
MiMo-VL-7B-RL [60]50.4863.3937.5747.3134.5963.8353.2374.6160.3356.97
VideoChat-Flash-7B [32]50.8272.0129.6324.6921.4139.3729.1779.3734.8728.07
InternVideo2.5 [73]51.9475.6328.2523.8124.0934.5127.0784.5733.6126.81
LongVA-7B-DPO [91]52.3667.9136.8130.7327.1745.0938.2379.3943.6737.03
Qwen2-VL-7B [68]54.0872.0136.1530.4926.6146.7133.6985.2743.3135.29
Aria [30]54.5572.9736.1330.1929.2344.2334.6386.2344.9934.57
Qwen2.5-VL-7B [2]56.5269.6143.4337.0733.1754.2742.2783.9150.1742.87
Keye-VL-8B-Preview [63]60.4464.4156.4762.5340.3773.6167.9970.1371.8371.03
Qwen2.5-VL-72B [2]61.7166.8556.5751.8746.1367.1354.1390.0463.6760.77
Kimi-VL-A3B-Thinking-2506 [62]61.8261.6761.9764.5347.9171.3769.5774.4771.2371.47
InternVL2.5-78B [7]62.3176.6148.0143.7738.6758.2146.1387.5354.4347.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.717, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Table 2. Evaluation results on VRBench across two evaluation metrics (outcome-level MCQ, process-level open-ended evaluation) and seven QA taxonomies (event attribution, counting problems, hypothetical reasoning, implicit inferences, information synopsis, event prediction, logical linkage). Bold values indicate the best results among all models, and underlined values are best results of LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.482, + 0.892 + ], + "angle": 0, + "content": "achieves the best outcome-level MCQ accuracy of \\(68.47\\%\\). For the VLMs, Gemini-2.0-Pro [61] achieves \\(74.61\\%\\) overall accuracy, surpassing all components by at least \\(5.93\\%\\). GPT-4o [47] and Claude-3.7-Sonnet [59] also show competitive results, which demonstrate \\(68.68\\%\\) and \\(68.15\\%\\) overall performance, respectively. For open-source models, InternVL2.5-78B [7], emerges as the best non-proprietary" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.785, + 0.774, + 0.801 + ], + "angle": 0, + "content": "VLMs with a \\(62.31\\%\\) overall accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We then delve into the specific results of each metric and taxonomy. For results in each evaluation stage, we find that most of LLMs and VLMs are capable of achieving high MCQ accuracy, yet compared to LLMs, VLMs struggle to demonstrate their reasoning steps and exhibit lower reasoning ratings. Through results across taxonomy, we found" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21660" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "that most of the large models are not proficient in counting problems since they require fine-grained visual perception, in contrast with other tasks. Since the perception of LLMs fully depends on the video summary and some VLMs only support a small number of frame inputs, they are unable to correctly perceive the original frames containing the target elements, resulting in the MCQ accuracy close to random guessing and low reasoning process ratings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.225, + 0.305, + 0.24 + ], + "angle": 0, + "content": "4.2. Ablations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.483, + 0.293 + ], + "angle": 0, + "content": "We then further discuss the observations from the evaluation results and highlight some insights on obtaining higher multi-step reasoning performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.483, + 0.429 + ], + "angle": 0, + "content": "Does the evaluation of VRBench align with Human Preference? In contrast to MCQs that compute accuracy via a deterministic method, the correctness of process-level open-ended evaluation highly depends on the judgment of LLM's performance, which may include hallucinations that could influence the evaluation reliability. To this end, we select a subset of 30 videos with 300 questions to perform process-level human evaluation, aiming to quantify the correlation between human and LLM assessment results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Specifically, the human annotators are asked to give process-level open-ended ratings following the same requirement for the judging LLM, i.e., four separated ratings are needed for each question to evaluate the logical coherence, similarity to ground-truth, factual accuracy, and expression clarity. We then compute the win ratio of pairwise comparison for each tested model following [24], where a model scores 1 if its rating is higher than its current opponent, 0 if lower, and 0.5 for a tie. Finally, we calculate the average win ratio of each model and the Spearman correlation coefficients \\((\\rho)\\) of the win ratios evaluated by LLMs and human annotators." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.824 + ], + "angle": 0, + "content": "We first investigate the human preference alignment of several LLMs. We select GPT-4o [47], DeepSeek-V3 [38], Qwen2.5-7B [83], and Qwen2.5-72B [83] as evaluation models and give them the same rating prompt for a fair comparison. As shown in Figure 4, Qwen2.5 shows low human preference alignment compared with DeepSeek-V3 and GPT-4o, which both have correlation coefficients greater than 0.8. Though the human alignment of GPT-4o is slightly better than DeepSeek-V3, we observe the evaluation cost of GPT-4o is approximately 10 times that of DeepSeek, which imposes a significant burden on the benchmark users. Hence, we adopt DeepSeek-V3 as the judging model for its optimal trade-off between performance and cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We also probe the evaluation correctness of each open-ended sub-metric. Following the same procedure, we compute the win ratio and correlation coefficient of DeepSeekV3 for each sub-metric, and the results are shown in Figure 4 (b-d). Results show the coefficients in all metrics are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "higher than 0.8, indicating a strong positive monotonic relationship between LLM and human evaluation results. Another alternative is to instruct the LLM to compute the final score based on the given weights of each sub-metric, which is denoted as DeepSeek w/o sep in Figure 4. Results show that adding this additional computing process diminishes the robustness of evaluation LLM, thus in practice, we ask the model to output all sub-metric rating separately and automatically compute the final score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.905, + 0.64 + ], + "angle": 0, + "content": "The role of System-2 thinking. o1-like LLMs, also denoted as system-2 models, have emerged with superior reasoning abilities in multi-disciplinary scenarios such as science and math. We further explore these models' performance on the narrative video benchmark. We both assess the role of open-source and proprietary o1-models, as well as the gap compared with their previous system-1 version. For proprietaries, we compare OpenAI o1-preview [48] with GPT-4o [47], and results show that the system-2 version o1 outperforms GPT-4o with \\(4.30\\%\\) overall accuracy. For open-source model DeepSeek-V3 [38] and R1 [11], we observe that DeepSeek-R1 achieves higher scores both on multiple-choice questions and open-ended evaluations. It is noted that compared with Qwen2.5-72B-Instruct [82], QwQ-32B [65] gets higher reasoning ratings yet lower outcome-level MCQ results. This shows that though some system-2 models are more proficient in multi-step thinking, their ability to converge lengthy thinking processes into correct answers still needs improvement. Recently, some o1-like training strategies have also been utilized by some VLMs, such as LongVA-7B-DPO [91], a model that utilizes direct preference optimization [52] for long video understanding. It achieves an overall performance of \\(52.36\\%\\) and surpasses its vanilla version, suggesting that training VLMs with different optimization methods is feasible and that developing system-2 VLM models is crucial for tackling multi-step reasoning questions in narrative videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Impact of Model Size. It is commonly believed that models with large scales are often accompanied by better perception and reasoning capabilities. To verify such claims, we conduct experiments on the models with different scales, such as Qwen2.5-7B [83] and Qwen2.5-72B, Qwen2.5-VL [2] with 7B and 72B parameters, and InternVL2.5-8B [7] and InternVL2.5-78B. As shown in Table 2, models with more parameters obtain higher overall accuracy (48.29% vs 53.51% for Qwen2.5, 56.52% vs 61.71% for Qwen2.5-VL, and 50.41% vs 62.31% for InternVL2.5), and the performance gap on process-level metrics is more apparent. This indicates that large-scale models are more likely to achieve advanced reasoning capabilities. However, we also noticed that models specifically trained on the reasoning corpus like QwQ-32B [65], can achieve comparable performance compared with its 72B Qwen2.5 counterpart (52.52% vs. 53.51% in overall accuracy) with a smaller" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21661" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "parameter amount. This suggests that training small-sized models with various preference optimization strategies and reasoning-centric corpora might improve their overall reasoning capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Long Context Helps. We observe a substantial performance gap between models with and without long context support. As shown in Table 2, models with long frame input tend to achieve higher overall accuracy. For example, Gemini-2.0-Pro [61] is capable of inference with 0.5 fps and supports large number frame inputs when answering questions of videos over 1 hour, and it achieves the optimal performance compared with all other fixed-length models. For the open-source models, Qwen2.5-VL-7B [2] and KimiVL-A3B-Thinking-2506 [62] with 256 frame input achieve \\(61.71\\%\\) and \\(61.82\\%\\) overall accuracy, which is \\(11.30\\%\\) and \\(11.41\\%\\) higher than the 8B InternVL-2.5 [7] model with 64 frames input. Since the average length of VRBench is 1.61 hours, the necessity of long context support is becoming more pronounced, so that models can perceive more plot elements that help the analysis of long video content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.653 + ], + "angle": 0, + "content": "LLMs vs. VLMs. Since tested LLMs are only provided with the automatically-generated video summary, yet VLMs are capable of directly perceiving the raw visual content, it is unfair to directly compare the numerical results of LLMs and VLMs. However, there are still some noteworthy phenomena that can be discussed. First, VLMs with coarse-grained visual input (e.g., 4 frames for H2OVL Mississippi-2B [17] and DeepSeek-VL2 [79]) perform significantly worse than several system-2 LLMs like DeepSeek-R1 [18] and Claude-3.7-Sonnet [59]. This indicates that randomly selecting a small number of frames is not enough to tackle narrative-based reasoning tasks that require long-range perception. Furthermore, VLMs with fine-grained visual perception, such as Gemini-2.0-Pro with a 0.5 fps perform better than top-tier LLMs on both process-level and outcome-level metrics, showing the inevitability of detailed visual content input for tackling questions in VRBench." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.363, + 0.683 + ], + "angle": 0, + "content": "4.3. Test-Time Scaling Exploration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We here probe models' performance when scaling test-time compute cost. Since we tend to prompt the models with a Chain-of-Thought template when tackling complex reasoning tasks, it is observed that system-2 models are capable of dramatically expanding the inference budget to improve their performance. To this end, we set a series of token limitations on these models and require the models to think with different instructions. Specifically, we select QwQ-32B [65] as the experimental LLM and Qwen2-VL-7B [82] as the testing VLM on a subset with 300 videos and 2,403 questions. For each model, several parallel experiments are conducted with increasing maximum token limits from 256 to 2048. We instruct the models to output the reasoning process as much as possible for large token limitations, and" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.091, + 0.891, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.272, + 0.907, + 0.302 + ], + "angle": 0, + "content": "Figure 5. Test-Time Scaling Results. We report the average accuracy of outcome-level MCQ and process-level open-ended ratings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.908, + 0.647 + ], + "angle": 0, + "content": "let them think concisely when the token limit is low. Since we aim to investigate the quality of models' initial reasoning process and their capability to generate the right answers through thinking, we report the overall scores that both evaluate the outcome accuracy and process quality. Results are shown in Figure 5, where QwQ shows a remarkable rating boost from \\(48.91\\%\\) to \\(61.34\\%\\) when setting a large token limit. On the contrary, the small-sized system-1 model Qwen2-VL-7B performs even worse when using a large token number limitation and instructing the model to output more thinking process. The model tends to output more ambiguous outputs that lead to the wrong answer. This observation suggests that models with large parameter scales and system-2 capabilities benefit from the test-time scaling strategy. It also provides the insight that developing more test-time scaling approaches that guide the model to generate long reasoning traces could be a feasible and promising way to unlock more potential capabilities of system-2 models, thereby benefiting complex tasks that require multi-step reasoning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.686, + 0.634, + 0.701 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper introduces VRBench, a comprehensive long-narrative video benchmark for evaluating multi-step reasoning. Through manually filtered narrative videos, high-quality stepwise annotations, and a multi-phase evaluation pipeline, VRBench distinguishes itself from other existing reasoning benchmarks and shows the robust capability to evaluate LLMs and VLMs both from the process and outcome perspectives. By assessing and analyzing 31 frontier large models, we thoroughly demonstrate the detailed performance of current reasoning models across various reasoning questions and provide valuable insights towards constructing more advanced multi-step reasoning models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21662" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.276, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.173 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160101) and Jiangsu Frontier Technology Research and Development Program (No. BF2024076)," + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.203, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.482, + 0.298 + ], + "angle": 0, + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.3, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.359, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[3] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.416, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[4] Guo Chen, Yicheng Liu, Yifei Huang, Yuping He, Baoqi Pei, Jilan Xu, Yali Wang, Tong Lu, and Limin Wang. Cg-bench: Clue-grounded question answering benchmark for long video understanding. arXiv preprint arXiv:2412.12075, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.488, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[5] Wenhu Chen, Ming Yin, Max Ku, Pan Lu, Yixin Wan, Xueguang Ma, Jianyu Xu, Xinyi Wang, and Tony Xia. Theoremq: A theorem-driven question answering dataset. arXiv preprint arXiv:2305.12524, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[6] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. In European Conference on Computer Vision, pages 179-195. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[7] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang-wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees GM Snoek, and Yuki M Asano. Tvbench: Redesigning video-language evaluation. arXiv preprint arXiv:2410.07752, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[10] Rocktim Jyoti Das, Simeon Emilov Hristov, Haonan Li, Dimitar Iliyanov Dimitrov, Ivan Koychev, and Preslav Nakov. *Exams-v: A multi-discipline multilingual multimodal exam benchmark for evaluating vision language models.* arXiv preprint arXiv:2403.10378, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[11] DeepL. Deepl translate: The world's most accurate translator. https://www deepl.com/en/Translator, 2025.4,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Google Deepmind. Gemini 2.0 flash thinking. https://deepmind.google/technologies/gemini/flash-thinking/, 2025. Accessed: 2025-01-21. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[13] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[14] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[15] Kehua Feng, Keyan Ding, Weijie Wang, Xiang Zhuang, Zeyuan Wang, Ming Qin, Yu Zhao, Jianhua Yao, Qiang Zhang, and Huajun Chen. Sciknoweval: Evaluating multilevel scientific knowledge of large language models. arXiv preprint arXiv:2406.09098, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[17] Shaikat Galib, Shanshan Wang, Guanshuo Xu, Pascal Pfeiffer, Ryan Chesler, Mark Landry, and Sri Satish Ambati. H2ovl-mississippi vision language models technical report. arXiv preprint arXiv:2410.13611, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[18] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[19] Songhao Han, Wei Huang, Hairong Shi, Le Zhuo, Xiu Su, Shifeng Zhang, Xu Zhou, Xiaojuan Qi, Yue Liao, and Si Liu. Videospeech: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection. arXiv preprint arXiv:2411.14794, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[20] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multi-discipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Kairui Hu, Penghao Wu, Fanyi Pu, Wang Xiao, Yuanhan Zhang, Xiang Yue, Bo Li, and Ziwei Liu. Video-mmmu: Evaluating knowledge acquisition from multi-discipline professional videos. arXiv preprint arXiv:2501.13826, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[23] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems, 36:62991-63010, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.091, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21663" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[25] Zhen Huang, Zengzhi Wang, Shijie Xia, Xuefeng Li, Haoyang Zou, Ruijie Xu, Run-Ze Fan, Lyumanshan Ye, Ethan Chern, Yixin Ye, et al. Olympic: Benchmarking multi-discipline cognitive reasoning for superintelligent ai. Advances in Neural Information Processing Systems, 37: 19209-19253, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[26] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livocodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[27] Zhihuan Jiang, Zhen Yang, Jinhao Chen, Zhengxiao Du, Weihan Wang, Bin Xu, Yuxiao Dong, and Jie Tang. Visscience: An extensive benchmark for evaluating k12 educational multi-modal scientific reasoning. arXiv preprint arXiv:2409.13730, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[28] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[29] Bozheng Li, Yongliang Wu, Yi Lu, Jiashuo Yu, Licheng Tang, Jiawang Cao, Wenqing Zhu, Yuyang Sun, Jay Wu, and Wenbo Zhu. Veu-bench: Towards comprehensive understanding of video editing. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13671-13680, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.593 + ], + "angle": 0, + "content": "[30] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[32] Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, et al. Videochat-flash: Hierarchical compression for long-context video modeling. arXiv preprint arXiv:2501.00574, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[33] Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning. arXiv preprint arXiv:2504.06958, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[34] Yunxin Li, Xinyu Chen, Baotian Hu, Longyue Wang, Haoyuan Shi, and Min Zhang. Videovista: A versatile benchmark for video understanding and reasoning. arXiv preprint arXiv:2406.11303, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[35] Zekun Li, Xianjun Yang, Kyuri Choi, Wanrong Zhu, Ryan Hsieh, HyeonJung Kim, Jin Hyuk Lim, Sungyoung Ji," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Byungju Lee, Xifeng Yan, et al. Mmsci: A multimodal multi-discipline dataset for phd-level scientific comprehension. In AI for Accelerated Materials Design-Vienna 2024, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[36] Zhenwen Liang, Kehan Guo, Gang Liu, Taicheng Guo, Yujun Zhou, Tianyu Yang, Jiajun Jiao, Renjie Pi, Jipeng Zhang, and Xiangliang Zhang. Scemqa: A scientific college entrance level multimodal question answering benchmark. arXiv preprint arXiv:2402.05138, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[37] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[38] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[39] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[40] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. Et bench: Towards open-ended event-level video-language understanding. arXiv preprint arXiv:2409.18111, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[41] Llama-3.3. Llama-3.3-70b-instruct. https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct, 2025.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[43] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[44] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[45] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[46] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[47] OpenAI. Hello gpt4-o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-13, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.902, + 0.872 + ], + "angle": 0, + "content": "[48] OpenAI. Introducing openai o1. https://openai.com/o1/, 2024.2,5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[49] OpenAI. Openai o3-mini. https://openai.com/index/openai-o3-mini/, 2025.2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21664" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[50] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36:42748-42761, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.176, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[52] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[53] Ruchit Rawal, Khalid Saifullah, Miquel Farre, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. arXiv preprint arXiv:2405.08813, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[54] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[55] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[56] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[57] Liangtai Sun, Yang Han, Zihan Zhao, Da Ma, Zhennan Shen, Baocai Chen, Lu Chen, and Kai Yu. Scieval: A multi-level large language model evaluation benchmark for scientific research. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19053-19061, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[58] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[59] Anthropic. Claude Team. Claude 3.7 sonnet. https://www.anthropic.com/claude/sonnet, 2025. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[60] Core Team, Zihao Yue, Zhenru Lin, Yifan Song, Weikun Wang, Shuhuai Ren, Shuhao Gu, Shicheng Li, Peidian Li, Liang Zhao, Lei Li, et al. Mimo-vl technical report. https://arxiv.org/abs/2506.03569, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[61] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chen-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "zhuang Du, Chu Wei, et al. Kimi-vl technical report. arXiv preprint arXiv:2504.07491, 2025. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[63] Kwai Keye Team, Biao Yang, Bin Wen, Changyi Liu, Chenglong Chu, Chengru Song, Chongling Rao, Chuan Yi, Da Li, Dunju Zang, et al. Kwai keye-vl technical report. arXiv preprint arXiv:2507.01949, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[64] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview/, 2024. Accessed: 2024-11-28. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[65] Qwen Team. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github.io/blog/qwq-32b/, 2025. Accessed: 2025-3-6. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[66] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamavol: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[69] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, et al. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[70] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[72] Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[73] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[74] Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, et al. Charxiv: Charting gaps in realistic chart understanding in multimodal llms. Advances in Neural Information Processing Systems, 37:113569-113697, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21665" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[75] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[76] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[77] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[78] Yifan Wu, Lutao Yan, Leixian Shen, Yunhai Wang, Nan Tang, and Yuyu Luo. Chartinsights: Evaluating multimodal large language models for low-level chart question answering. arXiv preprint arXiv:2405.07001, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[79] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[80] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[81] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-01: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[82] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report, 2024. URL: https://arxiv.org/abs/2407.10671, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[83] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[84] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. Advances in Neural Information Processing Systems, 37:57240-57261, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[85] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[86] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[87] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[88] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[89] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, et al. Cmmmu: A chinese massive multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2401.11944, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[90] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[91] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.374, + 0.905, + 0.443 + ], + "angle": 0, + "content": "[92] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.445, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[93] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.515, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[94] Wentao Zhu, Yufang Huang, Xiufeng Xie, Wenxian Liu, Jincan Deng, Debing Zhang, Zhangyang Wang, and Ji Liu. Autoshot: A short video dataset and state-of-the-art shot boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2238-2247, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21666" + } + ] +] \ No newline at end of file diff --git a/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_origin.pdf b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26472282c58e5aa1d6a21ad305eff86b18ee52d5 --- /dev/null +++ b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/13199d03-d2f0-4f02-9446-3b78b6cbc2ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffce093720b28afe65f749fa29398a8067088daa1ed12b66d4704a4eeac11010 +size 1418570 diff --git a/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/full.md b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..13fe033aedaf19f876576bf9476f2b4d55979d84 --- /dev/null +++ b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/full.md @@ -0,0 +1,296 @@ +# VRBench: A Benchmark for Multi-Step Reasoning in Long Narrative Videos + +Jiashuo Yu $^{1*}$ Yue Wu $^{1*}$ Meng Chu $^{1*}$ Zhifei Ren $^{1*}$ Zizheng Huang $^{2,3,1*}$ Pei Chu $^{1*}$ + +Ruijie Zhang $^{1}$ Yinan He $^{1}$ Qirui Li $^{1}$ Songze Li $^{1}$ Zhenxiang Li $^{1}$ Zhongying Tu $^{1}$ + +Conghui He1 Yu Qiao1 Yali Wang4,1⊗ Yi Wang1,3⊗ Limin Wang2,1⊗ + +$^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ Nanjing University $^{3}$ Shanghai Innovation Institute + +4Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences + +VRBench.github.io + +![](images/9b75860174e48a322566b434e2eea6279429859803f92425fc8b9f779f5ad7b1.jpg) +Figure 1. Overview of VRBench. We present VRBench, a long narrative video benchmark for multi-step reasoning. VRBench includes 960 manual-filtered narrative videos, covering 8 languages and 7 video categories that are suitable for reasoning about temporal relations. We also provide high-quality stepwise annotations for reasoning, which are labeled and reviewed by human experts. Each video incorporates 8-10 complex question-answer pairs, a multi-step reasoning chain, and fine-grained timestamps. To fully evaluate the capability of models in multi-step reasoning, we propose a multi-phase evaluation pipeline that assesses model results both from the process and outcome levels. Our VRBench is the first video reasoning benchmark that supports both multi-step annotation and evaluation. + +# Abstract + +We present VRBench, the first long narrative video benchmark crafted for evaluating large models' multi-step reasoning capabilities, addressing limitations in existing evaluations that overlook temporal reasoning and procedural validity. It comprises 960 long videos (with an average duration of 1.6 hours), along with 8,243 human-labeled multi-step question-answering pairs and 25,106 reasoning steps with timestamps. These videos are curated via a multistage filtering process including expert inter-rater reviewing to prioritize plot coherence. We develop a human-AI collaborative framework that generates coherent reasoning chains, each requiring multiple temporally grounded steps, + +spanning seven types (e.g., event attribution, implicit inference). VRBench designs a multi-phase evaluation pipeline that assesses models at both the outcome and process levels. Apart from the MCQs for the final results, we propose a progress-level LLM-guided scoring metric to evaluate the quality of the reasoning chain from multiple dimensions comprehensively. Through extensive evaluations of 12 LLMs and 19 VLMs on VRBench, we undertake a thorough analysis and provide valuable insights that advance the field of multi-step reasoning. + +# 1. Introduction + +The rapid evolution of vision language models (VLMs) has heightened the need for benchmarks that rigorously eval + +uate complex reasoning capabilities. While existing standards like GSM8K [8] focus on domain-specific knowledge in mathematics and science, they neglect a critical reasoning dimension: contextual analysis in visual narrative content. Real-world applications increasingly demand temporal reasoning across interconnected elements: tracking character dynamics in films, interpreting gameplay strategies, or understanding cause-and-effect chains in documentaries. This capability gap persists because current video benchmarks [16, 53, 70, 77] primarily assess single-step perception rather than sustained reasoning processes. + +We find three fundamental limitations in current evaluation paradigms: (1) an overemphasis on domain expertise rather than plot-driven reasoning, (2) the lack of temporally-grounded reasoning chains in annotations, and (3) Outcome-focused metrics that ignore procedural validity. To address these challenges, we present VRBench (as in Figure 1), the first benchmark specifically designed for multi-step reasoning in long-form narrative videos. Its construction involves the collection of narrative videos, a human-in-the-loop reasoning process method for annotation, and a multi-phase evaluation pipeline considering both procedure and outcome. + +In data construction, VRBench aggregates 960 meticulously selected videos (1.6h average duration) across seven narrative categories (e.g., movies, sports, travelogues), sourced through a multi-stage filtering process. Our curation pipeline combines automated retrieval with expert validation (inter-rater reliability $\rho = 0.82$ ), prioritizing plot coherence over domain-specific knowledge. This contrasts with existing benchmarks like MMVU [92] that emphasize disciplinary competence. To conduct multi-step reasoning chain annotation, we develop a human-AI collaborative approach, generating 8-10 QA pairs per video with explicit temporal grounding. Each question requires no less than 2 reasoning steps annotated with precise timestamps (Figure 3), validated through iterative expert review (95% inter-annotator agreement). The taxonomy spans seven reasoning types from event prediction to implicit inference, ensuring comprehensive coverage of narrative analysis skills. + +When evaluating models on VRBench, we propose a multi-phase evaluation pipeline that combines outcome verification with process analysis. Beyond conventional multiple-choice accuracy that evaluates the outcome-level performance, we further introduce the process-level metric LLM-guided scoring to evaluate the overall reasoning process quality. This dual approach reveals critical insights - for instance, while GPT-4o achieves $81.25\%$ outcome accuracy, its process rating scores drop to $56.11\%$ , exposing reasoning fragility. Evaluations of 31 state-of-the-art large models reveal proprietary VLMs with long-context support outperform text-only LLMs by $13.82\%$ absolute, emphasizing the importance of dense visual ground + +![](images/f51bb77bf0fc108a977640f5339c2cf941c34fb2de8defdcae4c944c9f6911bf.jpg) +Figure 2. An example of annotation in VRBench. For each question, VRBench provides the question-answer pair, multi-step reasoning chain, question type, and the start-to-end timestamps of the entire question as well as each reasoning step. + +# Questions: Why did the male protagonist arrange for his friend to rob female protagonist? [00:02:20->00:17:42] + +Step1: When the heroine first caught the hero and his friends, she acted very +righteously, and the hero fell in love with her. [00:02:20->00:05:05] +Step2: The female protagonist chats with the male protagonist and reveals that her +favorite type is a manly man. [00:05:06->00:07:58] + +Step3: The male protagonist asks his friends to plan how to show manliness and justice. [00:13:23~>00:16:01] + +Step4: The male protagonist arranges his friends to rob the female protagonist in an attempt to win her love. [00:16:09->00:17:42] + +# Answer: Because he wants to get the woman's love. + +# Question Type: Event Attribution + +ing. System-2 optimization strategies yield disproportionate gains in process metrics, and despite parameter parity, open-source VLMs lag behind proprietary counterparts by $12.30\%$ , suggesting architectural limitations beyond scale. Our ablation studies show strong alignment between human and LLM evaluations and quantify the impact of test-time scaling—doubling context windows improves QwQ-32B's [65] accuracy by $12.43\%$ . + +VRBench sets a new standard for evaluating narrative reasoning, offering insights distinct from knowledge-centric benchmarks. By separating domain expertise from contextual analysis, our work aids in developing models capable of sustained, human-like understanding in real-world video content. The entire suite, including annotation protocols and evaluation tools, has been fully open-sourced to advance reasoning research. + +# 2. Related Work + +Reasoning Large Models. As the capabilities of LLMs continue to expand, frontier models have demonstrated significant potential in addressing high-order reasoning tasks. OpenAI pioneered this advancement with the release of o1 [48], an LLM capable of sophisticated reasoning tasks. Subsequently, a series of proprietary [12, 49, 59] and open-source [3, 18, 65] LLMs designed for advanced reasoning have emerged. Furthermore, researchers have shifted their focus towards developing VLMs possessing similar higher-order reasoning capabilities. Previously, some VLMs [2, 7, 30, 32, 37, 71, 73, 88, 91] to a certain extent exhibited good reasoning proficiency due to their proficiency in tackling long-sequence context. More recently, a series of VLMs [14, 66, 69, 81] trained on data incorporating higher-order Chains-of-Thought in RL-based algorithms [33, 52, 55] have been developed, enabling them + +
Dataset#Size#QA pairs#Dur.(s)QA typesData SourceAnno StepEval StepEval TargetAnno TypeClueMultilingual
MMLU [21]14,07914,079/MCQMulti-DiscSingleSingleLLMMxx
MMLU-Pro [72]12,03212,032/MCQMulti-DiscSingleSingleLLMA+Mxx
LiveCodeBench [26]511511/MCQCodeMultiSingleLLMMxx
SciEval [57]15,90115,901/MCQ+OpenScienceSingleSingleLLMMxx
GSM8k [8]8,7928,792/MCQ+OpenMathMultiMultiLLMMxx
C-Eval [23]12,34212,342/MCQMulti-DiscSingleSingleLLMMxx
ScienceQA [54]10,33221,208/MCQScienceMultiSingleVLMMxx
VisScience [27]3,0003,000/MCQ+OpenScienceSingleSingleVLMA+Mxx
MMMU [85]11,50011,500/MCQ+OpenMulti-DiscSingleSingleVLMMxx
MMMU-Pro [86]3,4603,460/MCQMulti-DiscSingleSingleVLMA+Mxx
MathVista [42]6,1416,141/MCQ+OpenMathMultiSingleVLMMxx
MathVision [67]3,0403,040/MCQ+OpenMathMultiSingleVLMMxx
CharXiv [74]2,32311,615/OpenMulti-DiscSingleSingleVLMMxx
OlympicArena [25]7,57111,163/MCQ+OpenMulti-DiscMultiMultiVLMMxx
MVBench [31]3,6414,00016.0MCQOpen-DomainSingleSingleVLMA+Mxx
EgoSchema [43]5,0635,063180.0MCQEgocentricSingleSingleVLMAxx
LongVideoBench [77]3,7636,678473.0MCQOpen-DomainSingleSingleVLMMxx
LVBench [70]1031,5494,101MCQOpen-DomainSingleSingleVLMMxx
CGBench [4]1,21912,1291624.4MCQ+OpenOpen-DomainMultiMultiVLMM✓x
MLVU [93]1,7303,102930MCQNarrativeSingleSingleVLMMxx
VideoMME [16]9002,7001017.9MCQOpen-DomainSingleSingleVLMMx✓
Video-MMMU [22]300900506.2MCQMulti-DiscSingleSingleVLMMxx
MMWorld [20]1,9106,627108.0MCQMulti-DiscSingleSingleVLMA+Mxx
MMVU [92]1,5293,00051.4MCQ+OpenMulti-DiscMultiSingleVLMMxx
VRBench (Ours)9608,2435,796.0MCQ+OpenNarrativeMultiMultiLLM, VLMM✓✓
+ +Table 1. Comparison between VRBench and existing benchmarks. #Size is the number of text/images/videos, #Dur. means the average video duration, A and M indicate automatic and manual annotation type, respectively, and multi-disc denotes multi-disciplinary data source. Clue means clue-grounded annotation. Multilingual requires the number of data source languages to be greater than 2. + +to solve multidisciplinary, knowledge-intensive problems through a combination of fast and slow thinking processes. + +Reasoning Benchmarks. With the rapid advancement of LLMs and VLMs in reasoning capability, numerous reasoning benchmarks have been developed in text and image-text modalities to facilitate comprehensive evaluation. For the textual domain, several knowledge-driven benchmarks [5, 15, 21, 23, 57, 72] have been derived from exam and textbook data sources, and comprise expert-level questions that require multi-step reasoning. Meanwhile, some multimodal benchmarks also emerge to introduce multi-discipline tasks based on charts [44, 74, 78], plots [45], exam [10, 25, 36, 85, 89], or expert-level questions printed on the static images [27, 35, 54, 86]. It is noted that even in the image and text domains, benchmarks with process-level annotations [8, 87] are relatively scarce. For video understanding, evaluation also gradually shifts from short video clip perception [6, 9, 28, 29, 31, 39, 46, 50, 80] to long-form long video understanding [13, 16, 40, 43, 53, 56, 58, 70, 77, 90] and single-step reasoning. [4, 16, 19, 34, 76, 84]. More recently, some advances [13, 20, 92] propose several new video reasoning tasks in multi-disciplinary scenarios like healthcare, engineering, and science. Different from previous works, VRBench is the first narrative video-based benchmark that is purely used for multi-step reasoning eval + +uation. Table 1 further presents the detailed statistics of our dataset and distinguishes the difference between VRBench and existing text, image, and video reasoning benchmarks. + +# 3. Benchmark + +We present VRBench, a comprehensive multi-step reasoning benchmark consisting of a collection of long, multilingual, and narrative videos with corresponding question-answer pairs. Figure 1 gives the overall pipeline of the benchmark construction, and we illustrate it as follows. + +# 3.1. Video Curation + +We collect long narrative videos from YouTube, yielding an initial pool of over 10,000 public videos. Considering plausible reasoning in steps requires videos with rich and coherent plots, we source long-form footage using a cherry-picked tag set and a comprehensive criterion for narration. + +Plot-related Tags for Queries. We employ a manually curated set of 7 semantic tags (e.g., Film & Animation, Sport, Travel & Events) to retrieve videos with strong narrative potential. This tag set is developed through iterative validation by domain experts, explicitly excluding non-narrative categories such as news broadcasts and lecture recordings. Our analysis revealed that these excluded categories exhibit limited visual-semantic dynamics (e.g., static + +![](images/01a1f4f86ae78d2f1f2b5f4ca2438e27ce9a62d975e3b857e4f83723525d0e9f.jpg) + +![](images/fe11c9a4213c0bbaa61a589de2054c54a19a7b6d7af1151be4d9c7276eb11ff3.jpg) + +![](images/629ab1b1b4239bb663c922988fda19806527b64f756876e3d711adb4689a9af4.jpg) + +![](images/ddad08f6901c1c40f9553f14165c6c9dfd94deb9e73630dda394e2e47e0dd585.jpg) +Figure 3. Statistics of VRBench. We provide the detailed distribution of videos and annotations of VRBench, including video languages and durations, steps, types, and temporal duration of questions, as well as the token numbers of answers and the reasoning process. + +![](images/046cc163db7f6f9b589e856b6de421827f5ee2a6f603faa658e8776382ed5ee8.jpg) + +![](images/9c402527a1c58f4e9e7f8a5f7ee87f241137acf7dd2bf6c2dae32c420c1bf697.jpg) + +camera angles in talk formats) and minimal event progression, rendering them ineffective for benchmarking temporal reasoning capabilities. + +Criterion for Narration. Our sourcing standards include video duration, its language source, and user ratings. The duration requirement (a minimum of 20 minutes and an average of 1.61 hours) ensures adequate temporal context for constructing reasoning chains. Our language diversity strategy intentionally excludes English and Chinese to counterbalance existing dataset biases, aligning with established findings that linguistic variety improves benchmark generalizability. + +Upon the aforementioned two quantitative prerequisites, we organized a panel of 14 multilingual domain experts to evaluate candidate videos using a standardized 10-point scale based on plot coherence and richness. The scoring is guided by several instructions. Videos scoring below 7 are systematically excluded, resulting in a final curated set of 960 high-quality narratives. Detailed annotation protocols are documented in the supplementary materials. + +# 3.2. Stepwise Reasoning Annotations + +We label videos with reasoning steps via a two-stage human-in-the-loop framework. It first generates pseudo candidate QA pairs through automated pipelines, then employs expert-guided rewriting to curate high-quality annotations, ensuring rigorous adherence to benchmark speci + +fications while maintaining multimodal reasoning fidelity. To further improve annotation quality, we also implement a comprehensive quality assurance. + +Automatic Pipelines. We first employ AutoShot [94] to cut videos into several segments, and then use VideoChat2 [31] to caption them. For auditory content, we adopt whisper-large-v3 [51] to obtain speech transcripts, and DeepL [11] to translate them into English. The video captions and translated subtitles are then put into GPT-4o [47] to generate 6 pseudo QA pairs with a multi-step reasoning process. To ensure the quality of reasoning, we specify 7 multi-step reasoning types for narrative videos as: + +- Event prediction: Forecast subsequent events in the video timeline. +- Hypothetical reasoning: Deduce plausible scenario outcomes from stated premises. +- Event attribution: Determine causal origins or underlying motivations of video events. +- Implicit inference: Extract unstated temporal, emotional, or relational context from visual cues. +- Logical linkage: Establish event-mediated connections between visual/narrative elements. +- Information synopsis: Condense critical information across multimodal inputs. +- Counting problems: Quantify state changes through arithmetic/combinatorial analysis. + +Expert-guided Rewriting. We recruit and train 67 graduate students to generate 8-10 high-quality QA pairs per video, providing raw video footage, translated subtitles, and GPT-generated pre-annotations that offer contextual hints without meeting final benchmark standards. Each QA pair must satisfy four rigorous criteria: (1) Non-synopsis questions require $\geq 2$ timestamped reasoning steps (start/end times documented); (2) Temporal distribution constraints ( $\leq 4$ questions from $0 - 15\mathrm{min}$ , $\geq 3$ from $15 - 40\mathrm{min}$ , $\geq 1$ from $40 - 120\mathrm{min}$ ); (3) Coverage of $\geq 5$ predefined reasoning taxonomies (from 7 categories) ensuring diversity; (4) Mandatory multimodal grounding where solutions demand both visual analysis (excluding subtitle-only answers) and explicit reasoning (beyond basic perception). + +Comprehensive Quality Assurance. To ensure annotations meet our stringent multimodal standards, we implement a rigorous verification protocol: 10 trained reviewers validate each annotation, with non-compliant entries returned to annotators for iterative revisions until fully compliant. Annotators and reviewers are exclusively recruited from top-tier universities to ensure academic rigor. We further enforce quality through dual safeguards: a systematic $5\%$ random sampling audit across both annotation and review stages, coupled with full documentation of protocols, annotation guidelines, and quality assessment criteria in supplementary materials. + +# 3.3. Multi-Phase Evaluation Pipeline + +We benchmark VLMs and LLMs through a comprehensive multi-phase evaluation pipeline, which compares predictions against ground-truth annotations both at the process and outcome level. For the outcome-level evaluation stage, we adopt a multiple-choice question (MCQ) format, where false options are generated by DeepSeek-V3 [38] using carefully-designed prompts based on human-annotated answers. For the process-level stage, we propose the open-ended rating to fully evaluate the model's multi-step reasoning capability. Specifically, we adopt an LLM to evaluate the overall quality of the whole reasoning process through four 0-10 scores: logical coherence $(40\%)$ , similarity to ground truth $(40\%)$ , excluded for event prediction and hypothetical reasoning tasks), factual accuracy $(10\%)$ , and clarity $(10\%)$ . DeepSeek-V3 [38] serves as the judge due to its optimal balance between cost and human alignment (Section 4). For the event prediction and hypothetical reasoning tasks, we argue that there is no ground-truth reasoning steps since there might be multiple possible predictions, hence, we remove the similarity score and compute the final score with an 8:1:1 weight. All metrics are normalized to 0-100 scales, and we report the rankings by computing average scores across all metrics. + +![](images/4f6f89f22f1d9c212edacb9064aad1049e97b6d604835cc51b939f3aee0f2745.jpg) +Figure 4. Human preference alignment results. For each plot, we show the win ratios of three different tested LLMs evaluated by human experts and VRBench. We then fit them with a straight line and quantify the correlation by calculating the Spearman correlation coefficient. + +LLM Evaluation Support. For text-only LLMs, we convert videos to text inputs using Qwen2.5-72B-Instruct [83] to synthesize video captions and subtitles into fine-grained summaries. The process involves: 1) Dividing inputs into 5-minute chunks for duplicate removal and abstraction; 2) Merging chunk-level abstracts into coherent summaries with detailed audio-visual descriptions. These summaries enable LLM evaluations while addressing context window limits for VLMs. + +# 4. Experiments + +We test a number of both open-source and proprietary VLMs (e.g. Qwen2-VL [68], InternVL2.5 [7], InternVideo2.5 [73], GPT-4o [47], Gemini-2.0-Pro [61], etc.) and LLMs (DeepSeek-R1 [18], QwQ [65], OpenAI o1 [48], Claude-3.7-Sonnet [59], etc.) with a two-stage evaluation on VRBench. The configuration of each evaluated model is detailed in the supplementary material. + +Evaluation Protocols. Our protocol involves two phases. Given a question, 1) models produce multi-step reasoning using Chain-of-Thought [75] prompts. Then 2) models choose final answers from multiple-choice options. For answer extraction, final answers are parsed from response endings to mitigate verbose option analysis, following [31]. While reasoning process evaluations use judge LLMs with structured prompts and example-guided rating extraction. All evaluation prompts and implementation details are included in the supplementary materials. + +# 4.1. Main Results + +Table 2 shows the CoT evaluation results of LLMs and VLMs on VRBench. Among LLMs, we found the proprietary model Gemini-2.0-Flash-Thinking [12] achieves the optimal performance while other reasoning models such as OpenAI o1-preview [48] and DeepSeek-R1 [18] also demonstrate strong results. It is noted that o1-preview [48] + +
ModelOverallResults by MetricResults by Taxonomy
MCQ-OOE-PEACPHRIIISEPLL
LLMs
Proprietary Models
GPT-4o [47]55.4963.8747.1154.4132.6369.0361.4771.8366.3766.19
Claude-3.7-Sonnet [59]58.0762.9153.2355.8735.1169.8363.2972.6167.9167.89
o1-preview [48]60.1468.4751.8156.9735.8770.5164.8173.1368.2769.41
Gemini-2.0-Flash-Thinking [61]60.9767.5354.4158.6138.1172.0967.5774.1970.4371.13
Open-Source Models
QwQ-32B-preview [64]35.9027.5144.2934.3134.4146.2939.6727.9742.3940.61
InternLM3-8B-Instruct [3]47.8150.3145.3145.4734.8760.3351.7755.4756.9756.81
Qwen2.5-7B-Instruct [83]48.2952.6143.9746.9335.1761.8351.5754.8756.6357.03
Llama3.3-70B-Instruct [41]49.8452.5947.0947.6338.0763.8754.0354.8359.4756.97
QwQ-32B [65]52.5256.0149.0349.5336.0362.0154.0352.0358.0159.03
Qwen2.5-72B-Instruct [83]53.5160.4946.5351.0336.0167.0358.5367.8361.9363.03
DeepSeek-V3 [38]56.0664.7947.3354.0336.5769.9761.8369.5365.5365.43
DeepSeek-R1 [18]57.1364.1950.0756.1338.9172.8164.1368.0168.5367.93
VLMs
Proprietary Models
Claude-3.7-Sonnet [59]68.1580.0956.2163.1132.9772.6371.1375.3771.3370.27
GPT-4o [47]68.6881.2356.1366.6136.5176.6770.4778.0372.0373.17
Gemini-2.0-Pro [61]74.6183.2965.9371.0965.2181.0175.7387.1377.9375.89
Open-Source Models
DeepSeek-VL2 [79]31.5033.2729.7325.9322.4135.7331.7330.0133.2929.57
H2OVL Mississippi-2B [17]47.1552.3341.9735.1732.3751.7141.4160.0349.9741.91
Phi-3.5-Vision [1]48.5258.0339.0131.5328.0345.0337.0368.0344.0336.03
LongVA-7B [91]50.1467.8132.4727.6125.0738.6933.7776.9138.2732.07
InternVL2.5-8B [7]50.4169.3131.5126.6326.3137.9931.2185.9737.3730.69
MiMo-VL-7B-RL [60]50.4863.3937.5747.3134.5963.8353.2374.6160.3356.97
VideoChat-Flash-7B [32]50.8272.0129.6324.6921.4139.3729.1779.3734.8728.07
InternVideo2.5 [73]51.9475.6328.2523.8124.0934.5127.0784.5733.6126.81
LongVA-7B-DPO [91]52.3667.9136.8130.7327.1745.0938.2379.3943.6737.03
Qwen2-VL-7B [68]54.0872.0136.1530.4926.6146.7133.6985.2743.3135.29
Aria [30]54.5572.9736.1330.1929.2344.2334.6386.2344.9934.57
Qwen2.5-VL-7B [2]56.5269.6143.4337.0733.1754.2742.2783.9150.1742.87
Keye-VL-8B-Preview [63]60.4464.4156.4762.5340.3773.6167.9970.1371.8371.03
Qwen2.5-VL-72B [2]61.7166.8556.5751.8746.1367.1354.1390.0463.6760.77
Kimi-VL-A3B-Thinking-2506 [62]61.8261.6761.9764.5347.9171.3769.5774.4771.2371.47
InternVL2.5-78B [7]62.3176.6148.0143.7738.6758.2146.1387.5354.4347.57
+ +Table 2. Evaluation results on VRBench across two evaluation metrics (outcome-level MCQ, process-level open-ended evaluation) and seven QA taxonomies (event attribution, counting problems, hypothetical reasoning, implicit inferences, information synopsis, event prediction, logical linkage). Bold values indicate the best results among all models, and underlined values are best results of LLMs. + +achieves the best outcome-level MCQ accuracy of $68.47\%$ . For the VLMs, Gemini-2.0-Pro [61] achieves $74.61\%$ overall accuracy, surpassing all components by at least $5.93\%$ . GPT-4o [47] and Claude-3.7-Sonnet [59] also show competitive results, which demonstrate $68.68\%$ and $68.15\%$ overall performance, respectively. For open-source models, InternVL2.5-78B [7], emerges as the best non-proprietary + +VLMs with a $62.31\%$ overall accuracy. + +We then delve into the specific results of each metric and taxonomy. For results in each evaluation stage, we find that most of LLMs and VLMs are capable of achieving high MCQ accuracy, yet compared to LLMs, VLMs struggle to demonstrate their reasoning steps and exhibit lower reasoning ratings. Through results across taxonomy, we found + +that most of the large models are not proficient in counting problems since they require fine-grained visual perception, in contrast with other tasks. Since the perception of LLMs fully depends on the video summary and some VLMs only support a small number of frame inputs, they are unable to correctly perceive the original frames containing the target elements, resulting in the MCQ accuracy close to random guessing and low reasoning process ratings. + +# 4.2. Ablations and Analysis + +We then further discuss the observations from the evaluation results and highlight some insights on obtaining higher multi-step reasoning performance. + +Does the evaluation of VRBench align with Human Preference? In contrast to MCQs that compute accuracy via a deterministic method, the correctness of process-level open-ended evaluation highly depends on the judgment of LLM's performance, which may include hallucinations that could influence the evaluation reliability. To this end, we select a subset of 30 videos with 300 questions to perform process-level human evaluation, aiming to quantify the correlation between human and LLM assessment results. + +Specifically, the human annotators are asked to give process-level open-ended ratings following the same requirement for the judging LLM, i.e., four separated ratings are needed for each question to evaluate the logical coherence, similarity to ground-truth, factual accuracy, and expression clarity. We then compute the win ratio of pairwise comparison for each tested model following [24], where a model scores 1 if its rating is higher than its current opponent, 0 if lower, and 0.5 for a tie. Finally, we calculate the average win ratio of each model and the Spearman correlation coefficients $(\rho)$ of the win ratios evaluated by LLMs and human annotators. + +We first investigate the human preference alignment of several LLMs. We select GPT-4o [47], DeepSeek-V3 [38], Qwen2.5-7B [83], and Qwen2.5-72B [83] as evaluation models and give them the same rating prompt for a fair comparison. As shown in Figure 4, Qwen2.5 shows low human preference alignment compared with DeepSeek-V3 and GPT-4o, which both have correlation coefficients greater than 0.8. Though the human alignment of GPT-4o is slightly better than DeepSeek-V3, we observe the evaluation cost of GPT-4o is approximately 10 times that of DeepSeek, which imposes a significant burden on the benchmark users. Hence, we adopt DeepSeek-V3 as the judging model for its optimal trade-off between performance and cost. + +We also probe the evaluation correctness of each open-ended sub-metric. Following the same procedure, we compute the win ratio and correlation coefficient of DeepSeekV3 for each sub-metric, and the results are shown in Figure 4 (b-d). Results show the coefficients in all metrics are + +higher than 0.8, indicating a strong positive monotonic relationship between LLM and human evaluation results. Another alternative is to instruct the LLM to compute the final score based on the given weights of each sub-metric, which is denoted as DeepSeek w/o sep in Figure 4. Results show that adding this additional computing process diminishes the robustness of evaluation LLM, thus in practice, we ask the model to output all sub-metric rating separately and automatically compute the final score. + +The role of System-2 thinking. o1-like LLMs, also denoted as system-2 models, have emerged with superior reasoning abilities in multi-disciplinary scenarios such as science and math. We further explore these models' performance on the narrative video benchmark. We both assess the role of open-source and proprietary o1-models, as well as the gap compared with their previous system-1 version. For proprietaries, we compare OpenAI o1-preview [48] with GPT-4o [47], and results show that the system-2 version o1 outperforms GPT-4o with $4.30\%$ overall accuracy. For open-source model DeepSeek-V3 [38] and R1 [11], we observe that DeepSeek-R1 achieves higher scores both on multiple-choice questions and open-ended evaluations. It is noted that compared with Qwen2.5-72B-Instruct [82], QwQ-32B [65] gets higher reasoning ratings yet lower outcome-level MCQ results. This shows that though some system-2 models are more proficient in multi-step thinking, their ability to converge lengthy thinking processes into correct answers still needs improvement. Recently, some o1-like training strategies have also been utilized by some VLMs, such as LongVA-7B-DPO [91], a model that utilizes direct preference optimization [52] for long video understanding. It achieves an overall performance of $52.36\%$ and surpasses its vanilla version, suggesting that training VLMs with different optimization methods is feasible and that developing system-2 VLM models is crucial for tackling multi-step reasoning questions in narrative videos. + +Impact of Model Size. It is commonly believed that models with large scales are often accompanied by better perception and reasoning capabilities. To verify such claims, we conduct experiments on the models with different scales, such as Qwen2.5-7B [83] and Qwen2.5-72B, Qwen2.5-VL [2] with 7B and 72B parameters, and InternVL2.5-8B [7] and InternVL2.5-78B. As shown in Table 2, models with more parameters obtain higher overall accuracy (48.29% vs 53.51% for Qwen2.5, 56.52% vs 61.71% for Qwen2.5-VL, and 50.41% vs 62.31% for InternVL2.5), and the performance gap on process-level metrics is more apparent. This indicates that large-scale models are more likely to achieve advanced reasoning capabilities. However, we also noticed that models specifically trained on the reasoning corpus like QwQ-32B [65], can achieve comparable performance compared with its 72B Qwen2.5 counterpart (52.52% vs. 53.51% in overall accuracy) with a smaller + +parameter amount. This suggests that training small-sized models with various preference optimization strategies and reasoning-centric corpora might improve their overall reasoning capability. + +Long Context Helps. We observe a substantial performance gap between models with and without long context support. As shown in Table 2, models with long frame input tend to achieve higher overall accuracy. For example, Gemini-2.0-Pro [61] is capable of inference with 0.5 fps and supports large number frame inputs when answering questions of videos over 1 hour, and it achieves the optimal performance compared with all other fixed-length models. For the open-source models, Qwen2.5-VL-7B [2] and KimiVL-A3B-Thinking-2506 [62] with 256 frame input achieve $61.71\%$ and $61.82\%$ overall accuracy, which is $11.30\%$ and $11.41\%$ higher than the 8B InternVL-2.5 [7] model with 64 frames input. Since the average length of VRBench is 1.61 hours, the necessity of long context support is becoming more pronounced, so that models can perceive more plot elements that help the analysis of long video content. + +LLMs vs. VLMs. Since tested LLMs are only provided with the automatically-generated video summary, yet VLMs are capable of directly perceiving the raw visual content, it is unfair to directly compare the numerical results of LLMs and VLMs. However, there are still some noteworthy phenomena that can be discussed. First, VLMs with coarse-grained visual input (e.g., 4 frames for H2OVL Mississippi-2B [17] and DeepSeek-VL2 [79]) perform significantly worse than several system-2 LLMs like DeepSeek-R1 [18] and Claude-3.7-Sonnet [59]. This indicates that randomly selecting a small number of frames is not enough to tackle narrative-based reasoning tasks that require long-range perception. Furthermore, VLMs with fine-grained visual perception, such as Gemini-2.0-Pro with a 0.5 fps perform better than top-tier LLMs on both process-level and outcome-level metrics, showing the inevitability of detailed visual content input for tackling questions in VRBench. + +# 4.3. Test-Time Scaling Exploration + +We here probe models' performance when scaling test-time compute cost. Since we tend to prompt the models with a Chain-of-Thought template when tackling complex reasoning tasks, it is observed that system-2 models are capable of dramatically expanding the inference budget to improve their performance. To this end, we set a series of token limitations on these models and require the models to think with different instructions. Specifically, we select QwQ-32B [65] as the experimental LLM and Qwen2-VL-7B [82] as the testing VLM on a subset with 300 videos and 2,403 questions. For each model, several parallel experiments are conducted with increasing maximum token limits from 256 to 2048. We instruct the models to output the reasoning process as much as possible for large token limitations, and + +![](images/9f214471ff20dc28023bbddcd0fd23f4621a7b12fa38cd7c031b608e764a1971.jpg) +Figure 5. Test-Time Scaling Results. We report the average accuracy of outcome-level MCQ and process-level open-ended ratings. + +let them think concisely when the token limit is low. Since we aim to investigate the quality of models' initial reasoning process and their capability to generate the right answers through thinking, we report the overall scores that both evaluate the outcome accuracy and process quality. Results are shown in Figure 5, where QwQ shows a remarkable rating boost from $48.91\%$ to $61.34\%$ when setting a large token limit. On the contrary, the small-sized system-1 model Qwen2-VL-7B performs even worse when using a large token number limitation and instructing the model to output more thinking process. The model tends to output more ambiguous outputs that lead to the wrong answer. This observation suggests that models with large parameter scales and system-2 capabilities benefit from the test-time scaling strategy. It also provides the insight that developing more test-time scaling approaches that guide the model to generate long reasoning traces could be a feasible and promising way to unlock more potential capabilities of system-2 models, thereby benefiting complex tasks that require multi-step reasoning. + +# 5. Conclusion + +This paper introduces VRBench, a comprehensive long-narrative video benchmark for evaluating multi-step reasoning. Through manually filtered narrative videos, high-quality stepwise annotations, and a multi-phase evaluation pipeline, VRBench distinguishes itself from other existing reasoning benchmarks and shows the robust capability to evaluate LLMs and VLMs both from the process and outcome perspectives. By assessing and analyzing 31 frontier large models, we thoroughly demonstrate the detailed performance of current reasoning models across various reasoning questions and provide valuable insights towards constructing more advanced multi-step reasoning models. + +# 6. Acknowledgement. + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7, 8 +[3] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2, 6 +[4] Guo Chen, Yicheng Liu, Yifei Huang, Yuping He, Baoqi Pei, Jilan Xu, Yali Wang, Tong Lu, and Limin Wang. Cg-bench: Clue-grounded question answering benchmark for long video understanding. arXiv preprint arXiv:2412.12075, 2024.3 +[5] Wenhu Chen, Ming Yin, Max Ku, Pan Lu, Yixin Wan, Xueguang Ma, Jianyu Xu, Xinyi Wang, and Tony Xia. Theoremq: A theorem-driven question answering dataset. arXiv preprint arXiv:2305.12524, 2023. 3 +[6] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. In European Conference on Computer Vision, pages 179-195. Springer, 2024. 3 +[7] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang-wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 5, 6, 7, 8 +[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 3 +[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees GM Snoek, and Yuki M Asano. Tvbench: Redesigning video-language evaluation. arXiv preprint arXiv:2410.07752, 2024.3 +[10] Rocktim Jyoti Das, Simeon Emilov Hristov, Haonan Li, Dimitar Iliyanov Dimitrov, Ivan Koychev, and Preslav Nakov. *Exams-v: A multi-discipline multilingual multimodal exam benchmark for evaluating vision language models.* arXiv preprint arXiv:2403.10378, 2024. 3 + +This work is supported by the National Key R&D Program of China (No. 2022ZD0160101) and Jiangsu Frontier Technology Research and Development Program (No. BF2024076), +[11] DeepL. Deepl translate: The world's most accurate translator. https://www deepl.com/en/Translator, 2025.4,7 +[12] Google Deepmind. Gemini 2.0 flash thinking. https://deepmind.google/technologies/gemini/flash-thinking/, 2025. Accessed: 2025-01-21. 2, 5 +[13] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 3 +[14] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 2 +[15] Kehua Feng, Keyan Ding, Weijie Wang, Xiang Zhuang, Zeyuan Wang, Ming Qin, Yu Zhao, Jianhua Yao, Qiang Zhang, and Huajun Chen. Sciknoweval: Evaluating multilevel scientific knowledge of large language models. arXiv preprint arXiv:2406.09098, 2024. 3 +[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3 +[17] Shaikat Galib, Shanshan Wang, Guanshuo Xu, Pascal Pfeiffer, Ryan Chesler, Mark Landry, and Sri Satish Ambati. H2ovl-mississippi vision language models technical report. arXiv preprint arXiv:2410.13611, 2024. 6, 8 +[18] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2, 5, 6, 8 +[19] Songhao Han, Wei Huang, Hairong Shi, Le Zhuo, Xiu Su, Shifeng Zhang, Xu Zhou, Xiaojuan Qi, Yue Liao, and Si Liu. Videospeech: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection. arXiv preprint arXiv:2411.14794, 2024. 3 +[20] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multi-discipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3 +[21] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 3 +[22] Kairui Hu, Penghao Wu, Fanyi Pu, Wang Xiao, Yuanhan Zhang, Xiang Yue, Bo Li, and Ziwei Liu. Video-mmmu: Evaluating knowledge acquisition from multi-discipline professional videos. arXiv preprint arXiv:2501.13826, 2025. 3 +[23] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems, 36:62991-63010, 2023. 3 + +[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7 +[25] Zhen Huang, Zengzhi Wang, Shijie Xia, Xuefeng Li, Haoyang Zou, Ruijie Xu, Run-Ze Fan, Lyumanshan Ye, Ethan Chern, Yixin Ye, et al. Olympic: Benchmarking multi-discipline cognitive reasoning for superintelligent ai. Advances in Neural Information Processing Systems, 37: 19209-19253, 2025. 3 +[26] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livocodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024. 3 +[27] Zhihuan Jiang, Zhen Yang, Jinhao Chen, Zhengxiao Du, Weihan Wang, Bin Xu, Yuxiao Dong, and Jie Tang. Visscience: An extensive benchmark for evaluating k12 educational multi-modal scientific reasoning. arXiv preprint arXiv:2409.13730, 2024. 3 +[28] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 3 +[29] Bozheng Li, Yongliang Wu, Yi Lu, Jiashuo Yu, Licheng Tang, Jiawang Cao, Wenqing Zhu, Yuyang Sun, Jay Wu, and Wenbo Zhu. Veu-bench: Towards comprehensive understanding of video editing. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13671-13680, 2025. 3 +[30] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 2, 6 +[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 4, 5 +[32] Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, et al. Videochat-flash: Hierarchical compression for long-context video modeling. arXiv preprint arXiv:2501.00574, 2024. 2, 6 +[33] Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning. arXiv preprint arXiv:2504.06958, 2025. 2 +[34] Yunxin Li, Xinyu Chen, Baotian Hu, Longyue Wang, Haoyuan Shi, and Min Zhang. Videovista: A versatile benchmark for video understanding and reasoning. arXiv preprint arXiv:2406.11303, 2024. 3 +[35] Zekun Li, Xianjun Yang, Kyuri Choi, Wanrong Zhu, Ryan Hsieh, HyeonJung Kim, Jin Hyuk Lim, Sungyoung Ji, + +Byungju Lee, Xifeng Yan, et al. Mmsci: A multimodal multi-discipline dataset for phd-level scientific comprehension. In AI for Accelerated Materials Design-Vienna 2024, 2024.3 +[36] Zhenwen Liang, Kehan Guo, Gang Liu, Taicheng Guo, Yujun Zhou, Tianyu Yang, Jiajun Jiao, Renjie Pi, Jipeng Zhang, and Xiangliang Zhang. Scemqa: A scientific college entrance level multimodal question answering benchmark. arXiv preprint arXiv:2402.05138, 2024. 3 +[37] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 2 +[38] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 5, 6, 7 +[39] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 3 +[40] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. Et bench: Towards open-ended event-level video-language understanding. arXiv preprint arXiv:2409.18111, 2024. 3 +[41] Llama-3.3. Llama-3.3-70b-instruct. https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct, 2025.6 +[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 3 +[43] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 3 +[44] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3 +[45] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 3 +[46] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3 +[47] OpenAI. Hello gpt4-o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-13, 4, 5, 6, 7 +[48] OpenAI. Introducing openai o1. https://openai.com/o1/, 2024.2,5,6,7 +[49] OpenAI. Openai o3-mini. https://openai.com/index/openai-o3-mini/, 2025.2 + +[50] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36:42748-42761, 2023. 3 +[51] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 4 +[52] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 7 +[53] Ruchit Rawal, Khalid Saifullah, Miquel Farre, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. arXiv preprint arXiv:2405.08813, 2024. 2, 3 +[54] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3 +[55] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2 +[56] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 3 +[57] Liangtai Sun, Yang Han, Zihan Zhao, Da Ma, Zhennan Shen, Baocai Chen, Lu Chen, and Kai Yu. Scieval: A multi-level large language model evaluation benchmark for scientific research. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19053-19061, 2024. 3 +[58] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 3 +[59] Anthropic. Claude Team. Claude 3.7 sonnet. https://www.anthropic.com/claude/sonnet, 2025. 2, 5, 6, 8 +[60] Core Team, Zihao Yue, Zhenru Lin, Yifan Song, Weikun Wang, Shuhuai Ren, Shuhao Gu, Shicheng Li, Peidian Li, Liang Zhao, Lei Li, et al. Mimo-vl technical report. https://arxiv.org/abs/2506.03569, 2025. 6 +[61] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 6, 8 +[62] Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chen- + +zhuang Du, Chu Wei, et al. Kimi-vl technical report. arXiv preprint arXiv:2504.07491, 2025. 6, 8 +[63] Kwai Keye Team, Biao Yang, Bin Wen, Changyi Liu, Chenglong Chu, Chengru Song, Chongling Rao, Chuan Yi, Da Li, Dunju Zang, et al. Kwai keye-vl technical report. arXiv preprint arXiv:2507.01949, 2025. 6 +[64] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview/, 2024. Accessed: 2024-11-28. 6 +[65] Qwen Team. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github.io/blog/qwq-32b/, 2025. Accessed: 2025-3-6. 2, 5, 6, 7, 8 +[66] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamavol: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025. 2 +[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 3 +[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 6 +[69] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, et al. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024. 2 +[70] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 2, 3 +[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 2 +[72] Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 3 +[73] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 5, 6 +[74] Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, et al. Charxiv: Charting gaps in realistic chart understanding in multimodal llms. Advances in Neural Information Processing Systems, 37:113569-113697, 2025. 3 + +[75] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5 +[76] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 3 +[77] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 3 +[78] Yifan Wu, Lutao Yan, Leixian Shen, Yunhai Wang, Nan Tang, and Yuyu Luo. Chartinsights: Evaluating multimodal large language models for low-level chart question answering. arXiv preprint arXiv:2405.07001, 2024. 3 +[79] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 6, 8 +[80] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3 +[81] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-01: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 2 +[82] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report, 2024. URL: https://arxiv.org/abs/2407.10671, 2024. 7, 8 +[83] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5, 6, 7 +[84] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. Advances in Neural Information Processing Systems, 37:57240-57261, 2025. 3 +[85] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3 +[86] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3 +[87] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 3 + +[88] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 2 +[89] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, et al. Cmmmu: A chinese massive multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2401.11944, 2024. 3 +[90] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023.3 +[91] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2, 6, 7 +[92] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 2, 3 +[93] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 3 +[94] Wentao Zhu, Yufang Huang, Xiufeng Xie, Wenxian Liu, Jincan Deng, Debing Zhang, Zhangyang Wang, and Ji Liu. Autoshot: A short video dataset and state-of-the-art shot boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2238-2247, 2023. 4 \ No newline at end of file diff --git a/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/images.zip b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba12c8fc2c3a9477d922a8e5acc23bcbf187740b --- /dev/null +++ b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b75211cd565e6487d57488bae75fab617d261584bd48bd980f5f42e02a6d8602 +size 739942 diff --git a/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/layout.json b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f70911a10889e810e0e303bb7b7aea9a445a77ed --- /dev/null +++ b/2025/VRBench_ A Benchmark for Multi-Step Reasoning in Long Narrative Videos/layout.json @@ -0,0 +1,8356 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 103, + 542, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 103, + 542, + 121 + ], + "spans": [ + { + "bbox": [ + 68, + 103, + 542, + 121 + ], + "type": "text", + "content": "VRBench: A Benchmark for Multi-Step Reasoning in Long Narrative Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "spans": [ + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "text", + "content": "Jiashuo Yu" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "text", + "content": " Yue Wu" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "text", + "content": " Meng Chu" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "text", + "content": " Zhifei Ren" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "text", + "content": " Zizheng Huang" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{2,3,1*}" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "text", + "content": " Pei Chu" + }, + { + "bbox": [ + 81, + 141, + 531, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "spans": [ + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "text", + "content": "Ruijie Zhang" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "text", + "content": " Yinan He" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "text", + "content": " Qirui Li" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "text", + "content": " Songze Li" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "text", + "content": " Zhenxiang Li" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "text", + "content": " Zhongying Tu" + }, + { + "bbox": [ + 95, + 158, + 529, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 171, + 502, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 171, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 121, + 171, + 502, + 186 + ], + "type": "text", + "content": "Conghui He1 Yu Qiao1 Yali Wang4,1⊗ Yi Wang1,3⊗ Limin Wang2,1⊗" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "type": "text", + "content": "Nanjing University " + }, + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 185, + 547, + 199 + ], + "type": "text", + "content": "Shanghai Innovation Institute" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 199, + 491, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 199, + 491, + 213 + ], + "spans": [ + { + "bbox": [ + 121, + 199, + 491, + 213 + ], + "type": "text", + "content": "4Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 242, + 215, + 367, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 215, + 367, + 227 + ], + "spans": [ + { + "bbox": [ + 242, + 215, + 367, + 227 + ], + "type": "text", + "content": "VRBench.github.io" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 70, + 255, + 541, + 441 + ], + "blocks": [ + { + "bbox": [ + 70, + 255, + 541, + 441 + ], + "lines": [ + { + "bbox": [ + 70, + 255, + 541, + 441 + ], + "spans": [ + { + "bbox": [ + 70, + 255, + 541, + 441 + ], + "type": "image", + "image_path": "9b75860174e48a322566b434e2eea6279429859803f92425fc8b9f779f5ad7b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 451, + 555, + 517 + ], + "lines": [ + { + "bbox": [ + 54, + 451, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 451, + 555, + 517 + ], + "type": "text", + "content": "Figure 1. Overview of VRBench. We present VRBench, a long narrative video benchmark for multi-step reasoning. VRBench includes 960 manual-filtered narrative videos, covering 8 languages and 7 video categories that are suitable for reasoning about temporal relations. We also provide high-quality stepwise annotations for reasoning, which are labeled and reviewed by human experts. Each video incorporates 8-10 complex question-answer pairs, a multi-step reasoning chain, and fine-grained timestamps. To fully evaluate the capability of models in multi-step reasoning, we propose a multi-phase evaluation pipeline that assesses model results both from the process and outcome levels. Our VRBench is the first video reasoning benchmark that supports both multi-step annotation and evaluation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 528, + 200, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 528, + 200, + 540 + ], + "spans": [ + { + "bbox": [ + 151, + 528, + 200, + 540 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "type": "text", + "content": "We present VRBench, the first long narrative video benchmark crafted for evaluating large models' multi-step reasoning capabilities, addressing limitations in existing evaluations that overlook temporal reasoning and procedural validity. It comprises 960 long videos (with an average duration of 1.6 hours), along with 8,243 human-labeled multi-step question-answering pairs and 25,106 reasoning steps with timestamps. These videos are curated via a multistage filtering process including expert inter-rater reviewing to prioritize plot coherence. We develop a human-AI collaborative framework that generates coherent reasoning chains, each requiring multiple temporally grounded steps," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 649 + ], + "type": "text", + "content": "spanning seven types (e.g., event attribution, implicit inference). VRBench designs a multi-phase evaluation pipeline that assesses models at both the outcome and process levels. Apart from the MCQs for the final results, we propose a progress-level LLM-guided scoring metric to evaluate the quality of the reasoning chain from multiple dimensions comprehensively. Through extensive evaluations of 12 LLMs and 19 VLMs on VRBench, we undertake a thorough analysis and provide valuable insights that advance the field of multi-step reasoning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 395, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "The rapid evolution of vision language models (VLMs) has heightened the need for benchmarks that rigorously eval" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 222, + 713 + ], + "type": "text", + "content": "*equal contributions. corresponding authors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "21655" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "uate complex reasoning capabilities. While existing standards like GSM8K [8] focus on domain-specific knowledge in mathematics and science, they neglect a critical reasoning dimension: contextual analysis in visual narrative content. Real-world applications increasingly demand temporal reasoning across interconnected elements: tracking character dynamics in films, interpreting gameplay strategies, or understanding cause-and-effect chains in documentaries. This capability gap persists because current video benchmarks [16, 53, 70, 77] primarily assess single-step perception rather than sustained reasoning processes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 350 + ], + "type": "text", + "content": "We find three fundamental limitations in current evaluation paradigms: (1) an overemphasis on domain expertise rather than plot-driven reasoning, (2) the lack of temporally-grounded reasoning chains in annotations, and (3) Outcome-focused metrics that ignore procedural validity. To address these challenges, we present VRBench (as in Figure 1), the first benchmark specifically designed for multi-step reasoning in long-form narrative videos. Its construction involves the collection of narrative videos, a human-in-the-loop reasoning process method for annotation, and a multi-phase evaluation pipeline considering both procedure and outcome." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 352, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 555 + ], + "type": "text", + "content": "In data construction, VRBench aggregates 960 meticulously selected videos (1.6h average duration) across seven narrative categories (e.g., movies, sports, travelogues), sourced through a multi-stage filtering process. Our curation pipeline combines automated retrieval with expert validation (inter-rater reliability " + }, + { + "bbox": [ + 55, + 352, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\rho = 0.82" + }, + { + "bbox": [ + 55, + 352, + 296, + 555 + ], + "type": "text", + "content": "), prioritizing plot coherence over domain-specific knowledge. This contrasts with existing benchmarks like MMVU [92] that emphasize disciplinary competence. To conduct multi-step reasoning chain annotation, we develop a human-AI collaborative approach, generating 8-10 QA pairs per video with explicit temporal grounding. Each question requires no less than 2 reasoning steps annotated with precise timestamps (Figure 3), validated through iterative expert review (95% inter-annotator agreement). The taxonomy spans seven reasoning types from event prediction to implicit inference, ensuring comprehensive coverage of narrative analysis skills." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "When evaluating models on VRBench, we propose a multi-phase evaluation pipeline that combines outcome verification with process analysis. Beyond conventional multiple-choice accuracy that evaluates the outcome-level performance, we further introduce the process-level metric LLM-guided scoring to evaluate the overall reasoning process quality. This dual approach reveals critical insights - for instance, while GPT-4o achieves " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "81.25\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " outcome accuracy, its process rating scores drop to " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "56.11\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", exposing reasoning fragility. Evaluations of 31 state-of-the-art large models reveal proprietary VLMs with long-context support outperform text-only LLMs by " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "13.82\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " absolute, emphasizing the importance of dense visual ground" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 548, + 106 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 548, + 106 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 548, + 106 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 548, + 106 + ], + "type": "image", + "image_path": "f51bb77bf0fc108a977640f5339c2cf941c34fb2de8defdcae4c944c9f6911bf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 280 + ], + "type": "text", + "content": "Figure 2. An example of annotation in VRBench. For each question, VRBench provides the question-answer pair, multi-step reasoning chain, question type, and the start-to-end timestamps of the entire question as well as each reasoning step." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 109, + 541, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 109, + 541, + 129 + ], + "spans": [ + { + "bbox": [ + 326, + 109, + 541, + 129 + ], + "type": "text", + "content": "Questions: Why did the male protagonist arrange for his friend to rob female protagonist? [00:02:20->00:17:42]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 133, + 541, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 133, + 541, + 165 + ], + "spans": [ + { + "bbox": [ + 326, + 133, + 541, + 165 + ], + "type": "text", + "content": "Step1: When the heroine first caught the hero and his friends, she acted very \nrighteously, and the hero fell in love with her. [00:02:20->00:05:05] \nStep2: The female protagonist chats with the male protagonist and reveals that her \nfavorite type is a manly man. [00:05:06->00:07:58]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 164, + 541, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 164, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 326, + 164, + 541, + 178 + ], + "type": "text", + "content": "Step3: The male protagonist asks his friends to plan how to show manliness and justice. [00:13:23~>00:16:01]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 178, + 531, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 178, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 326, + 178, + 531, + 194 + ], + "type": "text", + "content": "Step4: The male protagonist arranges his friends to rob the female protagonist in an attempt to win her love. [00:16:09->00:17:42]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 199, + 487, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 199, + 487, + 209 + ], + "spans": [ + { + "bbox": [ + 326, + 199, + 487, + 209 + ], + "type": "text", + "content": "Answer: Because he wants to get the woman's love." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 214, + 433, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 214, + 433, + 224 + ], + "spans": [ + { + "bbox": [ + 326, + 214, + 433, + 224 + ], + "type": "text", + "content": "Question Type: Event Attribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 396 + ], + "type": "text", + "content": "ing. System-2 optimization strategies yield disproportionate gains in process metrics, and despite parameter parity, open-source VLMs lag behind proprietary counterparts by " + }, + { + "bbox": [ + 313, + 300, + 555, + 396 + ], + "type": "inline_equation", + "content": "12.30\\%" + }, + { + "bbox": [ + 313, + 300, + 555, + 396 + ], + "type": "text", + "content": ", suggesting architectural limitations beyond scale. Our ablation studies show strong alignment between human and LLM evaluations and quantify the impact of test-time scaling—doubling context windows improves QwQ-32B's [65] accuracy by " + }, + { + "bbox": [ + 313, + 300, + 555, + 396 + ], + "type": "inline_equation", + "content": "12.43\\%" + }, + { + "bbox": [ + 313, + 300, + 555, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 491 + ], + "type": "text", + "content": "VRBench sets a new standard for evaluating narrative reasoning, offering insights distinct from knowledge-centric benchmarks. By separating domain expertise from contextual analysis, our work aids in developing models capable of sustained, human-like understanding in real-world video content. The entire suite, including annotation protocols and evaluation tools, has been fully open-sourced to advance reasoning research." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 502, + 402, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 402, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 402, + 514 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "Reasoning Large Models. As the capabilities of LLMs continue to expand, frontier models have demonstrated significant potential in addressing high-order reasoning tasks. OpenAI pioneered this advancement with the release of o1 [48], an LLM capable of sophisticated reasoning tasks. Subsequently, a series of proprietary [12, 49, 59] and open-source [3, 18, 65] LLMs designed for advanced reasoning have emerged. Furthermore, researchers have shifted their focus towards developing VLMs possessing similar higher-order reasoning capabilities. Previously, some VLMs [2, 7, 30, 32, 37, 71, 73, 88, 91] to a certain extent exhibited good reasoning proficiency due to their proficiency in tackling long-sequence context. More recently, a series of VLMs [14, 66, 69, 81] trained on data incorporating higher-order Chains-of-Thought in RL-based algorithms [33, 52, 55] have been developed, enabling them" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21656" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 332 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 332 + ], + "type": "table", + "html": "
Dataset#Size#QA pairs#Dur.(s)QA typesData SourceAnno StepEval StepEval TargetAnno TypeClueMultilingual
MMLU [21]14,07914,079/MCQMulti-DiscSingleSingleLLMMxx
MMLU-Pro [72]12,03212,032/MCQMulti-DiscSingleSingleLLMA+Mxx
LiveCodeBench [26]511511/MCQCodeMultiSingleLLMMxx
SciEval [57]15,90115,901/MCQ+OpenScienceSingleSingleLLMMxx
GSM8k [8]8,7928,792/MCQ+OpenMathMultiMultiLLMMxx
C-Eval [23]12,34212,342/MCQMulti-DiscSingleSingleLLMMxx
ScienceQA [54]10,33221,208/MCQScienceMultiSingleVLMMxx
VisScience [27]3,0003,000/MCQ+OpenScienceSingleSingleVLMA+Mxx
MMMU [85]11,50011,500/MCQ+OpenMulti-DiscSingleSingleVLMMxx
MMMU-Pro [86]3,4603,460/MCQMulti-DiscSingleSingleVLMA+Mxx
MathVista [42]6,1416,141/MCQ+OpenMathMultiSingleVLMMxx
MathVision [67]3,0403,040/MCQ+OpenMathMultiSingleVLMMxx
CharXiv [74]2,32311,615/OpenMulti-DiscSingleSingleVLMMxx
OlympicArena [25]7,57111,163/MCQ+OpenMulti-DiscMultiMultiVLMMxx
MVBench [31]3,6414,00016.0MCQOpen-DomainSingleSingleVLMA+Mxx
EgoSchema [43]5,0635,063180.0MCQEgocentricSingleSingleVLMAxx
LongVideoBench [77]3,7636,678473.0MCQOpen-DomainSingleSingleVLMMxx
LVBench [70]1031,5494,101MCQOpen-DomainSingleSingleVLMMxx
CGBench [4]1,21912,1291624.4MCQ+OpenOpen-DomainMultiMultiVLMM✓x
MLVU [93]1,7303,102930MCQNarrativeSingleSingleVLMMxx
VideoMME [16]9002,7001017.9MCQOpen-DomainSingleSingleVLMMx✓
Video-MMMU [22]300900506.2MCQMulti-DiscSingleSingleVLMMxx
MMWorld [20]1,9106,627108.0MCQMulti-DiscSingleSingleVLMA+Mxx
MMVU [92]1,5293,00051.4MCQ+OpenMulti-DiscMultiSingleVLMMxx
VRBench (Ours)9608,2435,796.0MCQ+OpenNarrativeMultiMultiLLM, VLMM✓✓
", + "image_path": "6fd631ce4c824d5abdc0cb42c11fbe35ab3a9fe71e79a205384c69826ad97922.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 340, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 555, + 374 + ], + "type": "text", + "content": "Table 1. Comparison between VRBench and existing benchmarks. #Size is the number of text/images/videos, #Dur. means the average video duration, A and M indicate automatic and manual annotation type, respectively, and multi-disc denotes multi-disciplinary data source. Clue means clue-grounded annotation. Multilingual requires the number of data source languages to be greater than 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 395, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 294, + 418 + ], + "type": "text", + "content": "to solve multidisciplinary, knowledge-intensive problems through a combination of fast and slow thinking processes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 438, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 295, + 713 + ], + "type": "text", + "content": "Reasoning Benchmarks. With the rapid advancement of LLMs and VLMs in reasoning capability, numerous reasoning benchmarks have been developed in text and image-text modalities to facilitate comprehensive evaluation. For the textual domain, several knowledge-driven benchmarks [5, 15, 21, 23, 57, 72] have been derived from exam and textbook data sources, and comprise expert-level questions that require multi-step reasoning. Meanwhile, some multimodal benchmarks also emerge to introduce multi-discipline tasks based on charts [44, 74, 78], plots [45], exam [10, 25, 36, 85, 89], or expert-level questions printed on the static images [27, 35, 54, 86]. It is noted that even in the image and text domains, benchmarks with process-level annotations [8, 87] are relatively scarce. For video understanding, evaluation also gradually shifts from short video clip perception [6, 9, 28, 29, 31, 39, 46, 50, 80] to long-form long video understanding [13, 16, 40, 43, 53, 56, 58, 70, 77, 90] and single-step reasoning. [4, 16, 19, 34, 76, 84]. More recently, some advances [13, 20, 92] propose several new video reasoning tasks in multi-disciplinary scenarios like healthcare, engineering, and science. Different from previous works, VRBench is the first narrative video-based benchmark that is purely used for multi-step reasoning eval" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 431 + ], + "type": "text", + "content": "uation. Table 1 further presents the detailed statistics of our dataset and distinguishes the difference between VRBench and existing text, image, and video reasoning benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 441, + 389, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 389, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 389, + 452 + ], + "type": "text", + "content": "3. Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 460, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 520 + ], + "type": "text", + "content": "We present VRBench, a comprehensive multi-step reasoning benchmark consisting of a collection of long, multilingual, and narrative videos with corresponding question-answer pairs. Figure 1 gives the overall pipeline of the benchmark construction, and we illustrate it as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 526, + 409, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 409, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 409, + 538 + ], + "type": "text", + "content": "3.1. Video Curation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 604 + ], + "type": "text", + "content": "We collect long narrative videos from YouTube, yielding an initial pool of over 10,000 public videos. Considering plausible reasoning in steps requires videos with rich and coherent plots, we source long-form footage using a cherry-picked tag set and a comprehensive criterion for narration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Plot-related Tags for Queries. We employ a manually curated set of 7 semantic tags (e.g., Film & Animation, Sport, Travel & Events) to retrieve videos with strong narrative potential. This tag set is developed through iterative validation by domain experts, explicitly excluding non-narrative categories such as news broadcasts and lecture recordings. Our analysis revealed that these excluded categories exhibit limited visual-semantic dynamics (e.g., static" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21657" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 72, + 208, + 194 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 208, + 194 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 208, + 194 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 208, + 194 + ], + "type": "image", + "image_path": "01a1f4f86ae78d2f1f2b5f4ca2438e27ce9a62d975e3b857e4f83723525d0e9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 350, + 194 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 350, + 194 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 350, + 194 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 350, + 194 + ], + "type": "image", + "image_path": "fe11c9a4213c0bbaa61a589de2054c54a19a7b6d7af1151be4d9c7276eb11ff3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 372, + 72, + 524, + 194 + ], + "blocks": [ + { + "bbox": [ + 372, + 72, + 524, + 194 + ], + "lines": [ + { + "bbox": [ + 372, + 72, + 524, + 194 + ], + "spans": [ + { + "bbox": [ + 372, + 72, + 524, + 194 + ], + "type": "image", + "image_path": "629ab1b1b4239bb663c922988fda19806527b64f756876e3d711adb4689a9af4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 200, + 211, + 322 + ], + "blocks": [ + { + "bbox": [ + 82, + 200, + 211, + 322 + ], + "lines": [ + { + "bbox": [ + 82, + 200, + 211, + 322 + ], + "spans": [ + { + "bbox": [ + 82, + 200, + 211, + 322 + ], + "type": "image", + "image_path": "ddad08f6901c1c40f9553f14165c6c9dfd94deb9e73630dda394e2e47e0dd585.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 354 + ], + "type": "text", + "content": "Figure 3. Statistics of VRBench. We provide the detailed distribution of videos and annotations of VRBench, including video languages and durations, steps, types, and temporal duration of questions, as well as the token numbers of answers and the reasoning process." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 229, + 200, + 361, + 322 + ], + "blocks": [ + { + "bbox": [ + 229, + 200, + 361, + 322 + ], + "lines": [ + { + "bbox": [ + 229, + 200, + 361, + 322 + ], + "spans": [ + { + "bbox": [ + 229, + 200, + 361, + 322 + ], + "type": "image", + "image_path": "046cc163db7f6f9b589e856b6de421827f5ee2a6f603faa658e8776382ed5ee8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 373, + 201, + 523, + 322 + ], + "blocks": [ + { + "bbox": [ + 373, + 201, + 523, + 322 + ], + "lines": [ + { + "bbox": [ + 373, + 201, + 523, + 322 + ], + "spans": [ + { + "bbox": [ + 373, + 201, + 523, + 322 + ], + "type": "image", + "image_path": "9c402527a1c58f4e9e7f8a5f7ee87f241137acf7dd2bf6c2dae32c420c1bf697.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 375, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 411 + ], + "type": "text", + "content": "camera angles in talk formats) and minimal event progression, rendering them ineffective for benchmarking temporal reasoning capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 426, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 533 + ], + "type": "text", + "content": "Criterion for Narration. Our sourcing standards include video duration, its language source, and user ratings. The duration requirement (a minimum of 20 minutes and an average of 1.61 hours) ensures adequate temporal context for constructing reasoning chains. Our language diversity strategy intentionally excludes English and Chinese to counterbalance existing dataset biases, aligning with established findings that linguistic variety improves benchmark generalizability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": "Upon the aforementioned two quantitative prerequisites, we organized a panel of 14 multilingual domain experts to evaluate candidate videos using a standardized 10-point scale based on plot coherence and richness. The scoring is guided by several instructions. Videos scoring below 7 are systematically excluded, resulting in a final curated set of 960 high-quality narratives. Detailed annotation protocols are documented in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 636, + 231, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 231, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 231, + 649 + ], + "type": "text", + "content": "3.2. Stepwise Reasoning Annotations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "We label videos with reasoning steps via a two-stage human-in-the-loop framework. It first generates pseudo candidate QA pairs through automated pipelines, then employs expert-guided rewriting to curate high-quality annotations, ensuring rigorous adherence to benchmark speci" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 375, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 411 + ], + "type": "text", + "content": "fications while maintaining multimodal reasoning fidelity. To further improve annotation quality, we also implement a comprehensive quality assurance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 542 + ], + "type": "text", + "content": "Automatic Pipelines. We first employ AutoShot [94] to cut videos into several segments, and then use VideoChat2 [31] to caption them. For auditory content, we adopt whisper-large-v3 [51] to obtain speech transcripts, and DeepL [11] to translate them into English. The video captions and translated subtitles are then put into GPT-4o [47] to generate 6 pseudo QA pairs with a multi-step reasoning process. To ensure the quality of reasoning, we specify 7 multi-step reasoning types for narrative videos as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 544, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 544, + 553, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 553, + 566 + ], + "type": "text", + "content": "- Event prediction: Forecast subsequent events in the video timeline." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 568, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 568, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 314, + 568, + 553, + 591 + ], + "type": "text", + "content": "- Hypothetical reasoning: Deduce plausible scenario outcomes from stated premises." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 592, + 553, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 592, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 314, + 592, + 553, + 614 + ], + "type": "text", + "content": "- Event attribution: Determine causal origins or underlying motivations of video events." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 616, + 553, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 553, + 638 + ], + "type": "text", + "content": "- Implicit inference: Extract unstated temporal, emotional, or relational context from visual cues." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 639, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 639, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 314, + 639, + 553, + 662 + ], + "type": "text", + "content": "- Logical linkage: Establish event-mediated connections between visual/narrative elements." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 663, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 663, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 314, + 663, + 553, + 687 + ], + "type": "text", + "content": "- Information synopsis: Condense critical information across multimodal inputs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 688, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 688, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 314, + 688, + 553, + 711 + ], + "type": "text", + "content": "- Counting problems: Quantify state changes through arithmetic/combinatorial analysis." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21658" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "Expert-guided Rewriting. We recruit and train 67 graduate students to generate 8-10 high-quality QA pairs per video, providing raw video footage, translated subtitles, and GPT-generated pre-annotations that offer contextual hints without meeting final benchmark standards. Each QA pair must satisfy four rigorous criteria: (1) Non-synopsis questions require " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\geq 2" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " timestamped reasoning steps (start/end times documented); (2) Temporal distribution constraints (" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\leq 4" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " questions from " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "0 - 15\\mathrm{min}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\geq 3" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "15 - 40\\mathrm{min}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\geq 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "40 - 120\\mathrm{min}" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "); (3) Coverage of " + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\geq 5" + }, + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": " predefined reasoning taxonomies (from 7 categories) ensuring diversity; (4) Mandatory multimodal grounding where solutions demand both visual analysis (excluding subtitle-only answers) and explicit reasoning (beyond basic perception)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 263, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 407 + ], + "type": "text", + "content": "Comprehensive Quality Assurance. To ensure annotations meet our stringent multimodal standards, we implement a rigorous verification protocol: 10 trained reviewers validate each annotation, with non-compliant entries returned to annotators for iterative revisions until fully compliant. Annotators and reviewers are exclusively recruited from top-tier universities to ensure academic rigor. We further enforce quality through dual safeguards: a systematic " + }, + { + "bbox": [ + 55, + 263, + 296, + 407 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 263, + 296, + 407 + ], + "type": "text", + "content": " random sampling audit across both annotation and review stages, coupled with full documentation of protocols, annotation guidelines, and quality assessment criteria in supplementary materials." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 419, + 230, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 230, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 230, + 432 + ], + "type": "text", + "content": "3.3. Multi-Phase Evaluation Pipeline" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "content": "We benchmark VLMs and LLMs through a comprehensive multi-phase evaluation pipeline, which compares predictions against ground-truth annotations both at the process and outcome level. For the outcome-level evaluation stage, we adopt a multiple-choice question (MCQ) format, where false options are generated by DeepSeek-V3 [38] using carefully-designed prompts based on human-annotated answers. For the process-level stage, we propose the open-ended rating to fully evaluate the model's multi-step reasoning capability. Specifically, we adopt an LLM to evaluate the overall quality of the whole reasoning process through four 0-10 scores: logical coherence " + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "inline_equation", + "content": "(40\\%)" + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "content": ", similarity to ground truth " + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "inline_equation", + "content": "(40\\%)" + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "content": ", excluded for event prediction and hypothetical reasoning tasks), factual accuracy " + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "content": ", and clarity " + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "content": ". DeepSeek-V3 [38] serves as the judge due to its optimal balance between cost and human alignment (Section 4). For the event prediction and hypothetical reasoning tasks, we argue that there is no ground-truth reasoning steps since there might be multiple possible predictions, hence, we remove the similarity score and compute the final score with an 8:1:1 weight. All metrics are normalized to 0-100 scales, and we report the rankings by computing average scores across all metrics." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 552, + 193 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 552, + 193 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 552, + 193 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 552, + 193 + ], + "type": "image", + "image_path": "4f6f89f22f1d9c212edacb9064aad1049e97b6d604835cc51b939f3aee0f2745.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 202, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 257 + ], + "type": "text", + "content": "Figure 4. Human preference alignment results. For each plot, we show the win ratios of three different tested LLMs evaluated by human experts and VRBench. We then fit them with a straight line and quantify the correlation by calculating the Spearman correlation coefficient." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 277, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 384 + ], + "type": "text", + "content": "LLM Evaluation Support. For text-only LLMs, we convert videos to text inputs using Qwen2.5-72B-Instruct [83] to synthesize video captions and subtitles into fine-grained summaries. The process involves: 1) Dividing inputs into 5-minute chunks for duplicate removal and abstraction; 2) Merging chunk-level abstracts into coherent summaries with detailed audio-visual descriptions. These summaries enable LLM evaluations while addressing context window limits for VLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 394, + 395, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 395, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 395, + 407 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "type": "text", + "content": "We test a number of both open-source and proprietary VLMs (e.g. Qwen2-VL [68], InternVL2.5 [7], InternVideo2.5 [73], GPT-4o [47], Gemini-2.0-Pro [61], etc.) and LLMs (DeepSeek-R1 [18], QwQ [65], OpenAI o1 [48], Claude-3.7-Sonnet [59], etc.) with a two-stage evaluation on VRBench. The configuration of each evaluated model is detailed in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 556, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 618 + ], + "type": "text", + "content": "Evaluation Protocols. Our protocol involves two phases. Given a question, 1) models produce multi-step reasoning using Chain-of-Thought [75] prompts. Then 2) models choose final answers from multiple-choice options. For answer extraction, final answers are parsed from response endings to mitigate verbose option analysis, following [31]. While reasoning process evaluations use judge LLMs with structured prompts and example-guided rating extraction. All evaluation prompts and implementation details are included in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Table 2 shows the CoT evaluation results of LLMs and VLMs on VRBench. Among LLMs, we found the proprietary model Gemini-2.0-Flash-Thinking [12] achieves the optimal performance while other reasoning models such as OpenAI o1-preview [48] and DeepSeek-R1 [18] also demonstrate strong results. It is noted that o1-preview [48]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21659" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 69, + 555, + 559 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 555, + 559 + ], + "type": "table", + "html": "
ModelOverallResults by MetricResults by Taxonomy
MCQ-OOE-PEACPHRIIISEPLL
LLMs
Proprietary Models
GPT-4o [47]55.4963.8747.1154.4132.6369.0361.4771.8366.3766.19
Claude-3.7-Sonnet [59]58.0762.9153.2355.8735.1169.8363.2972.6167.9167.89
o1-preview [48]60.1468.4751.8156.9735.8770.5164.8173.1368.2769.41
Gemini-2.0-Flash-Thinking [61]60.9767.5354.4158.6138.1172.0967.5774.1970.4371.13
Open-Source Models
QwQ-32B-preview [64]35.9027.5144.2934.3134.4146.2939.6727.9742.3940.61
InternLM3-8B-Instruct [3]47.8150.3145.3145.4734.8760.3351.7755.4756.9756.81
Qwen2.5-7B-Instruct [83]48.2952.6143.9746.9335.1761.8351.5754.8756.6357.03
Llama3.3-70B-Instruct [41]49.8452.5947.0947.6338.0763.8754.0354.8359.4756.97
QwQ-32B [65]52.5256.0149.0349.5336.0362.0154.0352.0358.0159.03
Qwen2.5-72B-Instruct [83]53.5160.4946.5351.0336.0167.0358.5367.8361.9363.03
DeepSeek-V3 [38]56.0664.7947.3354.0336.5769.9761.8369.5365.5365.43
DeepSeek-R1 [18]57.1364.1950.0756.1338.9172.8164.1368.0168.5367.93
VLMs
Proprietary Models
Claude-3.7-Sonnet [59]68.1580.0956.2163.1132.9772.6371.1375.3771.3370.27
GPT-4o [47]68.6881.2356.1366.6136.5176.6770.4778.0372.0373.17
Gemini-2.0-Pro [61]74.6183.2965.9371.0965.2181.0175.7387.1377.9375.89
Open-Source Models
DeepSeek-VL2 [79]31.5033.2729.7325.9322.4135.7331.7330.0133.2929.57
H2OVL Mississippi-2B [17]47.1552.3341.9735.1732.3751.7141.4160.0349.9741.91
Phi-3.5-Vision [1]48.5258.0339.0131.5328.0345.0337.0368.0344.0336.03
LongVA-7B [91]50.1467.8132.4727.6125.0738.6933.7776.9138.2732.07
InternVL2.5-8B [7]50.4169.3131.5126.6326.3137.9931.2185.9737.3730.69
MiMo-VL-7B-RL [60]50.4863.3937.5747.3134.5963.8353.2374.6160.3356.97
VideoChat-Flash-7B [32]50.8272.0129.6324.6921.4139.3729.1779.3734.8728.07
InternVideo2.5 [73]51.9475.6328.2523.8124.0934.5127.0784.5733.6126.81
LongVA-7B-DPO [91]52.3667.9136.8130.7327.1745.0938.2379.3943.6737.03
Qwen2-VL-7B [68]54.0872.0136.1530.4926.6146.7133.6985.2743.3135.29
Aria [30]54.5572.9736.1330.1929.2344.2334.6386.2344.9934.57
Qwen2.5-VL-7B [2]56.5269.6143.4337.0733.1754.2742.2783.9150.1742.87
Keye-VL-8B-Preview [63]60.4464.4156.4762.5340.3773.6167.9970.1371.8371.03
Qwen2.5-VL-72B [2]61.7166.8556.5751.8746.1367.1354.1390.0463.6760.77
Kimi-VL-A3B-Thinking-2506 [62]61.8261.6761.9764.5347.9171.3769.5774.4771.2371.47
InternVL2.5-78B [7]62.3176.6148.0143.7738.6758.2146.1387.5354.4347.57
", + "image_path": "ae52fd76e13c844e68131919f9882f6caedda5c662fb6645364a1fa374943712.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 567, + 555, + 601 + ], + "lines": [ + { + "bbox": [ + 55, + 567, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 555, + 601 + ], + "type": "text", + "content": "Table 2. Evaluation results on VRBench across two evaluation metrics (outcome-level MCQ, process-level open-ended evaluation) and seven QA taxonomies (event attribution, counting problems, hypothetical reasoning, implicit inferences, information synopsis, event prediction, logical linkage). Bold values indicate the best results among all models, and underlined values are best results of LLMs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "text", + "content": "achieves the best outcome-level MCQ accuracy of " + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "inline_equation", + "content": "68.47\\%" + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "text", + "content": ". For the VLMs, Gemini-2.0-Pro [61] achieves " + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "inline_equation", + "content": "74.61\\%" + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "text", + "content": " overall accuracy, surpassing all components by at least " + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "inline_equation", + "content": "5.93\\%" + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "text", + "content": ". GPT-4o [47] and Claude-3.7-Sonnet [59] also show competitive results, which demonstrate " + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "inline_equation", + "content": "68.68\\%" + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "inline_equation", + "content": "68.15\\%" + }, + { + "bbox": [ + 55, + 621, + 294, + 706 + ], + "type": "text", + "content": " overall performance, respectively. For open-source models, InternVL2.5-78B [7], emerges as the best non-proprietary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 621, + 473, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 473, + 634 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 473, + 634 + ], + "type": "text", + "content": "VLMs with a " + }, + { + "bbox": [ + 314, + 621, + 473, + 634 + ], + "type": "inline_equation", + "content": "62.31\\%" + }, + { + "bbox": [ + 314, + 621, + 473, + 634 + ], + "type": "text", + "content": " overall accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "We then delve into the specific results of each metric and taxonomy. For results in each evaluation stage, we find that most of LLMs and VLMs are capable of achieving high MCQ accuracy, yet compared to LLMs, VLMs struggle to demonstrate their reasoning steps and exhibit lower reasoning ratings. Through results across taxonomy, we found" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21660" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "that most of the large models are not proficient in counting problems since they require fine-grained visual perception, in contrast with other tasks. Since the perception of LLMs fully depends on the video summary and some VLMs only support a small number of frame inputs, they are unable to correctly perceive the original frames containing the target elements, resulting in the MCQ accuracy close to random guessing and low reasoning process ratings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 186, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 186, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 186, + 190 + ], + "type": "text", + "content": "4.2. Ablations and Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 295, + 232 + ], + "type": "text", + "content": "We then further discuss the observations from the evaluation results and highlight some insights on obtaining higher multi-step reasoning performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 339 + ], + "type": "text", + "content": "Does the evaluation of VRBench align with Human Preference? In contrast to MCQs that compute accuracy via a deterministic method, the correctness of process-level open-ended evaluation highly depends on the judgment of LLM's performance, which may include hallucinations that could influence the evaluation reliability. To this end, we select a subset of 30 videos with 300 questions to perform process-level human evaluation, aiming to quantify the correlation between human and LLM assessment results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 341, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 483 + ], + "type": "text", + "content": "Specifically, the human annotators are asked to give process-level open-ended ratings following the same requirement for the judging LLM, i.e., four separated ratings are needed for each question to evaluate the logical coherence, similarity to ground-truth, factual accuracy, and expression clarity. We then compute the win ratio of pairwise comparison for each tested model following [24], where a model scores 1 if its rating is higher than its current opponent, 0 if lower, and 0.5 for a tie. Finally, we calculate the average win ratio of each model and the Spearman correlation coefficients " + }, + { + "bbox": [ + 55, + 341, + 295, + 483 + ], + "type": "inline_equation", + "content": "(\\rho)" + }, + { + "bbox": [ + 55, + 341, + 295, + 483 + ], + "type": "text", + "content": " of the win ratios evaluated by LLMs and human annotators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": "We first investigate the human preference alignment of several LLMs. We select GPT-4o [47], DeepSeek-V3 [38], Qwen2.5-7B [83], and Qwen2.5-72B [83] as evaluation models and give them the same rating prompt for a fair comparison. As shown in Figure 4, Qwen2.5 shows low human preference alignment compared with DeepSeek-V3 and GPT-4o, which both have correlation coefficients greater than 0.8. Though the human alignment of GPT-4o is slightly better than DeepSeek-V3, we observe the evaluation cost of GPT-4o is approximately 10 times that of DeepSeek, which imposes a significant burden on the benchmark users. Hence, we adopt DeepSeek-V3 as the judging model for its optimal trade-off between performance and cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "type": "text", + "content": "We also probe the evaluation correctness of each open-ended sub-metric. Following the same procedure, we compute the win ratio and correlation coefficient of DeepSeekV3 for each sub-metric, and the results are shown in Figure 4 (b-d). Results show the coefficients in all metrics are" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "higher than 0.8, indicating a strong positive monotonic relationship between LLM and human evaluation results. Another alternative is to instruct the LLM to compute the final score based on the given weights of each sub-metric, which is denoted as DeepSeek w/o sep in Figure 4. Results show that adding this additional computing process diminishes the robustness of evaluation LLM, thus in practice, we ask the model to output all sub-metric rating separately and automatically compute the final score." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 183, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 506 + ], + "type": "text", + "content": "The role of System-2 thinking. o1-like LLMs, also denoted as system-2 models, have emerged with superior reasoning abilities in multi-disciplinary scenarios such as science and math. We further explore these models' performance on the narrative video benchmark. We both assess the role of open-source and proprietary o1-models, as well as the gap compared with their previous system-1 version. For proprietaries, we compare OpenAI o1-preview [48] with GPT-4o [47], and results show that the system-2 version o1 outperforms GPT-4o with " + }, + { + "bbox": [ + 313, + 183, + 553, + 506 + ], + "type": "inline_equation", + "content": "4.30\\%" + }, + { + "bbox": [ + 313, + 183, + 553, + 506 + ], + "type": "text", + "content": " overall accuracy. For open-source model DeepSeek-V3 [38] and R1 [11], we observe that DeepSeek-R1 achieves higher scores both on multiple-choice questions and open-ended evaluations. It is noted that compared with Qwen2.5-72B-Instruct [82], QwQ-32B [65] gets higher reasoning ratings yet lower outcome-level MCQ results. This shows that though some system-2 models are more proficient in multi-step thinking, their ability to converge lengthy thinking processes into correct answers still needs improvement. Recently, some o1-like training strategies have also been utilized by some VLMs, such as LongVA-7B-DPO [91], a model that utilizes direct preference optimization [52] for long video understanding. It achieves an overall performance of " + }, + { + "bbox": [ + 313, + 183, + 553, + 506 + ], + "type": "inline_equation", + "content": "52.36\\%" + }, + { + "bbox": [ + 313, + 183, + 553, + 506 + ], + "type": "text", + "content": " and surpasses its vanilla version, suggesting that training VLMs with different optimization methods is feasible and that developing system-2 VLM models is crucial for tackling multi-step reasoning questions in narrative videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": "Impact of Model Size. It is commonly believed that models with large scales are often accompanied by better perception and reasoning capabilities. To verify such claims, we conduct experiments on the models with different scales, such as Qwen2.5-7B [83] and Qwen2.5-72B, Qwen2.5-VL [2] with 7B and 72B parameters, and InternVL2.5-8B [7] and InternVL2.5-78B. As shown in Table 2, models with more parameters obtain higher overall accuracy (48.29% vs 53.51% for Qwen2.5, 56.52% vs 61.71% for Qwen2.5-VL, and 50.41% vs 62.31% for InternVL2.5), and the performance gap on process-level metrics is more apparent. This indicates that large-scale models are more likely to achieve advanced reasoning capabilities. However, we also noticed that models specifically trained on the reasoning corpus like QwQ-32B [65], can achieve comparable performance compared with its 72B Qwen2.5 counterpart (52.52% vs. 53.51% in overall accuracy) with a smaller" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21661" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "parameter amount. This suggests that training small-sized models with various preference optimization strategies and reasoning-centric corpora might improve their overall reasoning capability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "content": "Long Context Helps. We observe a substantial performance gap between models with and without long context support. As shown in Table 2, models with long frame input tend to achieve higher overall accuracy. For example, Gemini-2.0-Pro [61] is capable of inference with 0.5 fps and supports large number frame inputs when answering questions of videos over 1 hour, and it achieves the optimal performance compared with all other fixed-length models. For the open-source models, Qwen2.5-VL-7B [2] and KimiVL-A3B-Thinking-2506 [62] with 256 frame input achieve " + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "inline_equation", + "content": "61.71\\%" + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "inline_equation", + "content": "61.82\\%" + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "content": " overall accuracy, which is " + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "inline_equation", + "content": "11.30\\%" + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "inline_equation", + "content": "11.41\\%" + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "content": " higher than the 8B InternVL-2.5 [7] model with 64 frames input. Since the average length of VRBench is 1.61 hours, the necessity of long context support is becoming more pronounced, so that models can perceive more plot elements that help the analysis of long video content." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 314, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 517 + ], + "type": "text", + "content": "LLMs vs. VLMs. Since tested LLMs are only provided with the automatically-generated video summary, yet VLMs are capable of directly perceiving the raw visual content, it is unfair to directly compare the numerical results of LLMs and VLMs. However, there are still some noteworthy phenomena that can be discussed. First, VLMs with coarse-grained visual input (e.g., 4 frames for H2OVL Mississippi-2B [17] and DeepSeek-VL2 [79]) perform significantly worse than several system-2 LLMs like DeepSeek-R1 [18] and Claude-3.7-Sonnet [59]. This indicates that randomly selecting a small number of frames is not enough to tackle narrative-based reasoning tasks that require long-range perception. Furthermore, VLMs with fine-grained visual perception, such as Gemini-2.0-Pro with a 0.5 fps perform better than top-tier LLMs on both process-level and outcome-level metrics, showing the inevitability of detailed visual content input for tackling questions in VRBench." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 527, + 222, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 222, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 222, + 540 + ], + "type": "text", + "content": "4.3. Test-Time Scaling Exploration" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "We here probe models' performance when scaling test-time compute cost. Since we tend to prompt the models with a Chain-of-Thought template when tackling complex reasoning tasks, it is observed that system-2 models are capable of dramatically expanding the inference budget to improve their performance. To this end, we set a series of token limitations on these models and require the models to think with different instructions. Specifically, we select QwQ-32B [65] as the experimental LLM and Qwen2-VL-7B [82] as the testing VLM on a subset with 300 videos and 2,403 questions. For each model, several parallel experiments are conducted with increasing maximum token limits from 256 to 2048. We instruct the models to output the reasoning process as much as possible for large token limitations, and" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 545, + 205 + ], + "type": "image", + "image_path": "9f214471ff20dc28023bbddcd0fd23f4621a7b12fa38cd7c031b608e764a1971.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 215, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 239 + ], + "type": "text", + "content": "Figure 5. Test-Time Scaling Results. We report the average accuracy of outcome-level MCQ and process-level open-ended ratings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 272, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 512 + ], + "type": "text", + "content": "let them think concisely when the token limit is low. Since we aim to investigate the quality of models' initial reasoning process and their capability to generate the right answers through thinking, we report the overall scores that both evaluate the outcome accuracy and process quality. Results are shown in Figure 5, where QwQ shows a remarkable rating boost from " + }, + { + "bbox": [ + 313, + 272, + 555, + 512 + ], + "type": "inline_equation", + "content": "48.91\\%" + }, + { + "bbox": [ + 313, + 272, + 555, + 512 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 272, + 555, + 512 + ], + "type": "inline_equation", + "content": "61.34\\%" + }, + { + "bbox": [ + 313, + 272, + 555, + 512 + ], + "type": "text", + "content": " when setting a large token limit. On the contrary, the small-sized system-1 model Qwen2-VL-7B performs even worse when using a large token number limitation and instructing the model to output more thinking process. The model tends to output more ambiguous outputs that lead to the wrong answer. This observation suggests that models with large parameter scales and system-2 capabilities benefit from the test-time scaling strategy. It also provides the insight that developing more test-time scaling approaches that guide the model to generate long reasoning traces could be a feasible and promising way to unlock more potential capabilities of system-2 models, thereby benefiting complex tasks that require multi-step reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 543, + 388, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 543, + 388, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 543, + 388, + 555 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "This paper introduces VRBench, a comprehensive long-narrative video benchmark for evaluating multi-step reasoning. Through manually filtered narrative videos, high-quality stepwise annotations, and a multi-phase evaluation pipeline, VRBench distinguishes itself from other existing reasoning benchmarks and shows the robust capability to evaluate LLMs and VLMs both from the process and outcome perspectives. By assessing and analyzing 31 frontier large models, we thoroughly demonstrate the detailed performance of current reasoning models across various reasoning questions and provide valuable insights towards constructing more advanced multi-step reasoning models." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21662" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "type": "text", + "content": "6. Acknowledgement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 160, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 180, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 294, + 236 + ], + "type": "text", + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 237, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 296, + 281 + ], + "type": "text", + "content": "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 284, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 284, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 284, + 296, + 327 + ], + "type": "text", + "content": "[3] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 329, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 329, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 329, + 296, + 384 + ], + "type": "text", + "content": "[4] Guo Chen, Yicheng Liu, Yifei Huang, Yuping He, Baoqi Pei, Jilan Xu, Yali Wang, Tong Lu, and Limin Wang. Cg-bench: Clue-grounded question answering benchmark for long video understanding. arXiv preprint arXiv:2412.12075, 2024.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 386, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 386, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 294, + 430 + ], + "type": "text", + "content": "[5] Wenhu Chen, Ming Yin, Max Ku, Pan Lu, Yixin Wan, Xueguang Ma, Jianyu Xu, Xinyi Wang, and Tony Xia. Theoremq: A theorem-driven question answering dataset. arXiv preprint arXiv:2305.12524, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "text", + "content": "[6] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. In European Conference on Computer Vision, pages 179-195. Springer, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 553 + ], + "type": "text", + "content": "[7] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhang-wei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 556, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 610 + ], + "type": "text", + "content": "[8] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 656 + ], + "type": "text", + "content": "[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees GM Snoek, and Yuki M Asano. Tvbench: Redesigning video-language evaluation. arXiv preprint arXiv:2410.07752, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Rocktim Jyoti Das, Simeon Emilov Hristov, Haonan Li, Dimitar Iliyanov Dimitrov, Ivan Koychev, and Preslav Nakov. *Exams-v: A multi-discipline multilingual multimodal exam benchmark for evaluating vision language models.* arXiv preprint arXiv:2403.10378, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 55, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 137 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160101) and Jiangsu Frontier Technology Research and Development Program (No. BF2024076)," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[11] DeepL. Deepl translate: The world's most accurate translator. https://www deepl.com/en/Translator, 2025.4,7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[12] Google Deepmind. Gemini 2.0 flash thinking. https://deepmind.google/technologies/gemini/flash-thinking/, 2025. Accessed: 2025-01-21. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[13] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. Advances in Neural Information Processing Systems, 37:89098-89124, 2025. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[14] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 293 + ], + "type": "text", + "content": "[15] Kehua Feng, Keyan Ding, Weijie Wang, Xiang Zhuang, Zeyuan Wang, Ming Qin, Yu Zhao, Jianhua Yao, Qiang Zhang, and Huajun Chen. Sciknoweval: Evaluating multilevel scientific knowledge of large language models. arXiv preprint arXiv:2406.09098, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 295, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 349 + ], + "type": "text", + "content": "[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[17] Shaikat Galib, Shanshan Wang, Guanshuo Xu, Pascal Pfeiffer, Ryan Chesler, Mark Landry, and Sri Satish Ambati. H2ovl-mississippi vision language models technical report. arXiv preprint arXiv:2410.13611, 2024. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[18] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 2, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "text", + "content": "[19] Songhao Han, Wei Huang, Hairong Shi, Le Zhuo, Xiu Su, Shifeng Zhang, Xu Zhou, Xiaojuan Qi, Yue Liao, and Si Liu. Videospeech: A large-scale chain-of-thought dataset for fine-grained video reasoning via core frame selection. arXiv preprint arXiv:2411.14794, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 505, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 558 + ], + "type": "text", + "content": "[20] Xuehai He, Weixi Feng, Kaizhi Zheng, Yujie Lu, Wanrong Zhu, Jiachen Li, Yue Fan, Jianfeng Wang, Linjie Li, Zhengyuan Yang, et al. Mmworld: Towards multi-discipline multi-faceted world model evaluation in videos. arXiv preprint arXiv:2406.08407, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "text", + "content": "[22] Kairui Hu, Penghao Wu, Fanyi Pu, Wang Xiao, Yuanhan Zhang, Xiang Yue, Bo Li, and Ziwei Liu. Video-mmmu: Evaluating knowledge acquisition from multi-discipline professional videos. arXiv preprint arXiv:2501.13826, 2025. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[23] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems, 36:62991-63010, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21663" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 205 + ], + "type": "text", + "content": "[25] Zhen Huang, Zengzhi Wang, Shijie Xia, Xuefeng Li, Haoyang Zou, Ruijie Xu, Run-Ze Fan, Lyumanshan Ye, Ethan Chern, Yixin Ye, et al. Olympic: Benchmarking multi-discipline cognitive reasoning for superintelligent ai. Advances in Neural Information Processing Systems, 37: 19209-19253, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 206, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 260 + ], + "type": "text", + "content": "[26] Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livocodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 315 + ], + "type": "text", + "content": "[27] Zhihuan Jiang, Zhen Yang, Jinhao Chen, Zhengxiao Du, Weihan Wang, Bin Xu, Yuxiao Dong, and Jie Tang. Visscience: An extensive benchmark for evaluating k12 educational multi-modal scientific reasoning. arXiv preprint arXiv:2409.13730, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 316, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 348 + ], + "type": "text", + "content": "[28] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 414 + ], + "type": "text", + "content": "[29] Bozheng Li, Yongliang Wu, Yi Lu, Jiashuo Yu, Licheng Tang, Jiawang Cao, Wenqing Zhu, Yuyang Sun, Jay Wu, and Wenbo Zhu. Veu-bench: Towards comprehensive understanding of video editing. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13671-13680, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 295, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 469 + ], + "type": "text", + "content": "[30] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "text", + "content": "[31] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 537, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 590 + ], + "type": "text", + "content": "[32] Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, et al. Videochat-flash: Hierarchical compression for long-context video modeling. arXiv preprint arXiv:2501.00574, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "text", + "content": "[33] Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforcement fine-tuning. arXiv preprint arXiv:2504.06958, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[34] Yunxin Li, Xinyu Chen, Baotian Hu, Longyue Wang, Haoyuan Shi, and Min Zhang. Videovista: A versatile benchmark for video understanding and reasoning. arXiv preprint arXiv:2406.11303, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "text", + "content": "[35] Zekun Li, Xianjun Yang, Kyuri Choi, Wanrong Zhu, Ryan Hsieh, HyeonJung Kim, Jin Hyuk Lim, Sungyoung Ji," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Byungju Lee, Xifeng Yan, et al. Mmsci: A multimodal multi-discipline dataset for phd-level scientific comprehension. In AI for Accelerated Materials Design-Vienna 2024, 2024.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[36] Zhenwen Liang, Kehan Guo, Gang Liu, Taicheng Guo, Yujun Zhou, Tianyu Yang, Jiajun Jiao, Renjie Pi, Jipeng Zhang, and Xiangliang Zhang. Scemqa: A scientific college entrance level multimodal question answering benchmark. arXiv preprint arXiv:2402.05138, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "text", + "content": "[37] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689-26699, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[38] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "type": "text", + "content": "[39] Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, and Lu Hou. Tempcompass: Do video llms really understand videos? arXiv preprint arXiv:2403.00476, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "text", + "content": "[40] Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. Et bench: Towards open-ended event-level video-language understanding. arXiv preprint arXiv:2409.18111, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "text", + "content": "[41] Llama-3.3. Llama-3.3-70b-instruct. https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct, 2025.6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "text", + "content": "[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[43] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[44] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[45] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "text", + "content": "[46] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. arXiv preprint arXiv:2311.16103, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[47] OpenAI. Hello gpt4-o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-13, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 552, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 552, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 552, + 690 + ], + "type": "text", + "content": "[48] OpenAI. Introducing openai o1. https://openai.com/o1/, 2024.2,5,6,7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[49] OpenAI. Openai o3-mini. https://openai.com/index/openai-o3-mini/, 2025.2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21664" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[50] Viorica Patraucean, Lucas Smaira, Ankush Gupta, Adria Recasens, Larisa Markeeva, Dylan Banarse, Skanda Koppula, Mateusz Malinowski, Yi Yang, Carl Doersch, et al. Perception test: A diagnostic benchmark for multimodal video models. Advances in Neural Information Processing Systems, 36:42748-42761, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 139, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 294, + 193 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "text", + "content": "[52] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[53] Ruchit Rawal, Khalid Saifullah, Miquel Farre, Ronen Basri, David Jacobs, Gowthami Somepalli, and Tom Goldstein. Cinepile: A long video question answering dataset and benchmark. arXiv preprint arXiv:2405.08813, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "text", + "content": "[54] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[55] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 437 + ], + "type": "text", + "content": "[56] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[57] Liangtai Sun, Yang Han, Zihan Zhao, Da Ma, Zhennan Shen, Baocai Chen, Lu Chen, and Kai Yu. Scieval: A multi-level large language model evaluation benchmark for scientific research. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 19053-19061, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "type": "text", + "content": "[58] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4631–4640, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "text", + "content": "[59] Anthropic. Claude Team. Claude 3.7 sonnet. https://www.anthropic.com/claude/sonnet, 2025. 2, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[60] Core Team, Zihao Yue, Zhenru Lin, Yifan Song, Weikun Wang, Shuhuai Ren, Shuhao Gu, Shicheng Li, Peidian Li, Liang Zhao, Lei Li, et al. Mimo-vl technical report. https://arxiv.org/abs/2506.03569, 2025. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[61] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[62] Kimi Team, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chen-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "zhuang Du, Chu Wei, et al. Kimi-vl technical report. arXiv preprint arXiv:2504.07491, 2025. 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[63] Kwai Keye Team, Biao Yang, Bin Wen, Changyi Liu, Chenglong Chu, Chengru Song, Chongling Rao, Chuan Yi, Da Li, Dunju Zang, et al. Kwai keye-vl technical report. arXiv preprint arXiv:2507.01949, 2025. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[64] Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown. https://qwenlm.github.io/blog/qwq-32b-preview/, 2024. Accessed: 2024-11-28. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[65] Qwen Team. Qwq-32b: Embracing the power of reinforcement learning. https://qwenlm.github.io/blog/qwq-32b/, 2025. Accessed: 2025-3-6. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[66] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamavol: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095-95169, 2025. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "text", + "content": "[69] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, et al. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[70] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "text", + "content": "[71] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 601 + ], + "type": "text", + "content": "[72] Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "text", + "content": "[73] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xiangyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Internvideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[74] Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, et al. Charxiv: Charting gaps in realistic chart understanding in multimodal llms. Advances in Neural Information Processing Systems, 37:113569-113697, 2025. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21665" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[75] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "type": "text", + "content": "[76] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[77] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[78] Yifan Wu, Lutao Yan, Leixian Shen, Yunhai Wang, Nan Tang, and Yuyu Luo. Chartinsights: Evaluating multimodal large language models for low-level chart question answering. arXiv preprint arXiv:2405.07001, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[79] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "text", + "content": "[80] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "text", + "content": "[81] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-01: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[82] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report, 2024. URL: https://arxiv.org/abs/2407.10671, 2024. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[83] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[84] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. Advances in Neural Information Processing Systems, 37:57240-57261, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 612 + ], + "type": "text", + "content": "[85] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[86] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[87] Zhongshen Zeng, Pengguang Chen, Shu Liu, Haiyun Jiang, and Jiaya Jia. Mr-gsm8k: A meta-reasoning benchmark for large language model evaluation. arXiv preprint arXiv:2312.17080, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 473 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[88] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[89] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, et al. Cmmmu: A chinese massive multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2401.11944, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 238 + ], + "type": "text", + "content": "[90] Hongjie Zhang, Yi Liu, Lu Dong, Yifei Huang, Zhen-Hua Ling, Yali Wang, Limin Wang, and Yu Qiao. Movqa: A benchmark of versatile question-answering for long-form movie understanding. arXiv preprint arXiv:2312.04817, 2023.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 294 + ], + "type": "text", + "content": "[91] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 296, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 296, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 553, + 350 + ], + "type": "text", + "content": "[92] Yilun Zhao, Lujing Xie, Haowei Zhang, Guo Gan, Yitao Long, Zhiyuan Hu, Tongyan Hu, Weiyuan Chen, Chuhan Li, Junyang Song, et al. Mmvu: Measuring expert-level multi-discipline video understanding. arXiv preprint arXiv:2501.12380, 2025. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 352, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 352, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 352, + 553, + 406 + ], + "type": "text", + "content": "[93] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 407, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 473 + ], + "type": "text", + "content": "[94] Wentao Zhu, Yufang Huang, Xiufeng Xie, Wenxian Liu, Jincan Deng, Debing Zhang, Zhangyang Wang, and Ji Liu. Autoshot: A short video dataset and state-of-the-art shot boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2238-2247, 2023. 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21666" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_content_list.json b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f83c0f786c7cddf11e23d3d31fabe11acb38633 --- /dev/null +++ b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_content_list.json @@ -0,0 +1,1876 @@ +[ + { + "type": "text", + "text": "VRM: Knowledge Distillation via Virtual Relation Matching", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weijia Zhang1", + "bbox": [ + 220, + 179, + 338, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fei Xie", + "bbox": [ + 393, + 180, + 460, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University", + "bbox": [ + 246, + 199, + 504, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weidong Cai", + "bbox": [ + 519, + 180, + 630, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chao $\\mathbf{Ma}^{1,*}$", + "bbox": [ + 687, + 180, + 779, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The University of Sydney", + "bbox": [ + 529, + 199, + 748, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{weijia.zhang, feixie, chaoma}@sjtu.edu.cn tom.cai@sydney.edu.au", + "bbox": [ + 205, + 218, + 785, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Knowledge distillation (KD) aims to transfer the knowledge of a more capable yet cumbersome teacher model to a lightweight student model. In recent years, relation-based KD methods have fallen behind, as their instance-matching counterparts dominate in performance. In this paper, we revive relational KD by identifying and tackling several key issues in relation-based methods, including their susceptibility to overfitting and spurious responses. Specifically, we transfer novelly constructed affinity graphs that compactly encapsulate a wealth of beneficial inter-sample, inter-class, and inter-view correlations by exploiting virtual views and relations as a new kind of knowledge. As a result, the student has access to richer guidance signals and stronger regularisation throughout the distillation process. To further mitigate the adverse impact of spurious responses, we prune the affinity graphs by dynamically detaching redundant and unreliable edges. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO datasets demonstrate the superior performance of the proposed virtual relation matching (VRM) method, where it consistently sets new state-of-the-art records over a range of models, architectures, tasks, and set-ups. For instance, VRM for the first time hits $74.0\\%$ accuracy for ResNet50 $\\rightarrow$ MobileNetV2 distillation on ImageNet, and improves DeiT by $14.44\\%$ on CIFAR-100 with a ResNet56 teacher. The code and models are released at https://github.com/VISION-SJTU/VRM.", + "bbox": [ + 89, + 301, + 483, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 726, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning is achieving incredible performance at the cost of increasing model complexity and overheads. As a consequence, large and cumbersome neural models struggle to work in resource-constrained environments. Knowledge distillation (KD) [27] has been proposed to address this issue by transferring the knowledge of larger and more capable models to smaller and lightweight ones that are resource-friendly. KD works by minimising the dis", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f457de3b44d1eb43aa33e863cc93e5ce4f2350c218ffe49ed23ec7eff009baa.jpg", + "image_caption": [ + "(a) Instance Matching" + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 705, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/840db0a4cabc73b6a2b1d4c181b55da4e6800a7a0d5e3641339bb5477e3f17cf.jpg", + "image_caption": [ + "(b) Relation Matching" + ], + "image_footnote": [], + "bbox": [ + 712, + 268, + 903, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ac77d54bb8a81d21d0854823bf132eb0efae85c267e7e13120e2ccd55a3bd0a.jpg", + "image_caption": [ + "(c) VRM (Dense)" + ], + "image_footnote": [], + "bbox": [ + 516, + 386, + 707, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea953baeec179d0a681f15a1ef6e9512c81bac526bf87d200963122154805cf9.jpg", + "image_caption": [ + "(d) VRM (Pruned)", + "Figure 1. Conceptual illustration of VRM compared to existing KD methods based on instance matching and relation matching." + ], + "image_footnote": [], + "bbox": [ + 712, + 386, + 903, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tance between compact representations of knowledge extracted from the teacher and student models. It has found widespread application across a spectrum of downstream tasks [18, 22, 30, 34, 62, 64, 67, 70, 75, 77, 83].", + "bbox": [ + 511, + 553, + 905, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "According to the type of knowledge representations to be transferred, KD methods can be broadly categorised into feature-based [56], logit-based [27], and relation-based [52] approaches. The former two directly match the feature maps or logit vectors produced by the teacher and student models for each training sample, which is essentially instance matching (IM). By contrast, relation matching (RM) methods construct and match structured relations extracted within a batch of model responses. A conceptual illustration is presented in Figs. 1a and 1b.", + "bbox": [ + 511, + 614, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instance matching has been the prevailing distillation approach in recent years. Popular KD benchmarks see a dominance by IM methods such as FCFD [45], NORM [47], and CRLD [76], with different downstream tasks successfully tackled by directly adopting IM-based distillation [6, 14, 62, 69]. Yet, recent studies discovered that relational knowledge is more robust to variations in neural architectures, data modalities, and tasks [52, 61]. Meanwhile, methods transferring relations have also achieved promising", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2707", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance for a range of tasks, including but not limited to segmentation [67] and detection [15, 32].", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite growing interest, relation-based methods still fall significantly short compared to their instance matching counterparts. Even the strongest RM method has been outperformed easily by recent IM solutions [29] (see Tabs. 1 and 2). RM-based methods also struggle with more challenging tasks such as object detection [29]. Moreover, previous RM solutions are primarily limited to matching inter-sample [29, 53, 61], inter-class [29], or inter-channel [46, 73] relations via simple Gram matrices. To our best knowledge, no different forms of relations other than these have been proposed since the work of DIST [29].", + "bbox": [ + 88, + 122, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper fills this gap with a new kind of relations for KD - inter-view relations (Fig. 1c), which seamlessly and compactly integrate with previous inter-sample and inter-class relations. Our designs are motivated by two important observations made about RM methods in a set of pilot experiments: 1) RM methods are more susceptible to overfitting than IM methods; 2) RM methods are subject to an adverse gradient propagation effect. We empirically find that incorporating richer and more diverse relations into the matching objective helps mitigate both issues.", + "bbox": [ + 89, + 291, + 482, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we generate virtual views of samples through simple transformations, followed by constructing virtual affinity graphs and transferring the virtual relations between real and virtual samples along the edges. In lieu of Gram matrices that suffer from significant knowledge loss [29, 53, 61], we preserve the raw relations along the secondary dimension as auxiliary knowledge which adds to the types and density of relational knowledge transferred. Moreover, we also prune our affinity graphs by striping away both redundant and unreliable edges to further alleviate the propagating gradients of spurious samples (Fig. 1d).", + "bbox": [ + 89, + 444, + 482, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above insights and remedies altogether lead to a novel Virtual Relation Matching (VRM) framework for knowledge distillation. VRM is conceptually simple, easy to implement, and devoid of complicated training procedures. It is capable of transferring rich, sophisticated knowledge robust to overfitting and spurious signals. VRM sets new state-of-the-art performance for different datasets, tasks, and settings. Perhaps more significant is that VRM makes relation-based methods regain competitiveness and back in the lead over instance matching approaches in different scenarios. To summarise, the contributions of this work include:", + "bbox": [ + 89, + 612, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We make an early effort to present comparative analyses of existing KD methods through the lens of training dynamics and sample-wise gradients, and identify overfitting and spurious gradient diffusion as two main cruxes in relational KD methods.", + "- We distill richer, more diverse relations by generating virtual views, constructing virtual affinity graphs, and" + ], + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "matching virtual relations. We also for the first time tackle relational KD with considerations of spurious samples and gradients by pruning redundant and unreliable edges, alongside designs to relax the matching criterion.", + "bbox": [ + 524, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present the streamlined VRM framework for knowledge distillation, with extensive experimental results on a diversity of models and tasks to highlight its superior performance, alongside rigorous analyses on the soundness and efficiency of our designs.", + "bbox": [ + 513, + 152, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 242, + 653, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "KD via instance matching. KD was first proposed in Hinton et al. [27], where a student is trained to mimic the prediction of a pre-trained teacher model for each sample. Follow-up works have mostly followed such instance matching (IM) paradigm, and can be categorised into logit (or prediction)-based and feature-based methods according to what is matched. Logit-based KD has evolved from using adaptive-softened logits [33, 41] to decoupling target- and non-target logits [24, 72, 81] and applying logit transformation [59, 82]. Others [50, 58] set up auxiliary ad-hoc networks between teacher and student to facilitate logit transfer. Feature-based methods minimise the distance between the feature maps [1, 25, 56] or salient regions within in features [23, 74] from specified layers in both teacher and student networks. Some also design sophisticated distillation paths [7, 42]. All these methods are based on instance-wise transfer of knowledge, and are herein referred to as the \"IM\" methods, as illustrated by Fig. 1a.", + "bbox": [ + 511, + 268, + 903, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "KD via relation matching. Some works transfer instead mutual relations mined amongst the network outputs extracted from a batch of training instances. These relation matching (RM) methods usually involve constructing network outputs into compact relation representations that encode rich higher-order information, as depicted in Fig. 1b. Different relation encoding functions mare, including inter-sample [29, 48, 52-54, 61], inter-class [29], inter-channel [46, 73, 78], inter-layer [48], and contrastive [60] relations. To date, IM solutions have dominated KD with their superior performance, leading top-performing relation-based methods by considerable margins. While many new IM methods are found within the last two years, frustratingly few RM solutions are being proposed. In this work, we strive to close this gap with a new kind of relations for KD — inter-view virtual relations, and revive relation-based KD by making it overtake its IM counterparts.", + "bbox": [ + 511, + 546, + 903, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning with virtual knowledge. While not a standalone research topic, learning with virtual knowledge finds relevance in a variety of learning-based problems. For instance, a commonly used paradigm in 3D vision tasks is to learn (or construct) from the raw data a virtual view or representation as auxiliary knowledge in solving the main task, rang-", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2708", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa3d16d809bfea323d73787a4bc4b71c47ced412ea933092802c9a519e8a3ac9.jpg", + "image_caption": [ + "(a) Increasing performance on training set." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 352, + 190 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b21167aac795e5f47ec07926028c316876f9d7dc6fc217c47de434bdb24de9c9.jpg", + "image_caption": [ + "(b) Plateaued performance on validation set.", + "(c) Diffusion of gradient perturbations within a batch.", + "Figure 2. Pilot studies that reveal the overfitting and spurious gradient diffusion issues with RM-based KD." + ], + "image_footnote": [], + "bbox": [ + 357, + 90, + 614, + 190 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e8b217402999478d55d20b29ac28241d31e164a4e76e5bccd35629a820ee062.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 90, + 900, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing from object reconstruction [5] and optical flow [2] to 3D semantic segmentation [36], monocular 3D object detection [13], and 3D GAN inversion [64]. More broadly, many data-efficient learning methods also share the spirit of utilising virtual knowledge. For instance, self-supervised learning methods generate virtual views of the unlabelled data to enable the learning of pretext tasks [11, 20]. Another popular paradigm is transformation-invariant representation learning [51, 57] in semi-supervised learning and domain adaptation. It enforces consistency between representations learnt for a raw sample and a virtual view of it. The virtual view is often obtained by applying semantic-preserving transformations to the raw sample [16, 17]. This work is more related to this later paradigm, but involves learning with virtual knowledge in a different context, via different approaches, and for a different problem.", + "bbox": [ + 88, + 239, + 483, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 494, + 228, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "KD methods generally employ a cross-entropy (CE) loss and a distillation loss to supervise student learning. The CE loss is computed between student logits $\\mathbf{z}_i^s$ for each sample and its ground-truth label $\\mathbf{y}_i$ . The distillation loss matches teacher and student outputs via a distance metric $\\phi(\\cdot)$ . In vanilla KD [27], $\\phi(\\cdot)$ is the Kullback-Leibler divergence (KLD) between teacher logits $\\mathbf{z}^t$ and student logits $\\mathbf{z}^s$ :", + "bbox": [ + 89, + 518, + 483, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {i} ^ {\\mathrm {K D}} = \\phi_ {\\mathrm {K L D}} \\left(\\mathbf {z} _ {i} ^ {s}, \\mathbf {z} _ {i} ^ {t}\\right) \\\\ = \\tau^ {2} \\sum_ {j = 1} ^ {C} \\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {t} / \\tau\\right) \\log \\frac {\\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {t} / \\tau\\right)}{\\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {s} / \\tau\\right)}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 630, + 482, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the Softmax operation with temperature parameter $\\tau$ , and $C$ is the number of classes. For feature-based methods, $\\phi (\\cdot)$ can be the mean squared error (MSE) between teacher and student feature maps for each instance [56], i.e., $\\mathcal{L}_i^{\\mathrm{KD}} = \\phi_{\\mathrm{MSE}}(\\mathbf{f}_i^s,\\mathbf{f}_i^t)$ .", + "bbox": [ + 89, + 704, + 482, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For relation-based methods, a relation encoding function $\\psi(\\cdot)$ first abstracts teacher or student outputs of all instances within a training batch into a relational representation, before applying $\\phi(\\cdot)$ to match these relational representations between the teacher and the student. For example, the KD objectives of DIST [29] take the general form of:", + "bbox": [ + 89, + 780, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {K D}} = \\phi \\left(\\psi \\left(\\mathbf {z} _ {1} ^ {s}, \\mathbf {z} _ {2} ^ {s}, \\dots , \\mathbf {z} _ {B} ^ {s}\\right), \\psi \\left(\\mathbf {z} _ {1} ^ {t}, \\mathbf {z} _ {2} ^ {t}, \\dots , \\mathbf {z} _ {B} ^ {t}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 878, + 482, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B$ is the batch size, and subscript $i$ of $\\mathcal{L}^{\\mathrm{KD}}$ is dropped as the loss is computed for a batch of instances. DIST uses both inter-class and inter-sample relation encoders as $\\psi (\\cdot)$ .", + "bbox": [ + 511, + 239, + 903, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Pilot Studies", + "text_level": 1, + "bbox": [ + 511, + 297, + 643, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training dynamics of KD methods. We examine the training dynamics of different relation-based methods on CIFAR-100 with ResNet32 $\\times$ 4 $\\rightarrow$ ResNet8 $\\times$ 4 as the teacher-student pair. In Figs. 2a and 2b, we immediately notice that relational methods achieve significantly higher training accuracy, but they only marginally lead or even fall short in test accuracy compared to IM-based KD [27]. We hypothesise that relational methods are more prone to overfitting. This is expected given that the optimality of IM matching (cond. $\\mathcal{A}$ ) implies the optimality of relation matching (cond. $\\mathcal{B}$ ), while the converse does not hold. Concretely, $\\mathcal{A} \\Rightarrow \\mathcal{B} \\land \\neg \\mathcal{B} \\Rightarrow \\neg \\mathcal{A}$ . In other words, relation matching is a weaker and less constrained objective than IM, which makes the student more readily fit the teaching signals and not generalise well. Thus, we conclude that $C1$ : relation matching methods are more prone to overfitting.", + "bbox": [ + 511, + 323, + 906, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gradient analysis of KD methods. We investigate the gradient patterns within a batch when a spurious sample produces a major misguiding signal. To this end, we first generate two random vectors $\\mathbf{x},\\mathbf{y}\\sim \\mathcal{N}(0,1)$ for $\\mathbf{x},\\mathbf{y}\\in \\mathbb{R}^{B\\times D}$ , where $B$ is the batch size and $D$ is the dimension of persample predictions, $\\mathbf{x}$ is taken as the sample-wise predictions, and $\\mathbf{y}$ the supervision signals. We then add a noise vector $\\epsilon = c\\cdot \\mathbf{z}$ to $\\mathbf{x}_t$ , where $\\mathbf{z}\\sim \\mathcal{N}(0,1)$ and $c$ is a scaling factor. In this case, $\\mathbf{x}_t + \\epsilon$ becomes a spurious prediction within our batch. We compute the loss from $\\mathbf{x}$ and $\\mathbf{y}$ using either IM or RM objectives, and consider the change in sample-wise gradients $\\mathbf{g}$ within the batch upon the injection of the spurious sample. Formally, we visualise:", + "bbox": [ + 511, + 570, + 906, + 767 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {g} = \\left[ \\| \\frac {\\partial \\mathcal {L}}{\\partial \\left(\\mathbf {x} _ {i} ^ {\\prime}\\right)} \\| _ {2} - \\| \\frac {\\partial \\mathcal {L}}{\\partial \\left(\\mathbf {x} _ {i}\\right)} \\| _ {2} \\right] _ {i = 1} ^ {B},\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 775, + 808, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} + \\epsilon \\cdot \\mathbb {I} (i = t), \\quad \\mathcal {L} \\in \\left\\{\\mathcal {L} _ {I M}, \\mathcal {L} _ {R M} \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 814, + 903, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Fig. 2c ( $B = 64$ and $t = 32$ ), when the IM objective is used, only the spurious sample receives a prominent gradient. Whereas for an RM objective, many other samples receive significant gradients as they are directly connected", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2709", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3da20e409478ffba022d966b916e2054d8d901941aec90a5beb4abec29a4013.jpg", + "image_caption": [ + "Figure 3. Alleviated overfitting and improved training dynamics." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 344, + 190 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c0212815eefb87d1b1b23fee3da4c880b319dcaa39c11e53cf646abaa94e1691.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 87, + 598, + 190 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72708e7921ac197d3176c5c1eebf20d25dfecbeb15f305b059d2c943384bd253.jpg", + "image_caption": [ + "Figure 4. Alleviated spurious gradients." + ], + "image_footnote": [], + "bbox": [ + 606, + 88, + 901, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to $\\mathbf{x}_t$ in the computational graph of the RM loss. In other words, the spurious signals produced by one malign prediction will propagate to and affect all samples within a batch (in fact, those closer to $\\mathbf{x}_t$ within the prediction manifold are more strongly affected). This means other sample-wise predictions will be significantly updated only to accommodate a malign prediction, even if they are already in relatively good shape. Through this investigation, we discover that $C2$ : relation matching methods are more prone to the adverse impact of spurious samples. We also present this pilot analysis on ImageNet training in Fig. 4.", + "bbox": [ + 88, + 220, + 482, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For $C1$ , common approaches to combat overfitting include the incorporation of richer learning signals and regularisation, which for RM-based KD methods means richer relations constructed and transferred. For $C2$ , an intuitive solution is to identify and suppress the effect of spurious predictions or relations or to slacken the matching criterion. Guided by these principles, we now proceed to formally build up our method step by step.", + "bbox": [ + 88, + 387, + 482, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Method", + "text_level": 1, + "bbox": [ + 89, + 521, + 181, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Constructing Inter-Sample Relations", + "text_level": 1, + "bbox": [ + 89, + 546, + 410, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first construct relation graph $\\mathcal{G}^{IS}$ that encodes intersample affinity within a batch of sample predictions $\\{\\mathbf{z}_i\\}_{i = 1}^B$ . Different from [52, 61], our relations are constructed from predicted logits which embed more compact categorical knowledge. We use the pairwise distance between instance-wise predictions within a batch as our measure of affinity. Existing methods leverage the Gram matrices [29, 53, 54, 61] to encode inter-sample relations, but we find that this leads to collapsed inter-class knowledge via the inner product operation. Instead, our pairwise distance preserves the inter-class knowledge along the secondary dimension (i.e., the class dimension), which enables such information to be explicitly transferred as auxiliary knowledge alongside the matching of inter-sample relations.", + "bbox": [ + 89, + 568, + 482, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, we have constructed a dense relation graph $\\mathcal{G}^{IS}$ , which comprises $B$ vertices and $B \\times B$ edges: $\\mathcal{G}^{IS} = (\\mathcal{V}^{IS}, \\mathcal{E}^{IS})$ . Each vertex in $\\mathcal{G}^{IS}$ represents the prediction vector $\\mathbf{z} \\in \\mathbb{R}^C$ for one instance within a batch, and is connected to all instances within the batch including itself. The attribute of edge $\\mathcal{E}_{i,j}^{IS}$ connecting vertices $i$ and $j$ describes the class-wise relations between the predictions of instances $i$ and $j$ . In practice, we can organise all edges into matrix", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{E}^{IS}\\in \\mathbb{R}^{B\\times B\\times C}$ , in which:", + "bbox": [ + 511, + 219, + 697, + 234 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {i, j} ^ {I S} = \\frac {\\mathbf {z} _ {i} - \\mathbf {z} _ {j}}{\\| \\mathbf {z} _ {i} - \\mathbf {z} _ {j} \\| _ {2}} \\in \\mathbb {R} ^ {C} \\quad \\text {f o r} \\quad i, j \\in [ 1, B ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 247, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we empirically find that normalisation along the secondary dimension helps regularise relations and improves performance. Concretely, $\\mathcal{E}$ encodes inter-sample class-wise relations within a batch of $B$ training samples. This design is different from and empirically more effective than previous relation formulations.", + "bbox": [ + 511, + 287, + 906, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. Constructing Inter-Class Relations", + "text_level": 1, + "bbox": [ + 511, + 387, + 815, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although $\\mathcal{G}^{IS}$ encodes rich inter-sample relations within a batch, it fails to explicitly model inter-class correlation patterns that are also beneficial structured knowledge [29]. Therefore, we propose to build a novel inter-class batchwise relation graph $\\mathcal{G}^{IC} = (\\mathcal{V}^{IC},\\mathcal{E}^{IC})$ , instead of the Gram matrices in the inner product space as in [29]. The construction of $\\mathcal{G}^{IC}$ mirrors that of $\\mathcal{G}^{IS}$ . Vertices in $\\mathcal{G}^{IC}$ are the class-wise logit vectors $\\mathbf{w}\\in \\mathbb{R}^B$ . Each edge in $\\mathcal{E}^{IC}\\in \\mathbb{R}^{C\\times C\\times B}$ embeds the pairwise difference between the $i$ -th and $j$ -th per-class vectors. Concretely:", + "bbox": [ + 511, + 407, + 906, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {i, j} ^ {I C} = \\frac {\\mathbf {w} _ {i} - \\mathbf {w} _ {j}}{\\| \\mathbf {w} _ {i} - \\mathbf {w} _ {j} \\| _ {2}} \\in \\mathbb {R} ^ {B} \\quad \\text {f o r} \\quad i, j \\in [ 1, C ]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 571, + 903, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our inter-class relations preserve the batch-wise discrepancies by treating them as a dimension of additional knowledge (reciprocal to the case of inter-sample relations), which is unlike any other previous methods [29]. We will demonstrate in Tab. 7 that our novel formulation of intersample and inter-class relations by preserving the raw affinity knowledge along the secondary dimension perform significantly better than previous relation encoders $\\psi(\\cdot)$ via Gram matrices [29, 52, 53, 61] or third-order angular distances [52].", + "bbox": [ + 511, + 612, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.3. Constructing Virtual Relations", + "text_level": 1, + "bbox": [ + 511, + 772, + 785, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each prediction $\\mathbf{z}_i$ within a batch $\\{\\mathbf{z}_i\\}_{i=1}^B$ , we create a virtual view of it, denoted as “ $\\tilde{\\mathbf{z}}_i$ \", by applying semantic-preserving transformations to original image $\\mathbf{x}_i$ . While other transformations are applicable, we choose RandAugment [17] that applies stochastic image transformations (see Supplementary Material for details). With our batch of predictions augmented into $\\{\\mathbf{z}_i, \\tilde{\\mathbf{z}}_i\\}_{i=1}^B$ , we can construct a", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2710", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2dde1c95c67306d36bb73fd00fc93f93592327bd45737ba9311dea63fddb3f87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Teacher\nStudentVenueResNet56 ResNet20ResNet32×4 WRN-40-2 WRN-40-2 VGG13 ShuffleNetV2ResNet32×4 VGG8ResNet32×4 VGG13 MobileNetV2ResNet50 MobileNetV2 ShuffleNetV2ResNet40-2
Teacher72.3479.4275.6175.6174.6479.42
Student69.0672.5073.2671.9870.3671.82
Feature-based
FitNets [56]ICLR'1569.2173.5073.5872.2471.0273.54
AT [74]ICLR'1770.5573.4474.0872.7771.4372.73
CRD [60]ICLR'2071.1675.5175.4874.1473.9475.65
SRRL [68]ICLR'2171.1375.3375.5974.1873.44-
PEFD [12]NeurIPS'2270.0776.0876.0274.9274.35-
TaT [42]CVPR'2271.5975.8976.0674.9774.39-
ReviewKD [10]CVPR'2171.8975.6376.1275.0974.8477.78
NORM [47]ICLR'2371.3576.4975.6574.8273.9578.32
FCFD [45]ICLR'2371.9676.6276.4375.4675.2278.18
Logit-based
KD [27]arXiv'1570.6673.3374.9273.5472.9874.45
TAKD [50]AAAI'2070.8373.8175.1273.7873.2374.82
CTKD [41]AAAI'2371.1973.7975.4573.9373.5275.31
NKD [72]ICCV'2370.4076.3575.2474.0774.8676.26
DKD [81]CVPR'2271.9776.3276.2474.8174.6877.07
LSKD [59]CVPR'2471.4376.6276.1174.3774.3675.56
TTM [82]ICLR'2471.8376.1776.2374.3274.3376.55
CRLD [76]MM'2472.1077.6076.4575.5875.2778.27
Relation-based
RKD [52]CVPR'1969.6171.9073.3572.2271.4873.21
CC [54]CVPR'1969.6372.9773.5672.2170.7171.29
SP [61]ICCV'1969.6772.9473.8372.4372.6874.56
ICKD [46]ICCV'2171.7675.2575.6474.3373.42-
DIST [29]†NeurIPS'2271.7576.31-74.73-77.35
VRM-72.0978.7677.4776.4676.1979.34
", + "bbox": [ + 135, + 88, + 864, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results for same- and different-model teacher-student pairs on CIFAR-100. †: using re-trained, stronger teachers.", + "bbox": [ + 135, + 443, + 857, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "larger inter-sample edge matrix $\\mathcal{E}^{IS}\\in \\mathbb{R}^{2B\\times 2B\\times C}$ and a larger inter-class edge matrix $\\mathcal{E}^{IC}\\in \\mathbb{R}^{C\\times C\\times 2B}$", + "bbox": [ + 88, + 476, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the perspective of sample views, our new $\\mathcal{E}^{IS}$ and $\\mathcal{E}^{IC}$ constructed from a batch of real and virtual samples essentially capture three types of knowledge, namely relations amongst real views (denoted as \"real-real\"), relations amongst virtual views (\"virtual-virtual\"), and relations between pairs of real and virtual views (\"real-virtual\"). For instance, a real-virtual edge that connects real vertex $m$ and virtual vertex $n$ in $\\mathcal{E}^{IS}$ is computed as:", + "bbox": [ + 88, + 508, + 482, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {m, n} ^ {I S} = \\frac {\\mathbf {z} _ {m} - \\tilde {\\mathbf {z}} _ {n}}{\\| \\mathbf {z} _ {m} - \\tilde {\\mathbf {z}} _ {n} \\| _ {2}} \\in \\mathbb {R} ^ {C}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 640, + 482, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.4. Pruning into Sparse Graphs", + "text_level": 1, + "bbox": [ + 89, + 681, + 344, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pruning redundant edges. The augmented $\\mathcal{E}^{IS}$ in Sec. 5.3 contains $2B\\times 2B$ edges, leading to quadrupled overheads. For better efficiency, we prune $\\mathcal{G}^{IS}$ into sparse graphs. Noticing that $\\mathcal{E}^{IS}$ is symmetric along its diagonals, we first prune its redundant half to save up to $50\\%$ edge count. We also remove intra-view edges to get a further $50\\%$ reduction, as we empirically find them redundant and harm knowledge transfer. For $\\mathcal{G}^{IC}$ , we decompose the augmented batch of predictions of size $2B$ into a real-view batch and a virtual-view batch, each of size $B$ , and in lieu use the inter-sample batch-wise affinity vectors between them as its vertices. Compared to their original intra-view formulation of size $2B$ in Sec. 5.3, this design encodes purely inter-view", + "bbox": [ + 89, + 703, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "affinity knowledge with halved parameters. With redundant edge pruning (REP), our graphs now become sparse and the remaining edges can again be rearranged into compact matrices: $\\mathcal{E}^{ISV} \\in \\mathbb{R}^{B \\times B \\times C}$ and $\\mathcal{E}^{ICV} \\in \\mathbb{R}^{C \\times C \\times B}$ .", + "bbox": [ + 511, + 478, + 906, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pruning unreliable edges. To mitigate the diffusive effect of spurious predictions discovered earlier, we further identify and prune the unreliable edges. In previous graph learning works, the absolute certainty of two vertices are often used to determine the reliability of an edge. For instance, REM [9] computes the reliability of an edge as the mean of the maximum predicted probabilities of two samples (i.e., two vertices). However, we argue this could bias the learning towards easy samples. Instead, we measure the discrepancy between two predictions. The larger this discrepancy, the more unreliable the relation constructed. The unreliable edge pruning (UEP) criterion is given by $\\mathcal{E}_{i,j}^{ISV} = \\emptyset$ if $\\mathrm{H}(\\mathbf{z}_i^s,\\tilde{\\mathbf{z}}_j^s) > P_m$ where $\\mathrm{H}(\\cdot)$ computes the joint entropy (JE) between two predictions and $P_{m}$ is the $m$ -th percentile within the batch. Note that the criterion is enforced on student predictions, resulting in adaptive and dynamic pruning as different edges get pruned in each iteration, which improves learning most of the time.", + "bbox": [ + 511, + 545, + 908, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.5. Full Objective", + "text_level": 1, + "bbox": [ + 511, + 830, + 660, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With $\\mathcal{G}^{IS}$ and $\\mathcal{G}^{IC}$ constructed for both teacher and student predictions, our VRM objective matches edge matrices $\\mathcal{E}^{ISV}$ and $\\mathcal{E}^{ICV}$ between teacher and student via distance", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2711", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ce7978ec40fd1f6f030fd6ce981ed98edebed3167630536cbaee6b9366995065.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Teacher StudentVenueResNet34 ResNet18ResNet50 MobileNetV1
Teacher Student73.31/91.4276.16/92.86
69.75/89.0768.87/88.76
Feature-based
AT [74]ICLR'1770.69/90.0169.56/89.33
OFD [25]ICCV'1970.81/89.9871.25/90.34
CRD [60]ICLR'2071.17/90.1371.37/90.41
CAT-KD [23]CVPR'2371.26/90.4572.24/91.13
SimKD [8]CVPR'2271.59/90.4872.25/90.86
ReviewKD [10]CVPR'2171.61/90.5172.56/91.00
SRRL [68]ICLR'2171.73/90.6072.49/90.92
PEFD [12]NeurIPS'2271.94/90.6873.16/91.24
RSD [78]ICCV'2572.18/90.7573.05/91.26
FCFD [45]ICLR'2372.24/90.7473.37/91.35
Logit-based
KD [27]arXiv'1570.66/89.8868.58/88.98
TAKD [50]AAAI'2070.78/90.1670.82/90.01
DKD [81]CVPR'2271.70/90.4172.05/91.05
SDD [63]CVPR'2471.14/90.0572.24/90.71
TTM [82]ICLR'2472.19/-73.09/-
LSKD [59]CVPR'2472.08/90.7473.22/91.59
CRLD [76]MM'2472.37/90.7673.53/91.43
Relation-based
RKD [52]CVPR'1971.34/90.3771.32/90.62
CC [54]CVPR'1970.74/--
ICKD [46]ICCV'2172.19/90.72-
DIST [29]NeurIPS'2272.07/90.4273.24/91.12
VRM-72.67/90.6874.16/91.78
", + "bbox": [ + 138, + 88, + 459, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metric $\\phi (\\cdot)$ , for which the Huber loss is used. Formally, $L_{vrm}^{ISV} = \\phi (\\mathcal{E}_S^{ISV},\\mathcal{E}_T^{ISV})$ and $L_{vrm}^{ICV} = \\phi (\\mathcal{E}_S^{ICV},\\mathcal{E}_T^{ICV})$ The full objective is a weighted combination of the CE loss and the proposed VRM losses:", + "bbox": [ + 89, + 439, + 482, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L _ {c e} + \\alpha L _ {v r m} ^ {I S V} + \\beta L _ {v r m} ^ {I C V} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 508, + 480, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{ce}$ is the CE loss applied to student's predictions of both real and virtual views and supervised by GT; $\\alpha$ and $\\beta$ are balancing scalars. A PyTorch-style pseudo-code for the computing of VRM losses is provided in Algorithm 1 in Supplementary Material.", + "bbox": [ + 89, + 537, + 483, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 627, + 223, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 651, + 297, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is evaluated on CIFAR-100 [35] and ImageNet [19] datasets for image classification, and MSCOCO [43] for object detection. All experimental configurations follow the standard practice in prior works. More details are provided in Supplementary Material.", + "bbox": [ + 89, + 672, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 757, + 230, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on CIFAR-100. The results for different KD methods on CIFAR-100 are shown in Tab. 1. For same-model KD, (left of Tab. 1), VRM surpasses all previous relation-based methods across all teacher-student pairs by large margins, including DIST [29]. Noticeably, VRM significantly outperforms the strongest feature-based method, FCFD [45]. On average, it also performs much better than top logit-based methods such as CRLD [76] and TTM [82].", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3e086eada0defab6125b3cd6c5d118e32ef6737082eb5e50bb575057889cabb.jpg", + "table_caption": [ + "Table 2. Results on ImageNet." + ], + "table_footnote": [], + "table_body": "
ResNet101→ResNet18ResNet50→MobileNetV2
APAP50AP75APAP50AP75
Teacher42.0462.4845.8840.2261.0243.81
Student33.2653.6135.2629.4748.8730.90
Feature-based
FitNets [56]34.4354.1636.7130.2049.8031.69
FGFI [62]35.4455.5138.1731.1650.6832.92
TAKD [50]34.5955.3537.1231.2651.0333.46
ReviewKD [10]36.7556.7234.0033.7153.1536.13
FCFD [45]37.3757.6040.3434.9755.0437.51
Logit-based
KD [27]33.9754.6636.6230.1350.2831.35
CTKD [41]34.5655.4336.9131.3952.3433.10
LSKD [59]---31.7452.7733.40
DKD [81]35.0556.6037.5432.3453.7734.01
Relation-based
VRM35.4656.9137.9332.6753.9634.48
", + "bbox": [ + 513, + 88, + 877, + 280 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/491eae01393c61f99a52c060a2f96b34b5b6ff124e9907387ffe9eea8739c296.jpg", + "table_caption": [ + "Table 3. Results on MS-COCO. All detectors are based on a Faster-RCNN [55] backbone." + ], + "table_footnote": [], + "table_body": "
TeacherStudentT.S.KDATSPLGAutoKDLSKDVRM
ResNet56DeiT-T70.4365.0873.2573.5167.3678.1578.5878.5579.52
ResNet56T2T-ViT-770.4369.3774.1574.0172.2678.3578.6278.4378.88
ResNet56PiT-T70.4373.5875.4776.0374.9778.4878.5178.7679.25
ResNet56PVT-T70.4369.2274.6677.0770.4877.4873.6078.4379.42
", + "bbox": [ + 500, + 333, + 895, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Results for cross-architecture distillation on CIFAR-100.", + "bbox": [ + 500, + 406, + 888, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These are significant, given that IM methods are known to be naturally more adapt at distilling between homogeneous model pairs. For different-model $KD$ (right of Tab. 1), which RM methods are supposed to be more competent with, VRM readily surpasses all RM and IM methods with notable margins. These results highlight VRM's versatility on both homogeneous and heterogeneous model pairs.", + "bbox": [ + 511, + 440, + 906, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on ImageNet. VRM surpasses all feature- and relation-based methods on this large-scale dataset, as shown in Tab. 2. The advantage of VRM is again more apparent over the heterogeneous pair (i.e., ResNet50 → MobileNetV2), whereby it for the first time hits $74.0\\%$ Top-1 accuracy. Notably, VRM outperforms strong competitors such as FCFD [45] and CRLD [76] with comparable or even less computational overheads as compared in Tab. 10.", + "bbox": [ + 511, + 553, + 906, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on MS-COCO. We demonstrate that VRM generalises to more challenging tasks by adapting it to object detection. From Tab. 3, VRM performs better than existing logit-based methods and competitively to feature-based methods. VRM is slightly behind top-performing feature-based methods such as FCFD [45] and ReviewKD [10]. This is a consequence of the very nature of the object detection task, where fine-grained contextual features play a vital role, making feature-based methods inherently better-off [62]. Nonetheless, we experimentally demonstrate that VRM is effective on object detection, whereby it improves the baseline by $2 - 3\\%$ AP, surpasses all logit-based methods, and is on par with strong feature-based methods.", + "bbox": [ + 511, + 681, + 908, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results for cross-architecture distillation. We follow the", + "bbox": [ + 513, + 885, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2712", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f37391f0572961781b805f213a899138618759a6b8ffcf51103f385f3331240a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ComponentRN32×4VGG13
RN8×4VGG8
Baseline72.5070.36
+ Match {EIS}75.6474.08
+ Match {EIS, IIC}75.9474.47
+ Match {EISV}77.6375.30
+ Match {EISV, IICV}78.3875.56
+ Pruning78.7676.19
", + "bbox": [ + 107, + 88, + 343, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d6fe391660ea9000943952d2441732830b09ea7cc4fb5d4cee5f58dbcf45a01f.jpg", + "table_caption": [ + "Table 5. Ablation of major designs." + ], + "table_footnote": [], + "table_body": "
Matching ChoiceRN32×4 RN8×4VGG13 VGG8RN50 MNV2
Match ∅72.5070.3668.33
Match {E}78.7676.1972.30
Match {E, V}78.6375.6070.97
Match {V}78.4675.7768.33
", + "bbox": [ + 367, + 92, + 596, + 176 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/53000251d04dcc574d08a5277d6b788d45661e46ff2a3a8b6cae0a97d82571f9.jpg", + "table_caption": [ + "Table 6. Effect of matching vertices." + ], + "table_footnote": [], + "table_body": "
ψ(·)Shape of Rel. MatrixMethodRN32×4 VGG13RN50
RN8×4VGG8MNV2
Gram matrices[B,B] & [C,C]CC, SP, PKT, DIST77.3175.1270.28
Angular relations[B,B,B]RKD77.1475.0969.65
ISV & ICV[B,B,C] & [C,C,B]VRM78.4675.3072.10
", + "bbox": [ + 616, + 88, + 888, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc63dc419c51672785c2d08b9e795142d515588dc7df4a77a7a7b5aaa234888e.jpg", + "table_caption": [ + "Table 7. Comparing different relation functions." + ], + "table_footnote": [], + "table_body": "
α3264128256512
Acc.78.0378.1378.7678.3277.84
β8163264128
Acc.78.4078.5178.7678.8678.60
n01234
Acc.77.9278.3878.7678.5578.43
", + "bbox": [ + 101, + 223, + 344, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Effect of different $\\alpha ,\\beta$ ,and $n$ .", + "bbox": [ + 99, + 327, + 336, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8057ee8253a401617275712487a0a801df9e72b95d91266ecb296ec831c94fd9.jpg", + "image_caption": [ + "Figure 5. Effect of different difficulty pairs." + ], + "image_footnote": [], + "bbox": [ + 351, + 220, + 604, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff7bbbfb266c386c21ebb690185ca2051f2f760e948fcd35de637e19d029ea83.jpg", + "image_caption": [ + "Figure 6. Robustness to varying $B$ ." + ], + "image_footnote": [], + "bbox": [ + 622, + 220, + 879, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "settings in [39] and [59] to perform CNN-to-ViT cross-architecture distillation. Tab. 4 shows that VRM sets new record under this set-up for different ViT architectures. Notably, by simply replacing the vanilla KD loss with the proposed VRM objective and making no other modifications, VRM improves DeiT-T's accuracy by $14.44\\%$ , leading the relation-based SP [61] by $12.16\\%$ and even surpassing AutoKD [40] with AutoML search.", + "bbox": [ + 88, + 359, + 483, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 494, + 254, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations of main design choices. In Tab. 5, starting from the baseline where only the CE loss is applied, we gradually incorporate each of our designs. As shown, each extra design consistently brings performance gains, which corroborate the validity of individual design choices.", + "bbox": [ + 89, + 517, + 483, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different relation encoding functions. Tab. 7 compares our proposed relation encoding scheme with existing ones. To demonstrate the superiority of our formulation, we substitute it with existing Gram matrices [29, 53, 54, 61] or angle-wise relations [52]. Results show that our inter-sample class-wise and inter-class sample-wise relations consistently yield much better performance than existing relations, which highlights VRM's contribution of a novel and superior type of relation encoding in addition to the introduction of virtual relation.", + "bbox": [ + 89, + 599, + 483, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of varying hyperparameters. VRM involves three major hyperparameters: $\\alpha$ and $\\beta$ , and $n$ . According to Tab. 8, VRM works generally well with $\\alpha$ and $\\beta$ in a reasonable range. Larger $\\alpha$ and $\\beta$ may produce better results for certain model pairs but worse results for others. For $n$ , a larger $n$ means more difficult images and larger inter-view prediction discrepancy. From Tab. 8, $n = 2$ works best, whereas other values also report competitive results.", + "bbox": [ + 89, + 757, + 483, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different real-virtual difficulty pairs. We also", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "experiment with different strength combinations for generating the real and virtual samples. The results are presented in Fig. 5, where \"Weak\" denotes the default transformation used in previous methods, and \"Strong\" denotes RandAugment with $n = 2$ . Overall, we conclude that 1) a moderate discrepancy leads to optimal performance, and 2) the discrepancy is significant to the success of VRM.", + "bbox": [ + 511, + 359, + 905, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to varying batch sizes. Relational methods are known to be sensitive to training batch size $B$ . We conduct experiments to examine how robust VRM is against varying $B$ . To adjust the learning rate accordingly, we consider two LR scaling rules: linear LR scaling [21] and square root LR scaling [11]. From Fig. 6, VRM yields competitive results across a wide range of $B$ values. In contrast, DIST [29] degrades significantly as $B$ grows.", + "bbox": [ + 511, + 472, + 906, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 601, + 679, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of training dynamics. Figs. 3a and 3b plot the per-epoch training and validation set accuracies throughout training. VRM maintains a all-time lead in both training and validation performance, with faster convergence. Besides, while VRM's lead in training performance tapers off towards the end of training, it remains more substantial, if not further enlarging, in validation performance, which highlights its superior generalisation properties. This analysis also shows our designs are effective in mitigating the issues with existing RM methods identified in the pilot studies.", + "bbox": [ + 511, + 622, + 906, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of spurious gradients. We revisit our second pilot study by analysing how spurious gradients are suppressed in real ImageNet training. From Fig. 4, similar gradient diffusion patterns are observed in real distillation. Samples semantically more similar to the sample with spurious predictions are more affected by spurious gradients (e.g., the dog image with a watery background resembles the bubbly texture in another sample that is actually a computer mouse.).", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2713", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e84ea35fef3f4f65b04d72ce95b992dfe3d0187817038425e616407067f19a22.jpg", + "image_caption": [ + "(a) RKD" + ], + "image_footnote": [], + "bbox": [ + 98, + 95, + 194, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63b01b1c402a6c63364bd81dce0e80a5d9866341c766c6e4279870c0ea8aa8bc.jpg", + "image_caption": [ + "(b) VRM", + "Figure 7. t-SNE (a-b) and loss landscape (c-d) visualisations for $\\mathrm{RN8\\times 4}$ students distilled with an $\\mathrm{RN32\\times 4}$ teacher via RKD and VRM." + ], + "image_footnote": [], + "bbox": [ + 200, + 95, + 295, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6138ea77c96853fdbdd4379a8f7942b8d118cf73678d2e190471b4df95241e85.jpg", + "image_caption": [ + "(c) RKD" + ], + "image_footnote": [], + "bbox": [ + 303, + 95, + 413, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18d18218b903574cecce0e17c870c55cd2061ae5253df19b4a98e6607c5783d4.jpg", + "image_caption": [ + "(d) VRM" + ], + "image_footnote": [], + "bbox": [ + 423, + 95, + 534, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/293d0b4c4036c64333aa98be682431c579e9b009937ba707b1a26755e951f227.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodKDFitNetsRKDICKDDISTCRDReviewKD
Train. Time (ms)24.926.731.028.027.241.239.9
Peak GPU Mem. (MB)32333033038133014181042
MethodSDDMLLDCRLDNORMPEFDFCFDVRM
Train. Time (ms)34.257.243.235.136.256.447.2
Peak GPU Mem. (MB)6905765511806701953579
", + "bbox": [ + 94, + 220, + 485, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualisation of embedding space. We conduct t-SNE analysis on student penultimate layer embeddings learnt via different methods. As presented in Fig. 7a and 7b, VRM leads to more compact per-class clusters with clearer inter-class separation and less stray points. These imply better student features distilled by VRM for the downstream task.", + "bbox": [ + 89, + 333, + 483, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualisation of loss landscape. We further analyse the generalisation and convergence properties of our method through the lens of visualised loss landscape [38]. From Fig. 7c and 7d, compared to RKD, VRM has a wider, flatter, and deeper region of minima – typical hint of better model generalisation and robustness; this wide convexity basin surrounded by salient pikes in various directions indicates excellent convergence properties. More visualisations are provided in Supplementary Material.", + "bbox": [ + 89, + 431, + 483, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vertex matching. Within the constructed graphs $\\mathcal{G}^{IS}$ and $\\mathcal{G}^{IC}$ , VRM matches edges $\\mathcal{E}^{ISV}$ and $\\mathcal{E}^{ICV}$ which carry relational knowledge. By extension, we can naturally expect vertices $\\nu$ to be also transferred. From Tab. 6, matching vertices is not as effective. The advantage of matching relations (i.e., edges) is more pronounced for heterogeneous KD pairs such as ResNet50 $\\rightarrow$ MobileNetV2. Adding vertex matching to the proposed edge matching does not improve but instead degrade performance. We argue this is because introducing vertex matching objectives make the matching criterion more stringent which is against our motivation of using more slackened matching.", + "bbox": [ + 89, + 575, + 483, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Role of transformation operations. We emphasise the strong performance of VRM is not a simple outcome of the transformations involved. This can be verified by the results in Fig. 5 and Tab. 9, where using only the default weak transformation (i.e., random crop and horizontal flip) for both views achieves highly competitive and even superior performance under VRM. In fact, in the case of \"Weak-Weak\", both views still differ because of the stochastic operations used. Essentially, the success of VRM lies exactly", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c9131fcb620a108edff397a910f14885bf0d2ec79ab70861b0d5c9255260e62a.jpg", + "table_caption": [ + "Table 10. Training efficiency of different distillation methods." + ], + "table_footnote": [], + "table_body": "
RealVirtualRN50→MNV2RN34→RN18
DefaultDefault74.16/91.7872.67/90.68
DefaultRandAug(n=0)74.10/91.7072.60/90.88
DefaultRandAug(n=1)74.13/91.6372.59/90.82
DefaultRandAug(n=2)74.13/91.6872.48/90.73
", + "bbox": [ + 568, + 99, + 887, + 171 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/33724832b6d61ae537652ef225ff4e12b0ed977c7c593d820866f860b4218477.jpg", + "table_caption": [ + "Table 9. Effect of different transformation strength pairs on ImageNet." + ], + "table_footnote": [], + "table_body": "
PruningRN32×4→RN8×4V13→V8W40-2→SNV2
Acc.tbatch(s)Mem. (MB)Acc.Acc.
w/o78.3861.434.375.5678.45
+ REP78.4646.313.675.3078.41
+ UEP78.7647.213.176.1978.62
", + "bbox": [ + 511, + 219, + 901, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11. Effectiveness and efficiency of pruning operations.", + "bbox": [ + 522, + 300, + 883, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the discrepancy of teacher and student predictions, where regularisation comes into crucial play. On smaller and easier datasets such as CIFAR-100, the gap between teacher and student predictions are relatively small, and manually enlarging this gap via extra transformations leads to promising outcomes. Whereas the more difficult ImageNet data already sees large teacher-student prediction discrepancy, such that it benefits less from further amplification.", + "bbox": [ + 511, + 334, + 903, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training efficiency. From Tab. 10, VRM is reasonably efficient compared to existing algorithms, including some leading feature- [12, 45, 47] and logit-based [33, 63] methods. VRM does not introduce any extra overheads at inference. We also breakdown the proposed pruning designs in Tab. 11 to understand how they individually impact the performance and efficiency of VRM. We see REP boosts efficiency at minimal performance cost, whereas UEP trades minimal efficiency cost for performance.", + "bbox": [ + 511, + 460, + 905, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More experiments, analyses, and discussions are provided in Supplementary Material.", + "bbox": [ + 511, + 601, + 903, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 645, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents VRM, a novel knowledge distillation framework that constructs and transfers virtual relations. Our designs are motivated by a set of pilot experiments, from which we identified two main cruxes with existing relation-based KD methods: their tendency to overfit and susceptibility to adverse gradient propagation. A series of tailored designs are developed and are shown to successfully mitigate these issues. We have conducted extensive experiments on different tasks and multiple datasets and verified VRM's validity and superiority in diverse settings, whereby it consistently achieves state-of-the-art performance. We hope that this work could renew the community's interest in relation-based knowledge distillation, and encourage more systematic reassessment of the design principles of such solutions.", + "bbox": [ + 511, + 670, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2714", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 165, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sungsoo Ahn, Shell Xu Hu, Andreas Damianou, Neil D. Lawrence, and Zhenwen Dai. Variational information distillation for knowledge transfer. In CVPR, 2019. 2, 1, 3", + "[2] Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning optical flow from still images. In CVPR, 2021. 3", + "[3] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. NeurIPS, 2014. 6", + "[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 6", + "[5] Joao Carreira, Abhishek Kar, Shubham Tulsiani, and Jitendra Malik. Virtual view networks for object reconstruction. In CVPR, 2015. 3", + "[6] Jiahao Chang, Shuo Wang, Hai-Ming Xu, Zehui Chen, Chenhongyi Yang, and Feng Zhao. Detrdistill: A universal knowledge distillation framework for detr-families. In ICCV, 2023. 1", + "[7] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2, 1, 3", + "[8] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, 2022. 6", + "[9] Peibin Chen, Tao Ma, Xu Qin, Weidi Xu, and Shuchang Zhou. Data-efficient semi-supervised learning by reliable edge mining. In CVPR, 2020. 5", + "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, 2021. 5, 6, 1, 2, 3", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 2020. 3, 7, 6", + "[12] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved feature distillation via projector ensemble. In NeurIPS, 2022. 5, 6, 8, 1, 2", + "[13] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 3", + "[14] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. In ICLR, 2023. 1", + "[15] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. *ICLR*, 2022. 2", + "[16] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In CVPR, 2019. 3" + ], + "bbox": [ + 93, + 189, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In NeurIPS, 2020. 3, 4", + "[18] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In CVPR, 2023. 1", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 2", + "[20] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 3", + "[21] P Goyal. Accurate, large minibatch sgd: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 7", + "[22] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In ICLR, 2024. 1", + "[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, 2023. 2, 6, 1, 3", + "[24] Zhiwei Hao, Jianyuan Guo, Kai Han, Yehui Tang, Han Hu, Yunhe Wang, and Chang Xu. One-for-all: Bridge the gap between heterogeneous architectures in knowledge distillation. NeurIPS, 2024. 2", + "[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2, 6, 1, 3", + "[26] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, 2019. 1, 2, 3", + "[27] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowlegde in a neural network. In arXiv:1503.02531, 2015. 1, 2, 3, 5, 6, 4", + "[28] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. In arXiv:1704.04861, 2017. 3", + "[29] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In NeurIPS, 2022. 2, 3, 4, 5, 6, 7, 1", + "[30] Yuge Huang, Jiaxiang Wu, Xingkun Xu, and Shouhong Ding. Evaluation-oriented knowledge distillation for deep face recognition. In CVPR, 2022. 1", + "[31] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. In arXiv:1707.01219, 2017. 1, 2, 3", + "[32] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: structural and temporal cross-modal distillation for multi-view 3d object detection. NeurIPS, 2024. 2", + "[33] Ying Jin, Jiaqi Wang, and Dahua Lin. Multi-level logit distillation. In CVPR, 2023. 2, 8", + "[34] Minha Kim, Shahroz Tariq, and Simon S Woo. Fretal: Generalizing deepfake detection using knowledge distillation and representation learning. In CVPR, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2715", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6, 2", + "[36] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In ECCV, 2020. 3", + "[37] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 6", + "[38] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 8", + "[39] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, 2022. 7, 1, 3", + "[40] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In ICCV, 2023. 7, 1, 3", + "[41] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, 2023. 2, 5, 6, 1, 3", + "[42] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge distillation via the target-aware transformer. In CVPR, 2022. 2, 5, 1", + "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6, 2", + "[44] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3", + "[45] Dongyang Liu, Meina Kan, Shiguang Shan, and Xilin Chen. Function-consistent feature distillation. In ICLR, 2023. 1, 5, 6, 8, 2, 3", + "[46] Li Liu, Qingle Huang, Sihao Lin, Hongwei Xie, Bing Wang, Xiaojun Chang, and Xiaodan Liang. Exploring inter-channel correlation for diversity-preserved knowledge distillation. In ICCV, 2021. 2, 5, 6, 1", + "[47] Xiaolong Liu, Lujun Li, Chao Li, and Anbang Yao. Norm: Knowledge distillation via n-to-one representation matching. In ICLR, 2023. 1, 5, 8, 2, 3", + "[48] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In CVPR, 2019. 2", + "[49] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. NeurIPS, 2021. 3", + "[50] Seyed-Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 2, 5, 6, 1, 3", + "[51] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 3", + "[52] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In IEEE TMM, 2019. 1, 2, 4, 5, 6, 7, 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2, 4, 7, 1, 3, 5", + "[54] Baoyun Peng, Xiao Jin, Jiaheng Liu, Shunfeng Zhou, Yichao Wu, Yu Liu, Dongsheng Li, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, 2019. 2, 4, 5, 6, 7, 1, 3", + "[55] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE T-PAMI, 39(6):1137-1149, 2016. 6", + "[56] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1, 2, 3, 5, 6", + "[57] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semisupervised learning with consistency and confidence. In NeurIPS, 2020. 3", + "[58] Wonchul Son, Jaemin Na, Junyong Choi, and Wonjun Hwang. Densely guided knowledge distillation using multiple teacher assistants. In ICCV, 2021. 2", + "[59] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, 2024. 2, 5, 6, 7, 1, 3", + "[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 5, 6, 1, 3", + "[61] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, 2019. 1, 2, 4, 5, 7, 3", + "[62] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 1, 6, 3", + "[63] Shicai Wei, Chunbo Luo, and Yang Luo. Scale decoupled distillation. In CVPR, 2024. 6, 8, 1, 3", + "[64] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In CVPR, 2023. 1, 3", + "[65] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 6", + "[66] Chuanguang Yang, Zhulin An, Linhang Cai, and Yongjun Xu. Hierarchical self-supervised augmented knowledge distillation. In IJCAI, 2021. 6", + "[67] Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, and Qian Zhang. Cross-image relational knowledge distillation for semantic segmentation. In CVPR, 2022. 1, 2", + "[68] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 5, 6, 1, 2", + "[69] Longrong Yang, Xianpan Zhou, Xuewei Li, Liang Qiao, Zheyang Li, Ziwei Yang, Gaoang Wang, and Xi Li. Bridging cross-task protocol inconsistency for distillation in dense object detection. In ICCV, 2023. 1", + "[70] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and global knowledge distillation for detectors. In CVPR, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2716", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked generative distillation. In ECCV, 2022. 1, 3", + "[72] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. In ICCV, 2023. 2, 5, 1, 3", + "[73] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Km. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2, 1", + "[74] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 5, 6, 1, 3", + "[75] Bo Zhang, Jiacheng Sui, and Li Niu. Foreground object search by distilling composite image feature. In ICCV, 2023. 1", + "[76] Weijia Zhang, Dongnan Liu, Weidong Cai, and Chao Ma. Cross-view consistency regularisation for knowledge distillation. In ACM MM, 2024. 1, 5, 6, 2, 3", + "[77] Weijia Zhang, Dongnan Liu, Chao Ma, and Weidong Cai. Alleviating foreground sparsity for semi-supervised monocular 3d object detection. In WACV, 2024. 1", + "[78] Weijia Zhang, Yuehao Liu, Wu Ran, and Chao Ma. Cross-architecture distillation made simple with redundancy suppression. ICCV, 2025. 2, 6, 1", + "[79] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 3", + "[80] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, 2018. 1, 2, 3", + "[81] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 2, 5, 6, 1, 3, 4", + "[82] Kaixiang Zheng and En-Hui Yang. Knowledge distillation based on transformed teaching matching. In ICLR, 2024. 2, 5, 6, 1, 3", + "[83] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In CVPR, 2023. 1" + ], + "bbox": [ + 91, + 92, + 480, + 688 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2717", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_model.json b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4522eac3597d6e2cfb26af5cb18aa51ef7cee62 --- /dev/null +++ b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_model.json @@ -0,0 +1,2873 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "VRM: Knowledge Distillation via Virtual Relation Matching" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.18, + 0.339, + 0.199 + ], + "angle": 0, + "content": "Weijia Zhang1" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.181, + 0.462, + 0.198 + ], + "angle": 0, + "content": "Fei Xie" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.2, + 0.505, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.181, + 0.632, + 0.199 + ], + "angle": 0, + "content": "Weidong Cai" + }, + { + "type": "text", + "bbox": [ + 0.688, + 0.181, + 0.781, + 0.198 + ], + "angle": 0, + "content": "Chao \\(\\mathbf{Ma}^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.2, + 0.749, + 0.218 + ], + "angle": 0, + "content": "2 The University of Sydney" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.219, + 0.787, + 0.234 + ], + "angle": 0, + "content": "{weijia.zhang, feixie, chaoma}@sjtu.edu.cn tom.cai@sydney.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.695 + ], + "angle": 0, + "content": "Knowledge distillation (KD) aims to transfer the knowledge of a more capable yet cumbersome teacher model to a lightweight student model. In recent years, relation-based KD methods have fallen behind, as their instance-matching counterparts dominate in performance. In this paper, we revive relational KD by identifying and tackling several key issues in relation-based methods, including their susceptibility to overfitting and spurious responses. Specifically, we transfer novelly constructed affinity graphs that compactly encapsulate a wealth of beneficial inter-sample, inter-class, and inter-view correlations by exploiting virtual views and relations as a new kind of knowledge. As a result, the student has access to richer guidance signals and stronger regularisation throughout the distillation process. To further mitigate the adverse impact of spurious responses, we prune the affinity graphs by dynamically detaching redundant and unreliable edges. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO datasets demonstrate the superior performance of the proposed virtual relation matching (VRM) method, where it consistently sets new state-of-the-art records over a range of models, architectures, tasks, and set-ups. For instance, VRM for the first time hits \\(74.0\\%\\) accuracy for ResNet50 \\(\\rightarrow\\) MobileNetV2 distillation on ImageNet, and improves DeiT by \\(14.44\\%\\) on CIFAR-100 with a ResNet56 teacher. The code and models are released at https://github.com/VISION-SJTU/VRM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.727, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Deep learning is achieving incredible performance at the cost of increasing model complexity and overheads. As a consequence, large and cumbersome neural models struggle to work in resource-constrained environments. Knowledge distillation (KD) [27] has been proposed to address this issue by transferring the knowledge of larger and more capable models to smaller and lightweight ones that are resource-friendly. KD works by minimising the dis" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.707, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.371, + 0.671, + 0.383 + ], + "angle": 0, + "content": "(a) Instance Matching" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.269, + 0.905, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.37, + 0.869, + 0.383 + ], + "angle": 0, + "content": "(b) Relation Matching" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.387, + 0.708, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.488, + 0.659, + 0.501 + ], + "angle": 0, + "content": "(c) VRM (Dense)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.387, + 0.905, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.488, + 0.859, + 0.501 + ], + "angle": 0, + "content": "(d) VRM (Pruned)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.506, + 0.907, + 0.534 + ], + "angle": 0, + "content": "Figure 1. Conceptual illustration of VRM compared to existing KD methods based on instance matching and relation matching." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.614 + ], + "angle": 0, + "content": "tance between compact representations of knowledge extracted from the teacher and student models. It has found widespread application across a spectrum of downstream tasks [18, 22, 30, 34, 62, 64, 67, 70, 75, 77, 83]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.765 + ], + "angle": 0, + "content": "According to the type of knowledge representations to be transferred, KD methods can be broadly categorised into feature-based [56], logit-based [27], and relation-based [52] approaches. The former two directly match the feature maps or logit vectors produced by the teacher and student models for each training sample, which is essentially instance matching (IM). By contrast, relation matching (RM) methods construct and match structured relations extracted within a batch of model responses. A conceptual illustration is presented in Figs. 1a and 1b." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Instance matching has been the prevailing distillation approach in recent years. Popular KD benchmarks see a dominance by IM methods such as FCFD [45], NORM [47], and CRLD [76], with different downstream tasks successfully tackled by directly adopting IM-based distillation [6, 14, 62, 69]. Yet, recent studies discovered that relational knowledge is more robust to variations in neural architectures, data modalities, and tasks [52, 61]. Meanwhile, methods transferring relations have also achieved promising" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2707" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "performance for a range of tasks, including but not limited to segmentation [67] and detection [15, 32]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.123, + 0.482, + 0.289 + ], + "angle": 0, + "content": "Despite growing interest, relation-based methods still fall significantly short compared to their instance matching counterparts. Even the strongest RM method has been outperformed easily by recent IM solutions [29] (see Tabs. 1 and 2). RM-based methods also struggle with more challenging tasks such as object detection [29]. Moreover, previous RM solutions are primarily limited to matching inter-sample [29, 53, 61], inter-class [29], or inter-channel [46, 73] relations via simple Gram matrices. To our best knowledge, no different forms of relations other than these have been proposed since the work of DIST [29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.483, + 0.442 + ], + "angle": 0, + "content": "This paper fills this gap with a new kind of relations for KD - inter-view relations (Fig. 1c), which seamlessly and compactly integrate with previous inter-sample and inter-class relations. Our designs are motivated by two important observations made about RM methods in a set of pilot experiments: 1) RM methods are more susceptible to overfitting than IM methods; 2) RM methods are subject to an adverse gradient propagation effect. We empirically find that incorporating richer and more diverse relations into the matching objective helps mitigate both issues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.611 + ], + "angle": 0, + "content": "To this end, we generate virtual views of samples through simple transformations, followed by constructing virtual affinity graphs and transferring the virtual relations between real and virtual samples along the edges. In lieu of Gram matrices that suffer from significant knowledge loss [29, 53, 61], we preserve the raw relations along the secondary dimension as auxiliary knowledge which adds to the types and density of relational knowledge transferred. Moreover, we also prune our affinity graphs by striping away both redundant and unreliable edges to further alleviate the propagating gradients of spurious samples (Fig. 1d)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.793 + ], + "angle": 0, + "content": "The above insights and remedies altogether lead to a novel Virtual Relation Matching (VRM) framework for knowledge distillation. VRM is conceptually simple, easy to implement, and devoid of complicated training procedures. It is capable of transferring rich, sophisticated knowledge robust to overfitting and spurious signals. VRM sets new state-of-the-art performance for different datasets, tasks, and settings. Perhaps more significant is that VRM makes relation-based methods regain competitiveness and back in the lead over instance matching approaches in different scenarios. To summarise, the contributions of this work include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.869 + ], + "angle": 0, + "content": "- We make an early effort to present comparative analyses of existing KD methods through the lens of training dynamics and sample-wise gradients, and identify overfitting and spurious gradient diffusion as two main cruxes in relational KD methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We distill richer, more diverse relations by generating virtual views, constructing virtual affinity graphs, and" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "matching virtual relations. We also for the first time tackle relational KD with considerations of spurious samples and gradients by pruning redundant and unreliable edges, alongside designs to relax the matching criterion." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- We present the streamlined VRM framework for knowledge distillation, with extensive experimental results on a diversity of models and tasks to highlight its superior performance, alongside rigorous analyses on the soundness and efficiency of our designs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.243, + 0.655, + 0.259 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.905, + 0.542 + ], + "angle": 0, + "content": "KD via instance matching. KD was first proposed in Hinton et al. [27], where a student is trained to mimic the prediction of a pre-trained teacher model for each sample. Follow-up works have mostly followed such instance matching (IM) paradigm, and can be categorised into logit (or prediction)-based and feature-based methods according to what is matched. Logit-based KD has evolved from using adaptive-softened logits [33, 41] to decoupling target- and non-target logits [24, 72, 81] and applying logit transformation [59, 82]. Others [50, 58] set up auxiliary ad-hoc networks between teacher and student to facilitate logit transfer. Feature-based methods minimise the distance between the feature maps [1, 25, 56] or salient regions within in features [23, 74] from specified layers in both teacher and student networks. Some also design sophisticated distillation paths [7, 42]. All these methods are based on instance-wise transfer of knowledge, and are herein referred to as the \"IM\" methods, as illustrated by Fig. 1a." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.547, + 0.905, + 0.804 + ], + "angle": 0, + "content": "KD via relation matching. Some works transfer instead mutual relations mined amongst the network outputs extracted from a batch of training instances. These relation matching (RM) methods usually involve constructing network outputs into compact relation representations that encode rich higher-order information, as depicted in Fig. 1b. Different relation encoding functions mare, including inter-sample [29, 48, 52-54, 61], inter-class [29], inter-channel [46, 73, 78], inter-layer [48], and contrastive [60] relations. To date, IM solutions have dominated KD with their superior performance, leading top-performing relation-based methods by considerable margins. While many new IM methods are found within the last two years, frustratingly few RM solutions are being proposed. In this work, we strive to close this gap with a new kind of relations for KD — inter-view virtual relations, and revive relation-based KD by making it overtake its IM counterparts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Learning with virtual knowledge. While not a standalone research topic, learning with virtual knowledge finds relevance in a variety of learning-based problems. For instance, a commonly used paradigm in 3D vision tasks is to learn (or construct) from the raw data a virtual view or representation as auxiliary knowledge in solving the main task, rang-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2708" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.353, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.192, + 0.337, + 0.203 + ], + "angle": 0, + "content": "(a) Increasing performance on training set." + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.091, + 0.615, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.192, + 0.603, + 0.203 + ], + "angle": 0, + "content": "(b) Plateaued performance on validation set." + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.092, + 0.901, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.191, + 0.902, + 0.203 + ], + "angle": 0, + "content": "(c) Diffusion of gradient perturbations within a batch." + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.207, + 0.815, + 0.221 + ], + "angle": 0, + "content": "Figure 2. Pilot studies that reveal the overfitting and spurious gradient diffusion issues with RM-based KD." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.24, + 0.485, + 0.483 + ], + "angle": 0, + "content": "ing from object reconstruction [5] and optical flow [2] to 3D semantic segmentation [36], monocular 3D object detection [13], and 3D GAN inversion [64]. More broadly, many data-efficient learning methods also share the spirit of utilising virtual knowledge. For instance, self-supervised learning methods generate virtual views of the unlabelled data to enable the learning of pretext tasks [11, 20]. Another popular paradigm is transformation-invariant representation learning [51, 57] in semi-supervised learning and domain adaptation. It enforces consistency between representations learnt for a raw sample and a virtual view of it. The virtual view is often obtained by applying semantic-preserving transformations to the raw sample [16, 17]. This work is more related to this later paradigm, but involves learning with virtual knowledge in a different context, via different approaches, and for a different problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.229, + 0.51 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.627 + ], + "angle": 0, + "content": "KD methods generally employ a cross-entropy (CE) loss and a distillation loss to supervise student learning. The CE loss is computed between student logits \\(\\mathbf{z}_i^s\\) for each sample and its ground-truth label \\(\\mathbf{y}_i\\). The distillation loss matches teacher and student outputs via a distance metric \\(\\phi(\\cdot)\\). In vanilla KD [27], \\(\\phi(\\cdot)\\) is the Kullback-Leibler divergence (KLD) between teacher logits \\(\\mathbf{z}^t\\) and student logits \\(\\mathbf{z}^s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.631, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {i} ^ {\\mathrm {K D}} = \\phi_ {\\mathrm {K L D}} \\left(\\mathbf {z} _ {i} ^ {s}, \\mathbf {z} _ {i} ^ {t}\\right) \\\\ = \\tau^ {2} \\sum_ {j = 1} ^ {C} \\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {t} / \\tau\\right) \\log \\frac {\\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {t} / \\tau\\right)}{\\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {s} / \\tau\\right)}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.781 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the Softmax operation with temperature parameter \\(\\tau\\), and \\(C\\) is the number of classes. For feature-based methods, \\(\\phi (\\cdot)\\) can be the mean squared error (MSE) between teacher and student feature maps for each instance [56], i.e., \\(\\mathcal{L}_i^{\\mathrm{KD}} = \\phi_{\\mathrm{MSE}}(\\mathbf{f}_i^s,\\mathbf{f}_i^t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.871 + ], + "angle": 0, + "content": "For relation-based methods, a relation encoding function \\(\\psi(\\cdot)\\) first abstracts teacher or student outputs of all instances within a training batch into a relational representation, before applying \\(\\phi(\\cdot)\\) to match these relational representations between the teacher and the student. For example, the KD objectives of DIST [29] take the general form of:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.88, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {K D}} = \\phi \\left(\\psi \\left(\\mathbf {z} _ {1} ^ {s}, \\mathbf {z} _ {2} ^ {s}, \\dots , \\mathbf {z} _ {B} ^ {s}\\right), \\psi \\left(\\mathbf {z} _ {1} ^ {t}, \\mathbf {z} _ {2} ^ {t}, \\dots , \\mathbf {z} _ {B} ^ {t}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.24, + 0.905, + 0.287 + ], + "angle": 0, + "content": "where \\(B\\) is the batch size, and subscript \\(i\\) of \\(\\mathcal{L}^{\\mathrm{KD}}\\) is dropped as the loss is computed for a batch of instances. DIST uses both inter-class and inter-sample relation encoders as \\(\\psi (\\cdot)\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.298, + 0.644, + 0.314 + ], + "angle": 0, + "content": "4. Pilot Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.566 + ], + "angle": 0, + "content": "Training dynamics of KD methods. We examine the training dynamics of different relation-based methods on CIFAR-100 with ResNet32 \\(\\times\\) 4 \\(\\rightarrow\\) ResNet8 \\(\\times\\) 4 as the teacher-student pair. In Figs. 2a and 2b, we immediately notice that relational methods achieve significantly higher training accuracy, but they only marginally lead or even fall short in test accuracy compared to IM-based KD [27]. We hypothesise that relational methods are more prone to overfitting. This is expected given that the optimality of IM matching (cond. \\(\\mathcal{A}\\)) implies the optimality of relation matching (cond. \\(\\mathcal{B}\\)), while the converse does not hold. Concretely, \\(\\mathcal{A} \\Rightarrow \\mathcal{B} \\land \\neg \\mathcal{B} \\Rightarrow \\neg \\mathcal{A}\\). In other words, relation matching is a weaker and less constrained objective than IM, which makes the student more readily fit the teaching signals and not generalise well. Thus, we conclude that \\(C1\\): relation matching methods are more prone to overfitting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.907, + 0.768 + ], + "angle": 0, + "content": "Gradient analysis of KD methods. We investigate the gradient patterns within a batch when a spurious sample produces a major misguiding signal. To this end, we first generate two random vectors \\(\\mathbf{x},\\mathbf{y}\\sim \\mathcal{N}(0,1)\\) for \\(\\mathbf{x},\\mathbf{y}\\in \\mathbb{R}^{B\\times D}\\), where \\(B\\) is the batch size and \\(D\\) is the dimension of persample predictions, \\(\\mathbf{x}\\) is taken as the sample-wise predictions, and \\(\\mathbf{y}\\) the supervision signals. We then add a noise vector \\(\\epsilon = c\\cdot \\mathbf{z}\\) to \\(\\mathbf{x}_t\\), where \\(\\mathbf{z}\\sim \\mathcal{N}(0,1)\\) and \\(c\\) is a scaling factor. In this case, \\(\\mathbf{x}_t + \\epsilon\\) becomes a spurious prediction within our batch. We compute the loss from \\(\\mathbf{x}\\) and \\(\\mathbf{y}\\) using either IM or RM objectives, and consider the change in sample-wise gradients \\(\\mathbf{g}\\) within the batch upon the injection of the spurious sample. Formally, we visualise:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.776, + 0.81, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {g} = \\left[ \\| \\frac {\\partial \\mathcal {L}}{\\partial \\left(\\mathbf {x} _ {i} ^ {\\prime}\\right)} \\| _ {2} - \\| \\frac {\\partial \\mathcal {L}}{\\partial \\left(\\mathbf {x} _ {i}\\right)} \\| _ {2} \\right] _ {i = 1} ^ {B},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.815, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\ns. t. \\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} + \\epsilon \\cdot \\mathbb {I} (i = t), \\quad \\mathcal {L} \\in \\left\\{\\mathcal {L} _ {I M}, \\mathcal {L} _ {R M} \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Fig. 2c (\\(B = 64\\) and \\(t = 32\\)), when the IM objective is used, only the spurious sample receives a prominent gradient. Whereas for an RM objective, many other samples receive significant gradients as they are directly connected" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2709" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.345, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.088, + 0.599, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.192, + 0.541, + 0.205 + ], + "angle": 0, + "content": "Figure 3. Alleviated overfitting and improved training dynamics." + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.089, + 0.903, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.192, + 0.871, + 0.206 + ], + "angle": 0, + "content": "Figure 4. Alleviated spurious gradients." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.222, + 0.483, + 0.388 + ], + "angle": 0, + "content": "to \\(\\mathbf{x}_t\\) in the computational graph of the RM loss. In other words, the spurious signals produced by one malign prediction will propagate to and affect all samples within a batch (in fact, those closer to \\(\\mathbf{x}_t\\) within the prediction manifold are more strongly affected). This means other sample-wise predictions will be significantly updated only to accommodate a malign prediction, even if they are already in relatively good shape. Through this investigation, we discover that \\(C2\\): relation matching methods are more prone to the adverse impact of spurious samples. We also present this pilot analysis on ImageNet training in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.483, + 0.51 + ], + "angle": 0, + "content": "For \\( C1 \\), common approaches to combat overfitting include the incorporation of richer learning signals and regularisation, which for RM-based KD methods means richer relations constructed and transferred. For \\( C2 \\), an intuitive solution is to identify and suppress the effect of spurious predictions or relations or to slacken the matching criterion. Guided by these principles, we now proceed to formally build up our method step by step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.182, + 0.537 + ], + "angle": 0, + "content": "5. Method" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.547, + 0.411, + 0.563 + ], + "angle": 0, + "content": "5.1. Constructing Inter-Sample Relations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.78 + ], + "angle": 0, + "content": "We first construct relation graph \\(\\mathcal{G}^{IS}\\) that encodes intersample affinity within a batch of sample predictions \\(\\{\\mathbf{z}_i\\}_{i = 1}^B\\). Different from [52, 61], our relations are constructed from predicted logits which embed more compact categorical knowledge. We use the pairwise distance between instance-wise predictions within a batch as our measure of affinity. Existing methods leverage the Gram matrices [29, 53, 54, 61] to encode inter-sample relations, but we find that this leads to collapsed inter-class knowledge via the inner product operation. Instead, our pairwise distance preserves the inter-class knowledge along the secondary dimension (i.e., the class dimension), which enables such information to be explicitly transferred as auxiliary knowledge alongside the matching of inter-sample relations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Thus, we have constructed a dense relation graph \\(\\mathcal{G}^{IS}\\), which comprises \\(B\\) vertices and \\(B \\times B\\) edges: \\(\\mathcal{G}^{IS} = (\\mathcal{V}^{IS}, \\mathcal{E}^{IS})\\). Each vertex in \\(\\mathcal{G}^{IS}\\) represents the prediction vector \\(\\mathbf{z} \\in \\mathbb{R}^C\\) for one instance within a batch, and is connected to all instances within the batch including itself. The attribute of edge \\(\\mathcal{E}_{i,j}^{IS}\\) connecting vertices \\(i\\) and \\(j\\) describes the class-wise relations between the predictions of instances \\(i\\) and \\(j\\). In practice, we can organise all edges into matrix" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.698, + 0.236 + ], + "angle": 0, + "content": "\\(\\mathcal{E}^{IS}\\in \\mathbb{R}^{B\\times B\\times C}\\) , in which:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.248, + 0.905, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {i, j} ^ {I S} = \\frac {\\mathbf {z} _ {i} - \\mathbf {z} _ {j}}{\\| \\mathbf {z} _ {i} - \\mathbf {z} _ {j} \\| _ {2}} \\in \\mathbb {R} ^ {C} \\quad \\text {f o r} \\quad i, j \\in [ 1, B ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.379 + ], + "angle": 0, + "content": "where we empirically find that normalisation along the secondary dimension helps regularise relations and improves performance. Concretely, \\(\\mathcal{E}\\) encodes inter-sample class-wise relations within a batch of \\(B\\) training samples. This design is different from and empirically more effective than previous relation formulations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.388, + 0.816, + 0.403 + ], + "angle": 0, + "content": "5.2. Constructing Inter-Class Relations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Although \\(\\mathcal{G}^{IS}\\) encodes rich inter-sample relations within a batch, it fails to explicitly model inter-class correlation patterns that are also beneficial structured knowledge [29]. Therefore, we propose to build a novel inter-class batchwise relation graph \\(\\mathcal{G}^{IC} = (\\mathcal{V}^{IC},\\mathcal{E}^{IC})\\), instead of the Gram matrices in the inner product space as in [29]. The construction of \\(\\mathcal{G}^{IC}\\) mirrors that of \\(\\mathcal{G}^{IS}\\). Vertices in \\(\\mathcal{G}^{IC}\\) are the class-wise logit vectors \\(\\mathbf{w}\\in \\mathbb{R}^B\\). Each edge in \\(\\mathcal{E}^{IC}\\in \\mathbb{R}^{C\\times C\\times B}\\) embeds the pairwise difference between the \\(i\\)-th and \\(j\\)-th per-class vectors. Concretely:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.572, + 0.905, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {i, j} ^ {I C} = \\frac {\\mathbf {w} _ {i} - \\mathbf {w} _ {j}}{\\| \\mathbf {w} _ {i} - \\mathbf {w} _ {j} \\| _ {2}} \\in \\mathbb {R} ^ {B} \\quad \\text {f o r} \\quad i, j \\in [ 1, C ]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Our inter-class relations preserve the batch-wise discrepancies by treating them as a dimension of additional knowledge (reciprocal to the case of inter-sample relations), which is unlike any other previous methods [29]. We will demonstrate in Tab. 7 that our novel formulation of intersample and inter-class relations by preserving the raw affinity knowledge along the secondary dimension perform significantly better than previous relation encoders \\(\\psi(\\cdot)\\) via Gram matrices [29, 52, 53, 61] or third-order angular distances [52]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.787, + 0.789 + ], + "angle": 0, + "content": "5.3. Constructing Virtual Relations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "For each prediction \\(\\mathbf{z}_i\\) within a batch \\(\\{\\mathbf{z}_i\\}_{i=1}^B\\), we create a virtual view of it, denoted as “\\(\\tilde{\\mathbf{z}}_i\\)\", by applying semantic-preserving transformations to original image \\(\\mathbf{x}_i\\). While other transformations are applicable, we choose RandAugment [17] that applies stochastic image transformations (see Supplementary Material for details). With our batch of predictions augmented into \\(\\{\\mathbf{z}_i, \\tilde{\\mathbf{z}}_i\\}_{i=1}^B\\), we can construct a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2710" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.865, + 0.438 + ], + "angle": 0, + "content": "
Teacher\nStudentVenueResNet56 ResNet20ResNet32×4 WRN-40-2 WRN-40-2 VGG13 ShuffleNetV2ResNet32×4 VGG8ResNet32×4 VGG13 MobileNetV2ResNet50 MobileNetV2 ShuffleNetV2ResNet40-2
Teacher72.3479.4275.6175.6174.6479.42
Student69.0672.5073.2671.9870.3671.82
Feature-based
FitNets [56]ICLR'1569.2173.5073.5872.2471.0273.54
AT [74]ICLR'1770.5573.4474.0872.7771.4372.73
CRD [60]ICLR'2071.1675.5175.4874.1473.9475.65
SRRL [68]ICLR'2171.1375.3375.5974.1873.44-
PEFD [12]NeurIPS'2270.0776.0876.0274.9274.35-
TaT [42]CVPR'2271.5975.8976.0674.9774.39-
ReviewKD [10]CVPR'2171.8975.6376.1275.0974.8477.78
NORM [47]ICLR'2371.3576.4975.6574.8273.9578.32
FCFD [45]ICLR'2371.9676.6276.4375.4675.2278.18
Logit-based
KD [27]arXiv'1570.6673.3374.9273.5472.9874.45
TAKD [50]AAAI'2070.8373.8175.1273.7873.2374.82
CTKD [41]AAAI'2371.1973.7975.4573.9373.5275.31
NKD [72]ICCV'2370.4076.3575.2474.0774.8676.26
DKD [81]CVPR'2271.9776.3276.2474.8174.6877.07
LSKD [59]CVPR'2471.4376.6276.1174.3774.3675.56
TTM [82]ICLR'2471.8376.1776.2374.3274.3376.55
CRLD [76]MM'2472.1077.6076.4575.5875.2778.27
Relation-based
RKD [52]CVPR'1969.6171.9073.3572.2271.4873.21
CC [54]CVPR'1969.6372.9773.5672.2170.7171.29
SP [61]ICCV'1969.6772.9473.8372.4372.6874.56
ICKD [46]ICCV'2171.7675.2575.6474.3373.42-
DIST [29]†NeurIPS'2271.7576.31-74.73-77.35
VRM-72.0978.7677.4776.4676.1979.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.444, + 0.859, + 0.459 + ], + "angle": 0, + "content": "Table 1. Results for same- and different-model teacher-student pairs on CIFAR-100. †: using re-trained, stronger teachers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.509 + ], + "angle": 0, + "content": "larger inter-sample edge matrix \\(\\mathcal{E}^{IS}\\in \\mathbb{R}^{2B\\times 2B\\times C}\\) and a larger inter-class edge matrix \\(\\mathcal{E}^{IC}\\in \\mathbb{R}^{C\\times C\\times 2B}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.483, + 0.631 + ], + "angle": 0, + "content": "From the perspective of sample views, our new \\(\\mathcal{E}^{IS}\\) and \\(\\mathcal{E}^{IC}\\) constructed from a batch of real and virtual samples essentially capture three types of knowledge, namely relations amongst real views (denoted as \"real-real\"), relations amongst virtual views (\"virtual-virtual\"), and relations between pairs of real and virtual views (\"real-virtual\"). For instance, a real-virtual edge that connects real vertex \\(m\\) and virtual vertex \\(n\\) in \\(\\mathcal{E}^{IS}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.641, + 0.483, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {m, n} ^ {I S} = \\frac {\\mathbf {z} _ {m} - \\tilde {\\mathbf {z}} _ {n}}{\\| \\mathbf {z} _ {m} - \\tilde {\\mathbf {z}} _ {n} \\| _ {2}} \\in \\mathbb {R} ^ {C}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.345, + 0.699 + ], + "angle": 0, + "content": "5.4. Pruning into Sparse Graphs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Pruning redundant edges. The augmented \\(\\mathcal{E}^{IS}\\) in Sec. 5.3 contains \\(2B\\times 2B\\) edges, leading to quadrupled overheads. For better efficiency, we prune \\(\\mathcal{G}^{IS}\\) into sparse graphs. Noticing that \\(\\mathcal{E}^{IS}\\) is symmetric along its diagonals, we first prune its redundant half to save up to \\(50\\%\\) edge count. We also remove intra-view edges to get a further \\(50\\%\\) reduction, as we empirically find them redundant and harm knowledge transfer. For \\(\\mathcal{G}^{IC}\\), we decompose the augmented batch of predictions of size \\(2B\\) into a real-view batch and a virtual-view batch, each of size \\(B\\), and in lieu use the inter-sample batch-wise affinity vectors between them as its vertices. Compared to their original intra-view formulation of size \\(2B\\) in Sec. 5.3, this design encodes purely inter-view" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.539 + ], + "angle": 0, + "content": "affinity knowledge with halved parameters. With redundant edge pruning (REP), our graphs now become sparse and the remaining edges can again be rearranged into compact matrices: \\(\\mathcal{E}^{ISV} \\in \\mathbb{R}^{B \\times B \\times C}\\) and \\(\\mathcal{E}^{ICV} \\in \\mathbb{R}^{C \\times C \\times B}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.909, + 0.82 + ], + "angle": 0, + "content": "Pruning unreliable edges. To mitigate the diffusive effect of spurious predictions discovered earlier, we further identify and prune the unreliable edges. In previous graph learning works, the absolute certainty of two vertices are often used to determine the reliability of an edge. For instance, REM [9] computes the reliability of an edge as the mean of the maximum predicted probabilities of two samples (i.e., two vertices). However, we argue this could bias the learning towards easy samples. Instead, we measure the discrepancy between two predictions. The larger this discrepancy, the more unreliable the relation constructed. The unreliable edge pruning (UEP) criterion is given by \\(\\mathcal{E}_{i,j}^{ISV} = \\emptyset\\) if \\(\\mathrm{H}(\\mathbf{z}_i^s,\\tilde{\\mathbf{z}}_j^s) > P_m\\) where \\(\\mathrm{H}(\\cdot)\\) computes the joint entropy (JE) between two predictions and \\(P_{m}\\) is the \\(m\\)-th percentile within the batch. Note that the criterion is enforced on student predictions, resulting in adaptive and dynamic pruning as different edges get pruned in each iteration, which improves learning most of the time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.661, + 0.849 + ], + "angle": 0, + "content": "5.5. Full Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.908, + 0.901 + ], + "angle": 0, + "content": "With \\(\\mathcal{G}^{IS}\\) and \\(\\mathcal{G}^{IC}\\) constructed for both teacher and student predictions, our VRM objective matches edge matrices \\(\\mathcal{E}^{ISV}\\) and \\(\\mathcal{E}^{ICV}\\) between teacher and student via distance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "2711" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.089, + 0.46, + 0.4 + ], + "angle": 0, + "content": "
Teacher StudentVenueResNet34 ResNet18ResNet50 MobileNetV1
Teacher Student73.31/91.4276.16/92.86
69.75/89.0768.87/88.76
Feature-based
AT [74]ICLR'1770.69/90.0169.56/89.33
OFD [25]ICCV'1970.81/89.9871.25/90.34
CRD [60]ICLR'2071.17/90.1371.37/90.41
CAT-KD [23]CVPR'2371.26/90.4572.24/91.13
SimKD [8]CVPR'2271.59/90.4872.25/90.86
ReviewKD [10]CVPR'2171.61/90.5172.56/91.00
SRRL [68]ICLR'2171.73/90.6072.49/90.92
PEFD [12]NeurIPS'2271.94/90.6873.16/91.24
RSD [78]ICCV'2572.18/90.7573.05/91.26
FCFD [45]ICLR'2372.24/90.7473.37/91.35
Logit-based
KD [27]arXiv'1570.66/89.8868.58/88.98
TAKD [50]AAAI'2070.78/90.1670.82/90.01
DKD [81]CVPR'2271.70/90.4172.05/91.05
SDD [63]CVPR'2471.14/90.0572.24/90.71
TTM [82]ICLR'2472.19/-73.09/-
LSKD [59]CVPR'2472.08/90.7473.22/91.59
CRLD [76]MM'2472.37/90.7673.53/91.43
Relation-based
RKD [52]CVPR'1971.34/90.3771.32/90.62
CC [54]CVPR'1970.74/--
ICKD [46]ICCV'2172.19/90.72-
DIST [29]NeurIPS'2272.07/90.4273.24/91.12
VRM-72.67/90.6874.16/91.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.207, + 0.407, + 0.39, + 0.421 + ], + "angle": 0, + "content": "Table 2. Results on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.501 + ], + "angle": 0, + "content": "metric \\(\\phi (\\cdot)\\) , for which the Huber loss is used. Formally, \\(L_{vrm}^{ISV} = \\phi (\\mathcal{E}_S^{ISV},\\mathcal{E}_T^{ISV})\\) and \\(L_{vrm}^{ICV} = \\phi (\\mathcal{E}_S^{ICV},\\mathcal{E}_T^{ICV})\\) The full objective is a weighted combination of the CE loss and the proposed VRM losses:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.51, + 0.482, + 0.529 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L _ {c e} + \\alpha L _ {v r m} ^ {I S V} + \\beta L _ {v r m} ^ {I C V} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.485, + 0.615 + ], + "angle": 0, + "content": "where \\( L_{ce} \\) is the CE loss applied to student's predictions of both real and virtual views and supervised by GT; \\( \\alpha \\) and \\( \\beta \\) are balancing scalars. A PyTorch-style pseudo-code for the computing of VRM losses is provided in Algorithm 1 in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.224, + 0.645 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.299, + 0.669 + ], + "angle": 0, + "content": "6.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Our method is evaluated on CIFAR-100 [35] and ImageNet [19] datasets for image classification, and MSCOCO [43] for object detection. All experimental configurations follow the standard practice in prior works. More details are provided in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.231, + 0.773 + ], + "angle": 0, + "content": "6.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Results on CIFAR-100. The results for different KD methods on CIFAR-100 are shown in Tab. 1. For same-model KD, (left of Tab. 1), VRM surpasses all previous relation-based methods across all teacher-student pairs by large margins, including DIST [29]. Noticeably, VRM significantly outperforms the strongest feature-based method, FCFD [45]. On average, it also performs much better than top logit-based methods such as CRLD [76] and TTM [82]." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.089, + 0.878, + 0.281 + ], + "angle": 0, + "content": "
ResNet101→ResNet18ResNet50→MobileNetV2
APAP50AP75APAP50AP75
Teacher42.0462.4845.8840.2261.0243.81
Student33.2653.6135.2629.4748.8730.90
Feature-based
FitNets [56]34.4354.1636.7130.2049.8031.69
FGFI [62]35.4455.5138.1731.1650.6832.92
TAKD [50]34.5955.3537.1231.2651.0333.46
ReviewKD [10]36.7556.7234.0033.7153.1536.13
FCFD [45]37.3757.6040.3434.9755.0437.51
Logit-based
KD [27]33.9754.6636.6230.1350.2831.35
CTKD [41]34.5655.4336.9131.3952.3433.10
LSKD [59]---31.7452.7733.40
DKD [81]35.0556.6037.5432.3453.7734.01
Relation-based
VRM35.4656.9137.9332.6753.9634.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.493, + 0.289, + 0.896, + 0.318 + ], + "angle": 0, + "content": "Table 3. Results on MS-COCO. All detectors are based on a Faster-RCNN [55] backbone." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.334, + 0.896, + 0.402 + ], + "angle": 0, + "content": "
TeacherStudentT.S.KDATSPLGAutoKDLSKDVRM
ResNet56DeiT-T70.4365.0873.2573.5167.3678.1578.5878.5579.52
ResNet56T2T-ViT-770.4369.3774.1574.0172.2678.3578.6278.4378.88
ResNet56PiT-T70.4373.5875.4776.0374.9778.4878.5178.7679.25
ResNet56PVT-T70.4369.2274.6677.0770.4877.4873.6078.4379.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.407, + 0.89, + 0.421 + ], + "angle": 0, + "content": "Table 4. Results for cross-architecture distillation on CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.547 + ], + "angle": 0, + "content": "These are significant, given that IM methods are known to be naturally more adapt at distilling between homogeneous model pairs. For different-model \\( KD \\) (right of Tab. 1), which RM methods are supposed to be more competent with, VRM readily surpasses all RM and IM methods with notable margins. These results highlight VRM's versatility on both homogeneous and heterogeneous model pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.676 + ], + "angle": 0, + "content": "Results on ImageNet. VRM surpasses all feature- and relation-based methods on this large-scale dataset, as shown in Tab. 2. The advantage of VRM is again more apparent over the heterogeneous pair (i.e., ResNet50 → MobileNetV2), whereby it for the first time hits \\(74.0\\%\\) Top-1 accuracy. Notably, VRM outperforms strong competitors such as FCFD [45] and CRLD [76] with comparable or even less computational overheads as compared in Tab. 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.909, + 0.879 + ], + "angle": 0, + "content": "Results on MS-COCO. We demonstrate that VRM generalises to more challenging tasks by adapting it to object detection. From Tab. 3, VRM performs better than existing logit-based methods and competitively to feature-based methods. VRM is slightly behind top-performing feature-based methods such as FCFD [45] and ReviewKD [10]. This is a consequence of the very nature of the object detection task, where fine-grained contextual features play a vital role, making feature-based methods inherently better-off [62]. Nonetheless, we experimentally demonstrate that VRM is effective on object detection, whereby it improves the baseline by \\(2 - 3\\%\\) AP, surpasses all logit-based methods, and is on par with strong feature-based methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Results for cross-architecture distillation. We follow the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2712" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.344, + 0.188 + ], + "angle": 0, + "content": "
ComponentRN32×4VGG13
RN8×4VGG8
Baseline72.5070.36
+ Match {EIS}75.6474.08
+ Match {EIS, IIC}75.9474.47
+ Match {EISV}77.6375.30
+ Match {EISV, IICV}78.3875.56
+ Pruning78.7676.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.191, + 0.33, + 0.205 + ], + "angle": 0, + "content": "Table 5. Ablation of major designs." + }, + { + "type": "table", + "bbox": [ + 0.368, + 0.093, + 0.597, + 0.178 + ], + "angle": 0, + "content": "
Matching ChoiceRN32×4 RN8×4VGG13 VGG8RN50 MNV2
Match ∅72.5070.3668.33
Match {E}78.7676.1972.30
Match {E, V}78.6375.6070.97
Match {V}78.4675.7768.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.371, + 0.191, + 0.59, + 0.206 + ], + "angle": 0, + "content": "Table 6. Effect of matching vertices." + }, + { + "type": "table", + "bbox": [ + 0.617, + 0.089, + 0.89, + 0.187 + ], + "angle": 0, + "content": "
ψ(·)Shape of Rel. MatrixMethodRN32×4 VGG13RN50
RN8×4VGG8MNV2
Gram matrices[B,B] & [C,C]CC, SP, PKT, DIST77.3175.1270.28
Angular relations[B,B,B]RKD77.1475.0969.65
ISV & ICV[B,B,C] & [C,C,B]VRM78.4675.3072.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.61, + 0.191, + 0.896, + 0.206 + ], + "angle": 0, + "content": "Table 7. Comparing different relation functions." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.224, + 0.345, + 0.324 + ], + "angle": 0, + "content": "
α3264128256512
Acc.78.0378.1378.7678.3277.84
β8163264128
Acc.78.4078.5178.7678.8678.60
n01234
Acc.77.9278.3878.7678.5578.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.328, + 0.338, + 0.342 + ], + "angle": 0, + "content": "Table 8. Effect of different \\( \\alpha ,\\beta \\) ,and \\( n \\) ." + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.222, + 0.605, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.328, + 0.604, + 0.342 + ], + "angle": 0, + "content": "Figure 5. Effect of different difficulty pairs." + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.222, + 0.88, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.328, + 0.858, + 0.343 + ], + "angle": 0, + "content": "Figure 6. Robustness to varying \\( B \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.484, + 0.483 + ], + "angle": 0, + "content": "settings in [39] and [59] to perform CNN-to-ViT cross-architecture distillation. Tab. 4 shows that VRM sets new record under this set-up for different ViT architectures. Notably, by simply replacing the vanilla KD loss with the proposed VRM objective and making no other modifications, VRM improves DeiT-T's accuracy by \\(14.44\\%\\), leading the relation-based SP [61] by \\(12.16\\%\\) and even surpassing AutoKD [40] with AutoML search." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.255, + 0.51 + ], + "angle": 0, + "content": "6.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Ablations of main design choices. In Tab. 5, starting from the baseline where only the CE loss is applied, we gradually incorporate each of our designs. As shown, each extra design consistently brings performance gains, which corroborate the validity of individual design choices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Effect of different relation encoding functions. Tab. 7 compares our proposed relation encoding scheme with existing ones. To demonstrate the superiority of our formulation, we substitute it with existing Gram matrices [29, 53, 54, 61] or angle-wise relations [52]. Results show that our inter-sample class-wise and inter-class sample-wise relations consistently yield much better performance than existing relations, which highlights VRM's contribution of a novel and superior type of relation encoding in addition to the introduction of virtual relation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Effect of varying hyperparameters. VRM involves three major hyperparameters: \\(\\alpha\\) and \\(\\beta\\), and \\(n\\). According to Tab. 8, VRM works generally well with \\(\\alpha\\) and \\(\\beta\\) in a reasonable range. Larger \\(\\alpha\\) and \\(\\beta\\) may produce better results for certain model pairs but worse results for others. For \\(n\\), a larger \\(n\\) means more difficult images and larger inter-view prediction discrepancy. From Tab. 8, \\(n = 2\\) works best, whereas other values also report competitive results." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effect of different real-virtual difficulty pairs. We also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.906, + 0.468 + ], + "angle": 0, + "content": "experiment with different strength combinations for generating the real and virtual samples. The results are presented in Fig. 5, where \"Weak\" denotes the default transformation used in previous methods, and \"Strong\" denotes RandAugment with \\( n = 2 \\). Overall, we conclude that 1) a moderate discrepancy leads to optimal performance, and 2) the discrepancy is significant to the success of VRM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.595 + ], + "angle": 0, + "content": "Robustness to varying batch sizes. Relational methods are known to be sensitive to training batch size \\( B \\). We conduct experiments to examine how robust VRM is against varying \\( B \\). To adjust the learning rate accordingly, we consider two LR scaling rules: linear LR scaling [21] and square root LR scaling [11]. From Fig. 6, VRM yields competitive results across a wide range of \\( B \\) values. In contrast, DIST [29] degrades significantly as \\( B \\) grows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.681, + 0.618 + ], + "angle": 0, + "content": "6.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Analysis of training dynamics. Figs. 3a and 3b plot the per-epoch training and validation set accuracies throughout training. VRM maintains a all-time lead in both training and validation performance, with faster convergence. Besides, while VRM's lead in training performance tapers off towards the end of training, it remains more substantial, if not further enlarging, in validation performance, which highlights its superior generalisation properties. This analysis also shows our designs are effective in mitigating the issues with existing RM methods identified in the pilot studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Analysis of spurious gradients. We revisit our second pilot study by analysing how spurious gradients are suppressed in real ImageNet training. From Fig. 4, similar gradient diffusion patterns are observed in real distillation. Samples semantically more similar to the sample with spurious predictions are more affected by spurious gradients (e.g., the dog image with a watery background resembles the bubbly texture in another sample that is actually a computer mouse.)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.096, + 0.195, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.167, + 0.166, + 0.177 + ], + "angle": 0, + "content": "(a) RKD" + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.096, + 0.297, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.167, + 0.269, + 0.177 + ], + "angle": 0, + "content": "(b) VRM" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.096, + 0.415, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.167, + 0.38, + 0.177 + ], + "angle": 0, + "content": "(c) RKD" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.096, + 0.535, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.167, + 0.501, + 0.177 + ], + "angle": 0, + "content": "(d) VRM" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.18, + 0.541, + 0.205 + ], + "angle": 0, + "content": "Figure 7. t-SNE (a-b) and loss landscape (c-d) visualisations for \\(\\mathrm{RN8\\times 4}\\) students distilled with an \\(\\mathrm{RN32\\times 4}\\) teacher via RKD and VRM." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.221, + 0.486, + 0.294 + ], + "angle": 0, + "content": "
MethodKDFitNetsRKDICKDDISTCRDReviewKD
Train. Time (ms)24.926.731.028.027.241.239.9
Peak GPU Mem. (MB)32333033038133014181042
MethodSDDMLLDCRLDNORMPEFDFCFDVRM
Train. Time (ms)34.257.243.235.136.256.447.2
Peak GPU Mem. (MB)6905765511806701953579
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.3, + 0.471, + 0.313 + ], + "angle": 0, + "content": "Table 10. Training efficiency of different distillation methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.425 + ], + "angle": 0, + "content": "Visualisation of embedding space. We conduct t-SNE analysis on student penultimate layer embeddings learnt via different methods. As presented in Fig. 7a and 7b, VRM leads to more compact per-class clusters with clearer inter-class separation and less stray points. These imply better student features distilled by VRM for the downstream task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Visualisation of loss landscape. We further analyse the generalisation and convergence properties of our method through the lens of visualised loss landscape [38]. From Fig. 7c and 7d, compared to RKD, VRM has a wider, flatter, and deeper region of minima – typical hint of better model generalisation and robustness; this wide convexity basin surrounded by salient pikes in various directions indicates excellent convergence properties. More visualisations are provided in Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Vertex matching. Within the constructed graphs \\(\\mathcal{G}^{IS}\\) and \\(\\mathcal{G}^{IC}\\), VRM matches edges \\(\\mathcal{E}^{ISV}\\) and \\(\\mathcal{E}^{ICV}\\) which carry relational knowledge. By extension, we can naturally expect vertices \\(\\nu\\) to be also transferred. From Tab. 6, matching vertices is not as effective. The advantage of matching relations (i.e., edges) is more pronounced for heterogeneous KD pairs such as ResNet50 \\(\\rightarrow\\) MobileNetV2. Adding vertex matching to the proposed edge matching does not improve but instead degrade performance. We argue this is because introducing vertex matching objectives make the matching criterion more stringent which is against our motivation of using more slackened matching." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Role of transformation operations. We emphasise the strong performance of VRM is not a simple outcome of the transformations involved. This can be verified by the results in Fig. 5 and Tab. 9, where using only the default weak transformation (i.e., random crop and horizontal flip) for both views achieves highly competitive and even superior performance under VRM. In fact, in the case of \"Weak-Weak\", both views still differ because of the stochastic operations used. Essentially, the success of VRM lies exactly" + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.1, + 0.888, + 0.172 + ], + "angle": 0, + "content": "
RealVirtualRN50→MNV2RN34→RN18
DefaultDefault74.16/91.7872.67/90.68
DefaultRandAug(n=0)74.10/91.7072.60/90.88
DefaultRandAug(n=1)74.13/91.6372.59/90.82
DefaultRandAug(n=2)74.13/91.6872.48/90.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.56, + 0.179, + 0.898, + 0.207 + ], + "angle": 0, + "content": "Table 9. Effect of different transformation strength pairs on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.22, + 0.902, + 0.292 + ], + "angle": 0, + "content": "
PruningRN32×4→RN8×4V13→V8W40-2→SNV2
Acc.tbatch(s)Mem. (MB)Acc.Acc.
w/o78.3861.434.375.5678.45
+ REP78.4646.313.675.3078.41
+ UEP78.7647.213.176.1978.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.524, + 0.301, + 0.884, + 0.315 + ], + "angle": 0, + "content": "Table 11. Effectiveness and efficiency of pruning operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.456 + ], + "angle": 0, + "content": "in the discrepancy of teacher and student predictions, where regularisation comes into crucial play. On smaller and easier datasets such as CIFAR-100, the gap between teacher and student predictions are relatively small, and manually enlarging this gap via extra transformations leads to promising outcomes. Whereas the more difficult ImageNet data already sees large teacher-student prediction discrepancy, such that it benefits less from further amplification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.597 + ], + "angle": 0, + "content": "Training efficiency. From Tab. 10, VRM is reasonably efficient compared to existing algorithms, including some leading feature- [12, 45, 47] and logit-based [33, 63] methods. VRM does not introduce any extra overheads at inference. We also breakdown the proposed pruning designs in Tab. 11 to understand how they individually impact the performance and efficiency of VRM. We see REP boosts efficiency at minimal performance cost, whereas UEP trades minimal efficiency cost for performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.602, + 0.905, + 0.634 + ], + "angle": 0, + "content": "More experiments, analyses, and discussions are provided in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.646, + 0.634, + 0.661 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.899 + ], + "angle": 0, + "content": "This paper presents VRM, a novel knowledge distillation framework that constructs and transfers virtual relations. Our designs are motivated by a set of pilot experiments, from which we identified two main cruxes with existing relation-based KD methods: their tendency to overfit and susceptibility to adverse gradient propagation. A series of tailored designs are developed and are shown to successfully mitigate these issues. We have conducted extensive experiments on different tasks and multiple datasets and verified VRM's validity and superiority in diverse settings, whereby it consistently achieves state-of-the-art performance. We hope that this work could renew the community's interest in relation-based knowledge distillation, and encourage more systematic reassessment of the design principles of such solutions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2714" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.166, + 0.188, + 0.182 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.19, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[1] Sungsoo Ahn, Shell Xu Hu, Andreas Damianou, Neil D. Lawrence, and Zhenwen Dai. Variational information distillation for knowledge transfer. In CVPR, 2019. 2, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.233, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[2] Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning optical flow from still images. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[3] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. NeurIPS, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[5] Joao Carreira, Abhishek Kar, Shubham Tulsiani, and Jitendra Malik. Virtual view networks for object reconstruction. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[6] Jiahao Chang, Shuo Wang, Hai-Ming Xu, Zehui Chen, Chenhongyi Yang, and Feng Zhao. Detrdistill: A universal knowledge distillation framework for detr-families. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[7] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[8] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.484, + 0.574 + ], + "angle": 0, + "content": "[9] Peibin Chen, Tao Ma, Xu Qin, Weidi Xu, and Shuchang Zhou. Data-efficient semi-supervised learning by reliable edge mining. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.576, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, 2021. 5, 6, 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.618, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 2020. 3, 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[12] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved feature distillation via projector ensemble. In NeurIPS, 2022. 5, 6, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[13] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[14] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[15] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. *ICLR*, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[16] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In CVPR, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.19, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In NeurIPS, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[18] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[20] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[21] P Goyal. Accurate, large minibatch sgd: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[22] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, 2023. 2, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[24] Zhiwei Hao, Jianyuan Guo, Kai Han, Yehui Tang, Han Hu, Yunhe Wang, and Chang Xu. One-for-all: Bridge the gap between heterogeneous architectures in knowledge distillation. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[26] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[27] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowlegde in a neural network. In arXiv:1503.02531, 2015. 1, 2, 3, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[28] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. In arXiv:1704.04861, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[29] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In NeurIPS, 2022. 2, 3, 4, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[30] Yuge Huang, Jiaxiang Wu, Xingkun Xu, and Shouhong Ding. Evaluation-oriented knowledge distillation for deep face recognition. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[31] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. In arXiv:1707.01219, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[32] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: structural and temporal cross-modal distillation for multi-view 3d object detection. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[33] Ying Jin, Jiaqi Wang, and Dahua Lin. Multi-level logit distillation. In CVPR, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[34] Minha Kim, Shahroz Tariq, and Simon S Woo. Fretal: Generalizing deepfake detection using knowledge distillation and representation learning. In CVPR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2715" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[35] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[36] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[37] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[38] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[39] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, 2022. 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[40] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In ICCV, 2023. 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[41] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, 2023. 2, 5, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[42] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge distillation via the target-aware transformer. In CVPR, 2022. 2, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[44] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[45] Dongyang Liu, Meina Kan, Shiguang Shan, and Xilin Chen. Function-consistent feature distillation. In ICLR, 2023. 1, 5, 6, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[46] Li Liu, Qingle Huang, Sihao Lin, Hongwei Xie, Bing Wang, Xiaojun Chang, and Xiaodan Liang. Exploring inter-channel correlation for diversity-preserved knowledge distillation. In ICCV, 2021. 2, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[47] Xiaolong Liu, Lujun Li, Chao Li, and Anbang Yao. Norm: Knowledge distillation via n-to-one representation matching. In ICLR, 2023. 1, 5, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[48] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[49] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[50] Seyed-Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 2, 5, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[51] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.858, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[52] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In IEEE TMM, 2019. 1, 2, 4, 5, 6, 7, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[53] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2, 4, 7, 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[54] Baoyun Peng, Xiao Jin, Jiaheng Liu, Shunfeng Zhou, Yichao Wu, Yu Liu, Dongsheng Li, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, 2019. 2, 4, 5, 6, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[55] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE T-PAMI, 39(6):1137-1149, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[56] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[57] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semisupervised learning with consistency and confidence. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[58] Wonchul Son, Jaemin Na, Junyong Choi, and Wonjun Hwang. Densely guided knowledge distillation using multiple teacher assistants. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[59] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, 2024. 2, 5, 6, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 5, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[61] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, 2019. 1, 2, 4, 5, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[62] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 1, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.552, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[63] Shicai Wei, Chunbo Luo, and Yang Luo. Scale decoupled distillation. In CVPR, 2024. 6, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[64] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In CVPR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[65] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[66] Chuanguang Yang, Zhulin An, Linhang Cai, and Yongjun Xu. Hierarchical self-supervised augmented knowledge distillation. In IJCAI, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[67] Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, and Qian Zhang. Cross-image relational knowledge distillation for semantic segmentation. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[68] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 5, 6, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[69] Longrong Yang, Xianpan Zhou, Xuewei Li, Liang Qiao, Zheyang Li, Ziwei Yang, Gaoang Wang, and Xi Li. Bridging cross-task protocol inconsistency for distillation in dense object detection. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[70] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and global knowledge distillation for detectors. In CVPR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2716" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[71] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked generative distillation. In ECCV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[72] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. In ICCV, 2023. 2, 5, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[73] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Km. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[74] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 5, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[75] Bo Zhang, Jiacheng Sui, and Li Niu. Foreground object search by distilling composite image feature. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[76] Weijia Zhang, Dongnan Liu, Weidong Cai, and Chao Ma. Cross-view consistency regularisation for knowledge distillation. In ACM MM, 2024. 1, 5, 6, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[77] Weijia Zhang, Dongnan Liu, Chao Ma, and Weidong Cai. Alleviating foreground sparsity for semi-supervised monocular 3d object detection. In WACV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[78] Weijia Zhang, Yuehao Liu, Wu Ran, and Chao Ma. Cross-architecture distillation made simple with redundancy suppression. ICCV, 2025. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[79] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[80] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[81] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 2, 5, 6, 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[82] Kaixiang Zheng and En-Hui Yang. Knowledge distillation based on transformed teaching matching. In ICLR, 2024. 2, 5, 6, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[83] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In CVPR, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2717" + } + ] +] \ No newline at end of file diff --git a/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_origin.pdf b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a802b8cab2a8a6a987bfabed7996714e1d238a1 --- /dev/null +++ b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/4752d372-1ed6-459b-bee6-6b202a028bae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:127d67e3304441ef246333542197e3e06c75e6bcf9f5036a9c620f106b88c028 +size 2137705 diff --git a/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/full.md b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..94e87c711e2c0fcbd817b30e66fdb76edd405dce --- /dev/null +++ b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/full.md @@ -0,0 +1,397 @@ +# VRM: Knowledge Distillation via Virtual Relation Matching + +Weijia Zhang1 + +Fei Xie + +$^{1}$ Shanghai Jiao Tong University + +Weidong Cai + +Chao $\mathbf{Ma}^{1,*}$ + +2 The University of Sydney + +{weijia.zhang, feixie, chaoma}@sjtu.edu.cn tom.cai@sydney.edu.au + +# Abstract + +Knowledge distillation (KD) aims to transfer the knowledge of a more capable yet cumbersome teacher model to a lightweight student model. In recent years, relation-based KD methods have fallen behind, as their instance-matching counterparts dominate in performance. In this paper, we revive relational KD by identifying and tackling several key issues in relation-based methods, including their susceptibility to overfitting and spurious responses. Specifically, we transfer novelly constructed affinity graphs that compactly encapsulate a wealth of beneficial inter-sample, inter-class, and inter-view correlations by exploiting virtual views and relations as a new kind of knowledge. As a result, the student has access to richer guidance signals and stronger regularisation throughout the distillation process. To further mitigate the adverse impact of spurious responses, we prune the affinity graphs by dynamically detaching redundant and unreliable edges. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO datasets demonstrate the superior performance of the proposed virtual relation matching (VRM) method, where it consistently sets new state-of-the-art records over a range of models, architectures, tasks, and set-ups. For instance, VRM for the first time hits $74.0\%$ accuracy for ResNet50 $\rightarrow$ MobileNetV2 distillation on ImageNet, and improves DeiT by $14.44\%$ on CIFAR-100 with a ResNet56 teacher. The code and models are released at https://github.com/VISION-SJTU/VRM. + +# 1. Introduction + +Deep learning is achieving incredible performance at the cost of increasing model complexity and overheads. As a consequence, large and cumbersome neural models struggle to work in resource-constrained environments. Knowledge distillation (KD) [27] has been proposed to address this issue by transferring the knowledge of larger and more capable models to smaller and lightweight ones that are resource-friendly. KD works by minimising the dis + +![](images/6f457de3b44d1eb43aa33e863cc93e5ce4f2350c218ffe49ed23ec7eff009baa.jpg) +(a) Instance Matching + +![](images/840db0a4cabc73b6a2b1d4c181b55da4e6800a7a0d5e3641339bb5477e3f17cf.jpg) +(b) Relation Matching + +![](images/8ac77d54bb8a81d21d0854823bf132eb0efae85c267e7e13120e2ccd55a3bd0a.jpg) +(c) VRM (Dense) + +![](images/ea953baeec179d0a681f15a1ef6e9512c81bac526bf87d200963122154805cf9.jpg) +(d) VRM (Pruned) +Figure 1. Conceptual illustration of VRM compared to existing KD methods based on instance matching and relation matching. + +tance between compact representations of knowledge extracted from the teacher and student models. It has found widespread application across a spectrum of downstream tasks [18, 22, 30, 34, 62, 64, 67, 70, 75, 77, 83]. + +According to the type of knowledge representations to be transferred, KD methods can be broadly categorised into feature-based [56], logit-based [27], and relation-based [52] approaches. The former two directly match the feature maps or logit vectors produced by the teacher and student models for each training sample, which is essentially instance matching (IM). By contrast, relation matching (RM) methods construct and match structured relations extracted within a batch of model responses. A conceptual illustration is presented in Figs. 1a and 1b. + +Instance matching has been the prevailing distillation approach in recent years. Popular KD benchmarks see a dominance by IM methods such as FCFD [45], NORM [47], and CRLD [76], with different downstream tasks successfully tackled by directly adopting IM-based distillation [6, 14, 62, 69]. Yet, recent studies discovered that relational knowledge is more robust to variations in neural architectures, data modalities, and tasks [52, 61]. Meanwhile, methods transferring relations have also achieved promising + +performance for a range of tasks, including but not limited to segmentation [67] and detection [15, 32]. + +Despite growing interest, relation-based methods still fall significantly short compared to their instance matching counterparts. Even the strongest RM method has been outperformed easily by recent IM solutions [29] (see Tabs. 1 and 2). RM-based methods also struggle with more challenging tasks such as object detection [29]. Moreover, previous RM solutions are primarily limited to matching inter-sample [29, 53, 61], inter-class [29], or inter-channel [46, 73] relations via simple Gram matrices. To our best knowledge, no different forms of relations other than these have been proposed since the work of DIST [29]. + +This paper fills this gap with a new kind of relations for KD - inter-view relations (Fig. 1c), which seamlessly and compactly integrate with previous inter-sample and inter-class relations. Our designs are motivated by two important observations made about RM methods in a set of pilot experiments: 1) RM methods are more susceptible to overfitting than IM methods; 2) RM methods are subject to an adverse gradient propagation effect. We empirically find that incorporating richer and more diverse relations into the matching objective helps mitigate both issues. + +To this end, we generate virtual views of samples through simple transformations, followed by constructing virtual affinity graphs and transferring the virtual relations between real and virtual samples along the edges. In lieu of Gram matrices that suffer from significant knowledge loss [29, 53, 61], we preserve the raw relations along the secondary dimension as auxiliary knowledge which adds to the types and density of relational knowledge transferred. Moreover, we also prune our affinity graphs by striping away both redundant and unreliable edges to further alleviate the propagating gradients of spurious samples (Fig. 1d). + +The above insights and remedies altogether lead to a novel Virtual Relation Matching (VRM) framework for knowledge distillation. VRM is conceptually simple, easy to implement, and devoid of complicated training procedures. It is capable of transferring rich, sophisticated knowledge robust to overfitting and spurious signals. VRM sets new state-of-the-art performance for different datasets, tasks, and settings. Perhaps more significant is that VRM makes relation-based methods regain competitiveness and back in the lead over instance matching approaches in different scenarios. To summarise, the contributions of this work include: + +- We make an early effort to present comparative analyses of existing KD methods through the lens of training dynamics and sample-wise gradients, and identify overfitting and spurious gradient diffusion as two main cruxes in relational KD methods. +- We distill richer, more diverse relations by generating virtual views, constructing virtual affinity graphs, and + +matching virtual relations. We also for the first time tackle relational KD with considerations of spurious samples and gradients by pruning redundant and unreliable edges, alongside designs to relax the matching criterion. + +- We present the streamlined VRM framework for knowledge distillation, with extensive experimental results on a diversity of models and tasks to highlight its superior performance, alongside rigorous analyses on the soundness and efficiency of our designs. + +# 2. Related Work + +KD via instance matching. KD was first proposed in Hinton et al. [27], where a student is trained to mimic the prediction of a pre-trained teacher model for each sample. Follow-up works have mostly followed such instance matching (IM) paradigm, and can be categorised into logit (or prediction)-based and feature-based methods according to what is matched. Logit-based KD has evolved from using adaptive-softened logits [33, 41] to decoupling target- and non-target logits [24, 72, 81] and applying logit transformation [59, 82]. Others [50, 58] set up auxiliary ad-hoc networks between teacher and student to facilitate logit transfer. Feature-based methods minimise the distance between the feature maps [1, 25, 56] or salient regions within in features [23, 74] from specified layers in both teacher and student networks. Some also design sophisticated distillation paths [7, 42]. All these methods are based on instance-wise transfer of knowledge, and are herein referred to as the "IM" methods, as illustrated by Fig. 1a. + +KD via relation matching. Some works transfer instead mutual relations mined amongst the network outputs extracted from a batch of training instances. These relation matching (RM) methods usually involve constructing network outputs into compact relation representations that encode rich higher-order information, as depicted in Fig. 1b. Different relation encoding functions mare, including inter-sample [29, 48, 52-54, 61], inter-class [29], inter-channel [46, 73, 78], inter-layer [48], and contrastive [60] relations. To date, IM solutions have dominated KD with their superior performance, leading top-performing relation-based methods by considerable margins. While many new IM methods are found within the last two years, frustratingly few RM solutions are being proposed. In this work, we strive to close this gap with a new kind of relations for KD — inter-view virtual relations, and revive relation-based KD by making it overtake its IM counterparts. + +Learning with virtual knowledge. While not a standalone research topic, learning with virtual knowledge finds relevance in a variety of learning-based problems. For instance, a commonly used paradigm in 3D vision tasks is to learn (or construct) from the raw data a virtual view or representation as auxiliary knowledge in solving the main task, rang- + +![](images/aa3d16d809bfea323d73787a4bc4b71c47ced412ea933092802c9a519e8a3ac9.jpg) +(a) Increasing performance on training set. + +![](images/b21167aac795e5f47ec07926028c316876f9d7dc6fc217c47de434bdb24de9c9.jpg) +(b) Plateaued performance on validation set. +(c) Diffusion of gradient perturbations within a batch. +Figure 2. Pilot studies that reveal the overfitting and spurious gradient diffusion issues with RM-based KD. + +![](images/7e8b217402999478d55d20b29ac28241d31e164a4e76e5bccd35629a820ee062.jpg) + +ing from object reconstruction [5] and optical flow [2] to 3D semantic segmentation [36], monocular 3D object detection [13], and 3D GAN inversion [64]. More broadly, many data-efficient learning methods also share the spirit of utilising virtual knowledge. For instance, self-supervised learning methods generate virtual views of the unlabelled data to enable the learning of pretext tasks [11, 20]. Another popular paradigm is transformation-invariant representation learning [51, 57] in semi-supervised learning and domain adaptation. It enforces consistency between representations learnt for a raw sample and a virtual view of it. The virtual view is often obtained by applying semantic-preserving transformations to the raw sample [16, 17]. This work is more related to this later paradigm, but involves learning with virtual knowledge in a different context, via different approaches, and for a different problem. + +# 3. Preliminaries + +KD methods generally employ a cross-entropy (CE) loss and a distillation loss to supervise student learning. The CE loss is computed between student logits $\mathbf{z}_i^s$ for each sample and its ground-truth label $\mathbf{y}_i$ . The distillation loss matches teacher and student outputs via a distance metric $\phi(\cdot)$ . In vanilla KD [27], $\phi(\cdot)$ is the Kullback-Leibler divergence (KLD) between teacher logits $\mathbf{z}^t$ and student logits $\mathbf{z}^s$ : + +$$ +\begin{array}{l} \mathcal {L} _ {i} ^ {\mathrm {K D}} = \phi_ {\mathrm {K L D}} \left(\mathbf {z} _ {i} ^ {s}, \mathbf {z} _ {i} ^ {t}\right) \\ = \tau^ {2} \sum_ {j = 1} ^ {C} \sigma_ {j} \left(\mathbf {z} _ {i} ^ {t} / \tau\right) \log \frac {\sigma_ {j} \left(\mathbf {z} _ {i} ^ {t} / \tau\right)}{\sigma_ {j} \left(\mathbf {z} _ {i} ^ {s} / \tau\right)}, \tag {1} \\ \end{array} +$$ + +where $\sigma (\cdot)$ is the Softmax operation with temperature parameter $\tau$ , and $C$ is the number of classes. For feature-based methods, $\phi (\cdot)$ can be the mean squared error (MSE) between teacher and student feature maps for each instance [56], i.e., $\mathcal{L}_i^{\mathrm{KD}} = \phi_{\mathrm{MSE}}(\mathbf{f}_i^s,\mathbf{f}_i^t)$ . + +For relation-based methods, a relation encoding function $\psi(\cdot)$ first abstracts teacher or student outputs of all instances within a training batch into a relational representation, before applying $\phi(\cdot)$ to match these relational representations between the teacher and the student. For example, the KD objectives of DIST [29] take the general form of: + +$$ +\mathcal {L} ^ {\mathrm {K D}} = \phi \left(\psi \left(\mathbf {z} _ {1} ^ {s}, \mathbf {z} _ {2} ^ {s}, \dots , \mathbf {z} _ {B} ^ {s}\right), \psi \left(\mathbf {z} _ {1} ^ {t}, \mathbf {z} _ {2} ^ {t}, \dots , \mathbf {z} _ {B} ^ {t}\right)\right), \tag {2} +$$ + +where $B$ is the batch size, and subscript $i$ of $\mathcal{L}^{\mathrm{KD}}$ is dropped as the loss is computed for a batch of instances. DIST uses both inter-class and inter-sample relation encoders as $\psi (\cdot)$ . + +# 4. Pilot Studies + +Training dynamics of KD methods. We examine the training dynamics of different relation-based methods on CIFAR-100 with ResNet32 $\times$ 4 $\rightarrow$ ResNet8 $\times$ 4 as the teacher-student pair. In Figs. 2a and 2b, we immediately notice that relational methods achieve significantly higher training accuracy, but they only marginally lead or even fall short in test accuracy compared to IM-based KD [27]. We hypothesise that relational methods are more prone to overfitting. This is expected given that the optimality of IM matching (cond. $\mathcal{A}$ ) implies the optimality of relation matching (cond. $\mathcal{B}$ ), while the converse does not hold. Concretely, $\mathcal{A} \Rightarrow \mathcal{B} \land \neg \mathcal{B} \Rightarrow \neg \mathcal{A}$ . In other words, relation matching is a weaker and less constrained objective than IM, which makes the student more readily fit the teaching signals and not generalise well. Thus, we conclude that $C1$ : relation matching methods are more prone to overfitting. + +Gradient analysis of KD methods. We investigate the gradient patterns within a batch when a spurious sample produces a major misguiding signal. To this end, we first generate two random vectors $\mathbf{x},\mathbf{y}\sim \mathcal{N}(0,1)$ for $\mathbf{x},\mathbf{y}\in \mathbb{R}^{B\times D}$ , where $B$ is the batch size and $D$ is the dimension of persample predictions, $\mathbf{x}$ is taken as the sample-wise predictions, and $\mathbf{y}$ the supervision signals. We then add a noise vector $\epsilon = c\cdot \mathbf{z}$ to $\mathbf{x}_t$ , where $\mathbf{z}\sim \mathcal{N}(0,1)$ and $c$ is a scaling factor. In this case, $\mathbf{x}_t + \epsilon$ becomes a spurious prediction within our batch. We compute the loss from $\mathbf{x}$ and $\mathbf{y}$ using either IM or RM objectives, and consider the change in sample-wise gradients $\mathbf{g}$ within the batch upon the injection of the spurious sample. Formally, we visualise: + +$$ +\Delta \mathbf {g} = \left[ \| \frac {\partial \mathcal {L}}{\partial \left(\mathbf {x} _ {i} ^ {\prime}\right)} \| _ {2} - \| \frac {\partial \mathcal {L}}{\partial \left(\mathbf {x} _ {i}\right)} \| _ {2} \right] _ {i = 1} ^ {B}, +$$ + +$$ +s. t. \mathbf {x} _ {i} ^ {\prime} = \mathbf {x} _ {i} + \epsilon \cdot \mathbb {I} (i = t), \quad \mathcal {L} \in \left\{\mathcal {L} _ {I M}, \mathcal {L} _ {R M} \right\}. \tag {3} +$$ + +In Fig. 2c ( $B = 64$ and $t = 32$ ), when the IM objective is used, only the spurious sample receives a prominent gradient. Whereas for an RM objective, many other samples receive significant gradients as they are directly connected + +![](images/b3da20e409478ffba022d966b916e2054d8d901941aec90a5beb4abec29a4013.jpg) +Figure 3. Alleviated overfitting and improved training dynamics. + +![](images/c0212815eefb87d1b1b23fee3da4c880b319dcaa39c11e53cf646abaa94e1691.jpg) + +![](images/72708e7921ac197d3176c5c1eebf20d25dfecbeb15f305b059d2c943384bd253.jpg) +Figure 4. Alleviated spurious gradients. + +to $\mathbf{x}_t$ in the computational graph of the RM loss. In other words, the spurious signals produced by one malign prediction will propagate to and affect all samples within a batch (in fact, those closer to $\mathbf{x}_t$ within the prediction manifold are more strongly affected). This means other sample-wise predictions will be significantly updated only to accommodate a malign prediction, even if they are already in relatively good shape. Through this investigation, we discover that $C2$ : relation matching methods are more prone to the adverse impact of spurious samples. We also present this pilot analysis on ImageNet training in Fig. 4. + +For $C1$ , common approaches to combat overfitting include the incorporation of richer learning signals and regularisation, which for RM-based KD methods means richer relations constructed and transferred. For $C2$ , an intuitive solution is to identify and suppress the effect of spurious predictions or relations or to slacken the matching criterion. Guided by these principles, we now proceed to formally build up our method step by step. + +# 5. Method + +# 5.1. Constructing Inter-Sample Relations + +We first construct relation graph $\mathcal{G}^{IS}$ that encodes intersample affinity within a batch of sample predictions $\{\mathbf{z}_i\}_{i = 1}^B$ . Different from [52, 61], our relations are constructed from predicted logits which embed more compact categorical knowledge. We use the pairwise distance between instance-wise predictions within a batch as our measure of affinity. Existing methods leverage the Gram matrices [29, 53, 54, 61] to encode inter-sample relations, but we find that this leads to collapsed inter-class knowledge via the inner product operation. Instead, our pairwise distance preserves the inter-class knowledge along the secondary dimension (i.e., the class dimension), which enables such information to be explicitly transferred as auxiliary knowledge alongside the matching of inter-sample relations. + +Thus, we have constructed a dense relation graph $\mathcal{G}^{IS}$ , which comprises $B$ vertices and $B \times B$ edges: $\mathcal{G}^{IS} = (\mathcal{V}^{IS}, \mathcal{E}^{IS})$ . Each vertex in $\mathcal{G}^{IS}$ represents the prediction vector $\mathbf{z} \in \mathbb{R}^C$ for one instance within a batch, and is connected to all instances within the batch including itself. The attribute of edge $\mathcal{E}_{i,j}^{IS}$ connecting vertices $i$ and $j$ describes the class-wise relations between the predictions of instances $i$ and $j$ . In practice, we can organise all edges into matrix + +$\mathcal{E}^{IS}\in \mathbb{R}^{B\times B\times C}$ , in which: + +$$ +\mathcal {E} _ {i, j} ^ {I S} = \frac {\mathbf {z} _ {i} - \mathbf {z} _ {j}}{\| \mathbf {z} _ {i} - \mathbf {z} _ {j} \| _ {2}} \in \mathbb {R} ^ {C} \quad \text {f o r} \quad i, j \in [ 1, B ], \tag {4} +$$ + +where we empirically find that normalisation along the secondary dimension helps regularise relations and improves performance. Concretely, $\mathcal{E}$ encodes inter-sample class-wise relations within a batch of $B$ training samples. This design is different from and empirically more effective than previous relation formulations. + +# 5.2. Constructing Inter-Class Relations + +Although $\mathcal{G}^{IS}$ encodes rich inter-sample relations within a batch, it fails to explicitly model inter-class correlation patterns that are also beneficial structured knowledge [29]. Therefore, we propose to build a novel inter-class batchwise relation graph $\mathcal{G}^{IC} = (\mathcal{V}^{IC},\mathcal{E}^{IC})$ , instead of the Gram matrices in the inner product space as in [29]. The construction of $\mathcal{G}^{IC}$ mirrors that of $\mathcal{G}^{IS}$ . Vertices in $\mathcal{G}^{IC}$ are the class-wise logit vectors $\mathbf{w}\in \mathbb{R}^B$ . Each edge in $\mathcal{E}^{IC}\in \mathbb{R}^{C\times C\times B}$ embeds the pairwise difference between the $i$ -th and $j$ -th per-class vectors. Concretely: + +$$ +\mathcal {E} _ {i, j} ^ {I C} = \frac {\mathbf {w} _ {i} - \mathbf {w} _ {j}}{\| \mathbf {w} _ {i} - \mathbf {w} _ {j} \| _ {2}} \in \mathbb {R} ^ {B} \quad \text {f o r} \quad i, j \in [ 1, C ]. \tag {5} +$$ + +Our inter-class relations preserve the batch-wise discrepancies by treating them as a dimension of additional knowledge (reciprocal to the case of inter-sample relations), which is unlike any other previous methods [29]. We will demonstrate in Tab. 7 that our novel formulation of intersample and inter-class relations by preserving the raw affinity knowledge along the secondary dimension perform significantly better than previous relation encoders $\psi(\cdot)$ via Gram matrices [29, 52, 53, 61] or third-order angular distances [52]. + +# 5.3. Constructing Virtual Relations + +For each prediction $\mathbf{z}_i$ within a batch $\{\mathbf{z}_i\}_{i=1}^B$ , we create a virtual view of it, denoted as “ $\tilde{\mathbf{z}}_i$ ", by applying semantic-preserving transformations to original image $\mathbf{x}_i$ . While other transformations are applicable, we choose RandAugment [17] that applies stochastic image transformations (see Supplementary Material for details). With our batch of predictions augmented into $\{\mathbf{z}_i, \tilde{\mathbf{z}}_i\}_{i=1}^B$ , we can construct a + +
Teacher +StudentVenueResNet56 ResNet20ResNet32×4 WRN-40-2 WRN-40-2 VGG13 ShuffleNetV2ResNet32×4 VGG8ResNet32×4 VGG13 MobileNetV2ResNet50 MobileNetV2 ShuffleNetV2ResNet40-2
Teacher72.3479.4275.6175.6174.6479.42
Student69.0672.5073.2671.9870.3671.82
Feature-based
FitNets [56]ICLR'1569.2173.5073.5872.2471.0273.54
AT [74]ICLR'1770.5573.4474.0872.7771.4372.73
CRD [60]ICLR'2071.1675.5175.4874.1473.9475.65
SRRL [68]ICLR'2171.1375.3375.5974.1873.44-
PEFD [12]NeurIPS'2270.0776.0876.0274.9274.35-
TaT [42]CVPR'2271.5975.8976.0674.9774.39-
ReviewKD [10]CVPR'2171.8975.6376.1275.0974.8477.78
NORM [47]ICLR'2371.3576.4975.6574.8273.9578.32
FCFD [45]ICLR'2371.9676.6276.4375.4675.2278.18
Logit-based
KD [27]arXiv'1570.6673.3374.9273.5472.9874.45
TAKD [50]AAAI'2070.8373.8175.1273.7873.2374.82
CTKD [41]AAAI'2371.1973.7975.4573.9373.5275.31
NKD [72]ICCV'2370.4076.3575.2474.0774.8676.26
DKD [81]CVPR'2271.9776.3276.2474.8174.6877.07
LSKD [59]CVPR'2471.4376.6276.1174.3774.3675.56
TTM [82]ICLR'2471.8376.1776.2374.3274.3376.55
CRLD [76]MM'2472.1077.6076.4575.5875.2778.27
Relation-based
RKD [52]CVPR'1969.6171.9073.3572.2271.4873.21
CC [54]CVPR'1969.6372.9773.5672.2170.7171.29
SP [61]ICCV'1969.6772.9473.8372.4372.6874.56
ICKD [46]ICCV'2171.7675.2575.6474.3373.42-
DIST [29]†NeurIPS'2271.7576.31-74.73-77.35
VRM-72.0978.7677.4776.4676.1979.34
+ +Table 1. Results for same- and different-model teacher-student pairs on CIFAR-100. †: using re-trained, stronger teachers. + +larger inter-sample edge matrix $\mathcal{E}^{IS}\in \mathbb{R}^{2B\times 2B\times C}$ and a larger inter-class edge matrix $\mathcal{E}^{IC}\in \mathbb{R}^{C\times C\times 2B}$ + +From the perspective of sample views, our new $\mathcal{E}^{IS}$ and $\mathcal{E}^{IC}$ constructed from a batch of real and virtual samples essentially capture three types of knowledge, namely relations amongst real views (denoted as "real-real"), relations amongst virtual views ("virtual-virtual"), and relations between pairs of real and virtual views ("real-virtual"). For instance, a real-virtual edge that connects real vertex $m$ and virtual vertex $n$ in $\mathcal{E}^{IS}$ is computed as: + +$$ +\mathcal {E} _ {m, n} ^ {I S} = \frac {\mathbf {z} _ {m} - \tilde {\mathbf {z}} _ {n}}{\| \mathbf {z} _ {m} - \tilde {\mathbf {z}} _ {n} \| _ {2}} \in \mathbb {R} ^ {C}. \tag {6} +$$ + +# 5.4. Pruning into Sparse Graphs + +Pruning redundant edges. The augmented $\mathcal{E}^{IS}$ in Sec. 5.3 contains $2B\times 2B$ edges, leading to quadrupled overheads. For better efficiency, we prune $\mathcal{G}^{IS}$ into sparse graphs. Noticing that $\mathcal{E}^{IS}$ is symmetric along its diagonals, we first prune its redundant half to save up to $50\%$ edge count. We also remove intra-view edges to get a further $50\%$ reduction, as we empirically find them redundant and harm knowledge transfer. For $\mathcal{G}^{IC}$ , we decompose the augmented batch of predictions of size $2B$ into a real-view batch and a virtual-view batch, each of size $B$ , and in lieu use the inter-sample batch-wise affinity vectors between them as its vertices. Compared to their original intra-view formulation of size $2B$ in Sec. 5.3, this design encodes purely inter-view + +affinity knowledge with halved parameters. With redundant edge pruning (REP), our graphs now become sparse and the remaining edges can again be rearranged into compact matrices: $\mathcal{E}^{ISV} \in \mathbb{R}^{B \times B \times C}$ and $\mathcal{E}^{ICV} \in \mathbb{R}^{C \times C \times B}$ . + +Pruning unreliable edges. To mitigate the diffusive effect of spurious predictions discovered earlier, we further identify and prune the unreliable edges. In previous graph learning works, the absolute certainty of two vertices are often used to determine the reliability of an edge. For instance, REM [9] computes the reliability of an edge as the mean of the maximum predicted probabilities of two samples (i.e., two vertices). However, we argue this could bias the learning towards easy samples. Instead, we measure the discrepancy between two predictions. The larger this discrepancy, the more unreliable the relation constructed. The unreliable edge pruning (UEP) criterion is given by $\mathcal{E}_{i,j}^{ISV} = \emptyset$ if $\mathrm{H}(\mathbf{z}_i^s,\tilde{\mathbf{z}}_j^s) > P_m$ where $\mathrm{H}(\cdot)$ computes the joint entropy (JE) between two predictions and $P_{m}$ is the $m$ -th percentile within the batch. Note that the criterion is enforced on student predictions, resulting in adaptive and dynamic pruning as different edges get pruned in each iteration, which improves learning most of the time. + +# 5.5. Full Objective + +With $\mathcal{G}^{IS}$ and $\mathcal{G}^{IC}$ constructed for both teacher and student predictions, our VRM objective matches edge matrices $\mathcal{E}^{ISV}$ and $\mathcal{E}^{ICV}$ between teacher and student via distance + +
Teacher StudentVenueResNet34 ResNet18ResNet50 MobileNetV1
Teacher Student73.31/91.4276.16/92.86
69.75/89.0768.87/88.76
Feature-based
AT [74]ICLR'1770.69/90.0169.56/89.33
OFD [25]ICCV'1970.81/89.9871.25/90.34
CRD [60]ICLR'2071.17/90.1371.37/90.41
CAT-KD [23]CVPR'2371.26/90.4572.24/91.13
SimKD [8]CVPR'2271.59/90.4872.25/90.86
ReviewKD [10]CVPR'2171.61/90.5172.56/91.00
SRRL [68]ICLR'2171.73/90.6072.49/90.92
PEFD [12]NeurIPS'2271.94/90.6873.16/91.24
RSD [78]ICCV'2572.18/90.7573.05/91.26
FCFD [45]ICLR'2372.24/90.7473.37/91.35
Logit-based
KD [27]arXiv'1570.66/89.8868.58/88.98
TAKD [50]AAAI'2070.78/90.1670.82/90.01
DKD [81]CVPR'2271.70/90.4172.05/91.05
SDD [63]CVPR'2471.14/90.0572.24/90.71
TTM [82]ICLR'2472.19/-73.09/-
LSKD [59]CVPR'2472.08/90.7473.22/91.59
CRLD [76]MM'2472.37/90.7673.53/91.43
Relation-based
RKD [52]CVPR'1971.34/90.3771.32/90.62
CC [54]CVPR'1970.74/--
ICKD [46]ICCV'2172.19/90.72-
DIST [29]NeurIPS'2272.07/90.4273.24/91.12
VRM-72.67/90.6874.16/91.78
+ +metric $\phi (\cdot)$ , for which the Huber loss is used. Formally, $L_{vrm}^{ISV} = \phi (\mathcal{E}_S^{ISV},\mathcal{E}_T^{ISV})$ and $L_{vrm}^{ICV} = \phi (\mathcal{E}_S^{ICV},\mathcal{E}_T^{ICV})$ The full objective is a weighted combination of the CE loss and the proposed VRM losses: + +$$ +L _ {t o t a l} = L _ {c e} + \alpha L _ {v r m} ^ {I S V} + \beta L _ {v r m} ^ {I C V} \tag {7} +$$ + +where $L_{ce}$ is the CE loss applied to student's predictions of both real and virtual views and supervised by GT; $\alpha$ and $\beta$ are balancing scalars. A PyTorch-style pseudo-code for the computing of VRM losses is provided in Algorithm 1 in Supplementary Material. + +# 6. Experiments + +# 6.1. Experimental Settings + +Our method is evaluated on CIFAR-100 [35] and ImageNet [19] datasets for image classification, and MSCOCO [43] for object detection. All experimental configurations follow the standard practice in prior works. More details are provided in Supplementary Material. + +# 6.2. Main Results + +Results on CIFAR-100. The results for different KD methods on CIFAR-100 are shown in Tab. 1. For same-model KD, (left of Tab. 1), VRM surpasses all previous relation-based methods across all teacher-student pairs by large margins, including DIST [29]. Noticeably, VRM significantly outperforms the strongest feature-based method, FCFD [45]. On average, it also performs much better than top logit-based methods such as CRLD [76] and TTM [82]. + +Table 2. Results on ImageNet. + +
ResNet101→ResNet18ResNet50→MobileNetV2
APAP50AP75APAP50AP75
Teacher42.0462.4845.8840.2261.0243.81
Student33.2653.6135.2629.4748.8730.90
Feature-based
FitNets [56]34.4354.1636.7130.2049.8031.69
FGFI [62]35.4455.5138.1731.1650.6832.92
TAKD [50]34.5955.3537.1231.2651.0333.46
ReviewKD [10]36.7556.7234.0033.7153.1536.13
FCFD [45]37.3757.6040.3434.9755.0437.51
Logit-based
KD [27]33.9754.6636.6230.1350.2831.35
CTKD [41]34.5655.4336.9131.3952.3433.10
LSKD [59]---31.7452.7733.40
DKD [81]35.0556.6037.5432.3453.7734.01
Relation-based
VRM35.4656.9137.9332.6753.9634.48
+ +Table 3. Results on MS-COCO. All detectors are based on a Faster-RCNN [55] backbone. + +
TeacherStudentT.S.KDATSPLGAutoKDLSKDVRM
ResNet56DeiT-T70.4365.0873.2573.5167.3678.1578.5878.5579.52
ResNet56T2T-ViT-770.4369.3774.1574.0172.2678.3578.6278.4378.88
ResNet56PiT-T70.4373.5875.4776.0374.9778.4878.5178.7679.25
ResNet56PVT-T70.4369.2274.6677.0770.4877.4873.6078.4379.42
+ +Table 4. Results for cross-architecture distillation on CIFAR-100. + +These are significant, given that IM methods are known to be naturally more adapt at distilling between homogeneous model pairs. For different-model $KD$ (right of Tab. 1), which RM methods are supposed to be more competent with, VRM readily surpasses all RM and IM methods with notable margins. These results highlight VRM's versatility on both homogeneous and heterogeneous model pairs. + +Results on ImageNet. VRM surpasses all feature- and relation-based methods on this large-scale dataset, as shown in Tab. 2. The advantage of VRM is again more apparent over the heterogeneous pair (i.e., ResNet50 → MobileNetV2), whereby it for the first time hits $74.0\%$ Top-1 accuracy. Notably, VRM outperforms strong competitors such as FCFD [45] and CRLD [76] with comparable or even less computational overheads as compared in Tab. 10. + +Results on MS-COCO. We demonstrate that VRM generalises to more challenging tasks by adapting it to object detection. From Tab. 3, VRM performs better than existing logit-based methods and competitively to feature-based methods. VRM is slightly behind top-performing feature-based methods such as FCFD [45] and ReviewKD [10]. This is a consequence of the very nature of the object detection task, where fine-grained contextual features play a vital role, making feature-based methods inherently better-off [62]. Nonetheless, we experimentally demonstrate that VRM is effective on object detection, whereby it improves the baseline by $2 - 3\%$ AP, surpasses all logit-based methods, and is on par with strong feature-based methods. + +Results for cross-architecture distillation. We follow the + +
ComponentRN32×4VGG13
RN8×4VGG8
Baseline72.5070.36
+ Match {EIS}75.6474.08
+ Match {EIS, IIC}75.9474.47
+ Match {EISV}77.6375.30
+ Match {EISV, IICV}78.3875.56
+ Pruning78.7676.19
+ +Table 5. Ablation of major designs. + +
Matching ChoiceRN32×4 RN8×4VGG13 VGG8RN50 MNV2
Match ∅72.5070.3668.33
Match {E}78.7676.1972.30
Match {E, V}78.6375.6070.97
Match {V}78.4675.7768.33
+ +Table 6. Effect of matching vertices. + +
ψ(·)Shape of Rel. MatrixMethodRN32×4 VGG13RN50
RN8×4VGG8MNV2
Gram matrices[B,B] & [C,C]CC, SP, PKT, DIST77.3175.1270.28
Angular relations[B,B,B]RKD77.1475.0969.65
ISV & ICV[B,B,C] & [C,C,B]VRM78.4675.3072.10
+ +Table 7. Comparing different relation functions. + +
α3264128256512
Acc.78.0378.1378.7678.3277.84
β8163264128
Acc.78.4078.5178.7678.8678.60
n01234
Acc.77.9278.3878.7678.5578.43
+ +Table 8. Effect of different $\alpha ,\beta$ ,and $n$ . + +![](images/8057ee8253a401617275712487a0a801df9e72b95d91266ecb296ec831c94fd9.jpg) +Figure 5. Effect of different difficulty pairs. + +![](images/ff7bbbfb266c386c21ebb690185ca2051f2f760e948fcd35de637e19d029ea83.jpg) +Figure 6. Robustness to varying $B$ . + +settings in [39] and [59] to perform CNN-to-ViT cross-architecture distillation. Tab. 4 shows that VRM sets new record under this set-up for different ViT architectures. Notably, by simply replacing the vanilla KD loss with the proposed VRM objective and making no other modifications, VRM improves DeiT-T's accuracy by $14.44\%$ , leading the relation-based SP [61] by $12.16\%$ and even surpassing AutoKD [40] with AutoML search. + +# 6.3. Ablation Studies + +Ablations of main design choices. In Tab. 5, starting from the baseline where only the CE loss is applied, we gradually incorporate each of our designs. As shown, each extra design consistently brings performance gains, which corroborate the validity of individual design choices. + +Effect of different relation encoding functions. Tab. 7 compares our proposed relation encoding scheme with existing ones. To demonstrate the superiority of our formulation, we substitute it with existing Gram matrices [29, 53, 54, 61] or angle-wise relations [52]. Results show that our inter-sample class-wise and inter-class sample-wise relations consistently yield much better performance than existing relations, which highlights VRM's contribution of a novel and superior type of relation encoding in addition to the introduction of virtual relation. + +Effect of varying hyperparameters. VRM involves three major hyperparameters: $\alpha$ and $\beta$ , and $n$ . According to Tab. 8, VRM works generally well with $\alpha$ and $\beta$ in a reasonable range. Larger $\alpha$ and $\beta$ may produce better results for certain model pairs but worse results for others. For $n$ , a larger $n$ means more difficult images and larger inter-view prediction discrepancy. From Tab. 8, $n = 2$ works best, whereas other values also report competitive results. + +Effect of different real-virtual difficulty pairs. We also + +experiment with different strength combinations for generating the real and virtual samples. The results are presented in Fig. 5, where "Weak" denotes the default transformation used in previous methods, and "Strong" denotes RandAugment with $n = 2$ . Overall, we conclude that 1) a moderate discrepancy leads to optimal performance, and 2) the discrepancy is significant to the success of VRM. + +Robustness to varying batch sizes. Relational methods are known to be sensitive to training batch size $B$ . We conduct experiments to examine how robust VRM is against varying $B$ . To adjust the learning rate accordingly, we consider two LR scaling rules: linear LR scaling [21] and square root LR scaling [11]. From Fig. 6, VRM yields competitive results across a wide range of $B$ values. In contrast, DIST [29] degrades significantly as $B$ grows. + +# 6.4. Further Analysis + +Analysis of training dynamics. Figs. 3a and 3b plot the per-epoch training and validation set accuracies throughout training. VRM maintains a all-time lead in both training and validation performance, with faster convergence. Besides, while VRM's lead in training performance tapers off towards the end of training, it remains more substantial, if not further enlarging, in validation performance, which highlights its superior generalisation properties. This analysis also shows our designs are effective in mitigating the issues with existing RM methods identified in the pilot studies. + +Analysis of spurious gradients. We revisit our second pilot study by analysing how spurious gradients are suppressed in real ImageNet training. From Fig. 4, similar gradient diffusion patterns are observed in real distillation. Samples semantically more similar to the sample with spurious predictions are more affected by spurious gradients (e.g., the dog image with a watery background resembles the bubbly texture in another sample that is actually a computer mouse.). + +![](images/e84ea35fef3f4f65b04d72ce95b992dfe3d0187817038425e616407067f19a22.jpg) +(a) RKD + +![](images/63b01b1c402a6c63364bd81dce0e80a5d9866341c766c6e4279870c0ea8aa8bc.jpg) +(b) VRM +Figure 7. t-SNE (a-b) and loss landscape (c-d) visualisations for $\mathrm{RN8\times 4}$ students distilled with an $\mathrm{RN32\times 4}$ teacher via RKD and VRM. + +![](images/6138ea77c96853fdbdd4379a8f7942b8d118cf73678d2e190471b4df95241e85.jpg) +(c) RKD + +![](images/18d18218b903574cecce0e17c870c55cd2061ae5253df19b4a98e6607c5783d4.jpg) +(d) VRM + +
MethodKDFitNetsRKDICKDDISTCRDReviewKD
Train. Time (ms)24.926.731.028.027.241.239.9
Peak GPU Mem. (MB)32333033038133014181042
MethodSDDMLLDCRLDNORMPEFDFCFDVRM
Train. Time (ms)34.257.243.235.136.256.447.2
Peak GPU Mem. (MB)6905765511806701953579
+ +Visualisation of embedding space. We conduct t-SNE analysis on student penultimate layer embeddings learnt via different methods. As presented in Fig. 7a and 7b, VRM leads to more compact per-class clusters with clearer inter-class separation and less stray points. These imply better student features distilled by VRM for the downstream task. + +Visualisation of loss landscape. We further analyse the generalisation and convergence properties of our method through the lens of visualised loss landscape [38]. From Fig. 7c and 7d, compared to RKD, VRM has a wider, flatter, and deeper region of minima – typical hint of better model generalisation and robustness; this wide convexity basin surrounded by salient pikes in various directions indicates excellent convergence properties. More visualisations are provided in Supplementary Material. + +Vertex matching. Within the constructed graphs $\mathcal{G}^{IS}$ and $\mathcal{G}^{IC}$ , VRM matches edges $\mathcal{E}^{ISV}$ and $\mathcal{E}^{ICV}$ which carry relational knowledge. By extension, we can naturally expect vertices $\nu$ to be also transferred. From Tab. 6, matching vertices is not as effective. The advantage of matching relations (i.e., edges) is more pronounced for heterogeneous KD pairs such as ResNet50 $\rightarrow$ MobileNetV2. Adding vertex matching to the proposed edge matching does not improve but instead degrade performance. We argue this is because introducing vertex matching objectives make the matching criterion more stringent which is against our motivation of using more slackened matching. + +Role of transformation operations. We emphasise the strong performance of VRM is not a simple outcome of the transformations involved. This can be verified by the results in Fig. 5 and Tab. 9, where using only the default weak transformation (i.e., random crop and horizontal flip) for both views achieves highly competitive and even superior performance under VRM. In fact, in the case of "Weak-Weak", both views still differ because of the stochastic operations used. Essentially, the success of VRM lies exactly + +Table 10. Training efficiency of different distillation methods. + +
RealVirtualRN50→MNV2RN34→RN18
DefaultDefault74.16/91.7872.67/90.68
DefaultRandAug(n=0)74.10/91.7072.60/90.88
DefaultRandAug(n=1)74.13/91.6372.59/90.82
DefaultRandAug(n=2)74.13/91.6872.48/90.73
+ +Table 9. Effect of different transformation strength pairs on ImageNet. + +
PruningRN32×4→RN8×4V13→V8W40-2→SNV2
Acc.tbatch(s)Mem. (MB)Acc.Acc.
w/o78.3861.434.375.5678.45
+ REP78.4646.313.675.3078.41
+ UEP78.7647.213.176.1978.62
+ +Table 11. Effectiveness and efficiency of pruning operations. + +in the discrepancy of teacher and student predictions, where regularisation comes into crucial play. On smaller and easier datasets such as CIFAR-100, the gap between teacher and student predictions are relatively small, and manually enlarging this gap via extra transformations leads to promising outcomes. Whereas the more difficult ImageNet data already sees large teacher-student prediction discrepancy, such that it benefits less from further amplification. + +Training efficiency. From Tab. 10, VRM is reasonably efficient compared to existing algorithms, including some leading feature- [12, 45, 47] and logit-based [33, 63] methods. VRM does not introduce any extra overheads at inference. We also breakdown the proposed pruning designs in Tab. 11 to understand how they individually impact the performance and efficiency of VRM. We see REP boosts efficiency at minimal performance cost, whereas UEP trades minimal efficiency cost for performance. + +More experiments, analyses, and discussions are provided in Supplementary Material. + +# 7. Conclusion + +This paper presents VRM, a novel knowledge distillation framework that constructs and transfers virtual relations. Our designs are motivated by a set of pilot experiments, from which we identified two main cruxes with existing relation-based KD methods: their tendency to overfit and susceptibility to adverse gradient propagation. A series of tailored designs are developed and are shown to successfully mitigate these issues. We have conducted extensive experiments on different tasks and multiple datasets and verified VRM's validity and superiority in diverse settings, whereby it consistently achieves state-of-the-art performance. We hope that this work could renew the community's interest in relation-based knowledge distillation, and encourage more systematic reassessment of the design principles of such solutions. + +Acknowledgements. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Sungsoo Ahn, Shell Xu Hu, Andreas Damianou, Neil D. Lawrence, and Zhenwen Dai. Variational information distillation for knowledge transfer. In CVPR, 2019. 2, 1, 3 +[2] Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning optical flow from still images. In CVPR, 2021. 3 +[3] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. NeurIPS, 2014. 6 +[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 6 +[5] Joao Carreira, Abhishek Kar, Shubham Tulsiani, and Jitendra Malik. Virtual view networks for object reconstruction. In CVPR, 2015. 3 +[6] Jiahao Chang, Shuo Wang, Hai-Ming Xu, Zehui Chen, Chenhongyi Yang, and Feng Zhao. Detrdistill: A universal knowledge distillation framework for detr-families. In ICCV, 2023. 1 +[7] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2, 1, 3 +[8] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, 2022. 6 +[9] Peibin Chen, Tao Ma, Xu Qin, Weidi Xu, and Shuchang Zhou. Data-efficient semi-supervised learning by reliable edge mining. In CVPR, 2020. 5 +[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, 2021. 5, 6, 1, 2, 3 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 2020. 3, 7, 6 +[12] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved feature distillation via projector ensemble. In NeurIPS, 2022. 5, 6, 8, 1, 2 +[13] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 3 +[14] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. In ICLR, 2023. 1 +[15] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. *ICLR*, 2022. 2 +[16] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In CVPR, 2019. 3 + +[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In NeurIPS, 2020. 3, 4 +[18] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In CVPR, 2023. 1 +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 2 +[20] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 3 +[21] P Goyal. Accurate, large minibatch sgd: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 7 +[22] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In ICLR, 2024. 1 +[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, 2023. 2, 6, 1, 3 +[24] Zhiwei Hao, Jianyuan Guo, Kai Han, Yehui Tang, Han Hu, Yunhe Wang, and Chang Xu. One-for-all: Bridge the gap between heterogeneous architectures in knowledge distillation. NeurIPS, 2024. 2 +[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2, 6, 1, 3 +[26] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, 2019. 1, 2, 3 +[27] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowlegde in a neural network. In arXiv:1503.02531, 2015. 1, 2, 3, 5, 6, 4 +[28] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. In arXiv:1704.04861, 2017. 3 +[29] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In NeurIPS, 2022. 2, 3, 4, 5, 6, 7, 1 +[30] Yuge Huang, Jiaxiang Wu, Xingkun Xu, and Shouhong Ding. Evaluation-oriented knowledge distillation for deep face recognition. In CVPR, 2022. 1 +[31] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. In arXiv:1707.01219, 2017. 1, 2, 3 +[32] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: structural and temporal cross-modal distillation for multi-view 3d object detection. NeurIPS, 2024. 2 +[33] Ying Jin, Jiaqi Wang, and Dahua Lin. Multi-level logit distillation. In CVPR, 2023. 2, 8 +[34] Minha Kim, Shahroz Tariq, and Simon S Woo. Fretal: Generalizing deepfake detection using knowledge distillation and representation learning. In CVPR, 2021. 1 + +[35] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6, 2 +[36] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In ECCV, 2020. 3 +[37] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 6 +[38] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 8 +[39] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, 2022. 7, 1, 3 +[40] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In ICCV, 2023. 7, 1, 3 +[41] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, 2023. 2, 5, 6, 1, 3 +[42] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge distillation via the target-aware transformer. In CVPR, 2022. 2, 5, 1 +[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6, 2 +[44] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3 +[45] Dongyang Liu, Meina Kan, Shiguang Shan, and Xilin Chen. Function-consistent feature distillation. In ICLR, 2023. 1, 5, 6, 8, 2, 3 +[46] Li Liu, Qingle Huang, Sihao Lin, Hongwei Xie, Bing Wang, Xiaojun Chang, and Xiaodan Liang. Exploring inter-channel correlation for diversity-preserved knowledge distillation. In ICCV, 2021. 2, 5, 6, 1 +[47] Xiaolong Liu, Lujun Li, Chao Li, and Anbang Yao. Norm: Knowledge distillation via n-to-one representation matching. In ICLR, 2023. 1, 5, 8, 2, 3 +[48] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In CVPR, 2019. 2 +[49] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. NeurIPS, 2021. 3 +[50] Seyed-Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 2, 5, 6, 1, 3 +[51] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 3 +[52] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In IEEE TMM, 2019. 1, 2, 4, 5, 6, 7, 3 + +[53] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2, 4, 7, 1, 3, 5 +[54] Baoyun Peng, Xiao Jin, Jiaheng Liu, Shunfeng Zhou, Yichao Wu, Yu Liu, Dongsheng Li, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, 2019. 2, 4, 5, 6, 7, 1, 3 +[55] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE T-PAMI, 39(6):1137-1149, 2016. 6 +[56] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1, 2, 3, 5, 6 +[57] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semisupervised learning with consistency and confidence. In NeurIPS, 2020. 3 +[58] Wonchul Son, Jaemin Na, Junyong Choi, and Wonjun Hwang. Densely guided knowledge distillation using multiple teacher assistants. In ICCV, 2021. 2 +[59] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, 2024. 2, 5, 6, 7, 1, 3 +[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 5, 6, 1, 3 +[61] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, 2019. 1, 2, 4, 5, 7, 3 +[62] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 1, 6, 3 +[63] Shicai Wei, Chunbo Luo, and Yang Luo. Scale decoupled distillation. In CVPR, 2024. 6, 8, 1, 3 +[64] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In CVPR, 2023. 1, 3 +[65] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 6 +[66] Chuanguang Yang, Zhulin An, Linhang Cai, and Yongjun Xu. Hierarchical self-supervised augmented knowledge distillation. In IJCAI, 2021. 6 +[67] Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, and Qian Zhang. Cross-image relational knowledge distillation for semantic segmentation. In CVPR, 2022. 1, 2 +[68] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 5, 6, 1, 2 +[69] Longrong Yang, Xianpan Zhou, Xuewei Li, Liang Qiao, Zheyang Li, Ziwei Yang, Gaoang Wang, and Xi Li. Bridging cross-task protocol inconsistency for distillation in dense object detection. In ICCV, 2023. 1 +[70] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and global knowledge distillation for detectors. In CVPR, 2022. 1 + +[71] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked generative distillation. In ECCV, 2022. 1, 3 +[72] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. In ICCV, 2023. 2, 5, 1, 3 +[73] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Km. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2, 1 +[74] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 5, 6, 1, 3 +[75] Bo Zhang, Jiacheng Sui, and Li Niu. Foreground object search by distilling composite image feature. In ICCV, 2023. 1 +[76] Weijia Zhang, Dongnan Liu, Weidong Cai, and Chao Ma. Cross-view consistency regularisation for knowledge distillation. In ACM MM, 2024. 1, 5, 6, 2, 3 +[77] Weijia Zhang, Dongnan Liu, Chao Ma, and Weidong Cai. Alleviating foreground sparsity for semi-supervised monocular 3d object detection. In WACV, 2024. 1 +[78] Weijia Zhang, Yuehao Liu, Wu Ran, and Chao Ma. Cross-architecture distillation made simple with redundancy suppression. ICCV, 2025. 2, 6, 1 +[79] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 3 +[80] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, 2018. 1, 2, 3 +[81] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 2, 5, 6, 1, 3, 4 +[82] Kaixiang Zheng and En-Hui Yang. Knowledge distillation based on transformed teaching matching. In ICLR, 2024. 2, 5, 6, 1, 3 +[83] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In CVPR, 2023. 1 \ No newline at end of file diff --git a/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/images.zip b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c9aa50b5d26747e5dc8902e0c1a07e982925624 --- /dev/null +++ b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4e2904c705de8da9516b9d44684058dd7ecb8f0fa2f3b5856f89dbc91eaa162 +size 699959 diff --git a/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/layout.json b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f6418131a90c88a554b367a50af08a01505ebe10 --- /dev/null +++ b/2025/VRM_ Knowledge Distillation via Virtual Relation Matching/layout.json @@ -0,0 +1,11783 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "text", + "content": "VRM: Knowledge Distillation via Virtual Relation Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 142, + 207, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 142, + 207, + 157 + ], + "spans": [ + { + "bbox": [ + 135, + 142, + 207, + 157 + ], + "type": "text", + "content": "Weijia Zhang1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 143, + 282, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 143, + 282, + 156 + ], + "spans": [ + { + "bbox": [ + 241, + 143, + 282, + 156 + ], + "type": "text", + "content": "Fei Xie" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 158, + 309, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 158, + 309, + 172 + ], + "spans": [ + { + "bbox": [ + 151, + 158, + 309, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 158, + 309, + 172 + ], + "type": "text", + "content": " Shanghai Jiao Tong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 143, + 386, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 143, + 386, + 157 + ], + "spans": [ + { + "bbox": [ + 318, + 143, + 386, + 157 + ], + "type": "text", + "content": "Weidong Cai" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 421, + 143, + 477, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 143, + 477, + 156 + ], + "spans": [ + { + "bbox": [ + 421, + 143, + 477, + 156 + ], + "type": "text", + "content": "Chao " + }, + { + "bbox": [ + 421, + 143, + 477, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{Ma}^{1,*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 158, + 458, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 158, + 458, + 172 + ], + "spans": [ + { + "bbox": [ + 324, + 158, + 458, + 172 + ], + "type": "text", + "content": "2 The University of Sydney" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 173, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 173, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 126, + 173, + 481, + 185 + ], + "type": "text", + "content": "{weijia.zhang, feixie, chaoma}@sjtu.edu.cn tom.cai@sydney.edu.au" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "content": "Knowledge distillation (KD) aims to transfer the knowledge of a more capable yet cumbersome teacher model to a lightweight student model. In recent years, relation-based KD methods have fallen behind, as their instance-matching counterparts dominate in performance. In this paper, we revive relational KD by identifying and tackling several key issues in relation-based methods, including their susceptibility to overfitting and spurious responses. Specifically, we transfer novelly constructed affinity graphs that compactly encapsulate a wealth of beneficial inter-sample, inter-class, and inter-view correlations by exploiting virtual views and relations as a new kind of knowledge. As a result, the student has access to richer guidance signals and stronger regularisation throughout the distillation process. To further mitigate the adverse impact of spurious responses, we prune the affinity graphs by dynamically detaching redundant and unreliable edges. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO datasets demonstrate the superior performance of the proposed virtual relation matching (VRM) method, where it consistently sets new state-of-the-art records over a range of models, architectures, tasks, and set-ups. For instance, VRM for the first time hits " + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "inline_equation", + "content": "74.0\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "content": " accuracy for ResNet50 " + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "content": " MobileNetV2 distillation on ImageNet, and improves DeiT by " + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "inline_equation", + "content": "14.44\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "content": " on CIFAR-100 with a ResNet56 teacher. The code and models are released at https://github.com/VISION-SJTU/VRM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 575, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 575, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 575, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "Deep learning is achieving incredible performance at the cost of increasing model complexity and overheads. As a consequence, large and cumbersome neural models struggle to work in resource-constrained environments. Knowledge distillation (KD) [27] has been proposed to address this issue by transferring the knowledge of larger and more capable models to smaller and lightweight ones that are resource-friendly. KD works by minimising the dis" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 432, + 291 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 432, + 291 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 432, + 291 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 432, + 291 + ], + "type": "image", + "image_path": "6f457de3b44d1eb43aa33e863cc93e5ce4f2350c218ffe49ed23ec7eff009baa.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 293, + 410, + 303 + ], + "lines": [ + { + "bbox": [ + 339, + 293, + 410, + 303 + ], + "spans": [ + { + "bbox": [ + 339, + 293, + 410, + 303 + ], + "type": "text", + "content": "(a) Instance Matching" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 213, + 553, + 291 + ], + "blocks": [ + { + "bbox": [ + 436, + 213, + 553, + 291 + ], + "lines": [ + { + "bbox": [ + 436, + 213, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 436, + 213, + 553, + 291 + ], + "type": "image", + "image_path": "840db0a4cabc73b6a2b1d4c181b55da4e6800a7a0d5e3641339bb5477e3f17cf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 293, + 531, + 303 + ], + "lines": [ + { + "bbox": [ + 459, + 293, + 531, + 303 + ], + "spans": [ + { + "bbox": [ + 459, + 293, + 531, + 303 + ], + "type": "text", + "content": "(b) Relation Matching" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 306, + 433, + 384 + ], + "blocks": [ + { + "bbox": [ + 316, + 306, + 433, + 384 + ], + "lines": [ + { + "bbox": [ + 316, + 306, + 433, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 433, + 384 + ], + "type": "image", + "image_path": "8ac77d54bb8a81d21d0854823bf132eb0efae85c267e7e13120e2ccd55a3bd0a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 386, + 403, + 396 + ], + "lines": [ + { + "bbox": [ + 345, + 386, + 403, + 396 + ], + "spans": [ + { + "bbox": [ + 345, + 386, + 403, + 396 + ], + "type": "text", + "content": "(c) VRM (Dense)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 436, + 306, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 436, + 306, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 436, + 306, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 436, + 306, + 553, + 384 + ], + "type": "image", + "image_path": "ea953baeec179d0a681f15a1ef6e9512c81bac526bf87d200963122154805cf9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 386, + 525, + 396 + ], + "lines": [ + { + "bbox": [ + 464, + 386, + 525, + 396 + ], + "spans": [ + { + "bbox": [ + 464, + 386, + 525, + 396 + ], + "type": "text", + "content": "(d) VRM (Pruned)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 400, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 422 + ], + "type": "text", + "content": "Figure 1. Conceptual illustration of VRM compared to existing KD methods based on instance matching and relation matching." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 486 + ], + "type": "text", + "content": "tance between compact representations of knowledge extracted from the teacher and student models. It has found widespread application across a spectrum of downstream tasks [18, 22, 30, 34, 62, 64, 67, 70, 75, 77, 83]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 487, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 605 + ], + "type": "text", + "content": "According to the type of knowledge representations to be transferred, KD methods can be broadly categorised into feature-based [56], logit-based [27], and relation-based [52] approaches. The former two directly match the feature maps or logit vectors produced by the teacher and student models for each training sample, which is essentially instance matching (IM). By contrast, relation matching (RM) methods construct and match structured relations extracted within a batch of model responses. A conceptual illustration is presented in Figs. 1a and 1b." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Instance matching has been the prevailing distillation approach in recent years. Popular KD benchmarks see a dominance by IM methods such as FCFD [45], NORM [47], and CRLD [76], with different downstream tasks successfully tackled by directly adopting IM-based distillation [6, 14, 62, 69]. Yet, recent studies discovered that relational knowledge is more robust to variations in neural architectures, data modalities, and tasks [52, 61]. Meanwhile, methods transferring relations have also achieved promising" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2707" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "performance for a range of tasks, including but not limited to segmentation [67] and detection [15, 32]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 97, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 97, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 97, + 294, + 228 + ], + "type": "text", + "content": "Despite growing interest, relation-based methods still fall significantly short compared to their instance matching counterparts. Even the strongest RM method has been outperformed easily by recent IM solutions [29] (see Tabs. 1 and 2). RM-based methods also struggle with more challenging tasks such as object detection [29]. Moreover, previous RM solutions are primarily limited to matching inter-sample [29, 53, 61], inter-class [29], or inter-channel [46, 73] relations via simple Gram matrices. To our best knowledge, no different forms of relations other than these have been proposed since the work of DIST [29]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 231, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 295, + 350 + ], + "type": "text", + "content": "This paper fills this gap with a new kind of relations for KD - inter-view relations (Fig. 1c), which seamlessly and compactly integrate with previous inter-sample and inter-class relations. Our designs are motivated by two important observations made about RM methods in a set of pilot experiments: 1) RM methods are more susceptible to overfitting than IM methods; 2) RM methods are subject to an adverse gradient propagation effect. We empirically find that incorporating richer and more diverse relations into the matching objective helps mitigate both issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 483 + ], + "type": "text", + "content": "To this end, we generate virtual views of samples through simple transformations, followed by constructing virtual affinity graphs and transferring the virtual relations between real and virtual samples along the edges. In lieu of Gram matrices that suffer from significant knowledge loss [29, 53, 61], we preserve the raw relations along the secondary dimension as auxiliary knowledge which adds to the types and density of relational knowledge transferred. Moreover, we also prune our affinity graphs by striping away both redundant and unreliable edges to further alleviate the propagating gradients of spurious samples (Fig. 1d)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 628 + ], + "type": "text", + "content": "The above insights and remedies altogether lead to a novel Virtual Relation Matching (VRM) framework for knowledge distillation. VRM is conceptually simple, easy to implement, and devoid of complicated training procedures. It is capable of transferring rich, sophisticated knowledge robust to overfitting and spurious signals. VRM sets new state-of-the-art performance for different datasets, tasks, and settings. Perhaps more significant is that VRM makes relation-based methods regain competitiveness and back in the lead over instance matching approaches in different scenarios. To summarise, the contributions of this work include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 630, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 688 + ], + "type": "text", + "content": "- We make an early effort to present comparative analyses of existing KD methods through the lens of training dynamics and sample-wise gradients, and identify overfitting and spurious gradient diffusion as two main cruxes in relational KD methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- We distill richer, more diverse relations by generating virtual views, constructing virtual affinity graphs, and" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 120 + ], + "type": "text", + "content": "matching virtual relations. We also for the first time tackle relational KD with considerations of spurious samples and gradients by pruning redundant and unreliable edges, alongside designs to relax the matching criterion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 121, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 180 + ], + "type": "text", + "content": "- We present the streamlined VRM framework for knowledge distillation, with extensive experimental results on a diversity of models and tasks to highlight its superior performance, alongside rigorous analyses on the soundness and efficiency of our designs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 192, + 400, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 400, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 400, + 205 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 553, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 429 + ], + "type": "text", + "content": "KD via instance matching. KD was first proposed in Hinton et al. [27], where a student is trained to mimic the prediction of a pre-trained teacher model for each sample. Follow-up works have mostly followed such instance matching (IM) paradigm, and can be categorised into logit (or prediction)-based and feature-based methods according to what is matched. Logit-based KD has evolved from using adaptive-softened logits [33, 41] to decoupling target- and non-target logits [24, 72, 81] and applying logit transformation [59, 82]. Others [50, 58] set up auxiliary ad-hoc networks between teacher and student to facilitate logit transfer. Feature-based methods minimise the distance between the feature maps [1, 25, 56] or salient regions within in features [23, 74] from specified layers in both teacher and student networks. Some also design sophisticated distillation paths [7, 42]. All these methods are based on instance-wise transfer of knowledge, and are herein referred to as the \"IM\" methods, as illustrated by Fig. 1a." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 433, + 553, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 553, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 553, + 636 + ], + "type": "text", + "content": "KD via relation matching. Some works transfer instead mutual relations mined amongst the network outputs extracted from a batch of training instances. These relation matching (RM) methods usually involve constructing network outputs into compact relation representations that encode rich higher-order information, as depicted in Fig. 1b. Different relation encoding functions mare, including inter-sample [29, 48, 52-54, 61], inter-class [29], inter-channel [46, 73, 78], inter-layer [48], and contrastive [60] relations. To date, IM solutions have dominated KD with their superior performance, leading top-performing relation-based methods by considerable margins. While many new IM methods are found within the last two years, frustratingly few RM solutions are being proposed. In this work, we strive to close this gap with a new kind of relations for KD — inter-view virtual relations, and revive relation-based KD by making it overtake its IM counterparts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Learning with virtual knowledge. While not a standalone research topic, learning with virtual knowledge finds relevance in a variety of learning-based problems. For instance, a commonly used paradigm in 3D vision tasks is to learn (or construct) from the raw data a virtual view or representation as auxiliary knowledge in solving the main task, rang-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2708" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 216, + 151 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 216, + 151 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 216, + 151 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 216, + 151 + ], + "type": "image", + "image_path": "aa3d16d809bfea323d73787a4bc4b71c47ced412ea933092802c9a519e8a3ac9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 152, + 206, + 160 + ], + "lines": [ + { + "bbox": [ + 67, + 152, + 206, + 160 + ], + "spans": [ + { + "bbox": [ + 67, + 152, + 206, + 160 + ], + "type": "text", + "content": "(a) Increasing performance on training set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 376, + 151 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 376, + 151 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 376, + 151 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 376, + 151 + ], + "type": "image", + "image_path": "b21167aac795e5f47ec07926028c316876f9d7dc6fc217c47de434bdb24de9c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 152, + 369, + 160 + ], + "lines": [ + { + "bbox": [ + 226, + 152, + 369, + 160 + ], + "spans": [ + { + "bbox": [ + 226, + 152, + 369, + 160 + ], + "type": "text", + "content": "(b) Plateaued performance on validation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 379, + 151, + 552, + 160 + ], + "lines": [ + { + "bbox": [ + 379, + 151, + 552, + 160 + ], + "spans": [ + { + "bbox": [ + 379, + 151, + 552, + 160 + ], + "type": "text", + "content": "(c) Diffusion of gradient perturbations within a batch." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 163, + 498, + 175 + ], + "lines": [ + { + "bbox": [ + 110, + 163, + 498, + 175 + ], + "spans": [ + { + "bbox": [ + 110, + 163, + 498, + 175 + ], + "type": "text", + "content": "Figure 2. Pilot studies that reveal the overfitting and spurious gradient diffusion issues with RM-based KD." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 551, + 148 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 551, + 148 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 551, + 148 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 551, + 148 + ], + "type": "image", + "image_path": "7e8b217402999478d55d20b29ac28241d31e164a4e76e5bccd35629a820ee062.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 190, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 296, + 382 + ], + "type": "text", + "content": "ing from object reconstruction [5] and optical flow [2] to 3D semantic segmentation [36], monocular 3D object detection [13], and 3D GAN inversion [64]. More broadly, many data-efficient learning methods also share the spirit of utilising virtual knowledge. For instance, self-supervised learning methods generate virtual views of the unlabelled data to enable the learning of pretext tasks [11, 20]. Another popular paradigm is transformation-invariant representation learning [51, 57] in semi-supervised learning and domain adaptation. It enforces consistency between representations learnt for a raw sample and a virtual view of it. The virtual view is often obtained by applying semantic-preserving transformations to the raw sample [16, 17]. This work is more related to this later paradigm, but involves learning with virtual knowledge in a different context, via different approaches, and for a different problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 392, + 140, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 140, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 140, + 403 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "content": "KD methods generally employ a cross-entropy (CE) loss and a distillation loss to supervise student learning. The CE loss is computed between student logits " + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^s" + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "content": " for each sample and its ground-truth label " + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "content": ". The distillation loss matches teacher and student outputs via a distance metric " + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "content": ". In vanilla KD [27], " + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "content": " is the Kullback-Leibler divergence (KLD) between teacher logits " + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "content": " and student logits " + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^s" + }, + { + "bbox": [ + 55, + 411, + 296, + 496 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 96, + 499, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 499, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 96, + 499, + 295, + 550 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {i} ^ {\\mathrm {K D}} = \\phi_ {\\mathrm {K L D}} \\left(\\mathbf {z} _ {i} ^ {s}, \\mathbf {z} _ {i} ^ {t}\\right) \\\\ = \\tau^ {2} \\sum_ {j = 1} ^ {C} \\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {t} / \\tau\\right) \\log \\frac {\\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {t} / \\tau\\right)}{\\sigma_ {j} \\left(\\mathbf {z} _ {i} ^ {s} / \\tau\\right)}, \\tag {1} \\\\ \\end{array}", + "image_path": "78176363707cd43acfd6f91b0c8fa386021e130a9dc9f26cd54d0467dd749533.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "content": " is the Softmax operation with temperature parameter " + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "content": " is the number of classes. For feature-based methods, " + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "content": " can be the mean squared error (MSE) between teacher and student feature maps for each instance [56], i.e., " + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_i^{\\mathrm{KD}} = \\phi_{\\mathrm{MSE}}(\\mathbf{f}_i^s,\\mathbf{f}_i^t)" + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 618, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 689 + ], + "type": "text", + "content": "For relation-based methods, a relation encoding function " + }, + { + "bbox": [ + 55, + 618, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 55, + 618, + 295, + 689 + ], + "type": "text", + "content": " first abstracts teacher or student outputs of all instances within a training batch into a relational representation, before applying " + }, + { + "bbox": [ + 55, + 618, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot)" + }, + { + "bbox": [ + 55, + 618, + 295, + 689 + ], + "type": "text", + "content": " to match these relational representations between the teacher and the student. For example, the KD objectives of DIST [29] take the general form of:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 80, + 696, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 696, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 80, + 696, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {K D}} = \\phi \\left(\\psi \\left(\\mathbf {z} _ {1} ^ {s}, \\mathbf {z} _ {2} ^ {s}, \\dots , \\mathbf {z} _ {B} ^ {s}\\right), \\psi \\left(\\mathbf {z} _ {1} ^ {t}, \\mathbf {z} _ {2} ^ {t}, \\dots , \\mathbf {z} _ {B} ^ {t}\\right)\\right), \\tag {2}", + "image_path": "b80d1001dc49976a53a2d1b988ea18c24d0c5c543619b228f9bc1160fdd11eed.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "text", + "content": " is the batch size, and subscript " + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{KD}}" + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "text", + "content": " is dropped as the loss is computed for a batch of instances. DIST uses both inter-class and inter-sample relation encoders as " + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 313, + 190, + 553, + 227 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 236, + 394, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 394, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 394, + 248 + ], + "type": "text", + "content": "4. Pilot Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": "Training dynamics of KD methods. We examine the training dynamics of different relation-based methods on CIFAR-100 with ResNet32 " + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": " 4 " + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": " ResNet8 " + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": " 4 as the teacher-student pair. In Figs. 2a and 2b, we immediately notice that relational methods achieve significantly higher training accuracy, but they only marginally lead or even fall short in test accuracy compared to IM-based KD [27]. We hypothesise that relational methods are more prone to overfitting. This is expected given that the optimality of IM matching (cond. " + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": ") implies the optimality of relation matching (cond. " + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": "), while the converse does not hold. Concretely, " + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\Rightarrow \\mathcal{B} \\land \\neg \\mathcal{B} \\Rightarrow \\neg \\mathcal{A}" + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": ". In other words, relation matching is a weaker and less constrained objective than IM, which makes the student more readily fit the teaching signals and not generalise well. Thus, we conclude that " + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "inline_equation", + "content": "C1" + }, + { + "bbox": [ + 313, + 256, + 555, + 448 + ], + "type": "text", + "content": ": relation matching methods are more prone to overfitting." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": "Gradient analysis of KD methods. We investigate the gradient patterns within a batch when a spurious sample produces a major misguiding signal. To this end, we first generate two random vectors " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{y}\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{y}\\in \\mathbb{R}^{B\\times D}" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " is the batch size and " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " is the dimension of persample predictions, " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " is taken as the sample-wise predictions, and " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " the supervision signals. We then add a noise vector " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\epsilon = c\\cdot \\mathbf{z}" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " is a scaling factor. In this case, " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t + \\epsilon" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " becomes a spurious prediction within our batch. We compute the loss from " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " using either IM or RM objectives, and consider the change in sample-wise gradients " + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 313, + 452, + 555, + 608 + ], + "type": "text", + "content": " within the batch upon the injection of the spurious sample. Formally, we visualise:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 614, + 495, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 614, + 495, + 643 + ], + "spans": [ + { + "bbox": [ + 343, + 614, + 495, + 643 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {g} = \\left[ \\| \\frac {\\partial \\mathcal {L}}{\\partial \\left(\\mathbf {x} _ {i} ^ {\\prime}\\right)} \\| _ {2} - \\| \\frac {\\partial \\mathcal {L}}{\\partial \\left(\\mathbf {x} _ {i}\\right)} \\| _ {2} \\right] _ {i = 1} ^ {B},", + "image_path": "0a1e2b3615ce0fa4a5ecb343ff97b3f3ceb51998a4df2632d9c17a05c3bef13d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 645, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 645, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 326, + 645, + 553, + 658 + ], + "type": "interline_equation", + "content": "s. t. \\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} + \\epsilon \\cdot \\mathbb {I} (i = t), \\quad \\mathcal {L} \\in \\left\\{\\mathcal {L} _ {I M}, \\mathcal {L} _ {R M} \\right\\}. \\tag {3}", + "image_path": "d99bc8bcb72739995a04baa13231326ff9767c170d2e36b371ec4d5b9c0aceb1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "In Fig. 2c (" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "B = 64" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "t = 32" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "), when the IM objective is used, only the spurious sample receives a prominent gradient. Whereas for an RM objective, many other samples receive significant gradients as they are directly connected" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2709" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 211, + 151 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 211, + 151 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 211, + 151 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 211, + 151 + ], + "type": "image", + "image_path": "b3da20e409478ffba022d966b916e2054d8d901941aec90a5beb4abec29a4013.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 152, + 331, + 162 + ], + "lines": [ + { + "bbox": [ + 95, + 152, + 331, + 162 + ], + "spans": [ + { + "bbox": [ + 95, + 152, + 331, + 162 + ], + "type": "text", + "content": "Figure 3. Alleviated overfitting and improved training dynamics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 69, + 366, + 151 + ], + "blocks": [ + { + "bbox": [ + 212, + 69, + 366, + 151 + ], + "lines": [ + { + "bbox": [ + 212, + 69, + 366, + 151 + ], + "spans": [ + { + "bbox": [ + 212, + 69, + 366, + 151 + ], + "type": "image", + "image_path": "c0212815eefb87d1b1b23fee3da4c880b319dcaa39c11e53cf646abaa94e1691.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 371, + 70, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 371, + 70, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 371, + 70, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 371, + 70, + 552, + 148 + ], + "type": "image", + "image_path": "72708e7921ac197d3176c5c1eebf20d25dfecbeb15f305b059d2c943384bd253.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 152, + 533, + 163 + ], + "lines": [ + { + "bbox": [ + 388, + 152, + 533, + 163 + ], + "spans": [ + { + "bbox": [ + 388, + 152, + 533, + 163 + ], + "type": "text", + "content": "Figure 4. Alleviated spurious gradients." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "text", + "content": " in the computational graph of the RM loss. In other words, the spurious signals produced by one malign prediction will propagate to and affect all samples within a batch (in fact, those closer to " + }, + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "text", + "content": " within the prediction manifold are more strongly affected). This means other sample-wise predictions will be significantly updated only to accommodate a malign prediction, even if they are already in relatively good shape. Through this investigation, we discover that " + }, + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "inline_equation", + "content": "C2" + }, + { + "bbox": [ + 54, + 175, + 295, + 307 + ], + "type": "text", + "content": ": relation matching methods are more prone to the adverse impact of spurious samples. We also present this pilot analysis on ImageNet training in Fig. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 307, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 295, + 403 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 54, + 307, + 295, + 403 + ], + "type": "inline_equation", + "content": "C1" + }, + { + "bbox": [ + 54, + 307, + 295, + 403 + ], + "type": "text", + "content": ", common approaches to combat overfitting include the incorporation of richer learning signals and regularisation, which for RM-based KD methods means richer relations constructed and transferred. For " + }, + { + "bbox": [ + 54, + 307, + 295, + 403 + ], + "type": "inline_equation", + "content": "C2" + }, + { + "bbox": [ + 54, + 307, + 295, + 403 + ], + "type": "text", + "content": ", an intuitive solution is to identify and suppress the effect of spurious predictions or relations or to slacken the matching criterion. Guided by these principles, we now proceed to formally build up our method step by step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 413, + 111, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 111, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 111, + 425 + ], + "type": "text", + "content": "5. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "type": "text", + "content": "5.1. Constructing Inter-Sample Relations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": "We first construct relation graph " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " that encodes intersample affinity within a batch of sample predictions " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i\\}_{i = 1}^B" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": ". Different from [52, 61], our relations are constructed from predicted logits which embed more compact categorical knowledge. We use the pairwise distance between instance-wise predictions within a batch as our measure of affinity. Existing methods leverage the Gram matrices [29, 53, 54, 61] to encode inter-sample relations, but we find that this leads to collapsed inter-class knowledge via the inner product operation. Instead, our pairwise distance preserves the inter-class knowledge along the secondary dimension (i.e., the class dimension), which enables such information to be explicitly transferred as auxiliary knowledge alongside the matching of inter-sample relations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Thus, we have constructed a dense relation graph " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", which comprises " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " vertices and " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "B \\times B" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " edges: " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS} = (\\mathcal{V}^{IS}, \\mathcal{E}^{IS})" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Each vertex in " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " represents the prediction vector " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " for one instance within a batch, and is connected to all instances within the batch including itself. The attribute of edge " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{i,j}^{IS}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " connecting vertices " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " describes the class-wise relations between the predictions of instances " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". In practice, we can organise all edges into matrix" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 174, + 427, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 427, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 427, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IS}\\in \\mathbb{R}^{B\\times B\\times C}" + }, + { + "bbox": [ + 313, + 174, + 427, + 186 + ], + "type": "text", + "content": " , in which:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 196, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 196, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 339, + 196, + 553, + 220 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {i, j} ^ {I S} = \\frac {\\mathbf {z} _ {i} - \\mathbf {z} _ {j}}{\\| \\mathbf {z} _ {i} - \\mathbf {z} _ {j} \\| _ {2}} \\in \\mathbb {R} ^ {C} \\quad \\text {f o r} \\quad i, j \\in [ 1, B ], \\tag {4}", + "image_path": "cfe73cbec03fe1b120b4bcd67544a3db63f289841d1564d629aa76dce6e52900.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": "where we empirically find that normalisation along the secondary dimension helps regularise relations and improves performance. Concretely, " + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": " encodes inter-sample class-wise relations within a batch of " + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": " training samples. This design is different from and empirically more effective than previous relation formulations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 307, + 499, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 499, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 499, + 319 + ], + "type": "text", + "content": "5.2. Constructing Inter-Class Relations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": "Although " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": " encodes rich inter-sample relations within a batch, it fails to explicitly model inter-class correlation patterns that are also beneficial structured knowledge [29]. Therefore, we propose to build a novel inter-class batchwise relation graph " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IC} = (\\mathcal{V}^{IC},\\mathcal{E}^{IC})" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": ", instead of the Gram matrices in the inner product space as in [29]. The construction of " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IC}" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": " mirrors that of " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": ". Vertices in " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IC}" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": " are the class-wise logit vectors " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{w}\\in \\mathbb{R}^B" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": ". Each edge in " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IC}\\in \\mathbb{R}^{C\\times C\\times B}" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": " embeds the pairwise difference between the " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 323, + 555, + 445 + ], + "type": "text", + "content": "-th per-class vectors. Concretely:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 453, + 553, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 453, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 330, + 453, + 553, + 477 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {i, j} ^ {I C} = \\frac {\\mathbf {w} _ {i} - \\mathbf {w} _ {j}}{\\| \\mathbf {w} _ {i} - \\mathbf {w} _ {j} \\| _ {2}} \\in \\mathbb {R} ^ {B} \\quad \\text {f o r} \\quad i, j \\in [ 1, C ]. \\tag {5}", + "image_path": "bb14057f8b5d117ef6ef6b907f82eaa83157aafc2dd0d55d65946a1e6cc9bc22.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "content": "Our inter-class relations preserve the batch-wise discrepancies by treating them as a dimension of additional knowledge (reciprocal to the case of inter-sample relations), which is unlike any other previous methods [29]. We will demonstrate in Tab. 7 that our novel formulation of intersample and inter-class relations by preserving the raw affinity knowledge along the secondary dimension perform significantly better than previous relation encoders " + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "content": " via Gram matrices [29, 52, 53, 61] or third-order angular distances [52]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 612, + 481, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 481, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 481, + 624 + ], + "type": "text", + "content": "5.3. Constructing Virtual Relations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "For each prediction " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": " within a batch " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i\\}_{i=1}^B" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", we create a virtual view of it, denoted as “" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_i" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "\", by applying semantic-preserving transformations to original image " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ". While other transformations are applicable, we choose RandAugment [17] that applies stochastic image transformations (see Supplementary Material for details). With our batch of predictions augmented into " + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i, \\tilde{\\mathbf{z}}_i\\}_{i=1}^B" + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": ", we can construct a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2710" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 346 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 346 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 346 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 346 + ], + "type": "table", + "html": "
Teacher\nStudentVenueResNet56 ResNet20ResNet32×4 WRN-40-2 WRN-40-2 VGG13 ShuffleNetV2ResNet32×4 VGG8ResNet32×4 VGG13 MobileNetV2ResNet50 MobileNetV2 ShuffleNetV2ResNet40-2
Teacher72.3479.4275.6175.6174.6479.42
Student69.0672.5073.2671.9870.3671.82
Feature-based
FitNets [56]ICLR'1569.2173.5073.5872.2471.0273.54
AT [74]ICLR'1770.5573.4474.0872.7771.4372.73
CRD [60]ICLR'2071.1675.5175.4874.1473.9475.65
SRRL [68]ICLR'2171.1375.3375.5974.1873.44-
PEFD [12]NeurIPS'2270.0776.0876.0274.9274.35-
TaT [42]CVPR'2271.5975.8976.0674.9774.39-
ReviewKD [10]CVPR'2171.8975.6376.1275.0974.8477.78
NORM [47]ICLR'2371.3576.4975.6574.8273.9578.32
FCFD [45]ICLR'2371.9676.6276.4375.4675.2278.18
Logit-based
KD [27]arXiv'1570.6673.3374.9273.5472.9874.45
TAKD [50]AAAI'2070.8373.8175.1273.7873.2374.82
CTKD [41]AAAI'2371.1973.7975.4573.9373.5275.31
NKD [72]ICCV'2370.4076.3575.2474.0774.8676.26
DKD [81]CVPR'2271.9776.3276.2474.8174.6877.07
LSKD [59]CVPR'2471.4376.6276.1174.3774.3675.56
TTM [82]ICLR'2471.8376.1776.2374.3274.3376.55
CRLD [76]MM'2472.1077.6076.4575.5875.2778.27
Relation-based
RKD [52]CVPR'1969.6171.9073.3572.2271.4873.21
CC [54]CVPR'1969.6372.9773.5672.2170.7171.29
SP [61]ICCV'1969.6772.9473.8372.4372.6874.56
ICKD [46]ICCV'2171.7675.2575.6474.3373.42-
DIST [29]†NeurIPS'2271.7576.31-74.73-77.35
VRM-72.0978.7677.4776.4676.1979.34
", + "image_path": "2dde1c95c67306d36bb73fd00fc93f93592327bd45737ba9311dea63fddb3f87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 351, + 525, + 363 + ], + "lines": [ + { + "bbox": [ + 83, + 351, + 525, + 363 + ], + "spans": [ + { + "bbox": [ + 83, + 351, + 525, + 363 + ], + "type": "text", + "content": "Table 1. Results for same- and different-model teacher-student pairs on CIFAR-100. †: using re-trained, stronger teachers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 377, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 403 + ], + "type": "text", + "content": "larger inter-sample edge matrix " + }, + { + "bbox": [ + 54, + 377, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IS}\\in \\mathbb{R}^{2B\\times 2B\\times C}" + }, + { + "bbox": [ + 54, + 377, + 295, + 403 + ], + "type": "text", + "content": " and a larger inter-class edge matrix " + }, + { + "bbox": [ + 54, + 377, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IC}\\in \\mathbb{R}^{C\\times C\\times 2B}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "text", + "content": "From the perspective of sample views, our new " + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IS}" + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IC}" + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "text", + "content": " constructed from a batch of real and virtual samples essentially capture three types of knowledge, namely relations amongst real views (denoted as \"real-real\"), relations amongst virtual views (\"virtual-virtual\"), and relations between pairs of real and virtual views (\"real-virtual\"). For instance, a real-virtual edge that connects real vertex " + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "text", + "content": " and virtual vertex " + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IS}" + }, + { + "bbox": [ + 54, + 403, + 295, + 499 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 507, + 295, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 507, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 116, + 507, + 295, + 533 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {m, n} ^ {I S} = \\frac {\\mathbf {z} _ {m} - \\tilde {\\mathbf {z}} _ {n}}{\\| \\mathbf {z} _ {m} - \\tilde {\\mathbf {z}} _ {n} \\| _ {2}} \\in \\mathbb {R} ^ {C}. \\tag {6}", + "image_path": "6a71d7467a014dee85deb13dd64a2f2e2090a7e6e0922968a88b36f7eee3124a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 211, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 211, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 211, + 553 + ], + "type": "text", + "content": "5.4. Pruning into Sparse Graphs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": "Pruning redundant edges. The augmented " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IS}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " in Sec. 5.3 contains " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "2B\\times 2B" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " edges, leading to quadrupled overheads. For better efficiency, we prune " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " into sparse graphs. Noticing that " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{IS}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " is symmetric along its diagonals, we first prune its redundant half to save up to " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " edge count. We also remove intra-view edges to get a further " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " reduction, as we empirically find them redundant and harm knowledge transfer. For " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IC}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": ", we decompose the augmented batch of predictions of size " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "2B" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " into a real-view batch and a virtual-view batch, each of size " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": ", and in lieu use the inter-sample batch-wise affinity vectors between them as its vertices. Compared to their original intra-view formulation of size " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "2B" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " in Sec. 5.3, this design encodes purely inter-view" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "content": "affinity knowledge with halved parameters. With redundant edge pruning (REP), our graphs now become sparse and the remaining edges can again be rearranged into compact matrices: " + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{ISV} \\in \\mathbb{R}^{B \\times B \\times C}" + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{ICV} \\in \\mathbb{R}^{C \\times C \\times B}" + }, + { + "bbox": [ + 313, + 379, + 555, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "text", + "content": "Pruning unreliable edges. To mitigate the diffusive effect of spurious predictions discovered earlier, we further identify and prune the unreliable edges. In previous graph learning works, the absolute certainty of two vertices are often used to determine the reliability of an edge. For instance, REM [9] computes the reliability of an edge as the mean of the maximum predicted probabilities of two samples (i.e., two vertices). However, we argue this could bias the learning towards easy samples. Instead, we measure the discrepancy between two predictions. The larger this discrepancy, the more unreliable the relation constructed. The unreliable edge pruning (UEP) criterion is given by " + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{i,j}^{ISV} = \\emptyset" + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(\\mathbf{z}_i^s,\\tilde{\\mathbf{z}}_j^s) > P_m" + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(\\cdot)" + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "text", + "content": " computes the joint entropy (JE) between two predictions and " + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "inline_equation", + "content": "P_{m}" + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 432, + 556, + 649 + ], + "type": "text", + "content": "-th percentile within the batch. Note that the criterion is enforced on student predictions, resulting in adaptive and dynamic pruning as different edges get pruned in each iteration, which improves learning most of the time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 658, + 404, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 404, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 404, + 672 + ], + "type": "text", + "content": "5.5. Full Objective" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IC}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " constructed for both teacher and student predictions, our VRM objective matches edge matrices " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{ISV}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{ICV}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " between teacher and student via distance" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "2711" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 281, + 316 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 281, + 316 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 281, + 316 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 281, + 316 + ], + "type": "table", + "html": "
Teacher StudentVenueResNet34 ResNet18ResNet50 MobileNetV1
Teacher Student73.31/91.4276.16/92.86
69.75/89.0768.87/88.76
Feature-based
AT [74]ICLR'1770.69/90.0169.56/89.33
OFD [25]ICCV'1970.81/89.9871.25/90.34
CRD [60]ICLR'2071.17/90.1371.37/90.41
CAT-KD [23]CVPR'2371.26/90.4572.24/91.13
SimKD [8]CVPR'2271.59/90.4872.25/90.86
ReviewKD [10]CVPR'2171.61/90.5172.56/91.00
SRRL [68]ICLR'2171.73/90.6072.49/90.92
PEFD [12]NeurIPS'2271.94/90.6873.16/91.24
RSD [78]ICCV'2572.18/90.7573.05/91.26
FCFD [45]ICLR'2372.24/90.7473.37/91.35
Logit-based
KD [27]arXiv'1570.66/89.8868.58/88.98
TAKD [50]AAAI'2070.78/90.1670.82/90.01
DKD [81]CVPR'2271.70/90.4172.05/91.05
SDD [63]CVPR'2471.14/90.0572.24/90.71
TTM [82]ICLR'2472.19/-73.09/-
LSKD [59]CVPR'2472.08/90.7473.22/91.59
CRLD [76]MM'2472.37/90.7673.53/91.43
Relation-based
RKD [52]CVPR'1971.34/90.3771.32/90.62
CC [54]CVPR'1970.74/--
ICKD [46]ICCV'2172.19/90.72-
DIST [29]NeurIPS'2272.07/90.4273.24/91.12
VRM-72.67/90.6874.16/91.78
", + "image_path": "ce7978ec40fd1f6f030fd6ce981ed98edebed3167630536cbaee6b9366995065.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "text", + "content": "metric " + }, + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "text", + "content": " , for which the Huber loss is used. Formally, " + }, + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "inline_equation", + "content": "L_{vrm}^{ISV} = \\phi (\\mathcal{E}_S^{ISV},\\mathcal{E}_T^{ISV})" + }, + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "inline_equation", + "content": "L_{vrm}^{ICV} = \\phi (\\mathcal{E}_S^{ICV},\\mathcal{E}_T^{ICV})" + }, + { + "bbox": [ + 55, + 348, + 295, + 396 + ], + "type": "text", + "content": " The full objective is a weighted combination of the CE loss and the proposed VRM losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 403, + 294, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 403, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 294, + 418 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L _ {c e} + \\alpha L _ {v r m} ^ {I S V} + \\beta L _ {v r m} ^ {I C V} \\tag {7}", + "image_path": "38ba2d709960d1bbf3cbb853da844445f0d4a156d823d5f95c3b214aa9ff1f1d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "text", + "content": " is the CE loss applied to student's predictions of both real and virtual views and supervised by GT; " + }, + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 426, + 296, + 487 + ], + "type": "text", + "content": " are balancing scalars. A PyTorch-style pseudo-code for the computing of VRM losses is provided in Algorithm 1 in Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 137, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 137, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 137, + 510 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 182, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 182, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 182, + 529 + ], + "type": "text", + "content": "6.1. Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 594 + ], + "type": "text", + "content": "Our method is evaluated on CIFAR-100 [35] and ImageNet [19] datasets for image classification, and MSCOCO [43] for object detection. All experimental configurations follow the standard practice in prior works. More details are provided in Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 141, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 141, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 141, + 612 + ], + "type": "text", + "content": "6.2. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "Results on CIFAR-100. The results for different KD methods on CIFAR-100 are shown in Tab. 1. For same-model KD, (left of Tab. 1), VRM surpasses all previous relation-based methods across all teacher-student pairs by large margins, including DIST [29]. Noticeably, VRM significantly outperforms the strongest feature-based method, FCFD [45]. On average, it also performs much better than top logit-based methods such as CRLD [76] and TTM [82]." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 314, + 70, + 537, + 222 + ], + "blocks": [ + { + "bbox": [ + 126, + 322, + 238, + 333 + ], + "lines": [ + { + "bbox": [ + 126, + 322, + 238, + 333 + ], + "spans": [ + { + "bbox": [ + 126, + 322, + 238, + 333 + ], + "type": "text", + "content": "Table 2. Results on ImageNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 70, + 537, + 222 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 537, + 222 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 537, + 222 + ], + "type": "table", + "html": "
ResNet101→ResNet18ResNet50→MobileNetV2
APAP50AP75APAP50AP75
Teacher42.0462.4845.8840.2261.0243.81
Student33.2653.6135.2629.4748.8730.90
Feature-based
FitNets [56]34.4354.1636.7130.2049.8031.69
FGFI [62]35.4455.5138.1731.1650.6832.92
TAKD [50]34.5955.3537.1231.2651.0333.46
ReviewKD [10]36.7556.7234.0033.7153.1536.13
FCFD [45]37.3757.6040.3434.9755.0437.51
Logit-based
KD [27]33.9754.6636.6230.1350.2831.35
CTKD [41]34.5655.4336.9131.3952.3433.10
LSKD [59]---31.7452.7733.40
DKD [81]35.0556.6037.5432.3453.7734.01
Relation-based
VRM35.4656.9137.9332.6753.9634.48
", + "image_path": "c3e086eada0defab6125b3cd6c5d118e32ef6737082eb5e50bb575057889cabb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 306, + 264, + 548, + 318 + ], + "blocks": [ + { + "bbox": [ + 301, + 228, + 548, + 251 + ], + "lines": [ + { + "bbox": [ + 301, + 228, + 548, + 251 + ], + "spans": [ + { + "bbox": [ + 301, + 228, + 548, + 251 + ], + "type": "text", + "content": "Table 3. Results on MS-COCO. All detectors are based on a Faster-RCNN [55] backbone." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 264, + 548, + 318 + ], + "lines": [ + { + "bbox": [ + 306, + 264, + 548, + 318 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 548, + 318 + ], + "type": "table", + "html": "
TeacherStudentT.S.KDATSPLGAutoKDLSKDVRM
ResNet56DeiT-T70.4365.0873.2573.5167.3678.1578.5878.5579.52
ResNet56T2T-ViT-770.4369.3774.1574.0172.2678.3578.6278.4378.88
ResNet56PiT-T70.4373.5875.4776.0374.9778.4878.5178.7679.25
ResNet56PVT-T70.4369.2274.6677.0770.4877.4873.6078.4379.42
", + "image_path": "491eae01393c61f99a52c060a2f96b34b5b6ff124e9907387ffe9eea8739c296.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 322, + 544, + 333 + ], + "lines": [ + { + "bbox": [ + 306, + 322, + 544, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 544, + 333 + ], + "type": "text", + "content": "Table 4. Results for cross-architecture distillation on CIFAR-100." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 349, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 433 + ], + "type": "text", + "content": "These are significant, given that IM methods are known to be naturally more adapt at distilling between homogeneous model pairs. For different-model " + }, + { + "bbox": [ + 313, + 349, + 555, + 433 + ], + "type": "inline_equation", + "content": "KD" + }, + { + "bbox": [ + 313, + 349, + 555, + 433 + ], + "type": "text", + "content": " (right of Tab. 1), which RM methods are supposed to be more competent with, VRM readily surpasses all RM and IM methods with notable margins. These results highlight VRM's versatility on both homogeneous and heterogeneous model pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 438, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 535 + ], + "type": "text", + "content": "Results on ImageNet. VRM surpasses all feature- and relation-based methods on this large-scale dataset, as shown in Tab. 2. The advantage of VRM is again more apparent over the heterogeneous pair (i.e., ResNet50 → MobileNetV2), whereby it for the first time hits " + }, + { + "bbox": [ + 313, + 438, + 555, + 535 + ], + "type": "inline_equation", + "content": "74.0\\%" + }, + { + "bbox": [ + 313, + 438, + 555, + 535 + ], + "type": "text", + "content": " Top-1 accuracy. Notably, VRM outperforms strong competitors such as FCFD [45] and CRLD [76] with comparable or even less computational overheads as compared in Tab. 10." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "type": "text", + "content": "Results on MS-COCO. We demonstrate that VRM generalises to more challenging tasks by adapting it to object detection. From Tab. 3, VRM performs better than existing logit-based methods and competitively to feature-based methods. VRM is slightly behind top-performing feature-based methods such as FCFD [45] and ReviewKD [10]. This is a consequence of the very nature of the object detection task, where fine-grained contextual features play a vital role, making feature-based methods inherently better-off [62]. Nonetheless, we experimentally demonstrate that VRM is effective on object detection, whereby it improves the baseline by " + }, + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "type": "inline_equation", + "content": "2 - 3\\%" + }, + { + "bbox": [ + 313, + 540, + 556, + 696 + ], + "type": "text", + "content": " AP, surpasses all logit-based methods, and is on par with strong feature-based methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "type": "text", + "content": "Results for cross-architecture distillation. We follow the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2712" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 210, + 148 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 210, + 148 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 210, + 148 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 210, + 148 + ], + "type": "table", + "html": "
ComponentRN32×4VGG13
RN8×4VGG8
Baseline72.5070.36
+ Match {EIS}75.6474.08
+ Match {EIS, IIC}75.9474.47
+ Match {EISV}77.6375.30
+ Match {EISV, IICV}78.3875.56
+ Pruning78.7676.19
", + "image_path": "f37391f0572961781b805f213a899138618759a6b8ffcf51103f385f3331240a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 225, + 73, + 365, + 140 + ], + "blocks": [ + { + "bbox": [ + 71, + 151, + 201, + 162 + ], + "lines": [ + { + "bbox": [ + 71, + 151, + 201, + 162 + ], + "spans": [ + { + "bbox": [ + 71, + 151, + 201, + 162 + ], + "type": "text", + "content": "Table 5. Ablation of major designs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 225, + 73, + 365, + 140 + ], + "lines": [ + { + "bbox": [ + 225, + 73, + 365, + 140 + ], + "spans": [ + { + "bbox": [ + 225, + 73, + 365, + 140 + ], + "type": "table", + "html": "
Matching ChoiceRN32×4 RN8×4VGG13 VGG8RN50 MNV2
Match ∅72.5070.3668.33
Match {E}78.7676.1972.30
Match {E, V}78.6375.6070.97
Match {V}78.4675.7768.33
", + "image_path": "d6fe391660ea9000943952d2441732830b09ea7cc4fb5d4cee5f58dbcf45a01f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 377, + 70, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 227, + 151, + 361, + 163 + ], + "lines": [ + { + "bbox": [ + 227, + 151, + 361, + 163 + ], + "spans": [ + { + "bbox": [ + 227, + 151, + 361, + 163 + ], + "type": "text", + "content": "Table 6. Effect of matching vertices." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 377, + 70, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 377, + 70, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 377, + 70, + 544, + 148 + ], + "type": "table", + "html": "
ψ(·)Shape of Rel. MatrixMethodRN32×4 VGG13RN50
RN8×4VGG8MNV2
Gram matrices[B,B] & [C,C]CC, SP, PKT, DIST77.3175.1270.28
Angular relations[B,B,B]RKD77.1475.0969.65
ISV & ICV[B,B,C] & [C,C,B]VRM78.4675.3072.10
", + "image_path": "53000251d04dcc574d08a5277d6b788d45661e46ff2a3a8b6cae0a97d82571f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 62, + 177, + 211, + 256 + ], + "blocks": [ + { + "bbox": [ + 373, + 151, + 548, + 163 + ], + "lines": [ + { + "bbox": [ + 373, + 151, + 548, + 163 + ], + "spans": [ + { + "bbox": [ + 373, + 151, + 548, + 163 + ], + "type": "text", + "content": "Table 7. Comparing different relation functions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 177, + 211, + 256 + ], + "lines": [ + { + "bbox": [ + 62, + 177, + 211, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 177, + 211, + 256 + ], + "type": "table", + "html": "
α3264128256512
Acc.78.0378.1378.7678.3277.84
β8163264128
Acc.78.4078.5178.7678.8678.60
n01234
Acc.77.9278.3878.7678.5578.43
", + "image_path": "cc63dc419c51672785c2d08b9e795142d515588dc7df4a77a7a7b5aaa234888e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 259, + 206, + 270 + ], + "lines": [ + { + "bbox": [ + 61, + 259, + 206, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 259, + 206, + 270 + ], + "type": "text", + "content": "Table 8. Effect of different " + }, + { + "bbox": [ + 61, + 259, + 206, + 270 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta" + }, + { + "bbox": [ + 61, + 259, + 206, + 270 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 61, + 259, + 206, + 270 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 259, + 206, + 270 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 215, + 175, + 370, + 259 + ], + "blocks": [ + { + "bbox": [ + 215, + 175, + 370, + 259 + ], + "lines": [ + { + "bbox": [ + 215, + 175, + 370, + 259 + ], + "spans": [ + { + "bbox": [ + 215, + 175, + 370, + 259 + ], + "type": "image", + "image_path": "8057ee8253a401617275712487a0a801df9e72b95d91266ecb296ec831c94fd9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 259, + 369, + 270 + ], + "lines": [ + { + "bbox": [ + 213, + 259, + 369, + 270 + ], + "spans": [ + { + "bbox": [ + 213, + 259, + 369, + 270 + ], + "type": "text", + "content": "Figure 5. Effect of different difficulty pairs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 381, + 175, + 538, + 259 + ], + "blocks": [ + { + "bbox": [ + 381, + 175, + 538, + 259 + ], + "lines": [ + { + "bbox": [ + 381, + 175, + 538, + 259 + ], + "spans": [ + { + "bbox": [ + 381, + 175, + 538, + 259 + ], + "type": "image", + "image_path": "ff7bbbfb266c386c21ebb690185ca2051f2f760e948fcd35de637e19d029ea83.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 259, + 525, + 271 + ], + "lines": [ + { + "bbox": [ + 395, + 259, + 525, + 271 + ], + "spans": [ + { + "bbox": [ + 395, + 259, + 525, + 271 + ], + "type": "text", + "content": "Figure 6. Robustness to varying " + }, + { + "bbox": [ + 395, + 259, + 525, + 271 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 395, + 259, + 525, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 285, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 296, + 382 + ], + "type": "text", + "content": "settings in [39] and [59] to perform CNN-to-ViT cross-architecture distillation. Tab. 4 shows that VRM sets new record under this set-up for different ViT architectures. Notably, by simply replacing the vanilla KD loss with the proposed VRM objective and making no other modifications, VRM improves DeiT-T's accuracy by " + }, + { + "bbox": [ + 54, + 285, + 296, + 382 + ], + "type": "inline_equation", + "content": "14.44\\%" + }, + { + "bbox": [ + 54, + 285, + 296, + 382 + ], + "type": "text", + "content": ", leading the relation-based SP [61] by " + }, + { + "bbox": [ + 54, + 285, + 296, + 382 + ], + "type": "inline_equation", + "content": "12.16\\%" + }, + { + "bbox": [ + 54, + 285, + 296, + 382 + ], + "type": "text", + "content": " and even surpassing AutoKD [40] with AutoML search." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 392, + 156, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 156, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 156, + 403 + ], + "type": "text", + "content": "6.3. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 410, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 471 + ], + "type": "text", + "content": "Ablations of main design choices. In Tab. 5, starting from the baseline where only the CE loss is applied, we gradually incorporate each of our designs. As shown, each extra design consistently brings performance gains, which corroborate the validity of individual design choices." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 475, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 594 + ], + "type": "text", + "content": "Effect of different relation encoding functions. Tab. 7 compares our proposed relation encoding scheme with existing ones. To demonstrate the superiority of our formulation, we substitute it with existing Gram matrices [29, 53, 54, 61] or angle-wise relations [52]. Results show that our inter-sample class-wise and inter-class sample-wise relations consistently yield much better performance than existing relations, which highlights VRM's contribution of a novel and superior type of relation encoding in addition to the introduction of virtual relation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": "Effect of varying hyperparameters. VRM involves three major hyperparameters: " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": ". According to Tab. 8, VRM works generally well with " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": " in a reasonable range. Larger " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": " may produce better results for certain model pairs but worse results for others. For " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": ", a larger " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": " means more difficult images and larger inter-view prediction discrepancy. From Tab. 8, " + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 55, + 600, + 296, + 696 + ], + "type": "text", + "content": " works best, whereas other values also report competitive results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Effect of different real-virtual difficulty pairs. We also" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": "experiment with different strength combinations for generating the real and virtual samples. The results are presented in Fig. 5, where \"Weak\" denotes the default transformation used in previous methods, and \"Strong\" denotes RandAugment with " + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 313, + 285, + 554, + 370 + ], + "type": "text", + "content": ". Overall, we conclude that 1) a moderate discrepancy leads to optimal performance, and 2) the discrepancy is significant to the success of VRM." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "content": "Robustness to varying batch sizes. Relational methods are known to be sensitive to training batch size " + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "content": ". We conduct experiments to examine how robust VRM is against varying " + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "content": ". To adjust the learning rate accordingly, we consider two LR scaling rules: linear LR scaling [21] and square root LR scaling [11]. From Fig. 6, VRM yields competitive results across a wide range of " + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "content": " values. In contrast, DIST [29] degrades significantly as " + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 374, + 555, + 471 + ], + "type": "text", + "content": " grows." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 476, + 416, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 416, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 416, + 489 + ], + "type": "text", + "content": "6.4. Further Analysis" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 493, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 613 + ], + "type": "text", + "content": "Analysis of training dynamics. Figs. 3a and 3b plot the per-epoch training and validation set accuracies throughout training. VRM maintains a all-time lead in both training and validation performance, with faster convergence. Besides, while VRM's lead in training performance tapers off towards the end of training, it remains more substantial, if not further enlarging, in validation performance, which highlights its superior generalisation properties. This analysis also shows our designs are effective in mitigating the issues with existing RM methods identified in the pilot studies." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Analysis of spurious gradients. We revisit our second pilot study by analysing how spurious gradients are suppressed in real ImageNet training. From Fig. 4, similar gradient diffusion patterns are observed in real distillation. Samples semantically more similar to the sample with spurious predictions are more affected by spurious gradients (e.g., the dog image with a watery background resembles the bubbly texture in another sample that is actually a computer mouse.)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2713" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 76, + 119, + 129 + ], + "blocks": [ + { + "bbox": [ + 60, + 76, + 119, + 129 + ], + "lines": [ + { + "bbox": [ + 60, + 76, + 119, + 129 + ], + "spans": [ + { + "bbox": [ + 60, + 76, + 119, + 129 + ], + "type": "image", + "image_path": "e84ea35fef3f4f65b04d72ce95b992dfe3d0187817038425e616407067f19a22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 132, + 101, + 140 + ], + "lines": [ + { + "bbox": [ + 74, + 132, + 101, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 132, + 101, + 140 + ], + "type": "text", + "content": "(a) RKD" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 76, + 181, + 129 + ], + "blocks": [ + { + "bbox": [ + 123, + 76, + 181, + 129 + ], + "lines": [ + { + "bbox": [ + 123, + 76, + 181, + 129 + ], + "spans": [ + { + "bbox": [ + 123, + 76, + 181, + 129 + ], + "type": "image", + "image_path": "63b01b1c402a6c63364bd81dce0e80a5d9866341c766c6e4279870c0ea8aa8bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 132, + 164, + 140 + ], + "lines": [ + { + "bbox": [ + 136, + 132, + 164, + 140 + ], + "spans": [ + { + "bbox": [ + 136, + 132, + 164, + 140 + ], + "type": "text", + "content": "(b) VRM" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 142, + 331, + 162 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 331, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 331, + 162 + ], + "type": "text", + "content": "Figure 7. t-SNE (a-b) and loss landscape (c-d) visualisations for " + }, + { + "bbox": [ + 55, + 142, + 331, + 162 + ], + "type": "inline_equation", + "content": "\\mathrm{RN8\\times 4}" + }, + { + "bbox": [ + 55, + 142, + 331, + 162 + ], + "type": "text", + "content": " students distilled with an " + }, + { + "bbox": [ + 55, + 142, + 331, + 162 + ], + "type": "inline_equation", + "content": "\\mathrm{RN32\\times 4}" + }, + { + "bbox": [ + 55, + 142, + 331, + 162 + ], + "type": "text", + "content": " teacher via RKD and VRM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 76, + 253, + 129 + ], + "blocks": [ + { + "bbox": [ + 186, + 76, + 253, + 129 + ], + "lines": [ + { + "bbox": [ + 186, + 76, + 253, + 129 + ], + "spans": [ + { + "bbox": [ + 186, + 76, + 253, + 129 + ], + "type": "image", + "image_path": "6138ea77c96853fdbdd4379a8f7942b8d118cf73678d2e190471b4df95241e85.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 132, + 232, + 140 + ], + "lines": [ + { + "bbox": [ + 205, + 132, + 232, + 140 + ], + "spans": [ + { + "bbox": [ + 205, + 132, + 232, + 140 + ], + "type": "text", + "content": "(c) RKD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 259, + 76, + 327, + 129 + ], + "blocks": [ + { + "bbox": [ + 259, + 76, + 327, + 129 + ], + "lines": [ + { + "bbox": [ + 259, + 76, + 327, + 129 + ], + "spans": [ + { + "bbox": [ + 259, + 76, + 327, + 129 + ], + "type": "image", + "image_path": "18d18218b903574cecce0e17c870c55cd2061ae5253df19b4a98e6607c5783d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 132, + 306, + 140 + ], + "lines": [ + { + "bbox": [ + 278, + 132, + 306, + 140 + ], + "spans": [ + { + "bbox": [ + 278, + 132, + 306, + 140 + ], + "type": "text", + "content": "(d) VRM" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 175, + 297, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 175, + 297, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 175, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 175, + 297, + 232 + ], + "type": "table", + "html": "
MethodKDFitNetsRKDICKDDISTCRDReviewKD
Train. Time (ms)24.926.731.028.027.241.239.9
Peak GPU Mem. (MB)32333033038133014181042
MethodSDDMLLDCRLDNORMPEFDFCFDVRM
Train. Time (ms)34.257.243.235.136.256.447.2
Peak GPU Mem. (MB)6905765511806701953579
", + "image_path": "293d0b4c4036c64333aa98be682431c579e9b009937ba707b1a26755e951f227.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 264, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 336 + ], + "type": "text", + "content": "Visualisation of embedding space. We conduct t-SNE analysis on student penultimate layer embeddings learnt via different methods. As presented in Fig. 7a and 7b, VRM leads to more compact per-class clusters with clearer inter-class separation and less stray points. These imply better student features distilled by VRM for the downstream task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": "Visualisation of loss landscape. We further analyse the generalisation and convergence properties of our method through the lens of visualised loss landscape [38]. From Fig. 7c and 7d, compared to RKD, VRM has a wider, flatter, and deeper region of minima – typical hint of better model generalisation and robustness; this wide convexity basin surrounded by salient pikes in various directions indicates excellent convergence properties. More visualisations are provided in Supplementary Material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": "Vertex matching. Within the constructed graphs " + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IS}" + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{IC}" + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": ", VRM matches edges " + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{ISV}" + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{ICV}" + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": " which carry relational knowledge. By extension, we can naturally expect vertices " + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": " to be also transferred. From Tab. 6, matching vertices is not as effective. The advantage of matching relations (i.e., edges) is more pronounced for heterogeneous KD pairs such as ResNet50 " + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": " MobileNetV2. Adding vertex matching to the proposed edge matching does not improve but instead degrade performance. We argue this is because introducing vertex matching objectives make the matching criterion more stringent which is against our motivation of using more slackened matching." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Role of transformation operations. We emphasise the strong performance of VRM is not a simple outcome of the transformations involved. This can be verified by the results in Fig. 5 and Tab. 9, where using only the default weak transformation (i.e., random crop and horizontal flip) for both views achieves highly competitive and even superior performance under VRM. In fact, in the case of \"Weak-Weak\", both views still differ because of the stochastic operations used. Essentially, the success of VRM lies exactly" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 348, + 79, + 543, + 136 + ], + "blocks": [ + { + "bbox": [ + 63, + 237, + 288, + 247 + ], + "lines": [ + { + "bbox": [ + 63, + 237, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 63, + 237, + 288, + 247 + ], + "type": "text", + "content": "Table 10. Training efficiency of different distillation methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 79, + 543, + 136 + ], + "lines": [ + { + "bbox": [ + 348, + 79, + 543, + 136 + ], + "spans": [ + { + "bbox": [ + 348, + 79, + 543, + 136 + ], + "type": "table", + "html": "
RealVirtualRN50→MNV2RN34→RN18
DefaultDefault74.16/91.7872.67/90.68
DefaultRandAug(n=0)74.10/91.7072.60/90.88
DefaultRandAug(n=1)74.13/91.6372.59/90.82
DefaultRandAug(n=2)74.13/91.6872.48/90.73
", + "image_path": "c9131fcb620a108edff397a910f14885bf0d2ec79ab70861b0d5c9255260e62a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 313, + 174, + 552, + 231 + ], + "blocks": [ + { + "bbox": [ + 342, + 141, + 549, + 163 + ], + "lines": [ + { + "bbox": [ + 342, + 141, + 549, + 163 + ], + "spans": [ + { + "bbox": [ + 342, + 141, + 549, + 163 + ], + "type": "text", + "content": "Table 9. Effect of different transformation strength pairs on ImageNet." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 174, + 552, + 231 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 552, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 552, + 231 + ], + "type": "table", + "html": "
PruningRN32×4→RN8×4V13→V8W40-2→SNV2
Acc.tbatch(s)Mem. (MB)Acc.Acc.
w/o78.3861.434.375.5678.45
+ REP78.4646.313.675.3078.41
+ UEP78.7647.213.176.1978.62
", + "image_path": "33724832b6d61ae537652ef225ff4e12b0ed977c7c593d820866f860b4218477.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 238, + 541, + 249 + ], + "lines": [ + { + "bbox": [ + 320, + 238, + 541, + 249 + ], + "spans": [ + { + "bbox": [ + 320, + 238, + 541, + 249 + ], + "type": "text", + "content": "Table 11. Effectiveness and efficiency of pruning operations." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 265, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 361 + ], + "type": "text", + "content": "in the discrepancy of teacher and student predictions, where regularisation comes into crucial play. On smaller and easier datasets such as CIFAR-100, the gap between teacher and student predictions are relatively small, and manually enlarging this gap via extra transformations leads to promising outcomes. Whereas the more difficult ImageNet data already sees large teacher-student prediction discrepancy, such that it benefits less from further amplification." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 365, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 472 + ], + "type": "text", + "content": "Training efficiency. From Tab. 10, VRM is reasonably efficient compared to existing algorithms, including some leading feature- [12, 45, 47] and logit-based [33, 63] methods. VRM does not introduce any extra overheads at inference. We also breakdown the proposed pruning designs in Tab. 11 to understand how they individually impact the performance and efficiency of VRM. We see REP boosts efficiency at minimal performance cost, whereas UEP trades minimal efficiency cost for performance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 476, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 553, + 502 + ], + "type": "text", + "content": "More experiments, analyses, and discussions are provided in Supplementary Material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 511, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 388, + 523 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 531, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 712 + ], + "type": "text", + "content": "This paper presents VRM, a novel knowledge distillation framework that constructs and transfers virtual relations. Our designs are motivated by a set of pilot experiments, from which we identified two main cruxes with existing relation-based KD methods: their tendency to overfit and susceptibility to adverse gradient propagation. A series of tailored designs are developed and are shown to successfully mitigate these issues. We have conducted extensive experiments on different tasks and multiple datasets and verified VRM's validity and superiority in diverse settings, whereby it consistently achieves state-of-the-art performance. We hope that this work could renew the community's interest in relation-based knowledge distillation, and encourage more systematic reassessment of the design principles of such solutions." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2714" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 150, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 150, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 150, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 150, + 296, + 183 + ], + "type": "text", + "content": "[1] Sungsoo Ahn, Shell Xu Hu, Andreas Damianou, Neil D. Lawrence, and Zhenwen Dai. Variational information distillation for knowledge transfer. In CVPR, 2019. 2, 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 184, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 206 + ], + "type": "text", + "content": "[2] Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning optical flow from still images. In CVPR, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 207, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 229 + ], + "type": "text", + "content": "[3] Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. NeurIPS, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "type": "text", + "content": "[4] David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation. arXiv preprint arXiv:2106.04732, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 276, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 296, + 308 + ], + "type": "text", + "content": "[5] Joao Carreira, Abhishek Kar, Shubham Tulsiani, and Jitendra Malik. Virtual view networks for object reconstruction. In CVPR, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 309, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 296, + 352 + ], + "type": "text", + "content": "[6] Jiahao Chang, Shuo Wang, Hai-Ming Xu, Zehui Chen, Chenhongyi Yang, and Feng Zhao. Detrdistill: A universal knowledge distillation framework for detr-families. In ICCV, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 354, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 387 + ], + "type": "text", + "content": "[7] Defang Chen, Jian-Ping Mei, Yuan Zhang, Can Wang, Zhe Wang, Yan Feng, and Chun Chen. Cross-layer distillation with semantic calibration. In AAAI, 2021. 2, 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 388, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 420 + ], + "type": "text", + "content": "[8] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 422, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 454 + ], + "type": "text", + "content": "[9] Peibin Chen, Tao Ma, Xu Qin, Weidi Xu, and Shuchang Zhou. Data-efficient semi-supervised learning by reliable edge mining. In CVPR, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 456, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 296, + 487 + ], + "type": "text", + "content": "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, 2021. 5, 6, 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 489, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 489, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 489, + 296, + 521 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 2020. 3, 7, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 523, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 296, + 555 + ], + "type": "text", + "content": "[12] Yudong Chen, Sen Wang, Jiajun Liu, Xuwei Xu, Frank de Hoog, and Zi Huang. Improved feature distillation via projector ensemble. In NeurIPS, 2022. 5, 6, 8, 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 557, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 589 + ], + "type": "text", + "content": "[13] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3d object detection in autonomous driving. In CVPR, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "type": "text", + "content": "[14] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. In ICLR, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "text", + "content": "[15] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learning spatial features for monocular 3d object detection. *ICLR*, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[16] Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation strategies from data. In CVPR, 2019. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In NeurIPS, 2020. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "text", + "content": "[18] Kaiwen Cui, Yingchen Yu, Fangneng Zhan, Shengcai Liao, Shijian Lu, and Eric P Xing. Kd-dlgan: Data limited image generation via knowledge distillation. In CVPR, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[20] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 208, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 229 + ], + "type": "text", + "content": "[21] P Goyal. Accurate, large minibatch sgd: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "text", + "content": "[22] Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In ICLR, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, 2023. 2, 6, 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[24] Zhiwei Hao, Jianyuan Guo, Kai Han, Yehui Tang, Han Hu, Yunhe Wang, and Chang Xu. One-for-all: Bridge the gap between heterogeneous architectures in knowledge distillation. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 2, 6, 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "text", + "content": "[26] Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. Knowledge transfer via distillation of activation boundaries formed by hidden neurons. In AAAI, 2019. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[27] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowlegde in a neural network. In arXiv:1503.02531, 2015. 1, 2, 3, 5, 6, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[28] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. In arXiv:1704.04861, 2017. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "text", + "content": "[29] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. In NeurIPS, 2022. 2, 3, 4, 5, 6, 7, 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[30] Yuge Huang, Jiaxiang Wu, Xingkun Xu, and Shouhong Ding. Evaluation-oriented knowledge distillation for deep face recognition. In CVPR, 2022. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 579, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 612 + ], + "type": "text", + "content": "[31] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. In arXiv:1707.01219, 2017. 1, 2, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "text", + "content": "[32] Sujin Jang, Dae Ung Jo, Sung Ju Hwang, Dongwook Lee, and Daehyun Ji. Stxd: structural and temporal cross-modal distillation for multi-view 3d object detection. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "type": "text", + "content": "[33] Ying Jin, Jiaqi Wang, and Dahua Lin. Multi-level logit distillation. In CVPR, 2023. 2, 8" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[34] Minha Kim, Shahroz Tariq, and Simon S Woo. Fretal: Generalizing deepfake detection using knowledge distillation and representation learning. In CVPR, 2021. 1" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2715" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[35] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 6, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 294, + 139 + ], + "type": "text", + "content": "[36] Abhijit Kundu, Xiaqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, and Caroline Pantofaru. Virtual multi-view fusion for 3d semantic segmentation. In ECCV, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 161 + ], + "type": "text", + "content": "[37] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In ICLR, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "type": "text", + "content": "[38] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. NeurIPS, 2018. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "text", + "content": "[39] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, 2022. 7, 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "text", + "content": "[40] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In ICCV, 2023. 7, 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 261, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 303 + ], + "type": "text", + "content": "[41] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, 2023. 2, 5, 6, 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[42] Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. Knowledge distillation via the target-aware transformer. In CVPR, 2022. 2, 5, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[44] Tsung-Yi Lin, Piotr Dólár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[45] Dongyang Liu, Meina Kan, Shiguang Shan, and Xilin Chen. Function-consistent feature distillation. In ICLR, 2023. 1, 5, 6, 8, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[46] Li Liu, Qingle Huang, Sihao Lin, Hongwei Xie, Bing Wang, Xiaojun Chang, and Xiaodan Liang. Exploring inter-channel correlation for diversity-preserved knowledge distillation. In ICCV, 2021. 2, 5, 6, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "text", + "content": "[47] Xiaolong Liu, Lujun Li, Chao Li, and Anbang Yao. Norm: Knowledge distillation via n-to-one representation matching. In ICLR, 2023. 1, 5, 8, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "text", + "content": "[48] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan. Knowledge distillation via instance relationship graph. In CVPR, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[49] Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, and Marco Nadai. Efficient training of visual transformers with small datasets. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[50] Seyed-Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 2, 5, 6, 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "text", + "content": "[51] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 679, + 294, + 712 + ], + "type": "text", + "content": "[52] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In IEEE TMM, 2019. 1, 2, 4, 5, 6, 7, 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[53] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 2, 4, 7, 1, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[54] Baoyun Peng, Xiao Jin, Jiaheng Liu, Shunfeng Zhou, Yichao Wu, Yu Liu, Dongsheng Li, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, 2019. 2, 4, 5, 6, 7, 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "text", + "content": "[55] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE T-PAMI, 39(6):1137-1149, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "type": "text", + "content": "[56] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[57] Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying semisupervised learning with consistency and confidence. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 282, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 316 + ], + "type": "text", + "content": "[58] Wonchul Son, Jaemin Na, Junyong Choi, and Wonjun Hwang. Densely guided knowledge distillation using multiple teacher assistants. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "text", + "content": "[59] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, 2024. 2, 5, 6, 7, 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "text", + "content": "[60] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. In ICLR, 2020. 2, 5, 6, 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 382, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 403 + ], + "type": "text", + "content": "[61] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, 2019. 1, 2, 4, 5, 7, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 436 + ], + "type": "text", + "content": "[62] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In CVPR, 2019. 1, 6, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 437, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 459 + ], + "type": "text", + "content": "[63] Shicai Wei, Chunbo Luo, and Yang Luo. Scale decoupled distillation. In CVPR, 2024. 6, 8, 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[64] Jiaxin Xie, Hao Ouyang, Jingtan Piao, Chenyang Lei, and Qifeng Chen. High-fidelity 3d gan inversion by pseudomulti-view optimization. In CVPR, 2023. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "text", + "content": "[65] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, 2020. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "text", + "content": "[66] Chuanguang Yang, Zhulin An, Linhang Cai, and Yongjun Xu. Hierarchical self-supervised augmented knowledge distillation. In IJCAI, 2021. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[67] Chuanguang Yang, Helong Zhou, Zhulin An, Xue Jiang, Yongjun Xu, and Qian Zhang. Cross-image relational knowledge distillation for semantic segmentation. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "text", + "content": "[68] Jing Yang, Brais Martinez, Adrian Bulat, and Georgios Tzimiropoulos. Knowledge distillation via softmax regression representation learning. In ICLR, 2021. 5, 6, 1, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[69] Longrong Yang, Xianpan Zhou, Xuewei Li, Liang Qiao, Zheyang Li, Ziwei Yang, Gaoang Wang, and Xi Li. Bridging cross-task protocol inconsistency for distillation in dense object detection. In ICCV, 2023. 1" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[70] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and global knowledge distillation for detectors. In CVPR, 2022. 1" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2716" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 545 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[71] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked generative distillation. In ECCV, 2022. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[72] Zhendong Yang, Ailing Zeng, Zhe Li, Tianke Zhang, Chun Yuan, and Yu Li. From knowledge distillation to self-knowledge distillation: A unified approach with normalized loss and customized soft labels. In ICCV, 2023. 2, 5, 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[73] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Km. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, 2017. 2, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[74] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 5, 6, 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "text", + "content": "[75] Bo Zhang, Jiacheng Sui, and Li Niu. Foreground object search by distilling composite image feature. In ICCV, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[76] Weijia Zhang, Dongnan Liu, Weidong Cai, and Chao Ma. Cross-view consistency regularisation for knowledge distillation. In ACM MM, 2024. 1, 5, 6, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[77] Weijia Zhang, Dongnan Liu, Chao Ma, and Weidong Cai. Alleviating foreground sparsity for semi-supervised monocular 3d object detection. In WACV, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "text", + "content": "[78] Weijia Zhang, Yuehao Liu, Wu Ran, and Chao Ma. Cross-architecture distillation made simple with redundancy suppression. ICCV, 2025. 2, 6, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "text", + "content": "[79] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 400, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 431 + ], + "type": "text", + "content": "[80] Ying Zhang, Tao Xiang, Timothy M. Hospedales, and Huchuan Lu. Deep mutual learning. In CVPR, 2018. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "type": "text", + "content": "[81] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, 2022. 2, 5, 6, 1, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 468, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 499 + ], + "type": "text", + "content": "[82] Kaixiang Zheng and En-Hui Yang. Knowledge distillation based on transformed teaching matching. In ICLR, 2024. 2, 5, 6, 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 502, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 294, + 545 + ], + "type": "text", + "content": "[83] Shengchao Zhou, Weizhou Liu, Chen Hu, Shuchang Zhou, and Chao Ma. Unidistill: A universal cross-modality knowledge distillation framework for 3d object detection in bird's-eye view. In CVPR, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2717" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_content_list.json b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c973fbaac3467aac576fc27156cc1de5e384bcb8 --- /dev/null +++ b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_content_list.json @@ -0,0 +1,1472 @@ +[ + { + "type": "text", + "text": "VSC: Visual Search Compositional Text-to-Image Diffusion Model", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Do Huu $\\mathrm{Dat}^{1*}$ Nam Hyeon-Woo² Po-Yuan Mao³ Tae-Hyun Oh²,⁴", + "bbox": [ + 210, + 191, + 785, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ VinUniversity $^{2}$ Dept. of EE, POSTECH $^{3}$ Academia Sinica $^{4}$ KAIST", + "bbox": [ + 199, + 222, + 797, + 241 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d28fd3a9091d57e8ff97628cb45d00453c997c2451d21b6dc79d704401c02bdc.jpg", + "image_caption": [ + "Figure 1. Left: Original pipeline and generated image of Diffusion Model. Right: VSC individually \"search\" for visual information of each binding pair by generating images for each separate pair (\"red car\" and \"yellow bicycle\"). Finally, VSC uses them as references to generate the correct compositional image." + ], + "image_footnote": [], + "bbox": [ + 140, + 284, + 867, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 465, + 326, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models have shown impressive capabilities in generating realistic visuals from natural-language prompts, yet they often struggle with accurately binding attributes to corresponding objects, especially in prompts containing multiple attribute-object pairs. This challenge primarily arises from the limitations of commonly used text encoders, such as CLIP, which can fail to encode complex linguistic relationships and modifiers effectively. Existing approaches have attempted to mitigate these issues through attention map control during inference and the use of layout information or fine-tuning during training, yet they face performance drops with increased prompt complexity. In this work, we introduce a novel compositional generation method that leverages pairwise image embeddings to improve attribute-object binding. Our approach decomposes complex prompts into sub-prompts, generates corresponding images, and computes visual prototypes that fuse with text embeddings to enhance representation. By applying segmentation-based localization training, we address cross-attention misalignment, achieving improved accuracy in binding multiple attributes to objects. Our approaches outperform existing compositional text-to-image diffusion models on the benchmark T2I CompBench, achieving better image quality, evaluated by humans, and emerging robustness under scaling the number", + "bbox": [ + 88, + 502, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of binding pairs in the prompt.", + "bbox": [ + 513, + 468, + 717, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 512, + 643, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compositionality is a fundamental ability that enables humans to naturally comprehend the world. It allows us to quickly recognize new objects through their components or construct complex sentences by combining words. This skill empowers us to make \"infinite use of finite mean\" [6], endlessly reusing and recombining familiar concepts in diverse ways. In terms of developing compositional AI systems, diffusion models rise as the most feasible component due to their inherent enormous mode coverage and flexible controllability. The diffusion model's possession of compositional generalization has been proved with synthetic data and the dataset CelebA, where the images are easily decomposed into gender and race [26]. Additionally, diffusion models for generating images from text prompts are capable of creating remarkably realistic visuals, allowing users to shape the content of generated images by crafting detailed and complex natural-language prompts.", + "bbox": [ + 509, + 537, + 908, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these capabilities, these models often fail to fully utilize compositionality and produce inaccurate images when provided with more intricate descriptions. A prevalent issue is improper binding, as illustrated in the left side of Fig. 1, where descriptive words fail to correctly influence the visual characteristics of specific nouns they are intended to modify. For instance, the prompt specifying \"A photo", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work is done as an intern at POSTECH", + "bbox": [ + 109, + 886, + 326, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of a red car and a yellow bicycle” may produce an image with mixed or misplaced attributes. One likely cause for these inconsistencies is the reliance on text encoders like CLIP within diffusion models. Although CLIP demonstrates strong performance in aligning images with textual descriptions, it faces significant challenges in handling complex linguistic structures. Specifically, it struggles with encoding hierarchical relationships and modifiers, which are crucial for understanding nuanced textual inputs. Previous studies [20, 33, 34] have highlighted that CLIP tends to simplify textual inputs by representing them as a “bag-of-concepts,” disregarding the intricate dependencies and relationships between words. These inherent weaknesses pose limitations for diffusion models that rely on CLIP for text encoding, as they hinder the model's capacity to fully capture and respect the precise linguistic relationships specified in prompts.", + "bbox": [ + 89, + 90, + 485, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Serial search theory [35] in cognitive psychology suggests that when humans encounter objects with multiple features (e.g., color and shape), our brain processes them individually to bind these features into a cohesive perception. The theory has been successfully applied to improve the capability of large vision-language models [3]. Noticeably, CLIP likewise strongly materializes compositionality in the single binding setting [20] and this property is well-preserved in the latent diffusion model (refer to Section 6.1). Providing spatial information such as bounding box, segmentation maps, blobs, [7, 10, 25, 40] can drastically improve the compositional generation by controlling the attention to be bounded in a specific regions while maintaining a reasonable layout. Those methods can be interpreted as disentangling the desired image into object-level generation. This leads to our motivation to generate images from every single attribute-object pair as references for improving the compositional generalization of the text-to-image diffusion model.", + "bbox": [ + 91, + 335, + 486, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, the diffusion model can merge and edit images, namely subject-driven image generation, by quickly fine-tuning the model [30] that encodes a unique identifier that refers to the subject. Notably, Xiao et al. [38] propose a tuning-free method via training the diffusion model alongside an image encoder and an MLP. Besides, training an MLP has also been shown to be effective in improving the compositionality of the text-to-image model. Therefore, taking the psychological theory as inspiration, with the above previous works in diffusion models, our key idea, illustrated on the right side of Fig. 1, is to develop a framework that generates images for each binding pair individually and ultimately fuses these images into a complete image from a given prompt containing several attribute-object binding pairs. The fusing process requires the use of an external MLP layer, so we create a synthetic dataset and fine-tune the MLP. Moreover, we extract the segmentation maps of the object and deploy cross-attention localization training as in [38] to fix the erroneous cross-attention maps, aiming to", + "bbox": [ + 89, + 613, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "address the extreme decrease while increasing the number of attribute-object pairs.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 532, + 121, + 872, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present VSC, a novel compositional text-to-image diffusion generation method based on a tuning-free, subject-driven image generation framework. Remarkably, VSC requires training only an MLP and the final layers of the image encoder.", + "- We leverage the remarkable ability of the pre-trained text-to-image diffusion model in handling single attribute binding prompts to create synthetic images as reference images throughout training and inference.", + "- We achieve state-of-the-art performance on the attribute binding benchmark compared to other compositional text-to-image diffusion models, and provide extensive analysis as well as ablation studies." + ], + "bbox": [ + 511, + 136, + 906, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compositional Text-to-Image Diffusion Model: In recent work, Huang et al. [15] introduce a benchmark to test compositionality in text-to-image models, highlighting the vulnerability of open-source models to simple compositional prompts. They also propose a fine-tuning approach to enhance compositionality in these models. Several existing studies suggest different methods for enhancing compositionality in text-to-image models in both inference time and training time. In terms of inference time, one popular way is controlling the attention maps of the cross-attention layers [1, 22, 28, 36], which would ensure the maps of binding tokens are overlapping. However, because the attention scores are illustrated to be erroneous [39], then scaling the number of pairs leads to a drastic decrease in performance in those works. Furthermore, combining the guidance in a structured approach [9, 21] also mitigates the binding problems.", + "bbox": [ + 511, + 372, + 908, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During training time, other works leverage the layout information (bounding box, segmentation)[10, 25, 40] to correctly represent the spatial information of the image and fine-tune the diffusion model to concentrate on the binding problem. Additionally, [15, 39] demonstrate the efficiency of fine-tuning the UNet or an MLP layer, which transforms the text embeddings. Our work aims to resolve the discrepancy between cross-attention maps and propose a new approach to tackling compositional generalization without the need for human-annotated layouts or direct latent space manipulation.", + "bbox": [ + 511, + 614, + 908, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies have explored compositionality in Vision-Language Models (VLMs) through structured scene [13, 14, 23] leverage scene graphs to enhance understanding and generation of complex visual scenes, while [8, 17] introduce compositional training and modular architectures to improve reasoning capabilities. These approaches are closely related to our goal of enabling compositional text-to-image generation, and further connect our method to the broader literature on structured VLM learning.", + "bbox": [ + 511, + 765, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19154", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb9fec195d78e3edebf15679a3748ed5d97132c76a5d5996a6806b9574fb353a.jpg", + "image_caption": [ + "Figure 2. The training pipeline of VSC: Given the input text prompt \"A blue apple and a green backpack\", VSC first generates images of each binding pair (\"blue apple\" and \"green backpack\"), then uses an image encoder to extract visual prototype features, which is later used to augment the text embeddings via a trainable MLP module. Finally, a frozen pre-trained diffusion model generates the image with the augmented text embeddings. Additionally, we compute the $L_{loc}$ loss between cross-attention maps and segmentation maps to enhance the precision of cross-attention maps." + ], + "image_footnote": [], + "bbox": [ + 116, + 90, + 857, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subject-driven Image Generation: Text-to-image personalization is a growing application where models are fine-tuned to generate unique renditions of subjects based on a few reference images [11, 19, 30]. Notable works like DreamBooth [30] and Custom Diffusion [19] excel in this area, generating highly personalized outputs by fine-tuning a pre-trained text-to-image diffusion model with a few reference images and a textual prompt in the form of \"a [v] [class name]\" or \"photo of a [v] [class name].\" In this format, [v] serves as the subject's unique identifier—usually a rare token with minimal inherent meaning—while [class name] represents a broad category for the subject.", + "bbox": [ + 88, + 345, + 485, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent approaches to subject-driven image generation focus on synthesizing new images of specific subjects using only a few reference photos but often require computationally expensive tuning, such as DreamBooth [30], Textual Inversion [24], and Custom Diffusion [19]. Some methods, like Tuning Encoder [12], aim to reduce fine-tuning steps by creating initial latent codes with pre-trained encoders but still face scalability challenges. Tuning-free alternatives like X&Fuse [18], ELITE [37], and InstantBooth [32] use embedding techniques to map reference images into text conditioning, although these methods primarily support single-subject generation. For multi-subject scenarios, Custom Diffusion [19] employs joint tuning, while others, like SpaText [2] and Collage Diffusion [31], rely on segmentation masks to define object layouts. However, these techniques struggle with subjects of similar types or require user-specified layouts, which limit flexibility. FastComposer [38] overcomes these limitations by implementing cross-attention localization for better subject separation. To the best of our knowledge, we are the first to utilize subject-driven image generation in solving attribute-object binding problems, fully operating on synthetic images as reference images.", + "bbox": [ + 89, + 569, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Approach", + "text_level": 1, + "bbox": [ + 511, + 344, + 661, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 369, + 653, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We build upon the Stable Diffusion (SD) model, a state-of-the-art framework comprising three main components: a variational autoencoder (VAE), a denoiser as U-Net in SD 1.4, SD 2.1 or Transformer in SD 3.5, and a text encoder. In this setup, the VAE encoder $E$ compresses an image $x$ into a smaller latent space representation $z$ , which is then perturbed with Gaussian noise $\\varepsilon$ during the forward diffusion process. The denoiser, parameterized by $\\theta$ , is trained to denoise the latent representation by predicting and removing the noise. Text prompts condition this denoising process via a cross-attention mechanism, where the text encoder $\\psi$ maps prompts $P$ to embeddings $\\psi(P)$ . During training, the model minimizes a noise loss function defined as:", + "bbox": [ + 511, + 391, + 906, + 589 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n o i s e}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), P, \\varepsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\psi (P)\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 599, + 906, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{t}$ is the latent variable at time step $t$ . At inference, a noise sample $z_{T}$ is drawn from $\\mathcal{N}(0,1)$ and gradually refined by the denoiser to recover $z_{0}$ , after which the VAE decoder $D$ reconstructs the final image $\\hat{x} = D(z_0)$ .", + "bbox": [ + 511, + 632, + 906, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The cross-attention mechanism, in both the U-Net and Transformer backbone, conditions the denoising process on the text prompt. This is achieved by transforming the latent code $z \\in \\mathbb{R}^{(h \\times w) \\times f}$ and text embeddings $c \\in \\mathbb{R}^{n \\times d}$ , where the text encoder $\\psi$ converts the text $P$ into a set of $d$ -dimensional embeddings. Both the latent code and the text embeddings are projected into Query, Key, and Value matrices: $Q = W^q z$ , $K = W^k c$ , and $V = W^v c$ , where $W^q \\in \\mathbb{R}^{f \\times d'}$ and $W^k$ , $W^v \\in \\mathbb{R}^{d \\times d'}$ are the learned projection matrices, and $d'$ is the shared dimension of Query, Key, and Value. The attention scores are computed as:", + "bbox": [ + 511, + 691, + 910, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d ^ {\\prime}}}\\right) \\in [ 0, 1 ] ^ {(h \\times w) \\times n} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 869, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "allowing the model to compute a weighted sum over the Value matrix, resulting in the cross-attention output $z_{\\mathrm{attn}} = AV \\in \\mathbb{R}^{(h \\times w) \\times d'}$ . This mechanism distributes textual information across the latent spatial dimensions, where $A[i,j,k]$ denotes the information flow from the $k$ -th text token to the latent pixel at position $(i,j)$ . This cross-attention structure provides a foundation for interpreting semantic information in the latent code.", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Visual Embedding Fusion", + "text_level": 1, + "bbox": [ + 89, + 218, + 326, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a prompt $P$ including multiple binding pairs $[a_1,o_1],[a_2,o_2],\\ldots ,[a_n,o_n]]$ having the corresponding index $\\mathcal{I} = [i_{a1},i_{o1},i_{a2},i_{o2},\\dots ,i_{an},i_{on}]$ , a pretrained image encoder $\\phi :\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^D$ and a lightweight MLP: $\\mathbb{R}^{2*D}\\rightarrow \\mathbb{R}^D$ . We first extract the pairs and generate $m$ images from each pair $[a_n,o_n]$ using SD, so we have the corresponding set of reference images $[r_1^1,\\dots r_1^m ],\\dots ,[r_n^1,\\dots ,r_n^m ]]$ . Then, the set of images is fed into the image encoder $\\phi$ to produce the image embeddings and we take the mean of each set to get the prototype representation of each binding pair according to the following equations:", + "bbox": [ + 89, + 241, + 486, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {j} ^ {k} = \\phi \\left(r _ {j} ^ {k}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 426, + 483, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {j} = \\frac {1}{m} \\sum_ {k = 0} ^ {m} v _ {j} ^ {k} \\quad \\text {f o r} \\quad j \\in \\mathcal {I} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 449, + 483, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resulting in the visual prototype features $[e_1, e_2, \\ldots, e_n]$ . Finally, we augment the text embeddings $c$ with those visual prototype features. For each token in $c$ :", + "bbox": [ + 89, + 494, + 485, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} c _ {i}, & i \\notin I \\\\ \\operatorname {M L P} ([ c _ {i}, e _ {j} ]), & i = i _ {j} \\in I \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 544, + 483, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cross-attention maps in the original Stable Diffusion are erroneous, leading to incorrect binding (refer to Section 6.3). Therefore, we adopt the localizing cross-attention loss from FastComposer [38] with minor modifications enhancing the alignment between the cross-attention maps of attribute and object. Let $\\mathcal{M} = \\{M_1,M_2,\\dots ,M_n\\}$ be the segmentation masks for the reference pairs existing in the denoising target. Define $A_{i_a} = A[:,:,i_a],A_{i_o} = A[:,:,i_o]\\in [0,1]^{h\\times w}$ as the cross-attention maps of the $i$ -th pairs, comprising one map for the attribute and another map for the object. We then enforce the both 2 maps $A_{i_a},A_{i_o}$ to approximate the segmentation mask $M_j$ for the $j$ -th subject token, i.e., $A_{i_a}\\approx A_{i_o}\\approx M_j$ . We use a balanced $L_{1}$ loss that minimizes the difference between the cross-attention map and the segmentation mask:", + "bbox": [ + 89, + 589, + 485, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {l o c}} = \\frac {1}{n} \\sum_ {j = 1} ^ {n} \\left[ \\left(\\operatorname {m e a n} \\left(A _ {i _ {a}} \\odot \\bar {M} _ {j}\\right) - \\operatorname {m e a n} \\left(A _ {i _ {a}} \\odot M _ {j}\\right)\\right) \\right. \\\\ \\left. + \\left(\\operatorname {m e a n} \\left(A _ {i _ {o}} \\odot \\bar {M} _ {j}\\right) - \\operatorname {m e a n} \\left(A _ {i _ {o}} \\odot M _ {j}\\right)\\right) \\right], \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 821, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{mean}(A) = \\frac{1}{h\\cdot w}\\sum_{i = 1}^{h}\\sum_{k = 1}^{w}A_{ik}$ . The overall training objective is then defined as:", + "bbox": [ + 511, + 89, + 906, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {n o i s e}} + \\lambda L _ {\\text {l o c}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 133, + 774, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Creation", + "text_level": 1, + "bbox": [ + 513, + 160, + 681, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We make use of the T2I-CompBench dataset [15], concentrating on three attribute categories—color, texture, and shape—each containing 700 prompts for training. To generate high-quality images, we use three generative models: Stable Diffusion 3.5 [29], Flux, and SynGen [28] to generate 300 images for each prompt. Then we use Mask2Former [5] for instance segmentation and OpenCLIP [16] to compute the alignment score of the segmentation and the given attribute-object exists in the prompt. Specifically, we run a greedy algorithm to assign the mask with the text embedding that has the highest alignment score. Additionally, our evaluation emphasizes the disentangled BLIP-Visual Question Answering (VQA) score introduced by Huang et al. [15] as a metric for image quality. The VQA score measures how precisely an image captures the compositional elements specified in a prompt and aligns more closely with human judgment compared to metrics like CLIP-Score [27]. Finally, we select 45 images with the best scores for each prompt, resulting in the final dataset consisting of 90,000 images and their precise segmentation.", + "bbox": [ + 511, + 181, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 498, + 645, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Evaluation Results", + "text_level": 1, + "bbox": [ + 513, + 523, + 694, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How well can VSC compose the object with its corresponding attribute? Huang et al. [15] has proposed the BLIP-vqa score to evaluate the composition capability. We adopt their evaluation method on attributes of color, shape, and texture. Furthermore, we compute the harmonic mean (HM) to unify the scores into one as follows:", + "bbox": [ + 511, + 544, + 906, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H M} = \\frac {3}{\\frac {1}{C} + \\frac {1}{T} + \\frac {1}{S}},\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 645, + 779, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C, T,$ and $S$ represent the individual performance scores for Color, Texture, and Shape, respectively. We evaluate the methods on T2I CompBench dataset.", + "bbox": [ + 511, + 688, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 lists the performance of baseline, ours, and previous work, such as Composable [21] and Structured [9]. In Stable Diffusion 1.4, our method achieves the highest scores across all three attributes (Color: 0.66, Texture: 0.61, Shape: 0.47), outperforming the other methods. In Stable Diffusion 2.1, our method also achieves the best performance, with scores of 0.74 for Color, 0.64 for Texture, and 0.53 for Shape, indicating significant improvements over other models like GORS-unbiased [15], WiCLIP [39], and SynGen [28]. Stable Diffusion 3.5 with VSC archives the highest performance, resulting in 0.85, 0.79, and 0.63 for Color,", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19156", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f1f6bbeb8e33c7a0b2f39f5d4b7e86b13d5c2d3e68f440792f3fd4cda6412721.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelColorTextureShapeHM
Stable Diffusion 1.4
Baseline0.370.410.350.375
WiCLP0.530.560.450.509
SynGen0.630.570.460.544
VSC0.660.610.470.568
Stable Diffusion 2.1
Baseline0.500.490.420.467
Composable0.400.360.320.357
Structured0.490.490.420.464
Attn-Exct0.640.590.450.547
GORS-unbiased0.640.600.450.550
WiCLP0.650.600.480.567
CONFORM0.680.610.490.5824
InitNO0.690.610.490.5849
SynGen0.710.610.500.594
VSC0.740.640.530.608
Stable Diffusion 3.5
Baseline0.760.670.530.639
Attn-Exct0.790.690.530.652
CONFORM0.800.690.530.654
InitNO0.790.700.540.660
SynGen0.820.740.590.703
VSC0.850.790.630.727
", + "bbox": [ + 91, + 88, + 480, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Texture, and Shape, respectively. The results demonstrate the effectiveness of our method in generating images with accurate color, texture, and shape fidelity across all three versions of Stable Diffusion. Our method can be applied to any diffusion model.", + "bbox": [ + 89, + 564, + 483, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Does VSC's performance drop with an increasing number of binding pairs in prompts? Previous cross-attention control methods [4, 22, 28] often experience a performance drop with an increasing number of binding pairs in prompts. Rassin et al [28] hypothesize that the loss in those cases becomes too large, making the image start to lose its visual quality. Given that the objective of those methods is to separate cross-attention maps, we believe that erroneous maps, as shown in [39], are the main factor causing failure. Thus, with the integration of the localizing cross-attention, we expect the cross-attention maps of VSC to be more precise and less affected by the aforementioned phenomenon. Thus, we evaluate VSC's performance under incrementing numbers of binding pairs. To set up the experiments, we provide all the individual attributes and objects, used in the T2I-CompBench validation set, and randomly combine them into the prompt with the template \"a [attribute] [object], ..., a", + "bbox": [ + 89, + 643, + 486, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bf081ceb519bd17db9dd15f437820e1b1b785510c8907852d8a425fbd5d32539.jpg", + "table_caption": [ + "Table 1. Quantitative results on T2I CompBench. We compare the state-of-the-art methods with ours on 3 categories of the benchmark dataset (color, texture, and shape). Additionally, we provide the harmonic mean. With the same backbone, VSC is consistently better than others. The best is Stable Diffusion 3.5 with VSC." + ], + "table_footnote": [], + "table_body": "
Method345
Stable Diffusion 2.1
SynGen0.67790.48460.1189
VSC0.68810.54790.1901
Stable Diffusion 3.5
SynGen0.80810.56640.1815
VSC0.81320.61830.2461
", + "bbox": [ + 575, + 87, + 846, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Quantitative score with Different Numbers of Binding Pairs. We randomly combine attributes and objects from T2I CompBench into single prompts to analyze compositions with varying complexity. Our method consistently performs well, even when handling a high number of bindings.", + "bbox": [ + 511, + 210, + 906, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[attribute] [object]\" with the desired number of pairs. Finally, we use chatGPT to correct the grammar of the prompt and avoid misleading the index in the fusing computation.", + "bbox": [ + 511, + 282, + 906, + 329 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7d21599f2030941456a189a1184244db3ffbdf03d6cfceb9573192fb138bec04.jpg", + "image_caption": [ + "Figure 3. Generation with the increasing number of binding pairs. We show the generated images when the prompt includes more pairs. We observe that the generated image consistently reflects the additional composition." + ], + "image_footnote": [], + "bbox": [ + 519, + 340, + 684, + 525 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/837bb444382fe90b261411a0b7b487b7e3910e3c3ac10c42647fc0bdb50690c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 340, + 794, + 517 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af5477fb9aa51ca54916f184c0e8f4597dec6a67682220f0aaeb90e4b3bbb7a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 340, + 903, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 presents a BLIP-vqa score comparison between two methods, SynGen, which updates the latent with cross-attention loss, and VSC across different numbers of attribute-object binding pairs (3, 4, and 5). For each binding pair configuration, our method outperforms SynGen. These results indicate that our method consistently performs better in handling complex attribute-object binding scenarios, with increasing accuracy benefits over SynGen as the complexity of the binding pairs increases. Additionally, we provide the qualitative results of when the prompts consisting of 3, 4, and 5 different binding pairs, respectively as shown in Fig. 3. Throughout the 3 images, the visual appearance of objects is consistent, with only one major change in the layout of the \"red book\" to fit into the image, highlighting the model's capability to generate coherent images while scaling the complexity of attribute-object bindings.", + "bbox": [ + 509, + 599, + 908, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Human Eval", + "text_level": 1, + "bbox": [ + 511, + 851, + 648, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Does VSC generate realistic images? We supplement this with human evaluations to assess the overall quality of gen", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19157", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/256947aaf5d6adf6648eabf17f7c57e7b42d2d729d550a725d5b0392f5085cbf.jpg", + "image_caption": [ + "Figure 4. Qualitative result of compositionally. Compared to the baselines, our method can generate the image with better composition. Specifically, for the last row, Stable Diffusion 3.5 and VSC with 2.1 and 3.5 reflect the prompt better than others, such as silver refrigerator and wood chairs. For the second row, VSC with 2.1 and 3.5 generates the mirror; Stable Diffusion 3.5 fails to reflect the mirror." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 901, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erated images, given that the lack of real reference images makes conventional metrics like the Fréchet Inception Distance (FID) score unsuitable for this task. We assess image quality using Amazon Mechanical Turk (AMT). Raters were given a multiple-choice task, which included four images of the same object cropped from generated images of the same prompt by different models reported in Table 3. We crop objects and compare them individually to discard the correlation of correct binding and visual quality. Raters could also select an option indicating that all images are \"equally good\" or \"equally bad\". We collected responses from ten raters and recorded the majority decision. If there was no clear majority, the result was categorized as \"no majority winner\". The values represent the fraction of the majority vote from raters, normalized to sum to 100. Furthermore, Figure.4 shows the qualitative results of compositional generation comparisons.", + "bbox": [ + 89, + 599, + 485, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Among the methods, VSC received the highest majority vote for visual appeal at $30.01\\%$ , indicating a strong prefer", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9dcbdc727876e5a9c50cb5d2471880e98f56717de6ac4b6b5b014c5dbf0ea105.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsImage Quality (%)
Stable Diffusion 3.512.61
A&E14.87
SynGen25.18
VSC30.01
No majority winner17.33
", + "bbox": [ + 562, + 595, + 859, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Human evaluation on image quality. Focusing on the visual quality regardless of binding correctness, we crop out the objects and compare them individually. Values are the fraction of the majority vote of raters, normalized to sum to 100. VSC is voted most, implying it generates visually appealing images.", + "bbox": [ + 511, + 702, + 906, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ence among raters. SynGen followed with $25.18\\%$ , while A&E and Stable Diffusion 3.5 scored lower at $14.87\\%$ and $12.61\\%$ , respectively. Finally, $17.33\\%$ of ratings resulted in no majority winner. The table. 3 demonstrates that VSC is most favored by raters in terms of realistic quality, outperforming the others, and fine-tuning on synthetic data does not lead to the degeneration of visual quality.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19158", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we examine the impact of scaling the synthetic training dataset, assess the transferability of different compositional types, and evaluate our method's inference cost with an increased number of binding pairs.", + "bbox": [ + 89, + 114, + 485, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Scaling Law of Synthetic Data", + "text_level": 1, + "bbox": [ + 89, + 181, + 359, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/710e27704ba89b9b0c460293c883ebf66c418263d215456109362cf2a4a98a85.jpg", + "image_caption": [ + "Figure 5. Scaling dataset: The accuracy of models on T2I-CompBench while training on different dataset sizes (visualized in log-scale), highlighting the positive impact of dataset scaling on model performance." + ], + "image_footnote": [], + "bbox": [ + 94, + 220, + 478, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We trained our model on six different subsets, each containing 5k, 10k, 20k, 30k, 60k, and 90k images, respectively, to analyze the impact of dataset scaling on model performance. As illustrated in Fig. 5, increasing the dataset size leads to a noticeable improvement in accuracy for each attribute, demonstrating that a larger dataset contributes to better model performance. Across all attribute categories—color, texture, and shape—as well as different backbone models, including SD2.1 and SD3.5, the performance steadily improves as the dataset grows from 5,000 to 30,000 images. Beyond this point, further scaling up to 90,000 images provides a marginal enhancement across all categories.", + "bbox": [ + 88, + 508, + 486, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Transferability of Compositionality", + "text_level": 1, + "bbox": [ + 89, + 696, + 398, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the model's ability to generalize compositionality across different groups, we trained it on images from the color group and tested its performance on both the texture and shape groups. Table 4 compares the performance of two models, a model trained on all compositional categories (\"Full-categories training\"), a model trained only on the color category (\"Color-only training\"), and the baseline — across three attributes. In both backbones, the \"Color-only training\" model slightly outperforms the \"Full-categories training\" in Color and importantly shows marginal advances in Texture and Shape compared to the baseline. This outcome suggests that as long as the pre-trained diffusion model", + "bbox": [ + 88, + 719, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e208b9ba460e610a2988b0ec521e0800fffa3c16cee778539e47b581e4c27bef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelColorTextureShape
Stable Diffusion 2.1
Baseline0.500.490.42
Full-categories training0.740.640.53
Color-only training0.770.600.46
Stable Diffusion 3.5
Baseline0.760.670.53
Full-categories training0.850.790.63
Color-only training0.890.740.58
", + "bbox": [ + 529, + 87, + 890, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Transferable compositionality: We show that training VSC only on \"Color\" category can improve the performance on the remaining categories. Color-only VSC outperforms the baseline on \"Texture\" and \"Shape\", demonstrating the transferability.", + "bbox": [ + 511, + 239, + 905, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "can exceptionally generate images from a single pair, VSC can control it in multi-pair settings even if the pair is not in the training set.", + "bbox": [ + 511, + 308, + 905, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Inference Time Cost", + "text_level": 1, + "bbox": [ + 513, + 361, + 705, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ed1e75d20a9dc354c0ca7d930ef186d79f305512248301d7ddebcf1c8b96fe1.jpg", + "image_caption": [ + "Figure 6. Inference cost. We show the inference time of VSC, SynGen, A&E, and the baseline according to the number of binding pairs in the prompt. Although VSC is the slowest in 2-pairs generation, VSC efficiently scales to be faster than SynGen and A&E after the number of pairs exceeds 4 and 5, respectively." + ], + "image_footnote": [], + "bbox": [ + 516, + 393, + 898, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VSC efficiently scales its inference time by generating all supporting images in parallel, making it more scalable than SynGen and Attend-and-Excite, and it requires roughly 1.1GB VRAM for each additional reference prompt. As visualized in Figure.6, across different numbers of binding pairs, the inference time of the VSC remains approximately 2.26 times (2.26x) the inference time of the default SD 3.5. In contrast, SynGen starts at 2.44x and Attend-and-Excite at 2.28x of the baseline when the prompt contains only two pairs. However, as the number of pairs increases, the inference cost rises significantly for SynGen and Attend-and-Excite, making VSC the fastest model when handling prompts with five pairs, excluding the baseline.", + "bbox": [ + 509, + 704, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19159", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 89, + 254, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Single binding generation", + "text_level": 1, + "bbox": [ + 89, + 114, + 323, + 131 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b5f6f5028093351c8ec039f65f95c594257abefd56d2efff8b785b6757ffea9e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelColorTextureShape
Stable Diffusion 2.10.94830.93980.8520
", + "bbox": [ + 109, + 148, + 464, + 185 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Single attribute-object prompt generation. We illustrate that the pre-trained Stable Diffusion 2.1 is already superior in generating images from any single pair in the T2I-CompBench.", + "bbox": [ + 89, + 194, + 482, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We extract all attribute-object pairs in T2I-CompBench [15], then for each pair, we sample 10 images and finally compute the BLIP-vqa score. The results in Table 5 demonstrate that SD2.1 already achieves outstanding accuracies on all categories in the single binding setting, discarding the need for data curation to be effectively used as reference images in both training and inference.", + "bbox": [ + 89, + 239, + 483, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Image Encoder", + "text_level": 1, + "bbox": [ + 89, + 356, + 243, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zarei et al [39] fine-tuned only an MLP and got a comparable result, so we ablate the importance of fine-tuning some late layers of the image encoder.", + "bbox": [ + 89, + 377, + 483, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4b66581083079e50d14906da878d98c982a962bcd8c65a3a41772e3187c068e.jpg", + "image_caption": [ + "Figure 7. Localization cross-attention Loss $(L_{Loc})$ . We depict the importance of fine-tuning the last few layers of the image encoder, as it helps the whole pipeline to descent the $L_{Loc}$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 436, + 475, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 7 shows the localization loss over training steps for different model configurations. The plot compares four setups: ViT-L/14 with trainable 2 layers, ViT-H/14 with 3 trainable layers, and 2 frozen configurations of ViT-L/14 and ViT-H/14. Both fine-tuning procedures make the localization loss converge to approximately $-0.04$ , while frozen models maintain high loss levels without any significant decrease after the training steps pass 5000 steps. The need to train the last 2-3 layers of the image encoder may stem from the nature of prototype embeddings. Unlike other subject-driven image generation frameworks, which are optimized to perfectly preserve the unique identity of a specific subject, our", + "bbox": [ + 88, + 719, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "focus is on the binding problem and on maintaining diversity in the generated images. Consequently, the cross-attention operation on prototype embeddings is less accurate compared to using exact-object embeddings.", + "bbox": [ + 511, + 90, + 908, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Cross-attention Maps", + "text_level": 1, + "bbox": [ + 511, + 167, + 718, + 184 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To support our previous claim about the high precision of the cross-attention map in Section. 4.1, we compare the cross-attention maps of the binding tokens obtained from VSC and the pre-trained SD2.1. Given a correct image, we use DDIM", + "bbox": [ + 511, + 193, + 906, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18d0fb84ad718f2da1c2c1815f221863f81da3f0be0c1500161657e291548937.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 271, + 671, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b8e9e5f9d5e1d36822ad65f332f6399428a7744269b0a664877e727b19603ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 268, + 787, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb4efa19b05a68b525d274d96729086bffe3f0fe0803a990456bd1cbd0053985.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 268, + 903, + 315 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a0163dcb04c8c7bf2db432fd6df8b94030ba89949f65318e7229b0adaef9679.jpg", + "image_caption": [ + "blue", + "Figure 8. Visualization of cross-attention maps. In VSC, the maps are more fine-grained, and the pairs' maps are well-aligned." + ], + "image_footnote": [], + "bbox": [ + 681, + 315, + 761, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb3c596c7aab5093341a1f99b525ffa712f0ef627ab0487603b712f0b4d908ce.jpg", + "image_caption": [ + "bench", + "green", + "bowl" + ], + "image_footnote": [], + "bbox": [ + 764, + 316, + 903, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to revert the image back to timestep 20, feed-forward to the UNet backbone, and visualize the cross-attention maps of both models at that timestep. Figure. 8 showcases cross-attention maps for a given prompt. In the SD2.1, attention maps exhibit inaccuracies, with tokens incorrectly focusing on unrelated pixels. In contrast, in VSC, objects and attributes align more precisely with their corresponding pixels even when the objects are overlapping, assisting our claims.", + "bbox": [ + 511, + 429, + 908, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 570, + 633, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work addresses the critical challenge of enhancing attribute-object binding in recent text-to-image generative models, especially diffusion models, by introducing a novel pairwise visual embedding fusion approach. By decomposing complex text prompts into sub-prompts, generating visual prototypes, and incorporating segmentation-based localization training, our proposed method noticeably improves compositional generalization and accuracy. The results demonstrate superior performance on the T2I Comp-Bench benchmark across various attributes, including color, texture, and shape, even under scaling complexity in binding pairs. Additionally, qualitative and quantitative evaluations validate the method's robustness and adaptability compared to the state-of-the-art models. These findings highlight the potential of combining linguistic and visual embeddings in improving generative models and paving the way for future advancements in creating highly compliant compositional images from complex textual prompts. Future work could explore expanding this approach to broader applications, such as complex scene generation or dynamic object interactions.", + "bbox": [ + 509, + 598, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19160", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2022-0-00124, No.RS-2022-II220124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities, No. RS-2024-00457882, National AI Research Lab Project). T.-H. Oh was partially supported by the KAIST Cross-Generation Collaborative Lab Project, Nam Hyeon-Woo was supported by the BK21 FOUR program through the Educational Institute for Intelligent Information Integration, and Huu Dat was supported by VinUni's Student Research Grant Program AY24-25.", + "bbox": [ + 89, + 114, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 310, + 187, + 325 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 2", + "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3", + "[3] Declan Campbell, Sunayana Rane, Tyler Giallanza, Nicolò De Sabbata, Kia Ghods, Amogh Joshi, Alexander Ku, Steven M Frankland, Thomas L Griffiths, Jonathan D Cohen, et al. Understanding the limits of vision language models through the lens of the binding problem. arXiv preprint arXiv:2411.00238, 2024. 2", + "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 5", + "[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 4", + "[6] Noam Chomsky. Aspects of the Theory of Syntax. Number 11. MIT press, 2014. 1", + "[7] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 2", + "[8] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2", + "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion" + ], + "bbox": [ + 99, + 335, + 486, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2, 4", + "[10] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3", + "[12] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3", + "[13] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darryll, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, pages 210-227. Springer, 2020. 2", + "[14] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2", + "[15] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 2, 4, 8", + "[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4", + "[17] Junmo Kang, Leonid Kurlinsky, Hongyin Luo, Zhen Wang, Jacob Hansen, James Glass, David Cox, Rameswar Panda, Rogerio Feris, and Alan Ritter. Self-moe: Towards compositional large language models with self-specialized experts. arXiv preprint arXiv:2406.12034, 2024. 2", + "[18] Yuval Kirstain, Omer Levy, and Adam Polyak. X&fuse: Fusing visual information in text-to-image generation. arXiv preprint arXiv:2303.01000, 2023. 3", + "[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3", + "[20] Martha Lewis, Nihal V Nayak, Peilin Yu, Qinan Yu, Jack Merullo, Stephen H Bach, and Ellie Pavlick. Does clip bind concepts? probing compositionality in large image models. arXiv preprint arXiv:2212.10537, 2022. 2", + "[21] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19161", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 2, 5", + "[23] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 2", + "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3", + "[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. arXiv preprint arXiv:2405.08246, 2024. 2", + "[26] Maya Okawa, Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Compositional abilities emerge multiplicatively: Exploring diffusion models on a synthetic task. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4", + "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3", + "[31] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4208-4217, 2024. 3", + "[32] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 3", + "[33] Michael Toker, Hadas Orgad, Mor Ventura, Dana Arad, and Yonatan Belinkov. Diffusion lens: Interpreting text encoders in text-to-image pipelines. arXiv preprint arXiv:2403.05846, 2024. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2", + "[35] Anne M Treisman and Garry Gelade. A feature-integration theory of attention. Cognitive psychology, 12(1):97-136, 1980. 2", + "[36] Ruichen Wang, Zekang Chen, Chen Chen, Jian Ma, Haonan Lu, and Xiaodong Lin. Compositional text-to-image synthesis with attention map control of diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5544-5552, 2024. 2", + "[37] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 3", + "[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 2, 3, 4", + "[39] Arman Zarei, Keivan Rezaei, Samyadeep Basu, Mehrdad Saberi, Mazda Moayeri, Priyatham Kattakinda, and Soheil Feizi. Understanding and mitigating compositional issues in text-to-image generative models. arXiv preprint arXiv:2406.07844, 2024. 2, 4, 5, 8", + "[40] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2" + ], + "bbox": [ + 516, + 90, + 906, + 556 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19162", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_model.json b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d7a211cba270bd322bb201e65e48dfdf9a76ee4 --- /dev/null +++ b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.834, + 0.152 + ], + "angle": 0, + "content": "VSC: Visual Search Compositional Text-to-Image Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.192, + 0.786, + 0.213 + ], + "angle": 0, + "content": "Do Huu \\(\\mathrm{Dat}^{1*}\\) Nam Hyeon-Woo² Po-Yuan Mao³ Tae-Hyun Oh²,⁴" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.223, + 0.798, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)VinUniversity \\(^{2}\\)Dept. of EE, POSTECH \\(^{3}\\)Academia Sinica \\(^{4}\\)KAIST" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.285, + 0.868, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.412, + 0.908, + 0.456 + ], + "angle": 0, + "content": "Figure 1. Left: Original pipeline and generated image of Diffusion Model. Right: VSC individually \"search\" for visual information of each binding pair by generating images for each separate pair (\"red car\" and \"yellow bicycle\"). Finally, VSC uses them as references to generate the correct compositional image." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.467, + 0.327, + 0.483 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.503, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Text-to-image diffusion models have shown impressive capabilities in generating realistic visuals from natural-language prompts, yet they often struggle with accurately binding attributes to corresponding objects, especially in prompts containing multiple attribute-object pairs. This challenge primarily arises from the limitations of commonly used text encoders, such as CLIP, which can fail to encode complex linguistic relationships and modifiers effectively. Existing approaches have attempted to mitigate these issues through attention map control during inference and the use of layout information or fine-tuning during training, yet they face performance drops with increased prompt complexity. In this work, we introduce a novel compositional generation method that leverages pairwise image embeddings to improve attribute-object binding. Our approach decomposes complex prompts into sub-prompts, generates corresponding images, and computes visual prototypes that fuse with text embeddings to enhance representation. By applying segmentation-based localization training, we address cross-attention misalignment, achieving improved accuracy in binding multiple attributes to objects. Our approaches outperform existing compositional text-to-image diffusion models on the benchmark T2I CompBench, achieving better image quality, evaluated by humans, and emerging robustness under scaling the number" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.469, + 0.718, + 0.485 + ], + "angle": 0, + "content": "of binding pairs in the prompt." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.513, + 0.644, + 0.529 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.538, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Compositionality is a fundamental ability that enables humans to naturally comprehend the world. It allows us to quickly recognize new objects through their components or construct complex sentences by combining words. This skill empowers us to make \"infinite use of finite mean\" [6], endlessly reusing and recombining familiar concepts in diverse ways. In terms of developing compositional AI systems, diffusion models rise as the most feasible component due to their inherent enormous mode coverage and flexible controllability. The diffusion model's possession of compositional generalization has been proved with synthetic data and the dataset CelebA, where the images are easily decomposed into gender and race [26]. Additionally, diffusion models for generating images from text prompts are capable of creating remarkably realistic visuals, allowing users to shape the content of generated images by crafting detailed and complex natural-language prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite these capabilities, these models often fail to fully utilize compositionality and produce inaccurate images when provided with more intricate descriptions. A prevalent issue is improper binding, as illustrated in the left side of Fig. 1, where descriptive words fail to correctly influence the visual characteristics of specific nouns they are intended to modify. For instance, the prompt specifying \"A photo" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.327, + 0.9 + ], + "angle": 0, + "content": "*Work is done as an intern at POSTECH" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19153" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.334 + ], + "angle": 0, + "content": "of a red car and a yellow bicycle” may produce an image with mixed or misplaced attributes. One likely cause for these inconsistencies is the reliance on text encoders like CLIP within diffusion models. Although CLIP demonstrates strong performance in aligning images with textual descriptions, it faces significant challenges in handling complex linguistic structures. Specifically, it struggles with encoding hierarchical relationships and modifiers, which are crucial for understanding nuanced textual inputs. Previous studies [20, 33, 34] have highlighted that CLIP tends to simplify textual inputs by representing them as a “bag-of-concepts,” disregarding the intricate dependencies and relationships between words. These inherent weaknesses pose limitations for diffusion models that rely on CLIP for text encoding, as they hinder the model's capacity to fully capture and respect the precise linguistic relationships specified in prompts." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.337, + 0.487, + 0.61 + ], + "angle": 0, + "content": "Serial search theory [35] in cognitive psychology suggests that when humans encounter objects with multiple features (e.g., color and shape), our brain processes them individually to bind these features into a cohesive perception. The theory has been successfully applied to improve the capability of large vision-language models [3]. Noticeably, CLIP likewise strongly materializes compositionality in the single binding setting [20] and this property is well-preserved in the latent diffusion model (refer to Section 6.1). Providing spatial information such as bounding box, segmentation maps, blobs, [7, 10, 25, 40] can drastically improve the compositional generation by controlling the attention to be bounded in a specific regions while maintaining a reasonable layout. Those methods can be interpreted as disentangling the desired image into object-level generation. This leads to our motivation to generate images from every single attribute-object pair as references for improving the compositional generalization of the text-to-image diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.486, + 0.902 + ], + "angle": 0, + "content": "On the other hand, the diffusion model can merge and edit images, namely subject-driven image generation, by quickly fine-tuning the model [30] that encodes a unique identifier that refers to the subject. Notably, Xiao et al. [38] propose a tuning-free method via training the diffusion model alongside an image encoder and an MLP. Besides, training an MLP has also been shown to be effective in improving the compositionality of the text-to-image model. Therefore, taking the psychological theory as inspiration, with the above previous works in diffusion models, our key idea, illustrated on the right side of Fig. 1, is to develop a framework that generates images for each binding pair individually and ultimately fuses these images into a complete image from a given prompt containing several attribute-object binding pairs. The fusing process requires the use of an external MLP layer, so we create a synthetic dataset and fine-tune the MLP. Moreover, we extract the segmentation maps of the object and deploy cross-attention localization training as in [38] to fix the erroneous cross-attention maps, aiming to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "address the extreme decrease while increasing the number of attribute-object pairs." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.122, + 0.874, + 0.136 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.907, + 0.212 + ], + "angle": 0, + "content": "- We present VSC, a novel compositional text-to-image diffusion generation method based on a tuning-free, subject-driven image generation framework. Remarkably, VSC requires training only an MLP and the final layers of the image encoder." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.907, + 0.272 + ], + "angle": 0, + "content": "- We leverage the remarkable ability of the pre-trained text-to-image diffusion model in handling single attribute binding prompts to create synthetic images as reference images throughout training and inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.332 + ], + "angle": 0, + "content": "- We achieve state-of-the-art performance on the attribute binding benchmark compared to other compositional text-to-image diffusion models, and provide extensive analysis as well as ablation studies." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.655, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.909, + 0.614 + ], + "angle": 0, + "content": "Compositional Text-to-Image Diffusion Model: In recent work, Huang et al. [15] introduce a benchmark to test compositionality in text-to-image models, highlighting the vulnerability of open-source models to simple compositional prompts. They also propose a fine-tuning approach to enhance compositionality in these models. Several existing studies suggest different methods for enhancing compositionality in text-to-image models in both inference time and training time. In terms of inference time, one popular way is controlling the attention maps of the cross-attention layers [1, 22, 28, 36], which would ensure the maps of binding tokens are overlapping. However, because the attention scores are illustrated to be erroneous [39], then scaling the number of pairs leads to a drastic decrease in performance in those works. Furthermore, combining the guidance in a structured approach [9, 21] also mitigates the binding problems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.909, + 0.765 + ], + "angle": 0, + "content": "During training time, other works leverage the layout information (bounding box, segmentation)[10, 25, 40] to correctly represent the spatial information of the image and fine-tune the diffusion model to concentrate on the binding problem. Additionally, [15, 39] demonstrate the efficiency of fine-tuning the UNet or an MLP layer, which transforms the text embeddings. Our work aims to resolve the discrepancy between cross-attention maps and propose a new approach to tackling compositional generalization without the need for human-annotated layouts or direct latent space manipulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Recent studies have explored compositionality in Vision-Language Models (VLMs) through structured scene [13, 14, 23] leverage scene graphs to enhance understanding and generation of complex visual scenes, while [8, 17] introduce compositional training and modular architectures to improve reasoning capabilities. These approaches are closely related to our goal of enabling compositional text-to-image generation, and further connect our method to the broader literature on structured VLM learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.092, + 0.859, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.263, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 2. The training pipeline of VSC: Given the input text prompt \"A blue apple and a green backpack\", VSC first generates images of each binding pair (\"blue apple\" and \"green backpack\"), then uses an image encoder to extract visual prototype features, which is later used to augment the text embeddings via a trainable MLP module. Finally, a frozen pre-trained diffusion model generates the image with the augmented text embeddings. Additionally, we compute the \\( L_{loc} \\) loss between cross-attention maps and segmentation maps to enhance the precision of cross-attention maps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.347, + 0.486, + 0.529 + ], + "angle": 0, + "content": "Subject-driven Image Generation: Text-to-image personalization is a growing application where models are fine-tuned to generate unique renditions of subjects based on a few reference images [11, 19, 30]. Notable works like DreamBooth [30] and Custom Diffusion [19] excel in this area, generating highly personalized outputs by fine-tuning a pre-trained text-to-image diffusion model with a few reference images and a textual prompt in the form of \"a [v] [class name]\" or \"photo of a [v] [class name].\" In this format, [v] serves as the subject's unique identifier—usually a rare token with minimal inherent meaning—while [class name] represents a broad category for the subject." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Recent approaches to subject-driven image generation focus on synthesizing new images of specific subjects using only a few reference photos but often require computationally expensive tuning, such as DreamBooth [30], Textual Inversion [24], and Custom Diffusion [19]. Some methods, like Tuning Encoder [12], aim to reduce fine-tuning steps by creating initial latent codes with pre-trained encoders but still face scalability challenges. Tuning-free alternatives like X&Fuse [18], ELITE [37], and InstantBooth [32] use embedding techniques to map reference images into text conditioning, although these methods primarily support single-subject generation. For multi-subject scenarios, Custom Diffusion [19] employs joint tuning, while others, like SpaText [2] and Collage Diffusion [31], rely on segmentation masks to define object layouts. However, these techniques struggle with subjects of similar types or require user-specified layouts, which limit flexibility. FastComposer [38] overcomes these limitations by implementing cross-attention localization for better subject separation. To the best of our knowledge, we are the first to utilize subject-driven image generation in solving attribute-object binding problems, fully operating on synthetic images as reference images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.662, + 0.363 + ], + "angle": 0, + "content": "3. Our Approach" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.37, + 0.655, + 0.385 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.908, + 0.59 + ], + "angle": 0, + "content": "We build upon the Stable Diffusion (SD) model, a state-of-the-art framework comprising three main components: a variational autoencoder (VAE), a denoiser as U-Net in SD 1.4, SD 2.1 or Transformer in SD 3.5, and a text encoder. In this setup, the VAE encoder \\( E \\) compresses an image \\( x \\) into a smaller latent space representation \\( z \\), which is then perturbed with Gaussian noise \\( \\varepsilon \\) during the forward diffusion process. The denoiser, parameterized by \\( \\theta \\), is trained to denoise the latent representation by predicting and removing the noise. Text prompts condition this denoising process via a cross-attention mechanism, where the text encoder \\( \\psi \\) maps prompts \\( P \\) to embeddings \\( \\psi(P) \\). During training, the model minimizes a noise loss function defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.6, + 0.907, + 0.632 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n o i s e}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), P, \\varepsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\psi (P)\\right) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.908, + 0.692 + ], + "angle": 0, + "content": "where \\( z_{t} \\) is the latent variable at time step \\( t \\). At inference, a noise sample \\( z_{T} \\) is drawn from \\( \\mathcal{N}(0,1) \\) and gradually refined by the denoiser to recover \\( z_{0} \\), after which the VAE decoder \\( D \\) reconstructs the final image \\( \\hat{x} = D(z_0) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.693, + 0.911, + 0.86 + ], + "angle": 0, + "content": "The cross-attention mechanism, in both the U-Net and Transformer backbone, conditions the denoising process on the text prompt. This is achieved by transforming the latent code \\(z \\in \\mathbb{R}^{(h \\times w) \\times f}\\) and text embeddings \\(c \\in \\mathbb{R}^{n \\times d}\\), where the text encoder \\(\\psi\\) converts the text \\(P\\) into a set of \\(d\\)-dimensional embeddings. Both the latent code and the text embeddings are projected into Query, Key, and Value matrices: \\(Q = W^q z\\), \\(K = W^k c\\), and \\(V = W^v c\\), where \\(W^q \\in \\mathbb{R}^{f \\times d'}\\) and \\(W^k\\), \\(W^v \\in \\mathbb{R}^{d \\times d'}\\) are the learned projection matrices, and \\(d'\\) is the shared dimension of Query, Key, and Value. The attention scores are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.87, + 0.908, + 0.906 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d ^ {\\prime}}}\\right) \\in [ 0, 1 ] ^ {(h \\times w) \\times n} \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19155" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.211 + ], + "angle": 0, + "content": "allowing the model to compute a weighted sum over the Value matrix, resulting in the cross-attention output \\( z_{\\mathrm{attn}} = AV \\in \\mathbb{R}^{(h \\times w) \\times d'} \\). This mechanism distributes textual information across the latent spatial dimensions, where \\( A[i,j,k] \\) denotes the information flow from the \\( k \\)-th text token to the latent pixel at position \\( (i,j) \\). This cross-attention structure provides a foundation for interpreting semantic information in the latent code." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.219, + 0.327, + 0.236 + ], + "angle": 0, + "content": "3.2. Visual Embedding Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.487, + 0.424 + ], + "angle": 0, + "content": "Given a prompt \\(P\\) including multiple binding pairs \\([a_1,o_1],[a_2,o_2],\\ldots ,[a_n,o_n]]\\) having the corresponding index \\(\\mathcal{I} = [i_{a1},i_{o1},i_{a2},i_{o2},\\dots ,i_{an},i_{on}]\\), a pretrained image encoder \\(\\phi :\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^D\\) and a lightweight MLP: \\(\\mathbb{R}^{2*D}\\rightarrow \\mathbb{R}^D\\). We first extract the pairs and generate \\(m\\) images from each pair \\([a_n,o_n]\\) using SD, so we have the corresponding set of reference images \\([r_1^1,\\dots r_1^m ],\\dots ,[r_n^1,\\dots ,r_n^m ]]\\). Then, the set of images is fed into the image encoder \\(\\phi\\) to produce the image embeddings and we take the mean of each set to get the prototype representation of each binding pair according to the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.427, + 0.484, + 0.448 + ], + "angle": 0, + "content": "\\[\nv _ {j} ^ {k} = \\phi \\left(r _ {j} ^ {k}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.45, + 0.484, + 0.489 + ], + "angle": 0, + "content": "\\[\ne _ {j} = \\frac {1}{m} \\sum_ {k = 0} ^ {m} v _ {j} ^ {k} \\quad \\text {f o r} \\quad j \\in \\mathcal {I} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.486, + 0.54 + ], + "angle": 0, + "content": "resulting in the visual prototype features \\([e_1, e_2, \\ldots, e_n]\\). Finally, we augment the text embeddings \\(c\\) with those visual prototype features. For each token in \\(c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.545, + 0.484, + 0.585 + ], + "angle": 0, + "content": "\\[\nc _ {i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} c _ {i}, & i \\notin I \\\\ \\operatorname {M L P} ([ c _ {i}, e _ {j} ]), & i = i _ {j} \\in I \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.486, + 0.817 + ], + "angle": 0, + "content": "The cross-attention maps in the original Stable Diffusion are erroneous, leading to incorrect binding (refer to Section 6.3). Therefore, we adopt the localizing cross-attention loss from FastComposer [38] with minor modifications enhancing the alignment between the cross-attention maps of attribute and object. Let \\(\\mathcal{M} = \\{M_1,M_2,\\dots ,M_n\\}\\) be the segmentation masks for the reference pairs existing in the denoising target. Define \\(A_{i_a} = A[:,:,i_a],A_{i_o} = A[:,:,i_o]\\in [0,1]^{h\\times w}\\) as the cross-attention maps of the \\(i\\) -th pairs, comprising one map for the attribute and another map for the object. We then enforce the both 2 maps \\(A_{i_a},A_{i_o}\\) to approximate the segmentation mask \\(M_j\\) for the \\(j\\) -th subject token, i.e., \\(A_{i_a}\\approx A_{i_o}\\approx M_j\\). We use a balanced \\(L_{1}\\) loss that minimizes the difference between the cross-attention map and the segmentation mask:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.822, + 0.484, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {l o c}} = \\frac {1}{n} \\sum_ {j = 1} ^ {n} \\left[ \\left(\\operatorname {m e a n} \\left(A _ {i _ {a}} \\odot \\bar {M} _ {j}\\right) - \\operatorname {m e a n} \\left(A _ {i _ {a}} \\odot M _ {j}\\right)\\right) \\right. \\\\ \\left. + \\left(\\operatorname {m e a n} \\left(A _ {i _ {o}} \\odot \\bar {M} _ {j}\\right) - \\operatorname {m e a n} \\left(A _ {i _ {o}} \\odot M _ {j}\\right)\\right) \\right], \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.908, + 0.122 + ], + "angle": 0, + "content": "where \\(\\mathrm{mean}(A) = \\frac{1}{h\\cdot w}\\sum_{i = 1}^{h}\\sum_{k = 1}^{w}A_{ik}\\). The overall training objective is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.134, + 0.776, + 0.15 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {n o i s e}} + \\lambda L _ {\\text {l o c}}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.161, + 0.682, + 0.176 + ], + "angle": 0, + "content": "3.3. Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.908, + 0.486 + ], + "angle": 0, + "content": "We make use of the T2I-CompBench dataset [15], concentrating on three attribute categories—color, texture, and shape—each containing 700 prompts for training. To generate high-quality images, we use three generative models: Stable Diffusion 3.5 [29], Flux, and SynGen [28] to generate 300 images for each prompt. Then we use Mask2Former [5] for instance segmentation and OpenCLIP [16] to compute the alignment score of the segmentation and the given attribute-object exists in the prompt. Specifically, we run a greedy algorithm to assign the mask with the text embedding that has the highest alignment score. Additionally, our evaluation emphasizes the disentangled BLIP-Visual Question Answering (VQA) score introduced by Huang et al. [15] as a metric for image quality. The VQA score measures how precisely an image captures the compositional elements specified in a prompt and aligns more closely with human judgment compared to metrics like CLIP-Score [27]. Finally, we select 45 images with the best scores for each prompt, resulting in the final dataset consisting of 90,000 images and their precise segmentation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.499, + 0.646, + 0.516 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.524, + 0.695, + 0.538 + ], + "angle": 0, + "content": "4.1. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.908, + 0.636 + ], + "angle": 0, + "content": "How well can VSC compose the object with its corresponding attribute? Huang et al. [15] has proposed the BLIP-vqa score to evaluate the composition capability. We adopt their evaluation method on attributes of color, shape, and texture. Furthermore, we compute the harmonic mean (HM) to unify the scores into one as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.646, + 0.78, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H M} = \\frac {3}{\\frac {1}{C} + \\frac {1}{T} + \\frac {1}{S}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.908, + 0.734 + ], + "angle": 0, + "content": "where \\(C, T,\\) and \\(S\\) represent the individual performance scores for Color, Texture, and Shape, respectively. We evaluate the methods on T2I CompBench dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Table 1 lists the performance of baseline, ours, and previous work, such as Composable [21] and Structured [9]. In Stable Diffusion 1.4, our method achieves the highest scores across all three attributes (Color: 0.66, Texture: 0.61, Shape: 0.47), outperforming the other methods. In Stable Diffusion 2.1, our method also achieves the best performance, with scores of 0.74 for Color, 0.64 for Texture, and 0.53 for Shape, indicating significant improvements over other models like GORS-unbiased [15], WiCLIP [39], and SynGen [28]. Stable Diffusion 3.5 with VSC archives the highest performance, resulting in 0.85, 0.79, and 0.63 for Color," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19156" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.476 + ], + "angle": 0, + "content": "
ModelColorTextureShapeHM
Stable Diffusion 1.4
Baseline0.370.410.350.375
WiCLP0.530.560.450.509
SynGen0.630.570.460.544
VSC0.660.610.470.568
Stable Diffusion 2.1
Baseline0.500.490.420.467
Composable0.400.360.320.357
Structured0.490.490.420.464
Attn-Exct0.640.590.450.547
GORS-unbiased0.640.600.450.550
WiCLP0.650.600.480.567
CONFORM0.680.610.490.5824
InitNO0.690.610.490.5849
SynGen0.710.610.500.594
VSC0.740.640.530.608
Stable Diffusion 3.5
Baseline0.760.670.530.639
Attn-Exct0.790.690.530.652
CONFORM0.800.690.530.654
InitNO0.790.700.540.660
SynGen0.820.740.590.703
VSC0.850.790.630.727
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.556 + ], + "angle": 0, + "content": "Table 1. Quantitative results on T2I CompBench. We compare the state-of-the-art methods with ours on 3 categories of the benchmark dataset (color, texture, and shape). Additionally, we provide the harmonic mean. With the same backbone, VSC is consistently better than others. The best is Stable Diffusion 3.5 with VSC." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.642 + ], + "angle": 0, + "content": "Texture, and Shape, respectively. The results demonstrate the effectiveness of our method in generating images with accurate color, texture, and shape fidelity across all three versions of Stable Diffusion. Our method can be applied to any diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Does VSC's performance drop with an increasing number of binding pairs in prompts? Previous cross-attention control methods [4, 22, 28] often experience a performance drop with an increasing number of binding pairs in prompts. Rassin et al [28] hypothesize that the loss in those cases becomes too large, making the image start to lose its visual quality. Given that the objective of those methods is to separate cross-attention maps, we believe that erroneous maps, as shown in [39], are the main factor causing failure. Thus, with the integration of the localizing cross-attention, we expect the cross-attention maps of VSC to be more precise and less affected by the aforementioned phenomenon. Thus, we evaluate VSC's performance under incrementing numbers of binding pairs. To set up the experiments, we provide all the individual attributes and objects, used in the T2I-CompBench validation set, and randomly combine them into the prompt with the template \"a [attribute] [object], ..., a" + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.088, + 0.848, + 0.201 + ], + "angle": 0, + "content": "
Method345
Stable Diffusion 2.1
SynGen0.67790.48460.1189
VSC0.68810.54790.1901
Stable Diffusion 3.5
SynGen0.80810.56640.1815
VSC0.81320.61830.2461
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.211, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Table 2. Quantitative score with Different Numbers of Binding Pairs. We randomly combine attributes and objects from T2I CompBench into single prompts to analyze compositions with varying complexity. Our method consistently performs well, even when handling a high number of bindings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.283, + 0.907, + 0.33 + ], + "angle": 0, + "content": "[attribute] [object]\" with the desired number of pairs. Finally, we use chatGPT to correct the grammar of the prompt and avoid misleading the index in the fusing computation." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.341, + 0.685, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.342, + 0.795, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.342, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Figure 3. Generation with the increasing number of binding pairs. We show the generated images when the prompt includes more pairs. We observe that the generated image consistently reflects the additional composition." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.601, + 0.909, + 0.844 + ], + "angle": 0, + "content": "Table 2 presents a BLIP-vqa score comparison between two methods, SynGen, which updates the latent with cross-attention loss, and VSC across different numbers of attribute-object binding pairs (3, 4, and 5). For each binding pair configuration, our method outperforms SynGen. These results indicate that our method consistently performs better in handling complex attribute-object binding scenarios, with increasing accuracy benefits over SynGen as the complexity of the binding pairs increases. Additionally, we provide the qualitative results of when the prompts consisting of 3, 4, and 5 different binding pairs, respectively as shown in Fig. 3. Throughout the 3 images, the visual appearance of objects is consistent, with only one major change in the layout of the \"red book\" to fit into the image, highlighting the model's capability to generate coherent images while scaling the complexity of attribute-object bindings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.65, + 0.866 + ], + "angle": 0, + "content": "4.1.1. Human Eval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Does VSC generate realistic images? We supplement this with human evaluations to assess the overall quality of gen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19157" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.903, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.544, + 0.908, + 0.586 + ], + "angle": 0, + "content": "Figure 4. Qualitative result of compositionally. Compared to the baselines, our method can generate the image with better composition. Specifically, for the last row, Stable Diffusion 3.5 and VSC with 2.1 and 3.5 reflect the prompt better than others, such as silver refrigerator and wood chairs. For the second row, VSC with 2.1 and 3.5 generates the mirror; Stable Diffusion 3.5 fails to reflect the mirror." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.486, + 0.844 + ], + "angle": 0, + "content": "erated images, given that the lack of real reference images makes conventional metrics like the Fréchet Inception Distance (FID) score unsuitable for this task. We assess image quality using Amazon Mechanical Turk (AMT). Raters were given a multiple-choice task, which included four images of the same object cropped from generated images of the same prompt by different models reported in Table 3. We crop objects and compare them individually to discard the correlation of correct binding and visual quality. Raters could also select an option indicating that all images are \"equally good\" or \"equally bad\". We collected responses from ten raters and recorded the majority decision. If there was no clear majority, the result was categorized as \"no majority winner\". The values represent the fraction of the majority vote from raters, normalized to sum to 100. Furthermore, Figure.4 shows the qualitative results of compositional generation comparisons." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Among the methods, VSC received the highest majority vote for visual appeal at \\(30.01\\%\\), indicating a strong prefer" + }, + { + "type": "table", + "bbox": [ + 0.563, + 0.596, + 0.86, + 0.692 + ], + "angle": 0, + "content": "
MethodsImage Quality (%)
Stable Diffusion 3.512.61
A&E14.87
SynGen25.18
VSC30.01
No majority winner17.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.703, + 0.908, + 0.773 + ], + "angle": 0, + "content": "Table 3. Human evaluation on image quality. Focusing on the visual quality regardless of binding correctness, we crop out the objects and compare them individually. Values are the fraction of the majority vote of raters, normalized to sum to 100. VSC is voted most, implying it generates visually appealing images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "ence among raters. SynGen followed with \\(25.18\\%\\), while A&E and Stable Diffusion 3.5 scored lower at \\(14.87\\%\\) and \\(12.61\\%\\), respectively. Finally, \\(17.33\\%\\) of ratings resulted in no majority winner. The table. 3 demonstrates that VSC is most favored by raters in terms of realistic quality, outperforming the others, and fine-tuning on synthetic data does not lead to the degeneration of visual quality." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19158" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.188, + 0.108 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.178 + ], + "angle": 0, + "content": "In this section, we examine the impact of scaling the synthetic training dataset, assess the transferability of different compositional types, and evaluate our method's inference cost with an increased number of binding pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.183, + 0.361, + 0.2 + ], + "angle": 0, + "content": "5.1. Scaling Law of Synthetic Data" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.221, + 0.48, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.452, + 0.485, + 0.509 + ], + "angle": 0, + "content": "Figure 5. Scaling dataset: The accuracy of models on T2I-CompBench while training on different dataset sizes (visualized in log-scale), highlighting the positive impact of dataset scaling on model performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.51, + 0.487, + 0.692 + ], + "angle": 0, + "content": "We trained our model on six different subsets, each containing 5k, 10k, 20k, 30k, 60k, and 90k images, respectively, to analyze the impact of dataset scaling on model performance. As illustrated in Fig. 5, increasing the dataset size leads to a noticeable improvement in accuracy for each attribute, demonstrating that a larger dataset contributes to better model performance. Across all attribute categories—color, texture, and shape—as well as different backbone models, including SD2.1 and SD3.5, the performance steadily improves as the dataset grows from 5,000 to 30,000 images. Beyond this point, further scaling up to 90,000 images provides a marginal enhancement across all categories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.4, + 0.715 + ], + "angle": 0, + "content": "5.2. Transferability of Compositionality" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To evaluate the model's ability to generalize compositionality across different groups, we trained it on images from the color group and tested its performance on both the texture and shape groups. Table 4 compares the performance of two models, a model trained on all compositional categories (\"Full-categories training\"), a model trained only on the color category (\"Color-only training\"), and the baseline — across three attributes. In both backbones, the \"Color-only training\" model slightly outperforms the \"Full-categories training\" in Color and importantly shows marginal advances in Texture and Shape compared to the baseline. This outcome suggests that as long as the pre-trained diffusion model" + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.088, + 0.892, + 0.231 + ], + "angle": 0, + "content": "
ModelColorTextureShape
Stable Diffusion 2.1
Baseline0.500.490.42
Full-categories training0.740.640.53
Color-only training0.770.600.46
Stable Diffusion 3.5
Baseline0.760.670.53
Full-categories training0.850.790.63
Color-only training0.890.740.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.241, + 0.906, + 0.298 + ], + "angle": 0, + "content": "Table 4. Transferable compositionality: We show that training VSC only on \"Color\" category can improve the performance on the remaining categories. Color-only VSC outperforms the baseline on \"Texture\" and \"Shape\", demonstrating the transferability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.906, + 0.355 + ], + "angle": 0, + "content": "can exceptionally generate images from a single pair, VSC can control it in multi-pair settings even if the pair is not in the training set." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.707, + 0.377 + ], + "angle": 0, + "content": "5.3. Inference Time Cost" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.394, + 0.9, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.63, + 0.908, + 0.701 + ], + "angle": 0, + "content": "Figure 6. Inference cost. We show the inference time of VSC, SynGen, A&E, and the baseline according to the number of binding pairs in the prompt. Although VSC is the slowest in 2-pairs generation, VSC efficiently scales to be faster than SynGen and A&E after the number of pairs exceeds 4 and 5, respectively." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "VSC efficiently scales its inference time by generating all supporting images in parallel, making it more scalable than SynGen and Attend-and-Excite, and it requires roughly 1.1GB VRAM for each additional reference prompt. As visualized in Figure.6, across different numbers of binding pairs, the inference time of the VSC remains approximately 2.26 times (2.26x) the inference time of the default SD 3.5. In contrast, SynGen starts at 2.44x and Attend-and-Excite at 2.28x of the baseline when the prompt contains only two pairs. However, as the number of pairs increases, the inference cost rises significantly for SynGen and Attend-and-Excite, making VSC the fastest model when handling prompts with five pairs, excluding the baseline." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19159" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.255, + 0.106 + ], + "angle": 0, + "content": "6. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.325, + 0.132 + ], + "angle": 0, + "content": "6.1. Single binding generation" + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.149, + 0.465, + 0.186 + ], + "angle": 0, + "content": "
ModelColorTextureShape
Stable Diffusion 2.10.94830.93980.8520
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.195, + 0.483, + 0.238 + ], + "angle": 0, + "content": "Table 5. Single attribute-object prompt generation. We illustrate that the pre-trained Stable Diffusion 2.1 is already superior in generating images from any single pair in the T2I-CompBench." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.484, + 0.348 + ], + "angle": 0, + "content": "We extract all attribute-object pairs in T2I-CompBench [15], then for each pair, we sample 10 images and finally compute the BLIP-vqa score. The results in Table 5 demonstrate that SD2.1 already achieves outstanding accuracies on all categories in the single binding setting, discarding the need for data curation to be effectively used as reference images in both training and inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.357, + 0.245, + 0.373 + ], + "angle": 0, + "content": "6.2. Image Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Zarei et al [39] fine-tuned only an MLP and got a comparable result, so we ablate the importance of fine-tuning some late layers of the image encoder." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.437, + 0.476, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Figure 7. Localization cross-attention Loss \\((L_{Loc})\\). We depict the importance of fine-tuning the last few layers of the image encoder, as it helps the whole pipeline to descent the \\(L_{Loc}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Fig. 7 shows the localization loss over training steps for different model configurations. The plot compares four setups: ViT-L/14 with trainable 2 layers, ViT-H/14 with 3 trainable layers, and 2 frozen configurations of ViT-L/14 and ViT-H/14. Both fine-tuning procedures make the localization loss converge to approximately \\(-0.04\\), while frozen models maintain high loss levels without any significant decrease after the training steps pass 5000 steps. The need to train the last 2-3 layers of the image encoder may stem from the nature of prototype embeddings. Unlike other subject-driven image generation frameworks, which are optimized to perfectly preserve the unique identity of a specific subject, our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.152 + ], + "angle": 0, + "content": "focus is on the binding problem and on maintaining diversity in the generated images. Consequently, the cross-attention operation on prototype embeddings is less accurate compared to using exact-object embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.169, + 0.719, + 0.185 + ], + "angle": 0, + "content": "6.3. Cross-attention Maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.194, + 0.908, + 0.255 + ], + "angle": 0, + "content": "To support our previous claim about the high precision of the cross-attention map in Section. 4.1, we compare the cross-attention maps of the binding tokens obtained from VSC and the pre-trained SD2.1. Given a correct image, we use DDIM" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.272, + 0.673, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.27, + 0.789, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.27, + 0.905, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.316, + 0.762, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.317, + 0.905, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.36, + 0.726, + 0.372 + ], + "angle": 0, + "content": "blue" + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.36, + 0.785, + 0.372 + ], + "angle": 0, + "content": "bench" + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.36, + 0.842, + 0.372 + ], + "angle": 0, + "content": "green" + }, + { + "type": "image_caption", + "bbox": [ + 0.857, + 0.36, + 0.893, + 0.371 + ], + "angle": 0, + "content": "bowl" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.392, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Figure 8. Visualization of cross-attention maps. In VSC, the maps are more fine-grained, and the pairs' maps are well-aligned." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.909, + 0.552 + ], + "angle": 0, + "content": "to revert the image back to timestep 20, feed-forward to the UNet backbone, and visualize the cross-attention maps of both models at that timestep. Figure. 8 showcases cross-attention maps for a given prompt. In the SD2.1, attention maps exhibit inaccuracies, with tokens incorrectly focusing on unrelated pixels. In contrast, in VSC, objects and attributes align more precisely with their corresponding pixels even when the objects are overlapping, assisting our claims." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.634, + 0.587 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work addresses the critical challenge of enhancing attribute-object binding in recent text-to-image generative models, especially diffusion models, by introducing a novel pairwise visual embedding fusion approach. By decomposing complex text prompts into sub-prompts, generating visual prototypes, and incorporating segmentation-based localization training, our proposed method noticeably improves compositional generalization and accuracy. The results demonstrate superior performance on the T2I Comp-Bench benchmark across various attributes, including color, texture, and shape, even under scaling complexity in binding pairs. Additionally, qualitative and quantitative evaluations validate the method's robustness and adaptability compared to the state-of-the-art models. These findings highlight the potential of combining linguistic and visual embeddings in improving generative models and paving the way for future advancements in creating highly compliant compositional images from complex textual prompts. Future work could explore expanding this approach to broader applications, such as complex scene generation or dynamic object interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19160" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.298 + ], + "angle": 0, + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2022-0-00124, No.RS-2022-II220124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities, No. RS-2024-00457882, National AI Research Lab Project). T.-H. Oh was partially supported by the KAIST Cross-Generation Collaborative Lab Project, Nam Hyeon-Woo was supported by the BK21 FOUR program through the Educational Institute for Intelligent Information Integration, and Huu Dat was supported by VinUni's Student Research Grant Program AY24-25." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.311, + 0.188, + 0.326 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.336, + 0.486, + 0.418 + ], + "angle": 0, + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.421, + 0.485, + 0.504 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.506, + 0.487, + 0.588 + ], + "angle": 0, + "content": "[3] Declan Campbell, Sunayana Rane, Tyler Giallanza, Nicolò De Sabbata, Kia Ghods, Amogh Joshi, Alexander Ku, Steven M Frankland, Thomas L Griffiths, Jonathan D Cohen, et al. Understanding the limits of vision language models through the lens of the binding problem. arXiv preprint arXiv:2411.00238, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.485, + 0.645 + ], + "angle": 0, + "content": "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.485, + 0.688 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[6] Noam Chomsky. Aspects of the Theory of Syntax. Number 11. MIT press, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[7] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[8] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.336, + 0.487, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.908, + 0.192 + ], + "angle": 0, + "content": "[10] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.908, + 0.263 + ], + "angle": 0, + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.907, + 0.323 + ], + "angle": 0, + "content": "[12] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.907, + 0.38 + ], + "angle": 0, + "content": "[13] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darryll, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, pages 210-227. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.907, + 0.452 + ], + "angle": 0, + "content": "[14] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[15] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.908, + 0.597 + ], + "angle": 0, + "content": "[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.908, + 0.669 + ], + "angle": 0, + "content": "[17] Junmo Kang, Leonid Kurlinsky, Hongyin Luo, Zhen Wang, Jacob Hansen, James Glass, David Cox, Rameswar Panda, Rogerio Feris, and Alan Ritter. Self-moe: Towards compositional large language models with self-specialized experts. arXiv preprint arXiv:2406.12034, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.907, + 0.713 + ], + "angle": 0, + "content": "[18] Yuval Kirstain, Omer Levy, and Adam Polyak. X&fuse: Fusing visual information in text-to-image generation. arXiv preprint arXiv:2303.01000, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.908, + 0.784 + ], + "angle": 0, + "content": "[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[20] Martha Lewis, Nihal V Nayak, Peilin Yu, Qinan Yu, Jack Merullo, Stephen H Bach, and Ellie Pavlick. Does clip bind concepts? probing compositionality in large image models. arXiv preprint arXiv:2212.10537, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[21] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19161" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[22] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[23] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. arXiv preprint arXiv:2405.08246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[26] Maya Okawa, Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Compositional abilities emerge multiplicatively: Exploring diffusion models on a synthetic task. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.484, + 0.567 + ], + "angle": 0, + "content": "[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.637 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[31] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4208-4217, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[32] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[33] Michael Toker, Hadas Orgad, Mor Ventura, Dana Arad, and Yonatan Belinkov. Diffusion lens: Interpreting text encoders in text-to-image pipelines. arXiv preprint arXiv:2403.05846, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.163 + ], + "angle": 0, + "content": "[34] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.204 + ], + "angle": 0, + "content": "[35] Anne M Treisman and Garry Gelade. A feature-integration theory of attention. Cognitive psychology, 12(1):97-136, 1980. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[36] Ruichen Wang, Zekang Chen, Chen Chen, Jian Ma, Haonan Lu, and Xiaodong Lin. Compositional text-to-image synthesis with attention map control of diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5544-5552, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.907, + 0.346 + ], + "angle": 0, + "content": "[37] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.907, + 0.415 + ], + "angle": 0, + "content": "[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.907, + 0.487 + ], + "angle": 0, + "content": "[39] Arman Zarei, Keivan Rezaei, Samyadeep Basu, Mehrdad Saberi, Mazda Moayeri, Priyatham Kattakinda, and Soheil Feizi. Understanding and mitigating compositional issues in text-to-image generative models. arXiv preprint arXiv:2406.07844, 2024. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.487, + 0.908, + 0.557 + ], + "angle": 0, + "content": "[40] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19162" + } + ] +] \ No newline at end of file diff --git a/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_origin.pdf b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cff32172a95b323c25b54219883daf91bc1f649a --- /dev/null +++ b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/9f76eb31-5853-4235-b96e-6306079810f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6cab619da629024ee35984a22665c061f4736acbd0f17b7ea68c2fe5ef45fa7 +size 1376448 diff --git a/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/full.md b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7c4931d26e02b2bab61e4b6b1d54af7f8f5015a --- /dev/null +++ b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/full.md @@ -0,0 +1,282 @@ +# VSC: Visual Search Compositional Text-to-Image Diffusion Model + +Do Huu $\mathrm{Dat}^{1*}$ Nam Hyeon-Woo² Po-Yuan Mao³ Tae-Hyun Oh²,⁴ + +$^{1}$ VinUniversity $^{2}$ Dept. of EE, POSTECH $^{3}$ Academia Sinica $^{4}$ KAIST + +![](images/d28fd3a9091d57e8ff97628cb45d00453c997c2451d21b6dc79d704401c02bdc.jpg) +Figure 1. Left: Original pipeline and generated image of Diffusion Model. Right: VSC individually "search" for visual information of each binding pair by generating images for each separate pair ("red car" and "yellow bicycle"). Finally, VSC uses them as references to generate the correct compositional image. + +# Abstract + +Text-to-image diffusion models have shown impressive capabilities in generating realistic visuals from natural-language prompts, yet they often struggle with accurately binding attributes to corresponding objects, especially in prompts containing multiple attribute-object pairs. This challenge primarily arises from the limitations of commonly used text encoders, such as CLIP, which can fail to encode complex linguistic relationships and modifiers effectively. Existing approaches have attempted to mitigate these issues through attention map control during inference and the use of layout information or fine-tuning during training, yet they face performance drops with increased prompt complexity. In this work, we introduce a novel compositional generation method that leverages pairwise image embeddings to improve attribute-object binding. Our approach decomposes complex prompts into sub-prompts, generates corresponding images, and computes visual prototypes that fuse with text embeddings to enhance representation. By applying segmentation-based localization training, we address cross-attention misalignment, achieving improved accuracy in binding multiple attributes to objects. Our approaches outperform existing compositional text-to-image diffusion models on the benchmark T2I CompBench, achieving better image quality, evaluated by humans, and emerging robustness under scaling the number + +of binding pairs in the prompt. + +# 1. Introduction + +Compositionality is a fundamental ability that enables humans to naturally comprehend the world. It allows us to quickly recognize new objects through their components or construct complex sentences by combining words. This skill empowers us to make "infinite use of finite mean" [6], endlessly reusing and recombining familiar concepts in diverse ways. In terms of developing compositional AI systems, diffusion models rise as the most feasible component due to their inherent enormous mode coverage and flexible controllability. The diffusion model's possession of compositional generalization has been proved with synthetic data and the dataset CelebA, where the images are easily decomposed into gender and race [26]. Additionally, diffusion models for generating images from text prompts are capable of creating remarkably realistic visuals, allowing users to shape the content of generated images by crafting detailed and complex natural-language prompts. + +Despite these capabilities, these models often fail to fully utilize compositionality and produce inaccurate images when provided with more intricate descriptions. A prevalent issue is improper binding, as illustrated in the left side of Fig. 1, where descriptive words fail to correctly influence the visual characteristics of specific nouns they are intended to modify. For instance, the prompt specifying "A photo + +of a red car and a yellow bicycle” may produce an image with mixed or misplaced attributes. One likely cause for these inconsistencies is the reliance on text encoders like CLIP within diffusion models. Although CLIP demonstrates strong performance in aligning images with textual descriptions, it faces significant challenges in handling complex linguistic structures. Specifically, it struggles with encoding hierarchical relationships and modifiers, which are crucial for understanding nuanced textual inputs. Previous studies [20, 33, 34] have highlighted that CLIP tends to simplify textual inputs by representing them as a “bag-of-concepts,” disregarding the intricate dependencies and relationships between words. These inherent weaknesses pose limitations for diffusion models that rely on CLIP for text encoding, as they hinder the model's capacity to fully capture and respect the precise linguistic relationships specified in prompts. + +Serial search theory [35] in cognitive psychology suggests that when humans encounter objects with multiple features (e.g., color and shape), our brain processes them individually to bind these features into a cohesive perception. The theory has been successfully applied to improve the capability of large vision-language models [3]. Noticeably, CLIP likewise strongly materializes compositionality in the single binding setting [20] and this property is well-preserved in the latent diffusion model (refer to Section 6.1). Providing spatial information such as bounding box, segmentation maps, blobs, [7, 10, 25, 40] can drastically improve the compositional generation by controlling the attention to be bounded in a specific regions while maintaining a reasonable layout. Those methods can be interpreted as disentangling the desired image into object-level generation. This leads to our motivation to generate images from every single attribute-object pair as references for improving the compositional generalization of the text-to-image diffusion model. + +On the other hand, the diffusion model can merge and edit images, namely subject-driven image generation, by quickly fine-tuning the model [30] that encodes a unique identifier that refers to the subject. Notably, Xiao et al. [38] propose a tuning-free method via training the diffusion model alongside an image encoder and an MLP. Besides, training an MLP has also been shown to be effective in improving the compositionality of the text-to-image model. Therefore, taking the psychological theory as inspiration, with the above previous works in diffusion models, our key idea, illustrated on the right side of Fig. 1, is to develop a framework that generates images for each binding pair individually and ultimately fuses these images into a complete image from a given prompt containing several attribute-object binding pairs. The fusing process requires the use of an external MLP layer, so we create a synthetic dataset and fine-tune the MLP. Moreover, we extract the segmentation maps of the object and deploy cross-attention localization training as in [38] to fix the erroneous cross-attention maps, aiming to + +address the extreme decrease while increasing the number of attribute-object pairs. + +Our main contributions are summarized as follows: + +- We present VSC, a novel compositional text-to-image diffusion generation method based on a tuning-free, subject-driven image generation framework. Remarkably, VSC requires training only an MLP and the final layers of the image encoder. +- We leverage the remarkable ability of the pre-trained text-to-image diffusion model in handling single attribute binding prompts to create synthetic images as reference images throughout training and inference. +- We achieve state-of-the-art performance on the attribute binding benchmark compared to other compositional text-to-image diffusion models, and provide extensive analysis as well as ablation studies. + +# 2. Related Work + +Compositional Text-to-Image Diffusion Model: In recent work, Huang et al. [15] introduce a benchmark to test compositionality in text-to-image models, highlighting the vulnerability of open-source models to simple compositional prompts. They also propose a fine-tuning approach to enhance compositionality in these models. Several existing studies suggest different methods for enhancing compositionality in text-to-image models in both inference time and training time. In terms of inference time, one popular way is controlling the attention maps of the cross-attention layers [1, 22, 28, 36], which would ensure the maps of binding tokens are overlapping. However, because the attention scores are illustrated to be erroneous [39], then scaling the number of pairs leads to a drastic decrease in performance in those works. Furthermore, combining the guidance in a structured approach [9, 21] also mitigates the binding problems. + +During training time, other works leverage the layout information (bounding box, segmentation)[10, 25, 40] to correctly represent the spatial information of the image and fine-tune the diffusion model to concentrate on the binding problem. Additionally, [15, 39] demonstrate the efficiency of fine-tuning the UNet or an MLP layer, which transforms the text embeddings. Our work aims to resolve the discrepancy between cross-attention maps and propose a new approach to tackling compositional generalization without the need for human-annotated layouts or direct latent space manipulation. + +Recent studies have explored compositionality in Vision-Language Models (VLMs) through structured scene [13, 14, 23] leverage scene graphs to enhance understanding and generation of complex visual scenes, while [8, 17] introduce compositional training and modular architectures to improve reasoning capabilities. These approaches are closely related to our goal of enabling compositional text-to-image generation, and further connect our method to the broader literature on structured VLM learning. + +![](images/fb9fec195d78e3edebf15679a3748ed5d97132c76a5d5996a6806b9574fb353a.jpg) +Figure 2. The training pipeline of VSC: Given the input text prompt "A blue apple and a green backpack", VSC first generates images of each binding pair ("blue apple" and "green backpack"), then uses an image encoder to extract visual prototype features, which is later used to augment the text embeddings via a trainable MLP module. Finally, a frozen pre-trained diffusion model generates the image with the augmented text embeddings. Additionally, we compute the $L_{loc}$ loss between cross-attention maps and segmentation maps to enhance the precision of cross-attention maps. + +Subject-driven Image Generation: Text-to-image personalization is a growing application where models are fine-tuned to generate unique renditions of subjects based on a few reference images [11, 19, 30]. Notable works like DreamBooth [30] and Custom Diffusion [19] excel in this area, generating highly personalized outputs by fine-tuning a pre-trained text-to-image diffusion model with a few reference images and a textual prompt in the form of "a [v] [class name]" or "photo of a [v] [class name]." In this format, [v] serves as the subject's unique identifier—usually a rare token with minimal inherent meaning—while [class name] represents a broad category for the subject. + +Recent approaches to subject-driven image generation focus on synthesizing new images of specific subjects using only a few reference photos but often require computationally expensive tuning, such as DreamBooth [30], Textual Inversion [24], and Custom Diffusion [19]. Some methods, like Tuning Encoder [12], aim to reduce fine-tuning steps by creating initial latent codes with pre-trained encoders but still face scalability challenges. Tuning-free alternatives like X&Fuse [18], ELITE [37], and InstantBooth [32] use embedding techniques to map reference images into text conditioning, although these methods primarily support single-subject generation. For multi-subject scenarios, Custom Diffusion [19] employs joint tuning, while others, like SpaText [2] and Collage Diffusion [31], rely on segmentation masks to define object layouts. However, these techniques struggle with subjects of similar types or require user-specified layouts, which limit flexibility. FastComposer [38] overcomes these limitations by implementing cross-attention localization for better subject separation. To the best of our knowledge, we are the first to utilize subject-driven image generation in solving attribute-object binding problems, fully operating on synthetic images as reference images. + +# 3. Our Approach + +# 3.1. Preliminaries + +We build upon the Stable Diffusion (SD) model, a state-of-the-art framework comprising three main components: a variational autoencoder (VAE), a denoiser as U-Net in SD 1.4, SD 2.1 or Transformer in SD 3.5, and a text encoder. In this setup, the VAE encoder $E$ compresses an image $x$ into a smaller latent space representation $z$ , which is then perturbed with Gaussian noise $\varepsilon$ during the forward diffusion process. The denoiser, parameterized by $\theta$ , is trained to denoise the latent representation by predicting and removing the noise. Text prompts condition this denoising process via a cross-attention mechanism, where the text encoder $\psi$ maps prompts $P$ to embeddings $\psi(P)$ . During training, the model minimizes a noise loss function defined as: + +$$ +L _ {\text {n o i s e}} = \mathbb {E} _ {z \sim \mathcal {E} (x), P, \varepsilon \sim \mathcal {N} (0, 1), t} \left[ \| \varepsilon - \varepsilon_ {\theta} \left(z _ {t}, t, \psi (P)\right) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $z_{t}$ is the latent variable at time step $t$ . At inference, a noise sample $z_{T}$ is drawn from $\mathcal{N}(0,1)$ and gradually refined by the denoiser to recover $z_{0}$ , after which the VAE decoder $D$ reconstructs the final image $\hat{x} = D(z_0)$ . + +The cross-attention mechanism, in both the U-Net and Transformer backbone, conditions the denoising process on the text prompt. This is achieved by transforming the latent code $z \in \mathbb{R}^{(h \times w) \times f}$ and text embeddings $c \in \mathbb{R}^{n \times d}$ , where the text encoder $\psi$ converts the text $P$ into a set of $d$ -dimensional embeddings. Both the latent code and the text embeddings are projected into Query, Key, and Value matrices: $Q = W^q z$ , $K = W^k c$ , and $V = W^v c$ , where $W^q \in \mathbb{R}^{f \times d'}$ and $W^k$ , $W^v \in \mathbb{R}^{d \times d'}$ are the learned projection matrices, and $d'$ is the shared dimension of Query, Key, and Value. The attention scores are computed as: + +$$ +A = \operatorname {S o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d ^ {\prime}}}\right) \in [ 0, 1 ] ^ {(h \times w) \times n} \tag {2} +$$ + +allowing the model to compute a weighted sum over the Value matrix, resulting in the cross-attention output $z_{\mathrm{attn}} = AV \in \mathbb{R}^{(h \times w) \times d'}$ . This mechanism distributes textual information across the latent spatial dimensions, where $A[i,j,k]$ denotes the information flow from the $k$ -th text token to the latent pixel at position $(i,j)$ . This cross-attention structure provides a foundation for interpreting semantic information in the latent code. + +# 3.2. Visual Embedding Fusion + +Given a prompt $P$ including multiple binding pairs $[a_1,o_1],[a_2,o_2],\ldots ,[a_n,o_n]]$ having the corresponding index $\mathcal{I} = [i_{a1},i_{o1},i_{a2},i_{o2},\dots ,i_{an},i_{on}]$ , a pretrained image encoder $\phi :\mathbb{R}^{C\times H\times W}\to \mathbb{R}^D$ and a lightweight MLP: $\mathbb{R}^{2*D}\rightarrow \mathbb{R}^D$ . We first extract the pairs and generate $m$ images from each pair $[a_n,o_n]$ using SD, so we have the corresponding set of reference images $[r_1^1,\dots r_1^m ],\dots ,[r_n^1,\dots ,r_n^m ]]$ . Then, the set of images is fed into the image encoder $\phi$ to produce the image embeddings and we take the mean of each set to get the prototype representation of each binding pair according to the following equations: + +$$ +v _ {j} ^ {k} = \phi \left(r _ {j} ^ {k}\right) \tag {3} +$$ + +$$ +e _ {j} = \frac {1}{m} \sum_ {k = 0} ^ {m} v _ {j} ^ {k} \quad \text {f o r} \quad j \in \mathcal {I} \tag {4} +$$ + +resulting in the visual prototype features $[e_1, e_2, \ldots, e_n]$ . Finally, we augment the text embeddings $c$ with those visual prototype features. For each token in $c$ : + +$$ +c _ {i} ^ {\prime} = \left\{ \begin{array}{l l} c _ {i}, & i \notin I \\ \operatorname {M L P} ([ c _ {i}, e _ {j} ]), & i = i _ {j} \in I \end{array} \right. \tag {5} +$$ + +The cross-attention maps in the original Stable Diffusion are erroneous, leading to incorrect binding (refer to Section 6.3). Therefore, we adopt the localizing cross-attention loss from FastComposer [38] with minor modifications enhancing the alignment between the cross-attention maps of attribute and object. Let $\mathcal{M} = \{M_1,M_2,\dots ,M_n\}$ be the segmentation masks for the reference pairs existing in the denoising target. Define $A_{i_a} = A[:,:,i_a],A_{i_o} = A[:,:,i_o]\in [0,1]^{h\times w}$ as the cross-attention maps of the $i$ -th pairs, comprising one map for the attribute and another map for the object. We then enforce the both 2 maps $A_{i_a},A_{i_o}$ to approximate the segmentation mask $M_j$ for the $j$ -th subject token, i.e., $A_{i_a}\approx A_{i_o}\approx M_j$ . We use a balanced $L_{1}$ loss that minimizes the difference between the cross-attention map and the segmentation mask: + +$$ +\begin{array}{l} L _ {\mathrm {l o c}} = \frac {1}{n} \sum_ {j = 1} ^ {n} \left[ \left(\operatorname {m e a n} \left(A _ {i _ {a}} \odot \bar {M} _ {j}\right) - \operatorname {m e a n} \left(A _ {i _ {a}} \odot M _ {j}\right)\right) \right. \\ \left. + \left(\operatorname {m e a n} \left(A _ {i _ {o}} \odot \bar {M} _ {j}\right) - \operatorname {m e a n} \left(A _ {i _ {o}} \odot M _ {j}\right)\right) \right], \tag {6} \\ \end{array} +$$ + +where $\mathrm{mean}(A) = \frac{1}{h\cdot w}\sum_{i = 1}^{h}\sum_{k = 1}^{w}A_{ik}$ . The overall training objective is then defined as: + +$$ +L = L _ {\text {n o i s e}} + \lambda L _ {\text {l o c}}. +$$ + +# 3.3. Dataset Creation + +We make use of the T2I-CompBench dataset [15], concentrating on three attribute categories—color, texture, and shape—each containing 700 prompts for training. To generate high-quality images, we use three generative models: Stable Diffusion 3.5 [29], Flux, and SynGen [28] to generate 300 images for each prompt. Then we use Mask2Former [5] for instance segmentation and OpenCLIP [16] to compute the alignment score of the segmentation and the given attribute-object exists in the prompt. Specifically, we run a greedy algorithm to assign the mask with the text embedding that has the highest alignment score. Additionally, our evaluation emphasizes the disentangled BLIP-Visual Question Answering (VQA) score introduced by Huang et al. [15] as a metric for image quality. The VQA score measures how precisely an image captures the compositional elements specified in a prompt and aligns more closely with human judgment compared to metrics like CLIP-Score [27]. Finally, we select 45 images with the best scores for each prompt, resulting in the final dataset consisting of 90,000 images and their precise segmentation. + +# 4. Experiments + +# 4.1. Evaluation Results + +How well can VSC compose the object with its corresponding attribute? Huang et al. [15] has proposed the BLIP-vqa score to evaluate the composition capability. We adopt their evaluation method on attributes of color, shape, and texture. Furthermore, we compute the harmonic mean (HM) to unify the scores into one as follows: + +$$ +\mathrm {H M} = \frac {3}{\frac {1}{C} + \frac {1}{T} + \frac {1}{S}}, +$$ + +where $C, T,$ and $S$ represent the individual performance scores for Color, Texture, and Shape, respectively. We evaluate the methods on T2I CompBench dataset. + +Table 1 lists the performance of baseline, ours, and previous work, such as Composable [21] and Structured [9]. In Stable Diffusion 1.4, our method achieves the highest scores across all three attributes (Color: 0.66, Texture: 0.61, Shape: 0.47), outperforming the other methods. In Stable Diffusion 2.1, our method also achieves the best performance, with scores of 0.74 for Color, 0.64 for Texture, and 0.53 for Shape, indicating significant improvements over other models like GORS-unbiased [15], WiCLIP [39], and SynGen [28]. Stable Diffusion 3.5 with VSC archives the highest performance, resulting in 0.85, 0.79, and 0.63 for Color, + +
ModelColorTextureShapeHM
Stable Diffusion 1.4
Baseline0.370.410.350.375
WiCLP0.530.560.450.509
SynGen0.630.570.460.544
VSC0.660.610.470.568
Stable Diffusion 2.1
Baseline0.500.490.420.467
Composable0.400.360.320.357
Structured0.490.490.420.464
Attn-Exct0.640.590.450.547
GORS-unbiased0.640.600.450.550
WiCLP0.650.600.480.567
CONFORM0.680.610.490.5824
InitNO0.690.610.490.5849
SynGen0.710.610.500.594
VSC0.740.640.530.608
Stable Diffusion 3.5
Baseline0.760.670.530.639
Attn-Exct0.790.690.530.652
CONFORM0.800.690.530.654
InitNO0.790.700.540.660
SynGen0.820.740.590.703
VSC0.850.790.630.727
+ +Texture, and Shape, respectively. The results demonstrate the effectiveness of our method in generating images with accurate color, texture, and shape fidelity across all three versions of Stable Diffusion. Our method can be applied to any diffusion model. + +Does VSC's performance drop with an increasing number of binding pairs in prompts? Previous cross-attention control methods [4, 22, 28] often experience a performance drop with an increasing number of binding pairs in prompts. Rassin et al [28] hypothesize that the loss in those cases becomes too large, making the image start to lose its visual quality. Given that the objective of those methods is to separate cross-attention maps, we believe that erroneous maps, as shown in [39], are the main factor causing failure. Thus, with the integration of the localizing cross-attention, we expect the cross-attention maps of VSC to be more precise and less affected by the aforementioned phenomenon. Thus, we evaluate VSC's performance under incrementing numbers of binding pairs. To set up the experiments, we provide all the individual attributes and objects, used in the T2I-CompBench validation set, and randomly combine them into the prompt with the template "a [attribute] [object], ..., a + +Table 1. Quantitative results on T2I CompBench. We compare the state-of-the-art methods with ours on 3 categories of the benchmark dataset (color, texture, and shape). Additionally, we provide the harmonic mean. With the same backbone, VSC is consistently better than others. The best is Stable Diffusion 3.5 with VSC. + +
Method345
Stable Diffusion 2.1
SynGen0.67790.48460.1189
VSC0.68810.54790.1901
Stable Diffusion 3.5
SynGen0.80810.56640.1815
VSC0.81320.61830.2461
+ +Table 2. Quantitative score with Different Numbers of Binding Pairs. We randomly combine attributes and objects from T2I CompBench into single prompts to analyze compositions with varying complexity. Our method consistently performs well, even when handling a high number of bindings. + +[attribute] [object]" with the desired number of pairs. Finally, we use chatGPT to correct the grammar of the prompt and avoid misleading the index in the fusing computation. + +![](images/7d21599f2030941456a189a1184244db3ffbdf03d6cfceb9573192fb138bec04.jpg) +Figure 3. Generation with the increasing number of binding pairs. We show the generated images when the prompt includes more pairs. We observe that the generated image consistently reflects the additional composition. + +![](images/837bb444382fe90b261411a0b7b487b7e3910e3c3ac10c42647fc0bdb50690c5.jpg) + +![](images/af5477fb9aa51ca54916f184c0e8f4597dec6a67682220f0aaeb90e4b3bbb7a8.jpg) + +Table 2 presents a BLIP-vqa score comparison between two methods, SynGen, which updates the latent with cross-attention loss, and VSC across different numbers of attribute-object binding pairs (3, 4, and 5). For each binding pair configuration, our method outperforms SynGen. These results indicate that our method consistently performs better in handling complex attribute-object binding scenarios, with increasing accuracy benefits over SynGen as the complexity of the binding pairs increases. Additionally, we provide the qualitative results of when the prompts consisting of 3, 4, and 5 different binding pairs, respectively as shown in Fig. 3. Throughout the 3 images, the visual appearance of objects is consistent, with only one major change in the layout of the "red book" to fit into the image, highlighting the model's capability to generate coherent images while scaling the complexity of attribute-object bindings. + +# 4.1.1. Human Eval + +Does VSC generate realistic images? We supplement this with human evaluations to assess the overall quality of gen + +![](images/256947aaf5d6adf6648eabf17f7c57e7b42d2d729d550a725d5b0392f5085cbf.jpg) +Figure 4. Qualitative result of compositionally. Compared to the baselines, our method can generate the image with better composition. Specifically, for the last row, Stable Diffusion 3.5 and VSC with 2.1 and 3.5 reflect the prompt better than others, such as silver refrigerator and wood chairs. For the second row, VSC with 2.1 and 3.5 generates the mirror; Stable Diffusion 3.5 fails to reflect the mirror. + +erated images, given that the lack of real reference images makes conventional metrics like the Fréchet Inception Distance (FID) score unsuitable for this task. We assess image quality using Amazon Mechanical Turk (AMT). Raters were given a multiple-choice task, which included four images of the same object cropped from generated images of the same prompt by different models reported in Table 3. We crop objects and compare them individually to discard the correlation of correct binding and visual quality. Raters could also select an option indicating that all images are "equally good" or "equally bad". We collected responses from ten raters and recorded the majority decision. If there was no clear majority, the result was categorized as "no majority winner". The values represent the fraction of the majority vote from raters, normalized to sum to 100. Furthermore, Figure.4 shows the qualitative results of compositional generation comparisons. + +Among the methods, VSC received the highest majority vote for visual appeal at $30.01\%$ , indicating a strong prefer + +
MethodsImage Quality (%)
Stable Diffusion 3.512.61
A&E14.87
SynGen25.18
VSC30.01
No majority winner17.33
+ +Table 3. Human evaluation on image quality. Focusing on the visual quality regardless of binding correctness, we crop out the objects and compare them individually. Values are the fraction of the majority vote of raters, normalized to sum to 100. VSC is voted most, implying it generates visually appealing images. + +ence among raters. SynGen followed with $25.18\%$ , while A&E and Stable Diffusion 3.5 scored lower at $14.87\%$ and $12.61\%$ , respectively. Finally, $17.33\%$ of ratings resulted in no majority winner. The table. 3 demonstrates that VSC is most favored by raters in terms of realistic quality, outperforming the others, and fine-tuning on synthetic data does not lead to the degeneration of visual quality. + +# 5. Analysis + +In this section, we examine the impact of scaling the synthetic training dataset, assess the transferability of different compositional types, and evaluate our method's inference cost with an increased number of binding pairs. + +# 5.1. Scaling Law of Synthetic Data + +![](images/710e27704ba89b9b0c460293c883ebf66c418263d215456109362cf2a4a98a85.jpg) +Figure 5. Scaling dataset: The accuracy of models on T2I-CompBench while training on different dataset sizes (visualized in log-scale), highlighting the positive impact of dataset scaling on model performance. + +We trained our model on six different subsets, each containing 5k, 10k, 20k, 30k, 60k, and 90k images, respectively, to analyze the impact of dataset scaling on model performance. As illustrated in Fig. 5, increasing the dataset size leads to a noticeable improvement in accuracy for each attribute, demonstrating that a larger dataset contributes to better model performance. Across all attribute categories—color, texture, and shape—as well as different backbone models, including SD2.1 and SD3.5, the performance steadily improves as the dataset grows from 5,000 to 30,000 images. Beyond this point, further scaling up to 90,000 images provides a marginal enhancement across all categories. + +# 5.2. Transferability of Compositionality + +To evaluate the model's ability to generalize compositionality across different groups, we trained it on images from the color group and tested its performance on both the texture and shape groups. Table 4 compares the performance of two models, a model trained on all compositional categories ("Full-categories training"), a model trained only on the color category ("Color-only training"), and the baseline — across three attributes. In both backbones, the "Color-only training" model slightly outperforms the "Full-categories training" in Color and importantly shows marginal advances in Texture and Shape compared to the baseline. This outcome suggests that as long as the pre-trained diffusion model + +
ModelColorTextureShape
Stable Diffusion 2.1
Baseline0.500.490.42
Full-categories training0.740.640.53
Color-only training0.770.600.46
Stable Diffusion 3.5
Baseline0.760.670.53
Full-categories training0.850.790.63
Color-only training0.890.740.58
+ +Table 4. Transferable compositionality: We show that training VSC only on "Color" category can improve the performance on the remaining categories. Color-only VSC outperforms the baseline on "Texture" and "Shape", demonstrating the transferability. + +can exceptionally generate images from a single pair, VSC can control it in multi-pair settings even if the pair is not in the training set. + +# 5.3. Inference Time Cost + +![](images/3ed1e75d20a9dc354c0ca7d930ef186d79f305512248301d7ddebcf1c8b96fe1.jpg) +Figure 6. Inference cost. We show the inference time of VSC, SynGen, A&E, and the baseline according to the number of binding pairs in the prompt. Although VSC is the slowest in 2-pairs generation, VSC efficiently scales to be faster than SynGen and A&E after the number of pairs exceeds 4 and 5, respectively. + +VSC efficiently scales its inference time by generating all supporting images in parallel, making it more scalable than SynGen and Attend-and-Excite, and it requires roughly 1.1GB VRAM for each additional reference prompt. As visualized in Figure.6, across different numbers of binding pairs, the inference time of the VSC remains approximately 2.26 times (2.26x) the inference time of the default SD 3.5. In contrast, SynGen starts at 2.44x and Attend-and-Excite at 2.28x of the baseline when the prompt contains only two pairs. However, as the number of pairs increases, the inference cost rises significantly for SynGen and Attend-and-Excite, making VSC the fastest model when handling prompts with five pairs, excluding the baseline. + +# 6. Ablation Studies + +# 6.1. Single binding generation + +
ModelColorTextureShape
Stable Diffusion 2.10.94830.93980.8520
+ +Table 5. Single attribute-object prompt generation. We illustrate that the pre-trained Stable Diffusion 2.1 is already superior in generating images from any single pair in the T2I-CompBench. + +We extract all attribute-object pairs in T2I-CompBench [15], then for each pair, we sample 10 images and finally compute the BLIP-vqa score. The results in Table 5 demonstrate that SD2.1 already achieves outstanding accuracies on all categories in the single binding setting, discarding the need for data curation to be effectively used as reference images in both training and inference. + +# 6.2. Image Encoder + +Zarei et al [39] fine-tuned only an MLP and got a comparable result, so we ablate the importance of fine-tuning some late layers of the image encoder. + +![](images/f4b66581083079e50d14906da878d98c982a962bcd8c65a3a41772e3187c068e.jpg) +Figure 7. Localization cross-attention Loss $(L_{Loc})$ . We depict the importance of fine-tuning the last few layers of the image encoder, as it helps the whole pipeline to descent the $L_{Loc}$ . + +Fig. 7 shows the localization loss over training steps for different model configurations. The plot compares four setups: ViT-L/14 with trainable 2 layers, ViT-H/14 with 3 trainable layers, and 2 frozen configurations of ViT-L/14 and ViT-H/14. Both fine-tuning procedures make the localization loss converge to approximately $-0.04$ , while frozen models maintain high loss levels without any significant decrease after the training steps pass 5000 steps. The need to train the last 2-3 layers of the image encoder may stem from the nature of prototype embeddings. Unlike other subject-driven image generation frameworks, which are optimized to perfectly preserve the unique identity of a specific subject, our + +focus is on the binding problem and on maintaining diversity in the generated images. Consequently, the cross-attention operation on prototype embeddings is less accurate compared to using exact-object embeddings. + +# 6.3. Cross-attention Maps + +To support our previous claim about the high precision of the cross-attention map in Section. 4.1, we compare the cross-attention maps of the binding tokens obtained from VSC and the pre-trained SD2.1. Given a correct image, we use DDIM + +![](images/18d0fb84ad718f2da1c2c1815f221863f81da3f0be0c1500161657e291548937.jpg) + +![](images/0b8e9e5f9d5e1d36822ad65f332f6399428a7744269b0a664877e727b19603ca.jpg) + +![](images/cb4efa19b05a68b525d274d96729086bffe3f0fe0803a990456bd1cbd0053985.jpg) + +![](images/3a0163dcb04c8c7bf2db432fd6df8b94030ba89949f65318e7229b0adaef9679.jpg) +blue +Figure 8. Visualization of cross-attention maps. In VSC, the maps are more fine-grained, and the pairs' maps are well-aligned. + +![](images/bb3c596c7aab5093341a1f99b525ffa712f0ef627ab0487603b712f0b4d908ce.jpg) +bench +green +bowl + +to revert the image back to timestep 20, feed-forward to the UNet backbone, and visualize the cross-attention maps of both models at that timestep. Figure. 8 showcases cross-attention maps for a given prompt. In the SD2.1, attention maps exhibit inaccuracies, with tokens incorrectly focusing on unrelated pixels. In contrast, in VSC, objects and attributes align more precisely with their corresponding pixels even when the objects are overlapping, assisting our claims. + +# 7. Conclusion + +This work addresses the critical challenge of enhancing attribute-object binding in recent text-to-image generative models, especially diffusion models, by introducing a novel pairwise visual embedding fusion approach. By decomposing complex text prompts into sub-prompts, generating visual prototypes, and incorporating segmentation-based localization training, our proposed method noticeably improves compositional generalization and accuracy. The results demonstrate superior performance on the T2I Comp-Bench benchmark across various attributes, including color, texture, and shape, even under scaling complexity in binding pairs. Additionally, qualitative and quantitative evaluations validate the method's robustness and adaptability compared to the state-of-the-art models. These findings highlight the potential of combining linguistic and visual embeddings in improving generative models and paving the way for future advancements in creating highly compliant compositional images from complex textual prompts. Future work could explore expanding this approach to broader applications, such as complex scene generation or dynamic object interactions. + +# Acknowledgement + +This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2022-0-00124, No.RS-2022-II220124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities, No. RS-2024-00457882, National AI Research Lab Project). T.-H. Oh was partially supported by the KAIST Cross-Generation Collaborative Lab Project, Nam Hyeon-Woo was supported by the BK21 FOUR program through the Educational Institute for Intelligent Information Integration, and Huu Dat was supported by VinUni's Student Research Grant Program AY24-25. + +# References + +[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 2 +[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3 +[3] Declan Campbell, Sunayana Rane, Tyler Giallanza, Nicolò De Sabbata, Kia Ghods, Amogh Joshi, Alexander Ku, Steven M Frankland, Thomas L Griffiths, Jonathan D Cohen, et al. Understanding the limits of vision language models through the lens of the binding problem. arXiv preprint arXiv:2411.00238, 2024. 2 +[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 5 +[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 4 +[6] Noam Chomsky. Aspects of the Theory of Syntax. Number 11. MIT press, 2014. 1 +[7] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 2 +[8] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2 +[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion + +guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2, 4 +[10] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3 +[12] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3 +[13] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darryll, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, pages 210-227. Springer, 2020. 2 +[14] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2 +[15] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 2, 4, 8 +[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4 +[17] Junmo Kang, Leonid Kurlinsky, Hongyin Luo, Zhen Wang, Jacob Hansen, James Glass, David Cox, Rameswar Panda, Rogerio Feris, and Alan Ritter. Self-moe: Towards compositional large language models with self-specialized experts. arXiv preprint arXiv:2406.12034, 2024. 2 +[18] Yuval Kirstain, Omer Levy, and Adam Polyak. X&fuse: Fusing visual information in text-to-image generation. arXiv preprint arXiv:2303.01000, 2023. 3 +[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3 +[20] Martha Lewis, Nihal V Nayak, Peilin Yu, Qinan Yu, Jack Merullo, Stephen H Bach, and Ellie Pavlick. Does clip bind concepts? probing compositionality in large image models. arXiv preprint arXiv:2212.10537, 2022. 2 +[21] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4 + +[22] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 2, 5 +[23] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 2 +[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3 +[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. arXiv preprint arXiv:2405.08246, 2024. 2 +[26] Maya Okawa, Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Compositional abilities emerge multiplicatively: Exploring diffusion models on a synthetic task. Advances in Neural Information Processing Systems, 36, 2024. 1 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4 +[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3 +[31] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4208-4217, 2024. 3 +[32] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 3 +[33] Michael Toker, Hadas Orgad, Mor Ventura, Dana Arad, and Yonatan Belinkov. Diffusion lens: Interpreting text encoders in text-to-image pipelines. arXiv preprint arXiv:2403.05846, 2024. 2 + +[34] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2 +[35] Anne M Treisman and Garry Gelade. A feature-integration theory of attention. Cognitive psychology, 12(1):97-136, 1980. 2 +[36] Ruichen Wang, Zekang Chen, Chen Chen, Jian Ma, Haonan Lu, and Xiaodong Lin. Compositional text-to-image synthesis with attention map control of diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5544-5552, 2024. 2 +[37] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 3 +[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 2, 3, 4 +[39] Arman Zarei, Keivan Rezaei, Samyadeep Basu, Mehrdad Saberi, Mazda Moayeri, Priyatham Kattakinda, and Soheil Feizi. Understanding and mitigating compositional issues in text-to-image generative models. arXiv preprint arXiv:2406.07844, 2024. 2, 4, 5, 8 +[40] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2 \ No newline at end of file diff --git a/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/images.zip b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1bfcd973265aede906c86e9bf5ca3ee4bf26a50 --- /dev/null +++ b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bbf8094349777ff15e555fd3e4e165fef790467dcde751a1482dcbe3b2de94c +size 636973 diff --git a/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/layout.json b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cca2d78743a6694765b359c3109379d031b10ae0 --- /dev/null +++ b/2025/VSC_ Visual Search Compositional Text-to-Image Diffusion Model/layout.json @@ -0,0 +1,7453 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 120 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 120 + ], + "type": "text", + "content": "VSC: Visual Search Compositional Text-to-Image Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 152, + 481, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 152, + 481, + 168 + ], + "spans": [ + { + "bbox": [ + 129, + 152, + 481, + 168 + ], + "type": "text", + "content": "Do Huu " + }, + { + "bbox": [ + 129, + 152, + 481, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{Dat}^{1*}" + }, + { + "bbox": [ + 129, + 152, + 481, + 168 + ], + "type": "text", + "content": " Nam Hyeon-Woo² Po-Yuan Mao³ Tae-Hyun Oh²,⁴" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "spans": [ + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "text", + "content": "VinUniversity " + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "text", + "content": "Dept. of EE, POSTECH " + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "text", + "content": "Academia Sinica " + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 122, + 176, + 488, + 191 + ], + "type": "text", + "content": "KAIST" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 86, + 225, + 531, + 315 + ], + "blocks": [ + { + "bbox": [ + 86, + 225, + 531, + 315 + ], + "lines": [ + { + "bbox": [ + 86, + 225, + 531, + 315 + ], + "spans": [ + { + "bbox": [ + 86, + 225, + 531, + 315 + ], + "type": "image", + "image_path": "d28fd3a9091d57e8ff97628cb45d00453c997c2451d21b6dc79d704401c02bdc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 326, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 326, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 555, + 361 + ], + "type": "text", + "content": "Figure 1. Left: Original pipeline and generated image of Diffusion Model. Right: VSC individually \"search\" for visual information of each binding pair by generating images for each separate pair (\"red car\" and \"yellow bicycle\"). Finally, VSC uses them as references to generate the correct compositional image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 369, + 200, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 369, + 200, + 382 + ], + "spans": [ + { + "bbox": [ + 151, + 369, + 200, + 382 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 398, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 398, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 398, + 297, + 685 + ], + "type": "text", + "content": "Text-to-image diffusion models have shown impressive capabilities in generating realistic visuals from natural-language prompts, yet they often struggle with accurately binding attributes to corresponding objects, especially in prompts containing multiple attribute-object pairs. This challenge primarily arises from the limitations of commonly used text encoders, such as CLIP, which can fail to encode complex linguistic relationships and modifiers effectively. Existing approaches have attempted to mitigate these issues through attention map control during inference and the use of layout information or fine-tuning during training, yet they face performance drops with increased prompt complexity. In this work, we introduce a novel compositional generation method that leverages pairwise image embeddings to improve attribute-object binding. Our approach decomposes complex prompts into sub-prompts, generates corresponding images, and computes visual prototypes that fuse with text embeddings to enhance representation. By applying segmentation-based localization training, we address cross-attention misalignment, achieving improved accuracy in binding multiple attributes to objects. Our approaches outperform existing compositional text-to-image diffusion models on the benchmark T2I CompBench, achieving better image quality, evaluated by humans, and emerging robustness under scaling the number" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 371, + 439, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 371, + 439, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 371, + 439, + 384 + ], + "type": "text", + "content": "of binding pairs in the prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 406, + 394, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 394, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 394, + 418 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 426, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 556, + 629 + ], + "type": "text", + "content": "Compositionality is a fundamental ability that enables humans to naturally comprehend the world. It allows us to quickly recognize new objects through their components or construct complex sentences by combining words. This skill empowers us to make \"infinite use of finite mean\" [6], endlessly reusing and recombining familiar concepts in diverse ways. In terms of developing compositional AI systems, diffusion models rise as the most feasible component due to their inherent enormous mode coverage and flexible controllability. The diffusion model's possession of compositional generalization has been proved with synthetic data and the dataset CelebA, where the images are easily decomposed into gender and race [26]. Additionally, diffusion models for generating images from text prompts are capable of creating remarkably realistic visuals, allowing users to shape the content of generated images by crafting detailed and complex natural-language prompts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Despite these capabilities, these models often fail to fully utilize compositionality and produce inaccurate images when provided with more intricate descriptions. A prevalent issue is improper binding, as illustrated in the left side of Fig. 1, where descriptive words fail to correctly influence the visual characteristics of specific nouns they are intended to modify. For instance, the prompt specifying \"A photo" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 200, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 200, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 200, + 712 + ], + "type": "text", + "content": "*Work is done as an intern at POSTECH" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19153" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 264 + ], + "type": "text", + "content": "of a red car and a yellow bicycle” may produce an image with mixed or misplaced attributes. One likely cause for these inconsistencies is the reliance on text encoders like CLIP within diffusion models. Although CLIP demonstrates strong performance in aligning images with textual descriptions, it faces significant challenges in handling complex linguistic structures. Specifically, it struggles with encoding hierarchical relationships and modifiers, which are crucial for understanding nuanced textual inputs. Previous studies [20, 33, 34] have highlighted that CLIP tends to simplify textual inputs by representing them as a “bag-of-concepts,” disregarding the intricate dependencies and relationships between words. These inherent weaknesses pose limitations for diffusion models that rely on CLIP for text encoding, as they hinder the model's capacity to fully capture and respect the precise linguistic relationships specified in prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 266, + 298, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 298, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 298, + 483 + ], + "type": "text", + "content": "Serial search theory [35] in cognitive psychology suggests that when humans encounter objects with multiple features (e.g., color and shape), our brain processes them individually to bind these features into a cohesive perception. The theory has been successfully applied to improve the capability of large vision-language models [3]. Noticeably, CLIP likewise strongly materializes compositionality in the single binding setting [20] and this property is well-preserved in the latent diffusion model (refer to Section 6.1). Providing spatial information such as bounding box, segmentation maps, blobs, [7, 10, 25, 40] can drastically improve the compositional generation by controlling the attention to be bounded in a specific regions while maintaining a reasonable layout. Those methods can be interpreted as disentangling the desired image into object-level generation. This leads to our motivation to generate images from every single attribute-object pair as references for improving the compositional generalization of the text-to-image diffusion model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 486, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 297, + 714 + ], + "type": "text", + "content": "On the other hand, the diffusion model can merge and edit images, namely subject-driven image generation, by quickly fine-tuning the model [30] that encodes a unique identifier that refers to the subject. Notably, Xiao et al. [38] propose a tuning-free method via training the diffusion model alongside an image encoder and an MLP. Besides, training an MLP has also been shown to be effective in improving the compositionality of the text-to-image model. Therefore, taking the psychological theory as inspiration, with the above previous works in diffusion models, our key idea, illustrated on the right side of Fig. 1, is to develop a framework that generates images for each binding pair individually and ultimately fuses these images into a complete image from a given prompt containing several attribute-object binding pairs. The fusing process requires the use of an external MLP layer, so we create a synthetic dataset and fine-tune the MLP. Moreover, we extract the segmentation maps of the object and deploy cross-attention localization training as in [38] to fix the erroneous cross-attention maps, aiming to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "address the extreme decrease while increasing the number of attribute-object pairs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 96, + 534, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 96, + 534, + 107 + ], + "spans": [ + { + "bbox": [ + 326, + 96, + 534, + 107 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 108, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 555, + 167 + ], + "type": "text", + "content": "- We present VSC, a novel compositional text-to-image diffusion generation method based on a tuning-free, subject-driven image generation framework. Remarkably, VSC requires training only an MLP and the final layers of the image encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 168, + 555, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 215 + ], + "type": "text", + "content": "- We leverage the remarkable ability of the pre-trained text-to-image diffusion model in handling single attribute binding prompts to create synthetic images as reference images throughout training and inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 262 + ], + "type": "text", + "content": "- We achieve state-of-the-art performance on the attribute binding benchmark compared to other compositional text-to-image diffusion models, and provide extensive analysis as well as ablation studies." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 295, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 556, + 486 + ], + "type": "text", + "content": "Compositional Text-to-Image Diffusion Model: In recent work, Huang et al. [15] introduce a benchmark to test compositionality in text-to-image models, highlighting the vulnerability of open-source models to simple compositional prompts. They also propose a fine-tuning approach to enhance compositionality in these models. Several existing studies suggest different methods for enhancing compositionality in text-to-image models in both inference time and training time. In terms of inference time, one popular way is controlling the attention maps of the cross-attention layers [1, 22, 28, 36], which would ensure the maps of binding tokens are overlapping. However, because the attention scores are illustrated to be erroneous [39], then scaling the number of pairs leads to a drastic decrease in performance in those works. Furthermore, combining the guidance in a structured approach [9, 21] also mitigates the binding problems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 487, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 605 + ], + "type": "text", + "content": "During training time, other works leverage the layout information (bounding box, segmentation)[10, 25, 40] to correctly represent the spatial information of the image and fine-tune the diffusion model to concentrate on the binding problem. Additionally, [15, 39] demonstrate the efficiency of fine-tuning the UNet or an MLP layer, which transforms the text embeddings. Our work aims to resolve the discrepancy between cross-attention maps and propose a new approach to tackling compositional generalization without the need for human-annotated layouts or direct latent space manipulation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "type": "text", + "content": "Recent studies have explored compositionality in Vision-Language Models (VLMs) through structured scene [13, 14, 23] leverage scene graphs to enhance understanding and generation of complex visual scenes, while [8, 17] introduce compositional training and modular architectures to improve reasoning capabilities. These approaches are closely related to our goal of enabling compositional text-to-image generation, and further connect our method to the broader literature on structured VLM learning." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19154" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 72, + 525, + 198 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 525, + 198 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 525, + 198 + ], + "type": "image", + "image_path": "fb9fec195d78e3edebf15679a3748ed5d97132c76a5d5996a6806b9574fb353a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 208, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 208, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. The training pipeline of VSC: Given the input text prompt \"A blue apple and a green backpack\", VSC first generates images of each binding pair (\"blue apple\" and \"green backpack\"), then uses an image encoder to extract visual prototype features, which is later used to augment the text embeddings via a trainable MLP module. Finally, a frozen pre-trained diffusion model generates the image with the augmented text embeddings. Additionally, we compute the " + }, + { + "bbox": [ + 54, + 208, + 555, + 266 + ], + "type": "inline_equation", + "content": "L_{loc}" + }, + { + "bbox": [ + 54, + 208, + 555, + 266 + ], + "type": "text", + "content": " loss between cross-attention maps and segmentation maps to enhance the precision of cross-attention maps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 274, + 297, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 274, + 297, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 297, + 418 + ], + "type": "text", + "content": "Subject-driven Image Generation: Text-to-image personalization is a growing application where models are fine-tuned to generate unique renditions of subjects based on a few reference images [11, 19, 30]. Notable works like DreamBooth [30] and Custom Diffusion [19] excel in this area, generating highly personalized outputs by fine-tuning a pre-trained text-to-image diffusion model with a few reference images and a textual prompt in the form of \"a [v] [class name]\" or \"photo of a [v] [class name].\" In this format, [v] serves as the subject's unique identifier—usually a rare token with minimal inherent meaning—while [class name] represents a broad category for the subject." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 451, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 297, + 714 + ], + "type": "text", + "content": "Recent approaches to subject-driven image generation focus on synthesizing new images of specific subjects using only a few reference photos but often require computationally expensive tuning, such as DreamBooth [30], Textual Inversion [24], and Custom Diffusion [19]. Some methods, like Tuning Encoder [12], aim to reduce fine-tuning steps by creating initial latent codes with pre-trained encoders but still face scalability challenges. Tuning-free alternatives like X&Fuse [18], ELITE [37], and InstantBooth [32] use embedding techniques to map reference images into text conditioning, although these methods primarily support single-subject generation. For multi-subject scenarios, Custom Diffusion [19] employs joint tuning, while others, like SpaText [2] and Collage Diffusion [31], rely on segmentation masks to define object layouts. However, these techniques struggle with subjects of similar types or require user-specified layouts, which limit flexibility. FastComposer [38] overcomes these limitations by implementing cross-attention localization for better subject separation. To the best of our knowledge, we are the first to utilize subject-driven image generation in solving attribute-object binding problems, fully operating on synthetic images as reference images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 273, + 405, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 405, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 405, + 287 + ], + "type": "text", + "content": "3. Our Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 293, + 400, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 400, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 400, + 304 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": "We build upon the Stable Diffusion (SD) model, a state-of-the-art framework comprising three main components: a variational autoencoder (VAE), a denoiser as U-Net in SD 1.4, SD 2.1 or Transformer in SD 3.5, and a text encoder. In this setup, the VAE encoder " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": " compresses an image " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": " into a smaller latent space representation " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": ", which is then perturbed with Gaussian noise " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": " during the forward diffusion process. The denoiser, parameterized by " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": ", is trained to denoise the latent representation by predicting and removing the noise. Text prompts condition this denoising process via a cross-attention mechanism, where the text encoder " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": " maps prompts " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": " to embeddings " + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\psi(P)" + }, + { + "bbox": [ + 313, + 310, + 555, + 467 + ], + "type": "text", + "content": ". During training, the model minimizes a noise loss function defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 475, + 555, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 475, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 320, + 475, + 555, + 500 + ], + "type": "interline_equation", + "content": "L _ {\\text {n o i s e}} = \\mathbb {E} _ {z \\sim \\mathcal {E} (x), P, \\varepsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, t, \\psi (P)\\right) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "190767a9d1d6f0e13584f3c06866f943e466ced04efb3761824c97a881227011.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": " is the latent variable at time step " + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": ". At inference, a noise sample " + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": " is drawn from " + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": " and gradually refined by the denoiser to recover " + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": ", after which the VAE decoder " + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": " reconstructs the final image " + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\hat{x} = D(z_0)" + }, + { + "bbox": [ + 313, + 501, + 555, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": "The cross-attention mechanism, in both the U-Net and Transformer backbone, conditions the denoising process on the text prompt. This is achieved by transforming the latent code " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{(h \\times w) \\times f}" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": " and text embeddings " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": ", where the text encoder " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": " converts the text " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": " into a set of " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": "-dimensional embeddings. Both the latent code and the text embeddings are projected into Query, Key, and Value matrices: " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "Q = W^q z" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "K = W^k c" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "V = W^v c" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "W^q \\in \\mathbb{R}^{f \\times d'}" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "W^k" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "W^v \\in \\mathbb{R}^{d \\times d'}" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": " are the learned projection matrices, and " + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 313, + 548, + 557, + 681 + ], + "type": "text", + "content": " is the shared dimension of Query, Key, and Value. The attention scores are computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 350, + 689, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 689, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 350, + 689, + 555, + 717 + ], + "type": "interline_equation", + "content": "A = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d ^ {\\prime}}}\\right) \\in [ 0, 1 ] ^ {(h \\times w) \\times n} \\tag {2}", + "image_path": "01eac794d16f10f5ef7da2813204c6dc08b55b9a9b98a3df3896174b1caae03a.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19155" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "allowing the model to compute a weighted sum over the Value matrix, resulting in the cross-attention output " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{attn}} = AV \\in \\mathbb{R}^{(h \\times w) \\times d'}" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": ". This mechanism distributes textual information across the latent spatial dimensions, where " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "A[i,j,k]" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " denotes the information flow from the " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "-th text token to the latent pixel at position " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": ". This cross-attention structure provides a foundation for interpreting semantic information in the latent code." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 173, + 200, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 200, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 200, + 186 + ], + "type": "text", + "content": "3.2. Visual Embedding Fusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": "Given a prompt " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": " including multiple binding pairs " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "[a_1,o_1],[a_2,o_2],\\ldots ,[a_n,o_n]]" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": " having the corresponding index " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = [i_{a1},i_{o1},i_{a2},i_{o2},\\dots ,i_{an},i_{on}]" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": ", a pretrained image encoder " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "\\phi :\\mathbb{R}^{C\\times H\\times W}\\to \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": " and a lightweight MLP: " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2*D}\\rightarrow \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": ". We first extract the pairs and generate " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": " images from each pair " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "[a_n,o_n]" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": " using SD, so we have the corresponding set of reference images " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "[r_1^1,\\dots r_1^m ],\\dots ,[r_n^1,\\dots ,r_n^m ]]" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": ". Then, the set of images is fed into the image encoder " + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 191, + 298, + 335 + ], + "type": "text", + "content": " to produce the image embeddings and we take the mean of each set to get the prototype representation of each binding pair according to the following equations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 338, + 296, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 338, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 116, + 338, + 296, + 354 + ], + "type": "interline_equation", + "content": "v _ {j} ^ {k} = \\phi \\left(r _ {j} ^ {k}\\right) \\tag {3}", + "image_path": "d495e7064fe729c30222ff01cd949388ab15b688b3e3fdd8dbdb720dab5a9840.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 356, + 296, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 356, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 296, + 387 + ], + "type": "interline_equation", + "content": "e _ {j} = \\frac {1}{m} \\sum_ {k = 0} ^ {m} v _ {j} ^ {k} \\quad \\text {f o r} \\quad j \\in \\mathcal {I} \\tag {4}", + "image_path": "f5fd452b547a3cf5a6c490d32e47ab4cc64df913a722b513b62449451cfac935.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "text", + "content": "resulting in the visual prototype features " + }, + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "inline_equation", + "content": "[e_1, e_2, \\ldots, e_n]" + }, + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "text", + "content": ". Finally, we augment the text embeddings " + }, + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "text", + "content": " with those visual prototype features. For each token in " + }, + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 392, + 297, + 427 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 431, + 296, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 431, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 102, + 431, + 296, + 463 + ], + "type": "interline_equation", + "content": "c _ {i} ^ {\\prime} = \\left\\{ \\begin{array}{l l} c _ {i}, & i \\notin I \\\\ \\operatorname {M L P} ([ c _ {i}, e _ {j} ]), & i = i _ {j} \\in I \\end{array} \\right. \\tag {5}", + "image_path": "24d1f3ce87ce10b1a242071f70347e87690b6ef903a88c042f950210334f57f4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": "The cross-attention maps in the original Stable Diffusion are erroneous, leading to incorrect binding (refer to Section 6.3). Therefore, we adopt the localizing cross-attention loss from FastComposer [38] with minor modifications enhancing the alignment between the cross-attention maps of attribute and object. Let " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{M_1,M_2,\\dots ,M_n\\}" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": " be the segmentation masks for the reference pairs existing in the denoising target. Define " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "A_{i_a} = A[:,:,i_a],A_{i_o} = A[:,:,i_o]\\in [0,1]^{h\\times w}" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": " as the cross-attention maps of the " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": " -th pairs, comprising one map for the attribute and another map for the object. We then enforce the both 2 maps " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "A_{i_a},A_{i_o}" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": " to approximate the segmentation mask " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "M_j" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": " -th subject token, i.e., " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "A_{i_a}\\approx A_{i_o}\\approx M_j" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": ". We use a balanced " + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 467, + 297, + 647 + ], + "type": "text", + "content": " loss that minimizes the difference between the cross-attention map and the segmentation mask:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 651, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 651, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 651, + 296, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {l o c}} = \\frac {1}{n} \\sum_ {j = 1} ^ {n} \\left[ \\left(\\operatorname {m e a n} \\left(A _ {i _ {a}} \\odot \\bar {M} _ {j}\\right) - \\operatorname {m e a n} \\left(A _ {i _ {a}} \\odot M _ {j}\\right)\\right) \\right. \\\\ \\left. + \\left(\\operatorname {m e a n} \\left(A _ {i _ {o}} \\odot \\bar {M} _ {j}\\right) - \\operatorname {m e a n} \\left(A _ {i _ {o}} \\odot M _ {j}\\right)\\right) \\right], \\tag {6} \\\\ \\end{array}", + "image_path": "8bdf864c2ade3c9b5dc077f38993ee5fa81c75cc7fcc041c8ac01aa3766df9c1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 71, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 71, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathrm{mean}(A) = \\frac{1}{h\\cdot w}\\sum_{i = 1}^{h}\\sum_{k = 1}^{w}A_{ik}" + }, + { + "bbox": [ + 313, + 71, + 555, + 96 + ], + "type": "text", + "content": ". The overall training objective is then defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 393, + 106, + 474, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 106, + 474, + 118 + ], + "spans": [ + { + "bbox": [ + 393, + 106, + 474, + 118 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {n o i s e}} + \\lambda L _ {\\text {l o c}}.", + "image_path": "c9118465d24c880cee7bd58a5499877b98907ea04bd835eafe61c7cb30a0c958.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 127, + 417, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 127, + 417, + 139 + ], + "spans": [ + { + "bbox": [ + 314, + 127, + 417, + 139 + ], + "type": "text", + "content": "3.3. Dataset Creation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 384 + ], + "type": "text", + "content": "We make use of the T2I-CompBench dataset [15], concentrating on three attribute categories—color, texture, and shape—each containing 700 prompts for training. To generate high-quality images, we use three generative models: Stable Diffusion 3.5 [29], Flux, and SynGen [28] to generate 300 images for each prompt. Then we use Mask2Former [5] for instance segmentation and OpenCLIP [16] to compute the alignment score of the segmentation and the given attribute-object exists in the prompt. Specifically, we run a greedy algorithm to assign the mask with the text embedding that has the highest alignment score. Additionally, our evaluation emphasizes the disentangled BLIP-Visual Question Answering (VQA) score introduced by Huang et al. [15] as a metric for image quality. The VQA score measures how precisely an image captures the compositional elements specified in a prompt and aligns more closely with human judgment compared to metrics like CLIP-Score [27]. Finally, we select 45 images with the best scores for each prompt, resulting in the final dataset consisting of 90,000 images and their precise segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 395, + 395, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 395, + 395, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 395, + 395, + 408 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 415, + 425, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 425, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 425, + 426 + ], + "type": "text", + "content": "4.1. Evaluation Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 503 + ], + "type": "text", + "content": "How well can VSC compose the object with its corresponding attribute? Huang et al. [15] has proposed the BLIP-vqa score to evaluate the composition capability. We adopt their evaluation method on attributes of color, shape, and texture. Furthermore, we compute the harmonic mean (HM) to unify the scores into one as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 390, + 511, + 477, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 511, + 477, + 539 + ], + "spans": [ + { + "bbox": [ + 390, + 511, + 477, + 539 + ], + "type": "interline_equation", + "content": "\\mathrm {H M} = \\frac {3}{\\frac {1}{C} + \\frac {1}{T} + \\frac {1}{S}},", + "image_path": "98efaf6d9a61e23433e3961d4c7d01cf5beaee84fb864af57c5757abe666f63b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 545, + 555, + 581 + ], + "type": "inline_equation", + "content": "C, T," + }, + { + "bbox": [ + 313, + 545, + 555, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 555, + 581 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 545, + 555, + 581 + ], + "type": "text", + "content": " represent the individual performance scores for Color, Texture, and Shape, respectively. We evaluate the methods on T2I CompBench dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Table 1 lists the performance of baseline, ours, and previous work, such as Composable [21] and Structured [9]. In Stable Diffusion 1.4, our method achieves the highest scores across all three attributes (Color: 0.66, Texture: 0.61, Shape: 0.47), outperforming the other methods. In Stable Diffusion 2.1, our method also achieves the best performance, with scores of 0.74 for Color, 0.64 for Texture, and 0.53 for Shape, indicating significant improvements over other models like GORS-unbiased [15], WiCLIP [39], and SynGen [28]. Stable Diffusion 3.5 with VSC archives the highest performance, resulting in 0.85, 0.79, and 0.63 for Color," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19156" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 376 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 376 + ], + "type": "table", + "html": "
ModelColorTextureShapeHM
Stable Diffusion 1.4
Baseline0.370.410.350.375
WiCLP0.530.560.450.509
SynGen0.630.570.460.544
VSC0.660.610.470.568
Stable Diffusion 2.1
Baseline0.500.490.420.467
Composable0.400.360.320.357
Structured0.490.490.420.464
Attn-Exct0.640.590.450.547
GORS-unbiased0.640.600.450.550
WiCLP0.650.600.480.567
CONFORM0.680.610.490.5824
InitNO0.690.610.490.5849
SynGen0.710.610.500.594
VSC0.740.640.530.608
Stable Diffusion 3.5
Baseline0.760.670.530.639
Attn-Exct0.790.690.530.652
CONFORM0.800.690.530.654
InitNO0.790.700.540.660
SynGen0.820.740.590.703
VSC0.850.790.630.727
", + "image_path": "f1f6bbeb8e33c7a0b2f39f5d4b7e86b13d5c2d3e68f440792f3fd4cda6412721.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 447, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 508 + ], + "type": "text", + "content": "Texture, and Shape, respectively. The results demonstrate the effectiveness of our method in generating images with accurate color, texture, and shape fidelity across all three versions of Stable Diffusion. Our method can be applied to any diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "type": "text", + "content": "Does VSC's performance drop with an increasing number of binding pairs in prompts? Previous cross-attention control methods [4, 22, 28] often experience a performance drop with an increasing number of binding pairs in prompts. Rassin et al [28] hypothesize that the loss in those cases becomes too large, making the image start to lose its visual quality. Given that the objective of those methods is to separate cross-attention maps, we believe that erroneous maps, as shown in [39], are the main factor causing failure. Thus, with the integration of the localizing cross-attention, we expect the cross-attention maps of VSC to be more precise and less affected by the aforementioned phenomenon. Thus, we evaluate VSC's performance under incrementing numbers of binding pairs. To set up the experiments, we provide all the individual attributes and objects, used in the T2I-CompBench validation set, and randomly combine them into the prompt with the template \"a [attribute] [object], ..., a" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 352, + 69, + 518, + 159 + ], + "blocks": [ + { + "bbox": [ + 55, + 385, + 296, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 440 + ], + "type": "text", + "content": "Table 1. Quantitative results on T2I CompBench. We compare the state-of-the-art methods with ours on 3 categories of the benchmark dataset (color, texture, and shape). Additionally, we provide the harmonic mean. With the same backbone, VSC is consistently better than others. The best is Stable Diffusion 3.5 with VSC." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 69, + 518, + 159 + ], + "lines": [ + { + "bbox": [ + 352, + 69, + 518, + 159 + ], + "spans": [ + { + "bbox": [ + 352, + 69, + 518, + 159 + ], + "type": "table", + "html": "
Method345
Stable Diffusion 2.1
SynGen0.67790.48460.1189
VSC0.68810.54790.1901
Stable Diffusion 3.5
SynGen0.80810.56640.1815
VSC0.81320.61830.2461
", + "image_path": "bf081ceb519bd17db9dd15f437820e1b1b785510c8907852d8a425fbd5d32539.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 167, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 222 + ], + "type": "text", + "content": "Table 2. Quantitative score with Different Numbers of Binding Pairs. We randomly combine attributes and objects from T2I CompBench into single prompts to analyze compositions with varying complexity. Our method consistently performs well, even when handling a high number of bindings." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 224, + 555, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 261 + ], + "type": "text", + "content": "[attribute] [object]\" with the desired number of pairs. Finally, we use chatGPT to correct the grammar of the prompt and avoid misleading the index in the fusing computation." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 270, + 419, + 416 + ], + "blocks": [ + { + "bbox": [ + 318, + 270, + 419, + 416 + ], + "lines": [ + { + "bbox": [ + 318, + 270, + 419, + 416 + ], + "spans": [ + { + "bbox": [ + 318, + 270, + 419, + 416 + ], + "type": "image", + "image_path": "7d21599f2030941456a189a1184244db3ffbdf03d6cfceb9573192fb138bec04.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 428, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 472 + ], + "type": "text", + "content": "Figure 3. Generation with the increasing number of binding pairs. We show the generated images when the prompt includes more pairs. We observe that the generated image consistently reflects the additional composition." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 421, + 270, + 486, + 410 + ], + "blocks": [ + { + "bbox": [ + 421, + 270, + 486, + 410 + ], + "lines": [ + { + "bbox": [ + 421, + 270, + 486, + 410 + ], + "spans": [ + { + "bbox": [ + 421, + 270, + 486, + 410 + ], + "type": "image", + "image_path": "837bb444382fe90b261411a0b7b487b7e3910e3c3ac10c42647fc0bdb50690c5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 489, + 270, + 553, + 408 + ], + "blocks": [ + { + "bbox": [ + 489, + 270, + 553, + 408 + ], + "lines": [ + { + "bbox": [ + 489, + 270, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 489, + 270, + 553, + 408 + ], + "type": "image", + "image_path": "af5477fb9aa51ca54916f184c0e8f4597dec6a67682220f0aaeb90e4b3bbb7a8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 475, + 556, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 475, + 556, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 475, + 556, + 668 + ], + "type": "text", + "content": "Table 2 presents a BLIP-vqa score comparison between two methods, SynGen, which updates the latent with cross-attention loss, and VSC across different numbers of attribute-object binding pairs (3, 4, and 5). For each binding pair configuration, our method outperforms SynGen. These results indicate that our method consistently performs better in handling complex attribute-object binding scenarios, with increasing accuracy benefits over SynGen as the complexity of the binding pairs increases. Additionally, we provide the qualitative results of when the prompts consisting of 3, 4, and 5 different binding pairs, respectively as shown in Fig. 3. Throughout the 3 images, the visual appearance of objects is consistent, with only one major change in the layout of the \"red book\" to fit into the image, highlighting the model's capability to generate coherent images while scaling the complexity of attribute-object bindings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 674, + 397, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 397, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 397, + 685 + ], + "type": "text", + "content": "4.1.1. Human Eval" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Does VSC generate realistic images? We supplement this with human evaluations to assess the overall quality of gen" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19157" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 552, + 423 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 552, + 423 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 552, + 423 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 552, + 423 + ], + "type": "image", + "image_path": "256947aaf5d6adf6648eabf17f7c57e7b42d2d729d550a725d5b0392f5085cbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 430, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 555, + 464 + ], + "type": "text", + "content": "Figure 4. Qualitative result of compositionally. Compared to the baselines, our method can generate the image with better composition. Specifically, for the last row, Stable Diffusion 3.5 and VSC with 2.1 and 3.5 reflect the prompt better than others, such as silver refrigerator and wood chairs. For the second row, VSC with 2.1 and 3.5 generates the mirror; Stable Diffusion 3.5 fails to reflect the mirror." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 475, + 297, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 297, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 297, + 668 + ], + "type": "text", + "content": "erated images, given that the lack of real reference images makes conventional metrics like the Fréchet Inception Distance (FID) score unsuitable for this task. We assess image quality using Amazon Mechanical Turk (AMT). Raters were given a multiple-choice task, which included four images of the same object cropped from generated images of the same prompt by different models reported in Table 3. We crop objects and compare them individually to discard the correlation of correct binding and visual quality. Raters could also select an option indicating that all images are \"equally good\" or \"equally bad\". We collected responses from ten raters and recorded the majority decision. If there was no clear majority, the result was categorized as \"no majority winner\". The values represent the fraction of the majority vote from raters, normalized to sum to 100. Furthermore, Figure.4 shows the qualitative results of compositional generation comparisons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Among the methods, VSC received the highest majority vote for visual appeal at " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "30.01\\%" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ", indicating a strong prefer" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 344, + 472, + 526, + 548 + ], + "blocks": [ + { + "bbox": [ + 344, + 472, + 526, + 548 + ], + "lines": [ + { + "bbox": [ + 344, + 472, + 526, + 548 + ], + "spans": [ + { + "bbox": [ + 344, + 472, + 526, + 548 + ], + "type": "table", + "html": "
MethodsImage Quality (%)
Stable Diffusion 3.512.61
A&E14.87
SynGen25.18
VSC30.01
No majority winner17.33
", + "image_path": "9dcbdc727876e5a9c50cb5d2471880e98f56717de6ac4b6b5b014c5dbf0ea105.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 556, + 555, + 612 + ], + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 612 + ], + "type": "text", + "content": "Table 3. Human evaluation on image quality. Focusing on the visual quality regardless of binding correctness, we crop out the objects and compare them individually. Values are the fraction of the majority vote of raters, normalized to sum to 100. VSC is voted most, implying it generates visually appealing images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "ence among raters. SynGen followed with " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "25.18\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", while A&E and Stable Diffusion 3.5 scored lower at " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "14.87\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "12.61\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", respectively. Finally, " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "17.33\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " of ratings resulted in no majority winner. The table. 3 demonstrates that VSC is most favored by raters in terms of realistic quality, outperforming the others, and fine-tuning on synthetic data does not lead to the degeneration of visual quality." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19158" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 85 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "content": "In this section, we examine the impact of scaling the synthetic training dataset, assess the transferability of different compositional types, and evaluate our method's inference cost with an increased number of binding pairs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 220, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 220, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 220, + 158 + ], + "type": "text", + "content": "5.1. Scaling Law of Synthetic Data" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 175, + 293, + 347 + ], + "blocks": [ + { + "bbox": [ + 58, + 175, + 293, + 347 + ], + "lines": [ + { + "bbox": [ + 58, + 175, + 293, + 347 + ], + "spans": [ + { + "bbox": [ + 58, + 175, + 293, + 347 + ], + "type": "image", + "image_path": "710e27704ba89b9b0c460293c883ebf66c418263d215456109362cf2a4a98a85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 296, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 403 + ], + "type": "text", + "content": "Figure 5. Scaling dataset: The accuracy of models on T2I-CompBench while training on different dataset sizes (visualized in log-scale), highlighting the positive impact of dataset scaling on model performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 403, + 298, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 298, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 298, + 548 + ], + "type": "text", + "content": "We trained our model on six different subsets, each containing 5k, 10k, 20k, 30k, 60k, and 90k images, respectively, to analyze the impact of dataset scaling on model performance. As illustrated in Fig. 5, increasing the dataset size leads to a noticeable improvement in accuracy for each attribute, demonstrating that a larger dataset contributes to better model performance. Across all attribute categories—color, texture, and shape—as well as different backbone models, including SD2.1 and SD3.5, the performance steadily improves as the dataset grows from 5,000 to 30,000 images. Beyond this point, further scaling up to 90,000 images provides a marginal enhancement across all categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 244, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 244, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 244, + 566 + ], + "type": "text", + "content": "5.2. Transferability of Compositionality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": "To evaluate the model's ability to generalize compositionality across different groups, we trained it on images from the color group and tested its performance on both the texture and shape groups. Table 4 compares the performance of two models, a model trained on all compositional categories (\"Full-categories training\"), a model trained only on the color category (\"Color-only training\"), and the baseline — across three attributes. In both backbones, the \"Color-only training\" model slightly outperforms the \"Full-categories training\" in Color and importantly shows marginal advances in Texture and Shape compared to the baseline. This outcome suggests that as long as the pre-trained diffusion model" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 324, + 69, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 324, + 69, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 324, + 69, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 324, + 69, + 545, + 182 + ], + "type": "table", + "html": "
ModelColorTextureShape
Stable Diffusion 2.1
Baseline0.500.490.42
Full-categories training0.740.640.53
Color-only training0.770.600.46
Stable Diffusion 3.5
Baseline0.760.670.53
Full-categories training0.850.790.63
Color-only training0.890.740.58
", + "image_path": "e208b9ba460e610a2988b0ec521e0800fffa3c16cee778539e47b581e4c27bef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 190, + 554, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 554, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 554, + 236 + ], + "type": "text", + "content": "Table 4. Transferable compositionality: We show that training VSC only on \"Color\" category can improve the performance on the remaining categories. Color-only VSC outperforms the baseline on \"Texture\" and \"Shape\", demonstrating the transferability." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 244, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 554, + 281 + ], + "type": "text", + "content": "can exceptionally generate images from a single pair, VSC can control it in multi-pair settings even if the pair is not in the training set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 286, + 432, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 432, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 432, + 298 + ], + "type": "text", + "content": "5.3. Inference Time Cost" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 312, + 550, + 490 + ], + "blocks": [ + { + "bbox": [ + 316, + 312, + 550, + 490 + ], + "lines": [ + { + "bbox": [ + 316, + 312, + 550, + 490 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 550, + 490 + ], + "type": "image", + "image_path": "3ed1e75d20a9dc354c0ca7d930ef186d79f305512248301d7ddebcf1c8b96fe1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 498, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 555 + ], + "type": "text", + "content": "Figure 6. Inference cost. We show the inference time of VSC, SynGen, A&E, and the baseline according to the number of binding pairs in the prompt. Although VSC is the slowest in 2-pairs generation, VSC efficiently scales to be faster than SynGen and A&E after the number of pairs exceeds 4 and 5, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "type": "text", + "content": "VSC efficiently scales its inference time by generating all supporting images in parallel, making it more scalable than SynGen and Attend-and-Excite, and it requires roughly 1.1GB VRAM for each additional reference prompt. As visualized in Figure.6, across different numbers of binding pairs, the inference time of the VSC remains approximately 2.26 times (2.26x) the inference time of the default SD 3.5. In contrast, SynGen starts at 2.44x and Attend-and-Excite at 2.28x of the baseline when the prompt contains only two pairs. However, as the number of pairs increases, the inference cost rises significantly for SynGen and Attend-and-Excite, making VSC the fastest model when handling prompts with five pairs, excluding the baseline." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19159" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 156, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 156, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 156, + 83 + ], + "type": "text", + "content": "6. Ablation Studies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 198, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 198, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 198, + 104 + ], + "type": "text", + "content": "6.1. Single binding generation" + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 118, + 284, + 147 + ], + "blocks": [ + { + "bbox": [ + 67, + 118, + 284, + 147 + ], + "lines": [ + { + "bbox": [ + 67, + 118, + 284, + 147 + ], + "spans": [ + { + "bbox": [ + 67, + 118, + 284, + 147 + ], + "type": "table", + "html": "
ModelColorTextureShape
Stable Diffusion 2.10.94830.93980.8520
", + "image_path": "b5f6f5028093351c8ec039f65f95c594257abefd56d2efff8b785b6757ffea9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 154, + 295, + 188 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 295, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 295, + 188 + ], + "type": "text", + "content": "Table 5. Single attribute-object prompt generation. We illustrate that the pre-trained Stable Diffusion 2.1 is already superior in generating images from any single pair in the T2I-CompBench." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 275 + ], + "type": "text", + "content": "We extract all attribute-object pairs in T2I-CompBench [15], then for each pair, we sample 10 images and finally compute the BLIP-vqa score. The results in Table 5 demonstrate that SD2.1 already achieves outstanding accuracies on all categories in the single binding setting, discarding the need for data curation to be effectively used as reference images in both training and inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 282, + 149, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 149, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 149, + 295 + ], + "type": "text", + "content": "6.2. Image Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 299, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 335 + ], + "type": "text", + "content": "Zarei et al [39] fine-tuned only an MLP and got a comparable result, so we ablate the importance of fine-tuning some late layers of the image encoder." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 346, + 291, + 522 + ], + "blocks": [ + { + "bbox": [ + 58, + 346, + 291, + 522 + ], + "lines": [ + { + "bbox": [ + 58, + 346, + 291, + 522 + ], + "spans": [ + { + "bbox": [ + 58, + 346, + 291, + 522 + ], + "type": "image", + "image_path": "f4b66581083079e50d14906da878d98c982a962bcd8c65a3a41772e3187c068e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "type": "text", + "content": "Figure 7. Localization cross-attention Loss " + }, + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "type": "inline_equation", + "content": "(L_{Loc})" + }, + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "type": "text", + "content": ". We depict the importance of fine-tuning the last few layers of the image encoder, as it helps the whole pipeline to descent the " + }, + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "type": "inline_equation", + "content": "L_{Loc}" + }, + { + "bbox": [ + 55, + 533, + 295, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": "Fig. 7 shows the localization loss over training steps for different model configurations. The plot compares four setups: ViT-L/14 with trainable 2 layers, ViT-H/14 with 3 trainable layers, and 2 frozen configurations of ViT-L/14 and ViT-H/14. Both fine-tuning procedures make the localization loss converge to approximately " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "-0.04" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": ", while frozen models maintain high loss levels without any significant decrease after the training steps pass 5000 steps. The need to train the last 2-3 layers of the image encoder may stem from the nature of prototype embeddings. Unlike other subject-driven image generation frameworks, which are optimized to perfectly preserve the unique identity of a specific subject, our" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 120 + ], + "type": "text", + "content": "focus is on the binding problem and on maintaining diversity in the generated images. Consequently, the cross-attention operation on prototype embeddings is less accurate compared to using exact-object embeddings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 440, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 440, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 440, + 146 + ], + "type": "text", + "content": "6.3. Cross-attention Maps" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 153, + 555, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 201 + ], + "type": "text", + "content": "To support our previous claim about the high precision of the cross-attention map in Section. 4.1, we compare the cross-attention maps of the binding tokens obtained from VSC and the pre-trained SD2.1. Given a correct image, we use DDIM" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 215, + 411, + 283 + ], + "blocks": [ + { + "bbox": [ + 315, + 215, + 411, + 283 + ], + "lines": [ + { + "bbox": [ + 315, + 215, + 411, + 283 + ], + "spans": [ + { + "bbox": [ + 315, + 215, + 411, + 283 + ], + "type": "image", + "image_path": "18d0fb84ad718f2da1c2c1815f221863f81da3f0be0c1500161657e291548937.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 416, + 213, + 482, + 250 + ], + "blocks": [ + { + "bbox": [ + 416, + 213, + 482, + 250 + ], + "lines": [ + { + "bbox": [ + 416, + 213, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 416, + 213, + 482, + 250 + ], + "type": "image", + "image_path": "0b8e9e5f9d5e1d36822ad65f332f6399428a7744269b0a664877e727b19603ca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 485, + 213, + 553, + 250 + ], + "blocks": [ + { + "bbox": [ + 485, + 213, + 553, + 250 + ], + "lines": [ + { + "bbox": [ + 485, + 213, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 485, + 213, + 553, + 250 + ], + "type": "image", + "image_path": "cb4efa19b05a68b525d274d96729086bffe3f0fe0803a990456bd1cbd0053985.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 417, + 250, + 466, + 283 + ], + "blocks": [ + { + "bbox": [ + 417, + 250, + 466, + 283 + ], + "lines": [ + { + "bbox": [ + 417, + 250, + 466, + 283 + ], + "spans": [ + { + "bbox": [ + 417, + 250, + 466, + 283 + ], + "type": "image", + "image_path": "3a0163dcb04c8c7bf2db432fd6df8b94030ba89949f65318e7229b0adaef9679.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 285, + 444, + 294 + ], + "lines": [ + { + "bbox": [ + 425, + 285, + 444, + 294 + ], + "spans": [ + { + "bbox": [ + 425, + 285, + 444, + 294 + ], + "type": "text", + "content": "blue" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "type": "text", + "content": "Figure 8. Visualization of cross-attention maps. In VSC, the maps are more fine-grained, and the pairs' maps are well-aligned." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 468, + 251, + 553, + 283 + ], + "blocks": [ + { + "bbox": [ + 468, + 251, + 553, + 283 + ], + "lines": [ + { + "bbox": [ + 468, + 251, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 468, + 251, + 553, + 283 + ], + "type": "image", + "image_path": "bb3c596c7aab5093341a1f99b525ffa712f0ef627ab0487603b712f0b4d908ce.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 285, + 480, + 294 + ], + "lines": [ + { + "bbox": [ + 453, + 285, + 480, + 294 + ], + "spans": [ + { + "bbox": [ + 453, + 285, + 480, + 294 + ], + "type": "text", + "content": "bench" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 491, + 285, + 515, + 294 + ], + "lines": [ + { + "bbox": [ + 491, + 285, + 515, + 294 + ], + "spans": [ + { + "bbox": [ + 491, + 285, + 515, + 294 + ], + "type": "text", + "content": "green" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 524, + 285, + 546, + 293 + ], + "lines": [ + { + "bbox": [ + 524, + 285, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 524, + 285, + 546, + 293 + ], + "type": "text", + "content": "bowl" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 340, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 556, + 437 + ], + "type": "text", + "content": "to revert the image back to timestep 20, feed-forward to the UNet backbone, and visualize the cross-attention maps of both models at that timestep. Figure. 8 showcases cross-attention maps for a given prompt. In the SD2.1, attention maps exhibit inaccuracies, with tokens incorrectly focusing on unrelated pixels. In contrast, in VSC, objects and attributes align more precisely with their corresponding pixels even when the objects are overlapping, assisting our claims." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 388, + 464 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 556, + 715 + ], + "type": "text", + "content": "This work addresses the critical challenge of enhancing attribute-object binding in recent text-to-image generative models, especially diffusion models, by introducing a novel pairwise visual embedding fusion approach. By decomposing complex text prompts into sub-prompts, generating visual prototypes, and incorporating segmentation-based localization training, our proposed method noticeably improves compositional generalization and accuracy. The results demonstrate superior performance on the T2I Comp-Bench benchmark across various attributes, including color, texture, and shape, even under scaling complexity in binding pairs. Additionally, qualitative and quantitative evaluations validate the method's robustness and adaptability compared to the state-of-the-art models. These findings highlight the potential of combining linguistic and visual embeddings in improving generative models and paving the way for future advancements in creating highly compliant compositional images from complex textual prompts. Future work could explore expanding this approach to broader applications, such as complex scene generation or dynamic object interactions." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19160" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2022-0-00124, No.RS-2022-II220124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities, No. RS-2024-00457882, National AI Research Lab Project). T.-H. Oh was partially supported by the KAIST Cross-Generation Collaborative Lab Project, Nam Hyeon-Woo was supported by the BK21 FOUR program through the Educational Institute for Intelligent Information Integration, and Huu Dat was supported by VinUni's Student Research Grant Program AY24-25." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 266, + 298, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 266, + 297, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 297, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 297, + 331 + ], + "type": "text", + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2283-2293, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 333, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 333, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 333, + 296, + 399 + ], + "type": "text", + "content": "[2] Omri Avrahami, Thomas Hayes, Oran Gafni, Sonal Gupta, Yaniv Taigman, Devi Parikh, Dani Lischinski, Ohad Fried, and Xi Yin. Spatext: Spatio-textual representation for controllable image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18370-18380, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 400, + 298, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 400, + 298, + 465 + ], + "spans": [ + { + "bbox": [ + 61, + 400, + 298, + 465 + ], + "type": "text", + "content": "[3] Declan Campbell, Sunayana Rane, Tyler Giallanza, Nicolò De Sabbata, Kia Ghods, Amogh Joshi, Alexander Ku, Steven M Frankland, Thomas L Griffiths, Jonathan D Cohen, et al. Understanding the limits of vision language models through the lens of the binding problem. arXiv preprint arXiv:2411.00238, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "text", + "content": "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 512, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 544 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 546, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 296, + 567 + ], + "type": "text", + "content": "[6] Noam Chomsky. Aspects of the Theory of Syntax. Number 11. MIT press, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 612 + ], + "type": "text", + "content": "[7] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 679 + ], + "type": "text", + "content": "[8] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 713 + ], + "type": "text", + "content": "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 98, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 555, + 152 + ], + "type": "text", + "content": "[10] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 154, + 555, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 555, + 208 + ], + "type": "text", + "content": "[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 211, + 555, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 555, + 255 + ], + "type": "text", + "content": "[12] Rinon Gal, Moab Arar, Yuval Atzmon, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Encoder-based domain tuning for fast personalization of text-to-image models. ACM Transactions on Graphics (TOG), 42(4):1-13, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 258, + 555, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 555, + 300 + ], + "type": "text", + "content": "[13] Roei Herzig, Amir Bar, Huijuan Xu, Gal Chechik, Trevor Darryll, and Amir Globerson. Learning canonical representations for scene graph to image generation. In European Conference on Computer Vision, pages 210-227. Springer, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 555, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 357 + ], + "type": "text", + "content": "[14] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. Incorporating structured representations into pretrained vision & language models using scene graphs. arXiv preprint arXiv:2305.06343, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 360, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 555, + 414 + ], + "type": "text", + "content": "[15] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 36:78723-78747, 2023. 2, 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 417, + 555, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 555, + 472 + ], + "type": "text", + "content": "[16] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 475, + 555, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 555, + 529 + ], + "type": "text", + "content": "[17] Junmo Kang, Leonid Kurlinsky, Hongyin Luo, Zhen Wang, Jacob Hansen, James Glass, David Cox, Rameswar Panda, Rogerio Feris, and Alan Ritter. Self-moe: Towards compositional large language models with self-specialized experts. arXiv preprint arXiv:2406.12034, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 532, + 555, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 555, + 564 + ], + "type": "text", + "content": "[18] Yuval Kirstain, Omer Levy, and Adam Polyak. X&fuse: Fusing visual information in text-to-image generation. arXiv preprint arXiv:2303.01000, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 567, + 555, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 620 + ], + "type": "text", + "content": "[19] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 623, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 555, + 667 + ], + "type": "text", + "content": "[20] Martha Lewis, Nihal V Nayak, Peilin Yu, Qinan Yu, Jack Merullo, Stephen H Bach, and Ellie Pavlick. Does clip bind concepts? probing compositionality in large image models. arXiv preprint arXiv:2212.10537, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[21] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2, 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19161" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[22] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9005–9014, 2024. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "text", + "content": "[23] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 296, + 239 + ], + "type": "text", + "content": "[24] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 282 + ], + "type": "text", + "content": "[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. arXiv preprint arXiv:2405.08246, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "text", + "content": "[26] Maya Okawa, Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Compositional abilities emerge multiplicatively: Exploring diffusion models on a synthetic task. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 392 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 449 + ], + "type": "text", + "content": "[28] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 504 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 568 + ], + "type": "text", + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 569, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 296, + 613 + ], + "type": "text", + "content": "[31] Vishnu Sarukkai, Linden Li, Arden Ma, Christopher Ré, and Kayvon Fatahalian. Collage diffusion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4208-4217, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 669 + ], + "type": "text", + "content": "[32] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[33] Michael Toker, Hadas Orgad, Mor Ventura, Dana Arad, and Yonatan Belinkov. Diffusion lens: Interpreting text encoders in text-to-image pipelines. arXiv preprint arXiv:2403.05846, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 441 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 129 + ], + "type": "text", + "content": "[34] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 161 + ], + "type": "text", + "content": "[35] Anne M Treisman and Garry Gelade. A feature-integration theory of attention. Cognitive psychology, 12(1):97-136, 1980. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 162, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 555, + 217 + ], + "type": "text", + "content": "[36] Ruichen Wang, Zekang Chen, Chen Chen, Jian Ma, Haonan Lu, and Xiaodong Lin. Compositional text-to-image synthesis with attention map control of diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5544-5552, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 218, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 555, + 274 + ], + "type": "text", + "content": "[37] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15943-15953, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 274, + 555, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 328 + ], + "type": "text", + "content": "[38] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision, pages 1-20, 2024. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 329, + 555, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 555, + 385 + ], + "type": "text", + "content": "[39] Arman Zarei, Keivan Rezaei, Samyadeep Basu, Mehrdad Saberi, Mazda Moayeri, Priyatham Kattakinda, and Soheil Feizi. Understanding and mitigating compositional issues in text-to-image generative models. arXiv preprint arXiv:2406.07844, 2024. 2, 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 385, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 385, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 385, + 555, + 441 + ], + "type": "text", + "content": "[40] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22490-22499, 2023. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19162" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_content_list.json b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04f70b8e8c54815a714f27d69a925f6e00fe4adc --- /dev/null +++ b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_content_list.json @@ -0,0 +1,1786 @@ +[ + { + "type": "text", + "text": "VSP: Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs", + "text_level": 1, + "bbox": [ + 111, + 128, + 885, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiucheng Wu $^{1}$ , Handong Zhao $^{2}$ , Michael Saxon $^{1}$ , Trung Bui $^{2}$ , William Yang Wang $^{1}$ , Yang Zhang $^{3}$ , Shiyu Chang $^{1}$ $^{1}$ UC Santa Barbara, $^{2}$ Adobe Research, $^{3}$ MIT-IBM Watson AI Lab", + "bbox": [ + 235, + 202, + 759, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qiucheng@ucsb.edu", + "bbox": [ + 421, + 258, + 575, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal large language models are an exciting emerging class of language models (LMs) that have merged classic LM capabilities with those of image processing systems. However, how these capabilities integrate is often not intuitive and warrants direct investigation. One understudied capability in MLLMs is visual spatial planning—the ability to comprehend the spatial arrangements of objects and devise action plans to achieve desired outcomes in visual scenes. It is unclear why MLLMs fall short on these tasks generally considered easy for humans, given their successes across other diverse scenarios. To this end, we introduce VSP, a benchmark that 1) evaluates the spatial planning capability in MLLMs in general, and 2) diagnoses this capability via finer-grained sub-tasks, including perception and reasoning, and measure the capabilities of models through these sub-tasks. Our evaluation confirms that both open-source and private MLLMs fail to generate effective plans for even simple spatial planning tasks. Evaluations on the fine-grained analytical tasks further reveal fundamental deficiencies in the models' visual perception and bottle-necks in reasoning abilities, explaining their worse performance in the general spatial planning tasks. Our work illuminates future directions for improving MLLMs' abilities in spatial planning. Our benchmark is publicly available1.", + "bbox": [ + 89, + 340, + 483, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 715, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of large language models has driven considerable growth in their capabilities to produce fluent text in many domains, generating outputs exhibiting potential \"reasoning\" and \"understanding\" abilities [12, 14, 30, 57]. Recently, multimodal large language models (MLLMs) have advanced on LLMs through training on native image inputs, to achieve impressive performance generating text describing and relating to input images [1, 8, 10,", + "bbox": [ + 89, + 741, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f215dad01c20b127c0ef840cea7cb79c943182576656799b9ef8815b0fb2116.jpg", + "image_caption": [ + "Figure 1. Overview of the VSP benchmark. The benchmark evaluates tasks that require MLLMs to perform complex reasoning by utilizing the spatial information they perceive, and diagnoses their failure modes using a series of carefully designed subtasks." + ], + "image_footnote": [], + "bbox": [ + 517, + 308, + 903, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "37, 54], with applications in image captioning, visual question answering, and visual reasoning [48, 64, 68, 72]. The swift evolution of MLLMs has enabled them to tackle increasingly sophisticated tasks that require multiple emerging abilities in complex scenarios. However, as model capabilities and deployment needs advance, the challenges in usefully evaluating them grow in kind.", + "bbox": [ + 511, + 643, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Planning is a fundamental capability in intelligent systems that is particularly contested in LMs [59]. Visual spatial planning refers to the task of comprehending the spatial arrangement of objects in a scene and designing action plans to achieve a desired outcome. For example, the classical maze problem can be considered a visual planning task, where an agent is given an input image describing the maze environment and is asked to produce a viable path to navigate the player from the starting position to the goal. This task requires two capabilities: image perception, which en", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/UCSB-NLP-Chang/Visual-Spatial-Planning", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2270", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/11f1776cd3b00804b50cf1c8807686ccb5c46e89ac089ffe902d477d5f01bf3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkCapabilitiesDiagnosing Incapabilities
Visual RecognitionSpatial Relation UnderstandingPlan with Spatial Info
MMMU [70]✓XXX
MathVision [60]✓XXX
MME [21]✓✓XX
SeedBench [34]✓✓XX
MM-Vet [69]✓✓XX
AlgoPuzzle [23]✓✓✓X
MMSI [71]✓✓✓X
VSP✓✓✓✓
", + "bbox": [ + 91, + 79, + 493, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison with representative existing benchmarks. Note that we focus on spatial reasoning related capabilities. ✓ indicates that the measurement is covered by the corresponding benchmark, while $\\times$ indicates not covered.", + "bbox": [ + 89, + 242, + 483, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ables the agent to understand the objects, environment and spatial relations present in the image, and reasoning, which enables the agent to perform strategic decision-making.", + "bbox": [ + 89, + 309, + 483, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual spatial planning is an important capability in many potential applications for MLLMs, such as navigating in complex environments with autonomous driving [40, 55] or manipulating objects with robotic hands [16, 29]. However, although there have been increasingly more benchmarks to evaluate the vision processing capabilities of MLLMs, few systematically investigate the underlying factors affecting their (in)ability to perform visual spatial planning tasks. As shown in Table 1, while most of the existing benchmarks identify the capability to recognize objects as a key feature of MLLM, quite a few benchmarks [60, 70] do not evaluate their ability to understand spatial relations, and even fewer benchmarks [23, 71] examine the ability to use spatial information in complex tasks, e.g., navigation on a road map based on the position towards goals. More importantly, it is often unclear why MLLMs perform so poorly on the spatial planning tasks that are generally considered easy for humans [23]. As a result, two research questions emerged: ① [Evaluation] How performant are MLLMs in performing visual planning tasks? ② [Diagnose] What are the bottleneck capabilities, e.g., perception or reasoning, that limit the performance of MLLMs in visual planning tasks?", + "bbox": [ + 91, + 357, + 483, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce Visual Spatial Planning (VSP), a benchmark specifically designed to evaluate and diagnose the spatial planning capabilities of MLLMs. As illustrated in Figure 1, the VSP benchmark consists of various environment settings designed to assess the capabilities of models in different scenarios. We focus on abstract scenarios to facilitate the design of diagnostic subtasks with consistent environments (Sec. 3.5). The first two scenarios, Maze Navigation and Blocks World, are basic environments developed from classical planning tasks. Additionally, we challenge the model's abilities in dynamic and realistic applications through the Collision and Google Map scenarios, respectively. All environments are fully observable through input", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images. The MLLMs are required to interpret the visual inputs, deduce the consequences of each action, and execute the designated tasks. To comprehensively evaluate the fine-grained capabilities needed for visual spatial planning, VSP includes 4.6K questions in 12 meticulously designed tasks that feature both simulated and photorealistic visual inputs. In addition to testing end-to-end spatial planning, these tasks test essential individual visual planning abilities, such as image perception and reasoning.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We apply the VSP benchmark to evaluate existing state-of-the-art MLLMs, including both open-source and private ones. Surprisingly, we find that even the most competitive MLLMs sometimes struggle in performing the simplest visual planning tasks, such as a 3x3 maze problem. Our fine-grained capability analysis further reveals that existing MLLMs have flaws in reasoning and bigger bottlenecks in perception. We believe the VSP benchmark highlights critical weaknesses in current MLLMs and sheds light on future directions for enhancing their spatial understanding and planning capabilities.", + "bbox": [ + 511, + 227, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 409, + 653, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. General Planning in LMs", + "text_level": 1, + "bbox": [ + 511, + 434, + 745, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Planning has been a central focus of research in AI. Traditional work in AI planning includes using formal languages to represent and solve planning problems [2], and developing algorithms like dynamic programming and reinforcement learning to explore environments and formulate viable plans [26, 53]. While these works mostly focus on predefined and restricted environments, recently, with the advancement of LMs, it has become intriguing to study whether LMs, with the potential to be general intelligent agents, can perform planning in different settings and environments [32, 33, 52]. Many works explore the best ways to activate the planning capabilities of LMs, including divide and conquer [49, 61, 66, 67], grounding outputs in admissible actions [4, 28], retrospecting and refining [41, 50], and leveraging external tools [25, 46]. Meanwhile, with the increasing capabilities of LMs, growing research efforts are now dedicated to benchmark their planning capabilities in various complex environments [58, 62, 63].", + "bbox": [ + 511, + 457, + 906, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Spatial and Visual Planning in LMs", + "text_level": 1, + "bbox": [ + 511, + 741, + 823, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many planning tasks in LMs involve understanding visual environments and comprehending spatial information. In embodied agent studies, LMs play a crucial role in grounding visual entities with references in open-domain instructions and formulating plans based on spatial constraints. Consequently, they are increasingly used in physically grounded scenarios such as object rearrangement [16, 29, 38], cooking [31, 47], navigation [4, 28, 42], etc. LMs are also used in AIGC to propose spatial arrangements of", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2271", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/433397f7e27764598c6d5c20b30321565b82e3befbe1bd5ffd9ab3798bdba5a6.jpg", + "image_caption": [ + "Figure 2. Overview of the Maze Navigation scenario." + ], + "image_footnote": [], + "bbox": [ + 189, + 73, + 810, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "entities following instructions [20]. While realistic planning tasks align with real needs, their complexity and expansive action spaces limit the analysis of LMs' detailed planning capabilities. Therefore, research also focuses on LMs' planning in abstract environments and games. For example, mystery blockworld is a dynamically generated set of blockworld tasks to test generalization in LMs [59]. Additionally, many text games have been introduced to test LMs' abilities in spatial understanding and imagination [3, 51, 62, 65]. While many of these studies transform visual information into text inputs, recent works also explore direct understanding and processing of visual inputs leveraging MLLMs [23, 71].", + "bbox": [ + 89, + 268, + 483, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Benchmarks for MLLMs", + "text_level": 1, + "bbox": [ + 89, + 472, + 320, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLLMs have inherited and advanced many intriguing features from text-only LMs [45, 65]. Benchmarks for MLLMs have rapidly emerged to evaluate performance in areas such as image content understanding [15, 21], perception [22, 56], knowledge [39, 60, 70], and reasoning [21, 36, 70]. While these benchmarks quantify MLLMs' abilities in many fields, their capabilities in spatial understanding and reaction are relatively under-explored. Some benchmarks cover spatial relations understanding [34, 69], but often overlook the ability to devise complex spatial action plans based on visual environment constraints. Recently, several benchmarks explore the capability of planning utilizing spatial information, such QA and navigation in virtual or realistic scenarios [17, 23, 44, 71]. While measurements on these benchmarks reveal some incapability of MLLMs, it is often unclear why they cannot perform well in these tasks. In this work, we focus on evaluating and diagnosing the ability to comprehend spatial arrangements of objects and devise action plans to achieve specific outcomes. Our benchmark highlights future directions for improving MLLMs towards models with general intelligence.", + "bbox": [ + 89, + 494, + 483, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Visual Spatial Planning Benchmark", + "text_level": 1, + "bbox": [ + 89, + 823, + 452, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of the Benchmark", + "text_level": 1, + "bbox": [ + 89, + 848, + 344, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this benchmark, our objectives are two-fold: ① quantify the visual spatial planning capabilities of MLLMs; and ②", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "uncover current capability bottlenecks that limit the effectiveness of MLLMs in visual spatial planning tasks. The first objective requires broader task designs. Specifically, the tasks should range from classical planning tasks [13, 58] to ones in those more dynamic and realistic environments. On top of that, the second objective requires deeper task designs. In particular, performing spatial planning in visual environments requires a series of cohesive steps. For example, to generate an accurate path to navigate a player to a goal, an agent needs to be able to correctly view and understand the visual map, reason to find which actions are safe or dangerous, and come up with a detailed plan to achieve the goal. Each of these steps could be challenging for a developing MLLM, and understanding which of these subtasks challenge them most will drive future improvement.", + "bbox": [ + 511, + 268, + 906, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we propose the Visual Spatial Planning (VSP) benchmark, with the objective of measuring and diagnosing the capabilities of MLLMs in producing accurate spatial plans in visual environments. VSP consists of four scenarios: ① the simulated Maze Navigation scenario, whose main task is to move a game character through a maze; ② the photo-realistic Blocks World scenario, whose main task is to move blocks from a starting configuration to a goal configuration; ③ the dynamic Collision scenario, whose main task is to determine if there is a danger of collision between two objects in the environment, and ④ the realistic Google Map scenario, whose main task is to find a path in real streets of New York City. In addition to the main task, VSP introduces four sub-tasks that focus on the individual capabilities needed for the main task:", + "bbox": [ + 511, + 496, + 908, + 720 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- T1. Single Object Perception - Determine the characteristics of a single object;", + "- T2. Spatial Relation Perception - Determine the relative positions of two objects;", + "- T3. Environment Perception – Find textual descriptions that describe the visual environment;", + "- T4. Spatial Reasoning - Determine the consequence of a series of actions or moves." + ], + "bbox": [ + 511, + 723, + 903, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sub-task details are designed specific to both scenarios. Furthermore, to demonstrate the model's performance under different levels of environmental complexity, we establish progressive difficulty settings for various tasks,", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2272", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/be5de3cc364b071b2f1cd8222a21a9b02ae981edd5bf736d4edb1a490980a7ae.jpg", + "image_caption": [ + "Figure 3. Overview of the Blocks World scenario." + ], + "image_footnote": [], + "bbox": [ + 189, + 90, + 808, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which are measured by parameters such as map size, minimum required number of actions, etc. The detailed task statistics are provided in appendix A. In what follows, we introduce each scenario in detail, as well as the data curation and the task creation processes.", + "bbox": [ + 89, + 299, + 483, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. The Maze Navigation Scenario", + "text_level": 1, + "bbox": [ + 89, + 383, + 362, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Maze Navigation scenario is inspired by the popular implementation [13] of a fully observable path-finding problem. As depicted in Figure 2 left, it simulates a classical grid world environment with a designated start and goal position, where part of the grids contain obstacles (the \"holes\") and cannot be passed through.", + "bbox": [ + 89, + 407, + 483, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The main spatial planning task and the four sub-tasks are defined as follows:", + "bbox": [ + 89, + 498, + 482, + 526 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Main Task (Spatial Planning) - Generate a safe path to navigate from the start grid to the goal;", + "- T1 (Single Object Perception) - Determine if a specified grid is safe;", + "- T2 (Spatial Relation Perception) - Find spatial relations between the player and the goal;", + "- T3 (Environment Perception) - Find the textual description that fits the visual environment;", + "- T4 (Spatial Reasoning) - Determine the consequence of a given action series." + ], + "bbox": [ + 89, + 527, + 482, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An example of input image and questions is demonstrated in Figure 2. Each task is equipped with progressive adjusted difficulty settings to comprehensively evaluate the model's capability. For Main Task and T1-T3, the difficulties are measured by the size of the map, ranging from 3x3 to 8x8, where a larger map introduces more challenges in correctly perceiving objects and planning accordingly. For task T4, since a longer path naturally introduces more challenges for reasoning, we adopt path length ranging from 1 to 9 as the difficulty measure. Please see Appendix A for the complete example of questions and answers in each task.", + "bbox": [ + 89, + 672, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. The Blocks World Scenario", + "text_level": 1, + "bbox": [ + 89, + 848, + 336, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Blocks World is a widely-adopted planning problem [24, 27, 58]. As depicted in Figure 3 left, in this sce", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nario, the agent is given images containing sets of blocks in unique colors. These blocks are stacked vertically, forming multiple stacks on the table. The agent is asked to turn the blocks from initial state to target state through a series of moving actions. For each action, the agent can only move the top block of any stack, providing it is moved to either the table or the top of another stack.", + "bbox": [ + 511, + 299, + 906, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, the main spatial planning task and the four sub-tasks are defined as follows:", + "bbox": [ + 511, + 405, + 906, + 434 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Main Task (Spatial Planning) - Form a moving plan to achieve the target state of block arrangement;", + "- T1 (Single Object Perception) - Determine the color of the block at a specific position;", + "- T2 (Spatial Relation Perception) - Determine the spatial relation between two blocks;", + "- T3 (Environment Perception) – Find the text representation that fits the visual environment;", + "- T4 (Spatial Reasoning) - Determine the consequence of a given moving plan." + ], + "bbox": [ + 511, + 436, + 903, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An example of input image and questions is demonstrated in Figure 3. Similar to Maze Navigation, each task is equipped with progressive adjusted difficulty. Specifically, in Main Task and T4, the difficulties are measured by the number of actions involved, ranging from 1 to 7, which quantifies the complexity of the action plan. On the other hand, for tasks T1-T3, which focus on perception, the difficulty is measured by the number of blocks presented in the image, ranging from 3 to 5. Please refer to Appendix A for complete examples of the question and answer in each task.", + "bbox": [ + 511, + 580, + 906, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. The Collision and Google Map Scenario", + "text_level": 1, + "bbox": [ + 511, + 757, + 856, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further experiment in a more dynamic (the collision scenario) and realistic (the Google map scenario) settings to explore the capabilities of models in more challenging cases. The input of these two scenarios are shown in the right panels of Figure 1. Please refer to Appendix A for the complete example of each scenario.", + "bbox": [ + 511, + 779, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Collision Scenario In this scenario, the input map is similar to Maze Navigation scenario. However, in this case,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2273", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b8e90779a670b9efe65d3fd1610a87e7b171c7feafd6215a52e4d00dd6164f7.jpg", + "image_caption": [ + "Figure 4. Benchmark creation process. Left: We prepare input images that fulfill the task requirements with different difficulties. Mid: We formulate input prompts consisting of interleaved texts and images. Right: We develop automatic evaluation process for each task." + ], + "image_footnote": [], + "bbox": [ + 148, + 90, + 415, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "... The following figure shows how the player, the holes, the lands, and the goals look like.", + "bbox": [ + 418, + 95, + 643, + 116 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 419, + 116, + 480, + 125 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/69e588ee4af1a93fba9dace713eba7e66cb75ac4fdb5545f8bc1ce9c9af99707.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 116, + 637, + 141 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "... We provide an example major illustrate the rules. $<$ IMAGE-2>In this provided example:", + "- The player is at ... \n- The goal is at ..." + ], + "bbox": [ + 419, + 141, + 575, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee9c20dc749b4755bd600632c54cb021f23c2ed8b73dcfb5e9c13b5705d4fc29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 95, + 848, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formulate Prompts", + "text_level": 1, + "bbox": [ + 457, + 204, + 586, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "... We provide an example map to further illustrate the rules.", + "bbox": [ + 419, + 229, + 625, + 250 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "", + "This example features four blocks arranged in three stacks:", + "- Stack 1: Purple block (alone)", + "- Stack 2: Blue block (alone)", + "- Stack 3: From bottom to top: Orange block, Red block ..." + ], + "bbox": [ + 419, + 253, + 640, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Auto Evaluation", + "text_level": 1, + "bbox": [ + 683, + 204, + 792, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "move(red, table)", + "bbox": [ + 683, + 227, + 764, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 658, + 239, + 834, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "move(red, yellow)", + "bbox": [ + 681, + 310, + 769, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 797, + 226, + 826, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 797, + 247, + 834, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 797, + 303, + 826, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the player is moving in an environment where a car is also present and moving. Given the moving information (speed, direction) of the player and the car, the model needs to determine the time the player and the car reaches the goal, and then determine if there is a collision danger here.", + "bbox": [ + 89, + 396, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Google Map Scenario In this scenario, the input image is a real Google map depicting the streets and avenues in New York City. The starting location and the goal are randomly chosen crossroads in the map and are marked on the image. The goal of the model is to find a path from the starting location to the goal. The model needs to output a path described by directions (north, east, west, south) and the number of blocks to traverse (e.g., head north for 2 blocks, then head east for 3 blocks).", + "bbox": [ + 89, + 473, + 483, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Benchmark Creation and Quality Control", + "text_level": 1, + "bbox": [ + 89, + 618, + 449, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Figure 4, we demonstrate the general process for benchmark creation and quality control. We use the Blocks World and Maze Navigation scenarios as examples.", + "bbox": [ + 89, + 641, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, in the left panel of Figure 4, we prepare the input images used for each task and scenario. In the Maze Navigation scenario, we generate input maps using the OpenAI Gym package [13], with modifications to ensure that the positions of the player, the goal, and the holes are all randomly generated. In the Blocks World scenario, we sample pairs of images from the BIRD dataset [24], ensuring there is at least one viable plan to move the blocks from the initial state to the target state. The images are prepared conditional on different levels of difficulty.", + "bbox": [ + 89, + 688, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, as shown in the center of Figure 4, we formulate input prompts for each task. The prompt consists of interleaved text and images to provide sufficient information. For example, for Maze Navigation, we include images", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to show the appearance of elements in the map and provide example maps to better illustrate how the models should interpret the map. We invite native speakers to refine the prompts so that they accurately describe the task requirements. The prompts are in Appendix A.", + "bbox": [ + 511, + 396, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, in the right panel of Figure 4, we evaluate the performance of MLLMs under each task. It is worth noting that the answer for each task is often not unique. For example, in the Blocks World scenario, there can be many ways to move the blocks to reach the target state. As such, we develop scripts to accurately evaluate each individual answer.", + "bbox": [ + 511, + 473, + 905, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the steps above, some tasks require extra steps to ensure meaningful questions, candidates, and answers. For example, in T4 of the Blocks World scenario, the input actions must cover various valid/invalid movements. The detailed steps we followed to create each task set are provided in Appendix A. We release all images, texts, and scripts to facilitate replication and scaling.", + "bbox": [ + 511, + 564, + 906, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 686, + 645, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present evaluation results of state-of-the-art MLLMs under our main tasks and sub-tasks. Our goal is to answer the following research questions: 1 How well can state-of-the-art MLLMs perform in the visual spatial planning tasks? 2 What are the bottleneck capabilities that limit the MLLMs in visual spatial planning tasks?", + "bbox": [ + 511, + 713, + 905, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Baselines", + "text_level": 1, + "bbox": [ + 511, + 816, + 622, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate various representative MLLMs including both private and open-source models.", + "bbox": [ + 511, + 839, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For private models, we include their different versions to observe how their abilities involved and what is the", + "bbox": [ + 511, + 869, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2274", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/da34fe18f1c9e2f982ff3dcc488e23445ffc9b4e8462b6d0dbbb0d17d7b95e7e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Difficulty levelMAZE NAVIGATIONBLOCKS WORLDCOLLISIONGOO MAPOVERALL
3456781357
Gemini-1.0 [54]0.310.260.150.060.140.100.100.140.000.010.130.000.1167
Gemini-2.0 [18]0.930.810.620.410.300.270.850.490.180.070.440.070.4533
Claude-3 [5]0.520.330.160.150.160.090.120.030.000.000.180.020.1467
Claude-3.7 [9]0.710.560.560.270.180.280.890.570.330.440.490.110.4492
GPT-4V [1]0.550.360.270.130.170.100.500.170.030.000.240.020.2117
GPT-4o [7]0.680.580.350.240.180.230.710.330.120.030.160.040.3042
LLaVA-Next [37]0.030.030.020.080.090.040.040.010.000.000.020.000.0300
InternLM [19]0.270.160.060.050.040.070.100.030.000.000.250.000.0858
InternLM-VL [19]0.150.140.080.040.020.050.020.000.000.000.220.010.0600
SPHINX [35]0.110.080.050.020.040.030.070.060.010.000.040.000.0425
LLaMA-3.2 [43]0.230.180.160.080.080.100.050.000.000.000.140.000.0850
Pixtral [6]0.320.200.170.100.060.060.210.110.010.000.220.030.1242
Qwen-2.5 [11]0.450.330.200.090.150.080.060.030.000.010.450.060.1592
", + "bbox": [ + 145, + 88, + 849, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Zero-shot success rates for the spatial planning task at various difficulty levels. Maze Navigation difficulty levels represent the maze's square grid length. Blocks World difficulty levels correspond to minimum steps to a solution. Results better than $30\\%$ are bolded.", + "bbox": [ + 89, + 295, + 906, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "state-of-the-art performance. We cover the following private models: ① Gemini [18, 54] has demonstrated remarkable capabilities in image understanding and reasoning. We adopt Gemini-1.0 (Gemini-1.0-pro-vision) and Gemini-2.0 (Gemini-2.0-flash) in our experiments. ② Claude [5, 9] is a family of MLLMs strong at advanced reasoning and vision analysis. We adopt Claude-3 (claude-3-sonnet-20240229) and Claude-3.7 (claude-3-7-sonnet-20250219) in the experiments. ③ GPT [1, 7] inherent strong text understanding capabilities from text-only models and is equipped with vision capabilities. We use GPT-4V (turbo-2024-04-09) and GPT-4o (gpt-4o-2024-05-13) for evaluation.", + "bbox": [ + 88, + 349, + 480, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, we cover the following open-source models: ④ LLaVA-Next [37] demonstrates strong reasoning abilities comparable to private models. We adopt LLAVA-v1.6-VICUNA-7B for evaluation. ⑤ InternLM-XComposer2 [19] shows strong ability to understand free-form text-image composition. We adopt internlm-xcomposer2-7b and internlm-xcomposer2-v1-7b, with the former focusing on general text-image composition and the latter focusing on VL benchmarks. ⑥ SPHINX [35] unfreezes the LM during pre-training to enhance cross-model alignment. We adopt SPHINX-v2-1k. ⑦ LLama-3.2 [43] exceeds on several image understanding tasks. We adopt llama-3.2-90B-Vision in our experiments. ⑧ Pixtral [6] is a recent MLLM that demonstrates strong performance across a series of multimodal tasks. We adopt pixtral-12b-2409. ⑨ Qwen [11] demonstrates strong object localization and visual recognition in several vision-language tasks. We adopt Qwen2.5-VL-7B-Instruct. We use the public released checkpoints and codes.", + "bbox": [ + 91, + 547, + 482, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Task (Spatial Planning) Evaluation", + "text_level": 1, + "bbox": [ + 89, + 847, + 441, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we present the main task results for the four scenarios, which reflect the general spatial planning capabilities of", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MLLMs. All the evaluation in this section is conducted under zero-shot setting. Evaluations with in-context learning and fine-tuning are presented in Sections 4.5 and 4.6.", + "bbox": [ + 511, + 349, + 906, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance is demonstrated in Table 2. We also present difficulty levels in the table, which is measured by the size of the map (3 represents $3 \\times 3$ maps) in Maze Navigation and by the minimum number of steps in Blocks World. From the table, we summarize our findings as follows:", + "bbox": [ + 511, + 398, + 906, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VSP highlights both the advancements and the challenges in MLLMs' spatial planning capabilities. Notably, private MLLMs demonstrate a clear trajectory of improvement, with earlier versions performing worse and later iterations showing advancements—though still exhibiting deficiencies. In fact, even some of the most capable models frequently make mistakes on tasks such as navigating a $7 \\times 7$ map or executing 3-step block-moving tasks, which are often trivial for humans. Meanwhile, open-source models face even greater difficulties and rarely succeed in these tasks. Overall, VSP demonstrates that while MLLMs have made progress in spatial planning, they still exhibit considerable room for improvement.", + "bbox": [ + 511, + 476, + 908, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MLLMs face significant difficulties with spatial planning in dynamic and realistic environments. Based on the experimental results, most models perform worse when tested in dynamic (Collision) and realistic (Google Map) settings. We identify two major reasons: First, the tasks in Collision are complex and typically require multiple capabilities. For example, to assess collision danger, the model must locate both the player and the car on the map, calculate the time needed for the player to reach the goal, and determine if the car will hit the player during that time. This poses significant challenges for the models. Second, the inputs in the Google Map are intricate and contain a series of irrelevant symbols, making it difficult for the model to accurately interpret the map. The models' performance in these environments suggests that current models are not yet", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2275", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66827513d759600fc796ebe01d54406d81d167c9d5bf66883c09fcfdab23f362.jpg", + "image_caption": [ + "Vision Input" + ], + "image_footnote": [], + "bbox": [ + 145, + 107, + 230, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44640806becebfc13ad3a4e82b604c489be384bb5698b7638f50b55ceb968ecf.jpg", + "image_caption": [ + "Pure Text Input" + ], + "image_footnote": [], + "bbox": [ + 236, + 113, + 408, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d787593d60533ef6c60e0074180fad41ade975b6bd0c23a4e09116e51b4d68db.jpg", + "image_caption": [ + "Maze Navigation Scenario", + "Table Input" + ], + "image_footnote": [], + "bbox": [ + 421, + 112, + 584, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e4b76d406dfdbea0ef3290da8471b2a3c1d4f60c12b458f672df71d7915be16.jpg", + "image_caption": [ + "Blocks World Scenario", + "Vision Input", + "Pure Text Input", + "Figure 5. The visual and corresponding textual inputs." + ], + "image_footnote": [], + "bbox": [ + 599, + 107, + 684, + 172 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Stack with purple block", + "- Stack with blue block", + "- Stack with orange block, red block, from bottom to top" + ], + "bbox": [ + 691, + 118, + 841, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "equipped to handle spatial reasoning in such complex environments. Consequently, we focus on the Maze Navigation and Blocks World scenarios in the following experiments to diagnose the models' hidden weaknesses in VSP tasks.", + "bbox": [ + 88, + 237, + 480, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quick performance decay as difficulty increases. We observe a significant drop in the success rates of MLLMs as task difficulty escalates. For example, GPT-4V may achieve a success rate of over $50\\%$ on 3x3 size maps, but this plummets to just $10\\%$ on 8x8 maps. Analyzing the impact of increased difficulty, we identify two major challenges for the models: First, increasing size of the map in Maze Navigation scenario could make it difficult for the model to accurately perceive the positions of elements within the map. Second, the increase in both map size and the number of steps required for moving blocks heightens the challenge for the model to reason deeply through the entire path and devise a complete, viable solution. In the following experiments, we focus on these two factors and provide in-depth diagnoses with subsequent tasks.", + "bbox": [ + 91, + 297, + 483, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Challenges in open-source models. Finally, we note that open-source models often face difficulties on these tasks. As will be analyzed in Section 4.3, their perception and reasoning capabilities present greater challenges. Besides, we note two other factors: ① Context length: Open-source models typically have shorter context windows compared to private models. Thus, these models may not have enough capacity to understand the complete inputs. For example, LLAVA-V1.6-VICUNA-7B is trained with a context window of 2048 tokens, while each image consumes 576 tokens. As such, when fed with multiple images and long texts in our tasks, the total token length may surpass training, resulting in poor performance. ② Multiple image input: Our tasks require the model to understand multiple images interleaved with text inputs, whereas some open-source models are only trained with single-image inputs, with the image positioned at the start of the input. To further explore their potential, we assess their performance after finetuning in Section 4.6. Also, we note that recent open-source MLLMs often avoid these two issues, which may contribute to the improved performance observed in Table 2.", + "bbox": [ + 91, + 523, + 483, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Diagnose the Perception and Reasoning", + "text_level": 1, + "bbox": [ + 89, + 848, + 429, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the previous observation, we identify that spatial perception and reasoning could be two important capabilities", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74543ce1100bfd9be267da631fe6be99067b2da683e633ee3dd321c4291e348d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4T1T2T3T4
Random Guess0.50.250.250.50.170.250.250.5
Gemini-1.0 [54]0.580.560.330.490.860.510.540.55
Gemini-2.0 [18]0.660.910.960.800.950.960.990.76
Claude-3 [5]0.450.670.320.610.430.530.490.66
Claude-3.7 [9]0.720.880.970.790.940.840.990.94
GPT-4V [1]0.560.270.460.560.730.800.700.71
GPT-4o [7]0.580.670.580.740.950.900.900.76
LLaVA-Next [37]0.490.270.210.540.220.210.240.55
InternLM [19]0.480.270.290.580.250.320.260.53
InternLM-VL [19]0.410.200.170.470.220.200.200.53
SPHINX [35]0.560.280.320.590.240.330.270.58
LLaMA-3.2 [43]0.530.290.330.570.280.360.290.56
Pixtral [6]0.440.350.330.510.490.720.630.57
Qwen-2.5 [11]0.580.620.310.500.530.460.790.56
", + "bbox": [ + 524, + 243, + 895, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Decomposed Capability Analysis. Results better than $70\\%$ are bolded. Each task consists of test with different difficulties. Please see Appendix F for the results for different difficulties.", + "bbox": [ + 511, + 452, + 903, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for an agent to successfully perform visual spatial planning. We show these examples in Appendix E. Next, we evaluate these two abilities through the remaining tasks T1-T4. Similar to previous settings, all the evaluation is conducted under zero-shot settings.", + "bbox": [ + 511, + 513, + 903, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 shows the decomposed capability results. Strong private models like Gemini-2.0 and Claude-3.7 show better performance across these tasks, demonstrating solid perception and reasoning capabilities, which explain their superior performance in the main tasks. However, they still exhibit some deficiencies, such as perceiving the location of a single object in Maze T1, highlighting a critical area for improvement. Furthermore, the performance of many open-source models is mostly close to random guessing on these tasks, indicating significant gaps compared to private models. One caveat is that while T4 focuses on reasoning capabilities, it still relies on the perception capabilities because the input still contains images. We perform further analysis to disentangle these two abilities in Section 4.4.", + "bbox": [ + 511, + 590, + 906, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. The Effects of Visual Input", + "text_level": 1, + "bbox": [ + 511, + 816, + 756, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Previous analysis shows that even strong MLLMs have clear deficiencies in various aspects of visual spatial planning. In this study, we focus on disentangling the effects of perception and reasoning by exploring the performance", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2276", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58a85caf84a5c335f70c5375d837680ff78bfab939266940a3c79f97ccf83c0f.jpg", + "image_caption": [ + "Figure 6. Result with the visual/textual input. When environment is described by text instead of image, the performance increases vastly." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 514, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9085b0ec41a60e3f4c12d2ede371886d1a75cec48526512f821eb7919bb1318e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 97, + 897, + 217 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c2608ee94a2970baf7f4ffdcacc028202c4daf951917018631771f42d088c29c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4MainT1T2T3T4Main
Gemini, 0-shot0.580.560.330.490.170.860.510.540.550.03
Gemini, 1-shot0.500.660.310.480.200.910.680.710.590.03
Gemini, 2-shot0.530.680.310.510.210.900.760.700.610.03
Gemini, 4-shot0.530.670.350.530.190.910.640.690.620.06
GPT-4V, 0-shot0.560.270.460.560.260.730.800.700.710.10
GPT-4V, 1-shot0.550.500.470.570.280.890.840.940.730.11
GPT-4V, 2-shot0.550.630.500.560.300.900.830.950.710.16
GPT-4V, 4-shot0.540.690.540.560.290.900.790.960.73-
", + "bbox": [ + 93, + 252, + 480, + 367 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f3d0162de5cf932f9f4e7de7ee06dd83ae9b8df4734dc425e8c69d354d6cdae.jpg", + "table_caption": [ + "Table 4. Effects of providing in-context examples." + ], + "table_footnote": [], + "table_body": "
ModelSettingMAZE NAVIGATIONBLOCKS OF WORLD
T1T2T3T4MainT1T2T3T4Main
LLaVAzero-shot0.490.270.210.540.050.220.210.240.550.01
fine-tune0.530.990.510.930.601.001.001.001.000.97
InternLMzero-shot0.480.270.290.580.110.250.320.260.530.00
fine-tune0.520.590.910.590.170.290.440.690.620.09
", + "bbox": [ + 93, + 388, + 480, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Fine-tuning results for open-source models.", + "bbox": [ + 129, + 464, + 441, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gain assuming the model had perfect perception.", + "bbox": [ + 89, + 492, + 413, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The key strategy here is to create a scenario where the model has acquired all necessary information that would typically be obtained through visual perception. To this end, for every input image, we produce the corresponding textual inputs and replace those images, as shown in Figure 5. For the Maze Navigation scenario, we use either pure text descriptions or tables to depict the image. For the Blocks World scenario, we use pure text descriptions. We do not use tables for the Blocks World scenario because the number of blocks in each horizontal stack is usually unequal, making it difficult to form a complete table. Appendix B includes complete examples with pure text or table input.", + "bbox": [ + 88, + 507, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are shown in Figure 6. We observe a clear performance improvement when using textual input across every task. This suggests image perception presents significant challenges for MLLMs, and poor perception ability is a key factor in the inferior performance observed in previous tasks. Meanwhile, note that even with textual input, Gemini still cannot achieve decent performance on tasks that require reasoning. This indicates its deficiencies in reasoning.", + "bbox": [ + 88, + 686, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. In-context Learning in Visual Spatial Planning", + "text_level": 1, + "bbox": [ + 89, + 818, + 482, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In-context learning is a widely-adopted method to enhance LM's reasoning ability [14]. In this analysis, we study if it boosts visual spatial planning capabilities. We included varying numbers of examples for Gemini-1.0 and GPT-4V", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(show in Appendix C) and show results in Table 4. There are two key observations: First, in-context examples make some potential contributions, but they are not significant. The examples only benefits in several sparse cases, such as T2 in Maze Navigation and T3 in Blocks World. Second, scaling examples generally does not help, as illustrated by the saturated performance in each task.", + "bbox": [ + 511, + 255, + 903, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Fine-tuning in VSP Tasks", + "text_level": 1, + "bbox": [ + 511, + 369, + 745, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we assess the capabilities of the open-source model through dedicated training. We performed LoRA fine-tuning on llava-v1.6-vicuna-7b and internlm-xcomposer2-7b. The models are trained on 10k data points (image-text pairs) for each task and scenario. We use the default hyperparameters provided in the official repository. More fine-tuning details can be found in Appendix D. The results, shown in Table 5, demonstrate clear performance improvements for both models across a series of tasks, highlighting their potential in spatial planning. Additionally, we observe that LLaVA shows greater improvement compared to InternLM, suggesting that different model architectures may exhibit varying levels of efficacy in spatial planning capabilities.", + "bbox": [ + 511, + 391, + 906, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 614, + 633, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present VSP, a benchmark measuring and diagnosing the visual spatial planning capabilities in MLLMs. VSP quantifies the model's performance through a series of carefully designed tasks, with main tasks testing general spatial planning abilities and sub-tasks diagnosing individual capabilities needed for the main task. We show that current models fail to generate effective plans for even simple spatial planning tasks, and further analyses expose their bottlenecks in spatial perception and reasoning abilities. Our work illuminates future directions for improving MLLMs' abilities in spatial planning.", + "bbox": [ + 511, + 641, + 906, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 820, + 684, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qiucheng Wu and Shiyu Chang acknowledge the support from National Science Foundation (NSF) Grant IIS-2338252, Grant IIS-2207052, and Grant IIS-2302730.", + "bbox": [ + 511, + 844, + 905, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2277", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6, 7, 23", + "[2] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins Sri, Anthony Barrett, Dave Christianson, et al. Pddll the planning domain definition language. Technical Report, Tech. Rep., 1998. 2", + "[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning. arXiv preprint arXiv:2310.03249, 2023. 3", + "[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2", + "[5] Anthropic AI. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 6, 7, 23", + "[6] Mistral AI. Announcing pixtral 12b. https://mistral.ai/news/pixtral-12b/, 2024. Accessed: 2024-09-27. 6, 7", + "[7] Open AI. Gpt-4o. https://openai.com/index/hello-gpt-4o/.6,7,23", + "[8] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1", + "[9] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025.6,7,23", + "[10] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1", + "[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 6, 7", + "[12] Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. 1", + "[13] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 3, 4, 5", + "[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8", + "[15] Sungguk Cha, Jusung Lee, Younghyun Lee, and Cheoljong Yang. Visually dehallucinative instruction generation: Know what you don't know. arXiv preprint arXiv:2402.09717, 2024.3", + "[16] Haonan Chang, Kai Gao, Kowndinya Boyalakuntla, Alex Lee, Baichuan Huang, Harish Udhaya Kumar, Jinjin Yu, and Abdeslam Boularias. Lgmcts: Language-guided montecarlo tree search for executable semantic object rearrangement. arXiv preprint arXiv:2309.15821, 2023. 2", + "[17] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 3", + "[18] Google Deepmind. Introducing gemini 2.0: our new ai model for the agentic era. https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/, 2024.6,7,23", + "[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6, 7", + "[20] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[21] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2, 3", + "[22] Wentao Ge, Shunian Chen, Guiming Chen, Junying Chen, Zhihong Chen, Shuo Yan, Chenghao Zhu, Ziyue Lin, Wenya Xie, Xidong Wang, et al. Mllm-bench, evaluating multimodal llms using gpt-4v. arXiv preprint arXiv:2311.13951, 2023. 3", + "[23] Deepanway Ghosal, Vernon Toh Yan Han, Chia Yew Ken, and Soujanya Poria. Are language models puzzle prodigies? algorithmic puzzles unveil serious challenges in multimodal reasoning. arXiv preprint arXiv:2403.03864, 2024. 2, 3", + "[24] Tejas Gokhale, Shailaja Sampat, Zhiyuan Fang, Yezhou Yang, and Chitta Baral. Blockworld revisited: Learning and reasoning to generate event-sequences from image pairs. arXiv preprint arXiv:1905.12042, 2019. 4, 5", + "[25] Lin Guan, Karthik Valmeekam, Sarath Sreedharan, and Subbarao Kambhampati. Leveraging pre-trained large language models to construct and utilize world models for model-based task planning. Advances in Neural Information Processing Systems, 36:79081-79094, 2023. 2", + "[26] Xiaoxiao Guo, Satinder Singh, Honglak Lee, Richard L Lewis, and Xiaoshi Wang. Deep learning for real-time" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2278", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "atari game play using offline monte-carlo tree search planning. Advances in neural information processing systems, 27, 2014. 2", + "[27] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992, 2023. 4", + "[28] Rishi Hazra, Pedro Zuidberg Dos Martires, and Luc De Raedt. Saycanpay: Heuristic planning with large language models using learnable domain knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 20123-20133, 2024. 2", + "[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 2", + "[30] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 1", + "[31] Frank Joublin, Antonello Ceravola, Pavel Smirnov, Felix Ocker, Joerg Deigmoeller, Anna Belardinelli, Chao Wang, Stephan Hasler, Daniel Tanneberg, and Michael Gienger. Copal: Corrective planning of robot actions with large language models. arXiv preprint arXiv:2310.07263, 2023. 2", + "[32] Subbarao Kambhampati. Can large language models reason and plan? Annals of the New York Academy of Sciences, 1534(1):15-18, 2024. 2", + "[33] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Kaya Stechly, Mudit Verma, Siddhant Bhambri, Lucas Saldyt, and Anil Murthy. Llms can't plan, but can help planning in llm-modulo frameworks. arXiv preprint arXiv:2402.01817, 2024. 2", + "[34] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 3", + "[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6, 7", + "[36] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 6, 7", + "[38] Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Yifan Xu, Xixuan Song, Shudan Zhang, Hanyu Lai, Xinyi Liu, Hanlin Zhao, et al. Visualagentbench: Towards large multimodal models as visual foundation agents. arXiv preprint arXiv:2408.06327, 2024. 2" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. arXiv e-prints, pages arXiv-2310, 2023. 3", + "[40] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2", + "[41] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[42] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2", + "[43] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 6, 7", + "[44] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024.3", + "[45] Zhangyang Qi, Ye Fang, Mengchen Zhang, Zeyi Sun, Tong Wu, Ziwei Liu, Dahua Lin, Jiaqi Wang, and Hengshuang Zhao. Gemini vs gpt-4v: A preliminary comparison and combination of vision-language models through qualitative cases. arXiv preprint arXiv:2312.15011, 2023. 3", + "[46] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Guoqing Du, Shiwei Shi, Hangyu Mao, Xingyu Zeng, and Rui Zhao. Tptu: Task planning and tool usage of large language model-based ai agents. arXiv preprint arXiv:2308.03427, 2023. 2", + "[47] Md Sadman Sakib and Yu Sun. From cooking recipes to robot task trees-improving planning correctness and task efficiency by leveraging llms with a knowledge network. arXiv preprint arXiv:2309.09181, 2023. 2", + "[48] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 1", + "[49] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[50] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[51] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2279", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768, 2020. 3", + "[52] Kaya Stechly, Karthik Valmeekam, and Subbarao Kambhampati. On the self-verification limitations of large language models on reasoning and planning tasks. arXiv preprint arXiv:2402.08115, 2024. 2", + "[53] Richard S Sutton. Planning by incremental dynamic programming. In Machine learning proceedings 1991, pages 353-357. Elsevier, 1991. 2", + "[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 6, 7, 23", + "[55] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2", + "[56] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209, 2024. 3", + "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[58] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large language models still can't plan (a benchmark for llms on planning and reasoning about change). arXiv preprint arXiv:2206.10498, 2022. 2, 3, 4", + "[59] Karthik Valmeekam, Matthew Marquez, Sarath Sreedharan, and Subbarao Kambhampati. On the planning abilities of large language models - a critical investigation. In Advances in Neural Information Processing Systems, pages 75993-76005. Curran Associates, Inc., 2023. 1, 3", + "[60] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2, 3", + "[61] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2", + "[62] Yue Wu, Xuan Tang, Tom M Mitchell, and Yuanzhi Li. Smartplay: A benchmark for llms as intelligent agents. arXiv preprint arXiv:2310.01557, 2023. 2, 3", + "[63] Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. Travelplanner: A benchmark for real-world planning with language agents. arXiv preprint arXiv:2402.01622, 2024. 2", + "[64] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3", + "[66] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2", + "[67] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[68] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, et al. Mmt-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask agi. arXiv preprint arXiv:2404.16006, 2024. 1", + "[69] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 3", + "[70] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2, 3", + "[71] Wenqi Zhang, Zhenglin Cheng, Yuanyu He, Mengna Wang, Yongliang Shen, Zeqi Tan, Guiyang Hou, Mingqian He, Yanna Ma, Weiming Lu, et al. Multimodal self-instruct: Synthetic abstract image and visual reasoning instruction using language model. arXiv preprint arXiv:2407.07053, 2024. 2, 3", + "[72] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1" + ], + "bbox": [ + 516, + 90, + 905, + 655 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2280", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_model.json b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..69a29b1fe565b09f729bbf5b2666adaf4fa17c9c --- /dev/null +++ b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_model.json @@ -0,0 +1,2840 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.13, + 0.887, + 0.177 + ], + "angle": 0, + "content": "VSP: Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.203, + 0.761, + 0.257 + ], + "angle": 0, + "content": "Qiucheng Wu\\(^{1}\\), Handong Zhao\\(^{2}\\), Michael Saxon\\(^{1}\\), Trung Bui\\(^{2}\\), William Yang Wang\\(^{1}\\), Yang Zhang\\(^{3}\\), Shiyu Chang\\(^{1}\\) \n\\(^{1}\\)UC Santa Barbara, \\(^{2}\\)Adobe Research, \\(^{3}\\)MIT-IBM Watson AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.26, + 0.576, + 0.274 + ], + "angle": 0, + "content": "qiucheng@ucsb.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.706 + ], + "angle": 0, + "content": "Multimodal large language models are an exciting emerging class of language models (LMs) that have merged classic LM capabilities with those of image processing systems. However, how these capabilities integrate is often not intuitive and warrants direct investigation. One understudied capability in MLLMs is visual spatial planning—the ability to comprehend the spatial arrangements of objects and devise action plans to achieve desired outcomes in visual scenes. It is unclear why MLLMs fall short on these tasks generally considered easy for humans, given their successes across other diverse scenarios. To this end, we introduce VSP, a benchmark that 1) evaluates the spatial planning capability in MLLMs in general, and 2) diagnoses this capability via finer-grained sub-tasks, including perception and reasoning, and measure the capabilities of models through these sub-tasks. Our evaluation confirms that both open-source and private MLLMs fail to generate effective plans for even simple spatial planning tasks. Evaluations on the fine-grained analytical tasks further reveal fundamental deficiencies in the models' visual perception and bottle-necks in reasoning abilities, explaining their worse performance in the general spatial planning tasks. Our work illuminates future directions for improving MLLMs' abilities in spatial planning. Our benchmark is publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.864 + ], + "angle": 0, + "content": "The rapid advancement of large language models has driven considerable growth in their capabilities to produce fluent text in many domains, generating outputs exhibiting potential \"reasoning\" and \"understanding\" abilities [12, 14, 30, 57]. Recently, multimodal large language models (MLLMs) have advanced on LLMs through training on native image inputs, to achieve impressive performance generating text describing and relating to input images [1, 8, 10," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.309, + 0.905, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.569, + 0.907, + 0.626 + ], + "angle": 0, + "content": "Figure 1. Overview of the VSP benchmark. The benchmark evaluates tasks that require MLLMs to perform complex reasoning by utilizing the spatial information they perceive, and diagnoses their failure modes using a series of carefully designed subtasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.749 + ], + "angle": 0, + "content": "37, 54], with applications in image captioning, visual question answering, and visual reasoning [48, 64, 68, 72]. The swift evolution of MLLMs has enabled them to tackle increasingly sophisticated tasks that require multiple emerging abilities in complex scenarios. However, as model capabilities and deployment needs advance, the challenges in usefully evaluating them grow in kind." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Planning is a fundamental capability in intelligent systems that is particularly contested in LMs [59]. Visual spatial planning refers to the task of comprehending the spatial arrangement of objects in a scene and designing action plans to achieve a desired outcome. For example, the classical maze problem can be considered a visual planning task, where an agent is given an input image describing the maze environment and is asked to produce a viable path to navigate the player from the starting position to the goal. This task requires two capabilities: image perception, which en" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/UCSB-NLP-Chang/Visual-Spatial-Planning" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2270" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.08, + 0.495, + 0.24 + ], + "angle": 0, + "content": "
BenchmarkCapabilitiesDiagnosing Incapabilities
Visual RecognitionSpatial Relation UnderstandingPlan with Spatial Info
MMMU [70]✓XXX
MathVision [60]✓XXX
MME [21]✓✓XX
SeedBench [34]✓✓XX
MM-Vet [69]✓✓XX
AlgoPuzzle [23]✓✓✓X
MMSI [71]✓✓✓X
VSP✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.298 + ], + "angle": 0, + "content": "Table 1. Comparison with representative existing benchmarks. Note that we focus on spatial reasoning related capabilities. ✓ indicates that the measurement is covered by the corresponding benchmark, while \\( \\times \\) indicates not covered." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.356 + ], + "angle": 0, + "content": "ables the agent to understand the objects, environment and spatial relations present in the image, and reasoning, which enables the agent to perform strategic decision-making." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Visual spatial planning is an important capability in many potential applications for MLLMs, such as navigating in complex environments with autonomous driving [40, 55] or manipulating objects with robotic hands [16, 29]. However, although there have been increasingly more benchmarks to evaluate the vision processing capabilities of MLLMs, few systematically investigate the underlying factors affecting their (in)ability to perform visual spatial planning tasks. As shown in Table 1, while most of the existing benchmarks identify the capability to recognize objects as a key feature of MLLM, quite a few benchmarks [60, 70] do not evaluate their ability to understand spatial relations, and even fewer benchmarks [23, 71] examine the ability to use spatial information in complex tasks, e.g., navigation on a road map based on the position towards goals. More importantly, it is often unclear why MLLMs perform so poorly on the spatial planning tasks that are generally considered easy for humans [23]. As a result, two research questions emerged: ① [Evaluation] How performant are MLLMs in performing visual planning tasks? ② [Diagnose] What are the bottleneck capabilities, e.g., perception or reasoning, that limit the performance of MLLMs in visual planning tasks?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To this end, we introduce Visual Spatial Planning (VSP), a benchmark specifically designed to evaluate and diagnose the spatial planning capabilities of MLLMs. As illustrated in Figure 1, the VSP benchmark consists of various environment settings designed to assess the capabilities of models in different scenarios. We focus on abstract scenarios to facilitate the design of diagnostic subtasks with consistent environments (Sec. 3.5). The first two scenarios, Maze Navigation and Blocks World, are basic environments developed from classical planning tasks. Additionally, we challenge the model's abilities in dynamic and realistic applications through the Collision and Google Map scenarios, respectively. All environments are fully observable through input" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "images. The MLLMs are required to interpret the visual inputs, deduce the consequences of each action, and execute the designated tasks. To comprehensively evaluate the fine-grained capabilities needed for visual spatial planning, VSP includes 4.6K questions in 12 meticulously designed tasks that feature both simulated and photorealistic visual inputs. In addition to testing end-to-end spatial planning, these tasks test essential individual visual planning abilities, such as image perception and reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.395 + ], + "angle": 0, + "content": "We apply the VSP benchmark to evaluate existing state-of-the-art MLLMs, including both open-source and private ones. Surprisingly, we find that even the most competitive MLLMs sometimes struggle in performing the simplest visual planning tasks, such as a 3x3 maze problem. Our fine-grained capability analysis further reveals that existing MLLMs have flaws in reasoning and bigger bottlenecks in perception. We believe the VSP benchmark highlights critical weaknesses in current MLLMs and sheds light on future directions for enhancing their spatial understanding and planning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.655, + 0.426 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.435, + 0.746, + 0.452 + ], + "angle": 0, + "content": "2.1. General Planning in LMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Planning has been a central focus of research in AI. Traditional work in AI planning includes using formal languages to represent and solve planning problems [2], and developing algorithms like dynamic programming and reinforcement learning to explore environments and formulate viable plans [26, 53]. While these works mostly focus on predefined and restricted environments, recently, with the advancement of LMs, it has become intriguing to study whether LMs, with the potential to be general intelligent agents, can perform planning in different settings and environments [32, 33, 52]. Many works explore the best ways to activate the planning capabilities of LMs, including divide and conquer [49, 61, 66, 67], grounding outputs in admissible actions [4, 28], retrospecting and refining [41, 50], and leveraging external tools [25, 46]. Meanwhile, with the increasing capabilities of LMs, growing research efforts are now dedicated to benchmark their planning capabilities in various complex environments [58, 62, 63]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.825, + 0.759 + ], + "angle": 0, + "content": "2.2. Spatial and Visual Planning in LMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Many planning tasks in LMs involve understanding visual environments and comprehending spatial information. In embodied agent studies, LMs play a crucial role in grounding visual entities with references in open-domain instructions and formulating plans based on spatial constraints. Consequently, they are increasingly used in physically grounded scenarios such as object rearrangement [16, 29, 38], cooking [31, 47], navigation [4, 28, 42], etc. LMs are also used in AIGC to propose spatial arrangements of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "2271" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.19, + 0.074, + 0.812, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.243, + 0.658, + 0.257 + ], + "angle": 0, + "content": "Figure 2. Overview of the Maze Navigation scenario." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.485, + 0.466 + ], + "angle": 0, + "content": "entities following instructions [20]. While realistic planning tasks align with real needs, their complexity and expansive action spaces limit the analysis of LMs' detailed planning capabilities. Therefore, research also focuses on LMs' planning in abstract environments and games. For example, mystery blockworld is a dynamically generated set of blockworld tasks to test generalization in LMs [59]. Additionally, many text games have been introduced to test LMs' abilities in spatial understanding and imagination [3, 51, 62, 65]. While many of these studies transform visual information into text inputs, recent works also explore direct understanding and processing of visual inputs leveraging MLLMs [23, 71]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.473, + 0.321, + 0.488 + ], + "angle": 0, + "content": "2.3. Benchmarks for MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.814 + ], + "angle": 0, + "content": "MLLMs have inherited and advanced many intriguing features from text-only LMs [45, 65]. Benchmarks for MLLMs have rapidly emerged to evaluate performance in areas such as image content understanding [15, 21], perception [22, 56], knowledge [39, 60, 70], and reasoning [21, 36, 70]. While these benchmarks quantify MLLMs' abilities in many fields, their capabilities in spatial understanding and reaction are relatively under-explored. Some benchmarks cover spatial relations understanding [34, 69], but often overlook the ability to devise complex spatial action plans based on visual environment constraints. Recently, several benchmarks explore the capability of planning utilizing spatial information, such QA and navigation in virtual or realistic scenarios [17, 23, 44, 71]. While measurements on these benchmarks reveal some incapability of MLLMs, it is often unclear why they cannot perform well in these tasks. In this work, we focus on evaluating and diagnosing the ability to comprehend spatial arrangements of objects and devise action plans to achieve specific outcomes. Our benchmark highlights future directions for improving MLLMs towards models with general intelligence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.453, + 0.841 + ], + "angle": 0, + "content": "3. The Visual Spatial Planning Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.345, + 0.864 + ], + "angle": 0, + "content": "3.1. Overview of the Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this benchmark, our objectives are two-fold: ① quantify the visual spatial planning capabilities of MLLMs; and ②" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.496 + ], + "angle": 0, + "content": "uncover current capability bottlenecks that limit the effectiveness of MLLMs in visual spatial planning tasks. The first objective requires broader task designs. Specifically, the tasks should range from classical planning tasks [13, 58] to ones in those more dynamic and realistic environments. On top of that, the second objective requires deeper task designs. In particular, performing spatial planning in visual environments requires a series of cohesive steps. For example, to generate an accurate path to navigate a player to a goal, an agent needs to be able to correctly view and understand the visual map, reason to find which actions are safe or dangerous, and come up with a detailed plan to achieve the goal. Each of these steps could be challenging for a developing MLLM, and understanding which of these subtasks challenge them most will drive future improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.909, + 0.722 + ], + "angle": 0, + "content": "To this end, we propose the Visual Spatial Planning (VSP) benchmark, with the objective of measuring and diagnosing the capabilities of MLLMs in producing accurate spatial plans in visual environments. VSP consists of four scenarios: ① the simulated Maze Navigation scenario, whose main task is to move a game character through a maze; ② the photo-realistic Blocks World scenario, whose main task is to move blocks from a starting configuration to a goal configuration; ③ the dynamic Collision scenario, whose main task is to determine if there is a danger of collision between two objects in the environment, and ④ the realistic Google Map scenario, whose main task is to find a path in real streets of New York City. In addition to the main task, VSP introduces four sub-tasks that focus on the individual capabilities needed for the main task:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.724, + 0.905, + 0.753 + ], + "angle": 0, + "content": "- T1. Single Object Perception - Determine the characteristics of a single object;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.753, + 0.905, + 0.782 + ], + "angle": 0, + "content": "- T2. Spatial Relation Perception - Determine the relative positions of two objects;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.782, + 0.905, + 0.81 + ], + "angle": 0, + "content": "- T3. Environment Perception – Find textual descriptions that describe the visual environment;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.839 + ], + "angle": 0, + "content": "- T4. Spatial Reasoning - Determine the consequence of a series of actions or moves." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.724, + 0.905, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The sub-task details are designed specific to both scenarios. Furthermore, to demonstrate the model's performance under different levels of environmental complexity, we establish progressive difficulty settings for various tasks," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2272" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.191, + 0.092, + 0.809, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.273, + 0.645, + 0.286 + ], + "angle": 0, + "content": "Figure 3. Overview of the Blocks World scenario." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.377 + ], + "angle": 0, + "content": "which are measured by parameters such as map size, minimum required number of actions, etc. The detailed task statistics are provided in appendix A. In what follows, we introduce each scenario in detail, as well as the data curation and the task creation processes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.385, + 0.364, + 0.401 + ], + "angle": 0, + "content": "3.2. The Maze Navigation Scenario" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.498 + ], + "angle": 0, + "content": "The Maze Navigation scenario is inspired by the popular implementation [13] of a fully observable path-finding problem. As depicted in Figure 2 left, it simulates a classical grid world environment with a designated start and goal position, where part of the grids contain obstacles (the \"holes\") and cannot be passed through." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.483, + 0.527 + ], + "angle": 0, + "content": "The main spatial planning task and the four sub-tasks are defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.529, + 0.483, + 0.558 + ], + "angle": 0, + "content": "- Main Task (Spatial Planning) - Generate a safe path to navigate from the start grid to the goal;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.558, + 0.483, + 0.586 + ], + "angle": 0, + "content": "- T1 (Single Object Perception) - Determine if a specified grid is safe;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.586, + 0.483, + 0.615 + ], + "angle": 0, + "content": "- T2 (Spatial Relation Perception) - Find spatial relations between the player and the goal;" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.643 + ], + "angle": 0, + "content": "- T3 (Environment Perception) - Find the textual description that fits the visual environment;" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.672 + ], + "angle": 0, + "content": "- T4 (Spatial Reasoning) - Determine the consequence of a given action series." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.529, + 0.483, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.84 + ], + "angle": 0, + "content": "An example of input image and questions is demonstrated in Figure 2. Each task is equipped with progressive adjusted difficulty settings to comprehensively evaluate the model's capability. For Main Task and T1-T3, the difficulties are measured by the size of the map, ranging from 3x3 to 8x8, where a larger map introduces more challenges in correctly perceiving objects and planning accordingly. For task T4, since a longer path naturally introduces more challenges for reasoning, we adopt path length ranging from 1 to 9 as the difficulty measure. Please see Appendix A for the complete example of questions and answers in each task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.338, + 0.863 + ], + "angle": 0, + "content": "3.3. The Blocks World Scenario" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The Blocks World is a widely-adopted planning problem [24, 27, 58]. As depicted in Figure 3 left, in this sce" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.405 + ], + "angle": 0, + "content": "nario, the agent is given images containing sets of blocks in unique colors. These blocks are stacked vertically, forming multiple stacks on the table. The agent is asked to turn the blocks from initial state to target state through a series of moving actions. For each action, the agent can only move the top block of any stack, providing it is moved to either the table or the top of another stack." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Similarly, the main spatial planning task and the four sub-tasks are defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.905, + 0.465 + ], + "angle": 0, + "content": "- Main Task (Spatial Planning) - Form a moving plan to achieve the target state of block arrangement;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.466, + 0.905, + 0.495 + ], + "angle": 0, + "content": "- T1 (Single Object Perception) - Determine the color of the block at a specific position;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.495, + 0.905, + 0.522 + ], + "angle": 0, + "content": "- T2 (Spatial Relation Perception) - Determine the spatial relation between two blocks;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.904, + 0.551 + ], + "angle": 0, + "content": "- T3 (Environment Perception) – Find the text representation that fits the visual environment;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.905, + 0.581 + ], + "angle": 0, + "content": "- T4 (Spatial Reasoning) - Determine the consequence of a given moving plan." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.437, + 0.905, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.733 + ], + "angle": 0, + "content": "An example of input image and questions is demonstrated in Figure 3. Similar to Maze Navigation, each task is equipped with progressive adjusted difficulty. Specifically, in Main Task and T4, the difficulties are measured by the number of actions involved, ranging from 1 to 7, which quantifies the complexity of the action plan. On the other hand, for tasks T1-T3, which focus on perception, the difficulty is measured by the number of blocks presented in the image, ranging from 3 to 5. Please refer to Appendix A for complete examples of the question and answer in each task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.857, + 0.774 + ], + "angle": 0, + "content": "3.4. The Collision and Google Map Scenario" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.87 + ], + "angle": 0, + "content": "We further experiment in a more dynamic (the collision scenario) and realistic (the Google map scenario) settings to explore the capabilities of models in more challenging cases. The input of these two scenarios are shown in the right panels of Figure 1. Please refer to Appendix A for the complete example of each scenario." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "- Collision Scenario In this scenario, the input map is similar to Maze Navigation scenario. However, in this case," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2273" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.092, + 0.416, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.096, + 0.645, + 0.117 + ], + "angle": 0, + "content": "... The following figure shows how the player, the holes, the lands, and the goals look like." + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.117, + 0.482, + 0.125 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.117, + 0.638, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.142, + 0.576, + 0.173 + ], + "angle": 0, + "content": "... We provide an example major illustrate the rules. \\(< \\) IMAGE-2>In this provided example:" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.173, + 0.525, + 0.193 + ], + "angle": 0, + "content": "- The player is at ... \n- The goal is at ..." + }, + { + "type": "list", + "bbox": [ + 0.421, + 0.142, + 0.576, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.096, + 0.849, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.458, + 0.205, + 0.588, + 0.219 + ], + "angle": 0, + "content": "Formulate Prompts" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.23, + 0.626, + 0.25 + ], + "angle": 0, + "content": "... We provide an example map to further illustrate the rules." + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.255, + 0.474, + 0.265 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.271, + 0.578, + 0.291 + ], + "angle": 0, + "content": "This example features four blocks arranged in three stacks:" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.291, + 0.57, + 0.3 + ], + "angle": 0, + "content": "- Stack 1: Purple block (alone)" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.301, + 0.561, + 0.311 + ], + "angle": 0, + "content": "- Stack 2: Blue block (alone)" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.311, + 0.642, + 0.33 + ], + "angle": 0, + "content": "- Stack 3: From bottom to top: Orange block, Red block ..." + }, + { + "type": "list", + "bbox": [ + 0.421, + 0.255, + 0.642, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.684, + 0.205, + 0.794, + 0.216 + ], + "angle": 0, + "content": "Auto Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.228, + 0.766, + 0.238 + ], + "angle": 0, + "content": "move(red, table)" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.24, + 0.835, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.683, + 0.311, + 0.771, + 0.321 + ], + "angle": 0, + "content": "move(red, yellow)" + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.227, + 0.827, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.248, + 0.835, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.304, + 0.827, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.351, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Figure 4. Benchmark creation process. Left: We prepare input images that fulfill the task requirements with different difficulties. Mid: We formulate input prompts consisting of interleaved texts and images. Right: We develop automatic evaluation process for each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.472 + ], + "angle": 0, + "content": "the player is moving in an environment where a car is also present and moving. Given the moving information (speed, direction) of the player and the car, the model needs to determine the time the player and the car reaches the goal, and then determine if there is a collision danger here." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.609 + ], + "angle": 0, + "content": "- Google Map Scenario In this scenario, the input image is a real Google map depicting the streets and avenues in New York City. The starting location and the goal are randomly chosen crossroads in the map and are marked on the image. The goal of the model is to find a path from the starting location to the goal. The model needs to output a path described by directions (north, east, west, south) and the number of blocks to traverse (e.g., head north for 2 blocks, then head east for 3 blocks)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.45, + 0.635 + ], + "angle": 0, + "content": "3.5. Benchmark Creation and Quality Control" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.687 + ], + "angle": 0, + "content": "In Figure 4, we demonstrate the general process for benchmark creation and quality control. We use the Blocks World and Maze Navigation scenarios as examples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.839 + ], + "angle": 0, + "content": "First, in the left panel of Figure 4, we prepare the input images used for each task and scenario. In the Maze Navigation scenario, we generate input maps using the OpenAI Gym package [13], with modifications to ensure that the positions of the player, the goal, and the holes are all randomly generated. In the Blocks World scenario, we sample pairs of images from the BIRD dataset [24], ensuring there is at least one viable plan to move the blocks from the initial state to the target state. The images are prepared conditional on different levels of difficulty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Second, as shown in the center of Figure 4, we formulate input prompts for each task. The prompt consists of interleaved text and images to provide sufficient information. For example, for Maze Navigation, we include images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.472 + ], + "angle": 0, + "content": "to show the appearance of elements in the map and provide example maps to better illustrate how the models should interpret the map. We invite native speakers to refine the prompts so that they accurately describe the task requirements. The prompts are in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.564 + ], + "angle": 0, + "content": "Finally, in the right panel of Figure 4, we evaluate the performance of MLLMs under each task. It is worth noting that the answer for each task is often not unique. For example, in the Blocks World scenario, there can be many ways to move the blocks to reach the target state. As such, we develop scripts to accurately evaluate each individual answer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.672 + ], + "angle": 0, + "content": "In addition to the steps above, some tasks require extra steps to ensure meaningful questions, candidates, and answers. For example, in T4 of the Blocks World scenario, the input actions must cover various valid/invalid movements. The detailed steps we followed to create each task set are provided in Appendix A. We release all images, texts, and scripts to facilitate replication and scaling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.646, + 0.705 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.906, + 0.805 + ], + "angle": 0, + "content": "In this section, we present evaluation results of state-of-the-art MLLMs under our main tasks and sub-tasks. Our goal is to answer the following research questions: 1 How well can state-of-the-art MLLMs perform in the visual spatial planning tasks? 2 What are the bottleneck capabilities that limit the MLLMs in visual spatial planning tasks?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.623, + 0.831 + ], + "angle": 0, + "content": "4.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.87 + ], + "angle": 0, + "content": "We evaluate various representative MLLMs including both private and open-source models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "For private models, we include their different versions to observe how their abilities involved and what is the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2274" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.146, + 0.089, + 0.851, + 0.285 + ], + "angle": 0, + "content": "
Difficulty levelMAZE NAVIGATIONBLOCKS WORLDCOLLISIONGOO MAPOVERALL
3456781357
Gemini-1.0 [54]0.310.260.150.060.140.100.100.140.000.010.130.000.1167
Gemini-2.0 [18]0.930.810.620.410.300.270.850.490.180.070.440.070.4533
Claude-3 [5]0.520.330.160.150.160.090.120.030.000.000.180.020.1467
Claude-3.7 [9]0.710.560.560.270.180.280.890.570.330.440.490.110.4492
GPT-4V [1]0.550.360.270.130.170.100.500.170.030.000.240.020.2117
GPT-4o [7]0.680.580.350.240.180.230.710.330.120.030.160.040.3042
LLaVA-Next [37]0.030.030.020.080.090.040.040.010.000.000.020.000.0300
InternLM [19]0.270.160.060.050.040.070.100.030.000.000.250.000.0858
InternLM-VL [19]0.150.140.080.040.020.050.020.000.000.000.220.010.0600
SPHINX [35]0.110.080.050.020.040.030.070.060.010.000.040.000.0425
LLaMA-3.2 [43]0.230.180.160.080.080.100.050.000.000.000.140.000.0850
Pixtral [6]0.320.200.170.100.060.060.210.110.010.000.220.030.1242
Qwen-2.5 [11]0.450.330.200.090.150.080.060.030.000.010.450.060.1592
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.296, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Table 2. Zero-shot success rates for the spatial planning task at various difficulty levels. Maze Navigation difficulty levels represent the maze's square grid length. Blocks World difficulty levels correspond to minimum steps to a solution. Results better than \\(30\\%\\) are bolded." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.482, + 0.546 + ], + "angle": 0, + "content": "state-of-the-art performance. We cover the following private models: ① Gemini [18, 54] has demonstrated remarkable capabilities in image understanding and reasoning. We adopt Gemini-1.0 (Gemini-1.0-pro-vision) and Gemini-2.0 (Gemini-2.0-flash) in our experiments. ② Claude [5, 9] is a family of MLLMs strong at advanced reasoning and vision analysis. We adopt Claude-3 (claude-3-sonnet-20240229) and Claude-3.7 (claude-3-7-sonnet-20250219) in the experiments. ③ GPT [1, 7] inherent strong text understanding capabilities from text-only models and is equipped with vision capabilities. We use GPT-4V (turbo-2024-04-09) and GPT-4o (gpt-4o-2024-05-13) for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.549, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Besides, we cover the following open-source models: ④ LLaVA-Next [37] demonstrates strong reasoning abilities comparable to private models. We adopt LLAVA-v1.6-VICUNA-7B for evaluation. ⑤ InternLM-XComposer2 [19] shows strong ability to understand free-form text-image composition. We adopt internlm-xcomposer2-7b and internlm-xcomposer2-v1-7b, with the former focusing on general text-image composition and the latter focusing on VL benchmarks. ⑥ SPHINX [35] unfreezes the LM during pre-training to enhance cross-model alignment. We adopt SPHINX-v2-1k. ⑦ LLama-3.2 [43] exceeds on several image understanding tasks. We adopt llama-3.2-90B-Vision in our experiments. ⑧ Pixtral [6] is a recent MLLM that demonstrates strong performance across a series of multimodal tasks. We adopt pixtral-12b-2409. ⑨ Qwen [11] demonstrates strong object localization and visual recognition in several vision-language tasks. We adopt Qwen2.5-VL-7B-Instruct. We use the public released checkpoints and codes." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.848, + 0.442, + 0.865 + ], + "angle": 0, + "content": "4.2. Main Task (Spatial Planning) Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "First, we present the main task results for the four scenarios, which reflect the general spatial planning capabilities of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.396 + ], + "angle": 0, + "content": "MLLMs. All the evaluation in this section is conducted under zero-shot setting. Evaluations with in-context learning and fine-tuning are presented in Sections 4.5 and 4.6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.474 + ], + "angle": 0, + "content": "The performance is demonstrated in Table 2. We also present difficulty levels in the table, which is measured by the size of the map (3 represents \\(3 \\times 3\\) maps) in Maze Navigation and by the minimum number of steps in Blocks World. From the table, we summarize our findings as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.909, + 0.672 + ], + "angle": 0, + "content": "VSP highlights both the advancements and the challenges in MLLMs' spatial planning capabilities. Notably, private MLLMs demonstrate a clear trajectory of improvement, with earlier versions performing worse and later iterations showing advancements—though still exhibiting deficiencies. In fact, even some of the most capable models frequently make mistakes on tasks such as navigating a \\(7 \\times 7\\) map or executing 3-step block-moving tasks, which are often trivial for humans. Meanwhile, open-source models face even greater difficulties and rarely succeed in these tasks. Overall, VSP demonstrates that while MLLMs have made progress in spatial planning, they still exhibit considerable room for improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "MLLMs face significant difficulties with spatial planning in dynamic and realistic environments. Based on the experimental results, most models perform worse when tested in dynamic (Collision) and realistic (Google Map) settings. We identify two major reasons: First, the tasks in Collision are complex and typically require multiple capabilities. For example, to assess collision danger, the model must locate both the player and the car on the map, calculate the time needed for the player to reach the goal, and determine if the car will hit the player during that time. This poses significant challenges for the models. Second, the inputs in the Google Map are intricate and contain a series of irrelevant symbols, making it difficult for the model to accurately interpret the map. The models' performance in these environments suggests that current models are not yet" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2275" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.09, + 0.455, + 0.103 + ], + "angle": 0, + "content": "Maze Navigation Scenario" + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.108, + 0.232, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.177, + 0.224, + 0.189 + ], + "angle": 0, + "content": "Vision Input" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.114, + 0.409, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.177, + 0.37, + 0.189 + ], + "angle": 0, + "content": "Pure Text Input" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.113, + 0.585, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.177, + 0.536, + 0.189 + ], + "angle": 0, + "content": "Table Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.09, + 0.801, + 0.103 + ], + "angle": 0, + "content": "Blocks World Scenario" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.108, + 0.685, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.177, + 0.678, + 0.189 + ], + "angle": 0, + "content": "Vision Input" + }, + { + "type": "text", + "bbox": [ + 0.692, + 0.119, + 0.818, + 0.13 + ], + "angle": 0, + "content": "- Stack with purple block" + }, + { + "type": "text", + "bbox": [ + 0.692, + 0.131, + 0.808, + 0.141 + ], + "angle": 0, + "content": "- Stack with blue block" + }, + { + "type": "text", + "bbox": [ + 0.692, + 0.142, + 0.842, + 0.163 + ], + "angle": 0, + "content": "- Stack with orange block, red block, from bottom to top" + }, + { + "type": "list", + "bbox": [ + 0.692, + 0.119, + 0.842, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.177, + 0.815, + 0.189 + ], + "angle": 0, + "content": "Pure Text Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.197, + 0.658, + 0.211 + ], + "angle": 0, + "content": "Figure 5. The visual and corresponding textual inputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.238, + 0.482, + 0.298 + ], + "angle": 0, + "content": "equipped to handle spatial reasoning in such complex environments. Consequently, we focus on the Maze Navigation and Blocks World scenarios in the following experiments to diagnose the models' hidden weaknesses in VSP tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.299, + 0.484, + 0.525 + ], + "angle": 0, + "content": "Quick performance decay as difficulty increases. We observe a significant drop in the success rates of MLLMs as task difficulty escalates. For example, GPT-4V may achieve a success rate of over \\(50\\%\\) on 3x3 size maps, but this plummets to just \\(10\\%\\) on 8x8 maps. Analyzing the impact of increased difficulty, we identify two major challenges for the models: First, increasing size of the map in Maze Navigation scenario could make it difficult for the model to accurately perceive the positions of elements within the map. Second, the increase in both map size and the number of steps required for moving blocks heightens the challenge for the model to reason deeply through the entire path and devise a complete, viable solution. In the following experiments, we focus on these two factors and provide in-depth diagnoses with subsequent tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.842 + ], + "angle": 0, + "content": "Challenges in open-source models. Finally, we note that open-source models often face difficulties on these tasks. As will be analyzed in Section 4.3, their perception and reasoning capabilities present greater challenges. Besides, we note two other factors: ① Context length: Open-source models typically have shorter context windows compared to private models. Thus, these models may not have enough capacity to understand the complete inputs. For example, LLAVA-V1.6-VICUNA-7B is trained with a context window of 2048 tokens, while each image consumes 576 tokens. As such, when fed with multiple images and long texts in our tasks, the total token length may surpass training, resulting in poor performance. ② Multiple image input: Our tasks require the model to understand multiple images interleaved with text inputs, whereas some open-source models are only trained with single-image inputs, with the image positioned at the start of the input. To further explore their potential, we assess their performance after finetuning in Section 4.6. Also, we note that recent open-source MLLMs often avoid these two issues, which may contribute to the improved performance observed in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.849, + 0.43, + 0.866 + ], + "angle": 0, + "content": "4.3. Diagnose the Perception and Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "From the previous observation, we identify that spatial perception and reasoning could be two important capabilities" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.244, + 0.897, + 0.442 + ], + "angle": 0, + "content": "
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4T1T2T3T4
Random Guess0.50.250.250.50.170.250.250.5
Gemini-1.0 [54]0.580.560.330.490.860.510.540.55
Gemini-2.0 [18]0.660.910.960.800.950.960.990.76
Claude-3 [5]0.450.670.320.610.430.530.490.66
Claude-3.7 [9]0.720.880.970.790.940.840.990.94
GPT-4V [1]0.560.270.460.560.730.800.700.71
GPT-4o [7]0.580.670.580.740.950.900.900.76
LLaVA-Next [37]0.490.270.210.540.220.210.240.55
InternLM [19]0.480.270.290.580.250.320.260.53
InternLM-VL [19]0.410.200.170.470.220.200.200.53
SPHINX [35]0.560.280.320.590.240.330.270.58
LLaMA-3.2 [43]0.530.290.330.570.280.360.290.56
Pixtral [6]0.440.350.330.510.490.720.630.57
Qwen-2.5 [11]0.580.620.310.500.530.460.790.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.453, + 0.905, + 0.495 + ], + "angle": 0, + "content": "Table 3. Decomposed Capability Analysis. Results better than \\(70\\%\\) are bolded. Each task consists of test with different difficulties. Please see Appendix F for the results for different difficulties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.905, + 0.59 + ], + "angle": 0, + "content": "for an agent to successfully perform visual spatial planning. We show these examples in Appendix E. Next, we evaluate these two abilities through the remaining tasks T1-T4. Similar to previous settings, all the evaluation is conducted under zero-shot settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.803 + ], + "angle": 0, + "content": "Table 3 shows the decomposed capability results. Strong private models like Gemini-2.0 and Claude-3.7 show better performance across these tasks, demonstrating solid perception and reasoning capabilities, which explain their superior performance in the main tasks. However, they still exhibit some deficiencies, such as perceiving the location of a single object in Maze T1, highlighting a critical area for improvement. Furthermore, the performance of many open-source models is mostly close to random guessing on these tasks, indicating significant gaps compared to private models. One caveat is that while T4 focuses on reasoning capabilities, it still relies on the perception capabilities because the input still contains images. We perform further analysis to disentangle these two abilities in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.757, + 0.833 + ], + "angle": 0, + "content": "4.4. The Effects of Visual Input" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Previous analysis shows that even strong MLLMs have clear deficiencies in various aspects of visual spatial planning. In this study, we focus on disentangling the effects of perception and reasoning by exploring the performance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2276" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.516, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.098, + 0.898, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.224, + 0.9, + 0.239 + ], + "angle": 0, + "content": "Figure 6. Result with the visual/textual input. When environment is described by text instead of image, the performance increases vastly." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.253, + 0.481, + 0.368 + ], + "angle": 0, + "content": "
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4MainT1T2T3T4Main
Gemini, 0-shot0.580.560.330.490.170.860.510.540.550.03
Gemini, 1-shot0.500.660.310.480.200.910.680.710.590.03
Gemini, 2-shot0.530.680.310.510.210.900.760.700.610.03
Gemini, 4-shot0.530.670.350.530.190.910.640.690.620.06
GPT-4V, 0-shot0.560.270.460.560.260.730.800.700.710.10
GPT-4V, 1-shot0.550.500.470.570.280.890.840.940.730.11
GPT-4V, 2-shot0.550.630.500.560.300.900.830.950.710.16
GPT-4V, 4-shot0.540.690.540.560.290.900.790.960.73-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.375, + 0.435, + 0.388 + ], + "angle": 0, + "content": "Table 4. Effects of providing in-context examples." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.389, + 0.481, + 0.458 + ], + "angle": 0, + "content": "
ModelSettingMAZE NAVIGATIONBLOCKS OF WORLD
T1T2T3T4MainT1T2T3T4Main
LLaVAzero-shot0.490.270.210.540.050.220.210.240.550.01
fine-tune0.530.990.510.930.601.001.001.001.000.97
InternLMzero-shot0.480.270.290.580.110.250.320.260.530.00
fine-tune0.520.590.910.590.170.290.440.690.620.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.465, + 0.442, + 0.479 + ], + "angle": 0, + "content": "Table 5. Fine-tuning results for open-source models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.414, + 0.508 + ], + "angle": 0, + "content": "gain assuming the model had perfect perception." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.688 + ], + "angle": 0, + "content": "The key strategy here is to create a scenario where the model has acquired all necessary information that would typically be obtained through visual perception. To this end, for every input image, we produce the corresponding textual inputs and replace those images, as shown in Figure 5. For the Maze Navigation scenario, we use either pure text descriptions or tables to depict the image. For the Blocks World scenario, we use pure text descriptions. We do not use tables for the Blocks World scenario because the number of blocks in each horizontal stack is usually unequal, making it difficult to form a complete table. Appendix B includes complete examples with pure text or table input." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.688, + 0.483, + 0.81 + ], + "angle": 0, + "content": "The results are shown in Figure 6. We observe a clear performance improvement when using textual input across every task. This suggests image perception presents significant challenges for MLLMs, and poor perception ability is a key factor in the inferior performance observed in previous tasks. Meanwhile, note that even with textual input, Gemini still cannot achieve decent performance on tasks that require reasoning. This indicates its deficiencies in reasoning." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.819, + 0.483, + 0.835 + ], + "angle": 0, + "content": "4.5. In-context Learning in Visual Spatial Planning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In-context learning is a widely-adopted method to enhance LM's reasoning ability [14]. In this analysis, we study if it boosts visual spatial planning capabilities. We included varying numbers of examples for Gemini-1.0 and GPT-4V" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.905, + 0.362 + ], + "angle": 0, + "content": "(show in Appendix C) and show results in Table 4. There are two key observations: First, in-context examples make some potential contributions, but they are not significant. The examples only benefits in several sparse cases, such as T2 in Maze Navigation and T3 in Blocks World. Second, scaling examples generally does not help, as illustrated by the saturated performance in each task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.37, + 0.746, + 0.387 + ], + "angle": 0, + "content": "4.6. Fine-tuning in VSP Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Finally, we assess the capabilities of the open-source model through dedicated training. We performed LoRA fine-tuning on llava-v1.6-vicuna-7b and internlm-xcomposer2-7b. The models are trained on 10k data points (image-text pairs) for each task and scenario. We use the default hyperparameters provided in the official repository. More fine-tuning details can be found in Appendix D. The results, shown in Table 5, demonstrate clear performance improvements for both models across a series of tasks, highlighting their potential in spatial planning. Additionally, we observe that LLaVA shows greater improvement compared to InternLM, suggesting that different model architectures may exhibit varying levels of efficacy in spatial planning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.616, + 0.634, + 0.632 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We present VSP, a benchmark measuring and diagnosing the visual spatial planning capabilities in MLLMs. VSP quantifies the model's performance through a series of carefully designed tasks, with main tasks testing general spatial planning abilities and sub-tasks diagnosing individual capabilities needed for the main task. We show that current models fail to generate effective plans for even simple spatial planning tasks, and further analyses expose their bottlenecks in spatial perception and reasoning abilities. Our work illuminates future directions for improving MLLMs' abilities in spatial planning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.685, + 0.838 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.906, + 0.887 + ], + "angle": 0, + "content": "Qiucheng Wu and Shiyu Chang acknowledge the support from National Science Foundation (NSF) Grant IIS-2338252, Grant IIS-2207052, and Grant IIS-2302730." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2277" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6, 7, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins Sri, Anthony Barrett, Dave Christianson, et al. Pddll the planning domain definition language. Technical Report, Tech. Rep., 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning. arXiv preprint arXiv:2310.03249, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[5] Anthropic AI. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 6, 7, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[6] Mistral AI. Announcing pixtral 12b. https://mistral.ai/news/pixtral-12b/, 2024. Accessed: 2024-09-27. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[7] Open AI. Gpt-4o. https://openai.com/index/hello-gpt-4o/.6,7,23" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[8] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[9] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025.6,7,23" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[10] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[13] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[15] Sungguk Cha, Jusung Lee, Younghyun Lee, and Cheoljong Yang. Visually dehallucinative instruction generation: Know what you don't know. arXiv preprint arXiv:2402.09717, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[16] Haonan Chang, Kai Gao, Kowndinya Boyalakuntla, Alex Lee, Baichuan Huang, Harish Udhaya Kumar, Jinjin Yu, and Abdeslam Boularias. Lgmcts: Language-guided montecarlo tree search for executable semantic object rearrangement. arXiv preprint arXiv:2309.15821, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[17] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Google Deepmind. Introducing gemini 2.0: our new ai model for the agentic era. https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/, 2024.6,7,23" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[20] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[21] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[22] Wentao Ge, Shunian Chen, Guiming Chen, Junying Chen, Zhihong Chen, Shuo Yan, Chenghao Zhu, Ziyue Lin, Wenya Xie, Xidong Wang, et al. Mllm-bench, evaluating multimodal llms using gpt-4v. arXiv preprint arXiv:2311.13951, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[23] Deepanway Ghosal, Vernon Toh Yan Han, Chia Yew Ken, and Soujanya Poria. Are language models puzzle prodigies? algorithmic puzzles unveil serious challenges in multimodal reasoning. arXiv preprint arXiv:2403.03864, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[24] Tejas Gokhale, Shailaja Sampat, Zhiyuan Fang, Yezhou Yang, and Chitta Baral. Blockworld revisited: Learning and reasoning to generate event-sequences from image pairs. arXiv preprint arXiv:1905.12042, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Lin Guan, Karthik Valmeekam, Sarath Sreedharan, and Subbarao Kambhampati. Leveraging pre-trained large language models to construct and utilize world models for model-based task planning. Advances in Neural Information Processing Systems, 36:79081-79094, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Xiaoxiao Guo, Satinder Singh, Honglak Lee, Richard L Lewis, and Xiaoshi Wang. Deep learning for real-time" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2278" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "atari game play using offline monte-carlo tree search planning. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[27] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[28] Rishi Hazra, Pedro Zuidberg Dos Martires, and Luc De Raedt. Saycanpay: Heuristic planning with large language models using learnable domain knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 20123-20133, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[30] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[31] Frank Joublin, Antonello Ceravola, Pavel Smirnov, Felix Ocker, Joerg Deigmoeller, Anna Belardinelli, Chao Wang, Stephan Hasler, Daniel Tanneberg, and Michael Gienger. Copal: Corrective planning of robot actions with large language models. arXiv preprint arXiv:2310.07263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[32] Subbarao Kambhampati. Can large language models reason and plan? Annals of the New York Academy of Sciences, 1534(1):15-18, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[33] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Kaya Stechly, Mudit Verma, Siddhant Bhambri, Lucas Saldyt, and Anil Murthy. Llms can't plan, but can help planning in llm-modulo frameworks. arXiv preprint arXiv:2402.01817, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[34] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[36] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Yifan Xu, Xixuan Song, Shudan Zhang, Hanyu Lai, Xinyi Liu, Hanlin Zhao, et al. Visualagentbench: Towards large multimodal models as visual foundation agents. arXiv preprint arXiv:2408.06327, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. arXiv e-prints, pages arXiv-2310, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[40] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[41] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[42] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[43] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[44] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[45] Zhangyang Qi, Ye Fang, Mengchen Zhang, Zeyi Sun, Tong Wu, Ziwei Liu, Dahua Lin, Jiaqi Wang, and Hengshuang Zhao. Gemini vs gpt-4v: A preliminary comparison and combination of vision-language models through qualitative cases. arXiv preprint arXiv:2312.15011, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[46] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Guoqing Du, Shiwei Shi, Hangyu Mao, Xingyu Zeng, and Rui Zhao. Tptu: Task planning and tool usage of large language model-based ai agents. arXiv preprint arXiv:2308.03427, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[47] Md Sadman Sakib and Yu Sun. From cooking recipes to robot task trees-improving planning correctness and task efficiency by leveraging llms with a knowledge network. arXiv preprint arXiv:2309.09181, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[48] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[49] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2279" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[52] Kaya Stechly, Karthik Valmeekam, and Subbarao Kambhampati. On the self-verification limitations of large language models on reasoning and planning tasks. arXiv preprint arXiv:2402.08115, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[53] Richard S Sutton. Planning by incremental dynamic programming. In Machine learning proceedings 1991, pages 353-357. Elsevier, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 6, 7, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[55] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[56] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[58] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large language models still can't plan (a benchmark for llms on planning and reasoning about change). arXiv preprint arXiv:2206.10498, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[59] Karthik Valmeekam, Matthew Marquez, Sarath Sreedharan, and Subbarao Kambhampati. On the planning abilities of large language models - a critical investigation. In Advances in Neural Information Processing Systems, pages 75993-76005. Curran Associates, Inc., 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[60] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[61] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[62] Yue Wu, Xuan Tang, Tom M Mitchell, and Yuanzhi Li. Smartplay: A benchmark for llms as intelligent agents. arXiv preprint arXiv:2310.01557, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[63] Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. Travelplanner: A benchmark for real-world planning with language agents. arXiv preprint arXiv:2402.01622, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[64] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[65] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[66] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[67] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[68] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, et al. Mmt-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask agi. arXiv preprint arXiv:2404.16006, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[69] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.906, + 0.514 + ], + "angle": 0, + "content": "[70] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.515, + 0.906, + 0.598 + ], + "angle": 0, + "content": "[71] Wenqi Zhang, Zhenglin Cheng, Yuanyu He, Mengna Wang, Yongliang Shen, Zeqi Tan, Guiyang Hou, Mingqian He, Yanna Ma, Weiming Lu, et al. Multimodal self-instruct: Synthetic abstract image and visual reasoning instruction using language model. arXiv preprint arXiv:2407.07053, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.906, + 0.656 + ], + "angle": 0, + "content": "[72] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2280" + } + ] +] \ No newline at end of file diff --git a/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_origin.pdf b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e48b1f472f388045e41ea00b1d7f84465c370680 --- /dev/null +++ b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/262f63b9-5c82-43f4-b645-d16941a910c1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76bd997d191c39f2e7aaa5ed61707a43066c589fb470abc2feaa38501f0ec831 +size 1435111 diff --git a/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/full.md b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e025ea3af327b112fb64449dcbc9670b1d8217e4 --- /dev/null +++ b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/full.md @@ -0,0 +1,352 @@ +# VSP: Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs + +Qiucheng Wu $^{1}$ , Handong Zhao $^{2}$ , Michael Saxon $^{1}$ , Trung Bui $^{2}$ , William Yang Wang $^{1}$ , Yang Zhang $^{3}$ , Shiyu Chang $^{1}$ $^{1}$ UC Santa Barbara, $^{2}$ Adobe Research, $^{3}$ MIT-IBM Watson AI Lab + +qiucheng@ucsb.edu + +# Abstract + +Multimodal large language models are an exciting emerging class of language models (LMs) that have merged classic LM capabilities with those of image processing systems. However, how these capabilities integrate is often not intuitive and warrants direct investigation. One understudied capability in MLLMs is visual spatial planning—the ability to comprehend the spatial arrangements of objects and devise action plans to achieve desired outcomes in visual scenes. It is unclear why MLLMs fall short on these tasks generally considered easy for humans, given their successes across other diverse scenarios. To this end, we introduce VSP, a benchmark that 1) evaluates the spatial planning capability in MLLMs in general, and 2) diagnoses this capability via finer-grained sub-tasks, including perception and reasoning, and measure the capabilities of models through these sub-tasks. Our evaluation confirms that both open-source and private MLLMs fail to generate effective plans for even simple spatial planning tasks. Evaluations on the fine-grained analytical tasks further reveal fundamental deficiencies in the models' visual perception and bottle-necks in reasoning abilities, explaining their worse performance in the general spatial planning tasks. Our work illuminates future directions for improving MLLMs' abilities in spatial planning. Our benchmark is publicly available1. + +# 1. Introduction + +The rapid advancement of large language models has driven considerable growth in their capabilities to produce fluent text in many domains, generating outputs exhibiting potential "reasoning" and "understanding" abilities [12, 14, 30, 57]. Recently, multimodal large language models (MLLMs) have advanced on LLMs through training on native image inputs, to achieve impressive performance generating text describing and relating to input images [1, 8, 10, + +![](images/4f215dad01c20b127c0ef840cea7cb79c943182576656799b9ef8815b0fb2116.jpg) +Figure 1. Overview of the VSP benchmark. The benchmark evaluates tasks that require MLLMs to perform complex reasoning by utilizing the spatial information they perceive, and diagnoses their failure modes using a series of carefully designed subtasks. + +37, 54], with applications in image captioning, visual question answering, and visual reasoning [48, 64, 68, 72]. The swift evolution of MLLMs has enabled them to tackle increasingly sophisticated tasks that require multiple emerging abilities in complex scenarios. However, as model capabilities and deployment needs advance, the challenges in usefully evaluating them grow in kind. + +Planning is a fundamental capability in intelligent systems that is particularly contested in LMs [59]. Visual spatial planning refers to the task of comprehending the spatial arrangement of objects in a scene and designing action plans to achieve a desired outcome. For example, the classical maze problem can be considered a visual planning task, where an agent is given an input image describing the maze environment and is asked to produce a viable path to navigate the player from the starting position to the goal. This task requires two capabilities: image perception, which en + +
BenchmarkCapabilitiesDiagnosing Incapabilities
Visual RecognitionSpatial Relation UnderstandingPlan with Spatial Info
MMMU [70]✓XXX
MathVision [60]✓XXX
MME [21]✓✓XX
SeedBench [34]✓✓XX
MM-Vet [69]✓✓XX
AlgoPuzzle [23]✓✓✓X
MMSI [71]✓✓✓X
VSP✓✓✓✓
+ +Table 1. Comparison with representative existing benchmarks. Note that we focus on spatial reasoning related capabilities. ✓ indicates that the measurement is covered by the corresponding benchmark, while $\times$ indicates not covered. + +ables the agent to understand the objects, environment and spatial relations present in the image, and reasoning, which enables the agent to perform strategic decision-making. + +Visual spatial planning is an important capability in many potential applications for MLLMs, such as navigating in complex environments with autonomous driving [40, 55] or manipulating objects with robotic hands [16, 29]. However, although there have been increasingly more benchmarks to evaluate the vision processing capabilities of MLLMs, few systematically investigate the underlying factors affecting their (in)ability to perform visual spatial planning tasks. As shown in Table 1, while most of the existing benchmarks identify the capability to recognize objects as a key feature of MLLM, quite a few benchmarks [60, 70] do not evaluate their ability to understand spatial relations, and even fewer benchmarks [23, 71] examine the ability to use spatial information in complex tasks, e.g., navigation on a road map based on the position towards goals. More importantly, it is often unclear why MLLMs perform so poorly on the spatial planning tasks that are generally considered easy for humans [23]. As a result, two research questions emerged: ① [Evaluation] How performant are MLLMs in performing visual planning tasks? ② [Diagnose] What are the bottleneck capabilities, e.g., perception or reasoning, that limit the performance of MLLMs in visual planning tasks? + +To this end, we introduce Visual Spatial Planning (VSP), a benchmark specifically designed to evaluate and diagnose the spatial planning capabilities of MLLMs. As illustrated in Figure 1, the VSP benchmark consists of various environment settings designed to assess the capabilities of models in different scenarios. We focus on abstract scenarios to facilitate the design of diagnostic subtasks with consistent environments (Sec. 3.5). The first two scenarios, Maze Navigation and Blocks World, are basic environments developed from classical planning tasks. Additionally, we challenge the model's abilities in dynamic and realistic applications through the Collision and Google Map scenarios, respectively. All environments are fully observable through input + +images. The MLLMs are required to interpret the visual inputs, deduce the consequences of each action, and execute the designated tasks. To comprehensively evaluate the fine-grained capabilities needed for visual spatial planning, VSP includes 4.6K questions in 12 meticulously designed tasks that feature both simulated and photorealistic visual inputs. In addition to testing end-to-end spatial planning, these tasks test essential individual visual planning abilities, such as image perception and reasoning. + +We apply the VSP benchmark to evaluate existing state-of-the-art MLLMs, including both open-source and private ones. Surprisingly, we find that even the most competitive MLLMs sometimes struggle in performing the simplest visual planning tasks, such as a 3x3 maze problem. Our fine-grained capability analysis further reveals that existing MLLMs have flaws in reasoning and bigger bottlenecks in perception. We believe the VSP benchmark highlights critical weaknesses in current MLLMs and sheds light on future directions for enhancing their spatial understanding and planning capabilities. + +# 2. Related Work + +# 2.1. General Planning in LMs + +Planning has been a central focus of research in AI. Traditional work in AI planning includes using formal languages to represent and solve planning problems [2], and developing algorithms like dynamic programming and reinforcement learning to explore environments and formulate viable plans [26, 53]. While these works mostly focus on predefined and restricted environments, recently, with the advancement of LMs, it has become intriguing to study whether LMs, with the potential to be general intelligent agents, can perform planning in different settings and environments [32, 33, 52]. Many works explore the best ways to activate the planning capabilities of LMs, including divide and conquer [49, 61, 66, 67], grounding outputs in admissible actions [4, 28], retrospecting and refining [41, 50], and leveraging external tools [25, 46]. Meanwhile, with the increasing capabilities of LMs, growing research efforts are now dedicated to benchmark their planning capabilities in various complex environments [58, 62, 63]. + +# 2.2. Spatial and Visual Planning in LMs + +Many planning tasks in LMs involve understanding visual environments and comprehending spatial information. In embodied agent studies, LMs play a crucial role in grounding visual entities with references in open-domain instructions and formulating plans based on spatial constraints. Consequently, they are increasingly used in physically grounded scenarios such as object rearrangement [16, 29, 38], cooking [31, 47], navigation [4, 28, 42], etc. LMs are also used in AIGC to propose spatial arrangements of + +![](images/433397f7e27764598c6d5c20b30321565b82e3befbe1bd5ffd9ab3798bdba5a6.jpg) +Figure 2. Overview of the Maze Navigation scenario. + +entities following instructions [20]. While realistic planning tasks align with real needs, their complexity and expansive action spaces limit the analysis of LMs' detailed planning capabilities. Therefore, research also focuses on LMs' planning in abstract environments and games. For example, mystery blockworld is a dynamically generated set of blockworld tasks to test generalization in LMs [59]. Additionally, many text games have been introduced to test LMs' abilities in spatial understanding and imagination [3, 51, 62, 65]. While many of these studies transform visual information into text inputs, recent works also explore direct understanding and processing of visual inputs leveraging MLLMs [23, 71]. + +# 2.3. Benchmarks for MLLMs + +MLLMs have inherited and advanced many intriguing features from text-only LMs [45, 65]. Benchmarks for MLLMs have rapidly emerged to evaluate performance in areas such as image content understanding [15, 21], perception [22, 56], knowledge [39, 60, 70], and reasoning [21, 36, 70]. While these benchmarks quantify MLLMs' abilities in many fields, their capabilities in spatial understanding and reaction are relatively under-explored. Some benchmarks cover spatial relations understanding [34, 69], but often overlook the ability to devise complex spatial action plans based on visual environment constraints. Recently, several benchmarks explore the capability of planning utilizing spatial information, such QA and navigation in virtual or realistic scenarios [17, 23, 44, 71]. While measurements on these benchmarks reveal some incapability of MLLMs, it is often unclear why they cannot perform well in these tasks. In this work, we focus on evaluating and diagnosing the ability to comprehend spatial arrangements of objects and devise action plans to achieve specific outcomes. Our benchmark highlights future directions for improving MLLMs towards models with general intelligence. + +# 3. The Visual Spatial Planning Benchmark + +# 3.1. Overview of the Benchmark + +In this benchmark, our objectives are two-fold: ① quantify the visual spatial planning capabilities of MLLMs; and ② + +uncover current capability bottlenecks that limit the effectiveness of MLLMs in visual spatial planning tasks. The first objective requires broader task designs. Specifically, the tasks should range from classical planning tasks [13, 58] to ones in those more dynamic and realistic environments. On top of that, the second objective requires deeper task designs. In particular, performing spatial planning in visual environments requires a series of cohesive steps. For example, to generate an accurate path to navigate a player to a goal, an agent needs to be able to correctly view and understand the visual map, reason to find which actions are safe or dangerous, and come up with a detailed plan to achieve the goal. Each of these steps could be challenging for a developing MLLM, and understanding which of these subtasks challenge them most will drive future improvement. + +To this end, we propose the Visual Spatial Planning (VSP) benchmark, with the objective of measuring and diagnosing the capabilities of MLLMs in producing accurate spatial plans in visual environments. VSP consists of four scenarios: ① the simulated Maze Navigation scenario, whose main task is to move a game character through a maze; ② the photo-realistic Blocks World scenario, whose main task is to move blocks from a starting configuration to a goal configuration; ③ the dynamic Collision scenario, whose main task is to determine if there is a danger of collision between two objects in the environment, and ④ the realistic Google Map scenario, whose main task is to find a path in real streets of New York City. In addition to the main task, VSP introduces four sub-tasks that focus on the individual capabilities needed for the main task: + +- T1. Single Object Perception - Determine the characteristics of a single object; +- T2. Spatial Relation Perception - Determine the relative positions of two objects; +- T3. Environment Perception – Find textual descriptions that describe the visual environment; +- T4. Spatial Reasoning - Determine the consequence of a series of actions or moves. + +The sub-task details are designed specific to both scenarios. Furthermore, to demonstrate the model's performance under different levels of environmental complexity, we establish progressive difficulty settings for various tasks, + +![](images/be5de3cc364b071b2f1cd8222a21a9b02ae981edd5bf736d4edb1a490980a7ae.jpg) +Figure 3. Overview of the Blocks World scenario. + +which are measured by parameters such as map size, minimum required number of actions, etc. The detailed task statistics are provided in appendix A. In what follows, we introduce each scenario in detail, as well as the data curation and the task creation processes. + +# 3.2. The Maze Navigation Scenario + +The Maze Navigation scenario is inspired by the popular implementation [13] of a fully observable path-finding problem. As depicted in Figure 2 left, it simulates a classical grid world environment with a designated start and goal position, where part of the grids contain obstacles (the "holes") and cannot be passed through. + +The main spatial planning task and the four sub-tasks are defined as follows: + +- Main Task (Spatial Planning) - Generate a safe path to navigate from the start grid to the goal; +- T1 (Single Object Perception) - Determine if a specified grid is safe; +- T2 (Spatial Relation Perception) - Find spatial relations between the player and the goal; +- T3 (Environment Perception) - Find the textual description that fits the visual environment; +- T4 (Spatial Reasoning) - Determine the consequence of a given action series. + +An example of input image and questions is demonstrated in Figure 2. Each task is equipped with progressive adjusted difficulty settings to comprehensively evaluate the model's capability. For Main Task and T1-T3, the difficulties are measured by the size of the map, ranging from 3x3 to 8x8, where a larger map introduces more challenges in correctly perceiving objects and planning accordingly. For task T4, since a longer path naturally introduces more challenges for reasoning, we adopt path length ranging from 1 to 9 as the difficulty measure. Please see Appendix A for the complete example of questions and answers in each task. + +# 3.3. The Blocks World Scenario + +The Blocks World is a widely-adopted planning problem [24, 27, 58]. As depicted in Figure 3 left, in this sce + +nario, the agent is given images containing sets of blocks in unique colors. These blocks are stacked vertically, forming multiple stacks on the table. The agent is asked to turn the blocks from initial state to target state through a series of moving actions. For each action, the agent can only move the top block of any stack, providing it is moved to either the table or the top of another stack. + +Similarly, the main spatial planning task and the four sub-tasks are defined as follows: + +- Main Task (Spatial Planning) - Form a moving plan to achieve the target state of block arrangement; +- T1 (Single Object Perception) - Determine the color of the block at a specific position; +- T2 (Spatial Relation Perception) - Determine the spatial relation between two blocks; +- T3 (Environment Perception) – Find the text representation that fits the visual environment; +- T4 (Spatial Reasoning) - Determine the consequence of a given moving plan. + +An example of input image and questions is demonstrated in Figure 3. Similar to Maze Navigation, each task is equipped with progressive adjusted difficulty. Specifically, in Main Task and T4, the difficulties are measured by the number of actions involved, ranging from 1 to 7, which quantifies the complexity of the action plan. On the other hand, for tasks T1-T3, which focus on perception, the difficulty is measured by the number of blocks presented in the image, ranging from 3 to 5. Please refer to Appendix A for complete examples of the question and answer in each task. + +# 3.4. The Collision and Google Map Scenario + +We further experiment in a more dynamic (the collision scenario) and realistic (the Google map scenario) settings to explore the capabilities of models in more challenging cases. The input of these two scenarios are shown in the right panels of Figure 1. Please refer to Appendix A for the complete example of each scenario. + +- Collision Scenario In this scenario, the input map is similar to Maze Navigation scenario. However, in this case, + +![](images/0b8e90779a670b9efe65d3fd1610a87e7b171c7feafd6215a52e4d00dd6164f7.jpg) +Figure 4. Benchmark creation process. Left: We prepare input images that fulfill the task requirements with different difficulties. Mid: We formulate input prompts consisting of interleaved texts and images. Right: We develop automatic evaluation process for each task. + +... The following figure shows how the player, the holes, the lands, and the goals look like. + + + +![](images/69e588ee4af1a93fba9dace713eba7e66cb75ac4fdb5545f8bc1ce9c9af99707.jpg) + +... We provide an example major illustrate the rules. $<$ IMAGE-2>In this provided example: +- The player is at ... +- The goal is at ... + +![](images/ee9c20dc749b4755bd600632c54cb021f23c2ed8b73dcfb5e9c13b5705d4fc29.jpg) + +# Formulate Prompts + +... We provide an example map to further illustrate the rules. + + +This example features four blocks arranged in three stacks: +- Stack 1: Purple block (alone) +- Stack 2: Blue block (alone) +- Stack 3: From bottom to top: Orange block, Red block ... + +# Auto Evaluation + +move(red, table) + +move(red, yellow) + +the player is moving in an environment where a car is also present and moving. Given the moving information (speed, direction) of the player and the car, the model needs to determine the time the player and the car reaches the goal, and then determine if there is a collision danger here. + +- Google Map Scenario In this scenario, the input image is a real Google map depicting the streets and avenues in New York City. The starting location and the goal are randomly chosen crossroads in the map and are marked on the image. The goal of the model is to find a path from the starting location to the goal. The model needs to output a path described by directions (north, east, west, south) and the number of blocks to traverse (e.g., head north for 2 blocks, then head east for 3 blocks). + +# 3.5. Benchmark Creation and Quality Control + +In Figure 4, we demonstrate the general process for benchmark creation and quality control. We use the Blocks World and Maze Navigation scenarios as examples. + +First, in the left panel of Figure 4, we prepare the input images used for each task and scenario. In the Maze Navigation scenario, we generate input maps using the OpenAI Gym package [13], with modifications to ensure that the positions of the player, the goal, and the holes are all randomly generated. In the Blocks World scenario, we sample pairs of images from the BIRD dataset [24], ensuring there is at least one viable plan to move the blocks from the initial state to the target state. The images are prepared conditional on different levels of difficulty. + +Second, as shown in the center of Figure 4, we formulate input prompts for each task. The prompt consists of interleaved text and images to provide sufficient information. For example, for Maze Navigation, we include images + +to show the appearance of elements in the map and provide example maps to better illustrate how the models should interpret the map. We invite native speakers to refine the prompts so that they accurately describe the task requirements. The prompts are in Appendix A. + +Finally, in the right panel of Figure 4, we evaluate the performance of MLLMs under each task. It is worth noting that the answer for each task is often not unique. For example, in the Blocks World scenario, there can be many ways to move the blocks to reach the target state. As such, we develop scripts to accurately evaluate each individual answer. + +In addition to the steps above, some tasks require extra steps to ensure meaningful questions, candidates, and answers. For example, in T4 of the Blocks World scenario, the input actions must cover various valid/invalid movements. The detailed steps we followed to create each task set are provided in Appendix A. We release all images, texts, and scripts to facilitate replication and scaling. + +# 4. Experiments + +In this section, we present evaluation results of state-of-the-art MLLMs under our main tasks and sub-tasks. Our goal is to answer the following research questions: 1 How well can state-of-the-art MLLMs perform in the visual spatial planning tasks? 2 What are the bottleneck capabilities that limit the MLLMs in visual spatial planning tasks? + +# 4.1. Baselines + +We evaluate various representative MLLMs including both private and open-source models. + +For private models, we include their different versions to observe how their abilities involved and what is the + +
Difficulty levelMAZE NAVIGATIONBLOCKS WORLDCOLLISIONGOO MAPOVERALL
3456781357
Gemini-1.0 [54]0.310.260.150.060.140.100.100.140.000.010.130.000.1167
Gemini-2.0 [18]0.930.810.620.410.300.270.850.490.180.070.440.070.4533
Claude-3 [5]0.520.330.160.150.160.090.120.030.000.000.180.020.1467
Claude-3.7 [9]0.710.560.560.270.180.280.890.570.330.440.490.110.4492
GPT-4V [1]0.550.360.270.130.170.100.500.170.030.000.240.020.2117
GPT-4o [7]0.680.580.350.240.180.230.710.330.120.030.160.040.3042
LLaVA-Next [37]0.030.030.020.080.090.040.040.010.000.000.020.000.0300
InternLM [19]0.270.160.060.050.040.070.100.030.000.000.250.000.0858
InternLM-VL [19]0.150.140.080.040.020.050.020.000.000.000.220.010.0600
SPHINX [35]0.110.080.050.020.040.030.070.060.010.000.040.000.0425
LLaMA-3.2 [43]0.230.180.160.080.080.100.050.000.000.000.140.000.0850
Pixtral [6]0.320.200.170.100.060.060.210.110.010.000.220.030.1242
Qwen-2.5 [11]0.450.330.200.090.150.080.060.030.000.010.450.060.1592
+ +Table 2. Zero-shot success rates for the spatial planning task at various difficulty levels. Maze Navigation difficulty levels represent the maze's square grid length. Blocks World difficulty levels correspond to minimum steps to a solution. Results better than $30\%$ are bolded. + +state-of-the-art performance. We cover the following private models: ① Gemini [18, 54] has demonstrated remarkable capabilities in image understanding and reasoning. We adopt Gemini-1.0 (Gemini-1.0-pro-vision) and Gemini-2.0 (Gemini-2.0-flash) in our experiments. ② Claude [5, 9] is a family of MLLMs strong at advanced reasoning and vision analysis. We adopt Claude-3 (claude-3-sonnet-20240229) and Claude-3.7 (claude-3-7-sonnet-20250219) in the experiments. ③ GPT [1, 7] inherent strong text understanding capabilities from text-only models and is equipped with vision capabilities. We use GPT-4V (turbo-2024-04-09) and GPT-4o (gpt-4o-2024-05-13) for evaluation. + +Besides, we cover the following open-source models: ④ LLaVA-Next [37] demonstrates strong reasoning abilities comparable to private models. We adopt LLAVA-v1.6-VICUNA-7B for evaluation. ⑤ InternLM-XComposer2 [19] shows strong ability to understand free-form text-image composition. We adopt internlm-xcomposer2-7b and internlm-xcomposer2-v1-7b, with the former focusing on general text-image composition and the latter focusing on VL benchmarks. ⑥ SPHINX [35] unfreezes the LM during pre-training to enhance cross-model alignment. We adopt SPHINX-v2-1k. ⑦ LLama-3.2 [43] exceeds on several image understanding tasks. We adopt llama-3.2-90B-Vision in our experiments. ⑧ Pixtral [6] is a recent MLLM that demonstrates strong performance across a series of multimodal tasks. We adopt pixtral-12b-2409. ⑨ Qwen [11] demonstrates strong object localization and visual recognition in several vision-language tasks. We adopt Qwen2.5-VL-7B-Instruct. We use the public released checkpoints and codes. + +# 4.2. Main Task (Spatial Planning) Evaluation + +First, we present the main task results for the four scenarios, which reflect the general spatial planning capabilities of + +MLLMs. All the evaluation in this section is conducted under zero-shot setting. Evaluations with in-context learning and fine-tuning are presented in Sections 4.5 and 4.6. + +The performance is demonstrated in Table 2. We also present difficulty levels in the table, which is measured by the size of the map (3 represents $3 \times 3$ maps) in Maze Navigation and by the minimum number of steps in Blocks World. From the table, we summarize our findings as follows: + +VSP highlights both the advancements and the challenges in MLLMs' spatial planning capabilities. Notably, private MLLMs demonstrate a clear trajectory of improvement, with earlier versions performing worse and later iterations showing advancements—though still exhibiting deficiencies. In fact, even some of the most capable models frequently make mistakes on tasks such as navigating a $7 \times 7$ map or executing 3-step block-moving tasks, which are often trivial for humans. Meanwhile, open-source models face even greater difficulties and rarely succeed in these tasks. Overall, VSP demonstrates that while MLLMs have made progress in spatial planning, they still exhibit considerable room for improvement. + +MLLMs face significant difficulties with spatial planning in dynamic and realistic environments. Based on the experimental results, most models perform worse when tested in dynamic (Collision) and realistic (Google Map) settings. We identify two major reasons: First, the tasks in Collision are complex and typically require multiple capabilities. For example, to assess collision danger, the model must locate both the player and the car on the map, calculate the time needed for the player to reach the goal, and determine if the car will hit the player during that time. This poses significant challenges for the models. Second, the inputs in the Google Map are intricate and contain a series of irrelevant symbols, making it difficult for the model to accurately interpret the map. The models' performance in these environments suggests that current models are not yet + +![](images/66827513d759600fc796ebe01d54406d81d167c9d5bf66883c09fcfdab23f362.jpg) +Vision Input + +![](images/44640806becebfc13ad3a4e82b604c489be384bb5698b7638f50b55ceb968ecf.jpg) +Pure Text Input + +![](images/d787593d60533ef6c60e0074180fad41ade975b6bd0c23a4e09116e51b4d68db.jpg) +Maze Navigation Scenario +Table Input + +![](images/6e4b76d406dfdbea0ef3290da8471b2a3c1d4f60c12b458f672df71d7915be16.jpg) +Blocks World Scenario +Vision Input +Pure Text Input +Figure 5. The visual and corresponding textual inputs. + +- Stack with purple block +- Stack with blue block +- Stack with orange block, red block, from bottom to top + +equipped to handle spatial reasoning in such complex environments. Consequently, we focus on the Maze Navigation and Blocks World scenarios in the following experiments to diagnose the models' hidden weaknesses in VSP tasks. + +Quick performance decay as difficulty increases. We observe a significant drop in the success rates of MLLMs as task difficulty escalates. For example, GPT-4V may achieve a success rate of over $50\%$ on 3x3 size maps, but this plummets to just $10\%$ on 8x8 maps. Analyzing the impact of increased difficulty, we identify two major challenges for the models: First, increasing size of the map in Maze Navigation scenario could make it difficult for the model to accurately perceive the positions of elements within the map. Second, the increase in both map size and the number of steps required for moving blocks heightens the challenge for the model to reason deeply through the entire path and devise a complete, viable solution. In the following experiments, we focus on these two factors and provide in-depth diagnoses with subsequent tasks. + +Challenges in open-source models. Finally, we note that open-source models often face difficulties on these tasks. As will be analyzed in Section 4.3, their perception and reasoning capabilities present greater challenges. Besides, we note two other factors: ① Context length: Open-source models typically have shorter context windows compared to private models. Thus, these models may not have enough capacity to understand the complete inputs. For example, LLAVA-V1.6-VICUNA-7B is trained with a context window of 2048 tokens, while each image consumes 576 tokens. As such, when fed with multiple images and long texts in our tasks, the total token length may surpass training, resulting in poor performance. ② Multiple image input: Our tasks require the model to understand multiple images interleaved with text inputs, whereas some open-source models are only trained with single-image inputs, with the image positioned at the start of the input. To further explore their potential, we assess their performance after finetuning in Section 4.6. Also, we note that recent open-source MLLMs often avoid these two issues, which may contribute to the improved performance observed in Table 2. + +# 4.3. Diagnose the Perception and Reasoning + +From the previous observation, we identify that spatial perception and reasoning could be two important capabilities + +
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4T1T2T3T4
Random Guess0.50.250.250.50.170.250.250.5
Gemini-1.0 [54]0.580.560.330.490.860.510.540.55
Gemini-2.0 [18]0.660.910.960.800.950.960.990.76
Claude-3 [5]0.450.670.320.610.430.530.490.66
Claude-3.7 [9]0.720.880.970.790.940.840.990.94
GPT-4V [1]0.560.270.460.560.730.800.700.71
GPT-4o [7]0.580.670.580.740.950.900.900.76
LLaVA-Next [37]0.490.270.210.540.220.210.240.55
InternLM [19]0.480.270.290.580.250.320.260.53
InternLM-VL [19]0.410.200.170.470.220.200.200.53
SPHINX [35]0.560.280.320.590.240.330.270.58
LLaMA-3.2 [43]0.530.290.330.570.280.360.290.56
Pixtral [6]0.440.350.330.510.490.720.630.57
Qwen-2.5 [11]0.580.620.310.500.530.460.790.56
+ +Table 3. Decomposed Capability Analysis. Results better than $70\%$ are bolded. Each task consists of test with different difficulties. Please see Appendix F for the results for different difficulties. + +for an agent to successfully perform visual spatial planning. We show these examples in Appendix E. Next, we evaluate these two abilities through the remaining tasks T1-T4. Similar to previous settings, all the evaluation is conducted under zero-shot settings. + +Table 3 shows the decomposed capability results. Strong private models like Gemini-2.0 and Claude-3.7 show better performance across these tasks, demonstrating solid perception and reasoning capabilities, which explain their superior performance in the main tasks. However, they still exhibit some deficiencies, such as perceiving the location of a single object in Maze T1, highlighting a critical area for improvement. Furthermore, the performance of many open-source models is mostly close to random guessing on these tasks, indicating significant gaps compared to private models. One caveat is that while T4 focuses on reasoning capabilities, it still relies on the perception capabilities because the input still contains images. We perform further analysis to disentangle these two abilities in Section 4.4. + +# 4.4. The Effects of Visual Input + +Previous analysis shows that even strong MLLMs have clear deficiencies in various aspects of visual spatial planning. In this study, we focus on disentangling the effects of perception and reasoning by exploring the performance + +![](images/58a85caf84a5c335f70c5375d837680ff78bfab939266940a3c79f97ccf83c0f.jpg) +Figure 6. Result with the visual/textual input. When environment is described by text instead of image, the performance increases vastly. + +![](images/9085b0ec41a60e3f4c12d2ede371886d1a75cec48526512f821eb7919bb1318e.jpg) + +
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4MainT1T2T3T4Main
Gemini, 0-shot0.580.560.330.490.170.860.510.540.550.03
Gemini, 1-shot0.500.660.310.480.200.910.680.710.590.03
Gemini, 2-shot0.530.680.310.510.210.900.760.700.610.03
Gemini, 4-shot0.530.670.350.530.190.910.640.690.620.06
GPT-4V, 0-shot0.560.270.460.560.260.730.800.700.710.10
GPT-4V, 1-shot0.550.500.470.570.280.890.840.940.730.11
GPT-4V, 2-shot0.550.630.500.560.300.900.830.950.710.16
GPT-4V, 4-shot0.540.690.540.560.290.900.790.960.73-
+ +Table 4. Effects of providing in-context examples. + +
ModelSettingMAZE NAVIGATIONBLOCKS OF WORLD
T1T2T3T4MainT1T2T3T4Main
LLaVAzero-shot0.490.270.210.540.050.220.210.240.550.01
fine-tune0.530.990.510.930.601.001.001.001.000.97
InternLMzero-shot0.480.270.290.580.110.250.320.260.530.00
fine-tune0.520.590.910.590.170.290.440.690.620.09
+ +Table 5. Fine-tuning results for open-source models. + +gain assuming the model had perfect perception. + +The key strategy here is to create a scenario where the model has acquired all necessary information that would typically be obtained through visual perception. To this end, for every input image, we produce the corresponding textual inputs and replace those images, as shown in Figure 5. For the Maze Navigation scenario, we use either pure text descriptions or tables to depict the image. For the Blocks World scenario, we use pure text descriptions. We do not use tables for the Blocks World scenario because the number of blocks in each horizontal stack is usually unequal, making it difficult to form a complete table. Appendix B includes complete examples with pure text or table input. + +The results are shown in Figure 6. We observe a clear performance improvement when using textual input across every task. This suggests image perception presents significant challenges for MLLMs, and poor perception ability is a key factor in the inferior performance observed in previous tasks. Meanwhile, note that even with textual input, Gemini still cannot achieve decent performance on tasks that require reasoning. This indicates its deficiencies in reasoning. + +# 4.5. In-context Learning in Visual Spatial Planning + +In-context learning is a widely-adopted method to enhance LM's reasoning ability [14]. In this analysis, we study if it boosts visual spatial planning capabilities. We included varying numbers of examples for Gemini-1.0 and GPT-4V + +(show in Appendix C) and show results in Table 4. There are two key observations: First, in-context examples make some potential contributions, but they are not significant. The examples only benefits in several sparse cases, such as T2 in Maze Navigation and T3 in Blocks World. Second, scaling examples generally does not help, as illustrated by the saturated performance in each task. + +# 4.6. Fine-tuning in VSP Tasks + +Finally, we assess the capabilities of the open-source model through dedicated training. We performed LoRA fine-tuning on llava-v1.6-vicuna-7b and internlm-xcomposer2-7b. The models are trained on 10k data points (image-text pairs) for each task and scenario. We use the default hyperparameters provided in the official repository. More fine-tuning details can be found in Appendix D. The results, shown in Table 5, demonstrate clear performance improvements for both models across a series of tasks, highlighting their potential in spatial planning. Additionally, we observe that LLaVA shows greater improvement compared to InternLM, suggesting that different model architectures may exhibit varying levels of efficacy in spatial planning capabilities. + +# 5. Conclusion + +We present VSP, a benchmark measuring and diagnosing the visual spatial planning capabilities in MLLMs. VSP quantifies the model's performance through a series of carefully designed tasks, with main tasks testing general spatial planning abilities and sub-tasks diagnosing individual capabilities needed for the main task. We show that current models fail to generate effective plans for even simple spatial planning tasks, and further analyses expose their bottlenecks in spatial perception and reasoning abilities. Our work illuminates future directions for improving MLLMs' abilities in spatial planning. + +# 6. Acknowledgment + +Qiucheng Wu and Shiyu Chang acknowledge the support from National Science Foundation (NSF) Grant IIS-2338252, Grant IIS-2207052, and Grant IIS-2302730. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6, 7, 23 +[2] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins Sri, Anthony Barrett, Dave Christianson, et al. Pddll the planning domain definition language. Technical Report, Tech. Rep., 1998. 2 +[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning. arXiv preprint arXiv:2310.03249, 2023. 3 +[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2 +[5] Anthropic AI. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 6, 7, 23 +[6] Mistral AI. Announcing pixtral 12b. https://mistral.ai/news/pixtral-12b/, 2024. Accessed: 2024-09-27. 6, 7 +[7] Open AI. Gpt-4o. https://openai.com/index/hello-gpt-4o/.6,7,23 +[8] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1 +[9] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025.6,7,23 +[10] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1 +[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 6, 7 +[12] Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. 1 +[13] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 3, 4, 5 +[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan + +tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8 +[15] Sungguk Cha, Jusung Lee, Younghyun Lee, and Cheoljong Yang. Visually dehallucinative instruction generation: Know what you don't know. arXiv preprint arXiv:2402.09717, 2024.3 +[16] Haonan Chang, Kai Gao, Kowndinya Boyalakuntla, Alex Lee, Baichuan Huang, Harish Udhaya Kumar, Jinjin Yu, and Abdeslam Boularias. Lgmcts: Language-guided montecarlo tree search for executable semantic object rearrangement. arXiv preprint arXiv:2309.15821, 2023. 2 +[17] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 3 +[18] Google Deepmind. Introducing gemini 2.0: our new ai model for the agentic era. https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/, 2024.6,7,23 +[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6, 7 +[20] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[21] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2, 3 +[22] Wentao Ge, Shunian Chen, Guiming Chen, Junying Chen, Zhihong Chen, Shuo Yan, Chenghao Zhu, Ziyue Lin, Wenya Xie, Xidong Wang, et al. Mllm-bench, evaluating multimodal llms using gpt-4v. arXiv preprint arXiv:2311.13951, 2023. 3 +[23] Deepanway Ghosal, Vernon Toh Yan Han, Chia Yew Ken, and Soujanya Poria. Are language models puzzle prodigies? algorithmic puzzles unveil serious challenges in multimodal reasoning. arXiv preprint arXiv:2403.03864, 2024. 2, 3 +[24] Tejas Gokhale, Shailaja Sampat, Zhiyuan Fang, Yezhou Yang, and Chitta Baral. Blockworld revisited: Learning and reasoning to generate event-sequences from image pairs. arXiv preprint arXiv:1905.12042, 2019. 4, 5 +[25] Lin Guan, Karthik Valmeekam, Sarath Sreedharan, and Subbarao Kambhampati. Leveraging pre-trained large language models to construct and utilize world models for model-based task planning. Advances in Neural Information Processing Systems, 36:79081-79094, 2023. 2 +[26] Xiaoxiao Guo, Satinder Singh, Honglak Lee, Richard L Lewis, and Xiaoshi Wang. Deep learning for real-time + +atari game play using offline monte-carlo tree search planning. Advances in neural information processing systems, 27, 2014. 2 +[27] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992, 2023. 4 +[28] Rishi Hazra, Pedro Zuidberg Dos Martires, and Luc De Raedt. Saycanpay: Heuristic planning with large language models using learnable domain knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 20123-20133, 2024. 2 +[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 2 +[30] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 1 +[31] Frank Joublin, Antonello Ceravola, Pavel Smirnov, Felix Ocker, Joerg Deigmoeller, Anna Belardinelli, Chao Wang, Stephan Hasler, Daniel Tanneberg, and Michael Gienger. Copal: Corrective planning of robot actions with large language models. arXiv preprint arXiv:2310.07263, 2023. 2 +[32] Subbarao Kambhampati. Can large language models reason and plan? Annals of the New York Academy of Sciences, 1534(1):15-18, 2024. 2 +[33] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Kaya Stechly, Mudit Verma, Siddhant Bhambri, Lucas Saldyt, and Anil Murthy. Llms can't plan, but can help planning in llm-modulo frameworks. arXiv preprint arXiv:2402.01817, 2024. 2 +[34] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 3 +[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6, 7 +[36] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3 +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 6, 7 +[38] Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Yifan Xu, Xixuan Song, Shudan Zhang, Hanyu Lai, Xinyi Liu, Hanlin Zhao, et al. Visualagentbench: Towards large multimodal models as visual foundation agents. arXiv preprint arXiv:2408.06327, 2024. 2 + +[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. arXiv e-prints, pages arXiv-2310, 2023. 3 +[40] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2 +[41] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36, 2024. 2 +[42] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2 +[43] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 6, 7 +[44] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024.3 +[45] Zhangyang Qi, Ye Fang, Mengchen Zhang, Zeyi Sun, Tong Wu, Ziwei Liu, Dahua Lin, Jiaqi Wang, and Hengshuang Zhao. Gemini vs gpt-4v: A preliminary comparison and combination of vision-language models through qualitative cases. arXiv preprint arXiv:2312.15011, 2023. 3 +[46] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Guoqing Du, Shiwei Shi, Hangyu Mao, Xingyu Zeng, and Rui Zhao. Tptu: Task planning and tool usage of large language model-based ai agents. arXiv preprint arXiv:2308.03427, 2023. 2 +[47] Md Sadman Sakib and Yu Sun. From cooking recipes to robot task trees-improving planning correctness and task efficiency by leveraging llms with a knowledge network. arXiv preprint arXiv:2309.09181, 2023. 2 +[48] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 1 +[49] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36, 2024. 2 +[50] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36, 2024. 2 +[51] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. + +Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768, 2020. 3 +[52] Kaya Stechly, Karthik Valmeekam, and Subbarao Kambhampati. On the self-verification limitations of large language models on reasoning and planning tasks. arXiv preprint arXiv:2402.08115, 2024. 2 +[53] Richard S Sutton. Planning by incremental dynamic programming. In Machine learning proceedings 1991, pages 353-357. Elsevier, 1991. 2 +[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 6, 7, 23 +[55] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2 +[56] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209, 2024. 3 +[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[58] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large language models still can't plan (a benchmark for llms on planning and reasoning about change). arXiv preprint arXiv:2206.10498, 2022. 2, 3, 4 +[59] Karthik Valmeekam, Matthew Marquez, Sarath Sreedharan, and Subbarao Kambhampati. On the planning abilities of large language models - a critical investigation. In Advances in Neural Information Processing Systems, pages 75993-76005. Curran Associates, Inc., 2023. 1, 3 +[60] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2, 3 +[61] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2 +[62] Yue Wu, Xuan Tang, Tom M Mitchell, and Yuanzhi Li. Smartplay: A benchmark for llms as intelligent agents. arXiv preprint arXiv:2310.01557, 2023. 2, 3 +[63] Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. Travelplanner: A benchmark for real-world planning with language agents. arXiv preprint arXiv:2402.01622, 2024. 2 +[64] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 1 + +[65] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3 +[66] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2 +[67] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[68] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, et al. Mmt-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask agi. arXiv preprint arXiv:2404.16006, 2024. 1 +[69] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 3 +[70] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2, 3 +[71] Wenqi Zhang, Zhenglin Cheng, Yuanyu He, Mengna Wang, Yongliang Shen, Zeqi Tan, Guiyang Hou, Mingqian He, Yanna Ma, Weiming Lu, et al. Multimodal self-instruct: Synthetic abstract image and visual reasoning instruction using language model. arXiv preprint arXiv:2407.07053, 2024. 2, 3 +[72] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1 \ No newline at end of file diff --git a/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/images.zip b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3964edee9023399d2b2a53c27666af0f49ff30b1 --- /dev/null +++ b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbd4b962034f9113046e62e2f442d814be0eebc3245e368a603a9ae720b8afd6 +size 643154 diff --git a/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/layout.json b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..71adede7f38eb3295a91403ad9b37594b2e7e6d0 --- /dev/null +++ b/2025/VSP_ Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs/layout.json @@ -0,0 +1,8906 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 102, + 542, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 102, + 542, + 140 + ], + "spans": [ + { + "bbox": [ + 68, + 102, + 542, + 140 + ], + "type": "text", + "content": "VSP: Diagnosing the Dual Challenges of Perception and Reasoning in Spatial Planning Tasks for MLLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "spans": [ + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": "Qiucheng Wu" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": ", Handong Zhao" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": ", Michael Saxon" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": ", Trung Bui" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": ", William Yang Wang" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": ", Yang Zhang" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": ", Shiyu Chang" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": "UC Santa Barbara, " + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": "Adobe Research, " + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 160, + 465, + 203 + ], + "type": "text", + "content": "MIT-IBM Watson AI Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 205, + 352, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 205, + 352, + 217 + ], + "spans": [ + { + "bbox": [ + 258, + 205, + 352, + 217 + ], + "type": "text", + "content": "qiucheng@ucsb.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 559 + ], + "type": "text", + "content": "Multimodal large language models are an exciting emerging class of language models (LMs) that have merged classic LM capabilities with those of image processing systems. However, how these capabilities integrate is often not intuitive and warrants direct investigation. One understudied capability in MLLMs is visual spatial planning—the ability to comprehend the spatial arrangements of objects and devise action plans to achieve desired outcomes in visual scenes. It is unclear why MLLMs fall short on these tasks generally considered easy for humans, given their successes across other diverse scenarios. To this end, we introduce VSP, a benchmark that 1) evaluates the spatial planning capability in MLLMs in general, and 2) diagnoses this capability via finer-grained sub-tasks, including perception and reasoning, and measure the capabilities of models through these sub-tasks. Our evaluation confirms that both open-source and private MLLMs fail to generate effective plans for even simple spatial planning tasks. Evaluations on the fine-grained analytical tasks further reveal fundamental deficiencies in the models' visual perception and bottle-necks in reasoning abilities, explaining their worse performance in the general spatial planning tasks. Our work illuminates future directions for improving MLLMs' abilities in spatial planning. Our benchmark is publicly available1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "content": "The rapid advancement of large language models has driven considerable growth in their capabilities to produce fluent text in many domains, generating outputs exhibiting potential \"reasoning\" and \"understanding\" abilities [12, 14, 30, 57]. Recently, multimodal large language models (MLLMs) have advanced on LLMs through training on native image inputs, to achieve impressive performance generating text describing and relating to input images [1, 8, 10," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 244, + 553, + 446 + ], + "blocks": [ + { + "bbox": [ + 317, + 244, + 553, + 446 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 553, + 446 + ], + "type": "image", + "image_path": "4f215dad01c20b127c0ef840cea7cb79c943182576656799b9ef8815b0fb2116.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 450, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 495 + ], + "type": "text", + "content": "Figure 1. Overview of the VSP benchmark. The benchmark evaluates tasks that require MLLMs to perform complex reasoning by utilizing the spatial information they perceive, and diagnoses their failure modes using a series of carefully designed subtasks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 593 + ], + "type": "text", + "content": "37, 54], with applications in image captioning, visual question answering, and visual reasoning [48, 64, 68, 72]. The swift evolution of MLLMs has enabled them to tackle increasingly sophisticated tasks that require multiple emerging abilities in complex scenarios. However, as model capabilities and deployment needs advance, the challenges in usefully evaluating them grow in kind." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Planning is a fundamental capability in intelligent systems that is particularly contested in LMs [59]. Visual spatial planning refers to the task of comprehending the spatial arrangement of objects in a scene and designing action plans to achieve a desired outcome. For example, the classical maze problem can be considered a visual planning task, where an agent is given an input image describing the maze environment and is asked to produce a viable path to navigate the player from the starting position to the goal. This task requires two capabilities: image perception, which en" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1https://github.com/UCSB-NLP-Chang/Visual-Spatial-Planning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2270" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 63, + 302, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 63, + 302, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 63, + 302, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 63, + 302, + 190 + ], + "type": "table", + "html": "
BenchmarkCapabilitiesDiagnosing Incapabilities
Visual RecognitionSpatial Relation UnderstandingPlan with Spatial Info
MMMU [70]✓XXX
MathVision [60]✓XXX
MME [21]✓✓XX
SeedBench [34]✓✓XX
MM-Vet [69]✓✓XX
AlgoPuzzle [23]✓✓✓X
MMSI [71]✓✓✓X
VSP✓✓✓✓
", + "image_path": "11f1776cd3b00804b50cf1c8807686ccb5c46e89ac089ffe902d477d5f01bf3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 296, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 236 + ], + "type": "text", + "content": "Table 1. Comparison with representative existing benchmarks. Note that we focus on spatial reasoning related capabilities. ✓ indicates that the measurement is covered by the corresponding benchmark, while " + }, + { + "bbox": [ + 55, + 192, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 192, + 296, + 236 + ], + "type": "text", + "content": " indicates not covered." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 245, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 281 + ], + "type": "text", + "content": "ables the agent to understand the objects, environment and spatial relations present in the image, and reasoning, which enables the agent to perform strategic decision-making." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 283, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 557 + ], + "type": "text", + "content": "Visual spatial planning is an important capability in many potential applications for MLLMs, such as navigating in complex environments with autonomous driving [40, 55] or manipulating objects with robotic hands [16, 29]. However, although there have been increasingly more benchmarks to evaluate the vision processing capabilities of MLLMs, few systematically investigate the underlying factors affecting their (in)ability to perform visual spatial planning tasks. As shown in Table 1, while most of the existing benchmarks identify the capability to recognize objects as a key feature of MLLM, quite a few benchmarks [60, 70] do not evaluate their ability to understand spatial relations, and even fewer benchmarks [23, 71] examine the ability to use spatial information in complex tasks, e.g., navigation on a road map based on the position towards goals. More importantly, it is often unclear why MLLMs perform so poorly on the spatial planning tasks that are generally considered easy for humans [23]. As a result, two research questions emerged: ① [Evaluation] How performant are MLLMs in performing visual planning tasks? ② [Diagnose] What are the bottleneck capabilities, e.g., perception or reasoning, that limit the performance of MLLMs in visual planning tasks?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "To this end, we introduce Visual Spatial Planning (VSP), a benchmark specifically designed to evaluate and diagnose the spatial planning capabilities of MLLMs. As illustrated in Figure 1, the VSP benchmark consists of various environment settings designed to assess the capabilities of models in different scenarios. We focus on abstract scenarios to facilitate the design of diagnostic subtasks with consistent environments (Sec. 3.5). The first two scenarios, Maze Navigation and Blocks World, are basic environments developed from classical planning tasks. Additionally, we challenge the model's abilities in dynamic and realistic applications through the Collision and Google Map scenarios, respectively. All environments are fully observable through input" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "images. The MLLMs are required to interpret the visual inputs, deduce the consequences of each action, and execute the designated tasks. To comprehensively evaluate the fine-grained capabilities needed for visual spatial planning, VSP includes 4.6K questions in 12 meticulously designed tasks that feature both simulated and photorealistic visual inputs. In addition to testing end-to-end spatial planning, these tasks test essential individual visual planning abilities, such as image perception and reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "content": "We apply the VSP benchmark to evaluate existing state-of-the-art MLLMs, including both open-source and private ones. Surprisingly, we find that even the most competitive MLLMs sometimes struggle in performing the simplest visual planning tasks, such as a 3x3 maze problem. Our fine-grained capability analysis further reveals that existing MLLMs have flaws in reasoning and bigger bottlenecks in perception. We believe the VSP benchmark highlights critical weaknesses in current MLLMs and sheds light on future directions for enhancing their spatial understanding and planning capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 324, + 400, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 400, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 400, + 337 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 344, + 456, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 456, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 456, + 357 + ], + "type": "text", + "content": "2.1. General Planning in LMs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 362, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 579 + ], + "type": "text", + "content": "Planning has been a central focus of research in AI. Traditional work in AI planning includes using formal languages to represent and solve planning problems [2], and developing algorithms like dynamic programming and reinforcement learning to explore environments and formulate viable plans [26, 53]. While these works mostly focus on predefined and restricted environments, recently, with the advancement of LMs, it has become intriguing to study whether LMs, with the potential to be general intelligent agents, can perform planning in different settings and environments [32, 33, 52]. Many works explore the best ways to activate the planning capabilities of LMs, including divide and conquer [49, 61, 66, 67], grounding outputs in admissible actions [4, 28], retrospecting and refining [41, 50], and leveraging external tools [25, 46]. Meanwhile, with the increasing capabilities of LMs, growing research efforts are now dedicated to benchmark their planning capabilities in various complex environments [58, 62, 63]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 587, + 504, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 504, + 601 + ], + "type": "text", + "content": "2.2. Spatial and Visual Planning in LMs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Many planning tasks in LMs involve understanding visual environments and comprehending spatial information. In embodied agent studies, LMs play a crucial role in grounding visual entities with references in open-domain instructions and formulating plans based on spatial constraints. Consequently, they are increasingly used in physically grounded scenarios such as object rearrangement [16, 29, 38], cooking [31, 47], navigation [4, 28, 42], etc. LMs are also used in AIGC to propose spatial arrangements of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "2271" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 58, + 496, + 187 + ], + "blocks": [ + { + "bbox": [ + 116, + 58, + 496, + 187 + ], + "lines": [ + { + "bbox": [ + 116, + 58, + 496, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 58, + 496, + 187 + ], + "type": "image", + "image_path": "433397f7e27764598c6d5c20b30321565b82e3befbe1bd5ffd9ab3798bdba5a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 192, + 402, + 203 + ], + "lines": [ + { + "bbox": [ + 207, + 192, + 402, + 203 + ], + "spans": [ + { + "bbox": [ + 207, + 192, + 402, + 203 + ], + "type": "text", + "content": "Figure 2. Overview of the Maze Navigation scenario." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 213, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 369 + ], + "type": "text", + "content": "entities following instructions [20]. While realistic planning tasks align with real needs, their complexity and expansive action spaces limit the analysis of LMs' detailed planning capabilities. Therefore, research also focuses on LMs' planning in abstract environments and games. For example, mystery blockworld is a dynamically generated set of blockworld tasks to test generalization in LMs [59]. Additionally, many text games have been introduced to test LMs' abilities in spatial understanding and imagination [3, 51, 62, 65]. While many of these studies transform visual information into text inputs, recent works also explore direct understanding and processing of visual inputs leveraging MLLMs [23, 71]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 196, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 196, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 196, + 386 + ], + "type": "text", + "content": "2.3. Benchmarks for MLLMs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 644 + ], + "type": "text", + "content": "MLLMs have inherited and advanced many intriguing features from text-only LMs [45, 65]. Benchmarks for MLLMs have rapidly emerged to evaluate performance in areas such as image content understanding [15, 21], perception [22, 56], knowledge [39, 60, 70], and reasoning [21, 36, 70]. While these benchmarks quantify MLLMs' abilities in many fields, their capabilities in spatial understanding and reaction are relatively under-explored. Some benchmarks cover spatial relations understanding [34, 69], but often overlook the ability to devise complex spatial action plans based on visual environment constraints. Recently, several benchmarks explore the capability of planning utilizing spatial information, such QA and navigation in virtual or realistic scenarios [17, 23, 44, 71]. While measurements on these benchmarks reveal some incapability of MLLMs, it is often unclear why they cannot perform well in these tasks. In this work, we focus on evaluating and diagnosing the ability to comprehend spatial arrangements of objects and devise action plans to achieve specific outcomes. Our benchmark highlights future directions for improving MLLMs towards models with general intelligence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 652, + 277, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 277, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 277, + 666 + ], + "type": "text", + "content": "3. The Visual Spatial Planning Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 211, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 211, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 211, + 684 + ], + "type": "text", + "content": "3.1. Overview of the Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In this benchmark, our objectives are two-fold: ① quantify the visual spatial planning capabilities of MLLMs; and ②" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 392 + ], + "type": "text", + "content": "uncover current capability bottlenecks that limit the effectiveness of MLLMs in visual spatial planning tasks. The first objective requires broader task designs. Specifically, the tasks should range from classical planning tasks [13, 58] to ones in those more dynamic and realistic environments. On top of that, the second objective requires deeper task designs. In particular, performing spatial planning in visual environments requires a series of cohesive steps. For example, to generate an accurate path to navigate a player to a goal, an agent needs to be able to correctly view and understand the visual map, reason to find which actions are safe or dangerous, and come up with a detailed plan to achieve the goal. Each of these steps could be challenging for a developing MLLM, and understanding which of these subtasks challenge them most will drive future improvement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 393, + 556, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 556, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 556, + 571 + ], + "type": "text", + "content": "To this end, we propose the Visual Spatial Planning (VSP) benchmark, with the objective of measuring and diagnosing the capabilities of MLLMs in producing accurate spatial plans in visual environments. VSP consists of four scenarios: ① the simulated Maze Navigation scenario, whose main task is to move a game character through a maze; ② the photo-realistic Blocks World scenario, whose main task is to move blocks from a starting configuration to a goal configuration; ③ the dynamic Collision scenario, whose main task is to determine if there is a danger of collision between two objects in the environment, and ④ the realistic Google Map scenario, whose main task is to find a path in real streets of New York City. In addition to the main task, VSP introduces four sub-tasks that focus on the individual capabilities needed for the main task:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 573, + 553, + 664 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 573, + 553, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 553, + 596 + ], + "type": "text", + "content": "- T1. Single Object Perception - Determine the characteristics of a single object;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 596, + 553, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 553, + 619 + ], + "type": "text", + "content": "- T2. Spatial Relation Perception - Determine the relative positions of two objects;" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 619, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 553, + 641 + ], + "type": "text", + "content": "- T3. Environment Perception – Find textual descriptions that describe the visual environment;" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 664 + ], + "type": "text", + "content": "- T4. Spatial Reasoning - Determine the consequence of a series of actions or moves." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "The sub-task details are designed specific to both scenarios. Furthermore, to demonstrate the model's performance under different levels of environmental complexity, we establish progressive difficulty settings for various tasks," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2272" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 72, + 495, + 208 + ], + "blocks": [ + { + "bbox": [ + 116, + 72, + 495, + 208 + ], + "lines": [ + { + "bbox": [ + 116, + 72, + 495, + 208 + ], + "spans": [ + { + "bbox": [ + 116, + 72, + 495, + 208 + ], + "type": "image", + "image_path": "be5de3cc364b071b2f1cd8222a21a9b02ae981edd5bf736d4edb1a490980a7ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 216, + 394, + 226 + ], + "lines": [ + { + "bbox": [ + 214, + 216, + 394, + 226 + ], + "spans": [ + { + "bbox": [ + 214, + 216, + 394, + 226 + ], + "type": "text", + "content": "Figure 3. Overview of the Blocks World scenario." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 237, + 296, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 298 + ], + "type": "text", + "content": "which are measured by parameters such as map size, minimum required number of actions, etc. The detailed task statistics are provided in appendix A. In what follows, we introduce each scenario in detail, as well as the data curation and the task creation processes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 222, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 222, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 222, + 317 + ], + "type": "text", + "content": "3.2. The Maze Navigation Scenario" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 394 + ], + "type": "text", + "content": "The Maze Navigation scenario is inspired by the popular implementation [13] of a fully observable path-finding problem. As depicted in Figure 2 left, it simulates a classical grid world environment with a designated start and goal position, where part of the grids contain obstacles (the \"holes\") and cannot be passed through." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 395, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 295, + 417 + ], + "type": "text", + "content": "The main spatial planning task and the four sub-tasks are defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 418, + 295, + 532 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 441 + ], + "type": "text", + "content": "- Main Task (Spatial Planning) - Generate a safe path to navigate from the start grid to the goal;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "type": "text", + "content": "- T1 (Single Object Perception) - Determine if a specified grid is safe;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "type": "text", + "content": "- T2 (Spatial Relation Perception) - Find spatial relations between the player and the goal;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 487, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 509 + ], + "type": "text", + "content": "- T3 (Environment Perception) - Find the textual description that fits the visual environment;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 532 + ], + "type": "text", + "content": "- T4 (Spatial Reasoning) - Determine the consequence of a given action series." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": "An example of input image and questions is demonstrated in Figure 2. Each task is equipped with progressive adjusted difficulty settings to comprehensively evaluate the model's capability. For Main Task and T1-T3, the difficulties are measured by the size of the map, ranging from 3x3 to 8x8, where a larger map introduces more challenges in correctly perceiving objects and planning accordingly. For task T4, since a longer path naturally introduces more challenges for reasoning, we adopt path length ranging from 1 to 9 as the difficulty measure. Please see Appendix A for the complete example of questions and answers in each task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 206, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 206, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 206, + 683 + ], + "type": "text", + "content": "3.3. The Blocks World Scenario" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The Blocks World is a widely-adopted planning problem [24, 27, 58]. As depicted in Figure 3 left, in this sce" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 237, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 320 + ], + "type": "text", + "content": "nario, the agent is given images containing sets of blocks in unique colors. These blocks are stacked vertically, forming multiple stacks on the table. The agent is asked to turn the blocks from initial state to target state through a series of moving actions. For each action, the agent can only move the top block of any stack, providing it is moved to either the table or the top of another stack." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 321, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 344 + ], + "type": "text", + "content": "Similarly, the main spatial planning task and the four sub-tasks are defined as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 346, + 553, + 460 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 313, + 346, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 553, + 368 + ], + "type": "text", + "content": "- Main Task (Spatial Planning) - Form a moving plan to achieve the target state of block arrangement;" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 369, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 392 + ], + "type": "text", + "content": "- T1 (Single Object Perception) - Determine the color of the block at a specific position;" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 392, + 553, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 553, + 413 + ], + "type": "text", + "content": "- T2 (Spatial Relation Perception) - Determine the spatial relation between two blocks;" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 414, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 553, + 436 + ], + "type": "text", + "content": "- T3 (Environment Perception) – Find the text representation that fits the visual environment;" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 437, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 460 + ], + "type": "text", + "content": "- T4 (Spatial Reasoning) - Determine the consequence of a given moving plan." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 460, + 555, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 580 + ], + "type": "text", + "content": "An example of input image and questions is demonstrated in Figure 3. Similar to Maze Navigation, each task is equipped with progressive adjusted difficulty. Specifically, in Main Task and T4, the difficulties are measured by the number of actions involved, ranging from 1 to 7, which quantifies the complexity of the action plan. On the other hand, for tasks T1-T3, which focus on perception, the difficulty is measured by the number of blocks presented in the image, ranging from 3 to 5. Please refer to Appendix A for complete examples of the question and answer in each task." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 600, + 524, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 524, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 524, + 613 + ], + "type": "text", + "content": "3.4. The Collision and Google Map Scenario" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "content": "We further experiment in a more dynamic (the collision scenario) and realistic (the Google map scenario) settings to explore the capabilities of models in more challenging cases. The input of these two scenarios are shown in the right panels of Figure 1. Please refer to Appendix A for the complete example of each scenario." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "- Collision Scenario In this scenario, the input map is similar to Maze Navigation scenario. However, in this case," + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2273" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 72, + 254, + 270 + ], + "blocks": [ + { + "bbox": [ + 91, + 72, + 254, + 270 + ], + "lines": [ + { + "bbox": [ + 91, + 72, + 254, + 270 + ], + "spans": [ + { + "bbox": [ + 91, + 72, + 254, + 270 + ], + "type": "image", + "image_path": "0b8e90779a670b9efe65d3fd1610a87e7b171c7feafd6215a52e4d00dd6164f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "text", + "content": "Figure 4. Benchmark creation process. Left: We prepare input images that fulfill the task requirements with different difficulties. Mid: We formulate input prompts consisting of interleaved texts and images. Right: We develop automatic evaluation process for each task." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 256, + 76, + 394, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 76, + 394, + 92 + ], + "spans": [ + { + "bbox": [ + 256, + 76, + 394, + 92 + ], + "type": "text", + "content": "... The following figure shows how the player, the holes, the lands, and the goals look like." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 257, + 92, + 294, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 92, + 294, + 99 + ], + "spans": [ + { + "bbox": [ + 257, + 92, + 294, + 99 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 340, + 92, + 390, + 112 + ], + "blocks": [ + { + "bbox": [ + 340, + 92, + 390, + 112 + ], + "lines": [ + { + "bbox": [ + 340, + 92, + 390, + 112 + ], + "spans": [ + { + "bbox": [ + 340, + 92, + 390, + 112 + ], + "type": "image", + "image_path": "69e588ee4af1a93fba9dace713eba7e66cb75ac4fdb5545f8bc1ce9c9af99707.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 257, + 112, + 352, + 152 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 257, + 112, + 352, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 112, + 352, + 137 + ], + "spans": [ + { + "bbox": [ + 257, + 112, + 352, + 137 + ], + "type": "text", + "content": "... We provide an example major illustrate the rules. " + }, + { + "bbox": [ + 257, + 112, + 352, + 137 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 257, + 112, + 352, + 137 + ], + "type": "text", + "content": " IMAGE-2>In this provided example:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 137, + 321, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 137, + 321, + 152 + ], + "spans": [ + { + "bbox": [ + 257, + 137, + 321, + 152 + ], + "type": "text", + "content": "- The player is at ... \n- The goal is at ..." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 394, + 76, + 519, + 270 + ], + "blocks": [ + { + "bbox": [ + 394, + 76, + 519, + 270 + ], + "lines": [ + { + "bbox": [ + 394, + 76, + 519, + 270 + ], + "spans": [ + { + "bbox": [ + 394, + 76, + 519, + 270 + ], + "type": "image", + "image_path": "ee9c20dc749b4755bd600632c54cb021f23c2ed8b73dcfb5e9c13b5705d4fc29.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 280, + 162, + 359, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 162, + 359, + 173 + ], + "spans": [ + { + "bbox": [ + 280, + 162, + 359, + 173 + ], + "type": "text", + "content": "Formulate Prompts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 257, + 182, + 383, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 182, + 383, + 198 + ], + "spans": [ + { + "bbox": [ + 257, + 182, + 383, + 198 + ], + "type": "text", + "content": "... We provide an example map to further illustrate the rules." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 257, + 201, + 392, + 261 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 257, + 201, + 290, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 201, + 290, + 209 + ], + "spans": [ + { + "bbox": [ + 257, + 201, + 290, + 209 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 257, + 214, + 353, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 214, + 353, + 230 + ], + "spans": [ + { + "bbox": [ + 257, + 214, + 353, + 230 + ], + "type": "text", + "content": "This example features four blocks arranged in three stacks:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 257, + 230, + 348, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 230, + 348, + 237 + ], + "spans": [ + { + "bbox": [ + 257, + 230, + 348, + 237 + ], + "type": "text", + "content": "- Stack 1: Purple block (alone)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 257, + 238, + 343, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 238, + 343, + 246 + ], + "spans": [ + { + "bbox": [ + 257, + 238, + 343, + 246 + ], + "type": "text", + "content": "- Stack 2: Blue block (alone)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 257, + 246, + 392, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 246, + 392, + 261 + ], + "spans": [ + { + "bbox": [ + 257, + 246, + 392, + 261 + ], + "type": "text", + "content": "- Stack 3: From bottom to top: Orange block, Red block ..." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 418, + 162, + 485, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 162, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 418, + 162, + 485, + 171 + ], + "type": "text", + "content": "Auto Evaluation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 418, + 180, + 468, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 180, + 468, + 188 + ], + "spans": [ + { + "bbox": [ + 418, + 180, + 468, + 188 + ], + "type": "text", + "content": "move(red, table)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 403, + 190, + 511, + 237 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 190, + 511, + 237 + ], + "spans": [ + { + "bbox": [ + 403, + 190, + 511, + 237 + ], + "type": "image", + "image_path": "0038e02a4e33061497381a3440a906df86cf492734c52ea25350d5f0e46ff437.jpg" + } + ] + } + ], + "index": 18, + "type": "text" + }, + { + "bbox": [ + 417, + 246, + 471, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 417, + 246, + 471, + 254 + ], + "spans": [ + { + "bbox": [ + 417, + 246, + 471, + 254 + ], + "type": "text", + "content": "move(red, yellow)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 488, + 179, + 506, + 194 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 179, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 488, + 179, + 506, + 194 + ], + "type": "image", + "image_path": "ab78ab3a92126b9334c8ea6c17593c83507269ddf13bbbcbd1ba231e8723b4c1.jpg" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "bbox": [ + 488, + 196, + 511, + 236 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 196, + 511, + 236 + ], + "spans": [ + { + "bbox": [ + 488, + 196, + 511, + 236 + ], + "type": "image", + "image_path": "c4e68bd8834dbb39ee1068cae783866cc253e91a1731879a5a613c602b2f688a.jpg" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "bbox": [ + 488, + 240, + 506, + 257 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 240, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 488, + 240, + 506, + 257 + ], + "type": "image", + "image_path": "d54d3ff34c3ff6faf05a86f59a0ef16aa2e73b136d8be5b7e016e99430454084.jpg" + } + ] + } + ], + "index": 22, + "type": "text" + }, + { + "bbox": [ + 55, + 314, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 373 + ], + "type": "text", + "content": "the player is moving in an environment where a car is also present and moving. Given the moving information (speed, direction) of the player and the car, the model needs to determine the time the player and the car reaches the goal, and then determine if there is a collision danger here." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 375, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 482 + ], + "type": "text", + "content": "- Google Map Scenario In this scenario, the input image is a real Google map depicting the streets and avenues in New York City. The starting location and the goal are randomly chosen crossroads in the map and are marked on the image. The goal of the model is to find a path from the starting location to the goal. The model needs to output a path described by directions (north, east, west, south) and the number of blocks to traverse (e.g., head north for 2 blocks, then head east for 3 blocks)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 490, + 275, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 275, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 275, + 502 + ], + "type": "text", + "content": "3.5. Benchmark Creation and Quality Control" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 508, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 544 + ], + "type": "text", + "content": "In Figure 4, we demonstrate the general process for benchmark creation and quality control. We use the Blocks World and Maze Navigation scenarios as examples." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "type": "text", + "content": "First, in the left panel of Figure 4, we prepare the input images used for each task and scenario. In the Maze Navigation scenario, we generate input maps using the OpenAI Gym package [13], with modifications to ensure that the positions of the player, the goal, and the holes are all randomly generated. In the Blocks World scenario, we sample pairs of images from the BIRD dataset [24], ensuring there is at least one viable plan to move the blocks from the initial state to the target state. The images are prepared conditional on different levels of difficulty." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Second, as shown in the center of Figure 4, we formulate input prompts for each task. The prompt consists of interleaved text and images to provide sufficient information. For example, for Maze Navigation, we include images" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": "to show the appearance of elements in the map and provide example maps to better illustrate how the models should interpret the map. We invite native speakers to refine the prompts so that they accurately describe the task requirements. The prompts are in Appendix A." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 375, + 554, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 446 + ], + "type": "text", + "content": "Finally, in the right panel of Figure 4, we evaluate the performance of MLLMs under each task. It is worth noting that the answer for each task is often not unique. For example, in the Blocks World scenario, there can be many ways to move the blocks to reach the target state. As such, we develop scripts to accurately evaluate each individual answer." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 447, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 532 + ], + "type": "text", + "content": "In addition to the steps above, some tasks require extra steps to ensure meaningful questions, candidates, and answers. For example, in T4 of the Blocks World scenario, the input actions must cover various valid/invalid movements. The detailed steps we followed to create each task set are provided in Appendix A. We release all images, texts, and scripts to facilitate replication and scaling." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 544, + 395, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 395, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 395, + 558 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 565, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 554, + 637 + ], + "type": "text", + "content": "In this section, we present evaluation results of state-of-the-art MLLMs under our main tasks and sub-tasks. Our goal is to answer the following research questions: 1 How well can state-of-the-art MLLMs perform in the visual spatial planning tasks? 2 What are the bottleneck capabilities that limit the MLLMs in visual spatial planning tasks?" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 647, + 381, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 381, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 381, + 658 + ], + "type": "text", + "content": "4.1. Baselines" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 665, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 689 + ], + "type": "text", + "content": "We evaluate various representative MLLMs including both private and open-source models." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "For private models, we include their different versions to observe how their abilities involved and what is the" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2274" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 70, + 520, + 225 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 520, + 225 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 520, + 225 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 520, + 225 + ], + "type": "table", + "html": "
Difficulty levelMAZE NAVIGATIONBLOCKS WORLDCOLLISIONGOO MAPOVERALL
3456781357
Gemini-1.0 [54]0.310.260.150.060.140.100.100.140.000.010.130.000.1167
Gemini-2.0 [18]0.930.810.620.410.300.270.850.490.180.070.440.070.4533
Claude-3 [5]0.520.330.160.150.160.090.120.030.000.000.180.020.1467
Claude-3.7 [9]0.710.560.560.270.180.280.890.570.330.440.490.110.4492
GPT-4V [1]0.550.360.270.130.170.100.500.170.030.000.240.020.2117
GPT-4o [7]0.680.580.350.240.180.230.710.330.120.030.160.040.3042
LLaVA-Next [37]0.030.030.020.080.090.040.040.010.000.000.020.000.0300
InternLM [19]0.270.160.060.050.040.070.100.030.000.000.250.000.0858
InternLM-VL [19]0.150.140.080.040.020.050.020.000.000.000.220.010.0600
SPHINX [35]0.110.080.050.020.040.030.070.060.010.000.040.000.0425
LLaMA-3.2 [43]0.230.180.160.080.080.100.050.000.000.000.140.000.0850
Pixtral [6]0.320.200.170.100.060.060.210.110.010.000.220.030.1242
Qwen-2.5 [11]0.450.330.200.090.150.080.060.030.000.010.450.060.1592
", + "image_path": "da34fe18f1c9e2f982ff3dcc488e23445ffc9b4e8462b6d0dbbb0d17d7b95e7e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "type": "text", + "content": "Table 2. Zero-shot success rates for the spatial planning task at various difficulty levels. Maze Navigation difficulty levels represent the maze's square grid length. Blocks World difficulty levels correspond to minimum steps to a solution. Results better than " + }, + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 234, + 555, + 257 + ], + "type": "text", + "content": " are bolded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 277, + 294, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 294, + 432 + ], + "type": "text", + "content": "state-of-the-art performance. We cover the following private models: ① Gemini [18, 54] has demonstrated remarkable capabilities in image understanding and reasoning. We adopt Gemini-1.0 (Gemini-1.0-pro-vision) and Gemini-2.0 (Gemini-2.0-flash) in our experiments. ② Claude [5, 9] is a family of MLLMs strong at advanced reasoning and vision analysis. We adopt Claude-3 (claude-3-sonnet-20240229) and Claude-3.7 (claude-3-7-sonnet-20250219) in the experiments. ③ GPT [1, 7] inherent strong text understanding capabilities from text-only models and is equipped with vision capabilities. We use GPT-4V (turbo-2024-04-09) and GPT-4o (gpt-4o-2024-05-13) for evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 434, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 661 + ], + "type": "text", + "content": "Besides, we cover the following open-source models: ④ LLaVA-Next [37] demonstrates strong reasoning abilities comparable to private models. We adopt LLAVA-v1.6-VICUNA-7B for evaluation. ⑤ InternLM-XComposer2 [19] shows strong ability to understand free-form text-image composition. We adopt internlm-xcomposer2-7b and internlm-xcomposer2-v1-7b, with the former focusing on general text-image composition and the latter focusing on VL benchmarks. ⑥ SPHINX [35] unfreezes the LM during pre-training to enhance cross-model alignment. We adopt SPHINX-v2-1k. ⑦ LLama-3.2 [43] exceeds on several image understanding tasks. We adopt llama-3.2-90B-Vision in our experiments. ⑧ Pixtral [6] is a recent MLLM that demonstrates strong performance across a series of multimodal tasks. We adopt pixtral-12b-2409. ⑨ Qwen [11] demonstrates strong object localization and visual recognition in several vision-language tasks. We adopt Qwen2.5-VL-7B-Instruct. We use the public released checkpoints and codes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 671, + 270, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 270, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 270, + 685 + ], + "type": "text", + "content": "4.2. Main Task (Spatial Planning) Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "First, we present the main task results for the four scenarios, which reflect the general spatial planning capabilities of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "content": "MLLMs. All the evaluation in this section is conducted under zero-shot setting. Evaluations with in-context learning and fine-tuning are presented in Sections 4.5 and 4.6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 555, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 375 + ], + "type": "text", + "content": "The performance is demonstrated in Table 2. We also present difficulty levels in the table, which is measured by the size of the map (3 represents " + }, + { + "bbox": [ + 313, + 316, + 555, + 375 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 316, + 555, + 375 + ], + "type": "text", + "content": " maps) in Maze Navigation and by the minimum number of steps in Blocks World. From the table, we summarize our findings as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 377, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 556, + 532 + ], + "type": "text", + "content": "VSP highlights both the advancements and the challenges in MLLMs' spatial planning capabilities. Notably, private MLLMs demonstrate a clear trajectory of improvement, with earlier versions performing worse and later iterations showing advancements—though still exhibiting deficiencies. In fact, even some of the most capable models frequently make mistakes on tasks such as navigating a " + }, + { + "bbox": [ + 313, + 377, + 556, + 532 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 313, + 377, + 556, + 532 + ], + "type": "text", + "content": " map or executing 3-step block-moving tasks, which are often trivial for humans. Meanwhile, open-source models face even greater difficulties and rarely succeed in these tasks. Overall, VSP demonstrates that while MLLMs have made progress in spatial planning, they still exhibit considerable room for improvement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "MLLMs face significant difficulties with spatial planning in dynamic and realistic environments. Based on the experimental results, most models perform worse when tested in dynamic (Collision) and realistic (Google Map) settings. We identify two major reasons: First, the tasks in Collision are complex and typically require multiple capabilities. For example, to assess collision danger, the model must locate both the player and the car on the map, calculate the time needed for the player to reach the goal, and determine if the car will hit the player during that time. This poses significant challenges for the models. Second, the inputs in the Google Map are intricate and contain a series of irrelevant symbols, making it difficult for the model to accurately interpret the map. The models' performance in these environments suggests that current models are not yet" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2275" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 85, + 141, + 137 + ], + "blocks": [ + { + "bbox": [ + 89, + 85, + 141, + 137 + ], + "lines": [ + { + "bbox": [ + 89, + 85, + 141, + 137 + ], + "spans": [ + { + "bbox": [ + 89, + 85, + 141, + 137 + ], + "type": "image", + "image_path": "66827513d759600fc796ebe01d54406d81d167c9d5bf66883c09fcfdab23f362.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 140, + 137, + 149 + ], + "lines": [ + { + "bbox": [ + 94, + 140, + 137, + 149 + ], + "spans": [ + { + "bbox": [ + 94, + 140, + 137, + 149 + ], + "type": "text", + "content": "Vision Input" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 145, + 90, + 250, + 137 + ], + "blocks": [ + { + "bbox": [ + 145, + 90, + 250, + 137 + ], + "lines": [ + { + "bbox": [ + 145, + 90, + 250, + 137 + ], + "spans": [ + { + "bbox": [ + 145, + 90, + 250, + 137 + ], + "type": "image", + "image_path": "44640806becebfc13ad3a4e82b604c489be384bb5698b7638f50b55ceb968ecf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 140, + 226, + 149 + ], + "lines": [ + { + "bbox": [ + 172, + 140, + 226, + 149 + ], + "spans": [ + { + "bbox": [ + 172, + 140, + 226, + 149 + ], + "type": "text", + "content": "Pure Text Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 258, + 89, + 358, + 137 + ], + "blocks": [ + { + "bbox": [ + 176, + 71, + 278, + 81 + ], + "lines": [ + { + "bbox": [ + 176, + 71, + 278, + 81 + ], + "spans": [ + { + "bbox": [ + 176, + 71, + 278, + 81 + ], + "type": "text", + "content": "Maze Navigation Scenario" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 89, + 358, + 137 + ], + "lines": [ + { + "bbox": [ + 258, + 89, + 358, + 137 + ], + "spans": [ + { + "bbox": [ + 258, + 89, + 358, + 137 + ], + "type": "image", + "image_path": "d787593d60533ef6c60e0074180fad41ade975b6bd0c23a4e09116e51b4d68db.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 140, + 328, + 149 + ], + "lines": [ + { + "bbox": [ + 288, + 140, + 328, + 149 + ], + "spans": [ + { + "bbox": [ + 288, + 140, + 328, + 149 + ], + "type": "text", + "content": "Table Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 367, + 85, + 419, + 137 + ], + "blocks": [ + { + "bbox": [ + 402, + 71, + 490, + 81 + ], + "lines": [ + { + "bbox": [ + 402, + 71, + 490, + 81 + ], + "spans": [ + { + "bbox": [ + 402, + 71, + 490, + 81 + ], + "type": "text", + "content": "Blocks World Scenario" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 367, + 85, + 419, + 137 + ], + "lines": [ + { + "bbox": [ + 367, + 85, + 419, + 137 + ], + "spans": [ + { + "bbox": [ + 367, + 85, + 419, + 137 + ], + "type": "image", + "image_path": "6e4b76d406dfdbea0ef3290da8471b2a3c1d4f60c12b458f672df71d7915be16.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 140, + 414, + 149 + ], + "lines": [ + { + "bbox": [ + 372, + 140, + 414, + 149 + ], + "spans": [ + { + "bbox": [ + 372, + 140, + 414, + 149 + ], + "type": "text", + "content": "Vision Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 140, + 498, + 149 + ], + "lines": [ + { + "bbox": [ + 444, + 140, + 498, + 149 + ], + "spans": [ + { + "bbox": [ + 444, + 140, + 498, + 149 + ], + "type": "text", + "content": "Pure Text Input" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 156, + 402, + 167 + ], + "lines": [ + { + "bbox": [ + 205, + 156, + 402, + 167 + ], + "spans": [ + { + "bbox": [ + 205, + 156, + 402, + 167 + ], + "type": "text", + "content": "Figure 5. The visual and corresponding textual inputs." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 423, + 94, + 515, + 129 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 423, + 94, + 500, + 102 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 94, + 500, + 102 + ], + "spans": [ + { + "bbox": [ + 423, + 94, + 500, + 102 + ], + "type": "text", + "content": "- Stack with purple block" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 423, + 103, + 494, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 103, + 494, + 111 + ], + "spans": [ + { + "bbox": [ + 423, + 103, + 494, + 111 + ], + "type": "text", + "content": "- Stack with blue block" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 423, + 112, + 515, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 112, + 515, + 129 + ], + "spans": [ + { + "bbox": [ + 423, + 112, + 515, + 129 + ], + "type": "text", + "content": "- Stack with orange block, red block, from bottom to top" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 188, + 294, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 188, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 54, + 188, + 294, + 236 + ], + "type": "text", + "content": "equipped to handle spatial reasoning in such complex environments. Consequently, we focus on the Maze Navigation and Blocks World scenarios in the following experiments to diagnose the models' hidden weaknesses in VSP tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 236, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 296, + 415 + ], + "type": "text", + "content": "Quick performance decay as difficulty increases. We observe a significant drop in the success rates of MLLMs as task difficulty escalates. For example, GPT-4V may achieve a success rate of over " + }, + { + "bbox": [ + 56, + 236, + 296, + 415 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 236, + 296, + 415 + ], + "type": "text", + "content": " on 3x3 size maps, but this plummets to just " + }, + { + "bbox": [ + 56, + 236, + 296, + 415 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 236, + 296, + 415 + ], + "type": "text", + "content": " on 8x8 maps. Analyzing the impact of increased difficulty, we identify two major challenges for the models: First, increasing size of the map in Maze Navigation scenario could make it difficult for the model to accurately perceive the positions of elements within the map. Second, the increase in both map size and the number of steps required for moving blocks heightens the challenge for the model to reason deeply through the entire path and devise a complete, viable solution. In the following experiments, we focus on these two factors and provide in-depth diagnoses with subsequent tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 415, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 666 + ], + "type": "text", + "content": "Challenges in open-source models. Finally, we note that open-source models often face difficulties on these tasks. As will be analyzed in Section 4.3, their perception and reasoning capabilities present greater challenges. Besides, we note two other factors: ① Context length: Open-source models typically have shorter context windows compared to private models. Thus, these models may not have enough capacity to understand the complete inputs. For example, LLAVA-V1.6-VICUNA-7B is trained with a context window of 2048 tokens, while each image consumes 576 tokens. As such, when fed with multiple images and long texts in our tasks, the total token length may surpass training, resulting in poor performance. ② Multiple image input: Our tasks require the model to understand multiple images interleaved with text inputs, whereas some open-source models are only trained with single-image inputs, with the image positioned at the start of the input. To further explore their potential, we assess their performance after finetuning in Section 4.6. Also, we note that recent open-source MLLMs often avoid these two issues, which may contribute to the improved performance observed in Table 2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 672, + 263, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 263, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 263, + 685 + ], + "type": "text", + "content": "4.3. Diagnose the Perception and Reasoning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "From the previous observation, we identify that spatial perception and reasoning could be two important capabilities" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 321, + 193, + 548, + 350 + ], + "blocks": [ + { + "bbox": [ + 321, + 193, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 321, + 193, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 321, + 193, + 548, + 350 + ], + "type": "table", + "html": "
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4T1T2T3T4
Random Guess0.50.250.250.50.170.250.250.5
Gemini-1.0 [54]0.580.560.330.490.860.510.540.55
Gemini-2.0 [18]0.660.910.960.800.950.960.990.76
Claude-3 [5]0.450.670.320.610.430.530.490.66
Claude-3.7 [9]0.720.880.970.790.940.840.990.94
GPT-4V [1]0.560.270.460.560.730.800.700.71
GPT-4o [7]0.580.670.580.740.950.900.900.76
LLaVA-Next [37]0.490.270.210.540.220.210.240.55
InternLM [19]0.480.270.290.580.250.320.260.53
InternLM-VL [19]0.410.200.170.470.220.200.200.53
SPHINX [35]0.560.280.320.590.240.330.270.58
LLaMA-3.2 [43]0.530.290.330.570.280.360.290.56
Pixtral [6]0.440.350.330.510.490.720.630.57
Qwen-2.5 [11]0.580.620.310.500.530.460.790.56
", + "image_path": "74543ce1100bfd9be267da631fe6be99067b2da683e633ee3dd321c4291e348d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 358, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 553, + 392 + ], + "type": "text", + "content": "Table 3. Decomposed Capability Analysis. Results better than " + }, + { + "bbox": [ + 313, + 358, + 553, + 392 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 358, + 553, + 392 + ], + "type": "text", + "content": " are bolded. Each task consists of test with different difficulties. Please see Appendix F for the results for different difficulties." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 407, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 467 + ], + "type": "text", + "content": "for an agent to successfully perform visual spatial planning. We show these examples in Appendix E. Next, we evaluate these two abilities through the remaining tasks T1-T4. Similar to previous settings, all the evaluation is conducted under zero-shot settings." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 468, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 635 + ], + "type": "text", + "content": "Table 3 shows the decomposed capability results. Strong private models like Gemini-2.0 and Claude-3.7 show better performance across these tasks, demonstrating solid perception and reasoning capabilities, which explain their superior performance in the main tasks. However, they still exhibit some deficiencies, such as perceiving the location of a single object in Maze T1, highlighting a critical area for improvement. Furthermore, the performance of many open-source models is mostly close to random guessing on these tasks, indicating significant gaps compared to private models. One caveat is that while T4 focuses on reasoning capabilities, it still relies on the perception capabilities because the input still contains images. We perform further analysis to disentangle these two abilities in Section 4.4." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 647, + 463, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 463, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 463, + 659 + ], + "type": "text", + "content": "4.4. The Effects of Visual Input" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Previous analysis shows that even strong MLLMs have clear deficiencies in various aspects of visual spatial planning. In this study, we focus on disentangling the effects of perception and reasoning by exploring the performance" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2276" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 315, + 173 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 315, + 173 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 315, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 315, + 173 + ], + "type": "image", + "image_path": "58a85caf84a5c335f70c5375d837680ff78bfab939266940a3c79f97ccf83c0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 177, + 550, + 189 + ], + "lines": [ + { + "bbox": [ + 57, + 177, + 550, + 189 + ], + "spans": [ + { + "bbox": [ + 57, + 177, + 550, + 189 + ], + "type": "text", + "content": "Figure 6. Result with the visual/textual input. When environment is described by text instead of image, the performance increases vastly." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 316, + 77, + 549, + 172 + ], + "blocks": [ + { + "bbox": [ + 316, + 77, + 549, + 172 + ], + "lines": [ + { + "bbox": [ + 316, + 77, + 549, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 77, + 549, + 172 + ], + "type": "image", + "image_path": "9085b0ec41a60e3f4c12d2ede371886d1a75cec48526512f821eb7919bb1318e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 200, + 294, + 291 + ], + "blocks": [ + { + "bbox": [ + 57, + 200, + 294, + 291 + ], + "lines": [ + { + "bbox": [ + 57, + 200, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 57, + 200, + 294, + 291 + ], + "type": "table", + "html": "
TaskMAZE NAVIGATIONBLOCKS WORLD
T1T2T3T4MainT1T2T3T4Main
Gemini, 0-shot0.580.560.330.490.170.860.510.540.550.03
Gemini, 1-shot0.500.660.310.480.200.910.680.710.590.03
Gemini, 2-shot0.530.680.310.510.210.900.760.700.610.03
Gemini, 4-shot0.530.670.350.530.190.910.640.690.620.06
GPT-4V, 0-shot0.560.270.460.560.260.730.800.700.710.10
GPT-4V, 1-shot0.550.500.470.570.280.890.840.940.730.11
GPT-4V, 2-shot0.550.630.500.560.300.900.830.950.710.16
GPT-4V, 4-shot0.540.690.540.560.290.900.790.960.73-
", + "image_path": "c2608ee94a2970baf7f4ffdcacc028202c4daf951917018631771f42d088c29c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 308, + 294, + 362 + ], + "blocks": [ + { + "bbox": [ + 84, + 297, + 266, + 307 + ], + "lines": [ + { + "bbox": [ + 84, + 297, + 266, + 307 + ], + "spans": [ + { + "bbox": [ + 84, + 297, + 266, + 307 + ], + "type": "text", + "content": "Table 4. Effects of providing in-context examples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 308, + 294, + 362 + ], + "lines": [ + { + "bbox": [ + 57, + 308, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 294, + 362 + ], + "type": "table", + "html": "
ModelSettingMAZE NAVIGATIONBLOCKS OF WORLD
T1T2T3T4MainT1T2T3T4Main
LLaVAzero-shot0.490.270.210.540.050.220.210.240.550.01
fine-tune0.530.990.510.930.601.001.001.001.000.97
InternLMzero-shot0.480.270.290.580.110.250.320.260.530.00
fine-tune0.520.590.910.590.170.290.440.690.620.09
", + "image_path": "0f3d0162de5cf932f9f4e7de7ee06dd83ae9b8df4734dc425e8c69d354d6cdae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 368, + 270, + 379 + ], + "lines": [ + { + "bbox": [ + 79, + 368, + 270, + 379 + ], + "spans": [ + { + "bbox": [ + 79, + 368, + 270, + 379 + ], + "type": "text", + "content": "Table 5. Fine-tuning results for open-source models." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 390, + 253, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 253, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 253, + 402 + ], + "type": "text", + "content": "gain assuming the model had perfect perception." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "content": "The key strategy here is to create a scenario where the model has acquired all necessary information that would typically be obtained through visual perception. To this end, for every input image, we produce the corresponding textual inputs and replace those images, as shown in Figure 5. For the Maze Navigation scenario, we use either pure text descriptions or tables to depict the image. For the Blocks World scenario, we use pure text descriptions. We do not use tables for the Blocks World scenario because the number of blocks in each horizontal stack is usually unequal, making it difficult to form a complete table. Appendix B includes complete examples with pure text or table input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 544, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 641 + ], + "type": "text", + "content": "The results are shown in Figure 6. We observe a clear performance improvement when using textual input across every task. This suggests image perception presents significant challenges for MLLMs, and poor perception ability is a key factor in the inferior performance observed in previous tasks. Meanwhile, note that even with textual input, Gemini still cannot achieve decent performance on tasks that require reasoning. This indicates its deficiencies in reasoning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 648, + 295, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 295, + 661 + ], + "type": "text", + "content": "4.5. In-context Learning in Visual Spatial Planning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "In-context learning is a widely-adopted method to enhance LM's reasoning ability [14]. In this analysis, we study if it boosts visual spatial planning capabilities. We included varying numbers of examples for Gemini-1.0 and GPT-4V" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 202, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 286 + ], + "type": "text", + "content": "(show in Appendix C) and show results in Table 4. There are two key observations: First, in-context examples make some potential contributions, but they are not significant. The examples only benefits in several sparse cases, such as T2 in Maze Navigation and T3 in Blocks World. Second, scaling examples generally does not help, as illustrated by the saturated performance in each task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 293, + 456, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 456, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 456, + 306 + ], + "type": "text", + "content": "4.6. Fine-tuning in VSP Tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 310, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 478 + ], + "type": "text", + "content": "Finally, we assess the capabilities of the open-source model through dedicated training. We performed LoRA fine-tuning on llava-v1.6-vicuna-7b and internlm-xcomposer2-7b. The models are trained on 10k data points (image-text pairs) for each task and scenario. We use the default hyperparameters provided in the official repository. More fine-tuning details can be found in Appendix D. The results, shown in Table 5, demonstrate clear performance improvements for both models across a series of tasks, highlighting their potential in spatial planning. Additionally, we observe that LLaVA shows greater improvement compared to InternLM, suggesting that different model architectures may exhibit varying levels of efficacy in spatial planning capabilities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "content": "We present VSP, a benchmark measuring and diagnosing the visual spatial planning capabilities in MLLMs. VSP quantifies the model's performance through a series of carefully designed tasks, with main tasks testing general spatial planning abilities and sub-tasks diagnosing individual capabilities needed for the main task. We show that current models fail to generate effective plans for even simple spatial planning tasks, and further analyses expose their bottlenecks in spatial perception and reasoning abilities. Our work illuminates future directions for improving MLLMs' abilities in spatial planning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 650, + 419, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 419, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 419, + 663 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 669, + 554, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 554, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 554, + 702 + ], + "type": "text", + "content": "Qiucheng Wu and Shiyu Chang acknowledge the support from National Science Foundation (NSF) Grant IIS-2338252, Grant IIS-2207052, and Grant IIS-2302730." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2277" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6, 7, 23" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Constructions Aeronautiques, Adele Howe, Craig Knoblock, ISI Drew McDermott, Ashwin Ram, Manuela Veloso, Daniel Weld, David Wilkins Sri, Anthony Barrett, Dave Christianson, et al. Pddll the planning domain definition language. Technical Report, Tech. Rep., 1998. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "text", + "content": "[3] Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning. arXiv preprint arXiv:2310.03249, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 301 + ], + "type": "text", + "content": "[4] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 323 + ], + "type": "text", + "content": "[5] Anthropic AI. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 6, 7, 23" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 357 + ], + "type": "text", + "content": "[6] Mistral AI. Announcing pixtral 12b. https://mistral.ai/news/pixtral-12b/, 2024. Accessed: 2024-09-27. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 379 + ], + "type": "text", + "content": "[7] Open AI. Gpt-4o. https://openai.com/index/hello-gpt-4o/.6,7,23" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 445 + ], + "type": "text", + "content": "[8] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "text", + "content": "[9] Anthropic. Claude 3.7 sonnet and claude code. https://www.anthropic.com/news/claude-3-7-sonnet, 2025.6,7,23" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 545 + ], + "type": "text", + "content": "[10] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[11] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "text", + "content": "[12] Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[13] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. 3, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 690, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 295, + 713 + ], + "type": "text", + "content": "[14] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakan" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "tan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[15] Sungguk Cha, Jusung Lee, Younghyun Lee, and Cheoljong Yang. Visually dehallucinative instruction generation: Know what you don't know. arXiv preprint arXiv:2402.09717, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[16] Haonan Chang, Kai Gao, Kowndinya Boyalakuntla, Alex Lee, Baichuan Huang, Harish Udhaya Kumar, Jinjin Yu, and Abdeslam Boularias. Lgmcts: Language-guided montecarlo tree search for executable semantic object rearrangement. arXiv preprint arXiv:2309.15821, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "text", + "content": "[17] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[18] Google Deepmind. Introducing gemini 2.0: our new ai model for the agentic era. https://blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/, 2024.6,7,23" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 374 + ], + "type": "text", + "content": "[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, et al. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420, 2024. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 432 + ], + "type": "text", + "content": "[20] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "text", + "content": "[21] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "text", + "content": "[22] Wentao Ge, Shunian Chen, Guiming Chen, Junying Chen, Zhihong Chen, Shuo Yan, Chenghao Zhu, Ziyue Lin, Wenya Xie, Xidong Wang, et al. Mllm-bench, evaluating multimodal llms using gpt-4v. arXiv preprint arXiv:2311.13951, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "type": "text", + "content": "[23] Deepanway Ghosal, Vernon Toh Yan Han, Chia Yew Ken, and Soujanya Poria. Are language models puzzle prodigies? algorithmic puzzles unveil serious challenges in multimodal reasoning. arXiv preprint arXiv:2403.03864, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[24] Tejas Gokhale, Shailaja Sampat, Zhiyuan Fang, Yezhou Yang, and Chitta Baral. Blockworld revisited: Learning and reasoning to generate event-sequences from image pairs. arXiv preprint arXiv:1905.12042, 2019. 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[25] Lin Guan, Karthik Valmeekam, Sarath Sreedharan, and Subbarao Kambhampati. Leveraging pre-trained large language models to construct and utilize world models for model-based task planning. Advances in Neural Information Processing Systems, 36:79081-79094, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[26] Xiaoxiao Guo, Satinder Singh, Honglak Lee, Richard L Lewis, and Xiaoshi Wang. Deep learning for real-time" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2278" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 105 + ], + "type": "text", + "content": "atari game play using offline monte-carlo tree search planning. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[27] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[28] Rishi Hazra, Pedro Zuidberg Dos Martires, and Luc De Raedt. Saycanpay: Heuristic planning with large language models using learnable domain knowledge. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 20123-20133, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[29] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. arXiv preprint arXiv:2311.17842, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[30] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "type": "text", + "content": "[31] Frank Joublin, Antonello Ceravola, Pavel Smirnov, Felix Ocker, Joerg Deigmoeller, Anna Belardinelli, Chao Wang, Stephan Hasler, Daniel Tanneberg, and Michael Gienger. Copal: Corrective planning of robot actions with large language models. arXiv preprint arXiv:2310.07263, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "text", + "content": "[32] Subbarao Kambhampati. Can large language models reason and plan? Annals of the New York Academy of Sciences, 1534(1):15-18, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "text", + "content": "[33] Subbarao Kambhampati, Karthik Valmeekam, Lin Guan, Kaya Stechly, Mudit Verma, Siddhant Bhambri, Lucas Saldyt, and Anil Murthy. Llms can't plan, but can help planning in llm-modulo frameworks. arXiv preprint arXiv:2402.01817, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "text", + "content": "[34] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 623 + ], + "type": "text", + "content": "[36] Fuxiao Liu, Tianrui Guan, Zongxia Li, Lichang Chen, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusion-bench: You see what you think? or you think what you see? an image-context reasoning benchmark challenging for gpt-4v (ision), llava-1.5, and other multi-modality models. arXiv preprint arXiv:2310.14566, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[38] Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Yifan Xu, Xixuan Song, Shudan Zhang, Hanyu Lai, Xinyi Liu, Hanlin Zhao, et al. Visualagentbench: Towards large multimodal models as visual foundation agents. arXiv preprint arXiv:2408.06327, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[39] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. arXiv e-prints, pages arXiv-2310, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "type": "text", + "content": "[40] Yingzi Ma, Yulong Cao, Jiachen Sun, Marco Pavone, and Chaowei Xiao. Dolphins: Multimodal language model for driving. arXiv preprint arXiv:2312.00438, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[41] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 231, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 306 + ], + "type": "text", + "content": "[42] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. Openeqa: Embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 341 + ], + "type": "text", + "content": "[43] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[44] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[45] Zhangyang Qi, Ye Fang, Mengchen Zhang, Zeyi Sun, Tong Wu, Ziwei Liu, Dahua Lin, Jiaqi Wang, and Hengshuang Zhao. Gemini vs gpt-4v: A preliminary comparison and combination of vision-language models through qualitative cases. arXiv preprint arXiv:2312.15011, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[46] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Guoqing Du, Shiwei Shi, Hangyu Mao, Xingyu Zeng, and Rui Zhao. Tptu: Task planning and tool usage of large language model-based ai agents. arXiv preprint arXiv:2308.03427, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[47] Md Sadman Sakib and Yu Sun. From cooking recipes to robot task trees-improving planning correctness and task efficiency by leveraging llms with a knowledge network. arXiv preprint arXiv:2309.09181, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "text", + "content": "[48] Zhenwei Shao, Zhou Yu, Meng Wang, and Jun Yu. Prompting large language models with answer heuristics for knowledge-based visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14974-14983, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[49] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[50] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[51] Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2279" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "text", + "content": "Alfworld: Aligning text and embodied environments for interactive learning. arXiv preprint arXiv:2010.03768, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "text", + "content": "[52] Kaya Stechly, Karthik Valmeekam, and Subbarao Kambhampati. On the self-verification limitations of large language models on reasoning and planning tasks. arXiv preprint arXiv:2402.08115, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[53] Richard S Sutton. Planning by incremental dynamic programming. In Machine learning proceedings 1991, pages 353-357. Elsevier, 1991. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 295, + 239 + ], + "type": "text", + "content": "[54] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 6, 7, 23" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "text", + "content": "[55] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 337 + ], + "type": "text", + "content": "[56] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. arXiv preprint arXiv:2401.06209, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "text", + "content": "[58] Karthik Valmeekam, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Large language models still can't plan (a benchmark for llms on planning and reasoning about change). arXiv preprint arXiv:2206.10498, 2022. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 491 + ], + "type": "text", + "content": "[59] Karthik Valmeekam, Matthew Marquez, Sarath Sreedharan, and Subbarao Kambhampati. On the planning abilities of large language models - a critical investigation. In Advances in Neural Information Processing Systems, pages 75993-76005. Curran Associates, Inc., 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 295, + 536 + ], + "type": "text", + "content": "[60] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "text", + "content": "[61] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 624 + ], + "type": "text", + "content": "[62] Yue Wu, Xuan Tang, Tom M Mitchell, and Yuanzhi Li. Smartplay: A benchmark for llms as intelligent agents. arXiv preprint arXiv:2310.01557, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "text", + "content": "[63] Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. Travelplanner: A benchmark for real-world planning with language agents. arXiv preprint arXiv:2402.01622, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[64] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 519 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[65] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 9(1):1, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "text", + "content": "[66] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "text", + "content": "[67] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 274 + ], + "type": "text", + "content": "[68] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, et al. Mmt-bench: A comprehensive multimodal benchmark for evaluating large vision-language models towards multitask agi. arXiv preprint arXiv:2404.16006, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 319 + ], + "type": "text", + "content": "[69] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 554, + 407 + ], + "type": "text", + "content": "[70] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 407, + 554, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 554, + 473 + ], + "type": "text", + "content": "[71] Wenqi Zhang, Zhenglin Cheng, Yuanyu He, Mengna Wang, Yongliang Shen, Zeqi Tan, Guiyang Hou, Mingqian He, Yanna Ma, Weiming Lu, et al. Multimodal self-instruct: Synthetic abstract image and visual reasoning instruction using language model. arXiv preprint arXiv:2407.07053, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 474, + 554, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 554, + 519 + ], + "type": "text", + "content": "[72] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 1" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2280" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_content_list.json b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6416480a7d8813a92b66b71afa11c1dcdef0bd61 --- /dev/null +++ b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_content_list.json @@ -0,0 +1,1890 @@ +[ + { + "type": "text", + "text": "VSRM: A Robust Mamba-Based Framework for Video Super-Resolution", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dinh Phu Tran", + "bbox": [ + 267, + 181, + 388, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dao Duy Hung", + "bbox": [ + 434, + 181, + 558, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daeyoung Kim", + "bbox": [ + 606, + 181, + 728, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{phutx2000, hicehehe, kimd}@kaist.ac.kr", + "bbox": [ + 326, + 208, + 671, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computing, KAIST, Republic of Korea", + "bbox": [ + 300, + 229, + 694, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 289, + 326, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution remains a major challenge in low-level vision tasks. To date, CNN- and Transformer-based methods have delivered impressive results. However, CNNs are limited by local receptive fields, while Transformers struggle with quadratic complexity, posing challenges for processing long sequences in VSR. Recently, Mamba has drawn attention for its long-sequence modeling, linear complexity, and large receptive fields. In this work, we propose VSRM, a novel Video Super-Resolution framework that leverages the power of Mamba. VSRM introduces Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to extract long-range spatio-temporal features and enhance receptive fields efficiently. To better align adjacent frames, we propose Deformable Cross-Mamba Alignment module. This module utilizes a deformable cross-mamba mechanism to make the compensation stage more dynamic and flexible, preventing feature distortions. Finally, we minimize the frequency domain gaps between reconstructed and ground-truth frames by proposing a simple yet effective Frequency Charbonnier-like loss that better preserves high-frequency content and enhances visual quality. Through extensive experiments, VSRM achieves state-of-the-art results on diverse benchmarks, establishing itself as a solid foundation for future research.", + "bbox": [ + 89, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 720, + 220, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video super-resolution (VSR) aims to produce higher resolution from low-resolution videos. VSR typically processes long sequences and benefits from large receptive fields, efficiently capturing inter-frames information [3, 15]. However, it incurs high computational costs, requiring a balance between performance and efficiency. VSR usually uses convolutional neural network (CNN) [3, 22, 56] or transformer [33, 46, 54, 58]. Tab. 1 compares the pros and cons of these methods. CNN-based methods cannot adjust based on input data and limit receptive fields [10, 48] to capture inter-frame", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "content. Transformer-based methods show strong performance due to the power of the attention mechanism. However, full attention has quadratic complexity [21, 36], which is impractical for long sequences, while local attention struggles with limited receptive fields [12]. Recently, Gu et al. introduced Mamba [16], which can model long-range information and data-dependent characteristics with linear complexity for NLP tasks. More recently, Mamba has been successfully applied in various vision tasks [7, 18, 41, 57]. As far as we know, no research has employed Mamba to address VSR. In this work, we propose VSRM, a novel Mamba-based framework for VSR. Specifically, we propose Dual Aggregation Mamba Block (DAMB), which contains $N$ Spatial-to-Temporal Mamba blocks (S2TMBs) followed by a Temporal-to-Spatial Mamba block (T2SMB) to model long-range spatio-temporal correlations for long sequences, enlarging large receptive fields. In S2TMB and T2SMB, we employ spatial-to-temporal and temporal-to-spatial scanning mechanisms to capture both spatial and temporal contexts.", + "bbox": [ + 511, + 290, + 906, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key challenge of VSR is effectively using the complementary information from adjacent frames, which may be misaligned due to object motions. To establish interframe correlations, many works employ the alignment module for VSR [3, 5, 46, 54]. However, most studies use bilinear or nearest-neighbor interpolation for spatial alignment due to its simplicity. Theoretically, using fixed weights (e.g., linear interpolation) causes feature distortions, leading to suboptimal results [4, 46]. To address this, Xu et al. [54] proposed an attention-based implicit interpolation for the alignment module. However, they computed interpolation in the fixed reference window, limiting flexibility and inaccuracy in aligning adjacent frames, especially in large motions between frames. Based on this observation, we introduce Deformable Cross-Mamba Alignment (DCA), which dynamically aligns neighboring frames using a deformable cross-mamba mechanism that allows information exchange between multiple selective scan modules.", + "bbox": [ + 511, + 595, + 908, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current VSR methods use pixel-based losses (e.g., L1/L2 loss [26, 34], charbonnier loss [5, 46]) or perception-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/496e03861e5b6a9e16cfbe067a4bc109c5369e8a3aa434cc36c84ebc36b99013.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLinear ComplexityReceptive FieldData-drivenNotable Method
CNN⊗⊗⊗BasicVSR [3]
Local Attention⊗⊗⊗IART [54]
Full Attention⊗⊗⊗-
Mamba⊗⊗⊗VSRM (ours)
", + "bbox": [ + 94, + 88, + 478, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of the backbone for VSR. VSRM reaches both linear complexity and a large receptive field.", + "bbox": [ + 89, + 176, + 482, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based losses (e.g., perceptual loss [24, 43], adversarial loss [29, 42]). Pixel-wise losses lead to overly smooth results due to their restriction in capturing perceptual similarity. Perception-based losses enhance visual quality by recovering finer details but introduce greater distortion. As VSR is an ill-posed task, gaps between super-resolution and ground-truth frames are inevitable, sometimes visible only in the frequency domain [1, 20]. Therefore, combining loss in the spatial and frequency domains may enhance the model's performance. We propose Frequency Charbonnier-like Loss (FCL), calculated from the real and imaginary parts of frequency components in Fourier space. FCL helps our model focus more on preserving and generating high-frequency details while maintaining low-frequency content. Contribution. 1) We successfully adapt Mamba-based model for VSR for the first time and propose a new framework, VSRM. VSRM introduces Dual Aggregation Mamba Block (DAMB), which aggregates spatial and temporal features from sequences by using multi-scan direction mechanisms to obtain powerful representation ability and enlarge the model's receptive field. 2) We design Deformable Cross-mamba Alignment (DCA) module to provide a more flexible and learnable alignment through a deformable window cross-mamba mechanism in the compensation stage. 3) We propose Frequency Charbonnier-like loss (FCL), which is computed in the frequency domain to preserve and generate high-frequency details better. 4) VSRM outperforms all state-of-the-art (SOTA) methods on various benchmarks and shows the potential as the solid backbone for VSR.", + "bbox": [ + 91, + 234, + 483, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 691, + 232, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Super-Resolution. 1) CNN-based methods [3, 25, 37, 50, 56] is the early approach in VSR. Initially, VSR-net [25] extended CNN-based networks to VSR. Following this, TDAN [50] introduced deformable convolution alignment to enhance the alignment module's performance. In contrast, BasicVSR [3] proposed an effective model consisting solely of generic components, residual blocks, and optical flow, which surpasses all previous methods. 2) Transformer-based methods [2, 33, 46, 54, 58] achieve superior performance in VSR. Cao et al. [2] proposed spatial-temporal convolutional self-attention to explore better locality features. VRT [33] used a temporal mutual self", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attention for a wider temporal receptive field and a parallel warp module to handle large motion. PSRT [46] introduced patch alignment that aligns image patches instead of pixels to maintain intra-patch relationships and preserve sub-pixel information. Based on this, Xu et al. [54] proposed an implicit alignment module based on cross-attention mechanism to adjust feature representations across frames, further improving performance. These methods applied window-based attention with small window to avoid the quadratic complexity of full attention but were limited by local receptive fields, affecting cross-frame information capture. Recent methods, like PSRT [46] and IART [54], focused on designing window-based alignment modules rather than improving the transformer backbone to enhance performance. They employed a fixed window for implicit motion compensation, limiting the flexibility of alignment modules, especially when minimal or excessive motion. To address these issues, this study proposes a new framework, VSRM, based on Mamba to leverage its advantage in large receptive fields and linear complexity. We also design Deformable Cross-mamba Alignment for more flexible motion compensation using a deformable mechanism to boost performance.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Model. SSMs [13, 16, 17, 47] show potential to replace Transformers. Gu et al. introduced S4 [17] that achieves linear complexity modeling of long sequences, followed by S5[47] with improved capacity and efficiency. Mamba [16], built upon SSMs, has recently surpassed Transformers on many benchmarks in the NLP domain due to its data-dependent characteristics and global receptive fields. In computer vision, Vim [59] adapted Mamba for images by converting them into sequences and using a bidirectional mamba block for global modeling. VideoMamba [30] processed video by rearranging frames into spatio-temporal sequences, achieving comparable results to Transformer-based methods. MambaIR [19] adopted Mamba for image restoration tasks, demonstrating that Mamba can work well for low-level vision tasks, such as video deblurring, denoising, etc. Expanding on that progress, we are pioneers in investigating the capabilities of Mamba for VSR, utilizing its large receptive fields and linear complexity.", + "bbox": [ + 511, + 426, + 906, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spectral bias. Many studies reveal spectral bias [39, 44, 49] that models bias toward low-frequency rather than high-frequency content. Meanwhile, high-frequency content (e.g., edges, textures) is vital in low-level vision tasks [14, 23, 27, 28]. Hence, Jiang et al. proposed Focal Frequency Loss (FFL) [23] computed in the frequency domain to emphasize high-frequency content and enhance visual quality. Based on this, Fuoli et al. [14] employed a GAN loss in Fourier space. In contrast, WHFL [27] used Wavelet transform, incorporating a module for multi-bandwidth analysis. These methods employ a frequency-adjusted reweight matrix to optimize the low and high-frequency elements. However, the reweight matrix does not", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14712", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/99d88d026dc4c7dedd45aa5916d292f0beadd65f98f4d7d7127dcba9768ea7d8.jpg", + "image_caption": [ + "Figure 1. The network architecture of our method. a) Video Super-Resolution based on Mamba (VSRM). b) Feature Propagation block (FPB) consists of Deformable Cross-mamba Alignment (DCA) and Dual Aggregation Mamba Blocks (DAMB). c) Spatial-to-Temporal Mamba block (S2TMB). d) Temporal-to-Spatial Mamba block (T2SMB)." + ], + "image_footnote": [], + "bbox": [ + 166, + 87, + 834, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "effectively balance the gradients of low and high frequencies. Furthermore, images mainly contain low-frequency content, so overemphasis can harm performance. Hence, we propose Frequency Charbonnier-like loss (FCL), which directly computes loss on the Real and Imaginary parts in Fourier space. Unlike FFL [23], FCL treats low- and high-frequency components equally.", + "bbox": [ + 89, + 434, + 483, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries: State Space Models", + "text_level": 1, + "bbox": [ + 89, + 556, + 403, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SSMs draw their inspiration from continuous linear time-invariant systems, which map a 1D function or sequence $x(t) \\in \\mathbb{R} \\to y(t) \\in \\mathbb{R}$ via a hidden state $h(t) \\in \\mathbb{R}^N$ . This is formulated as ordinary differential equations (ODE):", + "bbox": [ + 89, + 582, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = A h (t) + B x (t), y (t) = C h (t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 655, + 482, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $A \\in \\mathbb{R}^{N \\times N}$ is the evolution matrix, while $B \\in \\mathbb{R}^{N \\times 1}$ and $C \\in \\mathbb{R}^{1 \\times N}$ are the projection parameters. To model discrete sequences like images, we need to approximate them via discretization. SSMs utilize the zero-order hold method, incorporating a timescale parameter $\\Delta$ to map the continuous parameters $A$ , $B$ to discrete parameters $\\bar{A}$ , $\\bar{B}$ . The process of discretization is outlined as follows:", + "bbox": [ + 89, + 684, + 483, + 791 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\bar {A} h _ {t - 1} + \\bar {B} x _ {t}, y _ {t} = C h _ {t}, \\tag {2a}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 803, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {A} = e x p (\\Delta A), \\bar {B} = (\\Delta A) ^ {- 1} (e x p (\\Delta A - I)) \\Delta B, (2 \\mathrm {b})\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 823, + 482, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The parameters of these SSMs are independent of the data, which means that $A$ , $B$ , and $C$ remain constant regardless of the input, thereby restricting their adaptability", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in sequence modeling. Recently, Mamba [16] proposes selective SSMs (S6) with input data-dependent $x_{i} \\in \\mathbb{R}^{L}$ :", + "bbox": [ + 511, + 434, + 906, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nB _ {i} = S _ {B} x _ {i}, C _ {i} = S _ {c} x _ {i}, \\Delta_ {i} = S o f t p l u s (S _ {\\Delta} x _ {i}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 474, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S_B, S_C \\in \\mathbb{R}^{N \\times L}$ and $S_{\\Delta} \\in \\mathbb{R}^{L \\times L}$ are linear projection layers. $B_i$ and $C_i$ are utilized across for all channels of $x_i$ , $\\Delta x_i$ contains $\\Delta$ of $L$ channels, and $A$ remains consistent with earlier SSMs. Via the discretization presented in equations 3, $\\bar{A}$ and $\\bar{B}$ become input data-dependent.", + "bbox": [ + 511, + 498, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 588, + 648, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce VSRM architecture and its core components: Dual Aggregation Mamba block (DAMB) and Deformable Cross-mamba Alignment (DCA), as shown in Fig. 1. Furthermore, we present the training objectives to optimize our model.", + "bbox": [ + 511, + 613, + 906, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 696, + 624, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $I_{t}^{LR} \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ be the t-th low-resolution (LR) video frame, while $I_{t}^{HR} \\in \\mathbb{R}^{T \\times sH \\times sW \\times C}$ denotes the t-th corresponding ground-truth (GT or HR) frame. Here, $T, H, W, C$ , and $s$ indicate the total number of frames, height, width, channels, and the upscaling factor, respectively. Our aim is to generate the super-resolution video frame $I_{t}^{SR} \\in \\mathbb{R}^{T \\times sH \\times sW \\times C}$ using the consecutive $2N + 1$ frames $\\{I_{i = t - N}^{t + N}\\}$ , ensuring that $I_{t}^{SR}$ and $I_{t}^{HR}$ are as closely as possible. VSRM is depicted in Fig. 1a) includes two parts: Feature Extraction (Conv2d and Feature Propagation Block) and Upsampler (Reconstruction). Feature extraction extracts deep features from input frames and aligns and", + "bbox": [ + 511, + 717, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14713", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f9f6aae8a4c0bece79d5b087eef5cd9f33d3bba43405624c7bb8d529809b446.jpg", + "image_caption": [ + "Figure 2. Structure and scan directions of Mamba module. a) Spatial-to-Temporal Mamba (S2T-Mamba). b) Temporal-to-Spatial Mamba (T2S-Mamba). For simplicity, we exclude the initial normalization and the final residual." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 460, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fuses the features of adjacent frames to leverage inter-frame relationships. The upsampler upscales the fused spatiotemporal features to produce the output sequence frames.", + "bbox": [ + 89, + 301, + 483, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Dual Aggregation Mamba Block", + "text_level": 1, + "bbox": [ + 89, + 356, + 375, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dual Aggregation Mamba block (DAMB) is the core component of VSRM that consists of Spatial-to-Temporal Mamba block (S2TMB) and Temporal-to-Spatial Mamba block (T2SMB), as shown in Fig. 1(c, d). S2TMB and T2SMB are based on two modules: Spatial-to-Temporal Mamba (S2T-Mamba) and Temporal-to-Spatial Mamba (T2S-Mamba), with their structures shown in Fig. 2. By organizing $N$ S2TMBs followed by a T2SMB in DAMB, VSRM fully extracts spatio-temporal information in both spatial and temporal dimensions. Moreover, we propose Temporal Gated Feed-forward Network (TGFN) to explore spatial and temporal information better.", + "bbox": [ + 89, + 378, + 483, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial-to-Temporal Mamba. S2T-Mamba structure and scan directions are shown in Fig. 2(a). A 3D sequence ( $T$ frames) is converted into a 1D sequence and processed in two directions: forward (ForwardSSM(.)) and backward (BackwardSSM(.)) paths. This approach follows a mechanism that first scans spatially and then temporally, moving from frame 1 to frame $T$ . This enables both spatial and temporal awareness during processing. The original Mamba was designed for 1D sequences, so this bidirectional scan mechanism helps preserve spatial awareness. By employing the scanning process from spatial to temporal in both the forward and backward paths, our module is optimized for handling 3D sequences more effectively. Given an input $X_{in}$ , the output of S2T-Mamba is computed as:", + "bbox": [ + 89, + 560, + 483, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx = \\operatorname {L i n e a r} \\left(X _ {i n}\\right), z = \\sigma \\left(\\operatorname {L i n e a r} \\left(X _ {i n}\\right)\\right), \\tag {4a}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 781, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {1} = \\text {F o r w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {4b}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 801, + 480, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {2} = \\text {B a c k w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {4c}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 823, + 480, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS 2 T - M a m b a (x, z) = \\operatorname {L i n e a r} \\left(x _ {1} \\odot z + x _ {2} \\odot z\\right) \\tag {4d}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 843, + 480, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal-to-Spatial Mamba. T2S-Mamba structure and scan direction are shown in Fig. 2(b). T2S-Mamba also", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "converts 3D sequence into a 1D sequence. However, we only perform the scanning in forward direction for two reasons. First, forward scan in T2S-Mamba models feature changes over time, while adding backward scan disrupts learned dependencies, leading to a slight decrease in performance, and an increase in FLOPs. Second, our experiment shows that using only forward scan in T2S-Mamba yields better results than adding backward pass, shown in Tab. 5. We found that S2T-Mamba prioritizes extracting spatial rather than temporal information, leading to not extracting fully temporal information. In contrast, T2S-Mamba explicitly prioritizes scanning temporal rather than spatial information, allowing for a more comprehensive feature extraction. Given an input $X_{in}$ , the output of S2T-Mamba is calculated as:", + "bbox": [ + 511, + 90, + 906, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx = \\operatorname {L i n e a r} \\left(X _ {i n}\\right), z = \\sigma \\left(\\operatorname {L i n e a r} \\left(X _ {i n}\\right)\\right), \\tag {5a}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 329, + 903, + 345 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {1} = \\text {F o r w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {5b}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 349, + 903, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT 2 S - M a m b a (x, z) = \\operatorname {L i n e a r} \\left(x _ {1} \\odot z\\right) \\tag {5c}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 369, + 903, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Gated Feed-forward Network. Feed-forward network (FFN) [51] consists of a nonlinear activation positioned between two linear projection layers. It disregards spatial and temporal relationships and lacks feature alignment since it processes each input token individually. Moreover, the redundant information across channels hinders progress in learning feature representations. To address this issues, we propose Temporal Gated Feed-forward Network (TGFN) shown in Fig. 3. It has two main updates to FFN: 1) 3D depthwise convolutions (DW-3D Conv) and 2) a gating mechanism. We use DW-3D Conv to model relationships effectively from spatially and temporally neighboring pixel positions, which is vital in VSR. TGFN employs a gating mechanism by performing element-wise multiplication on two parallel feature branches obtained from the chunking method, with one of them passed by the GELU, which has become the preferred option in recent work [8, 31]. Given the input $X \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ , TGFN is formulated as:", + "bbox": [ + 511, + 397, + 906, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} = W _ {p} ^ {1} X, \\quad [ \\hat {X} _ {1}, \\hat {X} _ {2} ] = \\hat {X} \\tag {6a}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 679, + 903, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT G F N (X) = W _ {p} ^ {2} \\left(W _ {d} ^ {1} L N \\left(\\hat {X} _ {1}\\right) \\odot \\sigma \\left(W _ {d} ^ {2} L N \\left(\\hat {X} _ {2}\\right)\\right)\\right) \\tag {6b}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 703, + 903, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{p}^{1}, W_{p}^{2}$ are weights of linear projections; $W_{d}^{1}, W_{d}^{2}$ are weights of the 3D DWConv layers; $\\sigma$ is GELU activation, $C$ is the hidden dimension, LN is LayerNorm. Overall, TGFN adds nonlinear spatial and temporal information more effectively and optimizes the information flow through our pipeline.", + "bbox": [ + 511, + 731, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Deformable Cross-mamba Alignment", + "text_level": 1, + "bbox": [ + 511, + 833, + 836, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alignment modules in VSR often fail when pixel movement between consecutive frames is minimal or excessive. This requires an alignment module adaptable to varying pixel", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14714", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c1aab8293864997f4a89d1f695b60631c93dee61d8739bdf44be0bca0a155469.jpg", + "image_caption": [ + "Figure 3. Temporal Gated Feed-forward Network (TGFN)" + ], + "image_footnote": [], + "bbox": [ + 160, + 87, + 406, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "movement. Building upon [3, 5, 46], we design Deformable Cross-mamba Alignment (DCA), as shown in Fig. 4. DCA uses the pre-trained SpyNet [45] for flow estimation while incorporating a new motion compensation scheme. Specifically, we use a deformable window scheme to perform a more flexible warping operation in the compensation stage between frames according to inter-frame motion information. We then employ a cross-mamba module, which enables interactions between the target point and the values of the dynamic reference points to infer the aligned point, leading to more adaptive and accurate alignment results.", + "bbox": [ + 89, + 215, + 483, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each pixel $X(x,y)$ in the target frame $X$ (red point), let $(\\Delta_x,\\Delta_y)$ be the optical flow estimated between $X$ and the reference frame $X^{\\prime}$ . The aligned frame $\\bar{X}$ is obtained by a spatial warping function $\\mathcal{W}$ with its parameters $\\theta_{\\mathcal{W}}$ as:", + "bbox": [ + 89, + 382, + 483, + 443 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} (x, y) = \\mathcal {W} \\left(X ^ {\\prime}, (x + \\Delta x, y + \\Delta y); \\theta_ {\\mathcal {W}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 452, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the warping operation, we learn offset $\\epsilon_r$ from window $w$ to identify dynamic reference region $\\bar{r}$ . We extract $w \\in \\mathbb{R}^{H_w \\times W_w \\times C}$ , which is centered at the sampling position $(x', y') = (x + \\Delta x, y + \\Delta y)$ in $X'$ . Here, $H_w$ , $W_w$ , and $C$ are the height, width, and number of channels of the window, respectively, as shown in Fig. 4. Within this window, we initialize a reference region $r \\in \\mathbb{R}^{H_r \\times W_r \\times 2}$ , by iterating over $i \\in (0, 1, \\dots, H_r)$ and $j \\in (0, 1, \\dots, W_r)$ , as follows:", + "bbox": [ + 89, + 481, + 483, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (i, j) = X ^ {\\prime} \\left(\\lfloor x ^ {\\prime} \\rfloor - \\lfloor \\frac {W _ {r} - 1}{2} \\rfloor + i, \\lfloor y ^ {\\prime} \\rfloor - \\lfloor \\frac {H _ {r} - 1}{2} \\rfloor + j\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 612, + 482, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the dynamic reference region $\\bar{r}$ , we refine $r$ using a small learnable offset network $\\mathcal{S}(.)$ , which contains only convolution and activation layers (Fig. 4). This network learns the reference offset $\\epsilon_r$ from the features within the window $w$ . The final adaptive reference map is computed using bilinear interpolation $\\phi(.;.)$ as follows:", + "bbox": [ + 89, + 641, + 483, + 732 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {r} = \\phi (w; r + \\epsilon_ {r}), \\quad \\text {w i t h} \\epsilon_ {r} = S (w), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 743, + 482, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we apply implicit warping within the dynamic reference region to align each pixel by selectively cross-fusing reference point values with the target grid value at $X(x,y)$ using the cross-mamba module cross-mamba(.,.):", + "bbox": [ + 89, + 771, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {X} (x, y) = \\text {c r o s s - m a m b a} (R, Q), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 842, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R$ is the reference tensor and $Q$ is the target tensor, both augmented with positional information. The reference", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6621aaa41fc3a79f2bd20b77ea92c80c835c73b701c40ea0c9f2b06e22ebcf35.jpg", + "image_caption": [ + "Figure 4. Pipeline of our proposed Deformable Cross-mamba Alignment (DCA). We initialize a fixed reference region $r$ within the window $w$ . This region is then adjusted to a dynamic reference region $\\bar{r}$ using an offset $\\epsilon_r$ , which is learned from the features within $w$ . Finally, DCA learns the affinity through the cross-mamba module to calculate the final result (aligned point)." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tensor $R$ is computed by adding the flattened vector $\\bar{r}$ to the positional encoding $PE_{R}$ , and the target tensor $Q$ is obtained by adding the target value $X(x,y)$ to the positional embedding $PE_{Q}$ . We use sinusoidal positional encoding to capture high-frequency content, which maps continuous input coordinates into a high-dimensional space [39] (more details in the supp. file). Based on Eq. (2a), the cross-mamba module computes output $\\bar{X}$ at each step $t$ as:", + "bbox": [ + 511, + 359, + 906, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {t} = \\bar {A} _ {R} H _ {t - 1} + \\bar {B} _ {R} \\bar {R} _ {t}, \\bar {X} _ {t} = C _ {Q} H _ {t}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 492, + 906, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{A}_R$ , $\\bar{B}_R$ are matrices dependent on $R$ , and $C_Q$ is a cross-modal matrix dependent on $Q$ .", + "bbox": [ + 511, + 522, + 906, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Training Objectives", + "text_level": 1, + "bbox": [ + 511, + 564, + 702, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use two loss functions, Charbonnier loss (CL) [6] and Frequency Charbonnier-like loss (FCL). CL minimizes the spatial domain's pixel gaps, while FCL reduces the frequency domain's gaps to recover high-frequency content.", + "bbox": [ + 511, + 587, + 905, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CL is considered a local loss, measuring the absolute difference between corresponding pixels of two frames without considering global structures. CL is as follows:", + "bbox": [ + 511, + 648, + 905, + 693 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C L} = \\sqrt {\\| \\mathbf {I} _ {S R} - \\mathbf {I} _ {H R} \\| ^ {2} + \\epsilon^ {2}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 707, + 903, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, FCL utilizes Fast Fourier Transformation (FFT) [9] to convert frames from the spatial to the frequency domain. FFT utilizes the information from all pixels to obtain the FFT value for each spatial frequency. Thus, FCL is considered a global loss. Given an image $I \\in \\mathbb{R}^{H \\times W}$ of size $H \\times W$ , Discrete Fourier Transform $\\mathcal{F}$ is as follows:", + "bbox": [ + 511, + 738, + 906, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {X} (u, v) = \\mathcal {F} [ \\mathbf {I} ] = \\sum_ {x = 1} ^ {H} \\sum_ {y = 1} ^ {W} \\mathbf {I} (x, y) e ^ {- j 2 \\pi \\left(\\frac {u}{H} x + \\frac {v}{W} y\\right)} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 840, + 908, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "For simplicity, we only present the formulation for gray images.", + "bbox": [ + 529, + 886, + 874, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14715", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ae7bf1209ec806a5c4fa0172f36247f3d188da7d8bba0b060e494d42842a8f44.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFrames#Params (M)REDS4Vimeo-90K-TVid4
PSNRSSIMPSNRSSIMPSNRSSIM
EDVR [52]5/720.631.090.880037.610.948927.350.8264
VSR-T [2]5/732.631.190.881537.710.949427.360.8258
VRT [33]6/-30.731.600.8888----
PSRT-recurrent [46]6/-10.831.880.8964----
IART [54]6/-13.432.150.9010----
VSRM (ours)6/-17.132.430.9059----
BasicVSR++ [5]30/147.332.390.906937.790.950027.790.8400
VRT [33]16/735.632.190.900638.200.953027.930.8425
RVRT [32]30/1410.832.750.911338.150.952727.990.8462
PSRT-recurrent [46]16/1413.432.720.910638.270.953628.070.8485
IART [54]16/713.432.900.913838.140.952828.260.8517
VSRM (ours)16/717.133.110.916238.330.954328.440.8552
", + "bbox": [ + 158, + 88, + 831, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison (PSNR↑ and SSIM↑) of our VSRM with other state-of-the-art methods at a scale of 4x. The best and second-best results are highlighted in **bold** and **underlined**, and “-” indicates no test results.", + "bbox": [ + 89, + 325, + 906, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a471dfcfe95e3102147d6c2f48d57eca86c704ee3349cc3926e9fb6bfdeb6514.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 367, + 454, + 505 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8bc21ceb145dcbf13ddfd0ef987090625c9bffd2404adfa8fc09c3ad05ab67c2.jpg", + "image_caption": [ + "Nearest" + ], + "image_footnote": [], + "bbox": [ + 462, + 368, + 581, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a422594578d7392249a0ea86db8f3b5f2ec9f58217cd0b52aa3ed28fd0751cef.jpg", + "image_caption": [ + "VRT" + ], + "image_footnote": [], + "bbox": [ + 591, + 368, + 710, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d42a1ba3bbf4ad8f9d02aa37c893ac1d4aa67525129f329004ce3bc13c1dae8f.jpg", + "image_caption": [ + "PSRT-recurrent" + ], + "image_footnote": [], + "bbox": [ + 722, + 368, + 841, + 431 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd2c073459c0cce019827e5f75d0633c6125e853a0333dc2c0cccdf614c3181e.jpg", + "image_caption": [ + "Frame 026,Clip 000,REDS4", + "Frame 037, Clip Walk, Vid4", + "Figure 5. Qualitative comparisons on REDS4 and Vid4 datasets, VSRM achieves clearer and more precise results, revealing finer patterns." + ], + "image_footnote": [], + "bbox": [ + 158, + 513, + 452, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd5e9c0e0530663a2497a81f83cea79bb5c4f32efc0fa970bb563038807341d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 462, + 441, + 581, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c104266b210062afe17fff6e03e1aea1ed51e3bf072710b48a2a0742bc89bd17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 441, + 710, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b524355a9517c9af18ab4c86f3ddbb4453c9bd4bb6ae229ee3c3b5a21571e43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 441, + 841, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7888d498814b2355cc547e2f8b430e3f386e90a19fde289f13df0df7a24cf497.jpg", + "image_caption": [ + "IART", + "Nearest" + ], + "image_footnote": [], + "bbox": [ + 462, + 513, + 581, + 577 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2c4b508d3132bcaf5449cd7d9aa55da563186e41691bf0531bc9b4659d290a6.jpg", + "image_caption": [ + "VSRM (ours)" + ], + "image_footnote": [], + "bbox": [ + 591, + 513, + 710, + 577 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74191797c2065d203b4afc077c67a368c686c498fbdb9951f278932ab3a6f191.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 722, + 513, + 841, + 577 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d5f28409de5855caa26c365c307903fc44a5a72e7d73cbf14726f1aff35b6eed.jpg", + "image_caption": [ + "IART" + ], + "image_footnote": [], + "bbox": [ + 462, + 585, + 581, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1ac174075d28783ef63e9d3372e05ada0d324cb2e9e7ab38387dec42ad43402.jpg", + "image_caption": [ + "VRT", + "VSRM (ours)" + ], + "image_footnote": [], + "bbox": [ + 591, + 585, + 710, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/093407cb11ac512e408b7d582e98bf9d19c5e704662b527c59b7d43bbe28f992.jpg", + "image_caption": [ + "PSRT-recurrent", + "GT" + ], + "image_footnote": [], + "bbox": [ + 722, + 585, + 841, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{I}(x,y)$ is the pixel value at $(x,y)$ ; $\\mathcal{X}(u,v)$ is complex value at frequency $(u,v)$ ; $e$ and $j$ are Euler's number and the imaginary unit. $\\mathcal{F}(.)$ is a complex number with Real $(Re\\mathcal{F}(.))$ and Imaginary $(Im\\mathcal{F}(.))$ . We observe a limitation in calculating loss based on magnitude and phase. The magnitude calculation involves a square root, while the phase employs a nonlinear arctangent function. This creates discontinuities and increases sensitivity to numerical inaccuracies for VSR. To address this, we calculate frequency loss separately for the Real and Imaginary. This ensures that even minor changes in the spatial domain lead to proportional changes in these components, as they represent linear", + "bbox": [ + 89, + 707, + 485, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "decompositions of the signal in Fourier space. FCL implicitly captures phase information via R/I, letting SR/GT phase differences impact the loss. The formulation of FCL and total loss are as follows:", + "bbox": [ + 511, + 707, + 906, + 767 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F C L} = \\sum_ {i \\in \\{R e, I m \\}} \\lambda_ {i} \\sqrt {\\left\\| i \\mathcal {F} \\left(\\mathbf {I} _ {S R}\\right) - i \\mathcal {F} \\left(\\mathbf {I} _ {H R}\\right) \\right\\| ^ {2} + \\epsilon^ {2}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 775, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda \\mathcal {L} _ {C L} + \\mathcal {L} _ {F C L}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 818, + 903, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\lambda, \\lambda_{Re}$ , and $\\lambda_{Im}$ are hyperparameters; $\\lambda_{Re}, \\lambda_{Im} \\ll \\lambda$ . $\\lambda_{Re}$ , and $\\lambda_{Im}$ control the regularization strength of the high-frequency content in the output. In this work, we set $\\lambda = 1.0$ , $\\lambda_{Re} = \\lambda_{Im} = 0.02$ and $\\epsilon = 10^{-3}$ in our experiments.", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14716", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 113, + 289, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use REDS [40] and Vimeo-90K [55] as training datasets. Following [52], for REDS, we use $\\mathrm{REDS4}^2$ as a test set, and the remaining videos are used for training. For Vimeo-90K, Vid4 [35] and Vimeo-90K-T [55] are used as test sets. These two training sets have different motion movements. The motion in Vimeo-90K is generally small, whereas REDS contains large motions with more pixel movement. We study $\\times 4$ VSR task, use bicubic interpolation to produce LR video frames, and use PSNR/SSIM [53] as the evaluation metrics.", + "bbox": [ + 89, + 135, + 483, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 296, + 465, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our VSRM with the SOTA VSR methods: EDVR [52], VSR-T [2], VRT [33], BasicVSR++ [5], RVRT [32], PSRT-recurrent [46], IART [54].", + "bbox": [ + 89, + 318, + 482, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Tab. 2, VSRM achieves SOTA results across all benchmarks. Notably, VSRM surpasses VRT [33], a window-based transformer approach, by margins of up to 0.83dB and 0.90dB in PSNR for six and sixteen input frame settings on REDS4 while featuring a lower parameter count. Compared to the recent SOTA methods PSRT-recurrent[46] and IART [54], VSRM demonstrates improvements of 0.4dB and 0.28dB in PSNR for six-frame input setting and 0.39dB and 0.21dB in PSNR for sixteen-frame input setting on REDS4, though it has slightly more parameters. These findings indicate that our method balances well between complexity and performance. Additionally, VSRM is effective for slow-motion datasets, including Video-90K-T and Vid4, where our method also surpasses PSRT-recurrent [46] and IART [54] on these datasets for sixteen-frame input settings. This shows that our method works well in varying pixel movement and scenarios. Overall, these gains represent a substantial improvement in VSR.", + "bbox": [ + 89, + 364, + 483, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Results. Qualitative comparisons can be observed in Fig. 5 and Fig. 6. VSRM effectively captures intricate details. Specifically, VSRM is the only method that accurately restores the license plate numbers in the REDS4 dataset and the fine details of Spider-Man in the Vid4 dataset in Fig. 5. More examples of REDS4 and Vid4 datasets are shown in the supp. file. Fig. 6 shows the effective receptive field of our method in comparison to EDVR [52] (CNN-based) and IART [54] (Transformer-based) on REDS4. A local receptive field restricts the CNN-based methods. In contrast, Transformer-based methods can enhance the receptive field but still face limitations due to the window attention mechanism for lower computational costs. Our method stands out as it successfully achieves a significantly larger global receptive field, enhancing our model's performance in VSR.", + "bbox": [ + 89, + 651, + 483, + 878 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f1017c47218c32a92d81e23029a8215bbfb80f979e5ec56bd9e871d73f28caa.jpg", + "image_caption": [ + "Figure 6. Effective Receptive Field (ERF) [11, 38] visualization results. A more widely spread dark region signifies a larger receptive field. VSRM attains a considerable global receptive field." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 857, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 243, + 666, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies using VSRM with custom settings for efficiency to understand the contributions of each proposed module. Details settings are in the supp. file.", + "bbox": [ + 511, + 263, + 906, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/20f5fe52964f9cfb4417b6d97c032fa42dd9be069c544c717a6418d2ee578873.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPSNR (dB)Param. (M)FLOPs (G)
3D DW-Conv30.8419.49149.8
Window Attn.30.977.68152.4
Full Attn31.067.681018.1
Mamba (ours)31.098.61159.2
", + "bbox": [ + 532, + 321, + 885, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Mamba modules. Mamba modules are the core contribution of this study; they balance computational cost and high performance for inter-frame modeling. To verify its effectiveness, we replace these modules with convolutions from BasicVSR [3], window attention from IART [54], and global attention as in Tab. 3. Compared to Convolution and Window Attention methods, our method achieves notable performance gains, increasing PSNR by $0.25\\mathrm{dB}$ and $0.12\\mathrm{dB}$ , respectively, without adding significant computational overhead. Our method reaches a similar PSNR with Full Attention but is much more efficient regarding model complexity. Overall, Mamba modules balance computational efficiency and performance well.", + "bbox": [ + 511, + 424, + 906, + 621 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/264ee7ad85bc8a6b1565c0406b1230f2ff9addce7910ebbbed936d63846b09ca.jpg", + "table_caption": [ + "Table 3. Effective of Mamba modules." + ], + "table_footnote": [], + "table_body": "
StructurePSNR (dB)Param. (M)FLOPs (G)
w/o Align.30.878.53120.4
FGDA [32]30.928.70154.3
IA [54]31.008.57148.7
DCA (ours)31.098.61159.2
", + "bbox": [ + 540, + 631, + 877, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Effect of Deformable Cross-mamba Alignment module.", + "bbox": [ + 514, + 710, + 901, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of DCA. Tab. 4 compares our DCA with existing alignment modules. Removing the alignment module significantly decreases performance (↓ 0.22dB). Our DCA achieves higher PSNR than SOTA methods like FGDA [32] and IA [54] without incurring excessive computational overhead. Specifically, DCA achieves 0.17dB and 0.09dB higher PSNR than FGDA and IA, respectively, demonstrating the effectiveness of our proposed module. By leveraging the flexibility of the deformable window when computing cross-mamba, DCA enhances the precision of aligning adjacent frames to the target frame.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2Clips 000, 011, 015, 020 of REDS training set.", + "bbox": [ + 107, + 886, + 362, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14717", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/811896e2bb62af1bd611b893daeef0bb78fa36fea22ec459f32f8a8bd7d4d5cf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Structurew/o T2SMBw/ T2SMB†w/ T2SMB
PSNR (dB)30.9531.0231.09
Param. (M)7.878.658.61
FLOPs (G)155.6162.2159.2
", + "bbox": [ + 114, + 88, + 459, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of T2SMB. Tab. 5 shows the effect of the Temporal-to-Spatial Mamba block (T2SMB). Removing T2SMB considerably reduces the model's performance $(\\downarrow 0.14\\mathrm{dB})$ , highlighting the block's importance. This performance drop occurs because the Spatial-to-Temporal Mamba block (S2TMB) cannot fully extract temporal information from the sequential input features. The T2SMB block prioritizes and explicitly focuses on extracting temporal information, enabling more comprehensive final features to be generated. We also conducted an ablation study with T2SMB using both forward and backward scans. Our findings indicate that using only the forward scan for T2SMB yields better performance while requiring less computational effort.", + "bbox": [ + 89, + 200, + 483, + 398 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46d1943c0e6ccde326e545c801dc81ede6efb0ef9ea00ce2df16b5b949398102.jpg", + "table_caption": [ + "Table 5. Effect of Temporal-to-Spatial Mamba block. † indicates that both forward and backward scans are applied to T2SMB." + ], + "table_footnote": [], + "table_body": "
StructurePSNR (dB)Param. (M)FLOPs (G)
FFN [51]30.908.68136.2
TGFN w/o DWConv30.818.27124.7
TGFN w/o Split31.059.41186.1
TGFN (ours)31.098.61159.2
", + "bbox": [ + 94, + 410, + 486, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of TGFN. Tab. 6 shows the effect of the temporal gated feed-forward network (TGFN). We compare models using regular FFN [51], TGFN without 3D depthwise convolution (TGFN w/o DWConv), TGFN without split channel (TGFN w/o Split), and our TGFN. Compared to FFN, TGFN has a similar number of parameters and somewhat higher FLOPs while substantially improving performance ( $\\uparrow$ 0.19dB). Moreover, the performance suffers significantly ( $\\downarrow$ 0.28dB) when the 3D DWConv is removed from TGFN, emphasizing the necessity of both spatial and temporal information. Eliminating the split operation in TGFN leads to a slight decrease in PSNR but increases model size and complexity. This indicates that redundancy in channel features negatively impacts the models' performance.", + "bbox": [ + 89, + 513, + 483, + 726 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1124b4376994d5e9718a537e3ffd00a3f02d3d2cce7ec48ab804dcb9ad70794b.jpg", + "table_caption": [ + "Table 6. Effect of Temporal Gated Feed-forward Network." + ], + "table_footnote": [], + "table_body": "
LossFFL [23]\\( FSL^3 \\)[14]WHFL [27]FCL (ours)
00.020.2
PSNR31.0130.9930.9030.9731.0931.02
", + "bbox": [ + 106, + 737, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of FCL loss. Tab. 7 shows the effectiveness of Frequency Chabornier-like loss (FCL). We compare our loss with FFL [23], FSL [14], and WHFL [27] (default weight for these losses are 0.02 for fair comparison). Our", + "bbox": [ + 89, + 818, + 483, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FCL achieves a higher PSNR than FFL ( $\\uparrow 0.08\\mathrm{dB}$ ), indicating that balancing low- and high-frequency components and using Real and Imaginary parts for loss calculation yields more optimal results. Moreover, our loss demonstrates higher performance compared to FSL and WHFL, with clear margins. These results show the effectiveness of our proposed FCL compared to other frequency losses. Our method achieves the best results with FCL ( $\\lambda_{i} = 0.02$ where $i \\in \\{Re, Im\\}$ ). If FCL ( $\\lambda_{i} = 0$ ) is eliminated, performance substantially declines ( $\\downarrow 0.12\\mathrm{dB}$ ), highlighting the importance of this loss.", + "bbox": [ + 511, + 90, + 906, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Model Complexity Comparison", + "text_level": 1, + "bbox": [ + 511, + 263, + 792, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 8 reports comparing parameters, FLOPs, and runtime for VSRM and other VSR methods. The runtime is measured on a RTX A6000 48GB. Our VSRM is more efficient than EDVR when considering the number of parameters and FLOPs. Although it has a moderately higher complexity than Transformer-based methods that use window-based attention, these trade-offs are reasonable when balancing performance with model complexity. Furthermore, since the acceleration and optimization of Mamba are still being explored, our method has potential for further improvements.", + "bbox": [ + 511, + 287, + 906, + 439 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/506412586043fd94bf3bec64addaa116864654f8d1741d106485907a25ea19e5.jpg", + "table_caption": [ + "Table 7. Effect of Frequency Charbonnier-like loss (FCL). $i \\in \\{ Re, Im\\}$ and we set ${\\lambda }_{Re} = {\\lambda }_{Im}$ ." + ], + "table_footnote": [], + "table_body": "
MethodParam. (M)FLOPs (G)Runtime (ms)
VSR-T [2]32.6177.2-
EDVR [52]20.6323.6-
PSRT-recurrent [46]13.4164.6173
IART [54]13.4177.8180
VSRM (ours)17.1217.4223
", + "bbox": [ + 517, + 452, + 916, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Comparison of model complexity of different VSR models on the REDS4 dataset.", + "bbox": [ + 511, + 546, + 906, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 585, + 633, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce VSRM, which adapts Mamba for VSR for the first time and achieves state-of-the-art results on multiple benchmarks. Specifically, we propose Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to capture long-range spatio-temporal features effectively. Next, we propose a novel Deformable Cross-Mamba Alignment that uses a deformable cross-mamba mechanism to adapt to varying pixel movement flex to enhance model performance compared to previous alignment modules. Finally, we introduce a Frequency Charbonnier-like loss to reduce the gaps between reconstructed and ground truth frames in the frequency domain and better recover high-frequency content. This demonstrates further improvement by utilizing both spatial and frequency losses. VSRM has the potential to establish a solid baseline for upcoming research and shows great promise in substituting Transformer-based models. Furthermore, VSRM could be adaptable for other low-level vision tasks, including video colorization, deblurring, and denoising.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3FSL: Fourier Space Loss", + "bbox": [ + 107, + 886, + 250, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14718", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation(IITP)-Innovative Human Resource Development for Local Intellectualization program grant funded by the Korea government(MSIT)(IITP-2025-RS-2020-II201489). This work was also supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00573160), and the Technology Innovation Program (RS-2025-02222776) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea).", + "bbox": [ + 89, + 114, + 483, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was also supported by Hyundai Motor Chung Mong-Koo Global Scholarship to Dinh Phu Tran (first author) and generous GPU support by Korea Association for ICT Promotion.", + "bbox": [ + 89, + 280, + 483, + 339 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 364, + 187, + 382 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Runyuan Cai, Yue Ding, and Hongtao Lu. Freqnet: A frequency-domain image super-resolution network with discrete cosine transform. arXiv preprint arXiv:2111.10800, 2021. 2", + "[2] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2, 6, 7, 8", + "[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2, 5, 7", + "[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, pages 973-981, 2021. 1", + "[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1, 5, 6, 7", + "[6] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of 1st international conference on image processing, pages 168-172. IEEE, 1994. 5", + "[7] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.1", + "[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 4" + ], + "bbox": [ + 99, + 390, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] James W Cooley and John W Tukey. An algorithm for the machine calculation of complex fourier series. Mathematics of computation, 19(90):297-301, 1965. 5", + "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 1", + "[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 7", + "[12] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1", + "[13] Daniel Y. Fu, Tri Dao, Khaled K. Saab, Armin W. Thomas, Atri Rudra, and Christopher Ré. Hungry Hungry Hippos: Towards language modeling with state space models. In International Conference on Learning Representations, 2023. 2", + "[14] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 8", + "[15] Dario Fuoli, Martin Danelljan, Radu Timofte, and Luc Van Gool. Fast online video super-resolution with deformable attention pyramid. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1735-1744, 2023. 1", + "[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3", + "[17] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2", + "[18] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1", + "[19] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2", + "[20] Yukun Hao and Feihong Yu. Learning the frequency domain aliasing for real-world super-resolution. *Electronics*, 13(2): 250, 2024. 2", + "[21] Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. Transformer quality in linear time. In International conference on machine learning, pages 9099-9117. PMLR, 2022. 1", + "[22] Kui Jiang, Zhongyuan Wang, Peng Yi, and Junjun Jiang. Hierarchical dense recursive network for image superresolution. Pattern Recognition, 107:107475, 2020. 1", + "[23] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis." + ], + "bbox": [ + 514, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14719", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF international conference on computer vision, pages 13919-13929, 2021. 2, 3, 8", + "[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 2", + "[25] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE transactions on computational imaging, 2(2):109-122, 2016. 2", + "[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1", + "[27] Min Woo Kim and Nam Ik Cho. Whfl: Wavelet-domain high frequency loss for sketch-to-image translation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 744-754, 2023. 2, 8", + "[28] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 2", + "[29] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2", + "[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2", + "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 4", + "[32] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6, 7", + "[33] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Virt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 1, 2, 6, 7", + "[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1", + "[35] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE transactions on pattern analysis and machine intelligence, 36(2):346-360, 2013. 7" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34: 21297-21309, 2021. 1", + "[37] Jianping Luo, Shaofei Huang, and Yuan Yuan. Video superresolution using multi-scale pyramid 3d convolutional networks. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1882-1890, 2020. 2", + "[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 7", + "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5", + "[40] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 7", + "[41] Ali Nasiri-Sarvi, Vincent Quoc-Huy Trinh, Hassan Rivaz, and Mahdi S Hosseini. Vim4path: Self-supervised vision mamba for histopathology images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2024. 1", + "[42] Seong-Jin Park, Hyeongseok Son, Sunghyun Cho, Ki-Sang Hong, and Seungyong Lee. Srfeat: Single image superresolution with feature discrimination. In Proceedings of the European conference on computer vision (ECCV), pages 439-455, 2018. 2", + "[43] Mohammad Saeed Rad, Behzad Bozortabar, Urs-Viktor Marti, Max Basler, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Srobb: Targeted perceptual loss for single image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2710–2719, 2019. 2", + "[44] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 2", + "[45] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network, 2016. 5", + "[46] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 1, 2, 5, 6, 7, 8", + "[47] Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[48] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14720", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 1", + "[49] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2", + "[50] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2", + "[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 8", + "[52] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6, 7, 8", + "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[54] Kai Xu, Ziwei Yu, Xin Wang, Michael Bi Mi, and Angela Yao. Enhancing video super-resolution via implicit resampling-based alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2546-2555, 2024. 1, 2, 6, 7, 8", + "[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 7", + "[56] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 1, 2", + "[57] Yijun Yang, Zhaohu Xing, and Lei Zhu. Vivim: a video vision mamba for medical video object segmentation. arXiv preprint arXiv:2401.14168, 2024. 1", + "[58] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 1, 2", + "[59] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14721", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_model.json b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9851fa0dd40b932802ab4b81adee8a45a0e4eda6 --- /dev/null +++ b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_model.json @@ -0,0 +1,2631 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.131, + 0.868, + 0.152 + ], + "angle": 0, + "content": "VSRM: A Robust Mamba-Based Framework for Video Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.182, + 0.39, + 0.198 + ], + "angle": 0, + "content": "Dinh Phu Tran" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.182, + 0.559, + 0.199 + ], + "angle": 0, + "content": "Dao Duy Hung" + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.182, + 0.73, + 0.199 + ], + "angle": 0, + "content": "Daeyoung Kim" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.209, + 0.672, + 0.224 + ], + "angle": 0, + "content": "{phutx2000, hicehehe, kimd}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.23, + 0.695, + 0.248 + ], + "angle": 0, + "content": "School of Computing, KAIST, Republic of Korea" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.29, + 0.327, + 0.306 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Video super-resolution remains a major challenge in low-level vision tasks. To date, CNN- and Transformer-based methods have delivered impressive results. However, CNNs are limited by local receptive fields, while Transformers struggle with quadratic complexity, posing challenges for processing long sequences in VSR. Recently, Mamba has drawn attention for its long-sequence modeling, linear complexity, and large receptive fields. In this work, we propose VSRM, a novel Video Super-Resolution framework that leverages the power of Mamba. VSRM introduces Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to extract long-range spatio-temporal features and enhance receptive fields efficiently. To better align adjacent frames, we propose Deformable Cross-Mamba Alignment module. This module utilizes a deformable cross-mamba mechanism to make the compensation stage more dynamic and flexible, preventing feature distortions. Finally, we minimize the frequency domain gaps between reconstructed and ground-truth frames by proposing a simple yet effective Frequency Charbonnier-like loss that better preserves high-frequency content and enhances visual quality. Through extensive experiments, VSRM achieves state-of-the-art results on diverse benchmarks, establishing itself as a solid foundation for future research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.722, + 0.222, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Video super-resolution (VSR) aims to produce higher resolution from low-resolution videos. VSR typically processes long sequences and benefits from large receptive fields, efficiently capturing inter-frames information [3, 15]. However, it incurs high computational costs, requiring a balance between performance and efficiency. VSR usually uses convolutional neural network (CNN) [3, 22, 56] or transformer [33, 46, 54, 58]. Tab. 1 compares the pros and cons of these methods. CNN-based methods cannot adjust based on input data and limit receptive fields [10, 48] to capture inter-frame" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.908, + 0.593 + ], + "angle": 0, + "content": "content. Transformer-based methods show strong performance due to the power of the attention mechanism. However, full attention has quadratic complexity [21, 36], which is impractical for long sequences, while local attention struggles with limited receptive fields [12]. Recently, Gu et al. introduced Mamba [16], which can model long-range information and data-dependent characteristics with linear complexity for NLP tasks. More recently, Mamba has been successfully applied in various vision tasks [7, 18, 41, 57]. As far as we know, no research has employed Mamba to address VSR. In this work, we propose VSRM, a novel Mamba-based framework for VSR. Specifically, we propose Dual Aggregation Mamba Block (DAMB), which contains \\(N\\) Spatial-to-Temporal Mamba blocks (S2TMBs) followed by a Temporal-to-Spatial Mamba block (T2SMB) to model long-range spatio-temporal correlations for long sequences, enlarging large receptive fields. In S2TMB and T2SMB, we employ spatial-to-temporal and temporal-to-spatial scanning mechanisms to capture both spatial and temporal contexts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.868 + ], + "angle": 0, + "content": "The key challenge of VSR is effectively using the complementary information from adjacent frames, which may be misaligned due to object motions. To establish interframe correlations, many works employ the alignment module for VSR [3, 5, 46, 54]. However, most studies use bilinear or nearest-neighbor interpolation for spatial alignment due to its simplicity. Theoretically, using fixed weights (e.g., linear interpolation) causes feature distortions, leading to suboptimal results [4, 46]. To address this, Xu et al. [54] proposed an attention-based implicit interpolation for the alignment module. However, they computed interpolation in the fixed reference window, limiting flexibility and inaccuracy in aligning adjacent frames, especially in large motions between frames. Based on this observation, we introduce Deformable Cross-Mamba Alignment (DCA), which dynamically aligns neighboring frames using a deformable cross-mamba mechanism that allows information exchange between multiple selective scan modules." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Current VSR methods use pixel-based losses (e.g., L1/L2 loss [26, 34], charbonnier loss [5, 46]) or perception-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14711" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.479, + 0.173 + ], + "angle": 0, + "content": "
ModelLinear ComplexityReceptive FieldData-drivenNotable Method
CNN⊗⊗⊗BasicVSR [3]
Local Attention⊗⊗⊗IART [54]
Full Attention⊗⊗⊗-
Mamba⊗⊗⊗VSRM (ours)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.177, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Table 1. Comparison of the backbone for VSR. VSRM reaches both linear complexity and a large receptive field." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.236, + 0.485, + 0.675 + ], + "angle": 0, + "content": "based losses (e.g., perceptual loss [24, 43], adversarial loss [29, 42]). Pixel-wise losses lead to overly smooth results due to their restriction in capturing perceptual similarity. Perception-based losses enhance visual quality by recovering finer details but introduce greater distortion. As VSR is an ill-posed task, gaps between super-resolution and ground-truth frames are inevitable, sometimes visible only in the frequency domain [1, 20]. Therefore, combining loss in the spatial and frequency domains may enhance the model's performance. We propose Frequency Charbonnier-like Loss (FCL), calculated from the real and imaginary parts of frequency components in Fourier space. FCL helps our model focus more on preserving and generating high-frequency details while maintaining low-frequency content. Contribution. 1) We successfully adapt Mamba-based model for VSR for the first time and propose a new framework, VSRM. VSRM introduces Dual Aggregation Mamba Block (DAMB), which aggregates spatial and temporal features from sequences by using multi-scan direction mechanisms to obtain powerful representation ability and enlarge the model's receptive field. 2) We design Deformable Cross-mamba Alignment (DCA) module to provide a more flexible and learnable alignment through a deformable window cross-mamba mechanism in the compensation stage. 3) We propose Frequency Charbonnier-like loss (FCL), which is computed in the frequency domain to preserve and generate high-frequency details better. 4) VSRM outperforms all state-of-the-art (SOTA) methods on various benchmarks and shows the potential as the solid backbone for VSR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.233, + 0.709 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Video Super-Resolution. 1) CNN-based methods [3, 25, 37, 50, 56] is the early approach in VSR. Initially, VSR-net [25] extended CNN-based networks to VSR. Following this, TDAN [50] introduced deformable convolution alignment to enhance the alignment module's performance. In contrast, BasicVSR [3] proposed an effective model consisting solely of generic components, residual blocks, and optical flow, which surpasses all previous methods. 2) Transformer-based methods [2, 33, 46, 54, 58] achieve superior performance in VSR. Cao et al. [2] proposed spatial-temporal convolutional self-attention to explore better locality features. VRT [33] used a temporal mutual self" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "attention for a wider temporal receptive field and a parallel warp module to handle large motion. PSRT [46] introduced patch alignment that aligns image patches instead of pixels to maintain intra-patch relationships and preserve sub-pixel information. Based on this, Xu et al. [54] proposed an implicit alignment module based on cross-attention mechanism to adjust feature representations across frames, further improving performance. These methods applied window-based attention with small window to avoid the quadratic complexity of full attention but were limited by local receptive fields, affecting cross-frame information capture. Recent methods, like PSRT [46] and IART [54], focused on designing window-based alignment modules rather than improving the transformer backbone to enhance performance. They employed a fixed window for implicit motion compensation, limiting the flexibility of alignment modules, especially when minimal or excessive motion. To address these issues, this study proposes a new framework, VSRM, based on Mamba to leverage its advantage in large receptive fields and linear complexity. We also design Deformable Cross-mamba Alignment for more flexible motion compensation using a deformable mechanism to boost performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.701 + ], + "angle": 0, + "content": "State Space Model. SSMs [13, 16, 17, 47] show potential to replace Transformers. Gu et al. introduced S4 [17] that achieves linear complexity modeling of long sequences, followed by S5[47] with improved capacity and efficiency. Mamba [16], built upon SSMs, has recently surpassed Transformers on many benchmarks in the NLP domain due to its data-dependent characteristics and global receptive fields. In computer vision, Vim [59] adapted Mamba for images by converting them into sequences and using a bidirectional mamba block for global modeling. VideoMamba [30] processed video by rearranging frames into spatio-temporal sequences, achieving comparable results to Transformer-based methods. MambaIR [19] adopted Mamba for image restoration tasks, demonstrating that Mamba can work well for low-level vision tasks, such as video deblurring, denoising, etc. Expanding on that progress, we are pioneers in investigating the capabilities of Mamba for VSR, utilizing its large receptive fields and linear complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Spectral bias. Many studies reveal spectral bias [39, 44, 49] that models bias toward low-frequency rather than high-frequency content. Meanwhile, high-frequency content (e.g., edges, textures) is vital in low-level vision tasks [14, 23, 27, 28]. Hence, Jiang et al. proposed Focal Frequency Loss (FFL) [23] computed in the frequency domain to emphasize high-frequency content and enhance visual quality. Based on this, Fuoli et al. [14] employed a GAN loss in Fourier space. In contrast, WHFL [27] used Wavelet transform, incorporating a module for multi-bandwidth analysis. These methods employ a frequency-adjusted reweight matrix to optimize the low and high-frequency elements. However, the reweight matrix does not" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14712" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.168, + 0.088, + 0.835, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.367, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 1. The network architecture of our method. a) Video Super-Resolution based on Mamba (VSRM). b) Feature Propagation block (FPB) consists of Deformable Cross-mamba Alignment (DCA) and Dual Aggregation Mamba Blocks (DAMB). c) Spatial-to-Temporal Mamba block (S2TMB). d) Temporal-to-Spatial Mamba block (T2SMB)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.542 + ], + "angle": 0, + "content": "effectively balance the gradients of low and high frequencies. Furthermore, images mainly contain low-frequency content, so overemphasis can harm performance. Hence, we propose Frequency Charbonnier-like loss (FCL), which directly computes loss on the Real and Imaginary parts in Fourier space. Unlike FFL [23], FCL treats low- and high-frequency components equally." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.405, + 0.574 + ], + "angle": 0, + "content": "3. Preliminaries: State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.645 + ], + "angle": 0, + "content": "SSMs draw their inspiration from continuous linear time-invariant systems, which map a 1D function or sequence \\( x(t) \\in \\mathbb{R} \\to y(t) \\in \\mathbb{R} \\) via a hidden state \\( h(t) \\in \\mathbb{R}^N \\). This is formulated as ordinary differential equations (ODE):" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.656, + 0.483, + 0.673 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = A h (t) + B x (t), y (t) = C h (t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.792 + ], + "angle": 0, + "content": "where \\(A \\in \\mathbb{R}^{N \\times N}\\) is the evolution matrix, while \\(B \\in \\mathbb{R}^{N \\times 1}\\) and \\(C \\in \\mathbb{R}^{1 \\times N}\\) are the projection parameters. To model discrete sequences like images, we need to approximate them via discretization. SSMs utilize the zero-order hold method, incorporating a timescale parameter \\(\\Delta\\) to map the continuous parameters \\(A\\), \\(B\\) to discrete parameters \\(\\bar{A}\\), \\(\\bar{B}\\). The process of discretization is outlined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.804, + 0.483, + 0.822 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\bar {A} h _ {t - 1} + \\bar {B} x _ {t}, y _ {t} = C h _ {t}, \\tag {2a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.824, + 0.483, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\bar {A} = e x p (\\Delta A), \\bar {B} = (\\Delta A) ^ {- 1} (e x p (\\Delta A - I)) \\Delta B, (2 \\mathrm {b})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The parameters of these SSMs are independent of the data, which means that \\( A \\), \\( B \\), and \\( C \\) remain constant regardless of the input, thereby restricting their adaptability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.907, + 0.466 + ], + "angle": 0, + "content": "in sequence modeling. Recently, Mamba [16] proposes selective SSMs (S6) with input data-dependent \\( x_{i} \\in \\mathbb{R}^{L} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.475, + 0.907, + 0.492 + ], + "angle": 0, + "content": "\\[\nB _ {i} = S _ {B} x _ {i}, C _ {i} = S _ {c} x _ {i}, \\Delta_ {i} = S o f t p l u s (S _ {\\Delta} x _ {i}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.577 + ], + "angle": 0, + "content": "where \\(S_B, S_C \\in \\mathbb{R}^{N \\times L}\\) and \\(S_{\\Delta} \\in \\mathbb{R}^{L \\times L}\\) are linear projection layers. \\(B_i\\) and \\(C_i\\) are utilized across for all channels of \\(x_i\\), \\(\\Delta x_i\\) contains \\(\\Delta\\) of \\(L\\) channels, and \\(A\\) remains consistent with earlier SSMs. Via the discretization presented in equations 3, \\(\\bar{A}\\) and \\(\\bar{B}\\) become input data-dependent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.65, + 0.607 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.69 + ], + "angle": 0, + "content": "In this section, we introduce VSRM architecture and its core components: Dual Aggregation Mamba block (DAMB) and Deformable Cross-mamba Alignment (DCA), as shown in Fig. 1. Furthermore, we present the training objectives to optimize our model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.625, + 0.713 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Let \\( I_{t}^{LR} \\in \\mathbb{R}^{T \\times H \\times W \\times C} \\) be the t-th low-resolution (LR) video frame, while \\( I_{t}^{HR} \\in \\mathbb{R}^{T \\times sH \\times sW \\times C} \\) denotes the t-th corresponding ground-truth (GT or HR) frame. Here, \\( T, H, W, C \\), and \\( s \\) indicate the total number of frames, height, width, channels, and the upscaling factor, respectively. Our aim is to generate the super-resolution video frame \\( I_{t}^{SR} \\in \\mathbb{R}^{T \\times sH \\times sW \\times C} \\) using the consecutive \\( 2N + 1 \\) frames \\( \\{I_{i = t - N}^{t + N}\\} \\), ensuring that \\( I_{t}^{SR} \\) and \\( I_{t}^{HR} \\) are as closely as possible. VSRM is depicted in Fig. 1a) includes two parts: Feature Extraction (Conv2d and Feature Propagation Block) and Upsampler (Reconstruction). Feature extraction extracts deep features from input frames and aligns and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14713" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.462, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.22, + 0.485, + 0.276 + ], + "angle": 0, + "content": "Figure 2. Structure and scan directions of Mamba module. a) Spatial-to-Temporal Mamba (S2T-Mamba). b) Temporal-to-Spatial Mamba (T2S-Mamba). For simplicity, we exclude the initial normalization and the final residual." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.348 + ], + "angle": 0, + "content": "fuses the features of adjacent frames to leverage inter-frame relationships. The upsampler upscales the fused spatiotemporal features to produce the output sequence frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.357, + 0.377, + 0.373 + ], + "angle": 0, + "content": "4.2. Dual Aggregation Mamba Block" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Dual Aggregation Mamba block (DAMB) is the core component of VSRM that consists of Spatial-to-Temporal Mamba block (S2TMB) and Temporal-to-Spatial Mamba block (T2SMB), as shown in Fig. 1(c, d). S2TMB and T2SMB are based on two modules: Spatial-to-Temporal Mamba (S2T-Mamba) and Temporal-to-Spatial Mamba (T2S-Mamba), with their structures shown in Fig. 2. By organizing \\(N\\) S2TMBs followed by a T2SMB in DAMB, VSRM fully extracts spatio-temporal information in both spatial and temporal dimensions. Moreover, we propose Temporal Gated Feed-forward Network (TGFN) to explore spatial and temporal information better." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.772 + ], + "angle": 0, + "content": "Spatial-to-Temporal Mamba. S2T-Mamba structure and scan directions are shown in Fig. 2(a). A 3D sequence (\\(T\\) frames) is converted into a 1D sequence and processed in two directions: forward (ForwardSSM(.)) and backward (BackwardSSM(.)) paths. This approach follows a mechanism that first scans spatially and then temporally, moving from frame 1 to frame \\(T\\). This enables both spatial and temporal awareness during processing. The original Mamba was designed for 1D sequences, so this bidirectional scan mechanism helps preserve spatial awareness. By employing the scanning process from spatial to temporal in both the forward and backward paths, our module is optimized for handling 3D sequences more effectively. Given an input \\(X_{in}\\), the output of S2T-Mamba is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.782, + 0.483, + 0.799 + ], + "angle": 0, + "content": "\\[\nx = \\operatorname {L i n e a r} \\left(X _ {i n}\\right), z = \\sigma \\left(\\operatorname {L i n e a r} \\left(X _ {i n}\\right)\\right), \\tag {4a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.803, + 0.482, + 0.82 + ], + "angle": 0, + "content": "\\[\nx _ {1} = \\text {F o r w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {4b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.824, + 0.482, + 0.839 + ], + "angle": 0, + "content": "\\[\nx _ {2} = \\text {B a c k w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {4c}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.844, + 0.482, + 0.86 + ], + "angle": 0, + "content": "\\[\nS 2 T - M a m b a (x, z) = \\operatorname {L i n e a r} \\left(x _ {1} \\odot z + x _ {2} \\odot z\\right) \\tag {4d}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Temporal-to-Spatial Mamba. T2S-Mamba structure and scan direction are shown in Fig. 2(b). T2S-Mamba also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.318 + ], + "angle": 0, + "content": "converts 3D sequence into a 1D sequence. However, we only perform the scanning in forward direction for two reasons. First, forward scan in T2S-Mamba models feature changes over time, while adding backward scan disrupts learned dependencies, leading to a slight decrease in performance, and an increase in FLOPs. Second, our experiment shows that using only forward scan in T2S-Mamba yields better results than adding backward pass, shown in Tab. 5. We found that S2T-Mamba prioritizes extracting spatial rather than temporal information, leading to not extracting fully temporal information. In contrast, T2S-Mamba explicitly prioritizes scanning temporal rather than spatial information, allowing for a more comprehensive feature extraction. Given an input \\( X_{in} \\), the output of S2T-Mamba is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.33, + 0.905, + 0.346 + ], + "angle": 0, + "content": "\\[\nx = \\operatorname {L i n e a r} \\left(X _ {i n}\\right), z = \\sigma \\left(\\operatorname {L i n e a r} \\left(X _ {i n}\\right)\\right), \\tag {5a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.351, + 0.905, + 0.366 + ], + "angle": 0, + "content": "\\[\nx _ {1} = \\text {F o r w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {5b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.371, + 0.905, + 0.387 + ], + "angle": 0, + "content": "\\[\nT 2 S - M a m b a (x, z) = \\operatorname {L i n e a r} \\left(x _ {1} \\odot z\\right) \\tag {5c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Temporal Gated Feed-forward Network. Feed-forward network (FFN) [51] consists of a nonlinear activation positioned between two linear projection layers. It disregards spatial and temporal relationships and lacks feature alignment since it processes each input token individually. Moreover, the redundant information across channels hinders progress in learning feature representations. To address this issues, we propose Temporal Gated Feed-forward Network (TGFN) shown in Fig. 3. It has two main updates to FFN: 1) 3D depthwise convolutions (DW-3D Conv) and 2) a gating mechanism. We use DW-3D Conv to model relationships effectively from spatially and temporally neighboring pixel positions, which is vital in VSR. TGFN employs a gating mechanism by performing element-wise multiplication on two parallel feature branches obtained from the chunking method, with one of them passed by the GELU, which has become the preferred option in recent work [8, 31]. Given the input \\( X \\in \\mathbb{R}^{T \\times H \\times W \\times C} \\), TGFN is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.68, + 0.905, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\hat {X} = W _ {p} ^ {1} X, \\quad [ \\hat {X} _ {1}, \\hat {X} _ {2} ] = \\hat {X} \\tag {6a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.704, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\nT G F N (X) = W _ {p} ^ {2} \\left(W _ {d} ^ {1} L N \\left(\\hat {X} _ {1}\\right) \\odot \\sigma \\left(W _ {d} ^ {2} L N \\left(\\hat {X} _ {2}\\right)\\right)\\right) \\tag {6b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.825 + ], + "angle": 0, + "content": "where \\( W_{p}^{1}, W_{p}^{2} \\) are weights of linear projections; \\( W_{d}^{1}, W_{d}^{2} \\) are weights of the 3D DWConv layers; \\( \\sigma \\) is GELU activation, \\( C \\) is the hidden dimension, LN is LayerNorm. Overall, TGFN adds nonlinear spatial and temporal information more effectively and optimizes the information flow through our pipeline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.838, + 0.85 + ], + "angle": 0, + "content": "4.3. Deformable Cross-mamba Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Alignment modules in VSR often fail when pixel movement between consecutive frames is minimal or excessive. This requires an alignment module adaptable to varying pixel" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14714" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.088, + 0.408, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.175, + 0.461, + 0.19 + ], + "angle": 0, + "content": "Figure 3. Temporal Gated Feed-forward Network (TGFN)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.484, + 0.382 + ], + "angle": 0, + "content": "movement. Building upon [3, 5, 46], we design Deformable Cross-mamba Alignment (DCA), as shown in Fig. 4. DCA uses the pre-trained SpyNet [45] for flow estimation while incorporating a new motion compensation scheme. Specifically, we use a deformable window scheme to perform a more flexible warping operation in the compensation stage between frames according to inter-frame motion information. We then employ a cross-mamba module, which enables interactions between the target point and the values of the dynamic reference points to infer the aligned point, leading to more adaptive and accurate alignment results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.444 + ], + "angle": 0, + "content": "For each pixel \\(X(x,y)\\) in the target frame \\(X\\) (red point), let \\((\\Delta_x,\\Delta_y)\\) be the optical flow estimated between \\(X\\) and the reference frame \\(X^{\\prime}\\). The aligned frame \\(\\bar{X}\\) is obtained by a spatial warping function \\(\\mathcal{W}\\) with its parameters \\(\\theta_{\\mathcal{W}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.453, + 0.483, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\bar {X} (x, y) = \\mathcal {W} \\left(X ^ {\\prime}, (x + \\Delta x, y + \\Delta y); \\theta_ {\\mathcal {W}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.603 + ], + "angle": 0, + "content": "In the warping operation, we learn offset \\(\\epsilon_r\\) from window \\(w\\) to identify dynamic reference region \\(\\bar{r}\\). We extract \\(w \\in \\mathbb{R}^{H_w \\times W_w \\times C}\\), which is centered at the sampling position \\((x', y') = (x + \\Delta x, y + \\Delta y)\\) in \\(X'\\). Here, \\(H_w\\), \\(W_w\\), and \\(C\\) are the height, width, and number of channels of the window, respectively, as shown in Fig. 4. Within this window, we initialize a reference region \\(r \\in \\mathbb{R}^{H_r \\times W_r \\times 2}\\), by iterating over \\(i \\in (0, 1, \\dots, H_r)\\) and \\(j \\in (0, 1, \\dots, W_r)\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.613, + 0.483, + 0.632 + ], + "angle": 0, + "content": "\\[\nr (i, j) = X ^ {\\prime} \\left(\\lfloor x ^ {\\prime} \\rfloor - \\lfloor \\frac {W _ {r} - 1}{2} \\rfloor + i, \\lfloor y ^ {\\prime} \\rfloor - \\lfloor \\frac {H _ {r} - 1}{2} \\rfloor + j\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.733 + ], + "angle": 0, + "content": "To obtain the dynamic reference region \\(\\bar{r}\\), we refine \\(r\\) using a small learnable offset network \\(\\mathcal{S}(.)\\), which contains only convolution and activation layers (Fig. 4). This network learns the reference offset \\(\\epsilon_r\\) from the features within the window \\(w\\). The final adaptive reference map is computed using bilinear interpolation \\(\\phi(.;.)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.744, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\bar {r} = \\phi (w; r + \\epsilon_ {r}), \\quad \\text {w i t h} \\epsilon_ {r} = S (w), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.832 + ], + "angle": 0, + "content": "Finally, we apply implicit warping within the dynamic reference region to align each pixel by selectively cross-fusing reference point values with the target grid value at \\( X(x,y) \\) using the cross-mamba module cross-mamba(.,.):" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.843, + 0.483, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\bar {X} (x, y) = \\text {c r o s s - m a m b a} (R, Q), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(R\\) is the reference tensor and \\(Q\\) is the target tensor, both augmented with positional information. The reference" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.891, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.249, + 0.907, + 0.333 + ], + "angle": 0, + "content": "Figure 4. Pipeline of our proposed Deformable Cross-mamba Alignment (DCA). We initialize a fixed reference region \\( r \\) within the window \\( w \\). This region is then adjusted to a dynamic reference region \\( \\bar{r} \\) using an offset \\( \\epsilon_r \\), which is learned from the features within \\( w \\). Finally, DCA learns the affinity through the cross-mamba module to calculate the final result (aligned point)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.482 + ], + "angle": 0, + "content": "tensor \\( R \\) is computed by adding the flattened vector \\( \\bar{r} \\) to the positional encoding \\( PE_{R} \\), and the target tensor \\( Q \\) is obtained by adding the target value \\( X(x,y) \\) to the positional embedding \\( PE_{Q} \\). We use sinusoidal positional encoding to capture high-frequency content, which maps continuous input coordinates into a high-dimensional space [39] (more details in the supp. file). Based on Eq. (2a), the cross-mamba module computes output \\( \\bar{X} \\) at each step \\( t \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.493, + 0.907, + 0.512 + ], + "angle": 0, + "content": "\\[\nH _ {t} = \\bar {A} _ {R} H _ {t - 1} + \\bar {B} _ {R} \\bar {R} _ {t}, \\bar {X} _ {t} = C _ {Q} H _ {t}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.907, + 0.556 + ], + "angle": 0, + "content": "where \\(\\bar{A}_R\\), \\(\\bar{B}_R\\) are matrices dependent on \\(R\\), and \\(C_Q\\) is a cross-modal matrix dependent on \\(Q\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.704, + 0.582 + ], + "angle": 0, + "content": "4.4. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.906, + 0.649 + ], + "angle": 0, + "content": "We use two loss functions, Charbonnier loss (CL) [6] and Frequency Charbonnier-like loss (FCL). CL minimizes the spatial domain's pixel gaps, while FCL reduces the frequency domain's gaps to recover high-frequency content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.906, + 0.694 + ], + "angle": 0, + "content": "CL is considered a local loss, measuring the absolute difference between corresponding pixels of two frames without considering global structures. CL is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.708, + 0.905, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C L} = \\sqrt {\\| \\mathbf {I} _ {S R} - \\mathbf {I} _ {H R} \\| ^ {2} + \\epsilon^ {2}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.739, + 0.907, + 0.83 + ], + "angle": 0, + "content": "In contrast, FCL utilizes Fast Fourier Transformation (FFT) [9] to convert frames from the spatial to the frequency domain. FFT utilizes the information from all pixels to obtain the FFT value for each spatial frequency. Thus, FCL is considered a global loss. Given an image \\( I \\in \\mathbb{R}^{H \\times W} \\) of size \\( H \\times W \\), Discrete Fourier Transform \\( \\mathcal{F} \\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.841, + 0.91, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathcal {X} (u, v) = \\mathcal {F} [ \\mathbf {I} ] = \\sum_ {x = 1} ^ {H} \\sum_ {y = 1} ^ {W} \\mathbf {I} (x, y) e ^ {- j 2 \\pi \\left(\\frac {u}{H} x + \\frac {v}{W} y\\right)} \\tag {13}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.875, + 0.901 + ], + "angle": 0, + "content": "For simplicity, we only present the formulation for gray images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14715" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.833, + 0.324 + ], + "angle": 0, + "content": "
MethodFrames#Params (M)REDS4Vimeo-90K-TVid4
PSNRSSIMPSNRSSIMPSNRSSIM
EDVR [52]5/720.631.090.880037.610.948927.350.8264
VSR-T [2]5/732.631.190.881537.710.949427.360.8258
VRT [33]6/-30.731.600.8888----
PSRT-recurrent [46]6/-10.831.880.8964----
IART [54]6/-13.432.150.9010----
VSRM (ours)6/-17.132.430.9059----
BasicVSR++ [5]30/147.332.390.906937.790.950027.790.8400
VRT [33]16/735.632.190.900638.200.953027.930.8425
RVRT [32]30/1410.832.750.911338.150.952727.990.8462
PSRT-recurrent [46]16/1413.432.720.910638.270.953628.070.8485
IART [54]16/713.432.900.913838.140.952828.260.8517
VSRM (ours)16/717.133.110.916238.330.954328.440.8552
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.327, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison (PSNR↑ and SSIM↑) of our VSRM with other state-of-the-art methods at a scale of 4x. The best and second-best results are highlighted in **bold** and **underlined**, and “-” indicates no test results." + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.368, + 0.455, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.507, + 0.387, + 0.514 + ], + "angle": 0, + "content": "Frame 026,Clip 000,REDS4" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.369, + 0.582, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.433, + 0.535, + 0.44 + ], + "angle": 0, + "content": "Nearest" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.369, + 0.712, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.433, + 0.664, + 0.44 + ], + "angle": 0, + "content": "VRT" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.369, + 0.842, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.433, + 0.819, + 0.44 + ], + "angle": 0, + "content": "PSRT-recurrent" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.515, + 0.453, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.652, + 0.369, + 0.661 + ], + "angle": 0, + "content": "Frame 037, Clip Walk, Vid4" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.442, + 0.582, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.507, + 0.535, + 0.514 + ], + "angle": 0, + "content": "IART" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.442, + 0.712, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.507, + 0.683, + 0.514 + ], + "angle": 0, + "content": "VSRM (ours)" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.442, + 0.842, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.507, + 0.79, + 0.514 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.515, + 0.582, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.579, + 0.541, + 0.585 + ], + "angle": 0, + "content": "Nearest" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.515, + 0.712, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.579, + 0.662, + 0.586 + ], + "angle": 0, + "content": "VRT" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.515, + 0.842, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.579, + 0.819, + 0.586 + ], + "angle": 0, + "content": "PSRT-recurrent" + }, + { + "type": "image", + "bbox": [ + 0.464, + 0.587, + 0.582, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.652, + 0.535, + 0.659 + ], + "angle": 0, + "content": "IART" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.587, + 0.712, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.652, + 0.683, + 0.66 + ], + "angle": 0, + "content": "VSRM (ours)" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.587, + 0.842, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.652, + 0.791, + 0.659 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.668, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons on REDS4 and Vid4 datasets, VSRM achieves clearer and more precise results, revealing finer patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.486, + 0.89 + ], + "angle": 0, + "content": "\\(\\mathbf{I}(x,y)\\) is the pixel value at \\((x,y)\\); \\(\\mathcal{X}(u,v)\\) is complex value at frequency \\((u,v)\\); \\(e\\) and \\(j\\) are Euler's number and the imaginary unit. \\(\\mathcal{F}(.)\\) is a complex number with Real \\((Re\\mathcal{F}(.))\\) and Imaginary \\((Im\\mathcal{F}(.))\\). We observe a limitation in calculating loss based on magnitude and phase. The magnitude calculation involves a square root, while the phase employs a nonlinear arctangent function. This creates discontinuities and increases sensitivity to numerical inaccuracies for VSR. To address this, we calculate frequency loss separately for the Real and Imaginary. This ensures that even minor changes in the spatial domain lead to proportional changes in these components, as they represent linear" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.907, + 0.768 + ], + "angle": 0, + "content": "decompositions of the signal in Fourier space. FCL implicitly captures phase information via R/I, letting SR/GT phase differences impact the loss. The formulation of FCL and total loss are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.776, + 0.907, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F C L} = \\sum_ {i \\in \\{R e, I m \\}} \\lambda_ {i} \\sqrt {\\left\\| i \\mathcal {F} \\left(\\mathbf {I} _ {S R}\\right) - i \\mathcal {F} \\left(\\mathbf {I} _ {H R}\\right) \\right\\| ^ {2} + \\epsilon^ {2}} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.819, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda \\mathcal {L} _ {C L} + \\mathcal {L} _ {F C L}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "\\(\\lambda, \\lambda_{Re}\\), and \\(\\lambda_{Im}\\) are hyperparameters; \\(\\lambda_{Re}, \\lambda_{Im} \\ll \\lambda\\). \\(\\lambda_{Re}\\), and \\(\\lambda_{Im}\\) control the regularization strength of the high-frequency content in the output. In this work, we set \\(\\lambda = 1.0\\), \\(\\lambda_{Re} = \\lambda_{Im} = 0.02\\) and \\(\\epsilon = 10^{-3}\\) in our experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14716" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.114, + 0.29, + 0.129 + ], + "angle": 0, + "content": "5.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.136, + 0.484, + 0.288 + ], + "angle": 0, + "content": "We use REDS [40] and Vimeo-90K [55] as training datasets. Following [52], for REDS, we use \\(\\mathrm{REDS4}^2\\) as a test set, and the remaining videos are used for training. For Vimeo-90K, Vid4 [35] and Vimeo-90K-T [55] are used as test sets. These two training sets have different motion movements. The motion in Vimeo-90K is generally small, whereas REDS contains large motions with more pixel movement. We study \\(\\times 4\\) VSR task, use bicubic interpolation to produce LR video frames, and use PSNR/SSIM [53] as the evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.297, + 0.466, + 0.314 + ], + "angle": 0, + "content": "5.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.365 + ], + "angle": 0, + "content": "We compare our VSRM with the SOTA VSR methods: EDVR [52], VSR-T [2], VRT [33], BasicVSR++ [5], RVRT [32], PSRT-recurrent [46], IART [54]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Tab. 2, VSRM achieves SOTA results across all benchmarks. Notably, VSRM surpasses VRT [33], a window-based transformer approach, by margins of up to 0.83dB and 0.90dB in PSNR for six and sixteen input frame settings on REDS4 while featuring a lower parameter count. Compared to the recent SOTA methods PSRT-recurrent[46] and IART [54], VSRM demonstrates improvements of 0.4dB and 0.28dB in PSNR for six-frame input setting and 0.39dB and 0.21dB in PSNR for sixteen-frame input setting on REDS4, though it has slightly more parameters. These findings indicate that our method balances well between complexity and performance. Additionally, VSRM is effective for slow-motion datasets, including Video-90K-T and Vid4, where our method also surpasses PSRT-recurrent [46] and IART [54] on these datasets for sixteen-frame input settings. This shows that our method works well in varying pixel movement and scenarios. Overall, these gains represent a substantial improvement in VSR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Visual Results. Qualitative comparisons can be observed in Fig. 5 and Fig. 6. VSRM effectively captures intricate details. Specifically, VSRM is the only method that accurately restores the license plate numbers in the REDS4 dataset and the fine details of Spider-Man in the Vid4 dataset in Fig. 5. More examples of REDS4 and Vid4 datasets are shown in the supp. file. Fig. 6 shows the effective receptive field of our method in comparison to EDVR [52] (CNN-based) and IART [54] (Transformer-based) on REDS4. A local receptive field restricts the CNN-based methods. In contrast, Transformer-based methods can enhance the receptive field but still face limitations due to the window attention mechanism for lower computational costs. Our method stands out as it successfully achieves a significantly larger global receptive field, enhancing our model's performance in VSR." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.858, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.196, + 0.907, + 0.239 + ], + "angle": 0, + "content": "Figure 6. Effective Receptive Field (ERF) [11, 38] visualization results. A more widely spread dark region signifies a larger receptive field. VSRM attains a considerable global receptive field." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.244, + 0.667, + 0.26 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.312 + ], + "angle": 0, + "content": "We conduct ablation studies using VSRM with custom settings for efficiency to understand the contributions of each proposed module. Details settings are in the supp. file." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.322, + 0.887, + 0.399 + ], + "angle": 0, + "content": "
ModelPSNR (dB)Param. (M)FLOPs (G)
3D DW-Conv30.8419.49149.8
Window Attn.30.977.68152.4
Full Attn31.067.681018.1
Mamba (ours)31.098.61159.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.594, + 0.402, + 0.825, + 0.415 + ], + "angle": 0, + "content": "Table 3. Effective of Mamba modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Effect of Mamba modules. Mamba modules are the core contribution of this study; they balance computational cost and high performance for inter-frame modeling. To verify its effectiveness, we replace these modules with convolutions from BasicVSR [3], window attention from IART [54], and global attention as in Tab. 3. Compared to Convolution and Window Attention methods, our method achieves notable performance gains, increasing PSNR by \\(0.25\\mathrm{dB}\\) and \\(0.12\\mathrm{dB}\\), respectively, without adding significant computational overhead. Our method reaches a similar PSNR with Full Attention but is much more efficient regarding model complexity. Overall, Mamba modules balance computational efficiency and performance well." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.632, + 0.879, + 0.709 + ], + "angle": 0, + "content": "
StructurePSNR (dB)Param. (M)FLOPs (G)
w/o Align.30.878.53120.4
FGDA [32]30.928.70154.3
IA [54]31.008.57148.7
DCA (ours)31.098.61159.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.712, + 0.902, + 0.727 + ], + "angle": 0, + "content": "Table 4. Effect of Deformable Cross-mamba Alignment module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effect of DCA. Tab. 4 compares our DCA with existing alignment modules. Removing the alignment module significantly decreases performance (↓ 0.22dB). Our DCA achieves higher PSNR than SOTA methods like FGDA [32] and IA [54] without incurring excessive computational overhead. Specifically, DCA achieves 0.17dB and 0.09dB higher PSNR than FGDA and IA, respectively, demonstrating the effectiveness of our proposed module. By leveraging the flexibility of the deformable window when computing cross-mamba, DCA enhances the precision of aligning adjacent frames to the target frame." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.364, + 0.901 + ], + "angle": 0, + "content": "2Clips 000, 011, 015, 020 of REDS training set." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14717" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.46, + 0.152 + ], + "angle": 0, + "content": "
Structurew/o T2SMBw/ T2SMB†w/ T2SMB
PSNR (dB)30.9531.0231.09
Param. (M)7.878.658.61
FLOPs (G)155.6162.2159.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.155, + 0.483, + 0.185 + ], + "angle": 0, + "content": "Table 5. Effect of Temporal-to-Spatial Mamba block. † indicates that both forward and backward scans are applied to T2SMB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Effect of T2SMB. Tab. 5 shows the effect of the Temporal-to-Spatial Mamba block (T2SMB). Removing T2SMB considerably reduces the model's performance \\((\\downarrow 0.14\\mathrm{dB})\\), highlighting the block's importance. This performance drop occurs because the Spatial-to-Temporal Mamba block (S2TMB) cannot fully extract temporal information from the sequential input features. The T2SMB block prioritizes and explicitly focuses on extracting temporal information, enabling more comprehensive final features to be generated. We also conducted an ablation study with T2SMB using both forward and backward scans. Our findings indicate that using only the forward scan for T2SMB yields better performance while requiring less computational effort." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.411, + 0.487, + 0.488 + ], + "angle": 0, + "content": "
StructurePSNR (dB)Param. (M)FLOPs (G)
FFN [51]30.908.68136.2
TGFN w/o DWConv30.818.27124.7
TGFN w/o Split31.059.41186.1
TGFN (ours)31.098.61159.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.492, + 0.461, + 0.506 + ], + "angle": 0, + "content": "Table 6. Effect of Temporal Gated Feed-forward Network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.727 + ], + "angle": 0, + "content": "Effect of TGFN. Tab. 6 shows the effect of the temporal gated feed-forward network (TGFN). We compare models using regular FFN [51], TGFN without 3D depthwise convolution (TGFN w/o DWConv), TGFN without split channel (TGFN w/o Split), and our TGFN. Compared to FFN, TGFN has a similar number of parameters and somewhat higher FLOPs while substantially improving performance (\\(\\uparrow\\)0.19dB). Moreover, the performance suffers significantly (\\(\\downarrow\\)0.28dB) when the 3D DWConv is removed from TGFN, emphasizing the necessity of both spatial and temporal information. Eliminating the split operation in TGFN leads to a slight decrease in PSNR but increases model size and complexity. This indicates that redundancy in channel features negatively impacts the models' performance." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.738, + 0.47, + 0.788 + ], + "angle": 0, + "content": "
LossFFL [23]\\( FSL^3 \\)[14]WHFL [27]FCL (ours)
00.020.2
PSNR31.0130.9930.9030.9731.0931.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.791, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Table 7. Effect of Frequency Charbonnier-like loss (FCL). \\( i \\in \\{ Re, Im\\} \\) and we set \\( {\\lambda }_{Re} = {\\lambda }_{Im} \\) ." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Effect of FCL loss. Tab. 7 shows the effectiveness of Frequency Chabornier-like loss (FCL). We compare our loss with FFL [23], FSL [14], and WHFL [27] (default weight for these losses are 0.02 for fair comparison). Our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.258 + ], + "angle": 0, + "content": "FCL achieves a higher PSNR than FFL (\\(\\uparrow 0.08\\mathrm{dB}\\)), indicating that balancing low- and high-frequency components and using Real and Imaginary parts for loss calculation yields more optimal results. Moreover, our loss demonstrates higher performance compared to FSL and WHFL, with clear margins. These results show the effectiveness of our proposed FCL compared to other frequency losses. Our method achieves the best results with FCL (\\(\\lambda_{i} = 0.02\\) where \\(i \\in \\{Re, Im\\}\\)). If FCL (\\(\\lambda_{i} = 0\\)) is eliminated, performance substantially declines (\\(\\downarrow 0.12\\mathrm{dB}\\)), highlighting the importance of this loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.265, + 0.794, + 0.282 + ], + "angle": 0, + "content": "5.4. Model Complexity Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Tab. 8 reports comparing parameters, FLOPs, and runtime for VSRM and other VSR methods. The runtime is measured on a RTX A6000 48GB. Our VSRM is more efficient than EDVR when considering the number of parameters and FLOPs. Although it has a moderately higher complexity than Transformer-based methods that use window-based attention, these trade-offs are reasonable when balancing performance with model complexity. Furthermore, since the acceleration and optimization of Mamba are still being explored, our method has potential for further improvements." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.453, + 0.918, + 0.543 + ], + "angle": 0, + "content": "
MethodParam. (M)FLOPs (G)Runtime (ms)
VSR-T [2]32.6177.2-
EDVR [52]20.6323.6-
PSRT-recurrent [46]13.4164.6173
IART [54]13.4177.8180
VSRM (ours)17.1217.4223
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.547, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Table 8. Comparison of model complexity of different VSR models on the REDS4 dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.634, + 0.603 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce VSRM, which adapts Mamba for VSR for the first time and achieves state-of-the-art results on multiple benchmarks. Specifically, we propose Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to capture long-range spatio-temporal features effectively. Next, we propose a novel Deformable Cross-Mamba Alignment that uses a deformable cross-mamba mechanism to adapt to varying pixel movement flex to enhance model performance compared to previous alignment modules. Finally, we introduce a Frequency Charbonnier-like loss to reduce the gaps between reconstructed and ground truth frames in the frequency domain and better recover high-frequency content. This demonstrates further improvement by utilizing both spatial and frequency losses. VSRM has the potential to establish a solid baseline for upcoming research and shows great promise in substituting Transformer-based models. Furthermore, VSRM could be adaptable for other low-level vision tasks, including video colorization, deblurring, and denoising." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.251, + 0.901 + ], + "angle": 0, + "content": "3FSL: Fourier Space Loss" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14718" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.281 + ], + "angle": 0, + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation(IITP)-Innovative Human Resource Development for Local Intellectualization program grant funded by the Korea government(MSIT)(IITP-2025-RS-2020-II201489). This work was also supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00573160), and the Technology Innovation Program (RS-2025-02222776) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.34 + ], + "angle": 0, + "content": "This work was also supported by Hyundai Motor Chung Mong-Koo Global Scholarship to Dinh Phu Tran (first author) and generous GPU support by Korea Association for ICT Promotion." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.366, + 0.188, + 0.383 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.391, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[1] Runyuan Cai, Yue Ding, and Hongtao Lu. Freqnet: A frequency-domain image super-resolution network with discrete cosine transform. arXiv preprint arXiv:2111.10800, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.448, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[2] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.491, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, pages 973-981, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[6] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of 1st international conference on image processing, pages 168-172. IEEE, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[7] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.391, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[9] James W Cooley and John W Tukey. An algorithm for the machine calculation of complex fourier series. Mathematics of computation, 19(90):297-301, 1965. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.908, + 0.343 + ], + "angle": 0, + "content": "[12] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[13] Daniel Y. Fu, Tri Dao, Khaled K. Saab, Armin W. Thomas, Atri Rudra, and Christopher Ré. Hungry Hungry Hippos: Towards language modeling with state space models. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[14] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[15] Dario Fuoli, Martin Danelljan, Radu Timofte, and Luc Van Gool. Fast online video super-resolution with deformable attention pyramid. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1735-1744, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[17] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[18] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[19] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[20] Yukun Hao and Feihong Yu. Learning the frequency domain aliasing for real-world super-resolution. *Electronics*, 13(2): 250, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[21] Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. Transformer quality in linear time. In International conference on machine learning, pages 9099-9117. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[22] Kui Jiang, Zhongyuan Wang, Peng Yi, and Junjun Jiang. Hierarchical dense recursive network for image superresolution. Pattern Recognition, 107:107475, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[23] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14719" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF international conference on computer vision, pages 13919-13929, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[25] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE transactions on computational imaging, 2(2):109-122, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.306 + ], + "angle": 0, + "content": "[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.483, + 0.362 + ], + "angle": 0, + "content": "[27] Min Woo Kim and Nam Ik Cho. Whfl: Wavelet-domain high frequency loss for sketch-to-image translation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 744-754, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.364, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[28] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[29] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.533, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.577, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[32] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[33] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Virt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE transactions on pattern analysis and machine intelligence, 36(2):346-360, 2013. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[36] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34: 21297-21309, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[37] Jianping Luo, Shaofei Huang, and Yuan Yuan. Video superresolution using multi-scale pyramid 3d convolutional networks. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1882-1890, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[40] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[41] Ali Nasiri-Sarvi, Vincent Quoc-Huy Trinh, Hassan Rivaz, and Mahdi S Hosseini. Vim4path: Self-supervised vision mamba for histopathology images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[42] Seong-Jin Park, Hyeongseok Son, Sunghyun Cho, Ki-Sang Hong, and Seungyong Lee. Srfeat: Single image superresolution with feature discrimination. In Proceedings of the European conference on computer vision (ECCV), pages 439-455, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[43] Mohammad Saeed Rad, Behzad Bozortabar, Urs-Viktor Marti, Max Basler, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Srobb: Targeted perceptual loss for single image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2710–2719, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[44] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[45] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[46] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[47] Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14720" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[49] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[50] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.291, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[52] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.515 + ], + "angle": 0, + "content": "[54] Kai Xu, Ziwei Yu, Xin Wang, Michael Bi Mi, and Angela Yao. Enhancing video super-resolution via implicit resampling-based alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2546-2555, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.574, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[56] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[57] Yijun Yang, Zhaohu Xing, and Lei Zhu. Vivim: a video vision mamba for medical video object segmentation. arXiv preprint arXiv:2401.14168, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[58] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[59] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14721" + } + ] +] \ No newline at end of file diff --git a/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_origin.pdf b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4b63e9e005842f501d6337827f5d67e0f71c662 --- /dev/null +++ b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/6e341a1e-2a3f-4224-8262-4e1ca75e0c71_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83af4535a024de4fb0d7f59ed438b91f1f67a0dbfd8547920c2ccfa9aa86f39a +size 10385672 diff --git a/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/full.md b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc6025363f1b0ae8b72b76954f7eb8eef280b48c --- /dev/null +++ b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/full.md @@ -0,0 +1,399 @@ +# VSRM: A Robust Mamba-Based Framework for Video Super-Resolution + +Dinh Phu Tran + +Dao Duy Hung + +Daeyoung Kim + +{phutx2000, hicehehe, kimd}@kaist.ac.kr + +School of Computing, KAIST, Republic of Korea + +# Abstract + +Video super-resolution remains a major challenge in low-level vision tasks. To date, CNN- and Transformer-based methods have delivered impressive results. However, CNNs are limited by local receptive fields, while Transformers struggle with quadratic complexity, posing challenges for processing long sequences in VSR. Recently, Mamba has drawn attention for its long-sequence modeling, linear complexity, and large receptive fields. In this work, we propose VSRM, a novel Video Super-Resolution framework that leverages the power of Mamba. VSRM introduces Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to extract long-range spatio-temporal features and enhance receptive fields efficiently. To better align adjacent frames, we propose Deformable Cross-Mamba Alignment module. This module utilizes a deformable cross-mamba mechanism to make the compensation stage more dynamic and flexible, preventing feature distortions. Finally, we minimize the frequency domain gaps between reconstructed and ground-truth frames by proposing a simple yet effective Frequency Charbonnier-like loss that better preserves high-frequency content and enhances visual quality. Through extensive experiments, VSRM achieves state-of-the-art results on diverse benchmarks, establishing itself as a solid foundation for future research. + +# 1. Introduction + +Video super-resolution (VSR) aims to produce higher resolution from low-resolution videos. VSR typically processes long sequences and benefits from large receptive fields, efficiently capturing inter-frames information [3, 15]. However, it incurs high computational costs, requiring a balance between performance and efficiency. VSR usually uses convolutional neural network (CNN) [3, 22, 56] or transformer [33, 46, 54, 58]. Tab. 1 compares the pros and cons of these methods. CNN-based methods cannot adjust based on input data and limit receptive fields [10, 48] to capture inter-frame + +content. Transformer-based methods show strong performance due to the power of the attention mechanism. However, full attention has quadratic complexity [21, 36], which is impractical for long sequences, while local attention struggles with limited receptive fields [12]. Recently, Gu et al. introduced Mamba [16], which can model long-range information and data-dependent characteristics with linear complexity for NLP tasks. More recently, Mamba has been successfully applied in various vision tasks [7, 18, 41, 57]. As far as we know, no research has employed Mamba to address VSR. In this work, we propose VSRM, a novel Mamba-based framework for VSR. Specifically, we propose Dual Aggregation Mamba Block (DAMB), which contains $N$ Spatial-to-Temporal Mamba blocks (S2TMBs) followed by a Temporal-to-Spatial Mamba block (T2SMB) to model long-range spatio-temporal correlations for long sequences, enlarging large receptive fields. In S2TMB and T2SMB, we employ spatial-to-temporal and temporal-to-spatial scanning mechanisms to capture both spatial and temporal contexts. + +The key challenge of VSR is effectively using the complementary information from adjacent frames, which may be misaligned due to object motions. To establish interframe correlations, many works employ the alignment module for VSR [3, 5, 46, 54]. However, most studies use bilinear or nearest-neighbor interpolation for spatial alignment due to its simplicity. Theoretically, using fixed weights (e.g., linear interpolation) causes feature distortions, leading to suboptimal results [4, 46]. To address this, Xu et al. [54] proposed an attention-based implicit interpolation for the alignment module. However, they computed interpolation in the fixed reference window, limiting flexibility and inaccuracy in aligning adjacent frames, especially in large motions between frames. Based on this observation, we introduce Deformable Cross-Mamba Alignment (DCA), which dynamically aligns neighboring frames using a deformable cross-mamba mechanism that allows information exchange between multiple selective scan modules. + +Current VSR methods use pixel-based losses (e.g., L1/L2 loss [26, 34], charbonnier loss [5, 46]) or perception- + +
ModelLinear ComplexityReceptive FieldData-drivenNotable Method
CNN⊗⊗⊗BasicVSR [3]
Local Attention⊗⊗⊗IART [54]
Full Attention⊗⊗⊗-
Mamba⊗⊗⊗VSRM (ours)
+ +Table 1. Comparison of the backbone for VSR. VSRM reaches both linear complexity and a large receptive field. + +based losses (e.g., perceptual loss [24, 43], adversarial loss [29, 42]). Pixel-wise losses lead to overly smooth results due to their restriction in capturing perceptual similarity. Perception-based losses enhance visual quality by recovering finer details but introduce greater distortion. As VSR is an ill-posed task, gaps between super-resolution and ground-truth frames are inevitable, sometimes visible only in the frequency domain [1, 20]. Therefore, combining loss in the spatial and frequency domains may enhance the model's performance. We propose Frequency Charbonnier-like Loss (FCL), calculated from the real and imaginary parts of frequency components in Fourier space. FCL helps our model focus more on preserving and generating high-frequency details while maintaining low-frequency content. Contribution. 1) We successfully adapt Mamba-based model for VSR for the first time and propose a new framework, VSRM. VSRM introduces Dual Aggregation Mamba Block (DAMB), which aggregates spatial and temporal features from sequences by using multi-scan direction mechanisms to obtain powerful representation ability and enlarge the model's receptive field. 2) We design Deformable Cross-mamba Alignment (DCA) module to provide a more flexible and learnable alignment through a deformable window cross-mamba mechanism in the compensation stage. 3) We propose Frequency Charbonnier-like loss (FCL), which is computed in the frequency domain to preserve and generate high-frequency details better. 4) VSRM outperforms all state-of-the-art (SOTA) methods on various benchmarks and shows the potential as the solid backbone for VSR. + +# 2. Related Work + +Video Super-Resolution. 1) CNN-based methods [3, 25, 37, 50, 56] is the early approach in VSR. Initially, VSR-net [25] extended CNN-based networks to VSR. Following this, TDAN [50] introduced deformable convolution alignment to enhance the alignment module's performance. In contrast, BasicVSR [3] proposed an effective model consisting solely of generic components, residual blocks, and optical flow, which surpasses all previous methods. 2) Transformer-based methods [2, 33, 46, 54, 58] achieve superior performance in VSR. Cao et al. [2] proposed spatial-temporal convolutional self-attention to explore better locality features. VRT [33] used a temporal mutual self + +attention for a wider temporal receptive field and a parallel warp module to handle large motion. PSRT [46] introduced patch alignment that aligns image patches instead of pixels to maintain intra-patch relationships and preserve sub-pixel information. Based on this, Xu et al. [54] proposed an implicit alignment module based on cross-attention mechanism to adjust feature representations across frames, further improving performance. These methods applied window-based attention with small window to avoid the quadratic complexity of full attention but were limited by local receptive fields, affecting cross-frame information capture. Recent methods, like PSRT [46] and IART [54], focused on designing window-based alignment modules rather than improving the transformer backbone to enhance performance. They employed a fixed window for implicit motion compensation, limiting the flexibility of alignment modules, especially when minimal or excessive motion. To address these issues, this study proposes a new framework, VSRM, based on Mamba to leverage its advantage in large receptive fields and linear complexity. We also design Deformable Cross-mamba Alignment for more flexible motion compensation using a deformable mechanism to boost performance. + +State Space Model. SSMs [13, 16, 17, 47] show potential to replace Transformers. Gu et al. introduced S4 [17] that achieves linear complexity modeling of long sequences, followed by S5[47] with improved capacity and efficiency. Mamba [16], built upon SSMs, has recently surpassed Transformers on many benchmarks in the NLP domain due to its data-dependent characteristics and global receptive fields. In computer vision, Vim [59] adapted Mamba for images by converting them into sequences and using a bidirectional mamba block for global modeling. VideoMamba [30] processed video by rearranging frames into spatio-temporal sequences, achieving comparable results to Transformer-based methods. MambaIR [19] adopted Mamba for image restoration tasks, demonstrating that Mamba can work well for low-level vision tasks, such as video deblurring, denoising, etc. Expanding on that progress, we are pioneers in investigating the capabilities of Mamba for VSR, utilizing its large receptive fields and linear complexity. + +Spectral bias. Many studies reveal spectral bias [39, 44, 49] that models bias toward low-frequency rather than high-frequency content. Meanwhile, high-frequency content (e.g., edges, textures) is vital in low-level vision tasks [14, 23, 27, 28]. Hence, Jiang et al. proposed Focal Frequency Loss (FFL) [23] computed in the frequency domain to emphasize high-frequency content and enhance visual quality. Based on this, Fuoli et al. [14] employed a GAN loss in Fourier space. In contrast, WHFL [27] used Wavelet transform, incorporating a module for multi-bandwidth analysis. These methods employ a frequency-adjusted reweight matrix to optimize the low and high-frequency elements. However, the reweight matrix does not + +![](images/99d88d026dc4c7dedd45aa5916d292f0beadd65f98f4d7d7127dcba9768ea7d8.jpg) +Figure 1. The network architecture of our method. a) Video Super-Resolution based on Mamba (VSRM). b) Feature Propagation block (FPB) consists of Deformable Cross-mamba Alignment (DCA) and Dual Aggregation Mamba Blocks (DAMB). c) Spatial-to-Temporal Mamba block (S2TMB). d) Temporal-to-Spatial Mamba block (T2SMB). + +effectively balance the gradients of low and high frequencies. Furthermore, images mainly contain low-frequency content, so overemphasis can harm performance. Hence, we propose Frequency Charbonnier-like loss (FCL), which directly computes loss on the Real and Imaginary parts in Fourier space. Unlike FFL [23], FCL treats low- and high-frequency components equally. + +# 3. Preliminaries: State Space Models + +SSMs draw their inspiration from continuous linear time-invariant systems, which map a 1D function or sequence $x(t) \in \mathbb{R} \to y(t) \in \mathbb{R}$ via a hidden state $h(t) \in \mathbb{R}^N$ . This is formulated as ordinary differential equations (ODE): + +$$ +h ^ {\prime} (t) = A h (t) + B x (t), y (t) = C h (t), \tag {1} +$$ + +where $A \in \mathbb{R}^{N \times N}$ is the evolution matrix, while $B \in \mathbb{R}^{N \times 1}$ and $C \in \mathbb{R}^{1 \times N}$ are the projection parameters. To model discrete sequences like images, we need to approximate them via discretization. SSMs utilize the zero-order hold method, incorporating a timescale parameter $\Delta$ to map the continuous parameters $A$ , $B$ to discrete parameters $\bar{A}$ , $\bar{B}$ . The process of discretization is outlined as follows: + +$$ +h _ {t} = \bar {A} h _ {t - 1} + \bar {B} x _ {t}, y _ {t} = C h _ {t}, \tag {2a} +$$ + +$$ +\bar {A} = e x p (\Delta A), \bar {B} = (\Delta A) ^ {- 1} (e x p (\Delta A - I)) \Delta B, (2 \mathrm {b}) +$$ + +The parameters of these SSMs are independent of the data, which means that $A$ , $B$ , and $C$ remain constant regardless of the input, thereby restricting their adaptability + +in sequence modeling. Recently, Mamba [16] proposes selective SSMs (S6) with input data-dependent $x_{i} \in \mathbb{R}^{L}$ : + +$$ +B _ {i} = S _ {B} x _ {i}, C _ {i} = S _ {c} x _ {i}, \Delta_ {i} = S o f t p l u s (S _ {\Delta} x _ {i}), \tag {3} +$$ + +where $S_B, S_C \in \mathbb{R}^{N \times L}$ and $S_{\Delta} \in \mathbb{R}^{L \times L}$ are linear projection layers. $B_i$ and $C_i$ are utilized across for all channels of $x_i$ , $\Delta x_i$ contains $\Delta$ of $L$ channels, and $A$ remains consistent with earlier SSMs. Via the discretization presented in equations 3, $\bar{A}$ and $\bar{B}$ become input data-dependent. + +# 4. Methodology + +In this section, we introduce VSRM architecture and its core components: Dual Aggregation Mamba block (DAMB) and Deformable Cross-mamba Alignment (DCA), as shown in Fig. 1. Furthermore, we present the training objectives to optimize our model. + +# 4.1. Overview + +Let $I_{t}^{LR} \in \mathbb{R}^{T \times H \times W \times C}$ be the t-th low-resolution (LR) video frame, while $I_{t}^{HR} \in \mathbb{R}^{T \times sH \times sW \times C}$ denotes the t-th corresponding ground-truth (GT or HR) frame. Here, $T, H, W, C$ , and $s$ indicate the total number of frames, height, width, channels, and the upscaling factor, respectively. Our aim is to generate the super-resolution video frame $I_{t}^{SR} \in \mathbb{R}^{T \times sH \times sW \times C}$ using the consecutive $2N + 1$ frames $\{I_{i = t - N}^{t + N}\}$ , ensuring that $I_{t}^{SR}$ and $I_{t}^{HR}$ are as closely as possible. VSRM is depicted in Fig. 1a) includes two parts: Feature Extraction (Conv2d and Feature Propagation Block) and Upsampler (Reconstruction). Feature extraction extracts deep features from input frames and aligns and + +![](images/1f9f6aae8a4c0bece79d5b087eef5cd9f33d3bba43405624c7bb8d529809b446.jpg) +Figure 2. Structure and scan directions of Mamba module. a) Spatial-to-Temporal Mamba (S2T-Mamba). b) Temporal-to-Spatial Mamba (T2S-Mamba). For simplicity, we exclude the initial normalization and the final residual. + +fuses the features of adjacent frames to leverage inter-frame relationships. The upsampler upscales the fused spatiotemporal features to produce the output sequence frames. + +# 4.2. Dual Aggregation Mamba Block + +Dual Aggregation Mamba block (DAMB) is the core component of VSRM that consists of Spatial-to-Temporal Mamba block (S2TMB) and Temporal-to-Spatial Mamba block (T2SMB), as shown in Fig. 1(c, d). S2TMB and T2SMB are based on two modules: Spatial-to-Temporal Mamba (S2T-Mamba) and Temporal-to-Spatial Mamba (T2S-Mamba), with their structures shown in Fig. 2. By organizing $N$ S2TMBs followed by a T2SMB in DAMB, VSRM fully extracts spatio-temporal information in both spatial and temporal dimensions. Moreover, we propose Temporal Gated Feed-forward Network (TGFN) to explore spatial and temporal information better. + +Spatial-to-Temporal Mamba. S2T-Mamba structure and scan directions are shown in Fig. 2(a). A 3D sequence ( $T$ frames) is converted into a 1D sequence and processed in two directions: forward (ForwardSSM(.)) and backward (BackwardSSM(.)) paths. This approach follows a mechanism that first scans spatially and then temporally, moving from frame 1 to frame $T$ . This enables both spatial and temporal awareness during processing. The original Mamba was designed for 1D sequences, so this bidirectional scan mechanism helps preserve spatial awareness. By employing the scanning process from spatial to temporal in both the forward and backward paths, our module is optimized for handling 3D sequences more effectively. Given an input $X_{in}$ , the output of S2T-Mamba is computed as: + +$$ +x = \operatorname {L i n e a r} \left(X _ {i n}\right), z = \sigma \left(\operatorname {L i n e a r} \left(X _ {i n}\right)\right), \tag {4a} +$$ + +$$ +x _ {1} = \text {F o r w a r d S S M} (\operatorname {C o n v 1} d (x)), \tag {4b} +$$ + +$$ +x _ {2} = \text {B a c k w a r d S S M} (\operatorname {C o n v 1} d (x)), \tag {4c} +$$ + +$$ +S 2 T - M a m b a (x, z) = \operatorname {L i n e a r} \left(x _ {1} \odot z + x _ {2} \odot z\right) \tag {4d} +$$ + +Temporal-to-Spatial Mamba. T2S-Mamba structure and scan direction are shown in Fig. 2(b). T2S-Mamba also + +converts 3D sequence into a 1D sequence. However, we only perform the scanning in forward direction for two reasons. First, forward scan in T2S-Mamba models feature changes over time, while adding backward scan disrupts learned dependencies, leading to a slight decrease in performance, and an increase in FLOPs. Second, our experiment shows that using only forward scan in T2S-Mamba yields better results than adding backward pass, shown in Tab. 5. We found that S2T-Mamba prioritizes extracting spatial rather than temporal information, leading to not extracting fully temporal information. In contrast, T2S-Mamba explicitly prioritizes scanning temporal rather than spatial information, allowing for a more comprehensive feature extraction. Given an input $X_{in}$ , the output of S2T-Mamba is calculated as: + +$$ +x = \operatorname {L i n e a r} \left(X _ {i n}\right), z = \sigma \left(\operatorname {L i n e a r} \left(X _ {i n}\right)\right), \tag {5a} +$$ + +$$ +x _ {1} = \text {F o r w a r d S S M} (\operatorname {C o n v 1} d (x)), \tag {5b} +$$ + +$$ +T 2 S - M a m b a (x, z) = \operatorname {L i n e a r} \left(x _ {1} \odot z\right) \tag {5c} +$$ + +Temporal Gated Feed-forward Network. Feed-forward network (FFN) [51] consists of a nonlinear activation positioned between two linear projection layers. It disregards spatial and temporal relationships and lacks feature alignment since it processes each input token individually. Moreover, the redundant information across channels hinders progress in learning feature representations. To address this issues, we propose Temporal Gated Feed-forward Network (TGFN) shown in Fig. 3. It has two main updates to FFN: 1) 3D depthwise convolutions (DW-3D Conv) and 2) a gating mechanism. We use DW-3D Conv to model relationships effectively from spatially and temporally neighboring pixel positions, which is vital in VSR. TGFN employs a gating mechanism by performing element-wise multiplication on two parallel feature branches obtained from the chunking method, with one of them passed by the GELU, which has become the preferred option in recent work [8, 31]. Given the input $X \in \mathbb{R}^{T \times H \times W \times C}$ , TGFN is formulated as: + +$$ +\hat {X} = W _ {p} ^ {1} X, \quad [ \hat {X} _ {1}, \hat {X} _ {2} ] = \hat {X} \tag {6a} +$$ + +$$ +T G F N (X) = W _ {p} ^ {2} \left(W _ {d} ^ {1} L N \left(\hat {X} _ {1}\right) \odot \sigma \left(W _ {d} ^ {2} L N \left(\hat {X} _ {2}\right)\right)\right) \tag {6b} +$$ + +where $W_{p}^{1}, W_{p}^{2}$ are weights of linear projections; $W_{d}^{1}, W_{d}^{2}$ are weights of the 3D DWConv layers; $\sigma$ is GELU activation, $C$ is the hidden dimension, LN is LayerNorm. Overall, TGFN adds nonlinear spatial and temporal information more effectively and optimizes the information flow through our pipeline. + +# 4.3. Deformable Cross-mamba Alignment + +Alignment modules in VSR often fail when pixel movement between consecutive frames is minimal or excessive. This requires an alignment module adaptable to varying pixel + +![](images/c1aab8293864997f4a89d1f695b60631c93dee61d8739bdf44be0bca0a155469.jpg) +Figure 3. Temporal Gated Feed-forward Network (TGFN) + +movement. Building upon [3, 5, 46], we design Deformable Cross-mamba Alignment (DCA), as shown in Fig. 4. DCA uses the pre-trained SpyNet [45] for flow estimation while incorporating a new motion compensation scheme. Specifically, we use a deformable window scheme to perform a more flexible warping operation in the compensation stage between frames according to inter-frame motion information. We then employ a cross-mamba module, which enables interactions between the target point and the values of the dynamic reference points to infer the aligned point, leading to more adaptive and accurate alignment results. + +For each pixel $X(x,y)$ in the target frame $X$ (red point), let $(\Delta_x,\Delta_y)$ be the optical flow estimated between $X$ and the reference frame $X^{\prime}$ . The aligned frame $\bar{X}$ is obtained by a spatial warping function $\mathcal{W}$ with its parameters $\theta_{\mathcal{W}}$ as: + +$$ +\bar {X} (x, y) = \mathcal {W} \left(X ^ {\prime}, (x + \Delta x, y + \Delta y); \theta_ {\mathcal {W}}\right), \tag {7} +$$ + +In the warping operation, we learn offset $\epsilon_r$ from window $w$ to identify dynamic reference region $\bar{r}$ . We extract $w \in \mathbb{R}^{H_w \times W_w \times C}$ , which is centered at the sampling position $(x', y') = (x + \Delta x, y + \Delta y)$ in $X'$ . Here, $H_w$ , $W_w$ , and $C$ are the height, width, and number of channels of the window, respectively, as shown in Fig. 4. Within this window, we initialize a reference region $r \in \mathbb{R}^{H_r \times W_r \times 2}$ , by iterating over $i \in (0, 1, \dots, H_r)$ and $j \in (0, 1, \dots, W_r)$ , as follows: + +$$ +r (i, j) = X ^ {\prime} \left(\lfloor x ^ {\prime} \rfloor - \lfloor \frac {W _ {r} - 1}{2} \rfloor + i, \lfloor y ^ {\prime} \rfloor - \lfloor \frac {H _ {r} - 1}{2} \rfloor + j\right) \tag {8} +$$ + +To obtain the dynamic reference region $\bar{r}$ , we refine $r$ using a small learnable offset network $\mathcal{S}(.)$ , which contains only convolution and activation layers (Fig. 4). This network learns the reference offset $\epsilon_r$ from the features within the window $w$ . The final adaptive reference map is computed using bilinear interpolation $\phi(.;.)$ as follows: + +$$ +\bar {r} = \phi (w; r + \epsilon_ {r}), \quad \text {w i t h} \epsilon_ {r} = S (w), \tag {9} +$$ + +Finally, we apply implicit warping within the dynamic reference region to align each pixel by selectively cross-fusing reference point values with the target grid value at $X(x,y)$ using the cross-mamba module cross-mamba(.,.): + +$$ +\bar {X} (x, y) = \text {c r o s s - m a m b a} (R, Q), \tag {10} +$$ + +where $R$ is the reference tensor and $Q$ is the target tensor, both augmented with positional information. The reference + +![](images/6621aaa41fc3a79f2bd20b77ea92c80c835c73b701c40ea0c9f2b06e22ebcf35.jpg) +Figure 4. Pipeline of our proposed Deformable Cross-mamba Alignment (DCA). We initialize a fixed reference region $r$ within the window $w$ . This region is then adjusted to a dynamic reference region $\bar{r}$ using an offset $\epsilon_r$ , which is learned from the features within $w$ . Finally, DCA learns the affinity through the cross-mamba module to calculate the final result (aligned point). + +tensor $R$ is computed by adding the flattened vector $\bar{r}$ to the positional encoding $PE_{R}$ , and the target tensor $Q$ is obtained by adding the target value $X(x,y)$ to the positional embedding $PE_{Q}$ . We use sinusoidal positional encoding to capture high-frequency content, which maps continuous input coordinates into a high-dimensional space [39] (more details in the supp. file). Based on Eq. (2a), the cross-mamba module computes output $\bar{X}$ at each step $t$ as: + +$$ +H _ {t} = \bar {A} _ {R} H _ {t - 1} + \bar {B} _ {R} \bar {R} _ {t}, \bar {X} _ {t} = C _ {Q} H _ {t}, \tag {11} +$$ + +where $\bar{A}_R$ , $\bar{B}_R$ are matrices dependent on $R$ , and $C_Q$ is a cross-modal matrix dependent on $Q$ . + +# 4.4. Training Objectives + +We use two loss functions, Charbonnier loss (CL) [6] and Frequency Charbonnier-like loss (FCL). CL minimizes the spatial domain's pixel gaps, while FCL reduces the frequency domain's gaps to recover high-frequency content. + +CL is considered a local loss, measuring the absolute difference between corresponding pixels of two frames without considering global structures. CL is as follows: + +$$ +\mathcal {L} _ {C L} = \sqrt {\| \mathbf {I} _ {S R} - \mathbf {I} _ {H R} \| ^ {2} + \epsilon^ {2}}, \tag {12} +$$ + +In contrast, FCL utilizes Fast Fourier Transformation (FFT) [9] to convert frames from the spatial to the frequency domain. FFT utilizes the information from all pixels to obtain the FFT value for each spatial frequency. Thus, FCL is considered a global loss. Given an image $I \in \mathbb{R}^{H \times W}$ of size $H \times W$ , Discrete Fourier Transform $\mathcal{F}$ is as follows: + +$$ +\mathcal {X} (u, v) = \mathcal {F} [ \mathbf {I} ] = \sum_ {x = 1} ^ {H} \sum_ {y = 1} ^ {W} \mathbf {I} (x, y) e ^ {- j 2 \pi \left(\frac {u}{H} x + \frac {v}{W} y\right)} \tag {13} +$$ + +
MethodFrames#Params (M)REDS4Vimeo-90K-TVid4
PSNRSSIMPSNRSSIMPSNRSSIM
EDVR [52]5/720.631.090.880037.610.948927.350.8264
VSR-T [2]5/732.631.190.881537.710.949427.360.8258
VRT [33]6/-30.731.600.8888----
PSRT-recurrent [46]6/-10.831.880.8964----
IART [54]6/-13.432.150.9010----
VSRM (ours)6/-17.132.430.9059----
BasicVSR++ [5]30/147.332.390.906937.790.950027.790.8400
VRT [33]16/735.632.190.900638.200.953027.930.8425
RVRT [32]30/1410.832.750.911338.150.952727.990.8462
PSRT-recurrent [46]16/1413.432.720.910638.270.953628.070.8485
IART [54]16/713.432.900.913838.140.952828.260.8517
VSRM (ours)16/717.133.110.916238.330.954328.440.8552
+ +Table 2. Quantitative comparison (PSNR↑ and SSIM↑) of our VSRM with other state-of-the-art methods at a scale of 4x. The best and second-best results are highlighted in **bold** and **underlined**, and “-” indicates no test results. + +![](images/a471dfcfe95e3102147d6c2f48d57eca86c704ee3349cc3926e9fb6bfdeb6514.jpg) + +![](images/8bc21ceb145dcbf13ddfd0ef987090625c9bffd2404adfa8fc09c3ad05ab67c2.jpg) +Nearest + +![](images/a422594578d7392249a0ea86db8f3b5f2ec9f58217cd0b52aa3ed28fd0751cef.jpg) +VRT + +![](images/d42a1ba3bbf4ad8f9d02aa37c893ac1d4aa67525129f329004ce3bc13c1dae8f.jpg) +PSRT-recurrent + +![](images/cd2c073459c0cce019827e5f75d0633c6125e853a0333dc2c0cccdf614c3181e.jpg) +Frame 026,Clip 000,REDS4 +Frame 037, Clip Walk, Vid4 +Figure 5. Qualitative comparisons on REDS4 and Vid4 datasets, VSRM achieves clearer and more precise results, revealing finer patterns. + +![](images/cd5e9c0e0530663a2497a81f83cea79bb5c4f32efc0fa970bb563038807341d9.jpg) + +![](images/c104266b210062afe17fff6e03e1aea1ed51e3bf072710b48a2a0742bc89bd17.jpg) + +![](images/2b524355a9517c9af18ab4c86f3ddbb4453c9bd4bb6ae229ee3c3b5a21571e43.jpg) + +![](images/7888d498814b2355cc547e2f8b430e3f386e90a19fde289f13df0df7a24cf497.jpg) +IART +Nearest + +![](images/e2c4b508d3132bcaf5449cd7d9aa55da563186e41691bf0531bc9b4659d290a6.jpg) +VSRM (ours) + +![](images/74191797c2065d203b4afc077c67a368c686c498fbdb9951f278932ab3a6f191.jpg) +GT + +![](images/d5f28409de5855caa26c365c307903fc44a5a72e7d73cbf14726f1aff35b6eed.jpg) +IART + +![](images/d1ac174075d28783ef63e9d3372e05ada0d324cb2e9e7ab38387dec42ad43402.jpg) +VRT +VSRM (ours) + +![](images/093407cb11ac512e408b7d582e98bf9d19c5e704662b527c59b7d43bbe28f992.jpg) +PSRT-recurrent +GT + +$\mathbf{I}(x,y)$ is the pixel value at $(x,y)$ ; $\mathcal{X}(u,v)$ is complex value at frequency $(u,v)$ ; $e$ and $j$ are Euler's number and the imaginary unit. $\mathcal{F}(.)$ is a complex number with Real $(Re\mathcal{F}(.))$ and Imaginary $(Im\mathcal{F}(.))$ . We observe a limitation in calculating loss based on magnitude and phase. The magnitude calculation involves a square root, while the phase employs a nonlinear arctangent function. This creates discontinuities and increases sensitivity to numerical inaccuracies for VSR. To address this, we calculate frequency loss separately for the Real and Imaginary. This ensures that even minor changes in the spatial domain lead to proportional changes in these components, as they represent linear + +decompositions of the signal in Fourier space. FCL implicitly captures phase information via R/I, letting SR/GT phase differences impact the loss. The formulation of FCL and total loss are as follows: + +$$ +\mathcal {L} _ {F C L} = \sum_ {i \in \{R e, I m \}} \lambda_ {i} \sqrt {\left\| i \mathcal {F} \left(\mathbf {I} _ {S R}\right) - i \mathcal {F} \left(\mathbf {I} _ {H R}\right) \right\| ^ {2} + \epsilon^ {2}} \tag {14} +$$ + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda \mathcal {L} _ {C L} + \mathcal {L} _ {F C L}, \tag {15} +$$ + +$\lambda, \lambda_{Re}$ , and $\lambda_{Im}$ are hyperparameters; $\lambda_{Re}, \lambda_{Im} \ll \lambda$ . $\lambda_{Re}$ , and $\lambda_{Im}$ control the regularization strength of the high-frequency content in the output. In this work, we set $\lambda = 1.0$ , $\lambda_{Re} = \lambda_{Im} = 0.02$ and $\epsilon = 10^{-3}$ in our experiments. + +# 5. Experiments + +# 5.1. Datasets and Metrics + +We use REDS [40] and Vimeo-90K [55] as training datasets. Following [52], for REDS, we use $\mathrm{REDS4}^2$ as a test set, and the remaining videos are used for training. For Vimeo-90K, Vid4 [35] and Vimeo-90K-T [55] are used as test sets. These two training sets have different motion movements. The motion in Vimeo-90K is generally small, whereas REDS contains large motions with more pixel movement. We study $\times 4$ VSR task, use bicubic interpolation to produce LR video frames, and use PSNR/SSIM [53] as the evaluation metrics. + +# 5.2. Comparisons with State-of-the-Art Methods + +We compare our VSRM with the SOTA VSR methods: EDVR [52], VSR-T [2], VRT [33], BasicVSR++ [5], RVRT [32], PSRT-recurrent [46], IART [54]. + +Quantitative Results. As shown in Tab. 2, VSRM achieves SOTA results across all benchmarks. Notably, VSRM surpasses VRT [33], a window-based transformer approach, by margins of up to 0.83dB and 0.90dB in PSNR for six and sixteen input frame settings on REDS4 while featuring a lower parameter count. Compared to the recent SOTA methods PSRT-recurrent[46] and IART [54], VSRM demonstrates improvements of 0.4dB and 0.28dB in PSNR for six-frame input setting and 0.39dB and 0.21dB in PSNR for sixteen-frame input setting on REDS4, though it has slightly more parameters. These findings indicate that our method balances well between complexity and performance. Additionally, VSRM is effective for slow-motion datasets, including Video-90K-T and Vid4, where our method also surpasses PSRT-recurrent [46] and IART [54] on these datasets for sixteen-frame input settings. This shows that our method works well in varying pixel movement and scenarios. Overall, these gains represent a substantial improvement in VSR. + +Visual Results. Qualitative comparisons can be observed in Fig. 5 and Fig. 6. VSRM effectively captures intricate details. Specifically, VSRM is the only method that accurately restores the license plate numbers in the REDS4 dataset and the fine details of Spider-Man in the Vid4 dataset in Fig. 5. More examples of REDS4 and Vid4 datasets are shown in the supp. file. Fig. 6 shows the effective receptive field of our method in comparison to EDVR [52] (CNN-based) and IART [54] (Transformer-based) on REDS4. A local receptive field restricts the CNN-based methods. In contrast, Transformer-based methods can enhance the receptive field but still face limitations due to the window attention mechanism for lower computational costs. Our method stands out as it successfully achieves a significantly larger global receptive field, enhancing our model's performance in VSR. + +![](images/5f1017c47218c32a92d81e23029a8215bbfb80f979e5ec56bd9e871d73f28caa.jpg) +Figure 6. Effective Receptive Field (ERF) [11, 38] visualization results. A more widely spread dark region signifies a larger receptive field. VSRM attains a considerable global receptive field. + +# 5.3. Ablation Study + +We conduct ablation studies using VSRM with custom settings for efficiency to understand the contributions of each proposed module. Details settings are in the supp. file. + +
ModelPSNR (dB)Param. (M)FLOPs (G)
3D DW-Conv30.8419.49149.8
Window Attn.30.977.68152.4
Full Attn31.067.681018.1
Mamba (ours)31.098.61159.2
+ +Effect of Mamba modules. Mamba modules are the core contribution of this study; they balance computational cost and high performance for inter-frame modeling. To verify its effectiveness, we replace these modules with convolutions from BasicVSR [3], window attention from IART [54], and global attention as in Tab. 3. Compared to Convolution and Window Attention methods, our method achieves notable performance gains, increasing PSNR by $0.25\mathrm{dB}$ and $0.12\mathrm{dB}$ , respectively, without adding significant computational overhead. Our method reaches a similar PSNR with Full Attention but is much more efficient regarding model complexity. Overall, Mamba modules balance computational efficiency and performance well. + +Table 3. Effective of Mamba modules. + +
StructurePSNR (dB)Param. (M)FLOPs (G)
w/o Align.30.878.53120.4
FGDA [32]30.928.70154.3
IA [54]31.008.57148.7
DCA (ours)31.098.61159.2
+ +Table 4. Effect of Deformable Cross-mamba Alignment module. + +Effect of DCA. Tab. 4 compares our DCA with existing alignment modules. Removing the alignment module significantly decreases performance (↓ 0.22dB). Our DCA achieves higher PSNR than SOTA methods like FGDA [32] and IA [54] without incurring excessive computational overhead. Specifically, DCA achieves 0.17dB and 0.09dB higher PSNR than FGDA and IA, respectively, demonstrating the effectiveness of our proposed module. By leveraging the flexibility of the deformable window when computing cross-mamba, DCA enhances the precision of aligning adjacent frames to the target frame. + +
Structurew/o T2SMBw/ T2SMB†w/ T2SMB
PSNR (dB)30.9531.0231.09
Param. (M)7.878.658.61
FLOPs (G)155.6162.2159.2
+ +Effect of T2SMB. Tab. 5 shows the effect of the Temporal-to-Spatial Mamba block (T2SMB). Removing T2SMB considerably reduces the model's performance $(\downarrow 0.14\mathrm{dB})$ , highlighting the block's importance. This performance drop occurs because the Spatial-to-Temporal Mamba block (S2TMB) cannot fully extract temporal information from the sequential input features. The T2SMB block prioritizes and explicitly focuses on extracting temporal information, enabling more comprehensive final features to be generated. We also conducted an ablation study with T2SMB using both forward and backward scans. Our findings indicate that using only the forward scan for T2SMB yields better performance while requiring less computational effort. + +Table 5. Effect of Temporal-to-Spatial Mamba block. † indicates that both forward and backward scans are applied to T2SMB. + +
StructurePSNR (dB)Param. (M)FLOPs (G)
FFN [51]30.908.68136.2
TGFN w/o DWConv30.818.27124.7
TGFN w/o Split31.059.41186.1
TGFN (ours)31.098.61159.2
+ +Effect of TGFN. Tab. 6 shows the effect of the temporal gated feed-forward network (TGFN). We compare models using regular FFN [51], TGFN without 3D depthwise convolution (TGFN w/o DWConv), TGFN without split channel (TGFN w/o Split), and our TGFN. Compared to FFN, TGFN has a similar number of parameters and somewhat higher FLOPs while substantially improving performance ( $\uparrow$ 0.19dB). Moreover, the performance suffers significantly ( $\downarrow$ 0.28dB) when the 3D DWConv is removed from TGFN, emphasizing the necessity of both spatial and temporal information. Eliminating the split operation in TGFN leads to a slight decrease in PSNR but increases model size and complexity. This indicates that redundancy in channel features negatively impacts the models' performance. + +Table 6. Effect of Temporal Gated Feed-forward Network. + +
LossFFL [23]\( FSL^3 \)[14]WHFL [27]FCL (ours)
00.020.2
PSNR31.0130.9930.9030.9731.0931.02
+ +Effect of FCL loss. Tab. 7 shows the effectiveness of Frequency Chabornier-like loss (FCL). We compare our loss with FFL [23], FSL [14], and WHFL [27] (default weight for these losses are 0.02 for fair comparison). Our + +FCL achieves a higher PSNR than FFL ( $\uparrow 0.08\mathrm{dB}$ ), indicating that balancing low- and high-frequency components and using Real and Imaginary parts for loss calculation yields more optimal results. Moreover, our loss demonstrates higher performance compared to FSL and WHFL, with clear margins. These results show the effectiveness of our proposed FCL compared to other frequency losses. Our method achieves the best results with FCL ( $\lambda_{i} = 0.02$ where $i \in \{Re, Im\}$ ). If FCL ( $\lambda_{i} = 0$ ) is eliminated, performance substantially declines ( $\downarrow 0.12\mathrm{dB}$ ), highlighting the importance of this loss. + +# 5.4. Model Complexity Comparison + +Tab. 8 reports comparing parameters, FLOPs, and runtime for VSRM and other VSR methods. The runtime is measured on a RTX A6000 48GB. Our VSRM is more efficient than EDVR when considering the number of parameters and FLOPs. Although it has a moderately higher complexity than Transformer-based methods that use window-based attention, these trade-offs are reasonable when balancing performance with model complexity. Furthermore, since the acceleration and optimization of Mamba are still being explored, our method has potential for further improvements. + +Table 7. Effect of Frequency Charbonnier-like loss (FCL). $i \in \{ Re, Im\}$ and we set ${\lambda }_{Re} = {\lambda }_{Im}$ . + +
MethodParam. (M)FLOPs (G)Runtime (ms)
VSR-T [2]32.6177.2-
EDVR [52]20.6323.6-
PSRT-recurrent [46]13.4164.6173
IART [54]13.4177.8180
VSRM (ours)17.1217.4223
+ +Table 8. Comparison of model complexity of different VSR models on the REDS4 dataset. + +# 6. Conclusion + +In this paper, we introduce VSRM, which adapts Mamba for VSR for the first time and achieves state-of-the-art results on multiple benchmarks. Specifically, we propose Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to capture long-range spatio-temporal features effectively. Next, we propose a novel Deformable Cross-Mamba Alignment that uses a deformable cross-mamba mechanism to adapt to varying pixel movement flex to enhance model performance compared to previous alignment modules. Finally, we introduce a Frequency Charbonnier-like loss to reduce the gaps between reconstructed and ground truth frames in the frequency domain and better recover high-frequency content. This demonstrates further improvement by utilizing both spatial and frequency losses. VSRM has the potential to establish a solid baseline for upcoming research and shows great promise in substituting Transformer-based models. Furthermore, VSRM could be adaptable for other low-level vision tasks, including video colorization, deblurring, and denoising. + +# Acknowledgments + +This work was supported by the Institute of Information & Communications Technology Planning & Evaluation(IITP)-Innovative Human Resource Development for Local Intellectualization program grant funded by the Korea government(MSIT)(IITP-2025-RS-2020-II201489). This work was also supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00573160), and the Technology Innovation Program (RS-2025-02222776) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea). + +This work was also supported by Hyundai Motor Chung Mong-Koo Global Scholarship to Dinh Phu Tran (first author) and generous GPU support by Korea Association for ICT Promotion. + +# References + +[1] Runyuan Cai, Yue Ding, and Hongtao Lu. Freqnet: A frequency-domain image super-resolution network with discrete cosine transform. arXiv preprint arXiv:2111.10800, 2021. 2 +[2] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2, 6, 7, 8 +[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2, 5, 7 +[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, pages 973-981, 2021. 1 +[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1, 5, 6, 7 +[6] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of 1st international conference on image processing, pages 168-172. IEEE, 1994. 5 +[7] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.1 +[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 4 + +[9] James W Cooley and John W Tukey. An algorithm for the machine calculation of complex fourier series. Mathematics of computation, 19(90):297-301, 1965. 5 +[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 1 +[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 7 +[12] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1 +[13] Daniel Y. Fu, Tri Dao, Khaled K. Saab, Armin W. Thomas, Atri Rudra, and Christopher Ré. Hungry Hungry Hippos: Towards language modeling with state space models. In International Conference on Learning Representations, 2023. 2 +[14] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 8 +[15] Dario Fuoli, Martin Danelljan, Radu Timofte, and Luc Van Gool. Fast online video super-resolution with deformable attention pyramid. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1735-1744, 2023. 1 +[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3 +[17] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2 +[18] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1 +[19] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2 +[20] Yukun Hao and Feihong Yu. Learning the frequency domain aliasing for real-world super-resolution. *Electronics*, 13(2): 250, 2024. 2 +[21] Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. Transformer quality in linear time. In International conference on machine learning, pages 9099-9117. PMLR, 2022. 1 +[22] Kui Jiang, Zhongyuan Wang, Peng Yi, and Junjun Jiang. Hierarchical dense recursive network for image superresolution. Pattern Recognition, 107:107475, 2020. 1 +[23] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. + +In Proceedings of the IEEE/CVF international conference on computer vision, pages 13919-13929, 2021. 2, 3, 8 +[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 2 +[25] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE transactions on computational imaging, 2(2):109-122, 2016. 2 +[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1 +[27] Min Woo Kim and Nam Ik Cho. Whfl: Wavelet-domain high frequency loss for sketch-to-image translation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 744-754, 2023. 2, 8 +[28] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 2 +[29] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2 +[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2 +[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 4 +[32] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6, 7 +[33] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Virt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 1, 2, 6, 7 +[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1 +[35] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE transactions on pattern analysis and machine intelligence, 36(2):346-360, 2013. 7 + +[36] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34: 21297-21309, 2021. 1 +[37] Jianping Luo, Shaofei Huang, and Yuan Yuan. Video superresolution using multi-scale pyramid 3d convolutional networks. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1882-1890, 2020. 2 +[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 7 +[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5 +[40] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 7 +[41] Ali Nasiri-Sarvi, Vincent Quoc-Huy Trinh, Hassan Rivaz, and Mahdi S Hosseini. Vim4path: Self-supervised vision mamba for histopathology images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2024. 1 +[42] Seong-Jin Park, Hyeongseok Son, Sunghyun Cho, Ki-Sang Hong, and Seungyong Lee. Srfeat: Single image superresolution with feature discrimination. In Proceedings of the European conference on computer vision (ECCV), pages 439-455, 2018. 2 +[43] Mohammad Saeed Rad, Behzad Bozortabar, Urs-Viktor Marti, Max Basler, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Srobb: Targeted perceptual loss for single image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2710–2719, 2019. 2 +[44] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 2 +[45] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network, 2016. 5 +[46] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 1, 2, 5, 6, 7, 8 +[47] Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In The Eleventh International Conference on Learning Representations, 2023. 2 +[48] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional + +neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 1 +[49] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2 +[50] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2 +[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 8 +[52] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6, 7, 8 +[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[54] Kai Xu, Ziwei Yu, Xin Wang, Michael Bi Mi, and Angela Yao. Enhancing video super-resolution via implicit resampling-based alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2546-2555, 2024. 1, 2, 6, 7, 8 +[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 7 +[56] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 1, 2 +[57] Yijun Yang, Zhaohu Xing, and Lei Zhu. Vivim: a video vision mamba for medical video object segmentation. arXiv preprint arXiv:2401.14168, 2024. 1 +[58] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 1, 2 +[59] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*, 2024. 2 \ No newline at end of file diff --git a/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/images.zip b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..06bda78410ca227f3128d6e7d1ea4f9a00243125 --- /dev/null +++ b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26794f0b3bb912a23fe7a12b08ffc6b41170d85e03ffae840461a4be28fa610b +size 641864 diff --git a/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/layout.json b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf78508c07f9e5d3ac0db4f5c4f8c13ab3333636 --- /dev/null +++ b/2025/VSRM_ A Robust Mamba-Based Framework for Video Super-Resolution/layout.json @@ -0,0 +1,10707 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 120 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 120 + ], + "type": "text", + "content": "VSRM: A Robust Mamba-Based Framework for Video Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 144, + 238, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 144, + 238, + 156 + ], + "spans": [ + { + "bbox": [ + 164, + 144, + 238, + 156 + ], + "type": "text", + "content": "Dinh Phu Tran" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 266, + 144, + 342, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 144, + 342, + 157 + ], + "spans": [ + { + "bbox": [ + 266, + 144, + 342, + 157 + ], + "type": "text", + "content": "Dao Duy Hung" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 371, + 144, + 446, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 144, + 446, + 157 + ], + "spans": [ + { + "bbox": [ + 371, + 144, + 446, + 157 + ], + "type": "text", + "content": "Daeyoung Kim" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 165, + 411, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 165, + 411, + 177 + ], + "spans": [ + { + "bbox": [ + 200, + 165, + 411, + 177 + ], + "type": "text", + "content": "{phutx2000, hicehehe, kimd}@kaist.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 184, + 182, + 425, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 182, + 425, + 196 + ], + "spans": [ + { + "bbox": [ + 184, + 182, + 425, + 196 + ], + "type": "text", + "content": "School of Computing, KAIST, Republic of Korea" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 229, + 200, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 229, + 200, + 242 + ], + "spans": [ + { + "bbox": [ + 151, + 229, + 200, + 242 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "content": "Video super-resolution remains a major challenge in low-level vision tasks. To date, CNN- and Transformer-based methods have delivered impressive results. However, CNNs are limited by local receptive fields, while Transformers struggle with quadratic complexity, posing challenges for processing long sequences in VSR. Recently, Mamba has drawn attention for its long-sequence modeling, linear complexity, and large receptive fields. In this work, we propose VSRM, a novel Video Super-Resolution framework that leverages the power of Mamba. VSRM introduces Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to extract long-range spatio-temporal features and enhance receptive fields efficiently. To better align adjacent frames, we propose Deformable Cross-Mamba Alignment module. This module utilizes a deformable cross-mamba mechanism to make the compensation stage more dynamic and flexible, preventing feature distortions. Finally, we minimize the frequency domain gaps between reconstructed and ground-truth frames by proposing a simple yet effective Frequency Charbonnier-like loss that better preserves high-frequency content and enhances visual quality. Through extensive experiments, VSRM achieves state-of-the-art results on diverse benchmarks, establishing itself as a solid foundation for future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Video super-resolution (VSR) aims to produce higher resolution from low-resolution videos. VSR typically processes long sequences and benefits from large receptive fields, efficiently capturing inter-frames information [3, 15]. However, it incurs high computational costs, requiring a balance between performance and efficiency. VSR usually uses convolutional neural network (CNN) [3, 22, 56] or transformer [33, 46, 54, 58]. Tab. 1 compares the pros and cons of these methods. CNN-based methods cannot adjust based on input data and limit receptive fields [10, 48] to capture inter-frame" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 230, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 469 + ], + "type": "text", + "content": "content. Transformer-based methods show strong performance due to the power of the attention mechanism. However, full attention has quadratic complexity [21, 36], which is impractical for long sequences, while local attention struggles with limited receptive fields [12]. Recently, Gu et al. introduced Mamba [16], which can model long-range information and data-dependent characteristics with linear complexity for NLP tasks. More recently, Mamba has been successfully applied in various vision tasks [7, 18, 41, 57]. As far as we know, no research has employed Mamba to address VSR. In this work, we propose VSRM, a novel Mamba-based framework for VSR. Specifically, we propose Dual Aggregation Mamba Block (DAMB), which contains " + }, + { + "bbox": [ + 313, + 230, + 555, + 469 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 230, + 555, + 469 + ], + "type": "text", + "content": " Spatial-to-Temporal Mamba blocks (S2TMBs) followed by a Temporal-to-Spatial Mamba block (T2SMB) to model long-range spatio-temporal correlations for long sequences, enlarging large receptive fields. In S2TMB and T2SMB, we employ spatial-to-temporal and temporal-to-spatial scanning mechanisms to capture both spatial and temporal contexts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 687 + ], + "type": "text", + "content": "The key challenge of VSR is effectively using the complementary information from adjacent frames, which may be misaligned due to object motions. To establish interframe correlations, many works employ the alignment module for VSR [3, 5, 46, 54]. However, most studies use bilinear or nearest-neighbor interpolation for spatial alignment due to its simplicity. Theoretically, using fixed weights (e.g., linear interpolation) causes feature distortions, leading to suboptimal results [4, 46]. To address this, Xu et al. [54] proposed an attention-based implicit interpolation for the alignment module. However, they computed interpolation in the fixed reference window, limiting flexibility and inaccuracy in aligning adjacent frames, especially in large motions between frames. Based on this observation, we introduce Deformable Cross-Mamba Alignment (DCA), which dynamically aligns neighboring frames using a deformable cross-mamba mechanism that allows information exchange between multiple selective scan modules." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Current VSR methods use pixel-based losses (e.g., L1/L2 loss [26, 34], charbonnier loss [5, 46]) or perception-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14711" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 137 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 137 + ], + "type": "table", + "html": "
ModelLinear ComplexityReceptive FieldData-drivenNotable Method
CNN⊗⊗⊗BasicVSR [3]
Local Attention⊗⊗⊗IART [54]
Full Attention⊗⊗⊗-
Mamba⊗⊗⊗VSRM (ours)
", + "image_path": "496e03861e5b6a9e16cfbe067a4bc109c5369e8a3aa434cc36c84ebc36b99013.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "content": "Table 1. Comparison of the backbone for VSR. VSRM reaches both linear complexity and a large receptive field." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 186, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 534 + ], + "type": "text", + "content": "based losses (e.g., perceptual loss [24, 43], adversarial loss [29, 42]). Pixel-wise losses lead to overly smooth results due to their restriction in capturing perceptual similarity. Perception-based losses enhance visual quality by recovering finer details but introduce greater distortion. As VSR is an ill-posed task, gaps between super-resolution and ground-truth frames are inevitable, sometimes visible only in the frequency domain [1, 20]. Therefore, combining loss in the spatial and frequency domains may enhance the model's performance. We propose Frequency Charbonnier-like Loss (FCL), calculated from the real and imaginary parts of frequency components in Fourier space. FCL helps our model focus more on preserving and generating high-frequency details while maintaining low-frequency content. Contribution. 1) We successfully adapt Mamba-based model for VSR for the first time and propose a new framework, VSRM. VSRM introduces Dual Aggregation Mamba Block (DAMB), which aggregates spatial and temporal features from sequences by using multi-scan direction mechanisms to obtain powerful representation ability and enlarge the model's receptive field. 2) We design Deformable Cross-mamba Alignment (DCA) module to provide a more flexible and learnable alignment through a deformable window cross-mamba mechanism in the compensation stage. 3) We propose Frequency Charbonnier-like loss (FCL), which is computed in the frequency domain to preserve and generate high-frequency details better. 4) VSRM outperforms all state-of-the-art (SOTA) methods on various benchmarks and shows the potential as the solid backbone for VSR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 548, + 142, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 142, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 142, + 561 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Video Super-Resolution. 1) CNN-based methods [3, 25, 37, 50, 56] is the early approach in VSR. Initially, VSR-net [25] extended CNN-based networks to VSR. Following this, TDAN [50] introduced deformable convolution alignment to enhance the alignment module's performance. In contrast, BasicVSR [3] proposed an effective model consisting solely of generic components, residual blocks, and optical flow, which surpasses all previous methods. 2) Transformer-based methods [2, 33, 46, 54, 58] achieve superior performance in VSR. Cao et al. [2] proposed spatial-temporal convolutional self-attention to explore better locality features. VRT [33] used a temporal mutual self" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "attention for a wider temporal receptive field and a parallel warp module to handle large motion. PSRT [46] introduced patch alignment that aligns image patches instead of pixels to maintain intra-patch relationships and preserve sub-pixel information. Based on this, Xu et al. [54] proposed an implicit alignment module based on cross-attention mechanism to adjust feature representations across frames, further improving performance. These methods applied window-based attention with small window to avoid the quadratic complexity of full attention but were limited by local receptive fields, affecting cross-frame information capture. Recent methods, like PSRT [46] and IART [54], focused on designing window-based alignment modules rather than improving the transformer backbone to enhance performance. They employed a fixed window for implicit motion compensation, limiting the flexibility of alignment modules, especially when minimal or excessive motion. To address these issues, this study proposes a new framework, VSRM, based on Mamba to leverage its advantage in large receptive fields and linear complexity. We also design Deformable Cross-mamba Alignment for more flexible motion compensation using a deformable mechanism to boost performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 338, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 555 + ], + "type": "text", + "content": "State Space Model. SSMs [13, 16, 17, 47] show potential to replace Transformers. Gu et al. introduced S4 [17] that achieves linear complexity modeling of long sequences, followed by S5[47] with improved capacity and efficiency. Mamba [16], built upon SSMs, has recently surpassed Transformers on many benchmarks in the NLP domain due to its data-dependent characteristics and global receptive fields. In computer vision, Vim [59] adapted Mamba for images by converting them into sequences and using a bidirectional mamba block for global modeling. VideoMamba [30] processed video by rearranging frames into spatio-temporal sequences, achieving comparable results to Transformer-based methods. MambaIR [19] adopted Mamba for image restoration tasks, demonstrating that Mamba can work well for low-level vision tasks, such as video deblurring, denoising, etc. Expanding on that progress, we are pioneers in investigating the capabilities of Mamba for VSR, utilizing its large receptive fields and linear complexity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "Spectral bias. Many studies reveal spectral bias [39, 44, 49] that models bias toward low-frequency rather than high-frequency content. Meanwhile, high-frequency content (e.g., edges, textures) is vital in low-level vision tasks [14, 23, 27, 28]. Hence, Jiang et al. proposed Focal Frequency Loss (FFL) [23] computed in the frequency domain to emphasize high-frequency content and enhance visual quality. Based on this, Fuoli et al. [14] employed a GAN loss in Fourier space. In contrast, WHFL [27] used Wavelet transform, incorporating a module for multi-bandwidth analysis. These methods employ a frequency-adjusted reweight matrix to optimize the low and high-frequency elements. However, the reweight matrix does not" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14712" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 69, + 511, + 285 + ], + "blocks": [ + { + "bbox": [ + 102, + 69, + 511, + 285 + ], + "lines": [ + { + "bbox": [ + 102, + 69, + 511, + 285 + ], + "spans": [ + { + "bbox": [ + 102, + 69, + 511, + 285 + ], + "type": "image", + "image_path": "99d88d026dc4c7dedd45aa5916d292f0beadd65f98f4d7d7127dcba9768ea7d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 290, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 54, + 290, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 290, + 555, + 324 + ], + "type": "text", + "content": "Figure 1. The network architecture of our method. a) Video Super-Resolution based on Mamba (VSRM). b) Feature Propagation block (FPB) consists of Deformable Cross-mamba Alignment (DCA) and Dual Aggregation Mamba Blocks (DAMB). c) Spatial-to-Temporal Mamba block (S2TMB). d) Temporal-to-Spatial Mamba block (T2SMB)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 344, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 429 + ], + "type": "text", + "content": "effectively balance the gradients of low and high frequencies. Furthermore, images mainly contain low-frequency content, so overemphasis can harm performance. Hence, we propose Frequency Charbonnier-like loss (FCL), which directly computes loss on the Real and Imaginary parts in Fourier space. Unlike FFL [23], FCL treats low- and high-frequency components equally." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 247, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 247, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 247, + 454 + ], + "type": "text", + "content": "3. Preliminaries: State Space Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 510 + ], + "type": "text", + "content": "SSMs draw their inspiration from continuous linear time-invariant systems, which map a 1D function or sequence " + }, + { + "bbox": [ + 55, + 461, + 296, + 510 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R} \\to y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 461, + 296, + 510 + ], + "type": "text", + "content": " via a hidden state " + }, + { + "bbox": [ + 55, + 461, + 296, + 510 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 461, + 296, + 510 + ], + "type": "text", + "content": ". This is formulated as ordinary differential equations (ODE):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 519, + 295, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 519, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 93, + 519, + 295, + 533 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = A h (t) + B x (t), y (t) = C h (t), \\tag {1}", + "image_path": "b5d0103d3e5648c6e259f5c399d5bdbbf587dc2a3cdc1f2986338c8c72e9b291.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": " is the evolution matrix, while " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{1 \\times N}" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": " are the projection parameters. To model discrete sequences like images, we need to approximate them via discretization. SSMs utilize the zero-order hold method, incorporating a timescale parameter " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": " to map the continuous parameters " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": " to discrete parameters " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\bar{A}" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\bar{B}" + }, + { + "bbox": [ + 55, + 542, + 296, + 627 + ], + "type": "text", + "content": ". The process of discretization is outlined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 636, + 295, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 295, + 651 + ], + "type": "interline_equation", + "content": "h _ {t} = \\bar {A} h _ {t - 1} + \\bar {B} x _ {t}, y _ {t} = C h _ {t}, \\tag {2a}", + "image_path": "9ed138872a675b2d2b31dd55216df3c77c011d82ae29039565711bda424768ff.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 652, + 295, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 652, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 63, + 652, + 295, + 667 + ], + "type": "interline_equation", + "content": "\\bar {A} = e x p (\\Delta A), \\bar {B} = (\\Delta A) ^ {- 1} (e x p (\\Delta A - I)) \\Delta B, (2 \\mathrm {b})", + "image_path": "c677777b397bca48a5472751ae3c050da5d76c4d148394c573a81f2ce002fd25.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "The parameters of these SSMs are independent of the data, which means that " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " remain constant regardless of the input, thereby restricting their adaptability" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 344, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 369 + ], + "type": "text", + "content": "in sequence modeling. Recently, Mamba [16] proposes selective SSMs (S6) with input data-dependent " + }, + { + "bbox": [ + 313, + 344, + 555, + 369 + ], + "type": "inline_equation", + "content": "x_{i} \\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 313, + 344, + 555, + 369 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 376, + 555, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 376, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 326, + 376, + 555, + 389 + ], + "type": "interline_equation", + "content": "B _ {i} = S _ {B} x _ {i}, C _ {i} = S _ {c} x _ {i}, \\Delta_ {i} = S o f t p l u s (S _ {\\Delta} x _ {i}), \\tag {3}", + "image_path": "2897a53f60ed70b712c2d72f2fb4d0f1aba43ca75a8a9b763647ba9431140bc1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "S_B, S_C \\in \\mathbb{R}^{N \\times L}" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "S_{\\Delta} \\in \\mathbb{R}^{L \\times L}" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " are linear projection layers. " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "B_i" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " are utilized across for all channels of " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\Delta x_i" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " channels, and " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " remains consistent with earlier SSMs. Via the discretization presented in equations 3, " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\bar{A}" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\bar{B}" + }, + { + "bbox": [ + 313, + 395, + 555, + 456 + ], + "type": "text", + "content": " become input data-dependent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 466, + 397, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 397, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 397, + 480 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 546 + ], + "type": "text", + "content": "In this section, we introduce VSRM architecture and its core components: Dual Aggregation Mamba block (DAMB) and Deformable Cross-mamba Alignment (DCA), as shown in Fig. 1. Furthermore, we present the training objectives to optimize our model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 382, + 564 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{t}^{LR} \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": " be the t-th low-resolution (LR) video frame, while " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{t}^{HR} \\in \\mathbb{R}^{T \\times sH \\times sW \\times C}" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": " denotes the t-th corresponding ground-truth (GT or HR) frame. Here, " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "T, H, W, C" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": " indicate the total number of frames, height, width, channels, and the upscaling factor, respectively. Our aim is to generate the super-resolution video frame " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{t}^{SR} \\in \\mathbb{R}^{T \\times sH \\times sW \\times C}" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": " using the consecutive " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "2N + 1" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{I_{i = t - N}^{t + N}\\}" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": ", ensuring that " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{t}^{SR}" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "inline_equation", + "content": "I_{t}^{HR}" + }, + { + "bbox": [ + 313, + 568, + 555, + 713 + ], + "type": "text", + "content": " are as closely as possible. VSRM is depicted in Fig. 1a) includes two parts: Feature Extraction (Conv2d and Feature Propagation Block) and Upsampler (Reconstruction). Feature extraction extracts deep features from input frames and aligns and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14713" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 282, + 171 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 171 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 171 + ], + "type": "image", + "image_path": "1f9f6aae8a4c0bece79d5b087eef5cd9f33d3bba43405624c7bb8d529809b446.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 174, + 296, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 218 + ], + "type": "text", + "content": "Figure 2. Structure and scan directions of Mamba module. a) Spatial-to-Temporal Mamba (S2T-Mamba). b) Temporal-to-Spatial Mamba (T2S-Mamba). For simplicity, we exclude the initial normalization and the final residual." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "content": "fuses the features of adjacent frames to leverage inter-frame relationships. The upsampler upscales the fused spatiotemporal features to produce the output sequence frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 230, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 230, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 230, + 295 + ], + "type": "text", + "content": "4.2. Dual Aggregation Mamba Block" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 443 + ], + "type": "text", + "content": "Dual Aggregation Mamba block (DAMB) is the core component of VSRM that consists of Spatial-to-Temporal Mamba block (S2TMB) and Temporal-to-Spatial Mamba block (T2SMB), as shown in Fig. 1(c, d). S2TMB and T2SMB are based on two modules: Spatial-to-Temporal Mamba (S2T-Mamba) and Temporal-to-Spatial Mamba (T2S-Mamba), with their structures shown in Fig. 2. By organizing " + }, + { + "bbox": [ + 55, + 300, + 296, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 300, + 296, + 443 + ], + "type": "text", + "content": " S2TMBs followed by a T2SMB in DAMB, VSRM fully extracts spatio-temporal information in both spatial and temporal dimensions. Moreover, we propose Temporal Gated Feed-forward Network (TGFN) to explore spatial and temporal information better." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "text", + "content": "Spatial-to-Temporal Mamba. S2T-Mamba structure and scan directions are shown in Fig. 2(a). A 3D sequence (" + }, + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "text", + "content": " frames) is converted into a 1D sequence and processed in two directions: forward (ForwardSSM(.)) and backward (BackwardSSM(.)) paths. This approach follows a mechanism that first scans spatially and then temporally, moving from frame 1 to frame " + }, + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "text", + "content": ". This enables both spatial and temporal awareness during processing. The original Mamba was designed for 1D sequences, so this bidirectional scan mechanism helps preserve spatial awareness. By employing the scanning process from spatial to temporal in both the forward and backward paths, our module is optimized for handling 3D sequences more effectively. Given an input " + }, + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "inline_equation", + "content": "X_{in}" + }, + { + "bbox": [ + 55, + 444, + 296, + 611 + ], + "type": "text", + "content": ", the output of S2T-Mamba is computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 619, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 619, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 66, + 619, + 295, + 632 + ], + "type": "interline_equation", + "content": "x = \\operatorname {L i n e a r} \\left(X _ {i n}\\right), z = \\sigma \\left(\\operatorname {L i n e a r} \\left(X _ {i n}\\right)\\right), \\tag {4a}", + "image_path": "7f3f18df0c83e3de99e47d22886d724142129653e76d6647533b45546c5ec750.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 635, + 294, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 635, + 294, + 649 + ], + "spans": [ + { + "bbox": [ + 66, + 635, + 294, + 649 + ], + "type": "interline_equation", + "content": "x _ {1} = \\text {F o r w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {4b}", + "image_path": "414b337afd7c7eca20cb8415dc6f52fc09555bcb164c780c6c3937857d6acc40.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 652, + 294, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 652, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 66, + 652, + 294, + 664 + ], + "type": "interline_equation", + "content": "x _ {2} = \\text {B a c k w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {4c}", + "image_path": "c52c67ea0c9b7c2042665a26c438c087f28ce4e123c4864f8ff0cfb1eb3381df.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 668, + 294, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 668, + 294, + 681 + ], + "spans": [ + { + "bbox": [ + 66, + 668, + 294, + 681 + ], + "type": "interline_equation", + "content": "S 2 T - M a m b a (x, z) = \\operatorname {L i n e a r} \\left(x _ {1} \\odot z + x _ {2} \\odot z\\right) \\tag {4d}", + "image_path": "454cb3b5a474737f590fb3ea7398a486f35c6d677974533e407d5db731b58eb5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Temporal-to-Spatial Mamba. T2S-Mamba structure and scan direction are shown in Fig. 2(b). T2S-Mamba also" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "content": "converts 3D sequence into a 1D sequence. However, we only perform the scanning in forward direction for two reasons. First, forward scan in T2S-Mamba models feature changes over time, while adding backward scan disrupts learned dependencies, leading to a slight decrease in performance, and an increase in FLOPs. Second, our experiment shows that using only forward scan in T2S-Mamba yields better results than adding backward pass, shown in Tab. 5. We found that S2T-Mamba prioritizes extracting spatial rather than temporal information, leading to not extracting fully temporal information. In contrast, T2S-Mamba explicitly prioritizes scanning temporal rather than spatial information, allowing for a more comprehensive feature extraction. Given an input " + }, + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "inline_equation", + "content": "X_{in}" + }, + { + "bbox": [ + 313, + 72, + 555, + 251 + ], + "type": "text", + "content": ", the output of S2T-Mamba is calculated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 261, + 553, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 261, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 347, + 261, + 553, + 274 + ], + "type": "interline_equation", + "content": "x = \\operatorname {L i n e a r} \\left(X _ {i n}\\right), z = \\sigma \\left(\\operatorname {L i n e a r} \\left(X _ {i n}\\right)\\right), \\tag {5a}", + "image_path": "a968cd9bb40d564aca23a90006780ffd0b1f37371784881bd67c3982f4be7e73.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 277, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 277, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 347, + 277, + 553, + 289 + ], + "type": "interline_equation", + "content": "x _ {1} = \\text {F o r w a r d S S M} (\\operatorname {C o n v 1} d (x)), \\tag {5b}", + "image_path": "9e58b4a2f5addd9418482093fdfd75c2f3e12480d4855e149b0dd7e70de739e3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 293, + 553, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 293, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 347, + 293, + 553, + 306 + ], + "type": "interline_equation", + "content": "T 2 S - M a m b a (x, z) = \\operatorname {L i n e a r} \\left(x _ {1} \\odot z\\right) \\tag {5c}", + "image_path": "d051e267112c82aa5ce3f058b438a68fff8ac199e2c4c789bc9b863e22b82b27.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 315, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 531 + ], + "type": "text", + "content": "Temporal Gated Feed-forward Network. Feed-forward network (FFN) [51] consists of a nonlinear activation positioned between two linear projection layers. It disregards spatial and temporal relationships and lacks feature alignment since it processes each input token individually. Moreover, the redundant information across channels hinders progress in learning feature representations. To address this issues, we propose Temporal Gated Feed-forward Network (TGFN) shown in Fig. 3. It has two main updates to FFN: 1) 3D depthwise convolutions (DW-3D Conv) and 2) a gating mechanism. We use DW-3D Conv to model relationships effectively from spatially and temporally neighboring pixel positions, which is vital in VSR. TGFN employs a gating mechanism by performing element-wise multiplication on two parallel feature branches obtained from the chunking method, with one of them passed by the GELU, which has become the preferred option in recent work [8, 31]. Given the input " + }, + { + "bbox": [ + 313, + 315, + 555, + 531 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 315, + 555, + 531 + ], + "type": "text", + "content": ", TGFN is formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 538, + 553, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 538, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 323, + 538, + 553, + 555 + ], + "type": "interline_equation", + "content": "\\hat {X} = W _ {p} ^ {1} X, \\quad [ \\hat {X} _ {1}, \\hat {X} _ {2} ] = \\hat {X} \\tag {6a}", + "image_path": "6d2a32f37d95f950266326c2cb5bc0a5f8df9929020a322aa930c8aa47f46f23.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 322, + 557, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 557, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 322, + 557, + 553, + 572 + ], + "type": "interline_equation", + "content": "T G F N (X) = W _ {p} ^ {2} \\left(W _ {d} ^ {1} L N \\left(\\hat {X} _ {1}\\right) \\odot \\sigma \\left(W _ {d} ^ {2} L N \\left(\\hat {X} _ {2}\\right)\\right)\\right) \\tag {6b}", + "image_path": "0167e7efd9acfea611877ababce420dfd0d498f2adae24cbc18ca0d22bc5348f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "inline_equation", + "content": "W_{p}^{1}, W_{p}^{2}" + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "text", + "content": " are weights of linear projections; " + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "inline_equation", + "content": "W_{d}^{1}, W_{d}^{2}" + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "text", + "content": " are weights of the 3D DWConv layers; " + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "text", + "content": " is GELU activation, " + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 579, + 555, + 653 + ], + "type": "text", + "content": " is the hidden dimension, LN is LayerNorm. Overall, TGFN adds nonlinear spatial and temporal information more effectively and optimizes the information flow through our pipeline." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 660, + 512, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 512, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 512, + 673 + ], + "type": "text", + "content": "4.3. Deformable Cross-mamba Alignment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Alignment modules in VSR often fail when pixel movement between consecutive frames is minimal or excessive. This requires an alignment module adaptable to varying pixel" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14714" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 69, + 249, + 135 + ], + "blocks": [ + { + "bbox": [ + 98, + 69, + 249, + 135 + ], + "lines": [ + { + "bbox": [ + 98, + 69, + 249, + 135 + ], + "spans": [ + { + "bbox": [ + 98, + 69, + 249, + 135 + ], + "type": "image", + "image_path": "c1aab8293864997f4a89d1f695b60631c93dee61d8739bdf44be0bca0a155469.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 138, + 282, + 150 + ], + "lines": [ + { + "bbox": [ + 69, + 138, + 282, + 150 + ], + "spans": [ + { + "bbox": [ + 69, + 138, + 282, + 150 + ], + "type": "text", + "content": "Figure 3. Temporal Gated Feed-forward Network (TGFN)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 302 + ], + "type": "text", + "content": "movement. Building upon [3, 5, 46], we design Deformable Cross-mamba Alignment (DCA), as shown in Fig. 4. DCA uses the pre-trained SpyNet [45] for flow estimation while incorporating a new motion compensation scheme. Specifically, we use a deformable window scheme to perform a more flexible warping operation in the compensation stage between frames according to inter-frame motion information. We then employ a cross-mamba module, which enables interactions between the target point and the values of the dynamic reference points to infer the aligned point, leading to more adaptive and accurate alignment results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": "For each pixel " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "X(x,y)" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " in the target frame " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " (red point), let " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "(\\Delta_x,\\Delta_y)" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " be the optical flow estimated between " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " and the reference frame " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": ". The aligned frame " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\bar{X}" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " is obtained by a spatial warping function " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " with its parameters " + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{W}}" + }, + { + "bbox": [ + 55, + 303, + 296, + 351 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 358, + 295, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 358, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 85, + 358, + 295, + 373 + ], + "type": "interline_equation", + "content": "\\bar {X} (x, y) = \\mathcal {W} \\left(X ^ {\\prime}, (x + \\Delta x, y + \\Delta y); \\theta_ {\\mathcal {W}}\\right), \\tag {7}", + "image_path": "fbccc6370a98cc5fca14f66233b36afd649049292b0f8d72279a287aafc19280.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": "In the warping operation, we learn offset " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\epsilon_r" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": " from window " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": " to identify dynamic reference region " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "\\bar{r}" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": ". We extract " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{R}^{H_w \\times W_w \\times C}" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": ", which is centered at the sampling position " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "(x', y') = (x + \\Delta x, y + \\Delta y)" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "H_w" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "W_w" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": " are the height, width, and number of channels of the window, respectively, as shown in Fig. 4. Within this window, we initialize a reference region " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^{H_r \\times W_r \\times 2}" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": ", by iterating over " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "i \\in (0, 1, \\dots, H_r)" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "inline_equation", + "content": "j \\in (0, 1, \\dots, W_r)" + }, + { + "bbox": [ + 55, + 381, + 296, + 477 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 485, + 295, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 485, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 61, + 485, + 295, + 500 + ], + "type": "interline_equation", + "content": "r (i, j) = X ^ {\\prime} \\left(\\lfloor x ^ {\\prime} \\rfloor - \\lfloor \\frac {W _ {r} - 1}{2} \\rfloor + i, \\lfloor y ^ {\\prime} \\rfloor - \\lfloor \\frac {H _ {r} - 1}{2} \\rfloor + j\\right) \\tag {8}", + "image_path": "6e8e72e0861a3a40912403e6d6d1eacc9ebec117572b734b7f56d0537ab304a7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": "To obtain the dynamic reference region " + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "inline_equation", + "content": "\\bar{r}" + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": ", we refine " + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": " using a small learnable offset network " + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(.)" + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": ", which contains only convolution and activation layers (Fig. 4). This network learns the reference offset " + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "inline_equation", + "content": "\\epsilon_r" + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": " from the features within the window " + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": ". The final adaptive reference map is computed using bilinear interpolation " + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "inline_equation", + "content": "\\phi(.;.)" + }, + { + "bbox": [ + 55, + 508, + 296, + 580 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 589, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 589, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 97, + 589, + 295, + 602 + ], + "type": "interline_equation", + "content": "\\bar {r} = \\phi (w; r + \\epsilon_ {r}), \\quad \\text {w i t h} \\epsilon_ {r} = S (w), \\tag {9}", + "image_path": "de094b7c5d7decbf51630aff7faef6582a3aadfafd04c21143f8ad5fdab39771.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "content": "Finally, we apply implicit warping within the dynamic reference region to align each pixel by selectively cross-fusing reference point values with the target grid value at " + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "inline_equation", + "content": "X(x,y)" + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "content": " using the cross-mamba module cross-mamba(.,.):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 667, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 667, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 107, + 667, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\bar {X} (x, y) = \\text {c r o s s - m a m b a} (R, Q), \\tag {10}", + "image_path": "54e5c35daa47d279987927cfe0af4e02665f02a7ec0bf09676c86053804a8c9d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is the reference tensor and " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " is the target tensor, both augmented with positional information. The reference" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 545, + 194 + ], + "type": "image", + "image_path": "6621aaa41fc3a79f2bd20b77ea92c80c835c73b701c40ea0c9f2b06e22ebcf35.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "lines": [ + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "text", + "content": "Figure 4. Pipeline of our proposed Deformable Cross-mamba Alignment (DCA). We initialize a fixed reference region " + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "text", + "content": " within the window " + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "text", + "content": ". This region is then adjusted to a dynamic reference region " + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\bar{r}" + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "text", + "content": " using an offset " + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_r" + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "text", + "content": ", which is learned from the features within " + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 197, + 555, + 263 + ], + "type": "text", + "content": ". Finally, DCA learns the affinity through the cross-mamba module to calculate the final result (aligned point)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": "tensor " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": " is computed by adding the flattened vector " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\bar{r}" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": " to the positional encoding " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "PE_{R}" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": ", and the target tensor " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": " is obtained by adding the target value " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "X(x,y)" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": " to the positional embedding " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "PE_{Q}" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": ". We use sinusoidal positional encoding to capture high-frequency content, which maps continuous input coordinates into a high-dimensional space [39] (more details in the supp. file). Based on Eq. (2a), the cross-mamba module computes output " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\bar{X}" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": " at each step " + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 285, + 555, + 381 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 390, + 555, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 390, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 353, + 390, + 555, + 405 + ], + "type": "interline_equation", + "content": "H _ {t} = \\bar {A} _ {R} H _ {t - 1} + \\bar {B} _ {R} \\bar {R} _ {t}, \\bar {X} _ {t} = C _ {Q} H _ {t}, \\tag {11}", + "image_path": "df046c0a0aa6d8b061ae7439e0765ca59a6a55866d96337e72b38377fa3fb08d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\bar{A}_R" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\bar{B}_R" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": " are matrices dependent on " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "C_Q" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": " is a cross-modal matrix dependent on " + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 414, + 555, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 447, + 430, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 430, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 430, + 460 + ], + "type": "text", + "content": "4.4. Training Objectives" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 465, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 514 + ], + "type": "text", + "content": "We use two loss functions, Charbonnier loss (CL) [6] and Frequency Charbonnier-like loss (FCL). CL minimizes the spatial domain's pixel gaps, while FCL reduces the frequency domain's gaps to recover high-frequency content." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 514, + 554, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 549 + ], + "type": "text", + "content": "CL is considered a local loss, measuring the absolute difference between corresponding pixels of two frames without considering global structures. CL is as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 560, + 553, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 560, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 369, + 560, + 553, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C L} = \\sqrt {\\| \\mathbf {I} _ {S R} - \\mathbf {I} _ {H R} \\| ^ {2} + \\epsilon^ {2}}, \\tag {12}", + "image_path": "947f26a9b80b292109139d46fdc55594a82705bd04b6eafea26c5823947c3e3b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "text", + "content": "In contrast, FCL utilizes Fast Fourier Transformation (FFT) [9] to convert frames from the spatial to the frequency domain. FFT utilizes the information from all pixels to obtain the FFT value for each spatial frequency. Thus, FCL is considered a global loss. Given an image " + }, + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "text", + "content": ", Discrete Fourier Transform " + }, + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 585, + 555, + 657 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 666, + 556, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 666, + 556, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 666, + 556, + 693 + ], + "type": "interline_equation", + "content": "\\mathcal {X} (u, v) = \\mathcal {F} [ \\mathbf {I} ] = \\sum_ {x = 1} ^ {H} \\sum_ {y = 1} ^ {W} \\mathbf {I} (x, y) e ^ {- j 2 \\pi \\left(\\frac {u}{H} x + \\frac {v}{W} y\\right)} \\tag {13}", + "image_path": "ce2c07cd060956306ff799f33f1c1f5bc8fa99752bba91e2726cc62b21f69e6a.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 535, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 535, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 535, + 713 + ], + "type": "text", + "content": "For simplicity, we only present the formulation for gray images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14715" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 509, + 256 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 509, + 256 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 509, + 256 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 509, + 256 + ], + "type": "table", + "html": "
MethodFrames#Params (M)REDS4Vimeo-90K-TVid4
PSNRSSIMPSNRSSIMPSNRSSIM
EDVR [52]5/720.631.090.880037.610.948927.350.8264
VSR-T [2]5/732.631.190.881537.710.949427.360.8258
VRT [33]6/-30.731.600.8888----
PSRT-recurrent [46]6/-10.831.880.8964----
IART [54]6/-13.432.150.9010----
VSRM (ours)6/-17.132.430.9059----
BasicVSR++ [5]30/147.332.390.906937.790.950027.790.8400
VRT [33]16/735.632.190.900638.200.953027.930.8425
RVRT [32]30/1410.832.750.911338.150.952727.990.8462
PSRT-recurrent [46]16/1413.432.720.910638.270.953628.070.8485
IART [54]16/713.432.900.913838.140.952828.260.8517
VSRM (ours)16/717.133.110.916238.330.954328.440.8552
", + "image_path": "ae7bf1209ec806a5c4fa0172f36247f3d188da7d8bba0b060e494d42842a8f44.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 281 + ], + "type": "text", + "content": "Table 2. Quantitative comparison (PSNR↑ and SSIM↑) of our VSRM with other state-of-the-art methods at a scale of 4x. The best and second-best results are highlighted in **bold** and **underlined**, and “-” indicates no test results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 97, + 291, + 278, + 400 + ], + "blocks": [ + { + "bbox": [ + 97, + 291, + 278, + 400 + ], + "lines": [ + { + "bbox": [ + 97, + 291, + 278, + 400 + ], + "spans": [ + { + "bbox": [ + 97, + 291, + 278, + 400 + ], + "type": "image", + "image_path": "a471dfcfe95e3102147d6c2f48d57eca86c704ee3349cc3926e9fb6bfdeb6514.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 283, + 292, + 356, + 342 + ], + "blocks": [ + { + "bbox": [ + 283, + 292, + 356, + 342 + ], + "lines": [ + { + "bbox": [ + 283, + 292, + 356, + 342 + ], + "spans": [ + { + "bbox": [ + 283, + 292, + 356, + 342 + ], + "type": "image", + "image_path": "8bc21ceb145dcbf13ddfd0ef987090625c9bffd2404adfa8fc09c3ad05ab67c2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 342, + 327, + 348 + ], + "lines": [ + { + "bbox": [ + 304, + 342, + 327, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 327, + 348 + ], + "type": "text", + "content": "Nearest" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 292, + 435, + 342 + ], + "blocks": [ + { + "bbox": [ + 362, + 292, + 435, + 342 + ], + "lines": [ + { + "bbox": [ + 362, + 292, + 435, + 342 + ], + "spans": [ + { + "bbox": [ + 362, + 292, + 435, + 342 + ], + "type": "image", + "image_path": "a422594578d7392249a0ea86db8f3b5f2ec9f58217cd0b52aa3ed28fd0751cef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 342, + 406, + 348 + ], + "lines": [ + { + "bbox": [ + 392, + 342, + 406, + 348 + ], + "spans": [ + { + "bbox": [ + 392, + 342, + 406, + 348 + ], + "type": "text", + "content": "VRT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 442, + 292, + 515, + 342 + ], + "blocks": [ + { + "bbox": [ + 442, + 292, + 515, + 342 + ], + "lines": [ + { + "bbox": [ + 442, + 292, + 515, + 342 + ], + "spans": [ + { + "bbox": [ + 442, + 292, + 515, + 342 + ], + "type": "image", + "image_path": "d42a1ba3bbf4ad8f9d02aa37c893ac1d4aa67525129f329004ce3bc13c1dae8f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 342, + 501, + 348 + ], + "lines": [ + { + "bbox": [ + 457, + 342, + 501, + 348 + ], + "spans": [ + { + "bbox": [ + 457, + 342, + 501, + 348 + ], + "type": "text", + "content": "PSRT-recurrent" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 97, + 407, + 277, + 515 + ], + "blocks": [ + { + "bbox": [ + 155, + 401, + 236, + 407 + ], + "lines": [ + { + "bbox": [ + 155, + 401, + 236, + 407 + ], + "spans": [ + { + "bbox": [ + 155, + 401, + 236, + 407 + ], + "type": "text", + "content": "Frame 026,Clip 000,REDS4" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 97, + 407, + 277, + 515 + ], + "lines": [ + { + "bbox": [ + 97, + 407, + 277, + 515 + ], + "spans": [ + { + "bbox": [ + 97, + 407, + 277, + 515 + ], + "type": "image", + "image_path": "cd2c073459c0cce019827e5f75d0633c6125e853a0333dc2c0cccdf614c3181e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 516, + 225, + 523 + ], + "lines": [ + { + "bbox": [ + 149, + 516, + 225, + 523 + ], + "spans": [ + { + "bbox": [ + 149, + 516, + 225, + 523 + ], + "type": "text", + "content": "Frame 037, Clip Walk, Vid4" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 529, + 555, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 555, + 540 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons on REDS4 and Vid4 datasets, VSRM achieves clearer and more precise results, revealing finer patterns." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 283, + 350, + 356, + 399 + ], + "blocks": [ + { + "bbox": [ + 283, + 350, + 356, + 399 + ], + "lines": [ + { + "bbox": [ + 283, + 350, + 356, + 399 + ], + "spans": [ + { + "bbox": [ + 283, + 350, + 356, + 399 + ], + "type": "image", + "image_path": "cd5e9c0e0530663a2497a81f83cea79bb5c4f32efc0fa970bb563038807341d9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 362, + 350, + 435, + 399 + ], + "blocks": [ + { + "bbox": [ + 362, + 350, + 435, + 399 + ], + "lines": [ + { + "bbox": [ + 362, + 350, + 435, + 399 + ], + "spans": [ + { + "bbox": [ + 362, + 350, + 435, + 399 + ], + "type": "image", + "image_path": "c104266b210062afe17fff6e03e1aea1ed51e3bf072710b48a2a0742bc89bd17.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 442, + 350, + 515, + 399 + ], + "blocks": [ + { + "bbox": [ + 442, + 350, + 515, + 399 + ], + "lines": [ + { + "bbox": [ + 442, + 350, + 515, + 399 + ], + "spans": [ + { + "bbox": [ + 442, + 350, + 515, + 399 + ], + "type": "image", + "image_path": "2b524355a9517c9af18ab4c86f3ddbb4453c9bd4bb6ae229ee3c3b5a21571e43.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 283, + 407, + 356, + 457 + ], + "blocks": [ + { + "bbox": [ + 312, + 401, + 327, + 407 + ], + "lines": [ + { + "bbox": [ + 312, + 401, + 327, + 407 + ], + "spans": [ + { + "bbox": [ + 312, + 401, + 327, + 407 + ], + "type": "text", + "content": "IART" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 283, + 407, + 356, + 457 + ], + "lines": [ + { + "bbox": [ + 283, + 407, + 356, + 457 + ], + "spans": [ + { + "bbox": [ + 283, + 407, + 356, + 457 + ], + "type": "image", + "image_path": "7888d498814b2355cc547e2f8b430e3f386e90a19fde289f13df0df7a24cf497.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 458, + 331, + 463 + ], + "lines": [ + { + "bbox": [ + 308, + 458, + 331, + 463 + ], + "spans": [ + { + "bbox": [ + 308, + 458, + 331, + 463 + ], + "type": "text", + "content": "Nearest" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 362, + 407, + 435, + 457 + ], + "blocks": [ + { + "bbox": [ + 380, + 401, + 417, + 407 + ], + "lines": [ + { + "bbox": [ + 380, + 401, + 417, + 407 + ], + "spans": [ + { + "bbox": [ + 380, + 401, + 417, + 407 + ], + "type": "text", + "content": "VSRM (ours)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 407, + 435, + 457 + ], + "lines": [ + { + "bbox": [ + 362, + 407, + 435, + 457 + ], + "spans": [ + { + "bbox": [ + 362, + 407, + 435, + 457 + ], + "type": "image", + "image_path": "e2c4b508d3132bcaf5449cd7d9aa55da563186e41691bf0531bc9b4659d290a6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 442, + 407, + 515, + 457 + ], + "blocks": [ + { + "bbox": [ + 474, + 401, + 483, + 407 + ], + "lines": [ + { + "bbox": [ + 474, + 401, + 483, + 407 + ], + "spans": [ + { + "bbox": [ + 474, + 401, + 483, + 407 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 407, + 515, + 457 + ], + "lines": [ + { + "bbox": [ + 442, + 407, + 515, + 457 + ], + "spans": [ + { + "bbox": [ + 442, + 407, + 515, + 457 + ], + "type": "image", + "image_path": "74191797c2065d203b4afc077c67a368c686c498fbdb9951f278932ab3a6f191.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 283, + 464, + 356, + 515 + ], + "blocks": [ + { + "bbox": [ + 283, + 464, + 356, + 515 + ], + "lines": [ + { + "bbox": [ + 283, + 464, + 356, + 515 + ], + "spans": [ + { + "bbox": [ + 283, + 464, + 356, + 515 + ], + "type": "image", + "image_path": "d5f28409de5855caa26c365c307903fc44a5a72e7d73cbf14726f1aff35b6eed.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 516, + 327, + 521 + ], + "lines": [ + { + "bbox": [ + 312, + 516, + 327, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 516, + 327, + 521 + ], + "type": "text", + "content": "IART" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 362, + 464, + 435, + 515 + ], + "blocks": [ + { + "bbox": [ + 392, + 458, + 405, + 464 + ], + "lines": [ + { + "bbox": [ + 392, + 458, + 405, + 464 + ], + "spans": [ + { + "bbox": [ + 392, + 458, + 405, + 464 + ], + "type": "text", + "content": "VRT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 464, + 435, + 515 + ], + "lines": [ + { + "bbox": [ + 362, + 464, + 435, + 515 + ], + "spans": [ + { + "bbox": [ + 362, + 464, + 435, + 515 + ], + "type": "image", + "image_path": "d1ac174075d28783ef63e9d3372e05ada0d324cb2e9e7ab38387dec42ad43402.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 516, + 417, + 522 + ], + "lines": [ + { + "bbox": [ + 380, + 516, + 417, + 522 + ], + "spans": [ + { + "bbox": [ + 380, + 516, + 417, + 522 + ], + "type": "text", + "content": "VSRM (ours)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 442, + 464, + 515, + 515 + ], + "blocks": [ + { + "bbox": [ + 457, + 458, + 501, + 464 + ], + "lines": [ + { + "bbox": [ + 457, + 458, + 501, + 464 + ], + "spans": [ + { + "bbox": [ + 457, + 458, + 501, + 464 + ], + "type": "text", + "content": "PSRT-recurrent" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 464, + 515, + 515 + ], + "lines": [ + { + "bbox": [ + 442, + 464, + 515, + 515 + ], + "spans": [ + { + "bbox": [ + 442, + 464, + 515, + 515 + ], + "type": "image", + "image_path": "093407cb11ac512e408b7d582e98bf9d19c5e704662b527c59b7d43bbe28f992.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 516, + 484, + 521 + ], + "lines": [ + { + "bbox": [ + 474, + 516, + 484, + 521 + ], + "spans": [ + { + "bbox": [ + 474, + 516, + 484, + 521 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(x,y)" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": " is the pixel value at " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{X}(u,v)" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": " is complex value at frequency " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": " are Euler's number and the imaginary unit. " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(.)" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": " is a complex number with Real " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "(Re\\mathcal{F}(.))" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": " and Imaginary " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "(Im\\mathcal{F}(.))" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": ". We observe a limitation in calculating loss based on magnitude and phase. The magnitude calculation involves a square root, while the phase employs a nonlinear arctangent function. This creates discontinuities and increases sensitivity to numerical inaccuracies for VSR. To address this, we calculate frequency loss separately for the Real and Imaginary. This ensures that even minor changes in the spatial domain lead to proportional changes in these components, as they represent linear" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 560, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 555, + 608 + ], + "type": "text", + "content": "decompositions of the signal in Fourier space. FCL implicitly captures phase information via R/I, letting SR/GT phase differences impact the loss. The formulation of FCL and total loss are as follows:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 322, + 614, + 555, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 614, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 322, + 614, + 555, + 641 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F C L} = \\sum_ {i \\in \\{R e, I m \\}} \\lambda_ {i} \\sqrt {\\left\\| i \\mathcal {F} \\left(\\mathbf {I} _ {S R}\\right) - i \\mathcal {F} \\left(\\mathbf {I} _ {H R}\\right) \\right\\| ^ {2} + \\epsilon^ {2}} \\tag {14}", + "image_path": "1599dbce61bf06f265813fc92aa87628569a3e7ed143e0891c621f3db11176d5.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 381, + 648, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 648, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 381, + 648, + 553, + 661 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda \\mathcal {L} _ {C L} + \\mathcal {L} _ {F C L}, \\tag {15}", + "image_path": "1b6c3e49ea8b3f992bca1d3c49c7e3d7c9ed4a8b50f48d90a166629e8e34149d.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda, \\lambda_{Re}" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{Im}" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": " are hyperparameters; " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{Re}, \\lambda_{Im} \\ll \\lambda" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{Re}" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{Im}" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": " control the regularization strength of the high-frequency content in the output. In this work, we set " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 1.0" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{Re} = \\lambda_{Im} = 0.02" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon = 10^{-3}" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14716" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 177, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 177, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 177, + 102 + ], + "type": "text", + "content": "5.1. Datasets and Metrics" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 107, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 296, + 228 + ], + "type": "text", + "content": "We use REDS [40] and Vimeo-90K [55] as training datasets. Following [52], for REDS, we use " + }, + { + "bbox": [ + 55, + 107, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{REDS4}^2" + }, + { + "bbox": [ + 55, + 107, + 296, + 228 + ], + "type": "text", + "content": " as a test set, and the remaining videos are used for training. For Vimeo-90K, Vid4 [35] and Vimeo-90K-T [55] are used as test sets. These two training sets have different motion movements. The motion in Vimeo-90K is generally small, whereas REDS contains large motions with more pixel movement. We study " + }, + { + "bbox": [ + 55, + 107, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 107, + 296, + 228 + ], + "type": "text", + "content": " VSR task, use bicubic interpolation to produce LR video frames, and use PSNR/SSIM [53] as the evaluation metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 235, + 285, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 285, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 285, + 248 + ], + "type": "text", + "content": "5.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 252, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 289 + ], + "type": "text", + "content": "We compare our VSRM with the SOTA VSR methods: EDVR [52], VSR-T [2], VRT [33], BasicVSR++ [5], RVRT [32], PSRT-recurrent [46], IART [54]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 515 + ], + "type": "text", + "content": "Quantitative Results. As shown in Tab. 2, VSRM achieves SOTA results across all benchmarks. Notably, VSRM surpasses VRT [33], a window-based transformer approach, by margins of up to 0.83dB and 0.90dB in PSNR for six and sixteen input frame settings on REDS4 while featuring a lower parameter count. Compared to the recent SOTA methods PSRT-recurrent[46] and IART [54], VSRM demonstrates improvements of 0.4dB and 0.28dB in PSNR for six-frame input setting and 0.39dB and 0.21dB in PSNR for sixteen-frame input setting on REDS4, though it has slightly more parameters. These findings indicate that our method balances well between complexity and performance. Additionally, VSRM is effective for slow-motion datasets, including Video-90K-T and Vid4, where our method also surpasses PSRT-recurrent [46] and IART [54] on these datasets for sixteen-frame input settings. This shows that our method works well in varying pixel movement and scenarios. Overall, these gains represent a substantial improvement in VSR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 696 + ], + "type": "text", + "content": "Visual Results. Qualitative comparisons can be observed in Fig. 5 and Fig. 6. VSRM effectively captures intricate details. Specifically, VSRM is the only method that accurately restores the license plate numbers in the REDS4 dataset and the fine details of Spider-Man in the Vid4 dataset in Fig. 5. More examples of REDS4 and Vid4 datasets are shown in the supp. file. Fig. 6 shows the effective receptive field of our method in comparison to EDVR [52] (CNN-based) and IART [54] (Transformer-based) on REDS4. A local receptive field restricts the CNN-based methods. In contrast, Transformer-based methods can enhance the receptive field but still face limitations due to the window attention mechanism for lower computational costs. Our method stands out as it successfully achieves a significantly larger global receptive field, enhancing our model's performance in VSR." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 525, + 150 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 525, + 150 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 525, + 150 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 525, + 150 + ], + "type": "image", + "image_path": "5f1017c47218c32a92d81e23029a8215bbfb80f979e5ec56bd9e871d73f28caa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 155, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 555, + 189 + ], + "type": "text", + "content": "Figure 6. Effective Receptive Field (ERF) [11, 38] visualization results. A more widely spread dark region signifies a larger receptive field. VSRM attains a considerable global receptive field." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 408, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 408, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 408, + 205 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 209, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 247 + ], + "type": "text", + "content": "We conduct ablation studies using VSRM with custom settings for efficiency to understand the contributions of each proposed module. Details settings are in the supp. file." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 326, + 255, + 542, + 316 + ], + "blocks": [ + { + "bbox": [ + 326, + 255, + 542, + 316 + ], + "lines": [ + { + "bbox": [ + 326, + 255, + 542, + 316 + ], + "spans": [ + { + "bbox": [ + 326, + 255, + 542, + 316 + ], + "type": "table", + "html": "
ModelPSNR (dB)Param. (M)FLOPs (G)
3D DW-Conv30.8419.49149.8
Window Attn.30.977.68152.4
Full Attn31.067.681018.1
Mamba (ours)31.098.61159.2
", + "image_path": "20f5fe52964f9cfb4417b6d97c032fa42dd9be069c544c717a6418d2ee578873.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 336, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 492 + ], + "type": "text", + "content": "Effect of Mamba modules. Mamba modules are the core contribution of this study; they balance computational cost and high performance for inter-frame modeling. To verify its effectiveness, we replace these modules with convolutions from BasicVSR [3], window attention from IART [54], and global attention as in Tab. 3. Compared to Convolution and Window Attention methods, our method achieves notable performance gains, increasing PSNR by " + }, + { + "bbox": [ + 313, + 336, + 555, + 492 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 336, + 555, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 336, + 555, + 492 + ], + "type": "inline_equation", + "content": "0.12\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 336, + 555, + 492 + ], + "type": "text", + "content": ", respectively, without adding significant computational overhead. Our method reaches a similar PSNR with Full Attention but is much more efficient regarding model complexity. Overall, Mamba modules balance computational efficiency and performance well." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 331, + 500, + 537, + 561 + ], + "blocks": [ + { + "bbox": [ + 363, + 318, + 504, + 328 + ], + "lines": [ + { + "bbox": [ + 363, + 318, + 504, + 328 + ], + "spans": [ + { + "bbox": [ + 363, + 318, + 504, + 328 + ], + "type": "text", + "content": "Table 3. Effective of Mamba modules." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 500, + 537, + 561 + ], + "lines": [ + { + "bbox": [ + 331, + 500, + 537, + 561 + ], + "spans": [ + { + "bbox": [ + 331, + 500, + 537, + 561 + ], + "type": "table", + "html": "
StructurePSNR (dB)Param. (M)FLOPs (G)
w/o Align.30.878.53120.4
FGDA [32]30.928.70154.3
IA [54]31.008.57148.7
DCA (ours)31.098.61159.2
", + "image_path": "264ee7ad85bc8a6b1565c0406b1230f2ff9addce7910ebbbed936d63846b09ca.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 563, + 552, + 575 + ], + "lines": [ + { + "bbox": [ + 315, + 563, + 552, + 575 + ], + "spans": [ + { + "bbox": [ + 315, + 563, + 552, + 575 + ], + "type": "text", + "content": "Table 4. Effect of Deformable Cross-mamba Alignment module." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Effect of DCA. Tab. 4 compares our DCA with existing alignment modules. Removing the alignment module significantly decreases performance (↓ 0.22dB). Our DCA achieves higher PSNR than SOTA methods like FGDA [32] and IA [54] without incurring excessive computational overhead. Specifically, DCA achieves 0.17dB and 0.09dB higher PSNR than FGDA and IA, respectively, demonstrating the effectiveness of our proposed module. By leveraging the flexibility of the deformable window when computing cross-mamba, DCA enhances the precision of aligning adjacent frames to the target frame." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 222, + 713 + ], + "type": "text", + "content": "2Clips 000, 011, 015, 020 of REDS training set." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14717" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 281, + 120 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 281, + 120 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 281, + 120 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 281, + 120 + ], + "type": "table", + "html": "
Structurew/o T2SMBw/ T2SMB†w/ T2SMB
PSNR (dB)30.9531.0231.09
Param. (M)7.878.658.61
FLOPs (G)155.6162.2159.2
", + "image_path": "811896e2bb62af1bd611b893daeef0bb78fa36fea22ec459f32f8a8bd7d4d5cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 316 + ], + "type": "text", + "content": "Effect of T2SMB. Tab. 5 shows the effect of the Temporal-to-Spatial Mamba block (T2SMB). Removing T2SMB considerably reduces the model's performance " + }, + { + "bbox": [ + 55, + 159, + 296, + 316 + ], + "type": "inline_equation", + "content": "(\\downarrow 0.14\\mathrm{dB})" + }, + { + "bbox": [ + 55, + 159, + 296, + 316 + ], + "type": "text", + "content": ", highlighting the block's importance. This performance drop occurs because the Spatial-to-Temporal Mamba block (S2TMB) cannot fully extract temporal information from the sequential input features. The T2SMB block prioritizes and explicitly focuses on extracting temporal information, enabling more comprehensive final features to be generated. We also conducted an ablation study with T2SMB using both forward and backward scans. Our findings indicate that using only the forward scan for T2SMB yields better performance while requiring less computational effort." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 325, + 298, + 386 + ], + "blocks": [ + { + "bbox": [ + 55, + 122, + 295, + 146 + ], + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 146 + ], + "type": "text", + "content": "Table 5. Effect of Temporal-to-Spatial Mamba block. † indicates that both forward and backward scans are applied to T2SMB." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 325, + 298, + 386 + ], + "lines": [ + { + "bbox": [ + 58, + 325, + 298, + 386 + ], + "spans": [ + { + "bbox": [ + 58, + 325, + 298, + 386 + ], + "type": "table", + "html": "
StructurePSNR (dB)Param. (M)FLOPs (G)
FFN [51]30.908.68136.2
TGFN w/o DWConv30.818.27124.7
TGFN w/o Split31.059.41186.1
TGFN (ours)31.098.61159.2
", + "image_path": "46d1943c0e6ccde326e545c801dc81ede6efb0ef9ea00ce2df16b5b949398102.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 575 + ], + "type": "text", + "content": "Effect of TGFN. Tab. 6 shows the effect of the temporal gated feed-forward network (TGFN). We compare models using regular FFN [51], TGFN without 3D depthwise convolution (TGFN w/o DWConv), TGFN without split channel (TGFN w/o Split), and our TGFN. Compared to FFN, TGFN has a similar number of parameters and somewhat higher FLOPs while substantially improving performance (" + }, + { + "bbox": [ + 55, + 407, + 296, + 575 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 407, + 296, + 575 + ], + "type": "text", + "content": "0.19dB). Moreover, the performance suffers significantly (" + }, + { + "bbox": [ + 55, + 407, + 296, + 575 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 407, + 296, + 575 + ], + "type": "text", + "content": "0.28dB) when the 3D DWConv is removed from TGFN, emphasizing the necessity of both spatial and temporal information. Eliminating the split operation in TGFN leads to a slight decrease in PSNR but increases model size and complexity. This indicates that redundancy in channel features negatively impacts the models' performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 65, + 584, + 287, + 624 + ], + "blocks": [ + { + "bbox": [ + 69, + 389, + 282, + 400 + ], + "lines": [ + { + "bbox": [ + 69, + 389, + 282, + 400 + ], + "spans": [ + { + "bbox": [ + 69, + 389, + 282, + 400 + ], + "type": "text", + "content": "Table 6. Effect of Temporal Gated Feed-forward Network." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 584, + 287, + 624 + ], + "lines": [ + { + "bbox": [ + 65, + 584, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 65, + 584, + 287, + 624 + ], + "type": "table", + "html": "
LossFFL [23]\\( FSL^3 \\)[14]WHFL [27]FCL (ours)
00.020.2
PSNR31.0130.9930.9030.9731.0931.02
", + "image_path": "1124b4376994d5e9718a537e3ffd00a3f02d3d2cce7ec48ab804dcb9ad70794b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "type": "text", + "content": "Effect of FCL loss. Tab. 7 shows the effectiveness of Frequency Chabornier-like loss (FCL). We compare our loss with FFL [23], FSL [14], and WHFL [27] (default weight for these losses are 0.02 for fair comparison). Our" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "FCL achieves a higher PSNR than FFL (" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\uparrow 0.08\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "), indicating that balancing low- and high-frequency components and using Real and Imaginary parts for loss calculation yields more optimal results. Moreover, our loss demonstrates higher performance compared to FSL and WHFL, with clear margins. These results show the effectiveness of our proposed FCL compared to other frequency losses. Our method achieves the best results with FCL (" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\lambda_{i} = 0.02" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "i \\in \\{Re, Im\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "). If FCL (" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\lambda_{i} = 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": ") is eliminated, performance substantially declines (" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "\\downarrow 0.12\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "), highlighting the importance of this loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 209, + 485, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 485, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 485, + 223 + ], + "type": "text", + "content": "5.4. Model Complexity Comparison" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 348 + ], + "type": "text", + "content": "Tab. 8 reports comparing parameters, FLOPs, and runtime for VSRM and other VSR methods. The runtime is measured on a RTX A6000 48GB. Our VSRM is more efficient than EDVR when considering the number of parameters and FLOPs. Although it has a moderately higher complexity than Transformer-based methods that use window-based attention, these trade-offs are reasonable when balancing performance with model complexity. Furthermore, since the acceleration and optimization of Mamba are still being explored, our method has potential for further improvements." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 358, + 561, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 626, + 295, + 648 + ], + "lines": [ + { + "bbox": [ + 55, + 626, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 295, + 648 + ], + "type": "text", + "content": "Table 7. Effect of Frequency Charbonnier-like loss (FCL). " + }, + { + "bbox": [ + 55, + 626, + 295, + 648 + ], + "type": "inline_equation", + "content": "i \\in \\{ Re, Im\\}" + }, + { + "bbox": [ + 55, + 626, + 295, + 648 + ], + "type": "text", + "content": " and we set " + }, + { + "bbox": [ + 55, + 626, + 295, + 648 + ], + "type": "inline_equation", + "content": "{\\lambda }_{Re} = {\\lambda }_{Im}" + }, + { + "bbox": [ + 55, + 626, + 295, + 648 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 358, + 561, + 430 + ], + "lines": [ + { + "bbox": [ + 317, + 358, + 561, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 358, + 561, + 430 + ], + "type": "table", + "html": "
MethodParam. (M)FLOPs (G)Runtime (ms)
VSR-T [2]32.6177.2-
EDVR [52]20.6323.6-
PSRT-recurrent [46]13.4164.6173
IART [54]13.4177.8180
VSRM (ours)17.1217.4223
", + "image_path": "506412586043fd94bf3bec64addaa116864654f8d1741d106485907a25ea19e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 433, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 455 + ], + "type": "text", + "content": "Table 8. Comparison of model complexity of different VSR models on the REDS4 dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce VSRM, which adapts Mamba for VSR for the first time and achieves state-of-the-art results on multiple benchmarks. Specifically, we propose Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to capture long-range spatio-temporal features effectively. Next, we propose a novel Deformable Cross-Mamba Alignment that uses a deformable cross-mamba mechanism to adapt to varying pixel movement flex to enhance model performance compared to previous alignment modules. Finally, we introduce a Frequency Charbonnier-like loss to reduce the gaps between reconstructed and ground truth frames in the frequency domain and better recover high-frequency content. This demonstrates further improvement by utilizing both spatial and frequency losses. VSRM has the potential to establish a solid baseline for upcoming research and shows great promise in substituting Transformer-based models. Furthermore, VSRM could be adaptable for other low-level vision tasks, including video colorization, deblurring, and denoising." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 153, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 153, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 153, + 713 + ], + "type": "text", + "content": "3FSL: Fourier Space Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14718" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 222 + ], + "type": "text", + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation(IITP)-Innovative Human Resource Development for Local Intellectualization program grant funded by the Korea government(MSIT)(IITP-2025-RS-2020-II201489). This work was also supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00573160), and the Technology Innovation Program (RS-2025-02222776) funded By the Ministry of Trade, Industry & Energy(MOTIE, Korea)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 269 + ], + "type": "text", + "content": "This work was also supported by Hyundai Motor Chung Mong-Koo Global Scholarship to Dinh Phu Tran (first author) and generous GPU support by Korea Association for ICT Promotion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 289, + 115, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 115, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 115, + 303 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 309, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 309, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 309, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 309, + 295, + 353 + ], + "type": "text", + "content": "[1] Runyuan Cai, Yue Ding, and Hongtao Lu. Freqnet: A frequency-domain image super-resolution network with discrete cosine transform. arXiv preprint arXiv:2111.10800, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 354, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 296, + 388 + ], + "type": "text", + "content": "[2] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv preprint arXiv:2106.06847, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 388, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 388, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 61, + 388, + 296, + 443 + ], + "type": "text", + "content": "[3] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4947-4956, 2021. 1, 2, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 489 + ], + "type": "text", + "content": "[4] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, pages 973-981, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 545 + ], + "type": "text", + "content": "[5] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5972-5981, 2022. 1, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 296, + 600 + ], + "type": "text", + "content": "[6] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of 1st international conference on image processing, pages 168-172. IEEE, 1994. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "text", + "content": "[7] Guo Chen, Yifei Huang, Jilan Xu, Baoqi Pei, Zhe Chen, Zhiqi Li, Jiahao Wang, Kunchang Li, Tong Lu, and Limin Wang. Video mamba suite: State space model as a versatile alternative for video understanding. arXiv preprint arXiv:2403.09626, 2024.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "text", + "content": "[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12312-12321, 2023. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[9] James W Cooley and John W Tukey. An algorithm for the machine calculation of complex fourier series. Mathematics of computation, 19(90):297-301, 1965. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "text", + "content": "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 151, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 204 + ], + "type": "text", + "content": "[11] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963-11975, 2022. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 271 + ], + "type": "text", + "content": "[12] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "text", + "content": "[13] Daniel Y. Fu, Tri Dao, Khaled K. Saab, Armin W. Thomas, Atri Rudra, and Christopher Ré. Hungry Hungry Hippos: Towards language modeling with state space models. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[14] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "text", + "content": "[15] Dario Fuoli, Martin Danelljan, Radu Timofte, and Luc Van Gool. Fast online video super-resolution with deformable attention pyramid. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1735-1744, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 426, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 458 + ], + "type": "text", + "content": "[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[17] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "text", + "content": "[18] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "text", + "content": "[19] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 624 + ], + "type": "text", + "content": "[20] Yukun Hao and Feihong Yu. Learning the frequency domain aliasing for real-world super-resolution. *Electronics*, 13(2): 250, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 315, + 625, + 553, + 658 + ], + "type": "text", + "content": "[21] Weizhe Hua, Zihang Dai, Hanxiao Liu, and Quoc Le. Transformer quality in linear time. In International conference on machine learning, pages 9099-9117. PMLR, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[22] Kui Jiang, Zhongyuan Wang, Peng Yi, and Junjun Jiang. Hierarchical dense recursive network for image superresolution. Pattern Recognition, 107:107475, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[23] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14719" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF international conference on computer vision, pages 13919-13929, 2021. 2, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "type": "text", + "content": "[24] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 294, + 195 + ], + "type": "text", + "content": "[25] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K Katsaggelos. Video super-resolution with convolutional neural networks. IEEE transactions on computational imaging, 2(2):109-122, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 242 + ], + "type": "text", + "content": "[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 243, + 295, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 295, + 286 + ], + "type": "text", + "content": "[27] Min Woo Kim and Nam Ik Cho. Whfl: Wavelet-domain high frequency loss for sketch-to-image translation. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 744-754, 2023. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 288, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 288, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 57, + 288, + 294, + 342 + ], + "type": "text", + "content": "[28] Cansu Korkmaz, A Murat Tekalp, and Zafer Dogan. Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5926-5936, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 419 + ], + "type": "text", + "content": "[29] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 294, + 455 + ], + "type": "text", + "content": "[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 294, + 510 + ], + "type": "text", + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 576 + ], + "type": "text", + "content": "[32] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[33] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Virt: A video restoration transformer. IEEE Transactions on Image Processing, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "text", + "content": "[34] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[35] Ce Liu and Deqing Sun. On bayesian adaptive video super resolution. IEEE transactions on pattern analysis and machine intelligence, 36(2):346-360, 2013. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[36] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. Advances in Neural Information Processing Systems, 34: 21297-21309, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[37] Jianping Luo, Shaofei Huang, and Yuan Yuan. Video superresolution using multi-scale pyramid 3d convolutional networks. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1882-1890, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. Advances in neural information processing systems, 29, 2016. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "text", + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "type": "text", + "content": "[40] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[41] Ali Nasiri-Sarvi, Vincent Quoc-Huy Trinh, Hassan Rivaz, and Mahdi S Hosseini. Vim4path: Self-supervised vision mamba for histopathology images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6894-6903, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 446 + ], + "type": "text", + "content": "[42] Seong-Jin Park, Hyeongseok Son, Sunghyun Cho, Ki-Sang Hong, and Seungyong Lee. Srfeat: Single image superresolution with feature discrimination. In Proceedings of the European conference on computer vision (ECCV), pages 439-455, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 513 + ], + "type": "text", + "content": "[43] Mohammad Saeed Rad, Behzad Bozortabar, Urs-Viktor Marti, Max Basler, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Srobb: Targeted perceptual loss for single image super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2710–2719, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "text", + "content": "[44] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 592 + ], + "type": "text", + "content": "[45] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "text", + "content": "[46] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking alignment in video superresolution transformers. Advances in Neural Information Processing Systems, 35:36081-36093, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[47] Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. Simplified state space layers for sequence modeling. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[48] Hang Su, Varun Jampani, Deqing Sun, Orazio Gallo, Erik Learned-Miller, and Jan Kautz. Pixel-adaptive convolutional" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14720" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 654 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 106 + ], + "type": "text", + "content": "neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11166-11175, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "text", + "content": "[49] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "text", + "content": "[50] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3360-3369, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 230, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 230, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 295, + 251 + ], + "type": "text", + "content": "[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "type": "text", + "content": "[52] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 57, + 309, + 294, + 352 + ], + "type": "text", + "content": "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 407 + ], + "type": "text", + "content": "[54] Kai Xu, Ziwei Yu, Xin Wang, Michael Bi Mi, and Angela Yao. Enhancing video super-resolution via implicit resampling-based alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2546-2555, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 295, + 453 + ], + "type": "text", + "content": "[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 454, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 454, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 294, + 509 + ], + "type": "text", + "content": "[56] Xi Yang, Wangmeng Xiang, Hui Zeng, and Lei Zhang. Realworld video super-resolution: A benchmark dataset and a decomposition based learning scheme. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4781-4790, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 510, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 542 + ], + "type": "text", + "content": "[57] Yijun Yang, Zhaohu Xing, and Lei Zhu. Vivim: a video vision mamba for medical video object segmentation. arXiv preprint arXiv:2401.14168, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 599 + ], + "type": "text", + "content": "[58] Xingyu Zhou, Leheng Zhang, Xiaorui Zhao, Keze Wang, Leida Li, and Shuhang Gu. Video super-resolution transformer with masked inter&intra-frame attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25399-25408, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 600, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 654 + ], + "type": "text", + "content": "[59] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14721" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_content_list.json b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..76ff1b26409b323dec12028a11d25077d5ce9f26 --- /dev/null +++ b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_content_list.json @@ -0,0 +1,1560 @@ +[ + { + "type": "text", + "text": "VSSD: Vision Mamba with Non-Causal State Space Duality", + "text_level": 1, + "bbox": [ + 197, + 130, + 799, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuheng Shi $^{1}$ Mingjia Li $^{3}$ Minjing Dong $^{2}$ Chang Xu $^{1}$", + "bbox": [ + 261, + 179, + 733, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Sydney, 2City University of Hong Kong, 3Tianjin University", + "bbox": [ + 194, + 198, + 800, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yuhengshi99@gmail.com, mingjiali@tju.edu.cn, minjdong@cityu.edu.hk, c.xu@sydney.edu.au", + "bbox": [ + 114, + 219, + 877, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision transformers have significantly advanced the field of computer vision, offering robust modeling capabilities and global receptive field. However, their high computational demands limit their applicability in processing long sequences. To tackle this issue, State Space Models (SSMs) have gained prominence in vision tasks as they offer linear computational complexity. Recently, State Space Duality (SSD), an improved variant of SSMs, was introduced in Mamba2 to enhance model performance and efficiency. However, the inherent causal nature of SSD/SSMs restricts their applications in non-causal vision tasks. To address this limitation, we introduce Visual State Space Duality (VSSD) model, which has a non-causal format of SSD. Specifically, we propose to discard the magnitude of interactions between the hidden state and tokens while preserving their relative weights, which relieves the dependencies of token contribution on previous tokens. Together with the involvement of multi-scan strategies, we show that the scanning results can be integrated to achieve non-causality, which not only improves the performance of SSD in vision tasks but also enhances its efficiency. We conduct extensive experiments on various benchmarks including image classification, detection, and segmentation, where VSSD surpasses existing SOTA SSM-based models. Code is available at https://github.com/YuHengsss/VSSD.", + "bbox": [ + 89, + 300, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 708, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, vision transformers [8, 9, 18, 21, 35, 36, 48, 51], pioneered by the Vision Transformer (ViT) [9], have achieved tremendous success in the field of computer vision. Thanks to the global receptive field and the robust information modeling capabilities of the attention mechanism, models [3, 35, 64] based on vision transformers have advanced significantly in various tasks such as classification [7], detection [31], and segmentation [65], surpassing classic CNN-based models [22, 25, 26, 44]. However, the quadratic computational complexity of the attention mechanism makes it resource-intensive for tasks involving long", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sequences, which limits its broader application.", + "bbox": [ + 511, + 270, + 826, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, State Space Models (SSMs) [15-17, 45], exemplified by Mamba [14], have garnered considerable attention from researchers. The S6 block, in particular, offers a global receptive field and exhibits linear complexity with respect to sequence length, presenting an efficient alternative. Pioneering vision mamba models such as Vim [67] and VMamba [34] have been developed to apply SSMs to vision tasks. Afterward, many variants were proposed [27, 43, 47, 59], which flatten 2D feature maps into 1D sequences using different scanning routes, model them with the S6 block, and subsequently integrate the results in multiple scanning routes. These multi-scan approaches improve the performance of SSMs in vision tasks, achieving results competitive with those of both CNN-based and ViT-based methods. More recently, Mamba2 [6] has introduced further enhancements to the S6 block, proposing the concept of State Space Duality (SSD). Mamba2 treats the state space transition matrix as a scalar and expands the state space dimension, thereby enhancing model performance as well as training and inference efficiency.", + "bbox": [ + 511, + 290, + 906, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SSD/SSMs inherently possess causal properties, meaning that the prediction for a given token depends solely on the preceding tokens in the input sequence. However, a major concern arises when applying SSD/SSMs to vision tasks, as image data is naturally non-causal. While another concern is flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches. We provide an illustration in Fig. 1a to facilitate a more intuitive understanding of these two concerns. In this example, the central token within the flattened 1D sequences is restricted to accessing only previous tokens, unable to integrate information from subsequent tokens. Additionally, the token 1, which is adjacent to the central token in the 2D space, becomes distantly positioned in the 1D sequence, disrupting the natural structural relationships. A common practice in previous solutions [27, 34] is to increase the scanning routes on non-causal visual features, which alleviates these two concerns to some extent. Given these observations, an important question emerges: Is there a more effective and efficient way to apply SSD to non-causal vis", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10819", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da179a25f536b99305114d0eda81ff5b781a5314bdafda557d4a7dc11129d537.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 130, + 95, + 205, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d3c80d539e3098072c9de637070f545965113b1777f750b998275f7a51a55ca1.jpg", + "image_caption": [ + "Figure 1. (a) Two challenges when applying SSM/SSD to image data. (b) and (c) are comparisons on ImageNet. Our VSSD model achieves leading accuracy and efficiency compared to CNN-based ConvNeXt [37], ViT-based Swin Transformer [35], and SSM-based VMamba [34]. The throughput of all models is measured on an NVIDIA 4090 GPU using a batch size of 128." + ], + "image_footnote": [], + "bbox": [ + 205, + 95, + 281, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/903813502cdb547739bfdd74f662e02719ff12d3d263a881f1789a07e59efce7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 95, + 357, + 231 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/55a4ad91381d970b8dfffb389fd66f39a605cac187d1e668e72975cb70e84635.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 372, + 94, + 612, + 236 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4389ddc36e9956debf04295e2053f5115c9e98abf79c11769910e19cf97a65d5.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 616, + 94, + 857, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sion data compared to multi-scan methods?", + "bbox": [ + 89, + 303, + 379, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this question, our analysis of SSD reveals that treating the matrix $\\mathbf{A}$ as a scalar facilitates a straightforward transformation of the SSD into a non-causal and position-independent manner, which we denote as Non-Causal SSD (NC-SSD). Specifically, rather than using $\\mathbf{A}$ to determine the proportion of the hidden state to be retained, we employ it to dictate the extent of the current token's contribution to the hidden states. In this scenario, each token's contribution becomes self-referential. Based on this property, we demonstrate that the causal mask in SSD can be naturally removed without the need for specific scanning routes. This observation motivates us to develop a non-causal format of SSD, in which a single global hidden state can be derived to replace the previous token-wise hidden states, resulting in not only improved accuracy but also enhanced training and inference speeds. Unlike previous multi-scan methods [34, 67] that primarily alleviate the causal limitations of SSMs, our proposed NC-SSD also resolves the issue where flattening 2D feature maps into 1D sequences disrupts the continuity of adjacent tokens. Building on NC-SSD, we introduce our Visual State Space Duality (VSSD) model and demonstrate its superior effectiveness and efficiency relative to methods based on CNNs, ViTs, and SSMs, as illustrated in Fig. 1b and 1c. Compared to the recent proposed SSM-based VMamba [34], our VSSD model outperforms it by over $1\\%$ in top-1 accuracy on ImageNet-1K dataset [7] while keeping a similar computational cost.", + "bbox": [ + 91, + 321, + 483, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are twofold. First, we analyze the state space duality and demonstrate that it can be seamlessly converted to a non-causal mode. Based on this insight, we introduce the NC-SSD, which retains the global receptive field and linear complexity benefits of the original SSD while incorporating an inherent non-causal property. Second, utilizing NC-SSD as the foundational component, we propose the VSSD model and conduct extensive experiments to validate its effectiveness. With similar parameters and computational costs, our VSSD model outperforms other State-Of-The-Art (SOTA) SSM-based models across", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "several widely recognized benchmarks in classification, object detection, and segmentation.", + "bbox": [ + 511, + 303, + 906, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 345, + 653, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers. The introduction of Vision Transformers (ViTs) [8, 9, 35, 48, 51] has revitalized the field of computer vision, which was previously dominated by Convolutional Neural Networks (CNNs) [22, 24, 26, 29, 33, 37, 44, 47, 58]. However, the quadratic computational complexity of the self-attention mechanism in ViTs poses significant challenges when processing high-resolution images, demanding significant computational resources. To address this issue, different solutions have been proposed, including hierarchical architectures [8, 20, 35, 36, 51, 53], windowed attention [21, 35, 49, 66], and variants of self-attention [52, 55, 63]. Meanwhile, linear attentions [4, 19, 28, 40] have been successful in reducing the computational complexity to a linear scale by changing the computation order of query, key, and value in self-attention. Despite this advancement, the performance of linear attention remains inferior to that of the quadratic self-attention [50] and its variants [12, 21, 66].", + "bbox": [ + 511, + 371, + 906, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models. State Space Models (SSMs) [13-17, 45] have increasingly captured the attention of researchers due to their global receptive field and linear computational complexity. Mamba [14], a prominent example of SSMs, introduced the S6 block, achieving performance on par with or better than transformers in Nature Language Processing (NLP) benchmarks. Subsequent efforts [2, 10, 27, 30, 42, 61] have explored the adaptation of the S6 block to vision tasks, yielding competitive results compared to both CNNs and ViT-based models. A central challenge in developing Mamba-based vision models is adapting the inherently causal properties of the Mamba block for non-causal image data. The most direct approach involves using different scanning routes to flatten 2D feature maps into 1D sequences which are then modeled with the S6 block and integrated. Inspired by these considerations, various scanning routes have been employed and proven effective, as", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10820", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "evidenced by multiple studies [27, 34, 39, 43, 67]. More recently, Mamba2 [6] has highlighted the significant overlap between state space models and structured masked attention, identifying them as duals of each other, and introduced the concept of State Space Duality (SSD). Building on this foundation, we demonstrate that the SSD can be transformed into a non-causal mode through a straightforward transformation, without the need of specific scanning route.", + "bbox": [ + 89, + 90, + 483, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 241, + 181, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 266, + 230, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State Space Models. Classic State Space Models (SSMs) are used to describe the dynamics of a continuous system, transforming an input sequence $x(t) \\in \\mathbb{R}$ to a latent space representation $h(t) \\in \\mathbb{R}^N$ . This representation is then utilized to generate an output sequence $y(t) \\in \\mathbb{R}$ . The mathematical formulation of an SSM is structured as follows:", + "bbox": [ + 89, + 287, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = \\overset {\\circ} {\\mathbf {A}} h (t) + \\overset {\\circ} {\\mathbf {B}} x (t), y (t) = \\mathbf {C} h (t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 388, + 482, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathring{\\mathbf{A}}\\in \\mathbb{R}^{N\\times N}$ , $\\mathring{\\mathbf{B}}\\in \\mathbb{R}^{N\\times 1}$ and $\\mathbf{C}\\in \\mathbb{R}^{1\\times N}$ are parameters. To effectively integrate continuous SSMs into deep learning architectures, discretization is essential. This process involves introducing a timescale parameter $\\Delta \\in \\mathbb{R}$ and applying the zero-order hold (ZOH) technique for discretization. Through this approach, the continuous matrices $\\mathring{\\mathbf{A}}$ and $\\mathring{\\mathbf{B}}$ are transformed into their discrete counterparts, $\\mathbf{A}$ and $\\mathbf{B}$ . Consequently, Eq. 1 is redefined in a discrete format in Eq. 2, facilitating its application within modern computational frameworks:", + "bbox": [ + 89, + 424, + 483, + 583 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh (t) = \\mathbf {A} h (t - 1) + \\mathbf {B} x (t), y (t) = \\mathbf {C} h (t),\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 595, + 390, + 611 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\mathbf {A} = e ^ {\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}}, \\mathbf {B} = (\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}) ^ {- 1} (e ^ {\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}} - \\mathbf {I}) \\boldsymbol {\\Delta} \\mathring {\\mathbf {B}} \\approx \\boldsymbol {\\Delta} \\mathring {\\mathbf {B}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 612, + 482, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{I}$ denotes the identity matrix. Furthermore, the process of Eq. 2 could be implemented in a global convolution manner as:", + "bbox": [ + 89, + 646, + 483, + 691 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = x \\odot \\mathbf {K}, \\mathbf {K} = \\left(\\mathbf {C B}, \\mathbf {C A B}, \\dots , \\mathbf {C A} ^ {L - 1} \\mathbf {B}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 703, + 482, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{K} \\in \\mathbb{R}^{L}$ represents the convolution kernel. Recently, Mamba [14] makes the parameters $\\mathbf{B}, \\mathbf{C}$ , and $\\pmb{\\Delta}$ input-dependent. This modification addresses the limitations of the Linear Time Invariant (LTI) characteristics inherent in previous SSM models [13, 16], thus enhancing the adaptability and performance of SSMs.", + "bbox": [ + 89, + 732, + 483, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Non-Causal State Space Duality", + "text_level": 1, + "bbox": [ + 89, + 832, + 372, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mamba2 [6] introduced the State Space Duality (SSD) and simplified the matrix $\\mathbf{A}$ into a scalar. This special case of selective SSMs can be implemented in both linear and", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quadratic forms. Without loss of generality, the matrix transformation form of selective state space models is expressed as follows:", + "bbox": [ + 511, + 90, + 906, + 135 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} y (t) = \\sum_ {i = 1} ^ {t} \\mathbf {C} _ {t} ^ {T} \\mathbf {A} _ {t: i + 1} \\mathbf {B} _ {i} x (i), \\text {w h e r e} \\mathbf {A} _ {t: i} = \\prod_ {i} ^ {t} \\mathbf {A} _ {i}, \\\\ y = \\operatorname {S S M} (\\mathbf {A}, \\mathbf {B}, \\mathbf {C}) (x) = \\mathbf {F} x, \\text {w h e r e} \\mathbf {F} _ {j i} = \\mathbf {C} _ {j} ^ {T} \\mathbf {A} _ {j: i + 1} \\mathbf {B} _ {i}. \\end{array} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 143, + 911, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When $\\mathbf{A}_i$ is reduced to a scalar, the quadratic form of Eq. 4 can be reformulated as:", + "bbox": [ + 511, + 191, + 906, + 222 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = \\mathbf {F} x = \\mathbf {M} \\cdot (\\mathbf {C} ^ {T} \\mathbf {B}) x,\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 228, + 720, + 244 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\mathbf {M} _ {j i} = \\left\\{ \\begin{array}{l l} A _ {j} \\times \\dots \\times A _ {i + 1} & i < j \\\\ 1 & i = j \\\\ 0 & i > j, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 244, + 906, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while its linear form is denoted as:", + "bbox": [ + 511, + 300, + 745, + 313 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh (t) = A _ {t} h (t - 1) + \\mathbf {B} _ {t} x (t), y (t) = \\mathbf {C} _ {t} h (t). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 324, + 906, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To adapt SSMs for image data, the 2D feature maps should first be flattened into a 1D sequence of tokens which are then sequentially processed. Due to the causal nature of SSMs, where each token can only access previous tokens, information propagation is inherently unidirectional. This causal property leads to suboptimal performance when handling non-causal image data, a finding that has been corroborated by previous works [34, 67]. Furthermore, flattening the 2D feature maps into 1D sequences disrupts their intrinsic structural information. For instance, tokens that are adjacent in the 2D map might end up being far apart in the 1D sequence, resulting in a loss of performance in vision tasks. Since SSD is a variant of SSMs, adopting SSD for vision tasks presents similar challenges to those observed with SSMs:", + "bbox": [ + 511, + 349, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Challenge 1: The causal property of the model restricts the flow of information, preventing later tokens from influencing earlier ones.", + "- Challenge 2: Flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches during processing." + ], + "bbox": [ + 511, + 577, + 905, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of applying causal SSD to non-causal image data, it is instructive to revisit the linear formulation of SSD. In Eq. 6, the scalar $A_{t}$ modulates the influence of the previous hidden state $h(t - 1)$ and the information in current time step. In other words, the current hidden state $h(t)$ could be viewed as a linear combination of the previous hidden state and the current input, weighted by $A_{t}$ and 1, respectively. Therefore, if we discard the magnitude of these two terms and only preserve their relative weight, Eq. 6 could be rewritten as:", + "bbox": [ + 511, + 667, + 906, + 818 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh (t) = h (t - 1) + \\frac {1}{A _ {t}} \\mathbf {B} _ {t} x (t) = \\sum_ {i = 1} ^ {t} \\frac {1}{A _ {i}} \\mathbf {B} _ {i} x (i). \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 825, + 906, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this scenario, the contribution of a specific token to the current hidden state can be directly determined by itself as", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\frac{1}{A_i}$ , rather than through the cumulative multiplication of multiple coefficients. With each token's contribution becoming self-referential, Challenge 2 is only partially addressed, as the current token can access only a subset of tokens due to the issue discussed in Challenge 1.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address Challenge 1, previous SSM-based vision models frequently employed multi-scanning routes. Specifically, in the case of ViM [67], the token sequence was subjected to both forward and reverse scanning, enabling each token to access global information. Although multi-scan approaches mitigate the causal property of SSMs, they do not address Challenge 2, as the long-range decay characteristic of SSMs remains confined to a 1D format and does not extend into 2D. To enable the acquisition of global information and thus suit non-causal image data, we also begin with a bidirectional scanning strategy. And we demonstrate that the results of forward and reverse scanning of Eq. 7 can be integrated to effectively address the aforementioned two challenges simultaneously. Let $\\mathbf{H}_i = h(i)$ denote the hidden state of the $i_{th}$ token in the bidirectional scanning approach, from which we can easily derive:", + "bbox": [ + 89, + 167, + 482, + 409 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {i} = \\sum_ {j = 1} ^ {i} \\frac {1}{A _ {j}} \\mathbf {Z} _ {j} + \\sum_ {j = - L} ^ {- i} \\frac {1}{A _ {- j}} \\mathbf {Z} _ {- j} = \\sum_ {j = 1} ^ {L} \\frac {1}{A _ {j}} \\mathbf {Z} _ {j} + \\frac {1}{A _ {i}} \\mathbf {Z} _ {i} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 422, + 480, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Z}_j = \\mathbf{B}_j x(j)$ . Since $\\frac{1}{A_i} \\mathbf{Z}_i$ primarily affects the current token and provides redundant functionality with respect to the skip connection, we omit the bias term to streamline the model. In such case, Eq. 8 can be further simplified, resulting in all tokens sharing the same hidden state $\\mathbf{H} = \\sum_{j=1}^{L} \\frac{1}{A_j} \\mathbf{Z}_j$ . In such cases, the results from forward and reverse scanning can be seamlessly combined to establish a global context, effectively equivalent to removing the causal mask and transitioning to a non-causal format. Consequently, the first challenge associated with the causal property is resolved. Although the above results are derived from a bidirectional scanning approach, it is evident that in this non-causal format, different scanning routes yield consistent outcomes. Furthermore, as demonstrated in Eq. 8, the contribution of different tokens to the current hidden state is no longer related to their scanning distance. Therefore, processing a flattened 2D feature map into a 1D sequence no longer compromises the original structural relationships. To preserve the spatial relationships between patches, we simply incorporate 2D Rotary Position Embedding (RoPE) [46] into features $\\mathbf{C}$ and $\\mathbf{B}$ . This approach effectively addresses the second challenge, as the spatial relationships among patches are now encoded through positional embeddings rather than being determined by their sequential distance in the 1D representation.", + "bbox": [ + 89, + 476, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After revising the iteration rules for the hidden state space, we update the corresponding tensor contraction algorithm or einsum notation in the linear form, following the", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mamba2 framework [6]:", + "bbox": [ + 513, + 90, + 679, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\operatorname {c o n t r a c t} (\\mathrm {L D}, \\mathrm {L N} \\rightarrow \\mathrm {L N D}) (\\mathbf {X}, \\mathbf {B})\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 116, + 846, + 133 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\operatorname {c o n t r a c t} (\\mathrm {L L}, \\mathrm {L D N} \\rightarrow \\mathrm {N D}) (\\mathbf {M}, \\mathbf {Z}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 136, + 903, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\operatorname {c o n t r a c t} (\\mathrm {L N}, \\mathrm {N D} \\rightarrow \\mathrm {L D}) (\\mathbf {C}, \\mathbf {H}).\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 155, + 839, + 171 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This algorithm involves three steps: the first step expands the input $\\mathbf{X}$ using $\\mathbf{B}$ , the second step unrolls scalar SSM recurrences to create a global hidden state $\\mathbf{H}$ , and the final step contracts the hidden state $\\mathbf{H}$ with $\\mathbf{C}$ . For clarity, the initial two steps of SSD and NC-SSD are depicted in Fig. 2a. Compared to vanilla SSD, while the operation in the first step remains unchanged, the sequence length dimension in the hidden state $\\mathbf{H}$ is eliminated in the non-causal mode, as all tokens share the same hidden state. In the final step, the output $\\mathbf{Y}$ is produced through the matrix multiplication of $\\mathbf{C}$ and $\\mathbf{H}$ . Since $\\mathbf{M}_{i,j} = \\frac{1}{A_j}$ , the matrix $\\mathbf{M}$ could be simplified to vector of weights, where $\\mathbf{m}_j = \\frac{1}{A_j}$ . In this case, integrating $\\mathbf{m}$ with either $\\mathbf{X}$ or $\\mathbf{B}$ could further simplify the transformation of Eq. 9 to:", + "bbox": [ + 511, + 176, + 906, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\mathbf {C} \\left(\\left(\\mathbf {B} \\cdot \\mathbf {m}\\right) ^ {T} \\mathbf {X}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 402, + 903, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can be regarded as a variant of linear attention. However, it is worth noting that just as $A$ plays a distinguished role in Mamba2, the vector $\\mathbf{m}$ is also crucial, as demonstrated in our ablation studies. In practice, we directly use the learned $A$ instead of $\\frac{1}{A}$ since they share the same range of values. To gain a more intuitive understanding of the role of $\\mathbf{m}$ in Eq. 10, we visualize the average of $\\mathbf{m}$ across different heads as shown in Fig. 2b. Predominantly, $\\mathbf{m}$ focuses on foreground features, enabling the model to prioritize elements that are crucial for the task at hand.", + "bbox": [ + 511, + 431, + 906, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Visual State Space Duality Model", + "text_level": 1, + "bbox": [ + 511, + 590, + 805, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Block Design. To enhance the SSD block in Mamba2 for vision applications, several modifications have been implemented beyond merely substituting the SSD with an NC-SSD to develop our Visual State Space Duality (VSSD) block. When constructing the NC-SSD block, the causal convolution 1D is replaced by a Depth-Wise Convolution (DWConv) with a kernel size of three, in line with previous vision mamba works [27, 34]. Additionally, a Feed-Forward Network (FFN) is integrated subsequent to the NC-SSD block to facilitate enhanced information exchange across channels and to maintain alignment with the established practices of classical vision transformers [9, 35, 48]. Moreover, a Local Perception Unit (LPU) [18] is incorporated prior to the NC-SSD block and the FFN, augmenting the model's capability for local feature perception. Skip connections [22] are also implemented among different blocks.", + "bbox": [ + 511, + 613, + 906, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hybrid with Self-Attention. Mamba2 demonstrates that integrating SSD with standard Multi-head Self Attention", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10822", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da900e29701389a36f79fa19d60be70de7f54242f8ae752ca114c20eb27c47ca.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 331, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0c4f561dbd8191dfb2982de25762ea3c3b767a41fb88d5fa330d70939b453e29.jpg", + "image_caption": [ + "Figure 2. (a) Illustration of the Hidden State Generation Process for SSD and NC-SSD. During the hidden state update process, NC-SSD utilizes the scalar $A$ to determine the extent of information increment for the current token, in contrast to SSD where $A$ dictates the proportion of the hidden state to be retained. Unlike the SSD, which generates token-wise hidden states, the NC-SSD produces only a global hidden state to accommodate non-causal image data. (b) Visualization of input images alongside their corresponding heat maps, derived by averaging the vector $\\mathbf{m}$ across various heads in the NC-SSD." + ], + "image_footnote": [], + "bbox": [ + 338, + 89, + 571, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb87584c74a02ee162b29810ca7c29b65fc75b5827f72e06bfdc27932e265fc5.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 588, + 133, + 903, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(MSA) yields additional improvements. In a similar vein, our model incorporates self-attention. However, unlike Mamba2, which uniformly intersperses self-attention throughout the network, we strategically replace the NCSSD block with self-attention module exclusively in the last stage. This modification leverages the robust capabilities of self-attention in processing high-level features, as evidenced by prior works [12, 32, 41] in vision tasks.", + "bbox": [ + 88, + 347, + 482, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overlapped downsampling layers. As hierarchical vision transformers [35] and vision state space models [34] predominantly employ non-overlapped convolution for downsampling, recent studies [21, 54] have demonstrated that overlapped downsampling convolutions can introduce beneficial inductive biases. Consequently, we adopt overlapped convolutions in our model, following the manner used in previous works [12, 19]. To maintain the parameter count and computational FLOPs remain comparable, we have accordingly adjusted the depth of our model.", + "bbox": [ + 88, + 469, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Architecture. We develop our VSSD model in accordance with the methods discussed above. Mirroring the design principles of established vision backbones in previous works [34, 35, 37], our VSSD model is structured into four hierarchical stages. The first three stages employ the VSSD block, while the final stage incorporates the MSA block. Detailed architectures and illustration of the proposed model could be found in the appendix.", + "bbox": [ + 88, + 621, + 482, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 760, + 215, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Classification", + "text_level": 1, + "bbox": [ + 89, + 786, + 230, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Configurations. Our experiments are conducted using the ImageNet-1K dataset [7], consistent with methodologies from prior studies [34, 35]. Each model is subjected to a training regimen spanning 300 epochs, which includes a 20-epoch warm-up phase. Optimization is performed with AdamW, where the betas are set to (0.9, 0.999) and momen", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tum at 0.9. A cosine decay scheduler manages the learning rate and is combined with a weight decay rate of 0.05. To further refine model accuracy and generalization, we incorporate exponential moving average (EMA) techniques and apply label smoothing with a coefficient of 0.1. More detailed configurations could be found in our supplementary material.", + "bbox": [ + 511, + 347, + 906, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Performance Evaluation. Tab. 1 presents a comparative comparison of our VSSD models against CNNs, ViTs, and SSM-based frameworks on the ImageNet-1K dataset [7]. The VSSD-T model, with 28M parameters and 5.0G FLOPs, secures a top-1 accuracy of $83.8\\%$ , surpassing the similarly priced VMamba-T [34] by $1.2\\%$ . In comparison with models categorized as small and base, VSSD consistently outperforms its counterparts. Specifically, the VSSD-S model, which has 50M parameters and 8.1G FLOPs, achieves an accuracy of $84.6\\%$ , outdoing VMamba-S [34] by $1.0\\%$ . For the base-sized model variant, VSSD-B, which comprises 89M parameters and 16.1G FLOPs, achieves an accuracy of $85.4\\%$ , surpassing the VMamba-B [34] by $1.8\\%$ . Furthermore, we provide a comprehensive efficiency comparison between our VSSD and widely used architectures including Swin [35], ConvNeXt [37], and VMamba [34] in Tab. 2. Notably, compared to previous works, our VSSD achieves superior accuracy-throughput trade-off while substantially reducing memory consumption. We provide comparison with SOTA works in the appendix.", + "bbox": [ + 511, + 457, + 908, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effective Receptive Field Comparison. In addition to quantitative comparisons, we conducted a comparative analysis of the Effective Receptive Field (ERF) before and after training across various models, including CNN-based ResNet50 [22] and ConvNeXt-Tiny [37], attention-based Swin-Tiny [35] and DeiT-Small [48], and SSM-based VMamba-Tiny [34], along with our VSSD-Tiny. The ERF of the central pixel is plotted using the method proposed", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/138d282ff74901ef120470c1fdb0a47550f14c85b362642bd91947378e9dc1a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodType#Param.FLOPsTop-1 Acc(%)
Tiny Models
ConvNeXt-T [37]Conv29M4.5G82.1
MambaOut-T [62]Conv27M4.5G82.7
InternImage-T [52]Conv30M5.0G83.5
DeiT-S [48]Attn22M4.6G79.8
Swin-T [35]Attn29M4.5G81.3
PVTv2-B2 [53]Attn25M4.0G82.0
Focal-T [60]Attn29M4.9G82.2
CSwin-T [8]Attn23M4.3G82.7
NAT-T [21]Attn28M4.3G83.2
CMT-S [18]Attn25M4.0G83.5
VMamba-T [34]SSM30M4.9G82.6
LVMamba-T [27]SSM26M5.7G82.7
MSVMamba-T [43]SSM33M4.9G83.0
SpatialMamba-T [56]SSM27M4.5G83.5
VSSD-TSSD28M5.0G83.8
Small Models
ConvNeXt-S [37]Conv50M8.7G83.1
MambaOut-S [62]Conv48M9.0G84.1
Swin-S [35]Attn50M8.7G83.0
PVTv2-B3 [53]Attn45M6.9G83.2
Focal-S [60]Attn50M8.7G83.5
", + "bbox": [ + 98, + 88, + 480, + 444 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/edad4298cbd5eb6ea63b2693ad036c977f3f2eee53c0a3b06fa024b973657798.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodType#Param.FLOPsTop-1 Acc(%)
Small Models (cont.)
CSwin-S [8]Attn35M6.9G83.6
NAT-S [21]Attn51M7.8G83.7
SMT-B [32]Attn32M7.7G84.3
PMamba-L2 [59]SSM25M8.1G81.6
VMamba-S [34]SSM50M8.7G83.6
LVMamba-S [27]SSM50M11.4G83.7
MSVMamba-S [43]SSM50M8.8G84.1
SpatialMamba-S [56]SSM43M7.1G84.6
VSSD-SSSD50M8.1G84.6
Base Models
ConvNeXt-B [37]Conv89M15.4G83.8
MambaOut-B [62]Conv85M15.8G84.2
DeiT-B [48]Attn86M17.5G81.8
Swin-B [35]Attn88M15.4G83.5
CSwin-S [8]Attn78M15.0G84.2
NAT-B [21]Attn90M13.7G84.3
PMamba-L3 [59]SSM50M14.4G82.3
VMamba-B [34]SSM89M15.4G83.9
MSVMamba-B [43]SSM91M16.3G84.4
SpatialMamba-B [56]SSM96M15.8G85.3
VSSD-BSSD89M16.1G85.4
", + "bbox": [ + 511, + 88, + 898, + 444 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d9d86a947b92e43e862b944d57f7c630684f01cb745f8aa60592b1b7b62db259.jpg", + "table_caption": [ + "Table 1. Accuracy Comparison across Various Models on ImageNet-1K. The LAttn is the abbreviation of linear attention." + ], + "table_footnote": [], + "table_body": "
TinySwin SmallBaseTinyConvNeXtLargeVMambaTinyVSSD
SmallBaseTinySmallBaseTinySmallBaseTinySmall
Thru. (imgs/s)15709726541685108572839114567435311263826417
Memory (MB)2402259633621670175323803777320457807826194620513481
Acc. (%)81.383.083.582.183.183.884.382.683.683.983.884.685.4
", + "bbox": [ + 127, + 468, + 867, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Efficiency comparison of different architectures: Attention-based Swin Transformer, CNN-based ConvNeXt, SSM-based VMamba, and our proposed VSSD model. All models are evaluated with an input resolution of $224\\times 224$ and batch size of 128 on an NVIDIA RTX 4090 GPU, following the experimental settings in VMamba.", + "bbox": [ + 89, + 556, + 906, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in [38], utilizing 50 randomly selected images with a resolution of $1024 \\times 1024$ from the ImageNet-1K validation set. To demonstrate the effectiveness of the proposed NC-SSD, the techniques such as hybrid self-attention and overlapped downsampling layers are not employed in our VSSD model for this analysis. Notably, only our VSSD and DeiT consistently exhibit a global receptive field both before and after training. Under after-training, a distinct cross-shaped attenuation is observed in VMamba, whereas our approach effectively eliminates the impact of token spacing on the contribution of information.", + "bbox": [ + 89, + 608, + 482, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Object Detection and Instance Segmentation", + "text_level": 1, + "bbox": [ + 89, + 787, + 468, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Configurations. Our evaluation of the VSSD model utilizes the MS COCO dataset [31] within the Mask R-CNN framework [23] for tasks related to object detection and instance segmentation. All experiments are facilitated using the MMDetection library [1]. Consistent with prior studies [34, 35], during the training phase, images are adjusted", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "such that the shorter side measures 800 pixels, while the longer side does not exceed 1333 pixels. Optimization is carried out using the AdamW optimizer, with a set learning rate of 0.0001 and a batch size of 16. When adopting the standard \"1×\" training schedule, the learning rate is reduced by a factor of 0.1 at epochs 8 and 11 while the extended \"3× + MS\" schedule sees a reduction in the learning rate by the same factor at epochs 27 and 33, following Swin [35].", + "bbox": [ + 511, + 608, + 906, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance Evaluation. Tab. 3 details the comparative performance of our model against well-established CNNs, ViTs, and other SSM-based models. Our VSSD model demonstrates comparable or superior performance in various configurations. Remarkably, our VSSD-T model demonstrates a significant advantage, outperforming SwinT [35] by margins of +4.3 in box AP and +3.5 in mask AP. This superior performance extends to both small and base variants of our architecture. When evaluated under the extended \"3x\" training schedule, VSSD maintains its competitive edge across all metrics. Specifically, VSSD-S sur", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10824", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2e5cd666c279ca01b0470ed8d6f8213ba6d11c5412ec53b3ba2bf638c600c93.jpg", + "image_caption": [ + "Figure 3. Comparison of the Effective Receptive Field (ERF) among our VSSD, CNN-based models (ResNet [22] and ConvNeXt [37]), attention-based models (Swin [35] and DeiT [48]), and the SSM-based VMamba [34]. Our VSSD effectively eliminates the impact of token spacing on the contribution of information compared to SSM-based VMamba." + ], + "image_footnote": [], + "bbox": [ + 130, + 89, + 867, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91a2c6477271e5a934a280b279df45e179a10609895db6889c944f582308fbff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mask R-CNN 1x
Method\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^m\\)\\(AP^m_{50}\\)\\(AP^m_{75}\\)
Swin-T [35]42.765.246.839.362.242.2
ConvNeXt-T [37]44.266.648.340.163.342.8
MambaOut-T [62]45.167.349.641.064.144.1
MSVMamba-T [43]46.968.751.442.566.245.8
VMamba-T [34]47.369.352.042.766.445.9
VSSD-T47.069.551.642.866.546.1
Swin-S [35]44.866.648.940.963.244.2
ConvNeXt-S [37]45.467.950.041.865.245.1
MambaOut-S [62]47.469.152.442.766.146.2
MSVMamba-S [43]48.170.152.843.267.346.5
VMamba-S [34]48.770.053.443.767.347.0
VSSD-S48.971.153.843.968.147.4
Swin-B [35]46.9--42.3--
ConvNeXt-B [37]47.069.451.742.766.346.0
MambaOut-B [62]47.469.352.243.066.446.3
VMamba-B [34]49.270.953.943.967.747.6
VSSD-B49.972.154.744.869.148.3
Mask R-CNN 3x + MS
Method\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^m\\)\\(AP^m_{50}\\)\\(AP^m_{75}\\)
Swin-T [35]46.068.150.341.665.144.9
ConvNeXt-T [37]46.267.950.841.765.044.9
MSVMamba-T [43]48.769.853.343.467.246.8
VMamba-T [34]48.870.453.543.767.447.0
VSSD-T49.171.253.544.067.647.4
Swin-S [35]48.269.852.843.267.046.1
ConvNeXt-S [37]47.970.052.742.966.946.2
MSVMamba-S [43]49.770.954.344.268.047.9
VMamba-S [34]49.970.954.744.268.247.7
VSSD-S50.772.255.845.169.448.8
", + "bbox": [ + 91, + 345, + 495, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "passes VMamba-S by margins of 0.8 and 0.9 in box AP and", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2335f6c8a8cd20cfb54d6983dd1a7616cf7c1f1634487b8e5bf831c04550d098.jpg", + "table_caption": [ + "Table 3. Object Detection and Instance Segmentation with Mask R-CNN Framework [23] on MS COCO Dataset [31]. The FLOPs are tested with an input size of $1280 \\times 800$ ." + ], + "table_footnote": [], + "table_body": "
MethodmIoU SSmIoU MS#Param.FLOPs
Swin-T [35]44.445.860M945G
ConvNeXt-T [37]46.046.760M939G
Mambaout-T [62]47.448.654M938G
FocalNet-T [60]45.847.062M998G
MSVMamba-T [43]47.948.563M953G
VMamba-T [34]48.048.862M949G
VSSD-T48.749.759M953G
Swin-S [35]47.649.581M1038G
ConvNeXt-S [37]48.749.682M1027G
Mambaout-S [62]49.550.676M1032G
FocalNet-S [60]48.050.085M1130G
VMamba-S [34]50.651.276M1028G
VSSD-S50.351.081M1021G
Swin-B [35]48.149.7121M1188G
ConvNeXt-B [37]49.149.9122M1170G
Mambaout-B [62]49.651.0112M1178G
FocalNet-B [60]49.050.5126M1354G
VMamba-B [34]51.051.6122M1170G
VSSD-B51.552.2120M1190G
", + "bbox": [ + 535, + 345, + 887, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results of Semantic Segmentation on the ADE20K Dataset [65] using the UperNet Framework [57]. FLOPs for all models are computed using input dimensions of $512 \\times 2048$ . In the table, \"SS\" represents single-scale testing, while \"MS\" indicates multi-scale testing.", + "bbox": [ + 511, + 676, + 906, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mask AP, respectively.", + "bbox": [ + 513, + 762, + 665, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 787, + 730, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Configurations. In alignment with the approaches described in Swin [35] and VMamba [34], our experiments leverage the UperNet [57] framework, utilizing an ImageNet pre-trained backbone for initialization. The training regimen spans 160K iterations with a batch size of 16, executed using the MMSegmentation library [5]. The pri", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10825", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/17ec402a95b01ef0736cd3db12e8ba38b99d0e3c8601a363613cb574fc610373.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTop-1 Acc(%)Thru. (imgs/sec)Train Thru. (imgs/sec)Param. FLOPs
SSD82.0145559729.8M4.73G
Bi-SSD82.4137341030.3M4.85G
NC-SSD82.8147171729.8M4.80G
w/ Bias82.7128060629.8M4.80G
w/o m82.4151775829.7M4.74G
", + "bbox": [ + 91, + 88, + 444, + 208 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cec34fc4c2b4c931e9309eba9e7d0b3e7aa6bd8ac5404741f39fc4755ed6d206.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Op.Down.LayersTop-1#Params.FLOPs (G)Thru. (imgs/sec)Memory. (MB)
Acc(%)
SSDPatch2, 2, 10, 282.029.8 M4.714551751
+NC-SSDPatch2, 2, 10, 282.829.8 M4.814711807
+LPUConv2, 2, 8, 283.427.2 M4.912601941
+HybridConv2, 2, 8, 283.628.2 M5.012631946
+MESAConv2, 2, 8, 483.828.2 M5.012631946
", + "bbox": [ + 446, + 88, + 901, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 663, + 210, + 681, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. (a). Ablation for different SSD variants and the components in the proposed NC-SSD. (b). Ablation study of VSSD-Tiny on ImageNet-1K. Our NC-SSD outperforms vanilla SSD in terms of accuracy with similar inference throughput and memory cost. Other techniques further enhance the performance.", + "bbox": [ + 89, + 223, + 906, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mary experiments are performed with a standard input resolution of $512 \\times 512$ . To further evaluate the robustness of our model, Multi-Scale (MS) testing is implemented. Optimization is carried out using the AdamW optimizer, with the learning rate established at $6 \\times 10^{-5}$ .", + "bbox": [ + 88, + 273, + 482, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance Evaluation. Detailed performance for our model and competitors are displayed in Tab. 4, encompassing both single-scale and multi-scale testing. In the Tiny model category, our VSSD demonstrates compelling advantages over existing architectures. Specifically, VSSD-T surpasses Swin-T, ConvNeXt-T, and VMamba-T by margins of 4.3, 2.7, and $0.7\\mathrm{mIoU}$ under single-scale testing, and 3.9, 3.0, and $0.9\\mathrm{mIoU}$ with multi-scale testing, respectively. When scaled to the base variant, VSSD-B achieves an mIoU of 52.2 under multi-scale evaluation.", + "bbox": [ + 88, + 351, + 482, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 89, + 513, + 202, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed modules, we conducted detailed ablation experiments on the VSSD-Tiny model. Using the SSD block as the token mixer and patchified downsamplers (e.g. convolution with $4 \\times 4$ kernel and stride of 4 in stem) following Swin [35] and vanilla VMamaba [34], we established the baseline configuration, detailed in the first row of Tab. 5b. For throughput testing, we utilized an NVIDIA RTX 4090 GPU with a batch size of 128. SSD Mechanisms. In our ablation study for the token mixer, we explored different scanning routes for SSD. Specifically, the Bi-SSD is introduced, where we split the channels into two parts and reverse one part to create backward scanning sequences. These sequences with opposite scanning routes are then concatenated after the SSD block. As shown in Tab. 5a, our NC-SSD model outperforms both the vanilla SSD and Bi-SSD by $0.8\\%$ and $0.4\\%$ in top-1 accuracy, respectively. Moreover, the training throughput are significantly enhanced, with NC-SSD improving training throughput by over $50\\%$ compared to the Bi-SSD. Additionally, we conduct ablation studies on key components of NC-SSD, specifically examining the impact of the omitted bias term in Eq. 10 and the vector $\\mathbf{m}$ in Eq. 7. Incorporating the bias term leads to a marginal decrease of $0.1\\%$ in top-1 accuracy while adversely affecting both training and in", + "bbox": [ + 91, + 537, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ference efficiency. In contrast, removing vector $\\mathbf{m}$ slightly improves model throughput but results in a $0.4\\%$ reduction.", + "bbox": [ + 511, + 273, + 906, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LPU and Overlapped Downsampler. Our experiments in Tab. 5b demonstrate that incorporating overlapped downsampling layers and the local perception unit (LPU) introduces beneficial inductive bias and enhances local perception capabilities. These modifications improve the top-1 accuracy by $0.6\\%$ while marginally affecting inference throughput. To maintain comparable computational costs despite the additional parameters and FLOPs introduced by these operations, we strategically reduce the model depth.", + "bbox": [ + 511, + 304, + 905, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hybrid Architecture with Attention and MESA. Tab. 5b validates the effectiveness of our hybrid approach, which combines standard attention in the final stage with overlapped downsampling. Specifically, substituting NC-SSD with standard attention in the last stage yields a $0.2\\%$ accuracy improvement while slightly increasing the parameter count. Additionally, following MLLA [19], we incorporate MESA [11] to mitigate overfitting, resulting in a further $0.2\\%$ gain in top-1 accuracy.", + "bbox": [ + 511, + 440, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, our study introduces the NC-SSD, which redefines SSD by modifying the role of matrix $\\mathbf{A}$ and eliminating the causal mask. These adaptations facilitate a transition to a non-causal mode, significantly enhancing both accuracy and efficiency. Extensive experiments demonstrate its superiority over the vanilla SSD and its multi-scan-based variants. Furthermore, by integrating techniques such as hybrid standard attention and overlapped downsampling, our VSSD model achieves comparable or superior performance compared to well-established CNNs, ViTs, and Vision SSMs across several widely-used benchmarks.", + "bbox": [ + 511, + 613, + 906, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 792, + 702, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549.", + "bbox": [ + 511, + 816, + 905, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10826", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6", + "[2] Tianxiang Chen, Zhentao Tan, Tao Gong, Qi Chu, Yue Wu, Bin Liu, Jieping Ye, and Nenghai Yu. Mim-istd: Mamba-in-mamba for efficient infrared small target detection. arXiv preprint arXiv:2403.02148, 2024. 2", + "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1", + "[4] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 2", + "[5] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7", + "[6] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. In ICML, 2024. 1, 3, 4", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 5", + "[8] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 6", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4", + "[10] Chengbin Du, Yanxi Li, and Chang Xu. Understanding robustness of visual state space models for image classification. arXiv preprint arXiv:2403.10935, 2024. 2", + "[11] Jiawei Du, Daquan Zhou, Jiashi Feng, Vincent Tan, and Joel Tianyi Zhou. Sharpness-aware training for free. NeurIPS, 2022. 8", + "[12] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 2, 5", + "[13] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3", + "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 2020. 1", + "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3", + "[17] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. NeurIPS, 2021. 1, 2", + "[18] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 1, 4, 6", + "[19] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 2, 5, 8", + "[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, 2021. 2", + "[21] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 1, 2, 5, 6", + "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 2, 4, 5, 7", + "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 6, 7", + "[24] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2", + "[25] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 1", + "[26] Gao Huang, Zhuang Liu, Geoff Pleiss, Laurens Van Der Maaten, and Kilian Weinberger. Convolutional networks with dense connectivity. IEEE TPAMI, 2019. 1, 2", + "[27] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 1, 2, 3, 4, 6", + "[28] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 2", + "[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2", + "[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, 2024. 2", + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6", + "[33] Jing Liu, Bohan Zhuang, Zhuangwei Zhuang, Yong Guo, Junzhou Huang, Jinhui Zhu, and Mingkui Tan. Discrimination-aware network pruning for deep model compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2", + "[34] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 5, 6, 7, 8", + "[36] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 1, 2", + "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 2, 5, 6, 7", + "[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, 2016. 6", + "[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3", + "[40] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. arXiv preprint arXiv:2202.08791, 2022. 2", + "[41] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. Sg-former: Self-guided transformer with evolving token reallocation. In ICCV, 2023. 5", + "[42] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2", + "[43] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. In NeurIPS, 2024. 1, 3, 6, 7", + "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2", + "[45] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 1, 2", + "[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 4", + "[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019, 1, 2", + "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 4, 5, 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2", + "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2", + "[51] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 1, 2", + "[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 2, 6", + "[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022. 2, 6", + "[54] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 5", + "[55] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023. 2", + "[56] Chaodong Xiao, Minghan Li, Zhengqiang Zhang, Deyu Meng, and Lei Zhang. Spatial-mamba: Effective visual state space models via structure-aware state fusion. In ICLR, 2025. 6", + "[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 7", + "[58] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 2", + "[59] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 6", + "[60] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks. NeurIPS, 2022. 6, 7", + "[61] Yuhuan Yang, Chaofan Ma, Jiangchao Yao, Zhun Zhong, Ya Zhang, and Yanfeng Wang. Remember: Referring image segmentation with mamba twister. arXiv preprint arXiv:2403.17839, 2024. 2", + "[62] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 6, 7", + "[63] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 2", + "[64] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10828", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 1, 7", + "[66] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2", + "[67] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 1, 2, 3, 4" + ], + "bbox": [ + 91, + 90, + 480, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10829", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_model.json b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..efc5fc0a7dd7715aaf596a0114ea087cea6b5c44 --- /dev/null +++ b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.198, + 0.131, + 0.8, + 0.154 + ], + "angle": 0, + "content": "VSSD: Vision Mamba with Non-Causal State Space Duality" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.18, + 0.735, + 0.199 + ], + "angle": 0, + "content": "Yuheng Shi\\(^{1}\\) Mingjia Li\\(^{3}\\) Minjing Dong\\(^{2}\\) Chang Xu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.199, + 0.802, + 0.218 + ], + "angle": 0, + "content": "1University of Sydney, 2City University of Hong Kong, 3Tianjin University" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.22, + 0.878, + 0.235 + ], + "angle": 0, + "content": "yuhengshi99@gmail.com, mingjiali@tju.edu.cn, minjdong@cityu.edu.hk, c.xu@sydney.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.681 + ], + "angle": 0, + "content": "Vision transformers have significantly advanced the field of computer vision, offering robust modeling capabilities and global receptive field. However, their high computational demands limit their applicability in processing long sequences. To tackle this issue, State Space Models (SSMs) have gained prominence in vision tasks as they offer linear computational complexity. Recently, State Space Duality (SSD), an improved variant of SSMs, was introduced in Mamba2 to enhance model performance and efficiency. However, the inherent causal nature of SSD/SSMs restricts their applications in non-causal vision tasks. To address this limitation, we introduce Visual State Space Duality (VSSD) model, which has a non-causal format of SSD. Specifically, we propose to discard the magnitude of interactions between the hidden state and tokens while preserving their relative weights, which relieves the dependencies of token contribution on previous tokens. Together with the involvement of multi-scan strategies, we show that the scanning results can be integrated to achieve non-causality, which not only improves the performance of SSD in vision tasks but also enhances its efficiency. We conduct extensive experiments on various benchmarks including image classification, detection, and segmentation, where VSSD surpasses existing SOTA SSM-based models. Code is available at https://github.com/YuHengsss/VSSD." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.709, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In recent years, vision transformers [8, 9, 18, 21, 35, 36, 48, 51], pioneered by the Vision Transformer (ViT) [9], have achieved tremendous success in the field of computer vision. Thanks to the global receptive field and the robust information modeling capabilities of the attention mechanism, models [3, 35, 64] based on vision transformers have advanced significantly in various tasks such as classification [7], detection [31], and segmentation [65], surpassing classic CNN-based models [22, 25, 26, 44]. However, the quadratic computational complexity of the attention mechanism makes it resource-intensive for tasks involving long" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.271, + 0.828, + 0.287 + ], + "angle": 0, + "content": "sequences, which limits its broader application." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.908, + 0.595 + ], + "angle": 0, + "content": "Recently, State Space Models (SSMs) [15-17, 45], exemplified by Mamba [14], have garnered considerable attention from researchers. The S6 block, in particular, offers a global receptive field and exhibits linear complexity with respect to sequence length, presenting an efficient alternative. Pioneering vision mamba models such as Vim [67] and VMamba [34] have been developed to apply SSMs to vision tasks. Afterward, many variants were proposed [27, 43, 47, 59], which flatten 2D feature maps into 1D sequences using different scanning routes, model them with the S6 block, and subsequently integrate the results in multiple scanning routes. These multi-scan approaches improve the performance of SSMs in vision tasks, achieving results competitive with those of both CNN-based and ViT-based methods. More recently, Mamba2 [6] has introduced further enhancements to the S6 block, proposing the concept of State Space Duality (SSD). Mamba2 treats the state space transition matrix as a scalar and expands the state space dimension, thereby enhancing model performance as well as training and inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "SSD/SSMs inherently possess causal properties, meaning that the prediction for a given token depends solely on the preceding tokens in the input sequence. However, a major concern arises when applying SSD/SSMs to vision tasks, as image data is naturally non-causal. While another concern is flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches. We provide an illustration in Fig. 1a to facilitate a more intuitive understanding of these two concerns. In this example, the central token within the flattened 1D sequences is restricted to accessing only previous tokens, unable to integrate information from subsequent tokens. Additionally, the token 1, which is adjacent to the central token in the 2D space, becomes distantly positioned in the 1D sequence, disrupting the natural structural relationships. A common practice in previous solutions [27, 34] is to increase the scanning routes on non-causal visual features, which alleviates these two concerns to some extent. Given these observations, an important question emerges: Is there a more effective and efficient way to apply SSD to non-causal vis" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.096, + 0.206, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.097, + 0.282, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.097, + 0.358, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.241, + 0.255, + 0.254 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.095, + 0.614, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.24, + 0.524, + 0.253 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.095, + 0.859, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.241, + 0.761, + 0.254 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.908, + 0.299 + ], + "angle": 0, + "content": "Figure 1. (a) Two challenges when applying SSM/SSD to image data. (b) and (c) are comparisons on ImageNet. Our VSSD model achieves leading accuracy and efficiency compared to CNN-based ConvNeXt [37], ViT-based Swin Transformer [35], and SSM-based VMamba [34]. The throughput of all models is measured on an NVIDIA 4090 GPU using a batch size of 128." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.304, + 0.38, + 0.319 + ], + "angle": 0, + "content": "sion data compared to multi-scan methods?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.485, + 0.732 + ], + "angle": 0, + "content": "To address this question, our analysis of SSD reveals that treating the matrix \\(\\mathbf{A}\\) as a scalar facilitates a straightforward transformation of the SSD into a non-causal and position-independent manner, which we denote as Non-Causal SSD (NC-SSD). Specifically, rather than using \\(\\mathbf{A}\\) to determine the proportion of the hidden state to be retained, we employ it to dictate the extent of the current token's contribution to the hidden states. In this scenario, each token's contribution becomes self-referential. Based on this property, we demonstrate that the causal mask in SSD can be naturally removed without the need for specific scanning routes. This observation motivates us to develop a non-causal format of SSD, in which a single global hidden state can be derived to replace the previous token-wise hidden states, resulting in not only improved accuracy but also enhanced training and inference speeds. Unlike previous multi-scan methods [34, 67] that primarily alleviate the causal limitations of SSMs, our proposed NC-SSD also resolves the issue where flattening 2D feature maps into 1D sequences disrupts the continuity of adjacent tokens. Building on NC-SSD, we introduce our Visual State Space Duality (VSSD) model and demonstrate its superior effectiveness and efficiency relative to methods based on CNNs, ViTs, and SSMs, as illustrated in Fig. 1b and 1c. Compared to the recent proposed SSM-based VMamba [34], our VSSD model outperforms it by over \\(1\\%\\) in top-1 accuracy on ImageNet-1K dataset [7] while keeping a similar computational cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In summary, our contributions are twofold. First, we analyze the state space duality and demonstrate that it can be seamlessly converted to a non-causal mode. Based on this insight, we introduce the NC-SSD, which retains the global receptive field and linear complexity benefits of the original SSD while incorporating an inherent non-causal property. Second, utilizing NC-SSD as the foundational component, we propose the VSSD model and conduct extensive experiments to validate its effectiveness. With similar parameters and computational costs, our VSSD model outperforms other State-Of-The-Art (SOTA) SSM-based models across" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.334 + ], + "angle": 0, + "content": "several widely recognized benchmarks in classification, object detection, and segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.655, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Vision Transformers. The introduction of Vision Transformers (ViTs) [8, 9, 35, 48, 51] has revitalized the field of computer vision, which was previously dominated by Convolutional Neural Networks (CNNs) [22, 24, 26, 29, 33, 37, 44, 47, 58]. However, the quadratic computational complexity of the self-attention mechanism in ViTs poses significant challenges when processing high-resolution images, demanding significant computational resources. To address this issue, different solutions have been proposed, including hierarchical architectures [8, 20, 35, 36, 51, 53], windowed attention [21, 35, 49, 66], and variants of self-attention [52, 55, 63]. Meanwhile, linear attentions [4, 19, 28, 40] have been successful in reducing the computational complexity to a linear scale by changing the computation order of query, key, and value in self-attention. Despite this advancement, the performance of linear attention remains inferior to that of the quadratic self-attention [50] and its variants [12, 21, 66]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "State Space Models. State Space Models (SSMs) [13-17, 45] have increasingly captured the attention of researchers due to their global receptive field and linear computational complexity. Mamba [14], a prominent example of SSMs, introduced the S6 block, achieving performance on par with or better than transformers in Nature Language Processing (NLP) benchmarks. Subsequent efforts [2, 10, 27, 30, 42, 61] have explored the adaptation of the S6 block to vision tasks, yielding competitive results compared to both CNNs and ViT-based models. A central challenge in developing Mamba-based vision models is adapting the inherently causal properties of the Mamba block for non-causal image data. The most direct approach involves using different scanning routes to flatten 2D feature maps into 1D sequences which are then modeled with the S6 block and integrated. Inspired by these considerations, various scanning routes have been employed and proven effective, as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10820" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.226 + ], + "angle": 0, + "content": "evidenced by multiple studies [27, 34, 39, 43, 67]. More recently, Mamba2 [6] has highlighted the significant overlap between state space models and structured masked attention, identifying them as duals of each other, and introduced the concept of State Space Duality (SSD). Building on this foundation, we demonstrate that the SSD can be transformed into a non-causal mode through a straightforward transformation, without the need of specific scanning route." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.242, + 0.183, + 0.256 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.232, + 0.281 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.379 + ], + "angle": 0, + "content": "State Space Models. Classic State Space Models (SSMs) are used to describe the dynamics of a continuous system, transforming an input sequence \\( x(t) \\in \\mathbb{R} \\) to a latent space representation \\( h(t) \\in \\mathbb{R}^N \\). This representation is then utilized to generate an output sequence \\( y(t) \\in \\mathbb{R} \\). The mathematical formulation of an SSM is structured as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.39, + 0.483, + 0.412 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = \\overset {\\circ} {\\mathbf {A}} h (t) + \\overset {\\circ} {\\mathbf {B}} x (t), y (t) = \\mathbf {C} h (t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.584 + ], + "angle": 0, + "content": "where \\(\\mathring{\\mathbf{A}}\\in \\mathbb{R}^{N\\times N}\\), \\(\\mathring{\\mathbf{B}}\\in \\mathbb{R}^{N\\times 1}\\) and \\(\\mathbf{C}\\in \\mathbb{R}^{1\\times N}\\) are parameters. To effectively integrate continuous SSMs into deep learning architectures, discretization is essential. This process involves introducing a timescale parameter \\(\\Delta \\in \\mathbb{R}\\) and applying the zero-order hold (ZOH) technique for discretization. Through this approach, the continuous matrices \\(\\mathring{\\mathbf{A}}\\) and \\(\\mathring{\\mathbf{B}}\\) are transformed into their discrete counterparts, \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\). Consequently, Eq. 1 is redefined in a discrete format in Eq. 2, facilitating its application within modern computational frameworks:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.596, + 0.392, + 0.612 + ], + "angle": 0, + "content": "\\[\nh (t) = \\mathbf {A} h (t - 1) + \\mathbf {B} x (t), y (t) = \\mathbf {C} h (t),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.613, + 0.483, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\mathbf {A} = e ^ {\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}}, \\mathbf {B} = (\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}) ^ {- 1} (e ^ {\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}} - \\mathbf {I}) \\boldsymbol {\\Delta} \\mathring {\\mathbf {B}} \\approx \\boldsymbol {\\Delta} \\mathring {\\mathbf {B}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.692 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}\\) denotes the identity matrix. Furthermore, the process of Eq. 2 could be implemented in a global convolution manner as:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.704, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\ny = x \\odot \\mathbf {K}, \\mathbf {K} = \\left(\\mathbf {C B}, \\mathbf {C A B}, \\dots , \\mathbf {C A} ^ {L - 1} \\mathbf {B}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.484, + 0.824 + ], + "angle": 0, + "content": "where \\(\\mathbf{K} \\in \\mathbb{R}^{L}\\) represents the convolution kernel. Recently, Mamba [14] makes the parameters \\(\\mathbf{B}, \\mathbf{C}\\), and \\(\\pmb{\\Delta}\\) input-dependent. This modification addresses the limitations of the Linear Time Invariant (LTI) characteristics inherent in previous SSM models [13, 16], thus enhancing the adaptability and performance of SSMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.373, + 0.85 + ], + "angle": 0, + "content": "3.2. Non-Causal State Space Duality" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Mamba2 [6] introduced the State Space Duality (SSD) and simplified the matrix \\(\\mathbf{A}\\) into a scalar. This special case of selective SSMs can be implemented in both linear and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.136 + ], + "angle": 0, + "content": "quadratic forms. Without loss of generality, the matrix transformation form of selective state space models is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.145, + 0.912, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} y (t) = \\sum_ {i = 1} ^ {t} \\mathbf {C} _ {t} ^ {T} \\mathbf {A} _ {t: i + 1} \\mathbf {B} _ {i} x (i), \\text {w h e r e} \\mathbf {A} _ {t: i} = \\prod_ {i} ^ {t} \\mathbf {A} _ {i}, \\\\ y = \\operatorname {S S M} (\\mathbf {A}, \\mathbf {B}, \\mathbf {C}) (x) = \\mathbf {F} x, \\text {w h e r e} \\mathbf {F} _ {j i} = \\mathbf {C} _ {j} ^ {T} \\mathbf {A} _ {j: i + 1} \\mathbf {B} _ {i}. \\end{array} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.193, + 0.907, + 0.223 + ], + "angle": 0, + "content": "When \\(\\mathbf{A}_i\\) is reduced to a scalar, the quadratic form of Eq. 4 can be reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.229, + 0.722, + 0.246 + ], + "angle": 0, + "content": "\\[\ny = \\mathbf {F} x = \\mathbf {M} \\cdot (\\mathbf {C} ^ {T} \\mathbf {B}) x,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.245, + 0.907, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\mathbf {M} _ {j i} = \\left\\{ \\begin{array}{l l} A _ {j} \\times \\dots \\times A _ {i + 1} & i < j \\\\ 1 & i = j \\\\ 0 & i > j, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.301, + 0.746, + 0.314 + ], + "angle": 0, + "content": "while its linear form is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.325, + 0.907, + 0.343 + ], + "angle": 0, + "content": "\\[\nh (t) = A _ {t} h (t - 1) + \\mathbf {B} _ {t} x (t), y (t) = \\mathbf {C} _ {t} h (t). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.576 + ], + "angle": 0, + "content": "To adapt SSMs for image data, the 2D feature maps should first be flattened into a 1D sequence of tokens which are then sequentially processed. Due to the causal nature of SSMs, where each token can only access previous tokens, information propagation is inherently unidirectional. This causal property leads to suboptimal performance when handling non-causal image data, a finding that has been corroborated by previous works [34, 67]. Furthermore, flattening the 2D feature maps into 1D sequences disrupts their intrinsic structural information. For instance, tokens that are adjacent in the 2D map might end up being far apart in the 1D sequence, resulting in a loss of performance in vision tasks. Since SSD is a variant of SSMs, adopting SSD for vision tasks presents similar challenges to those observed with SSMs:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.578, + 0.906, + 0.622 + ], + "angle": 0, + "content": "- Challenge 1: The causal property of the model restricts the flow of information, preventing later tokens from influencing earlier ones." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.623, + 0.906, + 0.668 + ], + "angle": 0, + "content": "- Challenge 2: Flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches during processing." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.578, + 0.906, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.907, + 0.819 + ], + "angle": 0, + "content": "In the context of applying causal SSD to non-causal image data, it is instructive to revisit the linear formulation of SSD. In Eq. 6, the scalar \\( A_{t} \\) modulates the influence of the previous hidden state \\( h(t - 1) \\) and the information in current time step. In other words, the current hidden state \\( h(t) \\) could be viewed as a linear combination of the previous hidden state and the current input, weighted by \\( A_{t} \\) and 1, respectively. Therefore, if we discard the magnitude of these two terms and only preserve their relative weight, Eq. 6 could be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.826, + 0.907, + 0.867 + ], + "angle": 0, + "content": "\\[\nh (t) = h (t - 1) + \\frac {1}{A _ {t}} \\mathbf {B} _ {t} x (t) = \\sum_ {i = 1} ^ {t} \\frac {1}{A _ {i}} \\mathbf {B} _ {i} x (i). \\qquad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this scenario, the contribution of a specific token to the current hidden state can be directly determined by itself as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.168 + ], + "angle": 0, + "content": "\\(\\frac{1}{A_i}\\), rather than through the cumulative multiplication of multiple coefficients. With each token's contribution becoming self-referential, Challenge 2 is only partially addressed, as the current token can access only a subset of tokens due to the issue discussed in Challenge 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.41 + ], + "angle": 0, + "content": "To address Challenge 1, previous SSM-based vision models frequently employed multi-scanning routes. Specifically, in the case of ViM [67], the token sequence was subjected to both forward and reverse scanning, enabling each token to access global information. Although multi-scan approaches mitigate the causal property of SSMs, they do not address Challenge 2, as the long-range decay characteristic of SSMs remains confined to a 1D format and does not extend into 2D. To enable the acquisition of global information and thus suit non-causal image data, we also begin with a bidirectional scanning strategy. And we demonstrate that the results of forward and reverse scanning of Eq. 7 can be integrated to effectively address the aforementioned two challenges simultaneously. Let \\(\\mathbf{H}_i = h(i)\\) denote the hidden state of the \\(i_{th}\\) token in the bidirectional scanning approach, from which we can easily derive:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.423, + 0.482, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {i} = \\sum_ {j = 1} ^ {i} \\frac {1}{A _ {j}} \\mathbf {Z} _ {j} + \\sum_ {j = - L} ^ {- i} \\frac {1}{A _ {- j}} \\mathbf {Z} _ {- j} = \\sum_ {j = 1} ^ {L} \\frac {1}{A _ {j}} \\mathbf {Z} _ {j} + \\frac {1}{A _ {i}} \\mathbf {Z} _ {i} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.854 + ], + "angle": 0, + "content": "where \\(\\mathbf{Z}_j = \\mathbf{B}_j x(j)\\). Since \\(\\frac{1}{A_i} \\mathbf{Z}_i\\) primarily affects the current token and provides redundant functionality with respect to the skip connection, we omit the bias term to streamline the model. In such case, Eq. 8 can be further simplified, resulting in all tokens sharing the same hidden state \\(\\mathbf{H} = \\sum_{j=1}^{L} \\frac{1}{A_j} \\mathbf{Z}_j\\). In such cases, the results from forward and reverse scanning can be seamlessly combined to establish a global context, effectively equivalent to removing the causal mask and transitioning to a non-causal format. Consequently, the first challenge associated with the causal property is resolved. Although the above results are derived from a bidirectional scanning approach, it is evident that in this non-causal format, different scanning routes yield consistent outcomes. Furthermore, as demonstrated in Eq. 8, the contribution of different tokens to the current hidden state is no longer related to their scanning distance. Therefore, processing a flattened 2D feature map into a 1D sequence no longer compromises the original structural relationships. To preserve the spatial relationships between patches, we simply incorporate 2D Rotary Position Embedding (RoPE) [46] into features \\(\\mathbf{C}\\) and \\(\\mathbf{B}\\). This approach effectively addresses the second challenge, as the spatial relationships among patches are now encoded through positional embeddings rather than being determined by their sequential distance in the 1D representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "After revising the iteration rules for the hidden state space, we update the corresponding tensor contraction algorithm or einsum notation in the linear form, following the" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.68, + 0.106 + ], + "angle": 0, + "content": "Mamba2 framework [6]:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.117, + 0.847, + 0.134 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\operatorname {c o n t r a c t} (\\mathrm {L D}, \\mathrm {L N} \\rightarrow \\mathrm {L N D}) (\\mathbf {X}, \\mathbf {B})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.137, + 0.905, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\operatorname {c o n t r a c t} (\\mathrm {L L}, \\mathrm {L D N} \\rightarrow \\mathrm {N D}) (\\mathbf {M}, \\mathbf {Z}) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.156, + 0.84, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\operatorname {c o n t r a c t} (\\mathrm {L N}, \\mathrm {N D} \\rightarrow \\mathrm {L D}) (\\mathbf {C}, \\mathbf {H}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.178, + 0.907, + 0.395 + ], + "angle": 0, + "content": "This algorithm involves three steps: the first step expands the input \\(\\mathbf{X}\\) using \\(\\mathbf{B}\\), the second step unrolls scalar SSM recurrences to create a global hidden state \\(\\mathbf{H}\\), and the final step contracts the hidden state \\(\\mathbf{H}\\) with \\(\\mathbf{C}\\). For clarity, the initial two steps of SSD and NC-SSD are depicted in Fig. 2a. Compared to vanilla SSD, while the operation in the first step remains unchanged, the sequence length dimension in the hidden state \\(\\mathbf{H}\\) is eliminated in the non-causal mode, as all tokens share the same hidden state. In the final step, the output \\(\\mathbf{Y}\\) is produced through the matrix multiplication of \\(\\mathbf{C}\\) and \\(\\mathbf{H}\\). Since \\(\\mathbf{M}_{i,j} = \\frac{1}{A_j}\\), the matrix \\(\\mathbf{M}\\) could be simplified to vector of weights, where \\(\\mathbf{m}_j = \\frac{1}{A_j}\\). In this case, integrating \\(\\mathbf{m}\\) with either \\(\\mathbf{X}\\) or \\(\\mathbf{B}\\) could further simplify the transformation of Eq. 9 to:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.404, + 0.905, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\mathbf {C} \\left(\\left(\\mathbf {B} \\cdot \\mathbf {m}\\right) ^ {T} \\mathbf {X}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.583 + ], + "angle": 0, + "content": "which can be regarded as a variant of linear attention. However, it is worth noting that just as \\( A \\) plays a distinguished role in Mamba2, the vector \\( \\mathbf{m} \\) is also crucial, as demonstrated in our ablation studies. In practice, we directly use the learned \\( A \\) instead of \\( \\frac{1}{A} \\) since they share the same range of values. To gain a more intuitive understanding of the role of \\( \\mathbf{m} \\) in Eq. 10, we visualize the average of \\( \\mathbf{m} \\) across different heads as shown in Fig. 2b. Predominantly, \\( \\mathbf{m} \\) focuses on foreground features, enabling the model to prioritize elements that are crucial for the task at hand." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.807, + 0.608 + ], + "angle": 0, + "content": "3.3. Visual State Space Duality Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Block Design. To enhance the SSD block in Mamba2 for vision applications, several modifications have been implemented beyond merely substituting the SSD with an NC-SSD to develop our Visual State Space Duality (VSSD) block. When constructing the NC-SSD block, the causal convolution 1D is replaced by a Depth-Wise Convolution (DWConv) with a kernel size of three, in line with previous vision mamba works [27, 34]. Additionally, a Feed-Forward Network (FFN) is integrated subsequent to the NC-SSD block to facilitate enhanced information exchange across channels and to maintain alignment with the established practices of classical vision transformers [9, 35, 48]. Moreover, a Local Perception Unit (LPU) [18] is incorporated prior to the NC-SSD block and the FFN, augmenting the model's capability for local feature perception. Skip connections [22] are also implemented among different blocks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Hybrid with Self-Attention. Mamba2 demonstrates that integrating SSD with standard Multi-head Self Attention" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10822" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.332, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.09, + 0.573, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.261, + 0.347, + 0.271 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.135, + 0.904, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.26, + 0.755, + 0.271 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.273, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 2. (a) Illustration of the Hidden State Generation Process for SSD and NC-SSD. During the hidden state update process, NC-SSD utilizes the scalar \\( A \\) to determine the extent of information increment for the current token, in contrast to SSD where \\( A \\) dictates the proportion of the hidden state to be retained. Unlike the SSD, which generates token-wise hidden states, the NC-SSD produces only a global hidden state to accommodate non-causal image data. (b) Visualization of input images alongside their corresponding heat maps, derived by averaging the vector \\( \\mathbf{m} \\) across various heads in the NC-SSD." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.483, + 0.469 + ], + "angle": 0, + "content": "(MSA) yields additional improvements. In a similar vein, our model incorporates self-attention. However, unlike Mamba2, which uniformly intersperses self-attention throughout the network, we strategically replace the NCSSD block with self-attention module exclusively in the last stage. This modification leverages the robust capabilities of self-attention in processing high-level features, as evidenced by prior works [12, 32, 41] in vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.483, + 0.621 + ], + "angle": 0, + "content": "Overlapped downsampling layers. As hierarchical vision transformers [35] and vision state space models [34] predominantly employ non-overlapped convolution for downsampling, recent studies [21, 54] have demonstrated that overlapped downsampling convolutions can introduce beneficial inductive biases. Consequently, we adopt overlapped convolutions in our model, following the manner used in previous works [12, 19]. To maintain the parameter count and computational FLOPs remain comparable, we have accordingly adjusted the depth of our model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.622, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Overall Architecture. We develop our VSSD model in accordance with the methods discussed above. Mirroring the design principles of established vision backbones in previous works [34, 35, 37], our VSSD model is structured into four hierarchical stages. The first three stages employ the VSSD block, while the final stage incorporates the MSA block. Detailed architectures and illustration of the proposed model could be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.761, + 0.216, + 0.779 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.787, + 0.231, + 0.803 + ], + "angle": 0, + "content": "4.1. Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Configurations. Our experiments are conducted using the ImageNet-1K dataset [7], consistent with methodologies from prior studies [34, 35]. Each model is subjected to a training regimen spanning 300 epochs, which includes a 20-epoch warm-up phase. Optimization is performed with AdamW, where the betas are set to (0.9, 0.999) and momen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.453 + ], + "angle": 0, + "content": "tum at 0.9. A cosine decay scheduler manages the learning rate and is combined with a weight decay rate of 0.05. To further refine model accuracy and generalization, we incorporate exponential moving average (EMA) techniques and apply label smoothing with a coefficient of 0.1. More detailed configurations could be found in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Performance Evaluation. Tab. 1 presents a comparative comparison of our VSSD models against CNNs, ViTs, and SSM-based frameworks on the ImageNet-1K dataset [7]. The VSSD-T model, with 28M parameters and 5.0G FLOPs, secures a top-1 accuracy of \\(83.8\\%\\), surpassing the similarly priced VMamba-T [34] by \\(1.2\\%\\). In comparison with models categorized as small and base, VSSD consistently outperforms its counterparts. Specifically, the VSSD-S model, which has 50M parameters and 8.1G FLOPs, achieves an accuracy of \\(84.6\\%\\), outdoing VMamba-S [34] by \\(1.0\\%\\). For the base-sized model variant, VSSD-B, which comprises 89M parameters and 16.1G FLOPs, achieves an accuracy of \\(85.4\\%\\), surpassing the VMamba-B [34] by \\(1.8\\%\\). Furthermore, we provide a comprehensive efficiency comparison between our VSSD and widely used architectures including Swin [35], ConvNeXt [37], and VMamba [34] in Tab. 2. Notably, compared to previous works, our VSSD achieves superior accuracy-throughput trade-off while substantially reducing memory consumption. We provide comparison with SOTA works in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Effective Receptive Field Comparison. In addition to quantitative comparisons, we conducted a comparative analysis of the Effective Receptive Field (ERF) before and after training across various models, including CNN-based ResNet50 [22] and ConvNeXt-Tiny [37], attention-based Swin-Tiny [35] and DeiT-Small [48], and SSM-based VMamba-Tiny [34], along with our VSSD-Tiny. The ERF of the central pixel is plotted using the method proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10823" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.482, + 0.445 + ], + "angle": 0, + "content": "
MethodType#Param.FLOPsTop-1 Acc(%)
Tiny Models
ConvNeXt-T [37]Conv29M4.5G82.1
MambaOut-T [62]Conv27M4.5G82.7
InternImage-T [52]Conv30M5.0G83.5
DeiT-S [48]Attn22M4.6G79.8
Swin-T [35]Attn29M4.5G81.3
PVTv2-B2 [53]Attn25M4.0G82.0
Focal-T [60]Attn29M4.9G82.2
CSwin-T [8]Attn23M4.3G82.7
NAT-T [21]Attn28M4.3G83.2
CMT-S [18]Attn25M4.0G83.5
VMamba-T [34]SSM30M4.9G82.6
LVMamba-T [27]SSM26M5.7G82.7
MSVMamba-T [43]SSM33M4.9G83.0
SpatialMamba-T [56]SSM27M4.5G83.5
VSSD-TSSD28M5.0G83.8
Small Models
ConvNeXt-S [37]Conv50M8.7G83.1
MambaOut-S [62]Conv48M9.0G84.1
Swin-S [35]Attn50M8.7G83.0
PVTv2-B3 [53]Attn45M6.9G83.2
Focal-S [60]Attn50M8.7G83.5
" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.089, + 0.9, + 0.445 + ], + "angle": 0, + "content": "
MethodType#Param.FLOPsTop-1 Acc(%)
Small Models (cont.)
CSwin-S [8]Attn35M6.9G83.6
NAT-S [21]Attn51M7.8G83.7
SMT-B [32]Attn32M7.7G84.3
PMamba-L2 [59]SSM25M8.1G81.6
VMamba-S [34]SSM50M8.7G83.6
LVMamba-S [27]SSM50M11.4G83.7
MSVMamba-S [43]SSM50M8.8G84.1
SpatialMamba-S [56]SSM43M7.1G84.6
VSSD-SSSD50M8.1G84.6
Base Models
ConvNeXt-B [37]Conv89M15.4G83.8
MambaOut-B [62]Conv85M15.8G84.2
DeiT-B [48]Attn86M17.5G81.8
Swin-B [35]Attn88M15.4G83.5
CSwin-S [8]Attn78M15.0G84.2
NAT-B [21]Attn90M13.7G84.3
PMamba-L3 [59]SSM50M14.4G82.3
VMamba-B [34]SSM89M15.4G83.9
MSVMamba-B [43]SSM91M16.3G84.4
SpatialMamba-B [56]SSM96M15.8G85.3
VSSD-BSSD89M16.1G85.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.448, + 0.867, + 0.462 + ], + "angle": 0, + "content": "Table 1. Accuracy Comparison across Various Models on ImageNet-1K. The LAttn is the abbreviation of linear attention." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.469, + 0.868, + 0.554 + ], + "angle": 0, + "content": "
TinySwin SmallBaseTinyConvNeXtLargeVMambaTinyVSSD
SmallBaseTinySmallBaseTinySmallBaseTinySmall
Thru. (imgs/s)15709726541685108572839114567435311263826417
Memory (MB)2402259633621670175323803777320457807826194620513481
Acc. (%)81.383.083.582.183.183.884.382.683.683.983.884.685.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.558, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Table 2. Efficiency comparison of different architectures: Attention-based Swin Transformer, CNN-based ConvNeXt, SSM-based VMamba, and our proposed VSSD model. All models are evaluated with an input resolution of \\(224\\times 224\\) and batch size of 128 on an NVIDIA RTX 4090 GPU, following the experimental settings in VMamba." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.775 + ], + "angle": 0, + "content": "in [38], utilizing 50 randomly selected images with a resolution of \\(1024 \\times 1024\\) from the ImageNet-1K validation set. To demonstrate the effectiveness of the proposed NC-SSD, the techniques such as hybrid self-attention and overlapped downsampling layers are not employed in our VSSD model for this analysis. Notably, only our VSSD and DeiT consistently exhibit a global receptive field both before and after training. Under after-training, a distinct cross-shaped attenuation is observed in VMamba, whereas our approach effectively eliminates the impact of token spacing on the contribution of information." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.788, + 0.469, + 0.804 + ], + "angle": 0, + "content": "4.2. Object Detection and Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Configurations. Our evaluation of the VSSD model utilizes the MS COCO dataset [31] within the Mask R-CNN framework [23] for tasks related to object detection and instance segmentation. All experiments are facilitated using the MMDetection library [1]. Consistent with prior studies [34, 35], during the training phase, images are adjusted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.73 + ], + "angle": 0, + "content": "such that the shorter side measures 800 pixels, while the longer side does not exceed 1333 pixels. Optimization is carried out using the AdamW optimizer, with a set learning rate of 0.0001 and a batch size of 16. When adopting the standard \"1×\" training schedule, the learning rate is reduced by a factor of 0.1 at epochs 8 and 11 while the extended \"3× + MS\" schedule sees a reduction in the learning rate by the same factor at epochs 27 and 33, following Swin [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Performance Evaluation. Tab. 3 details the comparative performance of our model against well-established CNNs, ViTs, and other SSM-based models. Our VSSD model demonstrates comparable or superior performance in various configurations. Remarkably, our VSSD-T model demonstrates a significant advantage, outperforming SwinT [35] by margins of +4.3 in box AP and +3.5 in mask AP. This superior performance extends to both small and base variants of our architecture. When evaluated under the extended \"3x\" training schedule, VSSD maintains its competitive edge across all metrics. Specifically, VSSD-S sur" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10824" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.09, + 0.868, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.908, + 0.341 + ], + "angle": 0, + "content": "Figure 3. Comparison of the Effective Receptive Field (ERF) among our VSSD, CNN-based models (ResNet [22] and ConvNeXt [37]), attention-based models (Swin [35] and DeiT [48]), and the SSM-based VMamba [34]. Our VSSD effectively eliminates the impact of token spacing on the contribution of information compared to SSM-based VMamba." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.346, + 0.496, + 0.822 + ], + "angle": 0, + "content": "
Mask R-CNN 1x
Method\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^m\\)\\(AP^m_{50}\\)\\(AP^m_{75}\\)
Swin-T [35]42.765.246.839.362.242.2
ConvNeXt-T [37]44.266.648.340.163.342.8
MambaOut-T [62]45.167.349.641.064.144.1
MSVMamba-T [43]46.968.751.442.566.245.8
VMamba-T [34]47.369.352.042.766.445.9
VSSD-T47.069.551.642.866.546.1
Swin-S [35]44.866.648.940.963.244.2
ConvNeXt-S [37]45.467.950.041.865.245.1
MambaOut-S [62]47.469.152.442.766.146.2
MSVMamba-S [43]48.170.152.843.267.346.5
VMamba-S [34]48.770.053.443.767.347.0
VSSD-S48.971.153.843.968.147.4
Swin-B [35]46.9--42.3--
ConvNeXt-B [37]47.069.451.742.766.346.0
MambaOut-B [62]47.469.352.243.066.446.3
VMamba-B [34]49.270.953.943.967.747.6
VSSD-B49.972.154.744.869.148.3
Mask R-CNN 3x + MS
Method\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^m\\)\\(AP^m_{50}\\)\\(AP^m_{75}\\)
Swin-T [35]46.068.150.341.665.144.9
ConvNeXt-T [37]46.267.950.841.765.044.9
MSVMamba-T [43]48.769.853.343.467.246.8
VMamba-T [34]48.870.453.543.767.447.0
VSSD-T49.171.253.544.067.647.4
Swin-S [35]48.269.852.843.267.046.1
ConvNeXt-S [37]47.970.052.742.966.946.2
MSVMamba-S [43]49.770.954.344.268.047.9
VMamba-S [34]49.970.954.744.268.247.7
VSSD-S50.772.255.845.169.448.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.831, + 0.485, + 0.874 + ], + "angle": 0, + "content": "Table 3. Object Detection and Instance Segmentation with Mask R-CNN Framework [23] on MS COCO Dataset [31]. The FLOPs are tested with an input size of \\(1280 \\times 800\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "passes VMamba-S by margins of 0.8 and 0.9 in box AP and" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.346, + 0.888, + 0.667 + ], + "angle": 0, + "content": "
MethodmIoU SSmIoU MS#Param.FLOPs
Swin-T [35]44.445.860M945G
ConvNeXt-T [37]46.046.760M939G
Mambaout-T [62]47.448.654M938G
FocalNet-T [60]45.847.062M998G
MSVMamba-T [43]47.948.563M953G
VMamba-T [34]48.048.862M949G
VSSD-T48.749.759M953G
Swin-S [35]47.649.581M1038G
ConvNeXt-S [37]48.749.682M1027G
Mambaout-S [62]49.550.676M1032G
FocalNet-S [60]48.050.085M1130G
VMamba-S [34]50.651.276M1028G
VSSD-S50.351.081M1021G
Swin-B [35]48.149.7121M1188G
ConvNeXt-B [37]49.149.9122M1170G
Mambaout-B [62]49.651.0112M1178G
FocalNet-B [60]49.050.5126M1354G
VMamba-B [34]51.051.6122M1170G
VSSD-B51.552.2120M1190G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.677, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Table 4. Results of Semantic Segmentation on the ADE20K Dataset [65] using the UperNet Framework [57]. FLOPs for all models are computed using input dimensions of \\(512 \\times 2048\\). In the table, \"SS\" represents single-scale testing, while \"MS\" indicates multi-scale testing." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.763, + 0.666, + 0.779 + ], + "angle": 0, + "content": "mask AP, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.731, + 0.804 + ], + "angle": 0, + "content": "4.3. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Configurations. In alignment with the approaches described in Swin [35] and VMamba [34], our experiments leverage the UperNet [57] framework, utilizing an ImageNet pre-trained backbone for initialization. The training regimen spans 160K iterations with a batch size of 16, executed using the MMSegmentation library [5]. The pri" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10825" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.445, + 0.209 + ], + "angle": 0, + "content": "
MethodTop-1 Acc(%)Thru. (imgs/sec)Train Thru. (imgs/sec)Param. FLOPs
SSD82.0145559729.8M4.73G
Bi-SSD82.4137341030.3M4.85G
NC-SSD82.8147171729.8M4.80G
w/ Bias82.7128060629.8M4.80G
w/o m82.4151775829.7M4.74G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.259, + 0.212, + 0.278, + 0.222 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.447, + 0.089, + 0.902, + 0.209 + ], + "angle": 0, + "content": "
Op.Down.LayersTop-1#Params.FLOPs (G)Thru. (imgs/sec)Memory. (MB)
Acc(%)
SSDPatch2, 2, 10, 282.029.8 M4.714551751
+NC-SSDPatch2, 2, 10, 282.829.8 M4.814711807
+LPUConv2, 2, 8, 283.427.2 M4.912601941
+HybridConv2, 2, 8, 283.628.2 M5.012631946
+MESAConv2, 2, 8, 483.828.2 M5.012631946
" + }, + { + "type": "table_caption", + "bbox": [ + 0.664, + 0.211, + 0.682, + 0.222 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.224, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Table 5. (a). Ablation for different SSD variants and the components in the proposed NC-SSD. (b). Ablation study of VSSD-Tiny on ImageNet-1K. Our NC-SSD outperforms vanilla SSD in terms of accuracy with similar inference throughput and memory cost. Other techniques further enhance the performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.275, + 0.483, + 0.35 + ], + "angle": 0, + "content": "mary experiments are performed with a standard input resolution of \\(512 \\times 512\\). To further evaluate the robustness of our model, Multi-Scale (MS) testing is implemented. Optimization is carried out using the AdamW optimizer, with the learning rate established at \\(6 \\times 10^{-5}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.352, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Performance Evaluation. Detailed performance for our model and competitors are displayed in Tab. 4, encompassing both single-scale and multi-scale testing. In the Tiny model category, our VSSD demonstrates compelling advantages over existing architectures. Specifically, VSSD-T surpasses Swin-T, ConvNeXt-T, and VMamba-T by margins of 4.3, 2.7, and \\(0.7\\mathrm{mIoU}\\) under single-scale testing, and 3.9, 3.0, and \\(0.9\\mathrm{mIoU}\\) with multi-scale testing, respectively. When scaled to the base variant, VSSD-B achieves an mIoU of 52.2 under multi-scale evaluation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.515, + 0.203, + 0.53 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.538, + 0.485, + 0.902 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed modules, we conducted detailed ablation experiments on the VSSD-Tiny model. Using the SSD block as the token mixer and patchified downsamplers (e.g. convolution with \\(4 \\times 4\\) kernel and stride of 4 in stem) following Swin [35] and vanilla VMamaba [34], we established the baseline configuration, detailed in the first row of Tab. 5b. For throughput testing, we utilized an NVIDIA RTX 4090 GPU with a batch size of 128. SSD Mechanisms. In our ablation study for the token mixer, we explored different scanning routes for SSD. Specifically, the Bi-SSD is introduced, where we split the channels into two parts and reverse one part to create backward scanning sequences. These sequences with opposite scanning routes are then concatenated after the SSD block. As shown in Tab. 5a, our NC-SSD model outperforms both the vanilla SSD and Bi-SSD by \\(0.8\\%\\) and \\(0.4\\%\\) in top-1 accuracy, respectively. Moreover, the training throughput are significantly enhanced, with NC-SSD improving training throughput by over \\(50\\%\\) compared to the Bi-SSD. Additionally, we conduct ablation studies on key components of NC-SSD, specifically examining the impact of the omitted bias term in Eq. 10 and the vector \\(\\mathbf{m}\\) in Eq. 7. Incorporating the bias term leads to a marginal decrease of \\(0.1\\%\\) in top-1 accuracy while adversely affecting both training and in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.304 + ], + "angle": 0, + "content": "ference efficiency. In contrast, removing vector \\(\\mathbf{m}\\) slightly improves model throughput but results in a \\(0.4\\%\\) reduction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.441 + ], + "angle": 0, + "content": "LPU and Overlapped Downsampler. Our experiments in Tab. 5b demonstrate that incorporating overlapped downsampling layers and the local perception unit (LPU) introduces beneficial inductive bias and enhances local perception capabilities. These modifications improve the top-1 accuracy by \\(0.6\\%\\) while marginally affecting inference throughput. To maintain comparable computational costs despite the additional parameters and FLOPs introduced by these operations, we strategically reduce the model depth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Hybrid Architecture with Attention and MESA. Tab. 5b validates the effectiveness of our hybrid approach, which combines standard attention in the final stage with overlapped downsampling. Specifically, substituting NC-SSD with standard attention in the last stage yields a \\(0.2\\%\\) accuracy improvement while slightly increasing the parameter count. Additionally, following MLLA [19], we incorporate MESA [11] to mitigate overfitting, resulting in a further \\(0.2\\%\\) gain in top-1 accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.781 + ], + "angle": 0, + "content": "In conclusion, our study introduces the NC-SSD, which redefines SSD by modifying the role of matrix \\(\\mathbf{A}\\) and eliminating the causal mask. These adaptations facilitate a transition to a non-causal mode, significantly enhancing both accuracy and efficiency. Extensive experiments demonstrate its superiority over the vanilla SSD and its multi-scan-based variants. Furthermore, by integrating techniques such as hybrid standard attention and overlapped downsampling, our VSSD model achieves comparable or superior performance compared to well-established CNNs, ViTs, and Vision SSMs across several widely-used benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.793, + 0.703, + 0.81 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.906, + 0.888 + ], + "angle": 0, + "content": "This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10826" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[1] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Tianxiang Chen, Zhentao Tan, Tao Gong, Qi Chu, Yue Wu, Bin Liu, Jieping Ye, and Nenghai Yu. Mim-istd: Mamba-in-mamba for efficient infrared small target detection. arXiv preprint arXiv:2403.02148, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[4] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[5] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.435, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[6] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. In ICML, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.479, + 0.483, + 0.52 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[8] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.668, + 0.483, + 0.709 + ], + "angle": 0, + "content": "[10] Chengbin Du, Yanxi Li, and Chang Xu. Understanding robustness of visual state space models for image classification. arXiv preprint arXiv:2403.10935, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.713, + 0.483, + 0.753 + ], + "angle": 0, + "content": "[11] Jiawei Du, Daquan Zhou, Jiashi Feng, Vincent Tan, and Joel Tianyi Zhou. Sharpness-aware training for free. NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.757, + 0.483, + 0.797 + ], + "angle": 0, + "content": "[12] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[13] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[17] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[18] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[19] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.353, + 0.904, + 0.393 + ], + "angle": 0, + "content": "[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.396, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[21] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.44, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.513, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[24] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.904, + 0.611 + ], + "angle": 0, + "content": "[25] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.615, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[26] Gao Huang, Zhuang Liu, Geoff Pleiss, Laurens Van Der Maaten, and Kilian Weinberger. Convolutional networks with dense connectivity. IEEE TPAMI, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.658, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[27] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.716, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[28] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.759, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10827" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[32] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[33] Jing Liu, Bohan Zhuang, Zhuangwei Zhuang, Yong Guo, Junzhou Huang, Jinhui Zhu, and Mingkui Tan. Discrimination-aware network pruning for deep model compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[34] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[36] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[40] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. arXiv preprint arXiv:2202.08791, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[41] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. Sg-former: Self-guided transformer with evolving token reallocation. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[42] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[43] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. In NeurIPS, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[45] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[51] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[54] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[55] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[56] Chaodong Xiao, Minghan Li, Zhengqiang Zhang, Deyu Meng, and Lei Zhang. Spatial-mamba: Effective visual state space models via structure-aware state fusion. In ICLR, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[58] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[59] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[60] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks. NeurIPS, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[61] Yuhuan Yang, Chaofan Ma, Jiangchao Yao, Zhun Zhong, Ya Zhang, and Yanfeng Wang. Remember: Referring image segmentation with mamba twister. arXiv preprint arXiv:2403.17839, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[62] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[63] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[64] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10828" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[66] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[67] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10829" + } + ] +] \ No newline at end of file diff --git a/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_origin.pdf b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..26465c17b6bac72d415cc92c4330e68809642f90 --- /dev/null +++ b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/4e862f57-d9f1-4a7e-9559-4423a19c05c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:478a09666bd1e8f94610a6fc1bf155452aa26ca121cca4b97769525a8f992f3e +size 2851392 diff --git a/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/full.md b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60cd0cd00f2b38c441084df7ceb9406d80aaffb9 --- /dev/null +++ b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/full.md @@ -0,0 +1,330 @@ +# VSSD: Vision Mamba with Non-Causal State Space Duality + +Yuheng Shi $^{1}$ Mingjia Li $^{3}$ Minjing Dong $^{2}$ Chang Xu $^{1}$ + +1University of Sydney, 2City University of Hong Kong, 3Tianjin University + +yuhengshi99@gmail.com, mingjiali@tju.edu.cn, minjdong@cityu.edu.hk, c.xu@sydney.edu.au + +# Abstract + +Vision transformers have significantly advanced the field of computer vision, offering robust modeling capabilities and global receptive field. However, their high computational demands limit their applicability in processing long sequences. To tackle this issue, State Space Models (SSMs) have gained prominence in vision tasks as they offer linear computational complexity. Recently, State Space Duality (SSD), an improved variant of SSMs, was introduced in Mamba2 to enhance model performance and efficiency. However, the inherent causal nature of SSD/SSMs restricts their applications in non-causal vision tasks. To address this limitation, we introduce Visual State Space Duality (VSSD) model, which has a non-causal format of SSD. Specifically, we propose to discard the magnitude of interactions between the hidden state and tokens while preserving their relative weights, which relieves the dependencies of token contribution on previous tokens. Together with the involvement of multi-scan strategies, we show that the scanning results can be integrated to achieve non-causality, which not only improves the performance of SSD in vision tasks but also enhances its efficiency. We conduct extensive experiments on various benchmarks including image classification, detection, and segmentation, where VSSD surpasses existing SOTA SSM-based models. Code is available at https://github.com/YuHengsss/VSSD. + +# 1. Introduction + +In recent years, vision transformers [8, 9, 18, 21, 35, 36, 48, 51], pioneered by the Vision Transformer (ViT) [9], have achieved tremendous success in the field of computer vision. Thanks to the global receptive field and the robust information modeling capabilities of the attention mechanism, models [3, 35, 64] based on vision transformers have advanced significantly in various tasks such as classification [7], detection [31], and segmentation [65], surpassing classic CNN-based models [22, 25, 26, 44]. However, the quadratic computational complexity of the attention mechanism makes it resource-intensive for tasks involving long + +sequences, which limits its broader application. + +Recently, State Space Models (SSMs) [15-17, 45], exemplified by Mamba [14], have garnered considerable attention from researchers. The S6 block, in particular, offers a global receptive field and exhibits linear complexity with respect to sequence length, presenting an efficient alternative. Pioneering vision mamba models such as Vim [67] and VMamba [34] have been developed to apply SSMs to vision tasks. Afterward, many variants were proposed [27, 43, 47, 59], which flatten 2D feature maps into 1D sequences using different scanning routes, model them with the S6 block, and subsequently integrate the results in multiple scanning routes. These multi-scan approaches improve the performance of SSMs in vision tasks, achieving results competitive with those of both CNN-based and ViT-based methods. More recently, Mamba2 [6] has introduced further enhancements to the S6 block, proposing the concept of State Space Duality (SSD). Mamba2 treats the state space transition matrix as a scalar and expands the state space dimension, thereby enhancing model performance as well as training and inference efficiency. + +SSD/SSMs inherently possess causal properties, meaning that the prediction for a given token depends solely on the preceding tokens in the input sequence. However, a major concern arises when applying SSD/SSMs to vision tasks, as image data is naturally non-causal. While another concern is flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches. We provide an illustration in Fig. 1a to facilitate a more intuitive understanding of these two concerns. In this example, the central token within the flattened 1D sequences is restricted to accessing only previous tokens, unable to integrate information from subsequent tokens. Additionally, the token 1, which is adjacent to the central token in the 2D space, becomes distantly positioned in the 1D sequence, disrupting the natural structural relationships. A common practice in previous solutions [27, 34] is to increase the scanning routes on non-causal visual features, which alleviates these two concerns to some extent. Given these observations, an important question emerges: Is there a more effective and efficient way to apply SSD to non-causal vis + +![](images/da179a25f536b99305114d0eda81ff5b781a5314bdafda557d4a7dc11129d537.jpg) +(a) + +![](images/d3c80d539e3098072c9de637070f545965113b1777f750b998275f7a51a55ca1.jpg) +Figure 1. (a) Two challenges when applying SSM/SSD to image data. (b) and (c) are comparisons on ImageNet. Our VSSD model achieves leading accuracy and efficiency compared to CNN-based ConvNeXt [37], ViT-based Swin Transformer [35], and SSM-based VMamba [34]. The throughput of all models is measured on an NVIDIA 4090 GPU using a batch size of 128. + +![](images/903813502cdb547739bfdd74f662e02719ff12d3d263a881f1789a07e59efce7.jpg) + +![](images/55a4ad91381d970b8dfffb389fd66f39a605cac187d1e668e72975cb70e84635.jpg) +(b) + +![](images/4389ddc36e9956debf04295e2053f5115c9e98abf79c11769910e19cf97a65d5.jpg) +(c) + +sion data compared to multi-scan methods? + +To address this question, our analysis of SSD reveals that treating the matrix $\mathbf{A}$ as a scalar facilitates a straightforward transformation of the SSD into a non-causal and position-independent manner, which we denote as Non-Causal SSD (NC-SSD). Specifically, rather than using $\mathbf{A}$ to determine the proportion of the hidden state to be retained, we employ it to dictate the extent of the current token's contribution to the hidden states. In this scenario, each token's contribution becomes self-referential. Based on this property, we demonstrate that the causal mask in SSD can be naturally removed without the need for specific scanning routes. This observation motivates us to develop a non-causal format of SSD, in which a single global hidden state can be derived to replace the previous token-wise hidden states, resulting in not only improved accuracy but also enhanced training and inference speeds. Unlike previous multi-scan methods [34, 67] that primarily alleviate the causal limitations of SSMs, our proposed NC-SSD also resolves the issue where flattening 2D feature maps into 1D sequences disrupts the continuity of adjacent tokens. Building on NC-SSD, we introduce our Visual State Space Duality (VSSD) model and demonstrate its superior effectiveness and efficiency relative to methods based on CNNs, ViTs, and SSMs, as illustrated in Fig. 1b and 1c. Compared to the recent proposed SSM-based VMamba [34], our VSSD model outperforms it by over $1\%$ in top-1 accuracy on ImageNet-1K dataset [7] while keeping a similar computational cost. + +In summary, our contributions are twofold. First, we analyze the state space duality and demonstrate that it can be seamlessly converted to a non-causal mode. Based on this insight, we introduce the NC-SSD, which retains the global receptive field and linear complexity benefits of the original SSD while incorporating an inherent non-causal property. Second, utilizing NC-SSD as the foundational component, we propose the VSSD model and conduct extensive experiments to validate its effectiveness. With similar parameters and computational costs, our VSSD model outperforms other State-Of-The-Art (SOTA) SSM-based models across + +several widely recognized benchmarks in classification, object detection, and segmentation. + +# 2. Related Work + +Vision Transformers. The introduction of Vision Transformers (ViTs) [8, 9, 35, 48, 51] has revitalized the field of computer vision, which was previously dominated by Convolutional Neural Networks (CNNs) [22, 24, 26, 29, 33, 37, 44, 47, 58]. However, the quadratic computational complexity of the self-attention mechanism in ViTs poses significant challenges when processing high-resolution images, demanding significant computational resources. To address this issue, different solutions have been proposed, including hierarchical architectures [8, 20, 35, 36, 51, 53], windowed attention [21, 35, 49, 66], and variants of self-attention [52, 55, 63]. Meanwhile, linear attentions [4, 19, 28, 40] have been successful in reducing the computational complexity to a linear scale by changing the computation order of query, key, and value in self-attention. Despite this advancement, the performance of linear attention remains inferior to that of the quadratic self-attention [50] and its variants [12, 21, 66]. + +State Space Models. State Space Models (SSMs) [13-17, 45] have increasingly captured the attention of researchers due to their global receptive field and linear computational complexity. Mamba [14], a prominent example of SSMs, introduced the S6 block, achieving performance on par with or better than transformers in Nature Language Processing (NLP) benchmarks. Subsequent efforts [2, 10, 27, 30, 42, 61] have explored the adaptation of the S6 block to vision tasks, yielding competitive results compared to both CNNs and ViT-based models. A central challenge in developing Mamba-based vision models is adapting the inherently causal properties of the Mamba block for non-causal image data. The most direct approach involves using different scanning routes to flatten 2D feature maps into 1D sequences which are then modeled with the S6 block and integrated. Inspired by these considerations, various scanning routes have been employed and proven effective, as + +evidenced by multiple studies [27, 34, 39, 43, 67]. More recently, Mamba2 [6] has highlighted the significant overlap between state space models and structured masked attention, identifying them as duals of each other, and introduced the concept of State Space Duality (SSD). Building on this foundation, we demonstrate that the SSD can be transformed into a non-causal mode through a straightforward transformation, without the need of specific scanning route. + +# 3. Method + +# 3.1. Preliminaries + +State Space Models. Classic State Space Models (SSMs) are used to describe the dynamics of a continuous system, transforming an input sequence $x(t) \in \mathbb{R}$ to a latent space representation $h(t) \in \mathbb{R}^N$ . This representation is then utilized to generate an output sequence $y(t) \in \mathbb{R}$ . The mathematical formulation of an SSM is structured as follows: + +$$ +h ^ {\prime} (t) = \overset {\circ} {\mathbf {A}} h (t) + \overset {\circ} {\mathbf {B}} x (t), y (t) = \mathbf {C} h (t), \tag {1} +$$ + +where $\mathring{\mathbf{A}}\in \mathbb{R}^{N\times N}$ , $\mathring{\mathbf{B}}\in \mathbb{R}^{N\times 1}$ and $\mathbf{C}\in \mathbb{R}^{1\times N}$ are parameters. To effectively integrate continuous SSMs into deep learning architectures, discretization is essential. This process involves introducing a timescale parameter $\Delta \in \mathbb{R}$ and applying the zero-order hold (ZOH) technique for discretization. Through this approach, the continuous matrices $\mathring{\mathbf{A}}$ and $\mathring{\mathbf{B}}$ are transformed into their discrete counterparts, $\mathbf{A}$ and $\mathbf{B}$ . Consequently, Eq. 1 is redefined in a discrete format in Eq. 2, facilitating its application within modern computational frameworks: + +$$ +h (t) = \mathbf {A} h (t - 1) + \mathbf {B} x (t), y (t) = \mathbf {C} h (t), +$$ + +$$ +\text {w h e r e} \mathbf {A} = e ^ {\boldsymbol {\Delta} \mathring {\mathbf {A}}}, \mathbf {B} = (\boldsymbol {\Delta} \mathring {\mathbf {A}}) ^ {- 1} (e ^ {\boldsymbol {\Delta} \mathring {\mathbf {A}}} - \mathbf {I}) \boldsymbol {\Delta} \mathring {\mathbf {B}} \approx \boldsymbol {\Delta} \mathring {\mathbf {B}}, \tag {2} +$$ + +where $\mathbf{I}$ denotes the identity matrix. Furthermore, the process of Eq. 2 could be implemented in a global convolution manner as: + +$$ +y = x \odot \mathbf {K}, \mathbf {K} = \left(\mathbf {C B}, \mathbf {C A B}, \dots , \mathbf {C A} ^ {L - 1} \mathbf {B}\right), \tag {3} +$$ + +where $\mathbf{K} \in \mathbb{R}^{L}$ represents the convolution kernel. Recently, Mamba [14] makes the parameters $\mathbf{B}, \mathbf{C}$ , and $\pmb{\Delta}$ input-dependent. This modification addresses the limitations of the Linear Time Invariant (LTI) characteristics inherent in previous SSM models [13, 16], thus enhancing the adaptability and performance of SSMs. + +# 3.2. Non-Causal State Space Duality + +Mamba2 [6] introduced the State Space Duality (SSD) and simplified the matrix $\mathbf{A}$ into a scalar. This special case of selective SSMs can be implemented in both linear and + +quadratic forms. Without loss of generality, the matrix transformation form of selective state space models is expressed as follows: + +$$ +\begin{array}{l} y (t) = \sum_ {i = 1} ^ {t} \mathbf {C} _ {t} ^ {T} \mathbf {A} _ {t: i + 1} \mathbf {B} _ {i} x (i), \text {w h e r e} \mathbf {A} _ {t: i} = \prod_ {i} ^ {t} \mathbf {A} _ {i}, \\ y = \operatorname {S S M} (\mathbf {A}, \mathbf {B}, \mathbf {C}) (x) = \mathbf {F} x, \text {w h e r e} \mathbf {F} _ {j i} = \mathbf {C} _ {j} ^ {T} \mathbf {A} _ {j: i + 1} \mathbf {B} _ {i}. \end{array} \tag {4} +$$ + +When $\mathbf{A}_i$ is reduced to a scalar, the quadratic form of Eq. 4 can be reformulated as: + +$$ +y = \mathbf {F} x = \mathbf {M} \cdot (\mathbf {C} ^ {T} \mathbf {B}) x, +$$ + +$$ +\text {w h e r e} \mathbf {M} _ {j i} = \left\{ \begin{array}{l l} A _ {j} \times \dots \times A _ {i + 1} & i < j \\ 1 & i = j \\ 0 & i > j, \end{array} \right. \tag {5} +$$ + +while its linear form is denoted as: + +$$ +h (t) = A _ {t} h (t - 1) + \mathbf {B} _ {t} x (t), y (t) = \mathbf {C} _ {t} h (t). \tag {6} +$$ + +To adapt SSMs for image data, the 2D feature maps should first be flattened into a 1D sequence of tokens which are then sequentially processed. Due to the causal nature of SSMs, where each token can only access previous tokens, information propagation is inherently unidirectional. This causal property leads to suboptimal performance when handling non-causal image data, a finding that has been corroborated by previous works [34, 67]. Furthermore, flattening the 2D feature maps into 1D sequences disrupts their intrinsic structural information. For instance, tokens that are adjacent in the 2D map might end up being far apart in the 1D sequence, resulting in a loss of performance in vision tasks. Since SSD is a variant of SSMs, adopting SSD for vision tasks presents similar challenges to those observed with SSMs: + +- Challenge 1: The causal property of the model restricts the flow of information, preventing later tokens from influencing earlier ones. +- Challenge 2: Flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches during processing. + +In the context of applying causal SSD to non-causal image data, it is instructive to revisit the linear formulation of SSD. In Eq. 6, the scalar $A_{t}$ modulates the influence of the previous hidden state $h(t - 1)$ and the information in current time step. In other words, the current hidden state $h(t)$ could be viewed as a linear combination of the previous hidden state and the current input, weighted by $A_{t}$ and 1, respectively. Therefore, if we discard the magnitude of these two terms and only preserve their relative weight, Eq. 6 could be rewritten as: + +$$ +h (t) = h (t - 1) + \frac {1}{A _ {t}} \mathbf {B} _ {t} x (t) = \sum_ {i = 1} ^ {t} \frac {1}{A _ {i}} \mathbf {B} _ {i} x (i). \qquad (7) +$$ + +In this scenario, the contribution of a specific token to the current hidden state can be directly determined by itself as + +$\frac{1}{A_i}$ , rather than through the cumulative multiplication of multiple coefficients. With each token's contribution becoming self-referential, Challenge 2 is only partially addressed, as the current token can access only a subset of tokens due to the issue discussed in Challenge 1. + +To address Challenge 1, previous SSM-based vision models frequently employed multi-scanning routes. Specifically, in the case of ViM [67], the token sequence was subjected to both forward and reverse scanning, enabling each token to access global information. Although multi-scan approaches mitigate the causal property of SSMs, they do not address Challenge 2, as the long-range decay characteristic of SSMs remains confined to a 1D format and does not extend into 2D. To enable the acquisition of global information and thus suit non-causal image data, we also begin with a bidirectional scanning strategy. And we demonstrate that the results of forward and reverse scanning of Eq. 7 can be integrated to effectively address the aforementioned two challenges simultaneously. Let $\mathbf{H}_i = h(i)$ denote the hidden state of the $i_{th}$ token in the bidirectional scanning approach, from which we can easily derive: + +$$ +\mathbf {H} _ {i} = \sum_ {j = 1} ^ {i} \frac {1}{A _ {j}} \mathbf {Z} _ {j} + \sum_ {j = - L} ^ {- i} \frac {1}{A _ {- j}} \mathbf {Z} _ {- j} = \sum_ {j = 1} ^ {L} \frac {1}{A _ {j}} \mathbf {Z} _ {j} + \frac {1}{A _ {i}} \mathbf {Z} _ {i} \tag {8} +$$ + +where $\mathbf{Z}_j = \mathbf{B}_j x(j)$ . Since $\frac{1}{A_i} \mathbf{Z}_i$ primarily affects the current token and provides redundant functionality with respect to the skip connection, we omit the bias term to streamline the model. In such case, Eq. 8 can be further simplified, resulting in all tokens sharing the same hidden state $\mathbf{H} = \sum_{j=1}^{L} \frac{1}{A_j} \mathbf{Z}_j$ . In such cases, the results from forward and reverse scanning can be seamlessly combined to establish a global context, effectively equivalent to removing the causal mask and transitioning to a non-causal format. Consequently, the first challenge associated with the causal property is resolved. Although the above results are derived from a bidirectional scanning approach, it is evident that in this non-causal format, different scanning routes yield consistent outcomes. Furthermore, as demonstrated in Eq. 8, the contribution of different tokens to the current hidden state is no longer related to their scanning distance. Therefore, processing a flattened 2D feature map into a 1D sequence no longer compromises the original structural relationships. To preserve the spatial relationships between patches, we simply incorporate 2D Rotary Position Embedding (RoPE) [46] into features $\mathbf{C}$ and $\mathbf{B}$ . This approach effectively addresses the second challenge, as the spatial relationships among patches are now encoded through positional embeddings rather than being determined by their sequential distance in the 1D representation. + +After revising the iteration rules for the hidden state space, we update the corresponding tensor contraction algorithm or einsum notation in the linear form, following the + +Mamba2 framework [6]: + +$$ +\mathbf {Z} = \operatorname {c o n t r a c t} (\mathrm {L D}, \mathrm {L N} \rightarrow \mathrm {L N D}) (\mathbf {X}, \mathbf {B}) +$$ + +$$ +\mathbf {H} = \operatorname {c o n t r a c t} (\mathrm {L L}, \mathrm {L D N} \rightarrow \mathrm {N D}) (\mathbf {M}, \mathbf {Z}) \tag {9} +$$ + +$$ +\mathbf {Y} = \operatorname {c o n t r a c t} (\mathrm {L N}, \mathrm {N D} \rightarrow \mathrm {L D}) (\mathbf {C}, \mathbf {H}). +$$ + +This algorithm involves three steps: the first step expands the input $\mathbf{X}$ using $\mathbf{B}$ , the second step unrolls scalar SSM recurrences to create a global hidden state $\mathbf{H}$ , and the final step contracts the hidden state $\mathbf{H}$ with $\mathbf{C}$ . For clarity, the initial two steps of SSD and NC-SSD are depicted in Fig. 2a. Compared to vanilla SSD, while the operation in the first step remains unchanged, the sequence length dimension in the hidden state $\mathbf{H}$ is eliminated in the non-causal mode, as all tokens share the same hidden state. In the final step, the output $\mathbf{Y}$ is produced through the matrix multiplication of $\mathbf{C}$ and $\mathbf{H}$ . Since $\mathbf{M}_{i,j} = \frac{1}{A_j}$ , the matrix $\mathbf{M}$ could be simplified to vector of weights, where $\mathbf{m}_j = \frac{1}{A_j}$ . In this case, integrating $\mathbf{m}$ with either $\mathbf{X}$ or $\mathbf{B}$ could further simplify the transformation of Eq. 9 to: + +$$ +\mathbf {Y} = \mathbf {C} \left(\left(\mathbf {B} \cdot \mathbf {m}\right) ^ {T} \mathbf {X}\right), \tag {10} +$$ + +which can be regarded as a variant of linear attention. However, it is worth noting that just as $A$ plays a distinguished role in Mamba2, the vector $\mathbf{m}$ is also crucial, as demonstrated in our ablation studies. In practice, we directly use the learned $A$ instead of $\frac{1}{A}$ since they share the same range of values. To gain a more intuitive understanding of the role of $\mathbf{m}$ in Eq. 10, we visualize the average of $\mathbf{m}$ across different heads as shown in Fig. 2b. Predominantly, $\mathbf{m}$ focuses on foreground features, enabling the model to prioritize elements that are crucial for the task at hand. + +# 3.3. Visual State Space Duality Model + +Block Design. To enhance the SSD block in Mamba2 for vision applications, several modifications have been implemented beyond merely substituting the SSD with an NC-SSD to develop our Visual State Space Duality (VSSD) block. When constructing the NC-SSD block, the causal convolution 1D is replaced by a Depth-Wise Convolution (DWConv) with a kernel size of three, in line with previous vision mamba works [27, 34]. Additionally, a Feed-Forward Network (FFN) is integrated subsequent to the NC-SSD block to facilitate enhanced information exchange across channels and to maintain alignment with the established practices of classical vision transformers [9, 35, 48]. Moreover, a Local Perception Unit (LPU) [18] is incorporated prior to the NC-SSD block and the FFN, augmenting the model's capability for local feature perception. Skip connections [22] are also implemented among different blocks. + +Hybrid with Self-Attention. Mamba2 demonstrates that integrating SSD with standard Multi-head Self Attention + +![](images/da900e29701389a36f79fa19d60be70de7f54242f8ae752ca114c20eb27c47ca.jpg) +(a) + +![](images/0c4f561dbd8191dfb2982de25762ea3c3b767a41fb88d5fa330d70939b453e29.jpg) +Figure 2. (a) Illustration of the Hidden State Generation Process for SSD and NC-SSD. During the hidden state update process, NC-SSD utilizes the scalar $A$ to determine the extent of information increment for the current token, in contrast to SSD where $A$ dictates the proportion of the hidden state to be retained. Unlike the SSD, which generates token-wise hidden states, the NC-SSD produces only a global hidden state to accommodate non-causal image data. (b) Visualization of input images alongside their corresponding heat maps, derived by averaging the vector $\mathbf{m}$ across various heads in the NC-SSD. + +![](images/fb87584c74a02ee162b29810ca7c29b65fc75b5827f72e06bfdc27932e265fc5.jpg) +(b) + +(MSA) yields additional improvements. In a similar vein, our model incorporates self-attention. However, unlike Mamba2, which uniformly intersperses self-attention throughout the network, we strategically replace the NCSSD block with self-attention module exclusively in the last stage. This modification leverages the robust capabilities of self-attention in processing high-level features, as evidenced by prior works [12, 32, 41] in vision tasks. + +Overlapped downsampling layers. As hierarchical vision transformers [35] and vision state space models [34] predominantly employ non-overlapped convolution for downsampling, recent studies [21, 54] have demonstrated that overlapped downsampling convolutions can introduce beneficial inductive biases. Consequently, we adopt overlapped convolutions in our model, following the manner used in previous works [12, 19]. To maintain the parameter count and computational FLOPs remain comparable, we have accordingly adjusted the depth of our model. + +Overall Architecture. We develop our VSSD model in accordance with the methods discussed above. Mirroring the design principles of established vision backbones in previous works [34, 35, 37], our VSSD model is structured into four hierarchical stages. The first three stages employ the VSSD block, while the final stage incorporates the MSA block. Detailed architectures and illustration of the proposed model could be found in the appendix. + +# 4. Experiment + +# 4.1. Classification + +Configurations. Our experiments are conducted using the ImageNet-1K dataset [7], consistent with methodologies from prior studies [34, 35]. Each model is subjected to a training regimen spanning 300 epochs, which includes a 20-epoch warm-up phase. Optimization is performed with AdamW, where the betas are set to (0.9, 0.999) and momen + +tum at 0.9. A cosine decay scheduler manages the learning rate and is combined with a weight decay rate of 0.05. To further refine model accuracy and generalization, we incorporate exponential moving average (EMA) techniques and apply label smoothing with a coefficient of 0.1. More detailed configurations could be found in our supplementary material. + +Performance Evaluation. Tab. 1 presents a comparative comparison of our VSSD models against CNNs, ViTs, and SSM-based frameworks on the ImageNet-1K dataset [7]. The VSSD-T model, with 28M parameters and 5.0G FLOPs, secures a top-1 accuracy of $83.8\%$ , surpassing the similarly priced VMamba-T [34] by $1.2\%$ . In comparison with models categorized as small and base, VSSD consistently outperforms its counterparts. Specifically, the VSSD-S model, which has 50M parameters and 8.1G FLOPs, achieves an accuracy of $84.6\%$ , outdoing VMamba-S [34] by $1.0\%$ . For the base-sized model variant, VSSD-B, which comprises 89M parameters and 16.1G FLOPs, achieves an accuracy of $85.4\%$ , surpassing the VMamba-B [34] by $1.8\%$ . Furthermore, we provide a comprehensive efficiency comparison between our VSSD and widely used architectures including Swin [35], ConvNeXt [37], and VMamba [34] in Tab. 2. Notably, compared to previous works, our VSSD achieves superior accuracy-throughput trade-off while substantially reducing memory consumption. We provide comparison with SOTA works in the appendix. + +Effective Receptive Field Comparison. In addition to quantitative comparisons, we conducted a comparative analysis of the Effective Receptive Field (ERF) before and after training across various models, including CNN-based ResNet50 [22] and ConvNeXt-Tiny [37], attention-based Swin-Tiny [35] and DeiT-Small [48], and SSM-based VMamba-Tiny [34], along with our VSSD-Tiny. The ERF of the central pixel is plotted using the method proposed + +
MethodType#Param.FLOPsTop-1 Acc(%)
Tiny Models
ConvNeXt-T [37]Conv29M4.5G82.1
MambaOut-T [62]Conv27M4.5G82.7
InternImage-T [52]Conv30M5.0G83.5
DeiT-S [48]Attn22M4.6G79.8
Swin-T [35]Attn29M4.5G81.3
PVTv2-B2 [53]Attn25M4.0G82.0
Focal-T [60]Attn29M4.9G82.2
CSwin-T [8]Attn23M4.3G82.7
NAT-T [21]Attn28M4.3G83.2
CMT-S [18]Attn25M4.0G83.5
VMamba-T [34]SSM30M4.9G82.6
LVMamba-T [27]SSM26M5.7G82.7
MSVMamba-T [43]SSM33M4.9G83.0
SpatialMamba-T [56]SSM27M4.5G83.5
VSSD-TSSD28M5.0G83.8
Small Models
ConvNeXt-S [37]Conv50M8.7G83.1
MambaOut-S [62]Conv48M9.0G84.1
Swin-S [35]Attn50M8.7G83.0
PVTv2-B3 [53]Attn45M6.9G83.2
Focal-S [60]Attn50M8.7G83.5
+ +
MethodType#Param.FLOPsTop-1 Acc(%)
Small Models (cont.)
CSwin-S [8]Attn35M6.9G83.6
NAT-S [21]Attn51M7.8G83.7
SMT-B [32]Attn32M7.7G84.3
PMamba-L2 [59]SSM25M8.1G81.6
VMamba-S [34]SSM50M8.7G83.6
LVMamba-S [27]SSM50M11.4G83.7
MSVMamba-S [43]SSM50M8.8G84.1
SpatialMamba-S [56]SSM43M7.1G84.6
VSSD-SSSD50M8.1G84.6
Base Models
ConvNeXt-B [37]Conv89M15.4G83.8
MambaOut-B [62]Conv85M15.8G84.2
DeiT-B [48]Attn86M17.5G81.8
Swin-B [35]Attn88M15.4G83.5
CSwin-S [8]Attn78M15.0G84.2
NAT-B [21]Attn90M13.7G84.3
PMamba-L3 [59]SSM50M14.4G82.3
VMamba-B [34]SSM89M15.4G83.9
MSVMamba-B [43]SSM91M16.3G84.4
SpatialMamba-B [56]SSM96M15.8G85.3
VSSD-BSSD89M16.1G85.4
+ +Table 1. Accuracy Comparison across Various Models on ImageNet-1K. The LAttn is the abbreviation of linear attention. + +
TinySwin SmallBaseTinyConvNeXtLargeVMambaTinyVSSD
SmallBaseTinySmallBaseTinySmallBaseTinySmall
Thru. (imgs/s)15709726541685108572839114567435311263826417
Memory (MB)2402259633621670175323803777320457807826194620513481
Acc. (%)81.383.083.582.183.183.884.382.683.683.983.884.685.4
+ +Table 2. Efficiency comparison of different architectures: Attention-based Swin Transformer, CNN-based ConvNeXt, SSM-based VMamba, and our proposed VSSD model. All models are evaluated with an input resolution of $224\times 224$ and batch size of 128 on an NVIDIA RTX 4090 GPU, following the experimental settings in VMamba. + +in [38], utilizing 50 randomly selected images with a resolution of $1024 \times 1024$ from the ImageNet-1K validation set. To demonstrate the effectiveness of the proposed NC-SSD, the techniques such as hybrid self-attention and overlapped downsampling layers are not employed in our VSSD model for this analysis. Notably, only our VSSD and DeiT consistently exhibit a global receptive field both before and after training. Under after-training, a distinct cross-shaped attenuation is observed in VMamba, whereas our approach effectively eliminates the impact of token spacing on the contribution of information. + +# 4.2. Object Detection and Instance Segmentation + +Configurations. Our evaluation of the VSSD model utilizes the MS COCO dataset [31] within the Mask R-CNN framework [23] for tasks related to object detection and instance segmentation. All experiments are facilitated using the MMDetection library [1]. Consistent with prior studies [34, 35], during the training phase, images are adjusted + +such that the shorter side measures 800 pixels, while the longer side does not exceed 1333 pixels. Optimization is carried out using the AdamW optimizer, with a set learning rate of 0.0001 and a batch size of 16. When adopting the standard "1×" training schedule, the learning rate is reduced by a factor of 0.1 at epochs 8 and 11 while the extended "3× + MS" schedule sees a reduction in the learning rate by the same factor at epochs 27 and 33, following Swin [35]. + +Performance Evaluation. Tab. 3 details the comparative performance of our model against well-established CNNs, ViTs, and other SSM-based models. Our VSSD model demonstrates comparable or superior performance in various configurations. Remarkably, our VSSD-T model demonstrates a significant advantage, outperforming SwinT [35] by margins of +4.3 in box AP and +3.5 in mask AP. This superior performance extends to both small and base variants of our architecture. When evaluated under the extended "3x" training schedule, VSSD maintains its competitive edge across all metrics. Specifically, VSSD-S sur + +![](images/f2e5cd666c279ca01b0470ed8d6f8213ba6d11c5412ec53b3ba2bf638c600c93.jpg) +Figure 3. Comparison of the Effective Receptive Field (ERF) among our VSSD, CNN-based models (ResNet [22] and ConvNeXt [37]), attention-based models (Swin [35] and DeiT [48]), and the SSM-based VMamba [34]. Our VSSD effectively eliminates the impact of token spacing on the contribution of information compared to SSM-based VMamba. + +
Mask R-CNN 1x
Method\(AP^b\)\(AP^b_{50}\)\(AP^b_{75}\)\(AP^m\)\(AP^m_{50}\)\(AP^m_{75}\)
Swin-T [35]42.765.246.839.362.242.2
ConvNeXt-T [37]44.266.648.340.163.342.8
MambaOut-T [62]45.167.349.641.064.144.1
MSVMamba-T [43]46.968.751.442.566.245.8
VMamba-T [34]47.369.352.042.766.445.9
VSSD-T47.069.551.642.866.546.1
Swin-S [35]44.866.648.940.963.244.2
ConvNeXt-S [37]45.467.950.041.865.245.1
MambaOut-S [62]47.469.152.442.766.146.2
MSVMamba-S [43]48.170.152.843.267.346.5
VMamba-S [34]48.770.053.443.767.347.0
VSSD-S48.971.153.843.968.147.4
Swin-B [35]46.9--42.3--
ConvNeXt-B [37]47.069.451.742.766.346.0
MambaOut-B [62]47.469.352.243.066.446.3
VMamba-B [34]49.270.953.943.967.747.6
VSSD-B49.972.154.744.869.148.3
Mask R-CNN 3x + MS
Method\(AP^b\)\(AP^b_{50}\)\(AP^b_{75}\)\(AP^m\)\(AP^m_{50}\)\(AP^m_{75}\)
Swin-T [35]46.068.150.341.665.144.9
ConvNeXt-T [37]46.267.950.841.765.044.9
MSVMamba-T [43]48.769.853.343.467.246.8
VMamba-T [34]48.870.453.543.767.447.0
VSSD-T49.171.253.544.067.647.4
Swin-S [35]48.269.852.843.267.046.1
ConvNeXt-S [37]47.970.052.742.966.946.2
MSVMamba-S [43]49.770.954.344.268.047.9
VMamba-S [34]49.970.954.744.268.247.7
VSSD-S50.772.255.845.169.448.8
+ +passes VMamba-S by margins of 0.8 and 0.9 in box AP and + +Table 3. Object Detection and Instance Segmentation with Mask R-CNN Framework [23] on MS COCO Dataset [31]. The FLOPs are tested with an input size of $1280 \times 800$ . + +
MethodmIoU SSmIoU MS#Param.FLOPs
Swin-T [35]44.445.860M945G
ConvNeXt-T [37]46.046.760M939G
Mambaout-T [62]47.448.654M938G
FocalNet-T [60]45.847.062M998G
MSVMamba-T [43]47.948.563M953G
VMamba-T [34]48.048.862M949G
VSSD-T48.749.759M953G
Swin-S [35]47.649.581M1038G
ConvNeXt-S [37]48.749.682M1027G
Mambaout-S [62]49.550.676M1032G
FocalNet-S [60]48.050.085M1130G
VMamba-S [34]50.651.276M1028G
VSSD-S50.351.081M1021G
Swin-B [35]48.149.7121M1188G
ConvNeXt-B [37]49.149.9122M1170G
Mambaout-B [62]49.651.0112M1178G
FocalNet-B [60]49.050.5126M1354G
VMamba-B [34]51.051.6122M1170G
VSSD-B51.552.2120M1190G
+ +Table 4. Results of Semantic Segmentation on the ADE20K Dataset [65] using the UperNet Framework [57]. FLOPs for all models are computed using input dimensions of $512 \times 2048$ . In the table, "SS" represents single-scale testing, while "MS" indicates multi-scale testing. + +mask AP, respectively. + +# 4.3. Semantic Segmentation + +Configurations. In alignment with the approaches described in Swin [35] and VMamba [34], our experiments leverage the UperNet [57] framework, utilizing an ImageNet pre-trained backbone for initialization. The training regimen spans 160K iterations with a batch size of 16, executed using the MMSegmentation library [5]. The pri + +
MethodTop-1 Acc(%)Thru. (imgs/sec)Train Thru. (imgs/sec)Param. FLOPs
SSD82.0145559729.8M4.73G
Bi-SSD82.4137341030.3M4.85G
NC-SSD82.8147171729.8M4.80G
w/ Bias82.7128060629.8M4.80G
w/o m82.4151775829.7M4.74G
+ +(a) + +
Op.Down.LayersTop-1#Params.FLOPs (G)Thru. (imgs/sec)Memory. (MB)
Acc(%)
SSDPatch2, 2, 10, 282.029.8 M4.714551751
+NC-SSDPatch2, 2, 10, 282.829.8 M4.814711807
+LPUConv2, 2, 8, 283.427.2 M4.912601941
+HybridConv2, 2, 8, 283.628.2 M5.012631946
+MESAConv2, 2, 8, 483.828.2 M5.012631946
+ +(b) + +Table 5. (a). Ablation for different SSD variants and the components in the proposed NC-SSD. (b). Ablation study of VSSD-Tiny on ImageNet-1K. Our NC-SSD outperforms vanilla SSD in terms of accuracy with similar inference throughput and memory cost. Other techniques further enhance the performance. + +mary experiments are performed with a standard input resolution of $512 \times 512$ . To further evaluate the robustness of our model, Multi-Scale (MS) testing is implemented. Optimization is carried out using the AdamW optimizer, with the learning rate established at $6 \times 10^{-5}$ . + +Performance Evaluation. Detailed performance for our model and competitors are displayed in Tab. 4, encompassing both single-scale and multi-scale testing. In the Tiny model category, our VSSD demonstrates compelling advantages over existing architectures. Specifically, VSSD-T surpasses Swin-T, ConvNeXt-T, and VMamba-T by margins of 4.3, 2.7, and $0.7\mathrm{mIoU}$ under single-scale testing, and 3.9, 3.0, and $0.9\mathrm{mIoU}$ with multi-scale testing, respectively. When scaled to the base variant, VSSD-B achieves an mIoU of 52.2 under multi-scale evaluation. + +# 4.4. Ablations + +To validate the effectiveness of the proposed modules, we conducted detailed ablation experiments on the VSSD-Tiny model. Using the SSD block as the token mixer and patchified downsamplers (e.g. convolution with $4 \times 4$ kernel and stride of 4 in stem) following Swin [35] and vanilla VMamaba [34], we established the baseline configuration, detailed in the first row of Tab. 5b. For throughput testing, we utilized an NVIDIA RTX 4090 GPU with a batch size of 128. SSD Mechanisms. In our ablation study for the token mixer, we explored different scanning routes for SSD. Specifically, the Bi-SSD is introduced, where we split the channels into two parts and reverse one part to create backward scanning sequences. These sequences with opposite scanning routes are then concatenated after the SSD block. As shown in Tab. 5a, our NC-SSD model outperforms both the vanilla SSD and Bi-SSD by $0.8\%$ and $0.4\%$ in top-1 accuracy, respectively. Moreover, the training throughput are significantly enhanced, with NC-SSD improving training throughput by over $50\%$ compared to the Bi-SSD. Additionally, we conduct ablation studies on key components of NC-SSD, specifically examining the impact of the omitted bias term in Eq. 10 and the vector $\mathbf{m}$ in Eq. 7. Incorporating the bias term leads to a marginal decrease of $0.1\%$ in top-1 accuracy while adversely affecting both training and in + +ference efficiency. In contrast, removing vector $\mathbf{m}$ slightly improves model throughput but results in a $0.4\%$ reduction. + +LPU and Overlapped Downsampler. Our experiments in Tab. 5b demonstrate that incorporating overlapped downsampling layers and the local perception unit (LPU) introduces beneficial inductive bias and enhances local perception capabilities. These modifications improve the top-1 accuracy by $0.6\%$ while marginally affecting inference throughput. To maintain comparable computational costs despite the additional parameters and FLOPs introduced by these operations, we strategically reduce the model depth. + +Hybrid Architecture with Attention and MESA. Tab. 5b validates the effectiveness of our hybrid approach, which combines standard attention in the final stage with overlapped downsampling. Specifically, substituting NC-SSD with standard attention in the last stage yields a $0.2\%$ accuracy improvement while slightly increasing the parameter count. Additionally, following MLLA [19], we incorporate MESA [11] to mitigate overfitting, resulting in a further $0.2\%$ gain in top-1 accuracy. + +# 5. Conclusion + +In conclusion, our study introduces the NC-SSD, which redefines SSD by modifying the role of matrix $\mathbf{A}$ and eliminating the causal mask. These adaptations facilitate a transition to a non-causal mode, significantly enhancing both accuracy and efficiency. Extensive experiments demonstrate its superiority over the vanilla SSD and its multi-scan-based variants. Furthermore, by integrating techniques such as hybrid standard attention and overlapped downsampling, our VSSD model achieves comparable or superior performance compared to well-established CNNs, ViTs, and Vision SSMs across several widely-used benchmarks. + +# 6. Acknowledgements + +This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549. + +# References + +[1] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6 +[2] Tianxiang Chen, Zhentao Tan, Tao Gong, Qi Chu, Yue Wu, Bin Liu, Jieping Ye, and Nenghai Yu. Mim-istd: Mamba-in-mamba for efficient infrared small target detection. arXiv preprint arXiv:2403.02148, 2024. 2 +[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1 +[4] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 2 +[5] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7 +[6] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. In ICML, 2024. 1, 3, 4 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 5 +[8] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 6 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4 +[10] Chengbin Du, Yanxi Li, and Chang Xu. Understanding robustness of visual state space models for image classification. arXiv preprint arXiv:2403.10935, 2024. 2 +[11] Jiawei Du, Daquan Zhou, Jiashi Feng, Vincent Tan, and Joel Tianyi Zhou. Sharpness-aware training for free. NeurIPS, 2022. 8 +[12] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 2, 5 +[13] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3 +[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3 + +[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 2020. 1 +[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3 +[17] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. NeurIPS, 2021. 1, 2 +[18] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 1, 4, 6 +[19] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 2, 5, 8 +[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, 2021. 2 +[21] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 1, 2, 5, 6 +[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 2, 4, 5, 7 +[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 6, 7 +[24] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2 +[25] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 1 +[26] Gao Huang, Zhuang Liu, Geoff Pleiss, Laurens Van Der Maaten, and Kilian Weinberger. Convolutional networks with dense connectivity. IEEE TPAMI, 2019. 1, 2 +[27] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 1, 2, 3, 4, 6 +[28] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 2 +[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2 +[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, 2024. 2 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7 + +[32] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6 +[33] Jing Liu, Bohan Zhuang, Zhuangwei Zhuang, Yong Guo, Junzhou Huang, Jinhui Zhu, and Mingkui Tan. Discrimination-aware network pruning for deep model compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2 +[34] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 5, 6, 7, 8 +[36] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 1, 2 +[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 2, 5, 6, 7 +[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, 2016. 6 +[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3 +[40] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. arXiv preprint arXiv:2202.08791, 2022. 2 +[41] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. Sg-former: Self-guided transformer with evolving token reallocation. In ICCV, 2023. 5 +[42] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2 +[43] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. In NeurIPS, 2024. 1, 3, 6, 7 +[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2 +[45] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 1, 2 +[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 4 +[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019, 1, 2 +[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 4, 5, 6, 7 + +[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2 +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2 +[51] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 1, 2 +[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 2, 6 +[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022. 2, 6 +[54] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 5 +[55] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023. 2 +[56] Chaodong Xiao, Minghan Li, Zhengqiang Zhang, Deyu Meng, and Lei Zhang. Spatial-mamba: Effective visual state space models via structure-aware state fusion. In ICLR, 2025. 6 +[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 7 +[58] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 2 +[59] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 6 +[60] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks. NeurIPS, 2022. 6, 7 +[61] Yuhuan Yang, Chaofan Ma, Jiangchao Yao, Zhun Zhong, Ya Zhang, and Yanfeng Wang. Remember: Referring image segmentation with mamba twister. arXiv preprint arXiv:2403.17839, 2024. 2 +[62] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 6, 7 +[63] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 2 +[64] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1 + +[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 1, 7 +[66] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2 +[67] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 1, 2, 3, 4 \ No newline at end of file diff --git a/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/images.zip b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..057e2cfce8fb0dd71430efcdfc06c2b050e9a001 --- /dev/null +++ b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a32abd021c2461c8b685983a08ac9b5c57b0c5f2f0be3a4f454381e99e04561 +size 777372 diff --git a/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/layout.json b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2d49441fe6ab3bf1fd431e711a702615ad642b98 --- /dev/null +++ b/2025/VSSD_ Vision Mamba with Non-Causal State Space Duality/layout.json @@ -0,0 +1,8744 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "text", + "content": "VSSD: Vision Mamba with Non-Causal State Space Duality" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "spans": [ + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "text", + "content": "Yuheng Shi" + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "text", + "content": " Mingjia Li" + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "text", + "content": " Minjing Dong" + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "text", + "content": " Chang Xu" + }, + { + "bbox": [ + 160, + 142, + 449, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 157, + 490, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 157, + 490, + 172 + ], + "spans": [ + { + "bbox": [ + 119, + 157, + 490, + 172 + ], + "type": "text", + "content": "1University of Sydney, 2City University of Hong Kong, 3Tianjin University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 174, + 537, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 174, + 537, + 186 + ], + "spans": [ + { + "bbox": [ + 70, + 174, + 537, + 186 + ], + "type": "text", + "content": "yuhengshi99@gmail.com, mingjiali@tju.edu.cn, minjdong@cityu.edu.hk, c.xu@sydney.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 539 + ], + "type": "text", + "content": "Vision transformers have significantly advanced the field of computer vision, offering robust modeling capabilities and global receptive field. However, their high computational demands limit their applicability in processing long sequences. To tackle this issue, State Space Models (SSMs) have gained prominence in vision tasks as they offer linear computational complexity. Recently, State Space Duality (SSD), an improved variant of SSMs, was introduced in Mamba2 to enhance model performance and efficiency. However, the inherent causal nature of SSD/SSMs restricts their applications in non-causal vision tasks. To address this limitation, we introduce Visual State Space Duality (VSSD) model, which has a non-causal format of SSD. Specifically, we propose to discard the magnitude of interactions between the hidden state and tokens while preserving their relative weights, which relieves the dependencies of token contribution on previous tokens. Together with the involvement of multi-scan strategies, we show that the scanning results can be integrated to achieve non-causality, which not only improves the performance of SSD in vision tasks but also enhances its efficiency. We conduct extensive experiments on various benchmarks including image classification, detection, and segmentation, where VSSD surpasses existing SOTA SSM-based models. Code is available at https://github.com/YuHengsss/VSSD." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "In recent years, vision transformers [8, 9, 18, 21, 35, 36, 48, 51], pioneered by the Vision Transformer (ViT) [9], have achieved tremendous success in the field of computer vision. Thanks to the global receptive field and the robust information modeling capabilities of the attention mechanism, models [3, 35, 64] based on vision transformers have advanced significantly in various tasks such as classification [7], detection [31], and segmentation [65], surpassing classic CNN-based models [22, 25, 26, 44]. However, the quadratic computational complexity of the attention mechanism makes it resource-intensive for tasks involving long" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 214, + 506, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 506, + 227 + ], + "type": "text", + "content": "sequences, which limits its broader application." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 230, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 471 + ], + "type": "text", + "content": "Recently, State Space Models (SSMs) [15-17, 45], exemplified by Mamba [14], have garnered considerable attention from researchers. The S6 block, in particular, offers a global receptive field and exhibits linear complexity with respect to sequence length, presenting an efficient alternative. Pioneering vision mamba models such as Vim [67] and VMamba [34] have been developed to apply SSMs to vision tasks. Afterward, many variants were proposed [27, 43, 47, 59], which flatten 2D feature maps into 1D sequences using different scanning routes, model them with the S6 block, and subsequently integrate the results in multiple scanning routes. These multi-scan approaches improve the performance of SSMs in vision tasks, achieving results competitive with those of both CNN-based and ViT-based methods. More recently, Mamba2 [6] has introduced further enhancements to the S6 block, proposing the concept of State Space Duality (SSD). Mamba2 treats the state space transition matrix as a scalar and expands the state space dimension, thereby enhancing model performance as well as training and inference efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "SSD/SSMs inherently possess causal properties, meaning that the prediction for a given token depends solely on the preceding tokens in the input sequence. However, a major concern arises when applying SSD/SSMs to vision tasks, as image data is naturally non-causal. While another concern is flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches. We provide an illustration in Fig. 1a to facilitate a more intuitive understanding of these two concerns. In this example, the central token within the flattened 1D sequences is restricted to accessing only previous tokens, unable to integrate information from subsequent tokens. Additionally, the token 1, which is adjacent to the central token in the 2D space, becomes distantly positioned in the 1D sequence, disrupting the natural structural relationships. A common practice in previous solutions [27, 34] is to increase the scanning routes on non-causal visual features, which alleviates these two concerns to some extent. Given these observations, an important question emerges: Is there a more effective and efficient way to apply SSD to non-causal vis" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10819" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 76, + 126, + 183 + ], + "blocks": [ + { + "bbox": [ + 80, + 76, + 126, + 183 + ], + "lines": [ + { + "bbox": [ + 80, + 76, + 126, + 183 + ], + "spans": [ + { + "bbox": [ + 80, + 76, + 126, + 183 + ], + "type": "image", + "image_path": "da179a25f536b99305114d0eda81ff5b781a5314bdafda557d4a7dc11129d537.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 190, + 156, + 201 + ], + "lines": [ + { + "bbox": [ + 143, + 190, + 156, + 201 + ], + "spans": [ + { + "bbox": [ + 143, + 190, + 156, + 201 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 76, + 172, + 183 + ], + "blocks": [ + { + "bbox": [ + 126, + 76, + 172, + 183 + ], + "lines": [ + { + "bbox": [ + 126, + 76, + 172, + 183 + ], + "spans": [ + { + "bbox": [ + 126, + 76, + 172, + 183 + ], + "type": "image", + "image_path": "d3c80d539e3098072c9de637070f545965113b1777f750b998275f7a51a55ca1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 203, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 555, + 236 + ], + "type": "text", + "content": "Figure 1. (a) Two challenges when applying SSM/SSD to image data. (b) and (c) are comparisons on ImageNet. Our VSSD model achieves leading accuracy and efficiency compared to CNN-based ConvNeXt [37], ViT-based Swin Transformer [35], and SSM-based VMamba [34]. The throughput of all models is measured on an NVIDIA 4090 GPU using a batch size of 128." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 173, + 76, + 219, + 183 + ], + "blocks": [ + { + "bbox": [ + 173, + 76, + 219, + 183 + ], + "lines": [ + { + "bbox": [ + 173, + 76, + 219, + 183 + ], + "spans": [ + { + "bbox": [ + 173, + 76, + 219, + 183 + ], + "type": "image", + "image_path": "903813502cdb547739bfdd74f662e02719ff12d3d263a881f1789a07e59efce7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 75, + 375, + 187 + ], + "blocks": [ + { + "bbox": [ + 228, + 75, + 375, + 187 + ], + "lines": [ + { + "bbox": [ + 228, + 75, + 375, + 187 + ], + "spans": [ + { + "bbox": [ + 228, + 75, + 375, + 187 + ], + "type": "image", + "image_path": "55a4ad91381d970b8dfffb389fd66f39a605cac187d1e668e72975cb70e84635.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 190, + 320, + 200 + ], + "lines": [ + { + "bbox": [ + 306, + 190, + 320, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 320, + 200 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 377, + 75, + 525, + 188 + ], + "blocks": [ + { + "bbox": [ + 377, + 75, + 525, + 188 + ], + "lines": [ + { + "bbox": [ + 377, + 75, + 525, + 188 + ], + "spans": [ + { + "bbox": [ + 377, + 75, + 525, + 188 + ], + "type": "image", + "image_path": "4389ddc36e9956debf04295e2053f5115c9e98abf79c11769910e19cf97a65d5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 190, + 465, + 201 + ], + "lines": [ + { + "bbox": [ + 453, + 190, + 465, + 201 + ], + "spans": [ + { + "bbox": [ + 453, + 190, + 465, + 201 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 240, + 232, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 232, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 232, + 252 + ], + "type": "text", + "content": "sion data compared to multi-scan methods?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "text", + "content": "To address this question, our analysis of SSD reveals that treating the matrix " + }, + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "text", + "content": " as a scalar facilitates a straightforward transformation of the SSD into a non-causal and position-independent manner, which we denote as Non-Causal SSD (NC-SSD). Specifically, rather than using " + }, + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "text", + "content": " to determine the proportion of the hidden state to be retained, we employ it to dictate the extent of the current token's contribution to the hidden states. In this scenario, each token's contribution becomes self-referential. Based on this property, we demonstrate that the causal mask in SSD can be naturally removed without the need for specific scanning routes. This observation motivates us to develop a non-causal format of SSD, in which a single global hidden state can be derived to replace the previous token-wise hidden states, resulting in not only improved accuracy but also enhanced training and inference speeds. Unlike previous multi-scan methods [34, 67] that primarily alleviate the causal limitations of SSMs, our proposed NC-SSD also resolves the issue where flattening 2D feature maps into 1D sequences disrupts the continuity of adjacent tokens. Building on NC-SSD, we introduce our Visual State Space Duality (VSSD) model and demonstrate its superior effectiveness and efficiency relative to methods based on CNNs, ViTs, and SSMs, as illustrated in Fig. 1b and 1c. Compared to the recent proposed SSM-based VMamba [34], our VSSD model outperforms it by over " + }, + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 56, + 255, + 296, + 579 + ], + "type": "text", + "content": " in top-1 accuracy on ImageNet-1K dataset [7] while keeping a similar computational cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "In summary, our contributions are twofold. First, we analyze the state space duality and demonstrate that it can be seamlessly converted to a non-causal mode. Based on this insight, we introduce the NC-SSD, which retains the global receptive field and linear complexity benefits of the original SSD while incorporating an inherent non-causal property. Second, utilizing NC-SSD as the foundational component, we propose the VSSD model and conduct extensive experiments to validate its effectiveness. With similar parameters and computational costs, our VSSD model outperforms other State-Of-The-Art (SOTA) SSM-based models across" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 240, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 264 + ], + "type": "text", + "content": "several widely recognized benchmarks in classification, object detection, and segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 400, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "text", + "content": "Vision Transformers. The introduction of Vision Transformers (ViTs) [8, 9, 35, 48, 51] has revitalized the field of computer vision, which was previously dominated by Convolutional Neural Networks (CNNs) [22, 24, 26, 29, 33, 37, 44, 47, 58]. However, the quadratic computational complexity of the self-attention mechanism in ViTs poses significant challenges when processing high-resolution images, demanding significant computational resources. To address this issue, different solutions have been proposed, including hierarchical architectures [8, 20, 35, 36, 51, 53], windowed attention [21, 35, 49, 66], and variants of self-attention [52, 55, 63]. Meanwhile, linear attentions [4, 19, 28, 40] have been successful in reducing the computational complexity to a linear scale by changing the computation order of query, key, and value in self-attention. Despite this advancement, the performance of linear attention remains inferior to that of the quadratic self-attention [50] and its variants [12, 21, 66]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "State Space Models. State Space Models (SSMs) [13-17, 45] have increasingly captured the attention of researchers due to their global receptive field and linear computational complexity. Mamba [14], a prominent example of SSMs, introduced the S6 block, achieving performance on par with or better than transformers in Nature Language Processing (NLP) benchmarks. Subsequent efforts [2, 10, 27, 30, 42, 61] have explored the adaptation of the S6 block to vision tasks, yielding competitive results compared to both CNNs and ViT-based models. A central challenge in developing Mamba-based vision models is adapting the inherently causal properties of the Mamba block for non-causal image data. The most direct approach involves using different scanning routes to flatten 2D feature maps into 1D sequences which are then modeled with the S6 block and integrated. Inspired by these considerations, various scanning routes have been employed and proven effective, as" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10820" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 178 + ], + "type": "text", + "content": "evidenced by multiple studies [27, 34, 39, 43, 67]. More recently, Mamba2 [6] has highlighted the significant overlap between state space models and structured masked attention, identifying them as duals of each other, and introduced the concept of State Space Duality (SSD). Building on this foundation, we demonstrate that the SSD can be transformed into a non-causal mode through a straightforward transformation, without the need of specific scanning route." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 191, + 111, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 111, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 111, + 202 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 141, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 141, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 141, + 222 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": "State Space Models. Classic State Space Models (SSMs) are used to describe the dynamics of a continuous system, transforming an input sequence " + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": " to a latent space representation " + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": ". This representation is then utilized to generate an output sequence " + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "inline_equation", + "content": "y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": ". The mathematical formulation of an SSM is structured as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 308, + 295, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 308, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 93, + 308, + 295, + 326 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = \\overset {\\circ} {\\mathbf {A}} h (t) + \\overset {\\circ} {\\mathbf {B}} x (t), y (t) = \\mathbf {C} h (t), \\tag {1}", + "image_path": "34a85c28d19e7215ffd1ea04f81388b4d4009d83fc59bb637bf9281067ef7a18.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathring{\\mathbf{A}}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathring{\\mathbf{B}}\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{C}\\in \\mathbb{R}^{1\\times N}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": " are parameters. To effectively integrate continuous SSMs into deep learning architectures, discretization is essential. This process involves introducing a timescale parameter " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": " and applying the zero-order hold (ZOH) technique for discretization. Through this approach, the continuous matrices " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathring{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathring{\\mathbf{B}}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": " are transformed into their discrete counterparts, " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 336, + 296, + 462 + ], + "type": "text", + "content": ". Consequently, Eq. 1 is redefined in a discrete format in Eq. 2, facilitating its application within modern computational frameworks:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 60, + 472, + 239, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 472, + 239, + 484 + ], + "spans": [ + { + "bbox": [ + 60, + 472, + 239, + 484 + ], + "type": "interline_equation", + "content": "h (t) = \\mathbf {A} h (t - 1) + \\mathbf {B} x (t), y (t) = \\mathbf {C} h (t),", + "image_path": "e8d7dd5cb73b975b83c9f037616aa049e688d560da01e7ac649413b2abbb1184.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 485, + 295, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 485, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 60, + 485, + 295, + 512 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\mathbf {A} = e ^ {\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}}, \\mathbf {B} = (\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}) ^ {- 1} (e ^ {\\boldsymbol {\\Delta} \\mathring {\\mathbf {A}}} - \\mathbf {I}) \\boldsymbol {\\Delta} \\mathring {\\mathbf {B}} \\approx \\boldsymbol {\\Delta} \\mathring {\\mathbf {B}}, \\tag {2}", + "image_path": "d3306e9170593d0cb2be05cfbbf456a4fc3d79a3c7b931161c582a73eba49336.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 512, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 512, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 512, + 296, + 548 + ], + "type": "text", + "content": " denotes the identity matrix. Furthermore, the process of Eq. 2 could be implemented in a global convolution manner as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 557, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 557, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 68, + 557, + 295, + 572 + ], + "type": "interline_equation", + "content": "y = x \\odot \\mathbf {K}, \\mathbf {K} = \\left(\\mathbf {C B}, \\mathbf {C A B}, \\dots , \\mathbf {C A} ^ {L - 1} \\mathbf {B}\\right), \\tag {3}", + "image_path": "8ef02c9b3514adcd07adf824a2fab243e73f8745f646c18dff9cb04c83087c48.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "text", + "content": " represents the convolution kernel. Recently, Mamba [14] makes the parameters " + }, + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{B}, \\mathbf{C}" + }, + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}" + }, + { + "bbox": [ + 55, + 580, + 296, + 652 + ], + "type": "text", + "content": " input-dependent. This modification addresses the limitations of the Linear Time Invariant (LTI) characteristics inherent in previous SSM models [13, 16], thus enhancing the adaptability and performance of SSMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 659, + 228, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 228, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 228, + 673 + ], + "type": "text", + "content": "3.2. Non-Causal State Space Duality" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Mamba2 [6] introduced the State Space Duality (SSD) and simplified the matrix " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " into a scalar. This special case of selective SSMs can be implemented in both linear and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": "quadratic forms. Without loss of generality, the matrix transformation form of selective state space models is expressed as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 114, + 558, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 114, + 558, + 152 + ], + "spans": [ + { + "bbox": [ + 318, + 114, + 558, + 152 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} y (t) = \\sum_ {i = 1} ^ {t} \\mathbf {C} _ {t} ^ {T} \\mathbf {A} _ {t: i + 1} \\mathbf {B} _ {i} x (i), \\text {w h e r e} \\mathbf {A} _ {t: i} = \\prod_ {i} ^ {t} \\mathbf {A} _ {i}, \\\\ y = \\operatorname {S S M} (\\mathbf {A}, \\mathbf {B}, \\mathbf {C}) (x) = \\mathbf {F} x, \\text {w h e r e} \\mathbf {F} _ {j i} = \\mathbf {C} _ {j} ^ {T} \\mathbf {A} _ {j: i + 1} \\mathbf {B} _ {i}. \\end{array} \\tag {4}", + "image_path": "91b693e76eb60bb4097cec3a45b52cf6da1d659724a0ff6b29dc7cc33c95b150.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 152, + 555, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 176 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 152, + 555, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 152, + 555, + 176 + ], + "type": "text", + "content": " is reduced to a scalar, the quadratic form of Eq. 4 can be reformulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 335, + 181, + 441, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 181, + 441, + 194 + ], + "spans": [ + { + "bbox": [ + 335, + 181, + 441, + 194 + ], + "type": "interline_equation", + "content": "y = \\mathbf {F} x = \\mathbf {M} \\cdot (\\mathbf {C} ^ {T} \\mathbf {B}) x,", + "image_path": "61b189e146863e4d176a5637c4e85440656065732b3a15102f15a06b4a8bab0a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 194, + 555, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 194, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 336, + 194, + 555, + 232 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\mathbf {M} _ {j i} = \\left\\{ \\begin{array}{l l} A _ {j} \\times \\dots \\times A _ {i + 1} & i < j \\\\ 1 & i = j \\\\ 0 & i > j, \\end{array} \\right. \\tag {5}", + "image_path": "5e1c9f77a67b50a749b9aa1301467eee2805578ab678ac07a8c1ead430252106.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 238, + 456, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 456, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 456, + 248 + ], + "type": "text", + "content": "while its linear form is denoted as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 335, + 257, + 555, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 257, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 335, + 257, + 555, + 271 + ], + "type": "interline_equation", + "content": "h (t) = A _ {t} h (t - 1) + \\mathbf {B} _ {t} x (t), y (t) = \\mathbf {C} _ {t} h (t). \\tag {6}", + "image_path": "f79fdba1a1b6ae3ea3b74bc702e69df509b3ef47ab0aa0b297f8869e3b9b912e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "type": "text", + "content": "To adapt SSMs for image data, the 2D feature maps should first be flattened into a 1D sequence of tokens which are then sequentially processed. Due to the causal nature of SSMs, where each token can only access previous tokens, information propagation is inherently unidirectional. This causal property leads to suboptimal performance when handling non-causal image data, a finding that has been corroborated by previous works [34, 67]. Furthermore, flattening the 2D feature maps into 1D sequences disrupts their intrinsic structural information. For instance, tokens that are adjacent in the 2D map might end up being far apart in the 1D sequence, resulting in a loss of performance in vision tasks. Since SSD is a variant of SSMs, adopting SSD for vision tasks presents similar challenges to those observed with SSMs:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 314, + 457, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 457, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 314, + 457, + 554, + 492 + ], + "type": "text", + "content": "- Challenge 1: The causal property of the model restricts the flow of information, preventing later tokens from influencing earlier ones." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 554, + 529 + ], + "type": "text", + "content": "- Challenge 2: Flattening 2D feature maps into 1D sequences disrupts the inherent structural relationships among patches during processing." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "content": "In the context of applying causal SSD to non-causal image data, it is instructive to revisit the linear formulation of SSD. In Eq. 6, the scalar " + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "content": " modulates the influence of the previous hidden state " + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "inline_equation", + "content": "h(t - 1)" + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "content": " and the information in current time step. In other words, the current hidden state " + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "inline_equation", + "content": "h(t)" + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "content": " could be viewed as a linear combination of the previous hidden state and the current input, weighted by " + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "content": " and 1, respectively. Therefore, if we discard the magnitude of these two terms and only preserve their relative weight, Eq. 6 could be rewritten as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 330, + 654, + 555, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 654, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 330, + 654, + 555, + 686 + ], + "type": "interline_equation", + "content": "h (t) = h (t - 1) + \\frac {1}{A _ {t}} \\mathbf {B} _ {t} x (t) = \\sum_ {i = 1} ^ {t} \\frac {1}{A _ {i}} \\mathbf {B} _ {i} x (i). \\qquad (7)", + "image_path": "2345a2867fc6f708bc9b486fbb10de33e5a570e8755dc02ee49d32d8fb455841.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "In this scenario, the contribution of a specific token to the current hidden state can be directly determined by itself as" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10821" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\frac{1}{A_i}" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": ", rather than through the cumulative multiplication of multiple coefficients. With each token's contribution becoming self-referential, Challenge 2 is only partially addressed, as the current token can access only a subset of tokens due to the issue discussed in Challenge 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "text", + "content": "To address Challenge 1, previous SSM-based vision models frequently employed multi-scanning routes. Specifically, in the case of ViM [67], the token sequence was subjected to both forward and reverse scanning, enabling each token to access global information. Although multi-scan approaches mitigate the causal property of SSMs, they do not address Challenge 2, as the long-range decay characteristic of SSMs remains confined to a 1D format and does not extend into 2D. To enable the acquisition of global information and thus suit non-causal image data, we also begin with a bidirectional scanning strategy. And we demonstrate that the results of forward and reverse scanning of Eq. 7 can be integrated to effectively address the aforementioned two challenges simultaneously. Let " + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_i = h(i)" + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "text", + "content": " denote the hidden state of the " + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 55, + 133, + 295, + 324 + ], + "type": "text", + "content": " token in the bidirectional scanning approach, from which we can easily derive:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 335, + 294, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 335, + 294, + 377 + ], + "spans": [ + { + "bbox": [ + 58, + 335, + 294, + 377 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {i} = \\sum_ {j = 1} ^ {i} \\frac {1}{A _ {j}} \\mathbf {Z} _ {j} + \\sum_ {j = - L} ^ {- i} \\frac {1}{A _ {- j}} \\mathbf {Z} _ {- j} = \\sum_ {j = 1} ^ {L} \\frac {1}{A _ {j}} \\mathbf {Z} _ {j} + \\frac {1}{A _ {i}} \\mathbf {Z} _ {i} \\tag {8}", + "image_path": "055f37363d491175455a3352ff9778266dbf637dbd47d7652c748fbba8c5ab1f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_j = \\mathbf{B}_j x(j)" + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\frac{1}{A_i} \\mathbf{Z}_i" + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "text", + "content": " primarily affects the current token and provides redundant functionality with respect to the skip connection, we omit the bias term to streamline the model. In such case, Eq. 8 can be further simplified, resulting in all tokens sharing the same hidden state " + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = \\sum_{j=1}^{L} \\frac{1}{A_j} \\mathbf{Z}_j" + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "text", + "content": ". In such cases, the results from forward and reverse scanning can be seamlessly combined to establish a global context, effectively equivalent to removing the causal mask and transitioning to a non-causal format. Consequently, the first challenge associated with the causal property is resolved. Although the above results are derived from a bidirectional scanning approach, it is evident that in this non-causal format, different scanning routes yield consistent outcomes. Furthermore, as demonstrated in Eq. 8, the contribution of different tokens to the current hidden state is no longer related to their scanning distance. Therefore, processing a flattened 2D feature map into a 1D sequence no longer compromises the original structural relationships. To preserve the spatial relationships between patches, we simply incorporate 2D Rotary Position Embedding (RoPE) [46] into features " + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 377, + 295, + 676 + ], + "type": "text", + "content": ". This approach effectively addresses the second challenge, as the spatial relationships among patches are now encoded through positional embeddings rather than being determined by their sequential distance in the 1D representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "After revising the iteration rules for the hidden state space, we update the corresponding tensor contraction algorithm or einsum notation in the linear form, following the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 416, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 416, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 416, + 83 + ], + "type": "text", + "content": "Mamba2 framework [6]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 351, + 92, + 518, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 92, + 518, + 106 + ], + "spans": [ + { + "bbox": [ + 351, + 92, + 518, + 106 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\operatorname {c o n t r a c t} (\\mathrm {L D}, \\mathrm {L N} \\rightarrow \\mathrm {L N D}) (\\mathbf {X}, \\mathbf {B})", + "image_path": "08fc0028489d1cd0e8abb54ee58f07bdd5f4c984f01c8f73139a8aff3989e9f3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 350, + 108, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 108, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 350, + 108, + 553, + 121 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\operatorname {c o n t r a c t} (\\mathrm {L L}, \\mathrm {L D N} \\rightarrow \\mathrm {N D}) (\\mathbf {M}, \\mathbf {Z}) \\tag {9}", + "image_path": "c1c1f95be21b3d06b2886ab19b614485ef0b020cd15637ad3dcd690063218010.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 123, + 514, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 123, + 514, + 136 + ], + "spans": [ + { + "bbox": [ + 351, + 123, + 514, + 136 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\operatorname {c o n t r a c t} (\\mathrm {L N}, \\mathrm {N D} \\rightarrow \\mathrm {L D}) (\\mathbf {C}, \\mathbf {H}).", + "image_path": "f67f91bf128c7a812473d1fd07cd1e8406098c6d06a68428b6dfb33f36a6bd50.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": "This algorithm involves three steps: the first step expands the input " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": ", the second step unrolls scalar SSM recurrences to create a global hidden state " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": ", and the final step contracts the hidden state " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": ". For clarity, the initial two steps of SSD and NC-SSD are depicted in Fig. 2a. Compared to vanilla SSD, while the operation in the first step remains unchanged, the sequence length dimension in the hidden state " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " is eliminated in the non-causal mode, as all tokens share the same hidden state. In the final step, the output " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " is produced through the matrix multiplication of " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{i,j} = \\frac{1}{A_j}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": ", the matrix " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " could be simplified to vector of weights, where " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_j = \\frac{1}{A_j}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": ". In this case, integrating " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " with either " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 140, + 555, + 312 + ], + "type": "text", + "content": " could further simplify the transformation of Eq. 9 to:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 388, + 319, + 553, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 319, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 388, + 319, + 553, + 333 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\mathbf {C} \\left(\\left(\\mathbf {B} \\cdot \\mathbf {m}\\right) ^ {T} \\mathbf {X}\\right), \\tag {10}", + "image_path": "d9ae76815d4a664a419aaa536eff155940a2046695dda94adcf211fd8bda926d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": "which can be regarded as a variant of linear attention. However, it is worth noting that just as " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " plays a distinguished role in Mamba2, the vector " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " is also crucial, as demonstrated in our ablation studies. In practice, we directly use the learned " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\frac{1}{A}" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " since they share the same range of values. To gain a more intuitive understanding of the role of " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " in Eq. 10, we visualize the average of " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " across different heads as shown in Fig. 2b. Predominantly, " + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 342, + 555, + 461 + ], + "type": "text", + "content": " focuses on foreground features, enabling the model to prioritize elements that are crucial for the task at hand." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 493, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 493, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 493, + 481 + ], + "type": "text", + "content": "3.3. Visual State Space Duality Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 688 + ], + "type": "text", + "content": "Block Design. To enhance the SSD block in Mamba2 for vision applications, several modifications have been implemented beyond merely substituting the SSD with an NC-SSD to develop our Visual State Space Duality (VSSD) block. When constructing the NC-SSD block, the causal convolution 1D is replaced by a Depth-Wise Convolution (DWConv) with a kernel size of three, in line with previous vision mamba works [27, 34]. Additionally, a Feed-Forward Network (FFN) is integrated subsequent to the NC-SSD block to facilitate enhanced information exchange across channels and to maintain alignment with the established practices of classical vision transformers [9, 35, 48]. Moreover, a Local Perception Unit (LPU) [18] is incorporated prior to the NC-SSD block and the FFN, augmenting the model's capability for local feature perception. Skip connections [22] are also implemented among different blocks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Hybrid with Self-Attention. Mamba2 demonstrates that integrating SSD with standard Multi-head Self Attention" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10822" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 203, + 202 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 203, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 203, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 203, + 202 + ], + "type": "image", + "image_path": "da900e29701389a36f79fa19d60be70de7f54242f8ae752ca114c20eb27c47ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 206, + 212, + 214 + ], + "lines": [ + { + "bbox": [ + 201, + 206, + 212, + 214 + ], + "spans": [ + { + "bbox": [ + 201, + 206, + 212, + 214 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 207, + 71, + 350, + 202 + ], + "blocks": [ + { + "bbox": [ + 207, + 71, + 350, + 202 + ], + "lines": [ + { + "bbox": [ + 207, + 71, + 350, + 202 + ], + "spans": [ + { + "bbox": [ + 207, + 71, + 350, + 202 + ], + "type": "image", + "image_path": "0c4f561dbd8191dfb2982de25762ea3c3b767a41fb88d5fa330d70939b453e29.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "type": "text", + "content": "Figure 2. (a) Illustration of the Hidden State Generation Process for SSD and NC-SSD. During the hidden state update process, NC-SSD utilizes the scalar " + }, + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "type": "text", + "content": " to determine the extent of information increment for the current token, in contrast to SSD where " + }, + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "type": "text", + "content": " dictates the proportion of the hidden state to be retained. Unlike the SSD, which generates token-wise hidden states, the NC-SSD produces only a global hidden state to accommodate non-causal image data. (b) Visualization of input images alongside their corresponding heat maps, derived by averaging the vector " + }, + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 54, + 216, + 555, + 271 + ], + "type": "text", + "content": " across various heads in the NC-SSD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 360, + 106, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 360, + 106, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 360, + 106, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 360, + 106, + 553, + 202 + ], + "type": "image", + "image_path": "fb87584c74a02ee162b29810ca7c29b65fc75b5827f72e06bfdc27932e265fc5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 205, + 462, + 214 + ], + "lines": [ + { + "bbox": [ + 451, + 205, + 462, + 214 + ], + "spans": [ + { + "bbox": [ + 451, + 205, + 462, + 214 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 275, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 295, + 371 + ], + "type": "text", + "content": "(MSA) yields additional improvements. In a similar vein, our model incorporates self-attention. However, unlike Mamba2, which uniformly intersperses self-attention throughout the network, we strategically replace the NCSSD block with self-attention module exclusively in the last stage. This modification leverages the robust capabilities of self-attention in processing high-level features, as evidenced by prior works [12, 32, 41] in vision tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 372, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 295, + 491 + ], + "type": "text", + "content": "Overlapped downsampling layers. As hierarchical vision transformers [35] and vision state space models [34] predominantly employ non-overlapped convolution for downsampling, recent studies [21, 54] have demonstrated that overlapped downsampling convolutions can introduce beneficial inductive biases. Consequently, we adopt overlapped convolutions in our model, following the manner used in previous works [12, 19]. To maintain the parameter count and computational FLOPs remain comparable, we have accordingly adjusted the depth of our model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 492, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 492, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 492, + 295, + 589 + ], + "type": "text", + "content": "Overall Architecture. We develop our VSSD model in accordance with the methods discussed above. Mirroring the design principles of established vision backbones in previous works [34, 35, 37], our VSSD model is structured into four hierarchical stages. The first three stages employ the VSSD block, while the final stage incorporates the MSA block. Detailed architectures and illustration of the proposed model could be found in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 602, + 132, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 132, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 132, + 616 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 623, + 141, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 141, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 141, + 635 + ], + "type": "text", + "content": "4.1. Classification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Configurations. Our experiments are conducted using the ImageNet-1K dataset [7], consistent with methodologies from prior studies [34, 35]. Each model is subjected to a training regimen spanning 300 epochs, which includes a 20-epoch warm-up phase. Optimization is performed with AdamW, where the betas are set to (0.9, 0.999) and momen" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 275, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 358 + ], + "type": "text", + "content": "tum at 0.9. A cosine decay scheduler manages the learning rate and is combined with a weight decay rate of 0.05. To further refine model accuracy and generalization, we incorporate exponential moving average (EMA) techniques and apply label smoothing with a coefficient of 0.1. More detailed configurations could be found in our supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "content": "Performance Evaluation. Tab. 1 presents a comparative comparison of our VSSD models against CNNs, ViTs, and SSM-based frameworks on the ImageNet-1K dataset [7]. The VSSD-T model, with 28M parameters and 5.0G FLOPs, secures a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "inline_equation", + "content": "83.8\\%" + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "content": ", surpassing the similarly priced VMamba-T [34] by " + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "content": ". In comparison with models categorized as small and base, VSSD consistently outperforms its counterparts. Specifically, the VSSD-S model, which has 50M parameters and 8.1G FLOPs, achieves an accuracy of " + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "inline_equation", + "content": "84.6\\%" + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "content": ", outdoing VMamba-S [34] by " + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "content": ". For the base-sized model variant, VSSD-B, which comprises 89M parameters and 16.1G FLOPs, achieves an accuracy of " + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "inline_equation", + "content": "85.4\\%" + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "content": ", surpassing the VMamba-B [34] by " + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 362, + 556, + 615 + ], + "type": "text", + "content": ". Furthermore, we provide a comprehensive efficiency comparison between our VSSD and widely used architectures including Swin [35], ConvNeXt [37], and VMamba [34] in Tab. 2. Notably, compared to previous works, our VSSD achieves superior accuracy-throughput trade-off while substantially reducing memory consumption. We provide comparison with SOTA works in the appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Effective Receptive Field Comparison. In addition to quantitative comparisons, we conducted a comparative analysis of the Effective Receptive Field (ERF) before and after training across various models, including CNN-based ResNet50 [22] and ConvNeXt-Tiny [37], attention-based Swin-Tiny [35] and DeiT-Small [48], and SSM-based VMamba-Tiny [34], along with our VSSD-Tiny. The ERF of the central pixel is plotted using the method proposed" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10823" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 294, + 352 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 294, + 352 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 294, + 352 + ], + "type": "table", + "html": "
MethodType#Param.FLOPsTop-1 Acc(%)
Tiny Models
ConvNeXt-T [37]Conv29M4.5G82.1
MambaOut-T [62]Conv27M4.5G82.7
InternImage-T [52]Conv30M5.0G83.5
DeiT-S [48]Attn22M4.6G79.8
Swin-T [35]Attn29M4.5G81.3
PVTv2-B2 [53]Attn25M4.0G82.0
Focal-T [60]Attn29M4.9G82.2
CSwin-T [8]Attn23M4.3G82.7
NAT-T [21]Attn28M4.3G83.2
CMT-S [18]Attn25M4.0G83.5
VMamba-T [34]SSM30M4.9G82.6
LVMamba-T [27]SSM26M5.7G82.7
MSVMamba-T [43]SSM33M4.9G83.0
SpatialMamba-T [56]SSM27M4.5G83.5
VSSD-TSSD28M5.0G83.8
Small Models
ConvNeXt-S [37]Conv50M8.7G83.1
MambaOut-S [62]Conv48M9.0G84.1
Swin-S [35]Attn50M8.7G83.0
PVTv2-B3 [53]Attn45M6.9G83.2
Focal-S [60]Attn50M8.7G83.5
", + "image_path": "138d282ff74901ef120470c1fdb0a47550f14c85b362642bd91947378e9dc1a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 550, + 352 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 550, + 352 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 550, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 550, + 352 + ], + "type": "table", + "html": "
MethodType#Param.FLOPsTop-1 Acc(%)
Small Models (cont.)
CSwin-S [8]Attn35M6.9G83.6
NAT-S [21]Attn51M7.8G83.7
SMT-B [32]Attn32M7.7G84.3
PMamba-L2 [59]SSM25M8.1G81.6
VMamba-S [34]SSM50M8.7G83.6
LVMamba-S [27]SSM50M11.4G83.7
MSVMamba-S [43]SSM50M8.8G84.1
SpatialMamba-S [56]SSM43M7.1G84.6
VSSD-SSSD50M8.1G84.6
Base Models
ConvNeXt-B [37]Conv89M15.4G83.8
MambaOut-B [62]Conv85M15.8G84.2
DeiT-B [48]Attn86M17.5G81.8
Swin-B [35]Attn88M15.4G83.5
CSwin-S [8]Attn78M15.0G84.2
NAT-B [21]Attn90M13.7G84.3
PMamba-L3 [59]SSM50M14.4G82.3
VMamba-B [34]SSM89M15.4G83.9
MSVMamba-B [43]SSM91M16.3G84.4
SpatialMamba-B [56]SSM96M15.8G85.3
VSSD-BSSD89M16.1G85.4
", + "image_path": "edad4298cbd5eb6ea63b2693ad036c977f3f2eee53c0a3b06fa024b973657798.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 78, + 371, + 531, + 438 + ], + "blocks": [ + { + "bbox": [ + 78, + 354, + 530, + 365 + ], + "lines": [ + { + "bbox": [ + 78, + 354, + 530, + 365 + ], + "spans": [ + { + "bbox": [ + 78, + 354, + 530, + 365 + ], + "type": "text", + "content": "Table 1. Accuracy Comparison across Various Models on ImageNet-1K. The LAttn is the abbreviation of linear attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 371, + 531, + 438 + ], + "lines": [ + { + "bbox": [ + 78, + 371, + 531, + 438 + ], + "spans": [ + { + "bbox": [ + 78, + 371, + 531, + 438 + ], + "type": "table", + "html": "
TinySwin SmallBaseTinyConvNeXtLargeVMambaTinyVSSD
SmallBaseTinySmallBaseTinySmallBaseTinySmall
Thru. (imgs/s)15709726541685108572839114567435311263826417
Memory (MB)2402259633621670175323803777320457807826194620513481
Acc. (%)81.383.083.582.183.183.884.382.683.683.983.884.685.4
", + "image_path": "d9d86a947b92e43e862b944d57f7c630684f01cb745f8aa60592b1b7b62db259.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "type": "text", + "content": "Table 2. Efficiency comparison of different architectures: Attention-based Swin Transformer, CNN-based ConvNeXt, SSM-based VMamba, and our proposed VSSD model. All models are evaluated with an input resolution of " + }, + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 55, + 441, + 555, + 475 + ], + "type": "text", + "content": " and batch size of 128 on an NVIDIA RTX 4090 GPU, following the experimental settings in VMamba." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 482, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 613 + ], + "type": "text", + "content": "in [38], utilizing 50 randomly selected images with a resolution of " + }, + { + "bbox": [ + 55, + 482, + 295, + 613 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 482, + 295, + 613 + ], + "type": "text", + "content": " from the ImageNet-1K validation set. To demonstrate the effectiveness of the proposed NC-SSD, the techniques such as hybrid self-attention and overlapped downsampling layers are not employed in our VSSD model for this analysis. Notably, only our VSSD and DeiT consistently exhibit a global receptive field both before and after training. Under after-training, a distinct cross-shaped attenuation is observed in VMamba, whereas our approach effectively eliminates the impact of token spacing on the contribution of information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 287, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 287, + 636 + ], + "type": "text", + "content": "4.2. Object Detection and Instance Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Configurations. Our evaluation of the VSSD model utilizes the MS COCO dataset [31] within the Mask R-CNN framework [23] for tasks related to object detection and instance segmentation. All experiments are facilitated using the MMDetection library [1]. Consistent with prior studies [34, 35], during the training phase, images are adjusted" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 482, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 578 + ], + "type": "text", + "content": "such that the shorter side measures 800 pixels, while the longer side does not exceed 1333 pixels. Optimization is carried out using the AdamW optimizer, with a set learning rate of 0.0001 and a batch size of 16. When adopting the standard \"1×\" training schedule, the learning rate is reduced by a factor of 0.1 at epochs 8 and 11 while the extended \"3× + MS\" schedule sees a reduction in the learning rate by the same factor at epochs 27 and 33, following Swin [35]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Performance Evaluation. Tab. 3 details the comparative performance of our model against well-established CNNs, ViTs, and other SSM-based models. Our VSSD model demonstrates comparable or superior performance in various configurations. Remarkably, our VSSD-T model demonstrates a significant advantage, outperforming SwinT [35] by margins of +4.3 in box AP and +3.5 in mask AP. This superior performance extends to both small and base variants of our architecture. When evaluated under the extended \"3x\" training schedule, VSSD maintains its competitive edge across all metrics. Specifically, VSSD-S sur" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10824" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 531, + 232 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 531, + 232 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 531, + 232 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 531, + 232 + ], + "type": "image", + "image_path": "f2e5cd666c279ca01b0470ed8d6f8213ba6d11c5412ec53b3ba2bf638c600c93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "type": "text", + "content": "Figure 3. Comparison of the Effective Receptive Field (ERF) among our VSSD, CNN-based models (ResNet [22] and ConvNeXt [37]), attention-based models (Swin [35] and DeiT [48]), and the SSM-based VMamba [34]. Our VSSD effectively eliminates the impact of token spacing on the contribution of information compared to SSM-based VMamba." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 274, + 303, + 651 + ], + "blocks": [ + { + "bbox": [ + 56, + 274, + 303, + 651 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 303, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 303, + 651 + ], + "type": "table", + "html": "
Mask R-CNN 1x
Method\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^m\\)\\(AP^m_{50}\\)\\(AP^m_{75}\\)
Swin-T [35]42.765.246.839.362.242.2
ConvNeXt-T [37]44.266.648.340.163.342.8
MambaOut-T [62]45.167.349.641.064.144.1
MSVMamba-T [43]46.968.751.442.566.245.8
VMamba-T [34]47.369.352.042.766.445.9
VSSD-T47.069.551.642.866.546.1
Swin-S [35]44.866.648.940.963.244.2
ConvNeXt-S [37]45.467.950.041.865.245.1
MambaOut-S [62]47.469.152.442.766.146.2
MSVMamba-S [43]48.170.152.843.267.346.5
VMamba-S [34]48.770.053.443.767.347.0
VSSD-S48.971.153.843.968.147.4
Swin-B [35]46.9--42.3--
ConvNeXt-B [37]47.069.451.742.766.346.0
MambaOut-B [62]47.469.352.243.066.446.3
VMamba-B [34]49.270.953.943.967.747.6
VSSD-B49.972.154.744.869.148.3
Mask R-CNN 3x + MS
Method\\(AP^b\\)\\(AP^b_{50}\\)\\(AP^b_{75}\\)\\(AP^m\\)\\(AP^m_{50}\\)\\(AP^m_{75}\\)
Swin-T [35]46.068.150.341.665.144.9
ConvNeXt-T [37]46.267.950.841.765.044.9
MSVMamba-T [43]48.769.853.343.467.246.8
VMamba-T [34]48.870.453.543.767.447.0
VSSD-T49.171.253.544.067.647.4
Swin-S [35]48.269.852.843.267.046.1
ConvNeXt-S [37]47.970.052.742.966.946.2
MSVMamba-S [43]49.770.954.344.268.047.9
VMamba-S [34]49.970.954.744.268.247.7
VSSD-S50.772.255.845.169.448.8
", + "image_path": "91a2c6477271e5a934a280b279df45e179a10609895db6889c944f582308fbff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "passes VMamba-S by margins of 0.8 and 0.9 in box AP and" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 328, + 274, + 543, + 528 + ], + "blocks": [ + { + "bbox": [ + 55, + 658, + 296, + 692 + ], + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 692 + ], + "type": "text", + "content": "Table 3. Object Detection and Instance Segmentation with Mask R-CNN Framework [23] on MS COCO Dataset [31]. The FLOPs are tested with an input size of " + }, + { + "bbox": [ + 55, + 658, + 296, + 692 + ], + "type": "inline_equation", + "content": "1280 \\times 800" + }, + { + "bbox": [ + 55, + 658, + 296, + 692 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 274, + 543, + 528 + ], + "lines": [ + { + "bbox": [ + 328, + 274, + 543, + 528 + ], + "spans": [ + { + "bbox": [ + 328, + 274, + 543, + 528 + ], + "type": "table", + "html": "
MethodmIoU SSmIoU MS#Param.FLOPs
Swin-T [35]44.445.860M945G
ConvNeXt-T [37]46.046.760M939G
Mambaout-T [62]47.448.654M938G
FocalNet-T [60]45.847.062M998G
MSVMamba-T [43]47.948.563M953G
VMamba-T [34]48.048.862M949G
VSSD-T48.749.759M953G
Swin-S [35]47.649.581M1038G
ConvNeXt-S [37]48.749.682M1027G
Mambaout-S [62]49.550.676M1032G
FocalNet-S [60]48.050.085M1130G
VMamba-S [34]50.651.276M1028G
VSSD-S50.351.081M1021G
Swin-B [35]48.149.7121M1188G
ConvNeXt-B [37]49.149.9122M1170G
Mambaout-B [62]49.651.0112M1178G
FocalNet-B [60]49.050.5126M1354G
VMamba-B [34]51.051.6122M1170G
VSSD-B51.552.2120M1190G
", + "image_path": "2335f6c8a8cd20cfb54d6983dd1a7616cf7c1f1634487b8e5bf831c04550d098.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 536, + 555, + 591 + ], + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 591 + ], + "type": "text", + "content": "Table 4. Results of Semantic Segmentation on the ADE20K Dataset [65] using the UperNet Framework [57]. FLOPs for all models are computed using input dimensions of " + }, + { + "bbox": [ + 313, + 536, + 555, + 591 + ], + "type": "inline_equation", + "content": "512 \\times 2048" + }, + { + "bbox": [ + 313, + 536, + 555, + 591 + ], + "type": "text", + "content": ". In the table, \"SS\" represents single-scale testing, while \"MS\" indicates multi-scale testing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 604, + 407, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 604, + 407, + 616 + ], + "spans": [ + { + "bbox": [ + 314, + 604, + 407, + 616 + ], + "type": "text", + "content": "mask AP, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 624, + 447, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 447, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 447, + 636 + ], + "type": "text", + "content": "4.3. Semantic Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "Configurations. In alignment with the approaches described in Swin [35] and VMamba [34], our experiments leverage the UperNet [57] framework, utilizing an ImageNet pre-trained backbone for initialization. The training regimen spans 160K iterations with a batch size of 16, executed using the MMSegmentation library [5]. The pri" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10825" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 272, + 165 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 272, + 165 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 272, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 272, + 165 + ], + "type": "table", + "html": "
MethodTop-1 Acc(%)Thru. (imgs/sec)Train Thru. (imgs/sec)Param. FLOPs
SSD82.0145559729.8M4.73G
Bi-SSD82.4137341030.3M4.85G
NC-SSD82.8147171729.8M4.80G
w/ Bias82.7128060629.8M4.80G
w/o m82.4151775829.7M4.74G
", + "image_path": "17ec402a95b01ef0736cd3db12e8ba38b99d0e3c8601a363613cb574fc610373.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 273, + 70, + 552, + 165 + ], + "blocks": [ + { + "bbox": [ + 158, + 167, + 170, + 175 + ], + "lines": [ + { + "bbox": [ + 158, + 167, + 170, + 175 + ], + "spans": [ + { + "bbox": [ + 158, + 167, + 170, + 175 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 273, + 70, + 552, + 165 + ], + "lines": [ + { + "bbox": [ + 273, + 70, + 552, + 165 + ], + "spans": [ + { + "bbox": [ + 273, + 70, + 552, + 165 + ], + "type": "table", + "html": "
Op.Down.LayersTop-1#Params.FLOPs (G)Thru. (imgs/sec)Memory. (MB)
Acc(%)
SSDPatch2, 2, 10, 282.029.8 M4.714551751
+NC-SSDPatch2, 2, 10, 282.829.8 M4.814711807
+LPUConv2, 2, 8, 283.427.2 M4.912601941
+HybridConv2, 2, 8, 283.628.2 M5.012631946
+MESAConv2, 2, 8, 483.828.2 M5.012631946
", + "image_path": "cec34fc4c2b4c931e9309eba9e7d0b3e7aa6bd8ac5404741f39fc4755ed6d206.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 406, + 167, + 417, + 175 + ], + "lines": [ + { + "bbox": [ + 406, + 167, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 406, + 167, + 417, + 175 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 177, + 555, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 555, + 212 + ], + "type": "text", + "content": "Table 5. (a). Ablation for different SSD variants and the components in the proposed NC-SSD. (b). Ablation study of VSSD-Tiny on ImageNet-1K. Our NC-SSD outperforms vanilla SSD in terms of accuracy with similar inference throughput and memory cost. Other techniques further enhance the performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 217, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 295, + 277 + ], + "type": "text", + "content": "mary experiments are performed with a standard input resolution of " + }, + { + "bbox": [ + 54, + 217, + 295, + 277 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 217, + 295, + 277 + ], + "type": "text", + "content": ". To further evaluate the robustness of our model, Multi-Scale (MS) testing is implemented. Optimization is carried out using the AdamW optimizer, with the learning rate established at " + }, + { + "bbox": [ + 54, + 217, + 295, + 277 + ], + "type": "inline_equation", + "content": "6 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 217, + 295, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 278, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 278, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 295, + 397 + ], + "type": "text", + "content": "Performance Evaluation. Detailed performance for our model and competitors are displayed in Tab. 4, encompassing both single-scale and multi-scale testing. In the Tiny model category, our VSSD demonstrates compelling advantages over existing architectures. Specifically, VSSD-T surpasses Swin-T, ConvNeXt-T, and VMamba-T by margins of 4.3, 2.7, and " + }, + { + "bbox": [ + 54, + 278, + 295, + 397 + ], + "type": "inline_equation", + "content": "0.7\\mathrm{mIoU}" + }, + { + "bbox": [ + 54, + 278, + 295, + 397 + ], + "type": "text", + "content": " under single-scale testing, and 3.9, 3.0, and " + }, + { + "bbox": [ + 54, + 278, + 295, + 397 + ], + "type": "inline_equation", + "content": "0.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 54, + 278, + 295, + 397 + ], + "type": "text", + "content": " with multi-scale testing, respectively. When scaled to the base variant, VSSD-B achieves an mIoU of 52.2 under multi-scale evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 407, + 124, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 124, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 124, + 419 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed modules, we conducted detailed ablation experiments on the VSSD-Tiny model. Using the SSD block as the token mixer and patchified downsamplers (e.g. convolution with " + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "content": " kernel and stride of 4 in stem) following Swin [35] and vanilla VMamaba [34], we established the baseline configuration, detailed in the first row of Tab. 5b. For throughput testing, we utilized an NVIDIA RTX 4090 GPU with a batch size of 128. SSD Mechanisms. In our ablation study for the token mixer, we explored different scanning routes for SSD. Specifically, the Bi-SSD is introduced, where we split the channels into two parts and reverse one part to create backward scanning sequences. These sequences with opposite scanning routes are then concatenated after the SSD block. As shown in Tab. 5a, our NC-SSD model outperforms both the vanilla SSD and Bi-SSD by " + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "content": " in top-1 accuracy, respectively. Moreover, the training throughput are significantly enhanced, with NC-SSD improving training throughput by over " + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "content": " compared to the Bi-SSD. Additionally, we conduct ablation studies on key components of NC-SSD, specifically examining the impact of the omitted bias term in Eq. 10 and the vector " + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "content": " in Eq. 7. Incorporating the bias term leads to a marginal decrease of " + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 56, + 426, + 296, + 714 + ], + "type": "text", + "content": " in top-1 accuracy while adversely affecting both training and in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 217, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 240 + ], + "type": "text", + "content": "ference efficiency. In contrast, removing vector " + }, + { + "bbox": [ + 313, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 217, + 555, + 240 + ], + "type": "text", + "content": " slightly improves model throughput but results in a " + }, + { + "bbox": [ + 313, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 217, + 555, + 240 + ], + "type": "text", + "content": " reduction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 241, + 554, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 349 + ], + "type": "text", + "content": "LPU and Overlapped Downsampler. Our experiments in Tab. 5b demonstrate that incorporating overlapped downsampling layers and the local perception unit (LPU) introduces beneficial inductive bias and enhances local perception capabilities. These modifications improve the top-1 accuracy by " + }, + { + "bbox": [ + 313, + 241, + 554, + 349 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 313, + 241, + 554, + 349 + ], + "type": "text", + "content": " while marginally affecting inference throughput. To maintain comparable computational costs despite the additional parameters and FLOPs introduced by these operations, we strategically reduce the model depth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "text", + "content": "Hybrid Architecture with Attention and MESA. Tab. 5b validates the effectiveness of our hybrid approach, which combines standard attention in the final stage with overlapped downsampling. Specifically, substituting NC-SSD with standard attention in the last stage yields a " + }, + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "text", + "content": " accuracy improvement while slightly increasing the parameter count. Additionally, following MLLA [19], we incorporate MESA [11] to mitigate overfitting, resulting in a further " + }, + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "text", + "content": " gain in top-1 accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 618 + ], + "type": "text", + "content": "In conclusion, our study introduces the NC-SSD, which redefines SSD by modifying the role of matrix " + }, + { + "bbox": [ + 313, + 486, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 486, + 555, + 618 + ], + "type": "text", + "content": " and eliminating the causal mask. These adaptations facilitate a transition to a non-causal mode, significantly enhancing both accuracy and efficiency. Extensive experiments demonstrate its superiority over the vanilla SSD and its multi-scan-based variants. Furthermore, by integrating techniques such as hybrid standard attention and overlapped downsampling, our VSSD model achieves comparable or superior performance compared to well-established CNNs, ViTs, and Vision SSMs across several widely-used benchmarks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 628, + 430, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 430, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 430, + 641 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 554, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 703 + ], + "type": "text", + "content": "This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10826" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "text", + "content": "[1] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "text", + "content": "[2] Tianxiang Chen, Zhentao Tan, Tao Gong, Qi Chu, Yue Wu, Bin Liu, Jieping Ye, and Nenghai Yu. Mim-istd: Mamba-in-mamba for efficient infrared small target detection. arXiv preprint arXiv:2403.02148, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 241, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 296, + 295 + ], + "type": "text", + "content": "[4] Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamas Sarlos, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, et al. Rethinking attention with performers. arXiv preprint arXiv:2009.14794, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 342 + ], + "type": "text", + "content": "[5] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 344, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 344, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 295, + 376 + ], + "type": "text", + "content": "[6] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. In ICML, 2024. 1, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 379, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 379, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 379, + 295, + 411 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "text", + "content": "[8] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 460, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 295, + 525 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 529, + 295, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 529, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 57, + 529, + 295, + 561 + ], + "type": "text", + "content": "[10] Chengbin Du, Yanxi Li, and Chang Xu. Understanding robustness of visual state space models for image classification. arXiv preprint arXiv:2403.10935, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 564, + 295, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 564, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 57, + 564, + 295, + 596 + ], + "type": "text", + "content": "[11] Jiawei Du, Daquan Zhou, Jiashi Feng, Vincent Tan, and Joel Tianyi Zhou. Sharpness-aware training for free. NeurIPS, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 599, + 295, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 599, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 57, + 599, + 295, + 631 + ], + "type": "text", + "content": "[12] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "type": "text", + "content": "[13] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "text", + "content": "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "text", + "content": "[17] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 188, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 553, + 230 + ], + "type": "text", + "content": "[18] Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 1, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 234, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 277 + ], + "type": "text", + "content": "[19] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 2, 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 279, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 279, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 317, + 279, + 553, + 311 + ], + "type": "text", + "content": "[20] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 313, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 313, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 313, + 553, + 346 + ], + "type": "text", + "content": "[21] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 348, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 348, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 348, + 553, + 380 + ], + "type": "text", + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 383, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 403 + ], + "type": "text", + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 553, + 460 + ], + "type": "text", + "content": "[24] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 462, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 483 + ], + "type": "text", + "content": "[25] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 487, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 553, + 518 + ], + "type": "text", + "content": "[26] Gao Huang, Zhuang Liu, Geoff Pleiss, Laurens Van Der Maaten, and Kilian Weinberger. Convolutional networks with dense connectivity. IEEE TPAMI, 2019. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "type": "text", + "content": "[27] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 567, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 598 + ], + "type": "text", + "content": "[28] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In ICML, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 601, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 632 + ], + "type": "text", + "content": "[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 2012. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[30] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In ECCV, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 1, 6, 7" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10827" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[32] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 161 + ], + "type": "text", + "content": "[33] Jing Liu, Bohan Zhuang, Zhuangwei Zhuang, Yong Guo, Junzhou Huang, Jinhui Zhu, and Mingkui Tan. Discrimination-aware network pruning for deep model compression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[34] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. In NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[36] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "text", + "content": "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[38] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard S. Zemel. Understanding the effective receptive field in deep convolutional neural networks. In NeurIPS, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "text", + "content": "[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[40] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, Junjie Yan, Lingpeng Kong, and Yiran Zhong. cosformer: Rethinking softmax in attention. arXiv preprint arXiv:2202.08791, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[41] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. Sg-former: Self-guided transformer with evolving token reallocation. In ICCV, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "type": "text", + "content": "[42] Jiacheng Ruan and Suncheng Xiang. Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "text", + "content": "[43] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. In NeurIPS, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "text", + "content": "[44] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[45] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "text", + "content": "[47] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019, 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[49] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 140 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[51] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "text", + "content": "[52] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[54] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. arXiv preprint arXiv:2103.15808, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "type": "text", + "content": "[55] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Dat++: Spatially dynamic vision transformer with deformable attention. arXiv preprint arXiv:2309.01430, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "text", + "content": "[56] Chaodong Xiao, Minghan Li, Zhengqiang Zhang, Deyu Meng, and Lei Zhang. Spatial-mamba: Effective visual state space models via structure-aware state fusion. In ICLR, 2025. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "text", + "content": "[57] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "text", + "content": "[58] Saining Xie, Ross Girshick, Piotr Dolkar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 533 + ], + "type": "text", + "content": "[59] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. arXiv preprint arXiv:2403.17695, 2024. 1, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "text", + "content": "[60] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks. NeurIPS, 2022. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 556, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 600 + ], + "type": "text", + "content": "[61] Yuhuan Yang, Chaofan Ma, Jiangchao Yao, Zhun Zhong, Ya Zhang, and Yanfeng Wang. Remember: Referring image segmentation with mamba twister. arXiv preprint arXiv:2403.17839, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[62] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[63] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In CVPR, 2022. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[64] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 1" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10828" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 183 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[65] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 1, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[66] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "text", + "content": "[67] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10829" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_content_list.json b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2835fdfd6909ecdceaa616e76aeae96d7ac3b419 --- /dev/null +++ b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_content_list.json @@ -0,0 +1,2443 @@ +[ + { + "type": "text", + "text": "VTimeCoT: Thinking by Drawing for Video Temporal Grounding and Reasoning", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinglei Zhang $^{1,*}$ Yuanfan Guo $^{2,*}$ Rolandas Alexandros Potamias $^{3}$", + "bbox": [ + 223, + 179, + 772, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiankang Deng $^{3}$ Hang Xu $^{2}$ Chao Ma $^{1,\\dagger}$", + "bbox": [ + 328, + 199, + 666, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University $^{2}$ Noah's Ark Lab $^{3}$ Imperial College London", + "bbox": [ + 176, + 215, + 820, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangjinglei168,chaoma}@sjtu.edu.cn", + "bbox": [ + 334, + 237, + 661, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guoyuanfan1, xuHang}@huawei.com, {j.deng16, r.potamias}@imperial.ac.uk", + "bbox": [ + 181, + 253, + 810, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe77a4122404595f3d3153ae1e10d098fea5860feb41ce344586e51ce924620c.jpg", + "image_caption": [ + "How many streams did the cat cross in the video?" + ], + "image_footnote": [], + "bbox": [ + 96, + 304, + 480, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e418e21b15791e30915f398284f76a2bb4fa6041a11385dd17e212aa92a80617.jpg", + "image_caption": [ + "GPT4o:one" + ], + "image_footnote": [], + "bbox": [ + 106, + 387, + 130, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ea4872bac4f1be9fc5509a15c51c6695578eeebb9cb26d5b12eec6511916af9.jpg", + "image_caption": [ + "VTimeCoT + GPT40" + ], + "image_footnote": [], + "bbox": [ + 259, + 387, + 284, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca7f6f65b3e6b786d149aefbe4311a0c46d28c9722db93d660a119acca7431fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 387, + 320, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "THOUGHT 0: I can use the tool to draw progress bar and highlight the moments for 'streams the cat crossed'.", + "bbox": [ + 109, + 414, + 444, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ACTION 0:", + "bbox": [ + 109, + 441, + 161, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Call and ", + "bbox": [ + 109, + 452, + 401, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9dbb6183ff21fe52b936d4944c3b691ca4abe6d22f69ed2cbbd2f395f6ee8165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 470, + 127, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f086e473831ed1baf3b25f427a35ef1daeb0cd413e6de51e2cbd75879315ba6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 469, + 232, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/14b42001dbbd8d52f184e3b659849c8153c68c3e59428f90ea50fc0a53361666.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 470, + 354, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70e3d9cc003e995ea6f3501986a7a165305c70ff59e24198d45a63137fdca569.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 470, + 470, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "THOUGHT 1: The tool is helpful. I see the cat crossing streams at 52s to 68s and 71s to 76s. So it crossed two streams.", + "bbox": [ + 107, + 523, + 450, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/451949e5850d0b3ae0f65a22c4451172878935df6a79f2e332b199240d106eb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 304, + 900, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/446d95f8cbd00ff432d5384327f783f2ef44d1dec697669fe3db11b3d429cdef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 358, + 900, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e5376232fe887b572579d1ce7c4fd9f0958ff5fc0350909b889ab7b6aa10932d.jpg", + "image_caption": [ + "Figure 1. We propose VTimeCoT, a Visual Time Chain-of-Thought framework for video temporal grounding and reasoning. VTimeCoT constructs cross-modality reasoning across both video and text, which enables the MLLM to utilize progress bar tools to annotate the time progression and highlight the key relevant segments to answer complex temporal questions." + ], + "image_footnote": [], + "bbox": [ + 488, + 419, + 900, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 628, + 326, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, video question answering based on multimodal large language models (MLLM) has garnered considerable attention, due to the benefits from the substantial advancements in LLMs. However, these models have a notable deficiency in the domains of video temporal grounding and reasoning, posing challenges to the development of effective real-world video understanding systems. Inspired by how humans use video players to interact with the progress bar for video comprehension, we introduce VTimeCoT, a simple yet effective training-free framework, designed for high-performance video grounding and reasoning. The proposed framework incorporates two novel visual tools of the progress bar: a plug-and-play progress bar integration tool and a high-efficiency highlighting tool. In addition,", + "bbox": [ + 88, + 661, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to address the limitations of conventional text-based chain-of-thought (CoT) approaches, we introduce a visuotemporal CoT process that integrates cross-modality reasoning across both video and text. Our approach demonstrates significant performance improvements on both Qwen2VL-7B and GPT4o baselines in tasks of video temporal grounding and reasoning-based question answering. Finally, we showcase that the proposed framework achieves a compositional and interpretable reasoning process. Project page: https://vtimecot.github.io.", + "bbox": [ + 511, + 630, + 906, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 813, + 645, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video understanding is a longstanding problem in computer vision and has attracted more attention with the emergence of large language models (LLM) [5, 17, 40, 43]. Video QA is a representative task reflecting the video reasoning abl-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 114, + 886, + 364, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24203", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ity. Given that videos typically contain numerous events occurring at different time points, accurately answering questions is a highly challenging task. Recently, some end-to-end video understanding models based on multimodal large language models (MLLM) [20, 23, 26, 27, 56, 58] have been proposed to address video question-answering tasks, exhibiting remarkable capabilities. However, despite their ability to generate seemingly plausible results, these methods have been shown to exhibit a notable deficiency in temporal grounding [15] and fail to provide a temporal-grounded reasoning process.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advent of tool usage capabilities in LLMs, several agent-based video question-answering methods have been proposed [9, 33, 37, 44, 53], enabling step-by-step video understanding. These methods prompt the LLMs to sequentially reason and invoke external tool models (e.g., object detection and captioning models) to collect textual clues, ultimately inferring the final answer. This, so called, chain-of-thought approach offers significant potential, as it is compositional and training-free, demonstrating impressive zero-shot performance. However, such methods predominantly rely on language and captions as intermediaries, limiting their ability to directly capture the visuotemporal dynamics inherent in videos, especially in the case of long videos. Inspired by the time-progress bar commonly used in video players during human interactions, we identify it as an intuitive tool for grounding time and enhancing video comprehension. In contrast to written language, the progress bar, along with the concurrent visual content, directly conveys the concept of temporal progression. The video progress bar can represent temporal relationships through sequential positioning or more abstract temporal dependencies.", + "bbox": [ + 91, + 261, + 483, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a simple yet effective framework that empowers multimodal LLMs to utilize a video progress bar to construct a Visual Time Chain-of-Thought (VTimeCoT), facilitating long video temporal grounding and reasoning. Inspired by LLMs coupled with visual programming [12, 39], we propose enabling multimodal LLMs to generate code that overlays a progress bar on the video. First, we construct a frame-sync progress bar integration tool that can be invoked by the MLLM, which generates a progress bar displayed at the bottom of the video, marking the progress and annotating the timestamps on each frame. This is naturally adapted to long video and enables the MLLM to perceive the speed of temporal progress directly from the annotated video frames, adapting seamlessly to any frame-per-second (FPS) sampling rate. Second, to guide the model's attention on key relevant temporal segments, we propose a training-free and long-video-adaptive highlighting tool for the progress bar based on video-text similarity. Specifically, we leverage a robust video-text foundation model to identify the time intervals with the", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "top-k highest similarity and highlight them on the visual progress bar. Finally, we prompt the multimodal LLM to invoke the progress bar tools within a visuotemporal chain-of-thought process that integrates text, program, and video annotated with the progress bar to facilitate long video reasoning. At each thought step, the model performs cross-modality reasoning and determines whether to dynamically update the video memory while progressively inferring the answer. As shown in Fig. 1, to determine how many times a cat crosses a stream in the input video, the model first invokes tools to generate a progress bar, identifying the key timestamps corresponding to 'streams the cat crossed'. By analyzing the progress bar annotated frames, the model infers the exact timing of each crossing event and accurately determines the number of crossings. This approach enhances the model's temporal reasoning capabilities by seamlessly integrating temporal-grounded visual evidence.", + "bbox": [ + 511, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of the proposed framework across a wide range of video temporal grounding and video-QA tasks. The proposed method consistently outperforms the baseline models, Qwen2VL and GPT-4o in temporal grounding tasks, achieving significant performance gains with an average IoU improvement of $6.58\\%$ and $16.83\\%$ over GPT-4o across Charades-STA [11] and QVHighlights [19] benchmarks. Similarly, VTimeCoT significantly improves the accuracy of questions related to temporal retrieval, event counting, and event ordering in reasoning-based question-answering tasks. Specifically, the proposed method consistently surpasses state-of-the-art methods across Vript-RR [50] and VideoMME [10] benchmarks.", + "bbox": [ + 511, + 347, + 903, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, we present VTimeCoT, a training-free model for temporal reasoning based on the visual progress bar. Specifically:", + "bbox": [ + 511, + 559, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first, to the best of our knowledge, visual time chain-of-thought framework for video temporal grounding and reasoning. In contrast to previous methods that rely solely on textual reasoning, we leverage the visual progress bar as a medium, making a significant step towards real-world video understanding systems.", + "- The proposed reasoning framework leverages two novel tools from video progress bar integration and highlighting, enabling the MLLM to accurately perceive the times-tamp of each frame and identify key temporal segments.", + "- The proposed method enhances the performance of MLLMs by a large margin in video temporal grounding and reasoning tasks through a training-free approach, while also demonstrating strong reasoning capabilities." + ], + "bbox": [ + 511, + 606, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 829, + 653, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal LLM for Video. In contrast with traditional video understanding such as object detection, segmentation and pose estimation [34, 36, 41, 57], recent mul", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24204", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "timodal LLMs enable understanding through natural language. End-to-end video multimodal LLMs are trained on large-scale datasets, to align visual inputs with the language modality and integrate with large language models. Video-ChatGPT [31] leveraged CLIP [35] to extract spatial and temporal features from videos and integrate them with Vicuna [7]. Video-LLaMA [56] proposed a video Q-Former designed to enhance the modeling of temporal variations. Subsequently, a series of video multimodal LLM variants [6, 20, 22, 24–26, 48] have introduced advancements in enhancing video representation, expanding data scale, and refining training methods, leading to improved video question-answering performance. However, these end-to-end models operate as black boxes, directly generating answers without an explicit video reasoning process. In contrast, we propose a video understanding framework that incorporates a structured reasoning process.", + "bbox": [ + 89, + 90, + 480, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Programming and Tool Usage of LLM. Due to the strong code-generation capabilities of LLMs, several works [8, 12, 14, 29, 39, 52] attempted to prompt LLMs to call compositional visual tools to address visual question-answering tasks. VisProg [12] and ViperGPT [39] defined a variety of image foundation functions (e.g., object detection) and prompted the LLMs to generate code that invokes these functions for answering questions. MM-REACT [52] tackled visual tasks through multi-step reasoning guided by text prompts to the LLM, where each step involves thought and calling visual tools. More recently, several studies [9, 33, 37, 44, 53] proposed tool-usage frameworks tailored for video understanding. By designing specialized video tools and calling steps, these approaches extract textual cues from videos to facilitate question answering. In contrast to the aforementioned works that solely rely on text as the reasoning medium, we introduce a visual progress bar, enabling more effective integration of visual and temporal cues.", + "bbox": [ + 91, + 367, + 482, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Prompt for Multimodal LLM. LLMs have been shown to enhance their understanding capabilities by incorporating overlaid visual prompts on images (e.g., circles, keypoints) [3, 4, 13, 21, 32, 38, 46, 47, 51]. SoM [51] proposed to utilize SAM [18] segmentation model to overlay mask colors and labels on images as input of GPT-4V [2], demonstrating strong zero-shot performance on vision-grounding tasks. Hu et al. [13] introduced a chain-of-thought framework combined with visual prompts to draw markers on images during reasoning, effectively enhancing the image understanding performance of MLLMs. However, all previous methods are limited to spatial prompts, limiting their understanding capabilities. In this work, we propose to design the progress bar as visual-temporal prompts to facilitate temporal reasoning.", + "bbox": [ + 89, + 672, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose VTimeCoT, a general framework that empowers multimodal large language models using a visual progress bar as an intermediate reasoning step, facilitating temporal grounding and reasoning. Fig. 2 illustrates how our approach works. Given a raw video and a question, our method generates a chain-of-thought and invokes a set of tools for progress bar integration. By utilizing the visual cues from the progress bar as a medium, the model progressively infers the final answer.", + "bbox": [ + 511, + 114, + 906, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Visuotemporal Chain-of-Thought", + "text_level": 1, + "bbox": [ + 511, + 258, + 807, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In video understanding, particularly for long videos, existing MLLM methods can only provide static answers, which limits the model's ability to focus on critical scenes and results in significant shortcomings in real-world settings. Since the MLLMs fail to identify key scenes within a single forward pass, it is necessary to interact with the video to progressively infer the answer. Therefore, we propose a framework based on interactive mechanisms and dynamic video memory.", + "bbox": [ + 511, + 281, + 906, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework tackles video grounding and reasoning tasks through an iterative interaction with the MLLM, built upon a tool set leveraging the progress bar. Given a textual question and a video as input, VTimeCoT generates a sequence of thoughts and actions to dynamically update the video memory. By manipulating the video, it acquires the necessary information to answer the question. In this process, the MLLM reasons by plotting and analyzing the progress bar, seamlessly integrating both textual and visual reasoning into a Visuotemporal Chain-of-Thought. As illustrated in Fig. 2, VTimeCoT first provides the MLLM with a formatted prompt for initialization and executes actions at each time step $t$ . The tool set includes the , , , etc. The tool is utilized to trim specific segments when the MLLM determines that the video is too long to find the required information. The pseudo-code is presented in Algorithm 1, where $\\mathbb{V}$ and $\\mathbb{L}$ represent the modalities of video and language, respectively. In Algorithm 1, $p$ serves as a parser to retrieve specific sub-strings by keyword, and $c$ is a code builder that integrates the toolset library and generated code.", + "bbox": [ + 511, + 417, + 908, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initialization Prompt. To enable the MLLM to perform step-by-step reasoning and a structured output, we follow [13] and construct a set of initialization prompts. These prompts define the specific output structure (formatted by keywords of 'THOUGHT', 'ACTION', and 'TERMINATE') for each step and declare the Python tools that the model should invoke. The initialization prompts, along with the question and video frames, are then fed into the MLLM, triggering the iterative loop.", + "bbox": [ + 511, + 734, + 908, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thought. During this step, the model analyzes the historical context and video memory to generate its reasoning", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d9e22ee9e0ace6f86642d0065b302c600598867d6a2c306822daa9db50b36b28.jpg", + "image_caption": [ + "Figure 2. Overview of our method. On the left, we demonstrate how the framework iteratively generates thoughts and actions, which dynamically updates the video memory with an overlaid progress bar for reasoning. On the right, we illustrate two novel tools that integrate the frame-sync visual progress bar and highlight key moments." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the current step. It determines whether a direct answer can be provided directly in this step or if tool assistance is needed, and if so, which tools should be invoked.", + "bbox": [ + 88, + 440, + 482, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action. Following the thought step, the model generates an executable Python script utilizing the given tool functions. The framework then builds the code with the toolset library and executes the script to manipulate the video frames.", + "bbox": [ + 89, + 486, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Video Memory. Using the manipulated video frames we subsequently update the frames in the video memory. The updated context with the manipulated video frames is then fed into the MLLM to proceed with the next step of thought.", + "bbox": [ + 89, + 547, + 483, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To terminate the loop, the model determines at each step whether to cease reasoning and generate the 'TERMINATE' keyword to end. To prevent excessively prolonged reasoning, we impose a maximum step limit, enforcing termination at step $T$ .", + "bbox": [ + 89, + 625, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Frame-Sync Visual Progress Bar Integration", + "text_level": 1, + "bbox": [ + 89, + 710, + 468, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enabling MLLM to perceive the precise timestamp of each frame is a significant challenge and an unresolved problem, due to the diverse temporal sampling rates during MLLM training and inference without time. Although recent approaches [15, 16] have attempted to address this issue by incorporating temporal positional embeddings before the video encoder and introducing temporal-grounding training tasks, their flexibility is limited due to the scarcity of specific data and the need for fine-tuning. This issue is particularly pronounced in long videos, where the token limitations of MLLMs prevent them from accurately perceiving", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e0fb44169b00220a446443ae4f25ff1564008d043edb6676df7d3a41ed065d6a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Algorithm 1 Visuotemporal Chain-of-Thought
Input: {V, Q||V ∈ V, Q ∈ L}; I ∈ L;MLLM: V, L → L; Toolset: {T1, . . ., Tn} ∈ L
1: s ← 0; V0 ← V
2: H0 ← {I, Q}▷ Initialization prompt and query
3: while true do
4: ys ← MLLM (Vs, Hs)
5: THOUGHTs, ACTIONs, TERMINATEs ← p(ys)
6: if not TERMINATEs then
7: Cs ← c(ACTIONs, Toolset)▷ Build code
8: Vs+1 ← Cs(Vs)▷ Update video
9: Hs+1 ← {Hs, ys}▷ Update history
10: else
11: return ys▷ Respond final answer
12: end if
13: end while
", + "bbox": [ + 511, + 438, + 906, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the timestamp of each frame.", + "bbox": [ + 511, + 718, + 707, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by visual programming paradigms [12, 39], we propose a simple yet effective method to empower MLLMs to perceive the precise timestamp of each frame, by generating the video progress bar. Leveraging the code generation capabilities of MLLMs, our approach prompts the MLLM to invoke Python plotting tools and use their universal OCR and shape understanding capabilities to perceive time, without any additional training requirements. Specifically, given a raw video $V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ as input, this tool can generate the progress bar at the bottom of the original frames and return the annotated frames. To enable adaptation to ar", + "bbox": [ + 509, + 734, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24206", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "binary FPS sampling, we perform a frame-synchronization step to convert frame indexes to the real second time, as shown in Fig. 2. To enable MLLM to perceive seconds directly without any additional hour-minute-second conversions, we design the timestamps using a straightforward seconds format. To construct a robust visual prompt that can be easily interpreted from the MLLM, we employ compositional components using elongated rectangles, circles, and timestamps to plot the progress bar. Finally, the progress bar is placed at the bottom of raw frames, which can be expressed as:", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {t} ^ {\\prime} = V _ {t} \\oplus T _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 257, + 482, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ represents the vertical concatenation operation, $T_{t}$ denotes the image of the progress bar generated by OpenCV at time $t$ and $V_{t}^{\\prime}$ is the resulting frame. We wrap this integration process into a Python tool function, allowing MLLM to invoke it for generating the annotated video.", + "bbox": [ + 89, + 282, + 482, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Moment Highlight by Video-Text Similarity", + "text_level": 1, + "bbox": [ + 89, + 368, + 460, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although multimodal LLMs have made significant progress in video content captioning, recent studies [15, 16] have revealed substantial deficiencies in their temporal grounding capabilities. Existing works [15, 16] rely on constructing temporal grounding instructional data and require finetuning to emerge such capability. Besides, due to token limitations in MLLMs, they struggle to perform temporal grounding for long videos. In contrast, we propose a training-free and long-video-adaptive method to enhance the temporal grounding capabilities of MLLMs, leveraging the similarities from the robust video-text foundation models.", + "bbox": [ + 89, + 391, + 482, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a query and a video $V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}$ containing multiple events, our proposed tool estimates the query-relevant temporal segments, including their start and end times, and generates a highlighted visual prompt.", + "bbox": [ + 89, + 556, + 482, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moment Retrieval by Video-Text Similarity. To construct a robust zero-shot moment retrieval module, capable of generalizing across a wide range of videos, we employ the VideoCLIP-XL [42] foundation model as our embedding extractor. The video is initially processed at a high temporal resolution of $r$ FPS. We group every 8 frames into a video clip, resulting in $N$ clips. It should be noted that since the temporal retrieval module operates as an external component, it can efficiently handle high frame rates without being constrained by the token count of the MLLM. This makes our method adaptive to long-form video. These clips are then processed by the VideoCLIP-XL visual encoder, producing $N$ video embeddings. We use VideoCLIP-XL text encoder to extract the text embedding for the text query and we compute the cosine similarity between the query embedding and the video embeddings of each clip:", + "bbox": [ + 89, + 619, + 482, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} (i) = \\frac {e _ {i} ^ {v} \\cdot e ^ {q}}{\\left| e _ {i} ^ {v} \\right| \\left| e ^ {q} \\right|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 871, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_i^v$ denotes the embedding of the $i$ -th clip and $e^q$ is the query embedding. Using top- $k$ selection, we obtain the $k$ clips with the highest similarity and extract the start and end timestamps of each contiguous segment.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moment Highlight on the Progress Bar. Motivated by a series of works that demonstrated the effectiveness of visual cues as a medium for reasoning beyond text [3, 13, 21, 32, 46], we propose using moment-highlighting cues for video temporal reasoning. As shown in Fig. 2, we construct the highlight tool using OpenCV library to plot colorful highlight masks on specific intervals of the progress bar. This tool takes the video frames and time intervals as input and annotates the highlighted progress bar under the frames. To ensure that the MLLM can easily interpret the highlights, we utilize different colors between the mask and the progress bar. We wrap the aforementioned retrieval and highlighting process as a Python tool function, enabling the MLLM to invoke it to manipulate the input video frames.", + "bbox": [ + 511, + 152, + 906, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 376, + 645, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first present the implementation details, datasets and evaluation metrics. Subsequently, we assess the performance of the proposed framework on the video temporal grounding benchmarks (Sec. 4.1). To further evaluate the reasoning capabilities, we conducted extensive quantitative and qualitative analyses on the \"reasoning based on retrieving\" benchmark of Vript-RR (Sec. 4.2). In addition, we demonstrate the significant role our reasoning framework plays in enhancing long video question-answering performance (Sec. 4.3). Finally, we give the ablations of key modules (Sec. 4.4).", + "bbox": [ + 511, + 402, + 903, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In the framework, we conducted experiments on two core MLLMs, Qwen2VL-7B [43] and GPT4o-20240513 [17], as they represent the state-of-the-art open-source and closed-source MLLM models, respectively. We set the decoding temperature of the MLLM to 0. By default, the input for the MLLM consists of 32 frames uniformly sampled from the video. For video frames, $H$ and $W$ are resized to make the longer side 480 pixels. Our agent implementation is based on AutoGen [45]. The implementations of tool functions are adapted from Gradio [1]. For GPT4o, we utilize the OpenAI API service. The details of the MLLM prompts are provided in the supplementary materials. For VideoCLIP-XL we use FPS $r = 1$ to sample video frames and group every 8 frames as a clip to extract embedding. Regarding the top- $k$ selection of clips, $k$ is empirically set to 8. The maximum number of reasoning steps $T$ is set to 3.", + "bbox": [ + 511, + 568, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. To evaluate the proposed VTimeCoT method, we utilized four benchmark datasets, spanning a wide range of video temporal grounding, reasoning, and long-video question-answering tasks.", + "bbox": [ + 511, + 825, + 903, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Charades-STA [11] is a benchmark dataset for temporal", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1c3f026fbd1d4851d87059b90733015ab6d17d82817ed0e26044add37f65ff4c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodR1@0.3R1@0.5R1@0.7mIoU
VideoChat-7B [23]9.003.301.306.50
VideoLLaMA-7B [56]10.403.800.907.10
VideoChatGPT-7B [31]20.007.701.7013.70
LLAVA-Onevision-7B [20]33.0411.054.1120.98
VTimeLLM-7B [15]51.0027.5011.4031.20
VTimeLLM-13B [15]55.3034.3014.7034.60
Qwen2VL-7B [43]37.3112.854.1124.34
VTimeCoTQwen2VL-7B66.9638.7920.8343.41
GPT4o [17]63.7637.1214.6540.20
VTimeCoTGPT4o74.0651.0222.4546.78
", + "bbox": [ + 106, + 88, + 472, + 271 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c86945e2afc04894bc98b39b97eaef15f580717f7fb4fa3f210320c7659abdf.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on the Charades-STA dataset for temporal grounding. We report the mIoU and the recall performance at different IoU thresholds." + ], + "table_footnote": [], + "table_body": "
MethodR1@0.3R1@0.5R1@0.7mIoU
LLAVA-Onevision-7B [20]34.9117.919.725.53
VTimeLLM-7B [15]44.5825.039.2928.99
Qwen2VL-7B [43]31.8714.657.3522.77
VTimeCoTQwen2VL-7B67.5045.7925.1146.21
GPT4o [17]55.6135.6819.2937.66
VTimeCoTGPT4o79.3559.7433.8154.49
", + "bbox": [ + 106, + 339, + 472, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "grounding, composed of 1334 videos along with the corresponding start-end frame annotations of 3720 queries. The average video length is 30 seconds. We follow [15] and report mean IoU (mIoU) and recall@1, IoU $\\geq m(R@m)$ metrics, where $m$ includes 0.3, 0.5 and 0.7. IoU denotes the intersection over the union between the predicted and ground truth time segments.", + "bbox": [ + 89, + 547, + 482, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "QVHighlights [19] is a benchmark dataset for semantic relevance and temporal grounding from discontinuous time segments. It contains 1519 videos, with an average length of 150 seconds, and 1550 queries with time annotations, where each query spans multiple discontinuous time segments within a video. For evaluation, we report the commonly used mIoU and R@m metrics.", + "bbox": [ + 89, + 655, + 482, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vript-RR [50] is a challenging benchmark dataset for scene retrieval and multi-hop reasoning. The average length of the videos is 622 seconds. It contains 152 questions, each accompanied by a hint to locate the scene that the question refers to. The benchmark includes both multiple-choice and open-ended question-answering settings. We follow [50] and evaluate the open-ended accuracy using GPT4.", + "bbox": [ + 89, + 762, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VideoMME [10] is a video question-answering benchmark tailored for MLLMs, featuring diverse videos ranging from", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/60e75681c3acfc68915109eb2159fa383c62061e5864328faa2310506c197361.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on the QVHighlights dataset for temporal grounding of discontinuous segments. We report the mIoU and the recall performance at different IoU thresholds." + ], + "table_footnote": [], + "table_body": "
MethodsMulti-ChoiceOpen
VideoChatGPT [31]29.6017.80
Video-LLaMA [56]28.3014.50
VideoChat [23]22.4015.10
VideoChat2 [24]42.1022.40
ST-LLM [28]33.6026.30
PLLAVA 7B [48]55.3036.20
VILA-1.5 8B [27]55.3032.30
Qwen2VL-7B [43]59.8735.95
VTimeCoTQwen2VL-7B62.5041.45
GPT4o [17]70.3961.18
VTimeCoTGPT4o83.5568.42
", + "bbox": [ + 560, + 88, + 859, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison on the Vript-RR dataset. We report the response accuracy on both multiple-choice and open-ended question settings.", + "bbox": [ + 511, + 294, + 906, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "11 seconds to 1 hour in length. It contains 900 videos and 2700 question-answer pairs. We follow [10] to report accuracy metrics under two evaluation settings with subtitles and without subtitles.", + "bbox": [ + 511, + 364, + 906, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Temporal Grounding", + "text_level": 1, + "bbox": [ + 511, + 438, + 714, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure accurate temporal reasoning, it is essential to achieve robust and high-performance temporal grounding. In Tab. 1, we compare VTimeCoT with two state-of-the-art baselines, Qwen2VL-7B and GPT4o, to identify the start and end timestamps of the query event. Specifically, using both Qwen2VL-7B and GPT4o models as the core MLLM of our framework, we evaluate the performance of VTimeCoT to localize the query event. As can be easily observed, although these advanced MLLMs achieve state-of-the-art performance in video content recognition, they fall short in accurately grounding temporal boundaries for events in videos, facing difficulties in perceiving the precise timestamp and the boundaries of events. In contrast, the proposed method effectively addresses these challenges, providing accurate visual cues of the progress bar through the reasoning process that facilitates the grounding performance in videos. Note that VTimeCoT demonstrates superior performance without any additional training cost, even compared to the VTimeLLM model, which has been meticulously fine-tuned on extensive temporal grounding data.", + "bbox": [ + 511, + 460, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further evaluate the temporal grounding performance of the proposed and the baseline methods in the presence of multiple discontinuous time segments for each query, we utilized the QVHighlights benchmark dataset (Tab. 2). Given a query and a video, we request from each model to identify start-end timestamps of all temporal segments in the video that fit the query. As can be observed, VTime-CoT consistently enhances the performance compared to different MLLMs. On the contrary, baseline MLLM meth", + "bbox": [ + 511, + 765, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24208", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reasoning based on retrieving scenes", + "text_level": 1, + "bbox": [ + 166, + 99, + 410, + 112 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22ae648c1b8e507545ed2fbd81e1f13dc5919014f71f7020ce376d7f4f25660b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 121, + 480, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "? Hint: A scene in the video features a first-person perspective from a motorcyclist cruising on a highway. ... The motorcycle's chrome details, black leather accessories, the rider's gloved hands are noticeable.", + "bbox": [ + 99, + 181, + 439, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What is the round object on the bike in the middle bottom of the screen?", + "bbox": [ + 99, + 223, + 447, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(A) A GPS device", + "bbox": [ + 101, + 234, + 184, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(B) A sport camera", + "bbox": [ + 279, + 234, + 367, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(C) The motorcycle's headlight", + "bbox": [ + 101, + 244, + 243, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(D) The motorcycle's spedometer", + "bbox": [ + 279, + 244, + 433, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6a20ce201c7aad3b84205d472ec823da16927711521c1a1b96c481480360002.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison on Vript-RR benchmark. In contrast to GPT-4o which is misled by the irrelevant scenes, VTimeCoT reasons using the progress bar and highlights the key moments, leading to accurate answers." + ], + "image_footnote": [], + "bbox": [ + 101, + 271, + 129, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT40: (C) The motorcycle's headlight.", + "bbox": [ + 135, + 279, + 328, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e96aab5b6c8c4d153a2444d3354fde9bea27ec62754aefefca67d847258e7f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 275, + 465, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f60ecb40aab8634bea755dfb4513177b5a75c30fe8cf24aee5a014d8285a8447.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 93, + 531, + 111 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VTimeCoT + GPT40", + "bbox": [ + 535, + 95, + 661, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "THOUGHT0: I can use the progress_bar and highlight tool to highlight the key moments. Since the video is too long, I will cut the video into segments for a thorough review. Then I will display the highlighted frames to help me answer the question.", + "bbox": [ + 498, + 114, + 883, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ACTION 0:", + "text_level": 1, + "bbox": [ + 501, + 159, + 563, + 169 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sentence = 'A scene in the video features ...'", + "bbox": [ + 501, + 170, + 705, + 179 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames = progress_bar(frames)", + "bbox": [ + 501, + 179, + 640, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames, segment = highlight(frames, sentence, return_segment=True) \nframes = cut(frames, segment)", + "bbox": [ + 501, + 186, + 813, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for frame in frames:", + "bbox": [ + 501, + 203, + 598, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "display(frame)", + "bbox": [ + 521, + 210, + 589, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38551943a79adf478be4d4215b5ab39e4b0b5f0bdaf5641c47cafa4fcf1b62ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 830, + 184, + 888, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/214e3432346de37308894ffe776c07979fd129887fe4b6740753040cd8c5ad1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 224, + 589, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6982f63d0be24433d900774ff85e9813a1ec0a9fb1ef9485fcc6c499d7f4c8b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 224, + 681, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c013d4d9a6a3254278445b87f7e98f96f942882b5bcdb2696c57efa0aab0065c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 224, + 777, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/518e5f8b7721179ada3447eb2e4d69575395bb216c5e803186ba74fb2d7aab2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 224, + 870, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "THOUGHT 1: From the highlighted frames, I can observe the details of the motorcycle's features. The round object fits a motorcycle's speedmeter. ANSWER: (D) The motorcycle's speedometer. TERMINATE", + "bbox": [ + 501, + 272, + 851, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1", + "bbox": [ + 862, + 243, + 885, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal Counting", + "text_level": 1, + "bbox": [ + 200, + 366, + 346, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1de9e0c0058ed449ccd702978bce08ec7cf1f6a97f7dc63d38d78112056281c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 392, + 192, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df453bcaa0ab5e3aa19cae3c2f83433c6113c53d544fae8549ee3380e5613645.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 392, + 284, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18dfc0e110167516cdc5de5ff162d140495a817aa85cb850e3edb183aff782be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 392, + 380, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ffb3986c6a8f97c20d01950a0c6fbee82955d555f8126c594330f414fad0af0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 392, + 488, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da346dbd329ff689538753e3ce4c302cffd52f74392d2d930521ef62d9be286b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 446, + 130, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How many different kinds of animal faces are made in this video?", + "bbox": [ + 145, + 450, + 460, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A. 4", + "bbox": [ + 145, + 460, + 166, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B.", + "bbox": [ + 192, + 460, + 202, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "C. 5", + "bbox": [ + 236, + 460, + 254, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "D. 2", + "bbox": [ + 289, + 460, + 307, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eabaf48c67b154ee77f8be675dbd9b1c550e7c170631e9d938c32b8d2907b4a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 484, + 130, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT4o:A.4", + "bbox": [ + 145, + 489, + 214, + 501 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb86131d575065ddb3139aa3dac7063737b167d41e3e6bc12eaf1b279c1d6770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 489, + 264, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c31125daf2aa54b70df4b974c734d9988ad96b73e4146a19dd9469443ec3db44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 486, + 325, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VTimeCoT + GPT4o", + "bbox": [ + 331, + 489, + 462, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "THOUGHT 0: I can use the progress_bar and the highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for 'different kinds of animal faces' to highlight them. I will then display the highlighted frames to help me answer the question.", + "bbox": [ + 96, + 518, + 464, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ACTION 0:", + "text_level": 1, + "bbox": [ + 101, + 571, + 153, + 580 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/203b280ed43b686e591ad06c4982d8237173cc8b50cf16523140d0f9e6ed04db.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on VideoMME benchmark. Through reasoning that integrates the progress bar and highlights, VTimeCoT accurately answers questions related to temporal counting and order, which GPT-4o falls short." + ], + "image_footnote": [], + "bbox": [ + 161, + 571, + 173, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sentence = \"different kinds of animal faces\"", + "bbox": [ + 107, + 585, + 320, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames = progress_bar(frames)", + "bbox": [ + 107, + 594, + 250, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames = highlight(frame, sentence)", + "bbox": [ + 107, + 602, + 282, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for frame in frames:", + "bbox": [ + 107, + 611, + 205, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "display(frame)", + "bbox": [ + 127, + 619, + 196, + 626 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/515e82aaf7031981aab48ac2f627f6a38d95f68934020a48c60bfcf969457c70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 589, + 478, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ff557e762b51cfb1e13af427cdbb15ec9ededdce7eb296870671df08564ea13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 633, + 189, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a58c43ae90839f113e038e753c7154666c6fb7edbd745c752c2000e8a4efb59e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 633, + 284, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65cce358efdf273e18aed48cc53585089d5109f5036484dd5ca3bbf928c04b2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 633, + 379, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36486a1295d9999cb54b7498bd252ca2872436baa3de85f0612d84899d4e2b7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 635, + 483, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "THOUGHT 1: From the highlighted frames, I can count three different kinds of animal faces being made: a cat, a pig, and a dog. Therefore, the correct answer is 3. ANSWER: B TERMINATE", + "bbox": [ + 101, + 688, + 480, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal Order", + "text_level": 1, + "bbox": [ + 635, + 366, + 756, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb5677d6d17ea20abe0865f047f9db9b9942802b115989220fab96370409f755.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 392, + 596, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d521d0e788e868e4ddb8da0264314a1021785cbbc2776b73ddbb4d01dea6a54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 392, + 694, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f83eea998341443282a3c892c31d129e7d83b8327d6c1b7bb8879233e851c37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 392, + 787, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75014981f8ad94ff2f6d0fe17eef69cd5d95942b74c35221fa1225205a978c47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 392, + 887, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517e3b2316a670aa1d7f21a975b2cdfbc57d06ebc376ad92a7a696885b34851b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 446, + 535, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What was the exercise demonstrated after the squatting exercise in the video? A. Push-ups B. Plank C. Step-ups D. Pull-ups", + "bbox": [ + 550, + 450, + 874, + 470 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fcca75f1f70759257db5615476e7ae126eb4f11b0011694644299927fd2ca8c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 484, + 535, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT4o:B.Plank VTimeCoT+GPT4o", + "bbox": [ + 550, + 489, + 861, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "THOUGHT 0: I can use the progress_bar and highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for `exercise demonstrated after the squatting exercise` to highlight them. I will then display the highlighted frames to help me answer the question.", + "bbox": [ + 504, + 518, + 885, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ACTION 0:", + "text_level": 1, + "bbox": [ + 506, + 570, + 560, + 579 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61ebf001ece87ab30bbca97f8f045fddfa682cd5864d5cffb65f375c0ee910ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 562, + 569, + 581, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sentence = \"exercise demonstrated after the squatting exercise\" \nframes = progress_barimentary", + "bbox": [ + 514, + 585, + 820, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames = highlight(frame, sentence)", + "bbox": [ + 514, + 602, + 689, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for frame in frames:", + "bbox": [ + 514, + 611, + 612, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "display(frame)", + "bbox": [ + 535, + 619, + 602, + 626 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2550e143593791fc000b7e264c3fa10f255a6ca29d0d0836e24e5726a7ea912.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 590, + 885, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b44e7e7a84358bd05b66be4f345d45a3b49a69425e05bb4f2bf14a1a7096fd3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 632, + 598, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5879423196fdbe57056d48089100af2ac9729cc828027d6b7a4fef314778946e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 632, + 692, + 679 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6cae95118332a5a6f853a52ce6e9f13d0050cc593d4067af6b5057bbf239a574.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 632, + 787, + 679 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5864b1026c86667038a04f7baf29081d52fd852c61159ba3814473b0674fac8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 632, + 893, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "THOUGHT 1: The highlighted frames show the sequence of exercises demonstrated in the video. After the squatting exercise, the next", + "bbox": [ + 514, + 681, + 849, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstrated exercise is step-ups, as shown in the frames around the 21-second mark. ANSWER: C TERMINATE", + "bbox": [ + 514, + 704, + 874, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods struggle with discontinuous time segments, failing to identify multiple temporal spans and leading to degraded grounding performance.", + "bbox": [ + 89, + 796, + 482, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Long-Video Reasoning Based on Retrieving", + "text_level": 1, + "bbox": [ + 89, + 848, + 460, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 3, we evaluate the question-answering performance of the proposed and the baseline methods on the Vript-RR", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "benchmark, which contains questions that require scene retrieval. Given a scene hint and a question, the model must identify the specific scene to deliver an accurate response. We compared the proposed method with the state-of-the-art baseline methods including Qwen2VL-7B and GPT4o.", + "bbox": [ + 511, + 796, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be observed VTimeCoT achieves a significant per", + "bbox": [ + 514, + 871, + 903, + 887 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24209", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1675623bd2d99f74012aa7b15fb16437bdaa664035a15cb17b5f2b6f72765e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsFramesAccuracy
w/o subsw. subs
Gemini-1.5-Flash [40]1 fps70.375.0
GPT4o [17]38471.977.2
Gemini-1.5-Pro [40]1 fps75.081.3
Video-LLaVA-7B [26]839.941.6
VideoLLaMA2-7B [6]1654.956.4
LLAVA-Onevision-7B [20]3258.261.5
LongVILA-7B [49]25660.165.1
InternVL2.5-8B [5]6464.266.9
mPLUG-Owl3-7B [54]12859.368.1
LLaVA-Video-7B [58]6463.369.7
NVILA-8B [30]25664.270.0
VideoLLaMA3-7B [55]18066.270.3
GPT4o [17]3261.665.1
VTimeCoTgpt4o3264.273.6
", + "bbox": [ + 96, + 88, + 478, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance improvement, surpassing the baseline methods in both multiple-choice and open-ended settings.", + "bbox": [ + 89, + 416, + 482, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further illustrate the superiority of the proposed method in the task of scene retrieval and reasoning, we qualitatively compare GPT4o and VTimeCoT on the Vript-RR dataset in Fig. 3. Through multi-step reasoning that leverages the progress bar and highlights the key time interval, VTimeCoT accurately identifies the correct answer, while GPT-4o falls short.", + "bbox": [ + 89, + 446, + 483, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Long-Video Question Answering", + "text_level": 1, + "bbox": [ + 89, + 561, + 377, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A pivotal challenge in long-form video question answering lies in accurately inferring the frequency and sequential relationships of temporal events. Currently, end-to-end video MLLMs struggle to accurately localize events and capture their temporal dependencies across multiple frames, limiting their ability to generalize to real-world temporal understanding. In Tab. 4, we evaluate the performance of the proposed VTimeCoT model in question answering on long videos of the VideoMME benchmark dataset. Due to the prohibitive evaluation costs and to ensure a fair comparison, we re-evaluate GPT-4o using 32 frames. Utilizing robust temporal grounding and reasoning, VTimeCoT significantly outperforms the baseline methods in long-video question answering.", + "bbox": [ + 89, + 583, + 482, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effectiveness of the proposed method can be further validated in Fig. 4, where we compared the responses of the state-of-the-art GPT4o and VTimeCoT. In contrast to GPT4o, VTimeCoT can accurately identify the correct answers through its systematic, step-by-step reasoning process, even in challenging scenarios involving temporal counting and order discernment. It is also worth mention-", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/118e42a2a88bb8825b77813660eac0505360d2e073936225cac8add747023cce.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on the VideoMME dataset. We report the response accuracy under two settings: with subtitles and without subtitles." + ], + "table_footnote": [], + "table_body": "
MLLMCoTProgress BarHighlightQVHighlights (mIoU)Vript-RR (M-Acc)
GPT4o×××37.6670.39
GPT4o✓××41.8573.68
GPT4o✓✓×49.4076.32
GPT4o✓✓✓54.4983.55
", + "bbox": [ + 514, + 88, + 916, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablations on the key modules. M-Acc is the multi-choice accuracy on Vript-RR.", + "bbox": [ + 511, + 198, + 906, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing that VTimeCoT, apart from its superior performance in question-answering accuracy, demonstrates logical and interpretable reasoning steps.", + "bbox": [ + 511, + 253, + 906, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 511, + 309, + 624, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further investigate the contribution of each component in the proposed method, we conducted an ablation study. In Tab. 5, we evaluate the importance of the Chain-of-Thought (CoT), progress bar, and highlighting modules. As can be observed the GPT4o employing standard CoT without the use of the progress bar (i.e., solely on text-based step-by-step reasoning), results in significant performance degradation compared to the proposed method. To assess the effect of the proposed modules of the progress bar, we developed a configuration that utilizes only the progress bar tool without using the highlighting module. For completion, we also report the full VTimeCoT results, utilizing both the progress bar and highlighting tools. Both tools contribute to the performance improvement of the proposed method, which highlights the effectiveness of explicitly visualizing the precise timestamp and highlighting the relevant moments.", + "bbox": [ + 511, + 332, + 906, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose the first visual time framework, designed to formulate a visuotemporal chain-of-thought for video temporal grounding and reasoning. We introduce a plug-and-play tool of the progress bar to generate visual temporal cues from any video, that can enable multimodal large language models (MLLMs) to leverage tool-usage capabilities for accurately perceiving the speed of temporal progression. In addition, we propose a zero-shot temporal retrieval tool, built on top of a strong video-text foundational model, augmenting it with temporal grounding capabilities without any further training requirements. We integrate these two tools into a visuotemporal chain-of-thought framework to facilitate cross-modal reasoning between video and text. Through extensive experiments, we demonstrate that the proposed method surpasses previous state-of-the-art baselines in temporal grounding and reasoning-based question-answering benchmarks, demonstrating logical and interpretable reasoning steps.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24210", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 166, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. Gradio: Hassle-free sharing and testing of ml models in the wild. arXiv preprint arXiv:1906.02569, 2019. 5", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[3] Syeda Nahida Akter, Aman Madaan, Sangwu Lee, Yiming Yang, and Eric Nyberg. Self-imagine: Effective unimodal reasoning with multimodal models using self-imagination. arXiv preprint arXiv:2401.08025, 2024. 3, 5", + "[4] Mu Cai, Haotian Liu, Siva Karthik Mustikovela, Gregory P Meyer, Yuning Chai, Dennis Park, and Yong Jae Lee. Vipllava: Making large multimodal models understand arbitrary visual prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12914-12923, 2024. 3", + "[5] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1, 8", + "[6] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3, 8", + "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 3", + "[8] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 3", + "[9] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 2, 3", + "[10] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6", + "[11] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In" + ], + "bbox": [ + 93, + 191, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 5", + "[12] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 3, 4", + "[13] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3, 5", + "[14] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024. 3", + "[15] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 4, 5, 6", + "[16] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 4, 5", + "[17] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 5, 6, 8", + "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3", + "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 2, 6", + "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 8", + "[21] Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 3, 5", + "[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3", + "[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24211", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 6", + "[25] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024.", + "[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 3, 8", + "[27] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 2, 6", + "[28] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6", + "[29] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, et al. Llava-plus: Learning to use tools for creating multimodal agents. In European Conference on Computer Vision, pages 126-142. Springer, 2024. 3", + "[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 8", + "[31] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3, 6", + "[32] Sachit Menon, Richard Zemel, and Carl Vondrick. Whiteboard-of-thought: Thinking step-by-step across modalities. arXiv preprint arXiv:2406.14562, 2024. 3, 5", + "[33] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 2, 3", + "[34] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 12242-12254, 2025. 2", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman" + ], + "bbox": [ + 91, + 90, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2", + "[37] Chuyi Shang, Amos You, Sanjay Subramanian, Trevor Darrell, and Roei Herzig. Traveler: A modular multi-lmm agent framework for video question-answering. arXiv preprint arXiv:2404.01476, 2024. 2, 3", + "[38] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11987-11997, 2023. 3", + "[39] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 2, 3, 4", + "[40] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 8", + "[41] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 2", + "[42] Jiapeng Wang, Chengyu Wang, Kunzhe Huang, Jun Huang, and Lianwen Jin. Videoclip-xl: Advancing long description understanding for video clip models. arXiv preprint arXiv:2410.00741, 2024.5", + "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6", + "[44] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 2, 3", + "[45] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5", + "[46] Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. Mind's eye of llms: visualization-of-thought elicits spatial reasoning in large language models. Advances in Neural Information Processing Systems, 37:90277-90317, 2025. 3, 5", + "[47] Yixuan Wu, Yizhou Wang, Shixiang Tang, Wenhao Wu, Tong He, Wanli Ouyang, Philip Torr, and Jian Wu. Dettoolchain: A new prompting paradigm to unleash detection ability of mllm. In European Conference on Computer Vision, pages 164-182. Springer, 2024. 3", + "[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24212", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 8", + "[50] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 2, 6", + "[51] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3", + "[52] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3", + "[53] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In *Forty-first International Conference on Machine Learning*, 2024. 2, 3", + "[54] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8", + "[55] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 8", + "[56] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 3, 6", + "[57] Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandas Alexandros Potamias. Hawor: World-space hand motion reconstruction from egocentric videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1805-1815, 2025. 2", + "[58] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 8" + ], + "bbox": [ + 91, + 92, + 480, + 724 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24213", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_model.json b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4d922bbc070a812ce72c99056227a40b0c27e0b9 --- /dev/null +++ b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_model.json @@ -0,0 +1,3038 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "VTimeCoT: Thinking by Drawing for Video Temporal Grounding and Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.18, + 0.773, + 0.2 + ], + "angle": 0, + "content": "Jinglei Zhang\\(^{1,*}\\) Yuanfan Guo\\(^{2,*}\\) Rolandas Alexandros Potamias\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.2, + 0.667, + 0.217 + ], + "angle": 0, + "content": "Jiankang Deng\\(^{3}\\) Hang Xu\\(^{2}\\) Chao Ma\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.217, + 0.821, + 0.236 + ], + "angle": 0, + "content": "\\(^{1}\\) Shanghai Jiao Tong University \\(^{2}\\) Noah's Ark Lab \\(^{3}\\) Imperial College London" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.238, + 0.662, + 0.253 + ], + "angle": 0, + "content": "{zhangjinglei168,chaoma}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.255, + 0.812, + 0.27 + ], + "angle": 0, + "content": "{guoyuanfan1, xuHang}@huawei.com, {j.deng16, r.potamias}@imperial.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.305, + 0.481, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.369, + 0.348, + 0.38 + ], + "angle": 0, + "content": "How many streams did the cat cross in the video?" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.388, + 0.131, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.391, + 0.197, + 0.4 + ], + "angle": 0, + "content": "GPT4o:one" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.388, + 0.285, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.388, + 0.321, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.391, + 0.431, + 0.402 + ], + "angle": 0, + "content": "VTimeCoT + GPT40" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.415, + 0.446, + 0.435 + ], + "angle": 0, + "content": "THOUGHT 0: I can use the tool to draw progress bar and highlight the moments for 'streams the cat crossed'." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.443, + 0.162, + 0.453 + ], + "angle": 0, + "content": "ACTION 0:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.453, + 0.403, + 0.464 + ], + "angle": 0, + "content": "Call and " + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.472, + 0.128, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.47, + 0.233, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.471, + 0.355, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.471, + 0.471, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.524, + 0.451, + 0.544 + ], + "angle": 0, + "content": "THOUGHT 1: The tool is helpful. I see the cat crossing streams at 52s to 68s and 71s to 76s. So it crossed two streams." + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.305, + 0.901, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.359, + 0.901, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.42, + 0.901, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.574, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 1. We propose VTimeCoT, a Visual Time Chain-of-Thought framework for video temporal grounding and reasoning. VTimeCoT constructs cross-modality reasoning across both video and text, which enables the MLLM to utilize progress bar tools to annotate the time progression and highlight the key relevant segments to answer complex temporal questions." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.629, + 0.327, + 0.645 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.662, + 0.486, + 0.875 + ], + "angle": 0, + "content": "In recent years, video question answering based on multimodal large language models (MLLM) has garnered considerable attention, due to the benefits from the substantial advancements in LLMs. However, these models have a notable deficiency in the domains of video temporal grounding and reasoning, posing challenges to the development of effective real-world video understanding systems. Inspired by how humans use video players to interact with the progress bar for video comprehension, we introduce VTimeCoT, a simple yet effective training-free framework, designed for high-performance video grounding and reasoning. The proposed framework incorporates two novel visual tools of the progress bar: a plug-and-play progress bar integration tool and a high-efficiency highlighting tool. In addition," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.907, + 0.782 + ], + "angle": 0, + "content": "to address the limitations of conventional text-based chain-of-thought (CoT) approaches, we introduce a visuotemporal CoT process that integrates cross-modality reasoning across both video and text. Our approach demonstrates significant performance improvements on both Qwen2VL-7B and GPT4o baselines in tasks of video temporal grounding and reasoning-based question answering. Finally, we showcase that the proposed framework achieves a compositional and interpretable reasoning process. Project page: https://vtimecot.github.io." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.814, + 0.646, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Video understanding is a longstanding problem in computer vision and has attracted more attention with the emergence of large language models (LLM) [5, 17, 40, 43]. Video QA is a representative task reflecting the video reasoning abl-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.366, + 0.902 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24203" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "ity. Given that videos typically contain numerous events occurring at different time points, accurately answering questions is a highly challenging task. Recently, some end-to-end video understanding models based on multimodal large language models (MLLM) [20, 23, 26, 27, 56, 58] have been proposed to address video question-answering tasks, exhibiting remarkable capabilities. However, despite their ability to generate seemingly plausible results, these methods have been shown to exhibit a notable deficiency in temporal grounding [15] and fail to provide a temporal-grounded reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.485, + 0.595 + ], + "angle": 0, + "content": "With the advent of tool usage capabilities in LLMs, several agent-based video question-answering methods have been proposed [9, 33, 37, 44, 53], enabling step-by-step video understanding. These methods prompt the LLMs to sequentially reason and invoke external tool models (e.g., object detection and captioning models) to collect textual clues, ultimately inferring the final answer. This, so called, chain-of-thought approach offers significant potential, as it is compositional and training-free, demonstrating impressive zero-shot performance. However, such methods predominantly rely on language and captions as intermediaries, limiting their ability to directly capture the visuotemporal dynamics inherent in videos, especially in the case of long videos. Inspired by the time-progress bar commonly used in video players during human interactions, we identify it as an intuitive tool for grounding time and enhancing video comprehension. In contrast to written language, the progress bar, along with the concurrent visual content, directly conveys the concept of temporal progression. The video progress bar can represent temporal relationships through sequential positioning or more abstract temporal dependencies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a simple yet effective framework that empowers multimodal LLMs to utilize a video progress bar to construct a Visual Time Chain-of-Thought (VTimeCoT), facilitating long video temporal grounding and reasoning. Inspired by LLMs coupled with visual programming [12, 39], we propose enabling multimodal LLMs to generate code that overlays a progress bar on the video. First, we construct a frame-sync progress bar integration tool that can be invoked by the MLLM, which generates a progress bar displayed at the bottom of the video, marking the progress and annotating the timestamps on each frame. This is naturally adapted to long video and enables the MLLM to perceive the speed of temporal progress directly from the annotated video frames, adapting seamlessly to any frame-per-second (FPS) sampling rate. Second, to guide the model's attention on key relevant temporal segments, we propose a training-free and long-video-adaptive highlighting tool for the progress bar based on video-text similarity. Specifically, we leverage a robust video-text foundation model to identify the time intervals with the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.348 + ], + "angle": 0, + "content": "top-k highest similarity and highlight them on the visual progress bar. Finally, we prompt the multimodal LLM to invoke the progress bar tools within a visuotemporal chain-of-thought process that integrates text, program, and video annotated with the progress bar to facilitate long video reasoning. At each thought step, the model performs cross-modality reasoning and determines whether to dynamically update the video memory while progressively inferring the answer. As shown in Fig. 1, to determine how many times a cat crosses a stream in the input video, the model first invokes tools to generate a progress bar, identifying the key timestamps corresponding to 'streams the cat crossed'. By analyzing the progress bar annotated frames, the model infers the exact timing of each crossing event and accurately determines the number of crossings. This approach enhances the model's temporal reasoning capabilities by seamlessly integrating temporal-grounded visual evidence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.905, + 0.558 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of the proposed framework across a wide range of video temporal grounding and video-QA tasks. The proposed method consistently outperforms the baseline models, Qwen2VL and GPT-4o in temporal grounding tasks, achieving significant performance gains with an average IoU improvement of \\(6.58\\%\\) and \\(16.83\\%\\) over GPT-4o across Charades-STA [11] and QVHighlights [19] benchmarks. Similarly, VTimeCoT significantly improves the accuracy of questions related to temporal retrieval, event counting, and event ordering in reasoning-based question-answering tasks. Specifically, the proposed method consistently surpasses state-of-the-art methods across Vript-RR [50] and VideoMME [10] benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.905, + 0.605 + ], + "angle": 0, + "content": "To sum up, we present VTimeCoT, a training-free model for temporal reasoning based on the visual progress bar. Specifically:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.904, + 0.697 + ], + "angle": 0, + "content": "- We propose the first, to the best of our knowledge, visual time chain-of-thought framework for video temporal grounding and reasoning. In contrast to previous methods that rely solely on textual reasoning, we leverage the visual progress bar as a medium, making a significant step towards real-world video understanding systems." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.698, + 0.905, + 0.758 + ], + "angle": 0, + "content": "- The proposed reasoning framework leverages two novel tools from video progress bar integration and highlighting, enabling the MLLM to accurately perceive the times-tamp of each frame and identify key temporal segments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.759, + 0.905, + 0.818 + ], + "angle": 0, + "content": "- The proposed method enhances the performance of MLLMs by a large margin in video temporal grounding and reasoning tasks through a training-free approach, while also demonstrating strong reasoning capabilities." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.607, + 0.905, + 0.818 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.655, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Multimodal LLM for Video. In contrast with traditional video understanding such as object detection, segmentation and pose estimation [34, 36, 41, 57], recent mul" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24204" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.351 + ], + "angle": 0, + "content": "timodal LLMs enable understanding through natural language. End-to-end video multimodal LLMs are trained on large-scale datasets, to align visual inputs with the language modality and integrate with large language models. Video-ChatGPT [31] leveraged CLIP [35] to extract spatial and temporal features from videos and integrate them with Vicuna [7]. Video-LLaMA [56] proposed a video Q-Former designed to enhance the modeling of temporal variations. Subsequently, a series of video multimodal LLM variants [6, 20, 22, 24–26, 48] have introduced advancements in enhancing video representation, expanding data scale, and refining training methods, leading to improved video question-answering performance. However, these end-to-end models operate as black boxes, directly generating answers without an explicit video reasoning process. In contrast, we propose a video understanding framework that incorporates a structured reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.368, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Programming and Tool Usage of LLM. Due to the strong code-generation capabilities of LLMs, several works [8, 12, 14, 29, 39, 52] attempted to prompt LLMs to call compositional visual tools to address visual question-answering tasks. VisProg [12] and ViperGPT [39] defined a variety of image foundation functions (e.g., object detection) and prompted the LLMs to generate code that invokes these functions for answering questions. MM-REACT [52] tackled visual tasks through multi-step reasoning guided by text prompts to the LLM, where each step involves thought and calling visual tools. More recently, several studies [9, 33, 37, 44, 53] proposed tool-usage frameworks tailored for video understanding. By designing specialized video tools and calling steps, these approaches extract textual cues from videos to facilitate question answering. In contrast to the aforementioned works that solely rely on text as the reasoning medium, we introduce a visual progress bar, enabling more effective integration of visual and temporal cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Visual Prompt for Multimodal LLM. LLMs have been shown to enhance their understanding capabilities by incorporating overlaid visual prompts on images (e.g., circles, keypoints) [3, 4, 13, 21, 32, 38, 46, 47, 51]. SoM [51] proposed to utilize SAM [18] segmentation model to overlay mask colors and labels on images as input of GPT-4V [2], demonstrating strong zero-shot performance on vision-grounding tasks. Hu et al. [13] introduced a chain-of-thought framework combined with visual prompts to draw markers on images during reasoning, effectively enhancing the image understanding performance of MLLMs. However, all previous methods are limited to spatial prompts, limiting their understanding capabilities. In this work, we propose to design the progress bar as visual-temporal prompts to facilitate temporal reasoning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.252 + ], + "angle": 0, + "content": "We propose VTimeCoT, a general framework that empowers multimodal large language models using a visual progress bar as an intermediate reasoning step, facilitating temporal grounding and reasoning. Fig. 2 illustrates how our approach works. Given a raw video and a question, our method generates a chain-of-thought and invokes a set of tools for progress bar integration. By utilizing the visual cues from the progress bar as a medium, the model progressively infers the final answer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.26, + 0.808, + 0.277 + ], + "angle": 0, + "content": "3.1. Visuotemporal Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.907, + 0.418 + ], + "angle": 0, + "content": "In video understanding, particularly for long videos, existing MLLM methods can only provide static answers, which limits the model's ability to focus on critical scenes and results in significant shortcomings in real-world settings. Since the MLLMs fail to identify key scenes within a single forward pass, it is necessary to interact with the video to progressively infer the answer. Therefore, we propose a framework based on interactive mechanisms and dynamic video memory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Our framework tackles video grounding and reasoning tasks through an iterative interaction with the MLLM, built upon a tool set leveraging the progress bar. Given a textual question and a video as input, VTimeCoT generates a sequence of thoughts and actions to dynamically update the video memory. By manipulating the video, it acquires the necessary information to answer the question. In this process, the MLLM reasons by plotting and analyzing the progress bar, seamlessly integrating both textual and visual reasoning into a Visuotemporal Chain-of-Thought. As illustrated in Fig. 2, VTimeCoT first provides the MLLM with a formatted prompt for initialization and executes actions at each time step \\( t \\). The tool set includes the , , , etc. The tool is utilized to trim specific segments when the MLLM determines that the video is too long to find the required information. The pseudo-code is presented in Algorithm 1, where \\( \\mathbb{V} \\) and \\( \\mathbb{L} \\) represent the modalities of video and language, respectively. In Algorithm 1, \\( p \\) serves as a parser to retrieve specific sub-strings by keyword, and \\( c \\) is a code builder that integrates the toolset library and generated code." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.871 + ], + "angle": 0, + "content": "Initialization Prompt. To enable the MLLM to perform step-by-step reasoning and a structured output, we follow [13] and construct a set of initialization prompts. These prompts define the specific output structure (formatted by keywords of 'THOUGHT', 'ACTION', and 'TERMINATE') for each step and declare the Python tools that the model should invoke. The initialization prompts, along with the question and video frames, are then fed into the MLLM, triggering the iterative loop." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Thought. During this step, the model analyzes the historical context and video memory to generate its reasoning" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24205" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.372, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Overview of our method. On the left, we demonstrate how the framework iteratively generates thoughts and actions, which dynamically updates the video memory with an overlaid progress bar for reasoning. On the right, we illustrate two novel tools that integrate the frame-sync visual progress bar and highlight key moments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.483, + 0.486 + ], + "angle": 0, + "content": "for the current step. It determines whether a direct answer can be provided directly in this step or if tool assistance is needed, and if so, which tools should be invoked." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Action. Following the thought step, the model generates an executable Python script utilizing the given tool functions. The framework then builds the code with the toolset library and executes the script to manipulate the video frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.624 + ], + "angle": 0, + "content": "Dynamic Video Memory. Using the manipulated video frames we subsequently update the frames in the video memory. The updated context with the manipulated video frames is then fed into the MLLM to proceed with the next step of thought." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.701 + ], + "angle": 0, + "content": "To terminate the loop, the model determines at each step whether to cease reasoning and generate the 'TERMINATE' keyword to end. To prevent excessively prolonged reasoning, we impose a maximum step limit, enforcing termination at step \\( T \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.47, + 0.729 + ], + "angle": 0, + "content": "3.2. Frame-Sync Visual Progress Bar Integration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Enabling MLLM to perceive the precise timestamp of each frame is a significant challenge and an unresolved problem, due to the diverse temporal sampling rates during MLLM training and inference without time. Although recent approaches [15, 16] have attempted to address this issue by incorporating temporal positional embeddings before the video encoder and introducing temporal-grounding training tasks, their flexibility is limited due to the scarcity of specific data and the need for fine-tuning. This issue is particularly pronounced in long videos, where the token limitations of MLLMs prevent them from accurately perceiving" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.439, + 0.907, + 0.691 + ], + "angle": 0, + "content": "
Algorithm 1 Visuotemporal Chain-of-Thought
Input: {V, Q||V ∈ V, Q ∈ L}; I ∈ L;MLLM: V, L → L; Toolset: {T1, . . ., Tn} ∈ L
1: s ← 0; V0 ← V
2: H0 ← {I, Q}▷ Initialization prompt and query
3: while true do
4: ys ← MLLM (Vs, Hs)
5: THOUGHTs, ACTIONs, TERMINATEs ← p(ys)
6: if not TERMINATEs then
7: Cs ← c(ACTIONs, Toolset)▷ Build code
8: Vs+1 ← Cs(Vs)▷ Update video
9: Hs+1 ← {Hs, ys}▷ Update history
10: else
11: return ys▷ Respond final answer
12: end if
13: end while
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.719, + 0.708, + 0.733 + ], + "angle": 0, + "content": "the timestamp of each frame." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Inspired by visual programming paradigms [12, 39], we propose a simple yet effective method to empower MLLMs to perceive the precise timestamp of each frame, by generating the video progress bar. Leveraging the code generation capabilities of MLLMs, our approach prompts the MLLM to invoke Python plotting tools and use their universal OCR and shape understanding capabilities to perceive time, without any additional training requirements. Specifically, given a raw video \\( V \\in \\mathbb{R}^{T \\times H \\times W \\times 3} \\) as input, this tool can generate the progress bar at the bottom of the original frames and return the annotated frames. To enable adaptation to ar" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24206" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "binary FPS sampling, we perform a frame-synchronization step to convert frame indexes to the real second time, as shown in Fig. 2. To enable MLLM to perceive seconds directly without any additional hour-minute-second conversions, we design the timestamps using a straightforward seconds format. To construct a robust visual prompt that can be easily interpreted from the MLLM, we employ compositional components using elongated rectangles, circles, and timestamps to plot the progress bar. Finally, the progress bar is placed at the bottom of raw frames, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.258, + 0.483, + 0.275 + ], + "angle": 0, + "content": "\\[\nV _ {t} ^ {\\prime} = V _ {t} \\oplus T _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.359 + ], + "angle": 0, + "content": "where \\(\\oplus\\) represents the vertical concatenation operation, \\(T_{t}\\) denotes the image of the progress bar generated by OpenCV at time \\(t\\) and \\(V_{t}^{\\prime}\\) is the resulting frame. We wrap this integration process into a Python tool function, allowing MLLM to invoke it for generating the annotated video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.369, + 0.462, + 0.386 + ], + "angle": 0, + "content": "3.3. Moment Highlight by Video-Text Similarity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.483, + 0.557 + ], + "angle": 0, + "content": "Although multimodal LLMs have made significant progress in video content captioning, recent studies [15, 16] have revealed substantial deficiencies in their temporal grounding capabilities. Existing works [15, 16] rely on constructing temporal grounding instructional data and require finetuning to emerge such capability. Besides, due to token limitations in MLLMs, they struggle to perform temporal grounding for long videos. In contrast, we propose a training-free and long-video-adaptive method to enhance the temporal grounding capabilities of MLLMs, leveraging the similarities from the robust video-text foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.619 + ], + "angle": 0, + "content": "Given a query and a video \\(V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}\\) containing multiple events, our proposed tool estimates the query-relevant temporal segments, including their start and end times, and generates a highlighted visual prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Moment Retrieval by Video-Text Similarity. To construct a robust zero-shot moment retrieval module, capable of generalizing across a wide range of videos, we employ the VideoCLIP-XL [42] foundation model as our embedding extractor. The video is initially processed at a high temporal resolution of \\( r \\) FPS. We group every 8 frames into a video clip, resulting in \\( N \\) clips. It should be noted that since the temporal retrieval module operates as an external component, it can efficiently handle high frame rates without being constrained by the token count of the MLLM. This makes our method adaptive to long-form video. These clips are then processed by the VideoCLIP-XL visual encoder, producing \\( N \\) video embeddings. We use VideoCLIP-XL text encoder to extract the text embedding for the text query and we compute the cosine similarity between the query embedding and the video embeddings of each clip:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.872, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} (i) = \\frac {e _ {i} ^ {v} \\cdot e ^ {q}}{\\left| e _ {i} ^ {v} \\right| \\left| e ^ {q} \\right|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "where \\( e_i^v \\) denotes the embedding of the \\( i \\)-th clip and \\( e^q \\) is the query embedding. Using top-\\( k \\) selection, we obtain the \\( k \\) clips with the highest similarity and extract the start and end timestamps of each contiguous segment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.364 + ], + "angle": 0, + "content": "Moment Highlight on the Progress Bar. Motivated by a series of works that demonstrated the effectiveness of visual cues as a medium for reasoning beyond text [3, 13, 21, 32, 46], we propose using moment-highlighting cues for video temporal reasoning. As shown in Fig. 2, we construct the highlight tool using OpenCV library to plot colorful highlight masks on specific intervals of the progress bar. This tool takes the video frames and time intervals as input and annotates the highlighted progress bar under the frames. To ensure that the MLLM can easily interpret the highlights, we utilize different colors between the mask and the progress bar. We wrap the aforementioned retrieval and highlighting process as a Python tool function, enabling the MLLM to invoke it to manipulate the input video frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.377, + 0.646, + 0.394 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.905, + 0.569 + ], + "angle": 0, + "content": "In this section, we first present the implementation details, datasets and evaluation metrics. Subsequently, we assess the performance of the proposed framework on the video temporal grounding benchmarks (Sec. 4.1). To further evaluate the reasoning capabilities, we conducted extensive quantitative and qualitative analyses on the \"reasoning based on retrieving\" benchmark of Vript-RR (Sec. 4.2). In addition, we demonstrate the significant role our reasoning framework plays in enhancing long video question-answering performance (Sec. 4.3). Finally, we give the ablations of key modules (Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Implementation Details. In the framework, we conducted experiments on two core MLLMs, Qwen2VL-7B [43] and GPT4o-20240513 [17], as they represent the state-of-the-art open-source and closed-source MLLM models, respectively. We set the decoding temperature of the MLLM to 0. By default, the input for the MLLM consists of 32 frames uniformly sampled from the video. For video frames, \\( H \\) and \\( W \\) are resized to make the longer side 480 pixels. Our agent implementation is based on AutoGen [45]. The implementations of tool functions are adapted from Gradio [1]. For GPT4o, we utilize the OpenAI API service. The details of the MLLM prompts are provided in the supplementary materials. For VideoCLIP-XL we use FPS \\( r = 1 \\) to sample video frames and group every 8 frames as a clip to extract embedding. Regarding the top-\\( k \\) selection of clips, \\( k \\) is empirically set to 8. The maximum number of reasoning steps \\( T \\) is set to 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.886 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. To evaluate the proposed VTimeCoT method, we utilized four benchmark datasets, spanning a wide range of video temporal grounding, reasoning, and long-video question-answering tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Charades-STA [11] is a benchmark dataset for temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24207" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.473, + 0.272 + ], + "angle": 0, + "content": "
MethodR1@0.3R1@0.5R1@0.7mIoU
VideoChat-7B [23]9.003.301.306.50
VideoLLaMA-7B [56]10.403.800.907.10
VideoChatGPT-7B [31]20.007.701.7013.70
LLAVA-Onevision-7B [20]33.0411.054.1120.98
VTimeLLM-7B [15]51.0027.5011.4031.20
VTimeLLM-13B [15]55.3034.3014.7034.60
Qwen2VL-7B [43]37.3112.854.1124.34
VTimeCoTQwen2VL-7B66.9638.7920.8343.41
GPT4o [17]63.7637.1214.6540.20
VTimeCoTGPT4o74.0651.0222.4546.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.325 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on the Charades-STA dataset for temporal grounding. We report the mIoU and the recall performance at different IoU thresholds." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.34, + 0.473, + 0.468 + ], + "angle": 0, + "content": "
MethodR1@0.3R1@0.5R1@0.7mIoU
LLAVA-Onevision-7B [20]34.9117.919.725.53
VTimeLLM-7B [15]44.5825.039.2928.99
Qwen2VL-7B [43]31.8714.657.3522.77
VTimeCoTQwen2VL-7B67.5045.7925.1146.21
GPT4o [17]55.6135.6819.2937.66
VTimeCoTGPT4o79.3559.7433.8154.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.521 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on the QVHighlights dataset for temporal grounding of discontinuous segments. We report the mIoU and the recall performance at different IoU thresholds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.655 + ], + "angle": 0, + "content": "grounding, composed of 1334 videos along with the corresponding start-end frame annotations of 3720 queries. The average video length is 30 seconds. We follow [15] and report mean IoU (mIoU) and recall@1, IoU \\(\\geq m(R@m)\\) metrics, where \\(m\\) includes 0.3, 0.5 and 0.7. IoU denotes the intersection over the union between the predicted and ground truth time segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.762 + ], + "angle": 0, + "content": "QVHighlights [19] is a benchmark dataset for semantic relevance and temporal grounding from discontinuous time segments. It contains 1519 videos, with an average length of 150 seconds, and 1550 queries with time annotations, where each query spans multiple discontinuous time segments within a video. For evaluation, we report the commonly used mIoU and R@m metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Vript-RR [50] is a challenging benchmark dataset for scene retrieval and multi-hop reasoning. The average length of the videos is 622 seconds. It contains 152 questions, each accompanied by a hint to locate the scene that the question refers to. The benchmark includes both multiple-choice and open-ended question-answering settings. We follow [50] and evaluate the open-ended accuracy using GPT4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "VideoMME [10] is a video question-answering benchmark tailored for MLLMs, featuring diverse videos ranging from" + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.089, + 0.861, + 0.285 + ], + "angle": 0, + "content": "
MethodsMulti-ChoiceOpen
VideoChatGPT [31]29.6017.80
Video-LLaMA [56]28.3014.50
VideoChat [23]22.4015.10
VideoChat2 [24]42.1022.40
ST-LLM [28]33.6026.30
PLLAVA 7B [48]55.3036.20
VILA-1.5 8B [27]55.3032.30
Qwen2VL-7B [43]59.8735.95
VTimeCoTQwen2VL-7B62.5041.45
GPT4o [17]70.3961.18
VTimeCoTGPT4o83.5568.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.295, + 0.907, + 0.339 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on the Vript-RR dataset. We report the response accuracy on both multiple-choice and open-ended question settings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.907, + 0.427 + ], + "angle": 0, + "content": "11 seconds to 1 hour in length. It contains 900 videos and 2700 question-answer pairs. We follow [10] to report accuracy metrics under two evaluation settings with subtitles and without subtitles." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.715, + 0.456 + ], + "angle": 0, + "content": "4.1. Temporal Grounding" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.907, + 0.765 + ], + "angle": 0, + "content": "To ensure accurate temporal reasoning, it is essential to achieve robust and high-performance temporal grounding. In Tab. 1, we compare VTimeCoT with two state-of-the-art baselines, Qwen2VL-7B and GPT4o, to identify the start and end timestamps of the query event. Specifically, using both Qwen2VL-7B and GPT4o models as the core MLLM of our framework, we evaluate the performance of VTimeCoT to localize the query event. As can be easily observed, although these advanced MLLMs achieve state-of-the-art performance in video content recognition, they fall short in accurately grounding temporal boundaries for events in videos, facing difficulties in perceiving the precise timestamp and the boundaries of events. In contrast, the proposed method effectively addresses these challenges, providing accurate visual cues of the progress bar through the reasoning process that facilitates the grounding performance in videos. Note that VTimeCoT demonstrates superior performance without any additional training cost, even compared to the VTimeLLM model, which has been meticulously fine-tuned on extensive temporal grounding data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To further evaluate the temporal grounding performance of the proposed and the baseline methods in the presence of multiple discontinuous time segments for each query, we utilized the QVHighlights benchmark dataset (Tab. 2). Given a query and a video, we request from each model to identify start-end timestamps of all temporal segments in the video that fit the query. As can be observed, VTime-CoT consistently enhances the performance compared to different MLLMs. On the contrary, baseline MLLM meth" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24208" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.167, + 0.1, + 0.411, + 0.113 + ], + "angle": 0, + "content": "Reasoning based on retrieving scenes" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.122, + 0.482, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.182, + 0.441, + 0.216 + ], + "angle": 0, + "content": "? Hint: A scene in the video features a first-person perspective from a motorcyclist cruising on a highway. ... The motorcycle's chrome details, black leather accessories, the rider's gloved hands are noticeable." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.224, + 0.448, + 0.234 + ], + "angle": 0, + "content": "What is the round object on the bike in the middle bottom of the screen?" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.235, + 0.185, + 0.243 + ], + "angle": 0, + "content": "(A) A GPS device" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.235, + 0.369, + 0.244 + ], + "angle": 0, + "content": "(B) A sport camera" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.245, + 0.244, + 0.255 + ], + "angle": 0, + "content": "(C) The motorcycle's headlight" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.245, + 0.434, + 0.255 + ], + "angle": 0, + "content": "(D) The motorcycle's spedometer" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.272, + 0.13, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.28, + 0.329, + 0.293 + ], + "angle": 0, + "content": "GPT40: (C) The motorcycle's headlight." + }, + { + "type": "image", + "bbox": [ + 0.448, + 0.276, + 0.466, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.094, + 0.532, + 0.112 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.096, + 0.662, + 0.106 + ], + "angle": 0, + "content": "VTimeCoT + GPT40" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.884, + 0.154 + ], + "angle": 0, + "content": "THOUGHT0: I can use the progress_bar and highlight tool to highlight the key moments. Since the video is too long, I will cut the video into segments for a thorough review. Then I will display the highlighted frames to help me answer the question." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.16, + 0.564, + 0.17 + ], + "angle": 0, + "content": "ACTION 0:" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.171, + 0.707, + 0.18 + ], + "angle": 0, + "content": "sentence = 'A scene in the video features ...'" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.18, + 0.641, + 0.188 + ], + "angle": 0, + "content": "frames = progress_bar(frames)" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.188, + 0.815, + 0.204 + ], + "angle": 0, + "content": "frames, segment = highlight(frames, sentence, return_segment=True) \nframes = cut(frames, segment)" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.204, + 0.599, + 0.212 + ], + "angle": 0, + "content": "for frame in frames:" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.212, + 0.59, + 0.22 + ], + "angle": 0, + "content": "display(frame)" + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.185, + 0.89, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.226, + 0.59, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.226, + 0.683, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.226, + 0.778, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.226, + 0.871, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.273, + 0.852, + 0.302 + ], + "angle": 0, + "content": "THOUGHT 1: From the highlighted frames, I can observe the details of the motorcycle's features. The round object fits a motorcycle's speedmeter. ANSWER: (D) The motorcycle's speedometer. TERMINATE" + }, + { + "type": "text", + "bbox": [ + 0.864, + 0.244, + 0.886, + 0.25 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.907, + 0.347 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on Vript-RR benchmark. In contrast to GPT-4o which is misled by the irrelevant scenes, VTimeCoT reasons using the progress bar and highlights the key moments, leading to accurate answers." + }, + { + "type": "title", + "bbox": [ + 0.201, + 0.367, + 0.347, + 0.381 + ], + "angle": 0, + "content": "Temporal Counting" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.393, + 0.194, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.393, + 0.285, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.393, + 0.382, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.393, + 0.489, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.447, + 0.131, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.451, + 0.462, + 0.461 + ], + "angle": 0, + "content": "How many different kinds of animal faces are made in this video?" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.462, + 0.167, + 0.471 + ], + "angle": 0, + "content": "A. 4" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.462, + 0.204, + 0.47 + ], + "angle": 0, + "content": "B." + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.462, + 0.255, + 0.47 + ], + "angle": 0, + "content": "C. 5" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.462, + 0.308, + 0.47 + ], + "angle": 0, + "content": "D. 2" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.485, + 0.131, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.491, + 0.215, + 0.502 + ], + "angle": 0, + "content": "GPT4o:A.4" + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.49, + 0.266, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.487, + 0.326, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.49, + 0.463, + 0.502 + ], + "angle": 0, + "content": "VTimeCoT + GPT4o" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.52, + 0.465, + 0.559 + ], + "angle": 0, + "content": "THOUGHT 0: I can use the progress_bar and the highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for 'different kinds of animal faces' to highlight them. I will then display the highlighted frames to help me answer the question." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.572, + 0.154, + 0.581 + ], + "angle": 0, + "content": "ACTION 0:" + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.572, + 0.174, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.586, + 0.321, + 0.594 + ], + "angle": 0, + "content": "sentence = \"different kinds of animal faces\"" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.595, + 0.25, + 0.603 + ], + "angle": 0, + "content": "frames = progress_bar(frames)" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.603, + 0.283, + 0.611 + ], + "angle": 0, + "content": "frames = highlight(frame, sentence)" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.612, + 0.207, + 0.619 + ], + "angle": 0, + "content": "for frame in frames:" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.62, + 0.197, + 0.627 + ], + "angle": 0, + "content": "display(frame)" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.59, + 0.48, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.635, + 0.191, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.635, + 0.285, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.635, + 0.38, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.636, + 0.485, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.689, + 0.481, + 0.719 + ], + "angle": 0, + "content": "THOUGHT 1: From the highlighted frames, I can count three different kinds of animal faces being made: a cat, a pig, and a dog. Therefore, the correct answer is 3. ANSWER: B TERMINATE" + }, + { + "type": "title", + "bbox": [ + 0.636, + 0.367, + 0.757, + 0.381 + ], + "angle": 0, + "content": "Temporal Order" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.393, + 0.598, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.393, + 0.695, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.393, + 0.788, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.393, + 0.888, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.448, + 0.537, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.551, + 0.451, + 0.875, + 0.471 + ], + "angle": 0, + "content": "What was the exercise demonstrated after the squatting exercise in the video? A. Push-ups B. Plank C. Step-ups D. Pull-ups" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.485, + 0.537, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.552, + 0.491, + 0.862, + 0.507 + ], + "angle": 0, + "content": "GPT4o:B.Plank VTimeCoT+GPT4o" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.519, + 0.886, + 0.559 + ], + "angle": 0, + "content": "THOUGHT 0: I can use the progress_bar and highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for `exercise demonstrated after the squatting exercise` to highlight them. I will then display the highlighted frames to help me answer the question." + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.571, + 0.561, + 0.58 + ], + "angle": 0, + "content": "ACTION 0:" + }, + { + "type": "image", + "bbox": [ + 0.563, + 0.57, + 0.582, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.586, + 0.821, + 0.603 + ], + "angle": 0, + "content": "sentence = \"exercise demonstrated after the squatting exercise\" \nframes = progress_barimentary" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.603, + 0.69, + 0.611 + ], + "angle": 0, + "content": "frames = highlight(frame, sentence)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.612, + 0.613, + 0.619 + ], + "angle": 0, + "content": "for frame in frames:" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.62, + 0.603, + 0.627 + ], + "angle": 0, + "content": "display(frame)" + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.591, + 0.887, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.633, + 0.599, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.633, + 0.694, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.633, + 0.788, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.633, + 0.894, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.683, + 0.85, + 0.704 + ], + "angle": 0, + "content": "THOUGHT 1: The highlighted frames show the sequence of exercises demonstrated in the video. After the squatting exercise, the next" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.705, + 0.875, + 0.722 + ], + "angle": 0, + "content": "demonstrated exercise is step-ups, as shown in the frames around the 21-second mark. ANSWER: C TERMINATE" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.742, + 0.907, + 0.77 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on VideoMME benchmark. Through reasoning that integrates the progress bar and highlights, VTimeCoT accurately answers questions related to temporal counting and order, which GPT-4o falls short." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.483, + 0.843 + ], + "angle": 0, + "content": "ods struggle with discontinuous time segments, failing to identify multiple temporal spans and leading to degraded grounding performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.849, + 0.462, + 0.866 + ], + "angle": 0, + "content": "4.2. Long-Video Reasoning Based on Retrieving" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Tab. 3, we evaluate the question-answering performance of the proposed and the baseline methods on the Vript-RR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.907, + 0.871 + ], + "angle": 0, + "content": "benchmark, which contains questions that require scene retrieval. Given a scene hint and a question, the model must identify the specific scene to deliver an accurate response. We compared the proposed method with the state-of-the-art baseline methods including Qwen2VL-7B and GPT4o." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.872, + 0.905, + 0.888 + ], + "angle": 0, + "content": "As can be observed VTimeCoT achieves a significant per" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24209" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.339 + ], + "angle": 0, + "content": "
MethodsFramesAccuracy
w/o subsw. subs
Gemini-1.5-Flash [40]1 fps70.375.0
GPT4o [17]38471.977.2
Gemini-1.5-Pro [40]1 fps75.081.3
Video-LLaVA-7B [26]839.941.6
VideoLLaMA2-7B [6]1654.956.4
LLAVA-Onevision-7B [20]3258.261.5
LongVILA-7B [49]25660.165.1
InternVL2.5-8B [5]6464.266.9
mPLUG-Owl3-7B [54]12859.368.1
LLaVA-Video-7B [58]6463.369.7
NVILA-8B [30]25664.270.0
VideoLLaMA3-7B [55]18066.270.3
GPT4o [17]3261.665.1
VTimeCoTgpt4o3264.273.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on the VideoMME dataset. We report the response accuracy under two settings: with subtitles and without subtitles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.483, + 0.448 + ], + "angle": 0, + "content": "formance improvement, surpassing the baseline methods in both multiple-choice and open-ended settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.553 + ], + "angle": 0, + "content": "To further illustrate the superiority of the proposed method in the task of scene retrieval and reasoning, we qualitatively compare GPT4o and VTimeCoT on the Vript-RR dataset in Fig. 3. Through multi-step reasoning that leverages the progress bar and highlights the key time interval, VTimeCoT accurately identifies the correct answer, while GPT-4o falls short." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.379, + 0.579 + ], + "angle": 0, + "content": "4.3. Long-Video Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.795 + ], + "angle": 0, + "content": "A pivotal challenge in long-form video question answering lies in accurately inferring the frequency and sequential relationships of temporal events. Currently, end-to-end video MLLMs struggle to accurately localize events and capture their temporal dependencies across multiple frames, limiting their ability to generalize to real-world temporal understanding. In Tab. 4, we evaluate the performance of the proposed VTimeCoT model in question answering on long videos of the VideoMME benchmark dataset. Due to the prohibitive evaluation costs and to ensure a fair comparison, we re-evaluate GPT-4o using 32 frames. Utilizing robust temporal grounding and reasoning, VTimeCoT significantly outperforms the baseline methods in long-video question answering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The effectiveness of the proposed method can be further validated in Fig. 4, where we compared the responses of the state-of-the-art GPT4o and VTimeCoT. In contrast to GPT4o, VTimeCoT can accurately identify the correct answers through its systematic, step-by-step reasoning process, even in challenging scenarios involving temporal counting and order discernment. It is also worth mention-" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.918, + 0.189 + ], + "angle": 0, + "content": "
MLLMCoTProgress BarHighlightQVHighlights (mIoU)Vript-RR (M-Acc)
GPT4o×××37.6670.39
GPT4o✓××41.8573.68
GPT4o✓✓×49.4076.32
GPT4o✓✓✓54.4983.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.199, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Table 5. Ablations on the key modules. M-Acc is the multi-choice accuracy on Vript-RR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.301 + ], + "angle": 0, + "content": "ing that VTimeCoT, apart from its superior performance in question-answering accuracy, demonstrates logical and interpretable reasoning steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.31, + 0.625, + 0.325 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.589 + ], + "angle": 0, + "content": "To further investigate the contribution of each component in the proposed method, we conducted an ablation study. In Tab. 5, we evaluate the importance of the Chain-of-Thought (CoT), progress bar, and highlighting modules. As can be observed the GPT4o employing standard CoT without the use of the progress bar (i.e., solely on text-based step-by-step reasoning), results in significant performance degradation compared to the proposed method. To assess the effect of the proposed modules of the progress bar, we developed a configuration that utilizes only the progress bar tool without using the highlighting module. For completion, we also report the full VTimeCoT results, utilizing both the progress bar and highlighting tools. Both tools contribute to the performance improvement of the proposed method, which highlights the effectiveness of explicitly visualizing the precise timestamp and highlighting the relevant moments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.619 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose the first visual time framework, designed to formulate a visuotemporal chain-of-thought for video temporal grounding and reasoning. We introduce a plug-and-play tool of the progress bar to generate visual temporal cues from any video, that can enable multimodal large language models (MLLMs) to leverage tool-usage capabilities for accurately perceiving the speed of temporal progression. In addition, we propose a zero-shot temporal retrieval tool, built on top of a strong video-text foundational model, augmenting it with temporal grounding capabilities without any further training requirements. We integrate these two tools into a visuotemporal chain-of-thought framework to facilitate cross-modal reasoning between video and text. Through extensive experiments, we demonstrate that the proposed method surpasses previous state-of-the-art baselines in temporal grounding and reasoning-based question-answering benchmarks, demonstrating logical and interpretable reasoning steps." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24210" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.167, + 0.188, + 0.183 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[1] Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. Gradio: Hassle-free sharing and testing of ml models in the wild. arXiv preprint arXiv:1906.02569, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.249, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[3] Syeda Nahida Akter, Aman Madaan, Sangwu Lee, Yiming Yang, and Eric Nyberg. Self-imagine: Effective unimodal reasoning with multimodal models using self-imagination. arXiv preprint arXiv:2401.08025, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[4] Mu Cai, Haotian Liu, Siva Karthik Mustikovela, Gregory P Meyer, Yuning Chai, Dennis Park, and Yong Jae Lee. Vipllava: Making large multimodal models understand arbitrary visual prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12914-12923, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[6] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[8] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[9] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.192, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[13] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[14] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[15] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[16] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[17] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[21] Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24211" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[25] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[27] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[28] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[29] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, et al. Llava-plus: Learning to use tools for creating multimodal agents. In European Conference on Computer Vision, pages 126-142. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[31] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[32] Sachit Menon, Richard Zemel, and Carl Vondrick. Whiteboard-of-thought: Thinking step-by-step across modalities. arXiv preprint arXiv:2406.14562, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[33] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[34] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 12242-12254, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[37] Chuyi Shang, Amos You, Sanjay Subramanian, Trevor Darrell, and Roei Herzig. Traveler: A modular multi-lmm agent framework for video question-answering. arXiv preprint arXiv:2404.01476, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[38] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11987-11997, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.32 + ], + "angle": 0, + "content": "[39] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[40] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[41] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.504 + ], + "angle": 0, + "content": "[42] Jiapeng Wang, Chengyu Wang, Kunzhe Huang, Jun Huang, and Lianwen Jin. Videoclip-xl: Advancing long description understanding for video clip models. arXiv preprint arXiv:2410.00741, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.574 + ], + "angle": 0, + "content": "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.907, + 0.632 + ], + "angle": 0, + "content": "[44] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[45] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[46] Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. Mind's eye of llms: visualization-of-thought elicits spatial reasoning in large language models. Advances in Neural Information Processing Systems, 37:90277-90317, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[47] Yixuan Wu, Yizhou Wang, Shixiang Tang, Wenhao Wu, Tong He, Wanli Ouyang, Philip Torr, and Jian Wu. Dettoolchain: A new prompting paradigm to unleash detection ability of mllm. In European Conference on Computer Vision, pages 164-182. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24212" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[49] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[50] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[51] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[52] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[53] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In *Forty-first International Conference on Machine Learning*, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[54] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[55] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[56] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.683 + ], + "angle": 0, + "content": "[57] Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandas Alexandros Potamias. Hawor: World-space hand motion reconstruction from egocentric videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1805-1815, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[58] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24213" + } + ] +] \ No newline at end of file diff --git a/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_origin.pdf b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aaf7b902d553a0ea5250413d57d1c1a4c267e76c --- /dev/null +++ b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/340dea8a-103d-45f5-8ee9-2f2409e86042_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a222f824864d04e09ad7d9543989a124b53431db37c9fefcb3904df8a4901766 +size 1030252 diff --git a/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/full.md b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3c18460ddba1a40a8fe5624f270915dfcb288fe6 --- /dev/null +++ b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/full.md @@ -0,0 +1,449 @@ +# VTimeCoT: Thinking by Drawing for Video Temporal Grounding and Reasoning + +Jinglei Zhang $^{1,*}$ Yuanfan Guo $^{2,*}$ Rolandas Alexandros Potamias $^{3}$ + +Jiankang Deng $^{3}$ Hang Xu $^{2}$ Chao Ma $^{1,\dagger}$ + +$^{1}$ Shanghai Jiao Tong University $^{2}$ Noah's Ark Lab $^{3}$ Imperial College London + +{zhangjinglei168,chaoma}@sjtu.edu.cn + +{guoyuanfan1, xuHang}@huawei.com, {j.deng16, r.potamias}@imperial.ac.uk + +![](images/fe77a4122404595f3d3153ae1e10d098fea5860feb41ce344586e51ce924620c.jpg) +How many streams did the cat cross in the video? + +![](images/e418e21b15791e30915f398284f76a2bb4fa6041a11385dd17e212aa92a80617.jpg) +GPT4o:one + +![](images/7ea4872bac4f1be9fc5509a15c51c6695578eeebb9cb26d5b12eec6511916af9.jpg) +VTimeCoT + GPT40 + +![](images/ca7f6f65b3e6b786d149aefbe4311a0c46d28c9722db93d660a119acca7431fc.jpg) + +THOUGHT 0: I can use the tool to draw progress bar and highlight the moments for 'streams the cat crossed'. + +ACTION 0: + +Call and + +![](images/9dbb6183ff21fe52b936d4944c3b691ca4abe6d22f69ed2cbbd2f395f6ee8165.jpg) + +![](images/0f086e473831ed1baf3b25f427a35ef1daeb0cd413e6de51e2cbd75879315ba6.jpg) + +![](images/14b42001dbbd8d52f184e3b659849c8153c68c3e59428f90ea50fc0a53361666.jpg) + +![](images/70e3d9cc003e995ea6f3501986a7a165305c70ff59e24198d45a63137fdca569.jpg) + +THOUGHT 1: The tool is helpful. I see the cat crossing streams at 52s to 68s and 71s to 76s. So it crossed two streams. + +![](images/451949e5850d0b3ae0f65a22c4451172878935df6a79f2e332b199240d106eb3.jpg) + +![](images/446d95f8cbd00ff432d5384327f783f2ef44d1dec697669fe3db11b3d429cdef.jpg) + +![](images/e5376232fe887b572579d1ce7c4fd9f0958ff5fc0350909b889ab7b6aa10932d.jpg) +Figure 1. We propose VTimeCoT, a Visual Time Chain-of-Thought framework for video temporal grounding and reasoning. VTimeCoT constructs cross-modality reasoning across both video and text, which enables the MLLM to utilize progress bar tools to annotate the time progression and highlight the key relevant segments to answer complex temporal questions. + +# Abstract + +In recent years, video question answering based on multimodal large language models (MLLM) has garnered considerable attention, due to the benefits from the substantial advancements in LLMs. However, these models have a notable deficiency in the domains of video temporal grounding and reasoning, posing challenges to the development of effective real-world video understanding systems. Inspired by how humans use video players to interact with the progress bar for video comprehension, we introduce VTimeCoT, a simple yet effective training-free framework, designed for high-performance video grounding and reasoning. The proposed framework incorporates two novel visual tools of the progress bar: a plug-and-play progress bar integration tool and a high-efficiency highlighting tool. In addition, + +to address the limitations of conventional text-based chain-of-thought (CoT) approaches, we introduce a visuotemporal CoT process that integrates cross-modality reasoning across both video and text. Our approach demonstrates significant performance improvements on both Qwen2VL-7B and GPT4o baselines in tasks of video temporal grounding and reasoning-based question answering. Finally, we showcase that the proposed framework achieves a compositional and interpretable reasoning process. Project page: https://vtimecot.github.io. + +# 1. Introduction + +Video understanding is a longstanding problem in computer vision and has attracted more attention with the emergence of large language models (LLM) [5, 17, 40, 43]. Video QA is a representative task reflecting the video reasoning abl- + +ity. Given that videos typically contain numerous events occurring at different time points, accurately answering questions is a highly challenging task. Recently, some end-to-end video understanding models based on multimodal large language models (MLLM) [20, 23, 26, 27, 56, 58] have been proposed to address video question-answering tasks, exhibiting remarkable capabilities. However, despite their ability to generate seemingly plausible results, these methods have been shown to exhibit a notable deficiency in temporal grounding [15] and fail to provide a temporal-grounded reasoning process. + +With the advent of tool usage capabilities in LLMs, several agent-based video question-answering methods have been proposed [9, 33, 37, 44, 53], enabling step-by-step video understanding. These methods prompt the LLMs to sequentially reason and invoke external tool models (e.g., object detection and captioning models) to collect textual clues, ultimately inferring the final answer. This, so called, chain-of-thought approach offers significant potential, as it is compositional and training-free, demonstrating impressive zero-shot performance. However, such methods predominantly rely on language and captions as intermediaries, limiting their ability to directly capture the visuotemporal dynamics inherent in videos, especially in the case of long videos. Inspired by the time-progress bar commonly used in video players during human interactions, we identify it as an intuitive tool for grounding time and enhancing video comprehension. In contrast to written language, the progress bar, along with the concurrent visual content, directly conveys the concept of temporal progression. The video progress bar can represent temporal relationships through sequential positioning or more abstract temporal dependencies. + +In this work, we propose a simple yet effective framework that empowers multimodal LLMs to utilize a video progress bar to construct a Visual Time Chain-of-Thought (VTimeCoT), facilitating long video temporal grounding and reasoning. Inspired by LLMs coupled with visual programming [12, 39], we propose enabling multimodal LLMs to generate code that overlays a progress bar on the video. First, we construct a frame-sync progress bar integration tool that can be invoked by the MLLM, which generates a progress bar displayed at the bottom of the video, marking the progress and annotating the timestamps on each frame. This is naturally adapted to long video and enables the MLLM to perceive the speed of temporal progress directly from the annotated video frames, adapting seamlessly to any frame-per-second (FPS) sampling rate. Second, to guide the model's attention on key relevant temporal segments, we propose a training-free and long-video-adaptive highlighting tool for the progress bar based on video-text similarity. Specifically, we leverage a robust video-text foundation model to identify the time intervals with the + +top-k highest similarity and highlight them on the visual progress bar. Finally, we prompt the multimodal LLM to invoke the progress bar tools within a visuotemporal chain-of-thought process that integrates text, program, and video annotated with the progress bar to facilitate long video reasoning. At each thought step, the model performs cross-modality reasoning and determines whether to dynamically update the video memory while progressively inferring the answer. As shown in Fig. 1, to determine how many times a cat crosses a stream in the input video, the model first invokes tools to generate a progress bar, identifying the key timestamps corresponding to 'streams the cat crossed'. By analyzing the progress bar annotated frames, the model infers the exact timing of each crossing event and accurately determines the number of crossings. This approach enhances the model's temporal reasoning capabilities by seamlessly integrating temporal-grounded visual evidence. + +We demonstrate the effectiveness of the proposed framework across a wide range of video temporal grounding and video-QA tasks. The proposed method consistently outperforms the baseline models, Qwen2VL and GPT-4o in temporal grounding tasks, achieving significant performance gains with an average IoU improvement of $6.58\%$ and $16.83\%$ over GPT-4o across Charades-STA [11] and QVHighlights [19] benchmarks. Similarly, VTimeCoT significantly improves the accuracy of questions related to temporal retrieval, event counting, and event ordering in reasoning-based question-answering tasks. Specifically, the proposed method consistently surpasses state-of-the-art methods across Vript-RR [50] and VideoMME [10] benchmarks. + +To sum up, we present VTimeCoT, a training-free model for temporal reasoning based on the visual progress bar. Specifically: + +- We propose the first, to the best of our knowledge, visual time chain-of-thought framework for video temporal grounding and reasoning. In contrast to previous methods that rely solely on textual reasoning, we leverage the visual progress bar as a medium, making a significant step towards real-world video understanding systems. +- The proposed reasoning framework leverages two novel tools from video progress bar integration and highlighting, enabling the MLLM to accurately perceive the times-tamp of each frame and identify key temporal segments. +- The proposed method enhances the performance of MLLMs by a large margin in video temporal grounding and reasoning tasks through a training-free approach, while also demonstrating strong reasoning capabilities. + +# 2. Related Work + +Multimodal LLM for Video. In contrast with traditional video understanding such as object detection, segmentation and pose estimation [34, 36, 41, 57], recent mul + +timodal LLMs enable understanding through natural language. End-to-end video multimodal LLMs are trained on large-scale datasets, to align visual inputs with the language modality and integrate with large language models. Video-ChatGPT [31] leveraged CLIP [35] to extract spatial and temporal features from videos and integrate them with Vicuna [7]. Video-LLaMA [56] proposed a video Q-Former designed to enhance the modeling of temporal variations. Subsequently, a series of video multimodal LLM variants [6, 20, 22, 24–26, 48] have introduced advancements in enhancing video representation, expanding data scale, and refining training methods, leading to improved video question-answering performance. However, these end-to-end models operate as black boxes, directly generating answers without an explicit video reasoning process. In contrast, we propose a video understanding framework that incorporates a structured reasoning process. + +Programming and Tool Usage of LLM. Due to the strong code-generation capabilities of LLMs, several works [8, 12, 14, 29, 39, 52] attempted to prompt LLMs to call compositional visual tools to address visual question-answering tasks. VisProg [12] and ViperGPT [39] defined a variety of image foundation functions (e.g., object detection) and prompted the LLMs to generate code that invokes these functions for answering questions. MM-REACT [52] tackled visual tasks through multi-step reasoning guided by text prompts to the LLM, where each step involves thought and calling visual tools. More recently, several studies [9, 33, 37, 44, 53] proposed tool-usage frameworks tailored for video understanding. By designing specialized video tools and calling steps, these approaches extract textual cues from videos to facilitate question answering. In contrast to the aforementioned works that solely rely on text as the reasoning medium, we introduce a visual progress bar, enabling more effective integration of visual and temporal cues. + +Visual Prompt for Multimodal LLM. LLMs have been shown to enhance their understanding capabilities by incorporating overlaid visual prompts on images (e.g., circles, keypoints) [3, 4, 13, 21, 32, 38, 46, 47, 51]. SoM [51] proposed to utilize SAM [18] segmentation model to overlay mask colors and labels on images as input of GPT-4V [2], demonstrating strong zero-shot performance on vision-grounding tasks. Hu et al. [13] introduced a chain-of-thought framework combined with visual prompts to draw markers on images during reasoning, effectively enhancing the image understanding performance of MLLMs. However, all previous methods are limited to spatial prompts, limiting their understanding capabilities. In this work, we propose to design the progress bar as visual-temporal prompts to facilitate temporal reasoning. + +# 3. Method + +We propose VTimeCoT, a general framework that empowers multimodal large language models using a visual progress bar as an intermediate reasoning step, facilitating temporal grounding and reasoning. Fig. 2 illustrates how our approach works. Given a raw video and a question, our method generates a chain-of-thought and invokes a set of tools for progress bar integration. By utilizing the visual cues from the progress bar as a medium, the model progressively infers the final answer. + +# 3.1. Visuotemporal Chain-of-Thought + +In video understanding, particularly for long videos, existing MLLM methods can only provide static answers, which limits the model's ability to focus on critical scenes and results in significant shortcomings in real-world settings. Since the MLLMs fail to identify key scenes within a single forward pass, it is necessary to interact with the video to progressively infer the answer. Therefore, we propose a framework based on interactive mechanisms and dynamic video memory. + +Our framework tackles video grounding and reasoning tasks through an iterative interaction with the MLLM, built upon a tool set leveraging the progress bar. Given a textual question and a video as input, VTimeCoT generates a sequence of thoughts and actions to dynamically update the video memory. By manipulating the video, it acquires the necessary information to answer the question. In this process, the MLLM reasons by plotting and analyzing the progress bar, seamlessly integrating both textual and visual reasoning into a Visuotemporal Chain-of-Thought. As illustrated in Fig. 2, VTimeCoT first provides the MLLM with a formatted prompt for initialization and executes actions at each time step $t$ . The tool set includes the , , , etc. The tool is utilized to trim specific segments when the MLLM determines that the video is too long to find the required information. The pseudo-code is presented in Algorithm 1, where $\mathbb{V}$ and $\mathbb{L}$ represent the modalities of video and language, respectively. In Algorithm 1, $p$ serves as a parser to retrieve specific sub-strings by keyword, and $c$ is a code builder that integrates the toolset library and generated code. + +Initialization Prompt. To enable the MLLM to perform step-by-step reasoning and a structured output, we follow [13] and construct a set of initialization prompts. These prompts define the specific output structure (formatted by keywords of 'THOUGHT', 'ACTION', and 'TERMINATE') for each step and declare the Python tools that the model should invoke. The initialization prompts, along with the question and video frames, are then fed into the MLLM, triggering the iterative loop. + +Thought. During this step, the model analyzes the historical context and video memory to generate its reasoning + +![](images/d9e22ee9e0ace6f86642d0065b302c600598867d6a2c306822daa9db50b36b28.jpg) +Figure 2. Overview of our method. On the left, we demonstrate how the framework iteratively generates thoughts and actions, which dynamically updates the video memory with an overlaid progress bar for reasoning. On the right, we illustrate two novel tools that integrate the frame-sync visual progress bar and highlight key moments. + +for the current step. It determines whether a direct answer can be provided directly in this step or if tool assistance is needed, and if so, which tools should be invoked. + +Action. Following the thought step, the model generates an executable Python script utilizing the given tool functions. The framework then builds the code with the toolset library and executes the script to manipulate the video frames. + +Dynamic Video Memory. Using the manipulated video frames we subsequently update the frames in the video memory. The updated context with the manipulated video frames is then fed into the MLLM to proceed with the next step of thought. + +To terminate the loop, the model determines at each step whether to cease reasoning and generate the 'TERMINATE' keyword to end. To prevent excessively prolonged reasoning, we impose a maximum step limit, enforcing termination at step $T$ . + +# 3.2. Frame-Sync Visual Progress Bar Integration + +Enabling MLLM to perceive the precise timestamp of each frame is a significant challenge and an unresolved problem, due to the diverse temporal sampling rates during MLLM training and inference without time. Although recent approaches [15, 16] have attempted to address this issue by incorporating temporal positional embeddings before the video encoder and introducing temporal-grounding training tasks, their flexibility is limited due to the scarcity of specific data and the need for fine-tuning. This issue is particularly pronounced in long videos, where the token limitations of MLLMs prevent them from accurately perceiving + +
Algorithm 1 Visuotemporal Chain-of-Thought
Input: {V, Q||V ∈ V, Q ∈ L}; I ∈ L;MLLM: V, L → L; Toolset: {T1, . . ., Tn} ∈ L
1: s ← 0; V0 ← V
2: H0 ← {I, Q}▷ Initialization prompt and query
3: while true do
4: ys ← MLLM (Vs, Hs)
5: THOUGHTs, ACTIONs, TERMINATEs ← p(ys)
6: if not TERMINATEs then
7: Cs ← c(ACTIONs, Toolset)▷ Build code
8: Vs+1 ← Cs(Vs)▷ Update video
9: Hs+1 ← {Hs, ys}▷ Update history
10: else
11: return ys▷ Respond final answer
12: end if
13: end while
+ +the timestamp of each frame. + +Inspired by visual programming paradigms [12, 39], we propose a simple yet effective method to empower MLLMs to perceive the precise timestamp of each frame, by generating the video progress bar. Leveraging the code generation capabilities of MLLMs, our approach prompts the MLLM to invoke Python plotting tools and use their universal OCR and shape understanding capabilities to perceive time, without any additional training requirements. Specifically, given a raw video $V \in \mathbb{R}^{T \times H \times W \times 3}$ as input, this tool can generate the progress bar at the bottom of the original frames and return the annotated frames. To enable adaptation to ar + +binary FPS sampling, we perform a frame-synchronization step to convert frame indexes to the real second time, as shown in Fig. 2. To enable MLLM to perceive seconds directly without any additional hour-minute-second conversions, we design the timestamps using a straightforward seconds format. To construct a robust visual prompt that can be easily interpreted from the MLLM, we employ compositional components using elongated rectangles, circles, and timestamps to plot the progress bar. Finally, the progress bar is placed at the bottom of raw frames, which can be expressed as: + +$$ +V _ {t} ^ {\prime} = V _ {t} \oplus T _ {t}, \tag {1} +$$ + +where $\oplus$ represents the vertical concatenation operation, $T_{t}$ denotes the image of the progress bar generated by OpenCV at time $t$ and $V_{t}^{\prime}$ is the resulting frame. We wrap this integration process into a Python tool function, allowing MLLM to invoke it for generating the annotated video. + +# 3.3. Moment Highlight by Video-Text Similarity + +Although multimodal LLMs have made significant progress in video content captioning, recent studies [15, 16] have revealed substantial deficiencies in their temporal grounding capabilities. Existing works [15, 16] rely on constructing temporal grounding instructional data and require finetuning to emerge such capability. Besides, due to token limitations in MLLMs, they struggle to perform temporal grounding for long videos. In contrast, we propose a training-free and long-video-adaptive method to enhance the temporal grounding capabilities of MLLMs, leveraging the similarities from the robust video-text foundation models. + +Given a query and a video $V \in \mathbb{R}^{T \times H \times W \times 3}$ containing multiple events, our proposed tool estimates the query-relevant temporal segments, including their start and end times, and generates a highlighted visual prompt. + +Moment Retrieval by Video-Text Similarity. To construct a robust zero-shot moment retrieval module, capable of generalizing across a wide range of videos, we employ the VideoCLIP-XL [42] foundation model as our embedding extractor. The video is initially processed at a high temporal resolution of $r$ FPS. We group every 8 frames into a video clip, resulting in $N$ clips. It should be noted that since the temporal retrieval module operates as an external component, it can efficiently handle high frame rates without being constrained by the token count of the MLLM. This makes our method adaptive to long-form video. These clips are then processed by the VideoCLIP-XL visual encoder, producing $N$ video embeddings. We use VideoCLIP-XL text encoder to extract the text embedding for the text query and we compute the cosine similarity between the query embedding and the video embeddings of each clip: + +$$ +\operatorname {s i m} (i) = \frac {e _ {i} ^ {v} \cdot e ^ {q}}{\left| e _ {i} ^ {v} \right| \left| e ^ {q} \right|}, \tag {2} +$$ + +where $e_i^v$ denotes the embedding of the $i$ -th clip and $e^q$ is the query embedding. Using top- $k$ selection, we obtain the $k$ clips with the highest similarity and extract the start and end timestamps of each contiguous segment. + +Moment Highlight on the Progress Bar. Motivated by a series of works that demonstrated the effectiveness of visual cues as a medium for reasoning beyond text [3, 13, 21, 32, 46], we propose using moment-highlighting cues for video temporal reasoning. As shown in Fig. 2, we construct the highlight tool using OpenCV library to plot colorful highlight masks on specific intervals of the progress bar. This tool takes the video frames and time intervals as input and annotates the highlighted progress bar under the frames. To ensure that the MLLM can easily interpret the highlights, we utilize different colors between the mask and the progress bar. We wrap the aforementioned retrieval and highlighting process as a Python tool function, enabling the MLLM to invoke it to manipulate the input video frames. + +# 4. Experiments + +In this section, we first present the implementation details, datasets and evaluation metrics. Subsequently, we assess the performance of the proposed framework on the video temporal grounding benchmarks (Sec. 4.1). To further evaluate the reasoning capabilities, we conducted extensive quantitative and qualitative analyses on the "reasoning based on retrieving" benchmark of Vript-RR (Sec. 4.2). In addition, we demonstrate the significant role our reasoning framework plays in enhancing long video question-answering performance (Sec. 4.3). Finally, we give the ablations of key modules (Sec. 4.4). + +Implementation Details. In the framework, we conducted experiments on two core MLLMs, Qwen2VL-7B [43] and GPT4o-20240513 [17], as they represent the state-of-the-art open-source and closed-source MLLM models, respectively. We set the decoding temperature of the MLLM to 0. By default, the input for the MLLM consists of 32 frames uniformly sampled from the video. For video frames, $H$ and $W$ are resized to make the longer side 480 pixels. Our agent implementation is based on AutoGen [45]. The implementations of tool functions are adapted from Gradio [1]. For GPT4o, we utilize the OpenAI API service. The details of the MLLM prompts are provided in the supplementary materials. For VideoCLIP-XL we use FPS $r = 1$ to sample video frames and group every 8 frames as a clip to extract embedding. Regarding the top- $k$ selection of clips, $k$ is empirically set to 8. The maximum number of reasoning steps $T$ is set to 3. + +Datasets and Evaluation Metrics. To evaluate the proposed VTimeCoT method, we utilized four benchmark datasets, spanning a wide range of video temporal grounding, reasoning, and long-video question-answering tasks. + +Charades-STA [11] is a benchmark dataset for temporal + +
MethodR1@0.3R1@0.5R1@0.7mIoU
VideoChat-7B [23]9.003.301.306.50
VideoLLaMA-7B [56]10.403.800.907.10
VideoChatGPT-7B [31]20.007.701.7013.70
LLAVA-Onevision-7B [20]33.0411.054.1120.98
VTimeLLM-7B [15]51.0027.5011.4031.20
VTimeLLM-13B [15]55.3034.3014.7034.60
Qwen2VL-7B [43]37.3112.854.1124.34
VTimeCoTQwen2VL-7B66.9638.7920.8343.41
GPT4o [17]63.7637.1214.6540.20
VTimeCoTGPT4o74.0651.0222.4546.78
+ +Table 1. Quantitative comparison on the Charades-STA dataset for temporal grounding. We report the mIoU and the recall performance at different IoU thresholds. + +
MethodR1@0.3R1@0.5R1@0.7mIoU
LLAVA-Onevision-7B [20]34.9117.919.725.53
VTimeLLM-7B [15]44.5825.039.2928.99
Qwen2VL-7B [43]31.8714.657.3522.77
VTimeCoTQwen2VL-7B67.5045.7925.1146.21
GPT4o [17]55.6135.6819.2937.66
VTimeCoTGPT4o79.3559.7433.8154.49
+ +grounding, composed of 1334 videos along with the corresponding start-end frame annotations of 3720 queries. The average video length is 30 seconds. We follow [15] and report mean IoU (mIoU) and recall@1, IoU $\geq m(R@m)$ metrics, where $m$ includes 0.3, 0.5 and 0.7. IoU denotes the intersection over the union between the predicted and ground truth time segments. + +QVHighlights [19] is a benchmark dataset for semantic relevance and temporal grounding from discontinuous time segments. It contains 1519 videos, with an average length of 150 seconds, and 1550 queries with time annotations, where each query spans multiple discontinuous time segments within a video. For evaluation, we report the commonly used mIoU and R@m metrics. + +Vript-RR [50] is a challenging benchmark dataset for scene retrieval and multi-hop reasoning. The average length of the videos is 622 seconds. It contains 152 questions, each accompanied by a hint to locate the scene that the question refers to. The benchmark includes both multiple-choice and open-ended question-answering settings. We follow [50] and evaluate the open-ended accuracy using GPT4. + +VideoMME [10] is a video question-answering benchmark tailored for MLLMs, featuring diverse videos ranging from + +Table 2. Quantitative comparison on the QVHighlights dataset for temporal grounding of discontinuous segments. We report the mIoU and the recall performance at different IoU thresholds. + +
MethodsMulti-ChoiceOpen
VideoChatGPT [31]29.6017.80
Video-LLaMA [56]28.3014.50
VideoChat [23]22.4015.10
VideoChat2 [24]42.1022.40
ST-LLM [28]33.6026.30
PLLAVA 7B [48]55.3036.20
VILA-1.5 8B [27]55.3032.30
Qwen2VL-7B [43]59.8735.95
VTimeCoTQwen2VL-7B62.5041.45
GPT4o [17]70.3961.18
VTimeCoTGPT4o83.5568.42
+ +Table 3. Quantitative comparison on the Vript-RR dataset. We report the response accuracy on both multiple-choice and open-ended question settings. + +11 seconds to 1 hour in length. It contains 900 videos and 2700 question-answer pairs. We follow [10] to report accuracy metrics under two evaluation settings with subtitles and without subtitles. + +# 4.1. Temporal Grounding + +To ensure accurate temporal reasoning, it is essential to achieve robust and high-performance temporal grounding. In Tab. 1, we compare VTimeCoT with two state-of-the-art baselines, Qwen2VL-7B and GPT4o, to identify the start and end timestamps of the query event. Specifically, using both Qwen2VL-7B and GPT4o models as the core MLLM of our framework, we evaluate the performance of VTimeCoT to localize the query event. As can be easily observed, although these advanced MLLMs achieve state-of-the-art performance in video content recognition, they fall short in accurately grounding temporal boundaries for events in videos, facing difficulties in perceiving the precise timestamp and the boundaries of events. In contrast, the proposed method effectively addresses these challenges, providing accurate visual cues of the progress bar through the reasoning process that facilitates the grounding performance in videos. Note that VTimeCoT demonstrates superior performance without any additional training cost, even compared to the VTimeLLM model, which has been meticulously fine-tuned on extensive temporal grounding data. + +To further evaluate the temporal grounding performance of the proposed and the baseline methods in the presence of multiple discontinuous time segments for each query, we utilized the QVHighlights benchmark dataset (Tab. 2). Given a query and a video, we request from each model to identify start-end timestamps of all temporal segments in the video that fit the query. As can be observed, VTime-CoT consistently enhances the performance compared to different MLLMs. On the contrary, baseline MLLM meth + +# Reasoning based on retrieving scenes + +![](images/22ae648c1b8e507545ed2fbd81e1f13dc5919014f71f7020ce376d7f4f25660b.jpg) + +? Hint: A scene in the video features a first-person perspective from a motorcyclist cruising on a highway. ... The motorcycle's chrome details, black leather accessories, the rider's gloved hands are noticeable. + +What is the round object on the bike in the middle bottom of the screen? + +(A) A GPS device + +(B) A sport camera + +(C) The motorcycle's headlight + +(D) The motorcycle's spedometer + +![](images/c6a20ce201c7aad3b84205d472ec823da16927711521c1a1b96c481480360002.jpg) +Figure 3. Qualitative comparison on Vript-RR benchmark. In contrast to GPT-4o which is misled by the irrelevant scenes, VTimeCoT reasons using the progress bar and highlights the key moments, leading to accurate answers. + +GPT40: (C) The motorcycle's headlight. + +![](images/4e96aab5b6c8c4d153a2444d3354fde9bea27ec62754aefefca67d847258e7f9.jpg) + +![](images/f60ecb40aab8634bea755dfb4513177b5a75c30fe8cf24aee5a014d8285a8447.jpg) + +VTimeCoT + GPT40 + +THOUGHT0: I can use the progress_bar and highlight tool to highlight the key moments. Since the video is too long, I will cut the video into segments for a thorough review. Then I will display the highlighted frames to help me answer the question. + +# ACTION 0: + +sentence = 'A scene in the video features ...' + +frames = progress_bar(frames) + +frames, segment = highlight(frames, sentence, return_segment=True) +frames = cut(frames, segment) + +for frame in frames: + +display(frame) + +![](images/38551943a79adf478be4d4215b5ab39e4b0b5f0bdaf5641c47cafa4fcf1b62ff.jpg) + +![](images/214e3432346de37308894ffe776c07979fd129887fe4b6740753040cd8c5ad1e.jpg) + +![](images/6982f63d0be24433d900774ff85e9813a1ec0a9fb1ef9485fcc6c499d7f4c8b5.jpg) + +![](images/c013d4d9a6a3254278445b87f7e98f96f942882b5bcdb2696c57efa0aab0065c.jpg) + +![](images/518e5f8b7721179ada3447eb2e4d69575395bb216c5e803186ba74fb2d7aab2d.jpg) + +THOUGHT 1: From the highlighted frames, I can observe the details of the motorcycle's features. The round object fits a motorcycle's speedmeter. ANSWER: (D) The motorcycle's speedometer. TERMINATE + +1 + +# Temporal Counting + +![](images/1de9e0c0058ed449ccd702978bce08ec7cf1f6a97f7dc63d38d78112056281c6.jpg) + +![](images/df453bcaa0ab5e3aa19cae3c2f83433c6113c53d544fae8549ee3380e5613645.jpg) + +![](images/18dfc0e110167516cdc5de5ff162d140495a817aa85cb850e3edb183aff782be.jpg) + +![](images/ffb3986c6a8f97c20d01950a0c6fbee82955d555f8126c594330f414fad0af0d.jpg) + +![](images/da346dbd329ff689538753e3ce4c302cffd52f74392d2d930521ef62d9be286b.jpg) + +How many different kinds of animal faces are made in this video? + +A. 4 + +B. + +C. 5 + +D. 2 + +![](images/eabaf48c67b154ee77f8be675dbd9b1c550e7c170631e9d938c32b8d2907b4a2.jpg) + +GPT4o:A.4 + +![](images/eb86131d575065ddb3139aa3dac7063737b167d41e3e6bc12eaf1b279c1d6770.jpg) + +![](images/c31125daf2aa54b70df4b974c734d9988ad96b73e4146a19dd9469443ec3db44.jpg) + +VTimeCoT + GPT4o + +THOUGHT 0: I can use the progress_bar and the highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for 'different kinds of animal faces' to highlight them. I will then display the highlighted frames to help me answer the question. + +# ACTION 0: + +![](images/203b280ed43b686e591ad06c4982d8237173cc8b50cf16523140d0f9e6ed04db.jpg) +Figure 4. Qualitative comparison on VideoMME benchmark. Through reasoning that integrates the progress bar and highlights, VTimeCoT accurately answers questions related to temporal counting and order, which GPT-4o falls short. + +sentence = "different kinds of animal faces" + +frames = progress_bar(frames) + +frames = highlight(frame, sentence) + +for frame in frames: + +display(frame) + +![](images/515e82aaf7031981aab48ac2f627f6a38d95f68934020a48c60bfcf969457c70.jpg) + +![](images/9ff557e762b51cfb1e13af427cdbb15ec9ededdce7eb296870671df08564ea13.jpg) + +![](images/a58c43ae90839f113e038e753c7154666c6fb7edbd745c752c2000e8a4efb59e.jpg) + +![](images/65cce358efdf273e18aed48cc53585089d5109f5036484dd5ca3bbf928c04b2b.jpg) + +![](images/36486a1295d9999cb54b7498bd252ca2872436baa3de85f0612d84899d4e2b7a.jpg) + +THOUGHT 1: From the highlighted frames, I can count three different kinds of animal faces being made: a cat, a pig, and a dog. Therefore, the correct answer is 3. ANSWER: B TERMINATE + +# Temporal Order + +![](images/cb5677d6d17ea20abe0865f047f9db9b9942802b115989220fab96370409f755.jpg) + +![](images/5d521d0e788e868e4ddb8da0264314a1021785cbbc2776b73ddbb4d01dea6a54.jpg) + +![](images/7f83eea998341443282a3c892c31d129e7d83b8327d6c1b7bb8879233e851c37.jpg) + +![](images/75014981f8ad94ff2f6d0fe17eef69cd5d95942b74c35221fa1225205a978c47.jpg) + +![](images/517e3b2316a670aa1d7f21a975b2cdfbc57d06ebc376ad92a7a696885b34851b.jpg) + +What was the exercise demonstrated after the squatting exercise in the video? A. Push-ups B. Plank C. Step-ups D. Pull-ups + +![](images/fcca75f1f70759257db5615476e7ae126eb4f11b0011694644299927fd2ca8c6.jpg) + +GPT4o:B.Plank VTimeCoT+GPT4o + +THOUGHT 0: I can use the progress_bar and highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for `exercise demonstrated after the squatting exercise` to highlight them. I will then display the highlighted frames to help me answer the question. + +# ACTION 0: + +![](images/61ebf001ece87ab30bbca97f8f045fddfa682cd5864d5cffb65f375c0ee910ed.jpg) + +sentence = "exercise demonstrated after the squatting exercise" +frames = progress_barimentary + +frames = highlight(frame, sentence) + +for frame in frames: + +display(frame) + +![](images/f2550e143593791fc000b7e264c3fa10f255a6ca29d0d0836e24e5726a7ea912.jpg) + +![](images/b44e7e7a84358bd05b66be4f345d45a3b49a69425e05bb4f2bf14a1a7096fd3d.jpg) + +![](images/5879423196fdbe57056d48089100af2ac9729cc828027d6b7a4fef314778946e.jpg) + +![](images/6cae95118332a5a6f853a52ce6e9f13d0050cc593d4067af6b5057bbf239a574.jpg) + +![](images/5864b1026c86667038a04f7baf29081d52fd852c61159ba3814473b0674fac8b.jpg) + +THOUGHT 1: The highlighted frames show the sequence of exercises demonstrated in the video. After the squatting exercise, the next + +demonstrated exercise is step-ups, as shown in the frames around the 21-second mark. ANSWER: C TERMINATE + +ods struggle with discontinuous time segments, failing to identify multiple temporal spans and leading to degraded grounding performance. + +# 4.2. Long-Video Reasoning Based on Retrieving + +In Tab. 3, we evaluate the question-answering performance of the proposed and the baseline methods on the Vript-RR + +benchmark, which contains questions that require scene retrieval. Given a scene hint and a question, the model must identify the specific scene to deliver an accurate response. We compared the proposed method with the state-of-the-art baseline methods including Qwen2VL-7B and GPT4o. + +As can be observed VTimeCoT achieves a significant per + +
MethodsFramesAccuracy
w/o subsw. subs
Gemini-1.5-Flash [40]1 fps70.375.0
GPT4o [17]38471.977.2
Gemini-1.5-Pro [40]1 fps75.081.3
Video-LLaVA-7B [26]839.941.6
VideoLLaMA2-7B [6]1654.956.4
LLAVA-Onevision-7B [20]3258.261.5
LongVILA-7B [49]25660.165.1
InternVL2.5-8B [5]6464.266.9
mPLUG-Owl3-7B [54]12859.368.1
LLaVA-Video-7B [58]6463.369.7
NVILA-8B [30]25664.270.0
VideoLLaMA3-7B [55]18066.270.3
GPT4o [17]3261.665.1
VTimeCoTgpt4o3264.273.6
+ +formance improvement, surpassing the baseline methods in both multiple-choice and open-ended settings. + +To further illustrate the superiority of the proposed method in the task of scene retrieval and reasoning, we qualitatively compare GPT4o and VTimeCoT on the Vript-RR dataset in Fig. 3. Through multi-step reasoning that leverages the progress bar and highlights the key time interval, VTimeCoT accurately identifies the correct answer, while GPT-4o falls short. + +# 4.3. Long-Video Question Answering + +A pivotal challenge in long-form video question answering lies in accurately inferring the frequency and sequential relationships of temporal events. Currently, end-to-end video MLLMs struggle to accurately localize events and capture their temporal dependencies across multiple frames, limiting their ability to generalize to real-world temporal understanding. In Tab. 4, we evaluate the performance of the proposed VTimeCoT model in question answering on long videos of the VideoMME benchmark dataset. Due to the prohibitive evaluation costs and to ensure a fair comparison, we re-evaluate GPT-4o using 32 frames. Utilizing robust temporal grounding and reasoning, VTimeCoT significantly outperforms the baseline methods in long-video question answering. + +The effectiveness of the proposed method can be further validated in Fig. 4, where we compared the responses of the state-of-the-art GPT4o and VTimeCoT. In contrast to GPT4o, VTimeCoT can accurately identify the correct answers through its systematic, step-by-step reasoning process, even in challenging scenarios involving temporal counting and order discernment. It is also worth mention- + +Table 4. Quantitative comparison on the VideoMME dataset. We report the response accuracy under two settings: with subtitles and without subtitles. + +
MLLMCoTProgress BarHighlightQVHighlights (mIoU)Vript-RR (M-Acc)
GPT4o×××37.6670.39
GPT4o✓××41.8573.68
GPT4o✓✓×49.4076.32
GPT4o✓✓✓54.4983.55
+ +Table 5. Ablations on the key modules. M-Acc is the multi-choice accuracy on Vript-RR. + +ing that VTimeCoT, apart from its superior performance in question-answering accuracy, demonstrates logical and interpretable reasoning steps. + +# 4.4. Ablations + +To further investigate the contribution of each component in the proposed method, we conducted an ablation study. In Tab. 5, we evaluate the importance of the Chain-of-Thought (CoT), progress bar, and highlighting modules. As can be observed the GPT4o employing standard CoT without the use of the progress bar (i.e., solely on text-based step-by-step reasoning), results in significant performance degradation compared to the proposed method. To assess the effect of the proposed modules of the progress bar, we developed a configuration that utilizes only the progress bar tool without using the highlighting module. For completion, we also report the full VTimeCoT results, utilizing both the progress bar and highlighting tools. Both tools contribute to the performance improvement of the proposed method, which highlights the effectiveness of explicitly visualizing the precise timestamp and highlighting the relevant moments. + +# 5. Conclusion + +In this work, we propose the first visual time framework, designed to formulate a visuotemporal chain-of-thought for video temporal grounding and reasoning. We introduce a plug-and-play tool of the progress bar to generate visual temporal cues from any video, that can enable multimodal large language models (MLLMs) to leverage tool-usage capabilities for accurately perceiving the speed of temporal progression. In addition, we propose a zero-shot temporal retrieval tool, built on top of a strong video-text foundational model, augmenting it with temporal grounding capabilities without any further training requirements. We integrate these two tools into a visuotemporal chain-of-thought framework to facilitate cross-modal reasoning between video and text. Through extensive experiments, we demonstrate that the proposed method surpasses previous state-of-the-art baselines in temporal grounding and reasoning-based question-answering benchmarks, demonstrating logical and interpretable reasoning steps. + +Acknowledgments. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. Gradio: Hassle-free sharing and testing of ml models in the wild. arXiv preprint arXiv:1906.02569, 2019. 5 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[3] Syeda Nahida Akter, Aman Madaan, Sangwu Lee, Yiming Yang, and Eric Nyberg. Self-imagine: Effective unimodal reasoning with multimodal models using self-imagination. arXiv preprint arXiv:2401.08025, 2024. 3, 5 +[4] Mu Cai, Haotian Liu, Siva Karthik Mustikovela, Gregory P Meyer, Yuning Chai, Dennis Park, and Yong Jae Lee. Vipllava: Making large multimodal models understand arbitrary visual prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12914-12923, 2024. 3 +[5] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1, 8 +[6] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3, 8 +[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 3 +[8] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 3 +[9] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 2, 3 +[10] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6 +[11] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In + +Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 5 +[12] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 3, 4 +[13] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3, 5 +[14] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024. 3 +[15] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 4, 5, 6 +[16] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 4, 5 +[17] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 5, 6, 8 +[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3 +[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 2, 6 +[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 8 +[21] Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 3, 5 +[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3 +[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6 + +[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 6 +[25] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. +[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 3, 8 +[27] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 2, 6 +[28] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6 +[29] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, et al. Llava-plus: Learning to use tools for creating multimodal agents. In European Conference on Computer Vision, pages 126-142. Springer, 2024. 3 +[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 8 +[31] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3, 6 +[32] Sachit Menon, Richard Zemel, and Carl Vondrick. Whiteboard-of-thought: Thinking step-by-step across modalities. arXiv preprint arXiv:2406.14562, 2024. 3, 5 +[33] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 2, 3 +[34] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 12242-12254, 2025. 2 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman + +Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2 +[37] Chuyi Shang, Amos You, Sanjay Subramanian, Trevor Darrell, and Roei Herzig. Traveler: A modular multi-lmm agent framework for video question-answering. arXiv preprint arXiv:2404.01476, 2024. 2, 3 +[38] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11987-11997, 2023. 3 +[39] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 2, 3, 4 +[40] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 8 +[41] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 2 +[42] Jiapeng Wang, Chengyu Wang, Kunzhe Huang, Jun Huang, and Lianwen Jin. Videoclip-xl: Advancing long description understanding for video clip models. arXiv preprint arXiv:2410.00741, 2024.5 +[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6 +[44] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 2, 3 +[45] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5 +[46] Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. Mind's eye of llms: visualization-of-thought elicits spatial reasoning in large language models. Advances in Neural Information Processing Systems, 37:90277-90317, 2025. 3, 5 +[47] Yixuan Wu, Yizhou Wang, Shixiang Tang, Wenhao Wu, Tong He, Wanli Ouyang, Philip Torr, and Jian Wu. Dettoolchain: A new prompting paradigm to unleash detection ability of mllm. In European Conference on Computer Vision, pages 164-182. Springer, 2024. 3 +[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6 + +[49] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 8 +[50] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 2, 6 +[51] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3 +[52] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3 +[53] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In *Forty-first International Conference on Machine Learning*, 2024. 2, 3 +[54] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8 +[55] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 8 +[56] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 3, 6 +[57] Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandas Alexandros Potamias. Hawor: World-space hand motion reconstruction from egocentric videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1805-1815, 2025. 2 +[58] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 8 \ No newline at end of file diff --git a/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/images.zip b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d4729f99d311a46ac2e1620b49184455e10c580 --- /dev/null +++ b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15ddfac8dedb5130e581f81c3c68b2d5febd512a5b99af6b3cd2555a880239cb +size 564104 diff --git a/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/layout.json b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..601b3be9afd25a2a0c0035276f1271f33977de58 --- /dev/null +++ b/2025/VTimeCoT_ Thinking by Drawing for Video Temporal Grounding and Reasoning/layout.json @@ -0,0 +1,10469 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "VTimeCoT: Thinking by Drawing for Video Temporal Grounding and Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "spans": [ + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "type": "text", + "content": "Jinglei Zhang" + }, + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "type": "text", + "content": " Yuanfan Guo" + }, + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "type": "text", + "content": " Rolandas Alexandros Potamias" + }, + { + "bbox": [ + 137, + 142, + 473, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "spans": [ + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "type": "text", + "content": "Jiankang Deng" + }, + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "type": "text", + "content": " Hang Xu" + }, + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "type": "text", + "content": " Chao Ma" + }, + { + "bbox": [ + 201, + 158, + 408, + 171 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "text", + "content": " Shanghai Jiao Tong University " + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "text", + "content": " Noah's Ark Lab " + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 171, + 502, + 186 + ], + "type": "text", + "content": " Imperial College London" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 205, + 188, + 405, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 188, + 405, + 200 + ], + "spans": [ + { + "bbox": [ + 205, + 188, + 405, + 200 + ], + "type": "text", + "content": "{zhangjinglei168,chaoma}@sjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 201, + 496, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 201, + 496, + 213 + ], + "spans": [ + { + "bbox": [ + 111, + 201, + 496, + 213 + ], + "type": "text", + "content": "{guoyuanfan1, xuHang}@huawei.com, {j.deng16, r.potamias}@imperial.ac.uk" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 241, + 294, + 289 + ], + "blocks": [ + { + "bbox": [ + 59, + 241, + 294, + 289 + ], + "lines": [ + { + "bbox": [ + 59, + 241, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 241, + 294, + 289 + ], + "type": "image", + "image_path": "fe77a4122404595f3d3153ae1e10d098fea5860feb41ce344586e51ce924620c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 292, + 212, + 300 + ], + "lines": [ + { + "bbox": [ + 64, + 292, + 212, + 300 + ], + "spans": [ + { + "bbox": [ + 64, + 292, + 212, + 300 + ], + "type": "text", + "content": "How many streams did the cat cross in the video?" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 65, + 307, + 80, + 321 + ], + "blocks": [ + { + "bbox": [ + 65, + 307, + 80, + 321 + ], + "lines": [ + { + "bbox": [ + 65, + 307, + 80, + 321 + ], + "spans": [ + { + "bbox": [ + 65, + 307, + 80, + 321 + ], + "type": "image", + "image_path": "e418e21b15791e30915f398284f76a2bb4fa6041a11385dd17e212aa92a80617.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 309, + 120, + 316 + ], + "lines": [ + { + "bbox": [ + 86, + 309, + 120, + 316 + ], + "spans": [ + { + "bbox": [ + 86, + 309, + 120, + 316 + ], + "type": "text", + "content": "GPT4o:one" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 159, + 307, + 174, + 322 + ], + "blocks": [ + { + "bbox": [ + 159, + 307, + 174, + 322 + ], + "lines": [ + { + "bbox": [ + 159, + 307, + 174, + 322 + ], + "spans": [ + { + "bbox": [ + 159, + 307, + 174, + 322 + ], + "type": "image", + "image_path": "7ea4872bac4f1be9fc5509a15c51c6695578eeebb9cb26d5b12eec6511916af9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 309, + 263, + 318 + ], + "lines": [ + { + "bbox": [ + 208, + 309, + 263, + 318 + ], + "spans": [ + { + "bbox": [ + 208, + 309, + 263, + 318 + ], + "type": "text", + "content": "VTimeCoT + GPT40" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 179, + 307, + 196, + 323 + ], + "blocks": [ + { + "bbox": [ + 179, + 307, + 196, + 323 + ], + "lines": [ + { + "bbox": [ + 179, + 307, + 196, + 323 + ], + "spans": [ + { + "bbox": [ + 179, + 307, + 196, + 323 + ], + "type": "image", + "image_path": "ca7f6f65b3e6b786d149aefbe4311a0c46d28c9722db93d660a119acca7431fc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 328, + 272, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 328, + 272, + 344 + ], + "spans": [ + { + "bbox": [ + 67, + 328, + 272, + 344 + ], + "type": "text", + "content": "THOUGHT 0: I can use the tool to draw progress bar and highlight the moments for 'streams the cat crossed'." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 350, + 99, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 350, + 99, + 358 + ], + "spans": [ + { + "bbox": [ + 67, + 350, + 99, + 358 + ], + "type": "text", + "content": "ACTION 0:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 358, + 246, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 358, + 246, + 367 + ], + "spans": [ + { + "bbox": [ + 67, + 358, + 246, + 367 + ], + "type": "text", + "content": "Call and " + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 65, + 373, + 78, + 407 + ], + "blocks": [ + { + "bbox": [ + 65, + 373, + 78, + 407 + ], + "lines": [ + { + "bbox": [ + 65, + 373, + 78, + 407 + ], + "spans": [ + { + "bbox": [ + 65, + 373, + 78, + 407 + ], + "type": "image", + "image_path": "9dbb6183ff21fe52b936d4944c3b691ca4abe6d22f69ed2cbbd2f395f6ee8165.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 80, + 372, + 142, + 408 + ], + "blocks": [ + { + "bbox": [ + 80, + 372, + 142, + 408 + ], + "lines": [ + { + "bbox": [ + 80, + 372, + 142, + 408 + ], + "spans": [ + { + "bbox": [ + 80, + 372, + 142, + 408 + ], + "type": "image", + "image_path": "0f086e473831ed1baf3b25f427a35ef1daeb0cd413e6de51e2cbd75879315ba6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 154, + 373, + 217, + 408 + ], + "blocks": [ + { + "bbox": [ + 154, + 373, + 217, + 408 + ], + "lines": [ + { + "bbox": [ + 154, + 373, + 217, + 408 + ], + "spans": [ + { + "bbox": [ + 154, + 373, + 217, + 408 + ], + "type": "image", + "image_path": "14b42001dbbd8d52f184e3b659849c8153c68c3e59428f90ea50fc0a53361666.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 227, + 373, + 288, + 408 + ], + "blocks": [ + { + "bbox": [ + 227, + 373, + 288, + 408 + ], + "lines": [ + { + "bbox": [ + 227, + 373, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 227, + 373, + 288, + 408 + ], + "type": "image", + "image_path": "70e3d9cc003e995ea6f3501986a7a165305c70ff59e24198d45a63137fdca569.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 66, + 415, + 276, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 415, + 276, + 430 + ], + "spans": [ + { + "bbox": [ + 66, + 415, + 276, + 430 + ], + "type": "text", + "content": "THOUGHT 1: The tool is helpful. I see the cat crossing streams at 52s to 68s and 71s to 76s. So it crossed two streams." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 299, + 241, + 551, + 281 + ], + "blocks": [ + { + "bbox": [ + 299, + 241, + 551, + 281 + ], + "lines": [ + { + "bbox": [ + 299, + 241, + 551, + 281 + ], + "spans": [ + { + "bbox": [ + 299, + 241, + 551, + 281 + ], + "type": "image", + "image_path": "451949e5850d0b3ae0f65a22c4451172878935df6a79f2e332b199240d106eb3.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 299, + 284, + 551, + 329 + ], + "blocks": [ + { + "bbox": [ + 299, + 284, + 551, + 329 + ], + "lines": [ + { + "bbox": [ + 299, + 284, + 551, + 329 + ], + "spans": [ + { + "bbox": [ + 299, + 284, + 551, + 329 + ], + "type": "image", + "image_path": "446d95f8cbd00ff432d5384327f783f2ef44d1dec697669fe3db11b3d429cdef.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 299, + 332, + 551, + 441 + ], + "blocks": [ + { + "bbox": [ + 299, + 332, + 551, + 441 + ], + "lines": [ + { + "bbox": [ + 299, + 332, + 551, + 441 + ], + "spans": [ + { + "bbox": [ + 299, + 332, + 551, + 441 + ], + "type": "image", + "image_path": "e5376232fe887b572579d1ce7c4fd9f0958ff5fc0350909b889ab7b6aa10932d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 454, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 54, + 454, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 454, + 555, + 489 + ], + "type": "text", + "content": "Figure 1. We propose VTimeCoT, a Visual Time Chain-of-Thought framework for video temporal grounding and reasoning. VTimeCoT constructs cross-modality reasoning across both video and text, which enables the MLLM to utilize progress bar tools to annotate the time progression and highlight the key relevant segments to answer complex temporal questions." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 151, + 498, + 200, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 498, + 200, + 510 + ], + "spans": [ + { + "bbox": [ + 151, + 498, + 200, + 510 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 54, + 524, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 524, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 524, + 297, + 693 + ], + "type": "text", + "content": "In recent years, video question answering based on multimodal large language models (MLLM) has garnered considerable attention, due to the benefits from the substantial advancements in LLMs. However, these models have a notable deficiency in the domains of video temporal grounding and reasoning, posing challenges to the development of effective real-world video understanding systems. Inspired by how humans use video players to interact with the progress bar for video comprehension, we introduce VTimeCoT, a simple yet effective training-free framework, designed for high-performance video grounding and reasoning. The proposed framework incorporates two novel visual tools of the progress bar: a plug-and-play progress bar integration tool and a high-efficiency highlighting tool. In addition," + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 499, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 555, + 619 + ], + "type": "text", + "content": "to address the limitations of conventional text-based chain-of-thought (CoT) approaches, we introduce a visuotemporal CoT process that integrates cross-modality reasoning across both video and text. Our approach demonstrates significant performance improvements on both Qwen2VL-7B and GPT4o baselines in tasks of video temporal grounding and reasoning-based question answering. Finally, we showcase that the proposed framework achieves a compositional and interpretable reasoning process. Project page: https://vtimecot.github.io." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "spans": [ + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Video understanding is a longstanding problem in computer vision and has attracted more attention with the emergence of large language models (LLM) [5, 17, 40, 43]. Video QA is a representative task reflecting the video reasoning abl-" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 223, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 223, + 714 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 223, + 714 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24203" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "ity. Given that videos typically contain numerous events occurring at different time points, accurately answering questions is a highly challenging task. Recently, some end-to-end video understanding models based on multimodal large language models (MLLM) [20, 23, 26, 27, 56, 58] have been proposed to address video question-answering tasks, exhibiting remarkable capabilities. However, despite their ability to generate seemingly plausible results, these methods have been shown to exhibit a notable deficiency in temporal grounding [15] and fail to provide a temporal-grounded reasoning process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 471 + ], + "type": "text", + "content": "With the advent of tool usage capabilities in LLMs, several agent-based video question-answering methods have been proposed [9, 33, 37, 44, 53], enabling step-by-step video understanding. These methods prompt the LLMs to sequentially reason and invoke external tool models (e.g., object detection and captioning models) to collect textual clues, ultimately inferring the final answer. This, so called, chain-of-thought approach offers significant potential, as it is compositional and training-free, demonstrating impressive zero-shot performance. However, such methods predominantly rely on language and captions as intermediaries, limiting their ability to directly capture the visuotemporal dynamics inherent in videos, especially in the case of long videos. Inspired by the time-progress bar commonly used in video players during human interactions, we identify it as an intuitive tool for grounding time and enhancing video comprehension. In contrast to written language, the progress bar, along with the concurrent visual content, directly conveys the concept of temporal progression. The video progress bar can represent temporal relationships through sequential positioning or more abstract temporal dependencies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "In this work, we propose a simple yet effective framework that empowers multimodal LLMs to utilize a video progress bar to construct a Visual Time Chain-of-Thought (VTimeCoT), facilitating long video temporal grounding and reasoning. Inspired by LLMs coupled with visual programming [12, 39], we propose enabling multimodal LLMs to generate code that overlays a progress bar on the video. First, we construct a frame-sync progress bar integration tool that can be invoked by the MLLM, which generates a progress bar displayed at the bottom of the video, marking the progress and annotating the timestamps on each frame. This is naturally adapted to long video and enables the MLLM to perceive the speed of temporal progress directly from the annotated video frames, adapting seamlessly to any frame-per-second (FPS) sampling rate. Second, to guide the model's attention on key relevant temporal segments, we propose a training-free and long-video-adaptive highlighting tool for the progress bar based on video-text similarity. Specifically, we leverage a robust video-text foundation model to identify the time intervals with the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": "top-k highest similarity and highlight them on the visual progress bar. Finally, we prompt the multimodal LLM to invoke the progress bar tools within a visuotemporal chain-of-thought process that integrates text, program, and video annotated with the progress bar to facilitate long video reasoning. At each thought step, the model performs cross-modality reasoning and determines whether to dynamically update the video memory while progressively inferring the answer. As shown in Fig. 1, to determine how many times a cat crosses a stream in the input video, the model first invokes tools to generate a progress bar, identifying the key timestamps corresponding to 'streams the cat crossed'. By analyzing the progress bar annotated frames, the model infers the exact timing of each crossing event and accurately determines the number of crossings. This approach enhances the model's temporal reasoning capabilities by seamlessly integrating temporal-grounded visual evidence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 275, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 441 + ], + "type": "text", + "content": "We demonstrate the effectiveness of the proposed framework across a wide range of video temporal grounding and video-QA tasks. The proposed method consistently outperforms the baseline models, Qwen2VL and GPT-4o in temporal grounding tasks, achieving significant performance gains with an average IoU improvement of " + }, + { + "bbox": [ + 313, + 275, + 553, + 441 + ], + "type": "inline_equation", + "content": "6.58\\%" + }, + { + "bbox": [ + 313, + 275, + 553, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 275, + 553, + 441 + ], + "type": "inline_equation", + "content": "16.83\\%" + }, + { + "bbox": [ + 313, + 275, + 553, + 441 + ], + "type": "text", + "content": " over GPT-4o across Charades-STA [11] and QVHighlights [19] benchmarks. Similarly, VTimeCoT significantly improves the accuracy of questions related to temporal retrieval, event counting, and event ordering in reasoning-based question-answering tasks. Specifically, the proposed method consistently surpasses state-of-the-art methods across Vript-RR [50] and VideoMME [10] benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 553, + 479 + ], + "type": "text", + "content": "To sum up, we present VTimeCoT, a training-free model for temporal reasoning based on the visual progress bar. Specifically:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 480, + 553, + 647 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 480, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 553, + 552 + ], + "type": "text", + "content": "- We propose the first, to the best of our knowledge, visual time chain-of-thought framework for video temporal grounding and reasoning. In contrast to previous methods that rely solely on textual reasoning, we leverage the visual progress bar as a medium, making a significant step towards real-world video understanding systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 552, + 553, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 553, + 600 + ], + "type": "text", + "content": "- The proposed reasoning framework leverages two novel tools from video progress bar integration and highlighting, enabling the MLLM to accurately perceive the times-tamp of each frame and identify key temporal segments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 601, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 553, + 647 + ], + "type": "text", + "content": "- The proposed method enhances the performance of MLLMs by a large margin in video temporal grounding and reasoning tasks through a training-free approach, while also demonstrating strong reasoning capabilities." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Multimodal LLM for Video. In contrast with traditional video understanding such as object detection, segmentation and pose estimation [34, 36, 41, 57], recent mul" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24204" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 277 + ], + "type": "text", + "content": "timodal LLMs enable understanding through natural language. End-to-end video multimodal LLMs are trained on large-scale datasets, to align visual inputs with the language modality and integrate with large language models. Video-ChatGPT [31] leveraged CLIP [35] to extract spatial and temporal features from videos and integrate them with Vicuna [7]. Video-LLaMA [56] proposed a video Q-Former designed to enhance the modeling of temporal variations. Subsequently, a series of video multimodal LLM variants [6, 20, 22, 24–26, 48] have introduced advancements in enhancing video representation, expanding data scale, and refining training methods, leading to improved video question-answering performance. However, these end-to-end models operate as black boxes, directly generating answers without an explicit video reasoning process. In contrast, we propose a video understanding framework that incorporates a structured reasoning process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 291, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 295, + 518 + ], + "type": "text", + "content": "Programming and Tool Usage of LLM. Due to the strong code-generation capabilities of LLMs, several works [8, 12, 14, 29, 39, 52] attempted to prompt LLMs to call compositional visual tools to address visual question-answering tasks. VisProg [12] and ViperGPT [39] defined a variety of image foundation functions (e.g., object detection) and prompted the LLMs to generate code that invokes these functions for answering questions. MM-REACT [52] tackled visual tasks through multi-step reasoning guided by text prompts to the LLM, where each step involves thought and calling visual tools. More recently, several studies [9, 33, 37, 44, 53] proposed tool-usage frameworks tailored for video understanding. By designing specialized video tools and calling steps, these approaches extract textual cues from videos to facilitate question answering. In contrast to the aforementioned works that solely rely on text as the reasoning medium, we introduce a visual progress bar, enabling more effective integration of visual and temporal cues." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 714 + ], + "type": "text", + "content": "Visual Prompt for Multimodal LLM. LLMs have been shown to enhance their understanding capabilities by incorporating overlaid visual prompts on images (e.g., circles, keypoints) [3, 4, 13, 21, 32, 38, 46, 47, 51]. SoM [51] proposed to utilize SAM [18] segmentation model to overlay mask colors and labels on images as input of GPT-4V [2], demonstrating strong zero-shot performance on vision-grounding tasks. Hu et al. [13] introduced a chain-of-thought framework combined with visual prompts to draw markers on images during reasoning, effectively enhancing the image understanding performance of MLLMs. However, all previous methods are limited to spatial prompts, limiting their understanding capabilities. In this work, we propose to design the progress bar as visual-temporal prompts to facilitate temporal reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "type": "text", + "content": "We propose VTimeCoT, a general framework that empowers multimodal large language models using a visual progress bar as an intermediate reasoning step, facilitating temporal grounding and reasoning. Fig. 2 illustrates how our approach works. Given a raw video and a question, our method generates a chain-of-thought and invokes a set of tools for progress bar integration. By utilizing the visual cues from the progress bar as a medium, the model progressively infers the final answer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 205, + 494, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 494, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 494, + 219 + ], + "type": "text", + "content": "3.1. Visuotemporal Chain-of-Thought" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 223, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 331 + ], + "type": "text", + "content": "In video understanding, particularly for long videos, existing MLLM methods can only provide static answers, which limits the model's ability to focus on critical scenes and results in significant shortcomings in real-world settings. Since the MLLMs fail to identify key scenes within a single forward pass, it is necessary to interact with the video to progressively infer the answer. Therefore, we propose a framework based on interactive mechanisms and dynamic video memory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "content": "Our framework tackles video grounding and reasoning tasks through an iterative interaction with the MLLM, built upon a tool set leveraging the progress bar. Given a textual question and a video as input, VTimeCoT generates a sequence of thoughts and actions to dynamically update the video memory. By manipulating the video, it acquires the necessary information to answer the question. In this process, the MLLM reasons by plotting and analyzing the progress bar, seamlessly integrating both textual and visual reasoning into a Visuotemporal Chain-of-Thought. As illustrated in Fig. 2, VTimeCoT first provides the MLLM with a formatted prompt for initialization and executes actions at each time step " + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "content": ". The tool set includes the , , , etc. The tool is utilized to trim specific segments when the MLLM determines that the video is too long to find the required information. The pseudo-code is presented in Algorithm 1, where " + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "content": " represent the modalities of video and language, respectively. In Algorithm 1, " + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "content": " serves as a parser to retrieve specific sub-strings by keyword, and " + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "content": " is a code builder that integrates the toolset library and generated code." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 689 + ], + "type": "text", + "content": "Initialization Prompt. To enable the MLLM to perform step-by-step reasoning and a structured output, we follow [13] and construct a set of initialization prompts. These prompts define the specific output structure (formatted by keywords of 'THOUGHT', 'ACTION', and 'TERMINATE') for each step and declare the Python tools that the model should invoke. The initialization prompts, along with the question and video frames, are then fed into the MLLM, triggering the iterative loop." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Thought. During this step, the model analyzes the historical context and video memory to generate its reasoning" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24205" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 285 + ], + "type": "image", + "image_path": "d9e22ee9e0ace6f86642d0065b302c600598867d6a2c306822daa9db50b36b28.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 294, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 294, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 555, + 329 + ], + "type": "text", + "content": "Figure 2. Overview of our method. On the left, we demonstrate how the framework iteratively generates thoughts and actions, which dynamically updates the video memory with an overlaid progress bar for reasoning. On the right, we illustrate two novel tools that integrate the frame-sync visual progress bar and highlight key moments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 349, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 295, + 384 + ], + "type": "text", + "content": "for the current step. It determines whether a direct answer can be provided directly in this step or if tool assistance is needed, and if so, which tools should be invoked." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "text", + "content": "Action. Following the thought step, the model generates an executable Python script utilizing the given tool functions. The framework then builds the code with the toolset library and executes the script to manipulate the video frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 494 + ], + "type": "text", + "content": "Dynamic Video Memory. Using the manipulated video frames we subsequently update the frames in the video memory. The updated context with the manipulated video frames is then fed into the MLLM to proceed with the next step of thought." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "content": "To terminate the loop, the model determines at each step whether to cease reasoning and generate the 'TERMINATE' keyword to end. To prevent excessively prolonged reasoning, we impose a maximum step limit, enforcing termination at step " + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 287, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 287, + 577 + ], + "type": "text", + "content": "3.2. Frame-Sync Visual Progress Bar Integration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Enabling MLLM to perceive the precise timestamp of each frame is a significant challenge and an unresolved problem, due to the diverse temporal sampling rates during MLLM training and inference without time. Although recent approaches [15, 16] have attempted to address this issue by incorporating temporal positional embeddings before the video encoder and introducing temporal-grounding training tasks, their flexibility is limited due to the scarcity of specific data and the need for fine-tuning. This issue is particularly pronounced in long videos, where the token limitations of MLLMs prevent them from accurately perceiving" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 347, + 555, + 547 + ], + "blocks": [ + { + "bbox": [ + 313, + 347, + 555, + 547 + ], + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 547 + ], + "type": "table", + "html": "
Algorithm 1 Visuotemporal Chain-of-Thought
Input: {V, Q||V ∈ V, Q ∈ L}; I ∈ L;MLLM: V, L → L; Toolset: {T1, . . ., Tn} ∈ L
1: s ← 0; V0 ← V
2: H0 ← {I, Q}▷ Initialization prompt and query
3: while true do
4: ys ← MLLM (Vs, Hs)
5: THOUGHTs, ACTIONs, TERMINATEs ← p(ys)
6: if not TERMINATEs then
7: Cs ← c(ACTIONs, Toolset)▷ Build code
8: Vs+1 ← Cs(Vs)▷ Update video
9: Hs+1 ← {Hs, ys}▷ Update history
10: else
11: return ys▷ Respond final answer
12: end if
13: end while
", + "image_path": "e0fb44169b00220a446443ae4f25ff1564008d043edb6676df7d3a41ed065d6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 569, + 433, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 433, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 433, + 580 + ], + "type": "text", + "content": "the timestamp of each frame." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "content": "Inspired by visual programming paradigms [12, 39], we propose a simple yet effective method to empower MLLMs to perceive the precise timestamp of each frame, by generating the video progress bar. Leveraging the code generation capabilities of MLLMs, our approach prompts the MLLM to invoke Python plotting tools and use their universal OCR and shape understanding capabilities to perceive time, without any additional training requirements. Specifically, given a raw video " + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "content": " as input, this tool can generate the progress bar at the bottom of the original frames and return the annotated frames. To enable adaptation to ar" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24206" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "binary FPS sampling, we perform a frame-synchronization step to convert frame indexes to the real second time, as shown in Fig. 2. To enable MLLM to perceive seconds directly without any additional hour-minute-second conversions, we design the timestamps using a straightforward seconds format. To construct a robust visual prompt that can be easily interpreted from the MLLM, we employ compositional components using elongated rectangles, circles, and timestamps to plot the progress bar. Finally, the progress bar is placed at the bottom of raw frames, which can be expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 204, + 295, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 204, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 145, + 204, + 295, + 217 + ], + "type": "interline_equation", + "content": "V _ {t} ^ {\\prime} = V _ {t} \\oplus T _ {t}, \\tag {1}", + "image_path": "559e715bfcfb59e0146b76a29f901ca416f8f24f2f17e8821f1fcdeadd13a4f0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "text", + "content": " represents the vertical concatenation operation, " + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "text", + "content": " denotes the image of the progress bar generated by OpenCV at time " + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "inline_equation", + "content": "V_{t}^{\\prime}" + }, + { + "bbox": [ + 55, + 224, + 295, + 284 + ], + "type": "text", + "content": " is the resulting frame. We wrap this integration process into a Python tool function, allowing MLLM to invoke it for generating the annotated video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 282, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 282, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 282, + 305 + ], + "type": "text", + "content": "3.3. Moment Highlight by Video-Text Similarity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 310, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 295, + 441 + ], + "type": "text", + "content": "Although multimodal LLMs have made significant progress in video content captioning, recent studies [15, 16] have revealed substantial deficiencies in their temporal grounding capabilities. Existing works [15, 16] rely on constructing temporal grounding instructional data and require finetuning to emerge such capability. Besides, due to token limitations in MLLMs, they struggle to perform temporal grounding for long videos. In contrast, we propose a training-free and long-video-adaptive method to enhance the temporal grounding capabilities of MLLMs, leveraging the similarities from the robust video-text foundation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 490 + ], + "type": "text", + "content": "Given a query and a video " + }, + { + "bbox": [ + 55, + 441, + 295, + 490 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{T \\times H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 441, + 295, + 490 + ], + "type": "text", + "content": " containing multiple events, our proposed tool estimates the query-relevant temporal segments, including their start and end times, and generates a highlighted visual prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "text", + "content": "Moment Retrieval by Video-Text Similarity. To construct a robust zero-shot moment retrieval module, capable of generalizing across a wide range of videos, we employ the VideoCLIP-XL [42] foundation model as our embedding extractor. The video is initially processed at a high temporal resolution of " + }, + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "text", + "content": " FPS. We group every 8 frames into a video clip, resulting in " + }, + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "text", + "content": " clips. It should be noted that since the temporal retrieval module operates as an external component, it can efficiently handle high frame rates without being constrained by the token count of the MLLM. This makes our method adaptive to long-form video. These clips are then processed by the VideoCLIP-XL visual encoder, producing " + }, + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 491, + 295, + 682 + ], + "type": "text", + "content": " video embeddings. We use VideoCLIP-XL text encoder to extract the text embedding for the text query and we compute the cosine similarity between the query embedding and the video embeddings of each clip:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 690, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 690, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 136, + 690, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} (i) = \\frac {e _ {i} ^ {v} \\cdot e ^ {q}}{\\left| e _ {i} ^ {v} \\right| \\left| e ^ {q} \\right|}, \\tag {2}", + "image_path": "230ec75f94f75c516fd1abbbe427f4ef11c513cfa89d122d85715c9ca9129c2d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "e_i^v" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " denotes the embedding of the " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "-th clip and " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "e^q" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " is the query embedding. Using top-" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " selection, we obtain the " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " clips with the highest similarity and extract the start and end timestamps of each contiguous segment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 288 + ], + "type": "text", + "content": "Moment Highlight on the Progress Bar. Motivated by a series of works that demonstrated the effectiveness of visual cues as a medium for reasoning beyond text [3, 13, 21, 32, 46], we propose using moment-highlighting cues for video temporal reasoning. As shown in Fig. 2, we construct the highlight tool using OpenCV library to plot colorful highlight masks on specific intervals of the progress bar. This tool takes the video frames and time intervals as input and annotates the highlighted progress bar under the frames. To ensure that the MLLM can easily interpret the highlights, we utilize different colors between the mask and the progress bar. We wrap the aforementioned retrieval and highlighting process as a Python tool function, enabling the MLLM to invoke it to manipulate the input video frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 298, + 395, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 395, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 395, + 312 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 553, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 450 + ], + "type": "text", + "content": "In this section, we first present the implementation details, datasets and evaluation metrics. Subsequently, we assess the performance of the proposed framework on the video temporal grounding benchmarks (Sec. 4.1). To further evaluate the reasoning capabilities, we conducted extensive quantitative and qualitative analyses on the \"reasoning based on retrieving\" benchmark of Vript-RR (Sec. 4.2). In addition, we demonstrate the significant role our reasoning framework plays in enhancing long video question-answering performance (Sec. 4.3). Finally, we give the ablations of key modules (Sec. 4.4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "content": "Implementation Details. In the framework, we conducted experiments on two core MLLMs, Qwen2VL-7B [43] and GPT4o-20240513 [17], as they represent the state-of-the-art open-source and closed-source MLLM models, respectively. We set the decoding temperature of the MLLM to 0. By default, the input for the MLLM consists of 32 frames uniformly sampled from the video. For video frames, " + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "content": " are resized to make the longer side 480 pixels. Our agent implementation is based on AutoGen [45]. The implementations of tool functions are adapted from Gradio [1]. For GPT4o, we utilize the OpenAI API service. The details of the MLLM prompts are provided in the supplementary materials. For VideoCLIP-XL we use FPS " + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "content": " to sample video frames and group every 8 frames as a clip to extract embedding. Regarding the top-" + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "content": " selection of clips, " + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "content": " is empirically set to 8. The maximum number of reasoning steps " + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 450, + 555, + 653 + ], + "type": "text", + "content": " is set to 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 701 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. To evaluate the proposed VTimeCoT method, we utilized four benchmark datasets, spanning a wide range of video temporal grounding, reasoning, and long-video question-answering tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Charades-STA [11] is a benchmark dataset for temporal" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24207" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 289, + 215 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 289, + 215 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 289, + 215 + ], + "type": "table", + "html": "
MethodR1@0.3R1@0.5R1@0.7mIoU
VideoChat-7B [23]9.003.301.306.50
VideoLLaMA-7B [56]10.403.800.907.10
VideoChatGPT-7B [31]20.007.701.7013.70
LLAVA-Onevision-7B [20]33.0411.054.1120.98
VTimeLLM-7B [15]51.0027.5011.4031.20
VTimeLLM-13B [15]55.3034.3014.7034.60
Qwen2VL-7B [43]37.3112.854.1124.34
VTimeCoTQwen2VL-7B66.9638.7920.8343.41
GPT4o [17]63.7637.1214.6540.20
VTimeCoTGPT4o74.0651.0222.4546.78
", + "image_path": "1c3f026fbd1d4851d87059b90733015ab6d17d82817ed0e26044add37f65ff4c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 269, + 289, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 222, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 257 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on the Charades-STA dataset for temporal grounding. We report the mIoU and the recall performance at different IoU thresholds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 269, + 289, + 370 + ], + "lines": [ + { + "bbox": [ + 65, + 269, + 289, + 370 + ], + "spans": [ + { + "bbox": [ + 65, + 269, + 289, + 370 + ], + "type": "table", + "html": "
MethodR1@0.3R1@0.5R1@0.7mIoU
LLAVA-Onevision-7B [20]34.9117.919.725.53
VTimeLLM-7B [15]44.5825.039.2928.99
Qwen2VL-7B [43]31.8714.657.3522.77
VTimeCoTQwen2VL-7B67.5045.7925.1146.21
GPT4o [17]55.6135.6819.2937.66
VTimeCoTGPT4o79.3559.7433.8154.49
", + "image_path": "2c86945e2afc04894bc98b39b97eaef15f580717f7fb4fa3f210320c7659abdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": "grounding, composed of 1334 videos along with the corresponding start-end frame annotations of 3720 queries. The average video length is 30 seconds. We follow [15] and report mean IoU (mIoU) and recall@1, IoU " + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\geq m(R@m)" + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": " metrics, where " + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": " includes 0.3, 0.5 and 0.7. IoU denotes the intersection over the union between the predicted and ground truth time segments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 519, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 603 + ], + "type": "text", + "content": "QVHighlights [19] is a benchmark dataset for semantic relevance and temporal grounding from discontinuous time segments. It contains 1519 videos, with an average length of 150 seconds, and 1550 queries with time annotations, where each query spans multiple discontinuous time segments within a video. For evaluation, we report the commonly used mIoU and R@m metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 689 + ], + "type": "text", + "content": "Vript-RR [50] is a challenging benchmark dataset for scene retrieval and multi-hop reasoning. The average length of the videos is 622 seconds. It contains 152 questions, each accompanied by a hint to locate the scene that the question refers to. The benchmark includes both multiple-choice and open-ended question-answering settings. We follow [50] and evaluate the open-ended accuracy using GPT4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "VideoMME [10] is a video question-answering benchmark tailored for MLLMs, featuring diverse videos ranging from" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 343, + 70, + 526, + 225 + ], + "blocks": [ + { + "bbox": [ + 55, + 377, + 295, + 412 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 412 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on the QVHighlights dataset for temporal grounding of discontinuous segments. We report the mIoU and the recall performance at different IoU thresholds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 70, + 526, + 225 + ], + "lines": [ + { + "bbox": [ + 343, + 70, + 526, + 225 + ], + "spans": [ + { + "bbox": [ + 343, + 70, + 526, + 225 + ], + "type": "table", + "html": "
MethodsMulti-ChoiceOpen
VideoChatGPT [31]29.6017.80
Video-LLaMA [56]28.3014.50
VideoChat [23]22.4015.10
VideoChat2 [24]42.1022.40
ST-LLM [28]33.6026.30
PLLAVA 7B [48]55.3036.20
VILA-1.5 8B [27]55.3032.30
Qwen2VL-7B [43]59.8735.95
VTimeCoTQwen2VL-7B62.5041.45
GPT4o [17]70.3961.18
VTimeCoTGPT4o83.5568.42
", + "image_path": "60e75681c3acfc68915109eb2159fa383c62061e5864328faa2310506c197361.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 233, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 268 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on the Vript-RR dataset. We report the response accuracy on both multiple-choice and open-ended question settings." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 289, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 338 + ], + "type": "text", + "content": "11 seconds to 1 hour in length. It contains 900 videos and 2700 question-answer pairs. We follow [10] to report accuracy metrics under two evaluation settings with subtitles and without subtitles." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 347, + 437, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 437, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 437, + 361 + ], + "type": "text", + "content": "4.1. Temporal Grounding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 605 + ], + "type": "text", + "content": "To ensure accurate temporal reasoning, it is essential to achieve robust and high-performance temporal grounding. In Tab. 1, we compare VTimeCoT with two state-of-the-art baselines, Qwen2VL-7B and GPT4o, to identify the start and end timestamps of the query event. Specifically, using both Qwen2VL-7B and GPT4o models as the core MLLM of our framework, we evaluate the performance of VTimeCoT to localize the query event. As can be easily observed, although these advanced MLLMs achieve state-of-the-art performance in video content recognition, they fall short in accurately grounding temporal boundaries for events in videos, facing difficulties in perceiving the precise timestamp and the boundaries of events. In contrast, the proposed method effectively addresses these challenges, providing accurate visual cues of the progress bar through the reasoning process that facilitates the grounding performance in videos. Note that VTimeCoT demonstrates superior performance without any additional training cost, even compared to the VTimeLLM model, which has been meticulously fine-tuned on extensive temporal grounding data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 715 + ], + "type": "text", + "content": "To further evaluate the temporal grounding performance of the proposed and the baseline methods in the presence of multiple discontinuous time segments for each query, we utilized the QVHighlights benchmark dataset (Tab. 2). Given a query and a video, we request from each model to identify start-end timestamps of all temporal segments in the video that fit the query. As can be observed, VTime-CoT consistently enhances the performance compared to different MLLMs. On the contrary, baseline MLLM meth" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24208" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 102, + 79, + 251, + 89 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 79, + 251, + 89 + ], + "spans": [ + { + "bbox": [ + 102, + 79, + 251, + 89 + ], + "type": "text", + "content": "Reasoning based on retrieving scenes" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 96, + 294, + 137 + ], + "blocks": [ + { + "bbox": [ + 58, + 96, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 96, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 96, + 294, + 137 + ], + "type": "image", + "image_path": "22ae648c1b8e507545ed2fbd81e1f13dc5919014f71f7020ce376d7f4f25660b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 144, + 269, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 144, + 269, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 144, + 269, + 171 + ], + "type": "text", + "content": "? Hint: A scene in the video features a first-person perspective from a motorcyclist cruising on a highway. ... The motorcycle's chrome details, black leather accessories, the rider's gloved hands are noticeable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 177, + 274, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 274, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 274, + 185 + ], + "type": "text", + "content": "What is the round object on the bike in the middle bottom of the screen?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 186, + 113, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 113, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 113, + 192 + ], + "type": "text", + "content": "(A) A GPS device" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 186, + 225, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 186, + 225, + 193 + ], + "spans": [ + { + "bbox": [ + 171, + 186, + 225, + 193 + ], + "type": "text", + "content": "(B) A sport camera" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 194, + 149, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 149, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 149, + 201 + ], + "type": "text", + "content": "(C) The motorcycle's headlight" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 194, + 265, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 194, + 265, + 201 + ], + "spans": [ + { + "bbox": [ + 171, + 194, + 265, + 201 + ], + "type": "text", + "content": "(D) The motorcycle's spedometer" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 215, + 79, + 232 + ], + "blocks": [ + { + "bbox": [ + 62, + 215, + 79, + 232 + ], + "lines": [ + { + "bbox": [ + 62, + 215, + 79, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 79, + 232 + ], + "type": "image", + "image_path": "c6a20ce201c7aad3b84205d472ec823da16927711521c1a1b96c481480360002.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on Vript-RR benchmark. In contrast to GPT-4o which is misled by the irrelevant scenes, VTimeCoT reasons using the progress bar and highlights the key moments, leading to accurate answers." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 221, + 201, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 221, + 201, + 232 + ], + "spans": [ + { + "bbox": [ + 83, + 221, + 201, + 232 + ], + "type": "text", + "content": "GPT40: (C) The motorcycle's headlight." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 274, + 218, + 285, + 231 + ], + "blocks": [ + { + "bbox": [ + 274, + 218, + 285, + 231 + ], + "lines": [ + { + "bbox": [ + 274, + 218, + 285, + 231 + ], + "spans": [ + { + "bbox": [ + 274, + 218, + 285, + 231 + ], + "type": "image", + "image_path": "4e96aab5b6c8c4d153a2444d3354fde9bea27ec62754aefefca67d847258e7f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 74, + 325, + 88 + ], + "blocks": [ + { + "bbox": [ + 310, + 74, + 325, + 88 + ], + "lines": [ + { + "bbox": [ + 310, + 74, + 325, + 88 + ], + "spans": [ + { + "bbox": [ + 310, + 74, + 325, + 88 + ], + "type": "image", + "image_path": "f60ecb40aab8634bea755dfb4513177b5a75c30fe8cf24aee5a014d8285a8447.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 76, + 405, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 76, + 405, + 83 + ], + "spans": [ + { + "bbox": [ + 328, + 76, + 405, + 83 + ], + "type": "text", + "content": "VTimeCoT + GPT40" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 91, + 541, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 541, + 121 + ], + "type": "text", + "content": "THOUGHT0: I can use the progress_bar and highlight tool to highlight the key moments. Since the video is too long, I will cut the video into segments for a thorough review. Then I will display the highlighted frames to help me answer the question." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 126, + 345, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 126, + 345, + 134 + ], + "spans": [ + { + "bbox": [ + 307, + 126, + 345, + 134 + ], + "type": "text", + "content": "ACTION 0:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 135, + 432, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 135, + 432, + 142 + ], + "spans": [ + { + "bbox": [ + 307, + 135, + 432, + 142 + ], + "type": "text", + "content": "sentence = 'A scene in the video features ...'" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 142, + 392, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 392, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 392, + 148 + ], + "type": "text", + "content": "frames = progress_bar(frames)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 148, + 498, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 498, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 498, + 161 + ], + "type": "text", + "content": "frames, segment = highlight(frames, sentence, return_segment=True) \nframes = cut(frames, segment)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 161, + 366, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 366, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 366, + 167 + ], + "type": "text", + "content": "for frame in frames:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 167, + 361, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 167, + 361, + 174 + ], + "spans": [ + { + "bbox": [ + 319, + 167, + 361, + 174 + ], + "type": "text", + "content": "display(frame)" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 508, + 146, + 544, + 167 + ], + "blocks": [ + { + "bbox": [ + 508, + 146, + 544, + 167 + ], + "lines": [ + { + "bbox": [ + 508, + 146, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 508, + 146, + 544, + 167 + ], + "type": "image", + "image_path": "38551943a79adf478be4d4215b5ab39e4b0b5f0bdaf5641c47cafa4fcf1b62ff.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 304, + 178, + 361, + 215 + ], + "blocks": [ + { + "bbox": [ + 304, + 178, + 361, + 215 + ], + "lines": [ + { + "bbox": [ + 304, + 178, + 361, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 361, + 215 + ], + "type": "image", + "image_path": "214e3432346de37308894ffe776c07979fd129887fe4b6740753040cd8c5ad1e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 361, + 178, + 417, + 215 + ], + "blocks": [ + { + "bbox": [ + 361, + 178, + 417, + 215 + ], + "lines": [ + { + "bbox": [ + 361, + 178, + 417, + 215 + ], + "spans": [ + { + "bbox": [ + 361, + 178, + 417, + 215 + ], + "type": "image", + "image_path": "6982f63d0be24433d900774ff85e9813a1ec0a9fb1ef9485fcc6c499d7f4c8b5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 419, + 178, + 476, + 215 + ], + "blocks": [ + { + "bbox": [ + 419, + 178, + 476, + 215 + ], + "lines": [ + { + "bbox": [ + 419, + 178, + 476, + 215 + ], + "spans": [ + { + "bbox": [ + 419, + 178, + 476, + 215 + ], + "type": "image", + "image_path": "c013d4d9a6a3254278445b87f7e98f96f942882b5bcdb2696c57efa0aab0065c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 477, + 178, + 533, + 215 + ], + "blocks": [ + { + "bbox": [ + 477, + 178, + 533, + 215 + ], + "lines": [ + { + "bbox": [ + 477, + 178, + 533, + 215 + ], + "spans": [ + { + "bbox": [ + 477, + 178, + 533, + 215 + ], + "type": "image", + "image_path": "518e5f8b7721179ada3447eb2e4d69575395bb216c5e803186ba74fb2d7aab2d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 216, + 521, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 216, + 521, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 216, + 521, + 239 + ], + "type": "text", + "content": "THOUGHT 1: From the highlighted frames, I can observe the details of the motorcycle's features. The round object fits a motorcycle's speedmeter. ANSWER: (D) The motorcycle's speedometer. TERMINATE" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 528, + 193, + 542, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 528, + 193, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 528, + 193, + 542, + 198 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 123, + 290, + 212, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 290, + 212, + 301 + ], + "spans": [ + { + "bbox": [ + 123, + 290, + 212, + 301 + ], + "type": "text", + "content": "Temporal Counting" + } + ] + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 58, + 311, + 118, + 343 + ], + "blocks": [ + { + "bbox": [ + 58, + 311, + 118, + 343 + ], + "lines": [ + { + "bbox": [ + 58, + 311, + 118, + 343 + ], + "spans": [ + { + "bbox": [ + 58, + 311, + 118, + 343 + ], + "type": "image", + "image_path": "1de9e0c0058ed449ccd702978bce08ec7cf1f6a97f7dc63d38d78112056281c6.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 119, + 311, + 174, + 343 + ], + "blocks": [ + { + "bbox": [ + 119, + 311, + 174, + 343 + ], + "lines": [ + { + "bbox": [ + 119, + 311, + 174, + 343 + ], + "spans": [ + { + "bbox": [ + 119, + 311, + 174, + 343 + ], + "type": "image", + "image_path": "df453bcaa0ab5e3aa19cae3c2f83433c6113c53d544fae8549ee3380e5613645.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 175, + 311, + 233, + 343 + ], + "blocks": [ + { + "bbox": [ + 175, + 311, + 233, + 343 + ], + "lines": [ + { + "bbox": [ + 175, + 311, + 233, + 343 + ], + "spans": [ + { + "bbox": [ + 175, + 311, + 233, + 343 + ], + "type": "image", + "image_path": "18dfc0e110167516cdc5de5ff162d140495a817aa85cb850e3edb183aff782be.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 235, + 311, + 299, + 343 + ], + "blocks": [ + { + "bbox": [ + 235, + 311, + 299, + 343 + ], + "lines": [ + { + "bbox": [ + 235, + 311, + 299, + 343 + ], + "spans": [ + { + "bbox": [ + 235, + 311, + 299, + 343 + ], + "type": "image", + "image_path": "ffb3986c6a8f97c20d01950a0c6fbee82955d555f8126c594330f414fad0af0d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 58, + 354, + 80, + 376 + ], + "blocks": [ + { + "bbox": [ + 58, + 354, + 80, + 376 + ], + "lines": [ + { + "bbox": [ + 58, + 354, + 80, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 354, + 80, + 376 + ], + "type": "image", + "image_path": "da346dbd329ff689538753e3ce4c302cffd52f74392d2d930521ef62d9be286b.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 357, + 282, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 357, + 282, + 365 + ], + "spans": [ + { + "bbox": [ + 89, + 357, + 282, + 365 + ], + "type": "text", + "content": "How many different kinds of animal faces are made in this video?" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 365, + 102, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 365, + 102, + 373 + ], + "spans": [ + { + "bbox": [ + 89, + 365, + 102, + 373 + ], + "type": "text", + "content": "A. 4" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 118, + 365, + 124, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 365, + 124, + 372 + ], + "spans": [ + { + "bbox": [ + 118, + 365, + 124, + 372 + ], + "type": "text", + "content": "B." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 145, + 365, + 156, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 365, + 156, + 372 + ], + "spans": [ + { + "bbox": [ + 145, + 365, + 156, + 372 + ], + "type": "text", + "content": "C. 5" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 177, + 365, + 188, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 365, + 188, + 372 + ], + "spans": [ + { + "bbox": [ + 177, + 365, + 188, + 372 + ], + "type": "text", + "content": "D. 2" + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 63, + 384, + 80, + 400 + ], + "blocks": [ + { + "bbox": [ + 63, + 384, + 80, + 400 + ], + "lines": [ + { + "bbox": [ + 63, + 384, + 80, + 400 + ], + "spans": [ + { + "bbox": [ + 63, + 384, + 80, + 400 + ], + "type": "image", + "image_path": "eabaf48c67b154ee77f8be675dbd9b1c550e7c170631e9d938c32b8d2907b4a2.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 89, + 388, + 131, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 388, + 131, + 397 + ], + "spans": [ + { + "bbox": [ + 89, + 388, + 131, + 397 + ], + "type": "text", + "content": "GPT4o:A.4" + } + ] + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 150, + 388, + 162, + 399 + ], + "blocks": [ + { + "bbox": [ + 150, + 388, + 162, + 399 + ], + "lines": [ + { + "bbox": [ + 150, + 388, + 162, + 399 + ], + "spans": [ + { + "bbox": [ + 150, + 388, + 162, + 399 + ], + "type": "image", + "image_path": "eb86131d575065ddb3139aa3dac7063737b167d41e3e6bc12eaf1b279c1d6770.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 182, + 385, + 199, + 401 + ], + "blocks": [ + { + "bbox": [ + 182, + 385, + 199, + 401 + ], + "lines": [ + { + "bbox": [ + 182, + 385, + 199, + 401 + ], + "spans": [ + { + "bbox": [ + 182, + 385, + 199, + 401 + ], + "type": "image", + "image_path": "c31125daf2aa54b70df4b974c734d9988ad96b73e4146a19dd9469443ec3db44.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 203, + 388, + 283, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 388, + 283, + 397 + ], + "spans": [ + { + "bbox": [ + 203, + 388, + 283, + 397 + ], + "type": "text", + "content": "VTimeCoT + GPT4o" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 59, + 411, + 284, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 411, + 284, + 442 + ], + "spans": [ + { + "bbox": [ + 59, + 411, + 284, + 442 + ], + "type": "text", + "content": "THOUGHT 0: I can use the progress_bar and the highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for 'different kinds of animal faces' to highlight them. I will then display the highlighted frames to help me answer the question." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 62, + 453, + 94, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 94, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 94, + 460 + ], + "type": "text", + "content": "ACTION 0:" + } + ] + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 99, + 453, + 106, + 460 + ], + "blocks": [ + { + "bbox": [ + 99, + 453, + 106, + 460 + ], + "lines": [ + { + "bbox": [ + 99, + 453, + 106, + 460 + ], + "spans": [ + { + "bbox": [ + 99, + 453, + 106, + 460 + ], + "type": "image", + "image_path": "203b280ed43b686e591ad06c4982d8237173cc8b50cf16523140d0f9e6ed04db.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 587, + 555, + 609 + ], + "lines": [ + { + "bbox": [ + 55, + 587, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 555, + 609 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on VideoMME benchmark. Through reasoning that integrates the progress bar and highlights, VTimeCoT accurately answers questions related to temporal counting and order, which GPT-4o falls short." + } + ] + } + ], + "index": 81, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 66, + 464, + 196, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 464, + 196, + 470 + ], + "spans": [ + { + "bbox": [ + 66, + 464, + 196, + 470 + ], + "type": "text", + "content": "sentence = \"different kinds of animal faces\"" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 66, + 471, + 153, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 471, + 153, + 477 + ], + "spans": [ + { + "bbox": [ + 66, + 471, + 153, + 477 + ], + "type": "text", + "content": "frames = progress_bar(frames)" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 66, + 477, + 173, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 477, + 173, + 483 + ], + "spans": [ + { + "bbox": [ + 66, + 477, + 173, + 483 + ], + "type": "text", + "content": "frames = highlight(frame, sentence)" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 66, + 484, + 126, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 484, + 126, + 490 + ], + "spans": [ + { + "bbox": [ + 66, + 484, + 126, + 490 + ], + "type": "text", + "content": "for frame in frames:" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 78, + 491, + 120, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 491, + 120, + 496 + ], + "spans": [ + { + "bbox": [ + 78, + 491, + 120, + 496 + ], + "type": "text", + "content": "display(frame)" + } + ] + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 258, + 467, + 293, + 487 + ], + "blocks": [ + { + "bbox": [ + 258, + 467, + 293, + 487 + ], + "lines": [ + { + "bbox": [ + 258, + 467, + 293, + 487 + ], + "spans": [ + { + "bbox": [ + 258, + 467, + 293, + 487 + ], + "type": "image", + "image_path": "515e82aaf7031981aab48ac2f627f6a38d95f68934020a48c60bfcf969457c70.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 59, + 502, + 116, + 540 + ], + "blocks": [ + { + "bbox": [ + 59, + 502, + 116, + 540 + ], + "lines": [ + { + "bbox": [ + 59, + 502, + 116, + 540 + ], + "spans": [ + { + "bbox": [ + 59, + 502, + 116, + 540 + ], + "type": "image", + "image_path": "9ff557e762b51cfb1e13af427cdbb15ec9ededdce7eb296870671df08564ea13.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 118, + 502, + 174, + 540 + ], + "blocks": [ + { + "bbox": [ + 118, + 502, + 174, + 540 + ], + "lines": [ + { + "bbox": [ + 118, + 502, + 174, + 540 + ], + "spans": [ + { + "bbox": [ + 118, + 502, + 174, + 540 + ], + "type": "image", + "image_path": "a58c43ae90839f113e038e753c7154666c6fb7edbd745c752c2000e8a4efb59e.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 175, + 502, + 232, + 540 + ], + "blocks": [ + { + "bbox": [ + 175, + 502, + 232, + 540 + ], + "lines": [ + { + "bbox": [ + 175, + 502, + 232, + 540 + ], + "spans": [ + { + "bbox": [ + 175, + 502, + 232, + 540 + ], + "type": "image", + "image_path": "65cce358efdf273e18aed48cc53585089d5109f5036484dd5ca3bbf928c04b2b.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 233, + 503, + 296, + 541 + ], + "blocks": [ + { + "bbox": [ + 233, + 503, + 296, + 541 + ], + "lines": [ + { + "bbox": [ + 233, + 503, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 233, + 503, + 296, + 541 + ], + "type": "image", + "image_path": "36486a1295d9999cb54b7498bd252ca2872436baa3de85f0612d84899d4e2b7a.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "bbox": [ + 62, + 545, + 294, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 294, + 569 + ], + "type": "text", + "content": "THOUGHT 1: From the highlighted frames, I can count three different kinds of animal faces being made: a cat, a pig, and a dog. Therefore, the correct answer is 3. ANSWER: B TERMINATE" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 389, + 290, + 463, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 290, + 463, + 301 + ], + "spans": [ + { + "bbox": [ + 389, + 290, + 463, + 301 + ], + "type": "text", + "content": "Temporal Order" + } + ] + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 306, + 311, + 365, + 343 + ], + "blocks": [ + { + "bbox": [ + 306, + 311, + 365, + 343 + ], + "lines": [ + { + "bbox": [ + 306, + 311, + 365, + 343 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 365, + 343 + ], + "type": "image", + "image_path": "cb5677d6d17ea20abe0865f047f9db9b9942802b115989220fab96370409f755.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 366, + 311, + 425, + 343 + ], + "blocks": [ + { + "bbox": [ + 366, + 311, + 425, + 343 + ], + "lines": [ + { + "bbox": [ + 366, + 311, + 425, + 343 + ], + "spans": [ + { + "bbox": [ + 366, + 311, + 425, + 343 + ], + "type": "image", + "image_path": "5d521d0e788e868e4ddb8da0264314a1021785cbbc2776b73ddbb4d01dea6a54.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 425, + 311, + 482, + 343 + ], + "blocks": [ + { + "bbox": [ + 425, + 311, + 482, + 343 + ], + "lines": [ + { + "bbox": [ + 425, + 311, + 482, + 343 + ], + "spans": [ + { + "bbox": [ + 425, + 311, + 482, + 343 + ], + "type": "image", + "image_path": "7f83eea998341443282a3c892c31d129e7d83b8327d6c1b7bb8879233e851c37.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 483, + 311, + 543, + 343 + ], + "blocks": [ + { + "bbox": [ + 483, + 311, + 543, + 343 + ], + "lines": [ + { + "bbox": [ + 483, + 311, + 543, + 343 + ], + "spans": [ + { + "bbox": [ + 483, + 311, + 543, + 343 + ], + "type": "image", + "image_path": "75014981f8ad94ff2f6d0fe17eef69cd5d95942b74c35221fa1225205a978c47.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 306, + 354, + 328, + 376 + ], + "blocks": [ + { + "bbox": [ + 306, + 354, + 328, + 376 + ], + "lines": [ + { + "bbox": [ + 306, + 354, + 328, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 328, + 376 + ], + "type": "image", + "image_path": "517e3b2316a670aa1d7f21a975b2cdfbc57d06ebc376ad92a7a696885b34851b.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "bbox": [ + 337, + 357, + 535, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 357, + 535, + 373 + ], + "spans": [ + { + "bbox": [ + 337, + 357, + 535, + 373 + ], + "type": "text", + "content": "What was the exercise demonstrated after the squatting exercise in the video? A. Push-ups B. Plank C. Step-ups D. Pull-ups" + } + ] + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 311, + 384, + 328, + 401 + ], + "blocks": [ + { + "bbox": [ + 311, + 384, + 328, + 401 + ], + "lines": [ + { + "bbox": [ + 311, + 384, + 328, + 401 + ], + "spans": [ + { + "bbox": [ + 311, + 384, + 328, + 401 + ], + "type": "image", + "image_path": "fcca75f1f70759257db5615476e7ae126eb4f11b0011694644299927fd2ca8c6.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "bbox": [ + 337, + 388, + 527, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 388, + 527, + 401 + ], + "spans": [ + { + "bbox": [ + 337, + 388, + 527, + 401 + ], + "type": "text", + "content": "GPT4o:B.Plank VTimeCoT+GPT4o" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 309, + 411, + 542, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 411, + 542, + 442 + ], + "spans": [ + { + "bbox": [ + 309, + 411, + 542, + 442 + ], + "type": "text", + "content": "THOUGHT 0: I can use the progress_bar and highlight tool. This tool will draw a progress bar on the video frames and detect the moment proposals for `exercise demonstrated after the squatting exercise` to highlight them. I will then display the highlighted frames to help me answer the question." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 310, + 452, + 343, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 452, + 343, + 459 + ], + "spans": [ + { + "bbox": [ + 310, + 452, + 343, + 459 + ], + "type": "text", + "content": "ACTION 0:" + } + ] + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 344, + 451, + 356, + 462 + ], + "blocks": [ + { + "bbox": [ + 344, + 451, + 356, + 462 + ], + "lines": [ + { + "bbox": [ + 344, + 451, + 356, + 462 + ], + "spans": [ + { + "bbox": [ + 344, + 451, + 356, + 462 + ], + "type": "image", + "image_path": "61ebf001ece87ab30bbca97f8f045fddfa682cd5864d5cffb65f375c0ee910ed.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "bbox": [ + 315, + 464, + 502, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 464, + 502, + 477 + ], + "spans": [ + { + "bbox": [ + 315, + 464, + 502, + 477 + ], + "type": "text", + "content": "sentence = \"exercise demonstrated after the squatting exercise\" \nframes = progress_barimentary" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 315, + 477, + 422, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 477, + 422, + 483 + ], + "spans": [ + { + "bbox": [ + 315, + 477, + 422, + 483 + ], + "type": "text", + "content": "frames = highlight(frame, sentence)" + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 315, + 484, + 375, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 484, + 375, + 490 + ], + "spans": [ + { + "bbox": [ + 315, + 484, + 375, + 490 + ], + "type": "text", + "content": "for frame in frames:" + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 328, + 491, + 369, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 491, + 369, + 496 + ], + "spans": [ + { + "bbox": [ + 328, + 491, + 369, + 496 + ], + "type": "text", + "content": "display(frame)" + } + ] + } + ], + "index": 73 + }, + { + "type": "image", + "bbox": [ + 507, + 468, + 542, + 487 + ], + "blocks": [ + { + "bbox": [ + 507, + 468, + 542, + 487 + ], + "lines": [ + { + "bbox": [ + 507, + 468, + 542, + 487 + ], + "spans": [ + { + "bbox": [ + 507, + 468, + 542, + 487 + ], + "type": "image", + "image_path": "f2550e143593791fc000b7e264c3fa10f255a6ca29d0d0836e24e5726a7ea912.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "image_body" + } + ], + "index": 74 + }, + { + "type": "image", + "bbox": [ + 309, + 501, + 366, + 539 + ], + "blocks": [ + { + "bbox": [ + 309, + 501, + 366, + 539 + ], + "lines": [ + { + "bbox": [ + 309, + 501, + 366, + 539 + ], + "spans": [ + { + "bbox": [ + 309, + 501, + 366, + 539 + ], + "type": "image", + "image_path": "b44e7e7a84358bd05b66be4f345d45a3b49a69425e05bb4f2bf14a1a7096fd3d.jpg" + } + ] + } + ], + "index": 75, + "angle": 0, + "type": "image_body" + } + ], + "index": 75 + }, + { + "type": "image", + "bbox": [ + 367, + 501, + 424, + 538 + ], + "blocks": [ + { + "bbox": [ + 367, + 501, + 424, + 538 + ], + "lines": [ + { + "bbox": [ + 367, + 501, + 424, + 538 + ], + "spans": [ + { + "bbox": [ + 367, + 501, + 424, + 538 + ], + "type": "image", + "image_path": "5879423196fdbe57056d48089100af2ac9729cc828027d6b7a4fef314778946e.jpg" + } + ] + } + ], + "index": 76, + "angle": 0, + "type": "image_body" + } + ], + "index": 76 + }, + { + "type": "image", + "bbox": [ + 425, + 501, + 482, + 538 + ], + "blocks": [ + { + "bbox": [ + 425, + 501, + 482, + 538 + ], + "lines": [ + { + "bbox": [ + 425, + 501, + 482, + 538 + ], + "spans": [ + { + "bbox": [ + 425, + 501, + 482, + 538 + ], + "type": "image", + "image_path": "6cae95118332a5a6f853a52ce6e9f13d0050cc593d4067af6b5057bbf239a574.jpg" + } + ] + } + ], + "index": 77, + "angle": 0, + "type": "image_body" + } + ], + "index": 77 + }, + { + "type": "image", + "bbox": [ + 483, + 501, + 547, + 539 + ], + "blocks": [ + { + "bbox": [ + 483, + 501, + 547, + 539 + ], + "lines": [ + { + "bbox": [ + 483, + 501, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 483, + 501, + 547, + 539 + ], + "type": "image", + "image_path": "5864b1026c86667038a04f7baf29081d52fd852c61159ba3814473b0674fac8b.jpg" + } + ] + } + ], + "index": 78, + "angle": 0, + "type": "image_body" + } + ], + "index": 78 + }, + { + "bbox": [ + 315, + 540, + 520, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 540, + 520, + 557 + ], + "spans": [ + { + "bbox": [ + 315, + 540, + 520, + 557 + ], + "type": "text", + "content": "THOUGHT 1: The highlighted frames show the sequence of exercises demonstrated in the video. After the squatting exercise, the next" + } + ] + } + ], + "index": 79 + }, + { + "bbox": [ + 315, + 558, + 535, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 558, + 535, + 571 + ], + "spans": [ + { + "bbox": [ + 315, + 558, + 535, + 571 + ], + "type": "text", + "content": "demonstrated exercise is step-ups, as shown in the frames around the 21-second mark. ANSWER: C TERMINATE" + } + ] + } + ], + "index": 80 + }, + { + "bbox": [ + 55, + 631, + 295, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 295, + 667 + ], + "type": "text", + "content": "ods struggle with discontinuous time segments, failing to identify multiple temporal spans and leading to degraded grounding performance." + } + ] + } + ], + "index": 82 + }, + { + "bbox": [ + 55, + 672, + 282, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 282, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 282, + 685 + ], + "type": "text", + "content": "4.2. Long-Video Reasoning Based on Retrieving" + } + ] + } + ], + "index": 83 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In Tab. 3, we evaluate the question-answering performance of the proposed and the baseline methods on the Vript-RR" + } + ] + } + ], + "index": 84 + }, + { + "bbox": [ + 313, + 631, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 555, + 689 + ], + "type": "text", + "content": "benchmark, which contains questions that require scene retrieval. Given a scene hint and a question, the model must identify the specific scene to deliver an accurate response. We compared the proposed method with the state-of-the-art baseline methods including Qwen2VL-7B and GPT4o." + } + ] + } + ], + "index": 85 + }, + { + "bbox": [ + 315, + 690, + 553, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 690, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 315, + 690, + 553, + 703 + ], + "type": "text", + "content": "As can be observed VTimeCoT achieves a significant per" + } + ] + } + ], + "index": 86 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24209" + } + ] + } + ], + "index": 87 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 268 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 268 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 268 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 268 + ], + "type": "table", + "html": "
MethodsFramesAccuracy
w/o subsw. subs
Gemini-1.5-Flash [40]1 fps70.375.0
GPT4o [17]38471.977.2
Gemini-1.5-Pro [40]1 fps75.081.3
Video-LLaVA-7B [26]839.941.6
VideoLLaMA2-7B [6]1654.956.4
LLAVA-Onevision-7B [20]3258.261.5
LongVILA-7B [49]25660.165.1
InternVL2.5-8B [5]6464.266.9
mPLUG-Owl3-7B [54]12859.368.1
LLaVA-Video-7B [58]6463.369.7
NVILA-8B [30]25664.270.0
VideoLLaMA3-7B [55]18066.270.3
GPT4o [17]3261.665.1
VTimeCoTgpt4o3264.273.6
", + "image_path": "a1675623bd2d99f74012aa7b15fb16437bdaa664035a15cb17b5f2b6f72765e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 330, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 295, + 354 + ], + "type": "text", + "content": "formance improvement, surpassing the baseline methods in both multiple-choice and open-ended settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": "To further illustrate the superiority of the proposed method in the task of scene retrieval and reasoning, we qualitatively compare GPT4o and VTimeCoT on the Vript-RR dataset in Fig. 3. Through multi-step reasoning that leverages the progress bar and highlights the key time interval, VTimeCoT accurately identifies the correct answer, while GPT-4o falls short." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 231, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 231, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 231, + 458 + ], + "type": "text", + "content": "4.3. Long-Video Question Answering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 629 + ], + "type": "text", + "content": "A pivotal challenge in long-form video question answering lies in accurately inferring the frequency and sequential relationships of temporal events. Currently, end-to-end video MLLMs struggle to accurately localize events and capture their temporal dependencies across multiple frames, limiting their ability to generalize to real-world temporal understanding. In Tab. 4, we evaluate the performance of the proposed VTimeCoT model in question answering on long videos of the VideoMME benchmark dataset. Due to the prohibitive evaluation costs and to ensure a fair comparison, we re-evaluate GPT-4o using 32 frames. Utilizing robust temporal grounding and reasoning, VTimeCoT significantly outperforms the baseline methods in long-video question answering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "The effectiveness of the proposed method can be further validated in Fig. 4, where we compared the responses of the state-of-the-art GPT4o and VTimeCoT. In contrast to GPT4o, VTimeCoT can accurately identify the correct answers through its systematic, step-by-step reasoning process, even in challenging scenarios involving temporal counting and order discernment. It is also worth mention-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 561, + 149 + ], + "blocks": [ + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on the VideoMME dataset. We report the response accuracy under two settings: with subtitles and without subtitles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 561, + 149 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 561, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 561, + 149 + ], + "type": "table", + "html": "
MLLMCoTProgress BarHighlightQVHighlights (mIoU)Vript-RR (M-Acc)
GPT4o×××37.6670.39
GPT4o✓××41.8573.68
GPT4o✓✓×49.4076.32
GPT4o✓✓✓54.4983.55
", + "image_path": "118e42a2a88bb8825b77813660eac0505360d2e073936225cac8add747023cce.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 157, + 555, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 180 + ], + "type": "text", + "content": "Table 5. Ablations on the key modules. M-Acc is the multi-choice accuracy on Vript-RR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "type": "text", + "content": "ing that VTimeCoT, apart from its superior performance in question-answering accuracy, demonstrates logical and interpretable reasoning steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 245, + 382, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 382, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 382, + 257 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 263, + 555, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 466 + ], + "type": "text", + "content": "To further investigate the contribution of each component in the proposed method, we conducted an ablation study. In Tab. 5, we evaluate the importance of the Chain-of-Thought (CoT), progress bar, and highlighting modules. As can be observed the GPT4o employing standard CoT without the use of the progress bar (i.e., solely on text-based step-by-step reasoning), results in significant performance degradation compared to the proposed method. To assess the effect of the proposed modules of the progress bar, we developed a configuration that utilizes only the progress bar tool without using the highlighting module. For completion, we also report the full VTimeCoT results, utilizing both the progress bar and highlighting tools. Both tools contribute to the performance improvement of the proposed method, which highlights the effectiveness of explicitly visualizing the precise timestamp and highlighting the relevant moments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 478, + 388, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 490 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "In this work, we propose the first visual time framework, designed to formulate a visuotemporal chain-of-thought for video temporal grounding and reasoning. We introduce a plug-and-play tool of the progress bar to generate visual temporal cues from any video, that can enable multimodal large language models (MLLMs) to leverage tool-usage capabilities for accurately perceiving the speed of temporal progression. In addition, we propose a zero-shot temporal retrieval tool, built on top of a strong video-text foundational model, augmenting it with temporal grounding capabilities without any further training requirements. We integrate these two tools into a visuotemporal chain-of-thought framework to facilitate cross-modal reasoning between video and text. Through extensive experiments, we demonstrate that the proposed method surpasses previous state-of-the-art baselines in temporal grounding and reasoning-based question-answering benchmarks, demonstrating logical and interpretable reasoning steps." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24210" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by NSFC (62322113, 62376156), Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 152, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 296, + 195 + ], + "type": "text", + "content": "[1] Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. Gradio: Hassle-free sharing and testing of ml models in the wild. arXiv preprint arXiv:1906.02569, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 197, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 197, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 197, + 296, + 251 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 253, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 295, + 297 + ], + "type": "text", + "content": "[3] Syeda Nahida Akter, Aman Madaan, Sangwu Lee, Yiming Yang, and Eric Nyberg. Self-imagine: Effective unimodal reasoning with multimodal models using self-imagination. arXiv preprint arXiv:2401.08025, 2024. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 296, + 363 + ], + "type": "text", + "content": "[4] Mu Cai, Haotian Liu, Siva Karthik Mustikovela, Gregory P Meyer, Yuning Chai, Dennis Park, and Yong Jae Lee. Vipllava: Making large multimodal models understand arbitrary visual prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12914-12923, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 421 + ], + "type": "text", + "content": "[5] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "type": "text", + "content": "[6] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "text", + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 577 + ], + "type": "text", + "content": "[8] Rohan Choudhury, Koichiro Niinuma, Kris M Kitani, and László A Jeni. Zero-shot video question answering with procedural programs. arXiv preprint arXiv:2312.00937, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "text", + "content": "[9] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[10] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[11] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[12] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[13] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 274 + ], + "type": "text", + "content": "[14] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9590-9601, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 330 + ], + "type": "text", + "content": "[15] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "type": "text", + "content": "[16] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 431 + ], + "type": "text", + "content": "[17] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 5, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 487 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 533 + ], + "type": "text", + "content": "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[20] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "type": "text", + "content": "[21] Chengzu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "text", + "content": "[22] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 2, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24211" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[24] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "type": "text", + "content": "[25] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 229 + ], + "type": "text", + "content": "[26] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "type": "text", + "content": "[27] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 331 + ], + "type": "text", + "content": "[28] Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 387 + ], + "type": "text", + "content": "[29] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, et al. Llava-plus: Learning to use tools for creating multimodal agents. In European Conference on Computer Vision, pages 126-142. Springer, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "text", + "content": "[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, et al. Nvila: Efficient frontier visual language models. arXiv preprint arXiv:2412.04468, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 477 + ], + "type": "text", + "content": "[31] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 511 + ], + "type": "text", + "content": "[32] Sachit Menon, Richard Zemel, and Carl Vondrick. Whiteboard-of-thought: Thinking step-by-step across modalities. arXiv preprint arXiv:2406.14562, 2024. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 567 + ], + "type": "text", + "content": "[33] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 623 + ], + "type": "text", + "content": "[34] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 12242-12254, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[37] Chuyi Shang, Amos You, Sanjay Subramanian, Trevor Darrell, and Roei Herzig. Traveler: A modular multi-lmm agent framework for video question-answering. arXiv preprint arXiv:2404.01476, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "text", + "content": "[38] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11987-11997, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 555, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 253 + ], + "type": "text", + "content": "[39] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "text", + "content": "[40] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "type": "text", + "content": "[41] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 355, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 399 + ], + "type": "text", + "content": "[42] Jiapeng Wang, Chengyu Wang, Kunzhe Huang, Jun Huang, and Lianwen Jin. Videoclip-xl: Advancing long description understanding for video clip models. arXiv preprint arXiv:2410.00741, 2024.5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 555, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 454 + ], + "type": "text", + "content": "[43] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 555, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 500 + ], + "type": "text", + "content": "[44] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 555, + 555 + ], + "type": "text", + "content": "[45] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 555, + 612 + ], + "type": "text", + "content": "[46] Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. Mind's eye of llms: visualization-of-thought elicits spatial reasoning in large language models. Advances in Neural Information Processing Systems, 37:90277-90317, 2025. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 669 + ], + "type": "text", + "content": "[47] Yixuan Wu, Yizhou Wang, Shixiang Tang, Wenhao Wu, Tong He, Wanli Ouyang, Philip Torr, and Jian Wu. Dettoolchain: A new prompting paradigm to unleash detection ability of mllm. In European Conference on Computer Vision, pages 164-182. Springer, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24212" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 574 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[49] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[50] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[51] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[52] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[53] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In *Forty-first International Conference on Machine Learning*, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[54] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplugowl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "type": "text", + "content": "[55] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 484 + ], + "type": "text", + "content": "[56] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "text", + "content": "[57] Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandas Alexandros Potamias. Hawor: World-space hand motion reconstruction from egocentric videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1805-1815, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 574 + ], + "type": "text", + "content": "[58] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 2, 8" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24213" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_content_list.json b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2187f2e73bb1bed176e3d90a921442a054ab2824 --- /dev/null +++ b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "VAMBA: Understanding Hour-Long Videos with Hybrid Mamba-Transformers", + "text_level": 1, + "bbox": [ + 101, + 130, + 897, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weiming Ren $^{1,4}$ , Wentao Ma $^{2}$ , Huan Yang $^{3}$ , Cong Wei $^{1,4}$ , Ge Zhang $^{1,5}$ , Wenhu Chen $^{1,4}$ $^{1}$ University of Waterloo, $^{2}$ University of Toronto, $^{3}$ Kuaishou Technology, $^{4}$ Vector Institute, $^{5}$ M-A-P", + "bbox": [ + 109, + 179, + 887, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{w2ren,wenhuchen}@uwaterloo.ca，hyang@fastmail.com", + "bbox": [ + 276, + 218, + 720, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://tiger-ai-lab.github.io/Vamba/", + "bbox": [ + 310, + 234, + 679, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 263, + 326, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames $(640\\times 360)$ on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least $50\\%$ reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by $4.3\\%$ on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks.", + "bbox": [ + 88, + 296, + 485, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 661, + 220, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of large multimodal models (LMMs) has seen tremendous progress in recent years. The seminal work of LLaVA [45] successfully transferred the power of autoregressive next-token prediction from large language models (LLMs) [2, 5, 29, 50, 61] to the multimodal domain, establishing a new standard for visual understanding by autoregressively modelling visual and language tokens through causal transformer layers. Since then, transformer-based LMMs have been widely studied and enhanced to tackle a diverse range of tasks, such as high-resolution image understanding [34, 44, 63] and interleaved image-text reasoning [30, 35, 36]. Nevertheless, the quadratic complexity inherent in causal self-attention operations presents significant challenges for long-context inputs, making it difficult for", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17a515573ce5fe186ccc074041b0885395218a1fdc97ee11e002b101f513bb10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 263, + 903, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a085ff6a07ec325bd8a84af43ef36e1775977b79d260e07ca2a2ba876793e68.jpg", + "image_caption": [ + "Figure 1. VAMBA achieves strong long video understanding performance (42.1% on LVBench [64]) while being more computationally efficient compared to transformer-based LMMs." + ], + "image_footnote": [], + "bbox": [ + 517, + 430, + 901, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "transformer-based LMMs to effectively handle the task of understanding extremely long videos.", + "bbox": [ + 511, + 595, + 906, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key challenge for advanced LMMs when processing long video inputs is that they tend to encode each frame into a large number of vision tokens, which leads to substantial computational and memory costs during both training and inference. For example, Qwen2-VL [63] can only process 256 frames (360p) on a single GPU, which is far from sufficient for hour-long video understanding. To reduce the computational/memory cost, previous efforts have primarily focused on reducing the vision tokens in the input sequence. Several approaches [20, 38, 40, 72] utilize a Q-Former [37] to compress vision tokens. More recent methods, such as LongVU [55] and Video-XL [56], partition vision token sequences into chunks and employ adaptive token compression mechanisms to decrease the token count. Another line of work, such as InternVideo2.5 [66] and Orxy-1.5 [48], employs various mechanisms to evaluate the vision tokens' importance, and thus dropping or merging those less significant tokens [6] during training", + "bbox": [ + 509, + 628, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21197", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/760de367d39acb0675d3e5920af1632b762a21f5a922cce51fe8ef6868127cec.jpg", + "table_caption": [ + "Table 1. Theoretical time and memory complexity for transformer-based LMMs and VAMBA in pre-filling. $M$ and $N$ represent the number of vision and text tokens. $d$ denotes the hidden dimension." + ], + "table_footnote": [], + "table_body": "
ModelTime ComplexityMemory Complexity
TransformerO(d(M+N)2)O((M+N)2)
VAMBAO(dMN+d2M)O(MN)
", + "bbox": [ + 93, + 142, + 475, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and inference. Despite these advancements, key challenges in LMMs for long video understanding still persist. First, aggressive token reduction can lead to critical information loss, particularly when processing extremely long videos. Second, these methods still suffer from quadratic computational complexity as the number of input frames increases. Third, token reduction-based methods require additional overhead, which might increase the actual wall-clock time.", + "bbox": [ + 88, + 213, + 480, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we investigate an orthogonal direction to previous approaches: instead of compressing the video tokens, we seek to develop an alternative model architecture that improves the efficiency of processing video tokens during training and pre-filling stage of inference. We propose VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. The key insight of our method is that we can design efficient modules to approximate the causal self-attention operation for both text and video tokens in transformer-based LMMs. In particular, we propose to (1) utilize cross-attentionions to update text tokens based on video tokens, which is affordable due to the short length of text tokens, and (2) adopt Mamba-2 [17] to process the massive video tokens with linear complexity.", + "bbox": [ + 89, + 337, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assuming a combined input sequence of $M + N$ tokens, where $M$ is the number of video tokens and $N$ is the number of text tokens, we find that $M$ could be at least 100 times larger than $N$ on many long video tasks ( $M \\gg N$ ). Our model can reduce the training/prefilling computational complexity from $O(d(M + N)^2)$ to $O(dMN + d^2M)$ , where $d$ is the hidden dimension, according to Table 1. In practice, this theoretical improvement may not be fully realized due to the hardware underoptimization for Mamba [24]. Nevertheless, we still observe a more than $50\\%$ reduction in GPU memory usage and FLOPs/routine during training and inference for long video inputs, as shown in Figure 1. VAMBA can be efficiently trained using $8 \\times \\mathrm{A}800$ GPUs, whereas other efficient video LMMs such as LongVU [55] and LongLLaVA [65] require 64 and 24 GPUs for training, respectively. By performing a two-stage training, our VAMBA achieves a $4.3\\%$ improvement over the best efficient video LMMs on the challenging hour-scale video understanding benchmark LVBench [64]. On other long video understanding datasets like Video-MME [21], MLVU [76] and LongVideoBench [67], VAMBA also achieves top-notch performance.", + "bbox": [ + 89, + 550, + 483, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 89, + 885, + 416, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose VAMBA, a hybrid Mamba-Transformer model for hour-long video understanding. VAMBA's design features efficient modules such as Mamba-2 blocks and cross-attention layers, effectively reducing the computational overhead of transformer-based LMMs.", + "2. We conduct a comprehensive ablation study and show that employing Mamba-2 blocks and initializing crossattention weights from pretrained self-attention layers are crucial for achieving high performance in VAMBA.", + "3. Our extensive evaluations demonstrate that VAMBA achieves strong video understanding capabilities. Specifically, VAMBA achieves $4.3\\%$ improvement over state-of-the-art efficient video LMMs on the challenging hour-long benchmark LVBench." + ], + "bbox": [ + 506, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 315, + 651, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. State Space Models and Mamba", + "text_level": 1, + "bbox": [ + 513, + 340, + 797, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State space models (SSMs) are linear time-invariant systems for modelling continuous signals. A continuous SSM can be expressed by ordinary differential equations (ODEs):", + "bbox": [ + 511, + 362, + 905, + 407 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 417, + 849, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "here, $x(t)\\in \\mathbb{R}$ is the input signal, $h(t)\\in \\mathbb{R}^N$ is the $N$ -dimensional hidden state and $y(t)\\in \\mathbb{R}$ is the output signal. $\\mathbf{A}\\in \\mathbb{R}^{N\\times N}$ is the state transition matrix and $\\mathbf{B}\\in \\mathbb{R}^{N\\times 1}$ , $\\mathbf{C}\\in \\mathbb{R}^{1\\times N}$ are projection matrices.", + "bbox": [ + 511, + 444, + 905, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mamba [24], and its predecessor S4 [26], are discretized SSMs for discrete sequence modelling. The discretization of the SSM is done by the zero-order hold (ZOH) technique:", + "bbox": [ + 511, + 506, + 905, + 551 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\quad \\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} (\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}) \\cdot \\Delta \\mathbf {B},\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 560, + 893, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\overline{\\mathbf{A}},\\overline{\\mathbf{B}}$ are discretized state variables and $\\Delta$ is the step size. The discretized SSM can thus be rewritten as:", + "bbox": [ + 511, + 590, + 905, + 619 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\quad y _ {t} = \\mathbf {C} h _ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 630, + 823, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mamba proposes the selective scan algorithm that dynamically determines $\\mathbf{B},\\mathbf{C},\\Delta$ based on the sequence inputs. Following S4, the A matrix is initialized as the diagonalized HiPPO matrix [25]. These design choices are proven to be efficient and effective for modelling extremely long sequences. Mamba-2 [17] further simplifies the formulation of A to a scalar times identity structure. Mamba-2 supports multi-head SSM and allows a much larger state dimension $N$ than Mamba, while being faster during training.", + "bbox": [ + 511, + 657, + 905, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Transformer-based LMMs", + "text_level": 1, + "bbox": [ + 511, + 801, + 756, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We study how text and video tokens are being updated in transformers. As shown in Figure 2, we assume a standard case where video tokens are placed before text tokens and ignore the chat template tokens for simplicity. In each LMM decoder layer, the video and text tokens are updated as:", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21198", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3a517d881738d387cd2158cd71f584111048b44de0a22cb8e29a529cfd3650a4.jpg", + "image_caption": [ + "(a) Transformer-based Video LMMs", + "Figure 2. Overview of our VAMBA model architecture. Compared to transformer-based LMMs (left), we replace the costly causal self-attention operations with the more efficient cross-attention layers and Mamba blocks to achieve better efficiency." + ], + "image_footnote": [], + "bbox": [ + 145, + 93, + 480, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e5a6f9f66192c3a371d46041fe93b312764fab2cdc5a0ce68ee2f6ef28e73dcc.jpg", + "image_caption": [ + "(b) Our Vamba Model Architecture" + ], + "image_footnote": [], + "bbox": [ + 496, + 93, + 864, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- All video tokens get updated by an input layer normalization (LN) layer, a self-attention layer, a post LN layer and a final MLP layer. There are also residual connections for self-attention and MLP layers. The $i^{\\text{th}}$ video token $v_{i}$ computes self-attention based on the first $i$ video tokens:", + "bbox": [ + 89, + 460, + 482, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no _ {v _ {i}} = \\left(\\sigma \\left(\\frac {q _ {v _ {i}} \\mathbf {K} _ {[ v _ {1}: v _ {i} ]} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {[ v _ {1}: v _ {i} ]}\\right) \\mathbf {W} _ {o}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 542, + 482, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ denotes the softmax operation, $q_{v_i}$ is the query vector for $v_{i}$ and $\\mathbf{K}_{[v_1:v_i]}$ , $\\mathbf{V}_{[v_1:v_i]}$ are key and value matrices of the first $i$ video tokens. $d$ denotes the dimension of the query and key vectors and $\\mathbf{W}_o$ is the out projection matrix. $o_{v_i}$ represents the final output for token $v_{i}$ .", + "bbox": [ + 102, + 584, + 482, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- The text tokens share a similar update route to the video tokens. The self-attention updates are slightly different: for the $j^{\\text{th}}$ text token $t_j$ , it computes self-attention based on all video tokens and the first $j$ text tokens:", + "bbox": [ + 89, + 660, + 482, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\no _ {t _ {j}} = \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\left[ \\mathbf {K} _ {v} , \\mathbf {K} _ {[ t _ {1} : t _ {j} ]} \\right] ^ {\\top}}{\\sqrt {d}}\\right) \\left[ \\mathbf {V} _ {v}, \\mathbf {V} _ {[ t _ {1}: t _ {j} ]} \\right]\\right) \\mathbf {W} _ {o}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 724, + 482, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[\\mathbf{K}_v,\\mathbf{K}_{[t_1:t_j]}$ and $[\\mathbf{V}_v,\\mathbf{V}_{[t_1:t_j]}$ are the key and value matrices of the combination of all video tokens and text tokens up to $t_j$ .", + "bbox": [ + 102, + 763, + 482, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main computation overhead in the transformer-based LMMs comes from the quadratic complexity of the self-attention in the video tokens. To overcome this issue, we design a hybrid Mamba Transformer architecture to process text and video tokens differently. Our detailed model architecture designs are listed in the sections below.", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method: VAMBA", + "text_level": 1, + "bbox": [ + 511, + 460, + 718, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to devise a model $\\Theta'$ that preserves the performance of transformer $\\Theta$ , while being more efficient. We approach this problem by approximating the causal transformer layers in the pretrained video LMMs.", + "bbox": [ + 511, + 486, + 906, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Updating Text Tokens via Cross-Attentions", + "text_level": 1, + "bbox": [ + 511, + 559, + 879, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key idea of our method is to split the expensive self-attention operation over the entire video and text token sequence into two more efficient components. Since video tokens typically dominate the sequence while text tokens remain few, we maintain the self-attention mechanism exclusively for the text tokens and eliminate it for the video tokens. Instead, we add cross-attention layers that use text tokens as queries and video tokens as keys and values. As shown in Figure 2, this design updates the text tokens during the attention layers in our model as follows:", + "bbox": [ + 511, + 580, + 906, + 733 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} o _ {t _ {j}} = (1 - \\alpha) \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\mathbf {K} _ {v} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {v}\\right) \\mathbf {W} _ {o} ^ {c} \\quad (\\text {C r o s s - A t t e n t i o n}) \\\\ + \\alpha \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\mathbf {K} _ {[ t _ {1}: t _ {j} ]} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {[ t _ {1}: t _ {j} ]}\\right)\\left. \\right) \\mathbf {W} _ {o} ^ {s}, \\quad (\\text {S e l f - A t t e n t i o n}) \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 739, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_o^c$ and $\\mathbf{W}_o^s$ are the output projection matrices for the cross- and self-attention layers. $\\alpha \\in [0,1]$ is a learnable weighting parameter that balances the cross- and self-attention outputs. Assuming a total of $M$ video tokens and $N$ text tokens, the self-attention operations in the", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21199", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "transformer-based video LMMs have a (pre-filling) computational complexity of $O(d(M + N)^2)$ . Our self+cross-attention design has the complexity of $O(dN^2)$ (self-attention) and $O(dMN)$ (cross-attention), effectively reducing the pre-filling complexity to $O(dN^2 + dMN) \\sim O(dMN)$ , given that $M \\gg N \\approx d$ .", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our attention formulation can be viewed as an approximation of the causal self-attention operation (Equation 2), as each text token $q_{t_j}$ can still attend to all video tokens and the first $j$ text tokens. As the video and text tokens share the same channel dimension, the projection matrices in the cross-attention layer will have identical dimensions as in the self-attention layers. This allows us to consider two model design choices: we can either randomly initialize the weights in the cross-attention layers, or we can initialize the cross-attention layers using the self-attention layer weights (i.e. $\\mathbf{W}_o^c = \\mathbf{W}_o^s$ , same applies to the query, key and value projection matrices $\\mathbf{W}_q^c, \\mathbf{W}_k^c, \\mathbf{W}_v^c$ ). We add these two design choices to the VAMBA design space.", + "bbox": [ + 89, + 184, + 482, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Updating Video Tokens with Mamba Blocks", + "text_level": 1, + "bbox": [ + 89, + 398, + 465, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the self+cross-attention design significantly reduces the model's complexity, relying solely on cross-attention layers can compromise the model's representational power. Specifically, rather than updating video tokens through self-attention and MLP layers (i.e., through the causal transformer block, c.f. Section 2.2), the video tokens remain unchanged after the cross-attention layers. To address this limitation, we seek an efficient architecture to approximate the effects of the transformer blocks. Motivated by the success of the Mamba [17, 24] architectures in image and video modelling [39, 47, 77], we propose to employ Mamba blocks to effectively process the video tokens. As shown in Figure 2, the video token updates can be formulated as:", + "bbox": [ + 89, + 422, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\no _ {v _ {i}} = \\mathrm {M a m b a} (\\mathrm {L N} (v _ {i}), \\mathbf {h} _ {v _ {i - 1}}, \\overline {{\\mathbf {A}}}, \\overline {{\\mathbf {B}}}, \\mathbf {C}), \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 636, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{LN}(\\cdot)$ is the layer normalization operator and $\\mathbf{h}_{v_{i-1}}$ is the context (hidden states) at token $v_{i-1}$ . As a recurrent model, Mamba operates similarly to causal self-attention layers, enabling each video token to update itself based on all preceding tokens. Crucially, Mamba reduces the complexity of updating video tokens from $O(dM^2)$ in causal self-attention to $O(d^2M)$ . Combining the Mamba blocks and the self+cross-attention design, our model achieves an overall complexity of $O(dN^2 + dMN + d^2M) \\sim O(dMN + d^2M)$ , which is substantially lower than the quadratic complexity $O(d(M + N)^2)$ in the transformer-based video LMMs. In Section 3.3, we further detail how the Mamba layers are trained to approximate Equation 1.", + "bbox": [ + 89, + 671, + 482, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider employing Mamba or Mamba-2 blocks in Equation 4 in the VAMBA design space.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Paradigm", + "text_level": 1, + "bbox": [ + 513, + 90, + 697, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize a two-stage training strategy to optimize VAMBA, including a pretraining stage and an instruction-tuning stage. In the pretraining stage, the model is initialized with pretrained transformer-based LMM weights in all modules except for the newly introduced cross-attention and Mamba layers. We freeze the pretrained components and train only the new layers using image captioning data to restore the model's visual understanding capabilities. We explore two types of training objectives during pretraining: (1) the standard language modelling loss $\\mathcal{L}_{\\mathrm{LM}} = -\\frac{1}{T}\\sum_{t=1}^{T}\\log p(x_t|x_{Model IDCA from SA?Mamba Block TypeStage 1Stage 2G-VEvalLVBenchVidMMEMVBenchref. freetestw/o sub.testAXN/A75.723.747.640.9B✓N/A81.034.251.751.8C✓Mamba81.834.253.453.5D✓Mamba-282.235.354.153.5", + "bbox": [ + 89, + 154, + 480, + 282 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9c6b69ec55e76d7f1f130713d36a58d7e854cefb341a82936b0b9b9ef4390f19.jpg", + "table_caption": [ + "Table 3. Ablation study results on pretraining VAMBA models with a transformer-based LMM teacher and a distillation loss." + ], + "table_footnote": [], + "table_body": "
Model D + Distillλ value in Stage 1 Training
00.0010.010.512
G-VEval82.1981.0580.6873.6963.6547.61
", + "bbox": [ + 89, + 334, + 478, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "metric based on GPT-4o [51] to evaluate the quality of the captions. For instruction-tuning (stage 2), we evaluate model performances across three benchmarks: LVBench [64] for hour-long videos, Video-MME [21] for medium-to-long videos, and MVBench [40] for short videos. Ablation study results are shown in Table 2 and Table 3.", + "bbox": [ + 89, + 414, + 483, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Attention Layer Initialization Strategy According to Table 2, we find that it is crucial to initialize the cross-attention layer weights using the corresponding self-attention layer from the same decoder level. Compared to Model A, the dramatic performance boost observed in Model B across all metrics underscores the critical impact of this weight initialization strategy. We believe this observation stems from the fact that initializing the cross-attention layer weights from the self-attention layers allows our self+cross-attention operation (Equation 3) to more closely approximate the pretrained causal self-attention operation (Equation 2): once $\\mathbf{W}_q^c$ , $\\mathbf{W}_k^c$ , $\\mathbf{W}_v^c$ and $\\mathbf{W}_o^c$ are set equal to $\\mathbf{W}_q^s$ , $\\mathbf{W}_k^s$ , $\\mathbf{W}_v^s$ and $\\mathbf{W}_o^s$ , the only discrepancy between the two equations becomes their attention score matrices. This difference can be mitigated by our two-stage training process, making it easier for Model B to recover its multimodal understanding capabilities.", + "bbox": [ + 89, + 513, + 483, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mamba Block Design Choices Comparing Model B with Model C and D in Table 2, we observe that both Mamba and Mamba-2 blocks improve the model performances across all metrics, indicating that Mamba blocks bring extra representation power to the model's visual understanding capabilities. Furthermore, when comparing Model C and Model D, we find that Mamba-2 demonstrates superior image and video modelling capabilities. Despite having a simplified", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/35d22dd603d2c51823f54b03f29f60ea4c8ad694e6b302feabe8386efa7a93e0.jpg", + "table_caption": [ + "Table 4. Comparisons between baseline models and VAMBA on hour-long video understanding benchmarks. Bold: best results among efficient video LMMs. Underline: second-best." + ], + "table_footnote": [], + "table_body": "
ModelsSizeHour-Long Video Understanding
LVBenchHourVideoHourEval
testdevtest
Transformer-based LMMs
Gemini-1.5-Pro [59]-33.137.3-
GPT-4o [51]-34.719.6-
Qwen2-VL [63]7B42.033.853.0
Efficient LMMs
LLaVA-Mini [74]7B17.620.624.2
LongLLaVA [65]9B31.227.739.1
LongVU [55]7B37.830.846.8
Video-XL [56]7B36.833.047.1
VAMBA10B42.133.650.7
", + "bbox": [ + 514, + 142, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A matrix structure, Mamba-2 accommodates a larger state dimension than Mamba (64 versus 16 in our case), which likely contributes to its improved performance.", + "bbox": [ + 511, + 380, + 906, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretraining Distillation Loss We employ the pretrained Qwen2-VL-7B model as the teacher model and perform a series of pretraining experiments with the additional distillation loss based on our best model setting (Model D in Table 2). The results are shown in Table 3. In contrast to previous findings from cross-attention-based LLMs (e.g. CEPE [70]), where distillation loss is reported to be beneficial, we observe that incorporating an additional distillation loss does not further improve our model's performance, as the G-VEval score decreases for all $\\lambda > 0$ . Therefore, we exclude the distillation loss and only employ the language modelling loss in both stages in our final training setting.", + "bbox": [ + 511, + 434, + 908, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Evaluation Results", + "text_level": 1, + "bbox": [ + 511, + 627, + 741, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details Based on our ablation study (Table 2), we adopt Model D as our final design and initialize VAMBA from Qwen2-VL-7B. We source $\\sim 3\\mathrm{M}$ image caption data for pretraining and $\\sim 7\\mathrm{M}$ image and video QA data for instruction-tuning. The full implementation details can be found in Appendix 7.", + "bbox": [ + 511, + 648, + 908, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hour-Long Video Understanding We evaluate our model's ability to handle extremely long videos using two public benchmarks: LVBench [64] and HourVideo [8] (development set). We further compose a new benchmark called HourEval by selecting all questions related to videos longer than 30 minutes from Video-MME [21], MLVU [76] development set, and LongVideoBench [67] validation set. The average video lengths for LVBench, HourVideo, and HourEval are 68.4, 47.2, and 54.7 minutes, respectively. We compare our model against four efficient video LMMs:", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21201", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ff986fbd777c6de04864d542d83f7f7b457c992a858f1c9f4f7d98f44495470f.jpg", + "table_caption": [ + "Table 5. Comparisons between baseline models and VAMBA on medium-length or short video understanding benchmarks. Bold: best results among each section. Underline: second-best. * indicates the results based on our evaluation scripts." + ], + "table_footnote": [], + "table_body": "
ModelsSizeMedium-Length UnderstandingShort Video UnderstandingVideo Captioning
Video-MMEMLVULongVideoBenchMVBenchNExT-QADREAM-1K
w/o subtitlesm-avgvaltestmcF1
Proprietary Models
GPT-4V [2]-59.949.259.143.570.434.4
GPT-4o [51]-71.964.666.7-76.039.2
Gemini-1.5-Pro [59]-75.062.964.054.276.436.2
Open-source Transformer-based LMMs
VideoChat2 [40]7B39.547.939.351.978.626.6
ShareGPT4Video [12]7B39.946.439.751.2-19.5
LongVA [73]7B52.456.351.849.268.3-
Video-CCAM [20]9B50.358.5-64.6--
Kangaroo [46]8B56.061.054.861.1--
InternVL2 [15]8B54.048.1*51.8*66.4-26.9
LLaVA-OneVision [35]7B58.262.6*56.456.779.431.7
Qwen2-VL [63]7B63.364.2*52.4*67.0-29.6
Phi-4-Mini [1]5.6B55.060.146.760.4--
Open-source Efficient LMMs
LLaVA-Mini [74]7B40.344.319.3*44.547.6*22.9
LongLLaVA [65]9B51.653.3*42.1*54.672.2*24.6
LongVU [55]7B55.3*65.453.5*66.978.0*28.1
Video-XL [56]7B55.564.950.755.377.5*23.5
VAMBA10B57.865.955.960.478.128.1
", + "bbox": [ + 98, + 127, + 893, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LLaVA-Mini [74], LongLLaVA [65], Video-XL [56] and LongVU [55]. We also include the results from Qwen2-VL-7B (our baseline transformer-based LMM) as a reference.", + "bbox": [ + 89, + 510, + 482, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental results are shown in Table 4. Our VAMBA consistently outperforms all efficient video LMMs across the three hour-long video benchmarks, highlighting its exceptional ability to understand and reason over hour-scale videos. Notably, our model surpasses the baseline Qwen2-VL-7B on the LVBench benchmark, and its performance on HourVideo is also very close to Qwen2-VL-7B. These results underscore that our VAMBA is competitive with the best open-sourced transformer-based LMMs, while being significantly more efficient during training and inference.", + "bbox": [ + 88, + 556, + 482, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Medium-Length or Short Video Understanding To further demonstrate our model's generalizability across various video durations, we test VAMBA on several medium-length or short video understanding benchmarks. Specifically, we report multiple-choice question performance on Video-MME, MLVU, LongVideoBench, NExT-QA [68] and MVBench [40]. We also evaluate VAMBA on DREAM1K [62] for video captioning assessment.", + "bbox": [ + 89, + 718, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "According to Table 5, our VAMBA demonstrates superior performance across three medium-length video understanding benchmarks (with average video durations between 10-20 minutes), ranking first among efficient video LMMs", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on all metrics. Despite the newly integrated cross-attention and Mamba layers in VAMBA have only been trained on relatively smaller datasets, VAMBA's performance remains competitive with large-scale pretrained transformer-based LMMs. Furthermore, all our evaluations can be conducted on a single 80G GPU, while some transformer-based LMMs, such as Qwen2-VL-7B, require additional inference strategies like sliding window attention or ring attention over multiple GPUs to achieve optimal results.", + "bbox": [ + 511, + 510, + 906, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For short video understanding and video captioning benchmarks, VAMBA also achieves competitive performances, ranking first on NExT-QA and DREAM-1K and second on MVBench among efficient LMMs. Overall, our model delivers the best results on medium-length and long video benchmarks, demonstrating its strong ability to handle long-context video-language inputs.", + "bbox": [ + 511, + 648, + 908, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Runtime Efficiency Analysis", + "text_level": 1, + "bbox": [ + 511, + 770, + 767, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To understand our model's runtime efficiency gains over the baseline transformer-based LMM (Qwen2-VL-7B), we conduct an efficiency analysis for both training and inference. For model training, we ensure a consistent training environment by using 8 NVIDIA A800 80G GPUs for both models. All input videos are resized to a resolution of $640 \\times 360$ , and a batch size of 1 is maintained", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21202", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3feca9c59421931918412e9a4553f52726e2e7ab077cfa4c9592323f92c15b68.jpg", + "image_caption": [ + "(a) Training GPU Memory Usage Comparison" + ], + "image_footnote": [], + "bbox": [ + 96, + 107, + 478, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50cca20c3ce4702e9080e9b525804e9b1d14b58f6efd05273f758e096a57391f.jpg", + "image_caption": [ + "(b) Training Runtime Comparison (time for 1 training step)", + "Figure 3. Comparison of training GPU memory usage and runtime per training step between Qwen2-VL-7B and VAMBA." + ], + "image_footnote": [], + "bbox": [ + 96, + 252, + 478, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on each GPU. We apply standard optimization methods such as Flash-Attention 2 [16], DeepSpeed ZeRO-3 [52] and gradient checkpointing [13] for both models. Based on different numbers of input frames, we measure the average GPU memory across 8 GPUs and the runtime per training step during training. As shown in Figure 3a, although VAMBA and Qwen2-VL-7B share the same vision encoder design and generate vision tokens of identical sequence lengths, our model requires over $50\\%$ less training memory when processing videos with more than 16 frames. This efficiency gain allows us to handle a larger number of video frames during training (512 vs. 128). Furthermore, as depicted in Figure 3b, our efficient design also accelerates model training, achieving nearly twice the speedup per training step when working with more than 64 frames.", + "bbox": [ + 88, + 438, + 482, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For model inference, we focus on the pre-fill stage and analyze GPU memory usage and FLOPs for both models. All measurements are conducted on an input video with a resolution of $640 \\times 360$ . As shown in Figure 4a, VAMBA requires slightly more GPU memory than Qwen2-VL-7B when the number of input frames is low (fewer than 32 frames), due to the additional $\\sim 3\\mathrm{B}$ parameters in its crossattention and Mamba layers. However, VAMBA's memory usage increases more slowly as the frame increases, allowing it to handle four times as many frames on a single NVIDIA A800 80G GPU compared to Qwen2-VL-7B (256 vs. 1024). Regarding computational cost, VAMBA reduces FLOPs by $30\\%$ to $50\\%$ during inference (Figure 4b), demonstrating a significantly lower complexity than its transformer-based LMM counterparts.", + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98babf5f86533cc20a404a777a757e910274104e6d2218be930603d997538b7a.jpg", + "image_caption": [ + "(a) Pre-filling GPU Memory Usage Comparison" + ], + "image_footnote": [], + "bbox": [ + 517, + 107, + 903, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a131ca9eab62e084a6939cd2defbc8b21e612e41362fed1a070e0ba436fb181.jpg", + "image_caption": [ + "(b) Pre-filling FLOPs Comparison (measured using calflops)", + "Figure 4. Comparison of GPU memory usage and FLOPs between Qwen2-VL-7B and VAMBA during inference." + ], + "image_footnote": [], + "bbox": [ + 517, + 252, + 901, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Case Study", + "text_level": 1, + "bbox": [ + 511, + 417, + 635, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform a case study for Video-XL and VAMBA using different video inputs and show the results in Figure 5. The \"cookie\" example focuses on object property identification in a long video (36 minutes). Video-XL incorrectly identifies the cookie colors as green and white, possibly due to the video also containing scenes for making other desserts (e.g. white ice cream in the frame highlighted in red). On the other hand, VAMBA more accurately describes them as green and brown. In the \"magic trick\" example, Video-XL misinterprets the rubber band as a green ring, whereas VAMBA correctly describes the performer manipulating the rubber band around their fingers, making it appear to pass through. These comparisons highlight how VAMBA is better at retrieving relevant events in long videos and capturing fine-grained visual details and actions.", + "bbox": [ + 511, + 439, + 906, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 511, + 679, + 655, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMMs) Large Multimodal Models (LMMs) have rapidly evolved in recent years. Recent research has focused on improving LMM's instruction-following ability through better data curation [11, 35, 40, 44, 53, 75], building better vision encoders that can process images at higher resolutions [34, 44, 63], and extending LMMs to interleaved image [30, 34, 36] or video [42, 49, 73] understanding. To enhance instruction-following capabilities, LLaVA-1.5 [43] and LLaVA-NeXT [44] leverage GPT-4-generated multimodal data for knowledge-enhanced tuning. To optimize visual feature extraction, Idefics2 [34] leverages the NaViT [18] strategy and the SigLIP [71] vision encoder to support image inputs with native resolu", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c27668e7f0b7c47741c99762dac5d47be31ea5c124d10313c0e03dde36a4aa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 92, + 256, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58d2ca9b7ea72c1b8bbd8d6f8821ebbd55ebd80b5d70665d8b9fa24a419310a3.jpg", + "image_caption": [ + "Question: What are the colors of the cookies that the person is making?" + ], + "image_footnote": [], + "bbox": [ + 261, + 93, + 416, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5c71d377979acb30ee263547d5d90e2d851ed2df198c7d843b61b35659827ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 93, + 575, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4bbe69722d277735b454af9c9e62d9f52e3d08ff95c857c2667f37d8717e8478.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 93, + 735, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a7f4717494f37c8fdecef0c001e7e3a41168023e85835f28002fda0ab812145d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 93, + 895, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video-XL: The cookies have different colored fillings, with at least one being green and another one being white.", + "bbox": [ + 112, + 258, + 488, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vamba: The person is making cookies in two different colors, green and brown.", + "bbox": [ + 509, + 258, + 880, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d5c59c3a458e72a26ab0c727008cd1bd7d0aec6c17ca310388816d544537aad.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison between VAMBA and efficient LMM baselines. Red text denotes incorrect responses, while green text highlights the correct responses by our VAMBA." + ], + "image_footnote": [], + "bbox": [ + 99, + 301, + 256, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/397057d1cca01445e4b0fe53943e50f5fc7ee9547b6a9472c9b815f74590ac4a.jpg", + "image_caption": [ + "Question: Describe the magic trick." + ], + "image_footnote": [], + "bbox": [ + 261, + 301, + 416, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9085a9b7f018e6f524726f231f65c07c2a5ad02c7885f798bb25a7d69bbcbfba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 301, + 575, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/617c6a3c94026710fd5e0550ecaa7d9a9be61cb486a4eb28e9b56a263e3502fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 301, + 736, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b72dd582d0ed219f524b5ef1f7ba826a80ec9822d6ccb81a8cea2d1ac072148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 301, + 897, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video-XL: The magic trick involves a person holding a green ring and a piece of string... The person's fingers are spread out, and they seem to be guiding the ring around the string.", + "bbox": [ + 109, + 395, + 483, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vamba: ...The person demonstrates a magic trick by manipulating the rubber band around their fingers, making it appear to pass through their hand.", + "bbox": [ + 508, + 395, + 870, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion. Qwen2-VL [63] applied Multimodal RoPE (M-RoPE) to optimize feature extraction for varying input sizes.", + "bbox": [ + 89, + 484, + 482, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite significant advances, current state-of-the-art LMMs remain inefficient at handling long-context multimodal inputs due to the quadratic complexity of causal self-attention operations. To mitigate this issue, methods like Flamingo [3], OpenFlamingo [4], Idefics [33] and LLaMA 3.2 [19] interleave gated cross-attention layers with self-attention layers for vision-text modelling. mPlug-Owl3 [69] develops the hyper-attention layer that integrates self-attention with cross-attention to support multimodal understanding. However, these models generally underperform compared to LLaVA-like transformer-based LMMs. In this study, we identify that one possible reason is the lack of vision token updates for cross-attention-based models. Our VAMBA addresses this limitation by incorporating additional Mamba layers to update vision tokens.", + "bbox": [ + 89, + 517, + 483, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficient LLMs/LMMs To improve the efficiency of LMM reasoning, recent work [37, 38, 40, 55, 56, 72, 74] develops techniques to reduce the length of the vision token sequence. Video-LLaMA [72] and VideoChat [38, 40] use a Q-Former [37] to compress dense visual tokens into a more compact sequence. LLaVA-Mini [74] further reduces each image to only one token. Video-XL [56] introduces a Visual Summarization Token to summarize the visual information, while LongVU [55] designs a spatiotemporal adaptive compression mechanism to reduce the number of video tokens.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another research avenue, exemplified by the Mamba [17, 24] model series, focuses on developing linear-time sequence models to reduce complexity. Mamba-based [78] and Mamba-Transformer hybrid models [22, 41] have achieved notable success in the realm of pure language modelling. However, these efficient architectures have received little attention in the multimodal understanding domain. LongLLaVA [65] took an initial approach by incorporating the hybrid Jamba [41] model as the LMM decoder, yet its performance in long video understanding remains suboptimal. This shortfall is likely due to the absence of a strong hybrid LLM as the base model and the lack of large-scale training on multimodal instruction-following data.", + "bbox": [ + 511, + 486, + 906, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. By integrating cross-attention for text tokens and Mamba-2 blocks for video token updates, our approach reduces computational complexity and GPU memory usage while achieving competitive performance across long, medium, and short video benchmarks. Extensive evaluations on datasets such as LVBench demonstrate VAMBA's superiority over existing efficient video LMMs. For future work, since our model design is orthogonal to token reduction-based methods, we will focus on combining the strengths of both approaches to develop more efficient VAMBA variants.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21204", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, YenChun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Wei-jian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 6", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8", + "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8", + "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1", + "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 1", + "[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 1", + "[8] Keshigeyan Chandrasegaran, Agrim Gupta, Lea M Hadzic, Taran Kota, Jimming He, Cristóbal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li. Hourvideo: 1-hour video-language understanding. Advances in Neural Information Processing Systems, 37:53168-53197, 2025. 5, 2", + "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceed-" + ], + "bbox": [ + 99, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 4", + "[10] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for lite vision-language models. arXiv preprint arXiv:2402.11684, 2024. 1", + "[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 7, 1", + "[12] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. Advances in Neural Information Processing Systems, 37:19472-19495, 2025. 6, 1", + "[13] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 7, 1", + "[14] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 1", + "[15] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 6", + "[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 7, 1", + "[17] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 2, 4, 8", + "[18] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36:2252-2274, 2023. 7", + "[19] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8", + "[20] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 1, 6", + "[21] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21205", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Paolo Glorioso, Quentin Anthony, Yury Tokpanov, James Whittington, Jonathan Pilault, Adam Ibrahim, and Beren Millidge. Zamba: A compact 7b ssm hybrid model. arXiv preprint arXiv:2405.16712, 2024. 8", + "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2", + "[24] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 4, 8", + "[25] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2", + "[26] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2", + "[27] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 1", + "[28] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Gedas Bertasius, and Lorenzo Torresani. Bimba: Selective-scan compression for long-range video question answering. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29096-29107, 2025. 3", + "[29] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1", + "[30] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 7", + "[31] Jindong Jiang, Xiuyu Li, Zhijian Liu, Muyang Li, Guo Chen, Zhiqi Li, De-An Huang, Guilin Liu, Zhiding Yu, Kurt Keutzer, et al. Token-efficient long video understanding for multimodal llms. arXiv preprint arXiv:2503.04130, 2025. 3", + "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1", + "[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36:71683-71702, 2023. 8", + "[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 7" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 6, 7", + "[36] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 7", + "[37] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 8", + "[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 8", + "[39] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2024. 4", + "[40] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 5, 6, 7, 8, 2", + "[41] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 8", + "[42] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7", + "[43] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 7, 1", + "[44] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024. Accessed: 2025-02-14. 1, 7", + "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1", + "[46] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 6", + "[47] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model 2024. arXiv preprint arXiv:2401.10166, 2024. 4", + "[48] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21206", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1", + "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 1", + "[50] OpenAI. Chatgpt. https://openai.com/index/chatgpt/, 2023.1", + "[51] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.5,6", + "[52] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 7, 1", + "[53] Weiming Ren, Huan Yang, Jie Min, Cong Wei, and Wenhu Chen. Vista: Enhancing long-duration and high-resolution video understanding by video spatiotemporal augmentation. arXiv preprint arXiv:2412.00927, 2024. 7", + "[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1", + "[55] Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434, 2024. 1, 2, 5, 6, 8", + "[56] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 1, 5, 6, 8", + "[57] Vasu Singla, Kaiyu Yue, Sukriti Paul, Reza Shirkavand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav Bhatele, Gowthami Somepalli, and Tom Goldstein. From pixels to prose: A large dataset of dense image captions. arXiv preprint arXiv:2406.10328, 2024. 4", + "[58] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1", + "[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 5, 6", + "[60] Tony Cheng Tong, Sirui He, Zhiwen Shao, and Dit-Yan Yeung. G-veval: A versatile metric for evaluating image and video captions using gpt-4o. arXiv preprint arXiv:2412.13647, 2024. 4", + "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[62] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 6, 2", + "[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4, 5, 6, 7, 8", + "[64] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 1, 2, 5", + "[65] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. Longlava: Scaling multi-modal llms to 1000 images efficiently via a hybrid architecture. arXiv preprint arXiv:2409.02889, 2024. 2, 5, 6, 8, 1", + "[66] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 1", + "[67] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 5", + "[68] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6, 2", + "[69] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8", + "[70] Howard Yen. Long-context language modeling with parallel context encoding. Master's thesis, Princeton University, 2024. 5", + "[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 7", + "[72] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, Singapore, 2023. Association for Computational Linguistics. 1, 8", + "[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. arXiv preprint arXiv:2501.03895, 2025. 5, 6, 8, 1", + "[75] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 4, 7", + "[76] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5", + "[77] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 4", + "[78] Jingwei Zuo, Maksim Velikanov, Dhia Eddine Rhaiem, Ilyas Chahed, Younes Belkada, Guillaume Kunsch, and Hakim Hacid. Falcon mamba: The first competitive attention-free 7b language model. arXiv preprint arXiv:2410.05355, 2024. 8" + ], + "bbox": [ + 91, + 92, + 480, + 385 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21208", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_model.json b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ecce5dfc3f2ac0787d7e359d2828adb06ec319cc --- /dev/null +++ b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_model.json @@ -0,0 +1,2611 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.898, + 0.153 + ], + "angle": 0, + "content": "VAMBA: Understanding Hour-Long Videos with Hybrid Mamba-Transformers" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.18, + 0.888, + 0.217 + ], + "angle": 0, + "content": "Weiming Ren\\(^{1,4}\\), Wentao Ma\\(^{2}\\), Huan Yang\\(^{3}\\), Cong Wei\\(^{1,4}\\), Ge Zhang\\(^{1,5}\\), Wenhu Chen\\(^{1,4}\\) \n\\(^{1}\\)University of Waterloo, \\(^{2}\\)University of Toronto, \\(^{3}\\)Kuaishou Technology, \\(^{4}\\)Vector Institute, \\(^{5}\\)M-A-P" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.219, + 0.721, + 0.234 + ], + "angle": 0, + "content": "{w2ren,wenhuchen}@uwaterloo.ca，hyang@fastmail.com" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.236, + 0.68, + 0.251 + ], + "angle": 0, + "content": "https://tiger-ai-lab.github.io/Vamba/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.264, + 0.327, + 0.28 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.297, + 0.486, + 0.63 + ], + "angle": 0, + "content": "State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames \\((640\\times 360)\\) on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least \\(50\\%\\) reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by \\(4.3\\%\\) on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.662, + 0.222, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.901 + ], + "angle": 0, + "content": "The field of large multimodal models (LMMs) has seen tremendous progress in recent years. The seminal work of LLaVA [45] successfully transferred the power of autoregressive next-token prediction from large language models (LLMs) [2, 5, 29, 50, 61] to the multimodal domain, establishing a new standard for visual understanding by autoregressively modelling visual and language tokens through causal transformer layers. Since then, transformer-based LMMs have been widely studied and enhanced to tackle a diverse range of tasks, such as high-resolution image understanding [34, 44, 63] and interleaved image-text reasoning [30, 35, 36]. Nevertheless, the quadratic complexity inherent in causal self-attention operations presents significant challenges for long-context inputs, making it difficult for" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.264, + 0.904, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.431, + 0.902, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Figure 1. VAMBA achieves strong long video understanding performance (42.1% on LVBench [64]) while being more computationally efficient compared to transformer-based LMMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.627 + ], + "angle": 0, + "content": "transformer-based LMMs to effectively handle the task of understanding extremely long videos." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A key challenge for advanced LMMs when processing long video inputs is that they tend to encode each frame into a large number of vision tokens, which leads to substantial computational and memory costs during both training and inference. For example, Qwen2-VL [63] can only process 256 frames (360p) on a single GPU, which is far from sufficient for hour-long video understanding. To reduce the computational/memory cost, previous efforts have primarily focused on reducing the vision tokens in the input sequence. Several approaches [20, 38, 40, 72] utilize a Q-Former [37] to compress vision tokens. More recent methods, such as LongVU [55] and Video-XL [56], partition vision token sequences into chunks and employ adaptive token compression mechanisms to decrease the token count. Another line of work, such as InternVideo2.5 [66] and Orxy-1.5 [48], employs various mechanisms to evaluate the vision tokens' importance, and thus dropping or merging those less significant tokens [6] during training" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21197" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.132 + ], + "angle": 0, + "content": "Table 1. Theoretical time and memory complexity for transformer-based LMMs and VAMBA in pre-filling. \\( M \\) and \\( N \\) represent the number of vision and text tokens. \\( d \\) denotes the hidden dimension." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.143, + 0.477, + 0.202 + ], + "angle": 0, + "content": "
ModelTime ComplexityMemory Complexity
TransformerO(d(M+N)2)O((M+N)2)
VAMBAO(dMN+d2M)O(MN)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.214, + 0.482, + 0.336 + ], + "angle": 0, + "content": "and inference. Despite these advancements, key challenges in LMMs for long video understanding still persist. First, aggressive token reduction can lead to critical information loss, particularly when processing extremely long videos. Second, these methods still suffer from quadratic computational complexity as the number of input frames increases. Third, token reduction-based methods require additional overhead, which might increase the actual wall-clock time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.549 + ], + "angle": 0, + "content": "In this study, we investigate an orthogonal direction to previous approaches: instead of compressing the video tokens, we seek to develop an alternative model architecture that improves the efficiency of processing video tokens during training and pre-filling stage of inference. We propose VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. The key insight of our method is that we can design efficient modules to approximate the causal self-attention operation for both text and video tokens in transformer-based LMMs. In particular, we propose to (1) utilize cross-attentionions to update text tokens based on video tokens, which is affordable due to the short length of text tokens, and (2) adopt Mamba-2 [17] to process the massive video tokens with linear complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.884 + ], + "angle": 0, + "content": "Assuming a combined input sequence of \\( M + N \\) tokens, where \\( M \\) is the number of video tokens and \\( N \\) is the number of text tokens, we find that \\( M \\) could be at least 100 times larger than \\( N \\) on many long video tasks (\\( M \\gg N \\)). Our model can reduce the training/prefilling computational complexity from \\( O(d(M + N)^2) \\) to \\( O(dMN + d^2M) \\), where \\( d \\) is the hidden dimension, according to Table 1. In practice, this theoretical improvement may not be fully realized due to the hardware underoptimization for Mamba [24]. Nevertheless, we still observe a more than \\( 50\\% \\) reduction in GPU memory usage and FLOPs/routine during training and inference for long video inputs, as shown in Figure 1. VAMBA can be efficiently trained using \\( 8 \\times \\mathrm{A}800 \\) GPUs, whereas other efficient video LMMs such as LongVU [55] and LongLLaVA [65] require 64 and 24 GPUs for training, respectively. By performing a two-stage training, our VAMBA achieves a \\( 4.3\\% \\) improvement over the best efficient video LMMs on the challenging hour-scale video understanding benchmark LVBench [64]. On other long video understanding datasets like Video-MME [21], MLVU [76] and LongVideoBench [67], VAMBA also achieves top-notch performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.418, + 0.901 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "1. We propose VAMBA, a hybrid Mamba-Transformer model for hour-long video understanding. VAMBA's design features efficient modules such as Mamba-2 blocks and cross-attention layers, effectively reducing the computational overhead of transformer-based LMMs." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "2. We conduct a comprehensive ablation study and show that employing Mamba-2 blocks and initializing crossattention weights from pretrained self-attention layers are crucial for achieving high performance in VAMBA." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.228, + 0.905, + 0.303 + ], + "angle": 0, + "content": "3. Our extensive evaluations demonstrate that VAMBA achieves strong video understanding capabilities. Specifically, VAMBA achieves \\(4.3\\%\\) improvement over state-of-the-art efficient video LMMs on the challenging hour-long benchmark LVBench." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.316, + 0.652, + 0.332 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.341, + 0.798, + 0.357 + ], + "angle": 0, + "content": "2.1. State Space Models and Mamba" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.906, + 0.409 + ], + "angle": 0, + "content": "State space models (SSMs) are linear time-invariant systems for modelling continuous signals. A continuous SSM can be expressed by ordinary differential equations (ODEs):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.418, + 0.851, + 0.436 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.445, + 0.906, + 0.507 + ], + "angle": 0, + "content": "here, \\(x(t)\\in \\mathbb{R}\\) is the input signal, \\(h(t)\\in \\mathbb{R}^N\\) is the \\(N\\)-dimensional hidden state and \\(y(t)\\in \\mathbb{R}\\) is the output signal. \\(\\mathbf{A}\\in \\mathbb{R}^{N\\times N}\\) is the state transition matrix and \\(\\mathbf{B}\\in \\mathbb{R}^{N\\times 1}\\), \\(\\mathbf{C}\\in \\mathbb{R}^{1\\times N}\\) are projection matrices." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.906, + 0.553 + ], + "angle": 0, + "content": "Mamba [24], and its predecessor S4 [26], are discretized SSMs for discrete sequence modelling. The discretization of the SSM is done by the zero-order hold (ZOH) technique:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.561, + 0.895, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\quad \\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} (\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}) \\cdot \\Delta \\mathbf {B},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.591, + 0.906, + 0.621 + ], + "angle": 0, + "content": "where \\(\\overline{\\mathbf{A}},\\overline{\\mathbf{B}}\\) are discretized state variables and \\(\\Delta\\) is the step size. The discretized SSM can thus be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.631, + 0.824, + 0.649 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\quad y _ {t} = \\mathbf {C} h _ {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.797 + ], + "angle": 0, + "content": "Mamba proposes the selective scan algorithm that dynamically determines \\(\\mathbf{B},\\mathbf{C},\\Delta\\) based on the sequence inputs. Following S4, the A matrix is initialized as the diagonalized HiPPO matrix [25]. These design choices are proven to be efficient and effective for modelling extremely long sequences. Mamba-2 [17] further simplifies the formulation of A to a scalar times identity structure. Mamba-2 supports multi-head SSM and allows a much larger state dimension \\(N\\) than Mamba, while being faster during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.757, + 0.818 + ], + "angle": 0, + "content": "2.2. Transformer-based LMMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We study how text and video tokens are being updated in transformers. As shown in Figure 2, we assume a standard case where video tokens are placed before text tokens and ignore the chat template tokens for simplicity. In each LMM decoder layer, the video and text tokens are updated as:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21198" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.146, + 0.094, + 0.482, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.405, + 0.435, + 0.417 + ], + "angle": 0, + "content": "(a) Transformer-based Video LMMs" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.094, + 0.865, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.405, + 0.788, + 0.417 + ], + "angle": 0, + "content": "(b) Our Vamba Model Architecture" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.42, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Figure 2. Overview of our VAMBA model architecture. Compared to transformer-based LMMs (left), we replace the costly causal self-attention operations with the more efficient cross-attention layers and Mamba blocks to achieve better efficiency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.462, + 0.483, + 0.539 + ], + "angle": 0, + "content": "- All video tokens get updated by an input layer normalization (LN) layer, a self-attention layer, a post LN layer and a final MLP layer. There are also residual connections for self-attention and MLP layers. The \\( i^{\\text{th}} \\) video token \\( v_{i} \\) computes self-attention based on the first \\( i \\) video tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.543, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\no _ {v _ {i}} = \\left(\\sigma \\left(\\frac {q _ {v _ {i}} \\mathbf {K} _ {[ v _ {1}: v _ {i} ]} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {[ v _ {1}: v _ {i} ]}\\right) \\mathbf {W} _ {o}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.585, + 0.483, + 0.66 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) denotes the softmax operation, \\(q_{v_i}\\) is the query vector for \\(v_{i}\\) and \\(\\mathbf{K}_{[v_1:v_i]}\\), \\(\\mathbf{V}_{[v_1:v_i]}\\) are key and value matrices of the first \\(i\\) video tokens. \\(d\\) denotes the dimension of the query and key vectors and \\(\\mathbf{W}_o\\) is the out projection matrix. \\(o_{v_i}\\) represents the final output for token \\(v_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.661, + 0.483, + 0.72 + ], + "angle": 0, + "content": "- The text tokens share a similar update route to the video tokens. The self-attention updates are slightly different: for the \\(j^{\\text{th}}\\) text token \\(t_j\\), it computes self-attention based on all video tokens and the first \\(j\\) text tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.725, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\no _ {t _ {j}} = \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\left[ \\mathbf {K} _ {v} , \\mathbf {K} _ {[ t _ {1} : t _ {j} ]} \\right] ^ {\\top}}{\\sqrt {d}}\\right) \\left[ \\mathbf {V} _ {v}, \\mathbf {V} _ {[ t _ {1}: t _ {j} ]} \\right]\\right) \\mathbf {W} _ {o}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.765, + 0.483, + 0.811 + ], + "angle": 0, + "content": "where \\([\\mathbf{K}_v,\\mathbf{K}_{[t_1:t_j]}\\) and \\([\\mathbf{V}_v,\\mathbf{V}_{[t_1:t_j]}\\) are the key and value matrices of the combination of all video tokens and text tokens up to \\(t_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The main computation overhead in the transformer-based LMMs comes from the quadratic complexity of the self-attention in the video tokens. To overcome this issue, we design a hybrid Mamba Transformer architecture to process text and video tokens differently. Our detailed model architecture designs are listed in the sections below." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.461, + 0.72, + 0.476 + ], + "angle": 0, + "content": "3. Our Method: VAMBA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Our goal is to devise a model \\(\\Theta'\\) that preserves the performance of transformer \\(\\Theta\\), while being more efficient. We approach this problem by approximating the causal transformer layers in the pretrained video LMMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.88, + 0.575 + ], + "angle": 0, + "content": "3.1. Updating Text Tokens via Cross-Attentions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.734 + ], + "angle": 0, + "content": "The key idea of our method is to split the expensive self-attention operation over the entire video and text token sequence into two more efficient components. Since video tokens typically dominate the sequence while text tokens remain few, we maintain the self-attention mechanism exclusively for the text tokens and eliminate it for the video tokens. Instead, we add cross-attention layers that use text tokens as queries and video tokens as keys and values. As shown in Figure 2, this design updates the text tokens during the attention layers in our model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.74, + 0.905, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} o _ {t _ {j}} = (1 - \\alpha) \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\mathbf {K} _ {v} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {v}\\right) \\mathbf {W} _ {o} ^ {c} \\quad (\\text {C r o s s - A t t e n t i o n}) \\\\ + \\alpha \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\mathbf {K} _ {[ t _ {1}: t _ {j} ]} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {[ t _ {1}: t _ {j} ]}\\right)\\left. \\right) \\mathbf {W} _ {o} ^ {s}, \\quad (\\text {S e l f - A t t e n t i o n}) \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_o^c\\) and \\(\\mathbf{W}_o^s\\) are the output projection matrices for the cross- and self-attention layers. \\(\\alpha \\in [0,1]\\) is a learnable weighting parameter that balances the cross- and self-attention outputs. Assuming a total of \\(M\\) video tokens and \\(N\\) text tokens, the self-attention operations in the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21199" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "transformer-based video LMMs have a (pre-filling) computational complexity of \\( O(d(M + N)^2) \\). Our self+cross-attention design has the complexity of \\( O(dN^2) \\) (self-attention) and \\( O(dMN) \\) (cross-attention), effectively reducing the pre-filling complexity to \\( O(dN^2 + dMN) \\sim O(dMN) \\), given that \\( M \\gg N \\approx d \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Our attention formulation can be viewed as an approximation of the causal self-attention operation (Equation 2), as each text token \\( q_{t_j} \\) can still attend to all video tokens and the first \\( j \\) text tokens. As the video and text tokens share the same channel dimension, the projection matrices in the cross-attention layer will have identical dimensions as in the self-attention layers. This allows us to consider two model design choices: we can either randomly initialize the weights in the cross-attention layers, or we can initialize the cross-attention layers using the self-attention layer weights (i.e. \\( \\mathbf{W}_o^c = \\mathbf{W}_o^s \\), same applies to the query, key and value projection matrices \\( \\mathbf{W}_q^c, \\mathbf{W}_k^c, \\mathbf{W}_v^c \\)). We add these two design choices to the VAMBA design space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.399, + 0.466, + 0.416 + ], + "angle": 0, + "content": "3.2. Updating Video Tokens with Mamba Blocks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.621 + ], + "angle": 0, + "content": "While the self+cross-attention design significantly reduces the model's complexity, relying solely on cross-attention layers can compromise the model's representational power. Specifically, rather than updating video tokens through self-attention and MLP layers (i.e., through the causal transformer block, c.f. Section 2.2), the video tokens remain unchanged after the cross-attention layers. To address this limitation, we seek an efficient architecture to approximate the effects of the transformer blocks. Motivated by the success of the Mamba [17, 24] architectures in image and video modelling [39, 47, 77], we propose to employ Mamba blocks to effectively process the video tokens. As shown in Figure 2, the video token updates can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.637, + 0.483, + 0.657 + ], + "angle": 0, + "content": "\\[\no _ {v _ {i}} = \\mathrm {M a m b a} (\\mathrm {L N} (v _ {i}), \\mathbf {h} _ {v _ {i - 1}}, \\overline {{\\mathbf {A}}}, \\overline {{\\mathbf {B}}}, \\mathbf {C}), \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.868 + ], + "angle": 0, + "content": "where \\(\\mathrm{LN}(\\cdot)\\) is the layer normalization operator and \\(\\mathbf{h}_{v_{i-1}}\\) is the context (hidden states) at token \\(v_{i-1}\\). As a recurrent model, Mamba operates similarly to causal self-attention layers, enabling each video token to update itself based on all preceding tokens. Crucially, Mamba reduces the complexity of updating video tokens from \\(O(dM^2)\\) in causal self-attention to \\(O(d^2M)\\). Combining the Mamba blocks and the self+cross-attention design, our model achieves an overall complexity of \\(O(dN^2 + dMN + d^2M) \\sim O(dMN + d^2M)\\), which is substantially lower than the quadratic complexity \\(O(d(M + N)^2)\\) in the transformer-based video LMMs. In Section 3.3, we further detail how the Mamba layers are trained to approximate Equation 1." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We consider employing Mamba or Mamba-2 blocks in Equation 4 in the VAMBA design space." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.698, + 0.108 + ], + "angle": 0, + "content": "3.3. Training Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.415 + ], + "angle": 0, + "content": "We utilize a two-stage training strategy to optimize VAMBA, including a pretraining stage and an instruction-tuning stage. In the pretraining stage, the model is initialized with pretrained transformer-based LMM weights in all modules except for the newly introduced cross-attention and Mamba layers. We freeze the pretrained components and train only the new layers using image captioning data to restore the model's visual understanding capabilities. We explore two types of training objectives during pretraining: (1) the standard language modelling loss \\(\\mathcal{L}_{\\mathrm{LM}} = -\\frac{1}{T}\\sum_{t=1}^{T}\\log p(x_t|x_{Model IDCA from SA?Mamba Block TypeStage 1Stage 2G-VEvalLVBenchVidMMEMVBenchref. freetestw/o sub.testAXN/A75.723.747.640.9B✓N/A81.034.251.751.8C✓Mamba81.834.253.453.5D✓Mamba-282.235.354.153.5" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.297, + 0.485, + 0.327 + ], + "angle": 0, + "content": "Table 3. Ablation study results on pretraining VAMBA models with a transformer-based LMM teacher and a distillation loss." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.335, + 0.48, + 0.402 + ], + "angle": 0, + "content": "
Model D + Distillλ value in Stage 1 Training
00.0010.010.512
G-VEval82.1981.0580.6873.6963.6547.61
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.506 + ], + "angle": 0, + "content": "metric based on GPT-4o [51] to evaluate the quality of the captions. For instruction-tuning (stage 2), we evaluate model performances across three benchmarks: LVBench [64] for hour-long videos, Video-MME [21] for medium-to-long videos, and MVBench [40] for short videos. Ablation study results are shown in Table 2 and Table 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Cross-Attention Layer Initialization Strategy According to Table 2, we find that it is crucial to initialize the cross-attention layer weights using the corresponding self-attention layer from the same decoder level. Compared to Model A, the dramatic performance boost observed in Model B across all metrics underscores the critical impact of this weight initialization strategy. We believe this observation stems from the fact that initializing the cross-attention layer weights from the self-attention layers allows our self+cross-attention operation (Equation 3) to more closely approximate the pretrained causal self-attention operation (Equation 2): once \\(\\mathbf{W}_q^c\\), \\(\\mathbf{W}_k^c\\), \\(\\mathbf{W}_v^c\\) and \\(\\mathbf{W}_o^c\\) are set equal to \\(\\mathbf{W}_q^s\\), \\(\\mathbf{W}_k^s\\), \\(\\mathbf{W}_v^s\\) and \\(\\mathbf{W}_o^s\\), the only discrepancy between the two equations becomes their attention score matrices. This difference can be mitigated by our two-stage training process, making it easier for Model B to recover its multimodal understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Mamba Block Design Choices Comparing Model B with Model C and D in Table 2, we observe that both Mamba and Mamba-2 blocks improve the model performances across all metrics, indicating that Mamba blocks bring extra representation power to the model's visual understanding capabilities. Furthermore, when comparing Model C and Model D, we find that Mamba-2 demonstrates superior image and video modelling capabilities. Despite having a simplified" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 4. Comparisons between baseline models and VAMBA on hour-long video understanding benchmarks. Bold: best results among efficient video LMMs. Underline: second-best." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.143, + 0.905, + 0.366 + ], + "angle": 0, + "content": "
ModelsSizeHour-Long Video Understanding
LVBenchHourVideoHourEval
testdevtest
Transformer-based LMMs
Gemini-1.5-Pro [59]-33.137.3-
GPT-4o [51]-34.719.6-
Qwen2-VL [63]7B42.033.853.0
Efficient LMMs
LLaVA-Mini [74]7B17.620.624.2
LongLLaVA [65]9B31.227.739.1
LongVU [55]7B37.830.846.8
Video-XL [56]7B36.833.047.1
VAMBA10B42.133.650.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.427 + ], + "angle": 0, + "content": "A matrix structure, Mamba-2 accommodates a larger state dimension than Mamba (64 versus 16 in our case), which likely contributes to its improved performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.909, + 0.618 + ], + "angle": 0, + "content": "Pretraining Distillation Loss We employ the pretrained Qwen2-VL-7B model as the teacher model and perform a series of pretraining experiments with the additional distillation loss based on our best model setting (Model D in Table 2). The results are shown in Table 3. In contrast to previous findings from cross-attention-based LLMs (e.g. CEPE [70]), where distillation loss is reported to be beneficial, we observe that incorporating an additional distillation loss does not further improve our model's performance, as the G-VEval score decreases for all \\(\\lambda > 0\\). Therefore, we exclude the distillation loss and only employ the language modelling loss in both stages in our final training setting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.742, + 0.643 + ], + "angle": 0, + "content": "4.2. Main Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.909, + 0.743 + ], + "angle": 0, + "content": "Implementation Details Based on our ablation study (Table 2), we adopt Model D as our final design and initialize VAMBA from Qwen2-VL-7B. We source \\(\\sim 3\\mathrm{M}\\) image caption data for pretraining and \\(\\sim 7\\mathrm{M}\\) image and video QA data for instruction-tuning. The full implementation details can be found in Appendix 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Hour-Long Video Understanding We evaluate our model's ability to handle extremely long videos using two public benchmarks: LVBench [64] and HourVideo [8] (development set). We further compose a new benchmark called HourEval by selecting all questions related to videos longer than 30 minutes from Video-MME [21], MLVU [76] development set, and LongVideoBench [67] validation set. The average video lengths for LVBench, HourVideo, and HourEval are 68.4, 47.2, and 54.7 minutes, respectively. We compare our model against four efficient video LMMs:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "21201" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 5. Comparisons between baseline models and VAMBA on medium-length or short video understanding benchmarks. Bold: best results among each section. Underline: second-best. * indicates the results based on our evaluation scripts." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.128, + 0.894, + 0.499 + ], + "angle": 0, + "content": "
ModelsSizeMedium-Length UnderstandingShort Video UnderstandingVideo Captioning
Video-MMEMLVULongVideoBenchMVBenchNExT-QADREAM-1K
w/o subtitlesm-avgvaltestmcF1
Proprietary Models
GPT-4V [2]-59.949.259.143.570.434.4
GPT-4o [51]-71.964.666.7-76.039.2
Gemini-1.5-Pro [59]-75.062.964.054.276.436.2
Open-source Transformer-based LMMs
VideoChat2 [40]7B39.547.939.351.978.626.6
ShareGPT4Video [12]7B39.946.439.751.2-19.5
LongVA [73]7B52.456.351.849.268.3-
Video-CCAM [20]9B50.358.5-64.6--
Kangaroo [46]8B56.061.054.861.1--
InternVL2 [15]8B54.048.1*51.8*66.4-26.9
LLaVA-OneVision [35]7B58.262.6*56.456.779.431.7
Qwen2-VL [63]7B63.364.2*52.4*67.0-29.6
Phi-4-Mini [1]5.6B55.060.146.760.4--
Open-source Efficient LMMs
LLaVA-Mini [74]7B40.344.319.3*44.547.6*22.9
LongLLaVA [65]9B51.653.3*42.1*54.672.2*24.6
LongVU [55]7B55.3*65.453.5*66.978.0*28.1
Video-XL [56]7B55.564.950.755.377.5*23.5
VAMBA10B57.865.955.960.478.128.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.483, + 0.556 + ], + "angle": 0, + "content": "LLaVA-Mini [74], LongLLaVA [65], Video-XL [56] and LongVU [55]. We also include the results from Qwen2-VL-7B (our baseline transformer-based LMM) as a reference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.558, + 0.483, + 0.709 + ], + "angle": 0, + "content": "Experimental results are shown in Table 4. Our VAMBA consistently outperforms all efficient video LMMs across the three hour-long video benchmarks, highlighting its exceptional ability to understand and reason over hour-scale videos. Notably, our model surpasses the baseline Qwen2-VL-7B on the LVBench benchmark, and its performance on HourVideo is also very close to Qwen2-VL-7B. These results underscore that our VAMBA is competitive with the best open-sourced transformer-based LMMs, while being significantly more efficient during training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Medium-Length or Short Video Understanding To further demonstrate our model's generalizability across various video durations, we test VAMBA on several medium-length or short video understanding benchmarks. Specifically, we report multiple-choice question performance on Video-MME, MLVU, LongVideoBench, NExT-QA [68] and MVBench [40]. We also evaluate VAMBA on DREAM1K [62] for video captioning assessment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "According to Table 5, our VAMBA demonstrates superior performance across three medium-length video understanding benchmarks (with average video durations between 10-20 minutes), ranking first among efficient video LMMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.907, + 0.648 + ], + "angle": 0, + "content": "on all metrics. Despite the newly integrated cross-attention and Mamba layers in VAMBA have only been trained on relatively smaller datasets, VAMBA's performance remains competitive with large-scale pretrained transformer-based LMMs. Furthermore, all our evaluations can be conducted on a single 80G GPU, while some transformer-based LMMs, such as Qwen2-VL-7B, require additional inference strategies like sliding window attention or ring attention over multiple GPUs to achieve optimal results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.909, + 0.756 + ], + "angle": 0, + "content": "For short video understanding and video captioning benchmarks, VAMBA also achieves competitive performances, ranking first on NExT-QA and DREAM-1K and second on MVBench among efficient LMMs. Overall, our model delivers the best results on medium-length and long video benchmarks, demonstrating its strong ability to handle long-context video-language inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.771, + 0.768, + 0.789 + ], + "angle": 0, + "content": "4.3. Runtime Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To understand our model's runtime efficiency gains over the baseline transformer-based LMM (Qwen2-VL-7B), we conduct an efficiency analysis for both training and inference. For model training, we ensure a consistent training environment by using 8 NVIDIA A800 80G GPUs for both models. All input videos are resized to a resolution of \\(640 \\times 360\\), and a batch size of 1 is maintained" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21202" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.093, + 0.411, + 0.106 + ], + "angle": 0, + "content": "(a) Training GPU Memory Usage Comparison" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.108, + 0.48, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.239, + 0.443, + 0.252 + ], + "angle": 0, + "content": "(b) Training Runtime Comparison (time for 1 training step)" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.253, + 0.48, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.407 + ], + "angle": 0, + "content": "Figure 3. Comparison of training GPU memory usage and runtime per training step between Qwen2-VL-7B and VAMBA." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.483, + 0.667 + ], + "angle": 0, + "content": "on each GPU. We apply standard optimization methods such as Flash-Attention 2 [16], DeepSpeed ZeRO-3 [52] and gradient checkpointing [13] for both models. Based on different numbers of input frames, we measure the average GPU memory across 8 GPUs and the runtime per training step during training. As shown in Figure 3a, although VAMBA and Qwen2-VL-7B share the same vision encoder design and generate vision tokens of identical sequence lengths, our model requires over \\(50\\%\\) less training memory when processing videos with more than 16 frames. This efficiency gain allows us to handle a larger number of video frames during training (512 vs. 128). Furthermore, as depicted in Figure 3b, our efficient design also accelerates model training, achieving nearly twice the speedup per training step when working with more than 64 frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For model inference, we focus on the pre-fill stage and analyze GPU memory usage and FLOPs for both models. All measurements are conducted on an input video with a resolution of \\(640 \\times 360\\). As shown in Figure 4a, VAMBA requires slightly more GPU memory than Qwen2-VL-7B when the number of input frames is low (fewer than 32 frames), due to the additional \\(\\sim 3\\mathrm{B}\\) parameters in its crossattention and Mamba layers. However, VAMBA's memory usage increases more slowly as the frame increases, allowing it to handle four times as many frames on a single NVIDIA A800 80G GPU compared to Qwen2-VL-7B (256 vs. 1024). Regarding computational cost, VAMBA reduces FLOPs by \\(30\\%\\) to \\(50\\%\\) during inference (Figure 4b), demonstrating a significantly lower complexity than its transformer-based LMM counterparts." + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.093, + 0.838, + 0.107 + ], + "angle": 0, + "content": "(a) Pre-filling GPU Memory Usage Comparison" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.108, + 0.905, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.239, + 0.881, + 0.252 + ], + "angle": 0, + "content": "(b) Pre-filling FLOPs Comparison (measured using calflops)" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.253, + 0.902, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.378, + 0.905, + 0.406 + ], + "angle": 0, + "content": "Figure 4. Comparison of GPU memory usage and FLOPs between Qwen2-VL-7B and VAMBA during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.418, + 0.637, + 0.435 + ], + "angle": 0, + "content": "4.4. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.668 + ], + "angle": 0, + "content": "In this section, we perform a case study for Video-XL and VAMBA using different video inputs and show the results in Figure 5. The \"cookie\" example focuses on object property identification in a long video (36 minutes). Video-XL incorrectly identifies the cookie colors as green and white, possibly due to the video also containing scenes for making other desserts (e.g. white ice cream in the frame highlighted in red). On the other hand, VAMBA more accurately describes them as green and brown. In the \"magic trick\" example, Video-XL misinterprets the rubber band as a green ring, whereas VAMBA correctly describes the performer manipulating the rubber band around their fingers, making it appear to pass through. These comparisons highlight how VAMBA is better at retrieving relevant events in long videos and capturing fine-grained visual details and actions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.656, + 0.695 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Large Multimodal Models (LMMs) Large Multimodal Models (LMMs) have rapidly evolved in recent years. Recent research has focused on improving LMM's instruction-following ability through better data curation [11, 35, 40, 44, 53, 75], building better vision encoders that can process images at higher resolutions [34, 44, 63], and extending LMMs to interleaved image [30, 34, 36] or video [42, 49, 73] understanding. To enhance instruction-following capabilities, LLaVA-1.5 [43] and LLaVA-NeXT [44] leverage GPT-4-generated multimodal data for knowledge-enhanced tuning. To optimize visual feature extraction, Idefics2 [34] leverages the NaViT [18] strategy and the SigLIP [71] vision encoder to support image inputs with native resolu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.258, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.094, + 0.417, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.094, + 0.576, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.094, + 0.736, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.094, + 0.896, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.235, + 0.713, + 0.251 + ], + "angle": 0, + "content": "Question: What are the colors of the cookies that the person is making?" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.259, + 0.489, + 0.288 + ], + "angle": 0, + "content": "Video-XL: The cookies have different colored fillings, with at least one being green and another one being white." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.259, + 0.882, + 0.288 + ], + "angle": 0, + "content": "Vamba: The person is making cookies in two different colors, green and brown." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.302, + 0.258, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.302, + 0.418, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.302, + 0.576, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.302, + 0.737, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.302, + 0.898, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.374, + 0.606, + 0.388 + ], + "angle": 0, + "content": "Question: Describe the magic trick." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.396, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Video-XL: The magic trick involves a person holding a green ring and a piece of string... The person's fingers are spread out, and they seem to be guiding the ring around the string." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.396, + 0.872, + 0.439 + ], + "angle": 0, + "content": "Vamba: ...The person demonstrates a magic trick by manipulating the rubber band around their fingers, making it appear to pass through their hand." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.451, + 0.907, + 0.479 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison between VAMBA and efficient LMM baselines. Red text denotes incorrect responses, while green text highlights the correct responses by our VAMBA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.517 + ], + "angle": 0, + "content": "tion. Qwen2-VL [63] applied Multimodal RoPE (M-RoPE) to optimize feature extraction for varying input sizes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.744 + ], + "angle": 0, + "content": "Despite significant advances, current state-of-the-art LMMs remain inefficient at handling long-context multimodal inputs due to the quadratic complexity of causal self-attention operations. To mitigate this issue, methods like Flamingo [3], OpenFlamingo [4], Idefics [33] and LLaMA 3.2 [19] interleave gated cross-attention layers with self-attention layers for vision-text modelling. mPlug-Owl3 [69] develops the hyper-attention layer that integrates self-attention with cross-attention to support multimodal understanding. However, these models generally underperform compared to LLaVA-like transformer-based LMMs. In this study, we identify that one possible reason is the lack of vision token updates for cross-attention-based models. Our VAMBA addresses this limitation by incorporating additional Mamba layers to update vision tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Efficient LLMs/LMMs To improve the efficiency of LMM reasoning, recent work [37, 38, 40, 55, 56, 72, 74] develops techniques to reduce the length of the vision token sequence. Video-LLaMA [72] and VideoChat [38, 40] use a Q-Former [37] to compress dense visual tokens into a more compact sequence. LLaVA-Mini [74] further reduces each image to only one token. Video-XL [56] introduces a Visual Summarization Token to summarize the visual information, while LongVU [55] designs a spatiotemporal adaptive compression mechanism to reduce the number of video tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Another research avenue, exemplified by the Mamba [17, 24] model series, focuses on developing linear-time sequence models to reduce complexity. Mamba-based [78] and Mamba-Transformer hybrid models [22, 41] have achieved notable success in the realm of pure language modelling. However, these efficient architectures have received little attention in the multimodal understanding domain. LongLLaVA [65] took an initial approach by incorporating the hybrid Jamba [41] model as the LMM decoder, yet its performance in long video understanding remains suboptimal. This shortfall is likely due to the absence of a strong hybrid LLM as the base model and the lack of large-scale training on multimodal instruction-following data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We presented VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. By integrating cross-attention for text tokens and Mamba-2 blocks for video token updates, our approach reduces computational complexity and GPU memory usage while achieving competitive performance across long, medium, and short video benchmarks. Extensive evaluations on datasets such as LVBench demonstrate VAMBA's superiority over existing efficient video LMMs. For future work, since our model design is orthogonal to token reduction-based methods, we will focus on combining the strengths of both approaches to develop more efficient VAMBA variants." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21204" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.392 + ], + "angle": 0, + "content": "[1] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, YenChun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Wei-jian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.394, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[8] Keshigeyan Chandrasegaran, Agrim Gupta, Lea M Hadzic, Taran Kota, Jimming He, Cristóbal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li. Hourvideo: 1-hour video-language understanding. Advances in Neural Information Processing Systems, 37:53168-53197, 2025. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[10] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for lite vision-language models. arXiv preprint arXiv:2402.11684, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.34 + ], + "angle": 0, + "content": "[12] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. Advances in Neural Information Processing Systems, 37:19472-19495, 2025. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.904, + 0.385 + ], + "angle": 0, + "content": "[13] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.904, + 0.441 + ], + "angle": 0, + "content": "[14] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.904, + 0.524 + ], + "angle": 0, + "content": "[15] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[17] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[18] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36:2252-2274, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[19] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[20] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21205" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[22] Paolo Glorioso, Quentin Anthony, Yury Tokpanov, James Whittington, Jonathan Pilault, Adam Ibrahim, and Beren Millidge. Zamba: A compact 7b ssm hybrid model. arXiv preprint arXiv:2405.16712, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[24] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[25] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[26] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[27] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[28] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Gedas Bertasius, and Lorenzo Torresani. Bimba: Selective-scan compression for long-range video question answering. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29096-29107, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[29] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[30] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[31] Jindong Jiang, Xiuyu Li, Zhijian Liu, Muyang Li, Guo Chen, Zhiqi Li, De-An Huang, Guilin Liu, Zhiding Yu, Kurt Keutzer, et al. Token-efficient long video understanding for multimodal llms. arXiv preprint arXiv:2503.04130, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36:71683-71702, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[36] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[37] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[39] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[40] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 5, 6, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[41] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[42] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[43] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024. Accessed: 2025-02-14. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[46] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[47] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model 2024. arXiv preprint arXiv:2401.10166, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21206" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[50] OpenAI. Chatgpt. https://openai.com/index/chatgpt/, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[51] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.307 + ], + "angle": 0, + "content": "[52] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[53] Weiming Ren, Huan Yang, Jie Min, Cong Wei, and Wenhu Chen. Vista: Enhancing long-duration and high-resolution video understanding by video spatiotemporal augmentation. arXiv preprint arXiv:2412.00927, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.365, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[55] Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434, 2024. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[56] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[57] Vasu Singla, Kaiyu Yue, Sukriti Paul, Reza Shirkavand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav Bhatele, Gowthami Somepalli, and Tom Goldstein. From pixels to prose: A large dataset of dense image captions. arXiv preprint arXiv:2406.10328, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[58] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[60] Tony Cheng Tong, Sirui He, Zhiwen Shao, and Dit-Yan Yeung. G-veval: A versatile metric for evaluating image and video captions using gpt-4o. arXiv preprint arXiv:2412.13647, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[62] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.25 + ], + "angle": 0, + "content": "[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.253, + 0.905, + 0.309 + ], + "angle": 0, + "content": "[64] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.368 + ], + "angle": 0, + "content": "[65] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. Longlava: Scaling multi-modal llms to 1000 images efficiently via a hybrid architecture. arXiv preprint arXiv:2409.02889, 2024. 2, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.37, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[66] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[67] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[68] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[69] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[70] Howard Yen. Long-context language modeling with parallel context encoding. Master's thesis, Princeton University, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[72] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, Singapore, 2023. Association for Computational Linguistics. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21207" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[74] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. arXiv preprint arXiv:2501.03895, 2025. 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[75] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[76] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[77] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[78] Jingwei Zuo, Maksim Velikanov, Dhia Eddine Rhaiem, Ilyas Chahed, Younes Belkada, Guillaume Kunsch, and Hakim Hacid. Falcon mamba: The first competitive attention-free 7b language model. arXiv preprint arXiv:2410.05355, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21208" + } + ] +] \ No newline at end of file diff --git a/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_origin.pdf b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0abf9755cff2c2c7f9d3907db2f39905bcada5d0 --- /dev/null +++ b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/41032422-fba8-4ef6-8117-de06ce1069a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8370694a916f2e84ddf33f17e88c9132269c12322c00a018a63ecf573834dc60 +size 2618868 diff --git a/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/full.md b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3658a01aba4f93c101989b14aa530df077a06311 --- /dev/null +++ b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/full.md @@ -0,0 +1,350 @@ +# VAMBA: Understanding Hour-Long Videos with Hybrid Mamba-Transformers + +Weiming Ren $^{1,4}$ , Wentao Ma $^{2}$ , Huan Yang $^{3}$ , Cong Wei $^{1,4}$ , Ge Zhang $^{1,5}$ , Wenhu Chen $^{1,4}$ $^{1}$ University of Waterloo, $^{2}$ University of Toronto, $^{3}$ Kuaishou Technology, $^{4}$ Vector Institute, $^{5}$ M-A-P + +{w2ren,wenhuchen}@uwaterloo.ca，hyang@fastmail.com + +https://tiger-ai-lab.github.io/Vamba/ + +# Abstract + +State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames $(640\times 360)$ on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least $50\%$ reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by $4.3\%$ on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks. + +# 1. Introduction + +The field of large multimodal models (LMMs) has seen tremendous progress in recent years. The seminal work of LLaVA [45] successfully transferred the power of autoregressive next-token prediction from large language models (LLMs) [2, 5, 29, 50, 61] to the multimodal domain, establishing a new standard for visual understanding by autoregressively modelling visual and language tokens through causal transformer layers. Since then, transformer-based LMMs have been widely studied and enhanced to tackle a diverse range of tasks, such as high-resolution image understanding [34, 44, 63] and interleaved image-text reasoning [30, 35, 36]. Nevertheless, the quadratic complexity inherent in causal self-attention operations presents significant challenges for long-context inputs, making it difficult for + +![](images/17a515573ce5fe186ccc074041b0885395218a1fdc97ee11e002b101f513bb10.jpg) + +![](images/0a085ff6a07ec325bd8a84af43ef36e1775977b79d260e07ca2a2ba876793e68.jpg) +Figure 1. VAMBA achieves strong long video understanding performance (42.1% on LVBench [64]) while being more computationally efficient compared to transformer-based LMMs. + +transformer-based LMMs to effectively handle the task of understanding extremely long videos. + +A key challenge for advanced LMMs when processing long video inputs is that they tend to encode each frame into a large number of vision tokens, which leads to substantial computational and memory costs during both training and inference. For example, Qwen2-VL [63] can only process 256 frames (360p) on a single GPU, which is far from sufficient for hour-long video understanding. To reduce the computational/memory cost, previous efforts have primarily focused on reducing the vision tokens in the input sequence. Several approaches [20, 38, 40, 72] utilize a Q-Former [37] to compress vision tokens. More recent methods, such as LongVU [55] and Video-XL [56], partition vision token sequences into chunks and employ adaptive token compression mechanisms to decrease the token count. Another line of work, such as InternVideo2.5 [66] and Orxy-1.5 [48], employs various mechanisms to evaluate the vision tokens' importance, and thus dropping or merging those less significant tokens [6] during training + +Table 1. Theoretical time and memory complexity for transformer-based LMMs and VAMBA in pre-filling. $M$ and $N$ represent the number of vision and text tokens. $d$ denotes the hidden dimension. + +
ModelTime ComplexityMemory Complexity
TransformerO(d(M+N)2)O((M+N)2)
VAMBAO(dMN+d2M)O(MN)
+ +and inference. Despite these advancements, key challenges in LMMs for long video understanding still persist. First, aggressive token reduction can lead to critical information loss, particularly when processing extremely long videos. Second, these methods still suffer from quadratic computational complexity as the number of input frames increases. Third, token reduction-based methods require additional overhead, which might increase the actual wall-clock time. + +In this study, we investigate an orthogonal direction to previous approaches: instead of compressing the video tokens, we seek to develop an alternative model architecture that improves the efficiency of processing video tokens during training and pre-filling stage of inference. We propose VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. The key insight of our method is that we can design efficient modules to approximate the causal self-attention operation for both text and video tokens in transformer-based LMMs. In particular, we propose to (1) utilize cross-attentionions to update text tokens based on video tokens, which is affordable due to the short length of text tokens, and (2) adopt Mamba-2 [17] to process the massive video tokens with linear complexity. + +Assuming a combined input sequence of $M + N$ tokens, where $M$ is the number of video tokens and $N$ is the number of text tokens, we find that $M$ could be at least 100 times larger than $N$ on many long video tasks ( $M \gg N$ ). Our model can reduce the training/prefilling computational complexity from $O(d(M + N)^2)$ to $O(dMN + d^2M)$ , where $d$ is the hidden dimension, according to Table 1. In practice, this theoretical improvement may not be fully realized due to the hardware underoptimization for Mamba [24]. Nevertheless, we still observe a more than $50\%$ reduction in GPU memory usage and FLOPs/routine during training and inference for long video inputs, as shown in Figure 1. VAMBA can be efficiently trained using $8 \times \mathrm{A}800$ GPUs, whereas other efficient video LMMs such as LongVU [55] and LongLLaVA [65] require 64 and 24 GPUs for training, respectively. By performing a two-stage training, our VAMBA achieves a $4.3\%$ improvement over the best efficient video LMMs on the challenging hour-scale video understanding benchmark LVBench [64]. On other long video understanding datasets like Video-MME [21], MLVU [76] and LongVideoBench [67], VAMBA also achieves top-notch performance. + +Our contributions can be summarized as follows: + +1. We propose VAMBA, a hybrid Mamba-Transformer model for hour-long video understanding. VAMBA's design features efficient modules such as Mamba-2 blocks and cross-attention layers, effectively reducing the computational overhead of transformer-based LMMs. +2. We conduct a comprehensive ablation study and show that employing Mamba-2 blocks and initializing crossattention weights from pretrained self-attention layers are crucial for achieving high performance in VAMBA. +3. Our extensive evaluations demonstrate that VAMBA achieves strong video understanding capabilities. Specifically, VAMBA achieves $4.3\%$ improvement over state-of-the-art efficient video LMMs on the challenging hour-long benchmark LVBench. + +# 2. Preliminaries + +# 2.1. State Space Models and Mamba + +State space models (SSMs) are linear time-invariant systems for modelling continuous signals. A continuous SSM can be expressed by ordinary differential equations (ODEs): + +$$ +h ^ {\prime} (t) = \mathbf {A} h (t) + \mathbf {B} x (t), \quad y (t) = \mathbf {C} h (t), +$$ + +here, $x(t)\in \mathbb{R}$ is the input signal, $h(t)\in \mathbb{R}^N$ is the $N$ -dimensional hidden state and $y(t)\in \mathbb{R}$ is the output signal. $\mathbf{A}\in \mathbb{R}^{N\times N}$ is the state transition matrix and $\mathbf{B}\in \mathbb{R}^{N\times 1}$ , $\mathbf{C}\in \mathbb{R}^{1\times N}$ are projection matrices. + +Mamba [24], and its predecessor S4 [26], are discretized SSMs for discrete sequence modelling. The discretization of the SSM is done by the zero-order hold (ZOH) technique: + +$$ +\overline {{\mathbf {A}}} = \exp (\Delta \mathbf {A}), \quad \overline {{\mathbf {B}}} = (\Delta \mathbf {A}) ^ {- 1} (\exp (\Delta \mathbf {A}) - \mathbf {I}) \cdot \Delta \mathbf {B}, +$$ + +where $\overline{\mathbf{A}},\overline{\mathbf{B}}$ are discretized state variables and $\Delta$ is the step size. The discretized SSM can thus be rewritten as: + +$$ +h _ {t} = \overline {{\mathbf {A}}} h _ {t - 1} + \overline {{\mathbf {B}}} x _ {t}, \quad y _ {t} = \mathbf {C} h _ {t}. +$$ + +Mamba proposes the selective scan algorithm that dynamically determines $\mathbf{B},\mathbf{C},\Delta$ based on the sequence inputs. Following S4, the A matrix is initialized as the diagonalized HiPPO matrix [25]. These design choices are proven to be efficient and effective for modelling extremely long sequences. Mamba-2 [17] further simplifies the formulation of A to a scalar times identity structure. Mamba-2 supports multi-head SSM and allows a much larger state dimension $N$ than Mamba, while being faster during training. + +# 2.2. Transformer-based LMMs + +We study how text and video tokens are being updated in transformers. As shown in Figure 2, we assume a standard case where video tokens are placed before text tokens and ignore the chat template tokens for simplicity. In each LMM decoder layer, the video and text tokens are updated as: + +![](images/3a517d881738d387cd2158cd71f584111048b44de0a22cb8e29a529cfd3650a4.jpg) +(a) Transformer-based Video LMMs +Figure 2. Overview of our VAMBA model architecture. Compared to transformer-based LMMs (left), we replace the costly causal self-attention operations with the more efficient cross-attention layers and Mamba blocks to achieve better efficiency. + +![](images/e5a6f9f66192c3a371d46041fe93b312764fab2cdc5a0ce68ee2f6ef28e73dcc.jpg) +(b) Our Vamba Model Architecture + +- All video tokens get updated by an input layer normalization (LN) layer, a self-attention layer, a post LN layer and a final MLP layer. There are also residual connections for self-attention and MLP layers. The $i^{\text{th}}$ video token $v_{i}$ computes self-attention based on the first $i$ video tokens: + +$$ +o _ {v _ {i}} = \left(\sigma \left(\frac {q _ {v _ {i}} \mathbf {K} _ {[ v _ {1}: v _ {i} ]} ^ {\top}}{\sqrt {d}}\right) \mathbf {V} _ {[ v _ {1}: v _ {i} ]}\right) \mathbf {W} _ {o}, \tag {1} +$$ + +where $\sigma (\cdot)$ denotes the softmax operation, $q_{v_i}$ is the query vector for $v_{i}$ and $\mathbf{K}_{[v_1:v_i]}$ , $\mathbf{V}_{[v_1:v_i]}$ are key and value matrices of the first $i$ video tokens. $d$ denotes the dimension of the query and key vectors and $\mathbf{W}_o$ is the out projection matrix. $o_{v_i}$ represents the final output for token $v_{i}$ . + +- The text tokens share a similar update route to the video tokens. The self-attention updates are slightly different: for the $j^{\text{th}}$ text token $t_j$ , it computes self-attention based on all video tokens and the first $j$ text tokens: + +$$ +o _ {t _ {j}} = \left(\sigma \left(\frac {q _ {t _ {j}} \left[ \mathbf {K} _ {v} , \mathbf {K} _ {[ t _ {1} : t _ {j} ]} \right] ^ {\top}}{\sqrt {d}}\right) \left[ \mathbf {V} _ {v}, \mathbf {V} _ {[ t _ {1}: t _ {j} ]} \right]\right) \mathbf {W} _ {o}, \tag {2} +$$ + +where $[\mathbf{K}_v,\mathbf{K}_{[t_1:t_j]}$ and $[\mathbf{V}_v,\mathbf{V}_{[t_1:t_j]}$ are the key and value matrices of the combination of all video tokens and text tokens up to $t_j$ . + +The main computation overhead in the transformer-based LMMs comes from the quadratic complexity of the self-attention in the video tokens. To overcome this issue, we design a hybrid Mamba Transformer architecture to process text and video tokens differently. Our detailed model architecture designs are listed in the sections below. + +# 3. Our Method: VAMBA + +Our goal is to devise a model $\Theta'$ that preserves the performance of transformer $\Theta$ , while being more efficient. We approach this problem by approximating the causal transformer layers in the pretrained video LMMs. + +# 3.1. Updating Text Tokens via Cross-Attentions + +The key idea of our method is to split the expensive self-attention operation over the entire video and text token sequence into two more efficient components. Since video tokens typically dominate the sequence while text tokens remain few, we maintain the self-attention mechanism exclusively for the text tokens and eliminate it for the video tokens. Instead, we add cross-attention layers that use text tokens as queries and video tokens as keys and values. As shown in Figure 2, this design updates the text tokens during the attention layers in our model as follows: + +$$ +\begin{array}{l} o _ {t _ {j}} = (1 - \alpha) \left(\sigma \left(\frac {q _ {t _ {j}} \mathbf {K} _ {v} ^ {\top}}{\sqrt {d}}\right) \mathbf {V} _ {v}\right) \mathbf {W} _ {o} ^ {c} \quad (\text {C r o s s - A t t e n t i o n}) \\ + \alpha \left(\sigma \left(\frac {q _ {t _ {j}} \mathbf {K} _ {[ t _ {1}: t _ {j} ]} ^ {\top}}{\sqrt {d}}\right) \mathbf {V} _ {[ t _ {1}: t _ {j} ]}\right)\left. \right) \mathbf {W} _ {o} ^ {s}, \quad (\text {S e l f - A t t e n t i o n}) \tag {3} \\ \end{array} +$$ + +where $\mathbf{W}_o^c$ and $\mathbf{W}_o^s$ are the output projection matrices for the cross- and self-attention layers. $\alpha \in [0,1]$ is a learnable weighting parameter that balances the cross- and self-attention outputs. Assuming a total of $M$ video tokens and $N$ text tokens, the self-attention operations in the + +transformer-based video LMMs have a (pre-filling) computational complexity of $O(d(M + N)^2)$ . Our self+cross-attention design has the complexity of $O(dN^2)$ (self-attention) and $O(dMN)$ (cross-attention), effectively reducing the pre-filling complexity to $O(dN^2 + dMN) \sim O(dMN)$ , given that $M \gg N \approx d$ . + +Our attention formulation can be viewed as an approximation of the causal self-attention operation (Equation 2), as each text token $q_{t_j}$ can still attend to all video tokens and the first $j$ text tokens. As the video and text tokens share the same channel dimension, the projection matrices in the cross-attention layer will have identical dimensions as in the self-attention layers. This allows us to consider two model design choices: we can either randomly initialize the weights in the cross-attention layers, or we can initialize the cross-attention layers using the self-attention layer weights (i.e. $\mathbf{W}_o^c = \mathbf{W}_o^s$ , same applies to the query, key and value projection matrices $\mathbf{W}_q^c, \mathbf{W}_k^c, \mathbf{W}_v^c$ ). We add these two design choices to the VAMBA design space. + +# 3.2. Updating Video Tokens with Mamba Blocks + +While the self+cross-attention design significantly reduces the model's complexity, relying solely on cross-attention layers can compromise the model's representational power. Specifically, rather than updating video tokens through self-attention and MLP layers (i.e., through the causal transformer block, c.f. Section 2.2), the video tokens remain unchanged after the cross-attention layers. To address this limitation, we seek an efficient architecture to approximate the effects of the transformer blocks. Motivated by the success of the Mamba [17, 24] architectures in image and video modelling [39, 47, 77], we propose to employ Mamba blocks to effectively process the video tokens. As shown in Figure 2, the video token updates can be formulated as: + +$$ +o _ {v _ {i}} = \mathrm {M a m b a} (\mathrm {L N} (v _ {i}), \mathbf {h} _ {v _ {i - 1}}, \overline {{\mathbf {A}}}, \overline {{\mathbf {B}}}, \mathbf {C}), \qquad (4) +$$ + +where $\mathrm{LN}(\cdot)$ is the layer normalization operator and $\mathbf{h}_{v_{i-1}}$ is the context (hidden states) at token $v_{i-1}$ . As a recurrent model, Mamba operates similarly to causal self-attention layers, enabling each video token to update itself based on all preceding tokens. Crucially, Mamba reduces the complexity of updating video tokens from $O(dM^2)$ in causal self-attention to $O(d^2M)$ . Combining the Mamba blocks and the self+cross-attention design, our model achieves an overall complexity of $O(dN^2 + dMN + d^2M) \sim O(dMN + d^2M)$ , which is substantially lower than the quadratic complexity $O(d(M + N)^2)$ in the transformer-based video LMMs. In Section 3.3, we further detail how the Mamba layers are trained to approximate Equation 1. + +We consider employing Mamba or Mamba-2 blocks in Equation 4 in the VAMBA design space. + +# 3.3. Training Paradigm + +We utilize a two-stage training strategy to optimize VAMBA, including a pretraining stage and an instruction-tuning stage. In the pretraining stage, the model is initialized with pretrained transformer-based LMM weights in all modules except for the newly introduced cross-attention and Mamba layers. We freeze the pretrained components and train only the new layers using image captioning data to restore the model's visual understanding capabilities. We explore two types of training objectives during pretraining: (1) the standard language modelling loss $\mathcal{L}_{\mathrm{LM}} = -\frac{1}{T}\sum_{t=1}^{T}\log p(x_t|x_{Model IDCA from SA?Mamba Block TypeStage 1Stage 2G-VEvalLVBenchVidMMEMVBenchref. freetestw/o sub.testAXN/A75.723.747.640.9B✓N/A81.034.251.751.8C✓Mamba81.834.253.453.5D✓Mamba-282.235.354.153.5 + +Table 3. Ablation study results on pretraining VAMBA models with a transformer-based LMM teacher and a distillation loss. + +
Model D + Distillλ value in Stage 1 Training
00.0010.010.512
G-VEval82.1981.0580.6873.6963.6547.61
+ +metric based on GPT-4o [51] to evaluate the quality of the captions. For instruction-tuning (stage 2), we evaluate model performances across three benchmarks: LVBench [64] for hour-long videos, Video-MME [21] for medium-to-long videos, and MVBench [40] for short videos. Ablation study results are shown in Table 2 and Table 3. + +Cross-Attention Layer Initialization Strategy According to Table 2, we find that it is crucial to initialize the cross-attention layer weights using the corresponding self-attention layer from the same decoder level. Compared to Model A, the dramatic performance boost observed in Model B across all metrics underscores the critical impact of this weight initialization strategy. We believe this observation stems from the fact that initializing the cross-attention layer weights from the self-attention layers allows our self+cross-attention operation (Equation 3) to more closely approximate the pretrained causal self-attention operation (Equation 2): once $\mathbf{W}_q^c$ , $\mathbf{W}_k^c$ , $\mathbf{W}_v^c$ and $\mathbf{W}_o^c$ are set equal to $\mathbf{W}_q^s$ , $\mathbf{W}_k^s$ , $\mathbf{W}_v^s$ and $\mathbf{W}_o^s$ , the only discrepancy between the two equations becomes their attention score matrices. This difference can be mitigated by our two-stage training process, making it easier for Model B to recover its multimodal understanding capabilities. + +Mamba Block Design Choices Comparing Model B with Model C and D in Table 2, we observe that both Mamba and Mamba-2 blocks improve the model performances across all metrics, indicating that Mamba blocks bring extra representation power to the model's visual understanding capabilities. Furthermore, when comparing Model C and Model D, we find that Mamba-2 demonstrates superior image and video modelling capabilities. Despite having a simplified + +Table 4. Comparisons between baseline models and VAMBA on hour-long video understanding benchmarks. Bold: best results among efficient video LMMs. Underline: second-best. + +
ModelsSizeHour-Long Video Understanding
LVBenchHourVideoHourEval
testdevtest
Transformer-based LMMs
Gemini-1.5-Pro [59]-33.137.3-
GPT-4o [51]-34.719.6-
Qwen2-VL [63]7B42.033.853.0
Efficient LMMs
LLaVA-Mini [74]7B17.620.624.2
LongLLaVA [65]9B31.227.739.1
LongVU [55]7B37.830.846.8
Video-XL [56]7B36.833.047.1
VAMBA10B42.133.650.7
+ +A matrix structure, Mamba-2 accommodates a larger state dimension than Mamba (64 versus 16 in our case), which likely contributes to its improved performance. + +Pretraining Distillation Loss We employ the pretrained Qwen2-VL-7B model as the teacher model and perform a series of pretraining experiments with the additional distillation loss based on our best model setting (Model D in Table 2). The results are shown in Table 3. In contrast to previous findings from cross-attention-based LLMs (e.g. CEPE [70]), where distillation loss is reported to be beneficial, we observe that incorporating an additional distillation loss does not further improve our model's performance, as the G-VEval score decreases for all $\lambda > 0$ . Therefore, we exclude the distillation loss and only employ the language modelling loss in both stages in our final training setting. + +# 4.2. Main Evaluation Results + +Implementation Details Based on our ablation study (Table 2), we adopt Model D as our final design and initialize VAMBA from Qwen2-VL-7B. We source $\sim 3\mathrm{M}$ image caption data for pretraining and $\sim 7\mathrm{M}$ image and video QA data for instruction-tuning. The full implementation details can be found in Appendix 7. + +Hour-Long Video Understanding We evaluate our model's ability to handle extremely long videos using two public benchmarks: LVBench [64] and HourVideo [8] (development set). We further compose a new benchmark called HourEval by selecting all questions related to videos longer than 30 minutes from Video-MME [21], MLVU [76] development set, and LongVideoBench [67] validation set. The average video lengths for LVBench, HourVideo, and HourEval are 68.4, 47.2, and 54.7 minutes, respectively. We compare our model against four efficient video LMMs: + +Table 5. Comparisons between baseline models and VAMBA on medium-length or short video understanding benchmarks. Bold: best results among each section. Underline: second-best. * indicates the results based on our evaluation scripts. + +
ModelsSizeMedium-Length UnderstandingShort Video UnderstandingVideo Captioning
Video-MMEMLVULongVideoBenchMVBenchNExT-QADREAM-1K
w/o subtitlesm-avgvaltestmcF1
Proprietary Models
GPT-4V [2]-59.949.259.143.570.434.4
GPT-4o [51]-71.964.666.7-76.039.2
Gemini-1.5-Pro [59]-75.062.964.054.276.436.2
Open-source Transformer-based LMMs
VideoChat2 [40]7B39.547.939.351.978.626.6
ShareGPT4Video [12]7B39.946.439.751.2-19.5
LongVA [73]7B52.456.351.849.268.3-
Video-CCAM [20]9B50.358.5-64.6--
Kangaroo [46]8B56.061.054.861.1--
InternVL2 [15]8B54.048.1*51.8*66.4-26.9
LLaVA-OneVision [35]7B58.262.6*56.456.779.431.7
Qwen2-VL [63]7B63.364.2*52.4*67.0-29.6
Phi-4-Mini [1]5.6B55.060.146.760.4--
Open-source Efficient LMMs
LLaVA-Mini [74]7B40.344.319.3*44.547.6*22.9
LongLLaVA [65]9B51.653.3*42.1*54.672.2*24.6
LongVU [55]7B55.3*65.453.5*66.978.0*28.1
Video-XL [56]7B55.564.950.755.377.5*23.5
VAMBA10B57.865.955.960.478.128.1
+ +LLaVA-Mini [74], LongLLaVA [65], Video-XL [56] and LongVU [55]. We also include the results from Qwen2-VL-7B (our baseline transformer-based LMM) as a reference. + +Experimental results are shown in Table 4. Our VAMBA consistently outperforms all efficient video LMMs across the three hour-long video benchmarks, highlighting its exceptional ability to understand and reason over hour-scale videos. Notably, our model surpasses the baseline Qwen2-VL-7B on the LVBench benchmark, and its performance on HourVideo is also very close to Qwen2-VL-7B. These results underscore that our VAMBA is competitive with the best open-sourced transformer-based LMMs, while being significantly more efficient during training and inference. + +Medium-Length or Short Video Understanding To further demonstrate our model's generalizability across various video durations, we test VAMBA on several medium-length or short video understanding benchmarks. Specifically, we report multiple-choice question performance on Video-MME, MLVU, LongVideoBench, NExT-QA [68] and MVBench [40]. We also evaluate VAMBA on DREAM1K [62] for video captioning assessment. + +According to Table 5, our VAMBA demonstrates superior performance across three medium-length video understanding benchmarks (with average video durations between 10-20 minutes), ranking first among efficient video LMMs + +on all metrics. Despite the newly integrated cross-attention and Mamba layers in VAMBA have only been trained on relatively smaller datasets, VAMBA's performance remains competitive with large-scale pretrained transformer-based LMMs. Furthermore, all our evaluations can be conducted on a single 80G GPU, while some transformer-based LMMs, such as Qwen2-VL-7B, require additional inference strategies like sliding window attention or ring attention over multiple GPUs to achieve optimal results. + +For short video understanding and video captioning benchmarks, VAMBA also achieves competitive performances, ranking first on NExT-QA and DREAM-1K and second on MVBench among efficient LMMs. Overall, our model delivers the best results on medium-length and long video benchmarks, demonstrating its strong ability to handle long-context video-language inputs. + +# 4.3. Runtime Efficiency Analysis + +To understand our model's runtime efficiency gains over the baseline transformer-based LMM (Qwen2-VL-7B), we conduct an efficiency analysis for both training and inference. For model training, we ensure a consistent training environment by using 8 NVIDIA A800 80G GPUs for both models. All input videos are resized to a resolution of $640 \times 360$ , and a batch size of 1 is maintained + +![](images/3feca9c59421931918412e9a4553f52726e2e7ab077cfa4c9592323f92c15b68.jpg) +(a) Training GPU Memory Usage Comparison + +![](images/50cca20c3ce4702e9080e9b525804e9b1d14b58f6efd05273f758e096a57391f.jpg) +(b) Training Runtime Comparison (time for 1 training step) +Figure 3. Comparison of training GPU memory usage and runtime per training step between Qwen2-VL-7B and VAMBA. + +on each GPU. We apply standard optimization methods such as Flash-Attention 2 [16], DeepSpeed ZeRO-3 [52] and gradient checkpointing [13] for both models. Based on different numbers of input frames, we measure the average GPU memory across 8 GPUs and the runtime per training step during training. As shown in Figure 3a, although VAMBA and Qwen2-VL-7B share the same vision encoder design and generate vision tokens of identical sequence lengths, our model requires over $50\%$ less training memory when processing videos with more than 16 frames. This efficiency gain allows us to handle a larger number of video frames during training (512 vs. 128). Furthermore, as depicted in Figure 3b, our efficient design also accelerates model training, achieving nearly twice the speedup per training step when working with more than 64 frames. + +For model inference, we focus on the pre-fill stage and analyze GPU memory usage and FLOPs for both models. All measurements are conducted on an input video with a resolution of $640 \times 360$ . As shown in Figure 4a, VAMBA requires slightly more GPU memory than Qwen2-VL-7B when the number of input frames is low (fewer than 32 frames), due to the additional $\sim 3\mathrm{B}$ parameters in its crossattention and Mamba layers. However, VAMBA's memory usage increases more slowly as the frame increases, allowing it to handle four times as many frames on a single NVIDIA A800 80G GPU compared to Qwen2-VL-7B (256 vs. 1024). Regarding computational cost, VAMBA reduces FLOPs by $30\%$ to $50\%$ during inference (Figure 4b), demonstrating a significantly lower complexity than its transformer-based LMM counterparts. + +![](images/98babf5f86533cc20a404a777a757e910274104e6d2218be930603d997538b7a.jpg) +(a) Pre-filling GPU Memory Usage Comparison + +![](images/0a131ca9eab62e084a6939cd2defbc8b21e612e41362fed1a070e0ba436fb181.jpg) +(b) Pre-filling FLOPs Comparison (measured using calflops) +Figure 4. Comparison of GPU memory usage and FLOPs between Qwen2-VL-7B and VAMBA during inference. + +# 4.4. Case Study + +In this section, we perform a case study for Video-XL and VAMBA using different video inputs and show the results in Figure 5. The "cookie" example focuses on object property identification in a long video (36 minutes). Video-XL incorrectly identifies the cookie colors as green and white, possibly due to the video also containing scenes for making other desserts (e.g. white ice cream in the frame highlighted in red). On the other hand, VAMBA more accurately describes them as green and brown. In the "magic trick" example, Video-XL misinterprets the rubber band as a green ring, whereas VAMBA correctly describes the performer manipulating the rubber band around their fingers, making it appear to pass through. These comparisons highlight how VAMBA is better at retrieving relevant events in long videos and capturing fine-grained visual details and actions. + +# 5. Related Work + +Large Multimodal Models (LMMs) Large Multimodal Models (LMMs) have rapidly evolved in recent years. Recent research has focused on improving LMM's instruction-following ability through better data curation [11, 35, 40, 44, 53, 75], building better vision encoders that can process images at higher resolutions [34, 44, 63], and extending LMMs to interleaved image [30, 34, 36] or video [42, 49, 73] understanding. To enhance instruction-following capabilities, LLaVA-1.5 [43] and LLaVA-NeXT [44] leverage GPT-4-generated multimodal data for knowledge-enhanced tuning. To optimize visual feature extraction, Idefics2 [34] leverages the NaViT [18] strategy and the SigLIP [71] vision encoder to support image inputs with native resolu + +![](images/4c27668e7f0b7c47741c99762dac5d47be31ea5c124d10313c0e03dde36a4aa7.jpg) + +![](images/58d2ca9b7ea72c1b8bbd8d6f8821ebbd55ebd80b5d70665d8b9fa24a419310a3.jpg) +Question: What are the colors of the cookies that the person is making? + +![](images/a5c71d377979acb30ee263547d5d90e2d851ed2df198c7d843b61b35659827ac.jpg) + +![](images/4bbe69722d277735b454af9c9e62d9f52e3d08ff95c857c2667f37d8717e8478.jpg) + +![](images/a7f4717494f37c8fdecef0c001e7e3a41168023e85835f28002fda0ab812145d.jpg) + +Video-XL: The cookies have different colored fillings, with at least one being green and another one being white. + +Vamba: The person is making cookies in two different colors, green and brown. + +![](images/7d5c59c3a458e72a26ab0c727008cd1bd7d0aec6c17ca310388816d544537aad.jpg) +Figure 5. Qualitative comparison between VAMBA and efficient LMM baselines. Red text denotes incorrect responses, while green text highlights the correct responses by our VAMBA. + +![](images/397057d1cca01445e4b0fe53943e50f5fc7ee9547b6a9472c9b815f74590ac4a.jpg) +Question: Describe the magic trick. + +![](images/9085a9b7f018e6f524726f231f65c07c2a5ad02c7885f798bb25a7d69bbcbfba.jpg) + +![](images/617c6a3c94026710fd5e0550ecaa7d9a9be61cb486a4eb28e9b56a263e3502fc.jpg) + +![](images/1b72dd582d0ed219f524b5ef1f7ba826a80ec9822d6ccb81a8cea2d1ac072148.jpg) + +Video-XL: The magic trick involves a person holding a green ring and a piece of string... The person's fingers are spread out, and they seem to be guiding the ring around the string. + +Vamba: ...The person demonstrates a magic trick by manipulating the rubber band around their fingers, making it appear to pass through their hand. + +tion. Qwen2-VL [63] applied Multimodal RoPE (M-RoPE) to optimize feature extraction for varying input sizes. + +Despite significant advances, current state-of-the-art LMMs remain inefficient at handling long-context multimodal inputs due to the quadratic complexity of causal self-attention operations. To mitigate this issue, methods like Flamingo [3], OpenFlamingo [4], Idefics [33] and LLaMA 3.2 [19] interleave gated cross-attention layers with self-attention layers for vision-text modelling. mPlug-Owl3 [69] develops the hyper-attention layer that integrates self-attention with cross-attention to support multimodal understanding. However, these models generally underperform compared to LLaVA-like transformer-based LMMs. In this study, we identify that one possible reason is the lack of vision token updates for cross-attention-based models. Our VAMBA addresses this limitation by incorporating additional Mamba layers to update vision tokens. + +Efficient LLMs/LMMs To improve the efficiency of LMM reasoning, recent work [37, 38, 40, 55, 56, 72, 74] develops techniques to reduce the length of the vision token sequence. Video-LLaMA [72] and VideoChat [38, 40] use a Q-Former [37] to compress dense visual tokens into a more compact sequence. LLaVA-Mini [74] further reduces each image to only one token. Video-XL [56] introduces a Visual Summarization Token to summarize the visual information, while LongVU [55] designs a spatiotemporal adaptive compression mechanism to reduce the number of video tokens. + +Another research avenue, exemplified by the Mamba [17, 24] model series, focuses on developing linear-time sequence models to reduce complexity. Mamba-based [78] and Mamba-Transformer hybrid models [22, 41] have achieved notable success in the realm of pure language modelling. However, these efficient architectures have received little attention in the multimodal understanding domain. LongLLaVA [65] took an initial approach by incorporating the hybrid Jamba [41] model as the LMM decoder, yet its performance in long video understanding remains suboptimal. This shortfall is likely due to the absence of a strong hybrid LLM as the base model and the lack of large-scale training on multimodal instruction-following data. + +# 6. Conclusion + +We presented VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. By integrating cross-attention for text tokens and Mamba-2 blocks for video token updates, our approach reduces computational complexity and GPU memory usage while achieving competitive performance across long, medium, and short video benchmarks. Extensive evaluations on datasets such as LVBench demonstrate VAMBA's superiority over existing efficient video LMMs. For future work, since our model design is orthogonal to token reduction-based methods, we will focus on combining the strengths of both approaches to develop more efficient VAMBA variants. + +# References + +[1] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, YenChun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Wei-jian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 6 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8 +[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8 +[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1 +[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 1 +[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 1 +[8] Keshigeyan Chandrasegaran, Agrim Gupta, Lea M Hadzic, Taran Kota, Jimming He, Cristóbal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li. Hourvideo: 1-hour video-language understanding. Advances in Neural Information Processing Systems, 37:53168-53197, 2025. 5, 2 +[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceed- + +ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 4 +[10] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for lite vision-language models. arXiv preprint arXiv:2402.11684, 2024. 1 +[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 7, 1 +[12] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. Advances in Neural Information Processing Systems, 37:19472-19495, 2025. 6, 1 +[13] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 7, 1 +[14] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 1 +[15] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 6 +[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 7, 1 +[17] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 2, 4, 8 +[18] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36:2252-2274, 2023. 7 +[19] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8 +[20] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 1, 6 +[21] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5 + +[22] Paolo Glorioso, Quentin Anthony, Yury Tokpanov, James Whittington, Jonathan Pilault, Adam Ibrahim, and Beren Millidge. Zamba: A compact 7b ssm hybrid model. arXiv preprint arXiv:2405.16712, 2024. 8 +[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2 +[24] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 4, 8 +[25] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2 +[26] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2 +[27] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 1 +[28] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Gedas Bertasius, and Lorenzo Torresani. Bimba: Selective-scan compression for long-range video question answering. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29096-29107, 2025. 3 +[29] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1 +[30] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 7 +[31] Jindong Jiang, Xiuyu Li, Zhijian Liu, Muyang Li, Guo Chen, Zhiqi Li, De-An Huang, Guilin Liu, Zhiding Yu, Kurt Keutzer, et al. Token-efficient long video understanding for multimodal llms. arXiv preprint arXiv:2503.04130, 2025. 3 +[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1 +[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36:71683-71702, 2023. 8 +[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 7 + +[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 6, 7 +[36] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 7 +[37] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 8 +[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 8 +[39] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2024. 4 +[40] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 5, 6, 7, 8, 2 +[41] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 8 +[42] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7 +[43] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 7, 1 +[44] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024. Accessed: 2025-02-14. 1, 7 +[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1 +[46] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 6 +[47] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model 2024. arXiv preprint arXiv:2401.10166, 2024. 4 +[48] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand + +spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1 +[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 1 +[50] OpenAI. Chatgpt. https://openai.com/index/chatgpt/, 2023.1 +[51] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.5,6 +[52] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 7, 1 +[53] Weiming Ren, Huan Yang, Jie Min, Cong Wei, and Wenhu Chen. Vista: Enhancing long-duration and high-resolution video understanding by video spatiotemporal augmentation. arXiv preprint arXiv:2412.00927, 2024. 7 +[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1 +[55] Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434, 2024. 1, 2, 5, 6, 8 +[56] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 1, 5, 6, 8 +[57] Vasu Singla, Kaiyu Yue, Sukriti Paul, Reza Shirkavand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav Bhatele, Gowthami Somepalli, and Tom Goldstein. From pixels to prose: A large dataset of dense image captions. arXiv preprint arXiv:2406.10328, 2024. 4 +[58] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1 +[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 5, 6 +[60] Tony Cheng Tong, Sirui He, Zhiwen Shao, and Dit-Yan Yeung. G-veval: A versatile metric for evaluating image and video captions using gpt-4o. arXiv preprint arXiv:2412.13647, 2024. 4 +[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. + +Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[62] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 6, 2 +[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4, 5, 6, 7, 8 +[64] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 1, 2, 5 +[65] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. Longlava: Scaling multi-modal llms to 1000 images efficiently via a hybrid architecture. arXiv preprint arXiv:2409.02889, 2024. 2, 5, 6, 8, 1 +[66] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 1 +[67] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 5 +[68] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6, 2 +[69] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8 +[70] Howard Yen. Long-context language modeling with parallel context encoding. Master's thesis, Princeton University, 2024. 5 +[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 7 +[72] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, Singapore, 2023. Association for Computational Linguistics. 1, 8 +[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6, 7 + +[74] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. arXiv preprint arXiv:2501.03895, 2025. 5, 6, 8, 1 +[75] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 4, 7 +[76] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5 +[77] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 4 +[78] Jingwei Zuo, Maksim Velikanov, Dhia Eddine Rhaiem, Ilyas Chahed, Younes Belkada, Guillaume Kunsch, and Hakim Hacid. Falcon mamba: The first competitive attention-free 7b language model. arXiv preprint arXiv:2410.05355, 2024. 8 \ No newline at end of file diff --git a/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/images.zip b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..50c0ba067ab0fa734ce470b83d60af69f6250478 --- /dev/null +++ b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44421ee2524008cdcbe3c9cdec2718414b67b89ed9042cd3de66c2150ee097d1 +size 588003 diff --git a/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/layout.json b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18926a5b6d98044767f8a2c5b42aab0d4fdc41ad --- /dev/null +++ b/2025/Vamba_ Understanding Hour-Long Videos with Hybrid Mamba-Transformers/layout.json @@ -0,0 +1,10279 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 549, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 549, + 121 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 549, + 121 + ], + "type": "text", + "content": "VAMBA: Understanding Hour-Long Videos with Hybrid Mamba-Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": "Weiming Ren" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Wentao Ma" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Huan Yang" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Cong Wei" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Ge Zhang" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Wenhu Chen" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": "University of Waterloo, " + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": "University of Toronto, " + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": "Kuaishou Technology, " + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": "Vector Institute, " + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 142, + 543, + 171 + ], + "type": "text", + "content": "M-A-P" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 173, + 441, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 173, + 441, + 185 + ], + "spans": [ + { + "bbox": [ + 169, + 173, + 441, + 185 + ], + "type": "text", + "content": "{w2ren,wenhuchen}@uwaterloo.ca，hyang@fastmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 186, + 416, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 186, + 416, + 198 + ], + "spans": [ + { + "bbox": [ + 190, + 186, + 416, + 198 + ], + "type": "text", + "content": "https://tiger-ai-lab.github.io/Vamba/" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 209, + 200, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 209, + 200, + 221 + ], + "spans": [ + { + "bbox": [ + 151, + 209, + 200, + 221 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "text", + "content": "State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames " + }, + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "inline_equation", + "content": "(640\\times 360)" + }, + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "text", + "content": " on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least " + }, + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "text", + "content": " reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by " + }, + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 54, + 235, + 297, + 498 + ], + "type": "text", + "content": " on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 524, + 135, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 135, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 135, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "The field of large multimodal models (LMMs) has seen tremendous progress in recent years. The seminal work of LLaVA [45] successfully transferred the power of autoregressive next-token prediction from large language models (LLMs) [2, 5, 29, 50, 61] to the multimodal domain, establishing a new standard for visual understanding by autoregressively modelling visual and language tokens through causal transformer layers. Since then, transformer-based LMMs have been widely studied and enhanced to tackle a diverse range of tasks, such as high-resolution image understanding [34, 44, 63] and interleaved image-text reasoning [30, 35, 36]. Nevertheless, the quadratic complexity inherent in causal self-attention operations presents significant challenges for long-context inputs, making it difficult for" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 209, + 553, + 338 + ], + "blocks": [ + { + "bbox": [ + 317, + 209, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 338 + ], + "type": "image", + "image_path": "17a515573ce5fe186ccc074041b0885395218a1fdc97ee11e002b101f513bb10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 341, + 552, + 427 + ], + "blocks": [ + { + "bbox": [ + 317, + 341, + 552, + 427 + ], + "lines": [ + { + "bbox": [ + 317, + 341, + 552, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 341, + 552, + 427 + ], + "type": "image", + "image_path": "0a085ff6a07ec325bd8a84af43ef36e1775977b79d260e07ca2a2ba876793e68.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 429, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 462 + ], + "type": "text", + "content": "Figure 1. VAMBA achieves strong long video understanding performance (42.1% on LVBench [64]) while being more computationally efficient compared to transformer-based LMMs." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 472, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 496 + ], + "type": "text", + "content": "transformer-based LMMs to effectively handle the task of understanding extremely long videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 556, + 715 + ], + "type": "text", + "content": "A key challenge for advanced LMMs when processing long video inputs is that they tend to encode each frame into a large number of vision tokens, which leads to substantial computational and memory costs during both training and inference. For example, Qwen2-VL [63] can only process 256 frames (360p) on a single GPU, which is far from sufficient for hour-long video understanding. To reduce the computational/memory cost, previous efforts have primarily focused on reducing the vision tokens in the input sequence. Several approaches [20, 38, 40, 72] utilize a Q-Former [37] to compress vision tokens. More recent methods, such as LongVU [55] and Video-XL [56], partition vision token sequences into chunks and employ adaptive token compression mechanisms to decrease the token count. Another line of work, such as InternVideo2.5 [66] and Orxy-1.5 [48], employs various mechanisms to evaluate the vision tokens' importance, and thus dropping or merging those less significant tokens [6] during training" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21197" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 113, + 291, + 159 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "text", + "content": "Table 1. Theoretical time and memory complexity for transformer-based LMMs and VAMBA in pre-filling. " + }, + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "text", + "content": " represent the number of vision and text tokens. " + }, + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "text", + "content": " denotes the hidden dimension." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 113, + 291, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 113, + 291, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 113, + 291, + 159 + ], + "type": "table", + "html": "
ModelTime ComplexityMemory Complexity
TransformerO(d(M+N)2)O((M+N)2)
VAMBAO(dMN+d2M)O(MN)
", + "image_path": "760de367d39acb0675d3e5920af1632b762a21f5a922cce51fe8ef6868127cec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 169, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 169, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 169, + 294, + 266 + ], + "type": "text", + "content": "and inference. Despite these advancements, key challenges in LMMs for long video understanding still persist. First, aggressive token reduction can lead to critical information loss, particularly when processing extremely long videos. Second, these methods still suffer from quadratic computational complexity as the number of input frames increases. Third, token reduction-based methods require additional overhead, which might increase the actual wall-clock time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 267, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 434 + ], + "type": "text", + "content": "In this study, we investigate an orthogonal direction to previous approaches: instead of compressing the video tokens, we seek to develop an alternative model architecture that improves the efficiency of processing video tokens during training and pre-filling stage of inference. We propose VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. The key insight of our method is that we can design efficient modules to approximate the causal self-attention operation for both text and video tokens in transformer-based LMMs. In particular, we propose to (1) utilize cross-attentionions to update text tokens based on video tokens, which is affordable due to the short length of text tokens, and (2) adopt Mamba-2 [17] to process the massive video tokens with linear complexity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": "Assuming a combined input sequence of " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "M + N" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " tokens, where " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " is the number of video tokens and " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " is the number of text tokens, we find that " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " could be at least 100 times larger than " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " on many long video tasks (" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "M \\gg N" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": "). Our model can reduce the training/prefilling computational complexity from " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "O(d(M + N)^2)" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "O(dMN + d^2M)" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " is the hidden dimension, according to Table 1. In practice, this theoretical improvement may not be fully realized due to the hardware underoptimization for Mamba [24]. Nevertheless, we still observe a more than " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " reduction in GPU memory usage and FLOPs/routine during training and inference for long video inputs, as shown in Figure 1. VAMBA can be efficiently trained using " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{A}800" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " GPUs, whereas other efficient video LMMs such as LongVU [55] and LongLLaVA [65] require 64 and 24 GPUs for training, respectively. By performing a two-stage training, our VAMBA achieves a " + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 55, + 436, + 296, + 700 + ], + "type": "text", + "content": " improvement over the best efficient video LMMs on the challenging hour-scale video understanding benchmark LVBench [64]. On other long video understanding datasets like Video-MME [21], MLVU [76] and LongVideoBench [67], VAMBA also achieves top-notch performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 255, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 255, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 255, + 713 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 310, + 72, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 310, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 553, + 132 + ], + "type": "text", + "content": "1. We propose VAMBA, a hybrid Mamba-Transformer model for hour-long video understanding. VAMBA's design features efficient modules such as Mamba-2 blocks and cross-attention layers, effectively reducing the computational overhead of transformer-based LMMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 310, + 133, + 553, + 179 + ], + "type": "text", + "content": "2. We conduct a comprehensive ablation study and show that employing Mamba-2 blocks and initializing crossattention weights from pretrained self-attention layers are crucial for achieving high performance in VAMBA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 180, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 180, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 310, + 180, + 553, + 239 + ], + "type": "text", + "content": "3. Our extensive evaluations demonstrate that VAMBA achieves strong video understanding capabilities. Specifically, VAMBA achieves " + }, + { + "bbox": [ + 310, + 180, + 553, + 239 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 310, + 180, + 553, + 239 + ], + "type": "text", + "content": " improvement over state-of-the-art efficient video LMMs on the challenging hour-long benchmark LVBench." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 250, + 399, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 250, + 399, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 250, + 399, + 262 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 270, + 488, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 270, + 488, + 282 + ], + "spans": [ + { + "bbox": [ + 314, + 270, + 488, + 282 + ], + "type": "text", + "content": "2.1. State Space Models and Mamba" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "content": "State space models (SSMs) are linear time-invariant systems for modelling continuous signals. A continuous SSM can be expressed by ordinary differential equations (ODEs):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 331, + 520, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 331, + 520, + 345 + ], + "spans": [ + { + "bbox": [ + 347, + 331, + 520, + 345 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} (t) = \\mathbf {A} h (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C} h (t),", + "image_path": "46fc1f9a0c106874ecdbc4ef07839d7f5550a36053f044f3f9261bb116c84073.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "inline_equation", + "content": "x(t)\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": " is the input signal, " + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "inline_equation", + "content": "h(t)\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": "-dimensional hidden state and " + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "inline_equation", + "content": "y(t)\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": " is the output signal. " + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": " is the state transition matrix and " + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{B}\\in \\mathbb{R}^{N\\times 1}" + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{C}\\in \\mathbb{R}^{1\\times N}" + }, + { + "bbox": [ + 313, + 352, + 554, + 401 + ], + "type": "text", + "content": " are projection matrices." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "content": "Mamba [24], and its predecessor S4 [26], are discretized SSMs for discrete sequence modelling. The discretization of the SSM is done by the zero-order hold (ZOH) technique:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 444, + 547, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 444, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 319, + 444, + 547, + 460 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\quad \\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} (\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}) \\cdot \\Delta \\mathbf {B},", + "image_path": "a6e8ad88b99b4d8af5d90a09993cbbec3f228967a65eafcc2c3ba4b9f320fa06.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}},\\overline{\\mathbf{B}}" + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": " are discretized state variables and " + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 468, + 554, + 491 + ], + "type": "text", + "content": " is the step size. The discretized SSM can thus be rewritten as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 364, + 499, + 504, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 499, + 504, + 514 + ], + "spans": [ + { + "bbox": [ + 364, + 499, + 504, + 514 + ], + "type": "interline_equation", + "content": "h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\quad y _ {t} = \\mathbf {C} h _ {t}.", + "image_path": "aff561e33347969d6d3865c85d28895aa8217899f80bba82e97b2ba9bfb34d1f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 631 + ], + "type": "text", + "content": "Mamba proposes the selective scan algorithm that dynamically determines " + }, + { + "bbox": [ + 313, + 521, + 554, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{B},\\mathbf{C},\\Delta" + }, + { + "bbox": [ + 313, + 521, + 554, + 631 + ], + "type": "text", + "content": " based on the sequence inputs. Following S4, the A matrix is initialized as the diagonalized HiPPO matrix [25]. These design choices are proven to be efficient and effective for modelling extremely long sequences. Mamba-2 [17] further simplifies the formulation of A to a scalar times identity structure. Mamba-2 supports multi-head SSM and allows a much larger state dimension " + }, + { + "bbox": [ + 313, + 521, + 554, + 631 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 521, + 554, + 631 + ], + "type": "text", + "content": " than Mamba, while being faster during training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 635, + 463, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 463, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 463, + 647 + ], + "type": "text", + "content": "2.2. Transformer-based LMMs" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "We study how text and video tokens are being updated in transformers. As shown in Figure 2, we assume a standard case where video tokens are placed before text tokens and ignore the chat template tokens for simplicity. In each LMM decoder layer, the video and text tokens are updated as:" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21198" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 74, + 294, + 319 + ], + "blocks": [ + { + "bbox": [ + 89, + 74, + 294, + 319 + ], + "lines": [ + { + "bbox": [ + 89, + 74, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 74, + 294, + 319 + ], + "type": "image", + "image_path": "3a517d881738d387cd2158cd71f584111048b44de0a22cb8e29a529cfd3650a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 320, + 266, + 330 + ], + "lines": [ + { + "bbox": [ + 123, + 320, + 266, + 330 + ], + "spans": [ + { + "bbox": [ + 123, + 320, + 266, + 330 + ], + "type": "text", + "content": "(a) Transformer-based Video LMMs" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 332, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 55, + 332, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 555, + 355 + ], + "type": "text", + "content": "Figure 2. Overview of our VAMBA model architecture. Compared to transformer-based LMMs (left), we replace the costly causal self-attention operations with the more efficient cross-attention layers and Mamba blocks to achieve better efficiency." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 74, + 529, + 319 + ], + "blocks": [ + { + "bbox": [ + 304, + 74, + 529, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 74, + 529, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 74, + 529, + 319 + ], + "type": "image", + "image_path": "e5a6f9f66192c3a371d46041fe93b312764fab2cdc5a0ce68ee2f6ef28e73dcc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 320, + 482, + 330 + ], + "lines": [ + { + "bbox": [ + 343, + 320, + 482, + 330 + ], + "spans": [ + { + "bbox": [ + 343, + 320, + 482, + 330 + ], + "type": "text", + "content": "(b) Our Vamba Model Architecture" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "text", + "content": "- All video tokens get updated by an input layer normalization (LN) layer, a self-attention layer, a post LN layer and a final MLP layer. There are also residual connections for self-attention and MLP layers. The " + }, + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "text", + "content": " video token " + }, + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "text", + "content": " computes self-attention based on the first " + }, + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 365, + 295, + 426 + ], + "type": "text", + "content": " video tokens:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 430, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 430, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 295, + 459 + ], + "type": "interline_equation", + "content": "o _ {v _ {i}} = \\left(\\sigma \\left(\\frac {q _ {v _ {i}} \\mathbf {K} _ {[ v _ {1}: v _ {i} ]} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {[ v _ {1}: v _ {i} ]}\\right) \\mathbf {W} _ {o}, \\tag {1}", + "image_path": "f48d2d94e7ff951f0e7e7106d65deb6ba18a05fb514b42b76ba0a927ee1ecf68.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " denotes the softmax operation, " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "q_{v_i}" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " is the query vector for " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{[v_1:v_i]}" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{[v_1:v_i]}" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " are key and value matrices of the first " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " video tokens. " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " denotes the dimension of the query and key vectors and " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " is the out projection matrix. " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "o_{v_i}" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": " represents the final output for token " + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 63, + 463, + 295, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "text", + "content": "- The text tokens share a similar update route to the video tokens. The self-attention updates are slightly different: for the " + }, + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "inline_equation", + "content": "j^{\\text{th}}" + }, + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "text", + "content": " text token " + }, + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "text", + "content": ", it computes self-attention based on all video tokens and the first " + }, + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 523, + 295, + 570 + ], + "type": "text", + "content": " text tokens:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 574, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 574, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 75, + 574, + 295, + 602 + ], + "type": "interline_equation", + "content": "o _ {t _ {j}} = \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\left[ \\mathbf {K} _ {v} , \\mathbf {K} _ {[ t _ {1} : t _ {j} ]} \\right] ^ {\\top}}{\\sqrt {d}}\\right) \\left[ \\mathbf {V} _ {v}, \\mathbf {V} _ {[ t _ {1}: t _ {j} ]} \\right]\\right) \\mathbf {W} _ {o}, \\tag {2}", + "image_path": "e1659f96f6bb32a105bb292722a63d5e6d4ae5b819aa1372a86749d588b9edc2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "inline_equation", + "content": "[\\mathbf{K}_v,\\mathbf{K}_{[t_1:t_j]}" + }, + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "inline_equation", + "content": "[\\mathbf{V}_v,\\mathbf{V}_{[t_1:t_j]}" + }, + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "text", + "content": " are the key and value matrices of the combination of all video tokens and text tokens up to " + }, + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 63, + 605, + 295, + 642 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "The main computation overhead in the transformer-based LMMs comes from the quadratic complexity of the self-attention in the video tokens. To overcome this issue, we design a hybrid Mamba Transformer architecture to process text and video tokens differently. Our detailed model architecture designs are listed in the sections below." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 440, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 440, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 440, + 376 + ], + "type": "text", + "content": "3. Our Method: VAMBA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "content": "Our goal is to devise a model " + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\Theta'" + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "content": " that preserves the performance of transformer " + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 385, + 555, + 434 + ], + "type": "text", + "content": ", while being more efficient. We approach this problem by approximating the causal transformer layers in the pretrained video LMMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 443, + 538, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 538, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 538, + 455 + ], + "type": "text", + "content": "3.1. Updating Text Tokens via Cross-Attentions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 581 + ], + "type": "text", + "content": "The key idea of our method is to split the expensive self-attention operation over the entire video and text token sequence into two more efficient components. Since video tokens typically dominate the sequence while text tokens remain few, we maintain the self-attention mechanism exclusively for the text tokens and eliminate it for the video tokens. Instead, we add cross-attention layers that use text tokens as queries and video tokens as keys and values. As shown in Figure 2, this design updates the text tokens during the attention layers in our model as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 586, + 553, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 586, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 317, + 586, + 553, + 653 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} o _ {t _ {j}} = (1 - \\alpha) \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\mathbf {K} _ {v} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {v}\\right) \\mathbf {W} _ {o} ^ {c} \\quad (\\text {C r o s s - A t t e n t i o n}) \\\\ + \\alpha \\left(\\sigma \\left(\\frac {q _ {t _ {j}} \\mathbf {K} _ {[ t _ {1}: t _ {j} ]} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {[ t _ {1}: t _ {j} ]}\\right)\\left. \\right) \\mathbf {W} _ {o} ^ {s}, \\quad (\\text {S e l f - A t t e n t i o n}) \\tag {3} \\\\ \\end{array}", + "image_path": "ea388c445d4c62cf424693f6d02564e3e5c12e19beda4bef736db026aba31429.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o^c" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o^s" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " are the output projection matrices for the cross- and self-attention layers. " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " is a learnable weighting parameter that balances the cross- and self-attention outputs. Assuming a total of " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " video tokens and " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " text tokens, the self-attention operations in the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21199" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "transformer-based video LMMs have a (pre-filling) computational complexity of " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "O(d(M + N)^2)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ". Our self+cross-attention design has the complexity of " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "O(dN^2)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " (self-attention) and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "O(dMN)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " (cross-attention), effectively reducing the pre-filling complexity to " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "O(dN^2 + dMN) \\sim O(dMN)" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", given that " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "M \\gg N \\approx d" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "text", + "content": "Our attention formulation can be viewed as an approximation of the causal self-attention operation (Equation 2), as each text token " + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "inline_equation", + "content": "q_{t_j}" + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "text", + "content": " can still attend to all video tokens and the first " + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "text", + "content": " text tokens. As the video and text tokens share the same channel dimension, the projection matrices in the cross-attention layer will have identical dimensions as in the self-attention layers. This allows us to consider two model design choices: we can either randomly initialize the weights in the cross-attention layers, or we can initialize the cross-attention layers using the self-attention layer weights (i.e. " + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o^c = \\mathbf{W}_o^s" + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "text", + "content": ", same applies to the query, key and value projection matrices " + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^c, \\mathbf{W}_k^c, \\mathbf{W}_v^c" + }, + { + "bbox": [ + 55, + 146, + 295, + 301 + ], + "type": "text", + "content": "). We add these two design choices to the VAMBA design space." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 285, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 285, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 285, + 329 + ], + "type": "text", + "content": "3.2. Updating Video Tokens with Mamba Blocks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 491 + ], + "type": "text", + "content": "While the self+cross-attention design significantly reduces the model's complexity, relying solely on cross-attention layers can compromise the model's representational power. Specifically, rather than updating video tokens through self-attention and MLP layers (i.e., through the causal transformer block, c.f. Section 2.2), the video tokens remain unchanged after the cross-attention layers. To address this limitation, we seek an efficient architecture to approximate the effects of the transformer blocks. Motivated by the success of the Mamba [17, 24] architectures in image and video modelling [39, 47, 77], we propose to employ Mamba blocks to effectively process the video tokens. As shown in Figure 2, the video token updates can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 504, + 295, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 504, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 91, + 504, + 295, + 520 + ], + "type": "interline_equation", + "content": "o _ {v _ {i}} = \\mathrm {M a m b a} (\\mathrm {L N} (v _ {i}), \\mathbf {h} _ {v _ {i - 1}}, \\overline {{\\mathbf {A}}}, \\overline {{\\mathbf {B}}}, \\mathbf {C}), \\qquad (4)", + "image_path": "a8bec7530652541c614295407249b2efd06c904cd15b2102dfa3b789411b952b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": " is the layer normalization operator and " + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{v_{i-1}}" + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": " is the context (hidden states) at token " + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "inline_equation", + "content": "v_{i-1}" + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": ". As a recurrent model, Mamba operates similarly to causal self-attention layers, enabling each video token to update itself based on all preceding tokens. Crucially, Mamba reduces the complexity of updating video tokens from " + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "inline_equation", + "content": "O(dM^2)" + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": " in causal self-attention to " + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "inline_equation", + "content": "O(d^2M)" + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": ". Combining the Mamba blocks and the self+cross-attention design, our model achieves an overall complexity of " + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "inline_equation", + "content": "O(dN^2 + dMN + d^2M) \\sim O(dMN + d^2M)" + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": ", which is substantially lower than the quadratic complexity " + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "inline_equation", + "content": "O(d(M + N)^2)" + }, + { + "bbox": [ + 55, + 532, + 295, + 687 + ], + "type": "text", + "content": " in the transformer-based video LMMs. In Section 3.3, we further detail how the Mamba layers are trained to approximate Equation 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We consider employing Mamba or Mamba-2 blocks in Equation 4 in the VAMBA design space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 427, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 427, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 427, + 85 + ], + "type": "text", + "content": "3.3. Training Paradigm" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": "We utilize a two-stage training strategy to optimize VAMBA, including a pretraining stage and an instruction-tuning stage. In the pretraining stage, the model is initialized with pretrained transformer-based LMM weights in all modules except for the newly introduced cross-attention and Mamba layers. We freeze the pretrained components and train only the new layers using image captioning data to restore the model's visual understanding capabilities. We explore two types of training objectives during pretraining: (1) the standard language modelling loss " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LM}} = -\\frac{1}{T}\\sum_{t=1}^{T}\\log p(x_t|x_{Model IDCA from SA?Mamba Block TypeStage 1Stage 2G-VEvalLVBenchVidMMEMVBenchref. freetestw/o sub.testAXN/A75.723.747.640.9B✓N/A81.034.251.751.8C✓Mamba81.834.253.453.5D✓Mamba-282.235.354.153.5", + "image_path": "84f4043db987004ad2217cfe0f323879c1ac13a00b388e6bf4ae53f8e6c825de.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 55, + 265, + 293, + 318 + ], + "blocks": [ + { + "bbox": [ + 55, + 235, + 296, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 258 + ], + "type": "text", + "content": "Table 3. Ablation study results on pretraining VAMBA models with a transformer-based LMM teacher and a distillation loss." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 265, + 293, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 293, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 293, + 318 + ], + "type": "table", + "html": "
Model D + Distillλ value in Stage 1 Training
00.0010.010.512
G-VEval82.1981.0580.6873.6963.6547.61
", + "image_path": "9c6b69ec55e76d7f1f130713d36a58d7e854cefb341a82936b0b9b9ef4390f19.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 328, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 400 + ], + "type": "text", + "content": "metric based on GPT-4o [51] to evaluate the quality of the captions. For instruction-tuning (stage 2), we evaluate model performances across three benchmarks: LVBench [64] for hour-long videos, Video-MME [21] for medium-to-long videos, and MVBench [40] for short videos. Ablation study results are shown in Table 2 and Table 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": "Cross-Attention Layer Initialization Strategy According to Table 2, we find that it is crucial to initialize the cross-attention layer weights using the corresponding self-attention layer from the same decoder level. Compared to Model A, the dramatic performance boost observed in Model B across all metrics underscores the critical impact of this weight initialization strategy. We believe this observation stems from the fact that initializing the cross-attention layer weights from the self-attention layers allows our self+cross-attention operation (Equation 3) to more closely approximate the pretrained causal self-attention operation (Equation 2): once " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^c" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k^c" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v^c" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o^c" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": " are set equal to " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^s" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k^s" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v^s" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o^s" + }, + { + "bbox": [ + 55, + 407, + 296, + 612 + ], + "type": "text", + "content": ", the only discrepancy between the two equations becomes their attention score matrices. This difference can be mitigated by our two-stage training process, making it easier for Model B to recover its multimodal understanding capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Mamba Block Design Choices Comparing Model B with Model C and D in Table 2, we observe that both Mamba and Mamba-2 blocks improve the model performances across all metrics, indicating that Mamba blocks bring extra representation power to the model's visual understanding capabilities. Furthermore, when comparing Model C and Model D, we find that Mamba-2 demonstrates superior image and video modelling capabilities. Despite having a simplified" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 113, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 4. Comparisons between baseline models and VAMBA on hour-long video understanding benchmarks. Bold: best results among efficient video LMMs. Underline: second-best." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 113, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 315, + 113, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 315, + 113, + 553, + 289 + ], + "type": "table", + "html": "
ModelsSizeHour-Long Video Understanding
LVBenchHourVideoHourEval
testdevtest
Transformer-based LMMs
Gemini-1.5-Pro [59]-33.137.3-
GPT-4o [51]-34.719.6-
Qwen2-VL [63]7B42.033.853.0
Efficient LMMs
LLaVA-Mini [74]7B17.620.624.2
LongLLaVA [65]9B31.227.739.1
LongVU [55]7B37.830.846.8
Video-XL [56]7B36.833.047.1
VAMBA10B42.133.650.7
", + "image_path": "35d22dd603d2c51823f54b03f29f60ea4c8ad694e6b302feabe8386efa7a93e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 301, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 338 + ], + "type": "text", + "content": "A matrix structure, Mamba-2 accommodates a larger state dimension than Mamba (64 versus 16 in our case), which likely contributes to its improved performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 344, + 556, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 556, + 489 + ], + "type": "text", + "content": "Pretraining Distillation Loss We employ the pretrained Qwen2-VL-7B model as the teacher model and perform a series of pretraining experiments with the additional distillation loss based on our best model setting (Model D in Table 2). The results are shown in Table 3. In contrast to previous findings from cross-attention-based LLMs (e.g. CEPE [70]), where distillation loss is reported to be beneficial, we observe that incorporating an additional distillation loss does not further improve our model's performance, as the G-VEval score decreases for all " + }, + { + "bbox": [ + 313, + 344, + 556, + 489 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 313, + 344, + 556, + 489 + ], + "type": "text", + "content": ". Therefore, we exclude the distillation loss and only employ the language modelling loss in both stages in our final training setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 497, + 454, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 454, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 454, + 509 + ], + "type": "text", + "content": "4.2. Main Evaluation Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 514, + 556, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 588 + ], + "type": "text", + "content": "Implementation Details Based on our ablation study (Table 2), we adopt Model D as our final design and initialize VAMBA from Qwen2-VL-7B. We source " + }, + { + "bbox": [ + 313, + 514, + 556, + 588 + ], + "type": "inline_equation", + "content": "\\sim 3\\mathrm{M}" + }, + { + "bbox": [ + 313, + 514, + 556, + 588 + ], + "type": "text", + "content": " image caption data for pretraining and " + }, + { + "bbox": [ + 313, + 514, + 556, + 588 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{M}" + }, + { + "bbox": [ + 313, + 514, + 556, + 588 + ], + "type": "text", + "content": " image and video QA data for instruction-tuning. The full implementation details can be found in Appendix 7." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Hour-Long Video Understanding We evaluate our model's ability to handle extremely long videos using two public benchmarks: LVBench [64] and HourVideo [8] (development set). We further compose a new benchmark called HourEval by selecting all questions related to videos longer than 30 minutes from Video-MME [21], MLVU [76] development set, and LongVideoBench [67] validation set. The average video lengths for LVBench, HourVideo, and HourEval are 68.4, 47.2, and 54.7 minutes, respectively. We compare our model against four efficient video LMMs:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "21201" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 101, + 547, + 395 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 5. Comparisons between baseline models and VAMBA on medium-length or short video understanding benchmarks. Bold: best results among each section. Underline: second-best. * indicates the results based on our evaluation scripts." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 101, + 547, + 395 + ], + "lines": [ + { + "bbox": [ + 60, + 101, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 60, + 101, + 547, + 395 + ], + "type": "table", + "html": "
ModelsSizeMedium-Length UnderstandingShort Video UnderstandingVideo Captioning
Video-MMEMLVULongVideoBenchMVBenchNExT-QADREAM-1K
w/o subtitlesm-avgvaltestmcF1
Proprietary Models
GPT-4V [2]-59.949.259.143.570.434.4
GPT-4o [51]-71.964.666.7-76.039.2
Gemini-1.5-Pro [59]-75.062.964.054.276.436.2
Open-source Transformer-based LMMs
VideoChat2 [40]7B39.547.939.351.978.626.6
ShareGPT4Video [12]7B39.946.439.751.2-19.5
LongVA [73]7B52.456.351.849.268.3-
Video-CCAM [20]9B50.358.5-64.6--
Kangaroo [46]8B56.061.054.861.1--
InternVL2 [15]8B54.048.1*51.8*66.4-26.9
LLaVA-OneVision [35]7B58.262.6*56.456.779.431.7
Qwen2-VL [63]7B63.364.2*52.4*67.0-29.6
Phi-4-Mini [1]5.6B55.060.146.760.4--
Open-source Efficient LMMs
LLaVA-Mini [74]7B40.344.319.3*44.547.6*22.9
LongLLaVA [65]9B51.653.3*42.1*54.672.2*24.6
LongVU [55]7B55.3*65.453.5*66.978.0*28.1
Video-XL [56]7B55.564.950.755.377.5*23.5
VAMBA10B57.865.955.960.478.128.1
", + "image_path": "ff986fbd777c6de04864d542d83f7f7b457c992a858f1c9f4f7d98f44495470f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 404, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 295, + 440 + ], + "type": "text", + "content": "LLaVA-Mini [74], LongLLaVA [65], Video-XL [56] and LongVU [55]. We also include the results from Qwen2-VL-7B (our baseline transformer-based LMM) as a reference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 441, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 441, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 295, + 561 + ], + "type": "text", + "content": "Experimental results are shown in Table 4. Our VAMBA consistently outperforms all efficient video LMMs across the three hour-long video benchmarks, highlighting its exceptional ability to understand and reason over hour-scale videos. Notably, our model surpasses the baseline Qwen2-VL-7B on the LVBench benchmark, and its performance on HourVideo is also very close to Qwen2-VL-7B. These results underscore that our VAMBA is competitive with the best open-sourced transformer-based LMMs, while being significantly more efficient during training and inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 665 + ], + "type": "text", + "content": "Medium-Length or Short Video Understanding To further demonstrate our model's generalizability across various video durations, we test VAMBA on several medium-length or short video understanding benchmarks. Specifically, we report multiple-choice question performance on Video-MME, MLVU, LongVideoBench, NExT-QA [68] and MVBench [40]. We also evaluate VAMBA on DREAM1K [62] for video captioning assessment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "According to Table 5, our VAMBA demonstrates superior performance across three medium-length video understanding benchmarks (with average video durations between 10-20 minutes), ranking first among efficient video LMMs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 404, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 513 + ], + "type": "text", + "content": "on all metrics. Despite the newly integrated cross-attention and Mamba layers in VAMBA have only been trained on relatively smaller datasets, VAMBA's performance remains competitive with large-scale pretrained transformer-based LMMs. Furthermore, all our evaluations can be conducted on a single 80G GPU, while some transformer-based LMMs, such as Qwen2-VL-7B, require additional inference strategies like sliding window attention or ring attention over multiple GPUs to achieve optimal results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 598 + ], + "type": "text", + "content": "For short video understanding and video captioning benchmarks, VAMBA also achieves competitive performances, ranking first on NExT-QA and DREAM-1K and second on MVBench among efficient LMMs. Overall, our model delivers the best results on medium-length and long video benchmarks, demonstrating its strong ability to handle long-context video-language inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 610, + 470, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 470, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 470, + 624 + ], + "type": "text", + "content": "4.3. Runtime Efficiency Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "To understand our model's runtime efficiency gains over the baseline transformer-based LMM (Qwen2-VL-7B), we conduct an efficiency analysis for both training and inference. For model training, we ensure a consistent training environment by using 8 NVIDIA A800 80G GPUs for both models. All input videos are resized to a resolution of " + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "inline_equation", + "content": "640 \\times 360" + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": ", and a batch size of 1 is maintained" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21202" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 85, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 100, + 73, + 251, + 83 + ], + "lines": [ + { + "bbox": [ + 100, + 73, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 100, + 73, + 251, + 83 + ], + "type": "text", + "content": "(a) Training GPU Memory Usage Comparison" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 85, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 59, + 85, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 59, + 85, + 293, + 186 + ], + "type": "image", + "image_path": "3feca9c59421931918412e9a4553f52726e2e7ab077cfa4c9592323f92c15b68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 200, + 293, + 298 + ], + "blocks": [ + { + "bbox": [ + 80, + 189, + 271, + 199 + ], + "lines": [ + { + "bbox": [ + 80, + 189, + 271, + 199 + ], + "spans": [ + { + "bbox": [ + 80, + 189, + 271, + 199 + ], + "type": "text", + "content": "(b) Training Runtime Comparison (time for 1 training step)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 200, + 293, + 298 + ], + "lines": [ + { + "bbox": [ + 59, + 200, + 293, + 298 + ], + "spans": [ + { + "bbox": [ + 59, + 200, + 293, + 298 + ], + "type": "image", + "image_path": "50cca20c3ce4702e9080e9b525804e9b1d14b58f6efd05273f758e096a57391f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 295, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 322 + ], + "type": "text", + "content": "Figure 3. Comparison of training GPU memory usage and runtime per training step between Qwen2-VL-7B and VAMBA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 347, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 295, + 528 + ], + "type": "text", + "content": "on each GPU. We apply standard optimization methods such as Flash-Attention 2 [16], DeepSpeed ZeRO-3 [52] and gradient checkpointing [13] for both models. Based on different numbers of input frames, we measure the average GPU memory across 8 GPUs and the runtime per training step during training. As shown in Figure 3a, although VAMBA and Qwen2-VL-7B share the same vision encoder design and generate vision tokens of identical sequence lengths, our model requires over " + }, + { + "bbox": [ + 54, + 347, + 295, + 528 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 347, + 295, + 528 + ], + "type": "text", + "content": " less training memory when processing videos with more than 16 frames. This efficiency gain allows us to handle a larger number of video frames during training (512 vs. 128). Furthermore, as depicted in Figure 3b, our efficient design also accelerates model training, achieving nearly twice the speedup per training step when working with more than 64 frames." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": "For model inference, we focus on the pre-fill stage and analyze GPU memory usage and FLOPs for both models. All measurements are conducted on an input video with a resolution of " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "640 \\times 360" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": ". As shown in Figure 4a, VAMBA requires slightly more GPU memory than Qwen2-VL-7B when the number of input frames is low (fewer than 32 frames), due to the additional " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\sim 3\\mathrm{B}" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " parameters in its crossattention and Mamba layers. However, VAMBA's memory usage increases more slowly as the frame increases, allowing it to handle four times as many frames on a single NVIDIA A800 80G GPU compared to Qwen2-VL-7B (256 vs. 1024). Regarding computational cost, VAMBA reduces FLOPs by " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": " during inference (Figure 4b), demonstrating a significantly lower complexity than its transformer-based LMM counterparts." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 85, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 356, + 73, + 512, + 84 + ], + "lines": [ + { + "bbox": [ + 356, + 73, + 512, + 84 + ], + "spans": [ + { + "bbox": [ + 356, + 73, + 512, + 84 + ], + "type": "text", + "content": "(a) Pre-filling GPU Memory Usage Comparison" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 85, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 317, + 85, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 553, + 186 + ], + "type": "image", + "image_path": "98babf5f86533cc20a404a777a757e910274104e6d2218be930603d997538b7a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 200, + 552, + 297 + ], + "blocks": [ + { + "bbox": [ + 329, + 189, + 539, + 199 + ], + "lines": [ + { + "bbox": [ + 329, + 189, + 539, + 199 + ], + "spans": [ + { + "bbox": [ + 329, + 189, + 539, + 199 + ], + "type": "text", + "content": "(b) Pre-filling FLOPs Comparison (measured using calflops)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 200, + 552, + 297 + ], + "lines": [ + { + "bbox": [ + 317, + 200, + 552, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 200, + 552, + 297 + ], + "type": "image", + "image_path": "0a131ca9eab62e084a6939cd2defbc8b21e612e41362fed1a070e0ba436fb181.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 299, + 553, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 299, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 553, + 321 + ], + "type": "text", + "content": "Figure 4. Comparison of GPU memory usage and FLOPs between Qwen2-VL-7B and VAMBA during inference." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 389, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 389, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 389, + 344 + ], + "type": "text", + "content": "4.4. Case Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 348, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 529 + ], + "type": "text", + "content": "In this section, we perform a case study for Video-XL and VAMBA using different video inputs and show the results in Figure 5. The \"cookie\" example focuses on object property identification in a long video (36 minutes). Video-XL incorrectly identifies the cookie colors as green and white, possibly due to the video also containing scenes for making other desserts (e.g. white ice cream in the frame highlighted in red). On the other hand, VAMBA more accurately describes them as green and brown. In the \"magic trick\" example, Video-XL misinterprets the rubber band as a green ring, whereas VAMBA correctly describes the performer manipulating the rubber band around their fingers, making it appear to pass through. These comparisons highlight how VAMBA is better at retrieving relevant events in long videos and capturing fine-grained visual details and actions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 538, + 401, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 401, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 401, + 550 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "Large Multimodal Models (LMMs) Large Multimodal Models (LMMs) have rapidly evolved in recent years. Recent research has focused on improving LMM's instruction-following ability through better data curation [11, 35, 40, 44, 53, 75], building better vision encoders that can process images at higher resolutions [34, 44, 63], and extending LMMs to interleaved image [30, 34, 36] or video [42, 49, 73] understanding. To enhance instruction-following capabilities, LLaVA-1.5 [43] and LLaVA-NeXT [44] leverage GPT-4-generated multimodal data for knowledge-enhanced tuning. To optimize visual feature extraction, Idefics2 [34] leverages the NaViT [18] strategy and the SigLIP [71] vision encoder to support image inputs with native resolu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21203" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 157, + 183 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 157, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 157, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 157, + 183 + ], + "type": "image", + "image_path": "4c27668e7f0b7c47741c99762dac5d47be31ea5c124d10313c0e03dde36a4aa7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 160, + 74, + 255, + 183 + ], + "blocks": [ + { + "bbox": [ + 160, + 74, + 255, + 183 + ], + "lines": [ + { + "bbox": [ + 160, + 74, + 255, + 183 + ], + "spans": [ + { + "bbox": [ + 160, + 74, + 255, + 183 + ], + "type": "image", + "image_path": "58d2ca9b7ea72c1b8bbd8d6f8821ebbd55ebd80b5d70665d8b9fa24a419310a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 186, + 436, + 198 + ], + "lines": [ + { + "bbox": [ + 173, + 186, + 436, + 198 + ], + "spans": [ + { + "bbox": [ + 173, + 186, + 436, + 198 + ], + "type": "text", + "content": "Question: What are the colors of the cookies that the person is making?" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 258, + 74, + 352, + 183 + ], + "blocks": [ + { + "bbox": [ + 258, + 74, + 352, + 183 + ], + "lines": [ + { + "bbox": [ + 258, + 74, + 352, + 183 + ], + "spans": [ + { + "bbox": [ + 258, + 74, + 352, + 183 + ], + "type": "image", + "image_path": "a5c71d377979acb30ee263547d5d90e2d851ed2df198c7d843b61b35659827ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 356, + 74, + 450, + 183 + ], + "blocks": [ + { + "bbox": [ + 356, + 74, + 450, + 183 + ], + "lines": [ + { + "bbox": [ + 356, + 74, + 450, + 183 + ], + "spans": [ + { + "bbox": [ + 356, + 74, + 450, + 183 + ], + "type": "image", + "image_path": "4bbe69722d277735b454af9c9e62d9f52e3d08ff95c857c2667f37d8717e8478.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 453, + 74, + 548, + 183 + ], + "blocks": [ + { + "bbox": [ + 453, + 74, + 548, + 183 + ], + "lines": [ + { + "bbox": [ + 453, + 74, + 548, + 183 + ], + "spans": [ + { + "bbox": [ + 453, + 74, + 548, + 183 + ], + "type": "image", + "image_path": "a7f4717494f37c8fdecef0c001e7e3a41168023e85835f28002fda0ab812145d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 205, + 299, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 205, + 299, + 228 + ], + "spans": [ + { + "bbox": [ + 69, + 205, + 299, + 228 + ], + "type": "text", + "content": "Video-XL: The cookies have different colored fillings, with at least one being green and another one being white." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 205, + 539, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 205, + 539, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 205, + 539, + 228 + ], + "type": "text", + "content": "Vamba: The person is making cookies in two different colors, green and brown." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 61, + 239, + 157, + 294 + ], + "blocks": [ + { + "bbox": [ + 61, + 239, + 157, + 294 + ], + "lines": [ + { + "bbox": [ + 61, + 239, + 157, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 157, + 294 + ], + "type": "image", + "image_path": "7d5c59c3a458e72a26ab0c727008cd1bd7d0aec6c17ca310388816d544537aad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 555, + 379 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 555, + 379 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison between VAMBA and efficient LMM baselines. Red text denotes incorrect responses, while green text highlights the correct responses by our VAMBA." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 160, + 239, + 255, + 294 + ], + "blocks": [ + { + "bbox": [ + 160, + 239, + 255, + 294 + ], + "lines": [ + { + "bbox": [ + 160, + 239, + 255, + 294 + ], + "spans": [ + { + "bbox": [ + 160, + 239, + 255, + 294 + ], + "type": "image", + "image_path": "397057d1cca01445e4b0fe53943e50f5fc7ee9547b6a9472c9b815f74590ac4a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 296, + 370, + 307 + ], + "lines": [ + { + "bbox": [ + 238, + 296, + 370, + 307 + ], + "spans": [ + { + "bbox": [ + 238, + 296, + 370, + 307 + ], + "type": "text", + "content": "Question: Describe the magic trick." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 257, + 239, + 352, + 294 + ], + "blocks": [ + { + "bbox": [ + 257, + 239, + 352, + 294 + ], + "lines": [ + { + "bbox": [ + 257, + 239, + 352, + 294 + ], + "spans": [ + { + "bbox": [ + 257, + 239, + 352, + 294 + ], + "type": "image", + "image_path": "9085a9b7f018e6f524726f231f65c07c2a5ad02c7885f798bb25a7d69bbcbfba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 356, + 239, + 451, + 294 + ], + "blocks": [ + { + "bbox": [ + 356, + 239, + 451, + 294 + ], + "lines": [ + { + "bbox": [ + 356, + 239, + 451, + 294 + ], + "spans": [ + { + "bbox": [ + 356, + 239, + 451, + 294 + ], + "type": "image", + "image_path": "617c6a3c94026710fd5e0550ecaa7d9a9be61cb486a4eb28e9b56a263e3502fc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 452, + 239, + 549, + 294 + ], + "blocks": [ + { + "bbox": [ + 452, + 239, + 549, + 294 + ], + "lines": [ + { + "bbox": [ + 452, + 239, + 549, + 294 + ], + "spans": [ + { + "bbox": [ + 452, + 239, + 549, + 294 + ], + "type": "image", + "image_path": "1b72dd582d0ed219f524b5ef1f7ba826a80ec9822d6ccb81a8cea2d1ac072148.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 313, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 313, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 67, + 313, + 296, + 348 + ], + "type": "text", + "content": "Video-XL: The magic trick involves a person holding a green ring and a piece of string... The person's fingers are spread out, and they seem to be guiding the ring around the string." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 313, + 533, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 313, + 533, + 347 + ], + "spans": [ + { + "bbox": [ + 311, + 313, + 533, + 347 + ], + "type": "text", + "content": "Vamba: ...The person demonstrates a magic trick by manipulating the rubber band around their fingers, making it appear to pass through their hand." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 409 + ], + "type": "text", + "content": "tion. Qwen2-VL [63] applied Multimodal RoPE (M-RoPE) to optimize feature extraction for varying input sizes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 410, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 589 + ], + "type": "text", + "content": "Despite significant advances, current state-of-the-art LMMs remain inefficient at handling long-context multimodal inputs due to the quadratic complexity of causal self-attention operations. To mitigate this issue, methods like Flamingo [3], OpenFlamingo [4], Idefics [33] and LLaMA 3.2 [19] interleave gated cross-attention layers with self-attention layers for vision-text modelling. mPlug-Owl3 [69] develops the hyper-attention layer that integrates self-attention with cross-attention to support multimodal understanding. However, these models generally underperform compared to LLaVA-like transformer-based LMMs. In this study, we identify that one possible reason is the lack of vision token updates for cross-attention-based models. Our VAMBA addresses this limitation by incorporating additional Mamba layers to update vision tokens." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Efficient LLMs/LMMs To improve the efficiency of LMM reasoning, recent work [37, 38, 40, 55, 56, 72, 74] develops techniques to reduce the length of the vision token sequence. Video-LLaMA [72] and VideoChat [38, 40] use a Q-Former [37] to compress dense visual tokens into a more compact sequence. LLaVA-Mini [74] further reduces each image to only one token. Video-XL [56] introduces a Visual Summarization Token to summarize the visual information, while LongVU [55] designs a spatiotemporal adaptive compression mechanism to reduce the number of video tokens." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 385, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 540 + ], + "type": "text", + "content": "Another research avenue, exemplified by the Mamba [17, 24] model series, focuses on developing linear-time sequence models to reduce complexity. Mamba-based [78] and Mamba-Transformer hybrid models [22, 41] have achieved notable success in the realm of pure language modelling. However, these efficient architectures have received little attention in the multimodal understanding domain. LongLLaVA [65] took an initial approach by incorporating the hybrid Jamba [41] model as the LMM decoder, yet its performance in long video understanding remains suboptimal. This shortfall is likely due to the absence of a strong hybrid LLM as the base model and the lack of large-scale training on multimodal instruction-following data." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We presented VAMBA, a hybrid Mamba-Transformer model for efficient hour-long video understanding. By integrating cross-attention for text tokens and Mamba-2 blocks for video token updates, our approach reduces computational complexity and GPU memory usage while achieving competitive performance across long, medium, and short video benchmarks. Extensive evaluations on datasets such as LVBench demonstrate VAMBA's superiority over existing efficient video LMMs. For future work, since our model design is orthogonal to token reduction-based methods, we will focus on combining the strengths of both approaches to develop more efficient VAMBA variants." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21204" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 310 + ], + "type": "text", + "content": "[1] Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, YenChun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, Mei Gao, Min Gao, Amit Garg, Abhishek Goswami, Junheng Hao, Amr Hendy, Yuxuan Hu, Xin Jin, Mahmoud Khademi, Dongwoo Kim, Young Jin Kim, Gina Lee, Jinyu Li, Yunsheng Li, Chen Liang, Xihui Lin, Zeqi Lin, Mengchen Liu, Yang Liu, Gilsinia Lopez, Chong Luo, Piyush Madan, Vadim Mazalov, Ali Mousavi, Anh Nguyen, Jing Pan, Daniel Perez-Becker, Jacob Platin, Thomas Portet, Kai Qiu, Bo Ren, Liliang Ren, Sambuddha Roy, Ning Shang, Yelong Shen, Saksham Singhal, Subhojit Som, Xia Song, Tetyana Sych, Praneetha Vaddamanu, Shuohang Wang, Yiming Wang, Zhenghao Wang, Haibin Wu, Haoran Xu, Wei-jian Xu, Yifan Yang, Ziyi Yang, Donghan Yu, Ishmam Zabir, Jianwen Zhang, Li Lyna Zhang, Yunan Zhang, and Xiren Zhou. Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras, 2025. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 312, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 312, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 61, + 312, + 294, + 365 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 367, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 295, + 432 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 434, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 498 + ], + "type": "text", + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 546, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 294, + 578 + ], + "type": "text", + "content": "[6] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 580, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 294, + 612 + ], + "type": "text", + "content": "[7] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 614, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 677 + ], + "type": "text", + "content": "[8] Keshigeyan Chandrasegaran, Agrim Gupta, Lea M Hadzic, Taran Kota, Jimming He, Cristóbal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li. Hourvideo: 1-hour video-language understanding. Advances in Neural Information Processing Systems, 37:53168-53197, 2025. 5, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 294, + 713 + ], + "type": "text", + "content": "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceed-" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "text", + "content": "[10] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for lite vision-language models. arXiv preprint arXiv:2402.11684, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 205 + ], + "type": "text", + "content": "[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 7, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 206, + 553, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 269 + ], + "type": "text", + "content": "[12] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. Advances in Neural Information Processing Systems, 37:19472-19495, 2025. 6, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[13] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 7, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "text", + "content": "[14] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 4, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "text", + "content": "[15] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 416, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 447 + ], + "type": "text", + "content": "[16] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691, 2023. 7, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "text", + "content": "[17] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 2, 4, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 494, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 558 + ], + "type": "text", + "content": "[18] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36:2252-2274, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "text", + "content": "[19] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[20] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[21] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21205" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[22] Paolo Glorioso, Quentin Anthony, Yury Tokpanov, James Whittington, Jonathan Pilault, Adam Ibrahim, and Beren Millidge. Zamba: A compact 7b ssm hybrid model. arXiv preprint arXiv:2405.16712, 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 185 + ], + "type": "text", + "content": "[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[24] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[25] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[26] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[27] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. arXiv preprint arXiv:2402.11530, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "text", + "content": "[28] Md Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Gedas Bertasius, and Lorenzo Torresani. Bimba: Selective-scan compression for long-range video question answering. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 29096-29107, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "text", + "content": "[29] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[30] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. arXiv preprint arXiv:2405.01483, 2024. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 545 + ], + "type": "text", + "content": "[31] Jindong Jiang, Xiuyu Li, Zhijian Liu, Muyang Li, Guo Chen, Zhiqi Li, De-An Huang, Guilin Liu, Zhiding Yu, Kurt Keutzer, et al. Token-efficient long video understanding for multimodal llms. arXiv preprint arXiv:2503.04130, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 611 + ], + "type": "text", + "content": "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "text", + "content": "[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36:71683-71702, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[36] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "text", + "content": "[37] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[38] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 1, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[39] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. In European Conference on Computer Vision, pages 237-255. Springer, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "text", + "content": "[40] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 5, 6, 7, 8, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "text", + "content": "[41] Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[42] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[43] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 7, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 557 + ], + "type": "text", + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024. Accessed: 2025-02-14. 1, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[45] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "text", + "content": "[46] Jiajun Liu, Yibing Wang, Hanghang Ma, Xiaoping Wu, Xiaqi Ma, Xiaoming Wei, Jianbin Jiao, Enhua Wu, and Jie Hu. Kangaroo: A powerful video-language model supporting long-context video input. arXiv preprint arXiv:2408.15542, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[47] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model 2024. arXiv preprint arXiv:2401.10166, 2024. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[48] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21206" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "type": "text", + "content": "[50] OpenAI. Chatgpt. https://openai.com/index/chatgpt/, 2023.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 186 + ], + "type": "text", + "content": "[51] OpenAI. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.5,6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 186, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 243 + ], + "type": "text", + "content": "[52] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC20: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1-16. IEEE, 2020. 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 287 + ], + "type": "text", + "content": "[53] Weiming Ren, Huan Yang, Jie Min, Cong Wei, and Wenhu Chen. Vista: Enhancing long-duration and high-resolution video understanding by video spatiotemporal augmentation. arXiv preprint arXiv:2412.00927, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 289, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 294, + 343 + ], + "type": "text", + "content": "[54] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 409 + ], + "type": "text", + "content": "[55] Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434, 2024. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "text", + "content": "[56] Yan Shu, Peitian Zhang, Zheng Liu, Minghao Qin, Junjie Zhou, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 1, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "text", + "content": "[57] Vasu Singla, Kaiyu Yue, Sukriti Paul, Reza Shirkavand, Mayuka Jayawardhana, Alireza Ganjdanesh, Heng Huang, Abhinav Bhatele, Gowthami Somepalli, and Tom Goldstein. From pixels to prose: A large dataset of dense image captions. arXiv preprint arXiv:2406.10328, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "text", + "content": "[58] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 635 + ], + "type": "text", + "content": "[59] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[60] Tony Cheng Tong, Sirui He, Zhiwen Shao, and Dit-Yan Yeung. G-veval: A versatile metric for evaluating image and video captions using gpt-4o. arXiv preprint arXiv:2412.13647, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[61] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[62] Jiawei Wang, Liping Yuan, Yuchen Zhang, and Haomiao Sun. Tarsier: Recipes for training and evaluating large video description models. arXiv preprint arXiv:2407.00634, 2024. 6, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 143, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 198 + ], + "type": "text", + "content": "[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 200, + 553, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 553, + 244 + ], + "type": "text", + "content": "[64] Weihan Wang, Zehai He, Wenyi Hong, Yean Cheng, Xiaohan Zhang, Ji Qi, Xiaotao Gu, Shiyu Huang, Bin Xu, Yuxiao Dong, et al. Lvbench: An extreme long video understanding benchmark. arXiv preprint arXiv:2406.08035, 2024. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 246, + 553, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 291 + ], + "type": "text", + "content": "[65] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. Longlava: Scaling multi-modal llms to 1000 images efficiently via a hybrid architecture. arXiv preprint arXiv:2409.02889, 2024. 2, 5, 6, 8, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 293, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 346 + ], + "type": "text", + "content": "[66] Yi Wang, Xinhao Li, Ziang Yan, Yinan He, Jiashuo Yu, Xi-angyu Zeng, Chenting Wang, Changlian Ma, Haian Huang, Jianfei Gao, et al. Intervideo2. 5: Empowering video mllms with long and rich context modeling. arXiv preprint arXiv:2501.12386, 2025. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 349, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 393 + ], + "type": "text", + "content": "[67] Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828-28857, 2025. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 395, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 449 + ], + "type": "text", + "content": "[68] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 452, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 507 + ], + "type": "text", + "content": "[69] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 509, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 540 + ], + "type": "text", + "content": "[70] Howard Yen. Long-context language modeling with parallel context encoding. Master's thesis, Princeton University, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 589 + ], + "type": "text", + "content": "[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "type": "text", + "content": "[72] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 543–553, Singapore, 2023. Association for Computational Linguistics. 1, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[73] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 6, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21207" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 305 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[74] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large multimodal models with one vision token. arXiv preprint arXiv:2501.03895, 2025. 5, 6, 8, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[75] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "text", + "content": "[76] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[77] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 305 + ], + "type": "text", + "content": "[78] Jingwei Zuo, Maksim Velikanov, Dhia Eddine Rhaiem, Ilyas Chahed, Younes Belkada, Guillaume Kunsch, and Hakim Hacid. Falcon mamba: The first competitive attention-free 7b language model. arXiv preprint arXiv:2410.05355, 2024. 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21208" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_content_list.json b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0c843f4d52fcdaf08b30ce576b45f04c4295d929 --- /dev/null +++ b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_content_list.json @@ -0,0 +1,1696 @@ +[ + { + "type": "text", + "text": "Variance-Based Pruning for Accelerating and Compressing Trained Networks", + "text_level": 1, + "bbox": [ + 106, + 128, + 893, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Uranik Berisha12, Jens Mehnert1 and Alexandru Paul Condurache12", + "bbox": [ + 230, + 179, + 764, + 198 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1Automated Driving Research, Robert Bosch GmbH, 70469 Stuttgart, Germany", + "$^{2}$ Institute for Signal Processing, University of Lübeck, 23562 Lübeck, Germany" + ], + "bbox": [ + 259, + 200, + 733, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{Uranik.Berisha,JensEricMarkus.Mehnert,AlexandruPaul.Condurache}@de.bosch.com", + "bbox": [ + 156, + 236, + 834, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Increasingly expensive training of ever larger models such as Vision Transformers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge.", + "bbox": [ + 88, + 318, + 482, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over $70\\%$ of its original performance and requires only 10 epochs of fine-tuning to regain $99\\%$ of the original accuracy while simultaneously reducing MACs by $35\\%$ and model size by $36\\%$ , thus speeding up the model by $1.44 \\times$ .", + "bbox": [ + 89, + 500, + 483, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 709, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In both Computer Vision (CV) and Natural Language Processing (NLP), Transformers have become the dominant architecture [8, 21]. In CV, Vision Transformers (ViTs) [8] and their derivatives, such as Data-efficient image Transformers (DeiTs) [21], Swin-Transformers (Swins) [14] and ConvNeXts [15], achieve remarkable results but at significant costs in three main areas: training, storage, and inference. These large models require long training times, often spanning hundreds of epochs to achieve their performance, extensive memory to store their parameters, and high computational resources for inference causing high latency [21].", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fixing all three issues simultaneously remains especially challenging. For example, using the available collection of trained networks provided by the research community is a straightforward way to solve the first problem of costly training. However, this does not address the storage and inference costs, which can still make the use of these trained networks impractical for deployment on resource-constrained hardware [1]. To address these demands, various model pruning techniques have been proposed. These range from removing individual weights within layers in unstructured pruning [12] to removing entire neurons, layers, or blocks in structured pruning [1].", + "bbox": [ + 511, + 287, + 903, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While unstructured pruning can often retain a high degree of accuracy in trained networks, it is difficult to translate these theoretical gains into real-world speed-ups [1]. Since modern hardware is optimized for dense matrix multiplications, they cannot exploit the resulting sparsity effectively and therefore cannot solve the inference costs [1, 12].", + "bbox": [ + 511, + 470, + 903, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, structured pruning leads to straightforward reductions in both inference costs and memory, but again requires extensive retraining to recover the lost accuracy resulting from these major structural modifications. For instance, Global Vision Transformer Pruning with Hessian-Aware Saliency (NViT) [27] requires 300 epochs of fine-tuning despite using trained networks, highlighting how addressing memory and inference speeds alone can still demand large additional training overhead.", + "bbox": [ + 511, + 565, + 905, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, dynamic pruning approaches, such as token pruning, avoid permanently modifying the model architecture and instead dynamically reduce the number of features processed by the network [3, 19]. This minimizes the accuracy drop after pruning, indicating that the model preserved most important features and representation, thus providing a better starting point for subsequent fine-tuning. Token Merging (ToMe) [3] embraces this idea by merging tokens to reuse the pruned information, thus reducing the accuracy drop even further and allowing off-the-shelf deployment without any fine-tuning. While this solves the inference costs as well as the training costs, since the methods do not modify the model structure itself, the memory foot", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4973", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b74940e48b256274fae1d8c185235d3ee92e28525b305283a24a1edea1d4b689.jpg", + "image_caption": [ + "(a) We compute neuron-wise mean and variance of activations at the hidden layer of each MLP." + ], + "image_footnote": [], + "bbox": [ + 120, + 92, + 328, + 305 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9afc80bf2650b05c6d6475840a0765e1b7d3da968b26d255b997c3737b9d4311.jpg", + "image_caption": [ + "(b) We prune neurons with the smallest variances and replace their activation with the mean." + ], + "image_footnote": [], + "bbox": [ + 374, + 90, + 609, + 320 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d396469f25068fe47f42ee8df55f899860e48c449ca37d7276b05d0b7cd0495.jpg", + "image_caption": [ + "(c) We shift the mean activation of the removed neurons into the bias of the next layer.", + "Figure 1. Schematic visualization of the steps of Variance-Based Pruning. (a) Activation Statistics Computation, (b) Variance-Based Pruning, (c) Mean-Shift Compensation" + ], + "image_footnote": [], + "bbox": [ + 645, + 95, + 885, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "print remains unchanged [3]. This again is a limitation for hardware-constrained settings [1]. Fully resolving all three problems of training costs, storage, and inference speed simultaneously remains a challenge.", + "bbox": [ + 88, + 426, + 482, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we fill this gap by introducing Variance-Based Pruning (VBP), a simple and structured pruning approach that provides speed-ups and memory savings with minimal fine-tuning. In order to keep the method simple and allow deployment across a wide variety of architectures, we prune only the heavy Multi-Layer Perceptron (MLP) layers within the transformer blocks. In the first step, our approach leverages activation statistics gathered in the hidden layers of the MLP using Welford's algorithm [24] (see Fig. 1a). In the second step, these statistics are used to identify and remove the least impactful hidden neurons (see Fig. 1b). In the last step, which we refer to as Mean-Shift Compensation, we use the mean activation values of pruned neurons to mitigate accuracy drop by redistributing these mean contributions into the bias of the output layer (see Fig. 1c). This retains high levels of model performance and requires only minimal fine-tuning to then regain most of the original accuracy.", + "bbox": [ + 91, + 489, + 483, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "text_level": 1, + "bbox": [ + 89, + 777, + 415, + 791 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce VBP, a simple low-cost structured pruning method broadly applicable across various architectures to significantly reduce computational costs while maintaining performance.", + "- We apply our method to multiple architectures, including ViTs, Swin-Transformers, and ConvNeXts, demonstrating its generalizability." + ], + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We compare VBP to similarly straightforward neuron-importance measures adapted for structured pruning, as well as to state-of-the-art (SoTA) pruning approaches. We benchmark against NViT pruning, highlighting the benefits of our accuracy retention in constrained fine-tuning settings. We also apply VBP on top of ToMe demonstrating orthogonality and achieving $2 \\times$ speed-ups using our hybrid approach.", + "- We perform an extensive ablation study, demonstrating the efficacy of each component of our method and provide a sensitivity analysis for the variance as a pruning criterion." + ], + "bbox": [ + 513, + 426, + 906, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. RELATED WORK", + "text_level": 1, + "bbox": [ + 513, + 633, + 700, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Transformer Architectures", + "text_level": 1, + "bbox": [ + 513, + 662, + 756, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Initially introduced in NLP by [22], Transformers have largely replaced Convolutional Neural Network (CNN), since their adaption to CV by [8] with the ViT. ViTs processes images as a sequence of non-overlapping patches, and have since been further improved. DeiT [21] refined the training procedure to decrease the training costs and reduce the data reliance. More recent architectures [7, 9, 10], such as Swin [14] and the transformer-inspired ConvNeXt [15], which use hierarchical designs, have built upon this architecture to improve efficiency and scalability. However, despite these advances, these large models remain computationally demanding, in part due to their dense MLP blocks, which significantly impacts inference speed and memory requirements [22, 30].", + "bbox": [ + 511, + 688, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4974", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Model Compression", + "text_level": 1, + "bbox": [ + 89, + 90, + 281, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reduce the costs of large models, model compression efforts focus on scaling down different components of a network. Traditionally pruning aims to identify and remove redundant structures within the network [12, 26]. For example, magnitude pruning eliminates weights with the smallest absolute values, while SNIP and related methods [13, 25] identify crucial connections using gradient-based sensitivities. Structured pruning approaches, remove entire neurons, layers or filters producing smaller matrices that allow direct computational savings in current hardware [1, 5, 27-29, 32]. Most closely related to our method are works exploring variance-based criteria for structured pruning, such as Molchanov et al. [18], which performs dropout at the neuron level and can introduce sparsity in neurons with high dropout rates. Weinstein et al. [23] use these statistics in the initial training phase to remove neurons during training. These approaches operate on models before training, improving training time but still requiring the models to be trained from scratch. Contrary to that, NViT [27] utilizes trained ViTs. They apply iterative structured dimensionality reductions across the entire ViT architecture and significantly reduce costs. However, these structured modifications still have a substantial impact on performance, requiring high pruning iterations of 50 epochs and retraining costs of 300 epochs to regain accuracy.", + "bbox": [ + 89, + 112, + 483, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An alternative to structurally modifying the network is dynamic pruning, which adaptively selects and processes only a subset of features or combines features during inference to reduce costs. For CNNs, these have shown promising speed-accuracy trade-offs [4, 16]. Analogously for transformer-like models, token pruning reduces the number of tokens to be processed [3, 17, 19, 20]. For instance, Bolya et al. [3] merges similar tokens rather than pruning them, thereby conserving most of the information and retaining high accuracy before the fine-tuning. However, by design these methods cannot reduce the memory footprint.", + "bbox": [ + 89, + 489, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Our Approach", + "text_level": 1, + "bbox": [ + 89, + 667, + 240, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work builds upon these ideas by combining (i) the advantages of structured pruning by removing low-variance activation neurons with (ii) the advantages of dynamic pruning by incooperating their mean contribution into the bias of the last MLP layer. Our approach is based on the observation that neurons within MLP layers exhibit varying levels of contribution to the overall representation, as established in works on emerging modularity in trained Transformers for both NLP and CV [2, 31]. By applying our variance-based pruning approach to the widespread MLP layers, we provide a structured method applicable across a broad range of models. Our method is particularly beneficial for transformer-like architectures, where MLP layers heavily contribute to the computational overhead [22, 30].", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we detail the components involved in our proposed VBP approach and provide a mathematical derivation, justifying our pruning criterion and compensation strategy.", + "bbox": [ + 511, + 114, + 906, + 175 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method consists of three major steps:", + "bbox": [ + 511, + 176, + 790, + 190 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Activation Statistics Computation The neuron activation statistics are computed using Welford's algorithm.", + "2. Variance-Based Pruning The neurons to be pruned are selected based on the lowest activation variance.", + "3. Mean-Shift Compensation The mean activations of the selected neurons are added back into the output bias." + ], + "bbox": [ + 511, + 190, + 903, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Step 1: Activation Statistics Computation", + "text_level": 1, + "bbox": [ + 511, + 289, + 870, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we aim for a simple and widely adoptable method, we prune only the MLPs. This requires reducing the hidden layers of the network while keeping the input and output layers intact to ensure seamless integration with the rest of the model structure by maintaining dimensional consistency in the input and output.", + "bbox": [ + 511, + 311, + 906, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "W.l.o.g. we therefore consider a MLP with a single hidden layer of dimension $D_{\\mathrm{hid}}$ , that maps an input vector $\\mathbf{x} \\in \\mathbb{R}^{D_{\\mathrm{in}}}$ to an output vector $\\mathbf{y} \\in \\mathbb{R}^{D_{\\mathrm{out}}}$ . Using a pointwise nonlinear activation function $\\sigma(\\cdot)$ , this MLP can then be described as a function", + "bbox": [ + 511, + 402, + 905, + 477 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\sigma \\left(\\mathbf {W} _ {1} \\mathbf {x} + \\mathbf {b} _ {1}\\right), \\quad \\mathbf {y} = \\mathbf {W} _ {2} \\mathbf {h} + \\mathbf {b} _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 488, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_1 \\in \\mathbb{R}^{D_{\\mathrm{hid}} \\times D_{\\mathrm{in}}}$ , $\\mathbf{b}_1 \\in \\mathbb{R}^{D_{\\mathrm{hid}}}$ , $\\mathbf{W}_2 \\in \\mathbb{R}^{D_{\\mathrm{out}} \\times D_{\\mathrm{hid}}}$ , $\\mathbf{b}_2 \\in \\mathbb{R}^{D_{\\mathrm{out}}}$ . Since we deal with trained networks, these parameters $\\mathbf{W}_1, \\mathbf{b}_1, \\mathbf{W}_2, \\mathbf{b}_2$ remain fixed during pruning, which will become important in the Mean-Shift Compensation in Sec. 3.3.", + "bbox": [ + 511, + 513, + 905, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the hidden layer, we now want to calculate the perneuron mean and variance vectors", + "bbox": [ + 511, + 590, + 905, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} = \\left(\\mu_ {1}, \\dots , \\mu_ {D _ {\\mathrm {h i d}}}\\right) ^ {\\top} \\quad \\text {a n d} \\quad \\boldsymbol {\\sigma} ^ {2} = \\left(\\sigma_ {1} ^ {2}, \\dots , \\sigma_ {D _ {\\mathrm {h i d}}} ^ {2}\\right) ^ {\\top}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 628, + 895, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To compute these statistics, we capture the activations $\\mathbf{h}$ for every batch and use Welford's algorithm to compute a running mean $\\pmb{\\mu}^{(j)}$ and running second moment $\\mathbf{m}_2^{(j)}$ for an efficient online calculation of streaming data*. Given a total of $N$ data samples, let the $j$ -th sample $\\mathbf{h}^{(j)} \\in \\mathbb{R}^{D_{\\mathrm{hid}}}$ be the activation vector after the nonlinearity. Welford's algorithm ensures numerically stable calculations by updating $\\pmb{\\mu}^{(j)}$ and $\\mathbf{m}_2^{(j)}$ after observing the $j$ -th sample as follows:", + "bbox": [ + 511, + 659, + 906, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {(j)} = \\frac {j - 1}{j} \\boldsymbol {\\mu} ^ {(j - 1)} + \\frac {1}{j} \\mathbf {h} ^ {(j)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 794, + 906, + 827 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {2} ^ {(j)} = \\mathbf {m} _ {2} ^ {(j - 1)} + \\left(\\mathbf {h} ^ {(j)} - \\boldsymbol {\\mu} ^ {(j - 1)}\\right) \\odot \\left(\\mathbf {h} ^ {(j)} - \\boldsymbol {\\mu} ^ {(j)}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 835, + 903, + 856 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "*Our method does not require a minimum number of data samples for calculating activation statistics, but using an efficient online computation allows the use of as many data samples for pruning as available.", + "bbox": [ + 511, + 863, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4975", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ denotes elementwise multiplication. Once all $N$ samples are processed, the variance is given by:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\sigma} ^ {2} = \\frac {\\mathbf {m} _ {2} ^ {(N)}}{N - 1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 232, + 130, + 482, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Step 2: Variance-Based Pruning", + "text_level": 1, + "bbox": [ + 89, + 172, + 375, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using this variance vector $\\sigma^2 \\in \\mathbb{R}^{D_{\\mathrm{hid}}}$ , we rank neurons by their variance values, $\\sigma_i^2$ . Intuitively, if an activation rarely deviates from its mean $\\mu_i$ , the corresponding neuron contributes less to the expressiveness of the network. The pruning decision is formed across all hidden neurons in the network, to capture the neuron with the least variance across all layers.", + "bbox": [ + 89, + 194, + 483, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, for each MLP $l$ , let $D_{\\mathrm{hid}}^{(l)}$ denote the number of hidden neurons. For each neuron $i$ in MLP $l$ , we then simply select its activation variance $\\sigma_{l,i}^2$ as the corresponding pruning score. Given a pruning ratio $p \\in (0,1)$ , we then only need to form the global set of scores", + "bbox": [ + 89, + 301, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\bigcup_ {l \\in \\mathcal {L}} \\{\\sigma_ {l, i} ^ {2} \\mid i = 1, \\ldots , D _ {\\mathrm {h i d}} ^ {(l)} \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 388, + 397, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from which we select the $p\\%$ neurons with the smallest scores for pruning.", + "bbox": [ + 89, + 430, + 483, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimality from a Mean-Replacement Perspective To retain the contribution of the pruned network, we approximate the activation of a pruned neuron $h_i$ by replacing it with its mean $\\mu_i$ . This introduces an error proportional to $|h_i - \\mu_i|$ , which, by definition of the sample variance for any distribution, has an expected value:", + "bbox": [ + 89, + 478, + 483, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ (h _ {i} - \\mathbb {E} [ h _ {i} ]) ^ {2} ] = \\mathbb {E} [ (h _ {i} - \\mu_ {i}) ^ {2} ] = \\sigma_ {i} ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 582, + 421, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, pruning the neurons with the lowest variance results in the least reconstruction error, making variance the optimal metric in this context. This is intuitively confirmed by the fact that when the variance is zero, all activations are exactly equal to the mean.", + "bbox": [ + 89, + 606, + 483, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Step 3: Mean-Shift Compensation", + "text_level": 1, + "bbox": [ + 89, + 690, + 390, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intuitive approach of removing neurons and replacing their activations with $\\mu_{i}^{\\dagger}$ at inference time already reduces the output dimension of $\\mathbf{W}_1$ , but the reconstruction of the original embedding dimension $D_{hid}$ through replacement of the pruned activations still requires carrying out the full matrix multiplication for $\\mathbf{W}_2$ .", + "bbox": [ + 89, + 712, + 482, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead, we perform an equivalent transformation by shifting these mean values directly into the bias of the final linear layer $\\mathbf{b}_2$ , allowing us to reduce both the output dimension of $\\mathbf{W}_1$ as well as the input dimension of $\\mathbf{W}_2$ , doubling the cost savings.", + "bbox": [ + 89, + 803, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start from the usual two-layer linear mapping in the MLP block:", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathbf {W} _ {2} \\mathbf {h} + \\mathbf {b} _ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 121, + 769, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{h}$ is the hidden activation after pruning. For pruned neurons, we replace $h_j$ with its mean $\\mu_{j}$ , which we collect into a vector $\\Delta_{\\mu}\\in \\mathbb{R}^{D_{\\mathrm{hid}}}$", + "bbox": [ + 511, + 143, + 906, + 188 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(\\Delta_ {\\mu}) _ {j} = \\left\\{ \\begin{array}{l l} \\mu_ {j}, & j \\in \\mathcal {P}, \\\\ 0, & j \\not \\in \\mathcal {P}, \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 199, + 792, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}$ is the set of pruned neuron indices.", + "bbox": [ + 511, + 248, + 807, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of explicitly inserting these means into $\\mathbf{h}$ and then multiplying by $\\mathbf{W}_2$ , we exploit the linearity of the mapping:", + "bbox": [ + 511, + 263, + 906, + 309 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {2} \\mathbf {h} = \\mathbf {W} _ {2} \\left(\\mathbf {h} - \\Delta_ {\\mu}\\right) + \\mathbf {W} _ {2} \\Delta_ {\\mu}.\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 319, + 833, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second term is a constant vector, so we can shift it into the bias:", + "bbox": [ + 511, + 345, + 906, + 375 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {2} ^ {\\prime} = \\mathbf {b} _ {2} + \\mathbf {W} _ {2} \\Delta_ {\\mu},\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 376, + 787, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which results in an equivalent output:", + "bbox": [ + 511, + 398, + 764, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathbf {W} _ {2} (\\mathbf {h} - \\Delta_ {\\mu}) + \\mathbf {b} _ {2} ^ {\\prime}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 424, + 906, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because for each pruned index $j$ , we have replaced the activations $h_i$ by their means $h_j = \\mu_j$ , it follows that $(\\mathbf{h} - \\Delta_{\\mu})_j = 0$ . Consequently, we can both drop the row $j$ of $\\mathbf{W}_1$ as well as the column $j$ of $\\mathbf{W}_2$ without affecting the computation, reducing the hidden dimension by $|\\mathcal{P}|$ .", + "bbox": [ + 511, + 450, + 906, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Why Use Trained Networks? Our approach uses the fact that $\\mathbf{W}_1, \\mathbf{W}_2, \\mathbf{b}_1, \\mathbf{b}_2$ are constants in trained networks. This allows the mean to be easily shifted into the bias of the next layer, which in turn allows the low-variance neurons to be removed without requiring extensive additional training. Only minor fine-tuning is sufficient to adjust the remaining weights for best performance.", + "bbox": [ + 511, + 544, + 906, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 664, + 645, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we apply our VBP method to various transformers and similar architectures, namely DeiT, Swin, and ConvNeXt in the Tiny, Small, and Base sizes, all trained on the ImageNet-1k [6] dataset.", + "bbox": [ + 511, + 688, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate these models twice. Once immediately after pruning (before fine-tuning) to measure how much accuracy is retained, and the second time after a brief fine-tuning to measure the final accuracy. Fine-tuning is performed over merely 10 epochs using knowledge-distillation from the unmodified base model. We use the AdamW optimizer with an initial learning rate of $1.5\\mathrm{e - 5}$ , decayed by a cosine-annealing scheduler, and a batch size of 32. As regularization, we include a weight decay of 0.01. All experiments are performed on NVIDIA H200 Tensor Core GPUs.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ We omit the MLP indices for clarity.", + "bbox": [ + 114, + 886, + 316, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4976", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/cb5dc2e2fafd33afe645ab902ba27f473643b888c39626de46517ccf339cf498.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model\n(Pruning Rate)MACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRet.VBP
DeiT-T (45%)1.260.94 (-25.16%)5.724.12 (-27.97%)72.0249.77 (69.13%)70.08 (97.33%)
DeiT-S (50%)4.613.21 (-30.37%)22.0514.96 (-32.15%)79.7064.44 (80.85%)78.62 (98.64%)
DeiT-B (55%)17.5811.44 (-34.93%)86.5755.40 (-36.01%)81.7357.58 (70.48%)80.67 (98.74%)
DeiT-B (20%)17.5815.35 (-12.68%)86.5775.24 (-13.09%)81.7380.87 (98.98%)81.76 (100.07%)
Swin-T (45%)4.503.23 (-28.22%)28.2921.31 (-24.67%)80.9155.12 (68.13%)79.41 (98.15%)
Swin-S (50%)8.765.63 (-32.19%)49.6135.02 (-29.41%)83.0467.01 (80.70%)81.86 (98.58%)
Swin-B (55%)15.4610.22 (-33.89%)87.7756.29 (-35.87%)84.7162.61 (73.91%)83.61 (98.70%)
Swin-B (20%)15.4613.58 (-12.16%)87.7776.42 (-12.93%)84.7183.90 (99.04%)84.67 (99.95%)
", + "bbox": [ + 91, + 88, + 906, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 89, + 373, + 230, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As is common with various pruning techniques, the pruning rate is adapted for the size of the models. Larger models typically exhibit higher redundancy for similar tasks, and therefore allow for more pruning while maintaining accuracy. We report our results for DeiT and Swin in Tab. 1 and Tab. 2. Note that the pruning rate is applied to the MLPs only and does not indicate the total reduction in model size.", + "bbox": [ + 89, + 397, + 482, + 503 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the base-sized models, we prune all MLPs globally using a pruning rate of $55\\%$ keeping $99\\%$ of the original performance after fine-tuning, reducing the total model size of DeiT-Base, Swin-Base both by $36\\%$ . This corresponds to $35\\%$ , and $34\\%$ fewer MACs respectively and provides a speed-up of $1.44 \\times$ and $1.30 \\times$ times.", + "- The small-sized models have notably fewer parameters than the base models and are thus pruned $50\\%$ again reaching $99\\%$ of the original performance. This leads to a DeiT-Small size reduction of $32\\%$ and Swin-Small reduction of $29\\%$ with $30\\%$ and $32\\%$ MACs savings respectively for a total speed-up factor of $1.34 \\times$ and $1.29 \\times$ .", + "- Finally, for the DeiT- and Swin-Tiny models, we apply a lower pruning rate of $45\\%$ , achieving $97\\%$ and $98\\%$ final performance with $28\\%$ and $25\\%$ reductions in model size. This translates to $25\\%$ and $28\\%$ fewer MACs, along with speed-ups of $1.17 \\times$ and $1.20 \\times$ respectively.", + "- For the Base-Models, we demonstrate that when globally pruning the MLPs by $20\\%$ the accuracy retention directly after the one-shot structured pruning reaches $99\\%$ of the unpruned performance without any fine-tuning. This allows for an off-the-shelf $13\\%$ size and MACs reduction and a speed-up of $1.11 \\times$ for DeiT-Base and a similar reductions for Swin-Base" + ], + "bbox": [ + 89, + 505, + 482, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All further pruning rates ranging from $5\\%$ to $50\\%$ for all DeiT Models can be found in Appendix B.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3ce4f293e7f4237657a24787d695087662a3350a3a6d89c59ecb7edead085976.jpg", + "table_caption": [ + "Table 1. Results comparing the full baseline model with our VBP model using four metrics: MACs: computational operations, measured in billions of operations; and Parameters: the total model size in millions of parameters; Accuracy Retention (Ret.): retained accuracy after pruning, before fine-tuning; and Final Accuracy: accuracy after fine-tuning. Our method achieves competitive accuracy with significant reductions in MACs and parameters after fine-tuning, and even right after pruning, when reducing the pruning rate to $20\\%$ ." + ], + "table_footnote": [], + "table_body": "
Model (Pruning Rate)Tiny (45%)Small (50%)Base (55%)Base (20%)
DeiT1.171.341.441.11
Swin1.201.271.301.10
", + "bbox": [ + 514, + 371, + 906, + 449 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/83b6014170b97c8fc2e1be0c96df33855910f4a40966df8c54750da04be32383.jpg", + "table_caption": [ + "Table 2. Speed-ups of the pruned models relative to the baseline for different sizes and pruning rates." + ], + "table_footnote": [], + "table_body": "
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
DeiT-T1.265.72-72.02
Magnitude0.913.943.5668.49
SNIP0.913.9424.8969.10
VBP (ours)0.913.9439.5870.61
DeiT-S4.6122.05-79.70
Magnitude3.2114.964.0576.55
SNIP3.2114.9652.3977.27
VBP (ours)3.2114.9664.4478.62
DeiT-B17.5886.57-81.73
Magnitude12.058.240.3778.88
SNIP12.058.2453.2480.40
VBP (ours)12.058.2466.4080.99
", + "bbox": [ + 514, + 522, + 906, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Comparison of different one-shot pruning scores in structured pruning to our VBP approach at fixed $50\\%$ pruning rates. For constant MACs and parameters, our approach retains more accuracy (Ret.) and reaches a higher final accuracy compared to other pruning approaches.", + "bbox": [ + 511, + 779, + 906, + 851 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4977", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5f117d4cf1e81b805d40108f11653dad6db933d136c9f29313d70ad3d0a983c7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMACs (G)Param. (M)Speed-UpTop-1 Acc. (%)
Ret.Final
DeiT-B17.5886.57--81.73
ToMe-246.0286.572.1535.8575.74
ToMe-14 & VBP6.0858.242.0560.5380.09
ToMe-207.1486.571.8466.3178.66
ToMe-12 & VBP6.9058.241.8362.0980.29
ToMe-187.8886.571.7573.7979.97
ToMe-10 & VBP7.7358.241.7063.7280.45
", + "bbox": [ + 91, + 87, + 906, + 258 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5b1352b9605cb17c1777c1d68e1b65f41bc8f6e3f65bb49ea8da403de70f2c92.jpg", + "table_caption": [ + "Table 4. Comparison of ToMe for different token reduction rates of 24, 20, and 18, with the hybrid combination of ToMe using token reduction rates of 14, 12, and 10, and VBP using $50\\%$ pruning rates on top. Our hybrid method achieves significantly higher final accuracies throughout while also providing reductions in parameters, for similar MACs and speed-ups. The hybrid model thereby reaches speed-ups of up to $2.05 \\times$ while maintaining competitive performance." + ], + "table_footnote": [], + "table_body": "
ModelEpochsParam. (M)Top-1 Acc. (%)
Ret.Final
NViT (50 Ep.)5056.3781.1382.18
NViT (1 Ep.)156.8569.1081.92
VBP (ours)156.1872.3782.32
", + "bbox": [ + 91, + 349, + 483, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Comparison with NViT using two pruning durations of 50 epochs (as originally proposed) and 1 epoch (similar to our VBP approach), both followed by 10 epochs of fine-tuning. Our method achieves higher accuracy immediately after pruning, before fine-tuning (Ret.) compared to NViT with the same pruning duration, and reaches a higher final accuracy than the 50-epoch NViT model after fine-tuning for 10 epochs.", + "bbox": [ + 89, + 464, + 482, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparison", + "text_level": 1, + "bbox": [ + 89, + 604, + 326, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate direct comparisons of all models with different sizes across all experiments, we fix the pruning rate at $50\\%$ for all further experiments.", + "bbox": [ + 89, + 635, + 482, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of our selection criterion, by benchmarking against two other common and well-established one-shot-pruning baselines, namely Magnitude and SNIP, adapted for structured pruning. For a fixed pruning rate of $50\\%$ using all selection criteria, this produces identical parameter counts of 3.94, 14.96, and 58.24 million parameters, as well as fixed 0.91, 3.21, and 12.00 GMACs for the three model sizes. Our results indicate that VBP provides higher accuracy retention compared to the baselines, improving DeiT-Base by $13.16\\%$ , DeiT-Small by $12.05\\%$ , and DeiT-Tiny by $14.69\\%$ percentage points compared to SNIP. This difference in accuracy retention yields a final fine-tuned performance that is $0.59\\%$ to $1.51\\%$ percentage points higher after fine-tuning.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. SoTA Pruning Methods", + "text_level": 1, + "bbox": [ + 511, + 351, + 732, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further validate our approach, we compare against the SoTA structured transformer pruning method NViT (CVPR'23). Specifically, we apply our method to the NViT Base model (DeiT-Base distilled from RegNetY-160) and run NViT with a latency target of $70\\%$ , pruning once for 50 epochs (as per the published results) and once for one epoch (similar to our setup), with both scenarios followed by ten epochs of fine-tuning using the NViT settings. In a similar pruning setting, VBP achieves a $0.4\\%$ percentage points higher final accuracy, for similar number of parameters.", + "bbox": [ + 511, + 377, + 906, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we demonstrate the orthogonality of VBP to token-pruning methods by applying our method on top of the SoTA token-merging approach, ToMe (ICLR'23). We observe that the accuracy retention drops relative to the baseline remains consistent for both the VBP-pruned base DeiT model in Tab. 1 and the VBP- & ToMe-pruned model. This relative consistency also persists throughout fine-tuning, indicating that our method operates orthogonally to ToMe in terms of performance.", + "bbox": [ + 511, + 531, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding the MAC savings it is important to note that ToMe progressively reduces the throughput of tokens processed across the layers. Later layers therefore process less tokens, making pruning less effective. The relative MAC savings from our hybrid method therefore decrease as the network depth increases. Nonetheless, VBP additionally reduces the total parameter count, which cannot be achieved by simply increasing the ToMe token reduction rate.", + "bbox": [ + 511, + 670, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In fact, increasing the reduction rate to get an approximate $2 \\times$ speed-up yields similar MAC savings but reduces both the accuracy retention as well as the final accuracy significantly. Instead applying VBP on top of ToMe with a smaller reduction rate, allows more than $2 \\times$ speed-ups while still maintaining $98\\%$ of the original accuracy with additional parameter savings.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4978", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24725868ef51692c39d48ca35bea96b2c0aff464b75f33b0c8582b3bfdbd94ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Variance- BasedMean- ShiftTop-1 Acc. (%)
Ret.Final
X✓55.1980.23
✓X26.0480.62
✓✓66.4080.99
", + "bbox": [ + 91, + 88, + 482, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d08c1dddc8da297f901d0c2fee6bcc231b27ac79b4aae6ad3bdff18050085640.jpg", + "image_caption": [ + "Figure 2. Accuracy retention before fine-tuning and final accuracy after 10 epochs of fine-tuning for different pruning rates applied to DeiT-Base." + ], + "image_footnote": [], + "bbox": [ + 91, + 282, + 482, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15ff84dd38bb00e4ba02a35cb4ce4270f115ceacc4b68ea02cabcc7831246c15.jpg", + "image_caption": [ + "Figure 3. Activation variance for all neurons in the hidden MLP layers throughout the network and marked cumulative variances." + ], + "image_footnote": [], + "bbox": [ + 89, + 549, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 813, + 187, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the contribution of each individual component of our approach, we perform an ablation study in which we systematically modify the pruning criterion and the Mean-Shift Compensation. We report our results in Tab. 6.", + "bbox": [ + 89, + 840, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c062d431d17c732ddc94c20e789db90455d5873ae82fb96ce2e009cfb98260f2.jpg", + "table_caption": [ + "Table 6. Ablation study results showing the impact of Variance-Based Pruning, and Mean-Shift Compensation on the retained accuracy after pruning (Ret.), as well as the final: accuracy after fine-tuning. Both steps together yield the highest accuracy of $80.99\\%$ ." + ], + "table_footnote": [], + "table_body": "
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
Pre-Act. (eq. [23])12.058.240.4377.92
Post-Act. (ours)12.058.2466.4080.99
", + "bbox": [ + 514, + 88, + 905, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Comparison of different locations for the activation statistics computation in the network: before (Pre-Act.) as well as after the activation function (Post-Act.). For the same MACs and parameters, using pruning rates of $50\\%$ , gathering statistics post-activation significantly outperforms pre-activation in both accuracy retention (Ret.) as well as the final accuracy.", + "bbox": [ + 511, + 181, + 905, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first remove the variance-based pruning selection while still shifting the mean activation to the next bias. Removing the Variance-Based Pruning alone reduces the final accuracy by $0.76\\%$ percentage points.", + "bbox": [ + 511, + 299, + 903, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also remove the Mean-Shift Compensation while pruning using the variance-based criterion. This improves the final accuracy by $0.39\\%$ compared to using only the Mean-Shift Compensation.", + "bbox": [ + 511, + 363, + 905, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, the best results are achieved when both Variance-Based Pruning and Mean-Shift Compensation are used together. As discussed in Sec. 3.2, replacing the removed activations with their means is optimal when pruning based on variance. When both components are combined, the immediate accuracy retention increases by $11.21\\%$ percentage points, ultimately improving the fine-tuned accuracy by an additional $0.37\\%$ . This highlights the significance of both components.", + "bbox": [ + 511, + 426, + 906, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Sensitivity of Variance Thresholding", + "text_level": 1, + "bbox": [ + 511, + 584, + 830, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further examine the stability of the activation variance criterion by applying different pruning rates to the DeiT-Base model, and plotting the accuracy retention and final accuracy in Fig. 2. We observe that for smaller pruning rates (up to around $25\\%$ ), the accuracy retention is high enough to allow for off-the-shelf deployment without any fine-tuning.", + "bbox": [ + 511, + 609, + 905, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond these rates, the accuracy begins to degrade more noticeably, reflecting the non-uniform distribution of neuron variances, as seen in Fig. 3. Notably, to account for $10\\%$ of the cumulative variance across all neurons, nearly $60\\%$ of the lowest-variance neurons are needed. This is a disproportionately large share compared to the next $15\\%$ of neurons that also account for $10\\%$ of the total variance in a layer. Consequently, more aggressive pruning leads to an increasingly faster reduction in the total variance and thus expressiveness of the network. Nevertheless, significant performance issues arise only at very high pruning levels, validating the feasibility of our method over a broad range of pruning rates (see Appendix B).", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4979", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c3862ebc3699c761e63de9b172cf23ed0038d2faac53ffd8c6e002dea301248.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRetentionVBP
ConvNeXt-T4.472.96 (-33.8%)28.5912.61 (-55.9%)82.9016.80 (20.3%)81.30 (98.1%)
ConvNeXt-S8.715.11 (-41.3%)50.2223.50 (-53.2%)84.5730.92 (36.6%)82.82 (97.9%)
ConvNeXt-B15.388.91 (-42.1%)88.5941.32 (-53.4%)85.5157.10 (66.8%)83.40 (97.6%)
", + "bbox": [ + 91, + 88, + 906, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6e4f548a3fc4e6c87f030c1bb019aa970be5caa2cbc230a861b30835ae118a59.jpg", + "table_caption": [ + "Table 8. Results comparing the full ConvNeXt baseline models with the VBP models using a pruning rate of $50\\%$ . The MACs are reduced by up to $42\\%$ and parameters by over $50\\%$ while reaching competitive final accuracies of $98\\%$ original performance." + ], + "table_footnote": [], + "table_body": "
ModelTinySmallBase
ConvNeXt1.281.421.49
", + "bbox": [ + 91, + 250, + 483, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Speed-ups of the pruned models relative to the baseline for different-sized ConvNeXt models.", + "bbox": [ + 89, + 306, + 482, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Application Pre- vs. Post-Activation", + "text_level": 1, + "bbox": [ + 89, + 361, + 403, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While Sec. 3.2 shows that our mean-replacement strategy can be grounded in theory for any distribution, the question remains whether the variance should be measured before or after the activation function. The Central Limit Theorem confirms that pre-activation sums (composed of independent weight contributions) tend to approximate a normal distribution. However, once the nonlinearity $\\sigma$ is applied, this distribution changes. Consequently, the pre-activation variance does not necessarily correlate with the neuron importance for retaining accuracy in an already trained model.", + "bbox": [ + 88, + 383, + 482, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For instance, highly varying negative activations are compressed into range of about $(-0.2, 0)$ by GeLU, reducing their variance. As a result, such a neuron would be less likely to be pruned in a pre-activation setting and more likely to be pruned post-activation (see Appendix E).", + "bbox": [ + 88, + 535, + 482, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate this experimentally, we tested an alternative design in which pruning decisions are made based on the pre-activation variance in Tab. 7. While this approach has been successfully applied when pruning during training [23], we find that in a trained network setting, it significantly degrades performance. Therefore, measuring variance post-activation better captures neuron importance in already trained models, and helps maintain model performance despite structural modifications.", + "bbox": [ + 88, + 611, + 482, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Other Transformer-Like Architectures", + "text_level": 1, + "bbox": [ + 89, + 757, + 423, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our method is broadly applicable to any model containing MLP layers, network architectures that have more and larger MLPs benefit more from VBP. We therefore apply our pruning on ConvNeXt as a transformer-like architecture, that relies heavily on MLP layers, to further evaluate the generalizability of our method.", + "bbox": [ + 88, + 779, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We summarize our performance comparisons for ConvNeXt in Tab. 8. We achieve significant parameter reduc", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions of over $50\\%$ throughout all three sizes and reduced MACs by $34\\%$ for ConvNeXt-Tiny and over $40\\%$ for the larger sizes, while maintaining $98\\%$ of the original accuracy after 10 epochs of fine-tuning across the board. The corresponding speed-ups are listed in Tab. 9, and reach up to $1.49\\times$ for ConvNeXt-Base.", + "bbox": [ + 511, + 252, + 906, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Notably, we observe significantly less accuracy retention throughout the smaller sized models compared to the true transformers. We attribute this to the fact that the ConvNeXt architecture is built upon convolutions, which have significantly higher costs compared to their parameters. Consequently, since the MLPs take over a higher overall percentage of the total capacity, removing a similar fraction of neurons causes a relatively higher drop in accuracy retention. By contrast, transformers typically have less parameters in their MLP relatively speaking, thereby allowing for higher pruning rates while maintaining capacity.", + "bbox": [ + 511, + 347, + 908, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 540, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce VBP, a one-shot pruning method, designed to remove neurons based on their statistical importance in a single pruning operation while minimizing accuracy loss by integrating their mean activations in a Mean-Shift Compensation step.", + "bbox": [ + 511, + 569, + 906, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments demonstrate that VBP retains model performance significantly better than existing structured pruning methods. By taking advantage of trained networks, our approach provides a favourable starting point for the subsequent finetuning, eliminating the need for extensive retraining and thereby making it practical for low-cost deployment.", + "bbox": [ + 511, + 648, + 906, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we show that VBP operates orthogonally to SoTA token pruning methods, allowing it to be seamlessly combined with approaches like ToMe to achieve even greater computational savings.", + "bbox": [ + 511, + 760, + 906, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We believe that the simplicity and efficiency of our approach contributes to the democratization of deep learning by enabling a wider reuse of trained networks and hope it inspires further research in the CV community toward efficient model compression techniques.", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4980", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sajid Anwar, Kyuyeon Hwang, and Wonyong Sung. Structured pruning of deep convolutional neural networks. J. Emerg. Technol. Comput. Syst., 13(3), 2017. 1, 2, 3", + "[2] Uranik Berisha, Jens Mehnert, and Alexandru Condurache. Efficient data driven mixture-of-expert extraction from trained networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3", + "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 3", + "[4] Rinor Cakaj, Jens Mehnert, and Bin Yang. Cnn mixture-of-depths. In Computer Vision ACCV 2024: 17th Asian Conference on Computer Vision, Hanoi, Vietnam, December 812, 2024, Proceedings, Part VII, page 148166, Berlin, Heidelberg, 2024. Springer-Verlag. 3", + "[5] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19974-19988, 2021. 3", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4, 3", + "[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12114-12124. IEEE, 2022. 2", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 1, 2", + "[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6804-6815. IEEE, 2021. 2", + "[10] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Herve Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC," + ], + "bbox": [ + 93, + 114, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Canada, October 10-17, 2021, pages 12239-12249. IEEE, 2021. 2", + "[11] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4", + "[12] Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pages 598-605. Morgan Kaufmann, 1989. 1, 3", + "[13] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip H. S. Torr. Snip: single-shot network pruning based on connection sensitivity. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3", + "[14] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 1, 2", + "[15] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 1, 2", + "[16] Lukas Meiner, Jens Mehnert, and Alexandru Condurache. Data-free dynamic compression of cnns for tractable efficiency. In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 2: VISAPP, pages 196-208. INSTICC, SciTePress, 2025. 3", + "[17] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12299-12308. IEEE, 2022. 3", + "[18] Dmitry Molchanov, Arsenii Ashukha, and Dmitry P. Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2498-2507. PMLR, 2017. 3", + "[19] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 13937-13949, 2021. 1, 3", + "[20] David Raposo, Samuel Ritter, Blake A. Richards, Timothy P. Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models. CoRR, abs/2404.02258, 2024.3", + "[21] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4981", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 10347-10357. PMLR, 2021. 1, 2", + "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2, 3", + "[23] Berry Weinstein and Yonatan Belinkov. Variance pruning: Pruning language models via temporal neuron variance, 2025. Unpublished manuscript, OpenReview. 3, 7, 8", + "[24] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 2", + "[25] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Freezenet: Full performance by reduced storage costs. In Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, pages 685-701. Springer, 2020. 3", + "[26] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Interspace pruning: Using adaptive filter representations to improve training of sparse cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12527-12537, 2022. 3", + "[27] Huanrui Yang, Hongxu Yin, Maying Shen, Pavlo Molchanov, Hai Li, and Jan Kautz. Global vision transformer pruning with hessian-aware saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18547-18557. IEEE, 2023. 1, 3", + "[28] Fang Yu, Kun Huang, Meng Wang, Yuan Cheng, Wei Chu, and Li Cui. Width & depth pruning for vision transformers. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 3143-3151. AAAI Press, 2022.", + "[29] Lu Yu and Wei Xiang. X-pruner: explainable pruning for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24355-24363. IEEE, 2023. 3", + "[30] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 2, 3", + "[31] Zhengyan Zhang, Zhiyuan Zeng, Yankai Lin, Chaojun Xiao, Xiaozhi Wang, Xu Han, Zhiyuan Liu, Ruobing Xie, Maosong Sun, and Jie Zhou. Emergent modularity in pretrained transformers. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4066-4083, Toronto, Canada, 2023. Association for Computational Linguistics. 3" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[32] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning, 2021. 3", + "bbox": [ + 514, + 90, + 903, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4982", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_model.json b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..63f19244cf66860b2065d3540c706dd9215a31ba --- /dev/null +++ b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.895, + 0.152 + ], + "angle": 0, + "content": "Variance-Based Pruning for Accelerating and Compressing Trained Networks" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.18, + 0.766, + 0.199 + ], + "angle": 0, + "content": "Uranik Berisha12, Jens Mehnert1 and Alexandru Paul Condurache12" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.201, + 0.735, + 0.216 + ], + "angle": 0, + "content": "1Automated Driving Research, Robert Bosch GmbH, 70469 Stuttgart, Germany" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.219, + 0.735, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)Institute for Signal Processing, University of Lübeck, 23562 Lübeck, Germany" + }, + { + "type": "list", + "bbox": [ + 0.261, + 0.201, + 0.735, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.237, + 0.836, + 0.251 + ], + "angle": 0, + "content": "{Uranik.Berisha,JensEricMarkus.Mehnert,AlexandruPaul.Condurache}@de.bosch.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.483, + 0.499 + ], + "angle": 0, + "content": "Increasingly expensive training of ever larger models such as Vision Transformers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.681 + ], + "angle": 0, + "content": "To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over \\(70\\%\\) of its original performance and requires only 10 epochs of fine-tuning to regain \\(99\\%\\) of the original accuracy while simultaneously reducing MACs by \\(35\\%\\) and model size by \\(36\\%\\), thus speeding up the model by \\(1.44 \\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In both Computer Vision (CV) and Natural Language Processing (NLP), Transformers have become the dominant architecture [8, 21]. In CV, Vision Transformers (ViTs) [8] and their derivatives, such as Data-efficient image Transformers (DeiTs) [21], Swin-Transformers (Swins) [14] and ConvNeXts [15], achieve remarkable results but at significant costs in three main areas: training, storage, and inference. These large models require long training times, often spanning hundreds of epochs to achieve their performance, extensive memory to store their parameters, and high computational resources for inference causing high latency [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.47 + ], + "angle": 0, + "content": "Fixing all three issues simultaneously remains especially challenging. For example, using the available collection of trained networks provided by the research community is a straightforward way to solve the first problem of costly training. However, this does not address the storage and inference costs, which can still make the use of these trained networks impractical for deployment on resource-constrained hardware [1]. To address these demands, various model pruning techniques have been proposed. These range from removing individual weights within layers in unstructured pruning [12] to removing entire neurons, layers, or blocks in structured pruning [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.905, + 0.563 + ], + "angle": 0, + "content": "While unstructured pruning can often retain a high degree of accuracy in trained networks, it is difficult to translate these theoretical gains into real-world speed-ups [1]. Since modern hardware is optimized for dense matrix multiplications, they cannot exploit the resulting sparsity effectively and therefore cannot solve the inference costs [1, 12]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.702 + ], + "angle": 0, + "content": "In contrast, structured pruning leads to straightforward reductions in both inference costs and memory, but again requires extensive retraining to recover the lost accuracy resulting from these major structural modifications. For instance, Global Vision Transformer Pruning with Hessian-Aware Saliency (NViT) [27] requires 300 epochs of fine-tuning despite using trained networks, highlighting how addressing memory and inference speeds alone can still demand large additional training overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "On the other hand, dynamic pruning approaches, such as token pruning, avoid permanently modifying the model architecture and instead dynamically reduce the number of features processed by the network [3, 19]. This minimizes the accuracy drop after pruning, indicating that the model preserved most important features and representation, thus providing a better starting point for subsequent fine-tuning. Token Merging (ToMe) [3] embraces this idea by merging tokens to reuse the pruned information, thus reducing the accuracy drop even further and allowing off-the-shelf deployment without any fine-tuning. While this solves the inference costs as well as the training costs, since the methods do not modify the model structure itself, the memory foot" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4973" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.093, + 0.329, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.336, + 0.356, + 0.361 + ], + "angle": 0, + "content": "(a) We compute neuron-wise mean and variance of activations at the hidden layer of each MLP." + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.092, + 0.61, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.336, + 0.63, + 0.361 + ], + "angle": 0, + "content": "(b) We prune neurons with the smallest variances and replace their activation with the mean." + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.096, + 0.887, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.336, + 0.906, + 0.361 + ], + "angle": 0, + "content": "(c) We shift the mean activation of the removed neurons into the bias of the next layer." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.373, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Figure 1. Schematic visualization of the steps of Variance-Based Pruning. (a) Activation Statistics Computation, (b) Variance-Based Pruning, (c) Mean-Shift Compensation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.428, + 0.483, + 0.489 + ], + "angle": 0, + "content": "print remains unchanged [3]. This again is a limitation for hardware-constrained settings [1]. Fully resolving all three problems of training costs, storage, and inference speed simultaneously remains a challenge." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.491, + 0.485, + 0.765 + ], + "angle": 0, + "content": "In this work, we fill this gap by introducing Variance-Based Pruning (VBP), a simple and structured pruning approach that provides speed-ups and memory savings with minimal fine-tuning. In order to keep the method simple and allow deployment across a wide variety of architectures, we prune only the heavy Multi-Layer Perceptron (MLP) layers within the transformer blocks. In the first step, our approach leverages activation statistics gathered in the hidden layers of the MLP using Welford's algorithm [24] (see Fig. 1a). In the second step, these statistics are used to identify and remove the least impactful hidden neurons (see Fig. 1b). In the last step, which we refer to as Mean-Shift Compensation, we use the mean activation values of pruned neurons to mitigate accuracy drop by redistributing these mean contributions into the bias of the output layer (see Fig. 1c). This retains high levels of model performance and requires only minimal fine-tuning to then regain most of the original accuracy." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.779, + 0.416, + 0.792 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We introduce VBP, a simple low-cost structured pruning method broadly applicable across various architectures to significantly reduce computational costs while maintaining performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We apply our method to multiple architectures, including ViTs, Swin-Transformers, and ConvNeXts, demonstrating its generalizability." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.428, + 0.907, + 0.549 + ], + "angle": 0, + "content": "- We compare VBP to similarly straightforward neuron-importance measures adapted for structured pruning, as well as to state-of-the-art (SoTA) pruning approaches. We benchmark against NViT pruning, highlighting the benefits of our accuracy retention in constrained fine-tuning settings. We also apply VBP on top of ToMe demonstrating orthogonality and achieving \\(2 \\times\\) speed-ups using our hybrid approach." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.549, + 0.907, + 0.609 + ], + "angle": 0, + "content": "- We perform an extensive ablation study, demonstrating the efficacy of each component of our method and provide a sensitivity analysis for the variance as a pruning criterion." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.428, + 0.907, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.635, + 0.702, + 0.65 + ], + "angle": 0, + "content": "2. RELATED WORK" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.757, + 0.679 + ], + "angle": 0, + "content": "2.1. Transformer Architectures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Initially introduced in NLP by [22], Transformers have largely replaced Convolutional Neural Network (CNN), since their adaption to CV by [8] with the ViT. ViTs processes images as a sequence of non-overlapping patches, and have since been further improved. DeiT [21] refined the training procedure to decrease the training costs and reduce the data reliance. More recent architectures [7, 9, 10], such as Swin [14] and the transformer-inspired ConvNeXt [15], which use hierarchical designs, have built upon this architecture to improve efficiency and scalability. However, despite these advances, these large models remain computationally demanding, in part due to their dense MLP blocks, which significantly impacts inference speed and memory requirements [22, 30]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4974" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.282, + 0.107 + ], + "angle": 0, + "content": "2.2. Model Compression" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.49 + ], + "angle": 0, + "content": "To reduce the costs of large models, model compression efforts focus on scaling down different components of a network. Traditionally pruning aims to identify and remove redundant structures within the network [12, 26]. For example, magnitude pruning eliminates weights with the smallest absolute values, while SNIP and related methods [13, 25] identify crucial connections using gradient-based sensitivities. Structured pruning approaches, remove entire neurons, layers or filters producing smaller matrices that allow direct computational savings in current hardware [1, 5, 27-29, 32]. Most closely related to our method are works exploring variance-based criteria for structured pruning, such as Molchanov et al. [18], which performs dropout at the neuron level and can introduce sparsity in neurons with high dropout rates. Weinstein et al. [23] use these statistics in the initial training phase to remove neurons during training. These approaches operate on models before training, improving training time but still requiring the models to be trained from scratch. Contrary to that, NViT [27] utilizes trained ViTs. They apply iterative structured dimensionality reductions across the entire ViT architecture and significantly reduce costs. However, these structured modifications still have a substantial impact on performance, requiring high pruning iterations of 50 epochs and retraining costs of 300 epochs to regain accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.658 + ], + "angle": 0, + "content": "An alternative to structurally modifying the network is dynamic pruning, which adaptively selects and processes only a subset of features or combines features during inference to reduce costs. For CNNs, these have shown promising speed-accuracy trade-offs [4, 16]. Analogously for transformer-like models, token pruning reduces the number of tokens to be processed [3, 17, 19, 20]. For instance, Bolya et al. [3] merges similar tokens rather than pruning them, thereby conserving most of the information and retaining high accuracy before the fine-tuning. However, by design these methods cannot reduce the memory footprint." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.241, + 0.684 + ], + "angle": 0, + "content": "2.3. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our work builds upon these ideas by combining (i) the advantages of structured pruning by removing low-variance activation neurons with (ii) the advantages of dynamic pruning by incooperating their mean contribution into the bias of the last MLP layer. Our approach is based on the observation that neurons within MLP layers exhibit varying levels of contribution to the overall representation, as established in works on emerging modularity in trained Transformers for both NLP and CV [2, 31]. By applying our variance-based pruning approach to the widespread MLP layers, we provide a structured method applicable across a broad range of models. Our method is particularly beneficial for transformer-like architectures, where MLP layers heavily contribute to the computational overhead [22, 30]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.176 + ], + "angle": 0, + "content": "In this section, we detail the components involved in our proposed VBP approach and provide a mathematical derivation, justifying our pruning criterion and compensation strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.177, + 0.792, + 0.191 + ], + "angle": 0, + "content": "Our method consists of three major steps:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.191, + 0.905, + 0.221 + ], + "angle": 0, + "content": "1. Activation Statistics Computation The neuron activation statistics are computed using Welford's algorithm." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.221, + 0.905, + 0.25 + ], + "angle": 0, + "content": "2. Variance-Based Pruning The neurons to be pruned are selected based on the lowest activation variance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.251, + 0.905, + 0.281 + ], + "angle": 0, + "content": "3. Mean-Shift Compensation The mean activations of the selected neurons are added back into the output bias." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.191, + 0.905, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.29, + 0.871, + 0.307 + ], + "angle": 0, + "content": "3.1. Step 1: Activation Statistics Computation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.403 + ], + "angle": 0, + "content": "As we aim for a simple and widely adoptable method, we prune only the MLPs. This requires reducing the hidden layers of the network while keeping the input and output layers intact to ensure seamless integration with the rest of the model structure by maintaining dimensional consistency in the input and output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.478 + ], + "angle": 0, + "content": "W.l.o.g. we therefore consider a MLP with a single hidden layer of dimension \\(D_{\\mathrm{hid}}\\), that maps an input vector \\(\\mathbf{x} \\in \\mathbb{R}^{D_{\\mathrm{in}}}\\) to an output vector \\(\\mathbf{y} \\in \\mathbb{R}^{D_{\\mathrm{out}}}\\). Using a pointwise nonlinear activation function \\(\\sigma(\\cdot)\\), this MLP can then be described as a function" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.489, + 0.907, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\sigma \\left(\\mathbf {W} _ {1} \\mathbf {x} + \\mathbf {b} _ {1}\\right), \\quad \\mathbf {y} = \\mathbf {W} _ {2} \\mathbf {h} + \\mathbf {b} _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.906, + 0.591 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_1 \\in \\mathbb{R}^{D_{\\mathrm{hid}} \\times D_{\\mathrm{in}}}\\), \\(\\mathbf{b}_1 \\in \\mathbb{R}^{D_{\\mathrm{hid}}}\\), \\(\\mathbf{W}_2 \\in \\mathbb{R}^{D_{\\mathrm{out}} \\times D_{\\mathrm{hid}}}\\), \\(\\mathbf{b}_2 \\in \\mathbb{R}^{D_{\\mathrm{out}}}\\). Since we deal with trained networks, these parameters \\(\\mathbf{W}_1, \\mathbf{b}_1, \\mathbf{W}_2, \\mathbf{b}_2\\) remain fixed during pruning, which will become important in the Mean-Shift Compensation in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.592, + 0.906, + 0.622 + ], + "angle": 0, + "content": "For the hidden layer, we now want to calculate the perneuron mean and variance vectors" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.629, + 0.897, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} = \\left(\\mu_ {1}, \\dots , \\mu_ {D _ {\\mathrm {h i d}}}\\right) ^ {\\top} \\quad \\text {a n d} \\quad \\boldsymbol {\\sigma} ^ {2} = \\left(\\sigma_ {1} ^ {2}, \\dots , \\sigma_ {D _ {\\mathrm {h i d}}} ^ {2}\\right) ^ {\\top}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.786 + ], + "angle": 0, + "content": "To compute these statistics, we capture the activations \\(\\mathbf{h}\\) for every batch and use Welford's algorithm to compute a running mean \\(\\pmb{\\mu}^{(j)}\\) and running second moment \\(\\mathbf{m}_2^{(j)}\\) for an efficient online calculation of streaming data*. Given a total of \\(N\\) data samples, let the \\(j\\)-th sample \\(\\mathbf{h}^{(j)} \\in \\mathbb{R}^{D_{\\mathrm{hid}}}\\) be the activation vector after the nonlinearity. Welford's algorithm ensures numerically stable calculations by updating \\(\\pmb{\\mu}^{(j)}\\) and \\(\\mathbf{m}_2^{(j)}\\) after observing the \\(j\\)-th sample as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.795, + 0.907, + 0.828 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {(j)} = \\frac {j - 1}{j} \\boldsymbol {\\mu} ^ {(j - 1)} + \\frac {1}{j} \\mathbf {h} ^ {(j)} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.836, + 0.905, + 0.857 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {2} ^ {(j)} = \\mathbf {m} _ {2} ^ {(j - 1)} + \\left(\\mathbf {h} ^ {(j)} - \\boldsymbol {\\mu} ^ {(j - 1)}\\right) \\odot \\left(\\mathbf {h} ^ {(j)} - \\boldsymbol {\\mu} ^ {(j)}\\right) \\tag {3}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.864, + 0.907, + 0.901 + ], + "angle": 0, + "content": "*Our method does not require a minimum number of data samples for calculating activation statistics, but using an efficient online computation allows the use of as many data samples for pruning as available." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4975" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.123 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes elementwise multiplication. Once all \\(N\\) samples are processed, the variance is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.233, + 0.131, + 0.483, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\sigma} ^ {2} = \\frac {\\mathbf {m} _ {2} ^ {(N)}}{N - 1}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.174, + 0.377, + 0.191 + ], + "angle": 0, + "content": "3.2. Step 2: Variance-Based Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.195, + 0.484, + 0.301 + ], + "angle": 0, + "content": "Using this variance vector \\(\\sigma^2 \\in \\mathbb{R}^{D_{\\mathrm{hid}}}\\), we rank neurons by their variance values, \\(\\sigma_i^2\\). Intuitively, if an activation rarely deviates from its mean \\(\\mu_i\\), the corresponding neuron contributes less to the expressiveness of the network. The pruning decision is formed across all hidden neurons in the network, to capture the neuron with the least variance across all layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.38 + ], + "angle": 0, + "content": "Formally, for each MLP \\(l\\), let \\(D_{\\mathrm{hid}}^{(l)}\\) denote the number of hidden neurons. For each neuron \\(i\\) in MLP \\(l\\), we then simply select its activation variance \\(\\sigma_{l,i}^2\\) as the corresponding pruning score. Given a pruning ratio \\(p \\in (0,1)\\), we then only need to form the global set of scores" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.389, + 0.398, + 0.422 + ], + "angle": 0, + "content": "\\[\nS = \\bigcup_ {l \\in \\mathcal {L}} \\{\\sigma_ {l, i} ^ {2} \\mid i = 1, \\ldots , D _ {\\mathrm {h i d}} ^ {(l)} \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.462 + ], + "angle": 0, + "content": "from which we select the \\( p\\% \\) neurons with the smallest scores for pruning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.571 + ], + "angle": 0, + "content": "Optimality from a Mean-Replacement Perspective To retain the contribution of the pruned network, we approximate the activation of a pruned neuron \\( h_i \\) by replacing it with its mean \\( \\mu_i \\). This introduces an error proportional to \\( |h_i - \\mu_i| \\), which, by definition of the sample variance for any distribution, has an expected value:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.583, + 0.422, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ (h _ {i} - \\mathbb {E} [ h _ {i} ]) ^ {2} ] = \\mathbb {E} [ (h _ {i} - \\mu_ {i}) ^ {2} ] = \\sigma_ {i} ^ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Thus, pruning the neurons with the lowest variance results in the least reconstruction error, making variance the optimal metric in this context. This is intuitively confirmed by the fact that when the variance is zero, all activations are exactly equal to the mean." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.691, + 0.391, + 0.708 + ], + "angle": 0, + "content": "3.3. Step 3: Mean-Shift Compensation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.804 + ], + "angle": 0, + "content": "The intuitive approach of removing neurons and replacing their activations with \\(\\mu_{i}^{\\dagger}\\) at inference time already reduces the output dimension of \\(\\mathbf{W}_1\\), but the reconstruction of the original embedding dimension \\(D_{hid}\\) through replacement of the pruned activations still requires carrying out the full matrix multiplication for \\(\\mathbf{W}_2\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Instead, we perform an equivalent transformation by shifting these mean values directly into the bias of the final linear layer \\(\\mathbf{b}_2\\), allowing us to reduce both the output dimension of \\(\\mathbf{W}_1\\) as well as the input dimension of \\(\\mathbf{W}_2\\), doubling the cost savings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "We start from the usual two-layer linear mapping in the MLP block:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.122, + 0.771, + 0.138 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathbf {W} _ {2} \\mathbf {h} + \\mathbf {b} _ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.144, + 0.907, + 0.189 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}\\) is the hidden activation after pruning. For pruned neurons, we replace \\(h_j\\) with its mean \\(\\mu_{j}\\), which we collect into a vector \\(\\Delta_{\\mu}\\in \\mathbb{R}^{D_{\\mathrm{hid}}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.2, + 0.794, + 0.24 + ], + "angle": 0, + "content": "\\[\n(\\Delta_ {\\mu}) _ {j} = \\left\\{ \\begin{array}{l l} \\mu_ {j}, & j \\in \\mathcal {P}, \\\\ 0, & j \\not \\in \\mathcal {P}, \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.249, + 0.808, + 0.264 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}\\) is the set of pruned neuron indices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Instead of explicitly inserting these means into \\(\\mathbf{h}\\) and then multiplying by \\(\\mathbf{W}_2\\), we exploit the linearity of the mapping:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.32, + 0.834, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {2} \\mathbf {h} = \\mathbf {W} _ {2} \\left(\\mathbf {h} - \\Delta_ {\\mu}\\right) + \\mathbf {W} _ {2} \\Delta_ {\\mu}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.347, + 0.907, + 0.375 + ], + "angle": 0, + "content": "The second term is a constant vector, so we can shift it into the bias:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.377, + 0.789, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {2} ^ {\\prime} = \\mathbf {b} _ {2} + \\mathbf {W} _ {2} \\Delta_ {\\mu},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.399, + 0.766, + 0.415 + ], + "angle": 0, + "content": "which results in an equivalent output:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.425, + 0.907, + 0.443 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathbf {W} _ {2} (\\mathbf {h} - \\Delta_ {\\mu}) + \\mathbf {b} _ {2} ^ {\\prime}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.452, + 0.907, + 0.528 + ], + "angle": 0, + "content": "Because for each pruned index \\(j\\), we have replaced the activations \\(h_i\\) by their means \\(h_j = \\mu_j\\), it follows that \\((\\mathbf{h} - \\Delta_{\\mu})_j = 0\\). Consequently, we can both drop the row \\(j\\) of \\(\\mathbf{W}_1\\) as well as the column \\(j\\) of \\(\\mathbf{W}_2\\) without affecting the computation, reducing the hidden dimension by \\(|\\mathcal{P}|\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.907, + 0.652 + ], + "angle": 0, + "content": "Why Use Trained Networks? Our approach uses the fact that \\(\\mathbf{W}_1, \\mathbf{W}_2, \\mathbf{b}_1, \\mathbf{b}_2\\) are constants in trained networks. This allows the mean to be easily shifted into the bias of the next layer, which in turn allows the low-variance neurons to be removed without requiring extensive additional training. Only minor fine-tuning is sufficient to adjust the remaining weights for best performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.646, + 0.682 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.75 + ], + "angle": 0, + "content": "In this section, we apply our VBP method to various transformers and similar architectures, namely DeiT, Swin, and ConvNeXt in the Tiny, Small, and Base sizes, all trained on the ImageNet-1k [6] dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We evaluate these models twice. Once immediately after pruning (before fine-tuning) to measure how much accuracy is retained, and the second time after a brief fine-tuning to measure the final accuracy. Fine-tuning is performed over merely 10 epochs using knowledge-distillation from the unmodified base model. We use the AdamW optimizer with an initial learning rate of \\(1.5\\mathrm{e - 5}\\), decayed by a cosine-annealing scheduler, and a batch size of 32. As regularization, we include a weight decay of 0.01. All experiments are performed on NVIDIA H200 Tensor Core GPUs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.318, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) We omit the MLP indices for clarity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4976" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.907, + 0.281 + ], + "angle": 0, + "content": "
Model\n(Pruning Rate)MACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRet.VBP
DeiT-T (45%)1.260.94 (-25.16%)5.724.12 (-27.97%)72.0249.77 (69.13%)70.08 (97.33%)
DeiT-S (50%)4.613.21 (-30.37%)22.0514.96 (-32.15%)79.7064.44 (80.85%)78.62 (98.64%)
DeiT-B (55%)17.5811.44 (-34.93%)86.5755.40 (-36.01%)81.7357.58 (70.48%)80.67 (98.74%)
DeiT-B (20%)17.5815.35 (-12.68%)86.5775.24 (-13.09%)81.7380.87 (98.98%)81.76 (100.07%)
Swin-T (45%)4.503.23 (-28.22%)28.2921.31 (-24.67%)80.9155.12 (68.13%)79.41 (98.15%)
Swin-S (50%)8.765.63 (-32.19%)49.6135.02 (-29.41%)83.0467.01 (80.70%)81.86 (98.58%)
Swin-B (55%)15.4610.22 (-33.89%)87.7756.29 (-35.87%)84.7162.61 (73.91%)83.61 (98.70%)
Swin-B (20%)15.4613.58 (-12.16%)87.7776.42 (-12.93%)84.7183.90 (99.04%)84.67 (99.95%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.292, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Table 1. Results comparing the full baseline model with our VBP model using four metrics: MACs: computational operations, measured in billions of operations; and Parameters: the total model size in millions of parameters; Accuracy Retention (Ret.): retained accuracy after pruning, before fine-tuning; and Final Accuracy: accuracy after fine-tuning. Our method achieves competitive accuracy with significant reductions in MACs and parameters after fine-tuning, and even right after pruning, when reducing the pruning rate to \\(20\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.231, + 0.389 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.504 + ], + "angle": 0, + "content": "As is common with various pruning techniques, the pruning rate is adapted for the size of the models. Larger models typically exhibit higher redundancy for similar tasks, and therefore allow for more pruning while maintaining accuracy. We report our results for DeiT and Swin in Tab. 1 and Tab. 2. Note that the pruning rate is applied to the MLPs only and does not indicate the total reduction in model size." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.506, + 0.483, + 0.596 + ], + "angle": 0, + "content": "- For the base-sized models, we prune all MLPs globally using a pruning rate of \\(55\\%\\) keeping \\(99\\%\\) of the original performance after fine-tuning, reducing the total model size of DeiT-Base, Swin-Base both by \\(36\\%\\). This corresponds to \\(35\\%\\), and \\(34\\%\\) fewer MACs respectively and provides a speed-up of \\(1.44 \\times\\) and \\(1.30 \\times\\) times." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.597, + 0.483, + 0.687 + ], + "angle": 0, + "content": "- The small-sized models have notably fewer parameters than the base models and are thus pruned \\(50\\%\\) again reaching \\(99\\%\\) of the original performance. This leads to a DeiT-Small size reduction of \\(32\\%\\) and Swin-Small reduction of \\(29\\%\\) with \\(30\\%\\) and \\(32\\%\\) MACs savings respectively for a total speed-up factor of \\(1.34 \\times\\) and \\(1.29 \\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.688, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- Finally, for the DeiT- and Swin-Tiny models, we apply a lower pruning rate of \\(45\\%\\), achieving \\(97\\%\\) and \\(98\\%\\) final performance with \\(28\\%\\) and \\(25\\%\\) reductions in model size. This translates to \\(25\\%\\) and \\(28\\%\\) fewer MACs, along with speed-ups of \\(1.17 \\times\\) and \\(1.20 \\times\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.763, + 0.483, + 0.868 + ], + "angle": 0, + "content": "- For the Base-Models, we demonstrate that when globally pruning the MLPs by \\(20\\%\\) the accuracy retention directly after the one-shot structured pruning reaches \\(99\\%\\) of the unpruned performance without any fine-tuning. This allows for an off-the-shelf \\(13\\%\\) size and MACs reduction and a speed-up of \\(1.11 \\times\\) for DeiT-Base and a similar reductions for Swin-Base" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.506, + 0.483, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.903 + ], + "angle": 0, + "content": "All further pruning rates ranging from \\(5\\%\\) to \\(50\\%\\) for all DeiT Models can be found in Appendix B." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.372, + 0.907, + 0.45 + ], + "angle": 0, + "content": "
Model (Pruning Rate)Tiny (45%)Small (50%)Base (55%)Base (20%)
DeiT1.171.341.441.11
Swin1.201.271.301.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.459, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Table 2. Speed-ups of the pruned models relative to the baseline for different sizes and pruning rates." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.523, + 0.907, + 0.771 + ], + "angle": 0, + "content": "
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
DeiT-T1.265.72-72.02
Magnitude0.913.943.5668.49
SNIP0.913.9424.8969.10
VBP (ours)0.913.9439.5870.61
DeiT-S4.6122.05-79.70
Magnitude3.2114.964.0576.55
SNIP3.2114.9652.3977.27
VBP (ours)3.2114.9664.4478.62
DeiT-B17.5886.57-81.73
Magnitude12.058.240.3778.88
SNIP12.058.2453.2480.40
VBP (ours)12.058.2466.4080.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.852 + ], + "angle": 0, + "content": "Table 3. Comparison of different one-shot pruning scores in structured pruning to our VBP approach at fixed \\(50\\%\\) pruning rates. For constant MACs and parameters, our approach retains more accuracy (Ret.) and reaches a higher final accuracy compared to other pruning approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4977" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.088, + 0.907, + 0.26 + ], + "angle": 0, + "content": "
ModelMACs (G)Param. (M)Speed-UpTop-1 Acc. (%)
Ret.Final
DeiT-B17.5886.57--81.73
ToMe-246.0286.572.1535.8575.74
ToMe-14 & VBP6.0858.242.0560.5380.09
ToMe-207.1486.571.8466.3178.66
ToMe-12 & VBP6.9058.241.8362.0980.29
ToMe-187.8886.571.7573.7979.97
ToMe-10 & VBP7.7358.241.7063.7280.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.27, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Table 4. Comparison of ToMe for different token reduction rates of 24, 20, and 18, with the hybrid combination of ToMe using token reduction rates of 14, 12, and 10, and VBP using \\(50\\%\\) pruning rates on top. Our hybrid method achieves significantly higher final accuracies throughout while also providing reductions in parameters, for similar MACs and speed-ups. The hybrid model thereby reaches speed-ups of up to \\(2.05 \\times\\) while maintaining competitive performance." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.35, + 0.485, + 0.455 + ], + "angle": 0, + "content": "
ModelEpochsParam. (M)Top-1 Acc. (%)
Ret.Final
NViT (50 Ep.)5056.3781.1382.18
NViT (1 Ep.)156.8569.1081.92
VBP (ours)156.1872.3782.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.465, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Table 5. Comparison with NViT using two pruning durations of 50 epochs (as originally proposed) and 1 epoch (similar to our VBP approach), both followed by 10 epochs of fine-tuning. Our method achieves higher accuracy immediately after pruning, before fine-tuning (Ret.) compared to NViT with the same pruning duration, and reaches a higher final accuracy than the 50-epoch NViT model after fine-tuning for 10 epochs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.327, + 0.623 + ], + "angle": 0, + "content": "4.2. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.682 + ], + "angle": 0, + "content": "To facilitate direct comparisons of all models with different sizes across all experiments, we fix the pruning rate at \\(50\\%\\) for all further experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of our selection criterion, by benchmarking against two other common and well-established one-shot-pruning baselines, namely Magnitude and SNIP, adapted for structured pruning. For a fixed pruning rate of \\(50\\%\\) using all selection criteria, this produces identical parameter counts of 3.94, 14.96, and 58.24 million parameters, as well as fixed 0.91, 3.21, and 12.00 GMACs for the three model sizes. Our results indicate that VBP provides higher accuracy retention compared to the baselines, improving DeiT-Base by \\(13.16\\%\\), DeiT-Small by \\(12.05\\%\\), and DeiT-Tiny by \\(14.69\\%\\) percentage points compared to SNIP. This difference in accuracy retention yields a final fine-tuned performance that is \\(0.59\\%\\) to \\(1.51\\%\\) percentage points higher after fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.352, + 0.733, + 0.369 + ], + "angle": 0, + "content": "4.3. SoTA Pruning Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.529 + ], + "angle": 0, + "content": "To further validate our approach, we compare against the SoTA structured transformer pruning method NViT (CVPR'23). Specifically, we apply our method to the NViT Base model (DeiT-Base distilled from RegNetY-160) and run NViT with a latency target of \\(70\\%\\), pruning once for 50 epochs (as per the published results) and once for one epoch (similar to our setup), with both scenarios followed by ten epochs of fine-tuning using the NViT settings. In a similar pruning setting, VBP achieves a \\(0.4\\%\\) percentage points higher final accuracy, for similar number of parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Finally, we demonstrate the orthogonality of VBP to token-pruning methods by applying our method on top of the SoTA token-merging approach, ToMe (ICLR'23). We observe that the accuracy retention drops relative to the baseline remains consistent for both the VBP-pruned base DeiT model in Tab. 1 and the VBP- & ToMe-pruned model. This relative consistency also persists throughout fine-tuning, indicating that our method operates orthogonally to ToMe in terms of performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Regarding the MAC savings it is important to note that ToMe progressively reduces the throughput of tokens processed across the layers. Later layers therefore process less tokens, making pruning less effective. The relative MAC savings from our hybrid method therefore decrease as the network depth increases. Nonetheless, VBP additionally reduces the total parameter count, which cannot be achieved by simply increasing the ToMe token reduction rate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In fact, increasing the reduction rate to get an approximate \\(2 \\times\\) speed-up yields similar MAC savings but reduces both the accuracy retention as well as the final accuracy significantly. Instead applying VBP on top of ToMe with a smaller reduction rate, allows more than \\(2 \\times\\) speed-ups while still maintaining \\(98\\%\\) of the original accuracy with additional parameter savings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4978" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.483, + 0.187 + ], + "angle": 0, + "content": "
Variance- BasedMean- ShiftTop-1 Acc. (%)
Ret.Final
X✓55.1980.23
✓X26.0480.62
✓✓66.4080.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.197, + 0.483, + 0.267 + ], + "angle": 0, + "content": "Table 6. Ablation study results showing the impact of Variance-Based Pruning, and Mean-Shift Compensation on the retained accuracy after pruning (Ret.), as well as the final: accuracy after fine-tuning. Both steps together yield the highest accuracy of \\(80.99\\%\\)." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.283, + 0.483, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Figure 2. Accuracy retention before fine-tuning and final accuracy after 10 epochs of fine-tuning for different pruning rates applied to DeiT-Base." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.55, + 0.483, + 0.749 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.788 + ], + "angle": 0, + "content": "Figure 3. Activation variance for all neurons in the hidden MLP layers throughout the network and marked cumulative variances." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.188, + 0.831 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To assess the contribution of each individual component of our approach, we perform an ablation study in which we systematically modify the pruning criterion and the Mean-Shift Compensation. We report our results in Tab. 6." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.906, + 0.173 + ], + "angle": 0, + "content": "
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
Pre-Act. (eq. [23])12.058.240.4377.92
Post-Act. (ours)12.058.2466.4080.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.182, + 0.906, + 0.267 + ], + "angle": 0, + "content": "Table 7. Comparison of different locations for the activation statistics computation in the network: before (Pre-Act.) as well as after the activation function (Post-Act.). For the same MACs and parameters, using pruning rates of \\(50\\%\\), gathering statistics post-activation significantly outperforms pre-activation in both accuracy retention (Ret.) as well as the final accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.905, + 0.361 + ], + "angle": 0, + "content": "We first remove the variance-based pruning selection while still shifting the mean activation to the next bias. Removing the Variance-Based Pruning alone reduces the final accuracy by \\(0.76\\%\\) percentage points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.425 + ], + "angle": 0, + "content": "We also remove the Mean-Shift Compensation while pruning using the variance-based criterion. This improves the final accuracy by \\(0.39\\%\\) compared to using only the Mean-Shift Compensation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.565 + ], + "angle": 0, + "content": "However, the best results are achieved when both Variance-Based Pruning and Mean-Shift Compensation are used together. As discussed in Sec. 3.2, replacing the removed activations with their means is optimal when pruning based on variance. When both components are combined, the immediate accuracy retention increases by \\(11.21\\%\\) percentage points, ultimately improving the fine-tuned accuracy by an additional \\(0.37\\%\\). This highlights the significance of both components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.831, + 0.602 + ], + "angle": 0, + "content": "5.1. Sensitivity of Variance Thresholding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.702 + ], + "angle": 0, + "content": "We further examine the stability of the activation variance criterion by applying different pruning rates to the DeiT-Base model, and plotting the accuracy retention and final accuracy in Fig. 2. We observe that for smaller pruning rates (up to around \\(25\\%\\)), the accuracy retention is high enough to allow for off-the-shelf deployment without any fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Beyond these rates, the accuracy begins to degrade more noticeably, reflecting the non-uniform distribution of neuron variances, as seen in Fig. 3. Notably, to account for \\(10\\%\\) of the cumulative variance across all neurons, nearly \\(60\\%\\) of the lowest-variance neurons are needed. This is a disproportionately large share compared to the next \\(15\\%\\) of neurons that also account for \\(10\\%\\) of the total variance in a layer. Consequently, more aggressive pruning leads to an increasingly faster reduction in the total variance and thus expressiveness of the network. Nevertheless, significant performance issues arise only at very high pruning levels, validating the feasibility of our method over a broad range of pruning rates (see Appendix B)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4979" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.187 + ], + "angle": 0, + "content": "
ModelMACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRetentionVBP
ConvNeXt-T4.472.96 (-33.8%)28.5912.61 (-55.9%)82.9016.80 (20.3%)81.30 (98.1%)
ConvNeXt-S8.715.11 (-41.3%)50.2223.50 (-53.2%)84.5730.92 (36.6%)82.82 (97.9%)
ConvNeXt-B15.388.91 (-42.1%)88.5941.32 (-53.4%)85.5157.10 (66.8%)83.40 (97.6%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.197, + 0.907, + 0.227 + ], + "angle": 0, + "content": "Table 8. Results comparing the full ConvNeXt baseline models with the VBP models using a pruning rate of \\(50\\%\\). The MACs are reduced by up to \\(42\\%\\) and parameters by over \\(50\\%\\) while reaching competitive final accuracies of \\(98\\%\\) original performance." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.25, + 0.485, + 0.297 + ], + "angle": 0, + "content": "
ModelTinySmallBase
ConvNeXt1.281.421.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Table 9. Speed-ups of the pruned models relative to the baseline for different-sized ConvNeXt models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.405, + 0.379 + ], + "angle": 0, + "content": "5.2. Application Pre- vs. Post-Activation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.483, + 0.536 + ], + "angle": 0, + "content": "While Sec. 3.2 shows that our mean-replacement strategy can be grounded in theory for any distribution, the question remains whether the variance should be measured before or after the activation function. The Central Limit Theorem confirms that pre-activation sums (composed of independent weight contributions) tend to approximate a normal distribution. However, once the nonlinearity \\(\\sigma\\) is applied, this distribution changes. Consequently, the pre-activation variance does not necessarily correlate with the neuron importance for retaining accuracy in an already trained model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.483, + 0.612 + ], + "angle": 0, + "content": "For instance, highly varying negative activations are compressed into range of about \\((-0.2, 0)\\) by GeLU, reducing their variance. As a result, such a neuron would be less likely to be pruned in a pre-activation setting and more likely to be pruned post-activation (see Appendix E)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.483, + 0.748 + ], + "angle": 0, + "content": "To validate this experimentally, we tested an alternative design in which pruning decisions are made based on the pre-activation variance in Tab. 7. While this approach has been successfully applied when pruning during training [23], we find that in a trained network setting, it significantly degrades performance. Therefore, measuring variance post-activation better captures neuron importance in already trained models, and helps maintain model performance despite structural modifications." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.424, + 0.773 + ], + "angle": 0, + "content": "5.3. Other Transformer-Like Architectures" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.87 + ], + "angle": 0, + "content": "While our method is broadly applicable to any model containing MLP layers, network architectures that have more and larger MLPs benefit more from VBP. We therefore apply our pruning on ConvNeXt as a transformer-like architecture, that relies heavily on MLP layers, to further evaluate the generalizability of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We summarize our performance comparisons for ConvNeXt in Tab. 8. We achieve significant parameter reduc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.907, + 0.344 + ], + "angle": 0, + "content": "tions of over \\(50\\%\\) throughout all three sizes and reduced MACs by \\(34\\%\\) for ConvNeXt-Tiny and over \\(40\\%\\) for the larger sizes, while maintaining \\(98\\%\\) of the original accuracy after 10 epochs of fine-tuning across the board. The corresponding speed-ups are listed in Tab. 9, and reach up to \\(1.49\\times\\) for ConvNeXt-Base." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.909, + 0.515 + ], + "angle": 0, + "content": "Notably, we observe significantly less accuracy retention throughout the smaller sized models compared to the true transformers. We attribute this to the fact that the ConvNeXt architecture is built upon convolutions, which have significantly higher costs compared to their parameters. Consequently, since the MLPs take over a higher overall percentage of the total capacity, removing a similar fraction of neurons causes a relatively higher drop in accuracy retention. By contrast, transformers typically have less parameters in their MLP relatively speaking, thereby allowing for higher pruning rates while maintaining capacity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.541, + 0.634, + 0.557 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.647 + ], + "angle": 0, + "content": "We introduce VBP, a one-shot pruning method, designed to remove neurons based on their statistical importance in a single pruning operation while minimizing accuracy loss by integrating their mean activations in a Mean-Shift Compensation step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Our experiments demonstrate that VBP retains model performance significantly better than existing structured pruning methods. By taking advantage of trained networks, our approach provides a favourable starting point for the subsequent finetuning, eliminating the need for extensive retraining and thereby making it practical for low-cost deployment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.907, + 0.822 + ], + "angle": 0, + "content": "Furthermore, we show that VBP operates orthogonally to SoTA token pruning methods, allowing it to be seamlessly combined with approaches like ToMe to achieve even greater computational savings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We believe that the simplicity and efficiency of our approach contributes to the democratization of deep learning by enabling a wider reuse of trained networks and hope it inspires further research in the CV community toward efficient model compression techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4980" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Sajid Anwar, Kyuyeon Hwang, and Wonyong Sung. Structured pruning of deep convolutional neural networks. J. Emerg. Technol. Comput. Syst., 13(3), 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.482, + 0.225 + ], + "angle": 0, + "content": "[2] Uranik Berisha, Jens Mehnert, and Alexandru Condurache. Efficient data driven mixture-of-expert extraction from trained networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.482, + 0.31 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[4] Rinor Cakaj, Jens Mehnert, and Bin Yang. Cnn mixture-of-depths. In Computer Vision ACCV 2024: 17th Asian Conference on Computer Vision, Hanoi, Vietnam, December 812, 2024, Proceedings, Part VII, page 148166, Berlin, Heidelberg, 2024. Springer-Verlag. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.384, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[5] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19974-19988, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12114-12124. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6804-6815. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Herve Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Canada, October 10-17, 2021, pages 12239-12249. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[11] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[12] Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pages 598-605. Morgan Kaufmann, 1989. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[13] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip H. S. Torr. Snip: single-shot network pruning based on connection sensitivity. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.29, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[14] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[15] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.443, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[16] Lukas Meiner, Jens Mehnert, and Alexandru Condurache. Data-free dynamic compression of cnns for tractable efficiency. In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 2: VISAPP, pages 196-208. INSTICC, SciTePress, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[17] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12299-12308. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[18] Dmitry Molchanov, Arsenii Ashukha, and Dmitry P. Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2498-2507. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[19] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 13937-13949, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[20] David Raposo, Samuel Ritter, Blake A. Richards, Timothy P. Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models. CoRR, abs/2404.02258, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4981" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 10347-10357. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[23] Berry Weinstein and Yonatan Belinkov. Variance pruning: Pruning language models via temporal neuron variance, 2025. Unpublished manuscript, OpenReview. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[24] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[25] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Freezenet: Full performance by reduced storage costs. In Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, pages 685-701. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.393, + 0.484, + 0.463 + ], + "angle": 0, + "content": "[26] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Interspace pruning: Using adaptive filter representations to improve training of sparse cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12527-12537, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.465, + 0.483, + 0.547 + ], + "angle": 0, + "content": "[27] Huanrui Yang, Hongxu Yin, Maying Shen, Pavlo Molchanov, Hai Li, and Jan Kautz. Global vision transformer pruning with hessian-aware saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18547-18557. IEEE, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[28] Fang Yu, Kun Huang, Meng Wang, Yuan Cheng, Wei Chu, and Li Cui. Width & depth pruning for vision transformers. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 3143-3151. AAAI Press, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[29] Lu Yu and Wei Xiang. X-pruner: explainable pruning for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24355-24363. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[30] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[31] Zhengyan Zhang, Zhiyuan Zeng, Yankai Lin, Chaojun Xiao, Xiaozhi Wang, Xu Han, Zhiyuan Liu, Ruobing Xie, Maosong Sun, and Jie Zhou. Emergent modularity in pretrained transformers. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4066-4083, Toronto, Canada, 2023. Association for Computational Linguistics. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[32] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning, 2021. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4982" + } + ] +] \ No newline at end of file diff --git a/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_origin.pdf b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7a2bbc6110fa87f05e5f9856ab4de6c47280729 --- /dev/null +++ b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/5aed5135-0f01-4c2d-b7ee-d3f1987cb1ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bee2953a05b560cd3f4c80574b3c112eef54fa2ab8fc7df98b0d8a4b7d2f6a2a +size 333690 diff --git a/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/full.md b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7cd40fefcc43bf51f055c5d197e0d15c225e4d49 --- /dev/null +++ b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/full.md @@ -0,0 +1,330 @@ +# Variance-Based Pruning for Accelerating and Compressing Trained Networks + +Uranik Berisha12, Jens Mehnert1 and Alexandru Paul Condurache12 + +1Automated Driving Research, Robert Bosch GmbH, 70469 Stuttgart, Germany +$^{2}$ Institute for Signal Processing, University of Lübeck, 23562 Lübeck, Germany + +{Uranik.Berisha,JensEricMarkus.Mehnert,AlexandruPaul.Condurache}@de.bosch.com + +# Abstract + +Increasingly expensive training of ever larger models such as Vision Transformers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge. + +To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over $70\%$ of its original performance and requires only 10 epochs of fine-tuning to regain $99\%$ of the original accuracy while simultaneously reducing MACs by $35\%$ and model size by $36\%$ , thus speeding up the model by $1.44 \times$ . + +# 1. Introduction + +In both Computer Vision (CV) and Natural Language Processing (NLP), Transformers have become the dominant architecture [8, 21]. In CV, Vision Transformers (ViTs) [8] and their derivatives, such as Data-efficient image Transformers (DeiTs) [21], Swin-Transformers (Swins) [14] and ConvNeXts [15], achieve remarkable results but at significant costs in three main areas: training, storage, and inference. These large models require long training times, often spanning hundreds of epochs to achieve their performance, extensive memory to store their parameters, and high computational resources for inference causing high latency [21]. + +Fixing all three issues simultaneously remains especially challenging. For example, using the available collection of trained networks provided by the research community is a straightforward way to solve the first problem of costly training. However, this does not address the storage and inference costs, which can still make the use of these trained networks impractical for deployment on resource-constrained hardware [1]. To address these demands, various model pruning techniques have been proposed. These range from removing individual weights within layers in unstructured pruning [12] to removing entire neurons, layers, or blocks in structured pruning [1]. + +While unstructured pruning can often retain a high degree of accuracy in trained networks, it is difficult to translate these theoretical gains into real-world speed-ups [1]. Since modern hardware is optimized for dense matrix multiplications, they cannot exploit the resulting sparsity effectively and therefore cannot solve the inference costs [1, 12]. + +In contrast, structured pruning leads to straightforward reductions in both inference costs and memory, but again requires extensive retraining to recover the lost accuracy resulting from these major structural modifications. For instance, Global Vision Transformer Pruning with Hessian-Aware Saliency (NViT) [27] requires 300 epochs of fine-tuning despite using trained networks, highlighting how addressing memory and inference speeds alone can still demand large additional training overhead. + +On the other hand, dynamic pruning approaches, such as token pruning, avoid permanently modifying the model architecture and instead dynamically reduce the number of features processed by the network [3, 19]. This minimizes the accuracy drop after pruning, indicating that the model preserved most important features and representation, thus providing a better starting point for subsequent fine-tuning. Token Merging (ToMe) [3] embraces this idea by merging tokens to reuse the pruned information, thus reducing the accuracy drop even further and allowing off-the-shelf deployment without any fine-tuning. While this solves the inference costs as well as the training costs, since the methods do not modify the model structure itself, the memory foot + +![](images/b74940e48b256274fae1d8c185235d3ee92e28525b305283a24a1edea1d4b689.jpg) +(a) We compute neuron-wise mean and variance of activations at the hidden layer of each MLP. + +![](images/9afc80bf2650b05c6d6475840a0765e1b7d3da968b26d255b997c3737b9d4311.jpg) +(b) We prune neurons with the smallest variances and replace their activation with the mean. + +![](images/9d396469f25068fe47f42ee8df55f899860e48c449ca37d7276b05d0b7cd0495.jpg) +(c) We shift the mean activation of the removed neurons into the bias of the next layer. +Figure 1. Schematic visualization of the steps of Variance-Based Pruning. (a) Activation Statistics Computation, (b) Variance-Based Pruning, (c) Mean-Shift Compensation + +print remains unchanged [3]. This again is a limitation for hardware-constrained settings [1]. Fully resolving all three problems of training costs, storage, and inference speed simultaneously remains a challenge. + +In this work, we fill this gap by introducing Variance-Based Pruning (VBP), a simple and structured pruning approach that provides speed-ups and memory savings with minimal fine-tuning. In order to keep the method simple and allow deployment across a wide variety of architectures, we prune only the heavy Multi-Layer Perceptron (MLP) layers within the transformer blocks. In the first step, our approach leverages activation statistics gathered in the hidden layers of the MLP using Welford's algorithm [24] (see Fig. 1a). In the second step, these statistics are used to identify and remove the least impactful hidden neurons (see Fig. 1b). In the last step, which we refer to as Mean-Shift Compensation, we use the mean activation values of pruned neurons to mitigate accuracy drop by redistributing these mean contributions into the bias of the output layer (see Fig. 1c). This retains high levels of model performance and requires only minimal fine-tuning to then regain most of the original accuracy. + +# Our contributions are summarized as follows: + +- We introduce VBP, a simple low-cost structured pruning method broadly applicable across various architectures to significantly reduce computational costs while maintaining performance. +- We apply our method to multiple architectures, including ViTs, Swin-Transformers, and ConvNeXts, demonstrating its generalizability. + +- We compare VBP to similarly straightforward neuron-importance measures adapted for structured pruning, as well as to state-of-the-art (SoTA) pruning approaches. We benchmark against NViT pruning, highlighting the benefits of our accuracy retention in constrained fine-tuning settings. We also apply VBP on top of ToMe demonstrating orthogonality and achieving $2 \times$ speed-ups using our hybrid approach. +- We perform an extensive ablation study, demonstrating the efficacy of each component of our method and provide a sensitivity analysis for the variance as a pruning criterion. + +# 2. RELATED WORK + +# 2.1. Transformer Architectures + +Initially introduced in NLP by [22], Transformers have largely replaced Convolutional Neural Network (CNN), since their adaption to CV by [8] with the ViT. ViTs processes images as a sequence of non-overlapping patches, and have since been further improved. DeiT [21] refined the training procedure to decrease the training costs and reduce the data reliance. More recent architectures [7, 9, 10], such as Swin [14] and the transformer-inspired ConvNeXt [15], which use hierarchical designs, have built upon this architecture to improve efficiency and scalability. However, despite these advances, these large models remain computationally demanding, in part due to their dense MLP blocks, which significantly impacts inference speed and memory requirements [22, 30]. + +# 2.2. Model Compression + +To reduce the costs of large models, model compression efforts focus on scaling down different components of a network. Traditionally pruning aims to identify and remove redundant structures within the network [12, 26]. For example, magnitude pruning eliminates weights with the smallest absolute values, while SNIP and related methods [13, 25] identify crucial connections using gradient-based sensitivities. Structured pruning approaches, remove entire neurons, layers or filters producing smaller matrices that allow direct computational savings in current hardware [1, 5, 27-29, 32]. Most closely related to our method are works exploring variance-based criteria for structured pruning, such as Molchanov et al. [18], which performs dropout at the neuron level and can introduce sparsity in neurons with high dropout rates. Weinstein et al. [23] use these statistics in the initial training phase to remove neurons during training. These approaches operate on models before training, improving training time but still requiring the models to be trained from scratch. Contrary to that, NViT [27] utilizes trained ViTs. They apply iterative structured dimensionality reductions across the entire ViT architecture and significantly reduce costs. However, these structured modifications still have a substantial impact on performance, requiring high pruning iterations of 50 epochs and retraining costs of 300 epochs to regain accuracy. + +An alternative to structurally modifying the network is dynamic pruning, which adaptively selects and processes only a subset of features or combines features during inference to reduce costs. For CNNs, these have shown promising speed-accuracy trade-offs [4, 16]. Analogously for transformer-like models, token pruning reduces the number of tokens to be processed [3, 17, 19, 20]. For instance, Bolya et al. [3] merges similar tokens rather than pruning them, thereby conserving most of the information and retaining high accuracy before the fine-tuning. However, by design these methods cannot reduce the memory footprint. + +# 2.3. Our Approach + +Our work builds upon these ideas by combining (i) the advantages of structured pruning by removing low-variance activation neurons with (ii) the advantages of dynamic pruning by incooperating their mean contribution into the bias of the last MLP layer. Our approach is based on the observation that neurons within MLP layers exhibit varying levels of contribution to the overall representation, as established in works on emerging modularity in trained Transformers for both NLP and CV [2, 31]. By applying our variance-based pruning approach to the widespread MLP layers, we provide a structured method applicable across a broad range of models. Our method is particularly beneficial for transformer-like architectures, where MLP layers heavily contribute to the computational overhead [22, 30]. + +# 3. Methodology + +In this section, we detail the components involved in our proposed VBP approach and provide a mathematical derivation, justifying our pruning criterion and compensation strategy. + +Our method consists of three major steps: + +1. Activation Statistics Computation The neuron activation statistics are computed using Welford's algorithm. +2. Variance-Based Pruning The neurons to be pruned are selected based on the lowest activation variance. +3. Mean-Shift Compensation The mean activations of the selected neurons are added back into the output bias. + +# 3.1. Step 1: Activation Statistics Computation + +As we aim for a simple and widely adoptable method, we prune only the MLPs. This requires reducing the hidden layers of the network while keeping the input and output layers intact to ensure seamless integration with the rest of the model structure by maintaining dimensional consistency in the input and output. + +W.l.o.g. we therefore consider a MLP with a single hidden layer of dimension $D_{\mathrm{hid}}$ , that maps an input vector $\mathbf{x} \in \mathbb{R}^{D_{\mathrm{in}}}$ to an output vector $\mathbf{y} \in \mathbb{R}^{D_{\mathrm{out}}}$ . Using a pointwise nonlinear activation function $\sigma(\cdot)$ , this MLP can then be described as a function + +$$ +\mathbf {h} = \sigma \left(\mathbf {W} _ {1} \mathbf {x} + \mathbf {b} _ {1}\right), \quad \mathbf {y} = \mathbf {W} _ {2} \mathbf {h} + \mathbf {b} _ {2}, \tag {1} +$$ + +where $\mathbf{W}_1 \in \mathbb{R}^{D_{\mathrm{hid}} \times D_{\mathrm{in}}}$ , $\mathbf{b}_1 \in \mathbb{R}^{D_{\mathrm{hid}}}$ , $\mathbf{W}_2 \in \mathbb{R}^{D_{\mathrm{out}} \times D_{\mathrm{hid}}}$ , $\mathbf{b}_2 \in \mathbb{R}^{D_{\mathrm{out}}}$ . Since we deal with trained networks, these parameters $\mathbf{W}_1, \mathbf{b}_1, \mathbf{W}_2, \mathbf{b}_2$ remain fixed during pruning, which will become important in the Mean-Shift Compensation in Sec. 3.3. + +For the hidden layer, we now want to calculate the perneuron mean and variance vectors + +$$ +\boldsymbol {\mu} = \left(\mu_ {1}, \dots , \mu_ {D _ {\mathrm {h i d}}}\right) ^ {\top} \quad \text {a n d} \quad \boldsymbol {\sigma} ^ {2} = \left(\sigma_ {1} ^ {2}, \dots , \sigma_ {D _ {\mathrm {h i d}}} ^ {2}\right) ^ {\top} +$$ + +To compute these statistics, we capture the activations $\mathbf{h}$ for every batch and use Welford's algorithm to compute a running mean $\pmb{\mu}^{(j)}$ and running second moment $\mathbf{m}_2^{(j)}$ for an efficient online calculation of streaming data*. Given a total of $N$ data samples, let the $j$ -th sample $\mathbf{h}^{(j)} \in \mathbb{R}^{D_{\mathrm{hid}}}$ be the activation vector after the nonlinearity. Welford's algorithm ensures numerically stable calculations by updating $\pmb{\mu}^{(j)}$ and $\mathbf{m}_2^{(j)}$ after observing the $j$ -th sample as follows: + +$$ +\boldsymbol {\mu} ^ {(j)} = \frac {j - 1}{j} \boldsymbol {\mu} ^ {(j - 1)} + \frac {1}{j} \mathbf {h} ^ {(j)} \tag {2} +$$ + +$$ +\mathbf {m} _ {2} ^ {(j)} = \mathbf {m} _ {2} ^ {(j - 1)} + \left(\mathbf {h} ^ {(j)} - \boldsymbol {\mu} ^ {(j - 1)}\right) \odot \left(\mathbf {h} ^ {(j)} - \boldsymbol {\mu} ^ {(j)}\right) \tag {3} +$$ + +where $\odot$ denotes elementwise multiplication. Once all $N$ samples are processed, the variance is given by: + +$$ +\boldsymbol {\sigma} ^ {2} = \frac {\mathbf {m} _ {2} ^ {(N)}}{N - 1}. \tag {4} +$$ + +# 3.2. Step 2: Variance-Based Pruning + +Using this variance vector $\sigma^2 \in \mathbb{R}^{D_{\mathrm{hid}}}$ , we rank neurons by their variance values, $\sigma_i^2$ . Intuitively, if an activation rarely deviates from its mean $\mu_i$ , the corresponding neuron contributes less to the expressiveness of the network. The pruning decision is formed across all hidden neurons in the network, to capture the neuron with the least variance across all layers. + +Formally, for each MLP $l$ , let $D_{\mathrm{hid}}^{(l)}$ denote the number of hidden neurons. For each neuron $i$ in MLP $l$ , we then simply select its activation variance $\sigma_{l,i}^2$ as the corresponding pruning score. Given a pruning ratio $p \in (0,1)$ , we then only need to form the global set of scores + +$$ +S = \bigcup_ {l \in \mathcal {L}} \{\sigma_ {l, i} ^ {2} \mid i = 1, \ldots , D _ {\mathrm {h i d}} ^ {(l)} \}. +$$ + +from which we select the $p\%$ neurons with the smallest scores for pruning. + +Optimality from a Mean-Replacement Perspective To retain the contribution of the pruned network, we approximate the activation of a pruned neuron $h_i$ by replacing it with its mean $\mu_i$ . This introduces an error proportional to $|h_i - \mu_i|$ , which, by definition of the sample variance for any distribution, has an expected value: + +$$ +\mathbb {E} [ (h _ {i} - \mathbb {E} [ h _ {i} ]) ^ {2} ] = \mathbb {E} [ (h _ {i} - \mu_ {i}) ^ {2} ] = \sigma_ {i} ^ {2}. +$$ + +Thus, pruning the neurons with the lowest variance results in the least reconstruction error, making variance the optimal metric in this context. This is intuitively confirmed by the fact that when the variance is zero, all activations are exactly equal to the mean. + +# 3.3. Step 3: Mean-Shift Compensation + +The intuitive approach of removing neurons and replacing their activations with $\mu_{i}^{\dagger}$ at inference time already reduces the output dimension of $\mathbf{W}_1$ , but the reconstruction of the original embedding dimension $D_{hid}$ through replacement of the pruned activations still requires carrying out the full matrix multiplication for $\mathbf{W}_2$ . + +Instead, we perform an equivalent transformation by shifting these mean values directly into the bias of the final linear layer $\mathbf{b}_2$ , allowing us to reduce both the output dimension of $\mathbf{W}_1$ as well as the input dimension of $\mathbf{W}_2$ , doubling the cost savings. + +We start from the usual two-layer linear mapping in the MLP block: + +$$ +\mathbf {y} = \mathbf {W} _ {2} \mathbf {h} + \mathbf {b} _ {2}, +$$ + +where $\mathbf{h}$ is the hidden activation after pruning. For pruned neurons, we replace $h_j$ with its mean $\mu_{j}$ , which we collect into a vector $\Delta_{\mu}\in \mathbb{R}^{D_{\mathrm{hid}}}$ + +$$ +(\Delta_ {\mu}) _ {j} = \left\{ \begin{array}{l l} \mu_ {j}, & j \in \mathcal {P}, \\ 0, & j \not \in \mathcal {P}, \end{array} \right. +$$ + +where $\mathcal{P}$ is the set of pruned neuron indices. + +Instead of explicitly inserting these means into $\mathbf{h}$ and then multiplying by $\mathbf{W}_2$ , we exploit the linearity of the mapping: + +$$ +\mathbf {W} _ {2} \mathbf {h} = \mathbf {W} _ {2} \left(\mathbf {h} - \Delta_ {\mu}\right) + \mathbf {W} _ {2} \Delta_ {\mu}. +$$ + +The second term is a constant vector, so we can shift it into the bias: + +$$ +\mathbf {b} _ {2} ^ {\prime} = \mathbf {b} _ {2} + \mathbf {W} _ {2} \Delta_ {\mu}, +$$ + +which results in an equivalent output: + +$$ +\mathbf {y} = \mathbf {W} _ {2} (\mathbf {h} - \Delta_ {\mu}) + \mathbf {b} _ {2} ^ {\prime}. \tag {5} +$$ + +Because for each pruned index $j$ , we have replaced the activations $h_i$ by their means $h_j = \mu_j$ , it follows that $(\mathbf{h} - \Delta_{\mu})_j = 0$ . Consequently, we can both drop the row $j$ of $\mathbf{W}_1$ as well as the column $j$ of $\mathbf{W}_2$ without affecting the computation, reducing the hidden dimension by $|\mathcal{P}|$ . + +Why Use Trained Networks? Our approach uses the fact that $\mathbf{W}_1, \mathbf{W}_2, \mathbf{b}_1, \mathbf{b}_2$ are constants in trained networks. This allows the mean to be easily shifted into the bias of the next layer, which in turn allows the low-variance neurons to be removed without requiring extensive additional training. Only minor fine-tuning is sufficient to adjust the remaining weights for best performance. + +# 4. Experiments + +In this section, we apply our VBP method to various transformers and similar architectures, namely DeiT, Swin, and ConvNeXt in the Tiny, Small, and Base sizes, all trained on the ImageNet-1k [6] dataset. + +We evaluate these models twice. Once immediately after pruning (before fine-tuning) to measure how much accuracy is retained, and the second time after a brief fine-tuning to measure the final accuracy. Fine-tuning is performed over merely 10 epochs using knowledge-distillation from the unmodified base model. We use the AdamW optimizer with an initial learning rate of $1.5\mathrm{e - 5}$ , decayed by a cosine-annealing scheduler, and a batch size of 32. As regularization, we include a weight decay of 0.01. All experiments are performed on NVIDIA H200 Tensor Core GPUs. + +
Model +(Pruning Rate)MACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRet.VBP
DeiT-T (45%)1.260.94 (-25.16%)5.724.12 (-27.97%)72.0249.77 (69.13%)70.08 (97.33%)
DeiT-S (50%)4.613.21 (-30.37%)22.0514.96 (-32.15%)79.7064.44 (80.85%)78.62 (98.64%)
DeiT-B (55%)17.5811.44 (-34.93%)86.5755.40 (-36.01%)81.7357.58 (70.48%)80.67 (98.74%)
DeiT-B (20%)17.5815.35 (-12.68%)86.5775.24 (-13.09%)81.7380.87 (98.98%)81.76 (100.07%)
Swin-T (45%)4.503.23 (-28.22%)28.2921.31 (-24.67%)80.9155.12 (68.13%)79.41 (98.15%)
Swin-S (50%)8.765.63 (-32.19%)49.6135.02 (-29.41%)83.0467.01 (80.70%)81.86 (98.58%)
Swin-B (55%)15.4610.22 (-33.89%)87.7756.29 (-35.87%)84.7162.61 (73.91%)83.61 (98.70%)
Swin-B (20%)15.4613.58 (-12.16%)87.7776.42 (-12.93%)84.7183.90 (99.04%)84.67 (99.95%)
+ +# 4.1. Main Results + +As is common with various pruning techniques, the pruning rate is adapted for the size of the models. Larger models typically exhibit higher redundancy for similar tasks, and therefore allow for more pruning while maintaining accuracy. We report our results for DeiT and Swin in Tab. 1 and Tab. 2. Note that the pruning rate is applied to the MLPs only and does not indicate the total reduction in model size. + +- For the base-sized models, we prune all MLPs globally using a pruning rate of $55\%$ keeping $99\%$ of the original performance after fine-tuning, reducing the total model size of DeiT-Base, Swin-Base both by $36\%$ . This corresponds to $35\%$ , and $34\%$ fewer MACs respectively and provides a speed-up of $1.44 \times$ and $1.30 \times$ times. +- The small-sized models have notably fewer parameters than the base models and are thus pruned $50\%$ again reaching $99\%$ of the original performance. This leads to a DeiT-Small size reduction of $32\%$ and Swin-Small reduction of $29\%$ with $30\%$ and $32\%$ MACs savings respectively for a total speed-up factor of $1.34 \times$ and $1.29 \times$ . +- Finally, for the DeiT- and Swin-Tiny models, we apply a lower pruning rate of $45\%$ , achieving $97\%$ and $98\%$ final performance with $28\%$ and $25\%$ reductions in model size. This translates to $25\%$ and $28\%$ fewer MACs, along with speed-ups of $1.17 \times$ and $1.20 \times$ respectively. +- For the Base-Models, we demonstrate that when globally pruning the MLPs by $20\%$ the accuracy retention directly after the one-shot structured pruning reaches $99\%$ of the unpruned performance without any fine-tuning. This allows for an off-the-shelf $13\%$ size and MACs reduction and a speed-up of $1.11 \times$ for DeiT-Base and a similar reductions for Swin-Base + +All further pruning rates ranging from $5\%$ to $50\%$ for all DeiT Models can be found in Appendix B. + +Table 1. Results comparing the full baseline model with our VBP model using four metrics: MACs: computational operations, measured in billions of operations; and Parameters: the total model size in millions of parameters; Accuracy Retention (Ret.): retained accuracy after pruning, before fine-tuning; and Final Accuracy: accuracy after fine-tuning. Our method achieves competitive accuracy with significant reductions in MACs and parameters after fine-tuning, and even right after pruning, when reducing the pruning rate to $20\%$ . + +
Model (Pruning Rate)Tiny (45%)Small (50%)Base (55%)Base (20%)
DeiT1.171.341.441.11
Swin1.201.271.301.10
+ +Table 2. Speed-ups of the pruned models relative to the baseline for different sizes and pruning rates. + +
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
DeiT-T1.265.72-72.02
Magnitude0.913.943.5668.49
SNIP0.913.9424.8969.10
VBP (ours)0.913.9439.5870.61
DeiT-S4.6122.05-79.70
Magnitude3.2114.964.0576.55
SNIP3.2114.9652.3977.27
VBP (ours)3.2114.9664.4478.62
DeiT-B17.5886.57-81.73
Magnitude12.058.240.3778.88
SNIP12.058.2453.2480.40
VBP (ours)12.058.2466.4080.99
+ +Table 3. Comparison of different one-shot pruning scores in structured pruning to our VBP approach at fixed $50\%$ pruning rates. For constant MACs and parameters, our approach retains more accuracy (Ret.) and reaches a higher final accuracy compared to other pruning approaches. + +
ModelMACs (G)Param. (M)Speed-UpTop-1 Acc. (%)
Ret.Final
DeiT-B17.5886.57--81.73
ToMe-246.0286.572.1535.8575.74
ToMe-14 & VBP6.0858.242.0560.5380.09
ToMe-207.1486.571.8466.3178.66
ToMe-12 & VBP6.9058.241.8362.0980.29
ToMe-187.8886.571.7573.7979.97
ToMe-10 & VBP7.7358.241.7063.7280.45
+ +Table 4. Comparison of ToMe for different token reduction rates of 24, 20, and 18, with the hybrid combination of ToMe using token reduction rates of 14, 12, and 10, and VBP using $50\%$ pruning rates on top. Our hybrid method achieves significantly higher final accuracies throughout while also providing reductions in parameters, for similar MACs and speed-ups. The hybrid model thereby reaches speed-ups of up to $2.05 \times$ while maintaining competitive performance. + +
ModelEpochsParam. (M)Top-1 Acc. (%)
Ret.Final
NViT (50 Ep.)5056.3781.1382.18
NViT (1 Ep.)156.8569.1081.92
VBP (ours)156.1872.3782.32
+ +Table 5. Comparison with NViT using two pruning durations of 50 epochs (as originally proposed) and 1 epoch (similar to our VBP approach), both followed by 10 epochs of fine-tuning. Our method achieves higher accuracy immediately after pruning, before fine-tuning (Ret.) compared to NViT with the same pruning duration, and reaches a higher final accuracy than the 50-epoch NViT model after fine-tuning for 10 epochs. + +# 4.2. Performance Comparison + +To facilitate direct comparisons of all models with different sizes across all experiments, we fix the pruning rate at $50\%$ for all further experiments. + +We demonstrate the effectiveness of our selection criterion, by benchmarking against two other common and well-established one-shot-pruning baselines, namely Magnitude and SNIP, adapted for structured pruning. For a fixed pruning rate of $50\%$ using all selection criteria, this produces identical parameter counts of 3.94, 14.96, and 58.24 million parameters, as well as fixed 0.91, 3.21, and 12.00 GMACs for the three model sizes. Our results indicate that VBP provides higher accuracy retention compared to the baselines, improving DeiT-Base by $13.16\%$ , DeiT-Small by $12.05\%$ , and DeiT-Tiny by $14.69\%$ percentage points compared to SNIP. This difference in accuracy retention yields a final fine-tuned performance that is $0.59\%$ to $1.51\%$ percentage points higher after fine-tuning. + +# 4.3. SoTA Pruning Methods + +To further validate our approach, we compare against the SoTA structured transformer pruning method NViT (CVPR'23). Specifically, we apply our method to the NViT Base model (DeiT-Base distilled from RegNetY-160) and run NViT with a latency target of $70\%$ , pruning once for 50 epochs (as per the published results) and once for one epoch (similar to our setup), with both scenarios followed by ten epochs of fine-tuning using the NViT settings. In a similar pruning setting, VBP achieves a $0.4\%$ percentage points higher final accuracy, for similar number of parameters. + +Finally, we demonstrate the orthogonality of VBP to token-pruning methods by applying our method on top of the SoTA token-merging approach, ToMe (ICLR'23). We observe that the accuracy retention drops relative to the baseline remains consistent for both the VBP-pruned base DeiT model in Tab. 1 and the VBP- & ToMe-pruned model. This relative consistency also persists throughout fine-tuning, indicating that our method operates orthogonally to ToMe in terms of performance. + +Regarding the MAC savings it is important to note that ToMe progressively reduces the throughput of tokens processed across the layers. Later layers therefore process less tokens, making pruning less effective. The relative MAC savings from our hybrid method therefore decrease as the network depth increases. Nonetheless, VBP additionally reduces the total parameter count, which cannot be achieved by simply increasing the ToMe token reduction rate. + +In fact, increasing the reduction rate to get an approximate $2 \times$ speed-up yields similar MAC savings but reduces both the accuracy retention as well as the final accuracy significantly. Instead applying VBP on top of ToMe with a smaller reduction rate, allows more than $2 \times$ speed-ups while still maintaining $98\%$ of the original accuracy with additional parameter savings. + +
Variance- BasedMean- ShiftTop-1 Acc. (%)
Ret.Final
X✓55.1980.23
✓X26.0480.62
✓✓66.4080.99
+ +![](images/d08c1dddc8da297f901d0c2fee6bcc231b27ac79b4aae6ad3bdff18050085640.jpg) +Figure 2. Accuracy retention before fine-tuning and final accuracy after 10 epochs of fine-tuning for different pruning rates applied to DeiT-Base. + +![](images/15ff84dd38bb00e4ba02a35cb4ce4270f115ceacc4b68ea02cabcc7831246c15.jpg) +Figure 3. Activation variance for all neurons in the hidden MLP layers throughout the network and marked cumulative variances. + +# 5. Analysis + +To assess the contribution of each individual component of our approach, we perform an ablation study in which we systematically modify the pruning criterion and the Mean-Shift Compensation. We report our results in Tab. 6. + +Table 6. Ablation study results showing the impact of Variance-Based Pruning, and Mean-Shift Compensation on the retained accuracy after pruning (Ret.), as well as the final: accuracy after fine-tuning. Both steps together yield the highest accuracy of $80.99\%$ . + +
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
Pre-Act. (eq. [23])12.058.240.4377.92
Post-Act. (ours)12.058.2466.4080.99
+ +Table 7. Comparison of different locations for the activation statistics computation in the network: before (Pre-Act.) as well as after the activation function (Post-Act.). For the same MACs and parameters, using pruning rates of $50\%$ , gathering statistics post-activation significantly outperforms pre-activation in both accuracy retention (Ret.) as well as the final accuracy. + +We first remove the variance-based pruning selection while still shifting the mean activation to the next bias. Removing the Variance-Based Pruning alone reduces the final accuracy by $0.76\%$ percentage points. + +We also remove the Mean-Shift Compensation while pruning using the variance-based criterion. This improves the final accuracy by $0.39\%$ compared to using only the Mean-Shift Compensation. + +However, the best results are achieved when both Variance-Based Pruning and Mean-Shift Compensation are used together. As discussed in Sec. 3.2, replacing the removed activations with their means is optimal when pruning based on variance. When both components are combined, the immediate accuracy retention increases by $11.21\%$ percentage points, ultimately improving the fine-tuned accuracy by an additional $0.37\%$ . This highlights the significance of both components. + +# 5.1. Sensitivity of Variance Thresholding + +We further examine the stability of the activation variance criterion by applying different pruning rates to the DeiT-Base model, and plotting the accuracy retention and final accuracy in Fig. 2. We observe that for smaller pruning rates (up to around $25\%$ ), the accuracy retention is high enough to allow for off-the-shelf deployment without any fine-tuning. + +Beyond these rates, the accuracy begins to degrade more noticeably, reflecting the non-uniform distribution of neuron variances, as seen in Fig. 3. Notably, to account for $10\%$ of the cumulative variance across all neurons, nearly $60\%$ of the lowest-variance neurons are needed. This is a disproportionately large share compared to the next $15\%$ of neurons that also account for $10\%$ of the total variance in a layer. Consequently, more aggressive pruning leads to an increasingly faster reduction in the total variance and thus expressiveness of the network. Nevertheless, significant performance issues arise only at very high pruning levels, validating the feasibility of our method over a broad range of pruning rates (see Appendix B). + +
ModelMACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRetentionVBP
ConvNeXt-T4.472.96 (-33.8%)28.5912.61 (-55.9%)82.9016.80 (20.3%)81.30 (98.1%)
ConvNeXt-S8.715.11 (-41.3%)50.2223.50 (-53.2%)84.5730.92 (36.6%)82.82 (97.9%)
ConvNeXt-B15.388.91 (-42.1%)88.5941.32 (-53.4%)85.5157.10 (66.8%)83.40 (97.6%)
+ +Table 8. Results comparing the full ConvNeXt baseline models with the VBP models using a pruning rate of $50\%$ . The MACs are reduced by up to $42\%$ and parameters by over $50\%$ while reaching competitive final accuracies of $98\%$ original performance. + +
ModelTinySmallBase
ConvNeXt1.281.421.49
+ +Table 9. Speed-ups of the pruned models relative to the baseline for different-sized ConvNeXt models. + +# 5.2. Application Pre- vs. Post-Activation + +While Sec. 3.2 shows that our mean-replacement strategy can be grounded in theory for any distribution, the question remains whether the variance should be measured before or after the activation function. The Central Limit Theorem confirms that pre-activation sums (composed of independent weight contributions) tend to approximate a normal distribution. However, once the nonlinearity $\sigma$ is applied, this distribution changes. Consequently, the pre-activation variance does not necessarily correlate with the neuron importance for retaining accuracy in an already trained model. + +For instance, highly varying negative activations are compressed into range of about $(-0.2, 0)$ by GeLU, reducing their variance. As a result, such a neuron would be less likely to be pruned in a pre-activation setting and more likely to be pruned post-activation (see Appendix E). + +To validate this experimentally, we tested an alternative design in which pruning decisions are made based on the pre-activation variance in Tab. 7. While this approach has been successfully applied when pruning during training [23], we find that in a trained network setting, it significantly degrades performance. Therefore, measuring variance post-activation better captures neuron importance in already trained models, and helps maintain model performance despite structural modifications. + +# 5.3. Other Transformer-Like Architectures + +While our method is broadly applicable to any model containing MLP layers, network architectures that have more and larger MLPs benefit more from VBP. We therefore apply our pruning on ConvNeXt as a transformer-like architecture, that relies heavily on MLP layers, to further evaluate the generalizability of our method. + +We summarize our performance comparisons for ConvNeXt in Tab. 8. We achieve significant parameter reduc + +tions of over $50\%$ throughout all three sizes and reduced MACs by $34\%$ for ConvNeXt-Tiny and over $40\%$ for the larger sizes, while maintaining $98\%$ of the original accuracy after 10 epochs of fine-tuning across the board. The corresponding speed-ups are listed in Tab. 9, and reach up to $1.49\times$ for ConvNeXt-Base. + +Notably, we observe significantly less accuracy retention throughout the smaller sized models compared to the true transformers. We attribute this to the fact that the ConvNeXt architecture is built upon convolutions, which have significantly higher costs compared to their parameters. Consequently, since the MLPs take over a higher overall percentage of the total capacity, removing a similar fraction of neurons causes a relatively higher drop in accuracy retention. By contrast, transformers typically have less parameters in their MLP relatively speaking, thereby allowing for higher pruning rates while maintaining capacity. + +# 6. Conclusion + +We introduce VBP, a one-shot pruning method, designed to remove neurons based on their statistical importance in a single pruning operation while minimizing accuracy loss by integrating their mean activations in a Mean-Shift Compensation step. + +Our experiments demonstrate that VBP retains model performance significantly better than existing structured pruning methods. By taking advantage of trained networks, our approach provides a favourable starting point for the subsequent finetuning, eliminating the need for extensive retraining and thereby making it practical for low-cost deployment. + +Furthermore, we show that VBP operates orthogonally to SoTA token pruning methods, allowing it to be seamlessly combined with approaches like ToMe to achieve even greater computational savings. + +We believe that the simplicity and efficiency of our approach contributes to the democratization of deep learning by enabling a wider reuse of trained networks and hope it inspires further research in the CV community toward efficient model compression techniques. + +# References + +[1] Sajid Anwar, Kyuyeon Hwang, and Wonyong Sung. Structured pruning of deep convolutional neural networks. J. Emerg. Technol. Comput. Syst., 13(3), 2017. 1, 2, 3 +[2] Uranik Berisha, Jens Mehnert, and Alexandru Condurache. Efficient data driven mixture-of-expert extraction from trained networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3 +[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 3 +[4] Rinor Cakaj, Jens Mehnert, and Bin Yang. Cnn mixture-of-depths. In Computer Vision ACCV 2024: 17th Asian Conference on Computer Vision, Hanoi, Vietnam, December 812, 2024, Proceedings, Part VII, page 148166, Berlin, Heidelberg, 2024. Springer-Verlag. 3 +[5] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19974-19988, 2021. 3 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4, 3 +[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12114-12124. IEEE, 2022. 2 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 1, 2 +[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6804-6815. IEEE, 2021. 2 +[10] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Herve Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, + +Canada, October 10-17, 2021, pages 12239-12249. IEEE, 2021. 2 +[11] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4 +[12] Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pages 598-605. Morgan Kaufmann, 1989. 1, 3 +[13] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip H. S. Torr. Snip: single-shot network pruning based on connection sensitivity. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3 +[14] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 1, 2 +[15] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 1, 2 +[16] Lukas Meiner, Jens Mehnert, and Alexandru Condurache. Data-free dynamic compression of cnns for tractable efficiency. In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 2: VISAPP, pages 196-208. INSTICC, SciTePress, 2025. 3 +[17] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12299-12308. IEEE, 2022. 3 +[18] Dmitry Molchanov, Arsenii Ashukha, and Dmitry P. Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2498-2507. PMLR, 2017. 3 +[19] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 13937-13949, 2021. 1, 3 +[20] David Raposo, Samuel Ritter, Blake A. Richards, Timothy P. Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models. CoRR, abs/2404.02258, 2024.3 +[21] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference + +on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 10347-10357. PMLR, 2021. 1, 2 +[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2, 3 +[23] Berry Weinstein and Yonatan Belinkov. Variance pruning: Pruning language models via temporal neuron variance, 2025. Unpublished manuscript, OpenReview. 3, 7, 8 +[24] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 2 +[25] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Freezenet: Full performance by reduced storage costs. In Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, pages 685-701. Springer, 2020. 3 +[26] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Interspace pruning: Using adaptive filter representations to improve training of sparse cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12527-12537, 2022. 3 +[27] Huanrui Yang, Hongxu Yin, Maying Shen, Pavlo Molchanov, Hai Li, and Jan Kautz. Global vision transformer pruning with hessian-aware saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18547-18557. IEEE, 2023. 1, 3 +[28] Fang Yu, Kun Huang, Meng Wang, Yuan Cheng, Wei Chu, and Li Cui. Width & depth pruning for vision transformers. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 3143-3151. AAAI Press, 2022. +[29] Lu Yu and Wei Xiang. X-pruner: explainable pruning for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24355-24363. IEEE, 2023. 3 +[30] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 2, 3 +[31] Zhengyan Zhang, Zhiyuan Zeng, Yankai Lin, Chaojun Xiao, Xiaozhi Wang, Xu Han, Zhiyuan Liu, Ruobing Xie, Maosong Sun, and Jie Zhou. Emergent modularity in pretrained transformers. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4066-4083, Toronto, Canada, 2023. Association for Computational Linguistics. 3 + +[32] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning, 2021. 3 \ No newline at end of file diff --git a/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/images.zip b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..230c819e994fe50a5f12b1874e91b21dbad927ca --- /dev/null +++ b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e9e7b1d6b7b0ef20a84f3730bfa36e75aacd217723da709931b37c67eb7a45b +size 516377 diff --git a/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/layout.json b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e5e0e30242c01a495b6298afa30b896b1c9ded5 --- /dev/null +++ b/2025/Variance-Based Pruning for Accelerating and Compressing Trained Networks/layout.json @@ -0,0 +1,9304 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 547, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 547, + 120 + ], + "type": "text", + "content": "Variance-Based Pruning for Accelerating and Compressing Trained Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 468, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 468, + 157 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 468, + 157 + ], + "type": "text", + "content": "Uranik Berisha12, Jens Mehnert1 and Alexandru Paul Condurache12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 159, + 449, + 185 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 159, + 159, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 159, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 159, + 159, + 449, + 171 + ], + "type": "text", + "content": "1Automated Driving Research, Robert Bosch GmbH, 70469 Stuttgart, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 173, + 449, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 173, + 449, + 185 + ], + "spans": [ + { + "bbox": [ + 159, + 173, + 449, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 173, + 449, + 185 + ], + "type": "text", + "content": "Institute for Signal Processing, University of Lübeck, 23562 Lübeck, Germany" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 96, + 187, + 511, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 187, + 511, + 198 + ], + "spans": [ + { + "bbox": [ + 96, + 187, + 511, + 198 + ], + "type": "text", + "content": "{Uranik.Berisha,JensEricMarkus.Mehnert,AlexandruPaul.Condurache}@de.bosch.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 252, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 295, + 395 + ], + "type": "text", + "content": "Increasingly expensive training of ever larger models such as Vision Transformers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "text", + "content": "To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over " + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "text", + "content": " of its original performance and requires only 10 epochs of fine-tuning to regain " + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "text", + "content": " of the original accuracy while simultaneously reducing MACs by " + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "text", + "content": " and model size by " + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "text", + "content": ", thus speeding up the model by " + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "inline_equation", + "content": "1.44 \\times" + }, + { + "bbox": [ + 55, + 396, + 296, + 539 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "In both Computer Vision (CV) and Natural Language Processing (NLP), Transformers have become the dominant architecture [8, 21]. In CV, Vision Transformers (ViTs) [8] and their derivatives, such as Data-efficient image Transformers (DeiTs) [21], Swin-Transformers (Swins) [14] and ConvNeXts [15], achieve remarkable results but at significant costs in three main areas: training, storage, and inference. These large models require long training times, often spanning hundreds of epochs to achieve their performance, extensive memory to store their parameters, and high computational resources for inference causing high latency [21]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": "Fixing all three issues simultaneously remains especially challenging. For example, using the available collection of trained networks provided by the research community is a straightforward way to solve the first problem of costly training. However, this does not address the storage and inference costs, which can still make the use of these trained networks impractical for deployment on resource-constrained hardware [1]. To address these demands, various model pruning techniques have been proposed. These range from removing individual weights within layers in unstructured pruning [12] to removing entire neurons, layers, or blocks in structured pruning [1]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 373, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 553, + 445 + ], + "type": "text", + "content": "While unstructured pruning can often retain a high degree of accuracy in trained networks, it is difficult to translate these theoretical gains into real-world speed-ups [1]. Since modern hardware is optimized for dense matrix multiplications, they cannot exploit the resulting sparsity effectively and therefore cannot solve the inference costs [1, 12]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 448, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 555 + ], + "type": "text", + "content": "In contrast, structured pruning leads to straightforward reductions in both inference costs and memory, but again requires extensive retraining to recover the lost accuracy resulting from these major structural modifications. For instance, Global Vision Transformer Pruning with Hessian-Aware Saliency (NViT) [27] requires 300 epochs of fine-tuning despite using trained networks, highlighting how addressing memory and inference speeds alone can still demand large additional training overhead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "On the other hand, dynamic pruning approaches, such as token pruning, avoid permanently modifying the model architecture and instead dynamically reduce the number of features processed by the network [3, 19]. This minimizes the accuracy drop after pruning, indicating that the model preserved most important features and representation, thus providing a better starting point for subsequent fine-tuning. Token Merging (ToMe) [3] embraces this idea by merging tokens to reuse the pruned information, thus reducing the accuracy drop even further and allowing off-the-shelf deployment without any fine-tuning. While this solves the inference costs as well as the training costs, since the methods do not modify the model structure itself, the memory foot" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4973" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 73, + 201, + 242 + ], + "blocks": [ + { + "bbox": [ + 74, + 73, + 201, + 242 + ], + "lines": [ + { + "bbox": [ + 74, + 73, + 201, + 242 + ], + "spans": [ + { + "bbox": [ + 74, + 73, + 201, + 242 + ], + "type": "image", + "image_path": "b74940e48b256274fae1d8c185235d3ee92e28525b305283a24a1edea1d4b689.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 266, + 217, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 217, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 217, + 285 + ], + "type": "text", + "content": "(a) We compute neuron-wise mean and variance of activations at the hidden layer of each MLP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 229, + 72, + 373, + 254 + ], + "blocks": [ + { + "bbox": [ + 229, + 72, + 373, + 254 + ], + "lines": [ + { + "bbox": [ + 229, + 72, + 373, + 254 + ], + "spans": [ + { + "bbox": [ + 229, + 72, + 373, + 254 + ], + "type": "image", + "image_path": "9afc80bf2650b05c6d6475840a0765e1b7d3da968b26d255b997c3737b9d4311.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 266, + 385, + 285 + ], + "lines": [ + { + "bbox": [ + 223, + 266, + 385, + 285 + ], + "spans": [ + { + "bbox": [ + 223, + 266, + 385, + 285 + ], + "type": "text", + "content": "(b) We prune neurons with the smallest variances and replace their activation with the mean." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 395, + 76, + 542, + 257 + ], + "blocks": [ + { + "bbox": [ + 395, + 76, + 542, + 257 + ], + "lines": [ + { + "bbox": [ + 395, + 76, + 542, + 257 + ], + "spans": [ + { + "bbox": [ + 395, + 76, + 542, + 257 + ], + "type": "image", + "image_path": "9d396469f25068fe47f42ee8df55f899860e48c449ca37d7276b05d0b7cd0495.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 266, + 554, + 285 + ], + "lines": [ + { + "bbox": [ + 392, + 266, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 392, + 266, + 554, + 285 + ], + "type": "text", + "content": "(c) We shift the mean activation of the removed neurons into the bias of the next layer." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 295, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 295, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 555, + 319 + ], + "type": "text", + "content": "Figure 1. Schematic visualization of the steps of Variance-Based Pruning. (a) Activation Statistics Computation, (b) Variance-Based Pruning, (c) Mean-Shift Compensation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 338, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 295, + 387 + ], + "type": "text", + "content": "print remains unchanged [3]. This again is a limitation for hardware-constrained settings [1]. Fully resolving all three problems of training costs, storage, and inference speed simultaneously remains a challenge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 605 + ], + "type": "text", + "content": "In this work, we fill this gap by introducing Variance-Based Pruning (VBP), a simple and structured pruning approach that provides speed-ups and memory savings with minimal fine-tuning. In order to keep the method simple and allow deployment across a wide variety of architectures, we prune only the heavy Multi-Layer Perceptron (MLP) layers within the transformer blocks. In the first step, our approach leverages activation statistics gathered in the hidden layers of the MLP using Welford's algorithm [24] (see Fig. 1a). In the second step, these statistics are used to identify and remove the least impactful hidden neurons (see Fig. 1b). In the last step, which we refer to as Mean-Shift Compensation, we use the mean activation values of pruned neurons to mitigate accuracy drop by redistributing these mean contributions into the bias of the output layer (see Fig. 1c). This retains high levels of model performance and requires only minimal fine-tuning to then regain most of the original accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 254, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 254, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 254, + 627 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "content": "- We introduce VBP, a simple low-cost structured pruning method broadly applicable across various architectures to significantly reduce computational costs while maintaining performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "- We apply our method to multiple architectures, including ViTs, Swin-Transformers, and ConvNeXts, demonstrating its generalizability." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 338, + 555, + 482 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 338, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 555, + 434 + ], + "type": "text", + "content": "- We compare VBP to similarly straightforward neuron-importance measures adapted for structured pruning, as well as to state-of-the-art (SoTA) pruning approaches. We benchmark against NViT pruning, highlighting the benefits of our accuracy retention in constrained fine-tuning settings. We also apply VBP on top of ToMe demonstrating orthogonality and achieving " + }, + { + "bbox": [ + 314, + 338, + 555, + 434 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 314, + 338, + 555, + 434 + ], + "type": "text", + "content": " speed-ups using our hybrid approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 434, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 555, + 482 + ], + "type": "text", + "content": "- We perform an extensive ablation study, demonstrating the efficacy of each component of our method and provide a sensitivity analysis for the variance as a pruning criterion." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 502, + 429, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 429, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 429, + 514 + ], + "type": "text", + "content": "2. RELATED WORK" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 525, + 463, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 463, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 463, + 537 + ], + "type": "text", + "content": "2.1. Transformer Architectures" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 714 + ], + "type": "text", + "content": "Initially introduced in NLP by [22], Transformers have largely replaced Convolutional Neural Network (CNN), since their adaption to CV by [8] with the ViT. ViTs processes images as a sequence of non-overlapping patches, and have since been further improved. DeiT [21] refined the training procedure to decrease the training costs and reduce the data reliance. More recent architectures [7, 9, 10], such as Swin [14] and the transformer-inspired ConvNeXt [15], which use hierarchical designs, have built upon this architecture to improve efficiency and scalability. However, despite these advances, these large models remain computationally demanding, in part due to their dense MLP blocks, which significantly impacts inference speed and memory requirements [22, 30]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4974" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 172, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 172, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 172, + 84 + ], + "type": "text", + "content": "2.2. Model Compression" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 388 + ], + "type": "text", + "content": "To reduce the costs of large models, model compression efforts focus on scaling down different components of a network. Traditionally pruning aims to identify and remove redundant structures within the network [12, 26]. For example, magnitude pruning eliminates weights with the smallest absolute values, while SNIP and related methods [13, 25] identify crucial connections using gradient-based sensitivities. Structured pruning approaches, remove entire neurons, layers or filters producing smaller matrices that allow direct computational savings in current hardware [1, 5, 27-29, 32]. Most closely related to our method are works exploring variance-based criteria for structured pruning, such as Molchanov et al. [18], which performs dropout at the neuron level and can introduce sparsity in neurons with high dropout rates. Weinstein et al. [23] use these statistics in the initial training phase to remove neurons during training. These approaches operate on models before training, improving training time but still requiring the models to be trained from scratch. Contrary to that, NViT [27] utilizes trained ViTs. They apply iterative structured dimensionality reductions across the entire ViT architecture and significantly reduce costs. However, these structured modifications still have a substantial impact on performance, requiring high pruning iterations of 50 epochs and retraining costs of 300 epochs to regain accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 388, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 521 + ], + "type": "text", + "content": "An alternative to structurally modifying the network is dynamic pruning, which adaptively selects and processes only a subset of features or combines features during inference to reduce costs. For CNNs, these have shown promising speed-accuracy trade-offs [4, 16]. Analogously for transformer-like models, token pruning reduces the number of tokens to be processed [3, 17, 19, 20]. For instance, Bolya et al. [3] merges similar tokens rather than pruning them, thereby conserving most of the information and retaining high accuracy before the fine-tuning. However, by design these methods cannot reduce the memory footprint." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 147, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 147, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 147, + 541 + ], + "type": "text", + "content": "2.3. Our Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Our work builds upon these ideas by combining (i) the advantages of structured pruning by removing low-variance activation neurons with (ii) the advantages of dynamic pruning by incooperating their mean contribution into the bias of the last MLP layer. Our approach is based on the observation that neurons within MLP layers exhibit varying levels of contribution to the overall representation, as established in works on emerging modularity in trained Transformers for both NLP and CV [2, 31]. By applying our variance-based pruning approach to the widespread MLP layers, we provide a structured method applicable across a broad range of models. Our method is particularly beneficial for transformer-like architectures, where MLP layers heavily contribute to the computational overhead [22, 30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 555, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 139 + ], + "type": "text", + "content": "In this section, we detail the components involved in our proposed VBP approach and provide a mathematical derivation, justifying our pruning criterion and compensation strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 140, + 484, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 484, + 151 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 484, + 151 + ], + "type": "text", + "content": "Our method consists of three major steps:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 151, + 553, + 222 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 151, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 151, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 314, + 151, + 553, + 175 + ], + "type": "text", + "content": "1. Activation Statistics Computation The neuron activation statistics are computed using Welford's algorithm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 175, + 553, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 198 + ], + "type": "text", + "content": "2. Variance-Based Pruning The neurons to be pruned are selected based on the lowest activation variance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 198, + 553, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 222 + ], + "type": "text", + "content": "3. Mean-Shift Compensation The mean activations of the selected neurons are added back into the output bias." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 229, + 533, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 533, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 533, + 243 + ], + "type": "text", + "content": "3.1. Step 1: Activation Statistics Computation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "type": "text", + "content": "As we aim for a simple and widely adoptable method, we prune only the MLPs. This requires reducing the hidden layers of the network while keeping the input and output layers intact to ensure seamless integration with the rest of the model structure by maintaining dimensional consistency in the input and output." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "content": "W.l.o.g. we therefore consider a MLP with a single hidden layer of dimension " + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{hid}}" + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "content": ", that maps an input vector " + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{D_{\\mathrm{in}}}" + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "content": " to an output vector " + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^{D_{\\mathrm{out}}}" + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "content": ". Using a pointwise nonlinear activation function " + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "content": ", this MLP can then be described as a function" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 387, + 555, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 387, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 339, + 387, + 555, + 402 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\sigma \\left(\\mathbf {W} _ {1} \\mathbf {x} + \\mathbf {b} _ {1}\\right), \\quad \\mathbf {y} = \\mathbf {W} _ {2} \\mathbf {h} + \\mathbf {b} _ {2}, \\tag {1}", + "image_path": "aaeba908c42ece07fab03c8ea24b997ea867989a895c774913e41a4bb8189edb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1 \\in \\mathbb{R}^{D_{\\mathrm{hid}} \\times D_{\\mathrm{in}}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_1 \\in \\mathbb{R}^{D_{\\mathrm{hid}}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2 \\in \\mathbb{R}^{D_{\\mathrm{out}} \\times D_{\\mathrm{hid}}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_2 \\in \\mathbb{R}^{D_{\\mathrm{out}}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": ". Since we deal with trained networks, these parameters " + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1, \\mathbf{b}_1, \\mathbf{W}_2, \\mathbf{b}_2" + }, + { + "bbox": [ + 313, + 407, + 554, + 468 + ], + "type": "text", + "content": " remain fixed during pruning, which will become important in the Mean-Shift Compensation in Sec. 3.3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 492 + ], + "type": "text", + "content": "For the hidden layer, we now want to calculate the perneuron mean and variance vectors" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 498, + 548, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 498, + 548, + 516 + ], + "spans": [ + { + "bbox": [ + 318, + 498, + 548, + 516 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} = \\left(\\mu_ {1}, \\dots , \\mu_ {D _ {\\mathrm {h i d}}}\\right) ^ {\\top} \\quad \\text {a n d} \\quad \\boldsymbol {\\sigma} ^ {2} = \\left(\\sigma_ {1} ^ {2}, \\dots , \\sigma_ {D _ {\\mathrm {h i d}}} ^ {2}\\right) ^ {\\top}", + "image_path": "d1ed8ec13cee6a9b4108f0c79deb68e55a131ea15ce9fedba16bf0ab32af34d6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": "To compute these statistics, we capture the activations " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": " for every batch and use Welford's algorithm to compute a running mean " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{(j)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": " and running second moment " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_2^{(j)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": " for an efficient online calculation of streaming data*. Given a total of " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": " data samples, let the " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": "-th sample " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{h}^{(j)} \\in \\mathbb{R}^{D_{\\mathrm{hid}}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": " be the activation vector after the nonlinearity. Welford's algorithm ensures numerically stable calculations by updating " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{(j)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_2^{(j)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": " after observing the " + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 522, + 555, + 622 + ], + "type": "text", + "content": "-th sample as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 364, + 629, + 555, + 655 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 629, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 364, + 629, + 555, + 655 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {(j)} = \\frac {j - 1}{j} \\boldsymbol {\\mu} ^ {(j - 1)} + \\frac {1}{j} \\mathbf {h} ^ {(j)} \\tag {2}", + "image_path": "af1250c9dc1794bd373060219966c6ac73721e045fd91e701d3a35d743aaef14.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 319, + 662, + 553, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 662, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 319, + 662, + 553, + 678 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {2} ^ {(j)} = \\mathbf {m} _ {2} ^ {(j - 1)} + \\left(\\mathbf {h} ^ {(j)} - \\boldsymbol {\\mu} ^ {(j - 1)}\\right) \\odot \\left(\\mathbf {h} ^ {(j)} - \\boldsymbol {\\mu} ^ {(j)}\\right) \\tag {3}", + "image_path": "15589fe886c410cff7b0f8ccdf76fdf2ea0f3d90a9a2f11761c480b16a8ffec1.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 684, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 684, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 684, + 555, + 713 + ], + "type": "text", + "content": "*Our method does not require a minimum number of data samples for calculating activation statistics, but using an efficient online computation allows the use of as many data samples for pruning as available." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4975" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " denotes elementwise multiplication. Once all " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " samples are processed, the variance is given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 142, + 103, + 295, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 103, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 142, + 103, + 295, + 131 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\sigma} ^ {2} = \\frac {\\mathbf {m} _ {2} ^ {(N)}}{N - 1}. \\tag {4}", + "image_path": "81266b61a870e57e0f923825ba8e7d0312568a36251ca4cce019d563729c7af8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 230, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 230, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 230, + 151 + ], + "type": "text", + "content": "3.2. Step 2: Variance-Based Pruning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "text", + "content": "Using this variance vector " + }, + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "inline_equation", + "content": "\\sigma^2 \\in \\mathbb{R}^{D_{\\mathrm{hid}}}" + }, + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "text", + "content": ", we rank neurons by their variance values, " + }, + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "inline_equation", + "content": "\\sigma_i^2" + }, + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "text", + "content": ". Intuitively, if an activation rarely deviates from its mean " + }, + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "inline_equation", + "content": "\\mu_i" + }, + { + "bbox": [ + 55, + 154, + 296, + 238 + ], + "type": "text", + "content": ", the corresponding neuron contributes less to the expressiveness of the network. The pruning decision is formed across all hidden neurons in the network, to capture the neuron with the least variance across all layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "content": "Formally, for each MLP " + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{hid}}^{(l)}" + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "content": " denote the number of hidden neurons. For each neuron " + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "content": " in MLP " + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "content": ", we then simply select its activation variance " + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\sigma_{l,i}^2" + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "content": " as the corresponding pruning score. Given a pruning ratio " + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "inline_equation", + "content": "p \\in (0,1)" + }, + { + "bbox": [ + 55, + 239, + 296, + 300 + ], + "type": "text", + "content": ", we then only need to form the global set of scores" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 308, + 243, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 308, + 243, + 334 + ], + "spans": [ + { + "bbox": [ + 108, + 308, + 243, + 334 + ], + "type": "interline_equation", + "content": "S = \\bigcup_ {l \\in \\mathcal {L}} \\{\\sigma_ {l, i} ^ {2} \\mid i = 1, \\ldots , D _ {\\mathrm {h i d}} ^ {(l)} \\}.", + "image_path": "2c2504a5f0781af644b5018ab39efc5e3993110c00afd6af4a11fd429dad1ae8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 341, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 365 + ], + "type": "text", + "content": "from which we select the " + }, + { + "bbox": [ + 55, + 341, + 296, + 365 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 365 + ], + "type": "text", + "content": " neurons with the smallest scores for pruning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "content": "Optimality from a Mean-Replacement Perspective To retain the contribution of the pruned network, we approximate the activation of a pruned neuron " + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "content": " by replacing it with its mean " + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\mu_i" + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "content": ". This introduces an error proportional to " + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "inline_equation", + "content": "|h_i - \\mu_i|" + }, + { + "bbox": [ + 55, + 379, + 296, + 452 + ], + "type": "text", + "content": ", which, by definition of the sample variance for any distribution, has an expected value:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 92, + 461, + 258, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 461, + 258, + 476 + ], + "spans": [ + { + "bbox": [ + 92, + 461, + 258, + 476 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ (h _ {i} - \\mathbb {E} [ h _ {i} ]) ^ {2} ] = \\mathbb {E} [ (h _ {i} - \\mu_ {i}) ^ {2} ] = \\sigma_ {i} ^ {2}.", + "image_path": "2bf7d42aa6ee21f1d62df4aabbcefee1b2335e77f0f5ad8488648d8e40e9a192.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 540 + ], + "type": "text", + "content": "Thus, pruning the neurons with the lowest variance results in the least reconstruction error, making variance the optimal metric in this context. This is intuitively confirmed by the fact that when the variance is zero, all activations are exactly equal to the mean." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 547, + 239, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 239, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 239, + 560 + ], + "type": "text", + "content": "3.3. Step 3: Mean-Shift Compensation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "text", + "content": "The intuitive approach of removing neurons and replacing their activations with " + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mu_{i}^{\\dagger}" + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "text", + "content": " at inference time already reduces the output dimension of " + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1" + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "text", + "content": ", but the reconstruction of the original embedding dimension " + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "inline_equation", + "content": "D_{hid}" + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "text", + "content": " through replacement of the pruned activations still requires carrying out the full matrix multiplication for " + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2" + }, + { + "bbox": [ + 55, + 564, + 295, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "content": "Instead, we perform an equivalent transformation by shifting these mean values directly into the bias of the final linear layer " + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_2" + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "content": ", allowing us to reduce both the output dimension of " + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1" + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "content": " as well as the input dimension of " + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2" + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "content": ", doubling the cost savings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "We start from the usual two-layer linear mapping in the MLP block:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 397, + 96, + 471, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 96, + 471, + 109 + ], + "spans": [ + { + "bbox": [ + 397, + 96, + 471, + 109 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathbf {W} _ {2} \\mathbf {h} + \\mathbf {b} _ {2},", + "image_path": "cbfb9bd9b4a3352c5c66a584778295b070288189645bac23c97883cb84b921e9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "text", + "content": " is the hidden activation after pruning. For pruned neurons, we replace " + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "text", + "content": " with its mean " + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "inline_equation", + "content": "\\mu_{j}" + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "text", + "content": ", which we collect into a vector " + }, + { + "bbox": [ + 313, + 114, + 555, + 149 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mu}\\in \\mathbb{R}^{D_{\\mathrm{hid}}}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 158, + 485, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 158, + 485, + 190 + ], + "spans": [ + { + "bbox": [ + 383, + 158, + 485, + 190 + ], + "type": "interline_equation", + "content": "(\\Delta_ {\\mu}) _ {j} = \\left\\{ \\begin{array}{l l} \\mu_ {j}, & j \\in \\mathcal {P}, \\\\ 0, & j \\not \\in \\mathcal {P}, \\end{array} \\right.", + "image_path": "25130df2fbf07761752890ccd85bdcd5fb8ec02ddfe34768dfb9b41b99267fa5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 197, + 494, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 197, + 494, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 494, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 197, + 494, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 197, + 494, + 209 + ], + "type": "text", + "content": " is the set of pruned neuron indices." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 209, + 555, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 245 + ], + "type": "text", + "content": "Instead of explicitly inserting these means into " + }, + { + "bbox": [ + 313, + 209, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 313, + 209, + 555, + 245 + ], + "type": "text", + "content": " and then multiplying by " + }, + { + "bbox": [ + 313, + 209, + 555, + 245 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2" + }, + { + "bbox": [ + 313, + 209, + 555, + 245 + ], + "type": "text", + "content": ", we exploit the linearity of the mapping:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 358, + 253, + 510, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 253, + 510, + 268 + ], + "spans": [ + { + "bbox": [ + 358, + 253, + 510, + 268 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {2} \\mathbf {h} = \\mathbf {W} _ {2} \\left(\\mathbf {h} - \\Delta_ {\\mu}\\right) + \\mathbf {W} _ {2} \\Delta_ {\\mu}.", + "image_path": "d00309c655f0d01d89c97feea1ffc4372ffa2d01de9cd36facaf7bca2c0783c5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "type": "text", + "content": "The second term is a constant vector, so we can shift it into the bias:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 386, + 298, + 482, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 298, + 482, + 312 + ], + "spans": [ + { + "bbox": [ + 386, + 298, + 482, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {2} ^ {\\prime} = \\mathbf {b} _ {2} + \\mathbf {W} _ {2} \\Delta_ {\\mu},", + "image_path": "dbe93acdeb021fd129a84653ed599e57b57f3a4b72323796714de49e714d8f01.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 316, + 468, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 468, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 468, + 328 + ], + "type": "text", + "content": "which results in an equivalent output:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 373, + 336, + 555, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 336, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 373, + 336, + 555, + 350 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathbf {W} _ {2} (\\mathbf {h} - \\Delta_ {\\mu}) + \\mathbf {b} _ {2} ^ {\\prime}. \\tag {5}", + "image_path": "f2b88a1e79bcca6a83a0a0cc7787c9a4d504a36720cd0eaf7f085c05d8832a84.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": "Because for each pruned index " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": ", we have replaced the activations " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": " by their means " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "h_j = \\mu_j" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": ", it follows that " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{h} - \\Delta_{\\mu})_j = 0" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": ". Consequently, we can both drop the row " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": " as well as the column " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": " without affecting the computation, reducing the hidden dimension by " + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}|" + }, + { + "bbox": [ + 313, + 357, + 555, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 431, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 516 + ], + "type": "text", + "content": "Why Use Trained Networks? Our approach uses the fact that " + }, + { + "bbox": [ + 313, + 431, + 555, + 516 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1, \\mathbf{W}_2, \\mathbf{b}_1, \\mathbf{b}_2" + }, + { + "bbox": [ + 313, + 431, + 555, + 516 + ], + "type": "text", + "content": " are constants in trained networks. This allows the mean to be easily shifted into the bias of the next layer, which in turn allows the low-variance neurons to be removed without requiring extensive additional training. Only minor fine-tuning is sufficient to adjust the remaining weights for best performance." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 526, + 395, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 395, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 395, + 540 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 545, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 594 + ], + "type": "text", + "content": "In this section, we apply our VBP method to various transformers and similar architectures, namely DeiT, Swin, and ConvNeXt in the Tiny, Small, and Base sizes, all trained on the ImageNet-1k [6] dataset." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We evaluate these models twice. Once immediately after pruning (before fine-tuning) to measure how much accuracy is retained, and the second time after a brief fine-tuning to measure the final accuracy. Fine-tuning is performed over merely 10 epochs using knowledge-distillation from the unmodified base model. We use the AdamW optimizer with an initial learning rate of " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{e - 5}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", decayed by a cosine-annealing scheduler, and a batch size of 32. As regularization, we include a weight decay of 0.01. All experiments are performed on NVIDIA H200 Tensor Core GPUs." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 70, + 702, + 194, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 194, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 194, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 70, + 702, + 194, + 713 + ], + "type": "text", + "content": " We omit the MLP indices for clarity." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4976" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 222 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 222 + ], + "type": "table", + "html": "
Model\n(Pruning Rate)MACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRet.VBP
DeiT-T (45%)1.260.94 (-25.16%)5.724.12 (-27.97%)72.0249.77 (69.13%)70.08 (97.33%)
DeiT-S (50%)4.613.21 (-30.37%)22.0514.96 (-32.15%)79.7064.44 (80.85%)78.62 (98.64%)
DeiT-B (55%)17.5811.44 (-34.93%)86.5755.40 (-36.01%)81.7357.58 (70.48%)80.67 (98.74%)
DeiT-B (20%)17.5815.35 (-12.68%)86.5775.24 (-13.09%)81.7380.87 (98.98%)81.76 (100.07%)
Swin-T (45%)4.503.23 (-28.22%)28.2921.31 (-24.67%)80.9155.12 (68.13%)79.41 (98.15%)
Swin-S (50%)8.765.63 (-32.19%)49.6135.02 (-29.41%)83.0467.01 (80.70%)81.86 (98.58%)
Swin-B (55%)15.4610.22 (-33.89%)87.7756.29 (-35.87%)84.7162.61 (73.91%)83.61 (98.70%)
Swin-B (20%)15.4613.58 (-12.16%)87.7776.42 (-12.93%)84.7183.90 (99.04%)84.67 (99.95%)
", + "image_path": "cb5dc2e2fafd33afe645ab902ba27f473643b888c39626de46517ccf339cf498.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 296, + 141, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 141, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 141, + 308 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 399 + ], + "type": "text", + "content": "As is common with various pruning techniques, the pruning rate is adapted for the size of the models. Larger models typically exhibit higher redundancy for similar tasks, and therefore allow for more pruning while maintaining accuracy. We report our results for DeiT and Swin in Tab. 1 and Tab. 2. Note that the pruning rate is applied to the MLPs only and does not indicate the total reduction in model size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 295, + 687 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": "- For the base-sized models, we prune all MLPs globally using a pruning rate of " + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": " keeping " + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": " of the original performance after fine-tuning, reducing the total model size of DeiT-Base, Swin-Base both by " + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": ". This corresponds to " + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": " fewer MACs respectively and provides a speed-up of " + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "inline_equation", + "content": "1.44 \\times" + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "inline_equation", + "content": "1.30 \\times" + }, + { + "bbox": [ + 55, + 400, + 295, + 472 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "- The small-sized models have notably fewer parameters than the base models and are thus pruned " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " again reaching " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " of the original performance. This leads to a DeiT-Small size reduction of " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " and Swin-Small reduction of " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " MACs savings respectively for a total speed-up factor of " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "1.34 \\times" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "1.29 \\times" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": "- Finally, for the DeiT- and Swin-Tiny models, we apply a lower pruning rate of " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": ", achieving " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " final performance with " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " reductions in model size. This translates to " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " fewer MACs, along with speed-ups of " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "1.17 \\times" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "inline_equation", + "content": "1.20 \\times" + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "text", + "content": "- For the Base-Models, we demonstrate that when globally pruning the MLPs by " + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "text", + "content": " the accuracy retention directly after the one-shot structured pruning reaches " + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "text", + "content": " of the unpruned performance without any fine-tuning. This allows for an off-the-shelf " + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "text", + "content": " size and MACs reduction and a speed-up of " + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "inline_equation", + "content": "1.11 \\times" + }, + { + "bbox": [ + 55, + 604, + 295, + 687 + ], + "type": "text", + "content": " for DeiT-Base and a similar reductions for Swin-Base" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "All further pruning rates ranging from " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " for all DeiT Models can be found in Appendix B." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 294, + 555, + 356 + ], + "blocks": [ + { + "bbox": [ + 55, + 231, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 277 + ], + "type": "text", + "content": "Table 1. Results comparing the full baseline model with our VBP model using four metrics: MACs: computational operations, measured in billions of operations; and Parameters: the total model size in millions of parameters; Accuracy Retention (Ret.): retained accuracy after pruning, before fine-tuning; and Final Accuracy: accuracy after fine-tuning. Our method achieves competitive accuracy with significant reductions in MACs and parameters after fine-tuning, and even right after pruning, when reducing the pruning rate to " + }, + { + "bbox": [ + 55, + 231, + 555, + 277 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 231, + 555, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 294, + 555, + 356 + ], + "lines": [ + { + "bbox": [ + 315, + 294, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 315, + 294, + 555, + 356 + ], + "type": "table", + "html": "
Model (Pruning Rate)Tiny (45%)Small (50%)Base (55%)Base (20%)
DeiT1.171.341.441.11
Swin1.201.271.301.10
", + "image_path": "3ce4f293e7f4237657a24787d695087662a3350a3a6d89c59ecb7edead085976.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 414, + 555, + 610 + ], + "blocks": [ + { + "bbox": [ + 313, + 363, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 387 + ], + "type": "text", + "content": "Table 2. Speed-ups of the pruned models relative to the baseline for different sizes and pruning rates." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 414, + 555, + 610 + ], + "lines": [ + { + "bbox": [ + 315, + 414, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 315, + 414, + 555, + 610 + ], + "type": "table", + "html": "
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
DeiT-T1.265.72-72.02
Magnitude0.913.943.5668.49
SNIP0.913.9424.8969.10
VBP (ours)0.913.9439.5870.61
DeiT-S4.6122.05-79.70
Magnitude3.2114.964.0576.55
SNIP3.2114.9652.3977.27
VBP (ours)3.2114.9664.4478.62
DeiT-B17.5886.57-81.73
Magnitude12.058.240.3778.88
SNIP12.058.2453.2480.40
VBP (ours)12.058.2466.4080.99
", + "image_path": "83b6014170b97c8fc2e1be0c96df33855910f4a40966df8c54750da04be32383.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 674 + ], + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 674 + ], + "type": "text", + "content": "Table 3. Comparison of different one-shot pruning scores in structured pruning to our VBP approach at fixed " + }, + { + "bbox": [ + 313, + 617, + 555, + 674 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 617, + 555, + 674 + ], + "type": "text", + "content": " pruning rates. For constant MACs and parameters, our approach retains more accuracy (Ret.) and reaches a higher final accuracy compared to other pruning approaches." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4977" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 555, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 205 + ], + "type": "table", + "html": "
ModelMACs (G)Param. (M)Speed-UpTop-1 Acc. (%)
Ret.Final
DeiT-B17.5886.57--81.73
ToMe-246.0286.572.1535.8575.74
ToMe-14 & VBP6.0858.242.0560.5380.09
ToMe-207.1486.571.8466.3178.66
ToMe-12 & VBP6.9058.241.8362.0980.29
ToMe-187.8886.571.7573.7979.97
ToMe-10 & VBP7.7358.241.7063.7280.45
", + "image_path": "5f117d4cf1e81b805d40108f11653dad6db933d136c9f29313d70ad3d0a983c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 277, + 296, + 360 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": "Table 4. Comparison of ToMe for different token reduction rates of 24, 20, and 18, with the hybrid combination of ToMe using token reduction rates of 14, 12, and 10, and VBP using " + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": " pruning rates on top. Our hybrid method achieves significantly higher final accuracies throughout while also providing reductions in parameters, for similar MACs and speed-ups. The hybrid model thereby reaches speed-ups of up to " + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "inline_equation", + "content": "2.05 \\times" + }, + { + "bbox": [ + 55, + 213, + 555, + 258 + ], + "type": "text", + "content": " while maintaining competitive performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 277, + 296, + 360 + ], + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 360 + ], + "type": "table", + "html": "
ModelEpochsParam. (M)Top-1 Acc. (%)
Ret.Final
NViT (50 Ep.)5056.3781.1382.18
NViT (1 Ep.)156.8569.1081.92
VBP (ours)156.1872.3782.32
", + "image_path": "5b1352b9605cb17c1777c1d68e1b65f41bc8f6e3f65bb49ea8da403de70f2c92.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 368, + 295, + 445 + ], + "lines": [ + { + "bbox": [ + 55, + 368, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 295, + 445 + ], + "type": "text", + "content": "Table 5. Comparison with NViT using two pruning durations of 50 epochs (as originally proposed) and 1 epoch (similar to our VBP approach), both followed by 10 epochs of fine-tuning. Our method achieves higher accuracy immediately after pruning, before fine-tuning (Ret.) compared to NViT with the same pruning duration, and reaches a higher final accuracy than the 50-epoch NViT model after fine-tuning for 10 epochs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 479, + 200, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 200, + 493 + ], + "type": "text", + "content": "4.2. Performance Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "type": "text", + "content": "To facilitate direct comparisons of all models with different sizes across all experiments, we fix the pruning rate at " + }, + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 503, + 295, + 540 + ], + "type": "text", + "content": " for all further experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "We demonstrate the effectiveness of our selection criterion, by benchmarking against two other common and well-established one-shot-pruning baselines, namely Magnitude and SNIP, adapted for structured pruning. For a fixed pruning rate of " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " using all selection criteria, this produces identical parameter counts of 3.94, 14.96, and 58.24 million parameters, as well as fixed 0.91, 3.21, and 12.00 GMACs for the three model sizes. Our results indicate that VBP provides higher accuracy retention compared to the baselines, improving DeiT-Base by " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "13.16\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", DeiT-Small by " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "12.05\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", and DeiT-Tiny by " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "14.69\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " percentage points compared to SNIP. This difference in accuracy retention yields a final fine-tuned performance that is " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.59\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.51\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " percentage points higher after fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 278, + 448, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 448, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 448, + 292 + ], + "type": "text", + "content": "4.3. SoTA Pruning Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "text", + "content": "To further validate our approach, we compare against the SoTA structured transformer pruning method NViT (CVPR'23). Specifically, we apply our method to the NViT Base model (DeiT-Base distilled from RegNetY-160) and run NViT with a latency target of " + }, + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "text", + "content": ", pruning once for 50 epochs (as per the published results) and once for one epoch (similar to our setup), with both scenarios followed by ten epochs of fine-tuning using the NViT settings. In a similar pruning setting, VBP achieves a " + }, + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 313, + 299, + 555, + 418 + ], + "type": "text", + "content": " percentage points higher final accuracy, for similar number of parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 421, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 529 + ], + "type": "text", + "content": "Finally, we demonstrate the orthogonality of VBP to token-pruning methods by applying our method on top of the SoTA token-merging approach, ToMe (ICLR'23). We observe that the accuracy retention drops relative to the baseline remains consistent for both the VBP-pruned base DeiT model in Tab. 1 and the VBP- & ToMe-pruned model. This relative consistency also persists throughout fine-tuning, indicating that our method operates orthogonally to ToMe in terms of performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 531, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 628 + ], + "type": "text", + "content": "Regarding the MAC savings it is important to note that ToMe progressively reduces the throughput of tokens processed across the layers. Later layers therefore process less tokens, making pruning less effective. The relative MAC savings from our hybrid method therefore decrease as the network depth increases. Nonetheless, VBP additionally reduces the total parameter count, which cannot be achieved by simply increasing the ToMe token reduction rate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "In fact, increasing the reduction rate to get an approximate " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " speed-up yields similar MAC savings but reduces both the accuracy retention as well as the final accuracy significantly. Instead applying VBP on top of ToMe with a smaller reduction rate, allows more than " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " speed-ups while still maintaining " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " of the original accuracy with additional parameter savings." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4978" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 295, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 148 + ], + "type": "table", + "html": "
Variance- BasedMean- ShiftTop-1 Acc. (%)
Ret.Final
X✓55.1980.23
✓X26.0480.62
✓✓66.4080.99
", + "image_path": "24725868ef51692c39d48ca35bea96b2c0aff464b75f33b0c8582b3bfdbd94ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 224, + 295, + 381 + ], + "blocks": [ + { + "bbox": [ + 56, + 224, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 224, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 295, + 381 + ], + "type": "image", + "image_path": "d08c1dddc8da297f901d0c2fee6bcc231b27ac79b4aae6ad3bdff18050085640.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 389, + 295, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 422 + ], + "type": "text", + "content": "Figure 2. Accuracy retention before fine-tuning and final accuracy after 10 epochs of fine-tuning for different pruning rates applied to DeiT-Base." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 55, + 435, + 295, + 593 + ], + "blocks": [ + { + "bbox": [ + 55, + 435, + 295, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 593 + ], + "type": "image", + "image_path": "15ff84dd38bb00e4ba02a35cb4ce4270f115ceacc4b68ea02cabcc7831246c15.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "type": "text", + "content": "Figure 3. Activation variance for all neurons in the hidden MLP layers throughout the network and marked cumulative variances." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 644, + 115, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 115, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 115, + 658 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": "To assess the contribution of each individual component of our approach, we perform an ablation study in which we systematically modify the pruning criterion and the Mean-Shift Compensation. We report our results in Tab. 6." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 137 + ], + "blocks": [ + { + "bbox": [ + 55, + 156, + 295, + 211 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 211 + ], + "type": "text", + "content": "Table 6. Ablation study results showing the impact of Variance-Based Pruning, and Mean-Shift Compensation on the retained accuracy after pruning (Ret.), as well as the final: accuracy after fine-tuning. Both steps together yield the highest accuracy of " + }, + { + "bbox": [ + 55, + 156, + 295, + 211 + ], + "type": "inline_equation", + "content": "80.99\\%" + }, + { + "bbox": [ + 55, + 156, + 295, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 554, + 137 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 137 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 137 + ], + "type": "table", + "html": "
ModelMACs (G)Param. (M)Top-1 Acc. (%)
Ret.Final
Pre-Act. (eq. [23])12.058.240.4377.92
Post-Act. (ours)12.058.2466.4080.99
", + "image_path": "c062d431d17c732ddc94c20e789db90455d5873ae82fb96ce2e009cfb98260f2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 554, + 211 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 554, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 554, + 211 + ], + "type": "text", + "content": "Table 7. Comparison of different locations for the activation statistics computation in the network: before (Pre-Act.) as well as after the activation function (Post-Act.). For the same MACs and parameters, using pruning rates of " + }, + { + "bbox": [ + 313, + 144, + 554, + 211 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 144, + 554, + 211 + ], + "type": "text", + "content": ", gathering statistics post-activation significantly outperforms pre-activation in both accuracy retention (Ret.) as well as the final accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 237, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 553, + 285 + ], + "type": "text", + "content": "We first remove the variance-based pruning selection while still shifting the mean activation to the next bias. Removing the Variance-Based Pruning alone reduces the final accuracy by " + }, + { + "bbox": [ + 313, + 237, + 553, + 285 + ], + "type": "inline_equation", + "content": "0.76\\%" + }, + { + "bbox": [ + 313, + 237, + 553, + 285 + ], + "type": "text", + "content": " percentage points." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "text", + "content": "We also remove the Mean-Shift Compensation while pruning using the variance-based criterion. This improves the final accuracy by " + }, + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "inline_equation", + "content": "0.39\\%" + }, + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "text", + "content": " compared to using only the Mean-Shift Compensation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": "However, the best results are achieved when both Variance-Based Pruning and Mean-Shift Compensation are used together. As discussed in Sec. 3.2, replacing the removed activations with their means is optimal when pruning based on variance. When both components are combined, the immediate accuracy retention increases by " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "11.21\\%" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": " percentage points, ultimately improving the fine-tuned accuracy by an additional " + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "inline_equation", + "content": "0.37\\%" + }, + { + "bbox": [ + 313, + 338, + 555, + 447 + ], + "type": "text", + "content": ". This highlights the significance of both components." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 463, + 508, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 508, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 508, + 476 + ], + "type": "text", + "content": "5.1. Sensitivity of Variance Thresholding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "text", + "content": "We further examine the stability of the activation variance criterion by applying different pruning rates to the DeiT-Base model, and plotting the accuracy retention and final accuracy in Fig. 2. We observe that for smaller pruning rates (up to around " + }, + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 483, + 554, + 555 + ], + "type": "text", + "content": "), the accuracy retention is high enough to allow for off-the-shelf deployment without any fine-tuning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Beyond these rates, the accuracy begins to degrade more noticeably, reflecting the non-uniform distribution of neuron variances, as seen in Fig. 3. Notably, to account for " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of the cumulative variance across all neurons, nearly " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of the lowest-variance neurons are needed. This is a disproportionately large share compared to the next " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of neurons that also account for " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of the total variance in a layer. Consequently, more aggressive pruning leads to an increasingly faster reduction in the total variance and thus expressiveness of the network. Nevertheless, significant performance issues arise only at very high pruning levels, validating the feasibility of our method over a broad range of pruning rates (see Appendix B)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4979" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 148 + ], + "type": "table", + "html": "
ModelMACs (G)Parameters (M)Top-1 Acc. (%)
FullVBPFullVBPFullRetentionVBP
ConvNeXt-T4.472.96 (-33.8%)28.5912.61 (-55.9%)82.9016.80 (20.3%)81.30 (98.1%)
ConvNeXt-S8.715.11 (-41.3%)50.2223.50 (-53.2%)84.5730.92 (36.6%)82.82 (97.9%)
ConvNeXt-B15.388.91 (-42.1%)88.5941.32 (-53.4%)85.5157.10 (66.8%)83.40 (97.6%)
", + "image_path": "4c3862ebc3699c761e63de9b172cf23ed0038d2faac53ffd8c6e002dea301248.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 198, + 296, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "text", + "content": "Table 8. Results comparing the full ConvNeXt baseline models with the VBP models using a pruning rate of " + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "text", + "content": ". The MACs are reduced by up to " + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "text", + "content": " and parameters by over " + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "text", + "content": " while reaching competitive final accuracies of " + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 55, + 156, + 555, + 179 + ], + "type": "text", + "content": " original performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 198, + 296, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 235 + ], + "type": "table", + "html": "
ModelTinySmallBase
ConvNeXt1.281.421.49
", + "image_path": "6e4f548a3fc4e6c87f030c1bb019aa970be5caa2cbc230a861b30835ae118a59.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 243, + 295, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 266 + ], + "type": "text", + "content": "Table 9. Speed-ups of the pruned models relative to the baseline for different-sized ConvNeXt models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 286, + 247, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 247, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 247, + 300 + ], + "type": "text", + "content": "5.2. Application Pre- vs. Post-Activation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 304, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 424 + ], + "type": "text", + "content": "While Sec. 3.2 shows that our mean-replacement strategy can be grounded in theory for any distribution, the question remains whether the variance should be measured before or after the activation function. The Central Limit Theorem confirms that pre-activation sums (composed of independent weight contributions) tend to approximate a normal distribution. However, once the nonlinearity " + }, + { + "bbox": [ + 54, + 304, + 295, + 424 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 304, + 295, + 424 + ], + "type": "text", + "content": " is applied, this distribution changes. Consequently, the pre-activation variance does not necessarily correlate with the neuron importance for retaining accuracy in an already trained model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 424, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 295, + 484 + ], + "type": "text", + "content": "For instance, highly varying negative activations are compressed into range of about " + }, + { + "bbox": [ + 54, + 424, + 295, + 484 + ], + "type": "inline_equation", + "content": "(-0.2, 0)" + }, + { + "bbox": [ + 54, + 424, + 295, + 484 + ], + "type": "text", + "content": " by GeLU, reducing their variance. As a result, such a neuron would be less likely to be pruned in a pre-activation setting and more likely to be pruned post-activation (see Appendix E)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 484, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 295, + 592 + ], + "type": "text", + "content": "To validate this experimentally, we tested an alternative design in which pruning decisions are made based on the pre-activation variance in Tab. 7. While this approach has been successfully applied when pruning during training [23], we find that in a trained network setting, it significantly degrades performance. Therefore, measuring variance post-activation better captures neuron importance in already trained models, and helps maintain model performance despite structural modifications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 259, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 259, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 259, + 612 + ], + "type": "text", + "content": "5.3. Other Transformer-Like Architectures" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 689 + ], + "type": "text", + "content": "While our method is broadly applicable to any model containing MLP layers, network architectures that have more and larger MLPs benefit more from VBP. We therefore apply our pruning on ConvNeXt as a transformer-like architecture, that relies heavily on MLP layers, to further evaluate the generalizability of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We summarize our performance comparisons for ConvNeXt in Tab. 8. We achieve significant parameter reduc" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "text", + "content": "tions of over " + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "text", + "content": " throughout all three sizes and reduced MACs by " + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "text", + "content": " for ConvNeXt-Tiny and over " + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "text", + "content": " for the larger sizes, while maintaining " + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "text", + "content": " of the original accuracy after 10 epochs of fine-tuning across the board. The corresponding speed-ups are listed in Tab. 9, and reach up to " + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "inline_equation", + "content": "1.49\\times" + }, + { + "bbox": [ + 313, + 200, + 555, + 272 + ], + "type": "text", + "content": " for ConvNeXt-Base." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 275, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 556, + 407 + ], + "type": "text", + "content": "Notably, we observe significantly less accuracy retention throughout the smaller sized models compared to the true transformers. We attribute this to the fact that the ConvNeXt architecture is built upon convolutions, which have significantly higher costs compared to their parameters. Consequently, since the MLPs take over a higher overall percentage of the total capacity, removing a similar fraction of neurons causes a relatively higher drop in accuracy retention. By contrast, transformers typically have less parameters in their MLP relatively speaking, thereby allowing for higher pruning rates while maintaining capacity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 388, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 512 + ], + "type": "text", + "content": "We introduce VBP, a one-shot pruning method, designed to remove neurons based on their statistical importance in a single pruning operation while minimizing accuracy loss by integrating their mean activations in a Mean-Shift Compensation step." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 514, + 555, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 599 + ], + "type": "text", + "content": "Our experiments demonstrate that VBP retains model performance significantly better than existing structured pruning methods. By taking advantage of trained networks, our approach provides a favourable starting point for the subsequent finetuning, eliminating the need for extensive retraining and thereby making it practical for low-cost deployment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 602, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 555, + 651 + ], + "type": "text", + "content": "Furthermore, we show that VBP operates orthogonally to SoTA token pruning methods, allowing it to be seamlessly combined with approaches like ToMe to achieve even greater computational savings." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "We believe that the simplicity and efficiency of our approach contributes to the democratization of deep learning by enabling a wider reuse of trained networks and hope it inspires further research in the CV community toward efficient model compression techniques." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4980" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Sajid Anwar, Kyuyeon Hwang, and Wonyong Sung. Structured pruning of deep convolutional neural networks. J. Emerg. Technol. Comput. Syst., 13(3), 2017. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 294, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 294, + 178 + ], + "type": "text", + "content": "[2] Uranik Berisha, Jens Mehnert, and Alexandru Condurache. Efficient data driven mixture-of-expert extraction from trained networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 294, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 294, + 245 + ], + "type": "text", + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 303 + ], + "type": "text", + "content": "[4] Rinor Cakaj, Jens Mehnert, and Bin Yang. Cnn mixture-of-depths. In Computer Vision ACCV 2024: 17th Asian Conference on Computer Vision, Hanoi, Vietnam, December 812, 2024, Proceedings, Part VII, page 148166, Berlin, Heidelberg, 2024. Springer-Verlag. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 378 + ], + "type": "text", + "content": "[5] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19974-19988, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 381, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 294, + 424 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 4, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 426, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 294, + 501 + ], + "type": "text", + "content": "[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12114-12124. IEEE, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 503, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 294, + 590 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 592, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 294, + 656 + ], + "type": "text", + "content": "[9] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 6804-6815. IEEE, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[10] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Herve Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Canada, October 10-17, 2021, pages 12239-12249. IEEE, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "text", + "content": "[11] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[12] Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. In Advances in Neural Information Processing Systems 2, [NIPS Conference, Denver, Colorado, USA, November 27-30, 1989], pages 598-605. Morgan Kaufmann, 1989. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 228 + ], + "type": "text", + "content": "[13] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip H. S. Torr. Snip: single-shot network pruning based on connection sensitivity. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 229, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 229, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 229, + 553, + 293 + ], + "type": "text", + "content": "[14] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9992-10002. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[15] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "text", + "content": "[16] Lukas Meiner, Jens Mehnert, and Alexandru Condurache. Data-free dynamic compression of cnns for tractable efficiency. In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 2: VISAPP, pages 196-208. INSTICC, SciTePress, 2025. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "type": "text", + "content": "[17] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12299-12308. IEEE, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 483, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 536 + ], + "type": "text", + "content": "[18] Dmitry Molchanov, Arsenii Ashukha, and Dmitry P. Vetrov. Variational dropout sparsifies deep neural networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, pages 2498-2507. PMLR, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 537, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 613 + ], + "type": "text", + "content": "[19] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 13937-13949, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 615, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 667 + ], + "type": "text", + "content": "[20] David Raposo, Samuel Ritter, Blake A. Richards, Timothy P. Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models. CoRR, abs/2404.02258, 2024.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[21] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Proceedings of the 38th International Conference" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4981" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 10347-10357. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 173 + ], + "type": "text", + "content": "[22] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 207 + ], + "type": "text", + "content": "[23] Berry Weinstein and Yonatan Belinkov. Variance pruning: Pruning language models via temporal neuron variance, 2025. Unpublished manuscript, OpenReview. 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 240 + ], + "type": "text", + "content": "[24] B. P. Welford. Note on a method for calculating corrected sums of squares and products. Technometrics, 4(3):419-420, 1962. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 308 + ], + "type": "text", + "content": "[25] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Freezenet: Full performance by reduced storage costs. In Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, pages 685-701. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 311, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 311, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 296, + 366 + ], + "type": "text", + "content": "[26] Paul Wimmer, Jens Mehnert, and Alexandru Condurache. Interspace pruning: Using adaptive filter representations to improve training of sparse cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12527-12537, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 368, + 295, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 368, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 368, + 295, + 433 + ], + "type": "text", + "content": "[27] Huanrui Yang, Hongxu Yin, Maying Shen, Pavlo Molchanov, Hai Li, and Jan Kautz. Global vision transformer pruning with hessian-aware saliency. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18547-18557. IEEE, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 521 + ], + "type": "text", + "content": "[28] Fang Yu, Kun Huang, Meng Wang, Yuan Cheng, Wei Chu, and Li Cui. Width & depth pruning for vision transformers. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 3143-3151. AAAI Press, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 524, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 295, + 578 + ], + "type": "text", + "content": "[29] Lu Yu and Wei Xiang. X-pruner: explainable pruning for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24355-24363. IEEE, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "text", + "content": "[30] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 1204-1213. IEEE, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 637, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 295, + 712 + ], + "type": "text", + "content": "[31] Zhengyan Zhang, Zhiyuan Zeng, Yankai Lin, Chaojun Xiao, Xiaozhi Wang, Xu Han, Zhiyuan Liu, Ruobing Xie, Maosong Sun, and Jie Zhou. Emergent modularity in pretrained transformers. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4066-4083, Toronto, Canada, 2023. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "text", + "content": "[32] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4982" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_content_list.json b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc3d6d9066c9ee69fbab8c65ee6371811dc124ca --- /dev/null +++ b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_content_list.json @@ -0,0 +1,1471 @@ +[ + { + "type": "text", + "text": "Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision", + "text_level": 1, + "bbox": [ + 125, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuting He1 Shuo Li1,2,* \n1 Dept. of BME, 2 Dept. of CDS, Case Western Reserve University, US, Cleveland \n{yuting.he4, shuo.li11}@case.edu", + "bbox": [ + 169, + 179, + 823, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive learning (CL) has become a cornerstone of self-supervised pretraining (SSP) in foundation models; however, extending CL to pixel-wise representation—crucial for medical vision—remains an open problem. Standard CL formulates SSP as a binary optimization problem (binary CL) where the excessive pursuit of feature dispersion leads to an \"over-dispersion\" problem, breaking pixelwise feature correlation thus disrupting the intra-class distribution. Our vector CL reformulates CL as a vector regression problem, enabling dispersion quantification in pixel-wise pretraining via modeling feature distances in regressing displacement vectors. To implement this novel paradigm, we propose the COntrast in Vector Regression (COVER) framework. COVER establishes an extendable vector-based self-learning, enforces a consistent optimization flow from vector regression to distance modeling, and leverages a vector pyramid architecture for granularity adaptation, thus preserving pixel-wise feature correlations in SSP. Extensive experiments across 8 tasks, spanning 2 dimensions and 4 modalities, show that COVER significantly improves pixel-wise SSP, advancing generalizable medical visual foundation models. Codes will be publicly available at https://github.com/YutingHe-list/COVER.", + "bbox": [ + 88, + 301, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 680, + 220, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive learning (CL) [4, 7-9, 14, 37, 50] has become a cornerstone of self-supervised pretraining (SSP) [30] in medical visual foundation models (MVFMs) [6, 21, 33, 55, 62, 71]. However, applying CL in a pixel-wise manner, an approach pivotal for disentangling the clinically critical semantics within medical images [70], remains an open problem. Current preliminary works [6, 55, 62] adopt imagewise contrast strategies derived from natural image studies [7-9, 14, 37], which emphasize global representations [60]. In medical images, the global context is often similar due to the consistent anatomies in human body [34], lead", + "bbox": [ + 89, + 705, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c75613529c7360414ae6aadd7d0f15c5a3adef53081c9499986c8eccc2985e9.jpg", + "image_caption": [ + "Figure 1. Over-dispersion problem: Binary CL excessively pursues to disperse features, disrupting intra-class distribution and struggling models to disentangle the underlying semantics." + ], + "image_footnote": [], + "bbox": [ + 524, + 266, + 895, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing to a lack of global diversity that will cause CL models to converge to trivial solutions [45, 58]. Recent evidence highlights the critical importance of local features [29, 60, 64, 70] and leading results [15, 41, 65] in medical images increasingly incorporate pixel-wise methods. This observation raises a fundamental question: Can CL be reformulated as a truly pixel-wise SSP for MVFMs?", + "bbox": [ + 509, + 415, + 906, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An obstacle to answering this question is the overdispersion problem [23, 45, 58]. The standard CL formulates the SSP as a binary optimization problem (termed binary $CL$ ), where feature distances in positive or negative pairs are directly minimized or maximized [36], leading to excessive dispersion of features without degree modeling (Fig.1-b). However, recent findings indicate that excessive pursuit of dispersion will disrupt intra-class distribution [56, 58]. This rigid contrast imposes significant limitations on medical vision tasks that require high sensitivity to subtle variations and structures, like thin vessels and subtle lesions, which risk being blurred or indistinguishable (Fig.1-c). While some attempts [5, 63] softened the contrast or introduced regularization [58] to mitigate this problem, they remain fundamentally reliant on the binary CL.", + "bbox": [ + 509, + 522, + 908, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In pixel-wise pretraining, the over-dispersion problem renders the binary CL particularly ineffective (Fig.1 a-c). Pixel-wise features are distributed across the image grid, where semantics vary continuously [13], making them inherently correlated. Unlike discrete classification tasks, continuous signals lack a natural definition of distinct categories [16, 49], leading to ambiguity near proximity divisions. The excessive pursuit of dispersion breaks these correlations so that it aggravates the disruption of the inherent intra-class distribution, struggling pretrained models to", + "bbox": [ + 509, + 750, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: shuo.li11@case.edu", + "bbox": [ + 107, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "disentangle the underlying semantics hidden in pixels [2].", + "bbox": [ + 89, + 90, + 470, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivation: Modeling distances for pixel-wise features is essential to quantify their dispersion degrees. The binary CL follows a max-min contrast for feature distances, neglecting explicit quantification of dispersion (Fig.1-b). Ideally, suppose we can predefine a ground truth distance $\\alpha$ to indicate the desired feature dispersion, the model will be trained to align feature distances accordingly [47, 66]. Such a CL with explicit distance modeling would enable precise quantification of dispersion while preserving feature correlations (Fig.1-d). Unfortunately, in practice, these ground truth distances are task-dependent [40], making them unmeasurable in the SSP and mismatched with the general objectives of foundation models [3].", + "bbox": [ + 88, + 108, + 482, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes a solution for the over-dispersion problem in pixel-wise CL by formulating a vector contrastive learning. The key insight is that feature distances inherently encode semantic correspondences and can be represented as displacement vectors in image space (Fig.2). When pixel-wise features align on an image grid, such vectors (pointing from source to target coordinates) provide a structured way to model correspondences. Therefore, instead of directly minimizing $|\\alpha - d'|$ to model distances $d'$ in the embedding space, we formulate a function $\\mathcal{V}$ that relates distances to vectors $v'$ in the image space. This reformulates the CL as a vector regression problem: $|v - \\mathcal{V}(d')|$ (formulated in Sec.3.1), where $v$ is the ground truth vector. By leveraging the properties of vectors, this approach structurally regulates feature dispersion, preserving semantic continuity. Incorporating measurable magnitudes and directions from the image space, vector CL provides a principled way to guide CL with distance modeling.", + "bbox": [ + 91, + 306, + 482, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The implementation of vector CL remains unexplored, hindered by two questions: Q1. How to construct a self-learning paradigm with free ground truth vector $v$ that can be extended across diverse medical images? Q2. How to formulate the function $\\mathcal{V}$ to ensure a consistent optimization flow from vector regression to distance modeling while aligning with the properties of correspondence? Although preliminary studies [19, 22] explored vectors in medical SSP, they lacked a structured formulation to link distances and vectors, causing inconsistent optimization flows. For the first time, we formulate the vector CL and implement it as a COntrast in Vector Regression (COVER) framework, which models distance in pixel-wise CL, enabling quantitative dispersion in pixel-wise SSP via three key innovations:", + "bbox": [ + 89, + 580, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1) Self-Vector Regression (SeVR) (for $Q1$ ) constructs an extendable self-space transformation mechanism for vector CL with free ground truth vectors $v$ (Fig.3-a, Sec.3.2). Specifically, it generates a displacement vector field (DVF), a dense field of vectors on the image grid, capturing pixelwise correspondences between two transformed views of a medical image. The pixel-wise distances are modeled", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c67df685c15aacd495f309d74cfd98026a55844683228057bde80004b6f8e12b.jpg", + "image_caption": [ + "Figure 2. Motivation: Our vector CL unifies the distance modeling $|d - d'|$ and displacement vector regression $|v - \\mathcal{V}(d')|$ , enabling CL with quantitative dispersion via the optimization flow." + ], + "image_footnote": [], + "bbox": [ + 526, + 87, + 895, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "through vector regression supervised by the DVF. This synthesis-based process enables a scalable and annotation-free SSP algorithm, facilitating vector CL across a wide variety of medical images (free from paired images [19, 22]).", + "bbox": [ + 511, + 253, + 903, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2) Mixture of Vectors (MoV) (for $Q2$ ) formulates a mapping function with consistent optimization flow from vector regression to distance modeling (Fig.3-b, Sec.3.3) with two properties: i. Spatial continuity. It designs a vector embedding unit (VEU) that encodes continuous spatial relationships in a vector template to map pixel-wise feature distances to vectors, thus avoiding the artificial division [29, 60, 64] and preserving the feature correlation. ii. Correspondence ambiguity. It constructs a multi-vector integration (MVI) strategy that extracts multiple vectors indicating potential correspondences with diverse feature concerns, enhancing bias adaptability for pretrained models.", + "bbox": [ + 511, + 315, + 903, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3) Vector Pyramid Aggregation (VPA) (for $Q2$ ) formulates multiscalarity of correspondence via stacking the MoV in a pyramid-like architecture (Fig.3-c, Sec.3.4). It predicts vectors across multiple feature scales and fuses them to capture correspondences for semantic objects at varying scale levels. Therefore, it enables the vector CL on different scales with low computation cost, training the model to represent multiscale features and improving granularity adaptability for pretrained models.", + "bbox": [ + 511, + 496, + 905, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: 1) For the first time, we pioneer a novel CL paradigm, vector CL, overcoming the over-dispersion problem in standard binary CL while preserving intra-class distribution. 2) We propose the COVER framework, which implements the vector CL for an extendable pixel-wise SSP in MVFMs. 3) Our SeVR establishes a versatile self-learning paradigm extendable across diverse medical images. 4) Our MoV maps feature distances as vectors, enabling a consistent optimization flow that quantifies dispersion while maintaining feature correlations. 5) Our VPA captures multiscale correspondences, enhancing representations across varying granularity scales.", + "bbox": [ + 511, + 633, + 905, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 511, + 829, + 660, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1) Medical image self-supervised pretraining [30] (SSP), as a training algorithm without annotation, is effectively promoting the programming of medical visual foundation", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19828", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models [6, 21, 33, 71]. It will capture the posterior distribution of the underlying explanatory factors and make models easier to extract useful information [2], thus achieving the general application goal. It always utilizes the generative learning (GL) [28, 38, 54, 72] or contrastive learning (CL) [4, 7, 8, 14, 29, 37, 50, 60, 64, 67-69] paradigms that train model by generating generate meaningful information or contrasting feature pairs. In this paper, we limit our scope to medical visual CL and review the below topics that are relevant to the applications considered in the sequel.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2) Binary contrastive learning has become a dominant SSP paradigm that learns general representations without annotations. Early works [7, 14, 16] maximize the distance between positive pairs while minimizing distance with negatives, leveraging global representations in natural images and being implemented to medical images [6, 55, 62]. However, the lack of global diversity in medical images makes the global representations insufficient, causing model collapse [45, 58]. While dense binary CL [60, 64] learns dense representations for local features, it suffers from over-dispersion problem due to mini-max dispersion, distorting their correlations and intra-class distributions [56, 58].", + "bbox": [ + 89, + 242, + 482, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3) Distance modeling [47, 66] quantifies features' distances in embedding spaces, preserving their correlations. It focuses on learning similarity metrics and guiding feature alignment with specific distance [24, 46]. This approach aligns closely with human cognition in perceiving continuous objects [10]. This has proven valuable in capturing feature correlations for effective pixel-wise feature representation. However, when aiming for SSP with general representations, predefined distances become impractical [40], posing challenges to direct distance modeling.", + "bbox": [ + 89, + 424, + 482, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4) Vector regression was originally developed to learn correspondence vectors of visual objects, and has been widely applied in image registration [1, 17, 20, 57], geometric matching [11, 12, 43], etc. Some preliminary SSP works [28, 73] attempt to establish absolute correspondence between semantics and positions but overlook their relative variations. GVSL [19] and GEMINI [22] introduce correspondence learning into medical SSP by predicting vector fields that represent the dense correspondence between objects across images. However, the lack of a formulation for the mapping function between feature distances and vectors has limited its ability to exert control over pixel-wise representations. This paper addresses this gap by formulating the relationship between feature distances and their vectors.", + "bbox": [ + 89, + 574, + 482, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. COVER: COntrast in Vectors Regression", + "text_level": 1, + "bbox": [ + 89, + 799, + 459, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our COVER (Fig.3) implement the formulated vector CL (Sec.3.1) via establishing the consistent optimization flow between distance modeling and vector regression (MoV, Sec.3.3), modeling multi-scale representation (VPA, Sec3.4), and learning the vector regression in a self-learning", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pipeline (SeVR, Sec.3.2), thus yielding pixel-wise pretraining with quantitative dispersion in medical vision.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Formulation", + "text_level": 1, + "bbox": [ + 511, + 128, + 648, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Problem formulation The proposed vector CL reformulated the CL as a vector regression problem that models pixel-wise feature distances to quantify their dispersion degrees. It defines a mapping function $\\mathcal{V}$ that links pixel-wise features to displacement vectors with a consistent optimization flow. Given an input medical image $x$ from dataset $\\mathcal{D}$ , we apply a random spatial transformation function $\\mathcal{T}_{sp}$ to generate two augmented views $x_{a}, x_{b}$ with their spatial correspondence encoded in a displacement vector field (DVF) $\\psi_{ab} = \\{v^{i}\\}_{i\\in \\Omega}$ ( $\\Omega$ is the image grid), where $v^{i}$ represents the coordinate shift from $x_{a}$ to $x_{b}$ at pixel $i$ . The network $\\mathcal{N}_{\\theta}$ extracts dense feature maps $F_{a} = \\mathcal{N}_{\\theta}(x_{a})$ and $F_{b} = \\mathcal{N}_{\\theta}(x_{b})$ . The function $\\mathcal{V}$ then predicts the DVF $\\psi_{ab}' = \\{v'^i\\}_{i\\in \\Omega}$ from these features, aiming to minimize the disparity between $\\psi_{ab}'$ and $\\psi_{ab}$ . This is formulated as:", + "bbox": [ + 511, + 151, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {x \\sim \\mathcal {D}} [ | \\psi_ {a b} ^ {\\prime} - \\psi_ {a b} | ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 386, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e \\psi_ {a b} ^ {\\prime} = \\mathcal {V} \\left(F _ {a}, F _ {b}\\right), \\psi_ {a b} \\sim \\mathcal {T} _ {s p}, \\psi_ {a b} \\left(x _ {a}\\right) \\equiv x _ {b}.\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 412, + 883, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The $\\equiv$ means both sides have consistent space.", + "bbox": [ + 511, + 438, + 823, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Why does our vector CL drive distance modeling? Because the vector CL unifies feature distance modeling and displacement vector regression mathematically. It constructed an equivalency of vector regression and distance modeling, obtaining a consistent optimization flow, and thus enabling the ability to learn feature distances in the embedding space without ground truth distances. This process is formalized as follows:", + "bbox": [ + 511, + 453, + 906, + 573 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D i s t. \\quad M o d.: \\sum_ {i = 0} ^ {I} \\left| \\alpha^ {i} - d ^ {\\prime i} \\right|\\rightarrow 0 (2) \\\\ \\Leftrightarrow \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} \\left| \\alpha^ {i} - d ^ {i} \\right|\\rightarrow 0 (3) \\\\ \\Rightarrow \\left| \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} \\alpha^ {i} - \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} d ^ {\\prime i} \\right|\\rightarrow 0 (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 580, + 903, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V e c}. \\operatorname {R e g}.: \\Leftrightarrow | v - \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} d ^ {n i} | \\rightarrow 0. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 648, + 903, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ideal distance modeling in a set can be formulated as Equ.2, where the $\\alpha^i$ is the ground truth distance for feature pair $i$ in the set with $I$ distances. Our COVER defines this distance set as the distances on a field of the image grid (Sec.3.3) and embeds a vector template matrix $\\mathbb{V}$ to weight them for Equ.3 which is equivalent with Equ.2. Based on this, Equ.4 can be deduced that illustrates when the individual distance approaches its certain value, the whole will also approach the certain whole. Therefore, our COVER hypothesizes that existing a value $v$ (i.e., vector in DVF $\\psi_{ab}$ ) can make $\\sum_{i=0}^{I} \\mathbb{V}^i \\alpha^i = v$ , the Equ.5 will be derived from Equ.4, formulating distance modeling in vector regression. Compared with the binary CL, our vector CL makes a tighter generalization bound [32]. We have formulated and analyzed the theory foundations in the supplementary).", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19829", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4da739be2103bea3617726ba6fb6e49aa26171fef91c8e8ed3687624316f6c1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 85, + 867, + 246 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/23c49f305050799788d8e6b5a927ac4f5fe0df57eb6fb62ab1d9292b99170199.jpg", + "image_caption": [ + "Figure 3. Framework: Our COVER formulates the distance modeling as the regression of displacement vectors (coordinate correspondences), quantifying dispersion degree in pixel-wise CL. a) SeVR (Sec.3.2) implements vector CL in a self-space transformation process. b) MoV (Sec.3.3) formulates feature distances as vectors with the properties of spatial continuity (VEU (i)) and correspondence ambiguity (MVI (ii)). c) VPA (Sec.3.4) formulates multiscalarity of correspondence via integrating the multi-scale vectors for granularity adaptability." + ], + "image_footnote": [], + "bbox": [ + 135, + 247, + 584, + 409 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e7ffb6411fbcdde2b8c7eb4577b8c7330811388f85c50dd7465d87806aec901.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 247, + 864, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equivalency of vector regression and distance modeling. While vector regression is not strictly equivalent to conventional distance modeling, our approach of capturing overall distance distribution enhances representation generalizability. Although Equ.4 shows that the aggregated weighted error tends to zero, it does not guarantee that each individual error in Equ.3 vanishes. Fortunately, since the ground truth distance $\\alpha^i$ is task-dependent [40], our approach captures the overall distance distribution rather than enforcing exact matches for each feature pair. This flexibility allows the model to accommodate varying feature biases across different tasks, promoting a more adaptable representation.", + "bbox": [ + 88, + 484, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Self-vector regression for extendable self-learning", + "text_level": 1, + "bbox": [ + 89, + 680, + 482, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our SeVR (Fig.3-a) constructs and regresses the ground truth vector $v$ in a self-space transformation mechanism. Given a set of medical images $\\mathcal{D}$ , an image $x$ is randomly sampled and cropped $x \\sim \\mathcal{D}$ , and two transformation operations for appearance $t \\sim \\mathcal{T}_{ap}$ and space $\\psi_{ab} \\sim \\mathcal{T}_{sp}$ are sampled for appearance and space transformation functions $\\mathcal{T}_{ap}, \\mathcal{T}_{sp}$ . Especially, the space transformation operation is a DVF $\\psi_{ab}$ which transforms the space of the image $x$ and will play as the ground truth vectors $\\{v^i\\}_{i \\in \\Omega}$ in the field. Our COVER produces the transformation operations for the image $x$ and generates two views $x_a = t(x), x_b = \\psi_{ab}(x)$ with different space and appearance features. Two share-weighted networks $\\mathcal{N}_\\theta$ extract the pixel-wise features with", + "bbox": [ + 89, + 703, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L$ scale levels $F_{a} = \\{f_{a}^{l}\\}_{l = 0}^{L} = \\mathcal{N}_{\\theta}(x_{a}), F_{b} = \\{f_{b}^{l}\\}_{l = 0}^{L} = \\mathcal{N}_{\\theta}(x_{b})$ from these views for multi-scale representation. A function $\\mathcal{V}(\\cdot ,\\cdot)$ (VPA, Sec.3.4) formulates the relationship between the distances of these features $F_{a}, F_{b}$ , predicting the DVF $\\psi_{ab}^{\\prime}$ which indicates their correspondences.", + "bbox": [ + 511, + 484, + 906, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our COVER learns to regress the vector indicates the correspondence to drive the distance modeling, so a vector loss $\\mathcal{L}_{vec}$ calculates a mean absolute error between the predicted $\\psi_{ab}^{\\prime}$ and the ground truth DVF $\\psi_{ab}$ , i.e.,", + "bbox": [ + 511, + 561, + 906, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v e c} \\left(\\psi_ {a b}, \\psi_ {a b} ^ {\\prime}, \\epsilon_ {a b}\\right) = \\sum_ {i \\in \\{\\epsilon_ {a b} = 1 \\}} \\left| \\psi_ {a b} ^ {i} - \\psi_ {a b} ^ {\\prime i} \\right|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 633, + 906, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\epsilon_{ab}$ is a mask that eliminates the content mismatch of two views $x_{a}, x_{b}$ caused by the spatial transformation and is generated from the ground truth DVF $\\psi_{ab}$ . It enables the network to model the correspondence degrees between one feature and a set of features (illustrated in Sec.3.3) for position correspondence, i.e., vectors. An additional consistency loss $\\mathcal{L}_{con}$ is used to maintain the invariance of semantics in space transformation. It calculates a cosine similarity between the dense features from two views, i.e., $\\mathcal{L}_{con}(f_{ab}^{4}, f_{b}^{4}, \\epsilon_{ab}) = -\\sum_{i \\in \\{\\epsilon_{ab} = 1\\}} \\frac{f_{ab}^{4,i} \\cdot f_{b}^{4,i}}{||f_{ab}^{4,i}||||f_{b}^{4,i}||}$ , where the $f_{ab}^{4} = \\psi_{ab}(f_{a}^{4})$ is the transformed 4th level dense features $f_{a}^{4}$ for a consistent space of dense feature $f_{b}^{4}$ .", + "bbox": [ + 511, + 662, + 908, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The whole optimization of our COVER iteratively minimizes the arg $\\min_{\\theta} \\mathcal{L}_{\\text{COVER}} = \\mathcal{L}_{\\text{con}} + \\mathcal{L}_{\\text{vec}}$ for weights $\\theta$ . If it utilizes the SGD, its dynamics can be formulated", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19830", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as $\\theta^{*}\\gets \\theta -\\eta \\frac{\\partial\\mathcal{L}_{COVER}(\\theta,\\mathcal{D},\\mathcal{T}_{sp},\\mathcal{T}_{ap})}{\\partial\\theta}$ , where the $\\eta$ is the learning rate. More details are illustrated in Supplementary.", + "bbox": [ + 89, + 88, + 480, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion of properties Our SeVR constructs a self-space transformation mechanism, enabling an extendable self-supervised learning paradigm. It generates DVFs from random transformation as the ground truth vectors $v$ to drive the vector regression without annotation or externally paired data [19, 22]. This property makes it extendable to a wide variety of medical images for large-scale training simplifying data preparation and enhancing generality.", + "bbox": [ + 89, + 121, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Mixture of vectors with consistent optimization flow", + "text_level": 1, + "bbox": [ + 89, + 250, + 482, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a part of the function $\\mathcal{V}$ , our mixture of vectors (MoV, Fig.3-b) $\\mathcal{M}$ maps distances as vectors with consistent optimization flow. It constructs a vector embedding unit (VEU, b-i) and a multi-vector integration (MVI, b-ii) to encode the spatial continuity and correspondence ambiguity:", + "bbox": [ + 89, + 272, + 483, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vector embedding unit (VEU) $\\mathcal{U}(\\cdot, \\cdot)$ For a receptive field with $N \\times N$ size (For a clearer illustration, we take 2D situation as example.) on position $i$ of image grid, the distances between the center feature $f_{a}^{i}$ (from view $x_{a}$ ) and the set of features $f_{b}^{N \\times N} = \\{f_{b}^{(n,n)}\\}_{n=0}^{N}$ in the receptive field (from view $x_{b}$ ) are calculated. Inspired by the attention in transformer [53], it uses the scaled dot-product attention to model their similarity to $(0,1)$ thus achieving a distance map $D^{N \\times N}$ . To map the feature distances as vectors, we design a fixed vector template matrix $\\mathbb{V}^{N \\times N} = \\{\\mathbb{V}^{(n,n)}\\}_{n=0}^{N}$ that indicates the vectors in which the center coordinate pointing to the coordinates in the receptive field. This vector template $\\mathbb{V}^{N \\times N}$ is matrix multiplied with the distance map $D^{N \\times N}$ for a vector $v_{ab}^{\\prime i}$ . The function of our VEU $\\mathcal{U}(\\cdot, \\cdot)$ is", + "bbox": [ + 89, + 347, + 483, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {a b} ^ {\\prime i} = \\mathcal {U} \\left(f _ {a} ^ {i}, f _ {b} ^ {N \\times N}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {f _ {a} ^ {i} f _ {b} ^ {\\top N \\times N}}{\\tau}\\right) \\mathbb {V} ^ {N \\times N}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 568, + 482, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\tau$ is a scaling factor ( $\\tau = \\sqrt{c_k}$ in experiment following [53]). The predicted vector $v_{ab}^{\\prime i}$ is the displacement of feature $f_{a}^{i}$ to align the most similar features in $f_{b}^{N\\times N}$ .", + "bbox": [ + 89, + 599, + 482, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-vector integration (MVI) Our MVI predicts and integrates multiple vectors to adapt the ambiguous correspondence in a region with continuous semantics caused by varied feature concerns. The features $f_{a}, f_{b}$ with $C$ channels are divided into $J$ groups, and each group is produced by our VEU $\\mathcal{U}$ for a potential vector $v'^j = \\mathcal{U}(f_a[j], f_b[j])$ where $j$ is the index of group, thus achieving a vector group $\\mathbf{v}' = \\{v'^j\\}_j^J$ representing the correspondences in different feature concerns. Finally, it averages the vectors in the group for a displacement vector $u' = \\frac{1}{J}\\sum_{j=0}^{J}v'^j$ that is compatible with the ambiguity of correspondence. Calculating the displacement vectors for all positions on the image grid, our MoV will output the displacement vector field (DVF) $\\psi_{ab}' = \\{u_{ab}'\\}_{i\\in \\Omega}$ that indicates the correspondence of dense features between view $x_{a}$ and $x_{b}$ .", + "bbox": [ + 89, + 645, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion of properties Our MoV constructs a consistent optimization flow from vector regression to distance", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "modeling to quantify the dispersion. It encodes the properties of spatial continuity and correspondence ambiguity of the pixel-wise features, bringing two advantages: 1) Feature correlation preservation: Our VEU constructs a vector template $\\mathbb{V}$ that describes this continuous spatial relationship. This template $\\mathbb{V}$ maps the feature distances to vectors that avoid the artificial division [29, 60, 64] thus preserving the feature correlation. 2) Bias adaptability. Pixel-wise correspondence is uncertain and task-dependent, our MVI models this ambiguity to represent diverse feature concerns, enhancing bias adaptability in general applications.", + "bbox": [ + 511, + 90, + 905, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Vector pyramid aggregation adapts granularity", + "text_level": 1, + "bbox": [ + 511, + 266, + 893, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stacking our MoV $\\mathcal{M}$ in a pyramid-like architecture, our vector pyramid aggregation (VPA, Fig.3-c) formulates the final mapping function $\\nu$ to model the multiscalarity of correspondence efficiently, training a granularity adaptability. This architecture divides multi-level correspondences according to the scale levels of the dense features. By fusing these scale levels, the VPA will output the DVF $\\psi_{ab}^{\\prime}$ that models both global and local correspondences. In 0th level, dense features $f_{a}^{0},f_{b}^{0}$ are put into our MoV module for a 0th-level DVF $\\psi_{ab}^{\\prime 0} = \\mathcal{M}(f_{a}^{0},f_{b}^{0})$ with a global correspondence for high-level features. Then, in 1st level, the $\\psi_{ab}^{\\prime 0}$ transforms the 1st-level dense features $f_{a}^{1}$ to bridge the space gap between $f_{a}^{1}$ and $f_{b}^{1}$ on the 0th level. The features are further produced by our MoV and further fused with the $p s i_{ab}^{\\prime 0}$ for the 1st-level DVF $\\psi_{ab}^{\\prime 1} = \\mathcal{M}(\\psi_{ab}^{\\prime 0}(f_{a}^{1}),f_{b}^{1})\\odot \\psi_{ab}^{\\prime 0}$ , where the “ $\\odot$ ” is the fusion operation illustrated in Supplementary. The remaining levels are calculated similarly so that the final DVF $\\psi_{ab}^{\\prime}$ integrated multiscale vectors are output after a chain calculation, i.e.", + "bbox": [ + 511, + 287, + 906, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {a b} ^ {\\prime} = \\mathcal {V} (F _ {a}, F _ {b}) = H (\\{\\psi_ {a b} ^ {\\prime \\prime} \\} _ {l = 0} ^ {L}) \\quad (\\psi_ {a b} ^ {\\prime 0} = \\mathcal {M} (f _ {a} ^ {0}, f _ {b} ^ {0}))\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 582, + 901, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {a b} ^ {\\prime l} = \\mathcal {M} \\left(\\psi_ {a b} ^ {\\prime l - 1} \\left(f _ {a} ^ {l}\\right), f _ {b} ^ {l}\\right) \\odot \\psi_ {a b} ^ {\\prime l - 1}, l = 1, \\dots , L - 1, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 602, + 903, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $H$ is the chain calculation across all $L$ levels to output the final DVF $\\psi_{ab}^{\\prime}$ . Linear upsampling and value doubling are applied to bridge the scale gap of the DVFs.", + "bbox": [ + 511, + 628, + 905, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion of properties Correspondence occurs at multiple scales for different semantic granularity, so our VPA formulates this multiscalarity in a pyramid-like architecture, bringing the advantages: 1) Low computational cost: It enables the mapping function $\\mathcal{V}$ in a small receptive field at each level for a large whole receptive field reducing the computation. 2) Multi-scale representation: It learns the multi-scale features with multiple semantic granularities improving the granularity adaptability.", + "bbox": [ + 511, + 675, + 906, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparison study", + "text_level": 1, + "bbox": [ + 511, + 848, + 691, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Experiment protocol This study has made sufficient experiments to ensure a complete evaluation. Following is", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19831", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b77a2324897bdc4d6dbb0f5b24a83da5ac42fcad84c26cc6842667997a376729.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodsa) 2D evaluation pretrained on [59]b) 3D evaluation pretrained on [31]AVG Score %
\\(SCR_{S}^{25\\%}\\)\\(PDCX_{R_{C}}\\)\\(KiPA22_{S}^{2D}\\)\\(FIVES_{S}\\)\\(CANDI_{S}\\)\\(FeTA21_{S}\\)\\(KiPA22_{S}^{3D}\\)\\(STOIC_{C}\\)
-Scratch81.890.474.179.484.056.972.472.076.4
GLDenosing [54]83.9(+1.9)92.0(+1.6)60.3(-13.8)77.8(-1.6)83.7(-0.3)52.9(-4.0)70.0(-2.4)65.9(-6.1)73.3(-3.1)
In-painting [38]85.1(+3.3)93.9(+3.5)64.4(-9.7)78.9(-0.5)88.5(+4.5)54.4(-2.5)69.7(-2.7)67.2(-4.8)75.3(-1.1)
Models Genesis [72]86.1(+4.3)92.6(+2.2)66.6(-7.5)79.6(+0.2)88.7(+4.7)55.8(-1.1)75.8(+3.4)75.3(+3.3)77.6(+1.2)
Rotation [28]80.5(-1.3)89.9(-0.5)69.7(-4.4)80.3(+0.9)89.4(+5.4)58.7(+1.8)77.4(+5.0)68.8(-3.2)76.8(+0.4)
BCLSimSiam [8]87.2(+5.4)92.2(+1.8)72.6(-1.5)84.3(+4.9)87.3(+3.3)58.7(+1.8)83.8(+11.4)69.5(-2.5)79.5(+3.1)
BYOL [14]89.4(+7.6)86.6(-3.8)74.1(0)83.3(+3.9)89.7(+5.7)59.2(+2.3)83.6(+11.2)74.0(+2.0)80.0(+5.4)
SimCLR [7]89.0(+7.2)94.7(+4.3)74.4(+0.3)84.5(+5.1)89.2(+5.2)53.4(-3.5)78.9(+6.5)60.7(-11.3)78.1(+1.7)
MoCov2 [9]84.3(+2.5)93.2(+2.8)69.6(-4.5)80.7(+1.3)89.7(+5.7)61.5(+4.6)78.0(+5.6)74.8(+2.8)79.0(+2.6)
DeepCluster [4]84.0(+2.2)93.3(+2.9)72.7(-1.4)81.6(+2.2)89.8(+5.8)57.4(+0.5)79.7(+7.3)65.6(-6.4)78.0(+1.6)
DBCLVADeR [35]85.2(+3.4)92.8(+2.4)62.8(-11.3)78.9(-0.5)87.4(+3.4)43.1(-13.8)72.1(-0.3)73.2(+1.2)74.4(-2.0)
DenseCL [60]85.0(+3.2)92.4(+2.0)70.8(-3.3)79.2(-0.2)87.7(+3.7)43.7(-13.2)74.0(+1.6)58.8(-13.2)74.0(-2.4)
SetSim [61]85.2(+3.4)93.9(+3.5)70.8(-3.3)80.1(+0.7)88.4(+4.4)58.7(+1.8)73.5(+1.1)60.1(-11.9)76.3(-0.1)
DSC-PM [29]90.5(+8.7)91.8(+1.4)77.2(+3.1)83.8(+4.4)88.5(+4.5)52.2(-4.7)79.0(+6.6)59.5(-12.5)77.8(+1.2)
PixPro [64]91.5(+9.7)93.0(+2.6)73.6(-0.5)84.3(+4.9)89.9(+5.9)60.7(+3.8)80.0(+7.6)75.1(+3.1)81.0(+4.6)
Chaitanya et al. [5]87.3(+5.5)90.0(-0.4)76.5(+2.4)84.9(+5.5)87.4(+3.4)53.4(-3.5)70.7(-1.7)67.8(-4.2)77.3(+0.9)
VRGVSL [19]89.7(+7.9)92.5(+2.1)78.9(+4.8)86.2(+6.8)89.1(+5.1)62.6(+5.7)84.3(+11.9)75.4(+3.4)82.3(+5.9)
GEMINI [22]92.4(+10.6)92.9(+2.5)79.1(+5.0)85.3(+5.9)90.0(+6.0)61.7(+4.8)85.0(+12.6)79.5(+7.5)83.2(+6.8)
VCLCOVER (Ours)94.0(+12.2)95.9(+5.5)80.0(+5.9)87.2(+7.8)89.9(+5.9)63.6(+6.7)85.2(+12.8)80.4(+8.4)84.5(+8.1)
", + "bbox": [ + 94, + 87, + 903, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison study: Our COVER achieves the best performance compared with 17 methods on 8 downstream tasks across a) 2D and b) 3D, showing our great adaptation ability. \"AVG\" is the average rank of the row. The blue values are the improvement compared with the \"Scratch\" and the bold values are the highest score. \"S\" and \"C\" are the segmentation and classification tasks.", + "bbox": [ + 89, + 348, + 906, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the overview protocol and details are in our Supplementary.", + "bbox": [ + 89, + 400, + 480, + 414 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Materials: This study evaluated our COVER across 8 tasks across 2D and 3D evaluations with 4 modalities to showcase its versatility and advantages. a) $2D$ evaluation: It pretrained networks on the ChestX-ray8 dataset [59] containing 112,120 chest X-ray images. These pretrained models are fine-tuned on four tasks $(\\mathrm{SCR}_S^{25\\%}$ [52] (X-ray), $\\mathrm{PDCXR}_C$ [27] (X-ray), $\\mathrm{FIVES}_S$ [25] (fundus), $\\mathrm{KiPA22}_{S}^{2D}$ [18] (CT)). b) $3D$ evaluation: It pretrained networks on 837 3D T1 brain MR images from the PPMI database [31]. These pretrained models are then fine-tuned to three datasets with four tasks $(\\mathrm{CANDI}_S$ [26] (MR), $\\mathrm{FeTA21}_S$ [39] (MR), $\\mathrm{STOIC}_C$ [42] (CT), $\\mathrm{KiPA22}_{S}^{3D}$ [18] (CT)). The evaluations involve segmentation (S) and classification (C) tasks. To evaluate the data amount robustness, it used $25\\%$ of the training set in $\\mathrm{SCR}_S$ , and more data amount evaluations are performed in our analysis (Fig.4 b).", + "2) Comparisons: This study benchmarked COVER with 17 recent or classic methods across four categories: GL [28, 38, 54, 72], binary CL (BCL) [4, 7-9, 14], dense binary CL (DBCL) [5, 29, 35, 60, 61, 64], and vector regression (VR) [19, 22]. It used 2D and 3D U-Nets [44] as the backbone $\\mathcal{N}_{\\theta}$ (for the methods with global prediction, like MoCov2, it adopted the encoder part). For a fair comparison, full-parameter fine-tuning is employed in downstream tasks.", + "3) Implementation and evaluation metrics: Following [19], all tasks were implemented by PyTorch and optimized by Adam with a learning rate of $10^{-4}$ . The models were pretrained with $2 \\times 10^{5}$ iterations for pretraining tasks and fine-tuned with $4 \\times 10^{4}$ iterations for downstream tasks. Training and testing were implemented on NVIDIA A100 SXUM4 GPU with 40 GB memory. This study used the Dice coefficient (DSC) for segmentation tasks and area un" + ], + "bbox": [ + 88, + 415, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "der curve (AUC) for classification tasks [48].", + "bbox": [ + 511, + 400, + 812, + 414 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4.1.2 Quantitative analysis Our great adaptability for the downstream tasks shown in Tab. 1 demonstrates our COVER's powerful general representation for medical vision. Two observations can be discovered:", + "1) Cross-scale transferability. The SOTA performances on small and large targets show our powerful cross-scale transferability owing to the learned multiscalarity from our VPA a. For small structures like the vessels $(\\mathrm{FIVES}_S)$ and brain tissues (CANDI $s$ and $\\mathrm{FeTAS}_s$ ), our approach achieved $87.2\\%$ $89.9\\%$ ,and $63.6\\%$ DSC with the improvements of $7.8\\%$ $5.9\\%$ ,and $6.7\\%$ over the \"Scratch\". These gains stem from learning fine-grained representations at lowerscale levels. b. For large structures including the chest regions $(\\mathrm{SCR}_{S}^{25\\%})$ and kidney structures $(\\mathrm{KiPA22}_{S}^{2D}$ ,and $\\mathrm{KiPA22}_{S}^{3D})$ , our model achieved $94.0\\%$ $80.0\\%$ ,and $85.2\\%$ DSC improving by over $5 \\%$ compared to \"Scratch\". This is attributed to leveraging high-scale levels for coarse-grained representation. c. For classification tasks (PDCXR $c_{r}$ and STOIC), our COVER achieved AUC improvements of $5.5\\%$ and $8.4\\%$ by learning global representations at level 0. Although the GVSL and GEMINI also have improvements across scales, their inconsistent optimization flow limits the representation potential of pretraining data, resulting in an average score over $1 \\%$ lower than COVER. The BCL methods struggle to enhance all tasks, for example, the DenseCL improved $3.2\\%$ DSC on SCR, but it reduced $3.3\\%$ DSC on $\\mathrm{KiPA22}_{S}^{2D}$ . This arises from their over-dispersion, which introduces biases and compromises on incompatible tasks.", + "2) Cross-scene adaptability. The significant improvements on consistent $(\\mathrm{SCR}_S^{25\\%}$ , $\\mathrm{PDCXR}_C$ , $\\mathrm{CANDI}_S)$ and inconsistent (KiPA22 $_S^{2D}$ , FIVES $_S$ , FeTAS, KiPA22 $_S^{3D}$ , STOIC $_C$ ) scenes relative to the pretraining data illustrates" + ], + "bbox": [ + 511, + 415, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19832", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d541e0a43e5c8d83cde799220e17d9d1e0759fdf01579467189ce3b457ecffca.jpg", + "image_caption": [ + "Figure 4. Implementation analysis: Our COVER has great properties on the scaling of a) fine-tuned parameters, b) fine-tuning data, c) pretraining iterations, and d) model sizes." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our strong cross-scene adaptability. Across all tasks in both 2D and 3D evaluations, COVER outperformed the \"Scratch\" by more than $5\\%$ and achieved the highest average scores $(84.5\\%)$ . This adaptability arises from our distance modeling which quantifies the dispersion in embedding space, enabling the disentanglement of underlying explanatory factors hidden in low-level sensory data [2]. These learned factors capture the intrinsic properties of medical images, enhancing versatility across diverse scenes. Although most methods improve performance in consistent scenes by leveraging learned transformable knowledge, they struggle in inconsistent scenes. Notably, BCL and DBCL even degrade performance in some tasks due to their over-dispersion in binary CL, which amplifies biases in pretraining data. Especially, DenseCL and VADeR perform particularly poorly in inconsistent scenes like FeTA21 $_S$ , which reduce over $10\\%$ DSC.", + "bbox": [ + 88, + 263, + 482, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation study and model analysis", + "text_level": 1, + "bbox": [ + 89, + 530, + 390, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Component ablation The ablation study of the components in our COVER on the 2D evaluation of $\\mathrm{SCR}_S^{25\\%}$ task shows the effectiveness of our proposed innovations. This study took the learning of the consistency loss $\\mathcal{L}_{con}$ in pixel-wise as our base model, and gradually added our proposed components. The results are as follows:", + "bbox": [ + 89, + 551, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c c c c c} & \\text {base .} \\mathcal {L} _ {\\text {con}} & \\dots + \\text {VEU (SeVR)} & \\dots + \\text {VPA} & \\dots + \\text {MVI} \\\\ \\hline \\text {DSC} \\% & 91.8 & 92.9 & 93.4 & 94.0 \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 657, + 455, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The base model $\\mathcal{L}_{con}$ has $91.8\\%$ DSC that is higher than the BYOL's $89.4\\%$ owing to its pixel-wise representation. When adding our VEU, it effectively improves the discrimination, achieving a $1.1\\%$ improvement. Our VPA enhances the multi-scale representation so that the model further improves $0.5\\%$ DSC. Finally, our MVI models the ambiguity of the correspondence for potentially varied feature concerns, thus achieving the highest $94.0\\%$ DSC.", + "bbox": [ + 89, + 703, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Hyper-parameter ablation The receptive field $N$ in VEU, the amount of VEUs $J$ in MVI are two new hyperparameters in our COVER. This study took an ablation of them on the $\\mathrm{SCR}_S^{25\\%}$ via linear evaluation to demonstrate their properties. The results are as follows:", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2f039526514bc8b8d1b9c900c57db0d02d409d7befa34f905013bc65a2607b2.jpg", + "image_caption": [ + "Figure 5. Convergence analysis: a) Our COVER is able to promote convergence ability in fine-tuning. b) In the pretraining, the consistency learning causes representation collapse, and when adding our vector regression, the discrimination is enhanced." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 903, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/520c545af5cb179034a925e8ef7b08de99b968de6989d053c7e53a637c0ace75.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
N =3 × 35 × 57 × 79 × 9
DSC%44.948.454.835.1
J =[2,2,2,1,1][4,4,4,1,1][8,8,8,2,2][12,12,12,3,3]
DSC%55.356.354.848.6
", + "bbox": [ + 517, + 272, + 901, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the receptive field size $N$ , when enlarging it, its performance increases to $7 \\times 7$ and then decreases when it is $9 \\times 9$ . Because the increase of $N$ will adapt to large spatial correspondence, improving the pretrained representation. However, too large $N$ will introduce more ambiguous semantics, such as vascular with similar features, which will mislead the learning of correspondence. For the amount of VEUs $k$ , it also increases and then decreases. When it is [4,4,4,1,1] for the scale levels, the model achieves the best ability. Too many vectors will smooth the optimization, thus weakening the discrimination of learned features.", + "bbox": [ + 511, + 340, + 903, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3 Implementation analysis This experiment analyzed four implementation factors (Fig.4): a) Fine-tuned parameter amount: Our COVER effectively reduces the fine-tuned parameters. It only fine-tunes one U-Net module to achieve a very significant performance improvement (15.3% DSC higher than GVSL), quickly reaching the bottleneck level that $\\mathrm{SCR}_S^{25\\%}$ task can reach. b) Fine-tuning data amount: Our COVER effectively reduces the fine-tuning data requirement. Only utilizing 5% training data, our COVER can achieve the DSC close to 90% that GVSL and DenseCL have to use 25% and 100% data. c) pretraining iteration: With the progress of the training iterations, the performance (linear evaluation) on $\\mathrm{SCR}_S^{25\\%}$ of our COVER gradually improves and eventually tends to be flat. This illustrates that more training is beneficial for the vector CL, which makes the pretrained models capture more diverse features in detail. d) Model size: With the enlarging of model size, the larger capacity enables our COVER to capture more abundant feature information in the pretraining. This makes the model gain more powerful representability in downstream tasks only with a linear layer.", + "bbox": [ + 511, + 508, + 905, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.4 Convergence analysis This study analyzed the convergence performance of our COVER in both pretraining and fine-tuning phases (Fig.5): a) Fine-tuning: In downstream tasks, the network pretrained by our COVER exhibits superior convergence ability. Compared with the", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19833", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/288fc6cee09b2600efda5b1c01f168f210ebe3d84989ad83489f9b20c2dc3c85.jpg", + "image_caption": [ + "Figure 6. Cluster analysis: Our COVER has strong clustering capabilities by quantifying the dispersion degree of features." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 480, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\"Scratch\", GVSL, and DenseCL, COVER achieves a significantly lower loss value more quickly in both 2D segmentation (SCR) and 3D classification (STOIC) tasks. This advantage arises from our powerful pixel-wise feature learning that enables the pretrained network to capture the detailed information internal medical images. $b$ pretraining: COVER improves the discrimination of the representation. When relying solely on the consistency loss $\\mathcal{L}_{con}$ , the loss quickly collapses to zero, hindering subsequent knowledge learning. By integrating vector regression $\\mathcal{L}_{vec}$ , the model learns feature distances, stabilizing the consistency loss with fluctuations and enhancing the discrimination.", + "bbox": [ + 88, + 261, + 482, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.5 Computation efficiency This study calculated the FLOPs to evaluate the computation efficiency with the expanding of receptive field. For a significant analysis, we focused on the computation in head part, i.e., our VPA. It compared two methods: a. VPA; b. directly taking a large vector template at level 4 (named \"Direct\"). The results are:", + "bbox": [ + 89, + 443, + 483, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2df4300279be8ec292c9f8df774a0bfb0e74ccab85ca266a35f00fbeb3c2ebb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 549, + 483, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When expanding the whole receptive field, our VPA maintains a stable computational cost of 0.2, whereas the \"Direct\" approach shows an exponential increase. Especially when it is $121 \\times 121$ , the \"Direct\" has more than 52 times our consumption. Our VPA is able to expand the receptive field for a wider contrast with low additional consumption.", + "bbox": [ + 89, + 628, + 482, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.6 Distributed clusters from COVER The clusters show that our COVER has strong clustering capabilities by quantifying the dispersion of features. Using pretrained models from our 2D evaluation, this study extracted pixelwise features from the SCR and KiPA22 datasets and visualized their distributions via t-SNE [51] in Fig.6. It compared the clustering patterns of \"Scratch\", GVSL, and DenseCL with our COVER, revealing two key advantages: a) Continuous feature distribution: Our COVER predicts features that are distributed continuously in the embedded space. This continuity arises from our vector CL, which models the feature distance and avoids the excessive dis", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39ac4c0a47a9d8d885d5d8d139766b38abc90df31083185e1010b41a43db0172.jpg", + "image_caption": [ + "Figure 7. Multi-scale vectors analysis: Our VPA predicts multilevel correspondence for both global and detailed representation." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cretization that blurs small or subtle features. However, DenseCL produces a lot of discrete clusters with small gaps between them, i.e., over-dispersion, leading to fragmented representation and reduced transferability. b) Effective aggregation: Our COVER aggregates features with the same semantics, while other methods show mixed regions. In both SCR and KiPA22 datasets, \"Scratch\" and GVSL exhibit overlapped clusters, as shown in the enlarged patches. Although DenseCL performs well on SCR, it fails on KiPA22, mixing features in this inconsistent scene with the pretraining. Our COVER achieves effective aggregation across datasets because our modeling of distance mitigates the overfitting to the pretraining scene.", + "bbox": [ + 511, + 243, + 906, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.7 Analysis of multi-scale vectors in our VPA This study analyzes the vectors at each scale level in our VPA by visualizing spatial transformed images via the predicted DVFs (Fig.7) in our 2D (chest X-ray) and 3D (brain MR) evaluations. In level 0, the DVF predicted from global features is able to align the images on the whole, like the rotation, scaling, etc., driving the learning of global representation. With the expansion of the scales, the correspondences are gradually refined so the details between the images are aligned. Especially, in level 0, the images have high alignment degree on the boundary, so that the models will effectively learn the representation of small or subtle features.", + "bbox": [ + 511, + 443, + 908, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 646, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we reformulate the standard CL as a vector regression problem, vector CL, providing a solution for \"over-dispersion\" problem caused by binary CL in pixelwise medical visual pretraining. Our proposed COcontrast in VEctor Regression (COVER) framework implemented the reformulated vector CL, modeling feature distance for CL in a vector regression process. It enables quantitative dispersion of features which improves representation learning for medical vision applications. Although the unique properties of our vector CL have bright powerful performance, important future work is to involve a larger-scale medical image dataset in training, fostering the development of a large pixel-wise foundation model in the medical field. Beyond advancing MVFM [21] development, COVER opens new frontiers in representations for continuous signals.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19834", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This research was supported by the National Science Foundation grants (2306545) and National Institutes of Health grants (1R01HL177813-01, 1R01HL173186-01A1).", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 166, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 3", + "[2] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2, 3, 7", + "[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2", + "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 1, 3, 6", + "[5] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 1, 6", + "[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1, 3", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758, 2021. 3, 6", + "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6", + "[10] Oliver Contier, Chris I Baker, and Martin N Hebart. Distributed representations of behaviour-derived object dimensions in the human visual system. Nature Human Behaviour, 8(11):2179-2193, 2024. 3", + "[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Rethinking optical flow from geometric matching consistent perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1337-1347, 2023. 3", + "[12] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In Pro" + ], + "bbox": [ + 93, + 191, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27488-27497, 2024. 3", + "[13] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 1", + "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Pires, Zhaohan Guo, Mohammad Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In Neural Information Processing Systems, 2020. 1, 3, 6", + "[15] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial? Medical Image Analysis, 94:103086, 2024. 1", + "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3", + "[17] Yuting He, Tiantian Li, Rongjun Ge, Jian Yang, Youyong Kong, Jian Zhu, Huazhong Shu, Guanyu Yang, and Shuo Li. Few-shot learning for deformable medical image registration with perception-correspondence decoupling and reverse teaching. IEEE Journal of Biomedical and Health Informatics, 26(3):1177-1187, 2021. 3", + "[18] Yuting He, Guanyu Yang, Jian Yang, Rongjun Ge, Youyong Kong, Xiaomei Zhu, Shaobo Zhang, Pengfei Shao, Huazhong Shu, Jean-Louis Dillenseger, et al. Meta grayscale adaptive network for 3d integrated renal structures segmentation. Medical Image Analysis, 71:102055, 2021. 6", + "[19] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6", + "[20] Yuting He, Rongjun Ge, Xiaoming Qi, Yang Chen, Jia-song Wu, Jean-Louis Coatrioux, Guanyu Yang, and Shuo Li. Learning better registration to learn better few-shot medical image segmentation: Authenticity, diversity, and robustness. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2588-2601, 2024. 3", + "[21] Yuting He, Fuxiang Huang, Xinrui Jiang, Yuxiang Nie, Minghao Wang, Jiguang Wang, and Hao Chen. Foundation model for advancing healthcare: Challenges, opportunities, and future directions. arXiv preprint arXiv:2404.03264, 2024.1,3,8", + "[22] Yuting He, Boyu Wang, Rongjun Ge, Yang Chen, Guanyu Yang, and Shuo Li. Homeomorphism prior for false positive and negative problem in medical image dense contrastive representation learning, 2025. 2, 3, 5, 6", + "[23] John Hinde and Clarice GB Demétrio. Overdispersion: models and estimation. Computational statistics & data analysis, 27(2):151-170, 1998. 1", + "[24] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third inter" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19835", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 3", + "[25] Kai Jin, Xingru Huang, Jingxing Zhou, Yunxiang Li, Yan Yan, Yibao Sun, Qianni Zhang, Yaqi Wang, and Juan Ye. Fives: A fundus image dataset for artificial intelligence based vessel segmentation. Scientific data, 9(1):475, 2022. 6", + "[26] David N Kennedy, Christian Haselgrove, Steven M Hodge, Pallavi S Rane, Nikos Makris, and Jean A Frazier. Candishare: A resource for pediatric neuroimaging data. Neuroinformatics, 10(3):319, 2012. 6", + "[27] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. cell, 172(5):1122-1131, 2018. 6", + "[28] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 3, 6", + "[29] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 1, 2, 3, 5, 6", + "[30] Siladittya Manna, Saumik Bhattacharya, and Umapada Pal. Self-supervised visual representation learning for medical image analysis: A comprehensive survey. Transactions on Machine Learning Research. 1, 2", + "[31] Kenneth Marek, Danna Jennings, Shirley Lasch, Andrew Siderowf, Caroline Tanner, Tanya Simuni, Chris Coffey, Karl Kieburtz, Emily Flagg, Sohini Chowdhury, et al. The parkinson progression marker initiative (ppmi). Progress in neurobiology, 95(4):629-635, 2011. 6", + "[32] Mehryar Mohri. Foundations of machine learning, 2018. 3", + "[33] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1, 3", + "[34] Frank H Netter. Atlas of human anatomy, Professional Edition E-Book: including NetterReference. com Access with full downloadable image Bank. Elsevier health sciences, 2014. 1", + "[35] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 6", + "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1", + "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 1, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3, 6", + "[39] Kelly Payette, Hongwei Bran Li, Priscille de Dumast, Roxane Licandro, Hui Ji, Md Mahfuzur Rahman Siddiquee, Daguang Xu, Andriy Myronenko, Hao Liu, Yuchen Pei, et al. Fetal brain tissue annotation and segmentation challenge results. Medical image analysis, 88:102833, 2023. 6", + "[40] Guangzhi Qu, Salim Hariri, and Mazin Yousif. A new dependency and correlation analysis for features. IEEE Transactions on Knowledge and Data Engineering, 17(9):1199-1207, 2005. 2, 3, 4", + "[41] Quan Quan, Qingsong Yao, Heqin Zhu, and S Kevin Zhou. Igu-aug: Information-guided unsupervised augmentation and pixel-wise contrastive learning for medical image analysis. IEEE Transactions on Medical Imaging, 2024. 1", + "[42] Marie-Pierre Revel, Samia Boussouar, Constance de Margerie-Mellon, Inès Saab, Thibaut Lapotre, Dominique Mompoint, Guillaume Chassagnon, Audrey Milon, Mathieu Lederlin, Souhail Bennani, et al. Study of thoracic ct in疫情-19: The stoic project. Radiology, 301(1):E361-E370, 2021. 6", + "[43] Ignacio Rocco, Relja Arandjelovic, and Josef Sivic. Convolutional neural network architecture for geometric matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6148-6157, 2017. 3", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6", + "[45] Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning, pages 5628-5637. PMLR, 2019. 1, 3", + "[46] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 3", + "[47] Juan Luis Suárez, Salvador García, and Francisco Herrera. A tutorial on distance metric learning: Mathematical foundations, algorithms, experimental analysis, prospects and challenges. Neurocomputing, 425:300-322, 2021. 2, 3", + "[48] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15(1):1-28, 2015. 6", + "[49] Janine Thoma, Danda Pani Paudel, and Luc V Gool. Soft contrastive learning for visual localization. Advances in Neural Information Processing Systems, 33:11119-11130, 2020. 1", + "[50] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1, 3" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19836", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8", + "[52] Bram Van Ginneken, Mikkel B Stegmann, and Marco Loog. Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical image analysis, 10(1):19-40, 2006. 6", + "[53] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5", + "[54] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3, 6", + "[55] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature medicine, pages 1-12, 2024. 1, 3", + "[56] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 1, 3", + "[57] Haiqiao Wang, Dong Ni, and Yi Wang. Modet: Learning deformable image registration via motion decomposition transformer. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 740-749. Springer, 2023. 3", + "[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 1, 3", + "[59] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 6", + "[60] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 2022. 1, 2, 3, 5, 6", + "[61] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16590-16599, 2022. 6", + "[62] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Large-scale 3d medical image pre-training with geometric context priors. arXiv preprint arXiv:2410.09890, 2024. 1, 3", + "[63] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simpleyet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1, 2, 3, 5, 6", + "[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 1", + "[66] Liu Yang and Rong Jin. Distance metric learning: A comprehensive survey. Michigan State University, 2(2):4, 2006. 2, 3", + "[67] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence Staib, and James S Duncan. Simcvd: Simple contrastive voxel-wise representation distillation for semi-supervised medical image segmentation. IEEE Transactions on Medical Imaging, 41(9):2228-2237, 2022. 3", + "[68] Chenyu You, Weicheng Dai, Yifei Min, Fenglin Liu, David Clifton, S Kevin Zhou, Lawrence Staib, and James Duncan. Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. Advances in neural information processing systems, 36:9984-10021, 2023.", + "[69] Chenyu You, Weicheng Dai, Fenglin Liu, Yifei Min, Nicha C Dvornek, Xiaoxiao Li, David A Clifton, Lawrence Staib, and James S Duncan. Mine your own anatomy: Revisiting medical image segmentation with extremely limited labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[70] Yi Zhang, Yiji Mao, Xuanyu Lu, Xingyu Zou, Hao Huang, Xinyang Li, Jiayue Li, and Haixian Zhang. From single to universal: tiny lesion detection in medical imaging. Artificial Intelligence Review, 57(8):192, 2024. 1", + "[71] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3", + "[72] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 3, 6", + "[73] Jiuwen Zhu, Yuexiang Li, Yifan Hu, Kai Ma, S Kevin Zhou, and Yefeng Zheng. Rubik's cube+: A self-supervised feature learning framework for 3d medical image analysis. Medical image analysis, 64:101746, 2020. 3" + ], + "bbox": [ + 516, + 92, + 903, + 753 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19837", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_model.json b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ecbdb29a4652e6b2c6d3ab50930745ba59be24f --- /dev/null +++ b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.18, + 0.825, + 0.234 + ], + "angle": 0, + "content": "Yuting He1 Shuo Li1,2,* \n1 Dept. of BME, 2 Dept. of CDS, Case Western Reserve University, US, Cleveland \n{yuting.he4, shuo.li11}@case.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Contrastive learning (CL) has become a cornerstone of self-supervised pretraining (SSP) in foundation models; however, extending CL to pixel-wise representation—crucial for medical vision—remains an open problem. Standard CL formulates SSP as a binary optimization problem (binary CL) where the excessive pursuit of feature dispersion leads to an \"over-dispersion\" problem, breaking pixelwise feature correlation thus disrupting the intra-class distribution. Our vector CL reformulates CL as a vector regression problem, enabling dispersion quantification in pixel-wise pretraining via modeling feature distances in regressing displacement vectors. To implement this novel paradigm, we propose the COntrast in Vector Regression (COVER) framework. COVER establishes an extendable vector-based self-learning, enforces a consistent optimization flow from vector regression to distance modeling, and leverages a vector pyramid architecture for granularity adaptation, thus preserving pixel-wise feature correlations in SSP. Extensive experiments across 8 tasks, spanning 2 dimensions and 4 modalities, show that COVER significantly improves pixel-wise SSP, advancing generalizable medical visual foundation models. Codes will be publicly available at https://github.com/YutingHe-list/COVER." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.222, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Contrastive learning (CL) [4, 7-9, 14, 37, 50] has become a cornerstone of self-supervised pretraining (SSP) [30] in medical visual foundation models (MVFMs) [6, 21, 33, 55, 62, 71]. However, applying CL in a pixel-wise manner, an approach pivotal for disentangling the clinically critical semantics within medical images [70], remains an open problem. Current preliminary works [6, 55, 62] adopt imagewise contrast strategies derived from natural image studies [7-9, 14, 37], which emphasize global representations [60]. In medical images, the global context is often similar due to the consistent anatomies in human body [34], lead" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.267, + 0.897, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.363, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Figure 1. Over-dispersion problem: Binary CL excessively pursues to disperse features, disrupting intra-class distribution and struggling models to disentangle the underlying semantics." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.416, + 0.907, + 0.522 + ], + "angle": 0, + "content": "ing to a lack of global diversity that will cause CL models to converge to trivial solutions [45, 58]. Recent evidence highlights the critical importance of local features [29, 60, 64, 70] and leading results [15, 41, 65] in medical images increasingly incorporate pixel-wise methods. This observation raises a fundamental question: Can CL be reformulated as a truly pixel-wise SSP for MVFMs?" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.523, + 0.909, + 0.75 + ], + "angle": 0, + "content": "An obstacle to answering this question is the overdispersion problem [23, 45, 58]. The standard CL formulates the SSP as a binary optimization problem (termed binary \\(CL\\)), where feature distances in positive or negative pairs are directly minimized or maximized [36], leading to excessive dispersion of features without degree modeling (Fig.1-b). However, recent findings indicate that excessive pursuit of dispersion will disrupt intra-class distribution [56, 58]. This rigid contrast imposes significant limitations on medical vision tasks that require high sensitivity to subtle variations and structures, like thin vessels and subtle lesions, which risk being blurred or indistinguishable (Fig.1-c). While some attempts [5, 63] softened the contrast or introduced regularization [58] to mitigate this problem, they remain fundamentally reliant on the binary CL." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In pixel-wise pretraining, the over-dispersion problem renders the binary CL particularly ineffective (Fig.1 a-c). Pixel-wise features are distributed across the image grid, where semantics vary continuously [13], making them inherently correlated. Unlike discrete classification tasks, continuous signals lack a natural definition of distinct categories [16, 49], leading to ambiguity near proximity divisions. The excessive pursuit of dispersion breaks these correlations so that it aggravates the disruption of the inherent intra-class distribution, struggling pretrained models to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: shuo.li11@case.edu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.472, + 0.107 + ], + "angle": 0, + "content": "disentangle the underlying semantics hidden in pixels [2]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.109, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Motivation: Modeling distances for pixel-wise features is essential to quantify their dispersion degrees. The binary CL follows a max-min contrast for feature distances, neglecting explicit quantification of dispersion (Fig.1-b). Ideally, suppose we can predefine a ground truth distance \\(\\alpha\\) to indicate the desired feature dispersion, the model will be trained to align feature distances accordingly [47, 66]. Such a CL with explicit distance modeling would enable precise quantification of dispersion while preserving feature correlations (Fig.1-d). Unfortunately, in practice, these ground truth distances are task-dependent [40], making them unmeasurable in the SSP and mismatched with the general objectives of foundation models [3]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.58 + ], + "angle": 0, + "content": "This paper proposes a solution for the over-dispersion problem in pixel-wise CL by formulating a vector contrastive learning. The key insight is that feature distances inherently encode semantic correspondences and can be represented as displacement vectors in image space (Fig.2). When pixel-wise features align on an image grid, such vectors (pointing from source to target coordinates) provide a structured way to model correspondences. Therefore, instead of directly minimizing \\( |\\alpha - d'| \\) to model distances \\( d' \\) in the embedding space, we formulate a function \\( \\mathcal{V} \\) that relates distances to vectors \\( v' \\) in the image space. This reformulates the CL as a vector regression problem: \\( |v - \\mathcal{V}(d')| \\) (formulated in Sec.3.1), where \\( v \\) is the ground truth vector. By leveraging the properties of vectors, this approach structurally regulates feature dispersion, preserving semantic continuity. Incorporating measurable magnitudes and directions from the image space, vector CL provides a principled way to guide CL with distance modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.793 + ], + "angle": 0, + "content": "The implementation of vector CL remains unexplored, hindered by two questions: Q1. How to construct a self-learning paradigm with free ground truth vector \\( v \\) that can be extended across diverse medical images? Q2. How to formulate the function \\( \\mathcal{V} \\) to ensure a consistent optimization flow from vector regression to distance modeling while aligning with the properties of correspondence? Although preliminary studies [19, 22] explored vectors in medical SSP, they lacked a structured formulation to link distances and vectors, causing inconsistent optimization flows. For the first time, we formulate the vector CL and implement it as a COntrast in Vector Regression (COVER) framework, which models distance in pixel-wise CL, enabling quantitative dispersion in pixel-wise SSP via three key innovations:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1) Self-Vector Regression (SeVR) (for \\( Q1 \\)) constructs an extendable self-space transformation mechanism for vector CL with free ground truth vectors \\( v \\) (Fig.3-a, Sec.3.2). Specifically, it generates a displacement vector field (DVF), a dense field of vectors on the image grid, capturing pixelwise correspondences between two transformed views of a medical image. The pixel-wise distances are modeled" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.088, + 0.896, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.203, + 0.905, + 0.245 + ], + "angle": 0, + "content": "Figure 2. Motivation: Our vector CL unifies the distance modeling \\( |d - d'| \\) and displacement vector regression \\( |v - \\mathcal{V}(d')| \\), enabling CL with quantitative dispersion via the optimization flow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.905, + 0.315 + ], + "angle": 0, + "content": "through vector regression supervised by the DVF. This synthesis-based process enables a scalable and annotation-free SSP algorithm, facilitating vector CL across a wide variety of medical images (free from paired images [19, 22])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.905, + 0.496 + ], + "angle": 0, + "content": "2) Mixture of Vectors (MoV) (for \\(Q2\\)) formulates a mapping function with consistent optimization flow from vector regression to distance modeling (Fig.3-b, Sec.3.3) with two properties: i. Spatial continuity. It designs a vector embedding unit (VEU) that encodes continuous spatial relationships in a vector template to map pixel-wise feature distances to vectors, thus avoiding the artificial division [29, 60, 64] and preserving the feature correlation. ii. Correspondence ambiguity. It constructs a multi-vector integration (MVI) strategy that extracts multiple vectors indicating potential correspondences with diverse feature concerns, enhancing bias adaptability for pretrained models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.633 + ], + "angle": 0, + "content": "3) Vector Pyramid Aggregation (VPA) (for \\(Q2\\)) formulates multiscalarity of correspondence via stacking the MoV in a pyramid-like architecture (Fig.3-c, Sec.3.4). It predicts vectors across multiple feature scales and fuses them to capture correspondences for semantic objects at varying scale levels. Therefore, it enables the vector CL on different scales with low computation cost, training the model to represent multiscale features and improving granularity adaptability for pretrained models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.815 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: 1) For the first time, we pioneer a novel CL paradigm, vector CL, overcoming the over-dispersion problem in standard binary CL while preserving intra-class distribution. 2) We propose the COVER framework, which implements the vector CL for an extendable pixel-wise SSP in MVFMs. 3) Our SeVR establishes a versatile self-learning paradigm extendable across diverse medical images. 4) Our MoV maps feature distances as vectors, enabling a consistent optimization flow that quantifies dispersion while maintaining feature correlations. 5) Our VPA captures multiscale correspondences, enhancing representations across varying granularity scales." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.661, + 0.846 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "1) Medical image self-supervised pretraining [30] (SSP), as a training algorithm without annotation, is effectively promoting the programming of medical visual foundation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.242 + ], + "angle": 0, + "content": "models [6, 21, 33, 71]. It will capture the posterior distribution of the underlying explanatory factors and make models easier to extract useful information [2], thus achieving the general application goal. It always utilizes the generative learning (GL) [28, 38, 54, 72] or contrastive learning (CL) [4, 7, 8, 14, 29, 37, 50, 60, 64, 67-69] paradigms that train model by generating generate meaningful information or contrasting feature pairs. In this paper, we limit our scope to medical visual CL and review the below topics that are relevant to the applications considered in the sequel." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.424 + ], + "angle": 0, + "content": "2) Binary contrastive learning has become a dominant SSP paradigm that learns general representations without annotations. Early works [7, 14, 16] maximize the distance between positive pairs while minimizing distance with negatives, leveraging global representations in natural images and being implemented to medical images [6, 55, 62]. However, the lack of global diversity in medical images makes the global representations insufficient, causing model collapse [45, 58]. While dense binary CL [60, 64] learns dense representations for local features, it suffers from over-dispersion problem due to mini-max dispersion, distorting their correlations and intra-class distributions [56, 58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.575 + ], + "angle": 0, + "content": "3) Distance modeling [47, 66] quantifies features' distances in embedding spaces, preserving their correlations. It focuses on learning similarity metrics and guiding feature alignment with specific distance [24, 46]. This approach aligns closely with human cognition in perceiving continuous objects [10]. This has proven valuable in capturing feature correlations for effective pixel-wise feature representation. However, when aiming for SSP with general representations, predefined distances become impractical [40], posing challenges to direct distance modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.786 + ], + "angle": 0, + "content": "4) Vector regression was originally developed to learn correspondence vectors of visual objects, and has been widely applied in image registration [1, 17, 20, 57], geometric matching [11, 12, 43], etc. Some preliminary SSP works [28, 73] attempt to establish absolute correspondence between semantics and positions but overlook their relative variations. GVSL [19] and GEMINI [22] introduce correspondence learning into medical SSP by predicting vector fields that represent the dense correspondence between objects across images. However, the lack of a formulation for the mapping function between feature distances and vectors has limited its ability to exert control over pixel-wise representations. This paper addresses this gap by formulating the relationship between feature distances and their vectors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.8, + 0.46, + 0.817 + ], + "angle": 0, + "content": "3. COVER: COntrast in Vectors Regression" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our COVER (Fig.3) implement the formulated vector CL (Sec.3.1) via establishing the consistent optimization flow between distance modeling and vector regression (MoV, Sec.3.3), modeling multi-scale representation (VPA, Sec3.4), and learning the vector regression in a self-learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "pipeline (SeVR, Sec.3.2), thus yielding pixel-wise pretraining with quantitative dispersion in medical vision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.13, + 0.649, + 0.145 + ], + "angle": 0, + "content": "3.1. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.38 + ], + "angle": 0, + "content": "3.1.1 Problem formulation The proposed vector CL reformulated the CL as a vector regression problem that models pixel-wise feature distances to quantify their dispersion degrees. It defines a mapping function \\(\\mathcal{V}\\) that links pixel-wise features to displacement vectors with a consistent optimization flow. Given an input medical image \\(x\\) from dataset \\(\\mathcal{D}\\), we apply a random spatial transformation function \\(\\mathcal{T}_{sp}\\) to generate two augmented views \\(x_{a}, x_{b}\\) with their spatial correspondence encoded in a displacement vector field (DVF) \\(\\psi_{ab} = \\{v^{i}\\}_{i\\in \\Omega}\\) (\\(\\Omega\\) is the image grid), where \\(v^{i}\\) represents the coordinate shift from \\(x_{a}\\) to \\(x_{b}\\) at pixel \\(i\\). The network \\(\\mathcal{N}_{\\theta}\\) extracts dense feature maps \\(F_{a} = \\mathcal{N}_{\\theta}(x_{a})\\) and \\(F_{b} = \\mathcal{N}_{\\theta}(x_{b})\\). The function \\(\\mathcal{V}\\) then predicts the DVF \\(\\psi_{ab}' = \\{v'^i\\}_{i\\in \\Omega}\\) from these features, aiming to minimize the disparity between \\(\\psi_{ab}'\\) and \\(\\psi_{ab}\\). This is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.387, + 0.907, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {x \\sim \\mathcal {D}} [ | \\psi_ {a b} ^ {\\prime} - \\psi_ {a b} | ], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.413, + 0.885, + 0.43 + ], + "angle": 0, + "content": "\\[\nw h e r e \\psi_ {a b} ^ {\\prime} = \\mathcal {V} \\left(F _ {a}, F _ {b}\\right), \\psi_ {a b} \\sim \\mathcal {T} _ {s p}, \\psi_ {a b} \\left(x _ {a}\\right) \\equiv x _ {b}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.824, + 0.453 + ], + "angle": 0, + "content": "The \\(\\equiv\\) means both sides have consistent space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.574 + ], + "angle": 0, + "content": "3.1.2 Why does our vector CL drive distance modeling? Because the vector CL unifies feature distance modeling and displacement vector regression mathematically. It constructed an equivalency of vector regression and distance modeling, obtaining a consistent optimization flow, and thus enabling the ability to learn feature distances in the embedding space without ground truth distances. This process is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.582, + 0.905, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D i s t. \\quad M o d.: \\sum_ {i = 0} ^ {I} \\left| \\alpha^ {i} - d ^ {\\prime i} \\right|\\rightarrow 0 (2) \\\\ \\Leftrightarrow \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} \\left| \\alpha^ {i} - d ^ {i} \\right|\\rightarrow 0 (3) \\\\ \\Rightarrow \\left| \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} \\alpha^ {i} - \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} d ^ {\\prime i} \\right|\\rightarrow 0 (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.649, + 0.905, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V e c}. \\operatorname {R e g}.: \\Leftrightarrow | v - \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} d ^ {n i} | \\rightarrow 0. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The ideal distance modeling in a set can be formulated as Equ.2, where the \\(\\alpha^i\\) is the ground truth distance for feature pair \\(i\\) in the set with \\(I\\) distances. Our COVER defines this distance set as the distances on a field of the image grid (Sec.3.3) and embeds a vector template matrix \\(\\mathbb{V}\\) to weight them for Equ.3 which is equivalent with Equ.2. Based on this, Equ.4 can be deduced that illustrates when the individual distance approaches its certain value, the whole will also approach the certain whole. Therefore, our COVER hypothesizes that existing a value \\(v\\) (i.e., vector in DVF \\(\\psi_{ab}\\)) can make \\(\\sum_{i=0}^{I} \\mathbb{V}^i \\alpha^i = v\\), the Equ.5 will be derived from Equ.4, formulating distance modeling in vector regression. Compared with the binary CL, our vector CL makes a tighter generalization bound [32]. We have formulated and analyzed the theory foundations in the supplementary)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.087, + 0.868, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.248, + 0.585, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.248, + 0.865, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.421, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Figure 3. Framework: Our COVER formulates the distance modeling as the regression of displacement vectors (coordinate correspondences), quantifying dispersion degree in pixel-wise CL. a) SeVR (Sec.3.2) implements vector CL in a self-space transformation process. b) MoV (Sec.3.3) formulates feature distances as vectors with the properties of spatial continuity (VEU (i)) and correspondence ambiguity (MVI (ii)). c) VPA (Sec.3.4) formulates multiscalarity of correspondence via integrating the multi-scale vectors for granularity adaptability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.486, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Equivalency of vector regression and distance modeling. While vector regression is not strictly equivalent to conventional distance modeling, our approach of capturing overall distance distribution enhances representation generalizability. Although Equ.4 shows that the aggregated weighted error tends to zero, it does not guarantee that each individual error in Equ.3 vanishes. Fortunately, since the ground truth distance \\(\\alpha^i\\) is task-dependent [40], our approach captures the overall distance distribution rather than enforcing exact matches for each feature pair. This flexibility allows the model to accommodate varying feature biases across different tasks, promoting a more adaptable representation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.699 + ], + "angle": 0, + "content": "3.2. Self-vector regression for extendable self-learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our SeVR (Fig.3-a) constructs and regresses the ground truth vector \\( v \\) in a self-space transformation mechanism. Given a set of medical images \\( \\mathcal{D} \\), an image \\( x \\) is randomly sampled and cropped \\( x \\sim \\mathcal{D} \\), and two transformation operations for appearance \\( t \\sim \\mathcal{T}_{ap} \\) and space \\( \\psi_{ab} \\sim \\mathcal{T}_{sp} \\) are sampled for appearance and space transformation functions \\( \\mathcal{T}_{ap}, \\mathcal{T}_{sp} \\). Especially, the space transformation operation is a DVF \\( \\psi_{ab} \\) which transforms the space of the image \\( x \\) and will play as the ground truth vectors \\( \\{v^i\\}_{i \\in \\Omega} \\) in the field. Our COVER produces the transformation operations for the image \\( x \\) and generates two views \\( x_a = t(x), x_b = \\psi_{ab}(x) \\) with different space and appearance features. Two share-weighted networks \\( \\mathcal{N}_\\theta \\) extract the pixel-wise features with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.561 + ], + "angle": 0, + "content": "\\(L\\) scale levels \\(F_{a} = \\{f_{a}^{l}\\}_{l = 0}^{L} = \\mathcal{N}_{\\theta}(x_{a}), F_{b} = \\{f_{b}^{l}\\}_{l = 0}^{L} = \\mathcal{N}_{\\theta}(x_{b})\\) from these views for multi-scale representation. A function \\(\\mathcal{V}(\\cdot ,\\cdot)\\) (VPA, Sec.3.4) formulates the relationship between the distances of these features \\(F_{a}, F_{b}\\), predicting the DVF \\(\\psi_{ab}^{\\prime}\\) which indicates their correspondences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.623 + ], + "angle": 0, + "content": "Our COVER learns to regress the vector indicates the correspondence to drive the distance modeling, so a vector loss \\(\\mathcal{L}_{vec}\\) calculates a mean absolute error between the predicted \\(\\psi_{ab}^{\\prime}\\) and the ground truth DVF \\(\\psi_{ab}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.634, + 0.907, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v e c} \\left(\\psi_ {a b}, \\psi_ {a b} ^ {\\prime}, \\epsilon_ {a b}\\right) = \\sum_ {i \\in \\{\\epsilon_ {a b} = 1 \\}} \\left| \\psi_ {a b} ^ {i} - \\psi_ {a b} ^ {\\prime i} \\right|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.909, + 0.856 + ], + "angle": 0, + "content": "where the \\(\\epsilon_{ab}\\) is a mask that eliminates the content mismatch of two views \\(x_{a}, x_{b}\\) caused by the spatial transformation and is generated from the ground truth DVF \\(\\psi_{ab}\\). It enables the network to model the correspondence degrees between one feature and a set of features (illustrated in Sec.3.3) for position correspondence, i.e., vectors. An additional consistency loss \\(\\mathcal{L}_{con}\\) is used to maintain the invariance of semantics in space transformation. It calculates a cosine similarity between the dense features from two views, i.e., \\(\\mathcal{L}_{con}(f_{ab}^{4}, f_{b}^{4}, \\epsilon_{ab}) = -\\sum_{i \\in \\{\\epsilon_{ab} = 1\\}} \\frac{f_{ab}^{4,i} \\cdot f_{b}^{4,i}}{||f_{ab}^{4,i}||||f_{b}^{4,i}||}\\), where the \\(f_{ab}^{4} = \\psi_{ab}(f_{a}^{4})\\) is the transformed 4th level dense features \\(f_{a}^{4}\\) for a consistent space of dense feature \\(f_{b}^{4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "The whole optimization of our COVER iteratively minimizes the arg \\(\\min_{\\theta} \\mathcal{L}_{\\text{COVER}} = \\mathcal{L}_{\\text{con}} + \\mathcal{L}_{\\text{vec}}\\) for weights \\(\\theta\\). If it utilizes the SGD, its dynamics can be formulated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19830" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.122 + ], + "angle": 0, + "content": "as \\(\\theta^{*}\\gets \\theta -\\eta \\frac{\\partial\\mathcal{L}_{COVER}(\\theta,\\mathcal{D},\\mathcal{T}_{sp},\\mathcal{T}_{ap})}{\\partial\\theta}\\), where the \\(\\eta\\) is the learning rate. More details are illustrated in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Discussion of properties Our SeVR constructs a self-space transformation mechanism, enabling an extendable self-supervised learning paradigm. It generates DVFs from random transformation as the ground truth vectors \\( v \\) to drive the vector regression without annotation or externally paired data [19, 22]. This property makes it extendable to a wide variety of medical images for large-scale training simplifying data preparation and enhancing generality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.251, + 0.483, + 0.267 + ], + "angle": 0, + "content": "3.3. Mixture of vectors with consistent optimization flow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.348 + ], + "angle": 0, + "content": "As a part of the function \\(\\mathcal{V}\\), our mixture of vectors (MoV, Fig.3-b) \\(\\mathcal{M}\\) maps distances as vectors with consistent optimization flow. It constructs a vector embedding unit (VEU, b-i) and a multi-vector integration (MVI, b-ii) to encode the spatial continuity and correspondence ambiguity:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.564 + ], + "angle": 0, + "content": "Vector embedding unit (VEU) \\(\\mathcal{U}(\\cdot, \\cdot)\\) For a receptive field with \\(N \\times N\\) size (For a clearer illustration, we take 2D situation as example.) on position \\(i\\) of image grid, the distances between the center feature \\(f_{a}^{i}\\) (from view \\(x_{a}\\)) and the set of features \\(f_{b}^{N \\times N} = \\{f_{b}^{(n,n)}\\}_{n=0}^{N}\\) in the receptive field (from view \\(x_{b}\\)) are calculated. Inspired by the attention in transformer [53], it uses the scaled dot-product attention to model their similarity to \\((0,1)\\) thus achieving a distance map \\(D^{N \\times N}\\). To map the feature distances as vectors, we design a fixed vector template matrix \\(\\mathbb{V}^{N \\times N} = \\{\\mathbb{V}^{(n,n)}\\}_{n=0}^{N}\\) that indicates the vectors in which the center coordinate pointing to the coordinates in the receptive field. This vector template \\(\\mathbb{V}^{N \\times N}\\) is matrix multiplied with the distance map \\(D^{N \\times N}\\) for a vector \\(v_{ab}^{\\prime i}\\). The function of our VEU \\(\\mathcal{U}(\\cdot, \\cdot)\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.569, + 0.483, + 0.593 + ], + "angle": 0, + "content": "\\[\nv _ {a b} ^ {\\prime i} = \\mathcal {U} \\left(f _ {a} ^ {i}, f _ {b} ^ {N \\times N}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {f _ {a} ^ {i} f _ {b} ^ {\\top N \\times N}}{\\tau}\\right) \\mathbb {V} ^ {N \\times N}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.645 + ], + "angle": 0, + "content": "where the \\(\\tau\\) is a scaling factor (\\(\\tau = \\sqrt{c_k}\\) in experiment following [53]). The predicted vector \\(v_{ab}^{\\prime i}\\) is the displacement of feature \\(f_{a}^{i}\\) to align the most similar features in \\(f_{b}^{N\\times N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Multi-vector integration (MVI) Our MVI predicts and integrates multiple vectors to adapt the ambiguous correspondence in a region with continuous semantics caused by varied feature concerns. The features \\( f_{a}, f_{b} \\) with \\( C \\) channels are divided into \\( J \\) groups, and each group is produced by our VEU \\( \\mathcal{U} \\) for a potential vector \\( v'^j = \\mathcal{U}(f_a[j], f_b[j]) \\) where \\( j \\) is the index of group, thus achieving a vector group \\( \\mathbf{v}' = \\{v'^j\\}_j^J \\) representing the correspondences in different feature concerns. Finally, it averages the vectors in the group for a displacement vector \\( u' = \\frac{1}{J}\\sum_{j=0}^{J}v'^j \\) that is compatible with the ambiguity of correspondence. Calculating the displacement vectors for all positions on the image grid, our MoV will output the displacement vector field (DVF) \\( \\psi_{ab}' = \\{u_{ab}'\\}_{i\\in \\Omega} \\) that indicates the correspondence of dense features between view \\( x_{a} \\) and \\( x_{b} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Discussion of properties Our MoV constructs a consistent optimization flow from vector regression to distance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.259 + ], + "angle": 0, + "content": "modeling to quantify the dispersion. It encodes the properties of spatial continuity and correspondence ambiguity of the pixel-wise features, bringing two advantages: 1) Feature correlation preservation: Our VEU constructs a vector template \\(\\mathbb{V}\\) that describes this continuous spatial relationship. This template \\(\\mathbb{V}\\) maps the feature distances to vectors that avoid the artificial division [29, 60, 64] thus preserving the feature correlation. 2) Bias adaptability. Pixel-wise correspondence is uncertain and task-dependent, our MVI models this ambiguity to represent diverse feature concerns, enhancing bias adaptability in general applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.895, + 0.283 + ], + "angle": 0, + "content": "3.4. Vector pyramid aggregation adapts granularity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Stacking our MoV \\(\\mathcal{M}\\) in a pyramid-like architecture, our vector pyramid aggregation (VPA, Fig.3-c) formulates the final mapping function \\(\\nu\\) to model the multiscalarity of correspondence efficiently, training a granularity adaptability. This architecture divides multi-level correspondences according to the scale levels of the dense features. By fusing these scale levels, the VPA will output the DVF \\(\\psi_{ab}^{\\prime}\\) that models both global and local correspondences. In 0th level, dense features \\(f_{a}^{0},f_{b}^{0}\\) are put into our MoV module for a 0th-level DVF \\(\\psi_{ab}^{\\prime 0} = \\mathcal{M}(f_{a}^{0},f_{b}^{0})\\) with a global correspondence for high-level features. Then, in 1st level, the \\(\\psi_{ab}^{\\prime 0}\\) transforms the 1st-level dense features \\(f_{a}^{1}\\) to bridge the space gap between \\(f_{a}^{1}\\) and \\(f_{b}^{1}\\) on the 0th level. The features are further produced by our MoV and further fused with the \\(p s i_{ab}^{\\prime 0}\\) for the 1st-level DVF \\(\\psi_{ab}^{\\prime 1} = \\mathcal{M}(\\psi_{ab}^{\\prime 0}(f_{a}^{1}),f_{b}^{1})\\odot \\psi_{ab}^{\\prime 0}\\), where the “\\(\\odot\\)” is the fusion operation illustrated in Supplementary. The remaining levels are calculated similarly so that the final DVF \\(\\psi_{ab}^{\\prime}\\) integrated multiscale vectors are output after a chain calculation, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.583, + 0.903, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\psi_ {a b} ^ {\\prime} = \\mathcal {V} (F _ {a}, F _ {b}) = H (\\{\\psi_ {a b} ^ {\\prime \\prime} \\} _ {l = 0} ^ {L}) \\quad (\\psi_ {a b} ^ {\\prime 0} = \\mathcal {M} (f _ {a} ^ {0}, f _ {b} ^ {0}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.603, + 0.905, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\psi_ {a b} ^ {\\prime l} = \\mathcal {M} \\left(\\psi_ {a b} ^ {\\prime l - 1} \\left(f _ {a} ^ {l}\\right), f _ {b} ^ {l}\\right) \\odot \\psi_ {a b} ^ {\\prime l - 1}, l = 1, \\dots , L - 1, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.675 + ], + "angle": 0, + "content": "where the \\(H\\) is the chain calculation across all \\(L\\) levels to output the final DVF \\(\\psi_{ab}^{\\prime}\\). Linear upsampling and value doubling are applied to bridge the scale gap of the DVFs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Discussion of properties Correspondence occurs at multiple scales for different semantic granularity, so our VPA formulates this multiscalarity in a pyramid-like architecture, bringing the advantages: 1) Low computational cost: It enables the mapping function \\(\\mathcal{V}\\) in a small receptive field at each level for a large whole receptive field reducing the computation. 2) Multi-scale representation: It learns the multi-scale features with multiple semantic granularities improving the granularity adaptability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.692, + 0.866 + ], + "angle": 0, + "content": "4.1. Comparison study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "4.1.1 Experiment protocol This study has made sufficient experiments to ensure a complete evaluation. Following is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19831" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.904, + 0.34 + ], + "angle": 0, + "content": "
TypeMethodsa) 2D evaluation pretrained on [59]b) 3D evaluation pretrained on [31]AVG Score %
\\(SCR_{S}^{25\\%}\\)\\(PDCX_{R_{C}}\\)\\(KiPA22_{S}^{2D}\\)\\(FIVES_{S}\\)\\(CANDI_{S}\\)\\(FeTA21_{S}\\)\\(KiPA22_{S}^{3D}\\)\\(STOIC_{C}\\)
-Scratch81.890.474.179.484.056.972.472.076.4
GLDenosing [54]83.9(+1.9)92.0(+1.6)60.3(-13.8)77.8(-1.6)83.7(-0.3)52.9(-4.0)70.0(-2.4)65.9(-6.1)73.3(-3.1)
In-painting [38]85.1(+3.3)93.9(+3.5)64.4(-9.7)78.9(-0.5)88.5(+4.5)54.4(-2.5)69.7(-2.7)67.2(-4.8)75.3(-1.1)
Models Genesis [72]86.1(+4.3)92.6(+2.2)66.6(-7.5)79.6(+0.2)88.7(+4.7)55.8(-1.1)75.8(+3.4)75.3(+3.3)77.6(+1.2)
Rotation [28]80.5(-1.3)89.9(-0.5)69.7(-4.4)80.3(+0.9)89.4(+5.4)58.7(+1.8)77.4(+5.0)68.8(-3.2)76.8(+0.4)
BCLSimSiam [8]87.2(+5.4)92.2(+1.8)72.6(-1.5)84.3(+4.9)87.3(+3.3)58.7(+1.8)83.8(+11.4)69.5(-2.5)79.5(+3.1)
BYOL [14]89.4(+7.6)86.6(-3.8)74.1(0)83.3(+3.9)89.7(+5.7)59.2(+2.3)83.6(+11.2)74.0(+2.0)80.0(+5.4)
SimCLR [7]89.0(+7.2)94.7(+4.3)74.4(+0.3)84.5(+5.1)89.2(+5.2)53.4(-3.5)78.9(+6.5)60.7(-11.3)78.1(+1.7)
MoCov2 [9]84.3(+2.5)93.2(+2.8)69.6(-4.5)80.7(+1.3)89.7(+5.7)61.5(+4.6)78.0(+5.6)74.8(+2.8)79.0(+2.6)
DeepCluster [4]84.0(+2.2)93.3(+2.9)72.7(-1.4)81.6(+2.2)89.8(+5.8)57.4(+0.5)79.7(+7.3)65.6(-6.4)78.0(+1.6)
DBCLVADeR [35]85.2(+3.4)92.8(+2.4)62.8(-11.3)78.9(-0.5)87.4(+3.4)43.1(-13.8)72.1(-0.3)73.2(+1.2)74.4(-2.0)
DenseCL [60]85.0(+3.2)92.4(+2.0)70.8(-3.3)79.2(-0.2)87.7(+3.7)43.7(-13.2)74.0(+1.6)58.8(-13.2)74.0(-2.4)
SetSim [61]85.2(+3.4)93.9(+3.5)70.8(-3.3)80.1(+0.7)88.4(+4.4)58.7(+1.8)73.5(+1.1)60.1(-11.9)76.3(-0.1)
DSC-PM [29]90.5(+8.7)91.8(+1.4)77.2(+3.1)83.8(+4.4)88.5(+4.5)52.2(-4.7)79.0(+6.6)59.5(-12.5)77.8(+1.2)
PixPro [64]91.5(+9.7)93.0(+2.6)73.6(-0.5)84.3(+4.9)89.9(+5.9)60.7(+3.8)80.0(+7.6)75.1(+3.1)81.0(+4.6)
Chaitanya et al. [5]87.3(+5.5)90.0(-0.4)76.5(+2.4)84.9(+5.5)87.4(+3.4)53.4(-3.5)70.7(-1.7)67.8(-4.2)77.3(+0.9)
VRGVSL [19]89.7(+7.9)92.5(+2.1)78.9(+4.8)86.2(+6.8)89.1(+5.1)62.6(+5.7)84.3(+11.9)75.4(+3.4)82.3(+5.9)
GEMINI [22]92.4(+10.6)92.9(+2.5)79.1(+5.0)85.3(+5.9)90.0(+6.0)61.7(+4.8)85.0(+12.6)79.5(+7.5)83.2(+6.8)
VCLCOVER (Ours)94.0(+12.2)95.9(+5.5)80.0(+5.9)87.2(+7.8)89.9(+5.9)63.6(+6.7)85.2(+12.8)80.4(+8.4)84.5(+8.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.349, + 0.907, + 0.391 + ], + "angle": 0, + "content": "Table 1. Comparison study: Our COVER achieves the best performance compared with 17 methods on 8 downstream tasks across a) 2D and b) 3D, showing our great adaptation ability. \"AVG\" is the average rank of the row. The blue values are the improvement compared with the \"Scratch\" and the bold values are the highest score. \"S\" and \"C\" are the segmentation and classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.482, + 0.415 + ], + "angle": 0, + "content": "the overview protocol and details are in our Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.483, + 0.657 + ], + "angle": 0, + "content": "1) Materials: This study evaluated our COVER across 8 tasks across 2D and 3D evaluations with 4 modalities to showcase its versatility and advantages. a) \\(2D\\) evaluation: It pretrained networks on the ChestX-ray8 dataset [59] containing 112,120 chest X-ray images. These pretrained models are fine-tuned on four tasks \\((\\mathrm{SCR}_S^{25\\%}\\) [52] (X-ray), \\(\\mathrm{PDCXR}_C\\) [27] (X-ray), \\(\\mathrm{FIVES}_S\\) [25] (fundus), \\(\\mathrm{KiPA22}_{S}^{2D}\\) [18] (CT)). b) \\(3D\\) evaluation: It pretrained networks on 837 3D T1 brain MR images from the PPMI database [31]. These pretrained models are then fine-tuned to three datasets with four tasks \\((\\mathrm{CANDI}_S\\) [26] (MR), \\(\\mathrm{FeTA21}_S\\) [39] (MR), \\(\\mathrm{STOIC}_C\\) [42] (CT), \\(\\mathrm{KiPA22}_{S}^{3D}\\) [18] (CT)). The evaluations involve segmentation (S) and classification (C) tasks. To evaluate the data amount robustness, it used \\(25\\%\\) of the training set in \\(\\mathrm{SCR}_S\\), and more data amount evaluations are performed in our analysis (Fig.4 b)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.779 + ], + "angle": 0, + "content": "2) Comparisons: This study benchmarked COVER with 17 recent or classic methods across four categories: GL [28, 38, 54, 72], binary CL (BCL) [4, 7-9, 14], dense binary CL (DBCL) [5, 29, 35, 60, 61, 64], and vector regression (VR) [19, 22]. It used 2D and 3D U-Nets [44] as the backbone \\(\\mathcal{N}_{\\theta}\\) (for the methods with global prediction, like MoCov2, it adopted the encoder part). For a fair comparison, full-parameter fine-tuning is employed in downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3) Implementation and evaluation metrics: Following [19], all tasks were implemented by PyTorch and optimized by Adam with a learning rate of \\(10^{-4}\\). The models were pretrained with \\(2 \\times 10^{5}\\) iterations for pretraining tasks and fine-tuned with \\(4 \\times 10^{4}\\) iterations for downstream tasks. Training and testing were implemented on NVIDIA A100 SXUM4 GPU with 40 GB memory. This study used the Dice coefficient (DSC) for segmentation tasks and area un" + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.416, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.401, + 0.813, + 0.415 + ], + "angle": 0, + "content": "der curve (AUC) for classification tasks [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.475 + ], + "angle": 0, + "content": "4.1.2 Quantitative analysis Our great adaptability for the downstream tasks shown in Tab. 1 demonstrates our COVER's powerful general representation for medical vision. Two observations can be discovered:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.905, + 0.839 + ], + "angle": 0, + "content": "1) Cross-scale transferability. The SOTA performances on small and large targets show our powerful cross-scale transferability owing to the learned multiscalarity from our VPA a. For small structures like the vessels \\((\\mathrm{FIVES}_S)\\) and brain tissues (CANDI \\(s\\) and \\(\\mathrm{FeTAS}_s\\) ), our approach achieved \\(87.2\\%\\) \\(89.9\\%\\) ,and \\(63.6\\%\\) DSC with the improvements of \\(7.8\\%\\) \\(5.9\\%\\) ,and \\(6.7\\%\\) over the \"Scratch\". These gains stem from learning fine-grained representations at lowerscale levels. b. For large structures including the chest regions \\((\\mathrm{SCR}_{S}^{25\\%})\\) and kidney structures \\((\\mathrm{KiPA22}_{S}^{2D}\\) ,and \\(\\mathrm{KiPA22}_{S}^{3D})\\) , our model achieved \\(94.0\\%\\) \\(80.0\\%\\) ,and \\(85.2\\%\\) DSC improving by over \\(5 \\%\\) compared to \"Scratch\". This is attributed to leveraging high-scale levels for coarse-grained representation. c. For classification tasks (PDCXR \\(c_{r}\\) and STOIC), our COVER achieved AUC improvements of \\(5.5\\%\\) and \\(8.4\\%\\) by learning global representations at level 0. Although the GVSL and GEMINI also have improvements across scales, their inconsistent optimization flow limits the representation potential of pretraining data, resulting in an average score over \\(1 \\%\\) lower than COVER. The BCL methods struggle to enhance all tasks, for example, the DenseCL improved \\(3.2\\%\\) DSC on SCR, but it reduced \\(3.3\\%\\) DSC on \\(\\mathrm{KiPA22}_{S}^{2D}\\) . This arises from their over-dispersion, which introduces biases and compromises on incompatible tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2) Cross-scene adaptability. The significant improvements on consistent \\((\\mathrm{SCR}_S^{25\\%}\\), \\(\\mathrm{PDCXR}_C\\), \\(\\mathrm{CANDI}_S)\\) and inconsistent (KiPA22\\(_S^{2D}\\), FIVES\\(_S\\), FeTAS, KiPA22\\(_S^{3D}\\), STOIC\\(_C\\)) scenes relative to the pretraining data illustrates" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.212, + 0.483, + 0.255 + ], + "angle": 0, + "content": "Figure 4. Implementation analysis: Our COVER has great properties on the scaling of a) fine-tuned parameters, b) fine-tuning data, c) pretraining iterations, and d) model sizes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.264, + 0.483, + 0.52 + ], + "angle": 0, + "content": "our strong cross-scene adaptability. Across all tasks in both 2D and 3D evaluations, COVER outperformed the \"Scratch\" by more than \\(5\\%\\) and achieved the highest average scores \\((84.5\\%)\\). This adaptability arises from our distance modeling which quantifies the dispersion in embedding space, enabling the disentanglement of underlying explanatory factors hidden in low-level sensory data [2]. These learned factors capture the intrinsic properties of medical images, enhancing versatility across diverse scenes. Although most methods improve performance in consistent scenes by leveraging learned transformable knowledge, they struggle in inconsistent scenes. Notably, BCL and DBCL even degrade performance in some tasks due to their over-dispersion in binary CL, which amplifies biases in pretraining data. Especially, DenseCL and VADeR perform particularly poorly in inconsistent scenes like FeTA21\\(_S\\), which reduce over \\(10\\%\\) DSC." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.531, + 0.391, + 0.547 + ], + "angle": 0, + "content": "4.2. Ablation study and model analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.645 + ], + "angle": 0, + "content": "4.3.1 Component ablation The ablation study of the components in our COVER on the 2D evaluation of \\(\\mathrm{SCR}_S^{25\\%}\\) task shows the effectiveness of our proposed innovations. This study took the learning of the consistency loss \\(\\mathcal{L}_{con}\\) in pixel-wise as our base model, and gradually added our proposed components. The results are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.658, + 0.457, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c c c c c} & \\text {base .} \\mathcal {L} _ {\\text {con}} & \\dots + \\text {VEU (SeVR)} & \\dots + \\text {VPA} & \\dots + \\text {MVI} \\\\ \\hline \\text {DSC} \\% & 91.8 & 92.9 & 93.4 & 94.0 \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.825 + ], + "angle": 0, + "content": "The base model \\(\\mathcal{L}_{con}\\) has \\(91.8\\%\\) DSC that is higher than the BYOL's \\(89.4\\%\\) owing to its pixel-wise representation. When adding our VEU, it effectively improves the discrimination, achieving a \\(1.1\\%\\) improvement. Our VPA enhances the multi-scale representation so that the model further improves \\(0.5\\%\\) DSC. Finally, our MVI models the ambiguity of the correspondence for potentially varied feature concerns, thus achieving the highest \\(94.0\\%\\) DSC." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "4.3.2 Hyper-parameter ablation The receptive field \\(N\\) in VEU, the amount of VEUs \\(J\\) in MVI are two new hyperparameters in our COVER. This study took an ablation of them on the \\(\\mathrm{SCR}_S^{25\\%}\\) via linear evaluation to demonstrate their properties. The results are as follows:" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.905, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.212, + 0.905, + 0.268 + ], + "angle": 0, + "content": "Figure 5. Convergence analysis: a) Our COVER is able to promote convergence ability in fine-tuning. b) In the pretraining, the consistency learning causes representation collapse, and when adding our vector regression, the discrimination is enhanced." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.273, + 0.903, + 0.325 + ], + "angle": 0, + "content": "
N =3 × 35 × 57 × 79 × 9
DSC%44.948.454.835.1
J =[2,2,2,1,1][4,4,4,1,1][8,8,8,2,2][12,12,12,3,3]
DSC%55.356.354.848.6
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.507 + ], + "angle": 0, + "content": "For the receptive field size \\( N \\), when enlarging it, its performance increases to \\( 7 \\times 7 \\) and then decreases when it is \\( 9 \\times 9 \\). Because the increase of \\( N \\) will adapt to large spatial correspondence, improving the pretrained representation. However, too large \\( N \\) will introduce more ambiguous semantics, such as vascular with similar features, which will mislead the learning of correspondence. For the amount of VEUs \\( k \\), it also increases and then decreases. When it is [4,4,4,1,1] for the scale levels, the model achieves the best ability. Too many vectors will smooth the optimization, thus weakening the discrimination of learned features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.825 + ], + "angle": 0, + "content": "4.3.3 Implementation analysis This experiment analyzed four implementation factors (Fig.4): a) Fine-tuned parameter amount: Our COVER effectively reduces the fine-tuned parameters. It only fine-tunes one U-Net module to achieve a very significant performance improvement (15.3% DSC higher than GVSL), quickly reaching the bottleneck level that \\(\\mathrm{SCR}_S^{25\\%}\\) task can reach. b) Fine-tuning data amount: Our COVER effectively reduces the fine-tuning data requirement. Only utilizing 5% training data, our COVER can achieve the DSC close to 90% that GVSL and DenseCL have to use 25% and 100% data. c) pretraining iteration: With the progress of the training iterations, the performance (linear evaluation) on \\(\\mathrm{SCR}_S^{25\\%}\\) of our COVER gradually improves and eventually tends to be flat. This illustrates that more training is beneficial for the vector CL, which makes the pretrained models capture more diverse features in detail. d) Model size: With the enlarging of model size, the larger capacity enables our COVER to capture more abundant feature information in the pretraining. This makes the model gain more powerful representability in downstream tasks only with a linear layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "4.3.4 Convergence analysis This study analyzed the convergence performance of our COVER in both pretraining and fine-tuning phases (Fig.5): a) Fine-tuning: In downstream tasks, the network pretrained by our COVER exhibits superior convergence ability. Compared with the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.482, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.224, + 0.483, + 0.253 + ], + "angle": 0, + "content": "Figure 6. Cluster analysis: Our COVER has strong clustering capabilities by quantifying the dispersion degree of features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.262, + 0.483, + 0.444 + ], + "angle": 0, + "content": "\"Scratch\", GVSL, and DenseCL, COVER achieves a significantly lower loss value more quickly in both 2D segmentation (SCR) and 3D classification (STOIC) tasks. This advantage arises from our powerful pixel-wise feature learning that enables the pretrained network to capture the detailed information internal medical images. \\(b\\) pretraining: COVER improves the discrimination of the representation. When relying solely on the consistency loss \\(\\mathcal{L}_{con}\\), the loss quickly collapses to zero, hindering subsequent knowledge learning. By integrating vector regression \\(\\mathcal{L}_{vec}\\), the model learns feature distances, stabilizing the consistency loss with fluctuations and enhancing the discrimination." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.534 + ], + "angle": 0, + "content": "4.3.5 Computation efficiency This study calculated the FLOPs to evaluate the computation efficiency with the expanding of receptive field. For a significant analysis, we focused on the computation in head part, i.e., our VPA. It compared two methods: a. VPA; b. directly taking a large vector template at level 4 (named \"Direct\"). The results are:" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.55, + 0.485, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.719 + ], + "angle": 0, + "content": "When expanding the whole receptive field, our VPA maintains a stable computational cost of 0.2, whereas the \"Direct\" approach shows an exponential increase. Especially when it is \\(121 \\times 121\\), the \"Direct\" has more than 52 times our consumption. Our VPA is able to expand the receptive field for a wider contrast with low additional consumption." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "4.3.6 Distributed clusters from COVER The clusters show that our COVER has strong clustering capabilities by quantifying the dispersion of features. Using pretrained models from our 2D evaluation, this study extracted pixelwise features from the SCR and KiPA22 datasets and visualized their distributions via t-SNE [51] in Fig.6. It compared the clustering patterns of \"Scratch\", GVSL, and DenseCL with our COVER, revealing two key advantages: a) Continuous feature distribution: Our COVER predicts features that are distributed continuously in the embedded space. This continuity arises from our vector CL, which models the feature distance and avoids the excessive dis" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.202, + 0.905, + 0.231 + ], + "angle": 0, + "content": "Figure 7. Multi-scale vectors analysis: Our VPA predicts multilevel correspondence for both global and detailed representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.907, + 0.442 + ], + "angle": 0, + "content": "cretization that blurs small or subtle features. However, DenseCL produces a lot of discrete clusters with small gaps between them, i.e., over-dispersion, leading to fragmented representation and reduced transferability. b) Effective aggregation: Our COVER aggregates features with the same semantics, while other methods show mixed regions. In both SCR and KiPA22 datasets, \"Scratch\" and GVSL exhibit overlapped clusters, as shown in the enlarged patches. Although DenseCL performs well on SCR, it fails on KiPA22, mixing features in this inconsistent scene with the pretraining. Our COVER achieves effective aggregation across datasets because our modeling of distance mitigates the overfitting to the pretraining scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.909, + 0.626 + ], + "angle": 0, + "content": "4.3.7 Analysis of multi-scale vectors in our VPA This study analyzes the vectors at each scale level in our VPA by visualizing spatial transformed images via the predicted DVFs (Fig.7) in our 2D (chest X-ray) and 3D (brain MR) evaluations. In level 0, the DVF predicted from global features is able to align the images on the whole, like the rotation, scaling, etc., driving the learning of global representation. With the expansion of the scales, the correspondences are gradually refined so the details between the images are aligned. Especially, in level 0, the images have high alignment degree on the boundary, so that the models will effectively learn the representation of small or subtle features." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.647, + 0.634, + 0.663 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we reformulate the standard CL as a vector regression problem, vector CL, providing a solution for \"over-dispersion\" problem caused by binary CL in pixelwise medical visual pretraining. Our proposed COcontrast in VEctor Regression (COVER) framework implemented the reformulated vector CL, modeling feature distance for CL in a vector regression process. It enables quantitative dispersion of features which improves representation learning for medical vision applications. Although the unique properties of our vector CL have bright powerful performance, important future work is to involve a larger-scale medical image dataset in training, fostering the development of a large pixel-wise foundation model in the medical field. Beyond advancing MVFM [21] development, COVER opens new frontiers in representations for continuous signals." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19834" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgment This research was supported by the National Science Foundation grants (2306545) and National Institutes of Health grants (1R01HL177813-01, 1R01HL173186-01A1)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.167, + 0.188, + 0.183 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[1] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[2] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[5] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[10] Oliver Contier, Chris I Baker, and Martin N Hebart. Distributed representations of behaviour-derived object dimensions in the human visual system. Nature Human Behaviour, 8(11):2179-2193, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Rethinking optical flow from geometric matching consistent perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1337-1347, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.192, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27488-27497, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[13] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Pires, Zhaohan Guo, Mohammad Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In Neural Information Processing Systems, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[15] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial? Medical Image Analysis, 94:103086, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.386, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[17] Yuting He, Tiantian Li, Rongjun Ge, Jian Yang, Youyong Kong, Jian Zhu, Huazhong Shu, Guanyu Yang, and Shuo Li. Few-shot learning for deformable medical image registration with perception-correspondence decoupling and reverse teaching. IEEE Journal of Biomedical and Health Informatics, 26(3):1177-1187, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.471, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[18] Yuting He, Guanyu Yang, Jian Yang, Rongjun Ge, Youyong Kong, Xiaomei Zhu, Shaobo Zhang, Pengfei Shao, Huazhong Shu, Jean-Louis Dillenseger, et al. Meta grayscale adaptive network for 3d integrated renal structures segmentation. Medical Image Analysis, 71:102055, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[19] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[20] Yuting He, Rongjun Ge, Xiaoming Qi, Yang Chen, Jia-song Wu, Jean-Louis Coatrioux, Guanyu Yang, and Shuo Li. Learning better registration to learn better few-shot medical image segmentation: Authenticity, diversity, and robustness. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2588-2601, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[21] Yuting He, Fuxiang Huang, Xinrui Jiang, Yuxiang Nie, Minghao Wang, Jiguang Wang, and Hao Chen. Foundation model for advancing healthcare: Challenges, opportunities, and future directions. arXiv preprint arXiv:2404.03264, 2024.1,3,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[22] Yuting He, Boyu Wang, Rongjun Ge, Yang Chen, Guanyu Yang, and Shuo Li. Homeomorphism prior for false positive and negative problem in medical image dense contrastive representation learning, 2025. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] John Hinde and Clarice GB Demétrio. Overdispersion: models and estimation. Computational statistics & data analysis, 27(2):151-170, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third inter" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19835" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "national workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[25] Kai Jin, Xingru Huang, Jingxing Zhou, Yunxiang Li, Yan Yan, Yibao Sun, Qianni Zhang, Yaqi Wang, and Juan Ye. Fives: A fundus image dataset for artificial intelligence based vessel segmentation. Scientific data, 9(1):475, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[26] David N Kennedy, Christian Haselgrove, Steven M Hodge, Pallavi S Rane, Nikos Makris, and Jean A Frazier. Candishare: A resource for pediatric neuroimaging data. Neuroinformatics, 10(3):319, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[27] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. cell, 172(5):1122-1131, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[28] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[29] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[30] Siladittya Manna, Saumik Bhattacharya, and Umapada Pal. Self-supervised visual representation learning for medical image analysis: A comprehensive survey. Transactions on Machine Learning Research. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[31] Kenneth Marek, Danna Jennings, Shirley Lasch, Andrew Siderowf, Caroline Tanner, Tanya Simuni, Chris Coffey, Karl Kieburtz, Emily Flagg, Sohini Chowdhury, et al. The parkinson progression marker initiative (ppmi). Progress in neurobiology, 95(4):629-635, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.475, + 0.595 + ], + "angle": 0, + "content": "[32] Mehryar Mohri. Foundations of machine learning, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[33] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[34] Frank H Netter. Atlas of human anatomy, Professional Edition E-Book: including NetterReference. com Access with full downloadable image Bank. Elsevier health sciences, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[35] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[38] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[39] Kelly Payette, Hongwei Bran Li, Priscille de Dumast, Roxane Licandro, Hui Ji, Md Mahfuzur Rahman Siddiquee, Daguang Xu, Andriy Myronenko, Hao Liu, Yuchen Pei, et al. Fetal brain tissue annotation and segmentation challenge results. Medical image analysis, 88:102833, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[40] Guangzhi Qu, Salim Hariri, and Mazin Yousif. A new dependency and correlation analysis for features. IEEE Transactions on Knowledge and Data Engineering, 17(9):1199-1207, 2005. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[41] Quan Quan, Qingsong Yao, Heqin Zhu, and S Kevin Zhou. Igu-aug: Information-guided unsupervised augmentation and pixel-wise contrastive learning for medical image analysis. IEEE Transactions on Medical Imaging, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.351, + 0.905, + 0.435 + ], + "angle": 0, + "content": "[42] Marie-Pierre Revel, Samia Boussouar, Constance de Margerie-Mellon, Inès Saab, Thibaut Lapotre, Dominique Mompoint, Guillaume Chassagnon, Audrey Milon, Mathieu Lederlin, Souhail Bennani, et al. Study of thoracic ct in疫情-19: The stoic project. Radiology, 301(1):E361-E370, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.437, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[43] Ignacio Rocco, Relja Arandjelovic, and Josef Sivic. Convolutional neural network architecture for geometric matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6148-6157, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.496, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.638 + ], + "angle": 0, + "content": "[45] Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning, pages 5628-5637. PMLR, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.639, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[46] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.684, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[47] Juan Luis Suárez, Salvador García, and Francisco Herrera. A tutorial on distance metric learning: Mathematical foundations, algorithms, experimental analysis, prospects and challenges. Neurocomputing, 425:300-322, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.741, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[48] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15(1):1-28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[49] Janine Thoma, Danda Pani Paudel, and Luc V Gool. Soft contrastive learning for visual localization. Advances in Neural Information Processing Systems, 33:11119-11130, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19836" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[52] Bram Van Ginneken, Mikkel B Stegmann, and Marco Loog. Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical image analysis, 10(1):19-40, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[53] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[54] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[55] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature medicine, pages 1-12, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[56] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[57] Haiqiao Wang, Dong Ni, and Yi Wang. Modet: Learning deformable image registration via motion decomposition transformer. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 740-749. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[59] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[60] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[61] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16590-16599, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[62] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Large-scale 3d medical image pre-training with geometric context priors. arXiv preprint arXiv:2410.09890, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simpleyet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[64] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[66] Liu Yang and Rong Jin. Distance metric learning: A comprehensive survey. Michigan State University, 2(2):4, 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[67] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence Staib, and James S Duncan. Simcvd: Simple contrastive voxel-wise representation distillation for semi-supervised medical image segmentation. IEEE Transactions on Medical Imaging, 41(9):2228-2237, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[68] Chenyu You, Weicheng Dai, Yifei Min, Fenglin Liu, David Clifton, S Kevin Zhou, Lawrence Staib, and James Duncan. Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. Advances in neural information processing systems, 36:9984-10021, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[69] Chenyu You, Weicheng Dai, Fenglin Liu, Yifei Min, Nicha C Dvornek, Xiaoxiao Li, David A Clifton, Lawrence Staib, and James S Duncan. Mine your own anatomy: Revisiting medical image segmentation with extremely limited labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[70] Yi Zhang, Yiji Mao, Xuanyu Lu, Xingyu Zou, Hao Huang, Xinyang Li, Jiayue Li, and Haixian Zhang. From single to universal: tiny lesion detection in medical imaging. Artificial Intelligence Review, 57(8):192, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[71] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[72] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.754 + ], + "angle": 0, + "content": "[73] Jiuwen Zhu, Yuexiang Li, Yifan Hu, Kai Ma, S Kevin Zhou, and Yefeng Zheng. Rubik's cube+: A self-supervised feature learning framework for 3d medical image analysis. Medical image analysis, 64:101746, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19837" + } + ] +] \ No newline at end of file diff --git a/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_origin.pdf b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6233b2c659d9e104ed5e5b053103202348ee5421 --- /dev/null +++ b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/b863182b-c21f-4dc0-8b6a-73de7593a982_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d84719df8226332f5f2fed967f2694888ea503a08c24c47adc85e976ad1a5824 +size 2952240 diff --git a/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/full.md b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..070b0961fc5e9754b6caed764ddc91ddce5e4e6b --- /dev/null +++ b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/full.md @@ -0,0 +1,309 @@ +# Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision + +Yuting He1 Shuo Li1,2,* +1 Dept. of BME, 2 Dept. of CDS, Case Western Reserve University, US, Cleveland +{yuting.he4, shuo.li11}@case.edu + +# Abstract + +Contrastive learning (CL) has become a cornerstone of self-supervised pretraining (SSP) in foundation models; however, extending CL to pixel-wise representation—crucial for medical vision—remains an open problem. Standard CL formulates SSP as a binary optimization problem (binary CL) where the excessive pursuit of feature dispersion leads to an "over-dispersion" problem, breaking pixelwise feature correlation thus disrupting the intra-class distribution. Our vector CL reformulates CL as a vector regression problem, enabling dispersion quantification in pixel-wise pretraining via modeling feature distances in regressing displacement vectors. To implement this novel paradigm, we propose the COntrast in Vector Regression (COVER) framework. COVER establishes an extendable vector-based self-learning, enforces a consistent optimization flow from vector regression to distance modeling, and leverages a vector pyramid architecture for granularity adaptation, thus preserving pixel-wise feature correlations in SSP. Extensive experiments across 8 tasks, spanning 2 dimensions and 4 modalities, show that COVER significantly improves pixel-wise SSP, advancing generalizable medical visual foundation models. Codes will be publicly available at https://github.com/YutingHe-list/COVER. + +# 1. Introduction + +Contrastive learning (CL) [4, 7-9, 14, 37, 50] has become a cornerstone of self-supervised pretraining (SSP) [30] in medical visual foundation models (MVFMs) [6, 21, 33, 55, 62, 71]. However, applying CL in a pixel-wise manner, an approach pivotal for disentangling the clinically critical semantics within medical images [70], remains an open problem. Current preliminary works [6, 55, 62] adopt imagewise contrast strategies derived from natural image studies [7-9, 14, 37], which emphasize global representations [60]. In medical images, the global context is often similar due to the consistent anatomies in human body [34], lead + +![](images/2c75613529c7360414ae6aadd7d0f15c5a3adef53081c9499986c8eccc2985e9.jpg) +Figure 1. Over-dispersion problem: Binary CL excessively pursues to disperse features, disrupting intra-class distribution and struggling models to disentangle the underlying semantics. + +ing to a lack of global diversity that will cause CL models to converge to trivial solutions [45, 58]. Recent evidence highlights the critical importance of local features [29, 60, 64, 70] and leading results [15, 41, 65] in medical images increasingly incorporate pixel-wise methods. This observation raises a fundamental question: Can CL be reformulated as a truly pixel-wise SSP for MVFMs? + +An obstacle to answering this question is the overdispersion problem [23, 45, 58]. The standard CL formulates the SSP as a binary optimization problem (termed binary $CL$ ), where feature distances in positive or negative pairs are directly minimized or maximized [36], leading to excessive dispersion of features without degree modeling (Fig.1-b). However, recent findings indicate that excessive pursuit of dispersion will disrupt intra-class distribution [56, 58]. This rigid contrast imposes significant limitations on medical vision tasks that require high sensitivity to subtle variations and structures, like thin vessels and subtle lesions, which risk being blurred or indistinguishable (Fig.1-c). While some attempts [5, 63] softened the contrast or introduced regularization [58] to mitigate this problem, they remain fundamentally reliant on the binary CL. + +In pixel-wise pretraining, the over-dispersion problem renders the binary CL particularly ineffective (Fig.1 a-c). Pixel-wise features are distributed across the image grid, where semantics vary continuously [13], making them inherently correlated. Unlike discrete classification tasks, continuous signals lack a natural definition of distinct categories [16, 49], leading to ambiguity near proximity divisions. The excessive pursuit of dispersion breaks these correlations so that it aggravates the disruption of the inherent intra-class distribution, struggling pretrained models to + +disentangle the underlying semantics hidden in pixels [2]. + +Motivation: Modeling distances for pixel-wise features is essential to quantify their dispersion degrees. The binary CL follows a max-min contrast for feature distances, neglecting explicit quantification of dispersion (Fig.1-b). Ideally, suppose we can predefine a ground truth distance $\alpha$ to indicate the desired feature dispersion, the model will be trained to align feature distances accordingly [47, 66]. Such a CL with explicit distance modeling would enable precise quantification of dispersion while preserving feature correlations (Fig.1-d). Unfortunately, in practice, these ground truth distances are task-dependent [40], making them unmeasurable in the SSP and mismatched with the general objectives of foundation models [3]. + +This paper proposes a solution for the over-dispersion problem in pixel-wise CL by formulating a vector contrastive learning. The key insight is that feature distances inherently encode semantic correspondences and can be represented as displacement vectors in image space (Fig.2). When pixel-wise features align on an image grid, such vectors (pointing from source to target coordinates) provide a structured way to model correspondences. Therefore, instead of directly minimizing $|\alpha - d'|$ to model distances $d'$ in the embedding space, we formulate a function $\mathcal{V}$ that relates distances to vectors $v'$ in the image space. This reformulates the CL as a vector regression problem: $|v - \mathcal{V}(d')|$ (formulated in Sec.3.1), where $v$ is the ground truth vector. By leveraging the properties of vectors, this approach structurally regulates feature dispersion, preserving semantic continuity. Incorporating measurable magnitudes and directions from the image space, vector CL provides a principled way to guide CL with distance modeling. + +The implementation of vector CL remains unexplored, hindered by two questions: Q1. How to construct a self-learning paradigm with free ground truth vector $v$ that can be extended across diverse medical images? Q2. How to formulate the function $\mathcal{V}$ to ensure a consistent optimization flow from vector regression to distance modeling while aligning with the properties of correspondence? Although preliminary studies [19, 22] explored vectors in medical SSP, they lacked a structured formulation to link distances and vectors, causing inconsistent optimization flows. For the first time, we formulate the vector CL and implement it as a COntrast in Vector Regression (COVER) framework, which models distance in pixel-wise CL, enabling quantitative dispersion in pixel-wise SSP via three key innovations: + +1) Self-Vector Regression (SeVR) (for $Q1$ ) constructs an extendable self-space transformation mechanism for vector CL with free ground truth vectors $v$ (Fig.3-a, Sec.3.2). Specifically, it generates a displacement vector field (DVF), a dense field of vectors on the image grid, capturing pixelwise correspondences between two transformed views of a medical image. The pixel-wise distances are modeled + +![](images/c67df685c15aacd495f309d74cfd98026a55844683228057bde80004b6f8e12b.jpg) +Figure 2. Motivation: Our vector CL unifies the distance modeling $|d - d'|$ and displacement vector regression $|v - \mathcal{V}(d')|$ , enabling CL with quantitative dispersion via the optimization flow. + +through vector regression supervised by the DVF. This synthesis-based process enables a scalable and annotation-free SSP algorithm, facilitating vector CL across a wide variety of medical images (free from paired images [19, 22]). + +2) Mixture of Vectors (MoV) (for $Q2$ ) formulates a mapping function with consistent optimization flow from vector regression to distance modeling (Fig.3-b, Sec.3.3) with two properties: i. Spatial continuity. It designs a vector embedding unit (VEU) that encodes continuous spatial relationships in a vector template to map pixel-wise feature distances to vectors, thus avoiding the artificial division [29, 60, 64] and preserving the feature correlation. ii. Correspondence ambiguity. It constructs a multi-vector integration (MVI) strategy that extracts multiple vectors indicating potential correspondences with diverse feature concerns, enhancing bias adaptability for pretrained models. + +3) Vector Pyramid Aggregation (VPA) (for $Q2$ ) formulates multiscalarity of correspondence via stacking the MoV in a pyramid-like architecture (Fig.3-c, Sec.3.4). It predicts vectors across multiple feature scales and fuses them to capture correspondences for semantic objects at varying scale levels. Therefore, it enables the vector CL on different scales with low computation cost, training the model to represent multiscale features and improving granularity adaptability for pretrained models. + +Our contributions are summarized as follows: 1) For the first time, we pioneer a novel CL paradigm, vector CL, overcoming the over-dispersion problem in standard binary CL while preserving intra-class distribution. 2) We propose the COVER framework, which implements the vector CL for an extendable pixel-wise SSP in MVFMs. 3) Our SeVR establishes a versatile self-learning paradigm extendable across diverse medical images. 4) Our MoV maps feature distances as vectors, enabling a consistent optimization flow that quantifies dispersion while maintaining feature correlations. 5) Our VPA captures multiscale correspondences, enhancing representations across varying granularity scales. + +# 2. Related works + +1) Medical image self-supervised pretraining [30] (SSP), as a training algorithm without annotation, is effectively promoting the programming of medical visual foundation + +models [6, 21, 33, 71]. It will capture the posterior distribution of the underlying explanatory factors and make models easier to extract useful information [2], thus achieving the general application goal. It always utilizes the generative learning (GL) [28, 38, 54, 72] or contrastive learning (CL) [4, 7, 8, 14, 29, 37, 50, 60, 64, 67-69] paradigms that train model by generating generate meaningful information or contrasting feature pairs. In this paper, we limit our scope to medical visual CL and review the below topics that are relevant to the applications considered in the sequel. + +2) Binary contrastive learning has become a dominant SSP paradigm that learns general representations without annotations. Early works [7, 14, 16] maximize the distance between positive pairs while minimizing distance with negatives, leveraging global representations in natural images and being implemented to medical images [6, 55, 62]. However, the lack of global diversity in medical images makes the global representations insufficient, causing model collapse [45, 58]. While dense binary CL [60, 64] learns dense representations for local features, it suffers from over-dispersion problem due to mini-max dispersion, distorting their correlations and intra-class distributions [56, 58]. + +3) Distance modeling [47, 66] quantifies features' distances in embedding spaces, preserving their correlations. It focuses on learning similarity metrics and guiding feature alignment with specific distance [24, 46]. This approach aligns closely with human cognition in perceiving continuous objects [10]. This has proven valuable in capturing feature correlations for effective pixel-wise feature representation. However, when aiming for SSP with general representations, predefined distances become impractical [40], posing challenges to direct distance modeling. + +4) Vector regression was originally developed to learn correspondence vectors of visual objects, and has been widely applied in image registration [1, 17, 20, 57], geometric matching [11, 12, 43], etc. Some preliminary SSP works [28, 73] attempt to establish absolute correspondence between semantics and positions but overlook their relative variations. GVSL [19] and GEMINI [22] introduce correspondence learning into medical SSP by predicting vector fields that represent the dense correspondence between objects across images. However, the lack of a formulation for the mapping function between feature distances and vectors has limited its ability to exert control over pixel-wise representations. This paper addresses this gap by formulating the relationship between feature distances and their vectors. + +# 3. COVER: COntrast in Vectors Regression + +Our COVER (Fig.3) implement the formulated vector CL (Sec.3.1) via establishing the consistent optimization flow between distance modeling and vector regression (MoV, Sec.3.3), modeling multi-scale representation (VPA, Sec3.4), and learning the vector regression in a self-learning + +pipeline (SeVR, Sec.3.2), thus yielding pixel-wise pretraining with quantitative dispersion in medical vision. + +# 3.1. Formulation + +3.1.1 Problem formulation The proposed vector CL reformulated the CL as a vector regression problem that models pixel-wise feature distances to quantify their dispersion degrees. It defines a mapping function $\mathcal{V}$ that links pixel-wise features to displacement vectors with a consistent optimization flow. Given an input medical image $x$ from dataset $\mathcal{D}$ , we apply a random spatial transformation function $\mathcal{T}_{sp}$ to generate two augmented views $x_{a}, x_{b}$ with their spatial correspondence encoded in a displacement vector field (DVF) $\psi_{ab} = \{v^{i}\}_{i\in \Omega}$ ( $\Omega$ is the image grid), where $v^{i}$ represents the coordinate shift from $x_{a}$ to $x_{b}$ at pixel $i$ . The network $\mathcal{N}_{\theta}$ extracts dense feature maps $F_{a} = \mathcal{N}_{\theta}(x_{a})$ and $F_{b} = \mathcal{N}_{\theta}(x_{b})$ . The function $\mathcal{V}$ then predicts the DVF $\psi_{ab}' = \{v'^i\}_{i\in \Omega}$ from these features, aiming to minimize the disparity between $\psi_{ab}'$ and $\psi_{ab}$ . This is formulated as: + +$$ +\min _ {\theta} \mathbb {E} _ {x \sim \mathcal {D}} [ | \psi_ {a b} ^ {\prime} - \psi_ {a b} | ], \tag {1} +$$ + +$$ +w h e r e \psi_ {a b} ^ {\prime} = \mathcal {V} \left(F _ {a}, F _ {b}\right), \psi_ {a b} \sim \mathcal {T} _ {s p}, \psi_ {a b} \left(x _ {a}\right) \equiv x _ {b}. +$$ + +The $\equiv$ means both sides have consistent space. + +3.1.2 Why does our vector CL drive distance modeling? Because the vector CL unifies feature distance modeling and displacement vector regression mathematically. It constructed an equivalency of vector regression and distance modeling, obtaining a consistent optimization flow, and thus enabling the ability to learn feature distances in the embedding space without ground truth distances. This process is formalized as follows: + +$$ +\begin{array}{l} D i s t. \quad M o d.: \sum_ {i = 0} ^ {I} \left| \alpha^ {i} - d ^ {\prime i} \right|\rightarrow 0 (2) \\ \Leftrightarrow \sum_ {i = 0} ^ {I} \mathbb {V} ^ {i} \left| \alpha^ {i} - d ^ {i} \right|\rightarrow 0 (3) \\ \Rightarrow \left| \sum_ {i = 0} ^ {I} \mathbb {V} ^ {i} \alpha^ {i} - \sum_ {i = 0} ^ {I} \mathbb {V} ^ {i} d ^ {\prime i} \right|\rightarrow 0 (4) \\ \end{array} +$$ + +$$ +\operatorname {V e c}. \operatorname {R e g}.: \Leftrightarrow | v - \sum_ {i = 0} ^ {I} \mathbb {V} ^ {i} d ^ {n i} | \rightarrow 0. \tag {5} +$$ + +The ideal distance modeling in a set can be formulated as Equ.2, where the $\alpha^i$ is the ground truth distance for feature pair $i$ in the set with $I$ distances. Our COVER defines this distance set as the distances on a field of the image grid (Sec.3.3) and embeds a vector template matrix $\mathbb{V}$ to weight them for Equ.3 which is equivalent with Equ.2. Based on this, Equ.4 can be deduced that illustrates when the individual distance approaches its certain value, the whole will also approach the certain whole. Therefore, our COVER hypothesizes that existing a value $v$ (i.e., vector in DVF $\psi_{ab}$ ) can make $\sum_{i=0}^{I} \mathbb{V}^i \alpha^i = v$ , the Equ.5 will be derived from Equ.4, formulating distance modeling in vector regression. Compared with the binary CL, our vector CL makes a tighter generalization bound [32]. We have formulated and analyzed the theory foundations in the supplementary). + +![](images/4da739be2103bea3617726ba6fb6e49aa26171fef91c8e8ed3687624316f6c1e.jpg) + +![](images/23c49f305050799788d8e6b5a927ac4f5fe0df57eb6fb62ab1d9292b99170199.jpg) +Figure 3. Framework: Our COVER formulates the distance modeling as the regression of displacement vectors (coordinate correspondences), quantifying dispersion degree in pixel-wise CL. a) SeVR (Sec.3.2) implements vector CL in a self-space transformation process. b) MoV (Sec.3.3) formulates feature distances as vectors with the properties of spatial continuity (VEU (i)) and correspondence ambiguity (MVI (ii)). c) VPA (Sec.3.4) formulates multiscalarity of correspondence via integrating the multi-scale vectors for granularity adaptability. + +![](images/1e7ffb6411fbcdde2b8c7eb4577b8c7330811388f85c50dd7465d87806aec901.jpg) + +Equivalency of vector regression and distance modeling. While vector regression is not strictly equivalent to conventional distance modeling, our approach of capturing overall distance distribution enhances representation generalizability. Although Equ.4 shows that the aggregated weighted error tends to zero, it does not guarantee that each individual error in Equ.3 vanishes. Fortunately, since the ground truth distance $\alpha^i$ is task-dependent [40], our approach captures the overall distance distribution rather than enforcing exact matches for each feature pair. This flexibility allows the model to accommodate varying feature biases across different tasks, promoting a more adaptable representation. + +# 3.2. Self-vector regression for extendable self-learning + +Our SeVR (Fig.3-a) constructs and regresses the ground truth vector $v$ in a self-space transformation mechanism. Given a set of medical images $\mathcal{D}$ , an image $x$ is randomly sampled and cropped $x \sim \mathcal{D}$ , and two transformation operations for appearance $t \sim \mathcal{T}_{ap}$ and space $\psi_{ab} \sim \mathcal{T}_{sp}$ are sampled for appearance and space transformation functions $\mathcal{T}_{ap}, \mathcal{T}_{sp}$ . Especially, the space transformation operation is a DVF $\psi_{ab}$ which transforms the space of the image $x$ and will play as the ground truth vectors $\{v^i\}_{i \in \Omega}$ in the field. Our COVER produces the transformation operations for the image $x$ and generates two views $x_a = t(x), x_b = \psi_{ab}(x)$ with different space and appearance features. Two share-weighted networks $\mathcal{N}_\theta$ extract the pixel-wise features with + +$L$ scale levels $F_{a} = \{f_{a}^{l}\}_{l = 0}^{L} = \mathcal{N}_{\theta}(x_{a}), F_{b} = \{f_{b}^{l}\}_{l = 0}^{L} = \mathcal{N}_{\theta}(x_{b})$ from these views for multi-scale representation. A function $\mathcal{V}(\cdot ,\cdot)$ (VPA, Sec.3.4) formulates the relationship between the distances of these features $F_{a}, F_{b}$ , predicting the DVF $\psi_{ab}^{\prime}$ which indicates their correspondences. + +Our COVER learns to regress the vector indicates the correspondence to drive the distance modeling, so a vector loss $\mathcal{L}_{vec}$ calculates a mean absolute error between the predicted $\psi_{ab}^{\prime}$ and the ground truth DVF $\psi_{ab}$ , i.e., + +$$ +\mathcal {L} _ {v e c} \left(\psi_ {a b}, \psi_ {a b} ^ {\prime}, \epsilon_ {a b}\right) = \sum_ {i \in \{\epsilon_ {a b} = 1 \}} \left| \psi_ {a b} ^ {i} - \psi_ {a b} ^ {\prime i} \right|, \tag {6} +$$ + +where the $\epsilon_{ab}$ is a mask that eliminates the content mismatch of two views $x_{a}, x_{b}$ caused by the spatial transformation and is generated from the ground truth DVF $\psi_{ab}$ . It enables the network to model the correspondence degrees between one feature and a set of features (illustrated in Sec.3.3) for position correspondence, i.e., vectors. An additional consistency loss $\mathcal{L}_{con}$ is used to maintain the invariance of semantics in space transformation. It calculates a cosine similarity between the dense features from two views, i.e., $\mathcal{L}_{con}(f_{ab}^{4}, f_{b}^{4}, \epsilon_{ab}) = -\sum_{i \in \{\epsilon_{ab} = 1\}} \frac{f_{ab}^{4,i} \cdot f_{b}^{4,i}}{||f_{ab}^{4,i}||||f_{b}^{4,i}||}$ , where the $f_{ab}^{4} = \psi_{ab}(f_{a}^{4})$ is the transformed 4th level dense features $f_{a}^{4}$ for a consistent space of dense feature $f_{b}^{4}$ . + +The whole optimization of our COVER iteratively minimizes the arg $\min_{\theta} \mathcal{L}_{\text{COVER}} = \mathcal{L}_{\text{con}} + \mathcal{L}_{\text{vec}}$ for weights $\theta$ . If it utilizes the SGD, its dynamics can be formulated + +as $\theta^{*}\gets \theta -\eta \frac{\partial\mathcal{L}_{COVER}(\theta,\mathcal{D},\mathcal{T}_{sp},\mathcal{T}_{ap})}{\partial\theta}$ , where the $\eta$ is the learning rate. More details are illustrated in Supplementary. + +Discussion of properties Our SeVR constructs a self-space transformation mechanism, enabling an extendable self-supervised learning paradigm. It generates DVFs from random transformation as the ground truth vectors $v$ to drive the vector regression without annotation or externally paired data [19, 22]. This property makes it extendable to a wide variety of medical images for large-scale training simplifying data preparation and enhancing generality. + +# 3.3. Mixture of vectors with consistent optimization flow + +As a part of the function $\mathcal{V}$ , our mixture of vectors (MoV, Fig.3-b) $\mathcal{M}$ maps distances as vectors with consistent optimization flow. It constructs a vector embedding unit (VEU, b-i) and a multi-vector integration (MVI, b-ii) to encode the spatial continuity and correspondence ambiguity: + +Vector embedding unit (VEU) $\mathcal{U}(\cdot, \cdot)$ For a receptive field with $N \times N$ size (For a clearer illustration, we take 2D situation as example.) on position $i$ of image grid, the distances between the center feature $f_{a}^{i}$ (from view $x_{a}$ ) and the set of features $f_{b}^{N \times N} = \{f_{b}^{(n,n)}\}_{n=0}^{N}$ in the receptive field (from view $x_{b}$ ) are calculated. Inspired by the attention in transformer [53], it uses the scaled dot-product attention to model their similarity to $(0,1)$ thus achieving a distance map $D^{N \times N}$ . To map the feature distances as vectors, we design a fixed vector template matrix $\mathbb{V}^{N \times N} = \{\mathbb{V}^{(n,n)}\}_{n=0}^{N}$ that indicates the vectors in which the center coordinate pointing to the coordinates in the receptive field. This vector template $\mathbb{V}^{N \times N}$ is matrix multiplied with the distance map $D^{N \times N}$ for a vector $v_{ab}^{\prime i}$ . The function of our VEU $\mathcal{U}(\cdot, \cdot)$ is + +$$ +v _ {a b} ^ {\prime i} = \mathcal {U} \left(f _ {a} ^ {i}, f _ {b} ^ {N \times N}\right) = \operatorname {s o f t m a x} \left(\frac {f _ {a} ^ {i} f _ {b} ^ {\top N \times N}}{\tau}\right) \mathbb {V} ^ {N \times N}, \tag {7} +$$ + +where the $\tau$ is a scaling factor ( $\tau = \sqrt{c_k}$ in experiment following [53]). The predicted vector $v_{ab}^{\prime i}$ is the displacement of feature $f_{a}^{i}$ to align the most similar features in $f_{b}^{N\times N}$ . + +Multi-vector integration (MVI) Our MVI predicts and integrates multiple vectors to adapt the ambiguous correspondence in a region with continuous semantics caused by varied feature concerns. The features $f_{a}, f_{b}$ with $C$ channels are divided into $J$ groups, and each group is produced by our VEU $\mathcal{U}$ for a potential vector $v'^j = \mathcal{U}(f_a[j], f_b[j])$ where $j$ is the index of group, thus achieving a vector group $\mathbf{v}' = \{v'^j\}_j^J$ representing the correspondences in different feature concerns. Finally, it averages the vectors in the group for a displacement vector $u' = \frac{1}{J}\sum_{j=0}^{J}v'^j$ that is compatible with the ambiguity of correspondence. Calculating the displacement vectors for all positions on the image grid, our MoV will output the displacement vector field (DVF) $\psi_{ab}' = \{u_{ab}'\}_{i\in \Omega}$ that indicates the correspondence of dense features between view $x_{a}$ and $x_{b}$ . + +Discussion of properties Our MoV constructs a consistent optimization flow from vector regression to distance + +modeling to quantify the dispersion. It encodes the properties of spatial continuity and correspondence ambiguity of the pixel-wise features, bringing two advantages: 1) Feature correlation preservation: Our VEU constructs a vector template $\mathbb{V}$ that describes this continuous spatial relationship. This template $\mathbb{V}$ maps the feature distances to vectors that avoid the artificial division [29, 60, 64] thus preserving the feature correlation. 2) Bias adaptability. Pixel-wise correspondence is uncertain and task-dependent, our MVI models this ambiguity to represent diverse feature concerns, enhancing bias adaptability in general applications. + +# 3.4. Vector pyramid aggregation adapts granularity + +Stacking our MoV $\mathcal{M}$ in a pyramid-like architecture, our vector pyramid aggregation (VPA, Fig.3-c) formulates the final mapping function $\nu$ to model the multiscalarity of correspondence efficiently, training a granularity adaptability. This architecture divides multi-level correspondences according to the scale levels of the dense features. By fusing these scale levels, the VPA will output the DVF $\psi_{ab}^{\prime}$ that models both global and local correspondences. In 0th level, dense features $f_{a}^{0},f_{b}^{0}$ are put into our MoV module for a 0th-level DVF $\psi_{ab}^{\prime 0} = \mathcal{M}(f_{a}^{0},f_{b}^{0})$ with a global correspondence for high-level features. Then, in 1st level, the $\psi_{ab}^{\prime 0}$ transforms the 1st-level dense features $f_{a}^{1}$ to bridge the space gap between $f_{a}^{1}$ and $f_{b}^{1}$ on the 0th level. The features are further produced by our MoV and further fused with the $p s i_{ab}^{\prime 0}$ for the 1st-level DVF $\psi_{ab}^{\prime 1} = \mathcal{M}(\psi_{ab}^{\prime 0}(f_{a}^{1}),f_{b}^{1})\odot \psi_{ab}^{\prime 0}$ , where the “ $\odot$ ” is the fusion operation illustrated in Supplementary. The remaining levels are calculated similarly so that the final DVF $\psi_{ab}^{\prime}$ integrated multiscale vectors are output after a chain calculation, i.e. + +$$ +\psi_ {a b} ^ {\prime} = \mathcal {V} (F _ {a}, F _ {b}) = H (\{\psi_ {a b} ^ {\prime \prime} \} _ {l = 0} ^ {L}) \quad (\psi_ {a b} ^ {\prime 0} = \mathcal {M} (f _ {a} ^ {0}, f _ {b} ^ {0})) +$$ + +$$ +\psi_ {a b} ^ {\prime l} = \mathcal {M} \left(\psi_ {a b} ^ {\prime l - 1} \left(f _ {a} ^ {l}\right), f _ {b} ^ {l}\right) \odot \psi_ {a b} ^ {\prime l - 1}, l = 1, \dots , L - 1, \tag {8} +$$ + +where the $H$ is the chain calculation across all $L$ levels to output the final DVF $\psi_{ab}^{\prime}$ . Linear upsampling and value doubling are applied to bridge the scale gap of the DVFs. + +Discussion of properties Correspondence occurs at multiple scales for different semantic granularity, so our VPA formulates this multiscalarity in a pyramid-like architecture, bringing the advantages: 1) Low computational cost: It enables the mapping function $\mathcal{V}$ in a small receptive field at each level for a large whole receptive field reducing the computation. 2) Multi-scale representation: It learns the multi-scale features with multiple semantic granularities improving the granularity adaptability. + +# 4. Experiments + +# 4.1. Comparison study + +4.1.1 Experiment protocol This study has made sufficient experiments to ensure a complete evaluation. Following is + +
TypeMethodsa) 2D evaluation pretrained on [59]b) 3D evaluation pretrained on [31]AVG Score %
\(SCR_{S}^{25\%}\)\(PDCX_{R_{C}}\)\(KiPA22_{S}^{2D}\)\(FIVES_{S}\)\(CANDI_{S}\)\(FeTA21_{S}\)\(KiPA22_{S}^{3D}\)\(STOIC_{C}\)
-Scratch81.890.474.179.484.056.972.472.076.4
GLDenosing [54]83.9(+1.9)92.0(+1.6)60.3(-13.8)77.8(-1.6)83.7(-0.3)52.9(-4.0)70.0(-2.4)65.9(-6.1)73.3(-3.1)
In-painting [38]85.1(+3.3)93.9(+3.5)64.4(-9.7)78.9(-0.5)88.5(+4.5)54.4(-2.5)69.7(-2.7)67.2(-4.8)75.3(-1.1)
Models Genesis [72]86.1(+4.3)92.6(+2.2)66.6(-7.5)79.6(+0.2)88.7(+4.7)55.8(-1.1)75.8(+3.4)75.3(+3.3)77.6(+1.2)
Rotation [28]80.5(-1.3)89.9(-0.5)69.7(-4.4)80.3(+0.9)89.4(+5.4)58.7(+1.8)77.4(+5.0)68.8(-3.2)76.8(+0.4)
BCLSimSiam [8]87.2(+5.4)92.2(+1.8)72.6(-1.5)84.3(+4.9)87.3(+3.3)58.7(+1.8)83.8(+11.4)69.5(-2.5)79.5(+3.1)
BYOL [14]89.4(+7.6)86.6(-3.8)74.1(0)83.3(+3.9)89.7(+5.7)59.2(+2.3)83.6(+11.2)74.0(+2.0)80.0(+5.4)
SimCLR [7]89.0(+7.2)94.7(+4.3)74.4(+0.3)84.5(+5.1)89.2(+5.2)53.4(-3.5)78.9(+6.5)60.7(-11.3)78.1(+1.7)
MoCov2 [9]84.3(+2.5)93.2(+2.8)69.6(-4.5)80.7(+1.3)89.7(+5.7)61.5(+4.6)78.0(+5.6)74.8(+2.8)79.0(+2.6)
DeepCluster [4]84.0(+2.2)93.3(+2.9)72.7(-1.4)81.6(+2.2)89.8(+5.8)57.4(+0.5)79.7(+7.3)65.6(-6.4)78.0(+1.6)
DBCLVADeR [35]85.2(+3.4)92.8(+2.4)62.8(-11.3)78.9(-0.5)87.4(+3.4)43.1(-13.8)72.1(-0.3)73.2(+1.2)74.4(-2.0)
DenseCL [60]85.0(+3.2)92.4(+2.0)70.8(-3.3)79.2(-0.2)87.7(+3.7)43.7(-13.2)74.0(+1.6)58.8(-13.2)74.0(-2.4)
SetSim [61]85.2(+3.4)93.9(+3.5)70.8(-3.3)80.1(+0.7)88.4(+4.4)58.7(+1.8)73.5(+1.1)60.1(-11.9)76.3(-0.1)
DSC-PM [29]90.5(+8.7)91.8(+1.4)77.2(+3.1)83.8(+4.4)88.5(+4.5)52.2(-4.7)79.0(+6.6)59.5(-12.5)77.8(+1.2)
PixPro [64]91.5(+9.7)93.0(+2.6)73.6(-0.5)84.3(+4.9)89.9(+5.9)60.7(+3.8)80.0(+7.6)75.1(+3.1)81.0(+4.6)
Chaitanya et al. [5]87.3(+5.5)90.0(-0.4)76.5(+2.4)84.9(+5.5)87.4(+3.4)53.4(-3.5)70.7(-1.7)67.8(-4.2)77.3(+0.9)
VRGVSL [19]89.7(+7.9)92.5(+2.1)78.9(+4.8)86.2(+6.8)89.1(+5.1)62.6(+5.7)84.3(+11.9)75.4(+3.4)82.3(+5.9)
GEMINI [22]92.4(+10.6)92.9(+2.5)79.1(+5.0)85.3(+5.9)90.0(+6.0)61.7(+4.8)85.0(+12.6)79.5(+7.5)83.2(+6.8)
VCLCOVER (Ours)94.0(+12.2)95.9(+5.5)80.0(+5.9)87.2(+7.8)89.9(+5.9)63.6(+6.7)85.2(+12.8)80.4(+8.4)84.5(+8.1)
+ +Table 1. Comparison study: Our COVER achieves the best performance compared with 17 methods on 8 downstream tasks across a) 2D and b) 3D, showing our great adaptation ability. "AVG" is the average rank of the row. The blue values are the improvement compared with the "Scratch" and the bold values are the highest score. "S" and "C" are the segmentation and classification tasks. + +the overview protocol and details are in our Supplementary. + +1) Materials: This study evaluated our COVER across 8 tasks across 2D and 3D evaluations with 4 modalities to showcase its versatility and advantages. a) $2D$ evaluation: It pretrained networks on the ChestX-ray8 dataset [59] containing 112,120 chest X-ray images. These pretrained models are fine-tuned on four tasks $(\mathrm{SCR}_S^{25\%}$ [52] (X-ray), $\mathrm{PDCXR}_C$ [27] (X-ray), $\mathrm{FIVES}_S$ [25] (fundus), $\mathrm{KiPA22}_{S}^{2D}$ [18] (CT)). b) $3D$ evaluation: It pretrained networks on 837 3D T1 brain MR images from the PPMI database [31]. These pretrained models are then fine-tuned to three datasets with four tasks $(\mathrm{CANDI}_S$ [26] (MR), $\mathrm{FeTA21}_S$ [39] (MR), $\mathrm{STOIC}_C$ [42] (CT), $\mathrm{KiPA22}_{S}^{3D}$ [18] (CT)). The evaluations involve segmentation (S) and classification (C) tasks. To evaluate the data amount robustness, it used $25\%$ of the training set in $\mathrm{SCR}_S$ , and more data amount evaluations are performed in our analysis (Fig.4 b). +2) Comparisons: This study benchmarked COVER with 17 recent or classic methods across four categories: GL [28, 38, 54, 72], binary CL (BCL) [4, 7-9, 14], dense binary CL (DBCL) [5, 29, 35, 60, 61, 64], and vector regression (VR) [19, 22]. It used 2D and 3D U-Nets [44] as the backbone $\mathcal{N}_{\theta}$ (for the methods with global prediction, like MoCov2, it adopted the encoder part). For a fair comparison, full-parameter fine-tuning is employed in downstream tasks. +3) Implementation and evaluation metrics: Following [19], all tasks were implemented by PyTorch and optimized by Adam with a learning rate of $10^{-4}$ . The models were pretrained with $2 \times 10^{5}$ iterations for pretraining tasks and fine-tuned with $4 \times 10^{4}$ iterations for downstream tasks. Training and testing were implemented on NVIDIA A100 SXUM4 GPU with 40 GB memory. This study used the Dice coefficient (DSC) for segmentation tasks and area un + +der curve (AUC) for classification tasks [48]. + +4.1.2 Quantitative analysis Our great adaptability for the downstream tasks shown in Tab. 1 demonstrates our COVER's powerful general representation for medical vision. Two observations can be discovered: +1) Cross-scale transferability. The SOTA performances on small and large targets show our powerful cross-scale transferability owing to the learned multiscalarity from our VPA a. For small structures like the vessels $(\mathrm{FIVES}_S)$ and brain tissues (CANDI $s$ and $\mathrm{FeTAS}_s$ ), our approach achieved $87.2\%$ $89.9\%$ ,and $63.6\%$ DSC with the improvements of $7.8\%$ $5.9\%$ ,and $6.7\%$ over the "Scratch". These gains stem from learning fine-grained representations at lowerscale levels. b. For large structures including the chest regions $(\mathrm{SCR}_{S}^{25\%})$ and kidney structures $(\mathrm{KiPA22}_{S}^{2D}$ ,and $\mathrm{KiPA22}_{S}^{3D})$ , our model achieved $94.0\%$ $80.0\%$ ,and $85.2\%$ DSC improving by over $5 \%$ compared to "Scratch". This is attributed to leveraging high-scale levels for coarse-grained representation. c. For classification tasks (PDCXR $c_{r}$ and STOIC), our COVER achieved AUC improvements of $5.5\%$ and $8.4\%$ by learning global representations at level 0. Although the GVSL and GEMINI also have improvements across scales, their inconsistent optimization flow limits the representation potential of pretraining data, resulting in an average score over $1 \%$ lower than COVER. The BCL methods struggle to enhance all tasks, for example, the DenseCL improved $3.2\%$ DSC on SCR, but it reduced $3.3\%$ DSC on $\mathrm{KiPA22}_{S}^{2D}$ . This arises from their over-dispersion, which introduces biases and compromises on incompatible tasks. +2) Cross-scene adaptability. The significant improvements on consistent $(\mathrm{SCR}_S^{25\%}$ , $\mathrm{PDCXR}_C$ , $\mathrm{CANDI}_S)$ and inconsistent (KiPA22 $_S^{2D}$ , FIVES $_S$ , FeTAS, KiPA22 $_S^{3D}$ , STOIC $_C$ ) scenes relative to the pretraining data illustrates + +![](images/d541e0a43e5c8d83cde799220e17d9d1e0759fdf01579467189ce3b457ecffca.jpg) +Figure 4. Implementation analysis: Our COVER has great properties on the scaling of a) fine-tuned parameters, b) fine-tuning data, c) pretraining iterations, and d) model sizes. + +our strong cross-scene adaptability. Across all tasks in both 2D and 3D evaluations, COVER outperformed the "Scratch" by more than $5\%$ and achieved the highest average scores $(84.5\%)$ . This adaptability arises from our distance modeling which quantifies the dispersion in embedding space, enabling the disentanglement of underlying explanatory factors hidden in low-level sensory data [2]. These learned factors capture the intrinsic properties of medical images, enhancing versatility across diverse scenes. Although most methods improve performance in consistent scenes by leveraging learned transformable knowledge, they struggle in inconsistent scenes. Notably, BCL and DBCL even degrade performance in some tasks due to their over-dispersion in binary CL, which amplifies biases in pretraining data. Especially, DenseCL and VADeR perform particularly poorly in inconsistent scenes like FeTA21 $_S$ , which reduce over $10\%$ DSC. + +# 4.2. Ablation study and model analysis + +4.3.1 Component ablation The ablation study of the components in our COVER on the 2D evaluation of $\mathrm{SCR}_S^{25\%}$ task shows the effectiveness of our proposed innovations. This study took the learning of the consistency loss $\mathcal{L}_{con}$ in pixel-wise as our base model, and gradually added our proposed components. The results are as follows: + +$$ +\begin{array}{c c c c c} & \text {base .} \mathcal {L} _ {\text {con}} & \dots + \text {VEU (SeVR)} & \dots + \text {VPA} & \dots + \text {MVI} \\ \hline \text {DSC} \% & 91.8 & 92.9 & 93.4 & 94.0 \end{array} +$$ + +The base model $\mathcal{L}_{con}$ has $91.8\%$ DSC that is higher than the BYOL's $89.4\%$ owing to its pixel-wise representation. When adding our VEU, it effectively improves the discrimination, achieving a $1.1\%$ improvement. Our VPA enhances the multi-scale representation so that the model further improves $0.5\%$ DSC. Finally, our MVI models the ambiguity of the correspondence for potentially varied feature concerns, thus achieving the highest $94.0\%$ DSC. + +4.3.2 Hyper-parameter ablation The receptive field $N$ in VEU, the amount of VEUs $J$ in MVI are two new hyperparameters in our COVER. This study took an ablation of them on the $\mathrm{SCR}_S^{25\%}$ via linear evaluation to demonstrate their properties. The results are as follows: + +![](images/a2f039526514bc8b8d1b9c900c57db0d02d409d7befa34f905013bc65a2607b2.jpg) +Figure 5. Convergence analysis: a) Our COVER is able to promote convergence ability in fine-tuning. b) In the pretraining, the consistency learning causes representation collapse, and when adding our vector regression, the discrimination is enhanced. + +
N =3 × 35 × 57 × 79 × 9
DSC%44.948.454.835.1
J =[2,2,2,1,1][4,4,4,1,1][8,8,8,2,2][12,12,12,3,3]
DSC%55.356.354.848.6
+ +For the receptive field size $N$ , when enlarging it, its performance increases to $7 \times 7$ and then decreases when it is $9 \times 9$ . Because the increase of $N$ will adapt to large spatial correspondence, improving the pretrained representation. However, too large $N$ will introduce more ambiguous semantics, such as vascular with similar features, which will mislead the learning of correspondence. For the amount of VEUs $k$ , it also increases and then decreases. When it is [4,4,4,1,1] for the scale levels, the model achieves the best ability. Too many vectors will smooth the optimization, thus weakening the discrimination of learned features. + +4.3.3 Implementation analysis This experiment analyzed four implementation factors (Fig.4): a) Fine-tuned parameter amount: Our COVER effectively reduces the fine-tuned parameters. It only fine-tunes one U-Net module to achieve a very significant performance improvement (15.3% DSC higher than GVSL), quickly reaching the bottleneck level that $\mathrm{SCR}_S^{25\%}$ task can reach. b) Fine-tuning data amount: Our COVER effectively reduces the fine-tuning data requirement. Only utilizing 5% training data, our COVER can achieve the DSC close to 90% that GVSL and DenseCL have to use 25% and 100% data. c) pretraining iteration: With the progress of the training iterations, the performance (linear evaluation) on $\mathrm{SCR}_S^{25\%}$ of our COVER gradually improves and eventually tends to be flat. This illustrates that more training is beneficial for the vector CL, which makes the pretrained models capture more diverse features in detail. d) Model size: With the enlarging of model size, the larger capacity enables our COVER to capture more abundant feature information in the pretraining. This makes the model gain more powerful representability in downstream tasks only with a linear layer. + +4.3.4 Convergence analysis This study analyzed the convergence performance of our COVER in both pretraining and fine-tuning phases (Fig.5): a) Fine-tuning: In downstream tasks, the network pretrained by our COVER exhibits superior convergence ability. Compared with the + +![](images/288fc6cee09b2600efda5b1c01f168f210ebe3d84989ad83489f9b20c2dc3c85.jpg) +Figure 6. Cluster analysis: Our COVER has strong clustering capabilities by quantifying the dispersion degree of features. + +"Scratch", GVSL, and DenseCL, COVER achieves a significantly lower loss value more quickly in both 2D segmentation (SCR) and 3D classification (STOIC) tasks. This advantage arises from our powerful pixel-wise feature learning that enables the pretrained network to capture the detailed information internal medical images. $b$ pretraining: COVER improves the discrimination of the representation. When relying solely on the consistency loss $\mathcal{L}_{con}$ , the loss quickly collapses to zero, hindering subsequent knowledge learning. By integrating vector regression $\mathcal{L}_{vec}$ , the model learns feature distances, stabilizing the consistency loss with fluctuations and enhancing the discrimination. + +4.3.5 Computation efficiency This study calculated the FLOPs to evaluate the computation efficiency with the expanding of receptive field. For a significant analysis, we focused on the computation in head part, i.e., our VPA. It compared two methods: a. VPA; b. directly taking a large vector template at level 4 (named "Direct"). The results are: + +![](images/2df4300279be8ec292c9f8df774a0bfb0e74ccab85ca266a35f00fbeb3c2ebb1.jpg) + +When expanding the whole receptive field, our VPA maintains a stable computational cost of 0.2, whereas the "Direct" approach shows an exponential increase. Especially when it is $121 \times 121$ , the "Direct" has more than 52 times our consumption. Our VPA is able to expand the receptive field for a wider contrast with low additional consumption. + +4.3.6 Distributed clusters from COVER The clusters show that our COVER has strong clustering capabilities by quantifying the dispersion of features. Using pretrained models from our 2D evaluation, this study extracted pixelwise features from the SCR and KiPA22 datasets and visualized their distributions via t-SNE [51] in Fig.6. It compared the clustering patterns of "Scratch", GVSL, and DenseCL with our COVER, revealing two key advantages: a) Continuous feature distribution: Our COVER predicts features that are distributed continuously in the embedded space. This continuity arises from our vector CL, which models the feature distance and avoids the excessive dis + +![](images/39ac4c0a47a9d8d885d5d8d139766b38abc90df31083185e1010b41a43db0172.jpg) +Figure 7. Multi-scale vectors analysis: Our VPA predicts multilevel correspondence for both global and detailed representation. + +cretization that blurs small or subtle features. However, DenseCL produces a lot of discrete clusters with small gaps between them, i.e., over-dispersion, leading to fragmented representation and reduced transferability. b) Effective aggregation: Our COVER aggregates features with the same semantics, while other methods show mixed regions. In both SCR and KiPA22 datasets, "Scratch" and GVSL exhibit overlapped clusters, as shown in the enlarged patches. Although DenseCL performs well on SCR, it fails on KiPA22, mixing features in this inconsistent scene with the pretraining. Our COVER achieves effective aggregation across datasets because our modeling of distance mitigates the overfitting to the pretraining scene. + +4.3.7 Analysis of multi-scale vectors in our VPA This study analyzes the vectors at each scale level in our VPA by visualizing spatial transformed images via the predicted DVFs (Fig.7) in our 2D (chest X-ray) and 3D (brain MR) evaluations. In level 0, the DVF predicted from global features is able to align the images on the whole, like the rotation, scaling, etc., driving the learning of global representation. With the expansion of the scales, the correspondences are gradually refined so the details between the images are aligned. Especially, in level 0, the images have high alignment degree on the boundary, so that the models will effectively learn the representation of small or subtle features. + +# 5. Conclusion + +In this paper, we reformulate the standard CL as a vector regression problem, vector CL, providing a solution for "over-dispersion" problem caused by binary CL in pixelwise medical visual pretraining. Our proposed COcontrast in VEctor Regression (COVER) framework implemented the reformulated vector CL, modeling feature distance for CL in a vector regression process. It enables quantitative dispersion of features which improves representation learning for medical vision applications. Although the unique properties of our vector CL have bright powerful performance, important future work is to involve a larger-scale medical image dataset in training, fostering the development of a large pixel-wise foundation model in the medical field. Beyond advancing MVFM [21] development, COVER opens new frontiers in representations for continuous signals. + +Acknowledgment This research was supported by the National Science Foundation grants (2306545) and National Institutes of Health grants (1R01HL177813-01, 1R01HL173186-01A1). + +# References + +[1] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 3 +[2] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2, 3, 7 +[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2 +[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 1, 3, 6 +[5] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 1, 6 +[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1, 3 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6 +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758, 2021. 3, 6 +[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6 +[10] Oliver Contier, Chris I Baker, and Martin N Hebart. Distributed representations of behaviour-derived object dimensions in the human visual system. Nature Human Behaviour, 8(11):2179-2193, 2024. 3 +[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Rethinking optical flow from geometric matching consistent perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1337-1347, 2023. 3 +[12] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27488-27497, 2024. 3 +[13] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 1 +[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Pires, Zhaohan Guo, Mohammad Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In Neural Information Processing Systems, 2020. 1, 3, 6 +[15] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial? Medical Image Analysis, 94:103086, 2024. 1 +[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3 +[17] Yuting He, Tiantian Li, Rongjun Ge, Jian Yang, Youyong Kong, Jian Zhu, Huazhong Shu, Guanyu Yang, and Shuo Li. Few-shot learning for deformable medical image registration with perception-correspondence decoupling and reverse teaching. IEEE Journal of Biomedical and Health Informatics, 26(3):1177-1187, 2021. 3 +[18] Yuting He, Guanyu Yang, Jian Yang, Rongjun Ge, Youyong Kong, Xiaomei Zhu, Shaobo Zhang, Pengfei Shao, Huazhong Shu, Jean-Louis Dillenseger, et al. Meta grayscale adaptive network for 3d integrated renal structures segmentation. Medical Image Analysis, 71:102055, 2021. 6 +[19] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6 +[20] Yuting He, Rongjun Ge, Xiaoming Qi, Yang Chen, Jia-song Wu, Jean-Louis Coatrioux, Guanyu Yang, and Shuo Li. Learning better registration to learn better few-shot medical image segmentation: Authenticity, diversity, and robustness. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2588-2601, 2024. 3 +[21] Yuting He, Fuxiang Huang, Xinrui Jiang, Yuxiang Nie, Minghao Wang, Jiguang Wang, and Hao Chen. Foundation model for advancing healthcare: Challenges, opportunities, and future directions. arXiv preprint arXiv:2404.03264, 2024.1,3,8 +[22] Yuting He, Boyu Wang, Rongjun Ge, Yang Chen, Guanyu Yang, and Shuo Li. Homeomorphism prior for false positive and negative problem in medical image dense contrastive representation learning, 2025. 2, 3, 5, 6 +[23] John Hinde and Clarice GB Demétrio. Overdispersion: models and estimation. Computational statistics & data analysis, 27(2):151-170, 1998. 1 +[24] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third inter + +national workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 3 +[25] Kai Jin, Xingru Huang, Jingxing Zhou, Yunxiang Li, Yan Yan, Yibao Sun, Qianni Zhang, Yaqi Wang, and Juan Ye. Fives: A fundus image dataset for artificial intelligence based vessel segmentation. Scientific data, 9(1):475, 2022. 6 +[26] David N Kennedy, Christian Haselgrove, Steven M Hodge, Pallavi S Rane, Nikos Makris, and Jean A Frazier. Candishare: A resource for pediatric neuroimaging data. Neuroinformatics, 10(3):319, 2012. 6 +[27] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. cell, 172(5):1122-1131, 2018. 6 +[28] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 3, 6 +[29] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 1, 2, 3, 5, 6 +[30] Siladittya Manna, Saumik Bhattacharya, and Umapada Pal. Self-supervised visual representation learning for medical image analysis: A comprehensive survey. Transactions on Machine Learning Research. 1, 2 +[31] Kenneth Marek, Danna Jennings, Shirley Lasch, Andrew Siderowf, Caroline Tanner, Tanya Simuni, Chris Coffey, Karl Kieburtz, Emily Flagg, Sohini Chowdhury, et al. The parkinson progression marker initiative (ppmi). Progress in neurobiology, 95(4):629-635, 2011. 6 +[32] Mehryar Mohri. Foundations of machine learning, 2018. 3 +[33] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1, 3 +[34] Frank H Netter. Atlas of human anatomy, Professional Edition E-Book: including NetterReference. com Access with full downloadable image Bank. Elsevier health sciences, 2014. 1 +[35] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 6 +[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1 +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 1, 3 + +[38] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3, 6 +[39] Kelly Payette, Hongwei Bran Li, Priscille de Dumast, Roxane Licandro, Hui Ji, Md Mahfuzur Rahman Siddiquee, Daguang Xu, Andriy Myronenko, Hao Liu, Yuchen Pei, et al. Fetal brain tissue annotation and segmentation challenge results. Medical image analysis, 88:102833, 2023. 6 +[40] Guangzhi Qu, Salim Hariri, and Mazin Yousif. A new dependency and correlation analysis for features. IEEE Transactions on Knowledge and Data Engineering, 17(9):1199-1207, 2005. 2, 3, 4 +[41] Quan Quan, Qingsong Yao, Heqin Zhu, and S Kevin Zhou. Igu-aug: Information-guided unsupervised augmentation and pixel-wise contrastive learning for medical image analysis. IEEE Transactions on Medical Imaging, 2024. 1 +[42] Marie-Pierre Revel, Samia Boussouar, Constance de Margerie-Mellon, Inès Saab, Thibaut Lapotre, Dominique Mompoint, Guillaume Chassagnon, Audrey Milon, Mathieu Lederlin, Souhail Bennani, et al. Study of thoracic ct in疫情-19: The stoic project. Radiology, 301(1):E361-E370, 2021. 6 +[43] Ignacio Rocco, Relja Arandjelovic, and Josef Sivic. Convolutional neural network architecture for geometric matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6148-6157, 2017. 3 +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6 +[45] Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning, pages 5628-5637. PMLR, 2019. 1, 3 +[46] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 3 +[47] Juan Luis Suárez, Salvador García, and Francisco Herrera. A tutorial on distance metric learning: Mathematical foundations, algorithms, experimental analysis, prospects and challenges. Neurocomputing, 425:300-322, 2021. 2, 3 +[48] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15(1):1-28, 2015. 6 +[49] Janine Thoma, Danda Pani Paudel, and Luc V Gool. Soft contrastive learning for visual localization. Advances in Neural Information Processing Systems, 33:11119-11130, 2020. 1 +[50] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1, 3 + +[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8 +[52] Bram Van Ginneken, Mikkel B Stegmann, and Marco Loog. Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical image analysis, 10(1):19-40, 2006. 6 +[53] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5 +[54] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3, 6 +[55] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature medicine, pages 1-12, 2024. 1, 3 +[56] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 1, 3 +[57] Haiqiao Wang, Dong Ni, and Yi Wang. Modet: Learning deformable image registration via motion decomposition transformer. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 740-749. Springer, 2023. 3 +[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 1, 3 +[59] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 6 +[60] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 2022. 1, 2, 3, 5, 6 +[61] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16590-16599, 2022. 6 +[62] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Large-scale 3d medical image pre-training with geometric context priors. arXiv preprint arXiv:2410.09890, 2024. 1, 3 +[63] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simpleyet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1 + +[64] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1, 2, 3, 5, 6 +[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 1 +[66] Liu Yang and Rong Jin. Distance metric learning: A comprehensive survey. Michigan State University, 2(2):4, 2006. 2, 3 +[67] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence Staib, and James S Duncan. Simcvd: Simple contrastive voxel-wise representation distillation for semi-supervised medical image segmentation. IEEE Transactions on Medical Imaging, 41(9):2228-2237, 2022. 3 +[68] Chenyu You, Weicheng Dai, Yifei Min, Fenglin Liu, David Clifton, S Kevin Zhou, Lawrence Staib, and James Duncan. Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. Advances in neural information processing systems, 36:9984-10021, 2023. +[69] Chenyu You, Weicheng Dai, Fenglin Liu, Yifei Min, Nicha C Dvornek, Xiaoxiao Li, David A Clifton, Lawrence Staib, and James S Duncan. Mine your own anatomy: Revisiting medical image segmentation with extremely limited labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[70] Yi Zhang, Yiji Mao, Xuanyu Lu, Xingyu Zou, Hao Huang, Xinyang Li, Jiayue Li, and Haixian Zhang. From single to universal: tiny lesion detection in medical imaging. Artificial Intelligence Review, 57(8):192, 2024. 1 +[71] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3 +[72] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 3, 6 +[73] Jiuwen Zhu, Yuexiang Li, Yifan Hu, Kai Ma, S Kevin Zhou, and Yefeng Zheng. Rubik's cube+: A self-supervised feature learning framework for 3d medical image analysis. Medical image analysis, 64:101746, 2020. 3 \ No newline at end of file diff --git a/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/images.zip b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8b29e3e55b6c1cd45b9c2c431b35610d6df05c2 --- /dev/null +++ b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe4b4102816d770e66635ccf0d43a55a1546c153747b4d59d5b0ca81b604f88f +size 678915 diff --git a/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/layout.json b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85ef4015788d25c8927049a20f48f843ec0f060a --- /dev/null +++ b/2025/Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision/layout.json @@ -0,0 +1,10941 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "text", + "content": "Vector Contrastive Learning For Pixel-Wise Pretraining In Medical Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 142, + 504, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 142, + 504, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 504, + 185 + ], + "type": "text", + "content": "Yuting He1 Shuo Li1,2,* \n1 Dept. of BME, 2 Dept. of CDS, Case Western Reserve University, US, Cleveland \n{yuting.he4, shuo.li11}@case.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 515 + ], + "type": "text", + "content": "Contrastive learning (CL) has become a cornerstone of self-supervised pretraining (SSP) in foundation models; however, extending CL to pixel-wise representation—crucial for medical vision—remains an open problem. Standard CL formulates SSP as a binary optimization problem (binary CL) where the excessive pursuit of feature dispersion leads to an \"over-dispersion\" problem, breaking pixelwise feature correlation thus disrupting the intra-class distribution. Our vector CL reformulates CL as a vector regression problem, enabling dispersion quantification in pixel-wise pretraining via modeling feature distances in regressing displacement vectors. To implement this novel paradigm, we propose the COntrast in Vector Regression (COVER) framework. COVER establishes an extendable vector-based self-learning, enforces a consistent optimization flow from vector regression to distance modeling, and leverages a vector pyramid architecture for granularity adaptation, thus preserving pixel-wise feature correlations in SSP. Extensive experiments across 8 tasks, spanning 2 dimensions and 4 modalities, show that COVER significantly improves pixel-wise SSP, advancing generalizable medical visual foundation models. Codes will be publicly available at https://github.com/YutingHe-list/COVER." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "content": "Contrastive learning (CL) [4, 7-9, 14, 37, 50] has become a cornerstone of self-supervised pretraining (SSP) [30] in medical visual foundation models (MVFMs) [6, 21, 33, 55, 62, 71]. However, applying CL in a pixel-wise manner, an approach pivotal for disentangling the clinically critical semantics within medical images [70], remains an open problem. Current preliminary works [6, 55, 62] adopt imagewise contrast strategies derived from natural image studies [7-9, 14, 37], which emphasize global representations [60]. In medical images, the global context is often similar due to the consistent anatomies in human body [34], lead" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 211, + 548, + 280 + ], + "blocks": [ + { + "bbox": [ + 321, + 211, + 548, + 280 + ], + "lines": [ + { + "bbox": [ + 321, + 211, + 548, + 280 + ], + "spans": [ + { + "bbox": [ + 321, + 211, + 548, + 280 + ], + "type": "image", + "image_path": "2c75613529c7360414ae6aadd7d0f15c5a3adef53081c9499986c8eccc2985e9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 321 + ], + "type": "text", + "content": "Figure 1. Over-dispersion problem: Binary CL excessively pursues to disperse features, disrupting intra-class distribution and struggling models to disentangle the underlying semantics." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 329, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 329, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 312, + 329, + 555, + 413 + ], + "type": "text", + "content": "ing to a lack of global diversity that will cause CL models to converge to trivial solutions [45, 58]. Recent evidence highlights the critical importance of local features [29, 60, 64, 70] and leading results [15, 41, 65] in medical images increasingly incorporate pixel-wise methods. This observation raises a fundamental question: Can CL be reformulated as a truly pixel-wise SSP for MVFMs?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 414, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 414, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 414, + 556, + 594 + ], + "type": "text", + "content": "An obstacle to answering this question is the overdispersion problem [23, 45, 58]. The standard CL formulates the SSP as a binary optimization problem (termed binary " + }, + { + "bbox": [ + 312, + 414, + 556, + 594 + ], + "type": "inline_equation", + "content": "CL" + }, + { + "bbox": [ + 312, + 414, + 556, + 594 + ], + "type": "text", + "content": "), where feature distances in positive or negative pairs are directly minimized or maximized [36], leading to excessive dispersion of features without degree modeling (Fig.1-b). However, recent findings indicate that excessive pursuit of dispersion will disrupt intra-class distribution [56, 58]. This rigid contrast imposes significant limitations on medical vision tasks that require high sensitivity to subtle variations and structures, like thin vessels and subtle lesions, which risk being blurred or indistinguishable (Fig.1-c). While some attempts [5, 63] softened the contrast or introduced regularization [58] to mitigate this problem, they remain fundamentally reliant on the binary CL." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "content": "In pixel-wise pretraining, the over-dispersion problem renders the binary CL particularly ineffective (Fig.1 a-c). Pixel-wise features are distributed across the image grid, where semantics vary continuously [13], making them inherently correlated. Unlike discrete classification tasks, continuous signals lack a natural definition of distinct categories [16, 49], leading to ambiguity near proximity divisions. The excessive pursuit of dispersion breaks these correlations so that it aggravates the disruption of the inherent intra-class distribution, struggling pretrained models to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Corresponding author: shuo.li11@case.edu" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19827" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 288, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 288, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 288, + 84 + ], + "type": "text", + "content": "disentangle the underlying semantics hidden in pixels [2]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 86, + 295, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 86, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 86, + 295, + 242 + ], + "type": "text", + "content": "Motivation: Modeling distances for pixel-wise features is essential to quantify their dispersion degrees. The binary CL follows a max-min contrast for feature distances, neglecting explicit quantification of dispersion (Fig.1-b). Ideally, suppose we can predefine a ground truth distance " + }, + { + "bbox": [ + 54, + 86, + 295, + 242 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 86, + 295, + 242 + ], + "type": "text", + "content": " to indicate the desired feature dispersion, the model will be trained to align feature distances accordingly [47, 66]. Such a CL with explicit distance modeling would enable precise quantification of dispersion while preserving feature correlations (Fig.1-d). Unfortunately, in practice, these ground truth distances are task-dependent [40], making them unmeasurable in the SSP and mismatched with the general objectives of foundation models [3]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "content": "This paper proposes a solution for the over-dispersion problem in pixel-wise CL by formulating a vector contrastive learning. The key insight is that feature distances inherently encode semantic correspondences and can be represented as displacement vectors in image space (Fig.2). When pixel-wise features align on an image grid, such vectors (pointing from source to target coordinates) provide a structured way to model correspondences. Therefore, instead of directly minimizing " + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "inline_equation", + "content": "|\\alpha - d'|" + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "content": " to model distances " + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "content": " in the embedding space, we formulate a function " + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "content": " that relates distances to vectors " + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "content": " in the image space. This reformulates the CL as a vector regression problem: " + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "inline_equation", + "content": "|v - \\mathcal{V}(d')|" + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "content": " (formulated in Sec.3.1), where " + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 56, + 243, + 295, + 459 + ], + "type": "text", + "content": " is the ground truth vector. By leveraging the properties of vectors, this approach structurally regulates feature dispersion, preserving semantic continuity. Incorporating measurable magnitudes and directions from the image space, vector CL provides a principled way to guide CL with distance modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "text", + "content": "The implementation of vector CL remains unexplored, hindered by two questions: Q1. How to construct a self-learning paradigm with free ground truth vector " + }, + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "text", + "content": " that can be extended across diverse medical images? Q2. How to formulate the function " + }, + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 460, + 295, + 628 + ], + "type": "text", + "content": " to ensure a consistent optimization flow from vector regression to distance modeling while aligning with the properties of correspondence? Although preliminary studies [19, 22] explored vectors in medical SSP, they lacked a structured formulation to link distances and vectors, causing inconsistent optimization flows. For the first time, we formulate the vector CL and implement it as a COntrast in Vector Regression (COVER) framework, which models distance in pixel-wise CL, enabling quantitative dispersion in pixel-wise SSP via three key innovations:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "1) Self-Vector Regression (SeVR) (for " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "Q1" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": ") constructs an extendable self-space transformation mechanism for vector CL with free ground truth vectors " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " (Fig.3-a, Sec.3.2). Specifically, it generates a displacement vector field (DVF), a dense field of vectors on the image grid, capturing pixelwise correspondences between two transformed views of a medical image. The pixel-wise distances are modeled" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 69, + 548, + 152 + ], + "blocks": [ + { + "bbox": [ + 322, + 69, + 548, + 152 + ], + "lines": [ + { + "bbox": [ + 322, + 69, + 548, + 152 + ], + "spans": [ + { + "bbox": [ + 322, + 69, + 548, + 152 + ], + "type": "image", + "image_path": "c67df685c15aacd495f309d74cfd98026a55844683228057bde80004b6f8e12b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 160, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 194 + ], + "type": "text", + "content": "Figure 2. Motivation: Our vector CL unifies the distance modeling " + }, + { + "bbox": [ + 313, + 160, + 553, + 194 + ], + "type": "inline_equation", + "content": "|d - d'|" + }, + { + "bbox": [ + 313, + 160, + 553, + 194 + ], + "type": "text", + "content": " and displacement vector regression " + }, + { + "bbox": [ + 313, + 160, + 553, + 194 + ], + "type": "inline_equation", + "content": "|v - \\mathcal{V}(d')|" + }, + { + "bbox": [ + 313, + 160, + 553, + 194 + ], + "type": "text", + "content": ", enabling CL with quantitative dispersion via the optimization flow." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 201, + 553, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 249 + ], + "type": "text", + "content": "through vector regression supervised by the DVF. This synthesis-based process enables a scalable and annotation-free SSP algorithm, facilitating vector CL across a wide variety of medical images (free from paired images [19, 22])." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 250, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 553, + 392 + ], + "type": "text", + "content": "2) Mixture of Vectors (MoV) (for " + }, + { + "bbox": [ + 313, + 250, + 553, + 392 + ], + "type": "inline_equation", + "content": "Q2" + }, + { + "bbox": [ + 313, + 250, + 553, + 392 + ], + "type": "text", + "content": ") formulates a mapping function with consistent optimization flow from vector regression to distance modeling (Fig.3-b, Sec.3.3) with two properties: i. Spatial continuity. It designs a vector embedding unit (VEU) that encodes continuous spatial relationships in a vector template to map pixel-wise feature distances to vectors, thus avoiding the artificial division [29, 60, 64] and preserving the feature correlation. ii. Correspondence ambiguity. It constructs a multi-vector integration (MVI) strategy that extracts multiple vectors indicating potential correspondences with diverse feature concerns, enhancing bias adaptability for pretrained models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "type": "text", + "content": "3) Vector Pyramid Aggregation (VPA) (for " + }, + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "type": "inline_equation", + "content": "Q2" + }, + { + "bbox": [ + 313, + 393, + 554, + 501 + ], + "type": "text", + "content": ") formulates multiscalarity of correspondence via stacking the MoV in a pyramid-like architecture (Fig.3-c, Sec.3.4). It predicts vectors across multiple feature scales and fuses them to capture correspondences for semantic objects at varying scale levels. Therefore, it enables the vector CL on different scales with low computation cost, training the model to represent multiscale features and improving granularity adaptability for pretrained models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 554, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 645 + ], + "type": "text", + "content": "Our contributions are summarized as follows: 1) For the first time, we pioneer a novel CL paradigm, vector CL, overcoming the over-dispersion problem in standard binary CL while preserving intra-class distribution. 2) We propose the COVER framework, which implements the vector CL for an extendable pixel-wise SSP in MVFMs. 3) Our SeVR establishes a versatile self-learning paradigm extendable across diverse medical images. 4) Our MoV maps feature distances as vectors, enabling a consistent optimization flow that quantifies dispersion while maintaining feature correlations. 5) Our VPA captures multiscale correspondences, enhancing representations across varying granularity scales." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 657, + 404, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 404, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 404, + 670 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "1) Medical image self-supervised pretraining [30] (SSP), as a training algorithm without annotation, is effectively promoting the programming of medical visual foundation" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19828" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "models [6, 21, 33, 71]. It will capture the posterior distribution of the underlying explanatory factors and make models easier to extract useful information [2], thus achieving the general application goal. It always utilizes the generative learning (GL) [28, 38, 54, 72] or contrastive learning (CL) [4, 7, 8, 14, 29, 37, 50, 60, 64, 67-69] paradigms that train model by generating generate meaningful information or contrasting feature pairs. In this paper, we limit our scope to medical visual CL and review the below topics that are relevant to the applications considered in the sequel." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 335 + ], + "type": "text", + "content": "2) Binary contrastive learning has become a dominant SSP paradigm that learns general representations without annotations. Early works [7, 14, 16] maximize the distance between positive pairs while minimizing distance with negatives, leveraging global representations in natural images and being implemented to medical images [6, 55, 62]. However, the lack of global diversity in medical images makes the global representations insufficient, causing model collapse [45, 58]. While dense binary CL [60, 64] learns dense representations for local features, it suffers from over-dispersion problem due to mini-max dispersion, distorting their correlations and intra-class distributions [56, 58]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 336, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 455 + ], + "type": "text", + "content": "3) Distance modeling [47, 66] quantifies features' distances in embedding spaces, preserving their correlations. It focuses on learning similarity metrics and guiding feature alignment with specific distance [24, 46]. This approach aligns closely with human cognition in perceiving continuous objects [10]. This has proven valuable in capturing feature correlations for effective pixel-wise feature representation. However, when aiming for SSP with general representations, predefined distances become impractical [40], posing challenges to direct distance modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 622 + ], + "type": "text", + "content": "4) Vector regression was originally developed to learn correspondence vectors of visual objects, and has been widely applied in image registration [1, 17, 20, 57], geometric matching [11, 12, 43], etc. Some preliminary SSP works [28, 73] attempt to establish absolute correspondence between semantics and positions but overlook their relative variations. GVSL [19] and GEMINI [22] introduce correspondence learning into medical SSP by predicting vector fields that represent the dense correspondence between objects across images. However, the lack of a formulation for the mapping function between feature distances and vectors has limited its ability to exert control over pixel-wise representations. This paper addresses this gap by formulating the relationship between feature distances and their vectors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 633, + 281, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 281, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 281, + 647 + ], + "type": "text", + "content": "3. COVER: COntrast in Vectors Regression" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Our COVER (Fig.3) implement the formulated vector CL (Sec.3.1) via establishing the consistent optimization flow between distance modeling and vector regression (MoV, Sec.3.3), modeling multi-scale representation (VPA, Sec3.4), and learning the vector regression in a self-learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "pipeline (SeVR, Sec.3.2), thus yielding pixel-wise pretraining with quantitative dispersion in medical vision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 102, + 397, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 397, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 397, + 114 + ], + "type": "text", + "content": "3.1. Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": "3.1.1 Problem formulation The proposed vector CL reformulated the CL as a vector regression problem that models pixel-wise feature distances to quantify their dispersion degrees. It defines a mapping function " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " that links pixel-wise features to displacement vectors with a consistent optimization flow. Given an input medical image " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " from dataset " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": ", we apply a random spatial transformation function " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{sp}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " to generate two augmented views " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "x_{a}, x_{b}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " with their spatial correspondence encoded in a displacement vector field (DVF) " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\psi_{ab} = \\{v^{i}\\}_{i\\in \\Omega}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " is the image grid), where " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "v^{i}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " represents the coordinate shift from " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": ". The network " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\theta}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " extracts dense feature maps " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "F_{a} = \\mathcal{N}_{\\theta}(x_{a})" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "F_{b} = \\mathcal{N}_{\\theta}(x_{b})" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " then predicts the DVF " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\psi_{ab}' = \\{v'^i\\}_{i\\in \\Omega}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " from these features, aiming to minimize the disparity between " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\psi_{ab}'" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\psi_{ab}" + }, + { + "bbox": [ + 313, + 120, + 555, + 300 + ], + "type": "text", + "content": ". This is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 306, + 555, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 306, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 334, + 306, + 555, + 324 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {x \\sim \\mathcal {D}} [ | \\psi_ {a b} ^ {\\prime} - \\psi_ {a b} | ], \\tag {1}", + "image_path": "98988d10e3ebbbd58207ab92ea147a79535cd3167e19f6c19aaac98f82cefb2c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 327, + 327, + 541, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 327, + 541, + 340 + ], + "spans": [ + { + "bbox": [ + 327, + 327, + 541, + 340 + ], + "type": "interline_equation", + "content": "w h e r e \\psi_ {a b} ^ {\\prime} = \\mathcal {V} \\left(F _ {a}, F _ {b}\\right), \\psi_ {a b} \\sim \\mathcal {T} _ {s p}, \\psi_ {a b} \\left(x _ {a}\\right) \\equiv x _ {b}.", + "image_path": "a5ed56ae6aaaf4a0cb55e5d188ea5a329c2056f74bbc430ea503107c8e334252.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 347, + 504, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 504, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 504, + 358 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 313, + 347, + 504, + 358 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 313, + 347, + 504, + 358 + ], + "type": "text", + "content": " means both sides have consistent space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 359, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 454 + ], + "type": "text", + "content": "3.1.2 Why does our vector CL drive distance modeling? Because the vector CL unifies feature distance modeling and displacement vector regression mathematically. It constructed an equivalency of vector regression and distance modeling, obtaining a consistent optimization flow, and thus enabling the ability to learn feature distances in the embedding space without ground truth distances. This process is formalized as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 460, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 460, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 328, + 460, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D i s t. \\quad M o d.: \\sum_ {i = 0} ^ {I} \\left| \\alpha^ {i} - d ^ {\\prime i} \\right|\\rightarrow 0 (2) \\\\ \\Leftrightarrow \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} \\left| \\alpha^ {i} - d ^ {i} \\right|\\rightarrow 0 (3) \\\\ \\Rightarrow \\left| \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} \\alpha^ {i} - \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} d ^ {\\prime i} \\right|\\rightarrow 0 (4) \\\\ \\end{array}", + "image_path": "bc11ccf0b9106d2d1abea1c3cd7bdfdc3c707fda0911d78a565132a8c38e5251.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 514, + 553, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 514, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 335, + 514, + 553, + 528 + ], + "type": "interline_equation", + "content": "\\operatorname {V e c}. \\operatorname {R e g}.: \\Leftrightarrow | v - \\sum_ {i = 0} ^ {I} \\mathbb {V} ^ {i} d ^ {n i} | \\rightarrow 0. \\tag {5}", + "image_path": "f25a0394a6ebb1b771090fc7d5d4f4cc26a37a71624cb63e80f4103af1f874dd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "The ideal distance modeling in a set can be formulated as Equ.2, where the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha^i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " is the ground truth distance for feature pair " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " in the set with " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " distances. Our COVER defines this distance set as the distances on a field of the image grid (Sec.3.3) and embeds a vector template matrix " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " to weight them for Equ.3 which is equivalent with Equ.2. Based on this, Equ.4 can be deduced that illustrates when the individual distance approaches its certain value, the whole will also approach the certain whole. Therefore, our COVER hypothesizes that existing a value " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " (i.e., vector in DVF " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{ab}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ") can make " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sum_{i=0}^{I} \\mathbb{V}^i \\alpha^i = v" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", the Equ.5 will be derived from Equ.4, formulating distance modeling in vector regression. Compared with the binary CL, our vector CL makes a tighter generalization bound [32]. We have formulated and analyzed the theory foundations in the supplementary)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19829" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 531, + 195 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 531, + 195 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 531, + 195 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 531, + 195 + ], + "type": "image", + "image_path": "4da739be2103bea3617726ba6fb6e49aa26171fef91c8e8ed3687624316f6c1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 196, + 358, + 324 + ], + "blocks": [ + { + "bbox": [ + 83, + 196, + 358, + 324 + ], + "lines": [ + { + "bbox": [ + 83, + 196, + 358, + 324 + ], + "spans": [ + { + "bbox": [ + 83, + 196, + 358, + 324 + ], + "type": "image", + "image_path": "23c49f305050799788d8e6b5a927ac4f5fe0df57eb6fb62ab1d9292b99170199.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 333, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 54, + 333, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 555, + 378 + ], + "type": "text", + "content": "Figure 3. Framework: Our COVER formulates the distance modeling as the regression of displacement vectors (coordinate correspondences), quantifying dispersion degree in pixel-wise CL. a) SeVR (Sec.3.2) implements vector CL in a self-space transformation process. b) MoV (Sec.3.3) formulates feature distances as vectors with the properties of spatial continuity (VEU (i)) and correspondence ambiguity (MVI (ii)). c) VPA (Sec.3.4) formulates multiscalarity of correspondence via integrating the multi-scale vectors for granularity adaptability." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 359, + 196, + 529, + 324 + ], + "blocks": [ + { + "bbox": [ + 359, + 196, + 529, + 324 + ], + "lines": [ + { + "bbox": [ + 359, + 196, + 529, + 324 + ], + "spans": [ + { + "bbox": [ + 359, + 196, + 529, + 324 + ], + "type": "image", + "image_path": "1e7ffb6411fbcdde2b8c7eb4577b8c7330811388f85c50dd7465d87806aec901.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 384, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 296, + 529 + ], + "type": "text", + "content": "Equivalency of vector regression and distance modeling. While vector regression is not strictly equivalent to conventional distance modeling, our approach of capturing overall distance distribution enhances representation generalizability. Although Equ.4 shows that the aggregated weighted error tends to zero, it does not guarantee that each individual error in Equ.3 vanishes. Fortunately, since the ground truth distance " + }, + { + "bbox": [ + 54, + 384, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\alpha^i" + }, + { + "bbox": [ + 54, + 384, + 296, + 529 + ], + "type": "text", + "content": " is task-dependent [40], our approach captures the overall distance distribution rather than enforcing exact matches for each feature pair. This flexibility allows the model to accommodate varying feature biases across different tasks, promoting a more adaptable representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 295, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 553 + ], + "type": "text", + "content": "3.2. Self-vector regression for extendable self-learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": "Our SeVR (Fig.3-a) constructs and regresses the ground truth vector " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " in a self-space transformation mechanism. Given a set of medical images " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": ", an image " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " is randomly sampled and cropped " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "x \\sim \\mathcal{D}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": ", and two transformation operations for appearance " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{T}_{ap}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " and space " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\psi_{ab} \\sim \\mathcal{T}_{sp}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " are sampled for appearance and space transformation functions " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{ap}, \\mathcal{T}_{sp}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": ". Especially, the space transformation operation is a DVF " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\psi_{ab}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " which transforms the space of the image " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " and will play as the ground truth vectors " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{v^i\\}_{i \\in \\Omega}" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " in the field. Our COVER produces the transformation operations for the image " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " and generates two views " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_a = t(x), x_b = \\psi_{ab}(x)" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " with different space and appearance features. Two share-weighted networks " + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_\\theta" + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": " extract the pixel-wise features with" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "text", + "content": " scale levels " + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "inline_equation", + "content": "F_{a} = \\{f_{a}^{l}\\}_{l = 0}^{L} = \\mathcal{N}_{\\theta}(x_{a}), F_{b} = \\{f_{b}^{l}\\}_{l = 0}^{L} = \\mathcal{N}_{\\theta}(x_{b})" + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "text", + "content": " from these views for multi-scale representation. A function " + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "text", + "content": " (VPA, Sec.3.4) formulates the relationship between the distances of these features " + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "inline_equation", + "content": "F_{a}, F_{b}" + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "text", + "content": ", predicting the DVF " + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime}" + }, + { + "bbox": [ + 313, + 384, + 555, + 444 + ], + "type": "text", + "content": " which indicates their correspondences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "text", + "content": "Our COVER learns to regress the vector indicates the correspondence to drive the distance modeling, so a vector loss " + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vec}" + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "text", + "content": " calculates a mean absolute error between the predicted " + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime}" + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "text", + "content": " and the ground truth DVF " + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "inline_equation", + "content": "\\psi_{ab}" + }, + { + "bbox": [ + 313, + 445, + 555, + 493 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 502, + 555, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 502, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 331, + 502, + 555, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v e c} \\left(\\psi_ {a b}, \\psi_ {a b} ^ {\\prime}, \\epsilon_ {a b}\\right) = \\sum_ {i \\in \\{\\epsilon_ {a b} = 1 \\}} \\left| \\psi_ {a b} ^ {i} - \\psi_ {a b} ^ {\\prime i} \\right|, \\tag {6}", + "image_path": "5ec15975f0e55d7b8327b35330c1534758c2ea204cdb84a3f25fc38ccade1ea0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon_{ab}" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": " is a mask that eliminates the content mismatch of two views " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "x_{a}, x_{b}" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": " caused by the spatial transformation and is generated from the ground truth DVF " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\psi_{ab}" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": ". It enables the network to model the correspondence degrees between one feature and a set of features (illustrated in Sec.3.3) for position correspondence, i.e., vectors. An additional consistency loss " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": " is used to maintain the invariance of semantics in space transformation. It calculates a cosine similarity between the dense features from two views, i.e., " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}(f_{ab}^{4}, f_{b}^{4}, \\epsilon_{ab}) = -\\sum_{i \\in \\{\\epsilon_{ab} = 1\\}} \\frac{f_{ab}^{4,i} \\cdot f_{b}^{4,i}}{||f_{ab}^{4,i}||||f_{b}^{4,i}||}" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "f_{ab}^{4} = \\psi_{ab}(f_{a}^{4})" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": " is the transformed 4th level dense features " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "f_{a}^{4}" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": " for a consistent space of dense feature " + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "inline_equation", + "content": "f_{b}^{4}" + }, + { + "bbox": [ + 313, + 525, + 556, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "The whole optimization of our COVER iteratively minimizes the arg " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\min_{\\theta} \\mathcal{L}_{\\text{COVER}} = \\mathcal{L}_{\\text{con}} + \\mathcal{L}_{\\text{vec}}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " for weights " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ". If it utilizes the SGD, its dynamics can be formulated" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19830" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 96 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 55, + 70, + 294, + 96 + ], + "type": "inline_equation", + "content": "\\theta^{*}\\gets \\theta -\\eta \\frac{\\partial\\mathcal{L}_{COVER}(\\theta,\\mathcal{D},\\mathcal{T}_{sp},\\mathcal{T}_{ap})}{\\partial\\theta}" + }, + { + "bbox": [ + 55, + 70, + 294, + 96 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 55, + 70, + 294, + 96 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 70, + 294, + 96 + ], + "type": "text", + "content": " is the learning rate. More details are illustrated in Supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "type": "text", + "content": "Discussion of properties Our SeVR constructs a self-space transformation mechanism, enabling an extendable self-supervised learning paradigm. It generates DVFs from random transformation as the ground truth vectors " + }, + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 96, + 296, + 193 + ], + "type": "text", + "content": " to drive the vector regression without annotation or externally paired data [19, 22]. This property makes it extendable to a wide variety of medical images for large-scale training simplifying data preparation and enhancing generality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 295, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 211 + ], + "type": "text", + "content": "3.3. Mixture of vectors with consistent optimization flow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 275 + ], + "type": "text", + "content": "As a part of the function " + }, + { + "bbox": [ + 55, + 216, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 216, + 296, + 275 + ], + "type": "text", + "content": ", our mixture of vectors (MoV, Fig.3-b) " + }, + { + "bbox": [ + 55, + 216, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 216, + 296, + 275 + ], + "type": "text", + "content": " maps distances as vectors with consistent optimization flow. It constructs a vector embedding unit (VEU, b-i) and a multi-vector integration (MVI, b-ii) to encode the spatial continuity and correspondence ambiguity:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": "Vector embedding unit (VEU) " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " For a receptive field with " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " size (For a clearer illustration, we take 2D situation as example.) on position " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " of image grid, the distances between the center feature " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "f_{a}^{i}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " (from view " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": ") and the set of features " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "f_{b}^{N \\times N} = \\{f_{b}^{(n,n)}\\}_{n=0}^{N}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " in the receptive field (from view " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": ") are calculated. Inspired by the attention in transformer [53], it uses the scaled dot-product attention to model their similarity to " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " thus achieving a distance map " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "D^{N \\times N}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": ". To map the feature distances as vectors, we design a fixed vector template matrix " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbb{V}^{N \\times N} = \\{\\mathbb{V}^{(n,n)}\\}_{n=0}^{N}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " that indicates the vectors in which the center coordinate pointing to the coordinates in the receptive field. This vector template " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathbb{V}^{N \\times N}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " is matrix multiplied with the distance map " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "D^{N \\times N}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " for a vector " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "v_{ab}^{\\prime i}" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": ". The function of our VEU " + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 275, + 296, + 446 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 450, + 295, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 450, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 65, + 450, + 295, + 469 + ], + "type": "interline_equation", + "content": "v _ {a b} ^ {\\prime i} = \\mathcal {U} \\left(f _ {a} ^ {i}, f _ {b} ^ {N \\times N}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {f _ {a} ^ {i} f _ {b} ^ {\\top N \\times N}}{\\tau}\\right) \\mathbb {V} ^ {N \\times N}, \\tag {7}", + "image_path": "46619a61eb3405b73ed82e9e9c3cd27802001b19545a132c1eb528836b0b3cee.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " is a scaling factor (" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\tau = \\sqrt{c_k}" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " in experiment following [53]). The predicted vector " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "v_{ab}^{\\prime i}" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " is the displacement of feature " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "f_{a}^{i}" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " to align the most similar features in " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "f_{b}^{N\\times N}" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": "Multi-vector integration (MVI) Our MVI predicts and integrates multiple vectors to adapt the ambiguous correspondence in a region with continuous semantics caused by varied feature concerns. The features " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "f_{a}, f_{b}" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " channels are divided into " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " groups, and each group is produced by our VEU " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " for a potential vector " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "v'^j = \\mathcal{U}(f_a[j], f_b[j])" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " is the index of group, thus achieving a vector group " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{v}' = \\{v'^j\\}_j^J" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " representing the correspondences in different feature concerns. Finally, it averages the vectors in the group for a displacement vector " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "u' = \\frac{1}{J}\\sum_{j=0}^{J}v'^j" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " that is compatible with the ambiguity of correspondence. Calculating the displacement vectors for all positions on the image grid, our MoV will output the displacement vector field (DVF) " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\psi_{ab}' = \\{u_{ab}'\\}_{i\\in \\Omega}" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " that indicates the correspondence of dense features between view " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 55, + 511, + 295, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Discussion of properties Our MoV constructs a consistent optimization flow from vector regression to distance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 205 + ], + "type": "text", + "content": "modeling to quantify the dispersion. It encodes the properties of spatial continuity and correspondence ambiguity of the pixel-wise features, bringing two advantages: 1) Feature correlation preservation: Our VEU constructs a vector template " + }, + { + "bbox": [ + 313, + 72, + 554, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 313, + 72, + 554, + 205 + ], + "type": "text", + "content": " that describes this continuous spatial relationship. This template " + }, + { + "bbox": [ + 313, + 72, + 554, + 205 + ], + "type": "inline_equation", + "content": "\\mathbb{V}" + }, + { + "bbox": [ + 313, + 72, + 554, + 205 + ], + "type": "text", + "content": " maps the feature distances to vectors that avoid the artificial division [29, 60, 64] thus preserving the feature correlation. 2) Bias adaptability. Pixel-wise correspondence is uncertain and task-dependent, our MVI models this ambiguity to represent diverse feature concerns, enhancing bias adaptability in general applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 211, + 547, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 547, + 224 + ], + "type": "text", + "content": "3.4. Vector pyramid aggregation adapts granularity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": "Stacking our MoV " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " in a pyramid-like architecture, our vector pyramid aggregation (VPA, Fig.3-c) formulates the final mapping function " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " to model the multiscalarity of correspondence efficiently, training a granularity adaptability. This architecture divides multi-level correspondences according to the scale levels of the dense features. By fusing these scale levels, the VPA will output the DVF " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " that models both global and local correspondences. In 0th level, dense features " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{a}^{0},f_{b}^{0}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " are put into our MoV module for a 0th-level DVF " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime 0} = \\mathcal{M}(f_{a}^{0},f_{b}^{0})" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " with a global correspondence for high-level features. Then, in 1st level, the " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime 0}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " transforms the 1st-level dense features " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{a}^{1}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " to bridge the space gap between " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{a}^{1}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "f_{b}^{1}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " on the 0th level. The features are further produced by our MoV and further fused with the " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "p s i_{ab}^{\\prime 0}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " for the 1st-level DVF " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime 1} = \\mathcal{M}(\\psi_{ab}^{\\prime 0}(f_{a}^{1}),f_{b}^{1})\\odot \\psi_{ab}^{\\prime 0}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": ", where the “" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": "” is the fusion operation illustrated in Supplementary. The remaining levels are calculated similarly so that the final DVF " + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime}" + }, + { + "bbox": [ + 313, + 228, + 555, + 455 + ], + "type": "text", + "content": " integrated multiscale vectors are output after a chain calculation, i.e." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 461, + 552, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 461, + 552, + 475 + ], + "spans": [ + { + "bbox": [ + 321, + 461, + 552, + 475 + ], + "type": "interline_equation", + "content": "\\psi_ {a b} ^ {\\prime} = \\mathcal {V} (F _ {a}, F _ {b}) = H (\\{\\psi_ {a b} ^ {\\prime \\prime} \\} _ {l = 0} ^ {L}) \\quad (\\psi_ {a b} ^ {\\prime 0} = \\mathcal {M} (f _ {a} ^ {0}, f _ {b} ^ {0}))", + "image_path": "d2158f1b553e68f2fa475f93381ef2f32a58989a746622d28758c1f41dcc8bac.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 477, + 553, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 477, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 321, + 477, + 553, + 493 + ], + "type": "interline_equation", + "content": "\\psi_ {a b} ^ {\\prime l} = \\mathcal {M} \\left(\\psi_ {a b} ^ {\\prime l - 1} \\left(f _ {a} ^ {l}\\right), f _ {b} ^ {l}\\right) \\odot \\psi_ {a b} ^ {\\prime l - 1}, l = 1, \\dots , L - 1, \\tag {8}", + "image_path": "2461b8d9ed1da19903078cd0997b08133efd7460d69fbb282d7daa9eab122444.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "text", + "content": " is the chain calculation across all " + }, + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "text", + "content": " levels to output the final DVF " + }, + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\psi_{ab}^{\\prime}" + }, + { + "bbox": [ + 313, + 498, + 554, + 534 + ], + "type": "text", + "content": ". Linear upsampling and value doubling are applied to bridge the scale gap of the DVFs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 535, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 555, + 643 + ], + "type": "text", + "content": "Discussion of properties Correspondence occurs at multiple scales for different semantic granularity, so our VPA formulates this multiscalarity in a pyramid-like architecture, bringing the advantages: 1) Low computational cost: It enables the mapping function " + }, + { + "bbox": [ + 313, + 535, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 535, + 555, + 643 + ], + "type": "text", + "content": " in a small receptive field at each level for a large whole receptive field reducing the computation. 2) Multi-scale representation: It learns the multi-scale features with multiple semantic granularities improving the granularity adaptability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 672, + 423, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 423, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 423, + 685 + ], + "type": "text", + "content": "4.1. Comparison study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "4.1.1 Experiment protocol This study has made sufficient experiments to ensure a complete evaluation. Following is" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19831" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 553, + 269 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 269 + ], + "type": "table", + "html": "
TypeMethodsa) 2D evaluation pretrained on [59]b) 3D evaluation pretrained on [31]AVG Score %
\\(SCR_{S}^{25\\%}\\)\\(PDCX_{R_{C}}\\)\\(KiPA22_{S}^{2D}\\)\\(FIVES_{S}\\)\\(CANDI_{S}\\)\\(FeTA21_{S}\\)\\(KiPA22_{S}^{3D}\\)\\(STOIC_{C}\\)
-Scratch81.890.474.179.484.056.972.472.076.4
GLDenosing [54]83.9(+1.9)92.0(+1.6)60.3(-13.8)77.8(-1.6)83.7(-0.3)52.9(-4.0)70.0(-2.4)65.9(-6.1)73.3(-3.1)
In-painting [38]85.1(+3.3)93.9(+3.5)64.4(-9.7)78.9(-0.5)88.5(+4.5)54.4(-2.5)69.7(-2.7)67.2(-4.8)75.3(-1.1)
Models Genesis [72]86.1(+4.3)92.6(+2.2)66.6(-7.5)79.6(+0.2)88.7(+4.7)55.8(-1.1)75.8(+3.4)75.3(+3.3)77.6(+1.2)
Rotation [28]80.5(-1.3)89.9(-0.5)69.7(-4.4)80.3(+0.9)89.4(+5.4)58.7(+1.8)77.4(+5.0)68.8(-3.2)76.8(+0.4)
BCLSimSiam [8]87.2(+5.4)92.2(+1.8)72.6(-1.5)84.3(+4.9)87.3(+3.3)58.7(+1.8)83.8(+11.4)69.5(-2.5)79.5(+3.1)
BYOL [14]89.4(+7.6)86.6(-3.8)74.1(0)83.3(+3.9)89.7(+5.7)59.2(+2.3)83.6(+11.2)74.0(+2.0)80.0(+5.4)
SimCLR [7]89.0(+7.2)94.7(+4.3)74.4(+0.3)84.5(+5.1)89.2(+5.2)53.4(-3.5)78.9(+6.5)60.7(-11.3)78.1(+1.7)
MoCov2 [9]84.3(+2.5)93.2(+2.8)69.6(-4.5)80.7(+1.3)89.7(+5.7)61.5(+4.6)78.0(+5.6)74.8(+2.8)79.0(+2.6)
DeepCluster [4]84.0(+2.2)93.3(+2.9)72.7(-1.4)81.6(+2.2)89.8(+5.8)57.4(+0.5)79.7(+7.3)65.6(-6.4)78.0(+1.6)
DBCLVADeR [35]85.2(+3.4)92.8(+2.4)62.8(-11.3)78.9(-0.5)87.4(+3.4)43.1(-13.8)72.1(-0.3)73.2(+1.2)74.4(-2.0)
DenseCL [60]85.0(+3.2)92.4(+2.0)70.8(-3.3)79.2(-0.2)87.7(+3.7)43.7(-13.2)74.0(+1.6)58.8(-13.2)74.0(-2.4)
SetSim [61]85.2(+3.4)93.9(+3.5)70.8(-3.3)80.1(+0.7)88.4(+4.4)58.7(+1.8)73.5(+1.1)60.1(-11.9)76.3(-0.1)
DSC-PM [29]90.5(+8.7)91.8(+1.4)77.2(+3.1)83.8(+4.4)88.5(+4.5)52.2(-4.7)79.0(+6.6)59.5(-12.5)77.8(+1.2)
PixPro [64]91.5(+9.7)93.0(+2.6)73.6(-0.5)84.3(+4.9)89.9(+5.9)60.7(+3.8)80.0(+7.6)75.1(+3.1)81.0(+4.6)
Chaitanya et al. [5]87.3(+5.5)90.0(-0.4)76.5(+2.4)84.9(+5.5)87.4(+3.4)53.4(-3.5)70.7(-1.7)67.8(-4.2)77.3(+0.9)
VRGVSL [19]89.7(+7.9)92.5(+2.1)78.9(+4.8)86.2(+6.8)89.1(+5.1)62.6(+5.7)84.3(+11.9)75.4(+3.4)82.3(+5.9)
GEMINI [22]92.4(+10.6)92.9(+2.5)79.1(+5.0)85.3(+5.9)90.0(+6.0)61.7(+4.8)85.0(+12.6)79.5(+7.5)83.2(+6.8)
VCLCOVER (Ours)94.0(+12.2)95.9(+5.5)80.0(+5.9)87.2(+7.8)89.9(+5.9)63.6(+6.7)85.2(+12.8)80.4(+8.4)84.5(+8.1)
", + "image_path": "b77a2324897bdc4d6dbb0f5b24a83da5ac42fcad84c26cc6842667997a376729.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 555, + 309 + ], + "type": "text", + "content": "Table 1. Comparison study: Our COVER achieves the best performance compared with 17 methods on 8 downstream tasks across a) 2D and b) 3D, showing our great adaptation ability. \"AVG\" is the average rank of the row. The blue values are the improvement compared with the \"Scratch\" and the bold values are the highest score. \"S\" and \"C\" are the segmentation and classification tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 317, + 294, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 294, + 328 + ], + "type": "text", + "content": "the overview protocol and details are in our Supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 329, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": "1) Materials: This study evaluated our COVER across 8 tasks across 2D and 3D evaluations with 4 modalities to showcase its versatility and advantages. a) " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " evaluation: It pretrained networks on the ChestX-ray8 dataset [59] containing 112,120 chest X-ray images. These pretrained models are fine-tuned on four tasks " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "(\\mathrm{SCR}_S^{25\\%}" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [52] (X-ray), " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{PDCXR}_C" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [27] (X-ray), " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{FIVES}_S" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [25] (fundus), " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{KiPA22}_{S}^{2D}" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [18] (CT)). b) " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " evaluation: It pretrained networks on 837 3D T1 brain MR images from the PPMI database [31]. These pretrained models are then fine-tuned to three datasets with four tasks " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "(\\mathrm{CANDI}_S" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [26] (MR), " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{FeTA21}_S" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [39] (MR), " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{STOIC}_C" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [42] (CT), " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{KiPA22}_{S}^{3D}" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " [18] (CT)). The evaluations involve segmentation (S) and classification (C) tasks. To evaluate the data amount robustness, it used " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": " of the training set in " + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{SCR}_S" + }, + { + "bbox": [ + 54, + 329, + 295, + 520 + ], + "type": "text", + "content": ", and more data amount evaluations are performed in our analysis (Fig.4 b)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 521, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 616 + ], + "type": "text", + "content": "2) Comparisons: This study benchmarked COVER with 17 recent or classic methods across four categories: GL [28, 38, 54, 72], binary CL (BCL) [4, 7-9, 14], dense binary CL (DBCL) [5, 29, 35, 60, 61, 64], and vector regression (VR) [19, 22]. It used 2D and 3D U-Nets [44] as the backbone " + }, + { + "bbox": [ + 54, + 521, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\theta}" + }, + { + "bbox": [ + 54, + 521, + 295, + 616 + ], + "type": "text", + "content": " (for the methods with global prediction, like MoCov2, it adopted the encoder part). For a fair comparison, full-parameter fine-tuning is employed in downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": "3) Implementation and evaluation metrics: Following [19], all tasks were implemented by PyTorch and optimized by Adam with a learning rate of " + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": ". The models were pretrained with " + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 10^{5}" + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": " iterations for pretraining tasks and fine-tuned with " + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "inline_equation", + "content": "4 \\times 10^{4}" + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": " iterations for downstream tasks. Training and testing were implemented on NVIDIA A100 SXUM4 GPU with 40 GB memory. This study used the Dice coefficient (DSC) for segmentation tasks and area un" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 317, + 497, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 497, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 497, + 328 + ], + "type": "text", + "content": "der curve (AUC) for classification tasks [48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 329, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "content": "4.1.2 Quantitative analysis Our great adaptability for the downstream tasks shown in Tab. 1 demonstrates our COVER's powerful general representation for medical vision. Two observations can be discovered:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": "1) Cross-scale transferability. The SOTA performances on small and large targets show our powerful cross-scale transferability owing to the learned multiscalarity from our VPA a. For small structures like the vessels " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "(\\mathrm{FIVES}_S)" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " and brain tissues (CANDI " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{FeTAS}_s" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " ), our approach achieved " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "87.2\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "89.9\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "63.6\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " DSC with the improvements of " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " over the \"Scratch\". These gains stem from learning fine-grained representations at lowerscale levels. b. For large structures including the chest regions " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "(\\mathrm{SCR}_{S}^{25\\%})" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " and kidney structures " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "(\\mathrm{KiPA22}_{S}^{2D}" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{KiPA22}_{S}^{3D})" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " , our model achieved " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "94.0\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "80.0\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " DSC improving by over " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "5 \\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " compared to \"Scratch\". This is attributed to leveraging high-scale levels for coarse-grained representation. c. For classification tasks (PDCXR " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " and STOIC), our COVER achieved AUC improvements of " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " by learning global representations at level 0. Although the GVSL and GEMINI also have improvements across scales, their inconsistent optimization flow limits the representation potential of pretraining data, resulting in an average score over " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "1 \\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " lower than COVER. The BCL methods struggle to enhance all tasks, for example, the DenseCL improved " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " DSC on SCR, but it reduced " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " DSC on " + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{KiPA22}_{S}^{2D}" + }, + { + "bbox": [ + 313, + 378, + 553, + 664 + ], + "type": "text", + "content": " . This arises from their over-dispersion, which introduces biases and compromises on incompatible tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "2) Cross-scene adaptability. The significant improvements on consistent " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{SCR}_S^{25\\%}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PDCXR}_C" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CANDI}_S)" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " and inconsistent (KiPA22" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "_S^{2D}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", FIVES" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "_S" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", FeTAS, KiPA22" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "_S^{3D}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", STOIC" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "_C" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ") scenes relative to the pretraining data illustrates" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19832" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 159 + ], + "type": "image", + "image_path": "d541e0a43e5c8d83cde799220e17d9d1e0759fdf01579467189ce3b457ecffca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 167, + 295, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 295, + 201 + ], + "type": "text", + "content": "Figure 4. Implementation analysis: Our COVER has great properties on the scaling of a) fine-tuned parameters, b) fine-tuning data, c) pretraining iterations, and d) model sizes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "text", + "content": "our strong cross-scene adaptability. Across all tasks in both 2D and 3D evaluations, COVER outperformed the \"Scratch\" by more than " + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "text", + "content": " and achieved the highest average scores " + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "inline_equation", + "content": "(84.5\\%)" + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "text", + "content": ". This adaptability arises from our distance modeling which quantifies the dispersion in embedding space, enabling the disentanglement of underlying explanatory factors hidden in low-level sensory data [2]. These learned factors capture the intrinsic properties of medical images, enhancing versatility across diverse scenes. Although most methods improve performance in consistent scenes by leveraging learned transformable knowledge, they struggle in inconsistent scenes. Notably, BCL and DBCL even degrade performance in some tasks due to their over-dispersion in binary CL, which amplifies biases in pretraining data. Especially, DenseCL and VADeR perform particularly poorly in inconsistent scenes like FeTA21" + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "inline_equation", + "content": "_S" + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "text", + "content": ", which reduce over " + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 209, + 295, + 411 + ], + "type": "text", + "content": " DSC." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 420, + 239, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 239, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 239, + 433 + ], + "type": "text", + "content": "4.2. Ablation study and model analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": "4.3.1 Component ablation The ablation study of the components in our COVER on the 2D evaluation of " + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{SCR}_S^{25\\%}" + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": " task shows the effectiveness of our proposed innovations. This study took the learning of the consistency loss " + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": " in pixel-wise as our base model, and gradually added our proposed components. The results are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 521, + 279, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 521, + 279, + 542 + ], + "spans": [ + { + "bbox": [ + 69, + 521, + 279, + 542 + ], + "type": "interline_equation", + "content": "\\begin{array}{c c c c c} & \\text {base .} \\mathcal {L} _ {\\text {con}} & \\dots + \\text {VEU (SeVR)} & \\dots + \\text {VPA} & \\dots + \\text {MVI} \\\\ \\hline \\text {DSC} \\% & 91.8 & 92.9 & 93.4 & 94.0 \\end{array}", + "image_path": "836a501be4679d2cbd1b1e099dd3e0d08d529342da68503c864a364c86bb1907.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": "The base model " + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "inline_equation", + "content": "91.8\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": " DSC that is higher than the BYOL's " + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "inline_equation", + "content": "89.4\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": " owing to its pixel-wise representation. When adding our VEU, it effectively improves the discrimination, achieving a " + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": " improvement. Our VPA enhances the multi-scale representation so that the model further improves " + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": " DSC. Finally, our MVI models the ambiguity of the correspondence for potentially varied feature concerns, thus achieving the highest " + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "inline_equation", + "content": "94.0\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": " DSC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "4.3.2 Hyper-parameter ablation The receptive field " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " in VEU, the amount of VEUs " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " in MVI are two new hyperparameters in our COVER. This study took an ablation of them on the " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SCR}_S^{25\\%}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " via linear evaluation to demonstrate their properties. The results are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 553, + 160 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 553, + 160 + ], + "type": "image", + "image_path": "a2f039526514bc8b8d1b9c900c57db0d02d409d7befa34f905013bc65a2607b2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 167, + 553, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 212 + ], + "type": "text", + "content": "Figure 5. Convergence analysis: a) Our COVER is able to promote convergence ability in fine-tuning. b) In the pretraining, the consistency learning causes representation collapse, and when adding our vector regression, the discrimination is enhanced." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 216, + 552, + 257 + ], + "blocks": [ + { + "bbox": [ + 317, + 216, + 552, + 257 + ], + "lines": [ + { + "bbox": [ + 317, + 216, + 552, + 257 + ], + "spans": [ + { + "bbox": [ + 317, + 216, + 552, + 257 + ], + "type": "table", + "html": "
N =3 × 35 × 57 × 79 × 9
DSC%44.948.454.835.1
J =[2,2,2,1,1][4,4,4,1,1][8,8,8,2,2][12,12,12,3,3]
DSC%55.356.354.848.6
", + "image_path": "520c545af5cb179034a925e8ef7b08de99b968de6989d053c7e53a637c0ace75.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": "For the receptive field size " + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": ", when enlarging it, its performance increases to " + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": " and then decreases when it is " + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": ". Because the increase of " + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": " will adapt to large spatial correspondence, improving the pretrained representation. However, too large " + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": " will introduce more ambiguous semantics, such as vascular with similar features, which will mislead the learning of correspondence. For the amount of VEUs " + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": ", it also increases and then decreases. When it is [4,4,4,1,1] for the scale levels, the model achieves the best ability. Too many vectors will smooth the optimization, thus weakening the discrimination of learned features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 403, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 653 + ], + "type": "text", + "content": "4.3.3 Implementation analysis This experiment analyzed four implementation factors (Fig.4): a) Fine-tuned parameter amount: Our COVER effectively reduces the fine-tuned parameters. It only fine-tunes one U-Net module to achieve a very significant performance improvement (15.3% DSC higher than GVSL), quickly reaching the bottleneck level that " + }, + { + "bbox": [ + 313, + 403, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{SCR}_S^{25\\%}" + }, + { + "bbox": [ + 313, + 403, + 554, + 653 + ], + "type": "text", + "content": " task can reach. b) Fine-tuning data amount: Our COVER effectively reduces the fine-tuning data requirement. Only utilizing 5% training data, our COVER can achieve the DSC close to 90% that GVSL and DenseCL have to use 25% and 100% data. c) pretraining iteration: With the progress of the training iterations, the performance (linear evaluation) on " + }, + { + "bbox": [ + 313, + 403, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{SCR}_S^{25\\%}" + }, + { + "bbox": [ + 313, + 403, + 554, + 653 + ], + "type": "text", + "content": " of our COVER gradually improves and eventually tends to be flat. This illustrates that more training is beneficial for the vector CL, which makes the pretrained models capture more diverse features in detail. d) Model size: With the enlarging of model size, the larger capacity enables our COVER to capture more abundant feature information in the pretraining. This makes the model gain more powerful representability in downstream tasks only with a linear layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "4.3.4 Convergence analysis This study analyzed the convergence performance of our COVER in both pretraining and fine-tuning phases (Fig.5): a) Fine-tuning: In downstream tasks, the network pretrained by our COVER exhibits superior convergence ability. Compared with the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19833" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 294, + 169 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 294, + 169 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 294, + 169 + ], + "type": "image", + "image_path": "288fc6cee09b2600efda5b1c01f168f210ebe3d84989ad83489f9b20c2dc3c85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 177, + 295, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 295, + 200 + ], + "type": "text", + "content": "Figure 6. Cluster analysis: Our COVER has strong clustering capabilities by quantifying the dispersion degree of features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "text", + "content": "\"Scratch\", GVSL, and DenseCL, COVER achieves a significantly lower loss value more quickly in both 2D segmentation (SCR) and 3D classification (STOIC) tasks. This advantage arises from our powerful pixel-wise feature learning that enables the pretrained network to capture the detailed information internal medical images. " + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "text", + "content": " pretraining: COVER improves the discrimination of the representation. When relying solely on the consistency loss " + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "text", + "content": ", the loss quickly collapses to zero, hindering subsequent knowledge learning. By integrating vector regression " + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vec}" + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "text", + "content": ", the model learns feature distances, stabilizing the consistency loss with fluctuations and enhancing the discrimination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 422 + ], + "type": "text", + "content": "4.3.5 Computation efficiency This study calculated the FLOPs to evaluate the computation efficiency with the expanding of receptive field. For a significant analysis, we focused on the computation in head part, i.e., our VPA. It compared two methods: a. VPA; b. directly taking a large vector template at level 4 (named \"Direct\"). The results are:" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 435, + 296, + 482 + ], + "blocks": [ + { + "bbox": [ + 56, + 435, + 296, + 482 + ], + "lines": [ + { + "bbox": [ + 56, + 435, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 435, + 296, + 482 + ], + "type": "image", + "image_path": "2df4300279be8ec292c9f8df774a0bfb0e74ccab85ca266a35f00fbeb3c2ebb1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "text", + "content": "When expanding the whole receptive field, our VPA maintains a stable computational cost of 0.2, whereas the \"Direct\" approach shows an exponential increase. Especially when it is " + }, + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "inline_equation", + "content": "121 \\times 121" + }, + { + "bbox": [ + 55, + 498, + 295, + 569 + ], + "type": "text", + "content": ", the \"Direct\" has more than 52 times our consumption. Our VPA is able to expand the receptive field for a wider contrast with low additional consumption." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "4.3.6 Distributed clusters from COVER The clusters show that our COVER has strong clustering capabilities by quantifying the dispersion of features. Using pretrained models from our 2D evaluation, this study extracted pixelwise features from the SCR and KiPA22 datasets and visualized their distributions via t-SNE [51] in Fig.6. It compared the clustering patterns of \"Scratch\", GVSL, and DenseCL with our COVER, revealing two key advantages: a) Continuous feature distribution: Our COVER predicts features that are distributed continuously in the embedded space. This continuity arises from our vector CL, which models the feature distance and avoids the excessive dis" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 152 + ], + "type": "image", + "image_path": "39ac4c0a47a9d8d885d5d8d139766b38abc90df31083185e1010b41a43db0172.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 182 + ], + "type": "text", + "content": "Figure 7. Multi-scale vectors analysis: Our VPA predicts multilevel correspondence for both global and detailed representation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 350 + ], + "type": "text", + "content": "cretization that blurs small or subtle features. However, DenseCL produces a lot of discrete clusters with small gaps between them, i.e., over-dispersion, leading to fragmented representation and reduced transferability. b) Effective aggregation: Our COVER aggregates features with the same semantics, while other methods show mixed regions. In both SCR and KiPA22 datasets, \"Scratch\" and GVSL exhibit overlapped clusters, as shown in the enlarged patches. Although DenseCL performs well on SCR, it fails on KiPA22, mixing features in this inconsistent scene with the pretraining. Our COVER achieves effective aggregation across datasets because our modeling of distance mitigates the overfitting to the pretraining scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 351, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 556, + 495 + ], + "type": "text", + "content": "4.3.7 Analysis of multi-scale vectors in our VPA This study analyzes the vectors at each scale level in our VPA by visualizing spatial transformed images via the predicted DVFs (Fig.7) in our 2D (chest X-ray) and 3D (brain MR) evaluations. In level 0, the DVF predicted from global features is able to align the images on the whole, like the rotation, scaling, etc., driving the learning of global representation. With the expansion of the scales, the correspondences are gradually refined so the details between the images are aligned. Especially, in level 0, the images have high alignment degree on the boundary, so that the models will effectively learn the representation of small or subtle features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 512, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 512, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 512, + 388, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we reformulate the standard CL as a vector regression problem, vector CL, providing a solution for \"over-dispersion\" problem caused by binary CL in pixelwise medical visual pretraining. Our proposed COcontrast in VEctor Regression (COVER) framework implemented the reformulated vector CL, modeling feature distance for CL in a vector regression process. It enables quantitative dispersion of features which improves representation learning for medical vision applications. Although the unique properties of our vector CL have bright powerful performance, important future work is to involve a larger-scale medical image dataset in training, fostering the development of a large pixel-wise foundation model in the medical field. Beyond advancing MVFM [21] development, COVER opens new frontiers in representations for continuous signals." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19834" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgment This research was supported by the National Science Foundation grants (2306545) and National Institutes of Health grants (1R01HL177813-01, 1R01HL173186-01A1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 195 + ], + "type": "text", + "content": "[1] Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Voxelmorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788-1800, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 294, + 239 + ], + "type": "text", + "content": "[2] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 294, + 297 + ], + "type": "text", + "content": "[3] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 294, + 341 + ], + "type": "text", + "content": "[4] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In European Conference on Computer Vision, 2018. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 343, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 294, + 396 + ], + "type": "text", + "content": "[5] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 463 + ], + "type": "text", + "content": "[6] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 510 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 555 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15750-15758, 2021. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 588 + ], + "type": "text", + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 633 + ], + "type": "text", + "content": "[10] Oliver Contier, Chris I Baker, and Martin N Hebart. Distributed representations of behaviour-derived object dimensions in the human visual system. Nature Human Behaviour, 8(11):2179-2193, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[11] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Rethinking optical flow from geometric matching consistent perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1337-1347, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[12] Viktoria Ehm, Maolin Gao, Paul Roetzer, Marvin Eisenberger, Daniel Cremers, and Florian Bernard. Partial-to-partial shape matching with geometric consistency. In Pro" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27488-27497, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 126 + ], + "type": "text", + "content": "[13] Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 193 + ], + "type": "text", + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Pires, Zhaohan Guo, Mohammad Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In Neural Information Processing Systems, 2020. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 195, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 247 + ], + "type": "text", + "content": "[15] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Self-supervised learning for medical image analysis: Discriminative, restorative, or adversarial? Medical Image Analysis, 94:103086, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 251, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 303 + ], + "type": "text", + "content": "[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 305, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 305, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 305, + 553, + 370 + ], + "type": "text", + "content": "[17] Yuting He, Tiantian Li, Rongjun Ge, Jian Yang, Youyong Kong, Jian Zhu, Huazhong Shu, Guanyu Yang, and Shuo Li. Few-shot learning for deformable medical image registration with perception-correspondence decoupling and reverse teaching. IEEE Journal of Biomedical and Health Informatics, 26(3):1177-1187, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 373, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 373, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 373, + 553, + 425 + ], + "type": "text", + "content": "[18] Yuting He, Guanyu Yang, Jian Yang, Rongjun Ge, Youyong Kong, Xiaomei Zhu, Shaobo Zhang, Pengfei Shao, Huazhong Shu, Jean-Louis Dillenseger, et al. Meta grayscale adaptive network for 3d integrated renal structures segmentation. Medical Image Analysis, 71:102055, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 491 + ], + "type": "text", + "content": "[19] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "text", + "content": "[20] Yuting He, Rongjun Ge, Xiaoming Qi, Yang Chen, Jia-song Wu, Jean-Louis Coatrioux, Guanyu Yang, and Shuo Li. Learning better registration to learn better few-shot medical image segmentation: Authenticity, diversity, and robustness. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2588-2601, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 612 + ], + "type": "text", + "content": "[21] Yuting He, Fuxiang Huang, Xinrui Jiang, Yuxiang Nie, Minghao Wang, Jiguang Wang, and Hao Chen. Foundation model for advancing healthcare: Challenges, opportunities, and future directions. arXiv preprint arXiv:2404.03264, 2024.1,3,8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[22] Yuting He, Boyu Wang, Rongjun Ge, Yang Chen, Guanyu Yang, and Shuo Li. Homeomorphism prior for false positive and negative problem in medical image dense contrastive representation learning, 2025. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "text", + "content": "[23] John Hinde and Clarice GB Demétrio. Overdispersion: models and estimation. Computational statistics & data analysis, 27(2):151-170, 1998. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-based pattern recognition: third inter" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19835" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "national workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "text", + "content": "[25] Kai Jin, Xingru Huang, Jingxing Zhou, Yunxiang Li, Yan Yan, Yibao Sun, Qianni Zhang, Yaqi Wang, and Juan Ye. Fives: A fundus image dataset for artificial intelligence based vessel segmentation. Scientific data, 9(1):475, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[26] David N Kennedy, Christian Haselgrove, Steven M Hodge, Pallavi S Rane, Nikos Makris, and Jean A Frazier. Candishare: A resource for pediatric neuroimaging data. Neuroinformatics, 10(3):319, 2012. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 259 + ], + "type": "text", + "content": "[27] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. cell, 172(5):1122-1131, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[28] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), 2018. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 360 + ], + "type": "text", + "content": "[29] Xiaoni Li, Yu Zhou, Yifei Zhang, Aoting Zhang, Wei Wang, Ning Jiang, Haiying Wu, and Weiping Wang. Dense semantic contrast for self-supervised visual representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1368-1376, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[30] Siladittya Manna, Saumik Bhattacharya, and Umapada Pal. Self-supervised visual representation learning for medical image analysis: A comprehensive survey. Transactions on Machine Learning Research. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 459 + ], + "type": "text", + "content": "[31] Kenneth Marek, Danna Jennings, Shirley Lasch, Andrew Siderowf, Caroline Tanner, Tanya Simuni, Chris Coffey, Karl Kieburtz, Emily Flagg, Sohini Chowdhury, et al. The parkinson progression marker initiative (ppmi). Progress in neurobiology, 95(4):629-635, 2011. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 290, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 290, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 290, + 471 + ], + "type": "text", + "content": "[32] Mehryar Mohri. Foundations of machine learning, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[33] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "text", + "content": "[34] Frank H Netter. Atlas of human anatomy, Professional Edition E-Book: including NetterReference. com Access with full downloadable image Bank. Elsevier health sciences, 2014. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "type": "text", + "content": "[35] Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in Neural Information Processing Systems, 33:4489-4500, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "text", + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[38] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536-2544, 2016. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 185 + ], + "type": "text", + "content": "[39] Kelly Payette, Hongwei Bran Li, Priscille de Dumast, Roxane Licandro, Hui Ji, Md Mahfuzur Rahman Siddiquee, Daguang Xu, Andriy Myronenko, Hao Liu, Yuchen Pei, et al. Fetal brain tissue annotation and segmentation challenge results. Medical image analysis, 88:102833, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 231 + ], + "type": "text", + "content": "[40] Guangzhi Qu, Salim Hariri, and Mazin Yousif. A new dependency and correlation analysis for features. IEEE Transactions on Knowledge and Data Engineering, 17(9):1199-1207, 2005. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 277 + ], + "type": "text", + "content": "[41] Quan Quan, Qingsong Yao, Heqin Zhu, and S Kevin Zhou. Igu-aug: Information-guided unsupervised augmentation and pixel-wise contrastive learning for medical image analysis. IEEE Transactions on Medical Imaging, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 277, + 553, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 344 + ], + "type": "text", + "content": "[42] Marie-Pierre Revel, Samia Boussouar, Constance de Margerie-Mellon, Inès Saab, Thibaut Lapotre, Dominique Mompoint, Guillaume Chassagnon, Audrey Milon, Mathieu Lederlin, Souhail Bennani, et al. Study of thoracic ct in疫情-19: The stoic project. Radiology, 301(1):E361-E370, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 346, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 346, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 346, + 553, + 392 + ], + "type": "text", + "content": "[43] Ignacio Rocco, Relja Arandjelovic, and Josef Sivic. Convolutional neural network architecture for geometric matching. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6148-6157, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 392, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 392, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 392, + 553, + 448 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 505 + ], + "type": "text", + "content": "[45] Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning, pages 5628-5637. PMLR, 2019. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 506, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 506, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 553, + 540 + ], + "type": "text", + "content": "[46] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 541, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 553, + 586 + ], + "type": "text", + "content": "[47] Juan Luis Suárez, Salvador García, and Francisco Herrera. A tutorial on distance metric learning: Mathematical foundations, algorithms, experimental analysis, prospects and challenges. Neurocomputing, 425:300-322, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 586, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 586, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 586, + 553, + 621 + ], + "type": "text", + "content": "[48] Abdel Aziz Taha and Allan Hanbury. Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool. BMC medical imaging, 15(1):1-28, 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 622, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 667 + ], + "type": "text", + "content": "[49] Janine Thoma, Danda Pani Paudel, and Luc V Gool. Soft contrastive learning for visual localization. Advances in Neural Information Processing Systems, 33:11119-11130, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[50] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? Advances in Neural Information Processing Systems, 33:6827-6839, 2020. 1, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19836" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[51] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9 (11), 2008. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "text", + "content": "[52] Bram Van Ginneken, Mikkel B Stegmann, and Marco Loog. Segmentation of anatomical structures in chest radiographs using supervised methods: a comparative study on a public database. Medical image analysis, 10(1):19-40, 2006. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 173 + ], + "type": "text", + "content": "[53] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "type": "text", + "content": "[54] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 293 + ], + "type": "text", + "content": "[55] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature medicine, pages 1-12, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "text", + "content": "[56] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[57] Haiqiao Wang, Dong Ni, and Yi Wang. Modet: Learning deformable image registration via motion decomposition transformer. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 740-749. Springer, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[58] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 513 + ], + "type": "text", + "content": "[59] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "text", + "content": "[60] Xinlong Wang, Rufeng Zhang, Chunhua Shen, and Tao Kong. Densecl: A simple framework for self-supervised dense visual pre-training. Visual Informatics, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "text", + "content": "[61] Zhaoqing Wang, Qiang Li, Guoxin Zhang, Pengfei Wan, Wen Zheng, Nannan Wang, Mingming Gong, and Tongliang Liu. Exploring set similarity for dense self-supervised representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16590-16599, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[62] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Large-scale 3d medical image pre-training with geometric context priors. arXiv preprint arXiv:2410.09890, 2024. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[63] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simpleyet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 597 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "text", + "content": "[64] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 228 + ], + "type": "text", + "content": "[66] Liu Yang and Rong Jin. Distance metric learning: A comprehensive survey. Michigan State University, 2(2):4, 2006. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "text", + "content": "[67] Chenyu You, Yuan Zhou, Ruihan Zhao, Lawrence Staib, and James S Duncan. Simcvd: Simple contrastive voxel-wise representation distillation for semi-supervised medical image segmentation. IEEE Transactions on Medical Imaging, 41(9):2228-2237, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "text", + "content": "[68] Chenyu You, Weicheng Dai, Yifei Min, Fenglin Liu, David Clifton, S Kevin Zhou, Lawrence Staib, and James Duncan. Rethinking semi-supervised medical image segmentation: A variance-reduction perspective. Advances in neural information processing systems, 36:9984-10021, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "type": "text", + "content": "[69] Chenyu You, Weicheng Dai, Fenglin Liu, Yifei Min, Nicha C Dvornek, Xiaoxiao Li, David A Clifton, Lawrence Staib, and James S Duncan. Mine your own anatomy: Revisiting medical image segmentation with extremely limited labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "text", + "content": "[70] Yi Zhang, Yiji Mao, Xuanyu Lu, Xingyu Zou, Hao Huang, Xinyang Li, Jiayue Li, and Haixian Zhang. From single to universal: tiny lesion detection in medical imaging. Artificial Intelligence Review, 57(8):192, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "type": "text", + "content": "[71] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 520, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 552 + ], + "type": "text", + "content": "[72] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "text", + "content": "[73] Jiuwen Zhu, Yuexiang Li, Yifan Hu, Kai Ma, S Kevin Zhou, and Yefeng Zheng. Rubik's cube+: A self-supervised feature learning framework for 3d medical image analysis. Medical image analysis, 64:101746, 2020. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19837" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_content_list.json b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a2c9acfcd72cd496d9f71f85668e9f68e6c5502 --- /dev/null +++ b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_content_list.json @@ -0,0 +1,1478 @@ +[ + { + "type": "text", + "text": "VehicleMAE: View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Wang $^{1}$ , Zeyu Zhang $^{1}$ , Dong Wang $^{2,*}$ , Di Gai $^{1}$ , Xin Xiong $^{3}$ , Jiyang Xu $^{1}$ and Ruihua Zhou $^{2}$", + "bbox": [ + 143, + 202, + 872, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Mathematics and Computer Sciences, Nanchang University, Nanchang, China", + "bbox": [ + 156, + 220, + 859, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Software, Nanchang University, Nanchang, China", + "bbox": [ + 267, + 238, + 750, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China", + "bbox": [ + 130, + 256, + 883, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangqi,gaidi,xiongxinxx}@ncu.edu.cn，{zeyuzhang1010,dongwang,xujiyang,zrh}@email.ncu.edu.cn", + "bbox": [ + 101, + 277, + 910, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale pre-training technology has achieved remarkable performance in diversified object re-identification (Re-ID) downstream tasks. Nevertheless, to our best knowledge, the pre-training model specifically for vehicle Re-ID, which focuses on tackling the challenge of multi-view variations, has not been fully investigated. In this paper, we first leverage a diffusion model to build a large-scale vehicle Re-ID benchmark dataset, dubbed \"DiffVERI\", containing over 1700K images from abundant multi-view annotations. In terms of this dataset, we further present VehicleMAE, a novel masked image modeling pre-training paradigm that learns view-invariant representations by performing mutual-distillation in a self-supervised manner. To be specific, the pipeline of VehicleMAE unfolds two core modules, i.e., view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). Technically, VMIM consists of two homogeneous masked autoencoders (MAE) that simultaneously reconstruct the RGB pixels and multi-view semantic information of the specific vehicle body region via paired asymmetric mask sampling strategies. To progressively distill the knowledge of the model itself, PPMD considers the two MAEs in the current epoch and the previous one as the student models and the teacher models, respectively, which leverages the knowledge learned by the current student and the historical teacher for mutual feature-level distillation. Extensive experimental results have verified that the proposed pre-training paradigm on DiffVERI gains compelling downstream task performance for vehicle Re-ID.", + "bbox": [ + 88, + 358, + 485, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 823, + 223, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The purpose of vehicle re-identification (Re-ID)[3, 23, 53] is to match target vehicles with the same identity under vari", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ous views. Despite impressive performance in some current works, this task still faces a main obstacle of large intraclass discrepancy among vehicles of the same identity. The primary cause is that the unique geometrically rigid structure of the vehicle itself leads to dynamic appearances of the same instance from different views. Accordingly, it is imperative to enable the Re-ID model to efficiently learn view-invariant representations.", + "bbox": [ + 511, + 328, + 906, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94418750685e995e2bcd6f2d0c957feeaa5956cda2d23a96c39faf9db2da41e4.jpg", + "image_caption": [ + "Figure 1. Illustration of DiffVERI dataset: (a) Steps for synthesizing data and annotations. (b) Synthetic examples from DiffVERI dataset." + ], + "image_footnote": [], + "bbox": [ + 552, + 462, + 864, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the release of the previous generic vehicle Re-ID datasets (such as VeRi-776 [25], VehicleID [26], VERiWild [27], etc.), early efforts [10, 30, 44] preferred the use of vehicle keypoints and part-based bounding boxes for discriminative representation learning to achieve view alignment. Additionally, some follow-up works [22, 39] based on metric learning resort to vehicle attribute annotations and", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4701", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spatio-temporal information to narrow the appearance discrepancy across different views. Despite advanced progress on the aforesaid works, their training data are typically derived from real-world surveillance videos and only contain images within a fixed view range, which limits the inference ability of the Re-ID model for different views. An intuitive solution is to build a large-scale benchmark with a comprehensive view range to provide learnable view-invariant information for the Re-ID model while trading off labor costs. Until recently, there have been no existing works specifically investigating large-scale vehicle Re-ID benchmarks for a comprehensive range of views. To address the aforementioned limitations, we release a large-scale multiview dataset, named DiffVERI, as presented in Figure 1. In brief, we use the text-to-image diffusion model to synthesize vehicle data from various views. Figure 1(a) illustrates the construction process of DiffVERI, which includes three steps: multi-view data generation, data filtering, and multi-view semantic annotations. An example of generating multi-view data is also shown in Figure 1(b), and more the detailed visualizations of synthesized examples is included in the supplementary materials.", + "bbox": [ + 93, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the large domain gap between the generic ImageNet and vehicle data, ImageNet cannot be well-suited for various downstream tasks of vehicle Re-ID. Therefore, it is particularly crucial to obtain initial weights with strong generalization ability through task-specific pre-training. Some pioneering works based on self-supervised pre-training have been explored in the person-related field. For instance, Fu et al. [6] constructed a LUPerson benchmark and developed a pre-training model via contrastive learning. Another effective self-supervised approach is masked image modeling technology, which performs masking and pixel-level reconstruction on the backbone of vision transformer (ViT) to serve various downstream vision tasks. In contrast, to our best knowledge, there is currently no prior research on targeted large-scale pre-training for vehicle Re-ID. That is to say, mainstream vehicle Re-ID methods are still limited by the pre-training parameters from ImageNet.", + "bbox": [ + 93, + 426, + 480, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we have released a vehicle Re-ID benchmark, DiffVERI, with a comprehensive view range. Benefiting from DiffVERI, a novel masked image modeling pre-training paradigm called VehicleMAE is then proposed to serve the downstream task of vehicle Re-ID. To sum up, the main contributions of this work can be summarized as follows:", + "bbox": [ + 93, + 685, + 480, + 790 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We build a large-scale vehicle Re-ID benchmark, DiffVERI, by fine-tuning an off-the-shelf diffusion model, which incorporates over 1700K images and multi-view semantic annotations. To our knowledge, DiffVERI is by far the largest vehicle Re-ID dataset with diverse view data.", + "2. We begin the first attempt to investigate a large-" + ], + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scale pre-training solution for vehicle Re-ID tasks and therefore devise a novel masked image modeling pre-training paradigm, termed VehicleMAE, to jointly integrate the view-asymmetry masked image modeling (VMIM) module and the past-to-present mutual-distillation (PPMD) one.", + "bbox": [ + 535, + 90, + 903, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Comprehensive experiments have validated the extraordinary performance of our pre-training model in both supervised and unsupervised downstream tasks of vehicle Re-ID.", + "bbox": [ + 516, + 181, + 903, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 258, + 653, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large-scale Re-ID Datasets", + "text_level": 1, + "bbox": [ + 516, + 284, + 756, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent emergence of vehicle benchmark datasets has led to increasing attention to vehicle Re-ID technology. Existing efforts mainly follow two mainstream settings. The first setting, supervised learning [20, 32, 35], considers vehicle Re-ID as a representation learning problem exploit ID annotations to learn informative features to distinguish different vehicle instances. Early works [32, 35] focused on exploring potential correlations between global and local features to enhance the discriminative performance of Re-ID models. To distinguish vehicles with similar appearances from the same manufacturing, subsequent works attempted to feed additional attribute annotations, such as color, model, and camera ID, as auxiliary information to obtain robust identity-related features. Methods in the latter setting, i.e., unsupervised learning [4, 15, 28, 49], aim to fully learn discriminative representations of the training set without relying on any annotations. The vanilla unsupervised learning procedure updates the Re-ID model using pseudo labels generated by clustering algorithms as supervised information. For this purpose, previous works have tentatively proposed pseudo label refinement [4, 28] and contrastive learning [15, 49] to overcome the limitations of clustering-based label noise. In both settings, the aforementioned methods inevitably face the challenge of view variations due to the lack of learning view annotations. To tackle this, we first construct a large-scale vehicle dataset DiffVERI comprising multi-view annotations and then attempt to learn view-invariant features via pre-training.", + "bbox": [ + 516, + 306, + 903, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Self-supervised Pre-training", + "text_level": 1, + "bbox": [ + 516, + 742, + 764, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the focus of vision recognition tasks has shifted from achieving advanced performance in specific downstream tasks to designing generic large-scale pre-training techniques. Labeling large-scale datasets is undoubtedly labor-intensive, while utilizing the inherent information of the data itself for self-supervised learning is a more effective way. Pioneer self-supervised pre-training works [2, 11] attempted to learn robust representation via dynamic contrastive learning and knowledge distillation on unlabeled", + "bbox": [ + 516, + 763, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4702", + "bbox": [ + 483, + 946, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "large-scale datasets. As an illustration, in the Re-ID task, Fu et al. [6] constructed a large-scale person dataset called LU-Person and first designed a pre-training Re-ID model based on contrastive learning. With the advanced performance of vision transformer [5] in vision tasks, another more lightweight method, masked image modeling [18, 29, 50], has been systematically proven to be effective in the field of self-supervised pre-training. The representative work [12], i.e., masked autoencoders (MAE), presented an encoder-decoder architecture to perform pixel-level masking and reconstruction operations. Subsequently, some recent works, such as VideoMAE [38] and MultiMAE [1], have extended MAE to the fields of video and multiple vision tasks, respectively. However, to our knowledge, there is no frontier exploration applying large-scale pre-training to vehicle Re-ID research. Therefore, this paper first proposes VehicleMAE to leverage the MAE architecture for pre-training on the DiffVERI dataset, to provide a powerful pre-training model for generic vehicle Re-ID tasks.", + "bbox": [ + 89, + 90, + 483, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. DiffVERI", + "text_level": 1, + "bbox": [ + 89, + 398, + 197, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Owing to the remarkable success of diffusion models in the field of data generation [34, 47], a text-to-image diffusion model is employed to construct a large-scale vehicle Re-ID dataset, DiffVERI, with a comprehensive view. The construction of DiffVERI comprises the following three primary steps:", + "bbox": [ + 89, + 426, + 483, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-view data generation", + "text_level": 1, + "bbox": [ + 89, + 534, + 336, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The off-the-shelf diffusion model, termed DreamBooth [33], is used for fine-tuning on the VERI-Wild to produce highly realistic data. Specifically, a specific text description $T$ , that is, \"A photo of [V] vehicle\" is created for each vehicle instance, where [V] represents the identity-specific identifier to ensure ID consistency between the generated and input images. Given an initial noise of $\\varepsilon \\sim N(0,1)$ , the denoising process of the diffusion model is followed, and the model is then optimized by balancing the differences between the generated and original images. The noise $z_{t}$ at time $t$ in the latent space can be expressed as Eq. 1:", + "bbox": [ + 89, + 559, + 483, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} (x, \\varepsilon) = \\sqrt {\\tilde {a} _ {t}} \\cdot x + \\sqrt {1 - \\tilde {a} _ {t}} \\cdot \\varepsilon \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 741, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{a}_t$ is the average denoising factor for the diffusion process time $t$ , and $x$ represents the original image.", + "bbox": [ + 89, + 777, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, the text features extracted by the text encoder $\\Gamma$ of DreamBooth are used as a conditioning vector $c_{\\mathrm{pr}} = \\Gamma \\bigl (\"\\mathrm{a photo of vehicle}\"\\bigr)$ for the pre-trained text-to image diffusion model $x_{\\theta}$ to generate images $x_{pr} = x_{\\theta}(z_t,c_{pr})$ . As the objective function in the fine-tuning of the diffusion model process, squared error loss can be de", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fined as Eq. 2:", + "bbox": [ + 513, + 90, + 612, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {x, c, \\varepsilon , \\varepsilon^ {\\prime}, t} [ w _ {t} \\| x _ {\\theta} (z _ {t} (x, \\varepsilon), c) - x \\| _ {2} ^ {2} + \\\\ \\left. \\lambda w _ {t} ^ {\\prime} \\left\\| x _ {\\theta} \\left(z _ {t} ^ {\\prime} \\left(x _ {p r}, \\varepsilon^ {\\prime}\\right), c _ {p r}\\right) - x _ {p r} \\right\\| _ {2} ^ {2} \\right] \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 116, + 903, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{t}$ represents the quality term of sample generation, and $c$ is a conditioning vector derived from the text $T$ . $\\lambda$ is the relative weight that controls a specific identity vehicle. A flexible multi-view generation prompt, that is, \"A photo of [V] vehicle [O] rotated [X] degrees clockwise\", where [O] and [X] are variables that control the generation direction and rotation angle, respectively, and $[O] \\in left, right, [X] \\in 30, 60, 90, 120, 150, 180$ , is designed to generate multi-view vehicle images during the inference process.", + "bbox": [ + 511, + 180, + 906, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data filtering.", + "text_level": 1, + "bbox": [ + 511, + 340, + 656, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by ours insights, two types of data noise are inevitably found in the generated data: (1) part of the images exhibit pixel blurring and low resolution, and (2) the vehicle body occupies a small proportion of the entire image, increasing the difficulty in focusing on the details of each view. Images smaller than 10K in size are first empirically filtered out to address the aforementioned issues. Subsequently, we resort to YOLOv7 [40] to detect the bounding box of the vehicle body and then calculate the proportion ratio between the vehicle body and the entire image. Based on the specifications, images with a proportion ratio below $50\\%$ are further filtered.", + "bbox": [ + 511, + 363, + 906, + 544 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/47255859e0f7554f99f3fd9349a2eb5e616dc49765b8a6d491477df2c4569fd3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsImagesSourceMulti-view Semantic AnnotationsResolutionViews
VeRi-776[25]49,357RealX243 × 214Constrained
VehicleID[26]221,763RealX374 × 412Constrained
VeRi-Wild[27]416,314RealX415 × 354Constrained
VehicleX[48]192,150SyntheticX256 × 256Constrained
VRAI[42]137,613RealX349 × 234Constrained
DiffVERI(Ours)1,712,703Real-synthetic✓496 × 485Diverse
", + "bbox": [ + 516, + 555, + 906, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of the statistics between DiffVERI and other public vehicle Re-ID datasets. In contrast, DiffVERI is currently the largest multi-view vehicle Re-ID benchmark.", + "bbox": [ + 511, + 670, + 906, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Multi-view semantic annotations.", + "text_level": 1, + "bbox": [ + 511, + 742, + 803, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After dataset collection, 10 well-trained annotators were employed, and 1200 hours were spent annotating $20\\mathrm{k}$ images with multi-view semantic segmentation masks. In particular, each vehicle is uniformly parsed into a background region and four view regions (i.e., front, rear, side, and top views). Some images undeniably lack all four views or demonstrate partially occluded views. Accordingly, images for annotation that have comprehensive view information are selected. Considering the powerful zero-shot semantic", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4703", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b6c22317b955e3f3ec4b8cd11b355c4fc3593b1d04ba53bb3ea49736a306e0a.jpg", + "image_caption": [ + "Figure 2. Overall pipeline of the proposed VehicleMAE. Firstly, VMIM module utilizes two homogeneous encoder-decoders to reconstruct a pair of asymmetric masked-out patches for RGB pixels and multi-view semantic tasks. Subsequently, to fully develop the potential of homogeneous architecture, PPMD module exploits the mutual knowledge exchange between the historical and current models to enhance the generalization performance of the pre-training model." + ], + "image_footnote": [], + "bbox": [ + 148, + 99, + 854, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "segmentation capability of the SAM model [21], the annotated masks are utilized as the ground-truth, and SAM is fine-tuned through a lightweight LoRA adapter [16]. The fine-tuned SAM can efficiently perform multi-view semantic inference on the remaining unlabeled images, thereby obtaining multi-view annotations for the entire DiffVERI.", + "bbox": [ + 88, + 463, + 480, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the above steps, a large-scale vehicle Re-ID dataset, DiffVERI, is ultimately constructed, which includes raw VERI-Wild and synthesized multi-view data. Table 1 reviews some attributes of existing vehicle Re-ID datasets, including number of images, data sources, multiview semantic annotations, average image resolution, and range of views.", + "bbox": [ + 88, + 555, + 482, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. VehicleMAE", + "text_level": 1, + "bbox": [ + 89, + 676, + 220, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the success of MAE, we propose a novel masked image modeling pre-training paradigm, VehicleMAE, which is outlined in Figure 2, including two modules, namely view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). In VMIM module, given an image $I_{img}$ and multi-view annotation $I_{seg}$ , a pair of visible patch maps with non-overlapping patches is generated based on an asymmetric mask sampling strategy to ensure that two homogeneous masked autoencoders $(\\phi_1$ and $\\phi_2)$ can perform different degrees of self-supervised reconstruction tasks within one epoch. More specifically, $\\phi_1$ and $\\phi_2$ comprise its own encoder, RGB decoder, and view semantic decoder, that is,", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\phi_1 = \\{E_{\\phi_1},D_{\\phi_1}^{img},D_{\\phi_1}^{seg}\\}$ and $\\phi_{2} = \\{E_{\\phi_{2}},D_{\\phi_{2}}^{img},D_{\\phi_{2}}^{seg}\\}$ , respectively. The map of features output by the encoders is $\\{F_{\\phi_1}^{img},F_{\\phi_1}^{seg},F_{\\phi_2}^{img},F_{\\phi_2}^{seg}\\}$ , and the map of features output by the decoders is $\\{f_{\\phi_1}^{img},f_{\\phi_1}^{seg},f_{\\phi_2}^{img},f_{\\phi_2}^{seg}\\}$ . Then, the PPMD module enables two MAEs to fully exchange their current and past knowledge. In the current epoch $N$ , the decoders of $\\phi_1$ and $\\phi_2$ act as collaborative training for students to distill the predicted target features from each other. Meanwhile, in the previous epoch $N - 1$ , the two histories of $\\phi_1'$ and $\\phi_2'$ are regarded as teachers exchanging students $\\phi_1$ and $\\phi_2$ with each other and then distilling the intermediate features of encoders.", + "bbox": [ + 511, + 460, + 906, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. View-asymmetry Masked Image Modeling.", + "text_level": 1, + "bbox": [ + 511, + 664, + 877, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The fundamental principle of existing image-based MAEs is to perform pixel-level masking and reconstruction using an encoder-decoder architecture. When MAE is extended to vehicle Re-ID, we empirically argue that there exist two main defects: (1) learning robust representation is difficult due to random mask sampling and low-level pixel reconstruction; and (2) reconstructing the original image pixels inevitably involves some identity-unrelated redundant region, leading to an inefficient and resource-intensive masked image modeling task. Consequently, this paper further develops VMIM, which has two homogeneous MAEs $\\phi_{1}$ and $\\phi_{2}$ . As depicted in Figure 3, this VMIM design incorporates the following two submodules to tackle the aforementioned two issues encountered by vanilla MAE", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4704", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c5a41b115ce048202b829ea79dadde2311ee48430af7a8a8ed1c3a8a02bc759d.jpg", + "image_caption": [ + "Figure 3. VMIM consist of two submodules: (a) The asymmetric mask sampling strategy in VMIM module generates a pair of visible maps without overlapping patches to create diverse preservation clues for reconstruction tasks. (b) Illustration of the identity-related region reconstruction in terms of $\\phi_1$ . For the feature output of the decoder, RGB pixel and multi-view semantic reconstruction with MSE loss is performed only for the vehicle body region. Notably, another peer $\\phi_2$ also performs a series of reconstruction processes similar to $\\phi_1$ ." + ], + "image_footnote": [], + "bbox": [ + 140, + 95, + 491, + 232 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7933f87441b368d4be3bcd76da04e48ee2a1dd822425abeae3bf8d8b3d280fc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 95, + 857, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point-by-point:", + "bbox": [ + 89, + 332, + 194, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(i) Asymmetric mask sampling strategy. Initially, ours intention is to fully utilize the potential of self-supervised pre-training to serve downstream tasks with view variation by employing two homogeneous architectures. However, two twin MAEs will inevitably fail to expand the scope of their knowledge if the same masking and reconstruction operations are employed to facilitate equal knowledge exchange. Motivated by these analyses, an asymmetric mask sampling strategy is designed to generate paired visible maps without overlapping patches and encourage two MAEs to learn diverse knowledge, as shown in Figure 3(a). Concretely, the image $I_{img}$ and the corresponding multi-view mask $I_{seg}$ are simultaneously injected into the encoders of $E_{\\phi_1}$ and $E_{\\phi_2}$ , and are respectively divided into two sequences with $N$ patches $\\{P(I_{img}|\\phi_1); P(I_{seg}|\\phi_1)\\}$ and $\\{P(I_{img}|\\phi_2); P(I_{seg}|\\phi_2)\\}$ in spatial order. Subsequently, based on the semantic segmentation annotation of $I_{seg}$ , each sequence is further subdivided into two types: (1) subsequences $P(I^{id}|\\phi_1) = \\{P(I_{img}^{id}|\\phi_1); P(I_{seg}^{id}|\\phi_1)\\}$ and $P(I^{id}|\\phi_2) = \\{P(I_{img}^{id}|\\phi_2); P(I_{seg}^{id}|\\phi_2)\\}$ of the vehicle body region with size $N_{id}$ ; (2) subsequences $P(I^{bg}|\\phi_1) = \\{P(I_{img}^{bg}|\\phi_1); P(I_{seg}^{bg}|\\phi_1)\\}$ and $P(I^{bg}|\\phi_2) = \\{P(I_{img}^{bg}|\\phi_2); P(I_{seg}^{bg}|\\phi_2)\\}$ of the background region with size $N_{bg}$ , thereby also fulfilling $N_{id} + N_{bg} = N$ . The pair of visible patch maps $\\{M_{\\phi_1}, M_{\\phi_2}\\}$ obtained through the asymmetric mask sampling strategy follows two-point principles. The first point is to randomly mask the patches of the background region $P(I^{bg}|\\phi_1)$ and $P(I^{bg}|\\phi_2)$ as much as possible and preserve the vehicle body region $P(I^{id}|\\phi_1)$ and $P(I^{id}|\\phi_2)$ as visible clues for the reconstruction task. The second point is that no overlap simultaneously exists in the masking patches of two MAEs, that is, $M_{\\phi_1} \\cap M_{\\phi_2} = \\emptyset$ . Notably, $\\{M_{\\phi_1}, M_{\\phi_2}\\}$ will be will be complemented each other in RGB pixels and multi-view semantic reconstruction to enhance collaboration between the two types of tasks.", + "bbox": [ + 91, + 351, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(ii) Identity-related region reconstruction. Directly using the entire original image and a multi-view segmentation mask as the prediction target of reconstruction inevitably increases the computational cost and reduces the pre-training efficiency. Considering the aforementioned analysis, an identity-related region reconstruction method is designed. First, similar to the tokenizing and encoding operations of ordinary MAEs, the visible patch maps $\\{M_{\\phi_1}, M_{\\phi_2}\\}$ obtained by asymmetric mask sampling are tokenized and embedded into a sequence of tokens, and the masked patches will be replaced with a sequence of learnable masked tokens. Afterward, the visible and invisible tokens are combined and then inputted into the decoders $\\{D_{\\phi_1}^{img}, D_{\\phi_1}^{seg}\\}$ and $\\{D_{\\phi_2}^{img}, D_{\\phi_2}^{seg}\\}$ of the two MAEs. Ultimately, in the decoding stage, the ground-truth of two task-specific reconstructions is set with identity-related regions $P(I^{id}|\\phi_1)$ and $P(I^{id}|\\phi_2)$ , that is, the RGB pixels and multi-view semantic regions of the vehicle body, instead of the initial $I_{img}$ and $I_{seg}$ . Taking $\\phi_1$ as an example, Figure 3(b) illustrates the reconstruction process of the proposed method. Mean square error (MSE) is adopted to calculate the error of RGB pixels and multi-view semantics in the reconstructed body region. The loss function $L_{rec}$ of the entire reconstruction task can be defined by Eq. 3:", + "bbox": [ + 511, + 332, + 906, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c} = \\sum_ {m \\in \\{1, 2 \\}} \\sum_ {\\triangle \\in \\{i m g, s e g \\}} M S E \\left(R \\left(I _ {\\triangle} ^ {i d} \\mid \\phi_ {m}\\right), P \\left(I _ {\\triangle} ^ {i d} \\mid \\phi_ {m}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 705, + 911, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the term $\\phi_{m}$ and $m\\in \\{1,2\\}$ represent two models $\\phi_{1}$ and $\\phi_{2}$ , respectively. The term $\\triangle \\in \\{img,seg\\}$ denotes the reconstruction task of RGB pixels and multi-view semantics. $R(I_{\\triangle}^{id}|\\phi_m)$ and $P(I_{\\triangle}^{id}|\\phi_m)$ are the predicted results and ground-truth of the model $\\phi_{m}$ performing two reconstruction tasks, respectively. Correspondingly, function $MSE(\\cdot)$ is the MSE between the predicted target and ground-truth.", + "bbox": [ + 511, + 752, + 908, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4705", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Past-to-present Mutual-distillation.", + "text_level": 1, + "bbox": [ + 89, + 90, + 395, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the VMIM module can diversify the knowledge learned by $\\phi_{1}$ and $\\phi_{2}$ , empirical investigation revealed that this module cannot guarantee collaborative training between the two models and deep mining of their self-knowledge. Toward this end, this study aimed to establish the teacher-student role relationship, uniting the past and present knowledge of two models, and design a past-to-present mutual distillation method. More precisely, the core of this design is to conduct feature-level knowledge distillation through the collaboration of current models and the cross-guidance of historical models.", + "bbox": [ + 89, + 112, + 483, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Assuming the current pre-training is at epoch $N$ , $\\phi_1$ and $\\phi_2$ play the roles of two student models. They collaboratively learn and exchange the predicted features with each other based on asymmetric visible clues to enhance the multi-view generalization capability of the entire VehicleMAE. The predicted feature maps $\\{f_{\\phi_1}^{img}, f_{\\phi_1}^{seg}\\}$ and $\\{f_{\\phi_2}^{img}, f_{\\phi_2}^{seg}\\}$ of the decoder are extracted for $\\phi_1$ and $\\phi_2$ in two specific reconstruction tasks, respectively. Kullback-Leibler (KL) divergence is then adopted to minimize the predicted feature distribution error. The distillation losses $L_{pr}^{\\phi_1}$ and $L_{pr}^{\\phi_2}$ of the two MAEs in the current epoch $N$ are represented by Eq. 4:", + "bbox": [ + 89, + 279, + 483, + 463 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p r} ^ {\\phi_ {1}} = K L \\left(f _ {\\phi_ {2}} ^ {i m g} \\mid \\mid f _ {\\phi_ {1}} ^ {i m g}\\right) + K L \\left(f _ {\\phi_ {2}} ^ {s e g} \\mid \\mid f _ {\\phi_ {1}} ^ {s e g}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 468, + 483, + 494 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p r} ^ {\\phi_ {2}} = K L \\left(f _ {\\phi_ {1}} ^ {i m g} | | f _ {\\phi_ {2}} ^ {i m g}\\right) + K L \\left(f _ {\\phi_ {1}} ^ {s e g} | | f _ {\\phi_ {2}} ^ {s e g}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 491, + 433, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $KL(\\cdot)$ is the KL divergence function that measures the predicted features of two models.", + "bbox": [ + 89, + 515, + 482, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When $N\\geq 2$ , the two historical models $\\phi_1^{\\prime}$ and $\\phi_2^\\prime$ corresponding to epoch $N - 1$ in the past are regarded as teachers cross-guiding student models $\\phi_{1}$ and $\\phi_{2}$ of epoch $N$ to dynamically mine their own knowledge. Specifically, $\\phi_1^{\\prime}$ is set as the historical teacher for $\\phi_{2}$ and $\\phi_2^\\prime$ as the historical teacher for $\\phi_{1}$ to establish the teacher-student relationships from the past to the present. This mutual exchange of teacher-student relationships can further enhance the potential for mutual distillation and knowledge diversity. The encoded feature maps $\\{F_{\\phi_1'}^{img},F_{\\phi_1'}^{seg},F_{\\phi_2'}^{img},F_{\\phi_2'}^{seg}\\}$ and $\\{F_{\\phi_1}^{img},F_{\\phi_1}^{seg},F_{\\phi_2}^{img},F_{\\phi_2}^{seg}\\}$ output by the encoders of two historical teacher models and two current student models are employed to distill intermediate features due to the inability of self-supervised pre-training to achieve hard-label distillation. The distillation loss $L_{pa}^{\\phi_1}$ guided by $\\phi_2^\\prime$ for $\\phi_{1}$ and the distillation loss $L_{pa}^{\\phi_2}$ guided by $\\phi_1^\\prime$ for $\\phi_{2}$ are defined as Eq.5:", + "bbox": [ + 89, + 545, + 483, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p a} ^ {\\phi_ {1}} = K L \\left(F _ {\\phi_ {2} ^ {\\prime}} ^ {i m g} \\mid \\mid F _ {\\phi_ {1}} ^ {i m g}\\right) + K L \\left(F _ {\\phi_ {2} ^ {\\prime}} ^ {s e g} \\mid \\mid F _ {\\phi_ {1}} ^ {s e g}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 816, + 482, + 847 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p a} ^ {\\phi_ {2}} = K L (F _ {\\phi_ {1} ^ {\\prime}} ^ {i m g} | | F _ {\\phi_ {2}} ^ {i m g}) + K L (F _ {\\phi_ {1} ^ {\\prime}} ^ {s e g} | | F _ {\\phi_ {2}} ^ {s e g})\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 843, + 439, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the pre-training of VehicleMAE is completed, all decoders will be discarded and the learned knowledge will", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "be fused by solving the average parameters of the two encoders $E_{\\phi_1}$ and $E_{\\phi_2}$ . The process of obtaining the mean model $E[\\bar{\\theta}]$ can be written as Eq. 6:", + "bbox": [ + 511, + 90, + 905, + 138 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE [ \\bar {\\theta} ] = \\frac {E _ {\\phi_ {1}} [ \\theta_ {\\phi_ {1}} ] + E _ {\\phi_ {2}} [ \\theta_ {\\phi_ {2}} ]}{2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 148, + 903, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta_{\\phi_1}$ and $\\theta_{\\phi_2}$ are the parameter of $E_{\\phi_1}$ and $E_{\\phi_2}$ , respectively, and $\\bar{\\theta}$ is the parameter of the mean model $E[\\bar{\\theta}]$ . It is worth emphasizing that $E[\\bar{\\theta}]$ will be directly applied as a single pre-training model to downstream tasks of supervised and unsupervised vehicle Re-ID.", + "bbox": [ + 511, + 188, + 906, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training Objective", + "text_level": 1, + "bbox": [ + 511, + 276, + 694, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training objectives of VehicleMAE include two parts: the loss of the reconstruction task of two homogeneous MAEs in the VMIM module; in the PPMD module, the distillation loss $L_{pr} = L_{pr}^{\\phi_1} + L_{pr}^{\\phi_2}$ between the current two student models and the distillation loss $L_{pa} = L_{pa}^{\\phi_1} + L_{pa}^{\\phi_2}$ guided by the history teachers. Therefore, the total loss $L_{total}$ of the overall pre-training process can be defined by Eq. 7 as:", + "bbox": [ + 511, + 297, + 905, + 420 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L _ {r e c} + L _ {p r} + L _ {p a} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 421, + 903, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that, the pre-training mean model $E[\\bar{\\theta}]$ directly serves as supervised and unsupervised vehicle Re-ID downstream tasks and is supervised by cross-entropy and triplet losses.", + "bbox": [ + 511, + 444, + 905, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 518, + 645, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 544, + 812, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This paper utilizes two well-known vehicle Re-ID datasets: VeRi-776 [23] and VehicleID [3]. VehicleID dataset includes small, medium, and large-scale test sets. To evaluate the performance of the proposed method, mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank- $k$ ( $k = 1, 5$ ) are employed as the evaluation metrics.", + "bbox": [ + 511, + 566, + 905, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 681, + 730, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP-B/16 serves as the backbone network. Throughout the whole training process, the image size is set to $256 \\times 256$ , and models update parameters using the Adam optimizer. During the pre-training phase, all samples are divided into mini-batches of size 128 for 300 epochs of iterative learning. For downstream tasks, the batch size is set to 64, and all samples are trained for 50 epochs. Additionally, each image undergoes data augmentation with random horizontal flipping, padding, cropping, and erasing. In unsupervised tasks, DBSCAN is applied as the clustering algorithm to generate pseudo-labels. All experiments are conducted under the Ubuntu 22.04 operating system and Pytorch environment with 4 NVIDIA RTX A6000 GPUs.", + "bbox": [ + 511, + 703, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4706", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 90, + 243, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of different modules. The results in Table 2 demonstrate the effectiveness of all proposed modules. Three observations are summarized by us: (1) When we independently introduce the \"Pa\" or \"Pr\" into the \"Baseline\" to perform knowledge distillation, the significant performance improvement in Rank accuracy and mAP confirms that historical and current distillation can fully explore and exchange their own knowledge. (2) Whether using \"Asym\" in the mask sampling strategy during the encoding stage or introducing \"Id-rec\" in the reconstruction task during the decoding stage, it is superior to MAE that integrates a random mask strategy and complete pixel reconstruction. (3) VehicleMAE belongs to the ensemble of all modules and achieves the best performance, further verifying that VehicleMAE can efficiently pre-train two homogeneous MAEs in a collaborative manner.", + "bbox": [ + 89, + 113, + 485, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5cfd2d7eb641fd6110c29108e66bf1987cafa73c399ec95459ea42ed59d41b4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablation ModulesVeRi-776VehicleID
BaselinePaPrAsymId-recmAPRank-1Rank-1Rank-5
✓37.681.269.081.8
✓✓40.185.071.284.6
✓✓40.585.371.184.7
✓✓38.783.670.382.8
✓✓39.584.370.182.6
✓✓✓✓✓42.587.673.385.6
", + "bbox": [ + 94, + 367, + 485, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of different mask sampling strategies and ratios. Table 3 reports the impact of using asymmetric mask and random mask sampling strategies at different mask ratios. We observe that the performance of the proposed \"Asym\" is superior to \"Random\" for all three mask ratios. Especially at a mask rate of $75\\%$ , \"Asym\" achieves the highest mAP of $42.5\\%$ and Rank-1 accuracy of $87.6\\%$ . This is due to the visible maps generated by \"Asym\" having no overlapping patches, which encourage two MAEs to learn and exchange knowledge.", + "bbox": [ + 89, + 597, + 483, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of different distillation strategies. Table 4 investigates the impact of modifying VehicleMAE using various distillation architectures on Re-ID performance. Ours findings are two-fold: (1) Compared to two MAEs learning independently, using historical distillation or present distillation can ameliorate the pre-training model performance, especially when combined to achieve more effective knowledge exchange. (2) The mean encoder $E[\\bar{\\theta}]$ can accurately feedback and maintain the average level of two encoders, while also solving the difficulties of multiple model selec", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e61106b03bad235ffccfcf8c86c0a987c1fbe0f636476856cb6db928f652de7b.jpg", + "table_caption": [ + "Table 2. Performance comparison of different components in unsupervised vehicle ReID. \"Past\" and \"Pres\" refer to historical distillation and present distillation in PPDM, respectively. \"Asym\" and \"Id-rec\" correspond to the Asymmetric mask sampling strategy and Identity-related region reconstruction in VMIM." + ], + "table_footnote": [], + "table_body": "
Mask RatiosStrategiesVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
25%Random38.884.268.483.6
Asym40.685.570.484.2
50%Random39.284.668.983.8
Asym41.786.271.184.8
75%Random40.185.169.584.1
Asym42.587.673.385.6
", + "bbox": [ + 522, + 89, + 911, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion and deployment for downstream tasks.", + "bbox": [ + 511, + 330, + 799, + 345 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b06f7edf49b3920fa45b2cd123ad68526850af70d6c020983e33244c5c436f98.jpg", + "table_caption": [ + "Table 3. Effectiveness of different mask sampling strategies in serving unsupervised vehicle Re-ID tasks. \"Random\" and \"Asym\" express the random mask sampling strategy and the asymmetric mask sampling strategy, respectively." + ], + "table_footnote": [], + "table_body": "
Pre-trained EncodersVeRi-776VehicleID
IndPaPrPa+PrIndPaPrPa+Pr
Eφ184.384.784.887.170.370.871.073.2
Eφ284.484.785.186.970.370.770.973.1
E[θ]84.685.085.387.670.971.271.173.3
", + "bbox": [ + 517, + 356, + 906, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison Rank-1 accuracy of two homogeneous encoders $\\{E_{\\phi_1}, E_{\\phi_2}\\}$ and a mean encoder $E[\\tilde{\\theta}]$ using various distillation architectures in unsupervised downstream tasks. \"Ind\" denotes that $E_{\\phi_1}$ and $E_{\\phi_2}$ learning independently without using distillation. \"Pa\", \"Pr\", and \"Pa+Pr\" respectively signify the independent use of historical distillation, present distillation, and the simultaneous use of both them.", + "bbox": [ + 511, + 440, + 906, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Model Complexity. To further analyze the computational performance, we re-implement different self-supervised pre-training works on DiffVERI for model parameter sizes and floating-point operations (FLOPs), as shown in Table 5. For fairness, all methods except ConvNeXt v2, which relies on the ConvNeXt network architecture, use the CLIP-B/16 image encoder as the backbone. Although VehicleMAE adopts a mutual-distillation architecture, our pre-training model contains 57.3M parameters and only needs 14.7G multiply-accumulate operations for feature extraction. This is mainly due to the mean model design and that we follow the vanilla MAE without directly changing the encoder-decoder architecture. In contrast, both the fully convolutional mask autoencoder in ConvNeXt v2 and the hybrid convolution module in CMAE will increase the resource consumption of feature extraction. These results demonstrate that VehicleMAE achieves a favorable trade-off between model complexity and the performance of downstream tasks.", + "bbox": [ + 511, + 553, + 908, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 511, + 848, + 815, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Table 6, we compare the proposed method with state-of-the-art ones in both supervised and unsuper", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4707", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ff6c990473334b3c0eb40dc48e8c5a5d2279b70f3682b22f112c19e81f74309.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsFLOPsParams
ConvNeXt v2[45]15.4G89M
CMAE[18]19.4G104M
BEiT[43]15.7G82.5M
Ours14.7G57.3M
", + "bbox": [ + 163, + 88, + 403, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/82e1c86497b08c526a71cbbfa74f5148cb63c351b3397894d9255b91fd78d6f3.jpg", + "table_caption": [ + "Table 5. Comparison of different pre-training models in terms of parameter size and FLOPs." + ], + "table_footnote": [], + "table_body": "
MethodsVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
Supervised Learning
\\( A\\text{VER}{*}^{R}\\left\\lbrack {19}\\right\\rbrack \\)61.289.063.585.6
\\( VehicleNet*R\\left\\lbrack {54}\\right\\rbrack \\)83.496.879.592.0
\\( TransReID^{*V}\\left\\lbrack {14}\\right\\rbrack \\)82.097.1--
\\( GiT^{*V}\\left\\lbrack {35}\\right\\rbrack \\)80.396.977.9-
\\( LCNL^{*V}\\left\\lbrack {46}\\right\\rbrack \\)81.897.4--
\\( TANet^{*R}\\left\\lbrack {17}\\right\\rbrack \\)83.696.878.292.6
\\( CFA-Net^{*R}\\left\\lbrack {41}\\right\\rbrack \\)80.796.9--
\\( HCI-Net^{*R}\\left\\lbrack {36}\\right\\rbrack \\)83.896.676.491.2
\\( CLIP-ReID^{*C}\\left\\lbrack {24}\\right\\rbrack \\)83.397.478.192.7
\\( Ours^{\\dagger C} \\)87.697.485.994.9
Unsupervised Learning
\\( MMT^{*R}\\left\\lbrack {7}\\right\\rbrack \\)25.460.931.042.4
\\( SPCL^{*R}\\left\\lbrack {8}\\right\\rbrack \\)36.979.953.066.4
\\( RLCC^{*R}\\left\\lbrack {51}\\right\\rbrack \\)39.683.4--
\\( PPLR^{*R}\\left\\lbrack {4}\\right\\rbrack \\)41.685.6--
\\( VAPC TO^{*R}\\left\\lbrack {52}\\right\\rbrack \\)30.476.2--
\\( ICL^{*R}\\left\\lbrack {37}\\right\\rbrack \\)39.583.7--
\\( AdaMG^{*R}\\left\\lbrack {31}\\right\\rbrack \\)41.086.2--
\\( NNNI^{*R}\\left\\lbrack {9}\\right\\rbrack \\)42.386.3--
\\( STDA^{*R}\\left\\lbrack {13}\\right\\rbrack \\)42.387.4--
\\( MAE†^C\\left\\lbrack {12}\\right\\rbrack \\)37.681.269.081.8
\\( CL-MAE†^C\\left\\lbrack {29}\\right\\rbrack \\)41.084.070.882.6
\\( CMAE†^C\\left\\lbrack {18}\\right\\rbrack \\)41.684.771.184.9
\\( Ours^{\\dagger C} \\)42.587.673.385.6
", + "bbox": [ + 112, + 215, + 468, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Performance comparisons with the latest methods on supervised and unsupervised vehicle Re-ID tasks. “ $*^{R}$ , “ $*^{V}$ , and “ $*^{C}$ ” indicate using ResNet, ViT, and the image encoder of CLIP as backbones, respectively. “ $\\dagger^{C}$ ” denotes the implementation of pre-training using the image encoder of CLIP on DiffVERI. Note that experimental results are accurate to one decimal place, and this paper uses only the large-scale test set of VehicleID.", + "bbox": [ + 89, + 669, + 483, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vised downstream tasks. Compared to the most competitive methods among ResNet-based, ViT-based, and CLIP-based backbones, our method delivers the best mAP and Rank accuracy for both supervised and unsupervised settings. For fair comparison, we further implement several well-established MAE-based pre-training methods on DiffVERI. The noticeable performance improvement implies", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e90d2afdae2e949377b78ae246f199bf6ee3e205abffb5f1d19dc8820e21434d.jpg", + "image_caption": [ + "Figure 4. Visualization of patch-level feature distributions extracted by Baseline and using VehicleMAE pre-training. The class with the highest proportion of pixels within each patch in the patch-level ground-truth is defined as the final class corresponding to the patch. Different color patches represent different views or background regions" + ], + "image_footnote": [], + "bbox": [ + 542, + 89, + 879, + 203 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the masked image modeling based on VehicleMAE is better suited for downstream tasks of vehicle Re-ID.", + "bbox": [ + 511, + 325, + 906, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 368, + 705, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 displays two visualization examples of the feature distribution at the patch-level extracted by Baseline and VehicleMAE. Specifically, we feed the features of all patches into the K-Means algorithm to perform five class-based clustering operations (i.e., a background region and four view regions) to intuitively demonstrate the view discriminate ability of different pre-training models. It can be observed that our VehicleMAE is able to distinguish patches from different views and backgrounds in both examples, getting close to patch-level ground-truth. Remarkably, even in stricter occlusion scenarios, VehicleMAE accurately avoid interference from the power poles in example B, further completing the extraction of multi-view information. More details on qualitative visualization experiments are included in the supplementary materials.", + "bbox": [ + 511, + 390, + 906, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper releases DiffVERI, a large-scale multi-view vehicle Re-ID dataset for learning view-invariant representations, and proposes a masked image modeling pre-training paradigm termed VehicleMAE specially for vehicle Re-ID downstream tasks. VehicleMAE first proposes a VMIM module that attempts to apply two homogeneous MAEs to predict the RGB pixels and multi-view semantic clues of vehicles in pairs, thereby gaining diverse multi-view inference capabilities. Subsequently, to facilitate learning collaboratively, a PPMD module is designed to progressively exchange knowledge with each other. Extensive experiments demonstrate that equipping our pre-training model can achieve competitive performance in generic vehicle Re-ID downstream tasks. Future work contributes to further expanding VehicleMAE into a unified multimodal pre-training paradigm.", + "bbox": [ + 509, + 657, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4708", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In European Conference on Computer Vision, pages 348-367. Springer, 2022. 3", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2", + "[3] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 347–355, 2022. 1, 6", + "[4] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7308–7318, 2022. 2, 8", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[6] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750–14759, 2021. 2, 3", + "[7] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 8", + "[8] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. Advances in neural information processing systems, 33:11309-11321, 2020. 8", + "[9] Dongchen Han, Weifeng Liu, Mingchen Zou, and Baodi Liu. Non-contrastive nearest neighbor identity-guided method for unsupervised object re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2713–2723, 2022. 8", + "[10] Bing He, Jia Li, Yifan Zhao, and Yonghong Tian. Part-regularized near-duplicate vehicle re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3997–4005, 2019. 1" + ], + "bbox": [ + 93, + 232, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the National Natural Science Foundation of China under Grant No. 62461037 and the Jiangxi Provincial Natural Science Foundation under Grant Nos. 20224BAB212011, 20232BAB202051, 20232BAB212008, and 20242BAB25078.", + "[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 8", + "[13] Qiaolin He, Zihan Wang, Zhijie Zheng, and Haifeng Hu. Spatial and temporal dual-attention for unsupervised person re-identification. IEEE Transactions on Intelligent Transportation Systems, 2023. 8", + "[14] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013–15022, 2021. 8", + "[15] Zhijun He, Hongbo Zhao, Jianrong Wang, and Wenquan Feng. Multi-level progressive learning for unsupervised vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(4):4357-4371, 2022. 2", + "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4", + "[17] Wenbo Hu, Hongjian Zhan, Palaiahnakote Shivakumara, Umapada Pal, and Yue Lu. Tanet: Text region attention learning for vehicle re-identification. Engineering Applications of Artificial Intelligence, 133:108448, 2024. 8", + "[18] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 8", + "[19] Pirazh Khorramshahi, Amit Kumar, Neehar Peri, Sai Saketh Rambhatla, Jun-Cheng Chen, and Rama Chellappa. A dual-path model with adaptive attention for vehicle re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6132–6141, 2019. 8", + "[20] Pirazh Khorramshahi, Vineet Shenoy, and Rama Chellappa. Robust and scalable vehicle re-identification via self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5295–5304, 2023. 2", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4", + "[22] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE transactions on image processing, 31:5949-5962, 2022. 1", + "[23] Hongchao Li, Aihua Zheng, Liping Sun, and Yonglong Luo. Camera topology graph guided vehicle re-identification." + ], + "bbox": [ + 93, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4709", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Multimedia, 26:1565-1577, 2023. 1, 6", + "[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1405–1413, 2023. 8", + "[25] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2167-2175, 2016. 1, 3", + "[26] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In 2016 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2016. 1, 3", + "[27] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. Veri-wild: A large dataset and a new method for vehicle re-identification in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3235-3243, 2019. 1, 3", + "[28] Zefeng Lu, Ronghao Lin, Qiaolin He, and Haifeng Hu. Mask-aware pseudo label denoising for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 24(4):4333-4347, 2023. 2", + "[29] Neelu Madan, Nicolae-Catalin Ristea, Kamal Nasrollahi, Thomas B Moeslund, and Radu Tudor Ionescu. Cl-mae: Curriculum-learned masked autoencoders. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2492–2502, 2024. 3, 8", + "[30] Dechao Meng, Liang Li, Xuejing Liu, Yadong Li, Shijie Yang, Zheng-Jun Zha, Xingyu Gao, Shuhui Wang, and Qingming Huang. Parsing-based view-aware embedding network for vehicle re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7103–7112, 2020. 1", + "[31] Jinjia Peng, Guangqi Jiang, and Huibing Wang. Adaptive memorization with group labels for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5802-5813, 2023. 8", + "[32] Jiuchao Qian, Minting Pan, Wei Tong, Rob Law, and Edmond Q Wu. Urrnet: A unified relational reasoning network for vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(9):11156-11168, 2023. 2", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3", + "[34] Zhiyin Shao, Xinyu Zhang, Changxing Ding, Jian Wang, and Jingdong Wang. Unified pre-training with pseudo texts for text-to-image person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11174–11184, 2023. 3", + "[35] Fei Shen, Yi Xie, Jianqing Zhu, Xiaobin Zhu, and Huanqiang Zeng. Git: Graph interactive transformer for vehicle re-identification. IEEE Transactions on Image Processing, 32:1039–1051, 2023. 2, 8" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Ke Sun, Xiyu Pang, Meifeng Zheng, Xiushan Nie, Xi Li, Houren Zhou, and Yilong Yin. Heterogeneous context interaction network for vehicle re-identification. Neural Networks, 169:293–306, 2024. 8", + "[37] Qing Tang and Kang-Hyun Jo. Unsupervised object re-identification via instances correlation loss. In 2022 IEEE 20th International Conference on Industrial Informatics (IN-DIN), pages 135–139. IEEE, 2022. 8", + "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078-10093, 2022. 3", + "[39] Jingzheng Tu, Cailian Chen, Xiaolin Huang, Jianping He, and Xinping Guan. Dfr-st: Discriminative feature representation with spatio-temporal cues for vehicle re-identification. Pattern Recognition, 131:108887, 2022. 1", + "[40] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464-7475, 2023. 3", + "[41] Hai Wang, Yaqing Niu, Long Chen, Yicheng Li, Miguel Angel Sotelo, Zhixiong Li, and Yingfeng Cai. Dair-v2xreid: A new real-world vehicle-infrastructure cooperative re-id dataset and cross-shot feature aggregation network perception method. IEEE Transactions on Intelligent Transportation Systems, 25(8):9058-9068, 2024. 8", + "[42] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 460–469, 2019. 3", + "[43] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 8", + "[44] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In Proceedings of the IEEE international conference on computer vision, pages 379–387, 2017. 1", + "[45] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16133-16142, 2023. 8", + "[46] Mouxing Yang, Zhenyu Huang, and Xi Peng. Robust object re-identification with coupled noisy labels. International Journal of Computer Vision, pages 1–19, 2024. 8", + "[47] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4710", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492-4501, 2023. 3", + "[48] Yue Yao, Liang Zheng, Xiaodong Yang, Milind Naphade, and Tom Gedeon. Simulating content consistent vehicle datasets with attribute descent. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 775-791. Springer, 2020. 3", + "[49] Hongwei Zhang, Guoqing Zhang, Yuhao Chen, and Yuhui Zheng. Global relation-aware contrast learning for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8599–8610, 2022. 2", + "[50] Kevin Zhang and Zhiqiang Shen. i-mae: Are latent representations in masked autoencoders linearly separable? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7740-7749, 2024. 3", + "[51] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3436–3445, 2021. 8", + "[52] Aihua Zheng, Xia Sun, Chenglong Li, and Jin Tang. Aware progressive clustering for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(8):11422-11435, 2021. 8", + "[53] Aihua Zheng, Chaobin Zhang, Chenglong Li, Jin Tang, and Chang Tan. Multi-query vehicle re-identification: Viewpoint-conditioned network, unified dataset and new metric. IEEE Transactions on Image Processing, 32:5948–5960, 2023. 1", + "[54] Zhedong Zheng, Tao Ruan, Yunchao Wei, Yi Yang, and Tao Mei. Vehiclenet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia, 23:2683–2693, 2020. 8" + ], + "bbox": [ + 91, + 90, + 483, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4711", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_model.json b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9c258e99e782145b6c1aa3f4ba16db06d4ff0747 --- /dev/null +++ b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_model.json @@ -0,0 +1,2081 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "VehicleMAE: View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.873, + 0.222 + ], + "angle": 0, + "content": "Qi Wang\\(^{1}\\), Zeyu Zhang\\(^{1}\\), Dong Wang\\(^{2,*}\\), Di Gai\\(^{1}\\), Xin Xiong\\(^{3}\\), Jiyang Xu\\(^{1}\\) and Ruihua Zhou\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.222, + 0.86, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Mathematics and Computer Sciences, Nanchang University, Nanchang, China" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.239, + 0.75, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Software, Nanchang University, Nanchang, China" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.257, + 0.885, + 0.275 + ], + "angle": 0, + "content": "3The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.278, + 0.911, + 0.293 + ], + "angle": 0, + "content": "{wangqi,gaidi,xiongxinxx}@ncu.edu.cn，{zeyuzhang1010,dongwang,xujiyang,zrh}@email.ncu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.798 + ], + "angle": 0, + "content": "Large-scale pre-training technology has achieved remarkable performance in diversified object re-identification (Re-ID) downstream tasks. Nevertheless, to our best knowledge, the pre-training model specifically for vehicle Re-ID, which focuses on tackling the challenge of multi-view variations, has not been fully investigated. In this paper, we first leverage a diffusion model to build a large-scale vehicle Re-ID benchmark dataset, dubbed \"DiffVERI\", containing over 1700K images from abundant multi-view annotations. In terms of this dataset, we further present VehicleMAE, a novel masked image modeling pre-training paradigm that learns view-invariant representations by performing mutual-distillation in a self-supervised manner. To be specific, the pipeline of VehicleMAE unfolds two core modules, i.e., view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). Technically, VMIM consists of two homogeneous masked autoencoders (MAE) that simultaneously reconstruct the RGB pixels and multi-view semantic information of the specific vehicle body region via paired asymmetric mask sampling strategies. To progressively distill the knowledge of the model itself, PPMD considers the two MAEs in the current epoch and the previous one as the student models and the teacher models, respectively, which leverages the knowledge learned by the current student and the historical teacher for mutual feature-level distillation. Extensive experimental results have verified that the proposed pre-training paradigm on DiffVERI gains compelling downstream task performance for vehicle Re-ID." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.824, + 0.224, + 0.84 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.88 + ], + "angle": 0, + "content": "The purpose of vehicle re-identification (Re-ID)[3, 23, 53] is to match target vehicles with the same identity under vari" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.908, + 0.45 + ], + "angle": 0, + "content": "ous views. Despite impressive performance in some current works, this task still faces a main obstacle of large intraclass discrepancy among vehicles of the same identity. The primary cause is that the unique geometrically rigid structure of the vehicle itself leads to dynamic appearances of the same instance from different views. Accordingly, it is imperative to enable the Re-ID model to efficiently learn view-invariant representations." + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.463, + 0.865, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.736, + 0.908, + 0.778 + ], + "angle": 0, + "content": "Figure 1. Illustration of DiffVERI dataset: (a) Steps for synthesizing data and annotations. (b) Synthetic examples from DiffVERI dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "With the release of the previous generic vehicle Re-ID datasets (such as VeRi-776 [25], VehicleID [26], VERiWild [27], etc.), early efforts [10, 30, 44] preferred the use of vehicle keypoints and part-based bounding boxes for discriminative representation learning to achieve view alignment. Additionally, some follow-up works [22, 39] based on metric learning resort to vehicle attribute annotations and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4701" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.423 + ], + "angle": 0, + "content": "spatio-temporal information to narrow the appearance discrepancy across different views. Despite advanced progress on the aforesaid works, their training data are typically derived from real-world surveillance videos and only contain images within a fixed view range, which limits the inference ability of the Re-ID model for different views. An intuitive solution is to build a large-scale benchmark with a comprehensive view range to provide learnable view-invariant information for the Re-ID model while trading off labor costs. Until recently, there have been no existing works specifically investigating large-scale vehicle Re-ID benchmarks for a comprehensive range of views. To address the aforementioned limitations, we release a large-scale multiview dataset, named DiffVERI, as presented in Figure 1. In brief, we use the text-to-image diffusion model to synthesize vehicle data from various views. Figure 1(a) illustrates the construction process of DiffVERI, which includes three steps: multi-view data generation, data filtering, and multi-view semantic annotations. An example of generating multi-view data is also shown in Figure 1(b), and more the detailed visualizations of synthesized examples is included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.427, + 0.482, + 0.682 + ], + "angle": 0, + "content": "Due to the large domain gap between the generic ImageNet and vehicle data, ImageNet cannot be well-suited for various downstream tasks of vehicle Re-ID. Therefore, it is particularly crucial to obtain initial weights with strong generalization ability through task-specific pre-training. Some pioneering works based on self-supervised pre-training have been explored in the person-related field. For instance, Fu et al. [6] constructed a LUPerson benchmark and developed a pre-training model via contrastive learning. Another effective self-supervised approach is masked image modeling technology, which performs masking and pixel-level reconstruction on the backbone of vision transformer (ViT) to serve various downstream vision tasks. In contrast, to our best knowledge, there is currently no prior research on targeted large-scale pre-training for vehicle Re-ID. That is to say, mainstream vehicle Re-ID methods are still limited by the pre-training parameters from ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.686, + 0.482, + 0.791 + ], + "angle": 0, + "content": "To address the aforementioned issues, we have released a vehicle Re-ID benchmark, DiffVERI, with a comprehensive view range. Benefiting from DiffVERI, a novel masked image modeling pre-training paradigm called VehicleMAE is then proposed to serve the downstream task of vehicle Re-ID. To sum up, the main contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.795, + 0.481, + 0.884 + ], + "angle": 0, + "content": "1. We build a large-scale vehicle Re-ID benchmark, DiffVERI, by fine-tuning an off-the-shelf diffusion model, which incorporates over 1700K images and multi-view semantic annotations. To our knowledge, DiffVERI is by far the largest vehicle Re-ID dataset with diverse view data." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "2. We begin the first attempt to investigate a large-" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "scale pre-training solution for vehicle Re-ID tasks and therefore devise a novel masked image modeling pre-training paradigm, termed VehicleMAE, to jointly integrate the view-asymmetry masked image modeling (VMIM) module and the past-to-present mutual-distillation (PPMD) one." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.183, + 0.905, + 0.241 + ], + "angle": 0, + "content": "3. Comprehensive experiments have validated the extraordinary performance of our pre-training model in both supervised and unsupervised downstream tasks of vehicle Re-ID." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.259, + 0.654, + 0.274 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.285, + 0.758, + 0.3 + ], + "angle": 0, + "content": "2.1. Large-scale Re-ID Datasets" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.73 + ], + "angle": 0, + "content": "The recent emergence of vehicle benchmark datasets has led to increasing attention to vehicle Re-ID technology. Existing efforts mainly follow two mainstream settings. The first setting, supervised learning [20, 32, 35], considers vehicle Re-ID as a representation learning problem exploit ID annotations to learn informative features to distinguish different vehicle instances. Early works [32, 35] focused on exploring potential correlations between global and local features to enhance the discriminative performance of Re-ID models. To distinguish vehicles with similar appearances from the same manufacturing, subsequent works attempted to feed additional attribute annotations, such as color, model, and camera ID, as auxiliary information to obtain robust identity-related features. Methods in the latter setting, i.e., unsupervised learning [4, 15, 28, 49], aim to fully learn discriminative representations of the training set without relying on any annotations. The vanilla unsupervised learning procedure updates the Re-ID model using pseudo labels generated by clustering algorithms as supervised information. For this purpose, previous works have tentatively proposed pseudo label refinement [4, 28] and contrastive learning [15, 49] to overcome the limitations of clustering-based label noise. In both settings, the aforementioned methods inevitably face the challenge of view variations due to the lack of learning view annotations. To tackle this, we first construct a large-scale vehicle dataset DiffVERI comprising multi-view annotations and then attempt to learn view-invariant features via pre-training." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.743, + 0.765, + 0.758 + ], + "angle": 0, + "content": "2.2. Self-supervised Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.765, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Recently, the focus of vision recognition tasks has shifted from achieving advanced performance in specific downstream tasks to designing generic large-scale pre-training techniques. Labeling large-scale datasets is undoubtedly labor-intensive, while utilizing the inherent information of the data itself for self-supervised learning is a more effective way. Pioneer self-supervised pre-training works [2, 11] attempted to learn robust representation via dynamic contrastive learning and knowledge distillation on unlabeled" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.947, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4702" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.381 + ], + "angle": 0, + "content": "large-scale datasets. As an illustration, in the Re-ID task, Fu et al. [6] constructed a large-scale person dataset called LU-Person and first designed a pre-training Re-ID model based on contrastive learning. With the advanced performance of vision transformer [5] in vision tasks, another more lightweight method, masked image modeling [18, 29, 50], has been systematically proven to be effective in the field of self-supervised pre-training. The representative work [12], i.e., masked autoencoders (MAE), presented an encoder-decoder architecture to perform pixel-level masking and reconstruction operations. Subsequently, some recent works, such as VideoMAE [38] and MultiMAE [1], have extended MAE to the fields of video and multiple vision tasks, respectively. However, to our knowledge, there is no frontier exploration applying large-scale pre-training to vehicle Re-ID research. Therefore, this paper first proposes VehicleMAE to leverage the MAE architecture for pre-training on the DiffVERI dataset, to provide a powerful pre-training model for generic vehicle Re-ID tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.4, + 0.199, + 0.415 + ], + "angle": 0, + "content": "3. DiffVERI" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.52 + ], + "angle": 0, + "content": "Owing to the remarkable success of diffusion models in the field of data generation [34, 47], a text-to-image diffusion model is employed to construct a large-scale vehicle Re-ID dataset, DiffVERI, with a comprehensive view. The construction of DiffVERI comprises the following three primary steps:" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.338, + 0.551 + ], + "angle": 0, + "content": "3.1. Multi-view data generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.727 + ], + "angle": 0, + "content": "The off-the-shelf diffusion model, termed DreamBooth [33], is used for fine-tuning on the VERI-Wild to produce highly realistic data. Specifically, a specific text description \\( T \\), that is, \"A photo of [V] vehicle\" is created for each vehicle instance, where [V] represents the identity-specific identifier to ensure ID consistency between the generated and input images. Given an initial noise of \\( \\varepsilon \\sim N(0,1) \\), the denoising process of the diffusion model is followed, and the model is then optimized by balancing the differences between the generated and original images. The noise \\( z_{t} \\) at time \\( t \\) in the latent space can be expressed as Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.742, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\nz _ {t} (x, \\varepsilon) = \\sqrt {\\tilde {a} _ {t}} \\cdot x + \\sqrt {1 - \\tilde {a} _ {t}} \\cdot \\varepsilon \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.809 + ], + "angle": 0, + "content": "where \\(\\tilde{a}_t\\) is the average denoising factor for the diffusion process time \\(t\\), and \\(x\\) represents the original image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Subsequently, the text features extracted by the text encoder \\(\\Gamma\\) of DreamBooth are used as a conditioning vector \\(c_{\\mathrm{pr}} = \\Gamma \\bigl (\"\\mathrm{a photo of vehicle}\"\\bigr)\\) for the pre-trained text-to image diffusion model \\(x_{\\theta}\\) to generate images \\(x_{pr} = x_{\\theta}(z_t,c_{pr})\\). As the objective function in the fine-tuning of the diffusion model process, squared error loss can be de" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.614, + 0.107 + ], + "angle": 0, + "content": "fined as Eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.117, + 0.905, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {x, c, \\varepsilon , \\varepsilon^ {\\prime}, t} [ w _ {t} \\| x _ {\\theta} (z _ {t} (x, \\varepsilon), c) - x \\| _ {2} ^ {2} + \\\\ \\left. \\lambda w _ {t} ^ {\\prime} \\left\\| x _ {\\theta} \\left(z _ {t} ^ {\\prime} \\left(x _ {p r}, \\varepsilon^ {\\prime}\\right), c _ {p r}\\right) - x _ {p r} \\right\\| _ {2} ^ {2} \\right] \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.181, + 0.907, + 0.332 + ], + "angle": 0, + "content": "where \\( w_{t} \\) represents the quality term of sample generation, and \\( c \\) is a conditioning vector derived from the text \\( T \\). \\( \\lambda \\) is the relative weight that controls a specific identity vehicle. A flexible multi-view generation prompt, that is, \"A photo of [V] vehicle [O] rotated [X] degrees clockwise\", where [O] and [X] are variables that control the generation direction and rotation angle, respectively, and \\( [O] \\in left, right, [X] \\in 30, 60, 90, 120, 150, 180 \\), is designed to generate multi-view vehicle images during the inference process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.342, + 0.658, + 0.358 + ], + "angle": 0, + "content": "3.2. Data filtering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Motivated by ours insights, two types of data noise are inevitably found in the generated data: (1) part of the images exhibit pixel blurring and low resolution, and (2) the vehicle body occupies a small proportion of the entire image, increasing the difficulty in focusing on the details of each view. Images smaller than 10K in size are first empirically filtered out to address the aforementioned issues. Subsequently, we resort to YOLOv7 [40] to detect the bounding box of the vehicle body and then calculate the proportion ratio between the vehicle body and the entire image. Based on the specifications, images with a proportion ratio below \\(50\\%\\) are further filtered." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.556, + 0.907, + 0.661 + ], + "angle": 0, + "content": "
DatasetsImagesSourceMulti-view Semantic AnnotationsResolutionViews
VeRi-776[25]49,357RealX243 × 214Constrained
VehicleID[26]221,763RealX374 × 412Constrained
VeRi-Wild[27]416,314RealX415 × 354Constrained
VehicleX[48]192,150SyntheticX256 × 256Constrained
VRAI[42]137,613RealX349 × 234Constrained
DiffVERI(Ours)1,712,703Real-synthetic✓496 × 485Diverse
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.671, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Table 1. Comparison of the statistics between DiffVERI and other public vehicle Re-ID datasets. In contrast, DiffVERI is currently the largest multi-view vehicle Re-ID benchmark." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.805, + 0.758 + ], + "angle": 0, + "content": "3.3. Multi-view semantic annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "After dataset collection, 10 well-trained annotators were employed, and 1200 hours were spent annotating \\(20\\mathrm{k}\\) images with multi-view semantic segmentation masks. In particular, each vehicle is uniformly parsed into a background region and four view regions (i.e., front, rear, side, and top views). Some images undeniably lack all four views or demonstrate partially occluded views. Accordingly, images for annotation that have comprehensive view information are selected. Considering the powerful zero-shot semantic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4703" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.15, + 0.101, + 0.856, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.381, + 0.908, + 0.438 + ], + "angle": 0, + "content": "Figure 2. Overall pipeline of the proposed VehicleMAE. Firstly, VMIM module utilizes two homogeneous encoder-decoders to reconstruct a pair of asymmetric masked-out patches for RGB pixels and multi-view semantic tasks. Subsequently, to fully develop the potential of homogeneous architecture, PPMD module exploits the mutual knowledge exchange between the historical and current models to enhance the generalization performance of the pre-training model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.464, + 0.482, + 0.554 + ], + "angle": 0, + "content": "segmentation capability of the SAM model [21], the annotated masks are utilized as the ground-truth, and SAM is fine-tuned through a lightweight LoRA adapter [16]. The fine-tuned SAM can efficiently perform multi-view semantic inference on the remaining unlabeled images, thereby obtaining multi-view annotations for the entire DiffVERI." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.556, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Following the above steps, a large-scale vehicle Re-ID dataset, DiffVERI, is ultimately constructed, which includes raw VERI-Wild and synthesized multi-view data. Table 1 reviews some attributes of existing vehicle Re-ID datasets, including number of images, data sources, multiview semantic annotations, average image resolution, and range of views." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.678, + 0.221, + 0.694 + ], + "angle": 0, + "content": "4. VehicleMAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Motivated by the success of MAE, we propose a novel masked image modeling pre-training paradigm, VehicleMAE, which is outlined in Figure 2, including two modules, namely view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). In VMIM module, given an image \\(I_{img}\\) and multi-view annotation \\(I_{seg}\\), a pair of visible patch maps with non-overlapping patches is generated based on an asymmetric mask sampling strategy to ensure that two homogeneous masked autoencoders \\((\\phi_1\\) and \\(\\phi_2)\\) can perform different degrees of self-supervised reconstruction tasks within one epoch. More specifically, \\(\\phi_1\\) and \\(\\phi_2\\) comprise its own encoder, RGB decoder, and view semantic decoder, that is," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.648 + ], + "angle": 0, + "content": "\\(\\phi_1 = \\{E_{\\phi_1},D_{\\phi_1}^{img},D_{\\phi_1}^{seg}\\}\\) and \\(\\phi_{2} = \\{E_{\\phi_{2}},D_{\\phi_{2}}^{img},D_{\\phi_{2}}^{seg}\\}\\), respectively. The map of features output by the encoders is \\(\\{F_{\\phi_1}^{img},F_{\\phi_1}^{seg},F_{\\phi_2}^{img},F_{\\phi_2}^{seg}\\}\\), and the map of features output by the decoders is \\(\\{f_{\\phi_1}^{img},f_{\\phi_1}^{seg},f_{\\phi_2}^{img},f_{\\phi_2}^{seg}\\}\\). Then, the PPMD module enables two MAEs to fully exchange their current and past knowledge. In the current epoch \\(N\\), the decoders of \\(\\phi_1\\) and \\(\\phi_2\\) act as collaborative training for students to distill the predicted target features from each other. Meanwhile, in the previous epoch \\(N - 1\\), the two histories of \\(\\phi_1'\\) and \\(\\phi_2'\\) are regarded as teachers exchanging students \\(\\phi_1\\) and \\(\\phi_2\\) with each other and then distilling the intermediate features of encoders." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.879, + 0.683 + ], + "angle": 0, + "content": "4.1. View-asymmetry Masked Image Modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The fundamental principle of existing image-based MAEs is to perform pixel-level masking and reconstruction using an encoder-decoder architecture. When MAE is extended to vehicle Re-ID, we empirically argue that there exist two main defects: (1) learning robust representation is difficult due to random mask sampling and low-level pixel reconstruction; and (2) reconstructing the original image pixels inevitably involves some identity-unrelated redundant region, leading to an inefficient and resource-intensive masked image modeling task. Consequently, this paper further develops VMIM, which has two homogeneous MAEs \\(\\phi_{1}\\) and \\(\\phi_{2}\\). As depicted in Figure 3, this VMIM design incorporates the following two submodules to tackle the aforementioned two issues encountered by vanilla MAE" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4704" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.096, + 0.492, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.096, + 0.858, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.25, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 3. VMIM consist of two submodules: (a) The asymmetric mask sampling strategy in VMIM module generates a pair of visible maps without overlapping patches to create diverse preservation clues for reconstruction tasks. (b) Illustration of the identity-related region reconstruction in terms of \\(\\phi_1\\). For the feature output of the decoder, RGB pixel and multi-view semantic reconstruction with MSE loss is performed only for the vehicle body region. Notably, another peer \\(\\phi_2\\) also performs a series of reconstruction processes similar to \\(\\phi_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.196, + 0.349 + ], + "angle": 0, + "content": "point-by-point:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.352, + 0.485, + 0.9 + ], + "angle": 0, + "content": "(i) Asymmetric mask sampling strategy. Initially, ours intention is to fully utilize the potential of self-supervised pre-training to serve downstream tasks with view variation by employing two homogeneous architectures. However, two twin MAEs will inevitably fail to expand the scope of their knowledge if the same masking and reconstruction operations are employed to facilitate equal knowledge exchange. Motivated by these analyses, an asymmetric mask sampling strategy is designed to generate paired visible maps without overlapping patches and encourage two MAEs to learn diverse knowledge, as shown in Figure 3(a). Concretely, the image \\( I_{img} \\) and the corresponding multi-view mask \\( I_{seg} \\) are simultaneously injected into the encoders of \\( E_{\\phi_1} \\) and \\( E_{\\phi_2} \\), and are respectively divided into two sequences with \\( N \\) patches \\( \\{P(I_{img}|\\phi_1); P(I_{seg}|\\phi_1)\\} \\) and \\( \\{P(I_{img}|\\phi_2); P(I_{seg}|\\phi_2)\\} \\) in spatial order. Subsequently, based on the semantic segmentation annotation of \\( I_{seg} \\), each sequence is further subdivided into two types: (1) subsequences \\( P(I^{id}|\\phi_1) = \\{P(I_{img}^{id}|\\phi_1); P(I_{seg}^{id}|\\phi_1)\\} \\) and \\( P(I^{id}|\\phi_2) = \\{P(I_{img}^{id}|\\phi_2); P(I_{seg}^{id}|\\phi_2)\\} \\) of the vehicle body region with size \\( N_{id} \\); (2) subsequences \\( P(I^{bg}|\\phi_1) = \\{P(I_{img}^{bg}|\\phi_1); P(I_{seg}^{bg}|\\phi_1)\\} \\) and \\( P(I^{bg}|\\phi_2) = \\{P(I_{img}^{bg}|\\phi_2); P(I_{seg}^{bg}|\\phi_2)\\} \\) of the background region with size \\( N_{bg} \\), thereby also fulfilling \\( N_{id} + N_{bg} = N \\). The pair of visible patch maps \\( \\{M_{\\phi_1}, M_{\\phi_2}\\} \\) obtained through the asymmetric mask sampling strategy follows two-point principles. The first point is to randomly mask the patches of the background region \\( P(I^{bg}|\\phi_1) \\) and \\( P(I^{bg}|\\phi_2) \\) as much as possible and preserve the vehicle body region \\( P(I^{id}|\\phi_1) \\) and \\( P(I^{id}|\\phi_2) \\) as visible clues for the reconstruction task. The second point is that no overlap simultaneously exists in the masking patches of two MAEs, that is, \\( M_{\\phi_1} \\cap M_{\\phi_2} = \\emptyset \\). Notably, \\( \\{M_{\\phi_1}, M_{\\phi_2}\\} \\) will be will be complemented each other in RGB pixels and multi-view semantic reconstruction to enhance collaboration between the two types of tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.908, + 0.695 + ], + "angle": 0, + "content": "(ii) Identity-related region reconstruction. Directly using the entire original image and a multi-view segmentation mask as the prediction target of reconstruction inevitably increases the computational cost and reduces the pre-training efficiency. Considering the aforementioned analysis, an identity-related region reconstruction method is designed. First, similar to the tokenizing and encoding operations of ordinary MAEs, the visible patch maps \\(\\{M_{\\phi_1}, M_{\\phi_2}\\}\\) obtained by asymmetric mask sampling are tokenized and embedded into a sequence of tokens, and the masked patches will be replaced with a sequence of learnable masked tokens. Afterward, the visible and invisible tokens are combined and then inputted into the decoders \\(\\{D_{\\phi_1}^{img}, D_{\\phi_1}^{seg}\\}\\) and \\(\\{D_{\\phi_2}^{img}, D_{\\phi_2}^{seg}\\}\\) of the two MAEs. Ultimately, in the decoding stage, the ground-truth of two task-specific reconstructions is set with identity-related regions \\(P(I^{id}|\\phi_1)\\) and \\(P(I^{id}|\\phi_2)\\), that is, the RGB pixels and multi-view semantic regions of the vehicle body, instead of the initial \\(I_{img}\\) and \\(I_{seg}\\). Taking \\(\\phi_1\\) as an example, Figure 3(b) illustrates the reconstruction process of the proposed method. Mean square error (MSE) is adopted to calculate the error of RGB pixels and multi-view semantics in the reconstructed body region. The loss function \\(L_{rec}\\) of the entire reconstruction task can be defined by Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.706, + 0.912, + 0.753 + ], + "angle": 0, + "content": "\\[\nL _ {r e c} = \\sum_ {m \\in \\{1, 2 \\}} \\sum_ {\\triangle \\in \\{i m g, s e g \\}} M S E \\left(R \\left(I _ {\\triangle} ^ {i d} \\mid \\phi_ {m}\\right), P \\left(I _ {\\triangle} ^ {i d} \\mid \\phi_ {m}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.909, + 0.875 + ], + "angle": 0, + "content": "where the term \\(\\phi_{m}\\) and \\(m\\in \\{1,2\\}\\) represent two models \\(\\phi_{1}\\) and \\(\\phi_{2}\\), respectively. The term \\(\\triangle \\in \\{img,seg\\}\\) denotes the reconstruction task of RGB pixels and multi-view semantics. \\(R(I_{\\triangle}^{id}|\\phi_m)\\) and \\(P(I_{\\triangle}^{id}|\\phi_m)\\) are the predicted results and ground-truth of the model \\(\\phi_{m}\\) performing two reconstruction tasks, respectively. Correspondingly, function \\(MSE(\\cdot)\\) is the MSE between the predicted target and ground-truth." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4705" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.396, + 0.107 + ], + "angle": 0, + "content": "4.2. Past-to-present Mutual-distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.278 + ], + "angle": 0, + "content": "While the VMIM module can diversify the knowledge learned by \\(\\phi_{1}\\) and \\(\\phi_{2}\\), empirical investigation revealed that this module cannot guarantee collaborative training between the two models and deep mining of their self-knowledge. Toward this end, this study aimed to establish the teacher-student role relationship, uniting the past and present knowledge of two models, and design a past-to-present mutual distillation method. More precisely, the core of this design is to conduct feature-level knowledge distillation through the collaboration of current models and the cross-guidance of historical models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.464 + ], + "angle": 0, + "content": "Assuming the current pre-training is at epoch \\(N\\), \\(\\phi_1\\) and \\(\\phi_2\\) play the roles of two student models. They collaboratively learn and exchange the predicted features with each other based on asymmetric visible clues to enhance the multi-view generalization capability of the entire VehicleMAE. The predicted feature maps \\(\\{f_{\\phi_1}^{img}, f_{\\phi_1}^{seg}\\}\\) and \\(\\{f_{\\phi_2}^{img}, f_{\\phi_2}^{seg}\\}\\) of the decoder are extracted for \\(\\phi_1\\) and \\(\\phi_2\\) in two specific reconstruction tasks, respectively. Kullback-Leibler (KL) divergence is then adopted to minimize the predicted feature distribution error. The distillation losses \\(L_{pr}^{\\phi_1}\\) and \\(L_{pr}^{\\phi_2}\\) of the two MAEs in the current epoch \\(N\\) are represented by Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.469, + 0.484, + 0.496 + ], + "angle": 0, + "content": "\\[\nL _ {p r} ^ {\\phi_ {1}} = K L \\left(f _ {\\phi_ {2}} ^ {i m g} \\mid \\mid f _ {\\phi_ {1}} ^ {i m g}\\right) + K L \\left(f _ {\\phi_ {2}} ^ {s e g} \\mid \\mid f _ {\\phi_ {1}} ^ {s e g}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.492, + 0.434, + 0.512 + ], + "angle": 0, + "content": "\\[\nL _ {p r} ^ {\\phi_ {2}} = K L \\left(f _ {\\phi_ {1}} ^ {i m g} | | f _ {\\phi_ {2}} ^ {i m g}\\right) + K L \\left(f _ {\\phi_ {1}} ^ {s e g} | | f _ {\\phi_ {2}} ^ {s e g}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.546 + ], + "angle": 0, + "content": "where \\(KL(\\cdot)\\) is the KL divergence function that measures the predicted features of two models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.814 + ], + "angle": 0, + "content": "When \\(N\\geq 2\\) , the two historical models \\(\\phi_1^{\\prime}\\) and \\(\\phi_2^\\prime\\) corresponding to epoch \\(N - 1\\) in the past are regarded as teachers cross-guiding student models \\(\\phi_{1}\\) and \\(\\phi_{2}\\) of epoch \\(N\\) to dynamically mine their own knowledge. Specifically, \\(\\phi_1^{\\prime}\\) is set as the historical teacher for \\(\\phi_{2}\\) and \\(\\phi_2^\\prime\\) as the historical teacher for \\(\\phi_{1}\\) to establish the teacher-student relationships from the past to the present. This mutual exchange of teacher-student relationships can further enhance the potential for mutual distillation and knowledge diversity. The encoded feature maps \\(\\{F_{\\phi_1'}^{img},F_{\\phi_1'}^{seg},F_{\\phi_2'}^{img},F_{\\phi_2'}^{seg}\\}\\) and \\(\\{F_{\\phi_1}^{img},F_{\\phi_1}^{seg},F_{\\phi_2}^{img},F_{\\phi_2}^{seg}\\}\\) output by the encoders of two historical teacher models and two current student models are employed to distill intermediate features due to the inability of self-supervised pre-training to achieve hard-label distillation. The distillation loss \\(L_{pa}^{\\phi_1}\\) guided by \\(\\phi_2^\\prime\\) for \\(\\phi_{1}\\) and the distillation loss \\(L_{pa}^{\\phi_2}\\) guided by \\(\\phi_1^\\prime\\) for \\(\\phi_{2}\\) are defined as Eq.5:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.818, + 0.483, + 0.848 + ], + "angle": 0, + "content": "\\[\nL _ {p a} ^ {\\phi_ {1}} = K L \\left(F _ {\\phi_ {2} ^ {\\prime}} ^ {i m g} \\mid \\mid F _ {\\phi_ {1}} ^ {i m g}\\right) + K L \\left(F _ {\\phi_ {2} ^ {\\prime}} ^ {s e g} \\mid \\mid F _ {\\phi_ {1}} ^ {s e g}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.844, + 0.44, + 0.867 + ], + "angle": 0, + "content": "\\[\nL _ {p a} ^ {\\phi_ {2}} = K L (F _ {\\phi_ {1} ^ {\\prime}} ^ {i m g} | | F _ {\\phi_ {2}} ^ {i m g}) + K L (F _ {\\phi_ {1} ^ {\\prime}} ^ {s e g} | | F _ {\\phi_ {2}} ^ {s e g})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "After the pre-training of VehicleMAE is completed, all decoders will be discarded and the learned knowledge will" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.139 + ], + "angle": 0, + "content": "be fused by solving the average parameters of the two encoders \\( E_{\\phi_1} \\) and \\( E_{\\phi_2} \\). The process of obtaining the mean model \\( E[\\bar{\\theta}] \\) can be written as Eq. 6:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.149, + 0.905, + 0.18 + ], + "angle": 0, + "content": "\\[\nE [ \\bar {\\theta} ] = \\frac {E _ {\\phi_ {1}} [ \\theta_ {\\phi_ {1}} ] + E _ {\\phi_ {2}} [ \\theta_ {\\phi_ {2}} ]}{2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.907, + 0.268 + ], + "angle": 0, + "content": "where \\(\\theta_{\\phi_1}\\) and \\(\\theta_{\\phi_2}\\) are the parameter of \\(E_{\\phi_1}\\) and \\(E_{\\phi_2}\\), respectively, and \\(\\bar{\\theta}\\) is the parameter of the mean model \\(E[\\bar{\\theta}]\\). It is worth emphasizing that \\(E[\\bar{\\theta}]\\) will be directly applied as a single pre-training model to downstream tasks of supervised and unsupervised vehicle Re-ID." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.277, + 0.696, + 0.293 + ], + "angle": 0, + "content": "4.3. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.906, + 0.421 + ], + "angle": 0, + "content": "The training objectives of VehicleMAE include two parts: the loss of the reconstruction task of two homogeneous MAEs in the VMIM module; in the PPMD module, the distillation loss \\( L_{pr} = L_{pr}^{\\phi_1} + L_{pr}^{\\phi_2} \\) between the current two student models and the distillation loss \\( L_{pa} = L_{pa}^{\\phi_1} + L_{pa}^{\\phi_2} \\) guided by the history teachers. Therefore, the total loss \\( L_{total} \\) of the overall pre-training process can be defined by Eq. 7 as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.422, + 0.905, + 0.439 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L _ {r e c} + L _ {p r} + L _ {p a} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.906, + 0.504 + ], + "angle": 0, + "content": "Note that, the pre-training mean model \\( E[\\bar{\\theta}] \\) directly serves as supervised and unsupervised vehicle Re-ID downstream tasks and is supervised by cross-entropy and triplet losses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.646, + 0.537 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.813, + 0.56 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.672 + ], + "angle": 0, + "content": "This paper utilizes two well-known vehicle Re-ID datasets: VeRi-776 [23] and VehicleID [3]. VehicleID dataset includes small, medium, and large-scale test sets. To evaluate the performance of the proposed method, mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank-\\(k\\) (\\(k = 1, 5\\)) are employed as the evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.731, + 0.699 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.9 + ], + "angle": 0, + "content": "CLIP-B/16 serves as the backbone network. Throughout the whole training process, the image size is set to \\(256 \\times 256\\), and models update parameters using the Adam optimizer. During the pre-training phase, all samples are divided into mini-batches of size 128 for 300 epochs of iterative learning. For downstream tasks, the batch size is set to 64, and all samples are trained for 50 epochs. Additionally, each image undergoes data augmentation with random horizontal flipping, padding, cropping, and erasing. In unsupervised tasks, DBSCAN is applied as the clustering algorithm to generate pseudo-labels. All experiments are conducted under the Ubuntu 22.04 operating system and Pytorch environment with 4 NVIDIA RTX A6000 GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4706" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.355 + ], + "angle": 0, + "content": "Effects of different modules. The results in Table 2 demonstrate the effectiveness of all proposed modules. Three observations are summarized by us: (1) When we independently introduce the \"Pa\" or \"Pr\" into the \"Baseline\" to perform knowledge distillation, the significant performance improvement in Rank accuracy and mAP confirms that historical and current distillation can fully explore and exchange their own knowledge. (2) Whether using \"Asym\" in the mask sampling strategy during the encoding stage or introducing \"Id-rec\" in the reconstruction task during the decoding stage, it is superior to MAE that integrates a random mask strategy and complete pixel reconstruction. (3) VehicleMAE belongs to the ensemble of all modules and achieves the best performance, further verifying that VehicleMAE can efficiently pre-train two homogeneous MAEs in a collaborative manner." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.368, + 0.486, + 0.496 + ], + "angle": 0, + "content": "
Ablation ModulesVeRi-776VehicleID
BaselinePaPrAsymId-recmAPRank-1Rank-1Rank-5
✓37.681.269.081.8
✓✓40.185.071.284.6
✓✓40.585.371.184.7
✓✓38.783.670.382.8
✓✓39.584.370.182.6
✓✓✓✓✓42.587.673.385.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Table 2. Performance comparison of different components in unsupervised vehicle ReID. \"Past\" and \"Pres\" refer to historical distillation and present distillation in PPDM, respectively. \"Asym\" and \"Id-rec\" correspond to the Asymmetric mask sampling strategy and Identity-related region reconstruction in VMIM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Effects of different mask sampling strategies and ratios. Table 3 reports the impact of using asymmetric mask and random mask sampling strategies at different mask ratios. We observe that the performance of the proposed \"Asym\" is superior to \"Random\" for all three mask ratios. Especially at a mask rate of \\(75\\%\\), \"Asym\" achieves the highest mAP of \\(42.5\\%\\) and Rank-1 accuracy of \\(87.6\\%\\). This is due to the visible maps generated by \"Asym\" having no overlapping patches, which encourage two MAEs to learn and exchange knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effects of different distillation strategies. Table 4 investigates the impact of modifying VehicleMAE using various distillation architectures on Re-ID performance. Ours findings are two-fold: (1) Compared to two MAEs learning independently, using historical distillation or present distillation can ameliorate the pre-training model performance, especially when combined to achieve more effective knowledge exchange. (2) The mean encoder \\( E[\\bar{\\theta}] \\) can accurately feedback and maintain the average level of two encoders, while also solving the difficulties of multiple model selec" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.09, + 0.912, + 0.24 + ], + "angle": 0, + "content": "
Mask RatiosStrategiesVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
25%Random38.884.268.483.6
Asym40.685.570.484.2
50%Random39.284.668.983.8
Asym41.786.271.184.8
75%Random40.185.169.584.1
Asym42.587.673.385.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.25, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Table 3. Effectiveness of different mask sampling strategies in serving unsupervised vehicle Re-ID tasks. \"Random\" and \"Asym\" express the random mask sampling strategy and the asymmetric mask sampling strategy, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.331, + 0.8, + 0.347 + ], + "angle": 0, + "content": "tion and deployment for downstream tasks." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.357, + 0.907, + 0.43 + ], + "angle": 0, + "content": "
Pre-trained EncodersVeRi-776VehicleID
IndPaPrPa+PrIndPaPrPa+Pr
Eφ184.384.784.887.170.370.871.073.2
Eφ284.484.785.186.970.370.770.973.1
E[θ]84.685.085.387.670.971.271.173.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.538 + ], + "angle": 0, + "content": "Table 4. Comparison Rank-1 accuracy of two homogeneous encoders \\(\\{E_{\\phi_1}, E_{\\phi_2}\\}\\) and a mean encoder \\(E[\\tilde{\\theta}]\\) using various distillation architectures in unsupervised downstream tasks. \"Ind\" denotes that \\(E_{\\phi_1}\\) and \\(E_{\\phi_2}\\) learning independently without using distillation. \"Pa\", \"Pr\", and \"Pa+Pr\" respectively signify the independent use of historical distillation, present distillation, and the simultaneous use of both them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.841 + ], + "angle": 0, + "content": "Analysis of Model Complexity. To further analyze the computational performance, we re-implement different self-supervised pre-training works on DiffVERI for model parameter sizes and floating-point operations (FLOPs), as shown in Table 5. For fairness, all methods except ConvNeXt v2, which relies on the ConvNeXt network architecture, use the CLIP-B/16 image encoder as the backbone. Although VehicleMAE adopts a mutual-distillation architecture, our pre-training model contains 57.3M parameters and only needs 14.7G multiply-accumulate operations for feature extraction. This is mainly due to the mean model design and that we follow the vanilla MAE without directly changing the encoder-decoder architecture. In contrast, both the fully convolutional mask autoencoder in ConvNeXt v2 and the hybrid convolution module in CMAE will increase the resource consumption of feature extraction. These results demonstrate that VehicleMAE achieves a favorable trade-off between model complexity and the performance of downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.816, + 0.866 + ], + "angle": 0, + "content": "5.4. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Table 6, we compare the proposed method with state-of-the-art ones in both supervised and unsuper" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4707" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.165, + 0.089, + 0.404, + 0.164 + ], + "angle": 0, + "content": "
MethodsFLOPsParams
ConvNeXt v2[45]15.4G89M
CMAE[18]19.4G104M
BEiT[43]15.7G82.5M
Ours14.7G57.3M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.174, + 0.486, + 0.203 + ], + "angle": 0, + "content": "Table 5. Comparison of different pre-training models in terms of parameter size and FLOPs." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.217, + 0.47, + 0.659 + ], + "angle": 0, + "content": "
MethodsVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
Supervised Learning
\\( A\\text{VER}{*}^{R}\\left\\lbrack {19}\\right\\rbrack \\)61.289.063.585.6
\\( VehicleNet*R\\left\\lbrack {54}\\right\\rbrack \\)83.496.879.592.0
\\( TransReID^{*V}\\left\\lbrack {14}\\right\\rbrack \\)82.097.1--
\\( GiT^{*V}\\left\\lbrack {35}\\right\\rbrack \\)80.396.977.9-
\\( LCNL^{*V}\\left\\lbrack {46}\\right\\rbrack \\)81.897.4--
\\( TANet^{*R}\\left\\lbrack {17}\\right\\rbrack \\)83.696.878.292.6
\\( CFA-Net^{*R}\\left\\lbrack {41}\\right\\rbrack \\)80.796.9--
\\( HCI-Net^{*R}\\left\\lbrack {36}\\right\\rbrack \\)83.896.676.491.2
\\( CLIP-ReID^{*C}\\left\\lbrack {24}\\right\\rbrack \\)83.397.478.192.7
\\( Ours^{\\dagger C} \\)87.697.485.994.9
Unsupervised Learning
\\( MMT^{*R}\\left\\lbrack {7}\\right\\rbrack \\)25.460.931.042.4
\\( SPCL^{*R}\\left\\lbrack {8}\\right\\rbrack \\)36.979.953.066.4
\\( RLCC^{*R}\\left\\lbrack {51}\\right\\rbrack \\)39.683.4--
\\( PPLR^{*R}\\left\\lbrack {4}\\right\\rbrack \\)41.685.6--
\\( VAPC TO^{*R}\\left\\lbrack {52}\\right\\rbrack \\)30.476.2--
\\( ICL^{*R}\\left\\lbrack {37}\\right\\rbrack \\)39.583.7--
\\( AdaMG^{*R}\\left\\lbrack {31}\\right\\rbrack \\)41.086.2--
\\( NNNI^{*R}\\left\\lbrack {9}\\right\\rbrack \\)42.386.3--
\\( STDA^{*R}\\left\\lbrack {13}\\right\\rbrack \\)42.387.4--
\\( MAE†^C\\left\\lbrack {12}\\right\\rbrack \\)37.681.269.081.8
\\( CL-MAE†^C\\left\\lbrack {29}\\right\\rbrack \\)41.084.070.882.6
\\( CMAE†^C\\left\\lbrack {18}\\right\\rbrack \\)41.684.771.184.9
\\( Ours^{\\dagger C} \\)42.587.673.385.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.768 + ], + "angle": 0, + "content": "Table 6. Performance comparisons with the latest methods on supervised and unsupervised vehicle Re-ID tasks. “\\(*^{R}\\), “\\(*^{V}\\), and “\\(*^{C}\\)” indicate using ResNet, ViT, and the image encoder of CLIP as backbones, respectively. “\\(\\dagger^{C}\\)” denotes the implementation of pre-training using the image encoder of CLIP on DiffVERI. Note that experimental results are accurate to one decimal place, and this paper uses only the large-scale test set of VehicleID." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "vised downstream tasks. Compared to the most competitive methods among ResNet-based, ViT-based, and CLIP-based backbones, our method delivers the best mAP and Rank accuracy for both supervised and unsupervised settings. For fair comparison, we further implement several well-established MAE-based pre-training methods on DiffVERI. The noticeable performance improvement implies" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.09, + 0.88, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.216, + 0.907, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Visualization of patch-level feature distributions extracted by Baseline and using VehicleMAE pre-training. The class with the highest proportion of pixels within each patch in the patch-level ground-truth is defined as the final class corresponding to the patch. Different color patches represent different views or background regions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.357 + ], + "angle": 0, + "content": "that the masked image modeling based on VehicleMAE is better suited for downstream tasks of vehicle Re-ID." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.369, + 0.707, + 0.385 + ], + "angle": 0, + "content": "5.5. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.619 + ], + "angle": 0, + "content": "Figure 4 displays two visualization examples of the feature distribution at the patch-level extracted by Baseline and VehicleMAE. Specifically, we feed the features of all patches into the K-Means algorithm to perform five class-based clustering operations (i.e., a background region and four view regions) to intuitively demonstrate the view discriminate ability of different pre-training models. It can be observed that our VehicleMAE is able to distinguish patches from different views and backgrounds in both examples, getting close to patch-level ground-truth. Remarkably, even in stricter occlusion scenarios, VehicleMAE accurately avoid interference from the power poles in example B, further completing the extraction of multi-view information. More details on qualitative visualization experiments are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.649 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper releases DiffVERI, a large-scale multi-view vehicle Re-ID dataset for learning view-invariant representations, and proposes a masked image modeling pre-training paradigm termed VehicleMAE specially for vehicle Re-ID downstream tasks. VehicleMAE first proposes a VMIM module that attempts to apply two homogeneous MAEs to predict the RGB pixels and multi-view semantic clues of vehicles in pairs, thereby gaining diverse multi-view inference capabilities. Subsequently, to facilitate learning collaboratively, a PPMD module is designed to progressively exchange knowledge with each other. Extensive experiments demonstrate that equipping our pre-training model can achieve competitive performance in generic vehicle Re-ID downstream tasks. Future work contributes to further expanding VehicleMAE into a unified multimodal pre-training paradigm." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4708" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant No. 62461037 and the Jiangxi Provincial Natural Science Foundation under Grant Nos. 20224BAB212011, 20232BAB202051, 20232BAB212008, and 20242BAB25078." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[1] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In European Conference on Computer Vision, pages 348-367. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[3] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 347–355, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[4] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7308–7318, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[6] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750–14759, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[7] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. Advances in neural information processing systems, 33:11309-11321, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[9] Dongchen Han, Weifeng Liu, Mingchen Zou, and Baodi Liu. Non-contrastive nearest neighbor identity-guided method for unsupervised object re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2713–2723, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Bing He, Jia Li, Yifan Zhao, and Yonghong Tian. Part-regularized near-duplicate vehicle re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3997–4005, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.233, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[13] Qiaolin He, Zihan Wang, Zhijie Zheng, and Haifeng Hu. Spatial and temporal dual-attention for unsupervised person re-identification. IEEE Transactions on Intelligent Transportation Systems, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[14] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013–15022, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[15] Zhijun He, Hongbo Zhao, Jianrong Wang, and Wenquan Feng. Multi-level progressive learning for unsupervised vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(4):4357-4371, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[17] Wenbo Hu, Hongjian Zhan, Palaiahnakote Shivakumara, Umapada Pal, and Yue Lu. Tanet: Text region attention learning for vehicle re-identification. Engineering Applications of Artificial Intelligence, 133:108448, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[18] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Pirazh Khorramshahi, Amit Kumar, Neehar Peri, Sai Saketh Rambhatla, Jun-Cheng Chen, and Rama Chellappa. A dual-path model with adaptive attention for vehicle re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6132–6141, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[20] Pirazh Khorramshahi, Vineet Shenoy, and Rama Chellappa. Robust and scalable vehicle re-identification via self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5295–5304, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[22] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE transactions on image processing, 31:5949-5962, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[23] Hongchao Li, Aihua Zheng, Liping Sun, and Yonglong Luo. Camera topology graph guided vehicle re-identification." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4709" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "IEEE Transactions on Multimedia, 26:1565-1577, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1405–1413, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[25] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2167-2175, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[26] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In 2016 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[27] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. Veri-wild: A large dataset and a new method for vehicle re-identification in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3235-3243, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[28] Zefeng Lu, Ronghao Lin, Qiaolin He, and Haifeng Hu. Mask-aware pseudo label denoising for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 24(4):4333-4347, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[29] Neelu Madan, Nicolae-Catalin Ristea, Kamal Nasrollahi, Thomas B Moeslund, and Radu Tudor Ionescu. Cl-mae: Curriculum-learned masked autoencoders. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2492–2502, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[30] Dechao Meng, Liang Li, Xuejing Liu, Yadong Li, Shijie Yang, Zheng-Jun Zha, Xingyu Gao, Shuhui Wang, and Qingming Huang. Parsing-based view-aware embedding network for vehicle re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7103–7112, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[31] Jinjia Peng, Guangqi Jiang, and Huibing Wang. Adaptive memorization with group labels for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5802-5813, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[32] Jiuchao Qian, Minting Pan, Wei Tong, Rob Law, and Edmond Q Wu. Urrnet: A unified relational reasoning network for vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(9):11156-11168, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[34] Zhiyin Shao, Xinyu Zhang, Changxing Ding, Jian Wang, and Jingdong Wang. Unified pre-training with pseudo texts for text-to-image person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11174–11184, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Fei Shen, Yi Xie, Jianqing Zhu, Xiaobin Zhu, and Huanqiang Zeng. Git: Graph interactive transformer for vehicle re-identification. IEEE Transactions on Image Processing, 32:1039–1051, 2023. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[36] Ke Sun, Xiyu Pang, Meifeng Zheng, Xiushan Nie, Xi Li, Houren Zhou, and Yilong Yin. Heterogeneous context interaction network for vehicle re-identification. Neural Networks, 169:293–306, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[37] Qing Tang and Kang-Hyun Jo. Unsupervised object re-identification via instances correlation loss. In 2022 IEEE 20th International Conference on Industrial Informatics (IN-DIN), pages 135–139. IEEE, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078-10093, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.904, + 0.319 + ], + "angle": 0, + "content": "[39] Jingzheng Tu, Cailian Chen, Xiaolin Huang, Jianping He, and Xinping Guan. Dfr-st: Discriminative feature representation with spatio-temporal cues for vehicle re-identification. Pattern Recognition, 131:108887, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[40] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464-7475, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[41] Hai Wang, Yaqing Niu, Long Chen, Yicheng Li, Miguel Angel Sotelo, Zhixiong Li, and Yingfeng Cai. Dair-v2xreid: A new real-world vehicle-infrastructure cooperative re-id dataset and cross-shot feature aggregation network perception method. IEEE Transactions on Intelligent Transportation Systems, 25(8):9058-9068, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[42] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 460–469, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[43] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[44] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In Proceedings of the IEEE international conference on computer vision, pages 379–387, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[45] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16133-16142, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[46] Mouxing Yang, Zhenyu Huang, and Xi Peng. Robust object re-identification with coupled noisy labels. International Journal of Computer Vision, pages 1–19, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4710" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492-4501, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[48] Yue Yao, Liang Zheng, Xiaodong Yang, Milind Naphade, and Tom Gedeon. Simulating content consistent vehicle datasets with attribute descent. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 775-791. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[49] Hongwei Zhang, Guoqing Zhang, Yuhao Chen, and Yuhui Zheng. Global relation-aware contrast learning for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8599–8610, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[50] Kevin Zhang and Zhiqiang Shen. i-mae: Are latent representations in masked autoencoders linearly separable? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7740-7749, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[51] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3436–3445, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[52] Aihua Zheng, Xia Sun, Chenglong Li, and Jin Tang. Aware progressive clustering for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(8):11422-11435, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[53] Aihua Zheng, Chaobin Zhang, Chenglong Li, Jin Tang, and Chang Tan. Multi-query vehicle re-identification: Viewpoint-conditioned network, unified dataset and new metric. IEEE Transactions on Image Processing, 32:5948–5960, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[54] Zhedong Zheng, Tao Ruan, Yunchao Wei, Yi Yang, and Tao Mei. Vehiclenet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia, 23:2683–2693, 2020. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4711" + } + ] +] \ No newline at end of file diff --git a/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_origin.pdf b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e9840edd9671c2088930f93345815c4d421db06 --- /dev/null +++ b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/ec3290f4-caa9-4573-afb5-3440940874ee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:945dbbeffae6695bbbe2e1ed8f7ec660d4d8e87f30c404a0e7264275078efbd4 +size 22014483 diff --git a/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/full.md b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/full.md new file mode 100644 index 0000000000000000000000000000000000000000..53ddff065da6ab5ada3299a7af751a0c0e1ac205 --- /dev/null +++ b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/full.md @@ -0,0 +1,294 @@ +# VehicleMAE: View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders + +Qi Wang $^{1}$ , Zeyu Zhang $^{1}$ , Dong Wang $^{2,*}$ , Di Gai $^{1}$ , Xin Xiong $^{3}$ , Jiyang Xu $^{1}$ and Ruihua Zhou $^{2}$ + +$^{1}$ School of Mathematics and Computer Sciences, Nanchang University, Nanchang, China + +$^{2}$ School of Software, Nanchang University, Nanchang, China + +3The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China + +{wangqi,gaidi,xiongxinxx}@ncu.edu.cn，{zeyuzhang1010,dongwang,xujiyang,zrh}@email.ncu.edu.cn + +# Abstract + +Large-scale pre-training technology has achieved remarkable performance in diversified object re-identification (Re-ID) downstream tasks. Nevertheless, to our best knowledge, the pre-training model specifically for vehicle Re-ID, which focuses on tackling the challenge of multi-view variations, has not been fully investigated. In this paper, we first leverage a diffusion model to build a large-scale vehicle Re-ID benchmark dataset, dubbed "DiffVERI", containing over 1700K images from abundant multi-view annotations. In terms of this dataset, we further present VehicleMAE, a novel masked image modeling pre-training paradigm that learns view-invariant representations by performing mutual-distillation in a self-supervised manner. To be specific, the pipeline of VehicleMAE unfolds two core modules, i.e., view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). Technically, VMIM consists of two homogeneous masked autoencoders (MAE) that simultaneously reconstruct the RGB pixels and multi-view semantic information of the specific vehicle body region via paired asymmetric mask sampling strategies. To progressively distill the knowledge of the model itself, PPMD considers the two MAEs in the current epoch and the previous one as the student models and the teacher models, respectively, which leverages the knowledge learned by the current student and the historical teacher for mutual feature-level distillation. Extensive experimental results have verified that the proposed pre-training paradigm on DiffVERI gains compelling downstream task performance for vehicle Re-ID. + +# 1. Introduction + +The purpose of vehicle re-identification (Re-ID)[3, 23, 53] is to match target vehicles with the same identity under vari + +ous views. Despite impressive performance in some current works, this task still faces a main obstacle of large intraclass discrepancy among vehicles of the same identity. The primary cause is that the unique geometrically rigid structure of the vehicle itself leads to dynamic appearances of the same instance from different views. Accordingly, it is imperative to enable the Re-ID model to efficiently learn view-invariant representations. + +![](images/94418750685e995e2bcd6f2d0c957feeaa5956cda2d23a96c39faf9db2da41e4.jpg) +Figure 1. Illustration of DiffVERI dataset: (a) Steps for synthesizing data and annotations. (b) Synthetic examples from DiffVERI dataset. + +With the release of the previous generic vehicle Re-ID datasets (such as VeRi-776 [25], VehicleID [26], VERiWild [27], etc.), early efforts [10, 30, 44] preferred the use of vehicle keypoints and part-based bounding boxes for discriminative representation learning to achieve view alignment. Additionally, some follow-up works [22, 39] based on metric learning resort to vehicle attribute annotations and + +spatio-temporal information to narrow the appearance discrepancy across different views. Despite advanced progress on the aforesaid works, their training data are typically derived from real-world surveillance videos and only contain images within a fixed view range, which limits the inference ability of the Re-ID model for different views. An intuitive solution is to build a large-scale benchmark with a comprehensive view range to provide learnable view-invariant information for the Re-ID model while trading off labor costs. Until recently, there have been no existing works specifically investigating large-scale vehicle Re-ID benchmarks for a comprehensive range of views. To address the aforementioned limitations, we release a large-scale multiview dataset, named DiffVERI, as presented in Figure 1. In brief, we use the text-to-image diffusion model to synthesize vehicle data from various views. Figure 1(a) illustrates the construction process of DiffVERI, which includes three steps: multi-view data generation, data filtering, and multi-view semantic annotations. An example of generating multi-view data is also shown in Figure 1(b), and more the detailed visualizations of synthesized examples is included in the supplementary materials. + +Due to the large domain gap between the generic ImageNet and vehicle data, ImageNet cannot be well-suited for various downstream tasks of vehicle Re-ID. Therefore, it is particularly crucial to obtain initial weights with strong generalization ability through task-specific pre-training. Some pioneering works based on self-supervised pre-training have been explored in the person-related field. For instance, Fu et al. [6] constructed a LUPerson benchmark and developed a pre-training model via contrastive learning. Another effective self-supervised approach is masked image modeling technology, which performs masking and pixel-level reconstruction on the backbone of vision transformer (ViT) to serve various downstream vision tasks. In contrast, to our best knowledge, there is currently no prior research on targeted large-scale pre-training for vehicle Re-ID. That is to say, mainstream vehicle Re-ID methods are still limited by the pre-training parameters from ImageNet. + +To address the aforementioned issues, we have released a vehicle Re-ID benchmark, DiffVERI, with a comprehensive view range. Benefiting from DiffVERI, a novel masked image modeling pre-training paradigm called VehicleMAE is then proposed to serve the downstream task of vehicle Re-ID. To sum up, the main contributions of this work can be summarized as follows: + +1. We build a large-scale vehicle Re-ID benchmark, DiffVERI, by fine-tuning an off-the-shelf diffusion model, which incorporates over 1700K images and multi-view semantic annotations. To our knowledge, DiffVERI is by far the largest vehicle Re-ID dataset with diverse view data. +2. We begin the first attempt to investigate a large- + +scale pre-training solution for vehicle Re-ID tasks and therefore devise a novel masked image modeling pre-training paradigm, termed VehicleMAE, to jointly integrate the view-asymmetry masked image modeling (VMIM) module and the past-to-present mutual-distillation (PPMD) one. + +3. Comprehensive experiments have validated the extraordinary performance of our pre-training model in both supervised and unsupervised downstream tasks of vehicle Re-ID. + +# 2. Related Work + +# 2.1. Large-scale Re-ID Datasets + +The recent emergence of vehicle benchmark datasets has led to increasing attention to vehicle Re-ID technology. Existing efforts mainly follow two mainstream settings. The first setting, supervised learning [20, 32, 35], considers vehicle Re-ID as a representation learning problem exploit ID annotations to learn informative features to distinguish different vehicle instances. Early works [32, 35] focused on exploring potential correlations between global and local features to enhance the discriminative performance of Re-ID models. To distinguish vehicles with similar appearances from the same manufacturing, subsequent works attempted to feed additional attribute annotations, such as color, model, and camera ID, as auxiliary information to obtain robust identity-related features. Methods in the latter setting, i.e., unsupervised learning [4, 15, 28, 49], aim to fully learn discriminative representations of the training set without relying on any annotations. The vanilla unsupervised learning procedure updates the Re-ID model using pseudo labels generated by clustering algorithms as supervised information. For this purpose, previous works have tentatively proposed pseudo label refinement [4, 28] and contrastive learning [15, 49] to overcome the limitations of clustering-based label noise. In both settings, the aforementioned methods inevitably face the challenge of view variations due to the lack of learning view annotations. To tackle this, we first construct a large-scale vehicle dataset DiffVERI comprising multi-view annotations and then attempt to learn view-invariant features via pre-training. + +# 2.2. Self-supervised Pre-training + +Recently, the focus of vision recognition tasks has shifted from achieving advanced performance in specific downstream tasks to designing generic large-scale pre-training techniques. Labeling large-scale datasets is undoubtedly labor-intensive, while utilizing the inherent information of the data itself for self-supervised learning is a more effective way. Pioneer self-supervised pre-training works [2, 11] attempted to learn robust representation via dynamic contrastive learning and knowledge distillation on unlabeled + +large-scale datasets. As an illustration, in the Re-ID task, Fu et al. [6] constructed a large-scale person dataset called LU-Person and first designed a pre-training Re-ID model based on contrastive learning. With the advanced performance of vision transformer [5] in vision tasks, another more lightweight method, masked image modeling [18, 29, 50], has been systematically proven to be effective in the field of self-supervised pre-training. The representative work [12], i.e., masked autoencoders (MAE), presented an encoder-decoder architecture to perform pixel-level masking and reconstruction operations. Subsequently, some recent works, such as VideoMAE [38] and MultiMAE [1], have extended MAE to the fields of video and multiple vision tasks, respectively. However, to our knowledge, there is no frontier exploration applying large-scale pre-training to vehicle Re-ID research. Therefore, this paper first proposes VehicleMAE to leverage the MAE architecture for pre-training on the DiffVERI dataset, to provide a powerful pre-training model for generic vehicle Re-ID tasks. + +# 3. DiffVERI + +Owing to the remarkable success of diffusion models in the field of data generation [34, 47], a text-to-image diffusion model is employed to construct a large-scale vehicle Re-ID dataset, DiffVERI, with a comprehensive view. The construction of DiffVERI comprises the following three primary steps: + +# 3.1. Multi-view data generation + +The off-the-shelf diffusion model, termed DreamBooth [33], is used for fine-tuning on the VERI-Wild to produce highly realistic data. Specifically, a specific text description $T$ , that is, "A photo of [V] vehicle" is created for each vehicle instance, where [V] represents the identity-specific identifier to ensure ID consistency between the generated and input images. Given an initial noise of $\varepsilon \sim N(0,1)$ , the denoising process of the diffusion model is followed, and the model is then optimized by balancing the differences between the generated and original images. The noise $z_{t}$ at time $t$ in the latent space can be expressed as Eq. 1: + +$$ +z _ {t} (x, \varepsilon) = \sqrt {\tilde {a} _ {t}} \cdot x + \sqrt {1 - \tilde {a} _ {t}} \cdot \varepsilon \tag {1} +$$ + +where $\tilde{a}_t$ is the average denoising factor for the diffusion process time $t$ , and $x$ represents the original image. + +Subsequently, the text features extracted by the text encoder $\Gamma$ of DreamBooth are used as a conditioning vector $c_{\mathrm{pr}} = \Gamma \bigl ("\mathrm{a photo of vehicle}"\bigr)$ for the pre-trained text-to image diffusion model $x_{\theta}$ to generate images $x_{pr} = x_{\theta}(z_t,c_{pr})$ . As the objective function in the fine-tuning of the diffusion model process, squared error loss can be de + +fined as Eq. 2: + +$$ +\begin{array}{l} \mathbb {E} _ {x, c, \varepsilon , \varepsilon^ {\prime}, t} [ w _ {t} \| x _ {\theta} (z _ {t} (x, \varepsilon), c) - x \| _ {2} ^ {2} + \\ \left. \lambda w _ {t} ^ {\prime} \left\| x _ {\theta} \left(z _ {t} ^ {\prime} \left(x _ {p r}, \varepsilon^ {\prime}\right), c _ {p r}\right) - x _ {p r} \right\| _ {2} ^ {2} \right] \tag {2} \\ \end{array} +$$ + +where $w_{t}$ represents the quality term of sample generation, and $c$ is a conditioning vector derived from the text $T$ . $\lambda$ is the relative weight that controls a specific identity vehicle. A flexible multi-view generation prompt, that is, "A photo of [V] vehicle [O] rotated [X] degrees clockwise", where [O] and [X] are variables that control the generation direction and rotation angle, respectively, and $[O] \in left, right, [X] \in 30, 60, 90, 120, 150, 180$ , is designed to generate multi-view vehicle images during the inference process. + +# 3.2. Data filtering. + +Motivated by ours insights, two types of data noise are inevitably found in the generated data: (1) part of the images exhibit pixel blurring and low resolution, and (2) the vehicle body occupies a small proportion of the entire image, increasing the difficulty in focusing on the details of each view. Images smaller than 10K in size are first empirically filtered out to address the aforementioned issues. Subsequently, we resort to YOLOv7 [40] to detect the bounding box of the vehicle body and then calculate the proportion ratio between the vehicle body and the entire image. Based on the specifications, images with a proportion ratio below $50\%$ are further filtered. + +
DatasetsImagesSourceMulti-view Semantic AnnotationsResolutionViews
VeRi-776[25]49,357RealX243 × 214Constrained
VehicleID[26]221,763RealX374 × 412Constrained
VeRi-Wild[27]416,314RealX415 × 354Constrained
VehicleX[48]192,150SyntheticX256 × 256Constrained
VRAI[42]137,613RealX349 × 234Constrained
DiffVERI(Ours)1,712,703Real-synthetic✓496 × 485Diverse
+ +Table 1. Comparison of the statistics between DiffVERI and other public vehicle Re-ID datasets. In contrast, DiffVERI is currently the largest multi-view vehicle Re-ID benchmark. + +# 3.3. Multi-view semantic annotations. + +After dataset collection, 10 well-trained annotators were employed, and 1200 hours were spent annotating $20\mathrm{k}$ images with multi-view semantic segmentation masks. In particular, each vehicle is uniformly parsed into a background region and four view regions (i.e., front, rear, side, and top views). Some images undeniably lack all four views or demonstrate partially occluded views. Accordingly, images for annotation that have comprehensive view information are selected. Considering the powerful zero-shot semantic + +![](images/5b6c22317b955e3f3ec4b8cd11b355c4fc3593b1d04ba53bb3ea49736a306e0a.jpg) +Figure 2. Overall pipeline of the proposed VehicleMAE. Firstly, VMIM module utilizes two homogeneous encoder-decoders to reconstruct a pair of asymmetric masked-out patches for RGB pixels and multi-view semantic tasks. Subsequently, to fully develop the potential of homogeneous architecture, PPMD module exploits the mutual knowledge exchange between the historical and current models to enhance the generalization performance of the pre-training model. + +segmentation capability of the SAM model [21], the annotated masks are utilized as the ground-truth, and SAM is fine-tuned through a lightweight LoRA adapter [16]. The fine-tuned SAM can efficiently perform multi-view semantic inference on the remaining unlabeled images, thereby obtaining multi-view annotations for the entire DiffVERI. + +Following the above steps, a large-scale vehicle Re-ID dataset, DiffVERI, is ultimately constructed, which includes raw VERI-Wild and synthesized multi-view data. Table 1 reviews some attributes of existing vehicle Re-ID datasets, including number of images, data sources, multiview semantic annotations, average image resolution, and range of views. + +# 4. VehicleMAE + +Motivated by the success of MAE, we propose a novel masked image modeling pre-training paradigm, VehicleMAE, which is outlined in Figure 2, including two modules, namely view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). In VMIM module, given an image $I_{img}$ and multi-view annotation $I_{seg}$ , a pair of visible patch maps with non-overlapping patches is generated based on an asymmetric mask sampling strategy to ensure that two homogeneous masked autoencoders $(\phi_1$ and $\phi_2)$ can perform different degrees of self-supervised reconstruction tasks within one epoch. More specifically, $\phi_1$ and $\phi_2$ comprise its own encoder, RGB decoder, and view semantic decoder, that is, + +$\phi_1 = \{E_{\phi_1},D_{\phi_1}^{img},D_{\phi_1}^{seg}\}$ and $\phi_{2} = \{E_{\phi_{2}},D_{\phi_{2}}^{img},D_{\phi_{2}}^{seg}\}$ , respectively. The map of features output by the encoders is $\{F_{\phi_1}^{img},F_{\phi_1}^{seg},F_{\phi_2}^{img},F_{\phi_2}^{seg}\}$ , and the map of features output by the decoders is $\{f_{\phi_1}^{img},f_{\phi_1}^{seg},f_{\phi_2}^{img},f_{\phi_2}^{seg}\}$ . Then, the PPMD module enables two MAEs to fully exchange their current and past knowledge. In the current epoch $N$ , the decoders of $\phi_1$ and $\phi_2$ act as collaborative training for students to distill the predicted target features from each other. Meanwhile, in the previous epoch $N - 1$ , the two histories of $\phi_1'$ and $\phi_2'$ are regarded as teachers exchanging students $\phi_1$ and $\phi_2$ with each other and then distilling the intermediate features of encoders. + +# 4.1. View-asymmetry Masked Image Modeling. + +The fundamental principle of existing image-based MAEs is to perform pixel-level masking and reconstruction using an encoder-decoder architecture. When MAE is extended to vehicle Re-ID, we empirically argue that there exist two main defects: (1) learning robust representation is difficult due to random mask sampling and low-level pixel reconstruction; and (2) reconstructing the original image pixels inevitably involves some identity-unrelated redundant region, leading to an inefficient and resource-intensive masked image modeling task. Consequently, this paper further develops VMIM, which has two homogeneous MAEs $\phi_{1}$ and $\phi_{2}$ . As depicted in Figure 3, this VMIM design incorporates the following two submodules to tackle the aforementioned two issues encountered by vanilla MAE + +![](images/c5a41b115ce048202b829ea79dadde2311ee48430af7a8a8ed1c3a8a02bc759d.jpg) +Figure 3. VMIM consist of two submodules: (a) The asymmetric mask sampling strategy in VMIM module generates a pair of visible maps without overlapping patches to create diverse preservation clues for reconstruction tasks. (b) Illustration of the identity-related region reconstruction in terms of $\phi_1$ . For the feature output of the decoder, RGB pixel and multi-view semantic reconstruction with MSE loss is performed only for the vehicle body region. Notably, another peer $\phi_2$ also performs a series of reconstruction processes similar to $\phi_1$ . + +![](images/7933f87441b368d4be3bcd76da04e48ee2a1dd822425abeae3bf8d8b3d280fc6.jpg) + +point-by-point: + +(i) Asymmetric mask sampling strategy. Initially, ours intention is to fully utilize the potential of self-supervised pre-training to serve downstream tasks with view variation by employing two homogeneous architectures. However, two twin MAEs will inevitably fail to expand the scope of their knowledge if the same masking and reconstruction operations are employed to facilitate equal knowledge exchange. Motivated by these analyses, an asymmetric mask sampling strategy is designed to generate paired visible maps without overlapping patches and encourage two MAEs to learn diverse knowledge, as shown in Figure 3(a). Concretely, the image $I_{img}$ and the corresponding multi-view mask $I_{seg}$ are simultaneously injected into the encoders of $E_{\phi_1}$ and $E_{\phi_2}$ , and are respectively divided into two sequences with $N$ patches $\{P(I_{img}|\phi_1); P(I_{seg}|\phi_1)\}$ and $\{P(I_{img}|\phi_2); P(I_{seg}|\phi_2)\}$ in spatial order. Subsequently, based on the semantic segmentation annotation of $I_{seg}$ , each sequence is further subdivided into two types: (1) subsequences $P(I^{id}|\phi_1) = \{P(I_{img}^{id}|\phi_1); P(I_{seg}^{id}|\phi_1)\}$ and $P(I^{id}|\phi_2) = \{P(I_{img}^{id}|\phi_2); P(I_{seg}^{id}|\phi_2)\}$ of the vehicle body region with size $N_{id}$ ; (2) subsequences $P(I^{bg}|\phi_1) = \{P(I_{img}^{bg}|\phi_1); P(I_{seg}^{bg}|\phi_1)\}$ and $P(I^{bg}|\phi_2) = \{P(I_{img}^{bg}|\phi_2); P(I_{seg}^{bg}|\phi_2)\}$ of the background region with size $N_{bg}$ , thereby also fulfilling $N_{id} + N_{bg} = N$ . The pair of visible patch maps $\{M_{\phi_1}, M_{\phi_2}\}$ obtained through the asymmetric mask sampling strategy follows two-point principles. The first point is to randomly mask the patches of the background region $P(I^{bg}|\phi_1)$ and $P(I^{bg}|\phi_2)$ as much as possible and preserve the vehicle body region $P(I^{id}|\phi_1)$ and $P(I^{id}|\phi_2)$ as visible clues for the reconstruction task. The second point is that no overlap simultaneously exists in the masking patches of two MAEs, that is, $M_{\phi_1} \cap M_{\phi_2} = \emptyset$ . Notably, $\{M_{\phi_1}, M_{\phi_2}\}$ will be will be complemented each other in RGB pixels and multi-view semantic reconstruction to enhance collaboration between the two types of tasks. + +(ii) Identity-related region reconstruction. Directly using the entire original image and a multi-view segmentation mask as the prediction target of reconstruction inevitably increases the computational cost and reduces the pre-training efficiency. Considering the aforementioned analysis, an identity-related region reconstruction method is designed. First, similar to the tokenizing and encoding operations of ordinary MAEs, the visible patch maps $\{M_{\phi_1}, M_{\phi_2}\}$ obtained by asymmetric mask sampling are tokenized and embedded into a sequence of tokens, and the masked patches will be replaced with a sequence of learnable masked tokens. Afterward, the visible and invisible tokens are combined and then inputted into the decoders $\{D_{\phi_1}^{img}, D_{\phi_1}^{seg}\}$ and $\{D_{\phi_2}^{img}, D_{\phi_2}^{seg}\}$ of the two MAEs. Ultimately, in the decoding stage, the ground-truth of two task-specific reconstructions is set with identity-related regions $P(I^{id}|\phi_1)$ and $P(I^{id}|\phi_2)$ , that is, the RGB pixels and multi-view semantic regions of the vehicle body, instead of the initial $I_{img}$ and $I_{seg}$ . Taking $\phi_1$ as an example, Figure 3(b) illustrates the reconstruction process of the proposed method. Mean square error (MSE) is adopted to calculate the error of RGB pixels and multi-view semantics in the reconstructed body region. The loss function $L_{rec}$ of the entire reconstruction task can be defined by Eq. 3: + +$$ +L _ {r e c} = \sum_ {m \in \{1, 2 \}} \sum_ {\triangle \in \{i m g, s e g \}} M S E \left(R \left(I _ {\triangle} ^ {i d} \mid \phi_ {m}\right), P \left(I _ {\triangle} ^ {i d} \mid \phi_ {m}\right)\right) \tag {3} +$$ + +where the term $\phi_{m}$ and $m\in \{1,2\}$ represent two models $\phi_{1}$ and $\phi_{2}$ , respectively. The term $\triangle \in \{img,seg\}$ denotes the reconstruction task of RGB pixels and multi-view semantics. $R(I_{\triangle}^{id}|\phi_m)$ and $P(I_{\triangle}^{id}|\phi_m)$ are the predicted results and ground-truth of the model $\phi_{m}$ performing two reconstruction tasks, respectively. Correspondingly, function $MSE(\cdot)$ is the MSE between the predicted target and ground-truth. + +# 4.2. Past-to-present Mutual-distillation. + +While the VMIM module can diversify the knowledge learned by $\phi_{1}$ and $\phi_{2}$ , empirical investigation revealed that this module cannot guarantee collaborative training between the two models and deep mining of their self-knowledge. Toward this end, this study aimed to establish the teacher-student role relationship, uniting the past and present knowledge of two models, and design a past-to-present mutual distillation method. More precisely, the core of this design is to conduct feature-level knowledge distillation through the collaboration of current models and the cross-guidance of historical models. + +Assuming the current pre-training is at epoch $N$ , $\phi_1$ and $\phi_2$ play the roles of two student models. They collaboratively learn and exchange the predicted features with each other based on asymmetric visible clues to enhance the multi-view generalization capability of the entire VehicleMAE. The predicted feature maps $\{f_{\phi_1}^{img}, f_{\phi_1}^{seg}\}$ and $\{f_{\phi_2}^{img}, f_{\phi_2}^{seg}\}$ of the decoder are extracted for $\phi_1$ and $\phi_2$ in two specific reconstruction tasks, respectively. Kullback-Leibler (KL) divergence is then adopted to minimize the predicted feature distribution error. The distillation losses $L_{pr}^{\phi_1}$ and $L_{pr}^{\phi_2}$ of the two MAEs in the current epoch $N$ are represented by Eq. 4: + +$$ +L _ {p r} ^ {\phi_ {1}} = K L \left(f _ {\phi_ {2}} ^ {i m g} \mid \mid f _ {\phi_ {1}} ^ {i m g}\right) + K L \left(f _ {\phi_ {2}} ^ {s e g} \mid \mid f _ {\phi_ {1}} ^ {s e g}\right) \tag {4} +$$ + +$$ +L _ {p r} ^ {\phi_ {2}} = K L \left(f _ {\phi_ {1}} ^ {i m g} | | f _ {\phi_ {2}} ^ {i m g}\right) + K L \left(f _ {\phi_ {1}} ^ {s e g} | | f _ {\phi_ {2}} ^ {s e g}\right) +$$ + +where $KL(\cdot)$ is the KL divergence function that measures the predicted features of two models. + +When $N\geq 2$ , the two historical models $\phi_1^{\prime}$ and $\phi_2^\prime$ corresponding to epoch $N - 1$ in the past are regarded as teachers cross-guiding student models $\phi_{1}$ and $\phi_{2}$ of epoch $N$ to dynamically mine their own knowledge. Specifically, $\phi_1^{\prime}$ is set as the historical teacher for $\phi_{2}$ and $\phi_2^\prime$ as the historical teacher for $\phi_{1}$ to establish the teacher-student relationships from the past to the present. This mutual exchange of teacher-student relationships can further enhance the potential for mutual distillation and knowledge diversity. The encoded feature maps $\{F_{\phi_1'}^{img},F_{\phi_1'}^{seg},F_{\phi_2'}^{img},F_{\phi_2'}^{seg}\}$ and $\{F_{\phi_1}^{img},F_{\phi_1}^{seg},F_{\phi_2}^{img},F_{\phi_2}^{seg}\}$ output by the encoders of two historical teacher models and two current student models are employed to distill intermediate features due to the inability of self-supervised pre-training to achieve hard-label distillation. The distillation loss $L_{pa}^{\phi_1}$ guided by $\phi_2^\prime$ for $\phi_{1}$ and the distillation loss $L_{pa}^{\phi_2}$ guided by $\phi_1^\prime$ for $\phi_{2}$ are defined as Eq.5: + +$$ +L _ {p a} ^ {\phi_ {1}} = K L \left(F _ {\phi_ {2} ^ {\prime}} ^ {i m g} \mid \mid F _ {\phi_ {1}} ^ {i m g}\right) + K L \left(F _ {\phi_ {2} ^ {\prime}} ^ {s e g} \mid \mid F _ {\phi_ {1}} ^ {s e g}\right) \tag {5} +$$ + +$$ +L _ {p a} ^ {\phi_ {2}} = K L (F _ {\phi_ {1} ^ {\prime}} ^ {i m g} | | F _ {\phi_ {2}} ^ {i m g}) + K L (F _ {\phi_ {1} ^ {\prime}} ^ {s e g} | | F _ {\phi_ {2}} ^ {s e g}) +$$ + +After the pre-training of VehicleMAE is completed, all decoders will be discarded and the learned knowledge will + +be fused by solving the average parameters of the two encoders $E_{\phi_1}$ and $E_{\phi_2}$ . The process of obtaining the mean model $E[\bar{\theta}]$ can be written as Eq. 6: + +$$ +E [ \bar {\theta} ] = \frac {E _ {\phi_ {1}} [ \theta_ {\phi_ {1}} ] + E _ {\phi_ {2}} [ \theta_ {\phi_ {2}} ]}{2} \tag {6} +$$ + +where $\theta_{\phi_1}$ and $\theta_{\phi_2}$ are the parameter of $E_{\phi_1}$ and $E_{\phi_2}$ , respectively, and $\bar{\theta}$ is the parameter of the mean model $E[\bar{\theta}]$ . It is worth emphasizing that $E[\bar{\theta}]$ will be directly applied as a single pre-training model to downstream tasks of supervised and unsupervised vehicle Re-ID. + +# 4.3. Training Objective + +The training objectives of VehicleMAE include two parts: the loss of the reconstruction task of two homogeneous MAEs in the VMIM module; in the PPMD module, the distillation loss $L_{pr} = L_{pr}^{\phi_1} + L_{pr}^{\phi_2}$ between the current two student models and the distillation loss $L_{pa} = L_{pa}^{\phi_1} + L_{pa}^{\phi_2}$ guided by the history teachers. Therefore, the total loss $L_{total}$ of the overall pre-training process can be defined by Eq. 7 as: + +$$ +L _ {t o t a l} = L _ {r e c} + L _ {p r} + L _ {p a} \tag {7} +$$ + +Note that, the pre-training mean model $E[\bar{\theta}]$ directly serves as supervised and unsupervised vehicle Re-ID downstream tasks and is supervised by cross-entropy and triplet losses. + +# 5. Experiments + +# 5.1. Datasets and Evaluation Protocols + +This paper utilizes two well-known vehicle Re-ID datasets: VeRi-776 [23] and VehicleID [3]. VehicleID dataset includes small, medium, and large-scale test sets. To evaluate the performance of the proposed method, mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank- $k$ ( $k = 1, 5$ ) are employed as the evaluation metrics. + +# 5.2. Implementation Details + +CLIP-B/16 serves as the backbone network. Throughout the whole training process, the image size is set to $256 \times 256$ , and models update parameters using the Adam optimizer. During the pre-training phase, all samples are divided into mini-batches of size 128 for 300 epochs of iterative learning. For downstream tasks, the batch size is set to 64, and all samples are trained for 50 epochs. Additionally, each image undergoes data augmentation with random horizontal flipping, padding, cropping, and erasing. In unsupervised tasks, DBSCAN is applied as the clustering algorithm to generate pseudo-labels. All experiments are conducted under the Ubuntu 22.04 operating system and Pytorch environment with 4 NVIDIA RTX A6000 GPUs. + +# 5.3. Ablation Study + +Effects of different modules. The results in Table 2 demonstrate the effectiveness of all proposed modules. Three observations are summarized by us: (1) When we independently introduce the "Pa" or "Pr" into the "Baseline" to perform knowledge distillation, the significant performance improvement in Rank accuracy and mAP confirms that historical and current distillation can fully explore and exchange their own knowledge. (2) Whether using "Asym" in the mask sampling strategy during the encoding stage or introducing "Id-rec" in the reconstruction task during the decoding stage, it is superior to MAE that integrates a random mask strategy and complete pixel reconstruction. (3) VehicleMAE belongs to the ensemble of all modules and achieves the best performance, further verifying that VehicleMAE can efficiently pre-train two homogeneous MAEs in a collaborative manner. + +
Ablation ModulesVeRi-776VehicleID
BaselinePaPrAsymId-recmAPRank-1Rank-1Rank-5
✓37.681.269.081.8
✓✓40.185.071.284.6
✓✓40.585.371.184.7
✓✓38.783.670.382.8
✓✓39.584.370.182.6
✓✓✓✓✓42.587.673.385.6
+ +Effects of different mask sampling strategies and ratios. Table 3 reports the impact of using asymmetric mask and random mask sampling strategies at different mask ratios. We observe that the performance of the proposed "Asym" is superior to "Random" for all three mask ratios. Especially at a mask rate of $75\%$ , "Asym" achieves the highest mAP of $42.5\%$ and Rank-1 accuracy of $87.6\%$ . This is due to the visible maps generated by "Asym" having no overlapping patches, which encourage two MAEs to learn and exchange knowledge. + +Effects of different distillation strategies. Table 4 investigates the impact of modifying VehicleMAE using various distillation architectures on Re-ID performance. Ours findings are two-fold: (1) Compared to two MAEs learning independently, using historical distillation or present distillation can ameliorate the pre-training model performance, especially when combined to achieve more effective knowledge exchange. (2) The mean encoder $E[\bar{\theta}]$ can accurately feedback and maintain the average level of two encoders, while also solving the difficulties of multiple model selec + +Table 2. Performance comparison of different components in unsupervised vehicle ReID. "Past" and "Pres" refer to historical distillation and present distillation in PPDM, respectively. "Asym" and "Id-rec" correspond to the Asymmetric mask sampling strategy and Identity-related region reconstruction in VMIM. + +
Mask RatiosStrategiesVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
25%Random38.884.268.483.6
Asym40.685.570.484.2
50%Random39.284.668.983.8
Asym41.786.271.184.8
75%Random40.185.169.584.1
Asym42.587.673.385.6
+ +tion and deployment for downstream tasks. + +Table 3. Effectiveness of different mask sampling strategies in serving unsupervised vehicle Re-ID tasks. "Random" and "Asym" express the random mask sampling strategy and the asymmetric mask sampling strategy, respectively. + +
Pre-trained EncodersVeRi-776VehicleID
IndPaPrPa+PrIndPaPrPa+Pr
Eφ184.384.784.887.170.370.871.073.2
Eφ284.484.785.186.970.370.770.973.1
E[θ]84.685.085.387.670.971.271.173.3
+ +Table 4. Comparison Rank-1 accuracy of two homogeneous encoders $\{E_{\phi_1}, E_{\phi_2}\}$ and a mean encoder $E[\tilde{\theta}]$ using various distillation architectures in unsupervised downstream tasks. "Ind" denotes that $E_{\phi_1}$ and $E_{\phi_2}$ learning independently without using distillation. "Pa", "Pr", and "Pa+Pr" respectively signify the independent use of historical distillation, present distillation, and the simultaneous use of both them. + +Analysis of Model Complexity. To further analyze the computational performance, we re-implement different self-supervised pre-training works on DiffVERI for model parameter sizes and floating-point operations (FLOPs), as shown in Table 5. For fairness, all methods except ConvNeXt v2, which relies on the ConvNeXt network architecture, use the CLIP-B/16 image encoder as the backbone. Although VehicleMAE adopts a mutual-distillation architecture, our pre-training model contains 57.3M parameters and only needs 14.7G multiply-accumulate operations for feature extraction. This is mainly due to the mean model design and that we follow the vanilla MAE without directly changing the encoder-decoder architecture. In contrast, both the fully convolutional mask autoencoder in ConvNeXt v2 and the hybrid convolution module in CMAE will increase the resource consumption of feature extraction. These results demonstrate that VehicleMAE achieves a favorable trade-off between model complexity and the performance of downstream tasks. + +# 5.4. Comparison with State-of-the-arts + +As illustrated in Table 6, we compare the proposed method with state-of-the-art ones in both supervised and unsuper + +
MethodsFLOPsParams
ConvNeXt v2[45]15.4G89M
CMAE[18]19.4G104M
BEiT[43]15.7G82.5M
Ours14.7G57.3M
+ +Table 5. Comparison of different pre-training models in terms of parameter size and FLOPs. + +
MethodsVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
Supervised Learning
\( A\text{VER}{*}^{R}\left\lbrack {19}\right\rbrack \)61.289.063.585.6
\( VehicleNet*R\left\lbrack {54}\right\rbrack \)83.496.879.592.0
\( TransReID^{*V}\left\lbrack {14}\right\rbrack \)82.097.1--
\( GiT^{*V}\left\lbrack {35}\right\rbrack \)80.396.977.9-
\( LCNL^{*V}\left\lbrack {46}\right\rbrack \)81.897.4--
\( TANet^{*R}\left\lbrack {17}\right\rbrack \)83.696.878.292.6
\( CFA-Net^{*R}\left\lbrack {41}\right\rbrack \)80.796.9--
\( HCI-Net^{*R}\left\lbrack {36}\right\rbrack \)83.896.676.491.2
\( CLIP-ReID^{*C}\left\lbrack {24}\right\rbrack \)83.397.478.192.7
\( Ours^{\dagger C} \)87.697.485.994.9
Unsupervised Learning
\( MMT^{*R}\left\lbrack {7}\right\rbrack \)25.460.931.042.4
\( SPCL^{*R}\left\lbrack {8}\right\rbrack \)36.979.953.066.4
\( RLCC^{*R}\left\lbrack {51}\right\rbrack \)39.683.4--
\( PPLR^{*R}\left\lbrack {4}\right\rbrack \)41.685.6--
\( VAPC TO^{*R}\left\lbrack {52}\right\rbrack \)30.476.2--
\( ICL^{*R}\left\lbrack {37}\right\rbrack \)39.583.7--
\( AdaMG^{*R}\left\lbrack {31}\right\rbrack \)41.086.2--
\( NNNI^{*R}\left\lbrack {9}\right\rbrack \)42.386.3--
\( STDA^{*R}\left\lbrack {13}\right\rbrack \)42.387.4--
\( MAE†^C\left\lbrack {12}\right\rbrack \)37.681.269.081.8
\( CL-MAE†^C\left\lbrack {29}\right\rbrack \)41.084.070.882.6
\( CMAE†^C\left\lbrack {18}\right\rbrack \)41.684.771.184.9
\( Ours^{\dagger C} \)42.587.673.385.6
+ +Table 6. Performance comparisons with the latest methods on supervised and unsupervised vehicle Re-ID tasks. “ $*^{R}$ , “ $*^{V}$ , and “ $*^{C}$ ” indicate using ResNet, ViT, and the image encoder of CLIP as backbones, respectively. “ $\dagger^{C}$ ” denotes the implementation of pre-training using the image encoder of CLIP on DiffVERI. Note that experimental results are accurate to one decimal place, and this paper uses only the large-scale test set of VehicleID. + +vised downstream tasks. Compared to the most competitive methods among ResNet-based, ViT-based, and CLIP-based backbones, our method delivers the best mAP and Rank accuracy for both supervised and unsupervised settings. For fair comparison, we further implement several well-established MAE-based pre-training methods on DiffVERI. The noticeable performance improvement implies + +![](images/e90d2afdae2e949377b78ae246f199bf6ee3e205abffb5f1d19dc8820e21434d.jpg) +Figure 4. Visualization of patch-level feature distributions extracted by Baseline and using VehicleMAE pre-training. The class with the highest proportion of pixels within each patch in the patch-level ground-truth is defined as the final class corresponding to the patch. Different color patches represent different views or background regions + +that the masked image modeling based on VehicleMAE is better suited for downstream tasks of vehicle Re-ID. + +# 5.5. Qualitative Analysis + +Figure 4 displays two visualization examples of the feature distribution at the patch-level extracted by Baseline and VehicleMAE. Specifically, we feed the features of all patches into the K-Means algorithm to perform five class-based clustering operations (i.e., a background region and four view regions) to intuitively demonstrate the view discriminate ability of different pre-training models. It can be observed that our VehicleMAE is able to distinguish patches from different views and backgrounds in both examples, getting close to patch-level ground-truth. Remarkably, even in stricter occlusion scenarios, VehicleMAE accurately avoid interference from the power poles in example B, further completing the extraction of multi-view information. More details on qualitative visualization experiments are included in the supplementary materials. + +# 6. Conclusion + +This paper releases DiffVERI, a large-scale multi-view vehicle Re-ID dataset for learning view-invariant representations, and proposes a masked image modeling pre-training paradigm termed VehicleMAE specially for vehicle Re-ID downstream tasks. VehicleMAE first proposes a VMIM module that attempts to apply two homogeneous MAEs to predict the RGB pixels and multi-view semantic clues of vehicles in pairs, thereby gaining diverse multi-view inference capabilities. Subsequently, to facilitate learning collaboratively, a PPMD module is designed to progressively exchange knowledge with each other. Extensive experiments demonstrate that equipping our pre-training model can achieve competitive performance in generic vehicle Re-ID downstream tasks. Future work contributes to further expanding VehicleMAE into a unified multimodal pre-training paradigm. + +# Acknowledgements + +# References + +[1] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In European Conference on Computer Vision, pages 348-367. Springer, 2022. 3 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[3] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 347–355, 2022. 1, 6 +[4] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7308–7318, 2022. 2, 8 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[6] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750–14759, 2021. 2, 3 +[7] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 8 +[8] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. Advances in neural information processing systems, 33:11309-11321, 2020. 8 +[9] Dongchen Han, Weifeng Liu, Mingchen Zou, and Baodi Liu. Non-contrastive nearest neighbor identity-guided method for unsupervised object re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2713–2723, 2022. 8 +[10] Bing He, Jia Li, Yifan Zhao, and Yonghong Tian. Part-regularized near-duplicate vehicle re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3997–4005, 2019. 1 + +This work was supported by the National Natural Science Foundation of China under Grant No. 62461037 and the Jiangxi Provincial Natural Science Foundation under Grant Nos. 20224BAB212011, 20232BAB202051, 20232BAB212008, and 20242BAB25078. +[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 8 +[13] Qiaolin He, Zihan Wang, Zhijie Zheng, and Haifeng Hu. Spatial and temporal dual-attention for unsupervised person re-identification. IEEE Transactions on Intelligent Transportation Systems, 2023. 8 +[14] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013–15022, 2021. 8 +[15] Zhijun He, Hongbo Zhao, Jianrong Wang, and Wenquan Feng. Multi-level progressive learning for unsupervised vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(4):4357-4371, 2022. 2 +[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4 +[17] Wenbo Hu, Hongjian Zhan, Palaiahnakote Shivakumara, Umapada Pal, and Yue Lu. Tanet: Text region attention learning for vehicle re-identification. Engineering Applications of Artificial Intelligence, 133:108448, 2024. 8 +[18] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 8 +[19] Pirazh Khorramshahi, Amit Kumar, Neehar Peri, Sai Saketh Rambhatla, Jun-Cheng Chen, and Rama Chellappa. A dual-path model with adaptive attention for vehicle re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6132–6141, 2019. 8 +[20] Pirazh Khorramshahi, Vineet Shenoy, and Rama Chellappa. Robust and scalable vehicle re-identification via self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5295–5304, 2023. 2 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4 +[22] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE transactions on image processing, 31:5949-5962, 2022. 1 +[23] Hongchao Li, Aihua Zheng, Liping Sun, and Yonglong Luo. Camera topology graph guided vehicle re-identification. + +IEEE Transactions on Multimedia, 26:1565-1577, 2023. 1, 6 +[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1405–1413, 2023. 8 +[25] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2167-2175, 2016. 1, 3 +[26] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In 2016 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2016. 1, 3 +[27] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. Veri-wild: A large dataset and a new method for vehicle re-identification in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3235-3243, 2019. 1, 3 +[28] Zefeng Lu, Ronghao Lin, Qiaolin He, and Haifeng Hu. Mask-aware pseudo label denoising for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 24(4):4333-4347, 2023. 2 +[29] Neelu Madan, Nicolae-Catalin Ristea, Kamal Nasrollahi, Thomas B Moeslund, and Radu Tudor Ionescu. Cl-mae: Curriculum-learned masked autoencoders. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2492–2502, 2024. 3, 8 +[30] Dechao Meng, Liang Li, Xuejing Liu, Yadong Li, Shijie Yang, Zheng-Jun Zha, Xingyu Gao, Shuhui Wang, and Qingming Huang. Parsing-based view-aware embedding network for vehicle re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7103–7112, 2020. 1 +[31] Jinjia Peng, Guangqi Jiang, and Huibing Wang. Adaptive memorization with group labels for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5802-5813, 2023. 8 +[32] Jiuchao Qian, Minting Pan, Wei Tong, Rob Law, and Edmond Q Wu. Urrnet: A unified relational reasoning network for vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(9):11156-11168, 2023. 2 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3 +[34] Zhiyin Shao, Xinyu Zhang, Changxing Ding, Jian Wang, and Jingdong Wang. Unified pre-training with pseudo texts for text-to-image person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11174–11184, 2023. 3 +[35] Fei Shen, Yi Xie, Jianqing Zhu, Xiaobin Zhu, and Huanqiang Zeng. Git: Graph interactive transformer for vehicle re-identification. IEEE Transactions on Image Processing, 32:1039–1051, 2023. 2, 8 + +[36] Ke Sun, Xiyu Pang, Meifeng Zheng, Xiushan Nie, Xi Li, Houren Zhou, and Yilong Yin. Heterogeneous context interaction network for vehicle re-identification. Neural Networks, 169:293–306, 2024. 8 +[37] Qing Tang and Kang-Hyun Jo. Unsupervised object re-identification via instances correlation loss. In 2022 IEEE 20th International Conference on Industrial Informatics (IN-DIN), pages 135–139. IEEE, 2022. 8 +[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078-10093, 2022. 3 +[39] Jingzheng Tu, Cailian Chen, Xiaolin Huang, Jianping He, and Xinping Guan. Dfr-st: Discriminative feature representation with spatio-temporal cues for vehicle re-identification. Pattern Recognition, 131:108887, 2022. 1 +[40] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464-7475, 2023. 3 +[41] Hai Wang, Yaqing Niu, Long Chen, Yicheng Li, Miguel Angel Sotelo, Zhixiong Li, and Yingfeng Cai. Dair-v2xreid: A new real-world vehicle-infrastructure cooperative re-id dataset and cross-shot feature aggregation network perception method. IEEE Transactions on Intelligent Transportation Systems, 25(8):9058-9068, 2024. 8 +[42] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 460–469, 2019. 3 +[43] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 8 +[44] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In Proceedings of the IEEE international conference on computer vision, pages 379–387, 2017. 1 +[45] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16133-16142, 2023. 8 +[46] Mouxing Yang, Zhenyu Huang, and Xi Peng. Robust object re-identification with coupled noisy labels. International Journal of Computer Vision, pages 1–19, 2024. 8 +[47] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search + +benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492-4501, 2023. 3 +[48] Yue Yao, Liang Zheng, Xiaodong Yang, Milind Naphade, and Tom Gedeon. Simulating content consistent vehicle datasets with attribute descent. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 775-791. Springer, 2020. 3 +[49] Hongwei Zhang, Guoqing Zhang, Yuhao Chen, and Yuhui Zheng. Global relation-aware contrast learning for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8599–8610, 2022. 2 +[50] Kevin Zhang and Zhiqiang Shen. i-mae: Are latent representations in masked autoencoders linearly separable? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7740-7749, 2024. 3 +[51] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3436–3445, 2021. 8 +[52] Aihua Zheng, Xia Sun, Chenglong Li, and Jin Tang. Aware progressive clustering for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(8):11422-11435, 2021. 8 +[53] Aihua Zheng, Chaobin Zhang, Chenglong Li, Jin Tang, and Chang Tan. Multi-query vehicle re-identification: Viewpoint-conditioned network, unified dataset and new metric. IEEE Transactions on Image Processing, 32:5948–5960, 2023. 1 +[54] Zhedong Zheng, Tao Ruan, Yunchao Wei, Yi Yang, and Tao Mei. Vehiclenet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia, 23:2683–2693, 2020. 8 \ No newline at end of file diff --git a/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/images.zip b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8d875ad1b8a01eddc04f15724cbf9c14675b85d6 --- /dev/null +++ b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7eb9ccb721a5ff2580233f95e9ba7f75d8852e4e026aa64a89235df3ed16da5 +size 504049 diff --git a/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/layout.json b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba2760df155285a99d4d2fa812ebea1c2dc9112a --- /dev/null +++ b/2025/VehicleMAE_ View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders/layout.json @@ -0,0 +1,9311 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "text", + "content": "VehicleMAE: View-asymmetry Mutual Learning for Vehicle Re-identification Pre-training via Masked AutoEncoders" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "content": "Qi Wang" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "content": ", Zeyu Zhang" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "content": ", Dong Wang" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "content": ", Di Gai" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "content": ", Xin Xiong" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "content": ", Jiyang Xu" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "text", + "content": " and Ruihua Zhou" + }, + { + "bbox": [ + 88, + 160, + 534, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 175, + 526, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 175, + 526, + 189 + ], + "spans": [ + { + "bbox": [ + 96, + 175, + 526, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 175, + 526, + 189 + ], + "type": "text", + "content": "School of Mathematics and Computer Sciences, Nanchang University, Nanchang, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 189, + 459, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 189, + 459, + 203 + ], + "spans": [ + { + "bbox": [ + 164, + 189, + 459, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 189, + 459, + 203 + ], + "type": "text", + "content": "School of Software, Nanchang University, Nanchang, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 203, + 541, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 203, + 541, + 217 + ], + "spans": [ + { + "bbox": [ + 80, + 203, + 541, + 217 + ], + "type": "text", + "content": "3The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 220, + 557, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 557, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 557, + 232 + ], + "type": "text", + "content": "{wangqi,gaidi,xiongxinxx}@ncu.edu.cn，{zeyuzhang1010,dongwang,xujiyang,zrh}@email.ncu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 284, + 297, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 632 + ], + "type": "text", + "content": "Large-scale pre-training technology has achieved remarkable performance in diversified object re-identification (Re-ID) downstream tasks. Nevertheless, to our best knowledge, the pre-training model specifically for vehicle Re-ID, which focuses on tackling the challenge of multi-view variations, has not been fully investigated. In this paper, we first leverage a diffusion model to build a large-scale vehicle Re-ID benchmark dataset, dubbed \"DiffVERI\", containing over 1700K images from abundant multi-view annotations. In terms of this dataset, we further present VehicleMAE, a novel masked image modeling pre-training paradigm that learns view-invariant representations by performing mutual-distillation in a self-supervised manner. To be specific, the pipeline of VehicleMAE unfolds two core modules, i.e., view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). Technically, VMIM consists of two homogeneous masked autoencoders (MAE) that simultaneously reconstruct the RGB pixels and multi-view semantic information of the specific vehicle body region via paired asymmetric mask sampling strategies. To progressively distill the knowledge of the model itself, PPMD considers the two MAEs in the current epoch and the previous one as the student models and the teacher models, respectively, which leverages the knowledge learned by the current student and the historical teacher for mutual feature-level distillation. Extensive experimental results have verified that the proposed pre-training paradigm on DiffVERI gains compelling downstream task performance for vehicle Re-ID." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 652, + 137, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 652, + 137, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 652, + 137, + 665 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "The purpose of vehicle re-identification (Re-ID)[3, 23, 53] is to match target vehicles with the same identity under vari" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "text", + "content": "ous views. Despite impressive performance in some current works, this task still faces a main obstacle of large intraclass discrepancy among vehicles of the same identity. The primary cause is that the unique geometrically rigid structure of the vehicle itself leads to dynamic appearances of the same instance from different views. Accordingly, it is imperative to enable the Re-ID model to efficiently learn view-invariant representations." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 338, + 366, + 529, + 574 + ], + "blocks": [ + { + "bbox": [ + 338, + 366, + 529, + 574 + ], + "lines": [ + { + "bbox": [ + 338, + 366, + 529, + 574 + ], + "spans": [ + { + "bbox": [ + 338, + 366, + 529, + 574 + ], + "type": "image", + "image_path": "94418750685e995e2bcd6f2d0c957feeaa5956cda2d23a96c39faf9db2da41e4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 582, + 555, + 616 + ], + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 616 + ], + "type": "text", + "content": "Figure 1. Illustration of DiffVERI dataset: (a) Steps for synthesizing data and annotations. (b) Synthetic examples from DiffVERI dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "With the release of the previous generic vehicle Re-ID datasets (such as VeRi-776 [25], VehicleID [26], VERiWild [27], etc.), early efforts [10, 30, 44] preferred the use of vehicle keypoints and part-based bounding boxes for discriminative representation learning to achieve view alignment. Additionally, some follow-up works [22, 39] based on metric learning resort to vehicle attribute annotations and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4701" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": "spatio-temporal information to narrow the appearance discrepancy across different views. Despite advanced progress on the aforesaid works, their training data are typically derived from real-world surveillance videos and only contain images within a fixed view range, which limits the inference ability of the Re-ID model for different views. An intuitive solution is to build a large-scale benchmark with a comprehensive view range to provide learnable view-invariant information for the Re-ID model while trading off labor costs. Until recently, there have been no existing works specifically investigating large-scale vehicle Re-ID benchmarks for a comprehensive range of views. To address the aforementioned limitations, we release a large-scale multiview dataset, named DiffVERI, as presented in Figure 1. In brief, we use the text-to-image diffusion model to synthesize vehicle data from various views. Figure 1(a) illustrates the construction process of DiffVERI, which includes three steps: multi-view data generation, data filtering, and multi-view semantic annotations. An example of generating multi-view data is also shown in Figure 1(b), and more the detailed visualizations of synthesized examples is included in the supplementary materials." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 338, + 294, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 338, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 294, + 540 + ], + "type": "text", + "content": "Due to the large domain gap between the generic ImageNet and vehicle data, ImageNet cannot be well-suited for various downstream tasks of vehicle Re-ID. Therefore, it is particularly crucial to obtain initial weights with strong generalization ability through task-specific pre-training. Some pioneering works based on self-supervised pre-training have been explored in the person-related field. For instance, Fu et al. [6] constructed a LUPerson benchmark and developed a pre-training model via contrastive learning. Another effective self-supervised approach is masked image modeling technology, which performs masking and pixel-level reconstruction on the backbone of vision transformer (ViT) to serve various downstream vision tasks. In contrast, to our best knowledge, there is currently no prior research on targeted large-scale pre-training for vehicle Re-ID. That is to say, mainstream vehicle Re-ID methods are still limited by the pre-training parameters from ImageNet." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 543, + 294, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 543, + 294, + 626 + ], + "spans": [ + { + "bbox": [ + 57, + 543, + 294, + 626 + ], + "type": "text", + "content": "To address the aforementioned issues, we have released a vehicle Re-ID benchmark, DiffVERI, with a comprehensive view range. Benefiting from DiffVERI, a novel masked image modeling pre-training paradigm called VehicleMAE is then proposed to serve the downstream task of vehicle Re-ID. To sum up, the main contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 629, + 294, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 629, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 629, + 294, + 700 + ], + "type": "text", + "content": "1. We build a large-scale vehicle Re-ID benchmark, DiffVERI, by fine-tuning an off-the-shelf diffusion model, which incorporates over 1700K images and multi-view semantic annotations. To our knowledge, DiffVERI is by far the largest vehicle Re-ID dataset with diverse view data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "content": "2. We begin the first attempt to investigate a large-" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 328, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 553, + 143 + ], + "type": "text", + "content": "scale pre-training solution for vehicle Re-ID tasks and therefore devise a novel masked image modeling pre-training paradigm, termed VehicleMAE, to jointly integrate the view-asymmetry masked image modeling (VMIM) module and the past-to-present mutual-distillation (PPMD) one." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 144, + 553, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 144, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 316, + 144, + 553, + 190 + ], + "type": "text", + "content": "3. Comprehensive experiments have validated the extraordinary performance of our pre-training model in both supervised and unsupervised downstream tasks of vehicle Re-ID." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 205, + 400, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 400, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 400, + 217 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 225, + 463, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 225, + 463, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 225, + 463, + 237 + ], + "type": "text", + "content": "2.1. Large-scale Re-ID Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 243, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 578 + ], + "type": "text", + "content": "The recent emergence of vehicle benchmark datasets has led to increasing attention to vehicle Re-ID technology. Existing efforts mainly follow two mainstream settings. The first setting, supervised learning [20, 32, 35], considers vehicle Re-ID as a representation learning problem exploit ID annotations to learn informative features to distinguish different vehicle instances. Early works [32, 35] focused on exploring potential correlations between global and local features to enhance the discriminative performance of Re-ID models. To distinguish vehicles with similar appearances from the same manufacturing, subsequent works attempted to feed additional attribute annotations, such as color, model, and camera ID, as auxiliary information to obtain robust identity-related features. Methods in the latter setting, i.e., unsupervised learning [4, 15, 28, 49], aim to fully learn discriminative representations of the training set without relying on any annotations. The vanilla unsupervised learning procedure updates the Re-ID model using pseudo labels generated by clustering algorithms as supervised information. For this purpose, previous works have tentatively proposed pseudo label refinement [4, 28] and contrastive learning [15, 49] to overcome the limitations of clustering-based label noise. In both settings, the aforementioned methods inevitably face the challenge of view variations due to the lack of learning view annotations. To tackle this, we first construct a large-scale vehicle dataset DiffVERI comprising multi-view annotations and then attempt to learn view-invariant features via pre-training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 588, + 468, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 468, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 468, + 600 + ], + "type": "text", + "content": "2.2. Self-supervised Pre-training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 605, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 605, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 605, + 553, + 712 + ], + "type": "text", + "content": "Recently, the focus of vision recognition tasks has shifted from achieving advanced performance in specific downstream tasks to designing generic large-scale pre-training techniques. Labeling large-scale datasets is undoubtedly labor-intensive, while utilizing the inherent information of the data itself for self-supervised learning is a more effective way. Pioneer self-supervised pre-training works [2, 11] attempted to learn robust representation via dynamic contrastive learning and knowledge distillation on unlabeled" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "text", + "content": "4702" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "type": "text", + "content": "large-scale datasets. As an illustration, in the Re-ID task, Fu et al. [6] constructed a large-scale person dataset called LU-Person and first designed a pre-training Re-ID model based on contrastive learning. With the advanced performance of vision transformer [5] in vision tasks, another more lightweight method, masked image modeling [18, 29, 50], has been systematically proven to be effective in the field of self-supervised pre-training. The representative work [12], i.e., masked autoencoders (MAE), presented an encoder-decoder architecture to perform pixel-level masking and reconstruction operations. Subsequently, some recent works, such as VideoMAE [38] and MultiMAE [1], have extended MAE to the fields of video and multiple vision tasks, respectively. However, to our knowledge, there is no frontier exploration applying large-scale pre-training to vehicle Re-ID research. Therefore, this paper first proposes VehicleMAE to leverage the MAE architecture for pre-training on the DiffVERI dataset, to provide a powerful pre-training model for generic vehicle Re-ID tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 316, + 121, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 121, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 121, + 328 + ], + "type": "text", + "content": "3. DiffVERI" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 411 + ], + "type": "text", + "content": "Owing to the remarkable success of diffusion models in the field of data generation [34, 47], a text-to-image diffusion model is employed to construct a large-scale vehicle Re-ID dataset, DiffVERI, with a comprehensive view. The construction of DiffVERI comprises the following three primary steps:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 206, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 206, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 206, + 436 + ], + "type": "text", + "content": "3.1. Multi-view data generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "text", + "content": "The off-the-shelf diffusion model, termed DreamBooth [33], is used for fine-tuning on the VERI-Wild to produce highly realistic data. Specifically, a specific text description " + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "text", + "content": ", that is, \"A photo of [V] vehicle\" is created for each vehicle instance, where [V] represents the identity-specific identifier to ensure ID consistency between the generated and input images. Given an initial noise of " + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim N(0,1)" + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "text", + "content": ", the denoising process of the diffusion model is followed, and the model is then optimized by balancing the differences between the generated and original images. The noise " + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 443, + 296, + 575 + ], + "type": "text", + "content": " in the latent space can be expressed as Eq. 1:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 587, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 587, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 295, + 604 + ], + "type": "interline_equation", + "content": "z _ {t} (x, \\varepsilon) = \\sqrt {\\tilde {a} _ {t}} \\cdot x + \\sqrt {1 - \\tilde {a} _ {t}} \\cdot \\varepsilon \\tag {1}", + "image_path": "a6d2beaad5d24ca781da988b9db96089b6dc2939bc258881f053937592fc630e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\tilde{a}_t" + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": " is the average denoising factor for the diffusion process time " + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": " represents the original image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Subsequently, the text features extracted by the text encoder " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " of DreamBooth are used as a conditioning vector " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{pr}} = \\Gamma \\bigl (\"\\mathrm{a photo of vehicle}\"\\bigr)" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " for the pre-trained text-to image diffusion model " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{\\theta}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " to generate images " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{pr} = x_{\\theta}(z_t,c_{pr})" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". As the objective function in the fine-tuning of the diffusion model process, squared error loss can be de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 375, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 375, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 375, + 84 + ], + "type": "text", + "content": "fined as Eq. 2:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 92, + 553, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 92, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 355, + 92, + 553, + 135 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {x, c, \\varepsilon , \\varepsilon^ {\\prime}, t} [ w _ {t} \\| x _ {\\theta} (z _ {t} (x, \\varepsilon), c) - x \\| _ {2} ^ {2} + \\\\ \\left. \\lambda w _ {t} ^ {\\prime} \\left\\| x _ {\\theta} \\left(z _ {t} ^ {\\prime} \\left(x _ {p r}, \\varepsilon^ {\\prime}\\right), c _ {p r}\\right) - x _ {p r} \\right\\| _ {2} ^ {2} \\right] \\tag {2} \\\\ \\end{array}", + "image_path": "07a97b74c9b19aac7f47476a531ddfbcf07f821c998df86b5933f77107e2d855.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "text", + "content": " represents the quality term of sample generation, and " + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "text", + "content": " is a conditioning vector derived from the text " + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "text", + "content": " is the relative weight that controls a specific identity vehicle. A flexible multi-view generation prompt, that is, \"A photo of [V] vehicle [O] rotated [X] degrees clockwise\", where [O] and [X] are variables that control the generation direction and rotation angle, respectively, and " + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "inline_equation", + "content": "[O] \\in left, right, [X] \\in 30, 60, 90, 120, 150, 180" + }, + { + "bbox": [ + 313, + 143, + 555, + 262 + ], + "type": "text", + "content": ", is designed to generate multi-view vehicle images during the inference process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 402, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 402, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 402, + 283 + ], + "type": "text", + "content": "3.2. Data filtering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "type": "text", + "content": "Motivated by ours insights, two types of data noise are inevitably found in the generated data: (1) part of the images exhibit pixel blurring and low resolution, and (2) the vehicle body occupies a small proportion of the entire image, increasing the difficulty in focusing on the details of each view. Images smaller than 10K in size are first empirically filtered out to address the aforementioned issues. Subsequently, we resort to YOLOv7 [40] to detect the bounding box of the vehicle body and then calculate the proportion ratio between the vehicle body and the entire image. Based on the specifications, images with a proportion ratio below " + }, + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 288, + 555, + 431 + ], + "type": "text", + "content": " are further filtered." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 316, + 440, + 555, + 523 + ], + "blocks": [ + { + "bbox": [ + 316, + 440, + 555, + 523 + ], + "lines": [ + { + "bbox": [ + 316, + 440, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 555, + 523 + ], + "type": "table", + "html": "
DatasetsImagesSourceMulti-view Semantic AnnotationsResolutionViews
VeRi-776[25]49,357RealX243 × 214Constrained
VehicleID[26]221,763RealX374 × 412Constrained
VeRi-Wild[27]416,314RealX415 × 354Constrained
VehicleX[48]192,150SyntheticX256 × 256Constrained
VRAI[42]137,613RealX349 × 234Constrained
DiffVERI(Ours)1,712,703Real-synthetic✓496 × 485Diverse
", + "image_path": "47255859e0f7554f99f3fd9349a2eb5e616dc49765b8a6d491477df2c4569fd3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 564 + ], + "type": "text", + "content": "Table 1. Comparison of the statistics between DiffVERI and other public vehicle Re-ID datasets. In contrast, DiffVERI is currently the largest multi-view vehicle Re-ID benchmark." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 588, + 492, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 492, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 492, + 600 + ], + "type": "text", + "content": "3.3. Multi-view semantic annotations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "After dataset collection, 10 well-trained annotators were employed, and 1200 hours were spent annotating " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "20\\mathrm{k}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " images with multi-view semantic segmentation masks. In particular, each vehicle is uniformly parsed into a background region and four view regions (i.e., front, rear, side, and top views). Some images undeniably lack all four views or demonstrate partially occluded views. Accordingly, images for annotation that have comprehensive view information are selected. Considering the powerful zero-shot semantic" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4703" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 79, + 523, + 283 + ], + "blocks": [ + { + "bbox": [ + 91, + 79, + 523, + 283 + ], + "lines": [ + { + "bbox": [ + 91, + 79, + 523, + 283 + ], + "spans": [ + { + "bbox": [ + 91, + 79, + 523, + 283 + ], + "type": "image", + "image_path": "5b6c22317b955e3f3ec4b8cd11b355c4fc3593b1d04ba53bb3ea49736a306e0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 301, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 301, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 555, + 346 + ], + "type": "text", + "content": "Figure 2. Overall pipeline of the proposed VehicleMAE. Firstly, VMIM module utilizes two homogeneous encoder-decoders to reconstruct a pair of asymmetric masked-out patches for RGB pixels and multi-view semantic tasks. Subsequently, to fully develop the potential of homogeneous architecture, PPMD module exploits the mutual knowledge exchange between the historical and current models to enhance the generalization performance of the pre-training model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 367, + 294, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 367, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 367, + 294, + 438 + ], + "type": "text", + "content": "segmentation capability of the SAM model [21], the annotated masks are utilized as the ground-truth, and SAM is fine-tuned through a lightweight LoRA adapter [16]. The fine-tuned SAM can efficiently perform multi-view semantic inference on the remaining unlabeled images, thereby obtaining multi-view annotations for the entire DiffVERI." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 440, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 440, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 440, + 295, + 523 + ], + "type": "text", + "content": "Following the above steps, a large-scale vehicle Re-ID dataset, DiffVERI, is ultimately constructed, which includes raw VERI-Wild and synthesized multi-view data. Table 1 reviews some attributes of existing vehicle Re-ID datasets, including number of images, data sources, multiview semantic annotations, average image resolution, and range of views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 536, + 135, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 135, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 135, + 549 + ], + "type": "text", + "content": "4. VehicleMAE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Motivated by the success of MAE, we propose a novel masked image modeling pre-training paradigm, VehicleMAE, which is outlined in Figure 2, including two modules, namely view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). In VMIM module, given an image " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{img}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " and multi-view annotation " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{seg}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ", a pair of visible patch maps with non-overlapping patches is generated based on an asymmetric mask sampling strategy to ensure that two homogeneous masked autoencoders " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\phi_1" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi_2)" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " can perform different degrees of self-supervised reconstruction tasks within one epoch. More specifically, " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " comprise its own encoder, RGB decoder, and view semantic decoder, that is," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_1 = \\{E_{\\phi_1},D_{\\phi_1}^{img},D_{\\phi_1}^{seg}\\}" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_{2} = \\{E_{\\phi_{2}},D_{\\phi_{2}}^{img},D_{\\phi_{2}}^{seg}\\}" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": ", respectively. The map of features output by the encoders is " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\{F_{\\phi_1}^{img},F_{\\phi_1}^{seg},F_{\\phi_2}^{img},F_{\\phi_2}^{seg}\\}" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": ", and the map of features output by the decoders is " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\{f_{\\phi_1}^{img},f_{\\phi_1}^{seg},f_{\\phi_2}^{img},f_{\\phi_2}^{seg}\\}" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": ". Then, the PPMD module enables two MAEs to fully exchange their current and past knowledge. In the current epoch " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": ", the decoders of " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": " act as collaborative training for students to distill the predicted target features from each other. Meanwhile, in the previous epoch " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": ", the two histories of " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_1'" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_2'" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": " are regarded as teachers exchanging students " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 313, + 365, + 555, + 513 + ], + "type": "text", + "content": " with each other and then distilling the intermediate features of encoders." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 526, + 537, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 537, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 537, + 540 + ], + "type": "text", + "content": "4.1. View-asymmetry Masked Image Modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "The fundamental principle of existing image-based MAEs is to perform pixel-level masking and reconstruction using an encoder-decoder architecture. When MAE is extended to vehicle Re-ID, we empirically argue that there exist two main defects: (1) learning robust representation is difficult due to random mask sampling and low-level pixel reconstruction; and (2) reconstructing the original image pixels inevitably involves some identity-unrelated redundant region, leading to an inefficient and resource-intensive masked image modeling task. Consequently, this paper further develops VMIM, which has two homogeneous MAEs " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": ". As depicted in Figure 3, this VMIM design incorporates the following two submodules to tackle the aforementioned two issues encountered by vanilla MAE" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4704" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 76, + 301, + 184 + ], + "blocks": [ + { + "bbox": [ + 86, + 76, + 301, + 184 + ], + "lines": [ + { + "bbox": [ + 86, + 76, + 301, + 184 + ], + "spans": [ + { + "bbox": [ + 86, + 76, + 301, + 184 + ], + "type": "image", + "image_path": "c5a41b115ce048202b829ea79dadde2311ee48430af7a8a8ed1c3a8a02bc759d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "type": "text", + "content": "Figure 3. VMIM consist of two submodules: (a) The asymmetric mask sampling strategy in VMIM module generates a pair of visible maps without overlapping patches to create diverse preservation clues for reconstruction tasks. (b) Illustration of the identity-related region reconstruction in terms of " + }, + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "type": "text", + "content": ". For the feature output of the decoder, RGB pixel and multi-view semantic reconstruction with MSE loss is performed only for the vehicle body region. Notably, another peer " + }, + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "type": "text", + "content": " also performs a series of reconstruction processes similar to " + }, + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 54, + 198, + 555, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 76, + 525, + 184 + ], + "blocks": [ + { + "bbox": [ + 304, + 76, + 525, + 184 + ], + "lines": [ + { + "bbox": [ + 304, + 76, + 525, + 184 + ], + "spans": [ + { + "bbox": [ + 304, + 76, + 525, + 184 + ], + "type": "image", + "image_path": "7933f87441b368d4be3bcd76da04e48ee2a1dd822425abeae3bf8d8b3d280fc6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 119, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 119, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 119, + 276 + ], + "type": "text", + "content": "point-by-point:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": "(i) Asymmetric mask sampling strategy. Initially, ours intention is to fully utilize the potential of self-supervised pre-training to serve downstream tasks with view variation by employing two homogeneous architectures. However, two twin MAEs will inevitably fail to expand the scope of their knowledge if the same masking and reconstruction operations are employed to facilitate equal knowledge exchange. Motivated by these analyses, an asymmetric mask sampling strategy is designed to generate paired visible maps without overlapping patches and encourage two MAEs to learn diverse knowledge, as shown in Figure 3(a). Concretely, the image " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "I_{img}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " and the corresponding multi-view mask " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "I_{seg}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " are simultaneously injected into the encoders of " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "E_{\\phi_1}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "E_{\\phi_2}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": ", and are respectively divided into two sequences with " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " patches " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\{P(I_{img}|\\phi_1); P(I_{seg}|\\phi_1)\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\{P(I_{img}|\\phi_2); P(I_{seg}|\\phi_2)\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " in spatial order. Subsequently, based on the semantic segmentation annotation of " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "I_{seg}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": ", each sequence is further subdivided into two types: (1) subsequences " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{id}|\\phi_1) = \\{P(I_{img}^{id}|\\phi_1); P(I_{seg}^{id}|\\phi_1)\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{id}|\\phi_2) = \\{P(I_{img}^{id}|\\phi_2); P(I_{seg}^{id}|\\phi_2)\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " of the vehicle body region with size " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "N_{id}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": "; (2) subsequences " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{bg}|\\phi_1) = \\{P(I_{img}^{bg}|\\phi_1); P(I_{seg}^{bg}|\\phi_1)\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{bg}|\\phi_2) = \\{P(I_{img}^{bg}|\\phi_2); P(I_{seg}^{bg}|\\phi_2)\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " of the background region with size " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "N_{bg}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": ", thereby also fulfilling " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "N_{id} + N_{bg} = N" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": ". The pair of visible patch maps " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\{M_{\\phi_1}, M_{\\phi_2}\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " obtained through the asymmetric mask sampling strategy follows two-point principles. The first point is to randomly mask the patches of the background region " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{bg}|\\phi_1)" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{bg}|\\phi_2)" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " as much as possible and preserve the vehicle body region " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{id}|\\phi_1)" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "P(I^{id}|\\phi_2)" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " as visible clues for the reconstruction task. The second point is that no overlap simultaneously exists in the masking patches of two MAEs, that is, " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "M_{\\phi_1} \\cap M_{\\phi_2} = \\emptyset" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": ". Notably, " + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\{M_{\\phi_1}, M_{\\phi_2}\\}" + }, + { + "bbox": [ + 56, + 278, + 296, + 712 + ], + "type": "text", + "content": " will be will be complemented each other in RGB pixels and multi-view semantic reconstruction to enhance collaboration between the two types of tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": "(ii) Identity-related region reconstruction. Directly using the entire original image and a multi-view segmentation mask as the prediction target of reconstruction inevitably increases the computational cost and reduces the pre-training efficiency. Considering the aforementioned analysis, an identity-related region reconstruction method is designed. First, similar to the tokenizing and encoding operations of ordinary MAEs, the visible patch maps " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "\\{M_{\\phi_1}, M_{\\phi_2}\\}" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": " obtained by asymmetric mask sampling are tokenized and embedded into a sequence of tokens, and the masked patches will be replaced with a sequence of learnable masked tokens. Afterward, the visible and invisible tokens are combined and then inputted into the decoders " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "\\{D_{\\phi_1}^{img}, D_{\\phi_1}^{seg}\\}" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "\\{D_{\\phi_2}^{img}, D_{\\phi_2}^{seg}\\}" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": " of the two MAEs. Ultimately, in the decoding stage, the ground-truth of two task-specific reconstructions is set with identity-related regions " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "P(I^{id}|\\phi_1)" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "P(I^{id}|\\phi_2)" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": ", that is, the RGB pixels and multi-view semantic regions of the vehicle body, instead of the initial " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "I_{img}" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "I_{seg}" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": ". Taking " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": " as an example, Figure 3(b) illustrates the reconstruction process of the proposed method. Mean square error (MSE) is adopted to calculate the error of RGB pixels and multi-view semantics in the reconstructed body region. The loss function " + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "inline_equation", + "content": "L_{rec}" + }, + { + "bbox": [ + 313, + 263, + 555, + 550 + ], + "type": "text", + "content": " of the entire reconstruction task can be defined by Eq. 3:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 559, + 558, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 558, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 558, + 596 + ], + "type": "interline_equation", + "content": "L _ {r e c} = \\sum_ {m \\in \\{1, 2 \\}} \\sum_ {\\triangle \\in \\{i m g, s e g \\}} M S E \\left(R \\left(I _ {\\triangle} ^ {i d} \\mid \\phi_ {m}\\right), P \\left(I _ {\\triangle} ^ {i d} \\mid \\phi_ {m}\\right)\\right) \\tag {3}", + "image_path": "4abd513566ce45a67fc74b8365da834628cccc9b790894d0ffc3f8c9f46cfdb8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": "where the term " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "\\phi_{m}" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "m\\in \\{1,2\\}" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " represent two models " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": ", respectively. The term " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "\\triangle \\in \\{img,seg\\}" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " denotes the reconstruction task of RGB pixels and multi-view semantics. " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "R(I_{\\triangle}^{id}|\\phi_m)" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "P(I_{\\triangle}^{id}|\\phi_m)" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " are the predicted results and ground-truth of the model " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "\\phi_{m}" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " performing two reconstruction tasks, respectively. Correspondingly, function " + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "inline_equation", + "content": "MSE(\\cdot)" + }, + { + "bbox": [ + 313, + 596, + 556, + 693 + ], + "type": "text", + "content": " is the MSE between the predicted target and ground-truth." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4705" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 242, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 242, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 242, + 84 + ], + "type": "text", + "content": "4.2. Past-to-present Mutual-distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": "While the VMIM module can diversify the knowledge learned by " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": ", empirical investigation revealed that this module cannot guarantee collaborative training between the two models and deep mining of their self-knowledge. Toward this end, this study aimed to establish the teacher-student role relationship, uniting the past and present knowledge of two models, and design a past-to-present mutual distillation method. More precisely, the core of this design is to conduct feature-level knowledge distillation through the collaboration of current models and the cross-guidance of historical models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": "Assuming the current pre-training is at epoch " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " play the roles of two student models. They collaboratively learn and exchange the predicted features with each other based on asymmetric visible clues to enhance the multi-view generalization capability of the entire VehicleMAE. The predicted feature maps " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\{f_{\\phi_1}^{img}, f_{\\phi_1}^{seg}\\}" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\{f_{\\phi_2}^{img}, f_{\\phi_2}^{seg}\\}" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " of the decoder are extracted for " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\phi_1" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\phi_2" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " in two specific reconstruction tasks, respectively. Kullback-Leibler (KL) divergence is then adopted to minimize the predicted feature distribution error. The distillation losses " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "L_{pr}^{\\phi_1}" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "L_{pr}^{\\phi_2}" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " of the two MAEs in the current epoch " + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 221, + 296, + 367 + ], + "type": "text", + "content": " are represented by Eq. 4:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 371, + 296, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 371, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 87, + 371, + 296, + 392 + ], + "type": "interline_equation", + "content": "L _ {p r} ^ {\\phi_ {1}} = K L \\left(f _ {\\phi_ {2}} ^ {i m g} \\mid \\mid f _ {\\phi_ {1}} ^ {i m g}\\right) + K L \\left(f _ {\\phi_ {2}} ^ {s e g} \\mid \\mid f _ {\\phi_ {1}} ^ {s e g}\\right) \\tag {4}", + "image_path": "4c07738e55cddde2aaa5449e1f6ceb63b36548056f4a8db29156da69a1061190.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 389, + 265, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 389, + 265, + 405 + ], + "spans": [ + { + "bbox": [ + 87, + 389, + 265, + 405 + ], + "type": "interline_equation", + "content": "L _ {p r} ^ {\\phi_ {2}} = K L \\left(f _ {\\phi_ {1}} ^ {i m g} | | f _ {\\phi_ {2}} ^ {i m g}\\right) + K L \\left(f _ {\\phi_ {1}} ^ {s e g} | | f _ {\\phi_ {2}} ^ {s e g}\\right)", + "image_path": "cae39d0b1c7f1b6755adcc5d7651bf6459fd526cea7503e3e9f4485500a02d74.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 408, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 408, + 295, + 432 + ], + "type": "inline_equation", + "content": "KL(\\cdot)" + }, + { + "bbox": [ + 55, + 408, + 295, + 432 + ], + "type": "text", + "content": " is the KL divergence function that measures the predicted features of two models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "N\\geq 2" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " , the two historical models " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_1^{\\prime}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_2^\\prime" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " corresponding to epoch " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " in the past are regarded as teachers cross-guiding student models " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " of epoch " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " to dynamically mine their own knowledge. Specifically, " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_1^{\\prime}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " is set as the historical teacher for " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_2^\\prime" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " as the historical teacher for " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " to establish the teacher-student relationships from the past to the present. This mutual exchange of teacher-student relationships can further enhance the potential for mutual distillation and knowledge diversity. The encoded feature maps " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\{F_{\\phi_1'}^{img},F_{\\phi_1'}^{seg},F_{\\phi_2'}^{img},F_{\\phi_2'}^{seg}\\}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\{F_{\\phi_1}^{img},F_{\\phi_1}^{seg},F_{\\phi_2}^{img},F_{\\phi_2}^{seg}\\}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " output by the encoders of two historical teacher models and two current student models are employed to distill intermediate features due to the inability of self-supervised pre-training to achieve hard-label distillation. The distillation loss " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "L_{pa}^{\\phi_1}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " guided by " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_2^\\prime" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_{1}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " and the distillation loss " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "L_{pa}^{\\phi_2}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " guided by " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_1^\\prime" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "inline_equation", + "content": "\\phi_{2}" + }, + { + "bbox": [ + 55, + 432, + 296, + 644 + ], + "type": "text", + "content": " are defined as Eq.5:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 647, + 295, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 647, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 84, + 647, + 295, + 671 + ], + "type": "interline_equation", + "content": "L _ {p a} ^ {\\phi_ {1}} = K L \\left(F _ {\\phi_ {2} ^ {\\prime}} ^ {i m g} \\mid \\mid F _ {\\phi_ {1}} ^ {i m g}\\right) + K L \\left(F _ {\\phi_ {2} ^ {\\prime}} ^ {s e g} \\mid \\mid F _ {\\phi_ {1}} ^ {s e g}\\right) \\tag {5}", + "image_path": "59b6226fb24b0974b27d9d57d9f3c5962227330fda990983edfff9df66292c5f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 668, + 269, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 668, + 269, + 686 + ], + "spans": [ + { + "bbox": [ + 84, + 668, + 269, + 686 + ], + "type": "interline_equation", + "content": "L _ {p a} ^ {\\phi_ {2}} = K L (F _ {\\phi_ {1} ^ {\\prime}} ^ {i m g} | | F _ {\\phi_ {2}} ^ {i m g}) + K L (F _ {\\phi_ {1} ^ {\\prime}} ^ {s e g} | | F _ {\\phi_ {2}} ^ {s e g})", + "image_path": "772783f5576017b734d80fe0649d1da25d805e812d24d53dcd0daaa6d1774d16.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "After the pre-training of VehicleMAE is completed, all decoders will be discarded and the learned knowledge will" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "text", + "content": "be fused by solving the average parameters of the two encoders " + }, + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "inline_equation", + "content": "E_{\\phi_1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "inline_equation", + "content": "E_{\\phi_2}" + }, + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "text", + "content": ". The process of obtaining the mean model " + }, + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "inline_equation", + "content": "E[\\bar{\\theta}]" + }, + { + "bbox": [ + 313, + 72, + 554, + 110 + ], + "type": "text", + "content": " can be written as Eq. 6:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 118, + 553, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 118, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 369, + 118, + 553, + 142 + ], + "type": "interline_equation", + "content": "E [ \\bar {\\theta} ] = \\frac {E _ {\\phi_ {1}} [ \\theta_ {\\phi_ {1}} ] + E _ {\\phi_ {2}} [ \\theta_ {\\phi_ {2}} ]}{2} \\tag {6}", + "image_path": "674b67b4045484637905a0b9925b822f00fe26f9c43270ec82c9217304ae2c0f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\theta_{\\phi_1}" + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\theta_{\\phi_2}" + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": " are the parameter of " + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "inline_equation", + "content": "E_{\\phi_1}" + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "inline_equation", + "content": "E_{\\phi_2}" + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\bar{\\theta}" + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": " is the parameter of the mean model " + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "inline_equation", + "content": "E[\\bar{\\theta}]" + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": ". It is worth emphasizing that " + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "inline_equation", + "content": "E[\\bar{\\theta}]" + }, + { + "bbox": [ + 313, + 149, + 555, + 212 + ], + "type": "text", + "content": " will be directly applied as a single pre-training model to downstream tasks of supervised and unsupervised vehicle Re-ID." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 219, + 425, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 425, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 425, + 232 + ], + "type": "text", + "content": "4.3. Training Objective" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "text", + "content": "The training objectives of VehicleMAE include two parts: the loss of the reconstruction task of two homogeneous MAEs in the VMIM module; in the PPMD module, the distillation loss " + }, + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "inline_equation", + "content": "L_{pr} = L_{pr}^{\\phi_1} + L_{pr}^{\\phi_2}" + }, + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "text", + "content": " between the current two student models and the distillation loss " + }, + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "inline_equation", + "content": "L_{pa} = L_{pa}^{\\phi_1} + L_{pa}^{\\phi_2}" + }, + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "text", + "content": " guided by the history teachers. Therefore, the total loss " + }, + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "inline_equation", + "content": "L_{total}" + }, + { + "bbox": [ + 313, + 236, + 554, + 333 + ], + "type": "text", + "content": " of the overall pre-training process can be defined by Eq. 7 as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 377, + 334, + 553, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 334, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 377, + 334, + 553, + 347 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L _ {r e c} + L _ {p r} + L _ {p a} \\tag {7}", + "image_path": "a720e88f785a0b1089a3a33009521e5f8d97f2508ebf1bd8cb5d51bebe638524.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "text", + "content": "Note that, the pre-training mean model " + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "inline_equation", + "content": "E[\\bar{\\theta}]" + }, + { + "bbox": [ + 313, + 352, + 554, + 399 + ], + "type": "text", + "content": " directly serves as supervised and unsupervised vehicle Re-ID downstream tasks and is supervised by cross-entropy and triplet losses." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 411, + 395, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 395, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 395, + 425 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 431, + 497, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 497, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 497, + 443 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Protocols" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 449, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 532 + ], + "type": "text", + "content": "This paper utilizes two well-known vehicle Re-ID datasets: VeRi-776 [23] and VehicleID [3]. VehicleID dataset includes small, medium, and large-scale test sets. To evaluate the performance of the proposed method, mean Average Precision (mAP) and Cumulative Matching Characteristics (CMC) at Rank-" + }, + { + "bbox": [ + 313, + 449, + 554, + 532 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 449, + 554, + 532 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 449, + 554, + 532 + ], + "type": "inline_equation", + "content": "k = 1, 5" + }, + { + "bbox": [ + 313, + 449, + 554, + 532 + ], + "type": "text", + "content": ") are employed as the evaluation metrics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 540, + 447, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 447, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 447, + 553 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "content": "CLIP-B/16 serves as the backbone network. Throughout the whole training process, the image size is set to " + }, + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "content": ", and models update parameters using the Adam optimizer. During the pre-training phase, all samples are divided into mini-batches of size 128 for 300 epochs of iterative learning. For downstream tasks, the batch size is set to 64, and all samples are trained for 50 epochs. Additionally, each image undergoes data augmentation with random horizontal flipping, padding, cropping, and erasing. In unsupervised tasks, DBSCAN is applied as the clustering algorithm to generate pseudo-labels. All experiments are conducted under the Ubuntu 22.04 operating system and Pytorch environment with 4 NVIDIA RTX A6000 GPUs." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4706" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 149, + 85 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 281 + ], + "type": "text", + "content": "Effects of different modules. The results in Table 2 demonstrate the effectiveness of all proposed modules. Three observations are summarized by us: (1) When we independently introduce the \"Pa\" or \"Pr\" into the \"Baseline\" to perform knowledge distillation, the significant performance improvement in Rank accuracy and mAP confirms that historical and current distillation can fully explore and exchange their own knowledge. (2) Whether using \"Asym\" in the mask sampling strategy during the encoding stage or introducing \"Id-rec\" in the reconstruction task during the decoding stage, it is superior to MAE that integrates a random mask strategy and complete pixel reconstruction. (3) VehicleMAE belongs to the ensemble of all modules and achieves the best performance, further verifying that VehicleMAE can efficiently pre-train two homogeneous MAEs in a collaborative manner." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 291, + 297, + 392 + ], + "blocks": [ + { + "bbox": [ + 58, + 291, + 297, + 392 + ], + "lines": [ + { + "bbox": [ + 58, + 291, + 297, + 392 + ], + "spans": [ + { + "bbox": [ + 58, + 291, + 297, + 392 + ], + "type": "table", + "html": "
Ablation ModulesVeRi-776VehicleID
BaselinePaPrAsymId-recmAPRank-1Rank-1Rank-5
✓37.681.269.081.8
✓✓40.185.071.284.6
✓✓40.585.371.184.7
✓✓38.783.670.382.8
✓✓39.584.370.182.6
✓✓✓✓✓42.587.673.385.6
", + "image_path": "5cfd2d7eb641fd6110c29108e66bf1987cafa73c399ec95459ea42ed59d41b4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "text", + "content": "Effects of different mask sampling strategies and ratios. Table 3 reports the impact of using asymmetric mask and random mask sampling strategies at different mask ratios. We observe that the performance of the proposed \"Asym\" is superior to \"Random\" for all three mask ratios. Especially at a mask rate of " + }, + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "text", + "content": ", \"Asym\" achieves the highest mAP of " + }, + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "inline_equation", + "content": "42.5\\%" + }, + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "text", + "content": " and Rank-1 accuracy of " + }, + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "inline_equation", + "content": "87.6\\%" + }, + { + "bbox": [ + 55, + 473, + 296, + 593 + ], + "type": "text", + "content": ". This is due to the visible maps generated by \"Asym\" having no overlapping patches, which encourage two MAEs to learn and exchange knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Effects of different distillation strategies. Table 4 investigates the impact of modifying VehicleMAE using various distillation architectures on Re-ID performance. Ours findings are two-fold: (1) Compared to two MAEs learning independently, using historical distillation or present distillation can ameliorate the pre-training model performance, especially when combined to achieve more effective knowledge exchange. (2) The mean encoder " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "E[\\bar{\\theta}]" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " can accurately feedback and maintain the average level of two encoders, while also solving the difficulties of multiple model selec" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 71, + 558, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 400, + 296, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 456 + ], + "type": "text", + "content": "Table 2. Performance comparison of different components in unsupervised vehicle ReID. \"Past\" and \"Pres\" refer to historical distillation and present distillation in PPDM, respectively. \"Asym\" and \"Id-rec\" correspond to the Asymmetric mask sampling strategy and Identity-related region reconstruction in VMIM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 71, + 558, + 190 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 558, + 190 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 558, + 190 + ], + "type": "table", + "html": "
Mask RatiosStrategiesVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
25%Random38.884.268.483.6
Asym40.685.570.484.2
50%Random39.284.668.983.8
Asym41.786.271.184.8
75%Random40.185.169.584.1
Asym42.587.673.385.6
", + "image_path": "e61106b03bad235ffccfcf8c86c0a987c1fbe0f636476856cb6db928f652de7b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 262, + 489, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 489, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 489, + 274 + ], + "type": "text", + "content": "tion and deployment for downstream tasks." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 282, + 555, + 340 + ], + "blocks": [ + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "text", + "content": "Table 3. Effectiveness of different mask sampling strategies in serving unsupervised vehicle Re-ID tasks. \"Random\" and \"Asym\" express the random mask sampling strategy and the asymmetric mask sampling strategy, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 282, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 317, + 282, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 282, + 555, + 340 + ], + "type": "table", + "html": "
Pre-trained EncodersVeRi-776VehicleID
IndPaPrPa+PrIndPaPrPa+Pr
Eφ184.384.784.887.170.370.871.073.2
Eφ284.484.785.186.970.370.770.973.1
E[θ]84.685.085.387.670.971.271.173.3
", + "image_path": "b06f7edf49b3920fa45b2cd123ad68526850af70d6c020983e33244c5c436f98.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "text", + "content": "Table 4. Comparison Rank-1 accuracy of two homogeneous encoders " + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\{E_{\\phi_1}, E_{\\phi_2}\\}" + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "text", + "content": " and a mean encoder " + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "inline_equation", + "content": "E[\\tilde{\\theta}]" + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "text", + "content": " using various distillation architectures in unsupervised downstream tasks. \"Ind\" denotes that " + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "inline_equation", + "content": "E_{\\phi_1}" + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "inline_equation", + "content": "E_{\\phi_2}" + }, + { + "bbox": [ + 313, + 349, + 555, + 426 + ], + "type": "text", + "content": " learning independently without using distillation. \"Pa\", \"Pr\", and \"Pa+Pr\" respectively signify the independent use of historical distillation, present distillation, and the simultaneous use of both them." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 438, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 666 + ], + "type": "text", + "content": "Analysis of Model Complexity. To further analyze the computational performance, we re-implement different self-supervised pre-training works on DiffVERI for model parameter sizes and floating-point operations (FLOPs), as shown in Table 5. For fairness, all methods except ConvNeXt v2, which relies on the ConvNeXt network architecture, use the CLIP-B/16 image encoder as the backbone. Although VehicleMAE adopts a mutual-distillation architecture, our pre-training model contains 57.3M parameters and only needs 14.7G multiply-accumulate operations for feature extraction. This is mainly due to the mean model design and that we follow the vanilla MAE without directly changing the encoder-decoder architecture. In contrast, both the fully convolutional mask autoencoder in ConvNeXt v2 and the hybrid convolution module in CMAE will increase the resource consumption of feature extraction. These results demonstrate that VehicleMAE achieves a favorable trade-off between model complexity and the performance of downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 672, + 499, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 499, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 499, + 685 + ], + "type": "text", + "content": "5.4. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "As illustrated in Table 6, we compare the proposed method with state-of-the-art ones in both supervised and unsuper" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4707" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 70, + 247, + 129 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 247, + 129 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 247, + 129 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 247, + 129 + ], + "type": "table", + "html": "
MethodsFLOPsParams
ConvNeXt v2[45]15.4G89M
CMAE[18]19.4G104M
BEiT[43]15.7G82.5M
Ours14.7G57.3M
", + "image_path": "7ff6c990473334b3c0eb40dc48e8c5a5d2279b70f3682b22f112c19e81f74309.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 171, + 287, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 137, + 297, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 297, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 297, + 160 + ], + "type": "text", + "content": "Table 5. Comparison of different pre-training models in terms of parameter size and FLOPs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 171, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 69, + 171, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 69, + 171, + 287, + 521 + ], + "type": "table", + "html": "
MethodsVeRi-776VehicleID
mAPRank-1Rank-1Rank-5
Supervised Learning
\\( A\\text{VER}{*}^{R}\\left\\lbrack {19}\\right\\rbrack \\)61.289.063.585.6
\\( VehicleNet*R\\left\\lbrack {54}\\right\\rbrack \\)83.496.879.592.0
\\( TransReID^{*V}\\left\\lbrack {14}\\right\\rbrack \\)82.097.1--
\\( GiT^{*V}\\left\\lbrack {35}\\right\\rbrack \\)80.396.977.9-
\\( LCNL^{*V}\\left\\lbrack {46}\\right\\rbrack \\)81.897.4--
\\( TANet^{*R}\\left\\lbrack {17}\\right\\rbrack \\)83.696.878.292.6
\\( CFA-Net^{*R}\\left\\lbrack {41}\\right\\rbrack \\)80.796.9--
\\( HCI-Net^{*R}\\left\\lbrack {36}\\right\\rbrack \\)83.896.676.491.2
\\( CLIP-ReID^{*C}\\left\\lbrack {24}\\right\\rbrack \\)83.397.478.192.7
\\( Ours^{\\dagger C} \\)87.697.485.994.9
Unsupervised Learning
\\( MMT^{*R}\\left\\lbrack {7}\\right\\rbrack \\)25.460.931.042.4
\\( SPCL^{*R}\\left\\lbrack {8}\\right\\rbrack \\)36.979.953.066.4
\\( RLCC^{*R}\\left\\lbrack {51}\\right\\rbrack \\)39.683.4--
\\( PPLR^{*R}\\left\\lbrack {4}\\right\\rbrack \\)41.685.6--
\\( VAPC TO^{*R}\\left\\lbrack {52}\\right\\rbrack \\)30.476.2--
\\( ICL^{*R}\\left\\lbrack {37}\\right\\rbrack \\)39.583.7--
\\( AdaMG^{*R}\\left\\lbrack {31}\\right\\rbrack \\)41.086.2--
\\( NNNI^{*R}\\left\\lbrack {9}\\right\\rbrack \\)42.386.3--
\\( STDA^{*R}\\left\\lbrack {13}\\right\\rbrack \\)42.387.4--
\\( MAE†^C\\left\\lbrack {12}\\right\\rbrack \\)37.681.269.081.8
\\( CL-MAE†^C\\left\\lbrack {29}\\right\\rbrack \\)41.084.070.882.6
\\( CMAE†^C\\left\\lbrack {18}\\right\\rbrack \\)41.684.771.184.9
\\( Ours^{\\dagger C} \\)42.587.673.385.6
", + "image_path": "82e1c86497b08c526a71cbbfa74f5148cb63c351b3397894d9255b91fd78d6f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "text", + "content": "Table 6. Performance comparisons with the latest methods on supervised and unsupervised vehicle Re-ID tasks. “" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "inline_equation", + "content": "*^{R}" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "text", + "content": ", “" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "inline_equation", + "content": "*^{V}" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "text", + "content": ", and “" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "inline_equation", + "content": "*^{C}" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "text", + "content": "” indicate using ResNet, ViT, and the image encoder of CLIP as backbones, respectively. “" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\dagger^{C}" + }, + { + "bbox": [ + 55, + 530, + 296, + 608 + ], + "type": "text", + "content": "” denotes the implementation of pre-training using the image encoder of CLIP on DiffVERI. Note that experimental results are accurate to one decimal place, and this paper uses only the large-scale test set of VehicleID." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "vised downstream tasks. Compared to the most competitive methods among ResNet-based, ViT-based, and CLIP-based backbones, our method delivers the best mAP and Rank accuracy for both supervised and unsupervised settings. For fair comparison, we further implement several well-established MAE-based pre-training methods on DiffVERI. The noticeable performance improvement implies" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 332, + 71, + 538, + 161 + ], + "blocks": [ + { + "bbox": [ + 332, + 71, + 538, + 161 + ], + "lines": [ + { + "bbox": [ + 332, + 71, + 538, + 161 + ], + "spans": [ + { + "bbox": [ + 332, + 71, + 538, + 161 + ], + "type": "image", + "image_path": "e90d2afdae2e949377b78ae246f199bf6ee3e205abffb5f1d19dc8820e21434d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 171, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 237 + ], + "type": "text", + "content": "Figure 4. Visualization of patch-level feature distributions extracted by Baseline and using VehicleMAE pre-training. The class with the highest proportion of pixels within each patch in the patch-level ground-truth is defined as the final class corresponding to the patch. Different color patches represent different views or background regions" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 258, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 282 + ], + "type": "text", + "content": "that the masked image modeling based on VehicleMAE is better suited for downstream tasks of vehicle Re-ID." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 292, + 432, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 432, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 432, + 304 + ], + "type": "text", + "content": "5.5. Qualitative Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 309, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 490 + ], + "type": "text", + "content": "Figure 4 displays two visualization examples of the feature distribution at the patch-level extracted by Baseline and VehicleMAE. Specifically, we feed the features of all patches into the K-Means algorithm to perform five class-based clustering operations (i.e., a background region and four view regions) to intuitively demonstrate the view discriminate ability of different pre-training models. It can be observed that our VehicleMAE is able to distinguish patches from different views and backgrounds in both examples, getting close to patch-level ground-truth. Remarkably, even in stricter occlusion scenarios, VehicleMAE accurately avoid interference from the power poles in example B, further completing the extraction of multi-view information. More details on qualitative visualization experiments are included in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": "This paper releases DiffVERI, a large-scale multi-view vehicle Re-ID dataset for learning view-invariant representations, and proposes a masked image modeling pre-training paradigm termed VehicleMAE specially for vehicle Re-ID downstream tasks. VehicleMAE first proposes a VMIM module that attempts to apply two homogeneous MAEs to predict the RGB pixels and multi-view semantic clues of vehicles in pairs, thereby gaining diverse multi-view inference capabilities. Subsequently, to facilitate learning collaboratively, a PPMD module is designed to progressively exchange knowledge with each other. Extensive experiments demonstrate that equipping our pre-training model can achieve competitive performance in generic vehicle Re-ID downstream tasks. Future work contributes to further expanding VehicleMAE into a unified multimodal pre-training paradigm." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4708" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 228 + ], + "type": "text", + "content": "[1] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In European Conference on Computer Vision, pages 348-367. Springer, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 230, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 284 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 341 + ], + "type": "text", + "content": "[3] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 347–355, 2022. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 397 + ], + "type": "text", + "content": "[4] Yoonki Cho, Woo Jae Kim, Seunghoon Hong, and Sung-Eui Yoon. Part-based pseudo label refinement for unsupervised person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7308–7318, 2022. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 464 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 521 + ], + "type": "text", + "content": "[6] Dengpan Fu, Dongdong Chen, Jianmin Bao, Hao Yang, Lu Yuan, Lei Zhang, Houqiang Li, and Dong Chen. Unsupervised pre-training for person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14750–14759, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "text", + "content": "[7] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 612 + ], + "type": "text", + "content": "[8] Yixiao Ge, Feng Zhu, Dapeng Chen, Rui Zhao, et al. Self-paced contrastive learning with hybrid memory for domain adaptive object re-id. Advances in neural information processing systems, 33:11309-11321, 2020. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 667 + ], + "type": "text", + "content": "[9] Dongchen Han, Weifeng Liu, Mingchen Zou, and Baodi Liu. Non-contrastive nearest neighbor identity-guided method for unsupervised object re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(6):2713–2723, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "text", + "content": "[10] Bing He, Jia Li, Yifan Zhao, and Yonghong Tian. Part-regularized near-duplicate vehicle re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3997–4005, 2019. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant No. 62461037 and the Jiangxi Provincial Natural Science Foundation under Grant Nos. 20224BAB212011, 20232BAB202051, 20232BAB212008, and 20242BAB25078." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[11] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "text", + "content": "[13] Qiaolin He, Zihan Wang, Zhijie Zheng, and Haifeng Hu. Spatial and temporal dual-attention for unsupervised person re-identification. IEEE Transactions on Intelligent Transportation Systems, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 231, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 275 + ], + "type": "text", + "content": "[14] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013–15022, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[15] Zhijun He, Hongbo Zhao, Jianrong Wang, and Wenquan Feng. Multi-level progressive learning for unsupervised vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(4):4357-4371, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "type": "text", + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[17] Wenbo Hu, Hongjian Zhan, Palaiahnakote Shivakumara, Umapada Pal, and Yue Lu. Tanet: Text region attention learning for vehicle re-identification. Engineering Applications of Artificial Intelligence, 133:108448, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 477 + ], + "type": "text", + "content": "[18] Zhicheng Huang, Xiaojie Jin, Chengze Lu, Qibin Hou, Ming-Ming Cheng, Dongmei Fu, Xiaohui Shen, and Jiashi Feng. Contrastive masked autoencoders are stronger vision learners. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "text", + "content": "[19] Pirazh Khorramshahi, Amit Kumar, Neehar Peri, Sai Saketh Rambhatla, Jun-Cheng Chen, and Rama Chellappa. A dual-path model with adaptive attention for vehicle re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6132–6141, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[20] Pirazh Khorramshahi, Vineet Shenoy, and Rama Chellappa. Robust and scalable vehicle re-identification via self-supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5295–5304, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 645 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "text", + "content": "[22] Hongchao Li, Chenglong Li, Aihua Zheng, Jin Tang, and Bin Luo. Attribute and state guided structural embedding network for vehicle re-identification. IEEE transactions on image processing, 31:5949-5962, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[23] Hongchao Li, Aihua Zheng, Liping Sun, and Yonglong Luo. Camera topology graph guided vehicle re-identification." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4709" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "IEEE Transactions on Multimedia, 26:1565-1577, 2023. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[24] Siyuan Li, Li Sun, and Qingli Li. Clip-reid: exploiting vision-language model for image re-identification without concrete text labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1405–1413, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[25] Hongye Liu, Yonghong Tian, Yaowei Yang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2167-2175, 2016. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 295, + 239 + ], + "type": "text", + "content": "[26] Xinchen Liu, Wu Liu, Huadong Ma, and Huiyuan Fu. Large-scale vehicle re-identification in urban surveillance videos. In 2016 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2016. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 293 + ], + "type": "text", + "content": "[27] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. Veri-wild: A large dataset and a new method for vehicle re-identification in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3235-3243, 2019. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 295, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 338 + ], + "type": "text", + "content": "[28] Zefeng Lu, Ronghao Lin, Qiaolin He, and Haifeng Hu. Mask-aware pseudo label denoising for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 24(4):4333-4347, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 393 + ], + "type": "text", + "content": "[29] Neelu Madan, Nicolae-Catalin Ristea, Kamal Nasrollahi, Thomas B Moeslund, and Radu Tudor Ionescu. Cl-mae: Curriculum-learned masked autoencoders. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2492–2502, 2024. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 295, + 459 + ], + "type": "text", + "content": "[30] Dechao Meng, Liang Li, Xuejing Liu, Yadong Li, Shijie Yang, Zheng-Jun Zha, Xingyu Gao, Shuhui Wang, and Qingming Huang. Parsing-based view-aware embedding network for vehicle re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7103–7112, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "text", + "content": "[31] Jinjia Peng, Guangqi Jiang, and Huibing Wang. Adaptive memorization with group labels for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5802-5813, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "text", + "content": "[32] Jiuchao Qian, Minting Pan, Wei Tong, Rob Law, and Edmond Q Wu. Urrnet: A unified relational reasoning network for vehicle re-identification. IEEE Transactions on Vehicular Technology, 72(9):11156-11168, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[34] Zhiyin Shao, Xinyu Zhang, Changxing Ding, Jian Wang, and Jingdong Wang. Unified pre-training with pseudo texts for text-to-image person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11174–11184, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[35] Fei Shen, Yi Xie, Jianqing Zhu, Xiaobin Zhu, and Huanqiang Zeng. Git: Graph interactive transformer for vehicle re-identification. IEEE Transactions on Image Processing, 32:1039–1051, 2023. 2, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[36] Ke Sun, Xiyu Pang, Meifeng Zheng, Xiushan Nie, Xi Li, Houren Zhou, and Yilong Yin. Heterogeneous context interaction network for vehicle re-identification. Neural Networks, 169:293–306, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "text", + "content": "[37] Qing Tang and Kang-Hyun Jo. Unsupervised object re-identification via instances correlation loss. In 2022 IEEE 20th International Conference on Industrial Informatics (IN-DIN), pages 135–139. IEEE, 2022. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "text", + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078-10093, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[39] Jingzheng Tu, Cailian Chen, Xiaolin Huang, Jianping He, and Xinping Guan. Dfr-st: Discriminative feature representation with spatio-temporal cues for vehicle re-identification. Pattern Recognition, 131:108887, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "text", + "content": "[40] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464-7475, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "type": "text", + "content": "[41] Hai Wang, Yaqing Niu, Long Chen, Yicheng Li, Miguel Angel Sotelo, Zhixiong Li, and Yingfeng Cai. Dair-v2xreid: A new real-world vehicle-infrastructure cooperative re-id dataset and cross-shot feature aggregation network perception method. IEEE Transactions on Intelligent Transportation Systems, 25(8):9058-9068, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 433 + ], + "type": "text", + "content": "[42] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 460–469, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 510 + ], + "type": "text", + "content": "[43] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, and Furu Wei. Image as a foreign language: Beit pretraining for vision and vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19175-19186, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "text", + "content": "[44] Zhongdao Wang, Luming Tang, Xihui Liu, Zhuliang Yao, Shuai Yi, Jing Shao, Junjie Yan, Shengjin Wang, Hongsheng Li, and Xiaogang Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In Proceedings of the IEEE international conference on computer vision, pages 379–387, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "text", + "content": "[45] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16133-16142, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 646, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 679 + ], + "type": "text", + "content": "[46] Mouxing Yang, Zhenyu Huang, and Xi Peng. Robust object re-identification with coupled noisy labels. International Journal of Computer Vision, pages 1–19, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[47] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4710" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492-4501, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 162 + ], + "type": "text", + "content": "[48] Yue Yao, Liang Zheng, Xiaodong Yang, Milind Naphade, and Tom Gedeon. Simulating content consistent vehicle datasets with attribute descent. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 775-791. Springer, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[49] Hongwei Zhang, Guoqing Zhang, Yuhao Chen, and Yuhui Zheng. Global relation-aware contrast learning for unsupervised person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8599–8610, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "type": "text", + "content": "[50] Kevin Zhang and Zhiqiang Shen. i-mae: Are latent representations in masked autoencoders linearly separable? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7740-7749, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[51] Xiao Zhang, Yixiao Ge, Yu Qiao, and Hongsheng Li. Refining pseudo labels with clustering consensus over generations for unsupervised object re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3436–3445, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 363 + ], + "type": "text", + "content": "[52] Aihua Zheng, Xia Sun, Chenglong Li, and Jin Tang. Aware progressive clustering for unsupervised vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(8):11422-11435, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 418 + ], + "type": "text", + "content": "[53] Aihua Zheng, Chaobin Zhang, Chenglong Li, Jin Tang, and Chang Tan. Multi-query vehicle re-identification: Viewpoint-conditioned network, unified dataset and new metric. IEEE Transactions on Image Processing, 32:5948–5960, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 296, + 463 + ], + "type": "text", + "content": "[54] Zhedong Zheng, Tao Ruan, Yunchao Wei, Yi Yang, and Tao Mei. Vehiclenet: Learning robust visual representation for vehicle re-identification. IEEE Transactions on Multimedia, 23:2683–2693, 2020. 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4711" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_content_list.json b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea6be3743d114dbd27d017b2ffb952beff1f68e9 --- /dev/null +++ b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_content_list.json @@ -0,0 +1,1236 @@ +[ + { + "type": "text", + "text": "Verbalized Representation Learning for Interpretable Few-Shot Generalization", + "text_level": 1, + "bbox": [ + 99, + 130, + 897, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng-Fu Yang $^{1*}$ Da Yin $^{1*}$ Wenbo Hu $^{1}$ Heng Ji $^{2}$ Nanyun Peng $^{1}$ Bolei Zhou $^{1}$ Kai-Wei Chang $^{1}$ $^{1}$ University of California, Los Angeles $^{2}$ University of Illinois Urbana-Champaign", + "bbox": [ + 99, + 179, + 916, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cfyang@cs.ucla.edu", + "bbox": [ + 426, + 220, + 589, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2d3c95d96c71efe3946fbc88d9ec5631661e2b12cae9bb15c6c85703bcd3b71.jpg", + "image_caption": [ + "Figure 1. (a) Humans can identify key differences between similar species and recognize common traits within a class, even when appearance varies, and express these insights in concise language. Similarly, our Verbalized Representation Learning (VRL) extracts meaningful features by querying Vision-Language Models (VLMs) to capture inter-class discriminative features and intra-class commonalities. (b) These verbalized features are mapped to numerical vectors through the VLM and can be used by downstream classifiers. Our VRL demonstrates superior performance in few-shot classification compared to prior work, achieving a $24\\%$ absolute gain on iNaturalist and a $14\\%$ on Novel Object Classification, even with a much smaller model." + ], + "image_footnote": [], + "bbox": [ + 99, + 255, + 901, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 489, + 326, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans recognize objects after observing only a few examples, a remarkable capability enabled by their inherent language understanding of the real-world environment. Developing verbalized and interpretable representation can significantly improve model generalization in low-data settings. In this work, we propose Verbalized Representation Learning (VRL), a novel approach for automatically extracting human-interpretable features for object recognition using few-shot data. Our method uniquely captures interclass differences and intra-class commonalities in the form of natural language by employing a Vision-Language Model (VLM) to identify key discriminative features between different classes and shared characteristics within the same class. These verbalized features are then mapped to numeric vectors through the VLM. The resulting feature vectors can be further utilized to train and infer with downstream classifiers. Experimental results show that, at the same model scale, VRL achieves a $24\\%$ absolute improvement over prior state-of-the-art methods while using $95\\%$ less data and a smaller model. Furthermore, compared to human-labeled attributes, the features learned by VRL exhibit a $20\\%$ absolute gain when used for downstream classification tasks. $^{1}$", + "bbox": [ + 88, + 522, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 489, + 643, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans have a remarkable capability to recognize certain objects after seeing only a few examples. As suggested by [33], it is significantly enhanced by inherent language understanding. Language, with its representational strength, serves as a primary resource for conveying knowledge about visual objects. A single, precise description can effectively capture visual distinctions observed across different object categories. For example, in Fig. 1 (a), a human can identify the key difference between two fish species based on the pattern of their heads, even if both share a white stripe around the neck and a yellow body. As a result, we argue that developing verbalized features offers a valuable complement to improve model's generalization under low-resource conditions. In addition, incorporating language into image classification models also increases the interpretability of the visual system, facilitating transparent and reliable decision making and effective model auditing.", + "bbox": [ + 509, + 518, + 906, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prior works [15, 16] have explored integrating language into visual classifiers with a bottleneck of textual attributes, where the model first predicts relevant attributes and subsequently uses them to classify the image. However, these attributes often require human annotations or predefined ontologies. Development of Large-scale Vision-and-Language Models (VLMs) like CLIP [39] and GPT-4V [1] have enabled recent bottleneck methods [12, 32, 38, 48, 50]", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The code is available at: https://github.com/joeyy5588/VRL.", + "bbox": [ + 107, + 886, + 434, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1602", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to generate relevant attributes by leveraging the prior knowledge of VLMs learned during the pre-training stage. However, these attributes are often ungrounded for the testing images and rely on VLM's prior knowledge, resulting in low precision when applied to fine-grained or novel concept recognition where objects are underrepresented in the pre-training datasets.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose Verbalized Representation Learning (VRL) for automatic, human-interpretable feature extraction using only few-shot data. It applies to fine-grained and novel objects and could work with local models such as LLaVA [28]. Specifically, inspired by self-supervised representation learning (SSL) in computer vision, such as SimCLR [8], MoCo [19], SwAV [6], SimSiam [10], BYOL [18], CLIP-Event [29], and SOLO [11], we propose to leverage a VLM to capture the inter-class difference and intra-class commonality and articulate these findings in natural language, as illustrated in Fig. 1. Specifically, we cast the VLM to describe the key difference between two images from different classes, which would preserve the discriminative features and remove the redundant ones, such as the yellow body shared by both species on the left side of Fig. 1 (a). Conversely, we employ the VLM to list the key features that are shared by two images within the same class, extracting features that are robust to intra-class variance, such as the orange facial pattern observed in all of the Thalassoma Pavo. Notably, this process can be applied to any two images, whether from different classes or the same class, which allows us to exponentially scale the number of useful verbalized features with the image samples, enabling effective generalization even in few-shot settings.", + "bbox": [ + 91, + 200, + 482, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consequently, to obtain the numerical feature embedding of an image using our VRL, we employ a VLM to determine whether the image possesses the characteristics described by the verbalized features, as depicted in Fig. 1 (b). Each dimension in the resulting embedding represents a scalar value indicating the presence of the described feature. This approach effectively transforms the verbalized features into numeric representations that can serve as inputs to any classification methods, such as logistic regression, random forest, or MLP classifiers, allowing flexible and robust modeling based on the extracted interpretable features.", + "bbox": [ + 89, + 580, + 480, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on iNaturalist [47] and KikiBouba dataset [2], where the former includes objects that only have subtle differences between classes, while the latter contains novel objects that are nearly not present in the web-scale datasets. Compared to previous state-of-the-art (SoTA) baselines using 70B models, we achieve a $24\\%$ absolute improvement while using $95\\%$ less data and a much smaller model with 7B parameters. Against supervised finetuned baselines, we observe a $15\\%$ absolute improvement. Lastly, when compared to human annotated attributes, the", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attributes learned by our VRL demonstrate a $20\\%$ absolute gain. These results illustrate that the features extracted by VRL not only offer superior effectiveness but also exhibit strong robustness, providing a generalizable solution when adapting to fine-grained or novel classification tasks with limited data.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 189, + 653, + 204 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Natural Language for Image Classification. A common approach for integrating language into visual classifiers involves creating a concept bottleneck [26]. Early bottleneck methods have been extensively applied in few-shot or zero-shot classification models [7, 15-17, 23, 26, 27, 42]. However, these methods typically rely on manually annotated or predefined attributes, which can limit their adaptability to novel classes or fine-grained tasks. Recent advancements in VLMs have enabled researchers to directly sample interpretable descriptive features from these models [12, 32, 38, 45, 48-51]. This approach benefits from incorporating external knowledge embedded in the pretrained datasets. However, it often relies heavily on the prior knowledge encoded within these models, which can lead to the generation of ungrounded features and difficulty generalizing when the target data is not well-represented in the training datasets. In contrast, our method directly learns grounded, verbalized features from the visual data, allowing for automatic interpretable feature extraction that remains adaptable even for novel classes.", + "bbox": [ + 511, + 214, + 903, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpreting Model's Decision Process. There has been a long line of research aiming to improve the interpretability and explainability of deep models. Pioneering methods [3, 4, 13, 24, 43, 52] have tackled this challenge by visualizing learned features, categorizing maximally-activating inputs, and identifying key neurons that drive model decisions. More recently, efforts have been made to interpret model behavior using natural language [20, 21, 36, 44], enabling explanations that are more accessible and editable. However, the aforementioned methods are generally post-hoc, meaning they are applied after the model has been trained with abundant data. While post-hoc explanations can provide valuable insights, they may lack the ability to influence or shape the model's internal representations during training. Instead, our methods embed interpretability directly into the learning process by learning verbalized features. This enables the model to produce inherently explainable and contextually grounded representations without relying solely on retrospective analysis.", + "bbox": [ + 511, + 518, + 903, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Learning Models. Self-supervised learning (SSL) has emerged as a powerful approach in computer vision, enabling models to learn robust feature representations from unlabeled data. Contrastive learning [34] methods, such as SimCLR [8] and MoCo [19], achieve this goal by learning to discriminate between different samples (neg", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1603", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49ec2f97e95200572596d5b59352c13f5bc971403f9eb9444716c8cd482618e6.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 117, + 77, + 387, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3248b61fe0761815c8b7e7d1614d6f8279951fd77cac75e3a250e0104aaf1063.jpg", + "image_caption": [ + "Figure 2. The overview of our Verbalized Representation Learning (VRL) framework. (a) Given $N$ samples per class from $C$ different classes, VRL is able to generate a diverse set (exponentially scaling with $N \\times C$ ) of verbalized features by: 1) extracting key differences between samples from different classes, and 2) identifying commonalities shared among objects within the same class. (b) Given an image, a Vision-and-Language model (VLM) is employed to evaluate whether the image contains the characteristics described by the verbalized features. This process can map a set of verbalized features into numeric representations which can then be used for downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 388, + 77, + 671, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e8c179c428e489c9cafeea97b77c04b3a2ab6cce4a3ef76fcb9dd385cefc34fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 74, + 883, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ative pairs) and augmentations of the same sample (positive pairs). Chen et al. [10] have found that the key function of negative pairs is to prevent the model from learning collapsing features, where models produce constant or trivial outputs. As a result, subsequent works have relieved the need for negative samples by employing techniques like clustering [5, 6], momentum update [18], or stop gradient [10]. These methods focus on learning the underlying shared representations between the augmentations of the same sample. Our approach can be viewed as a variant of supervised contrastive learning [25], where the positive samples are drawn from different images from the same class. However, our method introduces two distinct advantages. First, it is data-efficient, as it does not require large datasets for gradient updates. Second, the expressiveness of verbalized features inherently prevents the model from collapsing to constant outputs, thereby ensuring the resulting features are meaningful and representative.", + "bbox": [ + 91, + 439, + 483, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 724, + 181, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Sec. 3.1, we first introduce the motivation and concept of verbalized representation learning (VRL) and then discuss how VRL automatically extracts interpretable, compact representations with few-shot data. Then, in Sec. 3.2, we outline the process of building a visual classifier using the derived features. Furthermore, we demonstrate that these extracted features are versatile and can be applied to arbitrary classification models, including but not limited to logistic regression, decision trees, and multi-layer perceptron (MLP) classifiers. The overall framework of our VRL", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is presented in Fig. 2.", + "bbox": [ + 511, + 439, + 658, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Verbalized Representation Learning", + "text_level": 1, + "bbox": [ + 511, + 464, + 828, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Humans can recognize objects after seeing only a few examples, a skill boosted by language understanding [33]. Language effectively conveys visual distinctions, even without many extra visual cues. Incorporating language also enhances image classification interpretability, enabling clearer, more reliable decisions. To this end, we propose a method that verbalizes key visual features in natural language. In this section, we focus on identifying which visual features are crucial for few-shot image classification and how to extract and verbalize them effectively, and how to map these verbalized features into vectors which would be later utilized during training and inference.", + "bbox": [ + 511, + 487, + 906, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "What visual features should be verbalized? Inspired by self-supervised representation learning methods, where model learns robust features through contrasting positive and negative pairs [8, 19] or maximizing the similarity of augmented views of the positive samples [6, 10, 18], we propose to verbalize the objective functions of these self-supervised learning methods with the help of VLMs such as LLaVA. Specifically, given a classification task with $C$ classes and $N$ -shot examples per class, each data point is defined as $(x,l)$ , where $x$ represents an image, and $l \\in \\{1,2,\\dots,C\\}$ is the categorical label indicating the class to which the image belongs. Our method leverages two types of paired images: positive pairs, defined as image pairs $(x_{i},x_{j})$ with the same label $(l_{i} = l_{j})$ , and negative pairs, which consist of images $(x_{i},x_{k})$ from different", + "bbox": [ + 511, + 672, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1604", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "classes $(l_i\\neq l_k)$ . Notably, with this pairing strategy, we can form $\\mathbf{C}_2^C\\times \\mathbf{C}_2^N$ distinct negative pairs, and $C\\times \\mathbf{C}_2^N$ positive pairs, which significantly increase the data utilization under the few-shot setting.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by contrastive learning methods [8, 19], for each negative pair sample, our approach emphasizes capturing inter-class differences by tasking the VLM to describe key distinguishing features between images from different classes, denoted as $y_{diff} = \\mathrm{VLM}(\\mathbf{x}_i,\\mathbf{x}_k,q_{diff})$ , where $l_i\\neq l_k$ and $q_{diff}$ denotes the query that captures the visual differences between the two input images $\\mathbf{x}_i$ and $\\mathbf{x}_k$ . For instance, in Fig. 2 (a), the model learns to distinguish two fish species by the coloration and the pattern of the fish. Conversely, for each positive pair, we draw inspirations from negative sample-free SSL methods [6, 10, 18], where we employ the VLM to capture intra-class commonalities, generating descriptions of the key shared features between two images from the same class, i.e., $y_{comm} = \\mathrm{VLM}(\\mathbf{x}_i,\\mathbf{x}_j,q_{comm})$ , where $l_i = l_j$ and $q_{comm}$ denotes the query that captures the visual commonalities between $\\mathbf{x}_i$ and $\\mathbf{x}_j$ . For example, in Fig. 2 (a), VRL identifies that both fish possess irregular orange spots around their face. We include the detailed prompt templates used in the above process in Appendix A.1.", + "bbox": [ + 91, + 154, + 482, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, $y_{diff}$ and $y_{comm}$ learn complimentary features. $y_{diff}$ captures the most discriminative features while filtering out redundant ones, such as the yellow body present in both types of fish, leading to less noisy features. While $y_{comm}$ is robust to the intra-class variance where it learns to identify the fish based on the shared features, like the pattern on its face, rather than the number of stripes on its back. In the later experimental section, we will further verify this assumption and show that these two features can yield the best performance when combined. In addition, as discussed in the earlier paragraph, even with only $N$ -shot samples, our VRL can sample a diverse set of features from $\\mathbf{C}_2^C \\times \\mathbf{C}_2^N$ negative and $C \\times \\mathbf{C}_2^N$ positive pairs, making our method excel under low-resource setting. Moreover, we empirically discover that sampling $C \\times N$ pairs for both negative and positive samples is sufficient to collect a robust set of features.", + "bbox": [ + 89, + 457, + 482, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How to map the verbalized features to numeric feature embeddings? To obtain feature vectors that can be later used to build visual classifiers, we transform the verbalized features into numeric vectors with the help of VLMs such as CLIP [39] and LLaVA [28]. Given an image and a set of verbalized features, which are in the form of language descriptions, we employ a VLM to determine whether the image possesses those features, as illustrated in Fig. 2 (b), resulting in a feature embedding $F$ , i.e., $F = \\mathrm{VLM}(y,x)$ . Each dimension of $F$ is produced by the VLM, indicating the presence of a certain verbalized feature from $y_{diff}$ and $y_{comm}$ or assessing the degree that the image has a certain", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature. Concretely, for generative VLMs like LLaVA, each verbalized feature is mapped to 0 or 1, based on whether the model infers the presence of the feature in the image. For VLMs like CLIP, feature embeddings are derived by calculating the similarity between each verbalized feature and the image using CLIP's visual encoders. This results in a continuous vector indicating the likelihood that the image contains each feature, with an optional similarity threshold to convert it into binary feature vectors.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training and Inference", + "text_level": 1, + "bbox": [ + 511, + 239, + 725, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. Once the numeric feature vectors are generated, they are used as input for training various visual classifiers, including, but not limited to, logistic regression, random forests, Naive Bayes, K-nearest neighbors (KNN), and MLP classifiers. These models learn to predict the class labels based on the interpretable feature vectors. This flexibility enables robust modeling that can be tailored to different applications. Moreover, this approach not only enables the construction of classifiers for novel concepts with few-shot data, but also provides insight into which visual features are most important for decision-making. For instance, we can extract feature importance from logistic regression or visualize decision paths in decision tree classifiers, enhancing the interpretability of our method.", + "bbox": [ + 511, + 262, + 903, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference. During inference, given a testing image, we generate its feature vector by following the approach described in Sec. 3.1. Specifically, we query the VLMs with the learned verbalized features to determine whether the image contains specific features, producing a numeric feature vector. This vector can either be continuous, representing the likelihood of a feature's presence, or binary, indicating the feature's absence or presence. The resulting vector is then passed to the trained classifier to make a prediction. Notably, to enhance the model robustness, we ensemble results from classifiers trained with different algorithms. Combining predictions from multiple classifiers is beneficial because it reduces the likelihood of overfitting to any specific algorithm's biases or weaknesses. The ensemble can use either hard or soft voting: In hard voting, each classifier makes a discrete prediction, and the final prediction is determined by a majority vote. In soft voting, the prediction logits from each classifier are averaged to produce the final output, allowing for more nuanced decision-making.", + "bbox": [ + 511, + 479, + 903, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 782, + 643, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our experiments are designed to answer the following research questions: (i) How well does VRL generalize to tasks requiring fine-grained recognition with few-shot data? (ii) Can VRL effectively adapt to tasks involving novel concepts that were not part of the VLM's pre-training resources, using only few-shot data? (iii) How effective is", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1605", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VRL compared to conventional few-shot adaptation algorithms like LoRA fine-tuning or in-context learning? (iv) Do different types of features extracted by VRL mutually benefit each other, and do the extracted features further improve on top of the other commonly used features? (v) How does the performance of automatically features extracted with VRL compare to human-labeled features?", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 89, + 205, + 382, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on iNaturalist and Kiki-Bouba to validate our method's effectiveness under varied scenarios, with all reported numbers representing classification accuracy (\\%). Although our primary analysis centers on these datasets, additional results for general few-shot classification on mini-ImageNet [41] are provided in the Appendix B.3.", + "bbox": [ + 89, + 220, + 480, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-Grained Species Classification. For fine-grained classification, we utilize the iNaturalist 2021 dataset [47], which comprises a diverse collection of images and annotations contributed by citizen scientists, spanning numerous species of animals, plants, and fungi. Each species in the training split contains between 200 and 300 images, while the validation split includes 10 images per species.", + "bbox": [ + 89, + 333, + 480, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [12], we experiment on images from five different families, each containing five to six species. These families are selected due to their challenging nature: distinguishing between species within the same family requires the model to identify complex features such as shapes and patterns, rather than simple color variations. The families used in our experiments are as follows: Lichen (fungi), Wrasse (fish), Wild rye (grass), Manzanita (berry shrubs), and Bulrush (herbs). For a detailed list of the species, please refer to the appendix. Notably, unlike [12], which utilizes a full training set containing between 200 and 300 images per species, our method uses on only 10 images per species to achieve generalization in a low-resource setting. We will report the baseline performance from their original paper as well as our reproduced results using the same limited data as our method for comparison.", + "bbox": [ + 89, + 439, + 482, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Novel Concept Classification. To test the model's generalizability on objects that have been rarely seen in the pretraining image resources, we evaluate our method using the Kiki-Bouba [2] dataset. The Kiki-Bouba experiment, originally introduced by [40], illustrates that people often associate specific shapes with different sounds. The dataset was constructed by prompting generative models trained to create 3D-rendered images from non-existent, meaningless words [2], providing a unique testbed for evaluating generalization to novel and abstract concepts. Following the approach in [12], we evaluate our method on two distinct splits, each containing five different classes. The first split consists of the classes bamaba, duludu, gaduga, lomulo, nomano, while the second split includes bouba, galaga,", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "kepike, kiki, maluma. In [12], the training set comprises 800 images per class, with 200 images per class for validation. In contrast, we adopt a similar low-resource setting as used for the iNaturalist dataset, with our method accessing only 10 training images per class. Baseline results for both settings will be reported in the table for comparison.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baseline Methods and Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 189, + 901, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our method, we compare it against several baselines, including the CLIP-based approach proposed in [12] and other baseline methods that utilize LLMs, such as LaBo [50], CBD [32] and LMMutate [12], which leverage LLM's pre-trained knowledge to generate attributes. Additionally, since our method leverages LLaVA-OneVision [28] to generate verbalized features, we perform quantitative comparisons with common approaches used to adapt LLaVA for downstream tasks, such as LoRA fine-tuning [22] and in-context learning [14]. We describe the details of each baseline method as follows.", + "bbox": [ + 511, + 209, + 903, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLIP-Based Methods. For CLIP-based baselines, we consider two variants. CLIP Class Name is a naive baseline where CLIP is employed to compute the similarity between the species' scientific name and the images. The best accuracy is reported as the highest value achieved among using the common name, the scientific name, or both names combined. The second variant involves Prompt Tuning with a CLIP encoder, where the text embeddings of specific class-related tokens are optimized through gradient descent. For both baselines, we present results as reported in [12].", + "bbox": [ + 511, + 382, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Language Bottleneck Methods. Classification by Description [32], LaBo [50], and LLM-Mutate [12] generate a list of attributes by prompting LLMs like GPT [35] and Llama [46] with the class name of the object. While this allows them to leverage the prior knowledge learned by LLMs, the generated features often lack grounding in the actual images. LaBo and LLM-Mutate alleviate this issue by using submodular optimization and evolutionary algorithm, respectively, to filter out ungrounded attributes. However, these methods still rely heavily on the capabilities and learned knowledge of the pre-trained LLMs. For LaBo, we adapt the official implementation for our datasets; for LLM-Mutate, we not only report the results from the original paper but also evaluate the method using the same limited data as our approach, ensuring a fair comparison.", + "bbox": [ + 511, + 541, + 905, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaVA-Based Methods. We include two common approaches for adopting LLaVA. The first baseline involves performing LoRA [22] fine-tuning (LLaVA-SFT), which has been demonstrated to be effective in scenarios with limited data. The second baseline is in-context learning [14] (LLaVA-ICL), where one exemplar image for each species or object is included in the prompt to guide the model's predictions. We study which of the LLaVA-based methods and VRL better utilizes LLaVA's capabilities.", + "bbox": [ + 511, + 776, + 905, + 912 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1606", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3684f3a1d776f7db2f4c36517900e6f4c1a47944ebab35108ba68c79016c0c61.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLichenWrasseWild RyeManzanitaBulrushAverage
Zero-Shot Methods
CLIP Class Name [12]23.332.032.026.026.027.86
Full Dataset Methods (200+ Images per Species)
CLIP Prompt Tuning [12]23.320.040.020.020.024.66
Classification by Description [32]30.034.036.028.020.029.60
LLM-Mutate-70B [12] (1-prompt)31.624.044.040.022.032.32
LLM-Mutate-70B [12] (10-prompt)48.344.058.058.042.050.06
Few-Shot Methods (10 Images per Species)
LLM-Mutate-7B† (10-prompt)35.048.038.044.026.038.20
LLM-Mutate-70B† (10-prompt)46.644.046.044.040.044.13
LaBo-7B† [50]45.048.044.042.036.043.00
LaBo-70B† [50]56.646.064.052.040.051.72
LLaVA-ICL-7B16.628.022.018.030.022.92
LLaVA-SFT-7B41.650.058.042.028.043.92
LLaVA-VRL-7B (Ours)58.348.074.066.046.058.46
LLaVA-VRL-72B (Ours)71.672.074.056.066.067.92
", + "bbox": [ + 133, + 75, + 862, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of classification accuracy (%) across different methods for fine-grained classification on iNaturalist. The table presents results for zero-shot, full-dataset (200+ images per species), and few-shot (10 images per species). Results marked with † denote values reproduced using the official implementation but restricted to the same few-shot data as our method.", + "bbox": [ + 89, + 351, + 903, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our Method. We implement VRL with LLaVA-OneVision for capturing visual difference and commonality features. We also select LLaVA-OneVision as the feature mapping model that assists in converting an image into a feature vector, and the base model for all the LLaVA-based baseline methods. For the few-shot training stage, we experiment with various classification methods including logistic regression or MLP. Unless otherwise specified, we utilize both the difference features $F_{diff}$ and commonality features $F_{comm}$ . We report the best performance achieved across different classifiers, selecting the optimal results from a single classifier.", + "bbox": [ + 88, + 409, + 480, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Fine-Grained Species Classification", + "text_level": 1, + "bbox": [ + 89, + 594, + 398, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the results for fine-grained species classification on the iNaturalist dataset in Table 1. We find that compared to previous SoTA methods, even with $95\\%$ less data (10 images per species compared to $200+$ images per species), and using a significantly smaller model (7B vs. 70B parameters), our method can already surpass the previous SoTA by $8\\%$ . When we increase the scale of our method to a comparable 72B model, we can further increase the improvements to nearly $18\\%$ . Moreover, when baseline methods are limited to the same data availability as our approach, our 7B and 72B models achieve improvements of $20\\%$ and $23\\%$ , respectively. These results highlight the advantage of using grounded descriptions to capture the subtle visual cues, especially when the model needs to discover visual nuances to discriminate species within the same family.", + "bbox": [ + 91, + 613, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare VRL against LLaVA-based baselines. Using the same backbone LLaVA model, our method showcases a $15\\%$ gain to LLaVA-SFT and a $36\\%$ advantage over the in-context learning method, LLaVA-ICL. In addition,", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VRL is easier to be scaled up than the baselines. For instance, VRL with the 72B model requires 8 GPUs with 48 GB of RAM, while fine-tuning the 13B model with LoRA already requires 8 GPUs with 80 GB of RAM. For in-context learning, including image examples for every category in the prompt significantly increases the memory usage of the method, making it infeasible to perform in-context learning on the 72B LLaVA with the 8 GPUs (48 GB). The model performance may also be constrained by the context window of LLaVA model. These results and analysis further validate the scalability and robustness of VRL comparing to those existing adaptation algorithms.", + "bbox": [ + 511, + 409, + 906, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Novel Concept Classification", + "text_level": 1, + "bbox": [ + 511, + 598, + 769, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For novel concept classification, we report the results on the Kiki-Bouba dataset in Table 2. Similar to the performance trend observed in the previous experiment, our method surpasses the previous SoTA by $14\\%$ with a smaller model, but with $99\\%$ less data (10 images per species compared to $800+$ images per species). Surprisingly, we observe that increasing the number of model parameters does not lead to better performance on this task. One possible explanation for this is that, rather than relying on prior knowledge learned during pre-training, the model needs to focus on learning new, critical features specific to these novel objects with unfamiliar shapes and patterns. Since the objects in this task are unseen during LLM training, existing models struggle to generate relevant attributes and fail to ground visual differences necessary for distinguishing objects from different classes. This hypothesis is further supported by the larger performance gap between our method and those that rely on LLM's pre-trained knowledge [32]. On this dataset, our method exceeds their performance by $50\\%$ , compared", + "bbox": [ + 511, + 614, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1607", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6a392247c41a85acfd5f385243260cd85a4e8666979dc275961a58104147880.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Accuracy (%) across different methods for novel concept classification on Kiki-Kouba dataset. v1 and v2 indicates two different splits described in Sec. 4.1. The table presents results for zero-shot, full-dataset, and few-shot. Results marked with $\\dagger$ denote values reproduced using the official implementation but restricted to the same few-shot data as our method.", + "to a $37\\%$ advantage on the previous dataset." + ], + "table_body": "
Methodv1v2Avg.
Zero-Shot Methods
CLIP Class Name38.738.838.75
Full Dataset Methods (800+ Images per Object)
CLIP Prompt Tuning16.755.636.15
Classification by Description28.836.832.80
LLM-Mutate-70B (1-prompt)50.347.849.05
LLM-Mutate-70B (10-prompt)79.259.469.30
Few-Shot Methods (10 Images per Object)
LLM-Mutate-7B† (10-prompt)63.359.261.25
LLM-Mutate-70B† (10-prompt)67.359.263.25
LLaVA-ICL-7B24.527.225.85
LLaVA-SFT-7B72.150.661.35
LLaVA-VRL-7B (Ours)89.476.683.00
LLaVA-VRL-72B (Ours)89.174.781.90
", + "bbox": [ + 102, + 75, + 470, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Additional Analysis", + "text_level": 1, + "bbox": [ + 89, + 438, + 277, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fusing Features Learned by VRL and Commonly Used Features. Table 3 summarizes the classification accuracy on iNaturalist achieved by fusing different types of visual features. Specifically, we study the two feature vectors: $F_{diff}$ and $F_{comm}$ learned via our VRL, which focusing on capturing regional nuances of the object. $F_{diff}$ captures the inter-class difference while $F_{comm}$ represents the shared features within the same class. We also consider the image features directly encoded by the vision encoder of CLIP, which often encode high-level semantics such as the object type, context, and other conceptual associations. As discussed in Sec. 3.1, these features are robust to different kinds of variance and could be benefited when combined for downstream tasks. This is evident from the $2\\%$ improvements when concatenating $F_{diff}$ and $F_{comm}$ , and an additional $6\\%$ gain after including the CLIP feature.", + "bbox": [ + 88, + 460, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also discover that ensemble classifiers trained separately on each feature type are more effective than simply concatenating all features to build a single classifier. This is primarily because these features are heterogeneous. For example, $F_{diff}$ and $F_{comm}$ are generally binary, with each element indicating the presence of a specific verbalized feature. Therefore, they tend to perform best when used with logistic regression models. Meanwhile, the features encoded by CLIP are continuous, which makes them more suitable for mapping to class labels through MLP models. Empirically, we demonstrate that the best configuration of ensembling can further improves the performance by $3\\%$ comparing to feature concatenation. We find that with three", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/20151383ba08af3505914a9043db02847917130b68bf3ed42a123d5afe15e577.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Accuracy $(\\%)$ when incorporating feature vectors learned from different methods. $F_{diff}$ and $F_{comm}$ denote the difference and commonality feature vectors learned from VRL. CLIP refers to the image features encoded by CLIP visual encoder. We compare two approaches: concatenating all features to train a single classifier and ensembling classifiers trained separately on each feature type. The visual classifiers include Logistic Regression (LR) and Multi-Layer Perceptron (MLP) models." + ], + "table_body": "
MethodFdiffFcommCLIPAvg.
Concat✓65.26
Concat✓58.32
Concat✓63.26
Concat✓✓67.92
Concat✓✓✓73.92
Ensemble (hard)LRLRLR68.32
Ensemble (hard)MLPMLPMLP74.26
Ensemble (hard)LRLRMLP71.60
Ensemble (soft)LRLRLR67.06
Ensemble (soft)MLPMLPMLP74.92
Ensemble (soft)LRLRMLP76.52
", + "bbox": [ + 535, + 75, + 883, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classifiers, hard-voting is prone to high variance, while soft-voting is more robust since it leverages the confidence of each classifier by averaging their probabilities. These experiments showcase that our approach is flexible and can be seamlessly integrated with existing representations to build a robust classification model through ensembling.", + "bbox": [ + 511, + 363, + 903, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Human-Labeled Features. To evaluate the performance of automatically extracted features compared to human-labeled features, we conduct experiments on the Kiki-Bouba dataset using the human-annotated attributes provided by [12]. From Table 4, we observe that VRL extracted features outperform human-labeled attributes across all configurations. Our best configuration, which combines both difference feature vector $F_{diff}$ and commonality feature vector $F_{comm}$ , surpasses the human-derived features by $20\\%$ . This underscores the effectiveness of our automatic feature extraction approach in capturing relevant and distinguishing characteristics for classification tasks. These results also suggest that, in scenarios with limited resources or new visual concepts, we can automate the process of feature extraction, reducing the need for manual human annotation.", + "bbox": [ + 511, + 458, + 906, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classifier Choices and Feature Mapping Models. We investigate the model performance using different choices for visual classifiers and feature mapping models. From Table 5, we notice that LLaVA, as a feature mapping model for obtaining binary feature vectors, achieves the best overall performance when paired with the logistic regression.", + "bbox": [ + 511, + 704, + 903, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.1, we can also leverage CLIP as a feature mapping model to convert images into features. When predicting, one can either set a threshold to determine if the image possesses a certain attribute, resulting in a binary feature, or preserve the similarity between the image and each attribute to form continuous features. Although the performance of using CLIP features is slightly lower", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1608", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fad7c3a0303193d993db182e43316478a00645c8d5290978d08932ea8001dd41.jpg", + "image_caption": [ + "Figure 3. Qualitative examples of the features extracted by VRL. We highlight the key attributes in bold. (a) Verbalized features extracted by comparing images from different classes. (b) Verbalized features extracted by comparing the images within the same class." + ], + "image_footnote": [], + "bbox": [ + 102, + 84, + 504, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1573def6c59fdf90683a5326ff1b74994857dd5d7d3d022787b08c1ecec77c67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 85, + 897, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1b9de4cc0a37101bf70c49045c36cec338f3317449aaca6181f5c8f9c8774723.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodv1v2Avg.
Human73.852.563.15
VRL-Fdiff88.475.281.80
VRL-Fcomm89.274.081.60
VRL-all89.476.683.00
", + "bbox": [ + 102, + 338, + 472, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compared to LLaVA, CLIP offers a notable advantage in terms of inference speed. This is because CLIP supports batch-wise operations when computing similarities between a set of images and a set of verbalized features, which offers our method the flexibility to scale with a larger dataset or an increasing number of object classes. It is worth noting that using CLIP as the feature mapping model still allows our method to surpass previous state-of-the-art results by significant margins, achieving improvements of $8\\%$ and $12\\%$ on the iNaturalist and Kiki-Bouba datasets, respectively. This highlights the universal applicability of VRL-extracted features, regardless of the feature mapping approach used.", + "bbox": [ + 88, + 476, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 665, + 284, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 presents qualitative examples of what verbalized features are extracted from image pairs. We list some of the features that have top-5 feature importance in our trained logistic regression model. For inter-class differences (Fig. 3 (a)), verbalized features are used to capture distinguishing characteristics between different species. For example, the Lichen (Fungi) category is distinguished based on the structure of the thallus, where one image describes a flat, crustlike thallus, and the other describes a complex, branch-like thallus. Similarly, in the Wild Rye (Grass) category, the differences between a slender stem with a feathery inflorescence and a more robust stem with a dense, elongated seed head are verbalized. For intra-class commonality (Fig. 3 (b)), verbalized features are employed to capture shared", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9de782fbcbe3b0c63f62a1d5eb736c94f62b2b7c5be876a7caa960fc35a38c04.jpg", + "table_caption": [ + "Table 4. Accuracy (%) of using human-labeled attributes and VRL extracted features on Kiki-Bouba dataset. Note that v1 and v2 indicates two different splits described in Sec. 4.1." + ], + "table_footnote": [], + "table_body": "
MethodRFLRMLP
iNaturalist (Prior Works: 50.1)
LLaVA64.967.466.0
CLIP (binary)52.258.055.2
CLIP (continuous)52.853.660.6
Kiki-Bouba (Prior Works: 69.3)
LLaVA83.6586.6585.65
CLIP (binary)79.2581.5081.10
CLIP (continuous)77.0577.1081.85
", + "bbox": [ + 545, + 338, + 874, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Accuracy (\\%) of different classifier choices (columns) and models used to map verbalized features to numeric vectors (rows). LLaVA and CLIP are utilized as mapping models to map an image to a feature vector.", + "bbox": [ + 511, + 474, + 906, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "traits within the same species. For example, the leaves of Manzanita (Berry) are consistently described as “ovate with a pointed tip and slightly serrated edges,” while the inflorescences of Bulrush (Herb) are uniformly described as having a “delicate, feathery appearance.” These verbalized features are used to guide the decision process by helping the model focus on important distinguishing features or shared traits within each class. Due to page limit, see Appendix B.5 for human study on the learned verbalized features.", + "bbox": [ + 511, + 537, + 906, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose Verbalized Representation Learning (VRL), which enables automatic interpretable feature extraction with few-shot samples. By leveraging VLMs, VRL generates verbalized features that capture both inter-class differences and intra-class commonalities. Our method not only enhances the model's adaptability with limited data but also provides transparency in the decision-making process, enabling easier interpretation of the features that influence predictions. Our experiments show that VRL outperforms prior approaches, achieving superior results while using significantly less data. This includes tasks like fine-grained recognition and novel concept adaptation, demonstrating its potential for real-world applications.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1609", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank anonymous reviewers and other members of UCLA-NLP+ group for their helpful comments. This work was partially supported by U.S. DARPA ECOLE Program No. #HR00112390060, ONR grant N00014-23-1-2780, DARPA ANSR program FA8750-23-2-0004, and Apple. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 314, + 187, + 329 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Morris Alper and Hadar Averbuch-Elor. Kiki or bouba? sound symbolism in vision-and-language models. Advances in Neural Information Processing Systems, 36, 2024. 2, 5", + "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2", + "[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2", + "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3", + "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 3, 4", + "[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 4", + "[9] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. arXiv preprint arXiv:2303.14123, 2023. 4" + ], + "bbox": [ + 99, + 338, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2, 3, 4", + "[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. Solo: A single transformer for scalable vision-language modeling. In Transactions on Machine Learning Research, 2024. 2", + "[12] Mia Chiquier, Utkarsh Mall, and Carl Vondrick. Evolving interpretable visual classifiers with large language models. arXiv preprint arXiv:2404.09941, 2024. 1, 2, 5, 6, 7", + "[13] Fahim Dalvi, Nadir Durrani, Hassan Sajjad, Yonatan Belinkov, Anthony Bau, and James Glass. What is one grain of sand in the desert? analyzing individual neurons in deep nlp models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6309-6317, 2019. 2", + "[14] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5", + "[15] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 1, 2", + "[16] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. Advances in neural information processing systems, 20, 2007. 1", + "[17] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 2", + "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 4", + "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 4", + "[20] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In Proceedings of the European conference on computer vision (ECCV), pages 264-279, 2018. 2", + "[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2", + "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5", + "[23] Shaoli Huang, Zhe Xu, Dacheng Tao, and Ya Zhang. Part-stacked cnn for fine-grained visual categorization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1173-1182, 2016. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1610", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Andrej Karpathy, Justin Johnson, and Li Fei-Fei. Visualizing and understanding recurrent networks. arXiv preprint arXiv:1506.02078, 2015. 2", + "[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3", + "[26] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 2", + "[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 2", + "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 4, 5", + "[29] Manling Li, Ruochen Xu, Shuohang Wang, Luowei Zhou, Xudong Lin, Chenguang Zhu, Michael Zeng, Heng Ji, and Shih-Fu Chang. Clip-event: Connecting text and images with event structures. In Proc. Conference on Computer Vision and Pattern Recognition (CVPR2022), 2022. 2", + "[30] Fan Liu, Wenwen Cai, Jian Huo, Chuanyi Zhang, De long Chen, and Jun Zhou. Making large vision language models to be good few-shot learners. arXiv preprint arXiv:2408.11297, 2024. 4", + "[31] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. arXiv preprint arXiv:2408.12469, 2024. 4", + "[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 1, 2, 5, 6", + "[33] Eduardo Navarrete, Michele Miozzo, and Francesca Peressotti. Language can shape the perception of oriented objects. Scientific reports, 10(1):8409, 2020. 1, 3", + "[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2", + "[35] OpenAI. ChatGPT. 2022. 5", + "[36] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8779-8788, 2018. 2", + "[37] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 3", + "[38] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 1, 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4", + "[40] Vilayanur S Ramachandran and Edward M Hubbard. Synaesthesia-a window into perception, thought and language. Journal of consciousness studies, 8(12):3-34, 2001. 5", + "[41] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5", + "[42] Bernardino Romera-Paredes and Philip Torr. An embarrassingly simple approach to zero-shot learning. In International conference on machine learning, pages 2152-2161. PMLR, 2015. 2", + "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2", + "[44] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[45] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2", + "[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5", + "[47] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5", + "[48] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3090-3100, 2023. 1, 2", + "[49] Cheng-Fu Yang, Yao-Hung Hubert Tsai, Wan-Cyuan Fan, Russ R Salakhutdinov, Louis-Philippe Morency, and Frank Wang. Paraphrasing is all you need for novel object captioning. Advances in Neural Information Processing Systems, 35:6492–6504, 2022." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1611", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 5, 6", + "[51] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2", + "[52] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2", + "[53] Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. Sglang: Efficient execution of structured language model programs. arXiv preprint arXiv:2312.07104, 2023. 1" + ], + "bbox": [ + 91, + 92, + 485, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1612", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_model.json b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..92b7e1883e631063f492b7e4043709ace3a71d84 --- /dev/null +++ b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_model.json @@ -0,0 +1,1828 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.898, + 0.153 + ], + "angle": 0, + "content": "Verbalized Representation Learning for Interpretable Few-Shot Generalization" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.918, + 0.219 + ], + "angle": 0, + "content": "Cheng-Fu Yang\\(^{1*}\\) Da Yin\\(^{1*}\\) Wenbo Hu\\(^{1}\\) Heng Ji\\(^{2}\\) Nanyun Peng\\(^{1}\\) Bolei Zhou\\(^{1}\\) Kai-Wei Chang\\(^{1}\\) \n\\(^{1}\\)University of California, Los Angeles \\(^{2}\\)University of Illinois Urbana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.221, + 0.591, + 0.234 + ], + "angle": 0, + "content": "cfyang@cs.ucla.edu" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.256, + 0.902, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.393, + 0.907, + 0.477 + ], + "angle": 0, + "content": "Figure 1. (a) Humans can identify key differences between similar species and recognize common traits within a class, even when appearance varies, and express these insights in concise language. Similarly, our Verbalized Representation Learning (VRL) extracts meaningful features by querying Vision-Language Models (VLMs) to capture inter-class discriminative features and intra-class commonalities. (b) These verbalized features are mapped to numerical vectors through the VLM and can be used by downstream classifiers. Our VRL demonstrates superior performance in few-shot classification compared to prior work, achieving a \\(24\\%\\) absolute gain on iNaturalist and a \\(14\\%\\) on Novel Object Classification, even with a much smaller model." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.49, + 0.327, + 0.506 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.486, + 0.871 + ], + "angle": 0, + "content": "Humans recognize objects after observing only a few examples, a remarkable capability enabled by their inherent language understanding of the real-world environment. Developing verbalized and interpretable representation can significantly improve model generalization in low-data settings. In this work, we propose Verbalized Representation Learning (VRL), a novel approach for automatically extracting human-interpretable features for object recognition using few-shot data. Our method uniquely captures interclass differences and intra-class commonalities in the form of natural language by employing a Vision-Language Model (VLM) to identify key discriminative features between different classes and shared characteristics within the same class. These verbalized features are then mapped to numeric vectors through the VLM. The resulting feature vectors can be further utilized to train and infer with downstream classifiers. Experimental results show that, at the same model scale, VRL achieves a \\(24\\%\\) absolute improvement over prior state-of-the-art methods while using \\(95\\%\\) less data and a smaller model. Furthermore, compared to human-labeled attributes, the features learned by VRL exhibit a \\(20\\%\\) absolute gain when used for downstream classification tasks.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.49, + 0.645, + 0.505 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.519, + 0.907, + 0.776 + ], + "angle": 0, + "content": "Humans have a remarkable capability to recognize certain objects after seeing only a few examples. As suggested by [33], it is significantly enhanced by inherent language understanding. Language, with its representational strength, serves as a primary resource for conveying knowledge about visual objects. A single, precise description can effectively capture visual distinctions observed across different object categories. For example, in Fig. 1 (a), a human can identify the key difference between two fish species based on the pattern of their heads, even if both share a white stripe around the neck and a yellow body. As a result, we argue that developing verbalized features offers a valuable complement to improve model's generalization under low-resource conditions. In addition, incorporating language into image classification models also increases the interpretability of the visual system, facilitating transparent and reliable decision making and effective model auditing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Prior works [15, 16] have explored integrating language into visual classifiers with a bottleneck of textual attributes, where the model first predicts relevant attributes and subsequently uses them to classify the image. However, these attributes often require human annotations or predefined ontologies. Development of Large-scale Vision-and-Language Models (VLMs) like CLIP [39] and GPT-4V [1] have enabled recent bottleneck methods [12, 32, 38, 48, 50]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.436, + 0.901 + ], + "angle": 0, + "content": "1The code is available at: https://github.com/joeyy5588/VRL." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "to generate relevant attributes by leveraging the prior knowledge of VLMs learned during the pre-training stage. However, these attributes are often ungrounded for the testing images and rely on VLM's prior knowledge, resulting in low precision when applied to fine-grained or novel concept recognition where objects are underrepresented in the pre-training datasets." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.483, + 0.578 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose Verbalized Representation Learning (VRL) for automatic, human-interpretable feature extraction using only few-shot data. It applies to fine-grained and novel objects and could work with local models such as LLaVA [28]. Specifically, inspired by self-supervised representation learning (SSL) in computer vision, such as SimCLR [8], MoCo [19], SwAV [6], SimSiam [10], BYOL [18], CLIP-Event [29], and SOLO [11], we propose to leverage a VLM to capture the inter-class difference and intra-class commonality and articulate these findings in natural language, as illustrated in Fig. 1. Specifically, we cast the VLM to describe the key difference between two images from different classes, which would preserve the discriminative features and remove the redundant ones, such as the yellow body shared by both species on the left side of Fig. 1 (a). Conversely, we employ the VLM to list the key features that are shared by two images within the same class, extracting features that are robust to intra-class variance, such as the orange facial pattern observed in all of the Thalassoma Pavo. Notably, this process can be applied to any two images, whether from different classes or the same class, which allows us to exponentially scale the number of useful verbalized features with the image samples, enabling effective generalization even in few-shot settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.482, + 0.746 + ], + "angle": 0, + "content": "Consequently, to obtain the numerical feature embedding of an image using our VRL, we employ a VLM to determine whether the image possesses the characteristics described by the verbalized features, as depicted in Fig. 1 (b). Each dimension in the resulting embedding represents a scalar value indicating the presence of the described feature. This approach effectively transforms the verbalized features into numeric representations that can serve as inputs to any classification methods, such as logistic regression, random forest, or MLP classifiers, allowing flexible and robust modeling based on the extracted interpretable features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments on iNaturalist [47] and KikiBouba dataset [2], where the former includes objects that only have subtle differences between classes, while the latter contains novel objects that are nearly not present in the web-scale datasets. Compared to previous state-of-the-art (SoTA) baselines using 70B models, we achieve a \\(24\\%\\) absolute improvement while using \\(95\\%\\) less data and a much smaller model with 7B parameters. Against supervised finetuned baselines, we observe a \\(15\\%\\) absolute improvement. Lastly, when compared to human annotated attributes, the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "attributes learned by our VRL demonstrate a \\(20\\%\\) absolute gain. These results illustrate that the features extracted by VRL not only offer superior effectiveness but also exhibit strong robustness, providing a generalizable solution when adapting to fine-grained or novel classification tasks with limited data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.19, + 0.655, + 0.205 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.518 + ], + "angle": 0, + "content": "Natural Language for Image Classification. A common approach for integrating language into visual classifiers involves creating a concept bottleneck [26]. Early bottleneck methods have been extensively applied in few-shot or zero-shot classification models [7, 15-17, 23, 26, 27, 42]. However, these methods typically rely on manually annotated or predefined attributes, which can limit their adaptability to novel classes or fine-grained tasks. Recent advancements in VLMs have enabled researchers to directly sample interpretable descriptive features from these models [12, 32, 38, 45, 48-51]. This approach benefits from incorporating external knowledge embedded in the pretrained datasets. However, it often relies heavily on the prior knowledge encoded within these models, which can lead to the generation of ungrounded features and difficulty generalizing when the target data is not well-represented in the training datasets. In contrast, our method directly learns grounded, verbalized features from the visual data, allowing for automatic interpretable feature extraction that remains adaptable even for novel classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.52, + 0.905, + 0.807 + ], + "angle": 0, + "content": "Interpreting Model's Decision Process. There has been a long line of research aiming to improve the interpretability and explainability of deep models. Pioneering methods [3, 4, 13, 24, 43, 52] have tackled this challenge by visualizing learned features, categorizing maximally-activating inputs, and identifying key neurons that drive model decisions. More recently, efforts have been made to interpret model behavior using natural language [20, 21, 36, 44], enabling explanations that are more accessible and editable. However, the aforementioned methods are generally post-hoc, meaning they are applied after the model has been trained with abundant data. While post-hoc explanations can provide valuable insights, they may lack the ability to influence or shape the model's internal representations during training. Instead, our methods embed interpretability directly into the learning process by learning verbalized features. This enables the model to produce inherently explainable and contextually grounded representations without relying solely on retrospective analysis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Self-Supervised Learning Models. Self-supervised learning (SSL) has emerged as a powerful approach in computer vision, enabling models to learn robust feature representations from unlabeled data. Contrastive learning [34] methods, such as SimCLR [8] and MoCo [19], achieve this goal by learning to discriminate between different samples (neg" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.078, + 0.388, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.078, + 0.673, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.336, + 0.397, + 0.346 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.075, + 0.885, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.358, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. The overview of our Verbalized Representation Learning (VRL) framework. (a) Given \\(N\\) samples per class from \\(C\\) different classes, VRL is able to generate a diverse set (exponentially scaling with \\(N \\times C\\)) of verbalized features by: 1) extracting key differences between samples from different classes, and 2) identifying commonalities shared among objects within the same class. (b) Given an image, a Vision-and-Language model (VLM) is employed to evaluate whether the image contains the characteristics described by the verbalized features. This process can map a set of verbalized features into numeric representations which can then be used for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.44, + 0.485, + 0.712 + ], + "angle": 0, + "content": "ative pairs) and augmentations of the same sample (positive pairs). Chen et al. [10] have found that the key function of negative pairs is to prevent the model from learning collapsing features, where models produce constant or trivial outputs. As a result, subsequent works have relieved the need for negative samples by employing techniques like clustering [5, 6], momentum update [18], or stop gradient [10]. These methods focus on learning the underlying shared representations between the augmentations of the same sample. Our approach can be viewed as a variant of supervised contrastive learning [25], where the positive samples are drawn from different images from the same class. However, our method introduces two distinct advantages. First, it is data-efficient, as it does not require large datasets for gradient updates. Second, the expressiveness of verbalized features inherently prevents the model from collapsing to constant outputs, thereby ensuring the resulting features are meaningful and representative." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.182, + 0.74 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Sec. 3.1, we first introduce the motivation and concept of verbalized representation learning (VRL) and then discuss how VRL automatically extracts interpretable, compact representations with few-shot data. Then, in Sec. 3.2, we outline the process of building a visual classifier using the derived features. Furthermore, we demonstrate that these extracted features are versatile and can be applied to arbitrary classification models, including but not limited to logistic regression, decision trees, and multi-layer perceptron (MLP) classifiers. The overall framework of our VRL" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.66, + 0.456 + ], + "angle": 0, + "content": "is presented in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.829, + 0.483 + ], + "angle": 0, + "content": "3.1. Verbalized Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.67 + ], + "angle": 0, + "content": "Humans can recognize objects after seeing only a few examples, a skill boosted by language understanding [33]. Language effectively conveys visual distinctions, even without many extra visual cues. Incorporating language also enhances image classification interpretability, enabling clearer, more reliable decisions. To this end, we propose a method that verbalizes key visual features in natural language. In this section, we focus on identifying which visual features are crucial for few-shot image classification and how to extract and verbalize them effectively, and how to map these verbalized features into vectors which would be later utilized during training and inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "What visual features should be verbalized? Inspired by self-supervised representation learning methods, where model learns robust features through contrasting positive and negative pairs [8, 19] or maximizing the similarity of augmented views of the positive samples [6, 10, 18], we propose to verbalize the objective functions of these self-supervised learning methods with the help of VLMs such as LLaVA. Specifically, given a classification task with \\( C \\) classes and \\( N \\)-shot examples per class, each data point is defined as \\( (x,l) \\), where \\( x \\) represents an image, and \\( l \\in \\{1,2,\\dots,C\\} \\) is the categorical label indicating the class to which the image belongs. Our method leverages two types of paired images: positive pairs, defined as image pairs \\( (x_{i},x_{j}) \\) with the same label \\( (l_{i} = l_{j}) \\), and negative pairs, which consist of images \\( (x_{i},x_{k}) \\) from different" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1604" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "classes \\((l_i\\neq l_k)\\). Notably, with this pairing strategy, we can form \\(\\mathbf{C}_2^C\\times \\mathbf{C}_2^N\\) distinct negative pairs, and \\(C\\times \\mathbf{C}_2^N\\) positive pairs, which significantly increase the data utilization under the few-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.155, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Inspired by contrastive learning methods [8, 19], for each negative pair sample, our approach emphasizes capturing inter-class differences by tasking the VLM to describe key distinguishing features between images from different classes, denoted as \\( y_{diff} = \\mathrm{VLM}(\\mathbf{x}_i,\\mathbf{x}_k,q_{diff}) \\), where \\( l_i\\neq l_k \\) and \\( q_{diff} \\) denotes the query that captures the visual differences between the two input images \\( \\mathbf{x}_i \\) and \\( \\mathbf{x}_k \\). For instance, in Fig. 2 (a), the model learns to distinguish two fish species by the coloration and the pattern of the fish. Conversely, for each positive pair, we draw inspirations from negative sample-free SSL methods [6, 10, 18], where we employ the VLM to capture intra-class commonalities, generating descriptions of the key shared features between two images from the same class, i.e., \\( y_{comm} = \\mathrm{VLM}(\\mathbf{x}_i,\\mathbf{x}_j,q_{comm}) \\), where \\( l_i = l_j \\) and \\( q_{comm} \\) denotes the query that captures the visual commonalities between \\( \\mathbf{x}_i \\) and \\( \\mathbf{x}_j \\). For example, in Fig. 2 (a), VRL identifies that both fish possess irregular orange spots around their face. We include the detailed prompt templates used in the above process in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Intuitively, \\( y_{diff} \\) and \\( y_{comm} \\) learn complimentary features. \\( y_{diff} \\) captures the most discriminative features while filtering out redundant ones, such as the yellow body present in both types of fish, leading to less noisy features. While \\( y_{comm} \\) is robust to the intra-class variance where it learns to identify the fish based on the shared features, like the pattern on its face, rather than the number of stripes on its back. In the later experimental section, we will further verify this assumption and show that these two features can yield the best performance when combined. In addition, as discussed in the earlier paragraph, even with only \\( N \\)-shot samples, our VRL can sample a diverse set of features from \\( \\mathbf{C}_2^C \\times \\mathbf{C}_2^N \\) negative and \\( C \\times \\mathbf{C}_2^N \\) positive pairs, making our method excel under low-resource setting. Moreover, we empirically discover that sampling \\( C \\times N \\) pairs for both negative and positive samples is sufficient to collect a robust set of features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "How to map the verbalized features to numeric feature embeddings? To obtain feature vectors that can be later used to build visual classifiers, we transform the verbalized features into numeric vectors with the help of VLMs such as CLIP [39] and LLaVA [28]. Given an image and a set of verbalized features, which are in the form of language descriptions, we employ a VLM to determine whether the image possesses those features, as illustrated in Fig. 2 (b), resulting in a feature embedding \\( F \\), i.e., \\( F = \\mathrm{VLM}(y,x) \\). Each dimension of \\( F \\) is produced by the VLM, indicating the presence of a certain verbalized feature from \\( y_{diff} \\) and \\( y_{comm} \\) or assessing the degree that the image has a certain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "feature. Concretely, for generative VLMs like LLaVA, each verbalized feature is mapped to 0 or 1, based on whether the model infers the presence of the feature in the image. For VLMs like CLIP, feature embeddings are derived by calculating the similarity between each verbalized feature and the image using CLIP's visual encoders. This results in a continuous vector indicating the likelihood that the image contains each feature, with an optional similarity threshold to convert it into binary feature vectors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.24, + 0.727, + 0.256 + ], + "angle": 0, + "content": "3.2. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.905, + 0.475 + ], + "angle": 0, + "content": "Training. Once the numeric feature vectors are generated, they are used as input for training various visual classifiers, including, but not limited to, logistic regression, random forests, Naive Bayes, K-nearest neighbors (KNN), and MLP classifiers. These models learn to predict the class labels based on the interpretable feature vectors. This flexibility enables robust modeling that can be tailored to different applications. Moreover, this approach not only enables the construction of classifiers for novel concepts with few-shot data, but also provides insight into which visual features are most important for decision-making. For instance, we can extract feature importance from logistic regression or visualize decision paths in decision tree classifiers, enhancing the interpretability of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.905, + 0.768 + ], + "angle": 0, + "content": "Inference. During inference, given a testing image, we generate its feature vector by following the approach described in Sec. 3.1. Specifically, we query the VLMs with the learned verbalized features to determine whether the image contains specific features, producing a numeric feature vector. This vector can either be continuous, representing the likelihood of a feature's presence, or binary, indicating the feature's absence or presence. The resulting vector is then passed to the trained classifier to make a prediction. Notably, to enhance the model robustness, we ensemble results from classifiers trained with different algorithms. Combining predictions from multiple classifiers is beneficial because it reduces the likelihood of overfitting to any specific algorithm's biases or weaknesses. The ensemble can use either hard or soft voting: In hard voting, each classifier makes a discrete prediction, and the final prediction is determined by a majority vote. In soft voting, the prediction logits from each classifier are averaged to produce the final output, allowing for more nuanced decision-making." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.645, + 0.801 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our experiments are designed to answer the following research questions: (i) How well does VRL generalize to tasks requiring fine-grained recognition with few-shot data? (ii) Can VRL effectively adapt to tasks involving novel concepts that were not part of the VLM's pre-training resources, using only few-shot data? (iii) How effective is" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1605" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "VRL compared to conventional few-shot adaptation algorithms like LoRA fine-tuning or in-context learning? (iv) Do different types of features extracted by VRL mutually benefit each other, and do the extracted features further improve on top of the other commonly used features? (v) How does the performance of automatically features extracted with VRL compare to human-labeled features?" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.206, + 0.383, + 0.219 + ], + "angle": 0, + "content": "4.1. Dataset and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.482, + 0.327 + ], + "angle": 0, + "content": "We conduct experiments on iNaturalist and Kiki-Bouba to validate our method's effectiveness under varied scenarios, with all reported numbers representing classification accuracy (\\%). Although our primary analysis centers on these datasets, additional results for general few-shot classification on mini-ImageNet [41] are provided in the Appendix B.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.482, + 0.44 + ], + "angle": 0, + "content": "Fine-Grained Species Classification. For fine-grained classification, we utilize the iNaturalist 2021 dataset [47], which comprises a diverse collection of images and annotations contributed by citizen scientists, spanning numerous species of animals, plants, and fungi. Each species in the training split contains between 200 and 300 images, while the validation split includes 10 images per species." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Following [12], we experiment on images from five different families, each containing five to six species. These families are selected due to their challenging nature: distinguishing between species within the same family requires the model to identify complex features such as shapes and patterns, rather than simple color variations. The families used in our experiments are as follows: Lichen (fungi), Wrasse (fish), Wild rye (grass), Manzanita (berry shrubs), and Bulrush (herbs). For a detailed list of the species, please refer to the appendix. Notably, unlike [12], which utilizes a full training set containing between 200 and 300 images per species, our method uses on only 10 images per species to achieve generalization in a low-resource setting. We will report the baseline performance from their original paper as well as our reproduced results using the same limited data as our method for comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Novel Concept Classification. To test the model's generalizability on objects that have been rarely seen in the pretraining image resources, we evaluate our method using the Kiki-Bouba [2] dataset. The Kiki-Bouba experiment, originally introduced by [40], illustrates that people often associate specific shapes with different sounds. The dataset was constructed by prompting generative models trained to create 3D-rendered images from non-existent, meaningless words [2], providing a unique testbed for evaluating generalization to novel and abstract concepts. Following the approach in [12], we evaluate our method on two distinct splits, each containing five different classes. The first split consists of the classes bamaba, duludu, gaduga, lomulo, nomano, while the second split includes bouba, galaga," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "kepike, kiki, maluma. In [12], the training set comprises 800 images per class, with 200 images per class for validation. In contrast, we adopt a similar low-resource setting as used for the iNaturalist dataset, with our method accessing only 10 training images per class. Baseline results for both settings will be reported in the table for comparison." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.19, + 0.903, + 0.206 + ], + "angle": 0, + "content": "4.2. Baseline Methods and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.905, + 0.375 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our method, we compare it against several baselines, including the CLIP-based approach proposed in [12] and other baseline methods that utilize LLMs, such as LaBo [50], CBD [32] and LMMutate [12], which leverage LLM's pre-trained knowledge to generate attributes. Additionally, since our method leverages LLaVA-OneVision [28] to generate verbalized features, we perform quantitative comparisons with common approaches used to adapt LLaVA for downstream tasks, such as LoRA fine-tuning [22] and in-context learning [14]. We describe the details of each baseline method as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.534 + ], + "angle": 0, + "content": "CLIP-Based Methods. For CLIP-based baselines, we consider two variants. CLIP Class Name is a naive baseline where CLIP is employed to compute the similarity between the species' scientific name and the images. The best accuracy is reported as the highest value achieved among using the common name, the scientific name, or both names combined. The second variant involves Prompt Tuning with a CLIP encoder, where the text embeddings of specific class-related tokens are optimized through gradient descent. For both baselines, we present results as reported in [12]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.906, + 0.77 + ], + "angle": 0, + "content": "Language Bottleneck Methods. Classification by Description [32], LaBo [50], and LLM-Mutate [12] generate a list of attributes by prompting LLMs like GPT [35] and Llama [46] with the class name of the object. While this allows them to leverage the prior knowledge learned by LLMs, the generated features often lack grounding in the actual images. LaBo and LLM-Mutate alleviate this issue by using submodular optimization and evolutionary algorithm, respectively, to filter out ungrounded attributes. However, these methods still rely heavily on the capabilities and learned knowledge of the pre-trained LLMs. For LaBo, we adapt the official implementation for our datasets; for LLM-Mutate, we not only report the results from the original paper but also evaluate the method using the same limited data as our approach, ensuring a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.906, + 0.913 + ], + "angle": 0, + "content": "LLaVA-Based Methods. We include two common approaches for adopting LLaVA. The first baseline involves performing LoRA [22] fine-tuning (LLaVA-SFT), which has been demonstrated to be effective in scenarios with limited data. The second baseline is in-context learning [14] (LLaVA-ICL), where one exemplar image for each species or object is included in the prompt to guide the model's predictions. We study which of the LLaVA-based methods and VRL better utilizes LLaVA's capabilities." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1606" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.077, + 0.864, + 0.35 + ], + "angle": 0, + "content": "
MethodLichenWrasseWild RyeManzanitaBulrushAverage
Zero-Shot Methods
CLIP Class Name [12]23.332.032.026.026.027.86
Full Dataset Methods (200+ Images per Species)
CLIP Prompt Tuning [12]23.320.040.020.020.024.66
Classification by Description [32]30.034.036.028.020.029.60
LLM-Mutate-70B [12] (1-prompt)31.624.044.040.022.032.32
LLM-Mutate-70B [12] (10-prompt)48.344.058.058.042.050.06
Few-Shot Methods (10 Images per Species)
LLM-Mutate-7B† (10-prompt)35.048.038.044.026.038.20
LLM-Mutate-70B† (10-prompt)46.644.046.044.040.044.13
LaBo-7B† [50]45.048.044.042.036.043.00
LaBo-70B† [50]56.646.064.052.040.051.72
LLaVA-ICL-7B16.628.022.018.030.022.92
LLaVA-SFT-7B41.650.058.042.028.043.92
LLaVA-VRL-7B (Ours)58.348.074.066.046.058.46
LLaVA-VRL-72B (Ours)71.672.074.056.066.067.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.352, + 0.905, + 0.392 + ], + "angle": 0, + "content": "Table 1. Comparison of classification accuracy (%) across different methods for fine-grained classification on iNaturalist. The table presents results for zero-shot, full-dataset (200+ images per species), and few-shot (10 images per species). Results marked with † denote values reproduced using the official implementation but restricted to the same few-shot data as our method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.41, + 0.482, + 0.59 + ], + "angle": 0, + "content": "Our Method. We implement VRL with LLaVA-OneVision for capturing visual difference and commonality features. We also select LLaVA-OneVision as the feature mapping model that assists in converting an image into a feature vector, and the base model for all the LLaVA-based baseline methods. For the few-shot training stage, we experiment with various classification methods including logistic regression or MLP. Unless otherwise specified, we utilize both the difference features \\( F_{diff} \\) and commonality features \\( F_{comm} \\). We report the best performance achieved across different classifiers, selecting the optimal results from a single classifier." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.595, + 0.399, + 0.611 + ], + "angle": 0, + "content": "4.3. Fine-Grained Species Classification" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.839 + ], + "angle": 0, + "content": "We present the results for fine-grained species classification on the iNaturalist dataset in Table 1. We find that compared to previous SoTA methods, even with \\(95\\%\\) less data (10 images per species compared to \\(200+\\) images per species), and using a significantly smaller model (7B vs. 70B parameters), our method can already surpass the previous SoTA by \\(8\\%\\). When we increase the scale of our method to a comparable 72B model, we can further increase the improvements to nearly \\(18\\%\\). Moreover, when baseline methods are limited to the same data availability as our approach, our 7B and 72B models achieve improvements of \\(20\\%\\) and \\(23\\%\\), respectively. These results highlight the advantage of using grounded descriptions to capture the subtle visual cues, especially when the model needs to discover visual nuances to discriminate species within the same family." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We also compare VRL against LLaVA-based baselines. Using the same backbone LLaVA model, our method showcases a \\(15\\%\\) gain to LLaVA-SFT and a \\(36\\%\\) advantage over the in-context learning method, LLaVA-ICL. In addition," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.593 + ], + "angle": 0, + "content": "VRL is easier to be scaled up than the baselines. For instance, VRL with the 72B model requires 8 GPUs with 48 GB of RAM, while fine-tuning the 13B model with LoRA already requires 8 GPUs with 80 GB of RAM. For in-context learning, including image examples for every category in the prompt significantly increases the memory usage of the method, making it infeasible to perform in-context learning on the 72B LLaVA with the 8 GPUs (48 GB). The model performance may also be constrained by the context window of LLaVA model. These results and analysis further validate the scalability and robustness of VRL comparing to those existing adaptation algorithms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.77, + 0.615 + ], + "angle": 0, + "content": "4.4. Novel Concept Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For novel concept classification, we report the results on the Kiki-Bouba dataset in Table 2. Similar to the performance trend observed in the previous experiment, our method surpasses the previous SoTA by \\(14\\%\\) with a smaller model, but with \\(99\\%\\) less data (10 images per species compared to \\(800+\\) images per species). Surprisingly, we observe that increasing the number of model parameters does not lead to better performance on this task. One possible explanation for this is that, rather than relying on prior knowledge learned during pre-training, the model needs to focus on learning new, critical features specific to these novel objects with unfamiliar shapes and patterns. Since the objects in this task are unseen during LLM training, existing models struggle to generate relevant attributes and fail to ground visual differences necessary for distinguishing objects from different classes. This hypothesis is further supported by the larger performance gap between our method and those that rely on LLM's pre-trained knowledge [32]. On this dataset, our method exceeds their performance by \\(50\\%\\), compared" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1607" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.077, + 0.472, + 0.322 + ], + "angle": 0, + "content": "
Methodv1v2Avg.
Zero-Shot Methods
CLIP Class Name38.738.838.75
Full Dataset Methods (800+ Images per Object)
CLIP Prompt Tuning16.755.636.15
Classification by Description28.836.832.80
LLM-Mutate-70B (1-prompt)50.347.849.05
LLM-Mutate-70B (10-prompt)79.259.469.30
Few-Shot Methods (10 Images per Object)
LLM-Mutate-7B† (10-prompt)63.359.261.25
LLM-Mutate-70B† (10-prompt)67.359.263.25
LLaVA-ICL-7B24.527.225.85
LLaVA-SFT-7B72.150.661.35
LLaVA-VRL-7B (Ours)89.476.683.00
LLaVA-VRL-72B (Ours)89.174.781.90
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.406 + ], + "angle": 0, + "content": "Table 2. Accuracy (%) across different methods for novel concept classification on Kiki-Kouba dataset. v1 and v2 indicates two different splits described in Sec. 4.1. The table presents results for zero-shot, full-dataset, and few-shot. Results marked with \\(\\dagger\\) denote values reproduced using the official implementation but restricted to the same few-shot data as our method." + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.412, + 0.382, + 0.428 + ], + "angle": 0, + "content": "to a \\(37\\%\\) advantage on the previous dataset." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.279, + 0.455 + ], + "angle": 0, + "content": "4.5. Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Fusing Features Learned by VRL and Commonly Used Features. Table 3 summarizes the classification accuracy on iNaturalist achieved by fusing different types of visual features. Specifically, we study the two feature vectors: \\( F_{diff} \\) and \\( F_{comm} \\) learned via our VRL, which focusing on capturing regional nuances of the object. \\( F_{diff} \\) captures the inter-class difference while \\( F_{comm} \\) represents the shared features within the same class. We also consider the image features directly encoded by the vision encoder of CLIP, which often encode high-level semantics such as the object type, context, and other conceptual associations. As discussed in Sec. 3.1, these features are robust to different kinds of variance and could be benefited when combined for downstream tasks. This is evident from the \\( 2\\% \\) improvements when concatenating \\( F_{diff} \\) and \\( F_{comm} \\), and an additional \\( 6\\% \\) gain after including the CLIP feature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We also discover that ensemble classifiers trained separately on each feature type are more effective than simply concatenating all features to build a single classifier. This is primarily because these features are heterogeneous. For example, \\( F_{diff} \\) and \\( F_{comm} \\) are generally binary, with each element indicating the presence of a specific verbalized feature. Therefore, they tend to perform best when used with logistic regression models. Meanwhile, the features encoded by CLIP are continuous, which makes them more suitable for mapping to class labels through MLP models. Empirically, we demonstrate that the best configuration of ensembling can further improves the performance by \\( 3\\% \\) comparing to feature concatenation. We find that with three" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.077, + 0.885, + 0.248 + ], + "angle": 0, + "content": "
MethodFdiffFcommCLIPAvg.
Concat✓65.26
Concat✓58.32
Concat✓63.26
Concat✓✓67.92
Concat✓✓✓73.92
Ensemble (hard)LRLRLR68.32
Ensemble (hard)MLPMLPMLP74.26
Ensemble (hard)LRLRMLP71.60
Ensemble (soft)LRLRLR67.06
Ensemble (soft)MLPMLPMLP74.92
Ensemble (soft)LRLRMLP76.52
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.25, + 0.905, + 0.36 + ], + "angle": 0, + "content": "Table 3. Accuracy \\((\\%)\\) when incorporating feature vectors learned from different methods. \\(F_{diff}\\) and \\(F_{comm}\\) denote the difference and commonality feature vectors learned from VRL. CLIP refers to the image features encoded by CLIP visual encoder. We compare two approaches: concatenating all features to train a single classifier and ensembling classifiers trained separately on each feature type. The visual classifiers include Logistic Regression (LR) and Multi-Layer Perceptron (MLP) models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.456 + ], + "angle": 0, + "content": "classifiers, hard-voting is prone to high variance, while soft-voting is more robust since it leverages the confidence of each classifier by averaging their probabilities. These experiments showcase that our approach is flexible and can be seamlessly integrated with existing representations to build a robust classification model through ensembling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.7 + ], + "angle": 0, + "content": "Comparison with Human-Labeled Features. To evaluate the performance of automatically extracted features compared to human-labeled features, we conduct experiments on the Kiki-Bouba dataset using the human-annotated attributes provided by [12]. From Table 4, we observe that VRL extracted features outperform human-labeled attributes across all configurations. Our best configuration, which combines both difference feature vector \\( F_{diff} \\) and commonality feature vector \\( F_{comm} \\), surpasses the human-derived features by \\( 20\\% \\). This underscores the effectiveness of our automatic feature extraction approach in capturing relevant and distinguishing characteristics for classification tasks. These results also suggest that, in scenarios with limited resources or new visual concepts, we can automate the process of feature extraction, reducing the need for manual human annotation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.795 + ], + "angle": 0, + "content": "Classifier Choices and Feature Mapping Models. We investigate the model performance using different choices for visual classifiers and feature mapping models. From Table 5, we notice that LLaVA, as a feature mapping model for obtaining binary feature vectors, achieves the best overall performance when paired with the logistic regression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.1, we can also leverage CLIP as a feature mapping model to convert images into features. When predicting, one can either set a threshold to determine if the image possesses a certain attribute, resulting in a binary feature, or preserve the similarity between the image and each attribute to form continuous features. Although the performance of using CLIP features is slightly lower" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.085, + 0.506, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.087, + 0.898, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.908, + 0.331 + ], + "angle": 0, + "content": "Figure 3. Qualitative examples of the features extracted by VRL. We highlight the key attributes in bold. (a) Verbalized features extracted by comparing images from different classes. (b) Verbalized features extracted by comparing the images within the same class." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.339, + 0.473, + 0.42 + ], + "angle": 0, + "content": "
Methodv1v2Avg.
Human73.852.563.15
VRL-Fdiff88.475.281.80
VRL-Fcomm89.274.081.60
VRL-all89.476.683.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.464 + ], + "angle": 0, + "content": "Table 4. Accuracy (%) of using human-labeled attributes and VRL extracted features on Kiki-Bouba dataset. Note that v1 and v2 indicates two different splits described in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.484, + 0.659 + ], + "angle": 0, + "content": "compared to LLaVA, CLIP offers a notable advantage in terms of inference speed. This is because CLIP supports batch-wise operations when computing similarities between a set of images and a set of verbalized features, which offers our method the flexibility to scale with a larger dataset or an increasing number of object classes. It is worth noting that using CLIP as the feature mapping model still allows our method to surpass previous state-of-the-art results by significant margins, achieving improvements of \\(8\\%\\) and \\(12\\%\\) on the iNaturalist and Kiki-Bouba datasets, respectively. This highlights the universal applicability of VRL-extracted features, regardless of the feature mapping approach used." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.285, + 0.683 + ], + "angle": 0, + "content": "4.6. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Figure 3 presents qualitative examples of what verbalized features are extracted from image pairs. We list some of the features that have top-5 feature importance in our trained logistic regression model. For inter-class differences (Fig. 3 (a)), verbalized features are used to capture distinguishing characteristics between different species. For example, the Lichen (Fungi) category is distinguished based on the structure of the thallus, where one image describes a flat, crustlike thallus, and the other describes a complex, branch-like thallus. Similarly, in the Wild Rye (Grass) category, the differences between a slender stem with a feathery inflorescence and a more robust stem with a dense, elongated seed head are verbalized. For intra-class commonality (Fig. 3 (b)), verbalized features are employed to capture shared" + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.339, + 0.875, + 0.476 + ], + "angle": 0, + "content": "
MethodRFLRMLP
iNaturalist (Prior Works: 50.1)
LLaVA64.967.466.0
CLIP (binary)52.258.055.2
CLIP (continuous)52.853.660.6
Kiki-Bouba (Prior Works: 69.3)
LLaVA83.6586.6585.65
CLIP (binary)79.2581.5081.10
CLIP (continuous)77.0577.1081.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.476, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Table 5. Accuracy (\\%) of different classifier choices (columns) and models used to map verbalized features to numeric vectors (rows). LLaVA and CLIP are utilized as mapping models to map an image to a feature vector." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.674 + ], + "angle": 0, + "content": "traits within the same species. For example, the leaves of Manzanita (Berry) are consistently described as “ovate with a pointed tip and slightly serrated edges,” while the inflorescences of Bulrush (Herb) are uniformly described as having a “delicate, feathery appearance.” These verbalized features are used to guide the decision process by helping the model focus on important distinguishing features or shared traits within each class. Due to page limit, see Appendix B.5 for human study on the learned verbalized features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose Verbalized Representation Learning (VRL), which enables automatic interpretable feature extraction with few-shot samples. By leveraging VLMs, VRL generates verbalized features that capture both inter-class differences and intra-class commonalities. Our method not only enhances the model's adaptability with limited data but also provides transparency in the decision-making process, enabling easier interpretation of the features that influence predictions. Our experiments show that VRL outperforms prior approaches, achieving superior results while using significantly less data. This includes tasks like fine-grained recognition and novel concept adaptation, demonstrating its potential for real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1609" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.284 + ], + "angle": 0, + "content": "We thank anonymous reviewers and other members of UCLA-NLP+ group for their helpful comments. This work was partially supported by U.S. DARPA ECOLE Program No. #HR00112390060, ONR grant N00014-23-1-2780, DARPA ANSR program FA8750-23-2-0004, and Apple. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.315, + 0.188, + 0.33 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.339, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.412, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[2] Morris Alper and Hadar Averbuch-Elor. Kiki or bouba? sound symbolism in vision-and-language models. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.456, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.528, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.6, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.729, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[9] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. arXiv preprint arXiv:2303.14123, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.339, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.904, + 0.191 + ], + "angle": 0, + "content": "[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. Solo: A single transformer for scalable vision-language modeling. In Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[12] Mia Chiquier, Utkarsh Mall, and Carl Vondrick. Evolving interpretable visual classifiers with large language models. arXiv preprint arXiv:2404.09941, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[13] Fahim Dalvi, Nadir Durrani, Hassan Sajjad, Yonatan Belinkov, Anthony Bau, and James Glass. What is one grain of sand in the desert? analyzing individual neurons in deep nlp models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6309-6317, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[14] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.417 + ], + "angle": 0, + "content": "[15] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.904, + 0.46 + ], + "angle": 0, + "content": "[16] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. Advances in neural information processing systems, 20, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.904, + 0.518 + ], + "angle": 0, + "content": "[17] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.521, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[20] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In Proceedings of the European conference on computer vision (ECCV), pages 264-279, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Shaoli Huang, Zhe Xu, Dacheng Tao, and Ya Zhang. Part-stacked cnn for fine-grained visual categorization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1173-1182, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1610" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[24] Andrej Karpathy, Justin Johnson, and Li Fei-Fei. Visualizing and understanding recurrent networks. arXiv preprint arXiv:1506.02078, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[26] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[29] Manling Li, Ruochen Xu, Shuohang Wang, Luowei Zhou, Xudong Lin, Chenguang Zhu, Michael Zeng, Heng Ji, and Shih-Fu Chang. Clip-event: Connecting text and images with event structures. In Proc. Conference on Computer Vision and Pattern Recognition (CVPR2022), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[30] Fan Liu, Wenwen Cai, Jian Huo, Chuanyi Zhang, De long Chen, and Jun Zhou. Making large vision language models to be good few-shot learners. arXiv preprint arXiv:2408.11297, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[31] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. arXiv preprint arXiv:2408.12469, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[33] Eduardo Navarrete, Michele Miozzo, and Francesca Peressotti. Language can shape the perception of oriented objects. Scientific reports, 10(1):8409, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.295, + 0.703 + ], + "angle": 0, + "content": "[35] OpenAI. ChatGPT. 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[36] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8779-8788, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[37] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[38] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Vilayanur S Ramachandran and Edward M Hubbard. Synaesthesia-a window into perception, thought and language. Journal of consciousness studies, 8(12):3-34, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[41] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[42] Bernardino Romera-Paredes and Philip Torr. An embarrassingly simple approach to zero-shot learning. In International conference on machine learning, pages 2152-2161. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[44] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.59 + ], + "angle": 0, + "content": "[45] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.592, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[47] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[48] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3090-3100, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Cheng-Fu Yang, Yao-Hung Hubert Tsai, Wan-Cyuan Fan, Russ R Salakhutdinov, Louis-Philippe Morency, and Frank Wang. Paraphrasing is all you need for novel object captioning. Advances in Neural Information Processing Systems, 35:6492–6504, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1611" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[50] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[51] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[52] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[53] Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. Sglang: Efficient execution of structured language model programs. arXiv preprint arXiv:2312.07104, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "1612" + } + ] +] \ No newline at end of file diff --git a/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_origin.pdf b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..75e81e8b03455300e3c75170d02581fd27c40e72 --- /dev/null +++ b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/e86880d5-4a52-49e7-83a5-bbf2d31e9dd1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7961d852e1729d40270c5ad68807cb375bb9e1be6385a1ab58928e7d7045ecc7 +size 9818770 diff --git a/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/full.md b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b21349bfa53589626293893176be321f4eb4514e --- /dev/null +++ b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/full.md @@ -0,0 +1,233 @@ +# Verbalized Representation Learning for Interpretable Few-Shot Generalization + +Cheng-Fu Yang $^{1*}$ Da Yin $^{1*}$ Wenbo Hu $^{1}$ Heng Ji $^{2}$ Nanyun Peng $^{1}$ Bolei Zhou $^{1}$ Kai-Wei Chang $^{1}$ $^{1}$ University of California, Los Angeles $^{2}$ University of Illinois Urbana-Champaign + +cfyang@cs.ucla.edu + +![](images/a2d3c95d96c71efe3946fbc88d9ec5631661e2b12cae9bb15c6c85703bcd3b71.jpg) +Figure 1. (a) Humans can identify key differences between similar species and recognize common traits within a class, even when appearance varies, and express these insights in concise language. Similarly, our Verbalized Representation Learning (VRL) extracts meaningful features by querying Vision-Language Models (VLMs) to capture inter-class discriminative features and intra-class commonalities. (b) These verbalized features are mapped to numerical vectors through the VLM and can be used by downstream classifiers. Our VRL demonstrates superior performance in few-shot classification compared to prior work, achieving a $24\%$ absolute gain on iNaturalist and a $14\%$ on Novel Object Classification, even with a much smaller model. + +# Abstract + +Humans recognize objects after observing only a few examples, a remarkable capability enabled by their inherent language understanding of the real-world environment. Developing verbalized and interpretable representation can significantly improve model generalization in low-data settings. In this work, we propose Verbalized Representation Learning (VRL), a novel approach for automatically extracting human-interpretable features for object recognition using few-shot data. Our method uniquely captures interclass differences and intra-class commonalities in the form of natural language by employing a Vision-Language Model (VLM) to identify key discriminative features between different classes and shared characteristics within the same class. These verbalized features are then mapped to numeric vectors through the VLM. The resulting feature vectors can be further utilized to train and infer with downstream classifiers. Experimental results show that, at the same model scale, VRL achieves a $24\%$ absolute improvement over prior state-of-the-art methods while using $95\%$ less data and a smaller model. Furthermore, compared to human-labeled attributes, the features learned by VRL exhibit a $20\%$ absolute gain when used for downstream classification tasks. $^{1}$ + +# 1. Introduction + +Humans have a remarkable capability to recognize certain objects after seeing only a few examples. As suggested by [33], it is significantly enhanced by inherent language understanding. Language, with its representational strength, serves as a primary resource for conveying knowledge about visual objects. A single, precise description can effectively capture visual distinctions observed across different object categories. For example, in Fig. 1 (a), a human can identify the key difference between two fish species based on the pattern of their heads, even if both share a white stripe around the neck and a yellow body. As a result, we argue that developing verbalized features offers a valuable complement to improve model's generalization under low-resource conditions. In addition, incorporating language into image classification models also increases the interpretability of the visual system, facilitating transparent and reliable decision making and effective model auditing. + +Prior works [15, 16] have explored integrating language into visual classifiers with a bottleneck of textual attributes, where the model first predicts relevant attributes and subsequently uses them to classify the image. However, these attributes often require human annotations or predefined ontologies. Development of Large-scale Vision-and-Language Models (VLMs) like CLIP [39] and GPT-4V [1] have enabled recent bottleneck methods [12, 32, 38, 48, 50] + +to generate relevant attributes by leveraging the prior knowledge of VLMs learned during the pre-training stage. However, these attributes are often ungrounded for the testing images and rely on VLM's prior knowledge, resulting in low precision when applied to fine-grained or novel concept recognition where objects are underrepresented in the pre-training datasets. + +To address the aforementioned challenges, we propose Verbalized Representation Learning (VRL) for automatic, human-interpretable feature extraction using only few-shot data. It applies to fine-grained and novel objects and could work with local models such as LLaVA [28]. Specifically, inspired by self-supervised representation learning (SSL) in computer vision, such as SimCLR [8], MoCo [19], SwAV [6], SimSiam [10], BYOL [18], CLIP-Event [29], and SOLO [11], we propose to leverage a VLM to capture the inter-class difference and intra-class commonality and articulate these findings in natural language, as illustrated in Fig. 1. Specifically, we cast the VLM to describe the key difference between two images from different classes, which would preserve the discriminative features and remove the redundant ones, such as the yellow body shared by both species on the left side of Fig. 1 (a). Conversely, we employ the VLM to list the key features that are shared by two images within the same class, extracting features that are robust to intra-class variance, such as the orange facial pattern observed in all of the Thalassoma Pavo. Notably, this process can be applied to any two images, whether from different classes or the same class, which allows us to exponentially scale the number of useful verbalized features with the image samples, enabling effective generalization even in few-shot settings. + +Consequently, to obtain the numerical feature embedding of an image using our VRL, we employ a VLM to determine whether the image possesses the characteristics described by the verbalized features, as depicted in Fig. 1 (b). Each dimension in the resulting embedding represents a scalar value indicating the presence of the described feature. This approach effectively transforms the verbalized features into numeric representations that can serve as inputs to any classification methods, such as logistic regression, random forest, or MLP classifiers, allowing flexible and robust modeling based on the extracted interpretable features. + +We conduct experiments on iNaturalist [47] and KikiBouba dataset [2], where the former includes objects that only have subtle differences between classes, while the latter contains novel objects that are nearly not present in the web-scale datasets. Compared to previous state-of-the-art (SoTA) baselines using 70B models, we achieve a $24\%$ absolute improvement while using $95\%$ less data and a much smaller model with 7B parameters. Against supervised finetuned baselines, we observe a $15\%$ absolute improvement. Lastly, when compared to human annotated attributes, the + +attributes learned by our VRL demonstrate a $20\%$ absolute gain. These results illustrate that the features extracted by VRL not only offer superior effectiveness but also exhibit strong robustness, providing a generalizable solution when adapting to fine-grained or novel classification tasks with limited data. + +# 2. Related Work + +Natural Language for Image Classification. A common approach for integrating language into visual classifiers involves creating a concept bottleneck [26]. Early bottleneck methods have been extensively applied in few-shot or zero-shot classification models [7, 15-17, 23, 26, 27, 42]. However, these methods typically rely on manually annotated or predefined attributes, which can limit their adaptability to novel classes or fine-grained tasks. Recent advancements in VLMs have enabled researchers to directly sample interpretable descriptive features from these models [12, 32, 38, 45, 48-51]. This approach benefits from incorporating external knowledge embedded in the pretrained datasets. However, it often relies heavily on the prior knowledge encoded within these models, which can lead to the generation of ungrounded features and difficulty generalizing when the target data is not well-represented in the training datasets. In contrast, our method directly learns grounded, verbalized features from the visual data, allowing for automatic interpretable feature extraction that remains adaptable even for novel classes. + +Interpreting Model's Decision Process. There has been a long line of research aiming to improve the interpretability and explainability of deep models. Pioneering methods [3, 4, 13, 24, 43, 52] have tackled this challenge by visualizing learned features, categorizing maximally-activating inputs, and identifying key neurons that drive model decisions. More recently, efforts have been made to interpret model behavior using natural language [20, 21, 36, 44], enabling explanations that are more accessible and editable. However, the aforementioned methods are generally post-hoc, meaning they are applied after the model has been trained with abundant data. While post-hoc explanations can provide valuable insights, they may lack the ability to influence or shape the model's internal representations during training. Instead, our methods embed interpretability directly into the learning process by learning verbalized features. This enables the model to produce inherently explainable and contextually grounded representations without relying solely on retrospective analysis. + +Self-Supervised Learning Models. Self-supervised learning (SSL) has emerged as a powerful approach in computer vision, enabling models to learn robust feature representations from unlabeled data. Contrastive learning [34] methods, such as SimCLR [8] and MoCo [19], achieve this goal by learning to discriminate between different samples (neg + +![](images/49ec2f97e95200572596d5b59352c13f5bc971403f9eb9444716c8cd482618e6.jpg) +(a) + +![](images/3248b61fe0761815c8b7e7d1614d6f8279951fd77cac75e3a250e0104aaf1063.jpg) +Figure 2. The overview of our Verbalized Representation Learning (VRL) framework. (a) Given $N$ samples per class from $C$ different classes, VRL is able to generate a diverse set (exponentially scaling with $N \times C$ ) of verbalized features by: 1) extracting key differences between samples from different classes, and 2) identifying commonalities shared among objects within the same class. (b) Given an image, a Vision-and-Language model (VLM) is employed to evaluate whether the image contains the characteristics described by the verbalized features. This process can map a set of verbalized features into numeric representations which can then be used for downstream tasks. + +![](images/e8c179c428e489c9cafeea97b77c04b3a2ab6cce4a3ef76fcb9dd385cefc34fa.jpg) + +ative pairs) and augmentations of the same sample (positive pairs). Chen et al. [10] have found that the key function of negative pairs is to prevent the model from learning collapsing features, where models produce constant or trivial outputs. As a result, subsequent works have relieved the need for negative samples by employing techniques like clustering [5, 6], momentum update [18], or stop gradient [10]. These methods focus on learning the underlying shared representations between the augmentations of the same sample. Our approach can be viewed as a variant of supervised contrastive learning [25], where the positive samples are drawn from different images from the same class. However, our method introduces two distinct advantages. First, it is data-efficient, as it does not require large datasets for gradient updates. Second, the expressiveness of verbalized features inherently prevents the model from collapsing to constant outputs, thereby ensuring the resulting features are meaningful and representative. + +# 3. Method + +In Sec. 3.1, we first introduce the motivation and concept of verbalized representation learning (VRL) and then discuss how VRL automatically extracts interpretable, compact representations with few-shot data. Then, in Sec. 3.2, we outline the process of building a visual classifier using the derived features. Furthermore, we demonstrate that these extracted features are versatile and can be applied to arbitrary classification models, including but not limited to logistic regression, decision trees, and multi-layer perceptron (MLP) classifiers. The overall framework of our VRL + +is presented in Fig. 2. + +# 3.1. Verbalized Representation Learning + +Humans can recognize objects after seeing only a few examples, a skill boosted by language understanding [33]. Language effectively conveys visual distinctions, even without many extra visual cues. Incorporating language also enhances image classification interpretability, enabling clearer, more reliable decisions. To this end, we propose a method that verbalizes key visual features in natural language. In this section, we focus on identifying which visual features are crucial for few-shot image classification and how to extract and verbalize them effectively, and how to map these verbalized features into vectors which would be later utilized during training and inference. + +What visual features should be verbalized? Inspired by self-supervised representation learning methods, where model learns robust features through contrasting positive and negative pairs [8, 19] or maximizing the similarity of augmented views of the positive samples [6, 10, 18], we propose to verbalize the objective functions of these self-supervised learning methods with the help of VLMs such as LLaVA. Specifically, given a classification task with $C$ classes and $N$ -shot examples per class, each data point is defined as $(x,l)$ , where $x$ represents an image, and $l \in \{1,2,\dots,C\}$ is the categorical label indicating the class to which the image belongs. Our method leverages two types of paired images: positive pairs, defined as image pairs $(x_{i},x_{j})$ with the same label $(l_{i} = l_{j})$ , and negative pairs, which consist of images $(x_{i},x_{k})$ from different + +classes $(l_i\neq l_k)$ . Notably, with this pairing strategy, we can form $\mathbf{C}_2^C\times \mathbf{C}_2^N$ distinct negative pairs, and $C\times \mathbf{C}_2^N$ positive pairs, which significantly increase the data utilization under the few-shot setting. + +Inspired by contrastive learning methods [8, 19], for each negative pair sample, our approach emphasizes capturing inter-class differences by tasking the VLM to describe key distinguishing features between images from different classes, denoted as $y_{diff} = \mathrm{VLM}(\mathbf{x}_i,\mathbf{x}_k,q_{diff})$ , where $l_i\neq l_k$ and $q_{diff}$ denotes the query that captures the visual differences between the two input images $\mathbf{x}_i$ and $\mathbf{x}_k$ . For instance, in Fig. 2 (a), the model learns to distinguish two fish species by the coloration and the pattern of the fish. Conversely, for each positive pair, we draw inspirations from negative sample-free SSL methods [6, 10, 18], where we employ the VLM to capture intra-class commonalities, generating descriptions of the key shared features between two images from the same class, i.e., $y_{comm} = \mathrm{VLM}(\mathbf{x}_i,\mathbf{x}_j,q_{comm})$ , where $l_i = l_j$ and $q_{comm}$ denotes the query that captures the visual commonalities between $\mathbf{x}_i$ and $\mathbf{x}_j$ . For example, in Fig. 2 (a), VRL identifies that both fish possess irregular orange spots around their face. We include the detailed prompt templates used in the above process in Appendix A.1. + +Intuitively, $y_{diff}$ and $y_{comm}$ learn complimentary features. $y_{diff}$ captures the most discriminative features while filtering out redundant ones, such as the yellow body present in both types of fish, leading to less noisy features. While $y_{comm}$ is robust to the intra-class variance where it learns to identify the fish based on the shared features, like the pattern on its face, rather than the number of stripes on its back. In the later experimental section, we will further verify this assumption and show that these two features can yield the best performance when combined. In addition, as discussed in the earlier paragraph, even with only $N$ -shot samples, our VRL can sample a diverse set of features from $\mathbf{C}_2^C \times \mathbf{C}_2^N$ negative and $C \times \mathbf{C}_2^N$ positive pairs, making our method excel under low-resource setting. Moreover, we empirically discover that sampling $C \times N$ pairs for both negative and positive samples is sufficient to collect a robust set of features. + +How to map the verbalized features to numeric feature embeddings? To obtain feature vectors that can be later used to build visual classifiers, we transform the verbalized features into numeric vectors with the help of VLMs such as CLIP [39] and LLaVA [28]. Given an image and a set of verbalized features, which are in the form of language descriptions, we employ a VLM to determine whether the image possesses those features, as illustrated in Fig. 2 (b), resulting in a feature embedding $F$ , i.e., $F = \mathrm{VLM}(y,x)$ . Each dimension of $F$ is produced by the VLM, indicating the presence of a certain verbalized feature from $y_{diff}$ and $y_{comm}$ or assessing the degree that the image has a certain + +feature. Concretely, for generative VLMs like LLaVA, each verbalized feature is mapped to 0 or 1, based on whether the model infers the presence of the feature in the image. For VLMs like CLIP, feature embeddings are derived by calculating the similarity between each verbalized feature and the image using CLIP's visual encoders. This results in a continuous vector indicating the likelihood that the image contains each feature, with an optional similarity threshold to convert it into binary feature vectors. + +# 3.2. Training and Inference + +Training. Once the numeric feature vectors are generated, they are used as input for training various visual classifiers, including, but not limited to, logistic regression, random forests, Naive Bayes, K-nearest neighbors (KNN), and MLP classifiers. These models learn to predict the class labels based on the interpretable feature vectors. This flexibility enables robust modeling that can be tailored to different applications. Moreover, this approach not only enables the construction of classifiers for novel concepts with few-shot data, but also provides insight into which visual features are most important for decision-making. For instance, we can extract feature importance from logistic regression or visualize decision paths in decision tree classifiers, enhancing the interpretability of our method. + +Inference. During inference, given a testing image, we generate its feature vector by following the approach described in Sec. 3.1. Specifically, we query the VLMs with the learned verbalized features to determine whether the image contains specific features, producing a numeric feature vector. This vector can either be continuous, representing the likelihood of a feature's presence, or binary, indicating the feature's absence or presence. The resulting vector is then passed to the trained classifier to make a prediction. Notably, to enhance the model robustness, we ensemble results from classifiers trained with different algorithms. Combining predictions from multiple classifiers is beneficial because it reduces the likelihood of overfitting to any specific algorithm's biases or weaknesses. The ensemble can use either hard or soft voting: In hard voting, each classifier makes a discrete prediction, and the final prediction is determined by a majority vote. In soft voting, the prediction logits from each classifier are averaged to produce the final output, allowing for more nuanced decision-making. + +# 4. Experiments + +Our experiments are designed to answer the following research questions: (i) How well does VRL generalize to tasks requiring fine-grained recognition with few-shot data? (ii) Can VRL effectively adapt to tasks involving novel concepts that were not part of the VLM's pre-training resources, using only few-shot data? (iii) How effective is + +VRL compared to conventional few-shot adaptation algorithms like LoRA fine-tuning or in-context learning? (iv) Do different types of features extracted by VRL mutually benefit each other, and do the extracted features further improve on top of the other commonly used features? (v) How does the performance of automatically features extracted with VRL compare to human-labeled features? + +# 4.1. Dataset and Evaluation Protocols + +We conduct experiments on iNaturalist and Kiki-Bouba to validate our method's effectiveness under varied scenarios, with all reported numbers representing classification accuracy (\%). Although our primary analysis centers on these datasets, additional results for general few-shot classification on mini-ImageNet [41] are provided in the Appendix B.3. + +Fine-Grained Species Classification. For fine-grained classification, we utilize the iNaturalist 2021 dataset [47], which comprises a diverse collection of images and annotations contributed by citizen scientists, spanning numerous species of animals, plants, and fungi. Each species in the training split contains between 200 and 300 images, while the validation split includes 10 images per species. + +Following [12], we experiment on images from five different families, each containing five to six species. These families are selected due to their challenging nature: distinguishing between species within the same family requires the model to identify complex features such as shapes and patterns, rather than simple color variations. The families used in our experiments are as follows: Lichen (fungi), Wrasse (fish), Wild rye (grass), Manzanita (berry shrubs), and Bulrush (herbs). For a detailed list of the species, please refer to the appendix. Notably, unlike [12], which utilizes a full training set containing between 200 and 300 images per species, our method uses on only 10 images per species to achieve generalization in a low-resource setting. We will report the baseline performance from their original paper as well as our reproduced results using the same limited data as our method for comparison. + +Novel Concept Classification. To test the model's generalizability on objects that have been rarely seen in the pretraining image resources, we evaluate our method using the Kiki-Bouba [2] dataset. The Kiki-Bouba experiment, originally introduced by [40], illustrates that people often associate specific shapes with different sounds. The dataset was constructed by prompting generative models trained to create 3D-rendered images from non-existent, meaningless words [2], providing a unique testbed for evaluating generalization to novel and abstract concepts. Following the approach in [12], we evaluate our method on two distinct splits, each containing five different classes. The first split consists of the classes bamaba, duludu, gaduga, lomulo, nomano, while the second split includes bouba, galaga, + +kepike, kiki, maluma. In [12], the training set comprises 800 images per class, with 200 images per class for validation. In contrast, we adopt a similar low-resource setting as used for the iNaturalist dataset, with our method accessing only 10 training images per class. Baseline results for both settings will be reported in the table for comparison. + +# 4.2. Baseline Methods and Implementation Details + +To demonstrate the effectiveness of our method, we compare it against several baselines, including the CLIP-based approach proposed in [12] and other baseline methods that utilize LLMs, such as LaBo [50], CBD [32] and LMMutate [12], which leverage LLM's pre-trained knowledge to generate attributes. Additionally, since our method leverages LLaVA-OneVision [28] to generate verbalized features, we perform quantitative comparisons with common approaches used to adapt LLaVA for downstream tasks, such as LoRA fine-tuning [22] and in-context learning [14]. We describe the details of each baseline method as follows. + +CLIP-Based Methods. For CLIP-based baselines, we consider two variants. CLIP Class Name is a naive baseline where CLIP is employed to compute the similarity between the species' scientific name and the images. The best accuracy is reported as the highest value achieved among using the common name, the scientific name, or both names combined. The second variant involves Prompt Tuning with a CLIP encoder, where the text embeddings of specific class-related tokens are optimized through gradient descent. For both baselines, we present results as reported in [12]. + +Language Bottleneck Methods. Classification by Description [32], LaBo [50], and LLM-Mutate [12] generate a list of attributes by prompting LLMs like GPT [35] and Llama [46] with the class name of the object. While this allows them to leverage the prior knowledge learned by LLMs, the generated features often lack grounding in the actual images. LaBo and LLM-Mutate alleviate this issue by using submodular optimization and evolutionary algorithm, respectively, to filter out ungrounded attributes. However, these methods still rely heavily on the capabilities and learned knowledge of the pre-trained LLMs. For LaBo, we adapt the official implementation for our datasets; for LLM-Mutate, we not only report the results from the original paper but also evaluate the method using the same limited data as our approach, ensuring a fair comparison. + +LLaVA-Based Methods. We include two common approaches for adopting LLaVA. The first baseline involves performing LoRA [22] fine-tuning (LLaVA-SFT), which has been demonstrated to be effective in scenarios with limited data. The second baseline is in-context learning [14] (LLaVA-ICL), where one exemplar image for each species or object is included in the prompt to guide the model's predictions. We study which of the LLaVA-based methods and VRL better utilizes LLaVA's capabilities. + +
MethodLichenWrasseWild RyeManzanitaBulrushAverage
Zero-Shot Methods
CLIP Class Name [12]23.332.032.026.026.027.86
Full Dataset Methods (200+ Images per Species)
CLIP Prompt Tuning [12]23.320.040.020.020.024.66
Classification by Description [32]30.034.036.028.020.029.60
LLM-Mutate-70B [12] (1-prompt)31.624.044.040.022.032.32
LLM-Mutate-70B [12] (10-prompt)48.344.058.058.042.050.06
Few-Shot Methods (10 Images per Species)
LLM-Mutate-7B† (10-prompt)35.048.038.044.026.038.20
LLM-Mutate-70B† (10-prompt)46.644.046.044.040.044.13
LaBo-7B† [50]45.048.044.042.036.043.00
LaBo-70B† [50]56.646.064.052.040.051.72
LLaVA-ICL-7B16.628.022.018.030.022.92
LLaVA-SFT-7B41.650.058.042.028.043.92
LLaVA-VRL-7B (Ours)58.348.074.066.046.058.46
LLaVA-VRL-72B (Ours)71.672.074.056.066.067.92
+ +Table 1. Comparison of classification accuracy (%) across different methods for fine-grained classification on iNaturalist. The table presents results for zero-shot, full-dataset (200+ images per species), and few-shot (10 images per species). Results marked with † denote values reproduced using the official implementation but restricted to the same few-shot data as our method. + +Our Method. We implement VRL with LLaVA-OneVision for capturing visual difference and commonality features. We also select LLaVA-OneVision as the feature mapping model that assists in converting an image into a feature vector, and the base model for all the LLaVA-based baseline methods. For the few-shot training stage, we experiment with various classification methods including logistic regression or MLP. Unless otherwise specified, we utilize both the difference features $F_{diff}$ and commonality features $F_{comm}$ . We report the best performance achieved across different classifiers, selecting the optimal results from a single classifier. + +# 4.3. Fine-Grained Species Classification + +We present the results for fine-grained species classification on the iNaturalist dataset in Table 1. We find that compared to previous SoTA methods, even with $95\%$ less data (10 images per species compared to $200+$ images per species), and using a significantly smaller model (7B vs. 70B parameters), our method can already surpass the previous SoTA by $8\%$ . When we increase the scale of our method to a comparable 72B model, we can further increase the improvements to nearly $18\%$ . Moreover, when baseline methods are limited to the same data availability as our approach, our 7B and 72B models achieve improvements of $20\%$ and $23\%$ , respectively. These results highlight the advantage of using grounded descriptions to capture the subtle visual cues, especially when the model needs to discover visual nuances to discriminate species within the same family. + +We also compare VRL against LLaVA-based baselines. Using the same backbone LLaVA model, our method showcases a $15\%$ gain to LLaVA-SFT and a $36\%$ advantage over the in-context learning method, LLaVA-ICL. In addition, + +VRL is easier to be scaled up than the baselines. For instance, VRL with the 72B model requires 8 GPUs with 48 GB of RAM, while fine-tuning the 13B model with LoRA already requires 8 GPUs with 80 GB of RAM. For in-context learning, including image examples for every category in the prompt significantly increases the memory usage of the method, making it infeasible to perform in-context learning on the 72B LLaVA with the 8 GPUs (48 GB). The model performance may also be constrained by the context window of LLaVA model. These results and analysis further validate the scalability and robustness of VRL comparing to those existing adaptation algorithms. + +# 4.4. Novel Concept Classification + +For novel concept classification, we report the results on the Kiki-Bouba dataset in Table 2. Similar to the performance trend observed in the previous experiment, our method surpasses the previous SoTA by $14\%$ with a smaller model, but with $99\%$ less data (10 images per species compared to $800+$ images per species). Surprisingly, we observe that increasing the number of model parameters does not lead to better performance on this task. One possible explanation for this is that, rather than relying on prior knowledge learned during pre-training, the model needs to focus on learning new, critical features specific to these novel objects with unfamiliar shapes and patterns. Since the objects in this task are unseen during LLM training, existing models struggle to generate relevant attributes and fail to ground visual differences necessary for distinguishing objects from different classes. This hypothesis is further supported by the larger performance gap between our method and those that rely on LLM's pre-trained knowledge [32]. On this dataset, our method exceeds their performance by $50\%$ , compared + +
Methodv1v2Avg.
Zero-Shot Methods
CLIP Class Name38.738.838.75
Full Dataset Methods (800+ Images per Object)
CLIP Prompt Tuning16.755.636.15
Classification by Description28.836.832.80
LLM-Mutate-70B (1-prompt)50.347.849.05
LLM-Mutate-70B (10-prompt)79.259.469.30
Few-Shot Methods (10 Images per Object)
LLM-Mutate-7B† (10-prompt)63.359.261.25
LLM-Mutate-70B† (10-prompt)67.359.263.25
LLaVA-ICL-7B24.527.225.85
LLaVA-SFT-7B72.150.661.35
LLaVA-VRL-7B (Ours)89.476.683.00
LLaVA-VRL-72B (Ours)89.174.781.90
+ +Table 2. Accuracy (%) across different methods for novel concept classification on Kiki-Kouba dataset. v1 and v2 indicates two different splits described in Sec. 4.1. The table presents results for zero-shot, full-dataset, and few-shot. Results marked with $\dagger$ denote values reproduced using the official implementation but restricted to the same few-shot data as our method. +to a $37\%$ advantage on the previous dataset. + +# 4.5. Additional Analysis + +Fusing Features Learned by VRL and Commonly Used Features. Table 3 summarizes the classification accuracy on iNaturalist achieved by fusing different types of visual features. Specifically, we study the two feature vectors: $F_{diff}$ and $F_{comm}$ learned via our VRL, which focusing on capturing regional nuances of the object. $F_{diff}$ captures the inter-class difference while $F_{comm}$ represents the shared features within the same class. We also consider the image features directly encoded by the vision encoder of CLIP, which often encode high-level semantics such as the object type, context, and other conceptual associations. As discussed in Sec. 3.1, these features are robust to different kinds of variance and could be benefited when combined for downstream tasks. This is evident from the $2\%$ improvements when concatenating $F_{diff}$ and $F_{comm}$ , and an additional $6\%$ gain after including the CLIP feature. + +We also discover that ensemble classifiers trained separately on each feature type are more effective than simply concatenating all features to build a single classifier. This is primarily because these features are heterogeneous. For example, $F_{diff}$ and $F_{comm}$ are generally binary, with each element indicating the presence of a specific verbalized feature. Therefore, they tend to perform best when used with logistic regression models. Meanwhile, the features encoded by CLIP are continuous, which makes them more suitable for mapping to class labels through MLP models. Empirically, we demonstrate that the best configuration of ensembling can further improves the performance by $3\%$ comparing to feature concatenation. We find that with three + +
MethodFdiffFcommCLIPAvg.
Concat✓65.26
Concat✓58.32
Concat✓63.26
Concat✓✓67.92
Concat✓✓✓73.92
Ensemble (hard)LRLRLR68.32
Ensemble (hard)MLPMLPMLP74.26
Ensemble (hard)LRLRMLP71.60
Ensemble (soft)LRLRLR67.06
Ensemble (soft)MLPMLPMLP74.92
Ensemble (soft)LRLRMLP76.52
+ +Table 3. Accuracy $(\%)$ when incorporating feature vectors learned from different methods. $F_{diff}$ and $F_{comm}$ denote the difference and commonality feature vectors learned from VRL. CLIP refers to the image features encoded by CLIP visual encoder. We compare two approaches: concatenating all features to train a single classifier and ensembling classifiers trained separately on each feature type. The visual classifiers include Logistic Regression (LR) and Multi-Layer Perceptron (MLP) models. + +classifiers, hard-voting is prone to high variance, while soft-voting is more robust since it leverages the confidence of each classifier by averaging their probabilities. These experiments showcase that our approach is flexible and can be seamlessly integrated with existing representations to build a robust classification model through ensembling. + +Comparison with Human-Labeled Features. To evaluate the performance of automatically extracted features compared to human-labeled features, we conduct experiments on the Kiki-Bouba dataset using the human-annotated attributes provided by [12]. From Table 4, we observe that VRL extracted features outperform human-labeled attributes across all configurations. Our best configuration, which combines both difference feature vector $F_{diff}$ and commonality feature vector $F_{comm}$ , surpasses the human-derived features by $20\%$ . This underscores the effectiveness of our automatic feature extraction approach in capturing relevant and distinguishing characteristics for classification tasks. These results also suggest that, in scenarios with limited resources or new visual concepts, we can automate the process of feature extraction, reducing the need for manual human annotation. + +Classifier Choices and Feature Mapping Models. We investigate the model performance using different choices for visual classifiers and feature mapping models. From Table 5, we notice that LLaVA, as a feature mapping model for obtaining binary feature vectors, achieves the best overall performance when paired with the logistic regression. + +As mentioned in Sec. 3.1, we can also leverage CLIP as a feature mapping model to convert images into features. When predicting, one can either set a threshold to determine if the image possesses a certain attribute, resulting in a binary feature, or preserve the similarity between the image and each attribute to form continuous features. Although the performance of using CLIP features is slightly lower + +![](images/fad7c3a0303193d993db182e43316478a00645c8d5290978d08932ea8001dd41.jpg) +Figure 3. Qualitative examples of the features extracted by VRL. We highlight the key attributes in bold. (a) Verbalized features extracted by comparing images from different classes. (b) Verbalized features extracted by comparing the images within the same class. + +![](images/1573def6c59fdf90683a5326ff1b74994857dd5d7d3d022787b08c1ecec77c67.jpg) + +
Methodv1v2Avg.
Human73.852.563.15
VRL-Fdiff88.475.281.80
VRL-Fcomm89.274.081.60
VRL-all89.476.683.00
+ +compared to LLaVA, CLIP offers a notable advantage in terms of inference speed. This is because CLIP supports batch-wise operations when computing similarities between a set of images and a set of verbalized features, which offers our method the flexibility to scale with a larger dataset or an increasing number of object classes. It is worth noting that using CLIP as the feature mapping model still allows our method to surpass previous state-of-the-art results by significant margins, achieving improvements of $8\%$ and $12\%$ on the iNaturalist and Kiki-Bouba datasets, respectively. This highlights the universal applicability of VRL-extracted features, regardless of the feature mapping approach used. + +# 4.6. Qualitative Analysis + +Figure 3 presents qualitative examples of what verbalized features are extracted from image pairs. We list some of the features that have top-5 feature importance in our trained logistic regression model. For inter-class differences (Fig. 3 (a)), verbalized features are used to capture distinguishing characteristics between different species. For example, the Lichen (Fungi) category is distinguished based on the structure of the thallus, where one image describes a flat, crustlike thallus, and the other describes a complex, branch-like thallus. Similarly, in the Wild Rye (Grass) category, the differences between a slender stem with a feathery inflorescence and a more robust stem with a dense, elongated seed head are verbalized. For intra-class commonality (Fig. 3 (b)), verbalized features are employed to capture shared + +Table 4. Accuracy (%) of using human-labeled attributes and VRL extracted features on Kiki-Bouba dataset. Note that v1 and v2 indicates two different splits described in Sec. 4.1. + +
MethodRFLRMLP
iNaturalist (Prior Works: 50.1)
LLaVA64.967.466.0
CLIP (binary)52.258.055.2
CLIP (continuous)52.853.660.6
Kiki-Bouba (Prior Works: 69.3)
LLaVA83.6586.6585.65
CLIP (binary)79.2581.5081.10
CLIP (continuous)77.0577.1081.85
+ +Table 5. Accuracy (\%) of different classifier choices (columns) and models used to map verbalized features to numeric vectors (rows). LLaVA and CLIP are utilized as mapping models to map an image to a feature vector. + +traits within the same species. For example, the leaves of Manzanita (Berry) are consistently described as “ovate with a pointed tip and slightly serrated edges,” while the inflorescences of Bulrush (Herb) are uniformly described as having a “delicate, feathery appearance.” These verbalized features are used to guide the decision process by helping the model focus on important distinguishing features or shared traits within each class. Due to page limit, see Appendix B.5 for human study on the learned verbalized features. + +# 5. Conclusion + +In this paper, we propose Verbalized Representation Learning (VRL), which enables automatic interpretable feature extraction with few-shot samples. By leveraging VLMs, VRL generates verbalized features that capture both inter-class differences and intra-class commonalities. Our method not only enhances the model's adaptability with limited data but also provides transparency in the decision-making process, enabling easier interpretation of the features that influence predictions. Our experiments show that VRL outperforms prior approaches, achieving superior results while using significantly less data. This includes tasks like fine-grained recognition and novel concept adaptation, demonstrating its potential for real-world applications. + +# Acknowledgements + +We thank anonymous reviewers and other members of UCLA-NLP+ group for their helpful comments. This work was partially supported by U.S. DARPA ECOLE Program No. #HR00112390060, ONR grant N00014-23-1-2780, DARPA ANSR program FA8750-23-2-0004, and Apple. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Morris Alper and Hadar Averbuch-Elor. Kiki or bouba? sound symbolism in vision-and-language models. Advances in Neural Information Processing Systems, 36, 2024. 2, 5 +[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2 +[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2 +[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3 +[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 3, 4 +[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 4 +[9] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. arXiv preprint arXiv:2303.14123, 2023. 4 + +[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2, 3, 4 +[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. Solo: A single transformer for scalable vision-language modeling. In Transactions on Machine Learning Research, 2024. 2 +[12] Mia Chiquier, Utkarsh Mall, and Carl Vondrick. Evolving interpretable visual classifiers with large language models. arXiv preprint arXiv:2404.09941, 2024. 1, 2, 5, 6, 7 +[13] Fahim Dalvi, Nadir Durrani, Hassan Sajjad, Yonatan Belinkov, Anthony Bau, and James Glass. What is one grain of sand in the desert? analyzing individual neurons in deep nlp models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6309-6317, 2019. 2 +[14] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5 +[15] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 1, 2 +[16] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. Advances in neural information processing systems, 20, 2007. 1 +[17] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 2 +[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 4 +[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 4 +[20] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In Proceedings of the European conference on computer vision (ECCV), pages 264-279, 2018. 2 +[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2 +[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[23] Shaoli Huang, Zhe Xu, Dacheng Tao, and Ya Zhang. Part-stacked cnn for fine-grained visual categorization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1173-1182, 2016. 2 + +[24] Andrej Karpathy, Justin Johnson, and Li Fei-Fei. Visualizing and understanding recurrent networks. arXiv preprint arXiv:1506.02078, 2015. 2 +[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3 +[26] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 2 +[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 2 +[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 4, 5 +[29] Manling Li, Ruochen Xu, Shuohang Wang, Luowei Zhou, Xudong Lin, Chenguang Zhu, Michael Zeng, Heng Ji, and Shih-Fu Chang. Clip-event: Connecting text and images with event structures. In Proc. Conference on Computer Vision and Pattern Recognition (CVPR2022), 2022. 2 +[30] Fan Liu, Wenwen Cai, Jian Huo, Chuanyi Zhang, De long Chen, and Jun Zhou. Making large vision language models to be good few-shot learners. arXiv preprint arXiv:2408.11297, 2024. 4 +[31] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. arXiv preprint arXiv:2408.12469, 2024. 4 +[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 1, 2, 5, 6 +[33] Eduardo Navarrete, Michele Miozzo, and Francesca Peressotti. Language can shape the perception of oriented objects. Scientific reports, 10(1):8409, 2020. 1, 3 +[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2 +[35] OpenAI. ChatGPT. 2022. 5 +[36] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8779-8788, 2018. 2 +[37] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 3 +[38] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 1, 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4 +[40] Vilayanur S Ramachandran and Edward M Hubbard. Synaesthesia-a window into perception, thought and language. Journal of consciousness studies, 8(12):3-34, 2001. 5 +[41] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5 +[42] Bernardino Romera-Paredes and Philip Torr. An embarrassingly simple approach to zero-shot learning. In International conference on machine learning, pages 2152-2161. PMLR, 2015. 2 +[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2 +[44] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[45] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2 +[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5 +[47] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5 +[48] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3090-3100, 2023. 1, 2 +[49] Cheng-Fu Yang, Yao-Hung Hubert Tsai, Wan-Cyuan Fan, Russ R Salakhutdinov, Louis-Philippe Morency, and Frank Wang. Paraphrasing is all you need for novel object captioning. Advances in Neural Information Processing Systems, 35:6492–6504, 2022. + +[50] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 5, 6 +[51] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2 +[52] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2 +[53] Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. Sglang: Efficient execution of structured language model programs. arXiv preprint arXiv:2312.07104, 2023. 1 \ No newline at end of file diff --git a/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/images.zip b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d53e8f2b5db458c8a2f9c5d25a82071aa7008de --- /dev/null +++ b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebc12cc2e9247f32c5a71e5c1a7884460ee9018c6a0ba481427aa0576ba44de5 +size 508583 diff --git a/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/layout.json b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c3e2d5efa2bcbdaf05b7e49df12dfe57c2fa8362 --- /dev/null +++ b/2025/Verbalized Representation Learning for Interpretable Few-Shot Generalization/layout.json @@ -0,0 +1,7348 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 549, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 549, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 549, + 121 + ], + "type": "text", + "content": "Verbalized Representation Learning for Interpretable Few-Shot Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": "Cheng-Fu Yang" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": " Da Yin" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": " Wenbo Hu" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": " Heng Ji" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": " Nanyun Peng" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": " Bolei Zhou" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": " Kai-Wei Chang" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": "University of California, Los Angeles " + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 561, + 173 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 175, + 361, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 175, + 361, + 185 + ], + "spans": [ + { + "bbox": [ + 261, + 175, + 361, + 185 + ], + "type": "text", + "content": "cfyang@cs.ucla.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 202, + 552, + 302 + ], + "blocks": [ + { + "bbox": [ + 61, + 202, + 552, + 302 + ], + "lines": [ + { + "bbox": [ + 61, + 202, + 552, + 302 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 552, + 302 + ], + "type": "image", + "image_path": "a2d3c95d96c71efe3946fbc88d9ec5631661e2b12cae9bb15c6c85703bcd3b71.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 311, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 54, + 311, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 555, + 377 + ], + "type": "text", + "content": "Figure 1. (a) Humans can identify key differences between similar species and recognize common traits within a class, even when appearance varies, and express these insights in concise language. Similarly, our Verbalized Representation Learning (VRL) extracts meaningful features by querying Vision-Language Models (VLMs) to capture inter-class discriminative features and intra-class commonalities. (b) These verbalized features are mapped to numerical vectors through the VLM and can be used by downstream classifiers. Our VRL demonstrates superior performance in few-shot classification compared to prior work, achieving a " + }, + { + "bbox": [ + 54, + 311, + 555, + 377 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 54, + 311, + 555, + 377 + ], + "type": "text", + "content": " absolute gain on iNaturalist and a " + }, + { + "bbox": [ + 54, + 311, + 555, + 377 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 54, + 311, + 555, + 377 + ], + "type": "text", + "content": " on Novel Object Classification, even with a much smaller model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 388, + 200, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 388, + 200, + 400 + ], + "spans": [ + { + "bbox": [ + 151, + 388, + 200, + 400 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "text", + "content": "Humans recognize objects after observing only a few examples, a remarkable capability enabled by their inherent language understanding of the real-world environment. Developing verbalized and interpretable representation can significantly improve model generalization in low-data settings. In this work, we propose Verbalized Representation Learning (VRL), a novel approach for automatically extracting human-interpretable features for object recognition using few-shot data. Our method uniquely captures interclass differences and intra-class commonalities in the form of natural language by employing a Vision-Language Model (VLM) to identify key discriminative features between different classes and shared characteristics within the same class. These verbalized features are then mapped to numeric vectors through the VLM. The resulting feature vectors can be further utilized to train and infer with downstream classifiers. Experimental results show that, at the same model scale, VRL achieves a " + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "text", + "content": " absolute improvement over prior state-of-the-art methods while using " + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "text", + "content": " less data and a smaller model. Furthermore, compared to human-labeled attributes, the features learned by VRL exhibit a " + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "text", + "content": " absolute gain when used for downstream classification tasks." + }, + { + "bbox": [ + 54, + 414, + 297, + 689 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 388, + 394, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 394, + 399 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 394, + 399 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 411, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 555, + 614 + ], + "type": "text", + "content": "Humans have a remarkable capability to recognize certain objects after seeing only a few examples. As suggested by [33], it is significantly enhanced by inherent language understanding. Language, with its representational strength, serves as a primary resource for conveying knowledge about visual objects. A single, precise description can effectively capture visual distinctions observed across different object categories. For example, in Fig. 1 (a), a human can identify the key difference between two fish species based on the pattern of their heads, even if both share a white stripe around the neck and a yellow body. As a result, we argue that developing verbalized features offers a valuable complement to improve model's generalization under low-resource conditions. In addition, incorporating language into image classification models also increases the interpretability of the visual system, facilitating transparent and reliable decision making and effective model auditing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Prior works [15, 16] have explored integrating language into visual classifiers with a bottleneck of textual attributes, where the model first predicts relevant attributes and subsequently uses them to classify the image. However, these attributes often require human annotations or predefined ontologies. Development of Large-scale Vision-and-Language Models (VLMs) like CLIP [39] and GPT-4V [1] have enabled recent bottleneck methods [12, 32, 38, 48, 50]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 266, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 266, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 266, + 713 + ], + "type": "text", + "content": "1The code is available at: https://github.com/joeyy5588/VRL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1602" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "to generate relevant attributes by leveraging the prior knowledge of VLMs learned during the pre-training stage. However, these attributes are often ungrounded for the testing images and rely on VLM's prior knowledge, resulting in low precision when applied to fine-grained or novel concept recognition where objects are underrepresented in the pre-training datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 295, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 295, + 457 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose Verbalized Representation Learning (VRL) for automatic, human-interpretable feature extraction using only few-shot data. It applies to fine-grained and novel objects and could work with local models such as LLaVA [28]. Specifically, inspired by self-supervised representation learning (SSL) in computer vision, such as SimCLR [8], MoCo [19], SwAV [6], SimSiam [10], BYOL [18], CLIP-Event [29], and SOLO [11], we propose to leverage a VLM to capture the inter-class difference and intra-class commonality and articulate these findings in natural language, as illustrated in Fig. 1. Specifically, we cast the VLM to describe the key difference between two images from different classes, which would preserve the discriminative features and remove the redundant ones, such as the yellow body shared by both species on the left side of Fig. 1 (a). Conversely, we employ the VLM to list the key features that are shared by two images within the same class, extracting features that are robust to intra-class variance, such as the orange facial pattern observed in all of the Thalassoma Pavo. Notably, this process can be applied to any two images, whether from different classes or the same class, which allows us to exponentially scale the number of useful verbalized features with the image samples, enabling effective generalization even in few-shot settings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 294, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 294, + 590 + ], + "type": "text", + "content": "Consequently, to obtain the numerical feature embedding of an image using our VRL, we employ a VLM to determine whether the image possesses the characteristics described by the verbalized features, as depicted in Fig. 1 (b). Each dimension in the resulting embedding represents a scalar value indicating the presence of the described feature. This approach effectively transforms the verbalized features into numeric representations that can serve as inputs to any classification methods, such as logistic regression, random forest, or MLP classifiers, allowing flexible and robust modeling based on the extracted interpretable features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "We conduct experiments on iNaturalist [47] and KikiBouba dataset [2], where the former includes objects that only have subtle differences between classes, while the latter contains novel objects that are nearly not present in the web-scale datasets. Compared to previous state-of-the-art (SoTA) baselines using 70B models, we achieve a " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " absolute improvement while using " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " less data and a much smaller model with 7B parameters. Against supervised finetuned baselines, we observe a " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " absolute improvement. Lastly, when compared to human annotated attributes, the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "attributes learned by our VRL demonstrate a " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " absolute gain. These results illustrate that the features extracted by VRL not only offer superior effectiveness but also exhibit strong robustness, providing a generalizable solution when adapting to fine-grained or novel classification tasks with limited data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 150, + 400, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 150, + 400, + 162 + ], + "spans": [ + { + "bbox": [ + 314, + 150, + 400, + 162 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 170, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 410 + ], + "type": "text", + "content": "Natural Language for Image Classification. A common approach for integrating language into visual classifiers involves creating a concept bottleneck [26]. Early bottleneck methods have been extensively applied in few-shot or zero-shot classification models [7, 15-17, 23, 26, 27, 42]. However, these methods typically rely on manually annotated or predefined attributes, which can limit their adaptability to novel classes or fine-grained tasks. Recent advancements in VLMs have enabled researchers to directly sample interpretable descriptive features from these models [12, 32, 38, 45, 48-51]. This approach benefits from incorporating external knowledge embedded in the pretrained datasets. However, it often relies heavily on the prior knowledge encoded within these models, which can lead to the generation of ungrounded features and difficulty generalizing when the target data is not well-represented in the training datasets. In contrast, our method directly learns grounded, verbalized features from the visual data, allowing for automatic interpretable feature extraction that remains adaptable even for novel classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 411, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 639 + ], + "type": "text", + "content": "Interpreting Model's Decision Process. There has been a long line of research aiming to improve the interpretability and explainability of deep models. Pioneering methods [3, 4, 13, 24, 43, 52] have tackled this challenge by visualizing learned features, categorizing maximally-activating inputs, and identifying key neurons that drive model decisions. More recently, efforts have been made to interpret model behavior using natural language [20, 21, 36, 44], enabling explanations that are more accessible and editable. However, the aforementioned methods are generally post-hoc, meaning they are applied after the model has been trained with abundant data. While post-hoc explanations can provide valuable insights, they may lack the ability to influence or shape the model's internal representations during training. Instead, our methods embed interpretability directly into the learning process by learning verbalized features. This enables the model to produce inherently explainable and contextually grounded representations without relying solely on retrospective analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Self-Supervised Learning Models. Self-supervised learning (SSL) has emerged as a powerful approach in computer vision, enabling models to learn robust feature representations from unlabeled data. Contrastive learning [34] methods, such as SimCLR [8] and MoCo [19], achieve this goal by learning to discriminate between different samples (neg" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1603" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 61, + 237, + 264 + ], + "blocks": [ + { + "bbox": [ + 72, + 61, + 237, + 264 + ], + "lines": [ + { + "bbox": [ + 72, + 61, + 237, + 264 + ], + "spans": [ + { + "bbox": [ + 72, + 61, + 237, + 264 + ], + "type": "image", + "image_path": "49ec2f97e95200572596d5b59352c13f5bc971403f9eb9444716c8cd482618e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 266, + 242, + 274 + ], + "lines": [ + { + "bbox": [ + 234, + 266, + 242, + 274 + ], + "spans": [ + { + "bbox": [ + 234, + 266, + 242, + 274 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 61, + 411, + 264 + ], + "blocks": [ + { + "bbox": [ + 238, + 61, + 411, + 264 + ], + "lines": [ + { + "bbox": [ + 238, + 61, + 411, + 264 + ], + "spans": [ + { + "bbox": [ + 238, + 61, + 411, + 264 + ], + "type": "image", + "image_path": "3248b61fe0761815c8b7e7d1614d6f8279951fd77cac75e3a250e0104aaf1063.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. The overview of our Verbalized Representation Learning (VRL) framework. (a) Given " + }, + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "type": "text", + "content": " samples per class from " + }, + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "type": "text", + "content": " different classes, VRL is able to generate a diverse set (exponentially scaling with " + }, + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "type": "inline_equation", + "content": "N \\times C" + }, + { + "bbox": [ + 54, + 283, + 555, + 340 + ], + "type": "text", + "content": ") of verbalized features by: 1) extracting key differences between samples from different classes, and 2) identifying commonalities shared among objects within the same class. (b) Given an image, a Vision-and-Language model (VLM) is employed to evaluate whether the image contains the characteristics described by the verbalized features. This process can map a set of verbalized features into numeric representations which can then be used for downstream tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 421, + 59, + 541, + 268 + ], + "blocks": [ + { + "bbox": [ + 421, + 59, + 541, + 268 + ], + "lines": [ + { + "bbox": [ + 421, + 59, + 541, + 268 + ], + "spans": [ + { + "bbox": [ + 421, + 59, + 541, + 268 + ], + "type": "image", + "image_path": "e8c179c428e489c9cafeea97b77c04b3a2ab6cce4a3ef76fcb9dd385cefc34fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 348, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 296, + 563 + ], + "type": "text", + "content": "ative pairs) and augmentations of the same sample (positive pairs). Chen et al. [10] have found that the key function of negative pairs is to prevent the model from learning collapsing features, where models produce constant or trivial outputs. As a result, subsequent works have relieved the need for negative samples by employing techniques like clustering [5, 6], momentum update [18], or stop gradient [10]. These methods focus on learning the underlying shared representations between the augmentations of the same sample. Our approach can be viewed as a variant of supervised contrastive learning [25], where the positive samples are drawn from different images from the same class. However, our method introduces two distinct advantages. First, it is data-efficient, as it does not require large datasets for gradient updates. Second, the expressiveness of verbalized features inherently prevents the model from collapsing to constant outputs, thereby ensuring the resulting features are meaningful and representative." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 111, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 111, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 111, + 586 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "In Sec. 3.1, we first introduce the motivation and concept of verbalized representation learning (VRL) and then discuss how VRL automatically extracts interpretable, compact representations with few-shot data. Then, in Sec. 3.2, we outline the process of building a visual classifier using the derived features. Furthermore, we demonstrate that these extracted features are versatile and can be applied to arbitrary classification models, including but not limited to logistic regression, decision trees, and multi-layer perceptron (MLP) classifiers. The overall framework of our VRL" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 348, + 403, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 403, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 403, + 361 + ], + "type": "text", + "content": "is presented in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 368, + 507, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 507, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 507, + 382 + ], + "type": "text", + "content": "3.1. Verbalized Representation Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 386, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 530 + ], + "type": "text", + "content": "Humans can recognize objects after seeing only a few examples, a skill boosted by language understanding [33]. Language effectively conveys visual distinctions, even without many extra visual cues. Incorporating language also enhances image classification interpretability, enabling clearer, more reliable decisions. To this end, we propose a method that verbalizes key visual features in natural language. In this section, we focus on identifying which visual features are crucial for few-shot image classification and how to extract and verbalize them effectively, and how to map these verbalized features into vectors which would be later utilized during training and inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "What visual features should be verbalized? Inspired by self-supervised representation learning methods, where model learns robust features through contrasting positive and negative pairs [8, 19] or maximizing the similarity of augmented views of the positive samples [6, 10, 18], we propose to verbalize the objective functions of these self-supervised learning methods with the help of VLMs such as LLaVA. Specifically, given a classification task with " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " classes and " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "-shot examples per class, each data point is defined as " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "(x,l)" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " represents an image, and " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "l \\in \\{1,2,\\dots,C\\}" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " is the categorical label indicating the class to which the image belongs. Our method leverages two types of paired images: positive pairs, defined as image pairs " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "(x_{i},x_{j})" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " with the same label " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "(l_{i} = l_{j})" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": ", and negative pairs, which consist of images " + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "inline_equation", + "content": "(x_{i},x_{k})" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": " from different" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1604" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "classes " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "(l_i\\neq l_k)" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ". Notably, with this pairing strategy, we can form " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2^C\\times \\mathbf{C}_2^N" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " distinct negative pairs, and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "C\\times \\mathbf{C}_2^N" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " positive pairs, which significantly increase the data utilization under the few-shot setting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": "Inspired by contrastive learning methods [8, 19], for each negative pair sample, our approach emphasizes capturing inter-class differences by tasking the VLM to describe key distinguishing features between images from different classes, denoted as " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "y_{diff} = \\mathrm{VLM}(\\mathbf{x}_i,\\mathbf{x}_k,q_{diff})" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "l_i\\neq l_k" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "q_{diff}" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": " denotes the query that captures the visual differences between the two input images " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": ". For instance, in Fig. 2 (a), the model learns to distinguish two fish species by the coloration and the pattern of the fish. Conversely, for each positive pair, we draw inspirations from negative sample-free SSL methods [6, 10, 18], where we employ the VLM to capture intra-class commonalities, generating descriptions of the key shared features between two images from the same class, i.e., " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "y_{comm} = \\mathrm{VLM}(\\mathbf{x}_i,\\mathbf{x}_j,q_{comm})" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "l_i = l_j" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "q_{comm}" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": " denotes the query that captures the visual commonalities between " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 56, + 122, + 295, + 361 + ], + "type": "text", + "content": ". For example, in Fig. 2 (a), VRL identifies that both fish possess irregular orange spots around their face. We include the detailed prompt templates used in the above process in Appendix A.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "y_{diff}" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "y_{comm}" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": " learn complimentary features. " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "y_{diff}" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": " captures the most discriminative features while filtering out redundant ones, such as the yellow body present in both types of fish, leading to less noisy features. While " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "y_{comm}" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": " is robust to the intra-class variance where it learns to identify the fish based on the shared features, like the pattern on its face, rather than the number of stripes on its back. In the later experimental section, we will further verify this assumption and show that these two features can yield the best performance when combined. In addition, as discussed in the earlier paragraph, even with only " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": "-shot samples, our VRL can sample a diverse set of features from " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2^C \\times \\mathbf{C}_2^N" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": " negative and " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "C \\times \\mathbf{C}_2^N" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": " positive pairs, making our method excel under low-resource setting. Moreover, we empirically discover that sampling " + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "inline_equation", + "content": "C \\times N" + }, + { + "bbox": [ + 55, + 362, + 295, + 564 + ], + "type": "text", + "content": " pairs for both negative and positive samples is sufficient to collect a robust set of features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "How to map the verbalized features to numeric feature embeddings? To obtain feature vectors that can be later used to build visual classifiers, we transform the verbalized features into numeric vectors with the help of VLMs such as CLIP [39] and LLaVA [28]. Given an image and a set of verbalized features, which are in the form of language descriptions, we employ a VLM to determine whether the image possesses those features, as illustrated in Fig. 2 (b), resulting in a feature embedding " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "F = \\mathrm{VLM}(y,x)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ". Each dimension of " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " is produced by the VLM, indicating the presence of a certain verbalized feature from " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "y_{diff}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "y_{comm}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " or assessing the degree that the image has a certain" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "feature. Concretely, for generative VLMs like LLaVA, each verbalized feature is mapped to 0 or 1, based on whether the model infers the presence of the feature in the image. For VLMs like CLIP, feature embeddings are derived by calculating the similarity between each verbalized feature and the image using CLIP's visual encoders. This results in a continuous vector indicating the likelihood that the image contains each feature, with an optional similarity threshold to convert it into binary feature vectors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 190, + 444, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 444, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 444, + 202 + ], + "type": "text", + "content": "3.2. Training and Inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 208, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 553, + 376 + ], + "type": "text", + "content": "Training. Once the numeric feature vectors are generated, they are used as input for training various visual classifiers, including, but not limited to, logistic regression, random forests, Naive Bayes, K-nearest neighbors (KNN), and MLP classifiers. These models learn to predict the class labels based on the interpretable feature vectors. This flexibility enables robust modeling that can be tailored to different applications. Moreover, this approach not only enables the construction of classifiers for novel concepts with few-shot data, but also provides insight into which visual features are most important for decision-making. For instance, we can extract feature importance from logistic regression or visualize decision paths in decision tree classifiers, enhancing the interpretability of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 380, + 553, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 608 + ], + "type": "text", + "content": "Inference. During inference, given a testing image, we generate its feature vector by following the approach described in Sec. 3.1. Specifically, we query the VLMs with the learned verbalized features to determine whether the image contains specific features, producing a numeric feature vector. This vector can either be continuous, representing the likelihood of a feature's presence, or binary, indicating the feature's absence or presence. The resulting vector is then passed to the trained classifier to make a prediction. Notably, to enhance the model robustness, we ensemble results from classifiers trained with different algorithms. Combining predictions from multiple classifiers is beneficial because it reduces the likelihood of overfitting to any specific algorithm's biases or weaknesses. The ensemble can use either hard or soft voting: In hard voting, each classifier makes a discrete prediction, and the final prediction is determined by a majority vote. In soft voting, the prediction logits from each classifier are averaged to produce the final output, allowing for more nuanced decision-making." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 620, + 394, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 394, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 394, + 634 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Our experiments are designed to answer the following research questions: (i) How well does VRL generalize to tasks requiring fine-grained recognition with few-shot data? (ii) Can VRL effectively adapt to tasks involving novel concepts that were not part of the VLM's pre-training resources, using only few-shot data? (iii) How effective is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1605" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "VRL compared to conventional few-shot adaptation algorithms like LoRA fine-tuning or in-context learning? (iv) Do different types of features extracted by VRL mutually benefit each other, and do the extracted features further improve on top of the other commonly used features? (v) How does the performance of automatically features extracted with VRL compare to human-labeled features?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 234, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 234, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 234, + 173 + ], + "type": "text", + "content": "4.1. Dataset and Evaluation Protocols" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 294, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 294, + 258 + ], + "type": "text", + "content": "We conduct experiments on iNaturalist and Kiki-Bouba to validate our method's effectiveness under varied scenarios, with all reported numbers representing classification accuracy (\\%). Although our primary analysis centers on these datasets, additional results for general few-shot classification on mini-ImageNet [41] are provided in the Appendix B.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 348 + ], + "type": "text", + "content": "Fine-Grained Species Classification. For fine-grained classification, we utilize the iNaturalist 2021 dataset [47], which comprises a diverse collection of images and annotations contributed by citizen scientists, spanning numerous species of animals, plants, and fungi. Each species in the training split contains between 200 and 300 images, while the validation split includes 10 images per species." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 539 + ], + "type": "text", + "content": "Following [12], we experiment on images from five different families, each containing five to six species. These families are selected due to their challenging nature: distinguishing between species within the same family requires the model to identify complex features such as shapes and patterns, rather than simple color variations. The families used in our experiments are as follows: Lichen (fungi), Wrasse (fish), Wild rye (grass), Manzanita (berry shrubs), and Bulrush (herbs). For a detailed list of the species, please refer to the appendix. Notably, unlike [12], which utilizes a full training set containing between 200 and 300 images per species, our method uses on only 10 images per species to achieve generalization in a low-resource setting. We will report the baseline performance from their original paper as well as our reproduced results using the same limited data as our method for comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Novel Concept Classification. To test the model's generalizability on objects that have been rarely seen in the pretraining image resources, we evaluate our method using the Kiki-Bouba [2] dataset. The Kiki-Bouba experiment, originally introduced by [40], illustrates that people often associate specific shapes with different sounds. The dataset was constructed by prompting generative models trained to create 3D-rendered images from non-existent, meaningless words [2], providing a unique testbed for evaluating generalization to novel and abstract concepts. Following the approach in [12], we evaluate our method on two distinct splits, each containing five different classes. The first split consists of the classes bamaba, duludu, gaduga, lomulo, nomano, while the second split includes bouba, galaga," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "kepike, kiki, maluma. In [12], the training set comprises 800 images per class, with 200 images per class for validation. In contrast, we adopt a similar low-resource setting as used for the iNaturalist dataset, with our method accessing only 10 training images per class. Baseline results for both settings will be reported in the table for comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 150, + 552, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 552, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 552, + 163 + ], + "type": "text", + "content": "4.2. Baseline Methods and Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 166, + 553, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 553, + 297 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our method, we compare it against several baselines, including the CLIP-based approach proposed in [12] and other baseline methods that utilize LLMs, such as LaBo [50], CBD [32] and LMMutate [12], which leverage LLM's pre-trained knowledge to generate attributes. Additionally, since our method leverages LLaVA-OneVision [28] to generate verbalized features, we perform quantitative comparisons with common approaches used to adapt LLaVA for downstream tasks, such as LoRA fine-tuning [22] and in-context learning [14]. We describe the details of each baseline method as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "type": "text", + "content": "CLIP-Based Methods. For CLIP-based baselines, we consider two variants. CLIP Class Name is a naive baseline where CLIP is employed to compute the similarity between the species' scientific name and the images. The best accuracy is reported as the highest value achieved among using the common name, the scientific name, or both names combined. The second variant involves Prompt Tuning with a CLIP encoder, where the text embeddings of specific class-related tokens are optimized through gradient descent. For both baselines, we present results as reported in [12]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 429, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 609 + ], + "type": "text", + "content": "Language Bottleneck Methods. Classification by Description [32], LaBo [50], and LLM-Mutate [12] generate a list of attributes by prompting LLMs like GPT [35] and Llama [46] with the class name of the object. While this allows them to leverage the prior knowledge learned by LLMs, the generated features often lack grounding in the actual images. LaBo and LLM-Mutate alleviate this issue by using submodular optimization and evolutionary algorithm, respectively, to filter out ungrounded attributes. However, these methods still rely heavily on the capabilities and learned knowledge of the pre-trained LLMs. For LaBo, we adapt the official implementation for our datasets; for LLM-Mutate, we not only report the results from the original paper but also evaluate the method using the same limited data as our approach, ensuring a fair comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 615, + 554, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 723 + ], + "type": "text", + "content": "LLaVA-Based Methods. We include two common approaches for adopting LLaVA. The first baseline involves performing LoRA [22] fine-tuning (LLaVA-SFT), which has been demonstrated to be effective in scenarios with limited data. The second baseline is in-context learning [14] (LLaVA-ICL), where one exemplar image for each species or object is included in the prompt to guide the model's predictions. We study which of the LLaVA-based methods and VRL better utilizes LLaVA's capabilities." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1606" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 60, + 528, + 277 + ], + "blocks": [ + { + "bbox": [ + 82, + 60, + 528, + 277 + ], + "lines": [ + { + "bbox": [ + 82, + 60, + 528, + 277 + ], + "spans": [ + { + "bbox": [ + 82, + 60, + 528, + 277 + ], + "type": "table", + "html": "
MethodLichenWrasseWild RyeManzanitaBulrushAverage
Zero-Shot Methods
CLIP Class Name [12]23.332.032.026.026.027.86
Full Dataset Methods (200+ Images per Species)
CLIP Prompt Tuning [12]23.320.040.020.020.024.66
Classification by Description [32]30.034.036.028.020.029.60
LLM-Mutate-70B [12] (1-prompt)31.624.044.040.022.032.32
LLM-Mutate-70B [12] (10-prompt)48.344.058.058.042.050.06
Few-Shot Methods (10 Images per Species)
LLM-Mutate-7B† (10-prompt)35.048.038.044.026.038.20
LLM-Mutate-70B† (10-prompt)46.644.046.044.040.044.13
LaBo-7B† [50]45.048.044.042.036.043.00
LaBo-70B† [50]56.646.064.052.040.051.72
LLaVA-ICL-7B16.628.022.018.030.022.92
LLaVA-SFT-7B41.650.058.042.028.043.92
LLaVA-VRL-7B (Ours)58.348.074.066.046.058.46
LLaVA-VRL-72B (Ours)71.672.074.056.066.067.92
", + "image_path": "3684f3a1d776f7db2f4c36517900e6f4c1a47944ebab35108ba68c79016c0c61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 553, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 278, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 553, + 310 + ], + "type": "text", + "content": "Table 1. Comparison of classification accuracy (%) across different methods for fine-grained classification on iNaturalist. The table presents results for zero-shot, full-dataset (200+ images per species), and few-shot (10 images per species). Results marked with † denote values reproduced using the official implementation but restricted to the same few-shot data as our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 324, + 294, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 324, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 294, + 467 + ], + "type": "text", + "content": "Our Method. We implement VRL with LLaVA-OneVision for capturing visual difference and commonality features. We also select LLaVA-OneVision as the feature mapping model that assists in converting an image into a feature vector, and the base model for all the LLaVA-based baseline methods. For the few-shot training stage, we experiment with various classification methods including logistic regression or MLP. Unless otherwise specified, we utilize both the difference features " + }, + { + "bbox": [ + 54, + 324, + 294, + 467 + ], + "type": "inline_equation", + "content": "F_{diff}" + }, + { + "bbox": [ + 54, + 324, + 294, + 467 + ], + "type": "text", + "content": " and commonality features " + }, + { + "bbox": [ + 54, + 324, + 294, + 467 + ], + "type": "inline_equation", + "content": "F_{comm}" + }, + { + "bbox": [ + 54, + 324, + 294, + 467 + ], + "type": "text", + "content": ". We report the best performance achieved across different classifiers, selecting the optimal results from a single classifier." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 471, + 244, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 244, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 244, + 483 + ], + "type": "text", + "content": "4.3. Fine-Grained Species Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "content": "We present the results for fine-grained species classification on the iNaturalist dataset in Table 1. We find that compared to previous SoTA methods, even with " + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "content": " less data (10 images per species compared to " + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "content": " images per species), and using a significantly smaller model (7B vs. 70B parameters), our method can already surpass the previous SoTA by " + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "content": ". When we increase the scale of our method to a comparable 72B model, we can further increase the improvements to nearly " + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "content": ". Moreover, when baseline methods are limited to the same data availability as our approach, our 7B and 72B models achieve improvements of " + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 56, + 486, + 295, + 664 + ], + "type": "text", + "content": ", respectively. These results highlight the advantage of using grounded descriptions to capture the subtle visual cues, especially when the model needs to discover visual nuances to discriminate species within the same family." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "We also compare VRL against LLaVA-based baselines. Using the same backbone LLaVA model, our method showcases a " + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": " gain to LLaVA-SFT and a " + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": " advantage over the in-context learning method, LLaVA-ICL. In addition," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 324, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 469 + ], + "type": "text", + "content": "VRL is easier to be scaled up than the baselines. For instance, VRL with the 72B model requires 8 GPUs with 48 GB of RAM, while fine-tuning the 13B model with LoRA already requires 8 GPUs with 80 GB of RAM. For in-context learning, including image examples for every category in the prompt significantly increases the memory usage of the method, making it infeasible to perform in-context learning on the 72B LLaVA with the 8 GPUs (48 GB). The model performance may also be constrained by the context window of LLaVA model. These results and analysis further validate the scalability and robustness of VRL comparing to those existing adaptation algorithms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 474, + 471, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 471, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 471, + 487 + ], + "type": "text", + "content": "4.4. Novel Concept Classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": "For novel concept classification, we report the results on the Kiki-Bouba dataset in Table 2. Similar to the performance trend observed in the previous experiment, our method surpasses the previous SoTA by " + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": " with a smaller model, but with " + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": " less data (10 images per species compared to " + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "inline_equation", + "content": "800+" + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": " images per species). Surprisingly, we observe that increasing the number of model parameters does not lead to better performance on this task. One possible explanation for this is that, rather than relying on prior knowledge learned during pre-training, the model needs to focus on learning new, critical features specific to these novel objects with unfamiliar shapes and patterns. Since the objects in this task are unseen during LLM training, existing models struggle to generate relevant attributes and fail to ground visual differences necessary for distinguishing objects from different classes. This hypothesis is further supported by the larger performance gap between our method and those that rely on LLM's pre-trained knowledge [32]. On this dataset, our method exceeds their performance by " + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 487, + 555, + 713 + ], + "type": "text", + "content": ", compared" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1607" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 60, + 288, + 255 + ], + "blocks": [ + { + "bbox": [ + 63, + 60, + 288, + 255 + ], + "lines": [ + { + "bbox": [ + 63, + 60, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 63, + 60, + 288, + 255 + ], + "type": "table", + "html": "
Methodv1v2Avg.
Zero-Shot Methods
CLIP Class Name38.738.838.75
Full Dataset Methods (800+ Images per Object)
CLIP Prompt Tuning16.755.636.15
Classification by Description28.836.832.80
LLM-Mutate-70B (1-prompt)50.347.849.05
LLM-Mutate-70B (10-prompt)79.259.469.30
Few-Shot Methods (10 Images per Object)
LLM-Mutate-7B† (10-prompt)63.359.261.25
LLM-Mutate-70B† (10-prompt)67.359.263.25
LLaVA-ICL-7B24.527.225.85
LLaVA-SFT-7B72.150.661.35
LLaVA-VRL-7B (Ours)89.476.683.00
LLaVA-VRL-72B (Ours)89.174.781.90
", + "image_path": "e6a392247c41a85acfd5f385243260cd85a4e8666979dc275961a58104147880.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "type": "text", + "content": "Table 2. Accuracy (%) across different methods for novel concept classification on Kiki-Kouba dataset. v1 and v2 indicates two different splits described in Sec. 4.1. The table presents results for zero-shot, full-dataset, and few-shot. Results marked with " + }, + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "type": "text", + "content": " denote values reproduced using the official implementation but restricted to the same few-shot data as our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 55, + 326, + 233, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 326, + 233, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 233, + 338 + ], + "type": "text", + "content": "to a " + }, + { + "bbox": [ + 55, + 326, + 233, + 338 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 55, + 326, + 233, + 338 + ], + "type": "text", + "content": " advantage on the previous dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 347, + 170, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 170, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 170, + 360 + ], + "type": "text", + "content": "4.5. Additional Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": "Fusing Features Learned by VRL and Commonly Used Features. Table 3 summarizes the classification accuracy on iNaturalist achieved by fusing different types of visual features. Specifically, we study the two feature vectors: " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "F_{diff}" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "F_{comm}" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": " learned via our VRL, which focusing on capturing regional nuances of the object. " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "F_{diff}" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": " captures the inter-class difference while " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "F_{comm}" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": " represents the shared features within the same class. We also consider the image features directly encoded by the vision encoder of CLIP, which often encode high-level semantics such as the object type, context, and other conceptual associations. As discussed in Sec. 3.1, these features are robust to different kinds of variance and could be benefited when combined for downstream tasks. This is evident from the " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": " improvements when concatenating " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "F_{diff}" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "F_{comm}" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": ", and an additional " + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 54, + 365, + 295, + 556 + ], + "type": "text", + "content": " gain after including the CLIP feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "We also discover that ensemble classifiers trained separately on each feature type are more effective than simply concatenating all features to build a single classifier. This is primarily because these features are heterogeneous. For example, " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "F_{diff}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "F_{comm}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " are generally binary, with each element indicating the presence of a specific verbalized feature. Therefore, they tend to perform best when used with logistic regression models. Meanwhile, the features encoded by CLIP are continuous, which makes them more suitable for mapping to class labels through MLP models. Empirically, we demonstrate that the best configuration of ensembling can further improves the performance by " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " comparing to feature concatenation. We find that with three" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 60, + 541, + 196 + ], + "blocks": [ + { + "bbox": [ + 328, + 60, + 541, + 196 + ], + "lines": [ + { + "bbox": [ + 328, + 60, + 541, + 196 + ], + "spans": [ + { + "bbox": [ + 328, + 60, + 541, + 196 + ], + "type": "table", + "html": "
MethodFdiffFcommCLIPAvg.
Concat✓65.26
Concat✓58.32
Concat✓63.26
Concat✓✓67.92
Concat✓✓✓73.92
Ensemble (hard)LRLRLR68.32
Ensemble (hard)MLPMLPMLP74.26
Ensemble (hard)LRLRMLP71.60
Ensemble (soft)LRLRLR67.06
Ensemble (soft)MLPMLPMLP74.92
Ensemble (soft)LRLRMLP76.52
", + "image_path": "20151383ba08af3505914a9043db02847917130b68bf3ed42a123d5afe15e577.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "type": "text", + "content": "Table 3. Accuracy " + }, + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "type": "text", + "content": " when incorporating feature vectors learned from different methods. " + }, + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "type": "inline_equation", + "content": "F_{diff}" + }, + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "type": "inline_equation", + "content": "F_{comm}" + }, + { + "bbox": [ + 313, + 198, + 553, + 285 + ], + "type": "text", + "content": " denote the difference and commonality feature vectors learned from VRL. CLIP refers to the image features encoded by CLIP visual encoder. We compare two approaches: concatenating all features to train a single classifier and ensembling classifiers trained separately on each feature type. The visual classifiers include Logistic Regression (LR) and Multi-Layer Perceptron (MLP) models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 361 + ], + "type": "text", + "content": "classifiers, hard-voting is prone to high variance, while soft-voting is more robust since it leverages the confidence of each classifier by averaging their probabilities. These experiments showcase that our approach is flexible and can be seamlessly integrated with existing representations to build a robust classification model through ensembling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "text", + "content": "Comparison with Human-Labeled Features. To evaluate the performance of automatically extracted features compared to human-labeled features, we conduct experiments on the Kiki-Bouba dataset using the human-annotated attributes provided by [12]. From Table 4, we observe that VRL extracted features outperform human-labeled attributes across all configurations. Our best configuration, which combines both difference feature vector " + }, + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "inline_equation", + "content": "F_{diff}" + }, + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "text", + "content": " and commonality feature vector " + }, + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "inline_equation", + "content": "F_{comm}" + }, + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "text", + "content": ", surpasses the human-derived features by " + }, + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 363, + 555, + 554 + ], + "type": "text", + "content": ". This underscores the effectiveness of our automatic feature extraction approach in capturing relevant and distinguishing characteristics for classification tasks. These results also suggest that, in scenarios with limited resources or new visual concepts, we can automate the process of feature extraction, reducing the need for manual human annotation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 629 + ], + "type": "text", + "content": "Classifier Choices and Feature Mapping Models. We investigate the model performance using different choices for visual classifiers and feature mapping models. From Table 5, we notice that LLaVA, as a feature mapping model for obtaining binary feature vectors, achieves the best overall performance when paired with the logistic regression." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "As mentioned in Sec. 3.1, we can also leverage CLIP as a feature mapping model to convert images into features. When predicting, one can either set a threshold to determine if the image possesses a certain attribute, resulting in a binary feature, or preserve the similarity between the image and each attribute to form continuous features. Although the performance of using CLIP features is slightly lower" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1608" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 67, + 309, + 228 + ], + "blocks": [ + { + "bbox": [ + 63, + 67, + 309, + 228 + ], + "lines": [ + { + "bbox": [ + 63, + 67, + 309, + 228 + ], + "spans": [ + { + "bbox": [ + 63, + 67, + 309, + 228 + ], + "type": "image", + "image_path": "fad7c3a0303193d993db182e43316478a00645c8d5290978d08932ea8001dd41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 262 + ], + "type": "text", + "content": "Figure 3. Qualitative examples of the features extracted by VRL. We highlight the key attributes in bold. (a) Verbalized features extracted by comparing images from different classes. (b) Verbalized features extracted by comparing the images within the same class." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 68, + 549, + 228 + ], + "blocks": [ + { + "bbox": [ + 312, + 68, + 549, + 228 + ], + "lines": [ + { + "bbox": [ + 312, + 68, + 549, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 68, + 549, + 228 + ], + "type": "image", + "image_path": "1573def6c59fdf90683a5326ff1b74994857dd5d7d3d022787b08c1ecec77c67.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 268, + 289, + 332 + ], + "blocks": [ + { + "bbox": [ + 63, + 268, + 289, + 332 + ], + "lines": [ + { + "bbox": [ + 63, + 268, + 289, + 332 + ], + "spans": [ + { + "bbox": [ + 63, + 268, + 289, + 332 + ], + "type": "table", + "html": "
Methodv1v2Avg.
Human73.852.563.15
VRL-Fdiff88.475.281.80
VRL-Fcomm89.274.081.60
VRL-all89.476.683.00
", + "image_path": "1b9de4cc0a37101bf70c49045c36cec338f3317449aaca6181f5c8f9c8774723.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 377, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 296, + 521 + ], + "type": "text", + "content": "compared to LLaVA, CLIP offers a notable advantage in terms of inference speed. This is because CLIP supports batch-wise operations when computing similarities between a set of images and a set of verbalized features, which offers our method the flexibility to scale with a larger dataset or an increasing number of object classes. It is worth noting that using CLIP as the feature mapping model still allows our method to surpass previous state-of-the-art results by significant margins, achieving improvements of " + }, + { + "bbox": [ + 54, + 377, + 296, + 521 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 54, + 377, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 377, + 296, + 521 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 54, + 377, + 296, + 521 + ], + "type": "text", + "content": " on the iNaturalist and Kiki-Bouba datasets, respectively. This highlights the universal applicability of VRL-extracted features, regardless of the feature mapping approach used." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 527, + 174, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 174, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 174, + 540 + ], + "type": "text", + "content": "4.6. Qualitative Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Figure 3 presents qualitative examples of what verbalized features are extracted from image pairs. We list some of the features that have top-5 feature importance in our trained logistic regression model. For inter-class differences (Fig. 3 (a)), verbalized features are used to capture distinguishing characteristics between different species. For example, the Lichen (Fungi) category is distinguished based on the structure of the thallus, where one image describes a flat, crustlike thallus, and the other describes a complex, branch-like thallus. Similarly, in the Wild Rye (Grass) category, the differences between a slender stem with a feathery inflorescence and a more robust stem with a dense, elongated seed head are verbalized. For intra-class commonality (Fig. 3 (b)), verbalized features are employed to capture shared" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 268, + 535, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 334, + 296, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 367 + ], + "type": "text", + "content": "Table 4. Accuracy (%) of using human-labeled attributes and VRL extracted features on Kiki-Bouba dataset. Note that v1 and v2 indicates two different splits described in Sec. 4.1." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 268, + 535, + 376 + ], + "lines": [ + { + "bbox": [ + 334, + 268, + 535, + 376 + ], + "spans": [ + { + "bbox": [ + 334, + 268, + 535, + 376 + ], + "type": "table", + "html": "
MethodRFLRMLP
iNaturalist (Prior Works: 50.1)
LLaVA64.967.466.0
CLIP (binary)52.258.055.2
CLIP (continuous)52.853.660.6
Kiki-Bouba (Prior Works: 69.3)
LLaVA83.6586.6585.65
CLIP (binary)79.2581.5081.10
CLIP (continuous)77.0577.1081.85
", + "image_path": "9de782fbcbe3b0c63f62a1d5eb736c94f62b2b7c5be876a7caa960fc35a38c04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 421 + ], + "type": "text", + "content": "Table 5. Accuracy (\\%) of different classifier choices (columns) and models used to map verbalized features to numeric vectors (rows). LLaVA and CLIP are utilized as mapping models to map an image to a feature vector." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "content": "traits within the same species. For example, the leaves of Manzanita (Berry) are consistently described as “ovate with a pointed tip and slightly serrated edges,” while the inflorescences of Bulrush (Herb) are uniformly described as having a “delicate, feathery appearance.” These verbalized features are used to guide the decision process by helping the model focus on important distinguishing features or shared traits within each class. Due to page limit, see Appendix B.5 for human study on the learned verbalized features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose Verbalized Representation Learning (VRL), which enables automatic interpretable feature extraction with few-shot samples. By leveraging VLMs, VRL generates verbalized features that capture both inter-class differences and intra-class commonalities. Our method not only enhances the model's adaptability with limited data but also provides transparency in the decision-making process, enabling easier interpretation of the features that influence predictions. Our experiments show that VRL outperforms prior approaches, achieving superior results while using significantly less data. This includes tasks like fine-grained recognition and novel concept adaptation, demonstrating its potential for real-world applications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1609" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "We thank anonymous reviewers and other members of UCLA-NLP+ group for their helpful comments. This work was partially supported by U.S. DARPA ECOLE Program No. #HR00112390060, ONR grant N00014-23-1-2780, DARPA ANSR program FA8750-23-2-0004, and Apple. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of DARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 249, + 115, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 249, + 115, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 115, + 261 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 268, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 268, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 268, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 268, + 294, + 323 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 326, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 326, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 61, + 326, + 295, + 358 + ], + "type": "text", + "content": "[2] Morris Alper and Hadar Averbuch-Elor. Kiki or bouba? sound symbolism in vision-and-language models. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 361, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 296, + 415 + ], + "type": "text", + "content": "[3] David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissection: Quantifying interpretability of deep visual representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6541-6549, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 418, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 296, + 471 + ], + "type": "text", + "content": "[4] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 475, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 294, + 518 + ], + "type": "text", + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 520, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 294, + 574 + ], + "type": "text", + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 577, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 577, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 577, + 295, + 621 + ], + "type": "text", + "content": "[7] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[9] Wentao Chen, Chenyang Si, Zhang Zhang, Liang Wang, Zilei Wang, and Tieniu Tan. Semantic prompt for few-shot image recognition. arXiv preprint arXiv:2303.14123, 2023. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 151 + ], + "type": "text", + "content": "[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. Solo: A single transformer for scalable vision-language modeling. In Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[12] Mia Chiquier, Utkarsh Mall, and Carl Vondrick. Evolving interpretable visual classifiers with large language models. arXiv preprint arXiv:2404.09941, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 241 + ], + "type": "text", + "content": "[13] Fahim Dalvi, Nadir Durrani, Hassan Sajjad, Yonatan Belinkov, Anthony Bau, and James Glass. What is one grain of sand in the desert? analyzing individual neurons in deep nlp models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6309-6317, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[14] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[15] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In 2009 IEEE conference on computer vision and pattern recognition, pages 1778-1785. IEEE, 2009. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[16] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. Advances in neural information processing systems, 20, 2007. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "type": "text", + "content": "[17] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 412, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 412, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 412, + 553, + 477 + ], + "type": "text", + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "type": "text", + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[20] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In Proceedings of the European conference on computer vision (ECCV), pages 264-279, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 623 + ], + "type": "text", + "content": "[21] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep visual features. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] Shaoli Huang, Zhe Xu, Dacheng Tao, and Ya Zhang. Part-stacked cnn for fine-grained visual categorization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1173-1182, 2016. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1610" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[24] Andrej Karpathy, Justin Johnson, and Li Fei-Fei. Visualizing and understanding recurrent networks. arXiv preprint arXiv:1506.02078, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[26] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "text", + "content": "[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE transactions on pattern analysis and machine intelligence, 36(3):453-465, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "text", + "content": "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 352 + ], + "type": "text", + "content": "[29] Manling Li, Ruochen Xu, Shuohang Wang, Luowei Zhou, Xudong Lin, Chenguang Zhu, Michael Zeng, Heng Ji, and Shih-Fu Chang. Clip-event: Connecting text and images with event structures. In Proc. Conference on Computer Vision and Pattern Recognition (CVPR2022), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "type": "text", + "content": "[30] Fan Liu, Wenwen Cai, Jian Huo, Chuanyi Zhang, De long Chen, and Jun Zhou. Making large vision language models to be good few-shot learners. arXiv preprint arXiv:2408.11297, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 442 + ], + "type": "text", + "content": "[31] Mushui Liu, Fangtai Wu, Bozheng Li, Ziqian Lu, Yunlong Yu, and Xi Li. Envisioning class entity reasoning by large language models for few-shot learning. arXiv preprint arXiv:2408.12469, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 296, + 476 + ], + "type": "text", + "content": "[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 510 + ], + "type": "text", + "content": "[33] Eduardo Navarrete, Michele Miozzo, and Francesca Peressotti. Language can shape the perception of oriented objects. Scientific reports, 10(1):8409, 2020. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 544 + ], + "type": "text", + "content": "[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 180, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 180, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 180, + 556 + ], + "type": "text", + "content": "[35] OpenAI. ChatGPT. 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 558, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 296, + 624 + ], + "type": "text", + "content": "[36] Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Marcus Rohrbach. Multimodal explanations: Justifying decisions and pointing to the evidence. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8779-8788, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "type": "text", + "content": "[37] Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, et al. Scikit-learn: Machine learning in python. the Journal of machine learning research, 12:2825–2830, 2011. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[38] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[40] Vilayanur S Ramachandran and Edward M Hubbard. Synaesthesia-a window into perception, thought and language. Journal of consciousness studies, 8(12):3-34, 2001. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "text", + "content": "[41] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[42] Bernardino Romera-Paredes and Philip Torr. An embarrassingly simple approach to zero-shot learning. In International conference on machine learning, pages 2152-2161. PMLR, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 288, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 354 + ], + "type": "text", + "content": "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 410 + ], + "type": "text", + "content": "[44] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 412, + 553, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 467 + ], + "type": "text", + "content": "[45] Chenming Shang, Shiji Zhou, Hengyuan Zhang, Xinzhe Ni, Yujiu Yang, and Yuwang Wang. Incremental residual concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11040, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 468, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 523 + ], + "type": "text", + "content": "[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 524, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 589 + ], + "type": "text", + "content": "[47] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 591, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 656 + ], + "type": "text", + "content": "[48] An Yan, Yu Wang, Yiwu Zhong, Chengyu Dong, Zexue He, Yujie Lu, William Yang Wang, Jingbo Shang, and Julian McAuley. Learning concise and descriptive attributes for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3090-3100, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[49] Cheng-Fu Yang, Yao-Hung Hubert Tsai, Wan-Cyuan Fan, Russ R Salakhutdinov, Louis-Philippe Morency, and Frank Wang. Paraphrasing is all you need for novel object captioning. Advances in Neural Information Processing Systems, 35:6492–6504, 2022." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1611" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "type": "text", + "content": "[50] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2023. 1, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 172 + ], + "type": "text", + "content": "[51] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. arXiv preprint arXiv:2205.15480, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 205 + ], + "type": "text", + "content": "[52] MD Zeiler. Visualizing and understanding convolutional networks. In European conference on computer vision/arXiv, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[53] Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E Gonzalez, et al. Sglang: Efficient execution of structured language model programs. arXiv preprint arXiv:2312.07104, 2023. 1" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1612" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_content_list.json b/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04f8a4451038fc9debac8cdc7a3f252faebf6bf6 --- /dev/null +++ b/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_content_list.json @@ -0,0 +1,1376 @@ +[ + { + "type": "text", + "text": "Versatile Transition Generation with Image-to-Video Diffusion", + "text_level": 1, + "bbox": [ + 181, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zuhao Yang $^{1*}$ Jiahui Zhang $^{1}$ Yingchen Yu $^{2}$ Shijian Lu $^{1\\dagger}$ Song Bai $^{2}$", + "bbox": [ + 209, + 179, + 784, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanyang Technological University $^{2}$ ByteDance Inc.", + "bbox": [ + 272, + 204, + 723, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://mwxely.github.io/projects/yang2025vtg/index", + "bbox": [ + 285, + 231, + 705, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8483849a237828a82a977654c85310005bc8c13829c0a98488a121268f2da6bc.jpg", + "image_caption": [ + "(a) \"a dog facing forward\" $\\rightarrow$ \"a dog turning his head to the right\"" + ], + "image_footnote": [], + "bbox": [ + 94, + 285, + 472, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e866d0401e18f5493c63449527279a3f69085053d9f580d68d19f526e7a3314.jpg", + "image_caption": [ + "(c) \"an airplane\" $\\rightarrow$ \"a cruise\"" + ], + "image_footnote": [], + "bbox": [ + 96, + 406, + 472, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6badb648b44404f8e188979e9cc5dd3ddf50a4edf74574b11298eb792c354739.jpg", + "image_caption": [ + "(b) \"a woman riding a horse\" $\\rightarrow$ \"a woman moving forward on a horse\"" + ], + "image_footnote": [], + "bbox": [ + 517, + 285, + 903, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1473901c7fea11184cdb67aec676be7fc83b51aa433663614e21aa4708bb2bb.jpg", + "image_caption": [ + "(d) \"a wooden house in the forest\" $\\rightarrow$ \"a wooden house in the snow\"", + "Figure 1. Illustration of Versatile Transition Generation (VTG). VTG is capable of performing four types of transition generation, namely, (a) object morphing, (b) motion prediction, (c) concept blending, and (d) scene transition, within a single uniform framework." + ], + "image_footnote": [], + "bbox": [ + 517, + 406, + 901, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 571, + 326, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leveraging text, images, structure maps, or motion trajectories as conditional guidance, diffusion models have achieved great success in automated and high-quality video generation. However, generating smooth and rational transition videos given the first and last video frames as well as descriptive text prompts is far underexplored. We present VTG, a Versatile Transition video Generation framework that can generate smooth, high-fidelity, and semantic-coherent video transitions. VTG introduces interpolation-based initialization that helps preserve object identity and handle abrupt content changes effectively. In addition, it incorporates dual-directional motion fine-tuning and representation alignment regularization that mitigate the limitations of the pre-trained image-to-video diffusion models in motion smoothness and generation fidelity, respectively. To evaluate VTG and facilitate future studies on unified transition generation, we collected TransitBench, a com", + "bbox": [ + 88, + 604, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prehensive benchmark for transition generation that covers two representative transition tasks including concept blending and scene transition. Extensive experiments show that VTG achieves superior transition performance consistently across the four tasks.", + "bbox": [ + 511, + 574, + 908, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 678, + 643, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of diffusion models [18, 43] in image synthesis [30, 34, 36, 38] has inspired a number of studies on diffusion-based video synthesis [19, 47, 58, 60]. Leveraging textual prompts, video frames, structure maps and even motion patterns, several studies [7, 41, 44, 48] have demonstrated impressive synthesis performance by generating realistic and high-fidelity videos automatically. Nevertheless, despite its significant value in various real-world tasks such as video and film production, generating high-quality transition videos conditioned on the first and last frames together with related text prompts remains largely underexplored.", + "bbox": [ + 509, + 703, + 906, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating realistic transition videos is a non-trivial task. A high-quality transition generator should meet at least four", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done while Zuhao Yang was interning at ByteDance.", + "bbox": [ + 107, + 875, + 473, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Shijian Lu is the corresponding author.", + "bbox": [ + 109, + 887, + 321, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16981", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "criteria: 1) semantic similarity to the input frames; 2) high fidelity to the input frames; 3) smoothness across the generated frames; and 4) alignment with the provided text prompts. Moreover, the research community in transition generation mostly relies on self-collected and well-filtered videos that are not accessible to the public [8, 54]. This further hinders the advancement of the research in transition video generation.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing studies tackle the challenge of transition generation via two representative approaches. The first focuses on morphing, given two images of topologically similar objects. Most recent methods [52, 56] harness various deep interpolation techniques for generating rational object-level transition. However, they generate intermittent images rather than temporally coherent video frames, losing the transition smoothness especially while handling moving objects. The second focuses on video frame interpolation. Most existing studies attempt to estimate intermediate optical flows [25, 27, 35, 55], or leverage frame conditioning at training time [4, 5, 22, 50]. However, they tend to generate irrational object transitions with abrupt content changes, struggle with producing long transition sequences, and require time-consuming training on large-scale motion videos. On top of the above, existing studies either employ deep interpolation for conceptual blending transitions [16, 51] or incorporate a random-mask condition layer for scene transition [8], lacking a uniform framework that can work for multiple transition generation tasks. This prompts a compelling inquiry: can we design a versatile transition generator that can handle various transition tasks with minimal adaptation across tasks?", + "bbox": [ + 91, + 215, + 486, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Figure 1, we propose four transition tasks to establish versatile transition generation, based on the types of input frames: (1) Object Morphing: The input frames are either the same object with different postures or different objects, as long as they are topologically similar; (2) Concept Blending: The input frames contain conceptually different objects (e.g., 'an airplane' and 'a cruise'); (3) Motion Prediction: The input frames can be seen as two moments in a video containing one (or more) moving object; (4) Scene Transition: The input frames are conceptually related scene images yet they either belong to two different domains (e.g., 'a wooden house in the forest' and 'a wooden house in the snow') or represent two distinct components of a scene (e.g., 'erupting volcano' and 'hot lava').", + "bbox": [ + 89, + 566, + 482, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design VTG, a versatile transition generation framework built upon image-to-video diffusion models. VTG features three designs for generating smooth, high-fidelity, and semantic-coherent transitions. First, we employ interpolation-based initialization that effectively mitigates abrupt content changes while handling input frames with significantly different contents. Specifically, we spherically interpolate latent Gaussian noises of the two input frames", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and text embeddings of corresponding transition captions and leverage two LoRA-integrated U-Nets [20] to capture object-level semantics in the denoising steps. Second, we adopt a lightweight fine-tuning strategy that merges the predicted forward and backward noises which greatly improves the transition smoothness in motion predictions. The bidirectional sampling can be achieved by fine-tuning the pretrained U-Net with a small collection of videos. Third, we introduce regularization with self-supervised visual encoder for video diffusion models to explicitly induce feature alignment, which helps learn meaningful representations and enhance the fidelity of generated transition videos. These components are logically cohesive and, in combination, distinguish VTG from prior work by enabling a single framework to handle diverse transition tasks with minimal task-specific adjustments. In addition, we collected TransitBench, a new benchmark of 200 pairs with first and last frames for concept blending and scene transition.", + "bbox": [ + 511, + 90, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work can be summarized in three aspects. First, we propose a unified task of versatile transition generation, aiming at smooth and rational transition for object morphing, concept blending, motion prediction, and scene transition. Second, we design VTG, a novel and versatile framework that can generate semantic-relevant, high-fidelity and temporally coherent video transitions effectively. Third, we introduce TransitBench, a curated dataset for benchmarking concept blending and scene transition. With TransitBench and other public benchmarks, we demonstrate, both qualitatively and quantitatively, that VTG outperforms the state-of-the-art consistently across the four transition generation tasks.", + "bbox": [ + 511, + 363, + 903, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 571, + 653, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Morphing. Image morphing has been a longstanding problem in the computer graphics community [1, 49, 61]. Conventional morphing techniques either utilize correspondence-driven bidirectional image warping [2, 3, 9, 26] or minimize the path energy on a Riemannian manifold [14, 29, 33, 40] to obtain image transitions. While producing seamless morphing, they either require massive human involvement or fail to create new content beyond the two given images. Recently, several studies [52, 56] explored the prior knowledge in image diffusion models for real-world morphing with more object categories. Differently, we exploit a pretrained video diffusion model for image morphing, naturally enabling coherent transitions across frames.", + "bbox": [ + 511, + 599, + 903, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Frame Interpolation. Video frame interpolation (VFI) has been extensively studied in the computer vision research community [11]. Classic VFI algorithms are closely related to optical flow prediction, which obtain interpolation by either forward warping [31] or backward", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16982", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "warping [21, 23, 25] based on the estimated flows. Several follow-up studies [22, 27, 35] focus on large motion problem in VFI. Recently, a line of research [12, 46] explores bounded video generation with image-to-video diffusion models under the guidance of the start and end frames. This approach fuses outputs from forward and backward paths by interpolating the predicted forward noise and its reversed counterpart. While generating coherent motion, it overlooks the identity preservation and struggles to produce natural transitions when the two input frames differ significantly in content.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transition Generation. Transition generation is a far underexplored problem, largely due to its open-ended nature. Typically, a transition generation framework leverages two input frames (i.e., the first and last frames) and a transition caption to produce a scene-level transition video that connects two distinct narrative moments. This has been explored in [8] that incorporates random masking to selectively suppress information from the original latent code, allowing capturing transition subtleties between frames. In addition, [50, 54] inject the first and last frames as additional conditions by concatenating them with noisy latent codes along the channel dimension. Recently, a concurrent study [57] explores transition video generation with imaged-to-video diffusion as well. Differently, we formulate the task by versatile transition generation by unifying four transition tasks under the same framework, which greatly expands the applicability and significantly enhances the quality of the generated transition videos.", + "bbox": [ + 89, + 272, + 483, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 556, + 181, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 580, + 230, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent diffusion models [36] are text-to-image diffusion models that operate in a compressed latent space. During forward diffusion process, it encodes an image sample $\\mathbf{x}_0$ into a latent code $\\mathbf{z}_0$ and corrupts $\\mathbf{z}_0$ by Gaussian noise at each time step $t = 1,\\dots ,T$ :", + "bbox": [ + 89, + 602, + 483, + 678 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 685, + 482, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ denotes the noise strength at time step $t$ and $\\mathbf{I}$ denotes the identity matrix with the same dimensions as $\\mathbf{z}_0$ . Given a text embedding $\\mathbf{c}$ , the backward denoising process can be parameterized as:", + "bbox": [ + 89, + 710, + 483, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 777, + 482, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}(\\mathbf{z}_t,\\mathbf{c},t)$ is output by a U-Net [37], and $\\Sigma_{\\theta}(\\mathbf{z}_t,\\mathbf{c},t)$ is determined by a noise scheduler (e.g., DDPM [18] or DDIM [42]). The model learns to predict noise via a parameterized noise estimator $\\epsilon_{\\theta}$ under the guidance of an objective function:", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\mathbb {E} _ {t \\sim [ 1, T ], \\mathbf {z} _ {0}, \\mathbf {c}, \\epsilon_ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework is built upon pretrained image-to-video diffusion models that incorporate temporal convolutions and temporal attention layers on top of latent diffusion models to establish temporal correlations between video frames. Given a video $\\mathbf{x}_0\\in \\mathbb{R}^{N\\times 3\\times H\\times W}$ , where $N$ is the number of frames and $H\\times W$ indicates the image resolution, each frame is first encoded into a latent representation $\\mathbf{z}_0\\in \\mathbb{R}^{N\\times C\\times h\\times w}$ via the VAE encoder $\\mathcal{E}_v$ . Both forward diffusion process and backward denoising process are then performed in the latent space. The generated video frames can be collectively obtained through the decoder $\\mathcal{D}_v$ . To inject image conditions, we fuse and incorporate text-conditioned and image-conditioned features by projecting the conditional image into a text-aligned embedding space. Meanwhile, the conditional image is concatenated with the initial per-frame noise to preserve more visual details.", + "bbox": [ + 511, + 90, + 906, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Interpolation-based Initialization", + "text_level": 1, + "bbox": [ + 511, + 340, + 805, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In diffusion models, the initial Gaussian noise determines the coarse structure that emerges in the early denoising steps, while high-frequency details are refined later. Since existing image-to-video diffusion models [8, 50] randomly initialize the latent code $\\mathbf{z}_0$ at inference time, each intermediate frame follows a distinct stochastic trajectory: colors, micro-textures, or even object pose shift slightly from frame to frame, yielding perceptual \"flicker\" [10].", + "bbox": [ + 511, + 362, + 906, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To suppress this drift, we correlate the frame-wise latents by interpolating between the endpoint noises instead of resampling them independently. Prior work [6] indicates that linear interpolation often yields intermediate latent norms that are extremely unlikely under a unit Gaussian, whereas spherical linear interpolation (SLERP) preserves the Euclidean norm and keeps samples on-distribution. Therefore, we leverage SLERP between the latent noises of two input frames $\\mathbf{z}_{t1}$ and $\\mathbf{z}_{tN}$ as follows:", + "bbox": [ + 511, + 484, + 908, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t n} = \\frac {\\sin ((1 - \\lambda_ {\\text {n o i s e}}) \\phi)}{\\sin \\phi} \\mathbf {z} _ {t 1} + \\frac {\\sin (\\lambda_ {\\text {n o i s e}} \\phi)}{\\sin \\phi} \\mathbf {z} _ {t N}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 625, + 906, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi = \\arccos \\left(\\frac{\\mathbf{z}_{t1}^T\\cdot\\mathbf{z}_{tN}}{\\|\\mathbf{z}_{t1}\\|_{\\|}\\mathbf{z}_{tN}\\|}\\right)$ , and $\\lambda_{\\mathrm{noise}}\\in [0,1]$ denotes the parameter for latent interpolation. As shown in Figure 2, we concatenate $\\mathbf{z}_{tn}$ with $\\mathbf{z}_{t1}$ and $\\mathbf{z}_{tN}$ along the frame dimension and inject them as the initial latent noise for DDIM sampling. It is worth noting that we only inject the interpolated latent noises at early denoising steps to preserve the appearance and motion priors from the image-to-video backbone as much as possible. With such meticulously designed latent injection strategy, our approach addresses the issue of random and abrupt transition frames.", + "bbox": [ + 511, + 666, + 908, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike GANs [15], whose latent space is structured and semantically meaningful, diffusion models operate in an unstructured noise space that does not explicitly encode high-level semantics [56]. This lack of structure can lead to artifacts when interpolating between semantically different in", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16983", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa27ac34aa3c8e70ab2e6846d79d0da738d1e1b4cc8fa4d40bf787a792e30614.jpg", + "image_caption": [ + "Figure 2. Inference framework of VTG. Our interpolation-based initialization features three designs: $①$ Interpolated Noise Injection, $②$ LoRA Interpolation, and $③$ Frame-aware Text Interpolation. VTG first converts two encoded input frames to latent noises via DDIM inversion. Then, it interpolates between the two latent noises and concatenate intermediate noises along the temporal dimension. To capture meaningful semantics and enable the transition between two conceptually different objects, we employ LoRA interpolation and text interpolation, respectively." + ], + "image_footnote": [], + "bbox": [ + 148, + 88, + 854, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "puts. To suppress such transition artifacts, we first train two LoRAs $\\Delta \\theta_{1}$ , $\\Delta \\theta_{N}$ given two input frames $x^{1}$ and $x^{N}$ , optimizing the following objective:", + "bbox": [ + 89, + 580, + 482, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\Delta \\theta_ {t}\\right) = \\mathbb {E} _ {\\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta + \\Delta \\theta_ {i}} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0 n} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, \\mathbf {c} _ {n}\\right) \\| ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 633, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_{0n}$ is the encoded latent vector of frame $n$ , and $\\mathbf{c}_n$ is the text embedding associated with transition caption $c^n$ . Then, we linearly interpolate the two LoRAs to fuse the semantics of two input frames:", + "bbox": [ + 89, + 657, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta = \\left(1 - \\lambda_ {l o r a}\\right) \\Delta \\theta_ {1} + \\lambda_ {l o r a} \\Delta \\theta_ {N}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 724, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{lora}$ denotes the parameter during lora interpolation.", + "bbox": [ + 89, + 750, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically, video diffusion models [8, 50] employ only one caption as the text condition. This hinders the transition generation between two conceptually different objects, causing abrupt content changes without generating intermediate frames with hybrid meanings. However, for imaged-to-video diffusion models, the text embedding of the entire frame sequence $c$ is integrated within the denoising U-Net via the cross-attention layer, without explicitly defining per-frame text embedding. A related still-image approach [52]", + "bbox": [ + 88, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "linearly blends the two text embeddings to morph between two static endpoints, whereas our method performs frame-aware SLERP across prompt sequences as follows:", + "bbox": [ + 511, + 580, + 906, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {\\lambda_ {t e x t}} = \\frac {\\sin ((1 - \\lambda_ {t e x t}) \\phi)}{\\sin \\phi} \\mathbf {c} _ {1} + \\frac {\\sin (\\lambda_ {t e x t} \\phi)}{\\sin \\phi} \\mathbf {c} _ {N}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 637, + 906, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{text} \\in [0,1]$ serves as the frame-aware coefficient to control the transition sequence; $\\mathbf{c}_1$ and $\\mathbf{c}_N$ denote the text embeddings of the first and last frames, respectively.", + "bbox": [ + 511, + 683, + 906, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Bidirectional Motion Prediction", + "text_level": 1, + "bbox": [ + 511, + 741, + 794, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments, we observed notable quality differences when reversing the order of input frames. This discrepancy arises from two main factors: (1) the model shows a bias toward resembling the initial input frame more closely due to conditional image leakage [59]; (2) existing image-to-video diffusion models [8, 50] are pretrained exclusively for forward motion prediction, while real-world motion is naturally asymmetric, resulting in ambiguity when predicting reverse motion.", + "bbox": [ + 511, + 763, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16984", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a844dcd3564d5a50c7a030960a63a525bbacd9a18de01bd94be556adfd4c9b48.jpg", + "image_caption": [ + "Figure 3. Training framework of VTG. In Bidirectional Motion Prediction (BMP), the noisy latent is flipped along temporal dimension, and self-attention maps undergo a 180-degree rotation to establish reversed motion-time correlations. Two U-Nets separately predict forward and backward motion, with the backward noise reversed again to fuse with the forward noise, ensuring a consistent motion path for iterative denoising. In Representation Alignment Regularization (RAR), we spatially patchify each frame independently, then aggregate the per-patch alignment loss across temporal dimension." + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 861, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [46], we simultaneously predict forward and backward motions to mitigate the ambiguity issue. Specifically, we rotate the temporal self-attention maps by 180 degrees both horizontally and vertically, causing a reversal of attention relationships as illustrated in Figure 3. We also reverse the original forward trajectory of video $\\{x^{1}, x^{2}, \\ldots, x^{N}\\}$ to $\\{x^{N}, x^{N - 1}, \\ldots, x^{1}\\}$ , given the two input frames $x^{1}$ and $x^{N}$ . Then, we obtain the reversed noisy latent $\\mathbf{z}_{t'}$ by flipping $\\mathbf{z}_t$ along the temporal dimension. After that, $\\mathbf{z}_{t'}$ is fed into the 3D U-Net for backward motion prediction.", + "bbox": [ + 89, + 487, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we employ a lightweight fine-tuning strategy that only updates the parameters of value and output matrices (i.e., $\\theta_{w,o}$ ) in the temporal attention layers. The objective $\\mathcal{L}_{BMP}$ is formulated by taking the L2 norm between the predicted noise and ground-truth reversed noise:", + "bbox": [ + 89, + 656, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B M P} = \\left\\| \\operatorname {f l i p} \\left(\\epsilon_ {t}\\right) - \\epsilon_ {\\theta_ {w, o}} \\left(\\mathbf {z} _ {t ^ {\\prime}}, \\mathbf {c}, t, A _ {i, j} ^ {\\prime}\\right) \\right\\| _ {2} ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 753, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The predicted backward noise is reversed again to fuse with forward noise. Specifically, we simply employ linear interpolation to ensure forward-backward consistency:", + "bbox": [ + 89, + 787, + 483, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} = \\left(1 - \\lambda_ {B M P}\\right) \\epsilon_ {t, i} + \\lambda_ {B M P} \\epsilon_ {t, N - i} ^ {\\prime}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 851, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{BMP} = 0.5$ denotes the weighting factor.", + "bbox": [ + 89, + 885, + 419, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Representation Alignment Regularization", + "text_level": 1, + "bbox": [ + 511, + 486, + 870, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Though previous techniques effectively mitigate abrupt content changes and motion-reversal ambiguity, our approach remains susceptible to generating blurry and low-fidelity transitions, especially when input frames contain fine-grained textures (e.g., bicycle spokes or fabric weave). Recent work [53] shows that diffusion latents inherently lack high-frequency semantics compared to self-supervised representations like DINOv2's [32]. Such mismatch tends to accumulate over video frames and appears as blur.", + "bbox": [ + 511, + 515, + 906, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome this, we propose to distill DINOv2 features back into the denoising trajectory of video diffusion. As illustrated in Figure 3, the video latent is first patched into $N$ sequences of $T$ tokens along the temporal dimension. For a N-frame video, the latent vector $\\mathbf{z}_n$ of $n$ -th frame has the shape of $C\\times h\\times w$ . Given a pre-defined patch size $s$ , the sequence length $T$ should be: $T = (h / s)\\times (w / s)$ . Then, we project the per-frame latent representation $\\mathbf{z}_n$ of video diffusion via a multilayer perceptron (MLP) to align with DINOv2 representation $\\mathbf{y}_{*} = \\mathcal{E}_{d}(\\mathbf{x}_{*})\\in \\mathbb{R}^{P\\times m}$ , where $\\mathbf{y}_{*}$ denotes the encoder output given a clean video frame $\\mathbf{x}_{*}$ ; $P$ and $m$ respectively denotes the number of patches and the embedding dimension of DINOv2 encoder $\\mathcal{E}_d$ . Let $y_{\\phi}(\\mathbf{h}_t)$ be the projected representation of intermediate diffusion latent feature $\\mathbf{h}_t$ , where $\\phi$ denotes the trainable parameters of the MLP. The regularization term $\\mathcal{L}_{RAR}$ is computed by", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16985", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aggregating patch-wise similarities across all video frames:", + "bbox": [ + 89, + 90, + 482, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R A R}} = - \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {t, \\mathbf {x} _ {*}, \\epsilon_ {t}} \\left[ \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\sin \\left(\\mathbf {y} _ {*} ^ {[ p ]}, y _ {\\phi} (\\mathbf {h} _ {t}) ^ {[ p ]}\\right) \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 128, + 482, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, the external encoder $\\mathcal{E}_d$ and MLP projector $y_{\\phi}$ are discarded.", + "bbox": [ + 89, + 161, + 482, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 205, + 223, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 229, + 279, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We leverage a small collection (i.e., 150) of high-quality in-house videos as our training data. These video samples encompass various motion (e.g., people running, airplanes taxiing) and appearance changes (e.g., a man transitioning from youth to old age). The learnable parameters of the backward U-Net are initialized to zero, while the other available weights are initialized from the pre-trained checkpoint of DynamiCrafter [50]. We utilize the AdamW optimizer with learning rate of $1 \\times e^{-5}$ . VTG is trained on 4 NVIDIA A100-80GB GPUs, taking $\\sim 20K$ iterations with batch size of 2.", + "bbox": [ + 89, + 251, + 482, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference Details. In our experiments, we use Stable Diffusion v2.1-base [36] as our interpolation backbone. When training LoRAs, we set the LoRA rank to 16 and train them for 200 steps using AdamW optimizer [28] with a learning rate of $2 \\times 10^{-4}$ . This requires only $\\sim 85$ s with 1 NVIDIA A100-80GB GPU. During inversion and sampling, we use the DDIM sampler [42] and set total time steps to the default value (i.e., 50) of the selected image-to-video models. We apply text classifier-free guidance [17] for all models with the same negative prompt \"Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms\" across all transition generations.", + "bbox": [ + 89, + 431, + 482, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Methods. We compare VTG with following methods: DiffMorpher [56], TVG [57], SEINE [8], DynamiCrafter [50], Generative Inbetweening [46], Text Embedding Interpolation (TEI) [45], Denoising Interpolation (DI) [16], and Attention Interpolation (AID) [16].", + "bbox": [ + 89, + 643, + 482, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Benchmarks. We employ several public datasets and our self-curated TransitBench to comprehensively evaluate the quality of the generated transition videos. Specifically, we leverage MorphBench [56] and TC-Bench [13] to evaluate object morphing and motion prediction. Concept blending and scene transition are rarely explored in the realm of generative models. Hence, there is a lack of specific evaluation benchmarks for these two tasks. Previously, [16] employs CIFAR-10 [24] and LAION-Aesthetics [39] as their evaluation benchmarks to test concept blending. While there are sufficient number of", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "samples in both datasets, CIFAR-10 only contains concepts from 10 classes, and LAION-Aesthetics consists entirely of synthetic images and lacks real-world samples. To this end, we present TransitBench, the first benchmark dataset for collectively assessing concept blending transitions of two distinct conceptual objects and scene transitions between two relevant scenarios. We collected 200 pairs of pictures (each pair forms the first and the last frames of one transition generation sample) of diverse content and styles, and evenly divide them into two categories: 1) concept-blending cases, and 2) scene-transition cases, both of which are obtained from web resources. We hope TransitBench can promote future studies on general transition generation.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 299, + 696, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the superiority of our method, we provide a visual comparison of VTG against five state-of-the-art baselines. Specifically, we compare our approach with DiffMorpher, Generative Inbetweening, TVG, SEINE, DynamamiCrafter across four transition tasks for uniformity. Figure 4 shows the generation results by different methods. For object morphing, the compared methods exhibit low-fidelity (e.g., DiffMorpher, DynamamiCrafter), oversaturation effect (e.g., TVG), and inaccurate semantics (e.g., Generative Inbetweening, SEINE). In contrast, VTG presents a natural transition, preserving the content of the first and last frames well. For motion prediction, Generative Inbetweening and SEINE generate inferior transitions as observed by their blurry intermediate frames. TVG yields unnatural transitions as it generates a front-facing image of a person, while the person's face is actually turned away. DiffMorpher shows low-fidelity transitions, and its generated video looks like a sequence of intermittent images without concerning the consistency between background and foreground objects. This is largely due to the missing temporal modeling in DiffMorpher's image diffusion backbone. VTG achieves comparable or even better results with DynamamiCrafter, validating that it can effectively preserve the appearance and motion priors in diffusion models. For concept blending, most video diffusion baselines present abrupt content changes, while other baselines show blurry (e.g., DiffMorpher) and ambiguous (e.g., TVG) intermediate frames. In contrast, VTG yields smooth transition between the lion and the truck with rational intermediate results (e.g., a lion-colored and lion-sized truck). For scene transition between two relevant scene images, existing methods suffer from abrupt content changes or low similarity (e.g., TVG), while VTG produces natural scene-level transition with pleasing effect.", + "bbox": [ + 511, + 321, + 906, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 847, + 707, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The objective of versatile transition generation is to generate a sequence of $(N - 2)$ intermedi-", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16986", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/032d98d612a2dbe24edff782b7194d13dc08b915212ef9ff448676a217916a34.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons on four transition tasks. For each task, the displayed images correspond to the 1st, 4th, 8th, 12th, and 16th video frames, respectively. Please refer to our project page for more video examples, which offers a more straightforward and comprehensive qualitative comparisons." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 883, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ate video frames, denoted as $x^{2:N - 1}$ , between two source frames $x^1$ and $x^N$ , with two captions describing the two frames. High-quality transition videos should simultaneously satisfy four criteria: semantic fidelity to the input endpoints, visual quality, temporal coherence, and perceptual", + "bbox": [ + 89, + 813, + 485, + 891 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "smoothness. To ensure fair comparisons, we performed extensive quantitative evaluations using experimental settings consistent with those prior methods as in their original paper. Specifically, we adopt the following evaluation metrics: (1) FID $(\\downarrow)$ and PPL $(\\downarrow)$ follow DiffMorpher [56]. FID com", + "bbox": [ + 511, + 813, + 908, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16987", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pares the first and last frames against the generated transition frames in the Inception feature space; PPL measures the average LPIPS fluctuation along an interpolation path, where lower values imply fewer perceptual jumps. (2) TCR $(\\uparrow)$ and TC-Score $(\\uparrow)$ are taken from the training-free TVG framework [57]. TCR is the fraction of video frames whose CLIP embedding stays within a pre-set similarity margin to either endpoint; TC-Score complements it with a mean cosine-similarity, offering a softer view of frame-level consistency. (3) Smoothness $(\\uparrow)$ is adapted from AID [16]. For every adjacent frame pair, we compute LPIPS and the Gini coefficient of these distances, reflecting how even the step-to-step changes are. We report its inverse so that a higher score denotes smoother transitions.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b4fedb4c59ef81f938cf54aa5997fc2409d5cb02e84f85ac308cc0b7ea07985.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMetamorphosisAnimation
FID (↓)PPL (↓)FID (↓)PPL (↓)
DiffMorpher [56]70.4918.1943.155.14
TVG [57]86.9235.1842.9912.46
SEINE [8]82.0347.7248.2516.26
DynamiCrafter [50]87.3242.0943.3111.16
VTG (Ours)67.3922.8039.165.14
", + "bbox": [ + 94, + 310, + 478, + 425 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c59634264d565be961d00caa04862aacc67c832bfc54a90be196b1fbdbec778d.jpg", + "table_caption": [ + "Table 1. Quantitative results on MorphBench. The best results are in bold; second-best are underlined." + ], + "table_footnote": [], + "table_body": "
MethodAttributeObjectBackground
TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)
DiffMorpher [56]41.820.84419.570.76550.000.819
SEINE [8]17.860.72010.480.6547.960.742
DynamiCrafter [50]16.550.74513.910.70725.560.795
TVG [57]41.820.87730.440.82238.890.864
VTG (Ours)42.780.89333.460.84950.000.883
", + "bbox": [ + 94, + 487, + 480, + 560 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4d8fbd4fcd775d3629797f705006cfa441225a63d0b86792e5b9e2a41179dfcd.jpg", + "table_caption": [ + "Table 2. Quantitative results on TC-Bench. The best results are in bold. Best viewed when zoomed in." + ], + "table_footnote": [], + "table_body": "
DatasetTEIDIAID-OAID-IOurs
CIFAR-100.75310.75640.78310.78610.7932
LAION-Aesthetics0.74240.75110.76430.81520.8215
", + "bbox": [ + 94, + 625, + 482, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Smoothness (↑) evaluation for Concept Blending. The best results are in bold.", + "bbox": [ + 89, + 689, + 482, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Results. As seen in Table 1 and Table 2, VTG outperforms existing methods with their original experimental setting, demonstrating its capability to generate high-quality, temporally coherent, and semantically consistent video transitions. It is worth noting that the only exception is Metamorphosis subset, where VTG achieves the second-best PPL (22.80). Nevertheless, VTG still demonstrates outstanding overall performance, achieving state-of-the-art FID and comparable PPL.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "According to AID [16], a well-interpolated concept-blending sequence should exhibit a gradual and smooth", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "transition. To offer a more comprehensive evaluation, we conducted experiments using the same benchmarks and metric as in AID. Please refer to Table 3 for details.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71e79c1636d39dd458cf4f41f926f41fccb8d006114a8eb4e6109c7f72c019f9.jpg", + "image_caption": [ + "Figure 5. The bar plot of user preference rates $(\\%)$ of different methods on four transition tasks. Best viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 516, + 161, + 903, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. To subjectively evaluate our approach, we further conduct a user study. We asked Amazon Mechanical Turk (AMT) workers to choose the best generated transition video from a set of five candidates. Given the two input frames and corresponding transition captions, AMT workers evaluate the videos based on three questions: 1) Which video is the most semantic-coherent? 2) Which video is the most temporally coherent? 3) Which video has the highest fidelity? To make a fair evaluation, we include static and obviously incorrect videos to identify random clicking. Each test example is assigned to 10 different workers for evaluation. We collected 150 valid responses for each transition task after filtering out those responses from random clicking workers. The results are shown in Figure 5, which illustrates that VTG obtains the highest preference rate with its superior quality of the generated transition videos compared to other baselines.", + "bbox": [ + 511, + 393, + 908, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we formulate the problem of versatile transition generation that encompasses four representative transition tasks, namely, object morphing, motion prediction, concept blending, and scene transition. Leveraging a pretrained image-to-video diffusion model, we design VTG, a unified Versatile Transition video Generator that can tackle the four tasks within a single framework. Both qualitative and quantitative experiments show that VTG achieves superior performance in generating semantically relevant, temporally coherent, and visually pleasing transition video frames, given a pair of text prompts and two input frames. With these advantages, VTG can be readily applied to real-world content creation scenarios, offering an efficient tool for generating high-quality transition in media production.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16988", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This study is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22.", + "bbox": [ + 89, + 116, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 179, + 187, + 195 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alyaa Qusay Aloraibi. Image morphing techniques: A review. Technium: Romanian Journal of Applied Sciences and Technology, 2023. 2", + "[2] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. Seminal Graphics Papers: Pushing the Boundaries, 1992. 2", + "[3] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2011. 2", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2", + "[6] Erik Bodin, Alexandru I Stere, Dragos D Margineantu, Carl Henrik Ek, and Henry Moss. Linear combinations of latents in generative models: subspaces and beyond. In ICLR, 2025. 3", + "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 1", + "[8] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 2, 3, 4, 6, 8", + "[9] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. TOG, 2012. 2", + "[10] Yitong Deng, Winnie Lin, Lingxiao Li, Dmitriy Smirnov, Ryan Burgert, Ning Yu, Vincent Dedun, and Mohammad H Taghavi. Infinite-resolution integral noise warping for diffusion models. In ICLR, 2025. 3", + "[11] Jiong Dong, Kaoru Ota, and Mianxiong Dong. Video frame interpolation: A comprehensive survey. TOMM, 2023. 2", + "[12] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In ECCV, 2024. 3", + "[13] Weixi Feng, Jiachen Li, Michael Saxon, Tsu-jui Fu, Wenhu Chen, and William Yang Wang. Tc-bench: Benchmarking temporal compositionality in text-to-video and image-to-video generation. In ACL, 2025. 6" + ], + "bbox": [ + 93, + 204, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, 2020. 2", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3", + "[16] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. In NeurIPS, 2024. 2, 6, 8", + "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 6", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3", + "[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2", + "[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In ECCV, 2022. 3", + "[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hólynski, Ben Poole, and Janne Kontkanen. Video interpolation with diffusion models. In CVPR, 2024. 2, 3", + "[23] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In CVPR, 2022. 3", + "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 6", + "[25] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 2, 3", + "[26] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. TOG, 2014. 2", + "[27] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In CVPR, 2024. 2, 3", + "[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2018. 6", + "[29] Michael I. Miller and Laurent Younes. Group actions, homeomorphisms, and matching: A general framework. *IJCV*, 2001. 2", + "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16989", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 1", + "[31] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 2", + "[32] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 5", + "[33] Marko Rajković. Geodesics, Splines, and Embeddings in Spaces of Images. PhD thesis, Universitäts-und Landesbibliothek Bonn, 2023. 2", + "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[35] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In ECCV, 2022. 2, 3", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6", + "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3", + "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1", + "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 6", + "[40] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In CVPR, 2010. 2", + "[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1", + "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. *ICLR*, 2021. 3, 6", + "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1", + "[44] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 1", + "[45] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. In ICMLW, 2023. 6", + "[46] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. In ICLR, 2025. 3, 5, 6" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1", + "[48] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH, 2024. 1", + "[49] George Wolberg. Image morphing: a survey. The visual computer, 1998. 2", + "[50] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3, 4, 6, 8", + "[51] Cheng Yang, Lijing Liang, and Zhixun Su. Realworld denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023. 2", + "[52] Zhaoyuan Yang, Zhengyang Yu, Zhiwei Xu, Jaskirat Singh, Jing Zhang, Dylan Campbell, Peter Tu, and Richard Hartley. Impus: Image morphing with perceptually-uniform sampling using diffusion models. In ICLR, 2024. 2, 4", + "[53] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. In ICLR, 2025. 5", + "[54] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In CVPR, 2024. 2, 3", + "[55] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In CVPR, 2023. 2", + "[56] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Xingang Pan, and Bo Dai. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In CVPR, 2024. 2, 3, 6, 7, 8", + "[57] Rui Zhang, Yaosen Chen, Yuegen Liu, Wei Wang, Xuming Wen, and Hongxia Wang. TVg: A training-free transition video generation method with diffusion models. TCVT, 2024. 3, 6, 8", + "[58] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 1", + "[59] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. In NeurIPS, 2024. 4", + "[60] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1", + "[61] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. IJAEMS, 2017. 2" + ], + "bbox": [ + 514, + 90, + 905, + 858 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16990", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_model.json b/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0db1db10f0ff58fe97aaf7b3dc49ec868593801 --- /dev/null +++ b/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "Versatile Transition Generation with Image-to-Video Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.18, + 0.785, + 0.201 + ], + "angle": 0, + "content": "Zuhao Yang\\(^{1*}\\) Jiahui Zhang\\(^{1}\\) Yingchen Yu\\(^{2}\\) Shijian Lu\\(^{1\\dagger}\\) Song Bai\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.205, + 0.725, + 0.225 + ], + "angle": 0, + "content": "\\(^{1}\\)Nanyang Technological University \\(^{2}\\)ByteDance Inc." + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.232, + 0.706, + 0.251 + ], + "angle": 0, + "content": "https://mwxely.github.io/projects/yang2025vtg/index" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.286, + 0.473, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.384, + 0.453, + 0.399 + ], + "angle": 0, + "content": "(a) \"a dog facing forward\" \\(\\rightarrow\\) \"a dog turning his head to the right\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.407, + 0.473, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.506, + 0.262, + 0.519 + ], + "angle": 0, + "content": "(c) \"an airplane\" \\(\\rightarrow\\) \"a cruise\"" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.286, + 0.905, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.384, + 0.89, + 0.399 + ], + "angle": 0, + "content": "(b) \"a woman riding a horse\" \\(\\rightarrow\\) \"a woman moving forward on a horse\"" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.407, + 0.903, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.506, + 0.893, + 0.519 + ], + "angle": 0, + "content": "(d) \"a wooden house in the forest\" \\(\\rightarrow\\) \"a wooden house in the snow\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.532, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Figure 1. Illustration of Versatile Transition Generation (VTG). VTG is capable of performing four types of transition generation, namely, (a) object morphing, (b) motion prediction, (c) concept blending, and (d) scene transition, within a single uniform framework." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.573, + 0.327, + 0.589 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.606, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Leveraging text, images, structure maps, or motion trajectories as conditional guidance, diffusion models have achieved great success in automated and high-quality video generation. However, generating smooth and rational transition videos given the first and last video frames as well as descriptive text prompts is far underexplored. We present VTG, a Versatile Transition video Generation framework that can generate smooth, high-fidelity, and semantic-coherent video transitions. VTG introduces interpolation-based initialization that helps preserve object identity and handle abrupt content changes effectively. In addition, it incorporates dual-directional motion fine-tuning and representation alignment regularization that mitigate the limitations of the pre-trained image-to-video diffusion models in motion smoothness and generation fidelity, respectively. To evaluate VTG and facilitate future studies on unified transition generation, we collected TransitBench, a com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.909, + 0.651 + ], + "angle": 0, + "content": "prehensive benchmark for transition generation that covers two representative transition tasks including concept blending and scene transition. Extensive experiments show that VTG achieves superior transition performance consistently across the four tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.679, + 0.645, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.907, + 0.87 + ], + "angle": 0, + "content": "The success of diffusion models [18, 43] in image synthesis [30, 34, 36, 38] has inspired a number of studies on diffusion-based video synthesis [19, 47, 58, 60]. Leveraging textual prompts, video frames, structure maps and even motion patterns, several studies [7, 41, 44, 48] have demonstrated impressive synthesis performance by generating realistic and high-fidelity videos automatically. Nevertheless, despite its significant value in various real-world tasks such as video and film production, generating high-quality transition videos conditioned on the first and last frames together with related text prompts remains largely underexplored." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Generating realistic transition videos is a non-trivial task. A high-quality transition generator should meet at least four" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.875, + 0.475, + 0.888 + ], + "angle": 0, + "content": "*This work was done while Zuhao Yang was interning at ByteDance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.323, + 0.901 + ], + "angle": 0, + "content": "Shijian Lu is the corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.875, + 0.475, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "criteria: 1) semantic similarity to the input frames; 2) high fidelity to the input frames; 3) smoothness across the generated frames; and 4) alignment with the provided text prompts. Moreover, the research community in transition generation mostly relies on self-collected and well-filtered videos that are not accessible to the public [8, 54]. This further hinders the advancement of the research in transition video generation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.216, + 0.487, + 0.562 + ], + "angle": 0, + "content": "Most existing studies tackle the challenge of transition generation via two representative approaches. The first focuses on morphing, given two images of topologically similar objects. Most recent methods [52, 56] harness various deep interpolation techniques for generating rational object-level transition. However, they generate intermittent images rather than temporally coherent video frames, losing the transition smoothness especially while handling moving objects. The second focuses on video frame interpolation. Most existing studies attempt to estimate intermediate optical flows [25, 27, 35, 55], or leverage frame conditioning at training time [4, 5, 22, 50]. However, they tend to generate irrational object transitions with abrupt content changes, struggle with producing long transition sequences, and require time-consuming training on large-scale motion videos. On top of the above, existing studies either employ deep interpolation for conceptual blending transitions [16, 51] or incorporate a random-mask condition layer for scene transition [8], lacking a uniform framework that can work for multiple transition generation tasks. This prompts a compelling inquiry: can we design a versatile transition generator that can handle various transition tasks with minimal adaptation across tasks?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.777 + ], + "angle": 0, + "content": "As illustrated in Figure 1, we propose four transition tasks to establish versatile transition generation, based on the types of input frames: (1) Object Morphing: The input frames are either the same object with different postures or different objects, as long as they are topologically similar; (2) Concept Blending: The input frames contain conceptually different objects (e.g., 'an airplane' and 'a cruise'); (3) Motion Prediction: The input frames can be seen as two moments in a video containing one (or more) moving object; (4) Scene Transition: The input frames are conceptually related scene images yet they either belong to two different domains (e.g., 'a wooden house in the forest' and 'a wooden house in the snow') or represent two distinct components of a scene (e.g., 'erupting volcano' and 'hot lava')." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We design VTG, a versatile transition generation framework built upon image-to-video diffusion models. VTG features three designs for generating smooth, high-fidelity, and semantic-coherent transitions. First, we employ interpolation-based initialization that effectively mitigates abrupt content changes while handling input frames with significantly different contents. Specifically, we spherically interpolate latent Gaussian noises of the two input frames" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.363 + ], + "angle": 0, + "content": "and text embeddings of corresponding transition captions and leverage two LoRA-integrated U-Nets [20] to capture object-level semantics in the denoising steps. Second, we adopt a lightweight fine-tuning strategy that merges the predicted forward and backward noises which greatly improves the transition smoothness in motion predictions. The bidirectional sampling can be achieved by fine-tuning the pretrained U-Net with a small collection of videos. Third, we introduce regularization with self-supervised visual encoder for video diffusion models to explicitly induce feature alignment, which helps learn meaningful representations and enhance the fidelity of generated transition videos. These components are logically cohesive and, in combination, distinguish VTG from prior work by enabling a single framework to handle diverse transition tasks with minimal task-specific adjustments. In addition, we collected TransitBench, a new benchmark of 200 pairs with first and last frames for concept blending and scene transition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.559 + ], + "angle": 0, + "content": "The contributions of this work can be summarized in three aspects. First, we propose a unified task of versatile transition generation, aiming at smooth and rational transition for object morphing, concept blending, motion prediction, and scene transition. Second, we design VTG, a novel and versatile framework that can generate semantic-relevant, high-fidelity and temporally coherent video transitions effectively. Third, we introduce TransitBench, a curated dataset for benchmarking concept blending and scene transition. With TransitBench and other public benchmarks, we demonstrate, both qualitatively and quantitatively, that VTG outperforms the state-of-the-art consistently across the four transition generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.655, + 0.589 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.905, + 0.81 + ], + "angle": 0, + "content": "Image Morphing. Image morphing has been a longstanding problem in the computer graphics community [1, 49, 61]. Conventional morphing techniques either utilize correspondence-driven bidirectional image warping [2, 3, 9, 26] or minimize the path energy on a Riemannian manifold [14, 29, 33, 40] to obtain image transitions. While producing seamless morphing, they either require massive human involvement or fail to create new content beyond the two given images. Recently, several studies [52, 56] explored the prior knowledge in image diffusion models for real-world morphing with more object categories. Differently, we exploit a pretrained video diffusion model for image morphing, naturally enabling coherent transitions across frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Video Frame Interpolation. Video frame interpolation (VFI) has been extensively studied in the computer vision research community [11]. Classic VFI algorithms are closely related to optical flow prediction, which obtain interpolation by either forward warping [31] or backward" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16982" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "warping [21, 23, 25] based on the estimated flows. Several follow-up studies [22, 27, 35] focus on large motion problem in VFI. Recently, a line of research [12, 46] explores bounded video generation with image-to-video diffusion models under the guidance of the start and end frames. This approach fuses outputs from forward and backward paths by interpolating the predicted forward noise and its reversed counterpart. While generating coherent motion, it overlooks the identity preservation and struggles to produce natural transitions when the two input frames differ significantly in content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.485, + 0.545 + ], + "angle": 0, + "content": "Transition Generation. Transition generation is a far underexplored problem, largely due to its open-ended nature. Typically, a transition generation framework leverages two input frames (i.e., the first and last frames) and a transition caption to produce a scene-level transition video that connects two distinct narrative moments. This has been explored in [8] that incorporates random masking to selectively suppress information from the original latent code, allowing capturing transition subtleties between frames. In addition, [50, 54] inject the first and last frames as additional conditions by concatenating them with noisy latent codes along the channel dimension. Recently, a concurrent study [57] explores transition video generation with imaged-to-video diffusion as well. Differently, we formulate the task by versatile transition generation by unifying four transition tasks under the same framework, which greatly expands the applicability and significantly enhances the quality of the generated transition videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.183, + 0.571 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.581, + 0.231, + 0.595 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.679 + ], + "angle": 0, + "content": "Latent diffusion models [36] are text-to-image diffusion models that operate in a compressed latent space. During forward diffusion process, it encodes an image sample \\(\\mathbf{x}_0\\) into a latent code \\(\\mathbf{z}_0\\) and corrupts \\(\\mathbf{z}_0\\) by Gaussian noise at each time step \\(t = 1,\\dots ,T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.686, + 0.483, + 0.705 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.772 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) denotes the noise strength at time step \\(t\\) and \\(\\mathbf{I}\\) denotes the identity matrix with the same dimensions as \\(\\mathbf{z}_0\\). Given a text embedding \\(\\mathbf{c}\\), the backward denoising process can be parameterized as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.779, + 0.483, + 0.796 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.484, + 0.879 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}(\\mathbf{z}_t,\\mathbf{c},t)\\) is output by a U-Net [37], and \\(\\Sigma_{\\theta}(\\mathbf{z}_t,\\mathbf{c},t)\\) is determined by a noise scheduler (e.g., DDPM [18] or DDIM [42]). The model learns to predict noise via a parameterized noise estimator \\(\\epsilon_{\\theta}\\) under the guidance of an objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\mathbb {E} _ {t \\sim [ 1, T ], \\mathbf {z} _ {0}, \\mathbf {c}, \\epsilon_ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.333 + ], + "angle": 0, + "content": "Our framework is built upon pretrained image-to-video diffusion models that incorporate temporal convolutions and temporal attention layers on top of latent diffusion models to establish temporal correlations between video frames. Given a video \\(\\mathbf{x}_0\\in \\mathbb{R}^{N\\times 3\\times H\\times W}\\), where \\(N\\) is the number of frames and \\(H\\times W\\) indicates the image resolution, each frame is first encoded into a latent representation \\(\\mathbf{z}_0\\in \\mathbb{R}^{N\\times C\\times h\\times w}\\) via the VAE encoder \\(\\mathcal{E}_v\\). Both forward diffusion process and backward denoising process are then performed in the latent space. The generated video frames can be collectively obtained through the decoder \\(\\mathcal{D}_v\\). To inject image conditions, we fuse and incorporate text-conditioned and image-conditioned features by projecting the conditional image into a text-aligned embedding space. Meanwhile, the conditional image is concatenated with the initial per-frame noise to preserve more visual details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.341, + 0.806, + 0.357 + ], + "angle": 0, + "content": "3.2. Interpolation-based Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.483 + ], + "angle": 0, + "content": "In diffusion models, the initial Gaussian noise determines the coarse structure that emerges in the early denoising steps, while high-frequency details are refined later. Since existing image-to-video diffusion models [8, 50] randomly initialize the latent code \\(\\mathbf{z}_0\\) at inference time, each intermediate frame follows a distinct stochastic trajectory: colors, micro-textures, or even object pose shift slightly from frame to frame, yielding perceptual \"flicker\" [10]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.909, + 0.621 + ], + "angle": 0, + "content": "To suppress this drift, we correlate the frame-wise latents by interpolating between the endpoint noises instead of resampling them independently. Prior work [6] indicates that linear interpolation often yields intermediate latent norms that are extremely unlikely under a unit Gaussian, whereas spherical linear interpolation (SLERP) preserves the Euclidean norm and keeps samples on-distribution. Therefore, we leverage SLERP between the latent noises of two input frames \\(\\mathbf{z}_{t1}\\) and \\(\\mathbf{z}_{tN}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.626, + 0.907, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t n} = \\frac {\\sin ((1 - \\lambda_ {\\text {n o i s e}}) \\phi)}{\\sin \\phi} \\mathbf {z} _ {t 1} + \\frac {\\sin (\\lambda_ {\\text {n o i s e}} \\phi)}{\\sin \\phi} \\mathbf {z} _ {t N}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.825 + ], + "angle": 0, + "content": "where \\(\\phi = \\arccos \\left(\\frac{\\mathbf{z}_{t1}^T\\cdot\\mathbf{z}_{tN}}{\\|\\mathbf{z}_{t1}\\|_{\\|}\\mathbf{z}_{tN}\\|}\\right)\\), and \\(\\lambda_{\\mathrm{noise}}\\in [0,1]\\) denotes the parameter for latent interpolation. As shown in Figure 2, we concatenate \\(\\mathbf{z}_{tn}\\) with \\(\\mathbf{z}_{t1}\\) and \\(\\mathbf{z}_{tN}\\) along the frame dimension and inject them as the initial latent noise for DDIM sampling. It is worth noting that we only inject the interpolated latent noises at early denoising steps to preserve the appearance and motion priors from the image-to-video backbone as much as possible. With such meticulously designed latent injection strategy, our approach addresses the issue of random and abrupt transition frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Unlike GANs [15], whose latent space is structured and semantically meaningful, diffusion models operate in an unstructured noise space that does not explicitly encode high-level semantics [56]. This lack of structure can lead to artifacts when interpolating between semantically different in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.089, + 0.855, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.483, + 0.908, + 0.555 + ], + "angle": 0, + "content": "Figure 2. Inference framework of VTG. Our interpolation-based initialization features three designs: \\(①\\) Interpolated Noise Injection, \\(②\\) LoRA Interpolation, and \\(③\\) Frame-aware Text Interpolation. VTG first converts two encoded input frames to latent noises via DDIM inversion. Then, it interpolates between the two latent noises and concatenate intermediate noises along the temporal dimension. To capture meaningful semantics and enable the transition between two conceptually different objects, we employ LoRA interpolation and text interpolation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.626 + ], + "angle": 0, + "content": "puts. To suppress such transition artifacts, we first train two LoRAs \\(\\Delta \\theta_{1}\\), \\(\\Delta \\theta_{N}\\) given two input frames \\(x^{1}\\) and \\(x^{N}\\), optimizing the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.634, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\Delta \\theta_ {t}\\right) = \\mathbb {E} _ {\\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta + \\Delta \\theta_ {i}} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0 n} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, \\mathbf {c} _ {n}\\right) \\| ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.718 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_{0n}\\) is the encoded latent vector of frame \\(n\\), and \\(\\mathbf{c}_n\\) is the text embedding associated with transition caption \\(c^n\\). Then, we linearly interpolate the two LoRAs to fuse the semantics of two input frames:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.726, + 0.483, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta = \\left(1 - \\lambda_ {l o r a}\\right) \\Delta \\theta_ {1} + \\lambda_ {l o r a} \\Delta \\theta_ {N}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.765 + ], + "angle": 0, + "content": "where \\(\\lambda_{lora}\\) denotes the parameter during lora interpolation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Typically, video diffusion models [8, 50] employ only one caption as the text condition. This hinders the transition generation between two conceptually different objects, causing abrupt content changes without generating intermediate frames with hybrid meanings. However, for imaged-to-video diffusion models, the text embedding of the entire frame sequence \\( c \\) is integrated within the denoising U-Net via the cross-attention layer, without explicitly defining per-frame text embedding. A related still-image approach [52]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.626 + ], + "angle": 0, + "content": "linearly blends the two text embeddings to morph between two static endpoints, whereas our method performs frame-aware SLERP across prompt sequences as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.638, + 0.907, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {\\lambda_ {t e x t}} = \\frac {\\sin ((1 - \\lambda_ {t e x t}) \\phi)}{\\sin \\phi} \\mathbf {c} _ {1} + \\frac {\\sin (\\lambda_ {t e x t} \\phi)}{\\sin \\phi} \\mathbf {c} _ {N}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.908, + 0.73 + ], + "angle": 0, + "content": "where \\(\\lambda_{text} \\in [0,1]\\) serves as the frame-aware coefficient to control the transition sequence; \\(\\mathbf{c}_1\\) and \\(\\mathbf{c}_N\\) denote the text embeddings of the first and last frames, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.795, + 0.757 + ], + "angle": 0, + "content": "3.3. Bidirectional Motion Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In our experiments, we observed notable quality differences when reversing the order of input frames. This discrepancy arises from two main factors: (1) the model shows a bias toward resembling the initial input frame more closely due to conditional image leakage [59]; (2) existing image-to-video diffusion models [8, 50] are pretrained exclusively for forward motion prediction, while real-world motion is naturally asymmetric, resulting in ambiguity when predicting reverse motion." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16984" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.862, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.391, + 0.908, + 0.462 + ], + "angle": 0, + "content": "Figure 3. Training framework of VTG. In Bidirectional Motion Prediction (BMP), the noisy latent is flipped along temporal dimension, and self-attention maps undergo a 180-degree rotation to establish reversed motion-time correlations. Two U-Nets separately predict forward and backward motion, with the backward noise reversed again to fuse with the forward noise, ensuring a consistent motion path for iterative denoising. In Representation Alignment Regularization (RAR), we spatially patchify each frame independently, then aggregate the per-patch alignment loss across temporal dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Inspired by [46], we simultaneously predict forward and backward motions to mitigate the ambiguity issue. Specifically, we rotate the temporal self-attention maps by 180 degrees both horizontally and vertically, causing a reversal of attention relationships as illustrated in Figure 3. We also reverse the original forward trajectory of video \\(\\{x^{1}, x^{2}, \\ldots, x^{N}\\}\\) to \\(\\{x^{N}, x^{N - 1}, \\ldots, x^{1}\\}\\), given the two input frames \\(x^{1}\\) and \\(x^{N}\\). Then, we obtain the reversed noisy latent \\(\\mathbf{z}_{t'}\\) by flipping \\(\\mathbf{z}_t\\) along the temporal dimension. After that, \\(\\mathbf{z}_{t'}\\) is fed into the 3D U-Net for backward motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.734 + ], + "angle": 0, + "content": "During training, we employ a lightweight fine-tuning strategy that only updates the parameters of value and output matrices (i.e., \\(\\theta_{w,o}\\)) in the temporal attention layers. The objective \\(\\mathcal{L}_{BMP}\\) is formulated by taking the L2 norm between the predicted noise and ground-truth reversed noise:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.755, + 0.483, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B M P} = \\left\\| \\operatorname {f l i p} \\left(\\epsilon_ {t}\\right) - \\epsilon_ {\\theta_ {w, o}} \\left(\\mathbf {z} _ {t ^ {\\prime}}, \\mathbf {c}, t, A _ {i, j} ^ {\\prime}\\right) \\right\\| _ {2} ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.835 + ], + "angle": 0, + "content": "The predicted backward noise is reversed again to fuse with forward noise. Specifically, we simply employ linear interpolation to ensure forward-backward consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.852, + 0.483, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} = \\left(1 - \\lambda_ {B M P}\\right) \\epsilon_ {t, i} + \\lambda_ {B M P} \\epsilon_ {t, N - i} ^ {\\prime}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.42, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{BMP} = 0.5\\) denotes the weighting factor." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.487, + 0.871, + 0.504 + ], + "angle": 0, + "content": "3.4. Representation Alignment Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Though previous techniques effectively mitigate abrupt content changes and motion-reversal ambiguity, our approach remains susceptible to generating blurry and low-fidelity transitions, especially when input frames contain fine-grained textures (e.g., bicycle spokes or fabric weave). Recent work [53] shows that diffusion latents inherently lack high-frequency semantics compared to self-supervised representations like DINOv2's [32]. Such mismatch tends to accumulate over video frames and appears as blur." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To overcome this, we propose to distill DINOv2 features back into the denoising trajectory of video diffusion. As illustrated in Figure 3, the video latent is first patched into \\(N\\) sequences of \\(T\\) tokens along the temporal dimension. For a N-frame video, the latent vector \\(\\mathbf{z}_n\\) of \\(n\\)-th frame has the shape of \\(C\\times h\\times w\\). Given a pre-defined patch size \\(s\\), the sequence length \\(T\\) should be: \\(T = (h / s)\\times (w / s)\\). Then, we project the per-frame latent representation \\(\\mathbf{z}_n\\) of video diffusion via a multilayer perceptron (MLP) to align with DINOv2 representation \\(\\mathbf{y}_{*} = \\mathcal{E}_{d}(\\mathbf{x}_{*})\\in \\mathbb{R}^{P\\times m}\\), where \\(\\mathbf{y}_{*}\\) denotes the encoder output given a clean video frame \\(\\mathbf{x}_{*}\\); \\(P\\) and \\(m\\) respectively denotes the number of patches and the embedding dimension of DINOv2 encoder \\(\\mathcal{E}_d\\). Let \\(y_{\\phi}(\\mathbf{h}_t)\\) be the projected representation of intermediate diffusion latent feature \\(\\mathbf{h}_t\\), where \\(\\phi\\) denotes the trainable parameters of the MLP. The regularization term \\(\\mathcal{L}_{RAR}\\) is computed by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16985" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.108 + ], + "angle": 0, + "content": "aggregating patch-wise similarities across all video frames:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.13, + 0.483, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R A R}} = - \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {t, \\mathbf {x} _ {*}, \\epsilon_ {t}} \\left[ \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\sin \\left(\\mathbf {y} _ {*} ^ {[ p ]}, y _ {\\phi} (\\mathbf {h} _ {t}) ^ {[ p ]}\\right) \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.162, + 0.483, + 0.192 + ], + "angle": 0, + "content": "During inference, the external encoder \\(\\mathcal{E}_d\\) and MLP projector \\(y_{\\phi}\\) are discarded." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.206, + 0.224, + 0.223 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.23, + 0.281, + 0.246 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Training Details. We leverage a small collection (i.e., 150) of high-quality in-house videos as our training data. These video samples encompass various motion (e.g., people running, airplanes taxiing) and appearance changes (e.g., a man transitioning from youth to old age). The learnable parameters of the backward U-Net are initialized to zero, while the other available weights are initialized from the pre-trained checkpoint of DynamiCrafter [50]. We utilize the AdamW optimizer with learning rate of \\(1 \\times e^{-5}\\). VTG is trained on 4 NVIDIA A100-80GB GPUs, taking \\(\\sim 20K\\) iterations with batch size of 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.63 + ], + "angle": 0, + "content": "Inference Details. In our experiments, we use Stable Diffusion v2.1-base [36] as our interpolation backbone. When training LoRAs, we set the LoRA rank to 16 and train them for 200 steps using AdamW optimizer [28] with a learning rate of \\(2 \\times 10^{-4}\\). This requires only \\(\\sim 85\\)s with 1 NVIDIA A100-80GB GPU. During inversion and sampling, we use the DDIM sampler [42] and set total time steps to the default value (i.e., 50) of the selected image-to-video models. We apply text classifier-free guidance [17] for all models with the same negative prompt \"Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms\" across all transition generations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.721 + ], + "angle": 0, + "content": "Comparison Methods. We compare VTG with following methods: DiffMorpher [56], TVG [57], SEINE [8], DynamiCrafter [50], Generative Inbetweening [46], Text Embedding Interpolation (TEI) [45], Denoising Interpolation (DI) [16], and Attention Interpolation (AID) [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation Benchmarks. We employ several public datasets and our self-curated TransitBench to comprehensively evaluate the quality of the generated transition videos. Specifically, we leverage MorphBench [56] and TC-Bench [13] to evaluate object morphing and motion prediction. Concept blending and scene transition are rarely explored in the realm of generative models. Hence, there is a lack of specific evaluation benchmarks for these two tasks. Previously, [16] employs CIFAR-10 [24] and LAION-Aesthetics [39] as their evaluation benchmarks to test concept blending. While there are sufficient number of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "samples in both datasets, CIFAR-10 only contains concepts from 10 classes, and LAION-Aesthetics consists entirely of synthetic images and lacks real-world samples. To this end, we present TransitBench, the first benchmark dataset for collectively assessing concept blending transitions of two distinct conceptual objects and scene transitions between two relevant scenarios. We collected 200 pairs of pictures (each pair forms the first and the last frames of one transition generation sample) of diverse content and styles, and evenly divide them into two categories: 1) concept-blending cases, and 2) scene-transition cases, both of which are obtained from web resources. We hope TransitBench can promote future studies on general transition generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.3, + 0.697, + 0.315 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.838 + ], + "angle": 0, + "content": "To demonstrate the superiority of our method, we provide a visual comparison of VTG against five state-of-the-art baselines. Specifically, we compare our approach with DiffMorpher, Generative Inbetweening, TVG, SEINE, DynamamiCrafter across four transition tasks for uniformity. Figure 4 shows the generation results by different methods. For object morphing, the compared methods exhibit low-fidelity (e.g., DiffMorpher, DynamamiCrafter), oversaturation effect (e.g., TVG), and inaccurate semantics (e.g., Generative Inbetweening, SEINE). In contrast, VTG presents a natural transition, preserving the content of the first and last frames well. For motion prediction, Generative Inbetweening and SEINE generate inferior transitions as observed by their blurry intermediate frames. TVG yields unnatural transitions as it generates a front-facing image of a person, while the person's face is actually turned away. DiffMorpher shows low-fidelity transitions, and its generated video looks like a sequence of intermittent images without concerning the consistency between background and foreground objects. This is largely due to the missing temporal modeling in DiffMorpher's image diffusion backbone. VTG achieves comparable or even better results with DynamamiCrafter, validating that it can effectively preserve the appearance and motion priors in diffusion models. For concept blending, most video diffusion baselines present abrupt content changes, while other baselines show blurry (e.g., DiffMorpher) and ambiguous (e.g., TVG) intermediate frames. In contrast, VTG yields smooth transition between the lion and the truck with rational intermediate results (e.g., a lion-colored and lion-sized truck). For scene transition between two relevant scene images, existing methods suffer from abrupt content changes or low similarity (e.g., TVG), while VTG produces natural scene-level transition with pleasing effect." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.708, + 0.864 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. The objective of versatile transition generation is to generate a sequence of \\((N - 2)\\) intermedi-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16986" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.885, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.746, + 0.908, + 0.788 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on four transition tasks. For each task, the displayed images correspond to the 1st, 4th, 8th, 12th, and 16th video frames, respectively. Please refer to our project page for more video examples, which offers a more straightforward and comprehensive qualitative comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.486, + 0.892 + ], + "angle": 0, + "content": "ate video frames, denoted as \\( x^{2:N - 1} \\), between two source frames \\( x^1 \\) and \\( x^N \\), with two captions describing the two frames. High-quality transition videos should simultaneously satisfy four criteria: semantic fidelity to the input endpoints, visual quality, temporal coherence, and perceptual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.814, + 0.909, + 0.892 + ], + "angle": 0, + "content": "smoothness. To ensure fair comparisons, we performed extensive quantitative evaluations using experimental settings consistent with those prior methods as in their original paper. Specifically, we adopt the following evaluation metrics: (1) FID \\((\\downarrow)\\) and PPL \\((\\downarrow)\\) follow DiffMorpher [56]. FID com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16987" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "pares the first and last frames against the generated transition frames in the Inception feature space; PPL measures the average LPIPS fluctuation along an interpolation path, where lower values imply fewer perceptual jumps. (2) TCR \\((\\uparrow)\\) and TC-Score \\((\\uparrow)\\) are taken from the training-free TVG framework [57]. TCR is the fraction of video frames whose CLIP embedding stays within a pre-set similarity margin to either endpoint; TC-Score complements it with a mean cosine-similarity, offering a softer view of frame-level consistency. (3) Smoothness \\((\\uparrow)\\) is adapted from AID [16]. For every adjacent frame pair, we compute LPIPS and the Gini coefficient of these distances, reflecting how even the step-to-step changes are. We report its inverse so that a higher score denotes smoother transitions." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.311, + 0.48, + 0.426 + ], + "angle": 0, + "content": "
MethodMetamorphosisAnimation
FID (↓)PPL (↓)FID (↓)PPL (↓)
DiffMorpher [56]70.4918.1943.155.14
TVG [57]86.9235.1842.9912.46
SEINE [8]82.0347.7248.2516.26
DynamiCrafter [50]87.3242.0943.3111.16
VTG (Ours)67.3922.8039.165.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.435, + 0.483, + 0.464 + ], + "angle": 0, + "content": "Table 1. Quantitative results on MorphBench. The best results are in bold; second-best are underlined." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.488, + 0.481, + 0.561 + ], + "angle": 0, + "content": "
MethodAttributeObjectBackground
TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)
DiffMorpher [56]41.820.84419.570.76550.000.819
SEINE [8]17.860.72010.480.6547.960.742
DynamiCrafter [50]16.550.74513.910.70725.560.795
TVG [57]41.820.87730.440.82238.890.864
VTG (Ours)42.780.89333.460.84950.000.883
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Table 2. Quantitative results on TC-Bench. The best results are in bold. Best viewed when zoomed in." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.625, + 0.483, + 0.68 + ], + "angle": 0, + "content": "
DatasetTEIDIAID-OAID-IOurs
CIFAR-100.75310.75640.78310.78610.7932
LAION-Aesthetics0.74240.75110.76430.81520.8215
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Table 3. Smoothness (↑) evaluation for Concept Blending. The best results are in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Quantitative Results. As seen in Table 1 and Table 2, VTG outperforms existing methods with their original experimental setting, demonstrating its capability to generate high-quality, temporally coherent, and semantically consistent video transitions. It is worth noting that the only exception is Metamorphosis subset, where VTG achieves the second-best PPL (22.80). Nevertheless, VTG still demonstrates outstanding overall performance, achieving state-of-the-art FID and comparable PPL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "According to AID [16], a well-interpolated concept-blending sequence should exhibit a gradual and smooth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "transition. To offer a more comprehensive evaluation, we conducted experiments using the same benchmarks and metric as in AID. Please refer to Table 3 for details." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.347, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Figure 5. The bar plot of user preference rates \\((\\%)\\) of different methods on four transition tasks. Best viewed when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.909, + 0.651 + ], + "angle": 0, + "content": "User Study. To subjectively evaluate our approach, we further conduct a user study. We asked Amazon Mechanical Turk (AMT) workers to choose the best generated transition video from a set of five candidates. Given the two input frames and corresponding transition captions, AMT workers evaluate the videos based on three questions: 1) Which video is the most semantic-coherent? 2) Which video is the most temporally coherent? 3) Which video has the highest fidelity? To make a fair evaluation, we include static and obviously incorrect videos to identify random clicking. Each test example is assigned to 10 different workers for evaluation. We collected 150 valid responses for each transition task after filtering out those responses from random clicking workers. The results are shown in Figure 5, which illustrates that VTG obtains the highest preference rate with its superior quality of the generated transition videos compared to other baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we formulate the problem of versatile transition generation that encompasses four representative transition tasks, namely, object morphing, motion prediction, concept blending, and scene transition. Leveraging a pretrained image-to-video diffusion model, we design VTG, a unified Versatile Transition video Generator that can tackle the four tasks within a single framework. Both qualitative and quantitative experiments show that VTG achieves superior performance in generating semantically relevant, temporally coherent, and visually pleasing transition video frames, given a pair of text prompts and two input frames. With these advantages, VTG can be readily applied to real-world content creation scenarios, offering an efficient tool for generating high-quality transition in media production." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16988" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This study is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.18, + 0.188, + 0.196 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.205, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[1] Alyaa Qusay Aloraibi. Image morphing techniques: A review. Technium: Romanian Journal of Applied Sciences and Technology, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.249, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[2] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. Seminal Graphics Papers: Pushing the Boundaries, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.484, + 0.335 + ], + "angle": 0, + "content": "[3] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.337, + 0.484, + 0.407 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.409, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[6] Erik Bodin, Alexandru I Stere, Dragos D Margineantu, Carl Henrik Ek, and Henry Moss. Linear combinations of latents in generative models: subspaces and beyond. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.583, + 0.484, + 0.652 + ], + "angle": 0, + "content": "[8] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.655, + 0.484, + 0.709 + ], + "angle": 0, + "content": "[9] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. TOG, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.713, + 0.484, + 0.768 + ], + "angle": 0, + "content": "[10] Yitong Deng, Winnie Lin, Lingxiao Li, Dmitriy Smirnov, Ryan Burgert, Ning Yu, Vincent Dedun, and Mohammad H Taghavi. Infinite-resolution integral noise warping for diffusion models. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.771, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[11] Jiong Dong, Kaoru Ota, and Mianxiong Dong. Video frame interpolation: A comprehensive survey. TOMM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[12] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Weixi Feng, Jiachen Li, Michael Saxon, Tsu-jui Fu, Wenhu Chen, and William Yang Wang. Tc-bench: Benchmarking temporal compositionality in text-to-video and image-to-video generation. In ACL, 2025. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.205, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[14] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[16] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. In NeurIPS, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.279 + ], + "angle": 0, + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.311, + 0.906, + 0.38 + ], + "angle": 0, + "content": "[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.382, + 0.905, + 0.436 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.439, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hólynski, Ben Poole, and Janne Kontkanen. Video interpolation with diffusion models. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[23] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.624 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[25] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.686, + 0.905, + 0.739 + ], + "angle": 0, + "content": "[26] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. TOG, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[27] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[29] Michael I. Miller and Laurent Younes. Group actions, homeomorphisms, and matching: A general framework. *IJCV*, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16989" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[31] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[32] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[33] Marko Rajković. Geodesics, Splines, and Embeddings in Spaces of Images. PhD thesis, Universitäts-und Landesbibliothek Bonn, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[35] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.623 + ], + "angle": 0, + "content": "[40] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In CVPR, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. *ICLR*, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[44] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[45] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. In ICMLW, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[46] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. In ICLR, 2025. 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[47] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[48] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.206, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[49] George Wolberg. Image morphing: a survey. The visual computer, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.234, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[50] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.291, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[51] Cheng Yang, Lijing Liang, and Zhixun Su. Realworld denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.334, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[52] Zhaoyuan Yang, Zhengyang Yu, Zhiwei Xu, Jaskirat Singh, Jing Zhang, Dylan Campbell, Peter Tu, and Richard Hartley. Impus: Image morphing with perceptually-uniform sampling using diffusion models. In ICLR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.391, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[53] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. In ICLR, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.448, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[54] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.49, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[55] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.546, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[56] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Xingang Pan, and Bo Dai. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In CVPR, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.59, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[57] Rui Zhang, Yaosen Chen, Yuegen Liu, Wei Wang, Xuming Wen, and Hongxia Wang. TVg: A training-free transition video generation method with diffusion models. TCVT, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.646, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[58] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.716, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[59] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. In NeurIPS, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.772, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[60] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.83, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[61] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. IJAEMS, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16990" + } + ] +] \ No newline at end of file diff --git a/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_origin.pdf b/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..66098eb012593aa0784112b515d9fc1e08c54741 --- /dev/null +++ b/2025/Versatile Transition Generation with Image-to-Video Diffusion/f3fc723c-8a6e-4204-9849-5843b598e850_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e65b5e85e16bec8423d6840015b9c35049907f1b25c2f9a2deed414fec79fd7e +size 3794974 diff --git a/2025/Versatile Transition Generation with Image-to-Video Diffusion/full.md b/2025/Versatile Transition Generation with Image-to-Video Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25a89a468038016e386dacc2971b45caab0b502d --- /dev/null +++ b/2025/Versatile Transition Generation with Image-to-Video Diffusion/full.md @@ -0,0 +1,286 @@ +# Versatile Transition Generation with Image-to-Video Diffusion + +Zuhao Yang $^{1*}$ Jiahui Zhang $^{1}$ Yingchen Yu $^{2}$ Shijian Lu $^{1\dagger}$ Song Bai $^{2}$ + +$^{1}$ Nanyang Technological University $^{2}$ ByteDance Inc. + +https://mwxely.github.io/projects/yang2025vtg/index + +![](images/8483849a237828a82a977654c85310005bc8c13829c0a98488a121268f2da6bc.jpg) +(a) "a dog facing forward" $\rightarrow$ "a dog turning his head to the right" + +![](images/9e866d0401e18f5493c63449527279a3f69085053d9f580d68d19f526e7a3314.jpg) +(c) "an airplane" $\rightarrow$ "a cruise" + +![](images/6badb648b44404f8e188979e9cc5dd3ddf50a4edf74574b11298eb792c354739.jpg) +(b) "a woman riding a horse" $\rightarrow$ "a woman moving forward on a horse" + +![](images/d1473901c7fea11184cdb67aec676be7fc83b51aa433663614e21aa4708bb2bb.jpg) +(d) "a wooden house in the forest" $\rightarrow$ "a wooden house in the snow" +Figure 1. Illustration of Versatile Transition Generation (VTG). VTG is capable of performing four types of transition generation, namely, (a) object morphing, (b) motion prediction, (c) concept blending, and (d) scene transition, within a single uniform framework. + +# Abstract + +Leveraging text, images, structure maps, or motion trajectories as conditional guidance, diffusion models have achieved great success in automated and high-quality video generation. However, generating smooth and rational transition videos given the first and last video frames as well as descriptive text prompts is far underexplored. We present VTG, a Versatile Transition video Generation framework that can generate smooth, high-fidelity, and semantic-coherent video transitions. VTG introduces interpolation-based initialization that helps preserve object identity and handle abrupt content changes effectively. In addition, it incorporates dual-directional motion fine-tuning and representation alignment regularization that mitigate the limitations of the pre-trained image-to-video diffusion models in motion smoothness and generation fidelity, respectively. To evaluate VTG and facilitate future studies on unified transition generation, we collected TransitBench, a com + +prehensive benchmark for transition generation that covers two representative transition tasks including concept blending and scene transition. Extensive experiments show that VTG achieves superior transition performance consistently across the four tasks. + +# 1. Introduction + +The success of diffusion models [18, 43] in image synthesis [30, 34, 36, 38] has inspired a number of studies on diffusion-based video synthesis [19, 47, 58, 60]. Leveraging textual prompts, video frames, structure maps and even motion patterns, several studies [7, 41, 44, 48] have demonstrated impressive synthesis performance by generating realistic and high-fidelity videos automatically. Nevertheless, despite its significant value in various real-world tasks such as video and film production, generating high-quality transition videos conditioned on the first and last frames together with related text prompts remains largely underexplored. + +Generating realistic transition videos is a non-trivial task. A high-quality transition generator should meet at least four + +criteria: 1) semantic similarity to the input frames; 2) high fidelity to the input frames; 3) smoothness across the generated frames; and 4) alignment with the provided text prompts. Moreover, the research community in transition generation mostly relies on self-collected and well-filtered videos that are not accessible to the public [8, 54]. This further hinders the advancement of the research in transition video generation. + +Most existing studies tackle the challenge of transition generation via two representative approaches. The first focuses on morphing, given two images of topologically similar objects. Most recent methods [52, 56] harness various deep interpolation techniques for generating rational object-level transition. However, they generate intermittent images rather than temporally coherent video frames, losing the transition smoothness especially while handling moving objects. The second focuses on video frame interpolation. Most existing studies attempt to estimate intermediate optical flows [25, 27, 35, 55], or leverage frame conditioning at training time [4, 5, 22, 50]. However, they tend to generate irrational object transitions with abrupt content changes, struggle with producing long transition sequences, and require time-consuming training on large-scale motion videos. On top of the above, existing studies either employ deep interpolation for conceptual blending transitions [16, 51] or incorporate a random-mask condition layer for scene transition [8], lacking a uniform framework that can work for multiple transition generation tasks. This prompts a compelling inquiry: can we design a versatile transition generator that can handle various transition tasks with minimal adaptation across tasks? + +As illustrated in Figure 1, we propose four transition tasks to establish versatile transition generation, based on the types of input frames: (1) Object Morphing: The input frames are either the same object with different postures or different objects, as long as they are topologically similar; (2) Concept Blending: The input frames contain conceptually different objects (e.g., 'an airplane' and 'a cruise'); (3) Motion Prediction: The input frames can be seen as two moments in a video containing one (or more) moving object; (4) Scene Transition: The input frames are conceptually related scene images yet they either belong to two different domains (e.g., 'a wooden house in the forest' and 'a wooden house in the snow') or represent two distinct components of a scene (e.g., 'erupting volcano' and 'hot lava'). + +We design VTG, a versatile transition generation framework built upon image-to-video diffusion models. VTG features three designs for generating smooth, high-fidelity, and semantic-coherent transitions. First, we employ interpolation-based initialization that effectively mitigates abrupt content changes while handling input frames with significantly different contents. Specifically, we spherically interpolate latent Gaussian noises of the two input frames + +and text embeddings of corresponding transition captions and leverage two LoRA-integrated U-Nets [20] to capture object-level semantics in the denoising steps. Second, we adopt a lightweight fine-tuning strategy that merges the predicted forward and backward noises which greatly improves the transition smoothness in motion predictions. The bidirectional sampling can be achieved by fine-tuning the pretrained U-Net with a small collection of videos. Third, we introduce regularization with self-supervised visual encoder for video diffusion models to explicitly induce feature alignment, which helps learn meaningful representations and enhance the fidelity of generated transition videos. These components are logically cohesive and, in combination, distinguish VTG from prior work by enabling a single framework to handle diverse transition tasks with minimal task-specific adjustments. In addition, we collected TransitBench, a new benchmark of 200 pairs with first and last frames for concept blending and scene transition. + +The contributions of this work can be summarized in three aspects. First, we propose a unified task of versatile transition generation, aiming at smooth and rational transition for object morphing, concept blending, motion prediction, and scene transition. Second, we design VTG, a novel and versatile framework that can generate semantic-relevant, high-fidelity and temporally coherent video transitions effectively. Third, we introduce TransitBench, a curated dataset for benchmarking concept blending and scene transition. With TransitBench and other public benchmarks, we demonstrate, both qualitatively and quantitatively, that VTG outperforms the state-of-the-art consistently across the four transition generation tasks. + +# 2. Related Work + +Image Morphing. Image morphing has been a longstanding problem in the computer graphics community [1, 49, 61]. Conventional morphing techniques either utilize correspondence-driven bidirectional image warping [2, 3, 9, 26] or minimize the path energy on a Riemannian manifold [14, 29, 33, 40] to obtain image transitions. While producing seamless morphing, they either require massive human involvement or fail to create new content beyond the two given images. Recently, several studies [52, 56] explored the prior knowledge in image diffusion models for real-world morphing with more object categories. Differently, we exploit a pretrained video diffusion model for image morphing, naturally enabling coherent transitions across frames. + +Video Frame Interpolation. Video frame interpolation (VFI) has been extensively studied in the computer vision research community [11]. Classic VFI algorithms are closely related to optical flow prediction, which obtain interpolation by either forward warping [31] or backward + +warping [21, 23, 25] based on the estimated flows. Several follow-up studies [22, 27, 35] focus on large motion problem in VFI. Recently, a line of research [12, 46] explores bounded video generation with image-to-video diffusion models under the guidance of the start and end frames. This approach fuses outputs from forward and backward paths by interpolating the predicted forward noise and its reversed counterpart. While generating coherent motion, it overlooks the identity preservation and struggles to produce natural transitions when the two input frames differ significantly in content. + +Transition Generation. Transition generation is a far underexplored problem, largely due to its open-ended nature. Typically, a transition generation framework leverages two input frames (i.e., the first and last frames) and a transition caption to produce a scene-level transition video that connects two distinct narrative moments. This has been explored in [8] that incorporates random masking to selectively suppress information from the original latent code, allowing capturing transition subtleties between frames. In addition, [50, 54] inject the first and last frames as additional conditions by concatenating them with noisy latent codes along the channel dimension. Recently, a concurrent study [57] explores transition video generation with imaged-to-video diffusion as well. Differently, we formulate the task by versatile transition generation by unifying four transition tasks under the same framework, which greatly expands the applicability and significantly enhances the quality of the generated transition videos. + +# 3. Method + +# 3.1. Preliminaries + +Latent diffusion models [36] are text-to-image diffusion models that operate in a compressed latent space. During forward diffusion process, it encodes an image sample $\mathbf{x}_0$ into a latent code $\mathbf{z}_0$ and corrupts $\mathbf{z}_0$ by Gaussian noise at each time step $t = 1,\dots ,T$ : + +$$ +q \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {z} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {z} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\beta_{t}$ denotes the noise strength at time step $t$ and $\mathbf{I}$ denotes the identity matrix with the same dimensions as $\mathbf{z}_0$ . Given a text embedding $\mathbf{c}$ , the backward denoising process can be parameterized as: + +$$ +p _ {\theta} \left(\mathbf {z} _ {t - 1} \mid \mathbf {z} _ {t}\right) = \mathcal {N} \left(\mathbf {z} _ {t - 1}; \mu_ {\theta} \left(\mathbf {z} _ {t}, \mathbf {c}, t\right), \Sigma_ {\theta} \left(\mathbf {z} _ {t}, \mathbf {c}, t\right)\right), \tag {2} +$$ + +where $\mu_{\theta}(\mathbf{z}_t,\mathbf{c},t)$ is output by a U-Net [37], and $\Sigma_{\theta}(\mathbf{z}_t,\mathbf{c},t)$ is determined by a noise scheduler (e.g., DDPM [18] or DDIM [42]). The model learns to predict noise via a parameterized noise estimator $\epsilon_{\theta}$ under the guidance of an objective function: + +$$ +\mathcal {L} _ {t} = \mathbb {E} _ {t \sim [ 1, T ], \mathbf {z} _ {0}, \mathbf {c}, \epsilon_ {t} \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \| \epsilon_ {t} - \epsilon_ {\theta} (\mathbf {z} _ {t}, \mathbf {c}, t) \| _ {2} ^ {2}. \tag {3} +$$ + +Our framework is built upon pretrained image-to-video diffusion models that incorporate temporal convolutions and temporal attention layers on top of latent diffusion models to establish temporal correlations between video frames. Given a video $\mathbf{x}_0\in \mathbb{R}^{N\times 3\times H\times W}$ , where $N$ is the number of frames and $H\times W$ indicates the image resolution, each frame is first encoded into a latent representation $\mathbf{z}_0\in \mathbb{R}^{N\times C\times h\times w}$ via the VAE encoder $\mathcal{E}_v$ . Both forward diffusion process and backward denoising process are then performed in the latent space. The generated video frames can be collectively obtained through the decoder $\mathcal{D}_v$ . To inject image conditions, we fuse and incorporate text-conditioned and image-conditioned features by projecting the conditional image into a text-aligned embedding space. Meanwhile, the conditional image is concatenated with the initial per-frame noise to preserve more visual details. + +# 3.2. Interpolation-based Initialization + +In diffusion models, the initial Gaussian noise determines the coarse structure that emerges in the early denoising steps, while high-frequency details are refined later. Since existing image-to-video diffusion models [8, 50] randomly initialize the latent code $\mathbf{z}_0$ at inference time, each intermediate frame follows a distinct stochastic trajectory: colors, micro-textures, or even object pose shift slightly from frame to frame, yielding perceptual "flicker" [10]. + +To suppress this drift, we correlate the frame-wise latents by interpolating between the endpoint noises instead of resampling them independently. Prior work [6] indicates that linear interpolation often yields intermediate latent norms that are extremely unlikely under a unit Gaussian, whereas spherical linear interpolation (SLERP) preserves the Euclidean norm and keeps samples on-distribution. Therefore, we leverage SLERP between the latent noises of two input frames $\mathbf{z}_{t1}$ and $\mathbf{z}_{tN}$ as follows: + +$$ +\mathbf {z} _ {t n} = \frac {\sin ((1 - \lambda_ {\text {n o i s e}}) \phi)}{\sin \phi} \mathbf {z} _ {t 1} + \frac {\sin (\lambda_ {\text {n o i s e}} \phi)}{\sin \phi} \mathbf {z} _ {t N}, \tag {4} +$$ + +where $\phi = \arccos \left(\frac{\mathbf{z}_{t1}^T\cdot\mathbf{z}_{tN}}{\|\mathbf{z}_{t1}\|_{\|}\mathbf{z}_{tN}\|}\right)$ , and $\lambda_{\mathrm{noise}}\in [0,1]$ denotes the parameter for latent interpolation. As shown in Figure 2, we concatenate $\mathbf{z}_{tn}$ with $\mathbf{z}_{t1}$ and $\mathbf{z}_{tN}$ along the frame dimension and inject them as the initial latent noise for DDIM sampling. It is worth noting that we only inject the interpolated latent noises at early denoising steps to preserve the appearance and motion priors from the image-to-video backbone as much as possible. With such meticulously designed latent injection strategy, our approach addresses the issue of random and abrupt transition frames. + +Unlike GANs [15], whose latent space is structured and semantically meaningful, diffusion models operate in an unstructured noise space that does not explicitly encode high-level semantics [56]. This lack of structure can lead to artifacts when interpolating between semantically different in + +![](images/fa27ac34aa3c8e70ab2e6846d79d0da738d1e1b4cc8fa4d40bf787a792e30614.jpg) +Figure 2. Inference framework of VTG. Our interpolation-based initialization features three designs: $①$ Interpolated Noise Injection, $②$ LoRA Interpolation, and $③$ Frame-aware Text Interpolation. VTG first converts two encoded input frames to latent noises via DDIM inversion. Then, it interpolates between the two latent noises and concatenate intermediate noises along the temporal dimension. To capture meaningful semantics and enable the transition between two conceptually different objects, we employ LoRA interpolation and text interpolation, respectively. + +puts. To suppress such transition artifacts, we first train two LoRAs $\Delta \theta_{1}$ , $\Delta \theta_{N}$ given two input frames $x^{1}$ and $x^{N}$ , optimizing the following objective: + +$$ +\mathcal {L} \left(\Delta \theta_ {t}\right) = \mathbb {E} _ {\epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta + \Delta \theta_ {i}} \left(\sqrt {\bar {\alpha} _ {t}} \mathbf {z} _ {0 n} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t, \mathbf {c} _ {n}\right) \| ^ {2} \right], \tag {5} +$$ + +where $\mathbf{z}_{0n}$ is the encoded latent vector of frame $n$ , and $\mathbf{c}_n$ is the text embedding associated with transition caption $c^n$ . Then, we linearly interpolate the two LoRAs to fuse the semantics of two input frames: + +$$ +\Delta \theta = \left(1 - \lambda_ {l o r a}\right) \Delta \theta_ {1} + \lambda_ {l o r a} \Delta \theta_ {N}, \tag {6} +$$ + +where $\lambda_{lora}$ denotes the parameter during lora interpolation. + +Typically, video diffusion models [8, 50] employ only one caption as the text condition. This hinders the transition generation between two conceptually different objects, causing abrupt content changes without generating intermediate frames with hybrid meanings. However, for imaged-to-video diffusion models, the text embedding of the entire frame sequence $c$ is integrated within the denoising U-Net via the cross-attention layer, without explicitly defining per-frame text embedding. A related still-image approach [52] + +linearly blends the two text embeddings to morph between two static endpoints, whereas our method performs frame-aware SLERP across prompt sequences as follows: + +$$ +\mathbf {c} _ {\lambda_ {t e x t}} = \frac {\sin ((1 - \lambda_ {t e x t}) \phi)}{\sin \phi} \mathbf {c} _ {1} + \frac {\sin (\lambda_ {t e x t} \phi)}{\sin \phi} \mathbf {c} _ {N}, \tag {7} +$$ + +where $\lambda_{text} \in [0,1]$ serves as the frame-aware coefficient to control the transition sequence; $\mathbf{c}_1$ and $\mathbf{c}_N$ denote the text embeddings of the first and last frames, respectively. + +# 3.3. Bidirectional Motion Prediction + +In our experiments, we observed notable quality differences when reversing the order of input frames. This discrepancy arises from two main factors: (1) the model shows a bias toward resembling the initial input frame more closely due to conditional image leakage [59]; (2) existing image-to-video diffusion models [8, 50] are pretrained exclusively for forward motion prediction, while real-world motion is naturally asymmetric, resulting in ambiguity when predicting reverse motion. + +![](images/a844dcd3564d5a50c7a030960a63a525bbacd9a18de01bd94be556adfd4c9b48.jpg) +Figure 3. Training framework of VTG. In Bidirectional Motion Prediction (BMP), the noisy latent is flipped along temporal dimension, and self-attention maps undergo a 180-degree rotation to establish reversed motion-time correlations. Two U-Nets separately predict forward and backward motion, with the backward noise reversed again to fuse with the forward noise, ensuring a consistent motion path for iterative denoising. In Representation Alignment Regularization (RAR), we spatially patchify each frame independently, then aggregate the per-patch alignment loss across temporal dimension. + +Inspired by [46], we simultaneously predict forward and backward motions to mitigate the ambiguity issue. Specifically, we rotate the temporal self-attention maps by 180 degrees both horizontally and vertically, causing a reversal of attention relationships as illustrated in Figure 3. We also reverse the original forward trajectory of video $\{x^{1}, x^{2}, \ldots, x^{N}\}$ to $\{x^{N}, x^{N - 1}, \ldots, x^{1}\}$ , given the two input frames $x^{1}$ and $x^{N}$ . Then, we obtain the reversed noisy latent $\mathbf{z}_{t'}$ by flipping $\mathbf{z}_t$ along the temporal dimension. After that, $\mathbf{z}_{t'}$ is fed into the 3D U-Net for backward motion prediction. + +During training, we employ a lightweight fine-tuning strategy that only updates the parameters of value and output matrices (i.e., $\theta_{w,o}$ ) in the temporal attention layers. The objective $\mathcal{L}_{BMP}$ is formulated by taking the L2 norm between the predicted noise and ground-truth reversed noise: + +$$ +\mathcal {L} _ {B M P} = \left\| \operatorname {f l i p} \left(\epsilon_ {t}\right) - \epsilon_ {\theta_ {w, o}} \left(\mathbf {z} _ {t ^ {\prime}}, \mathbf {c}, t, A _ {i, j} ^ {\prime}\right) \right\| _ {2} ^ {2}. \tag {8} +$$ + +The predicted backward noise is reversed again to fuse with forward noise. Specifically, we simply employ linear interpolation to ensure forward-backward consistency: + +$$ +\epsilon_ {t} = \left(1 - \lambda_ {B M P}\right) \epsilon_ {t, i} + \lambda_ {B M P} \epsilon_ {t, N - i} ^ {\prime}, \tag {9} +$$ + +where $\lambda_{BMP} = 0.5$ denotes the weighting factor. + +# 3.4. Representation Alignment Regularization + +Though previous techniques effectively mitigate abrupt content changes and motion-reversal ambiguity, our approach remains susceptible to generating blurry and low-fidelity transitions, especially when input frames contain fine-grained textures (e.g., bicycle spokes or fabric weave). Recent work [53] shows that diffusion latents inherently lack high-frequency semantics compared to self-supervised representations like DINOv2's [32]. Such mismatch tends to accumulate over video frames and appears as blur. + +To overcome this, we propose to distill DINOv2 features back into the denoising trajectory of video diffusion. As illustrated in Figure 3, the video latent is first patched into $N$ sequences of $T$ tokens along the temporal dimension. For a N-frame video, the latent vector $\mathbf{z}_n$ of $n$ -th frame has the shape of $C\times h\times w$ . Given a pre-defined patch size $s$ , the sequence length $T$ should be: $T = (h / s)\times (w / s)$ . Then, we project the per-frame latent representation $\mathbf{z}_n$ of video diffusion via a multilayer perceptron (MLP) to align with DINOv2 representation $\mathbf{y}_{*} = \mathcal{E}_{d}(\mathbf{x}_{*})\in \mathbb{R}^{P\times m}$ , where $\mathbf{y}_{*}$ denotes the encoder output given a clean video frame $\mathbf{x}_{*}$ ; $P$ and $m$ respectively denotes the number of patches and the embedding dimension of DINOv2 encoder $\mathcal{E}_d$ . Let $y_{\phi}(\mathbf{h}_t)$ be the projected representation of intermediate diffusion latent feature $\mathbf{h}_t$ , where $\phi$ denotes the trainable parameters of the MLP. The regularization term $\mathcal{L}_{RAR}$ is computed by + +aggregating patch-wise similarities across all video frames: + +$$ +\mathcal {L} _ {\mathrm {R A R}} = - \sum_ {n = 1} ^ {N} \mathbb {E} _ {t, \mathbf {x} _ {*}, \epsilon_ {t}} \left[ \frac {1}{P} \sum_ {p = 1} ^ {P} \sin \left(\mathbf {y} _ {*} ^ {[ p ]}, y _ {\phi} (\mathbf {h} _ {t}) ^ {[ p ]}\right) \right]. \tag {10} +$$ + +During inference, the external encoder $\mathcal{E}_d$ and MLP projector $y_{\phi}$ are discarded. + +# 4. Experiments + +# 4.1. Experimental Setup + +Training Details. We leverage a small collection (i.e., 150) of high-quality in-house videos as our training data. These video samples encompass various motion (e.g., people running, airplanes taxiing) and appearance changes (e.g., a man transitioning from youth to old age). The learnable parameters of the backward U-Net are initialized to zero, while the other available weights are initialized from the pre-trained checkpoint of DynamiCrafter [50]. We utilize the AdamW optimizer with learning rate of $1 \times e^{-5}$ . VTG is trained on 4 NVIDIA A100-80GB GPUs, taking $\sim 20K$ iterations with batch size of 2. + +Inference Details. In our experiments, we use Stable Diffusion v2.1-base [36] as our interpolation backbone. When training LoRAs, we set the LoRA rank to 16 and train them for 200 steps using AdamW optimizer [28] with a learning rate of $2 \times 10^{-4}$ . This requires only $\sim 85$ s with 1 NVIDIA A100-80GB GPU. During inversion and sampling, we use the DDIM sampler [42] and set total time steps to the default value (i.e., 50) of the selected image-to-video models. We apply text classifier-free guidance [17] for all models with the same negative prompt "Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms" across all transition generations. + +Comparison Methods. We compare VTG with following methods: DiffMorpher [56], TVG [57], SEINE [8], DynamiCrafter [50], Generative Inbetweening [46], Text Embedding Interpolation (TEI) [45], Denoising Interpolation (DI) [16], and Attention Interpolation (AID) [16]. + +Evaluation Benchmarks. We employ several public datasets and our self-curated TransitBench to comprehensively evaluate the quality of the generated transition videos. Specifically, we leverage MorphBench [56] and TC-Bench [13] to evaluate object morphing and motion prediction. Concept blending and scene transition are rarely explored in the realm of generative models. Hence, there is a lack of specific evaluation benchmarks for these two tasks. Previously, [16] employs CIFAR-10 [24] and LAION-Aesthetics [39] as their evaluation benchmarks to test concept blending. While there are sufficient number of + +samples in both datasets, CIFAR-10 only contains concepts from 10 classes, and LAION-Aesthetics consists entirely of synthetic images and lacks real-world samples. To this end, we present TransitBench, the first benchmark dataset for collectively assessing concept blending transitions of two distinct conceptual objects and scene transitions between two relevant scenarios. We collected 200 pairs of pictures (each pair forms the first and the last frames of one transition generation sample) of diverse content and styles, and evenly divide them into two categories: 1) concept-blending cases, and 2) scene-transition cases, both of which are obtained from web resources. We hope TransitBench can promote future studies on general transition generation. + +# 4.2. Qualitative Results + +To demonstrate the superiority of our method, we provide a visual comparison of VTG against five state-of-the-art baselines. Specifically, we compare our approach with DiffMorpher, Generative Inbetweening, TVG, SEINE, DynamamiCrafter across four transition tasks for uniformity. Figure 4 shows the generation results by different methods. For object morphing, the compared methods exhibit low-fidelity (e.g., DiffMorpher, DynamamiCrafter), oversaturation effect (e.g., TVG), and inaccurate semantics (e.g., Generative Inbetweening, SEINE). In contrast, VTG presents a natural transition, preserving the content of the first and last frames well. For motion prediction, Generative Inbetweening and SEINE generate inferior transitions as observed by their blurry intermediate frames. TVG yields unnatural transitions as it generates a front-facing image of a person, while the person's face is actually turned away. DiffMorpher shows low-fidelity transitions, and its generated video looks like a sequence of intermittent images without concerning the consistency between background and foreground objects. This is largely due to the missing temporal modeling in DiffMorpher's image diffusion backbone. VTG achieves comparable or even better results with DynamamiCrafter, validating that it can effectively preserve the appearance and motion priors in diffusion models. For concept blending, most video diffusion baselines present abrupt content changes, while other baselines show blurry (e.g., DiffMorpher) and ambiguous (e.g., TVG) intermediate frames. In contrast, VTG yields smooth transition between the lion and the truck with rational intermediate results (e.g., a lion-colored and lion-sized truck). For scene transition between two relevant scene images, existing methods suffer from abrupt content changes or low similarity (e.g., TVG), while VTG produces natural scene-level transition with pleasing effect. + +# 4.3.Quantitative Results + +Evaluation Metrics. The objective of versatile transition generation is to generate a sequence of $(N - 2)$ intermedi- + +![](images/032d98d612a2dbe24edff782b7194d13dc08b915212ef9ff448676a217916a34.jpg) +Figure 4. Qualitative comparisons on four transition tasks. For each task, the displayed images correspond to the 1st, 4th, 8th, 12th, and 16th video frames, respectively. Please refer to our project page for more video examples, which offers a more straightforward and comprehensive qualitative comparisons. + +ate video frames, denoted as $x^{2:N - 1}$ , between two source frames $x^1$ and $x^N$ , with two captions describing the two frames. High-quality transition videos should simultaneously satisfy four criteria: semantic fidelity to the input endpoints, visual quality, temporal coherence, and perceptual + +smoothness. To ensure fair comparisons, we performed extensive quantitative evaluations using experimental settings consistent with those prior methods as in their original paper. Specifically, we adopt the following evaluation metrics: (1) FID $(\downarrow)$ and PPL $(\downarrow)$ follow DiffMorpher [56]. FID com + +pares the first and last frames against the generated transition frames in the Inception feature space; PPL measures the average LPIPS fluctuation along an interpolation path, where lower values imply fewer perceptual jumps. (2) TCR $(\uparrow)$ and TC-Score $(\uparrow)$ are taken from the training-free TVG framework [57]. TCR is the fraction of video frames whose CLIP embedding stays within a pre-set similarity margin to either endpoint; TC-Score complements it with a mean cosine-similarity, offering a softer view of frame-level consistency. (3) Smoothness $(\uparrow)$ is adapted from AID [16]. For every adjacent frame pair, we compute LPIPS and the Gini coefficient of these distances, reflecting how even the step-to-step changes are. We report its inverse so that a higher score denotes smoother transitions. + +
MethodMetamorphosisAnimation
FID (↓)PPL (↓)FID (↓)PPL (↓)
DiffMorpher [56]70.4918.1943.155.14
TVG [57]86.9235.1842.9912.46
SEINE [8]82.0347.7248.2516.26
DynamiCrafter [50]87.3242.0943.3111.16
VTG (Ours)67.3922.8039.165.14
+ +Table 1. Quantitative results on MorphBench. The best results are in bold; second-best are underlined. + +
MethodAttributeObjectBackground
TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)
DiffMorpher [56]41.820.84419.570.76550.000.819
SEINE [8]17.860.72010.480.6547.960.742
DynamiCrafter [50]16.550.74513.910.70725.560.795
TVG [57]41.820.87730.440.82238.890.864
VTG (Ours)42.780.89333.460.84950.000.883
+ +Table 2. Quantitative results on TC-Bench. The best results are in bold. Best viewed when zoomed in. + +
DatasetTEIDIAID-OAID-IOurs
CIFAR-100.75310.75640.78310.78610.7932
LAION-Aesthetics0.74240.75110.76430.81520.8215
+ +Table 3. Smoothness (↑) evaluation for Concept Blending. The best results are in bold. + +Quantitative Results. As seen in Table 1 and Table 2, VTG outperforms existing methods with their original experimental setting, demonstrating its capability to generate high-quality, temporally coherent, and semantically consistent video transitions. It is worth noting that the only exception is Metamorphosis subset, where VTG achieves the second-best PPL (22.80). Nevertheless, VTG still demonstrates outstanding overall performance, achieving state-of-the-art FID and comparable PPL. + +According to AID [16], a well-interpolated concept-blending sequence should exhibit a gradual and smooth + +transition. To offer a more comprehensive evaluation, we conducted experiments using the same benchmarks and metric as in AID. Please refer to Table 3 for details. + +![](images/71e79c1636d39dd458cf4f41f926f41fccb8d006114a8eb4e6109c7f72c019f9.jpg) +Figure 5. The bar plot of user preference rates $(\%)$ of different methods on four transition tasks. Best viewed when zoomed in. + +User Study. To subjectively evaluate our approach, we further conduct a user study. We asked Amazon Mechanical Turk (AMT) workers to choose the best generated transition video from a set of five candidates. Given the two input frames and corresponding transition captions, AMT workers evaluate the videos based on three questions: 1) Which video is the most semantic-coherent? 2) Which video is the most temporally coherent? 3) Which video has the highest fidelity? To make a fair evaluation, we include static and obviously incorrect videos to identify random clicking. Each test example is assigned to 10 different workers for evaluation. We collected 150 valid responses for each transition task after filtering out those responses from random clicking workers. The results are shown in Figure 5, which illustrates that VTG obtains the highest preference rate with its superior quality of the generated transition videos compared to other baselines. + +# 5. Conclusion + +In this paper, we formulate the problem of versatile transition generation that encompasses four representative transition tasks, namely, object morphing, motion prediction, concept blending, and scene transition. Leveraging a pretrained image-to-video diffusion model, we design VTG, a unified Versatile Transition video Generator that can tackle the four tasks within a single framework. Both qualitative and quantitative experiments show that VTG achieves superior performance in generating semantically relevant, temporally coherent, and visually pleasing transition video frames, given a pair of text prompts and two input frames. With these advantages, VTG can be readily applied to real-world content creation scenarios, offering an efficient tool for generating high-quality transition in media production. + +# Acknowledgements + +This study is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22. + +# References + +[1] Alyaa Qusay Aloraibi. Image morphing techniques: A review. Technium: Romanian Journal of Applied Sciences and Technology, 2023. 2 +[2] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. Seminal Graphics Papers: Pushing the Boundaries, 1992. 2 +[3] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2011. 2 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2 +[6] Erik Bodin, Alexandru I Stere, Dragos D Margineantu, Carl Henrik Ek, and Henry Moss. Linear combinations of latents in generative models: subspaces and beyond. In ICLR, 2025. 3 +[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 1 +[8] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 2, 3, 4, 6, 8 +[9] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. TOG, 2012. 2 +[10] Yitong Deng, Winnie Lin, Lingxiao Li, Dmitriy Smirnov, Ryan Burgert, Ning Yu, Vincent Dedun, and Mohammad H Taghavi. Infinite-resolution integral noise warping for diffusion models. In ICLR, 2025. 3 +[11] Jiong Dong, Kaoru Ota, and Mianxiong Dong. Video frame interpolation: A comprehensive survey. TOMM, 2023. 2 +[12] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In ECCV, 2024. 3 +[13] Weixi Feng, Jiachen Li, Michael Saxon, Tsu-jui Fu, Wenhu Chen, and William Yang Wang. Tc-bench: Benchmarking temporal compositionality in text-to-video and image-to-video generation. In ACL, 2025. 6 + +[14] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, 2020. 2 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3 +[16] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. In NeurIPS, 2024. 2, 6, 8 +[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 6 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3 +[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2 +[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In ECCV, 2022. 3 +[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hólynski, Ben Poole, and Janne Kontkanen. Video interpolation with diffusion models. In CVPR, 2024. 2, 3 +[23] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In CVPR, 2022. 3 +[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 6 +[25] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 2, 3 +[26] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. TOG, 2014. 2 +[27] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In CVPR, 2024. 2, 3 +[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2018. 6 +[29] Michael I. Miller and Laurent Younes. Group actions, homeomorphisms, and matching: A general framework. *IJCV*, 2001. 2 +[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and + +Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 1 +[31] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 2 +[32] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 5 +[33] Marko Rajković. Geodesics, Splines, and Embeddings in Spaces of Images. PhD thesis, Universitäts-und Landesbibliothek Bonn, 2023. 2 +[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[35] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In ECCV, 2022. 2, 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6 +[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3 +[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1 +[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 6 +[40] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In CVPR, 2010. 2 +[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1 +[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. *ICLR*, 2021. 3, 6 +[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1 +[44] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 1 +[45] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. In ICMLW, 2023. 6 +[46] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. In ICLR, 2025. 3, 5, 6 + +[47] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1 +[48] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH, 2024. 1 +[49] George Wolberg. Image morphing: a survey. The visual computer, 1998. 2 +[50] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3, 4, 6, 8 +[51] Cheng Yang, Lijing Liang, and Zhixun Su. Realworld denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023. 2 +[52] Zhaoyuan Yang, Zhengyang Yu, Zhiwei Xu, Jaskirat Singh, Jing Zhang, Dylan Campbell, Peter Tu, and Richard Hartley. Impus: Image morphing with perceptually-uniform sampling using diffusion models. In ICLR, 2024. 2, 4 +[53] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. In ICLR, 2025. 5 +[54] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In CVPR, 2024. 2, 3 +[55] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In CVPR, 2023. 2 +[56] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Xingang Pan, and Bo Dai. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In CVPR, 2024. 2, 3, 6, 7, 8 +[57] Rui Zhang, Yaosen Chen, Yuegen Liu, Wei Wang, Xuming Wen, and Hongxia Wang. TVg: A training-free transition video generation method with diffusion models. TCVT, 2024. 3, 6, 8 +[58] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 1 +[59] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. In NeurIPS, 2024. 4 +[60] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1 +[61] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. IJAEMS, 2017. 2 \ No newline at end of file diff --git a/2025/Versatile Transition Generation with Image-to-Video Diffusion/images.zip b/2025/Versatile Transition Generation with Image-to-Video Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44f7c1c85141aafc66a2feae715aa7b02f3b0c77 --- /dev/null +++ b/2025/Versatile Transition Generation with Image-to-Video Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba19466ad127e3323216c887a2574a5d9b493c49555e4659ebaca542257be420 +size 760805 diff --git a/2025/Versatile Transition Generation with Image-to-Video Diffusion/layout.json b/2025/Versatile Transition Generation with Image-to-Video Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c24f0f44fc482e130970ed7f78c368028dcb4340 --- /dev/null +++ b/2025/Versatile Transition Generation with Image-to-Video Diffusion/layout.json @@ -0,0 +1,8230 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 499, + 121 + ], + "type": "text", + "content": "Versatile Transition Generation with Image-to-Video Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "spans": [ + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "text", + "content": "Zuhao Yang" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "text", + "content": " Jiahui Zhang" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "text", + "content": " Yingchen Yu" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "text", + "content": " Shijian Lu" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "text", + "content": " Song Bai" + }, + { + "bbox": [ + 128, + 142, + 480, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 162, + 443, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 162, + 443, + 178 + ], + "spans": [ + { + "bbox": [ + 167, + 162, + 443, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 167, + 162, + 443, + 178 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 167, + 162, + 443, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 167, + 162, + 443, + 178 + ], + "type": "text", + "content": "ByteDance Inc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 183, + 432, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 183, + 432, + 198 + ], + "spans": [ + { + "bbox": [ + 175, + 183, + 432, + 198 + ], + "type": "text", + "content": "https://mwxely.github.io/projects/yang2025vtg/index" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 226, + 289, + 299 + ], + "blocks": [ + { + "bbox": [ + 58, + 226, + 289, + 299 + ], + "lines": [ + { + "bbox": [ + 58, + 226, + 289, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 226, + 289, + 299 + ], + "type": "image", + "image_path": "8483849a237828a82a977654c85310005bc8c13829c0a98488a121268f2da6bc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 304, + 277, + 316 + ], + "lines": [ + { + "bbox": [ + 59, + 304, + 277, + 316 + ], + "spans": [ + { + "bbox": [ + 59, + 304, + 277, + 316 + ], + "type": "text", + "content": "(a) \"a dog facing forward\" " + }, + { + "bbox": [ + 59, + 304, + 277, + 316 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 59, + 304, + 277, + 316 + ], + "type": "text", + "content": " \"a dog turning his head to the right\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 322, + 289, + 395 + ], + "blocks": [ + { + "bbox": [ + 59, + 322, + 289, + 395 + ], + "lines": [ + { + "bbox": [ + 59, + 322, + 289, + 395 + ], + "spans": [ + { + "bbox": [ + 59, + 322, + 289, + 395 + ], + "type": "image", + "image_path": "9e866d0401e18f5493c63449527279a3f69085053d9f580d68d19f526e7a3314.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 400, + 160, + 411 + ], + "lines": [ + { + "bbox": [ + 59, + 400, + 160, + 411 + ], + "spans": [ + { + "bbox": [ + 59, + 400, + 160, + 411 + ], + "type": "text", + "content": "(c) \"an airplane\" " + }, + { + "bbox": [ + 59, + 400, + 160, + 411 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 59, + 400, + 160, + 411 + ], + "type": "text", + "content": " \"a cruise\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 226, + 553, + 299 + ], + "blocks": [ + { + "bbox": [ + 317, + 226, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 317, + 226, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 226, + 553, + 299 + ], + "type": "image", + "image_path": "6badb648b44404f8e188979e9cc5dd3ddf50a4edf74574b11298eb792c354739.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 304, + 544, + 316 + ], + "lines": [ + { + "bbox": [ + 318, + 304, + 544, + 316 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 544, + 316 + ], + "type": "text", + "content": "(b) \"a woman riding a horse\" " + }, + { + "bbox": [ + 318, + 304, + 544, + 316 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 318, + 304, + 544, + 316 + ], + "type": "text", + "content": " \"a woman moving forward on a horse\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 322, + 552, + 395 + ], + "blocks": [ + { + "bbox": [ + 317, + 322, + 552, + 395 + ], + "lines": [ + { + "bbox": [ + 317, + 322, + 552, + 395 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 552, + 395 + ], + "type": "image", + "image_path": "d1473901c7fea11184cdb67aec676be7fc83b51aa433663614e21aa4708bb2bb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 400, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 318, + 400, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 318, + 400, + 546, + 411 + ], + "type": "text", + "content": "(d) \"a wooden house in the forest\" " + }, + { + "bbox": [ + 318, + 400, + 546, + 411 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 318, + 400, + 546, + 411 + ], + "type": "text", + "content": " \"a wooden house in the snow\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 421, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 54, + 421, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 54, + 421, + 555, + 444 + ], + "type": "text", + "content": "Figure 1. Illustration of Versatile Transition Generation (VTG). VTG is capable of performing four types of transition generation, namely, (a) object morphing, (b) motion prediction, (c) concept blending, and (d) scene transition, within a single uniform framework." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 453, + 200, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 453, + 200, + 466 + ], + "spans": [ + { + "bbox": [ + 151, + 453, + 200, + 466 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 479, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 297, + 684 + ], + "type": "text", + "content": "Leveraging text, images, structure maps, or motion trajectories as conditional guidance, diffusion models have achieved great success in automated and high-quality video generation. However, generating smooth and rational transition videos given the first and last video frames as well as descriptive text prompts is far underexplored. We present VTG, a Versatile Transition video Generation framework that can generate smooth, high-fidelity, and semantic-coherent video transitions. VTG introduces interpolation-based initialization that helps preserve object identity and handle abrupt content changes effectively. In addition, it incorporates dual-directional motion fine-tuning and representation alignment regularization that mitigate the limitations of the pre-trained image-to-video diffusion models in motion smoothness and generation fidelity, respectively. To evaluate VTG and facilitate future studies on unified transition generation, we collected TransitBench, a com" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 455, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 556, + 515 + ], + "type": "text", + "content": "prehensive benchmark for transition generation that covers two representative transition tasks including concept blending and scene transition. Extensive experiments show that VTG achieves superior transition performance consistently across the four tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 537, + 394, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 537, + 394, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 537, + 394, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 557, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 555, + 689 + ], + "type": "text", + "content": "The success of diffusion models [18, 43] in image synthesis [30, 34, 36, 38] has inspired a number of studies on diffusion-based video synthesis [19, 47, 58, 60]. Leveraging textual prompts, video frames, structure maps and even motion patterns, several studies [7, 41, 44, 48] have demonstrated impressive synthesis performance by generating realistic and high-fidelity videos automatically. Nevertheless, despite its significant value in various real-world tasks such as video and film production, generating high-quality transition videos conditioned on the first and last frames together with related text prompts remains largely underexplored." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Generating realistic transition videos is a non-trivial task. A high-quality transition generator should meet at least four" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 290, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 290, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 290, + 703 + ], + "type": "text", + "content": "*This work was done while Zuhao Yang was interning at ByteDance." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 67, + 703, + 197, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 197, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 197, + 713 + ], + "type": "text", + "content": "Shijian Lu is the corresponding author." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16981" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "criteria: 1) semantic similarity to the input frames; 2) high fidelity to the input frames; 3) smoothness across the generated frames; and 4) alignment with the provided text prompts. Moreover, the research community in transition generation mostly relies on self-collected and well-filtered videos that are not accessible to the public [8, 54]. This further hinders the advancement of the research in transition video generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 298, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 298, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 298, + 445 + ], + "type": "text", + "content": "Most existing studies tackle the challenge of transition generation via two representative approaches. The first focuses on morphing, given two images of topologically similar objects. Most recent methods [52, 56] harness various deep interpolation techniques for generating rational object-level transition. However, they generate intermittent images rather than temporally coherent video frames, losing the transition smoothness especially while handling moving objects. The second focuses on video frame interpolation. Most existing studies attempt to estimate intermediate optical flows [25, 27, 35, 55], or leverage frame conditioning at training time [4, 5, 22, 50]. However, they tend to generate irrational object transitions with abrupt content changes, struggle with producing long transition sequences, and require time-consuming training on large-scale motion videos. On top of the above, existing studies either employ deep interpolation for conceptual blending transitions [16, 51] or incorporate a random-mask condition layer for scene transition [8], lacking a uniform framework that can work for multiple transition generation tasks. This prompts a compelling inquiry: can we design a versatile transition generator that can handle various transition tasks with minimal adaptation across tasks?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 449, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 615 + ], + "type": "text", + "content": "As illustrated in Figure 1, we propose four transition tasks to establish versatile transition generation, based on the types of input frames: (1) Object Morphing: The input frames are either the same object with different postures or different objects, as long as they are topologically similar; (2) Concept Blending: The input frames contain conceptually different objects (e.g., 'an airplane' and 'a cruise'); (3) Motion Prediction: The input frames can be seen as two moments in a video containing one (or more) moving object; (4) Scene Transition: The input frames are conceptually related scene images yet they either belong to two different domains (e.g., 'a wooden house in the forest' and 'a wooden house in the snow') or represent two distinct components of a scene (e.g., 'erupting volcano' and 'hot lava')." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "We design VTG, a versatile transition generation framework built upon image-to-video diffusion models. VTG features three designs for generating smooth, high-fidelity, and semantic-coherent transitions. First, we employ interpolation-based initialization that effectively mitigates abrupt content changes while handling input frames with significantly different contents. Specifically, we spherically interpolate latent Gaussian noises of the two input frames" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": "and text embeddings of corresponding transition captions and leverage two LoRA-integrated U-Nets [20] to capture object-level semantics in the denoising steps. Second, we adopt a lightweight fine-tuning strategy that merges the predicted forward and backward noises which greatly improves the transition smoothness in motion predictions. The bidirectional sampling can be achieved by fine-tuning the pretrained U-Net with a small collection of videos. Third, we introduce regularization with self-supervised visual encoder for video diffusion models to explicitly induce feature alignment, which helps learn meaningful representations and enhance the fidelity of generated transition videos. These components are logically cohesive and, in combination, distinguish VTG from prior work by enabling a single framework to handle diverse transition tasks with minimal task-specific adjustments. In addition, we collected TransitBench, a new benchmark of 200 pairs with first and last frames for concept blending and scene transition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 288, + 553, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 442 + ], + "type": "text", + "content": "The contributions of this work can be summarized in three aspects. First, we propose a unified task of versatile transition generation, aiming at smooth and rational transition for object morphing, concept blending, motion prediction, and scene transition. Second, we design VTG, a novel and versatile framework that can generate semantic-relevant, high-fidelity and temporally coherent video transitions effectively. Third, we introduce TransitBench, a curated dataset for benchmarking concept blending and scene transition. With TransitBench and other public benchmarks, we demonstrate, both qualitatively and quantitatively, that VTG outperforms the state-of-the-art consistently across the four transition generation tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 453, + 400, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 400, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 400, + 466 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 475, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 641 + ], + "type": "text", + "content": "Image Morphing. Image morphing has been a longstanding problem in the computer graphics community [1, 49, 61]. Conventional morphing techniques either utilize correspondence-driven bidirectional image warping [2, 3, 9, 26] or minimize the path energy on a Riemannian manifold [14, 29, 33, 40] to obtain image transitions. While producing seamless morphing, they either require massive human involvement or fail to create new content beyond the two given images. Recently, several studies [52, 56] explored the prior knowledge in image diffusion models for real-world morphing with more object categories. Differently, we exploit a pretrained video diffusion model for image morphing, naturally enabling coherent transitions across frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Video Frame Interpolation. Video frame interpolation (VFI) has been extensively studied in the computer vision research community [11]. Classic VFI algorithms are closely related to optical flow prediction, which obtain interpolation by either forward warping [31] or backward" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16982" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "warping [21, 23, 25] based on the estimated flows. Several follow-up studies [22, 27, 35] focus on large motion problem in VFI. Recently, a line of research [12, 46] explores bounded video generation with image-to-video diffusion models under the guidance of the start and end frames. This approach fuses outputs from forward and backward paths by interpolating the predicted forward noise and its reversed counterpart. While generating coherent motion, it overlooks the identity preservation and struggles to produce natural transitions when the two input frames differ significantly in content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 431 + ], + "type": "text", + "content": "Transition Generation. Transition generation is a far underexplored problem, largely due to its open-ended nature. Typically, a transition generation framework leverages two input frames (i.e., the first and last frames) and a transition caption to produce a scene-level transition video that connects two distinct narrative moments. This has been explored in [8] that incorporates random masking to selectively suppress information from the original latent code, allowing capturing transition subtleties between frames. In addition, [50, 54] inject the first and last frames as additional conditions by concatenating them with noisy latent codes along the channel dimension. Recently, a concurrent study [57] explores transition video generation with imaged-to-video diffusion as well. Differently, we formulate the task by versatile transition generation by unifying four transition tasks under the same framework, which greatly expands the applicability and significantly enhances the quality of the generated transition videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 441, + 111, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 111, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 111, + 452 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 141, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 141, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 141, + 471 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "text", + "content": "Latent diffusion models [36] are text-to-image diffusion models that operate in a compressed latent space. During forward diffusion process, it encodes an image sample " + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "text", + "content": " into a latent code " + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "text", + "content": " and corrupts " + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "text", + "content": " by Gaussian noise at each time step " + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 55, + 477, + 296, + 537 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 543, + 295, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 543, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 92, + 543, + 295, + 558 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "26d8d9ca78ae1b024ea881a69e247c6005b725c1069ef4d65d1aeb7a47660000.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": " denotes the noise strength at time step " + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": " denotes the identity matrix with the same dimensions as " + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": ". Given a text embedding " + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": ", the backward denoising process can be parameterized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 616, + 295, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 295, + 630 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {z} _ {t - 1} \\mid \\mathbf {z} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right), \\Sigma_ {\\theta} \\left(\\mathbf {z} _ {t}, \\mathbf {c}, t\\right)\\right), \\tag {2}", + "image_path": "e364b5ba5653aa2752362d9b8c15b4b2ea3e54c4bafc23a1bba1bc353f35d1df.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{z}_t,\\mathbf{c},t)" + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": " is output by a U-Net [37], and " + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}(\\mathbf{z}_t,\\mathbf{c},t)" + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": " is determined by a noise scheduler (e.g., DDPM [18] or DDIM [42]). The model learns to predict noise via a parameterized noise estimator " + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": " under the guidance of an objective function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 66, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\mathbb {E} _ {t \\sim [ 1, T ], \\mathbf {z} _ {0}, \\mathbf {c}, \\epsilon_ {t} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\| \\epsilon_ {t} - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, \\mathbf {c}, t) \\| _ {2} ^ {2}. \\tag {3}", + "image_path": "ce87747af1515ca5910bbd3b8cd18c1c8722e830db4a20fbfc06e6f8cbf28d84.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": "Our framework is built upon pretrained image-to-video diffusion models that incorporate temporal convolutions and temporal attention layers on top of latent diffusion models to establish temporal correlations between video frames. Given a video " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\in \\mathbb{R}^{N\\times 3\\times H\\times W}" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " is the number of frames and " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " indicates the image resolution, each frame is first encoded into a latent representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0\\in \\mathbb{R}^{N\\times C\\times h\\times w}" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": " via the VAE encoder " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": ". Both forward diffusion process and backward denoising process are then performed in the latent space. The generated video frames can be collectively obtained through the decoder " + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": ". To inject image conditions, we fuse and incorporate text-conditioned and image-conditioned features by projecting the conditional image into a text-aligned embedding space. Meanwhile, the conditional image is concatenated with the initial per-frame noise to preserve more visual details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 493, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 493, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 493, + 282 + ], + "type": "text", + "content": "3.2. Interpolation-based Initialization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 287, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 382 + ], + "type": "text", + "content": "In diffusion models, the initial Gaussian noise determines the coarse structure that emerges in the early denoising steps, while high-frequency details are refined later. Since existing image-to-video diffusion models [8, 50] randomly initialize the latent code " + }, + { + "bbox": [ + 313, + 287, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 313, + 287, + 555, + 382 + ], + "type": "text", + "content": " at inference time, each intermediate frame follows a distinct stochastic trajectory: colors, micro-textures, or even object pose shift slightly from frame to frame, yielding perceptual \"flicker\" [10]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 491 + ], + "type": "text", + "content": "To suppress this drift, we correlate the frame-wise latents by interpolating between the endpoint noises instead of resampling them independently. Prior work [6] indicates that linear interpolation often yields intermediate latent norms that are extremely unlikely under a unit Gaussian, whereas spherical linear interpolation (SLERP) preserves the Euclidean norm and keeps samples on-distribution. Therefore, we leverage SLERP between the latent noises of two input frames " + }, + { + "bbox": [ + 313, + 384, + 556, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t1}" + }, + { + "bbox": [ + 313, + 384, + 556, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 384, + 556, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{tN}" + }, + { + "bbox": [ + 313, + 384, + 556, + 491 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 495, + 555, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 495, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 324, + 495, + 555, + 521 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t n} = \\frac {\\sin ((1 - \\lambda_ {\\text {n o i s e}}) \\phi)}{\\sin \\phi} \\mathbf {z} _ {t 1} + \\frac {\\sin (\\lambda_ {\\text {n o i s e}} \\phi)}{\\sin \\phi} \\mathbf {z} _ {t N}, \\tag {4}", + "image_path": "bd6a23ed47e2bddc8fea0039cc0533537a2da605f1deb22ee12ccaf06e380aad.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\phi = \\arccos \\left(\\frac{\\mathbf{z}_{t1}^T\\cdot\\mathbf{z}_{tN}}{\\|\\mathbf{z}_{t1}\\|_{\\|}\\mathbf{z}_{tN}\\|}\\right)" + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{noise}}\\in [0,1]" + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "text", + "content": " denotes the parameter for latent interpolation. As shown in Figure 2, we concatenate " + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{tn}" + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t1}" + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{tN}" + }, + { + "bbox": [ + 313, + 528, + 556, + 653 + ], + "type": "text", + "content": " along the frame dimension and inject them as the initial latent noise for DDIM sampling. It is worth noting that we only inject the interpolated latent noises at early denoising steps to preserve the appearance and motion priors from the image-to-video backbone as much as possible. With such meticulously designed latent injection strategy, our approach addresses the issue of random and abrupt transition frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Unlike GANs [15], whose latent space is structured and semantically meaningful, diffusion models operate in an unstructured noise space that does not explicitly encode high-level semantics [56]. This lack of structure can lead to artifacts when interpolating between semantically different in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16983" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 70, + 523, + 373 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 523, + 373 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 523, + 373 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 523, + 373 + ], + "type": "image", + "image_path": "fa27ac34aa3c8e70ab2e6846d79d0da738d1e1b4cc8fa4d40bf787a792e30614.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": "Figure 2. Inference framework of VTG. Our interpolation-based initialization features three designs: " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " Interpolated Noise Injection, " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " LoRA Interpolation, and " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " Frame-aware Text Interpolation. VTG first converts two encoded input frames to latent noises via DDIM inversion. Then, it interpolates between the two latent noises and concatenate intermediate noises along the temporal dimension. To capture meaningful semantics and enable the transition between two conceptually different objects, we employ LoRA interpolation and text interpolation, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": "puts. To suppress such transition artifacts, we first train two LoRAs " + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{1}" + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{N}" + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": " given two input frames " + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "inline_equation", + "content": "x^{1}" + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "inline_equation", + "content": "x^{N}" + }, + { + "bbox": [ + 55, + 460, + 295, + 495 + ], + "type": "text", + "content": ", optimizing the following objective:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 502, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 502, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 61, + 502, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\Delta \\theta_ {t}\\right) = \\mathbb {E} _ {\\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta + \\Delta \\theta_ {i}} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0 n} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, \\mathbf {c} _ {n}\\right) \\| ^ {2} \\right], \\tag {5}", + "image_path": "93f33b0b5d170d96e2dd43836694c9bc145dd797bae3e065af273f266c724153.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0n}" + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": " is the encoded latent vector of frame " + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": " is the text embedding associated with transition caption " + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "inline_equation", + "content": "c^n" + }, + { + "bbox": [ + 55, + 521, + 296, + 568 + ], + "type": "text", + "content": ". Then, we linearly interpolate the two LoRAs to fuse the semantics of two input frames:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 574, + 295, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 574, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 100, + 574, + 295, + 588 + ], + "type": "interline_equation", + "content": "\\Delta \\theta = \\left(1 - \\lambda_ {l o r a}\\right) \\Delta \\theta_ {1} + \\lambda_ {l o r a} \\Delta \\theta_ {N}, \\tag {6}", + "image_path": "bb27f19e630c907e8855589a512bbc30a4fef16aa17e12cd2ddffddceb003da9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{lora}" + }, + { + "bbox": [ + 55, + 594, + 295, + 605 + ], + "type": "text", + "content": " denotes the parameter during lora interpolation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "text", + "content": "Typically, video diffusion models [8, 50] employ only one caption as the text condition. This hinders the transition generation between two conceptually different objects, causing abrupt content changes without generating intermediate frames with hybrid meanings. However, for imaged-to-video diffusion models, the text embedding of the entire frame sequence " + }, + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "text", + "content": " is integrated within the denoising U-Net via the cross-attention layer, without explicitly defining per-frame text embedding. A related still-image approach [52]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 495 + ], + "type": "text", + "content": "linearly blends the two text embeddings to morph between two static endpoints, whereas our method performs frame-aware SLERP across prompt sequences as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 505, + 555, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 505, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 328, + 505, + 555, + 530 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {\\lambda_ {t e x t}} = \\frac {\\sin ((1 - \\lambda_ {t e x t}) \\phi)}{\\sin \\phi} \\mathbf {c} _ {1} + \\frac {\\sin (\\lambda_ {t e x t} \\phi)}{\\sin \\phi} \\mathbf {c} _ {N}, \\tag {7}", + "image_path": "07986b6202a384a408eca81acc4e16d3e614270f0a016aafe01b2d3a208b7bba.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\lambda_{text} \\in [0,1]" + }, + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "text", + "content": " serves as the frame-aware coefficient to control the transition sequence; " + }, + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_1" + }, + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_N" + }, + { + "bbox": [ + 313, + 541, + 555, + 578 + ], + "type": "text", + "content": " denote the text embeddings of the first and last frames, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 587, + 486, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 486, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 486, + 599 + ], + "type": "text", + "content": "3.3. Bidirectional Motion Prediction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "content": "In our experiments, we observed notable quality differences when reversing the order of input frames. This discrepancy arises from two main factors: (1) the model shows a bias toward resembling the initial input frame more closely due to conditional image leakage [59]; (2) existing image-to-video diffusion models [8, 50] are pretrained exclusively for forward motion prediction, while real-world motion is naturally asymmetric, resulting in ambiguity when predicting reverse motion." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16984" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 527, + 300 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 527, + 300 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 527, + 300 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 527, + 300 + ], + "type": "image", + "image_path": "a844dcd3564d5a50c7a030960a63a525bbacd9a18de01bd94be556adfd4c9b48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 309, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 54, + 309, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 555, + 365 + ], + "type": "text", + "content": "Figure 3. Training framework of VTG. In Bidirectional Motion Prediction (BMP), the noisy latent is flipped along temporal dimension, and self-attention maps undergo a 180-degree rotation to establish reversed motion-time correlations. Two U-Nets separately predict forward and backward motion, with the backward noise reversed again to fuse with the forward noise, ensuring a consistent motion path for iterative denoising. In Representation Alignment Regularization (RAR), we spatially patchify each frame independently, then aggregate the per-patch alignment loss across temporal dimension." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": "Inspired by [46], we simultaneously predict forward and backward motions to mitigate the ambiguity issue. Specifically, we rotate the temporal self-attention maps by 180 degrees both horizontally and vertically, causing a reversal of attention relationships as illustrated in Figure 3. We also reverse the original forward trajectory of video " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\{x^{1}, x^{2}, \\ldots, x^{N}\\}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\{x^{N}, x^{N - 1}, \\ldots, x^{1}\\}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": ", given the two input frames " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "x^{1}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "x^{N}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": ". Then, we obtain the reversed noisy latent " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t'}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " by flipping " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " along the temporal dimension. After that, " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t'}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " is fed into the 3D U-Net for backward motion prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 520, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 581 + ], + "type": "text", + "content": "During training, we employ a lightweight fine-tuning strategy that only updates the parameters of value and output matrices (i.e., " + }, + { + "bbox": [ + 55, + 520, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\theta_{w,o}" + }, + { + "bbox": [ + 55, + 520, + 296, + 581 + ], + "type": "text", + "content": ") in the temporal attention layers. The objective " + }, + { + "bbox": [ + 55, + 520, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BMP}" + }, + { + "bbox": [ + 55, + 520, + 296, + 581 + ], + "type": "text", + "content": " is formulated by taking the L2 norm between the predicted noise and ground-truth reversed noise:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 597, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 597, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 82, + 597, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B M P} = \\left\\| \\operatorname {f l i p} \\left(\\epsilon_ {t}\\right) - \\epsilon_ {\\theta_ {w, o}} \\left(\\mathbf {z} _ {t ^ {\\prime}}, \\mathbf {c}, t, A _ {i, j} ^ {\\prime}\\right) \\right\\| _ {2} ^ {2}. \\tag {8}", + "image_path": "d7c137eb59e117487323679dae653b283e5cc2ca2717877be2a0548aeba2ad95.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 661 + ], + "type": "text", + "content": "The predicted backward noise is reversed again to fuse with forward noise. Specifically, we simply employ linear interpolation to ensure forward-backward consistency:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 674, + 295, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 674, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 96, + 674, + 295, + 689 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} = \\left(1 - \\lambda_ {B M P}\\right) \\epsilon_ {t, i} + \\lambda_ {B M P} \\epsilon_ {t, N - i} ^ {\\prime}, \\tag {9}", + "image_path": "232663267dc49b801d79dfc9a4e43f06d2605cf357f60f80872ce7e3c25565c1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 257, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 257, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 257, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 257, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{BMP} = 0.5" + }, + { + "bbox": [ + 55, + 701, + 257, + 714 + ], + "type": "text", + "content": " denotes the weighting factor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 385, + 533, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 533, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 533, + 399 + ], + "type": "text", + "content": "3.4. Representation Alignment Regularization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 408, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 517 + ], + "type": "text", + "content": "Though previous techniques effectively mitigate abrupt content changes and motion-reversal ambiguity, our approach remains susceptible to generating blurry and low-fidelity transitions, especially when input frames contain fine-grained textures (e.g., bicycle spokes or fabric weave). Recent work [53] shows that diffusion latents inherently lack high-frequency semantics compared to self-supervised representations like DINOv2's [32]. Such mismatch tends to accumulate over video frames and appears as blur." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "To overcome this, we propose to distill DINOv2 features back into the denoising trajectory of video diffusion. As illustrated in Figure 3, the video latent is first patched into " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " sequences of " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " tokens along the temporal dimension. For a N-frame video, the latent vector " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "-th frame has the shape of " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "C\\times h\\times w" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ". Given a pre-defined patch size " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ", the sequence length " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " should be: " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "T = (h / s)\\times (w / s)" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ". Then, we project the per-frame latent representation " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " of video diffusion via a multilayer perceptron (MLP) to align with DINOv2 representation " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{*} = \\mathcal{E}_{d}(\\mathbf{x}_{*})\\in \\mathbb{R}^{P\\times m}" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{*}" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " denotes the encoder output given a clean video frame " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{*}" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " respectively denotes the number of patches and the embedding dimension of DINOv2 encoder " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_d" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "y_{\\phi}(\\mathbf{h}_t)" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " be the projected representation of intermediate diffusion latent feature " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " denotes the trainable parameters of the MLP. The regularization term " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{RAR}" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " is computed by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16985" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": "aggregating patch-wise similarities across all video frames:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 102, + 295, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 295, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R A R}} = - \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {t, \\mathbf {x} _ {*}, \\epsilon_ {t}} \\left[ \\frac {1}{P} \\sum_ {p = 1} ^ {P} \\sin \\left(\\mathbf {y} _ {*} ^ {[ p ]}, y _ {\\phi} (\\mathbf {h} _ {t}) ^ {[ p ]}\\right) \\right]. \\tag {10}", + "image_path": "d3bcd1bea20a68fa4f176ecdea910350e5c21cd9bcb609b148b6563dc2c74499.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 128, + 295, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 152 + ], + "type": "text", + "content": "During inference, the external encoder " + }, + { + "bbox": [ + 55, + 128, + 295, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_d" + }, + { + "bbox": [ + 55, + 128, + 295, + 152 + ], + "type": "text", + "content": " and MLP projector " + }, + { + "bbox": [ + 55, + 128, + 295, + 152 + ], + "type": "inline_equation", + "content": "y_{\\phi}" + }, + { + "bbox": [ + 55, + 128, + 295, + 152 + ], + "type": "text", + "content": " are discarded." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 163, + 137, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 137, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 137, + 176 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 182, + 171, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 171, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 171, + 194 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "text", + "content": "Training Details. We leverage a small collection (i.e., 150) of high-quality in-house videos as our training data. These video samples encompass various motion (e.g., people running, airplanes taxiing) and appearance changes (e.g., a man transitioning from youth to old age). The learnable parameters of the backward U-Net are initialized to zero, while the other available weights are initialized from the pre-trained checkpoint of DynamiCrafter [50]. We utilize the AdamW optimizer with learning rate of " + }, + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "inline_equation", + "content": "1 \\times e^{-5}" + }, + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "text", + "content": ". VTG is trained on 4 NVIDIA A100-80GB GPUs, taking " + }, + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\sim 20K" + }, + { + "bbox": [ + 55, + 199, + 295, + 331 + ], + "type": "text", + "content": " iterations with batch size of 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "content": "Inference Details. In our experiments, we use Stable Diffusion v2.1-base [36] as our interpolation backbone. When training LoRAs, we set the LoRA rank to 16 and train them for 200 steps using AdamW optimizer [28] with a learning rate of " + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "content": ". This requires only " + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\sim 85" + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "content": "s with 1 NVIDIA A100-80GB GPU. During inversion and sampling, we use the DDIM sampler [42] and set total time steps to the default value (i.e., 50) of the selected image-to-video models. We apply text classifier-free guidance [17] for all models with the same negative prompt \"Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms\" across all transition generations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 571 + ], + "type": "text", + "content": "Comparison Methods. We compare VTG with following methods: DiffMorpher [56], TVG [57], SEINE [8], DynamiCrafter [50], Generative Inbetweening [46], Text Embedding Interpolation (TEI) [45], Denoising Interpolation (DI) [16], and Attention Interpolation (AID) [16]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Evaluation Benchmarks. We employ several public datasets and our self-curated TransitBench to comprehensively evaluate the quality of the generated transition videos. Specifically, we leverage MorphBench [56] and TC-Bench [13] to evaluate object morphing and motion prediction. Concept blending and scene transition are rarely explored in the realm of generative models. Hence, there is a lack of specific evaluation benchmarks for these two tasks. Previously, [16] employs CIFAR-10 [24] and LAION-Aesthetics [39] as their evaluation benchmarks to test concept blending. While there are sufficient number of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "samples in both datasets, CIFAR-10 only contains concepts from 10 classes, and LAION-Aesthetics consists entirely of synthetic images and lacks real-world samples. To this end, we present TransitBench, the first benchmark dataset for collectively assessing concept blending transitions of two distinct conceptual objects and scene transitions between two relevant scenarios. We collected 200 pairs of pictures (each pair forms the first and the last frames of one transition generation sample) of diverse content and styles, and evenly divide them into two categories: 1) concept-blending cases, and 2) scene-transition cases, both of which are obtained from web resources. We hope TransitBench can promote future studies on general transition generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 237, + 426, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 426, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 426, + 249 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 255, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 663 + ], + "type": "text", + "content": "To demonstrate the superiority of our method, we provide a visual comparison of VTG against five state-of-the-art baselines. Specifically, we compare our approach with DiffMorpher, Generative Inbetweening, TVG, SEINE, DynamamiCrafter across four transition tasks for uniformity. Figure 4 shows the generation results by different methods. For object morphing, the compared methods exhibit low-fidelity (e.g., DiffMorpher, DynamamiCrafter), oversaturation effect (e.g., TVG), and inaccurate semantics (e.g., Generative Inbetweening, SEINE). In contrast, VTG presents a natural transition, preserving the content of the first and last frames well. For motion prediction, Generative Inbetweening and SEINE generate inferior transitions as observed by their blurry intermediate frames. TVG yields unnatural transitions as it generates a front-facing image of a person, while the person's face is actually turned away. DiffMorpher shows low-fidelity transitions, and its generated video looks like a sequence of intermittent images without concerning the consistency between background and foreground objects. This is largely due to the missing temporal modeling in DiffMorpher's image diffusion backbone. VTG achieves comparable or even better results with DynamamiCrafter, validating that it can effectively preserve the appearance and motion priors in diffusion models. For concept blending, most video diffusion baselines present abrupt content changes, while other baselines show blurry (e.g., DiffMorpher) and ambiguous (e.g., TVG) intermediate frames. In contrast, VTG yields smooth transition between the lion and the truck with rational intermediate results (e.g., a lion-colored and lion-sized truck). For scene transition between two relevant scene images, existing methods suffer from abrupt content changes or low similarity (e.g., TVG), while VTG produces natural scene-level transition with pleasing effect." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 671, + 433, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 433, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 433, + 684 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. The objective of versatile transition generation is to generate a sequence of " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "(N - 2)" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " intermedi-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16986" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 541, + 582 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 541, + 582 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 541, + 582 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 541, + 582 + ], + "type": "image", + "image_path": "032d98d612a2dbe24edff782b7194d13dc08b915212ef9ff448676a217916a34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 590, + 555, + 624 + ], + "lines": [ + { + "bbox": [ + 55, + 590, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 555, + 624 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on four transition tasks. For each task, the displayed images correspond to the 1st, 4th, 8th, 12th, and 16th video frames, respectively. Please refer to our project page for more video examples, which offers a more straightforward and comprehensive qualitative comparisons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "text", + "content": "ate video frames, denoted as " + }, + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "inline_equation", + "content": "x^{2:N - 1}" + }, + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "text", + "content": ", between two source frames " + }, + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "inline_equation", + "content": "x^1" + }, + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "inline_equation", + "content": "x^N" + }, + { + "bbox": [ + 55, + 644, + 297, + 706 + ], + "type": "text", + "content": ", with two captions describing the two frames. High-quality transition videos should simultaneously satisfy four criteria: semantic fidelity to the input endpoints, visual quality, temporal coherence, and perceptual" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "text", + "content": "smoothness. To ensure fair comparisons, we performed extensive quantitative evaluations using experimental settings consistent with those prior methods as in their original paper. Specifically, we adopt the following evaluation metrics: (1) FID " + }, + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "text", + "content": " and PPL " + }, + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 313, + 644, + 556, + 706 + ], + "type": "text", + "content": " follow DiffMorpher [56]. FID com" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16987" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "pares the first and last frames against the generated transition frames in the Inception feature space; PPL measures the average LPIPS fluctuation along an interpolation path, where lower values imply fewer perceptual jumps. (2) TCR " + }, + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": " and TC-Score " + }, + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": " are taken from the training-free TVG framework [57]. TCR is the fraction of video frames whose CLIP embedding stays within a pre-set similarity margin to either endpoint; TC-Score complements it with a mean cosine-similarity, offering a softer view of frame-level consistency. (3) Smoothness " + }, + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": " is adapted from AID [16]. For every adjacent frame pair, we compute LPIPS and the Gini coefficient of these distances, reflecting how even the step-to-step changes are. We report its inverse so that a higher score denotes smoother transitions." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 246, + 293, + 337 + ], + "blocks": [ + { + "bbox": [ + 58, + 246, + 293, + 337 + ], + "lines": [ + { + "bbox": [ + 58, + 246, + 293, + 337 + ], + "spans": [ + { + "bbox": [ + 58, + 246, + 293, + 337 + ], + "type": "table", + "html": "
MethodMetamorphosisAnimation
FID (↓)PPL (↓)FID (↓)PPL (↓)
DiffMorpher [56]70.4918.1943.155.14
TVG [57]86.9235.1842.9912.46
SEINE [8]82.0347.7248.2516.26
DynamiCrafter [50]87.3242.0943.3111.16
VTG (Ours)67.3922.8039.165.14
", + "image_path": "9b4fedb4c59ef81f938cf54aa5997fc2409d5cb02e84f85ac308cc0b7ea07985.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 386, + 294, + 444 + ], + "blocks": [ + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "type": "text", + "content": "Table 1. Quantitative results on MorphBench. The best results are in bold; second-best are underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 386, + 294, + 444 + ], + "lines": [ + { + "bbox": [ + 58, + 386, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 58, + 386, + 294, + 444 + ], + "type": "table", + "html": "
MethodAttributeObjectBackground
TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)TCR (↑)TC-Score (↑)
DiffMorpher [56]41.820.84419.570.76550.000.819
SEINE [8]17.860.72010.480.6547.960.742
DynamiCrafter [50]16.550.74513.910.70725.560.795
TVG [57]41.820.87730.440.82238.890.864
VTG (Ours)42.780.89333.460.84950.000.883
", + "image_path": "c59634264d565be961d00caa04862aacc67c832bfc54a90be196b1fbdbec778d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 495, + 295, + 538 + ], + "blocks": [ + { + "bbox": [ + 55, + 452, + 295, + 474 + ], + "lines": [ + { + "bbox": [ + 55, + 452, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 295, + 474 + ], + "type": "text", + "content": "Table 2. Quantitative results on TC-Bench. The best results are in bold. Best viewed when zoomed in." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 495, + 295, + 538 + ], + "lines": [ + { + "bbox": [ + 58, + 495, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 58, + 495, + 295, + 538 + ], + "type": "table", + "html": "
DatasetTEIDIAID-OAID-IOurs
CIFAR-100.75310.75640.78310.78610.7932
LAION-Aesthetics0.74240.75110.76430.81520.8215
", + "image_path": "4d8fbd4fcd775d3629797f705006cfa441225a63d0b86792e5b9e2a41179dfcd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 295, + 568 + ], + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 568 + ], + "type": "text", + "content": "Table 3. Smoothness (↑) evaluation for Concept Blending. The best results are in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "Quantitative Results. As seen in Table 1 and Table 2, VTG outperforms existing methods with their original experimental setting, demonstrating its capability to generate high-quality, temporally coherent, and semantically consistent video transitions. It is worth noting that the only exception is Metamorphosis subset, where VTG achieves the second-best PPL (22.80). Nevertheless, VTG still demonstrates outstanding overall performance, achieving state-of-the-art FID and comparable PPL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "According to AID [16], a well-interpolated concept-blending sequence should exhibit a gradual and smooth" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "transition. To offer a more comprehensive evaluation, we conducted experiments using the same benchmarks and metric as in AID. Please refer to Table 3 for details." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 128, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 316, + 128, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 266 + ], + "type": "image", + "image_path": "71e79c1636d39dd458cf4f41f926f41fccb8d006114a8eb4e6109c7f72c019f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "type": "text", + "content": "Figure 5. The bar plot of user preference rates " + }, + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 274, + 555, + 297 + ], + "type": "text", + "content": " of different methods on four transition tasks. Best viewed when zoomed in." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 556, + 515 + ], + "type": "text", + "content": "User Study. To subjectively evaluate our approach, we further conduct a user study. We asked Amazon Mechanical Turk (AMT) workers to choose the best generated transition video from a set of five candidates. Given the two input frames and corresponding transition captions, AMT workers evaluate the videos based on three questions: 1) Which video is the most semantic-coherent? 2) Which video is the most temporally coherent? 3) Which video has the highest fidelity? To make a fair evaluation, we include static and obviously incorrect videos to identify random clicking. Each test example is assigned to 10 different workers for evaluation. We collected 150 valid responses for each transition task after filtering out those responses from random clicking workers. The results are shown in Figure 5, which illustrates that VTG obtains the highest preference rate with its superior quality of the generated transition videos compared to other baselines." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we formulate the problem of versatile transition generation that encompasses four representative transition tasks, namely, object morphing, motion prediction, concept blending, and scene transition. Leveraging a pretrained image-to-video diffusion model, we design VTG, a unified Versatile Transition video Generator that can tackle the four tasks within a single framework. Both qualitative and quantitative experiments show that VTG achieves superior performance in generating semantically relevant, temporally coherent, and visually pleasing transition video frames, given a pair of text prompts and two input frames. With these advantages, VTG can be readily applied to real-world content creation scenarios, offering an efficient tool for generating high-quality transition in media production." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16988" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 128 + ], + "type": "text", + "content": "This study is funded by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 115, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 115, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 115, + 155 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 162, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 162, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 296, + 195 + ], + "type": "text", + "content": "[1] Alyaa Qusay Aloraibi. Image morphing techniques: A review. Technium: Romanian Journal of Applied Sciences and Technology, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 197, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 197, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 61, + 197, + 296, + 229 + ], + "type": "text", + "content": "[2] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. Seminal Graphics Papers: Pushing the Boundaries, 1992. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 232, + 296, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 296, + 265 + ], + "type": "text", + "content": "[3] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 266, + 296, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 296, + 322 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 323, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 296, + 367 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "text", + "content": "[6] Erik Bodin, Alexandru I Stere, Dragos D Margineantu, Carl Henrik Ek, and Henry Moss. Linear combinations of latents in generative models: subspaces and beyond. In ICLR, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 459 + ], + "type": "text", + "content": "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 461, + 296, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 461, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 62, + 461, + 296, + 516 + ], + "type": "text", + "content": "[8] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 518, + 296, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 518, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 62, + 518, + 296, + 561 + ], + "type": "text", + "content": "[9] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. TOG, 2012. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 564, + 296, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 564, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 57, + 564, + 296, + 608 + ], + "type": "text", + "content": "[10] Yitong Deng, Winnie Lin, Lingxiao Li, Dmitriy Smirnov, Ryan Burgert, Ning Yu, Vincent Dedun, and Mohammad H Taghavi. Infinite-resolution integral noise warping for diffusion models. In ICLR, 2025. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 610, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 610, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 610, + 296, + 632 + ], + "type": "text", + "content": "[11] Jiong Dong, Kaoru Ota, and Mianxiong Dong. Video frame interpolation: A comprehensive survey. TOMM, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "text", + "content": "[12] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In ECCV, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[13] Weixi Feng, Jiachen Li, Michael Saxon, Tsu-jui Fu, Wenhu Chen, and William Yang Wang. Tc-bench: Benchmarking temporal compositionality in text-to-video and image-to-video generation. In ACL, 2025. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[14] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 161 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 164, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 554, + 196 + ], + "type": "text", + "content": "[16] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. In NeurIPS, 2024. 2, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 220 + ], + "type": "text", + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshop, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 222, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 243 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 246, + 554, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 246, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 246, + 554, + 300 + ], + "type": "text", + "content": "[19] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 302, + 553, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 302, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 316, + 302, + 553, + 345 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 347, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 347, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 347, + 553, + 380 + ], + "type": "text", + "content": "[21] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In ECCV, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "text", + "content": "[22] Siddhant Jain, Daniel Watson, Eric Tabellion, Aleksander Hólynski, Ben Poole, and Janne Kontkanen. Video interpolation with diffusion models. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 417, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 460 + ], + "type": "text", + "content": "[23] Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 462, + 553, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 494 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Toronto, ON, Canada, 2009. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 496, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 553, + 540 + ], + "type": "text", + "content": "[25] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 543, + 553, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 543, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 317, + 543, + 553, + 585 + ], + "type": "text", + "content": "[26] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. TOG, 2014. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 588, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 620 + ], + "type": "text", + "content": "[27] Chunxu Liu, Guozhen Zhang, Rui Zhao, and Limin Wang. Sparse global matching for video frame interpolation with large motion. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 623, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 655 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2018. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "text", + "content": "[29] Michael I. Miller and Laurent Younes. Group actions, homeomorphisms, and matching: A general framework. *IJCV*, 2001. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16989" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "text", + "content": "[31] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[32] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "text", + "content": "[33] Marko Rajković. Geodesics, Splines, and Embeddings in Spaces of Images. PhD thesis, Universitäts-und Landesbibliothek Bonn, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "text", + "content": "[35] Fitsum Reda, Janne Kontkanen, Eric Tabellion, Deqing Sun, Caroline Pantofaru, and Brian Curless. Film: Frame interpolation for large motion. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 327 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "text", + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 493 + ], + "type": "text", + "content": "[40] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In CVPR, 2010. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[41] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In ICLR, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 559 + ], + "type": "text", + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. *ICLR*, 2021. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "text", + "content": "[44] Shuyuan Tu, Qi Dai, Zhi-Qi Cheng, Han Hu, Xintong Han, Zuxuan Wu, and Yu-Gang Jiang. Motioneditor: Editing video motion via content-aware diffusion. In CVPR, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "text", + "content": "[45] Clinton Wang and Polina Golland. Interpolating between images with diffusion models. In ICMLW, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[46] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. In ICLR, 2025. 3, 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 680 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 117 + ], + "type": "text", + "content": "[47] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 554, + 162 + ], + "type": "text", + "content": "[48] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 163, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 554, + 185 + ], + "type": "text", + "content": "[49] George Wolberg. Image morphing: a survey. The visual computer, 1998. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 185, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 185, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 315, + 185, + 554, + 229 + ], + "type": "text", + "content": "[50] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. In ECCV, 2024. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 230, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 554, + 262 + ], + "type": "text", + "content": "[51] Cheng Yang, Lijing Liang, and Zhixun Su. Realworld denoising via diffusion model. arXiv preprint arXiv:2305.04457, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 554, + 308 + ], + "type": "text", + "content": "[52] Zhaoyuan Yang, Zhengyang Yu, Zhiwei Xu, Jaskirat Singh, Jing Zhang, Dylan Campbell, Peter Tu, and Richard Hartley. Impus: Image morphing with perceptually-uniform sampling using diffusion models. In ICLR, 2024. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 309, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 309, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 309, + 554, + 354 + ], + "type": "text", + "content": "[53] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. In ICLR, 2025. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 354, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 354, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 554, + 388 + ], + "type": "text", + "content": "[54] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 388, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 554, + 431 + ], + "type": "text", + "content": "[55] Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and appearance via inter-frame attention for efficient video frame interpolation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 432, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 432, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 315, + 432, + 554, + 466 + ], + "type": "text", + "content": "[56] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Xingang Pan, and Bo Dai. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In CVPR, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 467, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 467, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 554, + 510 + ], + "type": "text", + "content": "[57] Rui Zhang, Yaosen Chen, Yuegen Liu, Wei Wang, Xuming Wen, and Hongxia Wang. TVg: A training-free transition video generation method with diffusion models. TCVT, 2024. 3, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 511, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 511, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 315, + 511, + 554, + 566 + ], + "type": "text", + "content": "[58] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 567, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 567, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 315, + 567, + 554, + 611 + ], + "type": "text", + "content": "[59] Min Zhao, Hongzhou Zhu, Chengdong Xiang, Kaiwen Zheng, Chongxuan Li, and Jun Zhu. Identifying and solving conditional image leakage in image-to-video diffusion model. In NeurIPS, 2024. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 611, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 611, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 315, + 611, + 554, + 656 + ], + "type": "text", + "content": "[60] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 657, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 657, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 315, + 657, + 554, + 680 + ], + "type": "text", + "content": "[61] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. IJAEMS, 2017. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16990" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_content_list.json b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9f29e894403a0a44ba6acdd92e0bcbd8628199c6 --- /dev/null +++ b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_content_list.json @@ -0,0 +1,1618 @@ +[ + { + "type": "text", + "text": "VertexRegen: Mesh Generation with Continuous Level of Detail", + "text_level": 1, + "bbox": [ + 176, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiang Zhang $^{1*}$ Yawar Siddiqui $^{2}$ Armen Avetisyan $^{2}$ Chris Xie $^{2}$ Jakob Engel $^{2}$ Henry Howard-Jenkins $^{2}$ $^{1}$ UC San Diego $^{2}$ Meta Reality Labs Research", + "bbox": [ + 220, + 178, + 774, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0420cc4dd6766a5a06076fe8de70bc226cfe6c1a1d1e21f181e313f6c11f6e2.jpg", + "image_caption": [ + "Figure 1. Mesh generation process of VertexRegen (coarse-to-fine process) vs. previous work (partial-to-complete process). Prior work generates meshes face by face, with the step count corresponding to the face count in the figure. In contrast, VertexRegen produces meshes with a continuously increasing level of detail, where each step yields a valid mesh $\\mathcal{M}_k$ ." + ], + "image_footnote": [], + "bbox": [ + 99, + 250, + 890, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 551, + 328, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce VertexRegen, a novel mesh generation framework that enables generation at a continuous level of detail. Existing autoregressive methods generate meshes in a partial-to-complete manner and thus intermediate steps of generation represent incomplete structures. VertexRegen takes inspiration from progressive meshes and reformulates the process as the reversal of edge collapse, i.e. vertex split, learned through a generative model. Experimental results demonstrate that VertexRegen produces meshes of comparable quality to state-of-the-art methods while uniquely offering anytime generation with the flexibility to halt at any step to yield valid meshes with varying levels of detail.", + "bbox": [ + 88, + 583, + 485, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 792, + 223, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meshes are essential for 3D asset representation and are widely used in industries such as film, design, and gaming due to their compatibility with most 3D software and", + "bbox": [ + 89, + 819, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hardware. However, manually creating high-quality meshes is costly and time-consuming, prompting extensive research into automated 3D content creation. Largely, these approaches have used alternative representations such as neural fields [11, 27, 31, 38], voxels [34, 51] or point clouds [54, 59] which are later post-processed to meshes instead of direct mesh modeling. Unfortunately, these post-processed meshes often exhibit poor topology, overtessellation, and artifacts, lacking the quality of artist-crafted meshes.", + "bbox": [ + 511, + 553, + 906, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, there has been a surge of approaches that directly generate meshes, using autoregressive models to represent meshes as sequences of triangles [2-4, 15, 37, 43]. These methods capture the high-fidelity and aesthetic qualities of artist-created works without the need for post-generation conversion. Although significant progress has been made in improving tokenization schemes [2, 4, 37] and network architectures [15], these approaches do not fundamentally alter a key characteristic of generation: namely, to produce a valid mesh, the full sequence must be generated to completion. Consequently, these methods offer no control over the level of detail during generation; early stopping results in a mesh with missing faces. Simple extensions,", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work conducted while the author was an intern at Meta. \nProject page: https://vertexregen.github.io", + "bbox": [ + 114, + 875, + 423, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12570", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as face-count conditioning tokens in EdgeRunner [43], have been proposed to introduce some coarse control over the detail by pre-specifying a target face count. However, each generation still must be completed in its entirety to yield a complete mesh, and thus a single generation still offers only a single level of detail.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To allow for a continuous level of detail generation, we take inspiration from Hoppe's progressive mesh formulation [17]. Progressive meshes use two reversible operations to transition between levels of detail. The edge collapse operation simplifies a mesh into a coarser one by reducing one edge at a time. The vertex split operation reverses this simplification to add more detail by using information stored during edge collapse. By starting with a coarse mesh and keeping edge collapse records, this approach creates an efficient, lossless representation of the original mesh, allowing for continuous resolution adjustments by applying any number of vertex split operations.", + "bbox": [ + 89, + 181, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we build on this progressive mesh representation by learning the vertex split, i.e. reversing the edge collapse operation, as a generative problem. This allows the resulting mesh generation to inherit the properties of progressive meshes and provides a solution for anytime mesh generation, where the process can be stopped early to yield a coarser mesh rather than an incomplete one. By properly serializing the vertex split sequence, the entire generation process can be modeled using a Transformer trained with a next-token prediction objective, a widely adopted paradigm in mesh generation. We evaluate VertexRegen on the task of unconditional mesh generation, demonstrating superior results both qualitatively and quantitatively. We further provide examples of shape-conditioned generation. These experimental results illustrate the ability of our method to generate compelling meshes at a continuous resolution.", + "bbox": [ + 89, + 363, + 482, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 604, + 434, + 618 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Inspired by Hoppe's progressive meshes [17], we reframe mesh generation as the reversal of edge collapse operations, i.e. generating vertex splits.", + "- We formulate a token-efficient parameterization of a progressive mesh, through a half-edge data structure.", + "- We propose VertexRegen for continuous level of detail mesh generation. VertexRegen creates meshes in a coarse-to-fine fashion, rather than partial-to-complete, uniquely providing a solution for anytime generation." + ], + "bbox": [ + 91, + 621, + 482, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 768, + 232, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Mesh Generation. Recent advancements in 3D shape generation have explored various representations, including point-clouds [30, 44, 52, 59], signed distance functions (SDFs) [7, 19, 22, 23, 36, 53, 57], neural radiance fields (NeRFs) [16, 18, 46, 55], and Gaussian splatting [14, 42]. These implicit representations require iso-surface extraction techniques [24, 35, 47] to output meshes, often resulting", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in over-tessellated and excessively smooth outputs, which pose challenges for downstream applications such as geometric processing and manipulation.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, direct mesh generation inherently produces structured, well-defined geometry without the need for postprocessing or surface extraction, making it an increasingly prominent approach in recent years. Early methods tackle this task by generating meshes from surface patches [13], deforming ellipsoids [45], predicting mesh graphs [8], or employing binary space partitioning [6]. More recent techniques leverage generative models, particularly diffusion models and sequence-based approaches. PolyDiff [1] applies discrete diffusion, while PolyGen [29] autoregressively predicts vertices and faces using two separate networks. The sequence modeling paradigm has been further refined by representing the entire mesh as a single sequence. MeshGPT [37] introduces a tokenization scheme based on a vector quantized variational auto-encoder (VQVAE), while MeshXL [2] directly models discretized triangle soup sequences without compression. MeshAnything [4] extends MeshGPT by incorporating a point-cloud encoder for shape-conditioned generation. PivotMesh [49] introduces a hierarchical approach, generating pivot vertices before producing the full mesh. Further research has focused on optimizing tokenization and sequence modeling. Adjacency-aware compression techniques [3, 43, 50] improve tokenization efficiency and enable higher face counts within fixed context windows, while Meshtron [15] leverages Hourglass Transformers and sliding window attention to scale MeshXL sequences more effectively.", + "bbox": [ + 511, + 137, + 903, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, most existing approaches follow a partial-to-complete paradigm, where mesh regions are constructed sequentially at a uniform level of detail. Our work instead adopts a coarse-to-fine approach, continuously increasing the level of detail as new tokens are generated, achieving better control over geometric complexity.", + "bbox": [ + 511, + 545, + 903, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Level of Detail Representations. Level of Detail [26] (LOD) is a widely used technique in computer graphics to optimize rendering performance by reducing the complexity of 3D models based on their size, distance from the camera, or importance in a scene. Different LOD strategies depend on the underlying representation of 3D shapes. For meshes, progressive meshes [17] generate a mesh sequence starting from a coarse base model, gradually refining it through a series of transformations that incrementally add detail. Traditional mesh simplification methods [12, 20] can be adopted to construct such sequences by iteratively reducing polygon count while preserving geometric fidelity. Progressive simplicial complexes [32] extend progressive meshes to handle arbitrary meshes, including non-manifold and non-orientable surfaces.", + "bbox": [ + 511, + 657, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies have incorporated neural networks for", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6d5a953d4abc6248f36699478fa034c38125043fcbb81d4ce61d00df28cee9f2.jpg", + "image_caption": [ + "Figure 2. Illustration of edge collapse and its inverse operation, i.e. vertex split. During edge collapse, vertex $v_{t}$ is collapsed into $v_{s}$ , resulting in two degenerate triangles (shaded in yellow), which forms two new edges in the result mesh." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 478, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LOD representations. Neural progressive meshes [5] propose a learned approach supporting LOD with a subdivision-based encoder-decoder. Similar representations have been explored for reconstruction approaches with signed distance functions (SDFs) [39-41] and neural fields [28]. However, these approaches are not generative; they typically encode and decode existing meshes [5] or optimize the LOD representation per shape [39]. In contrast, our approach is purely generative. VertexRegen learns to create meshes from scratch, progressively adding detail through autoregressive sequence generation. This allows us to generate new, high-fidelity meshes without relying on pre-existing structures, distinguishing our method from both traditional and neural LOD techniques.", + "bbox": [ + 89, + 263, + 483, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 489, + 181, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Progressive Meshes Overview", + "text_level": 1, + "bbox": [ + 89, + 513, + 354, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A progressive mesh [17] (PM) proposes an efficient and lossless continuous-resolution representation for arbitrary triangle meshes. The representation is built off two observations: (i) that a single mesh transformation, edge collapse, is sufficient for effective simplification of meshes; (ii) that edge collapse transformations are invertible via a vertex split operation. In the following, we provide details of the edge collapse operation and its inversion with vertex splits, each illustrated in Fig. 2.", + "bbox": [ + 89, + 537, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The edge collapse operation, $\\operatorname{ecol}(v_s, v_t)$ , unifies two adjacent vertices $v_s$ and $v_t$ into a single vertex $v_s$ . The operation results in the vanishing of two faces, $\\{v_t, v_s, v_l\\}$ and $\\{v_s, v_t, v_r\\}$ , as well as the vertex $v_t$ . In the general case, a new position for $v_s$ is also specified; however, for the case of half-edge collapse, the original position for $v_s$ is kept.", + "bbox": [ + 89, + 672, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An initial mesh, $\\mathcal{M}$ , can be simplified into a more coarse mesh, $\\mathcal{M}_n$ , by $n$ successive edge collapse operations:", + "bbox": [ + 89, + 763, + 482, + 795 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {0} \\xleftarrow {\\mathrm {e c o l} _ {0}} \\mathcal {M} _ {1} \\xleftarrow {\\mathrm {e c o l} _ {1}} \\dots \\xleftarrow {\\mathrm {e c o l} _ {n - 1}} \\mathcal {M}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 806, + 433, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The order in which the edge collapse operations are performed is determined such that each successive edge collapse results in the minimum increase in Quadratic Error Metrics (QEM) [12] with respect to the original mesh. This", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "process can be repeated until there are no candidate edge collapses, e.g. without flipping face normals.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A vertex split, $\\mathrm{vsplit}(v_s,v_l,v_r,v_t)$ , defines the inverse of edge collapse. It restores $v_{t}$ and the two faces, $\\{v_{t},v_{s},v_{l}\\}$ and $\\{v_{s},v_{t},v_{r}\\}$ , which vanished during edge collapse, as well as $v_{s}$ to its original position. In case $(v_{s},v_{t})$ is a boundary edge, either $v_{l}$ or $v_{r}$ will not exist, and there is only one face restored during the vertex split.", + "bbox": [ + 511, + 121, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The combination of edge collapse and its inverse vertex splits enables the progressive mesh representation, $\\mathrm{PM}(\\mathcal{M}) = (\\mathcal{M}_0,\\{\\mathrm{vsplit}_0,\\dots ,\\mathrm{vsplit}_{n - 1}\\})$ to express an arbitrary triangle mesh, $\\mathcal{M}$ , as a combination of a coarse mesh, $\\mathcal{M}_0$ , obtained through edge collapse, and the sequence of $n$ vertex split records required to reverse them:", + "bbox": [ + 511, + 212, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {0} \\xrightarrow {\\mathrm {v s p l i t} _ {0}} \\mathcal {M} _ {1} \\xrightarrow {\\mathrm {v s p l i t} _ {1}} \\dots \\xrightarrow {\\mathrm {v s p l i t} _ {n - 1}} \\mathcal {M}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 315, + 870, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. VertexRegen", + "text_level": 1, + "bbox": [ + 511, + 349, + 648, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional progressive meshes require starting an initial detailed mesh to be simplified into a coarse mesh, $\\mathcal{M}_0$ , and recording each of these collapse steps as vertex split records to form the representation. We observe that the formation of the PM representation serves effectively as forward (edge collapse) and reverse (vertex split) processes that enable transition between detailed and coarse meshes.", + "bbox": [ + 511, + 371, + 905, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VertexRegen frames the creation of a detailed mesh as the generation of a PM representation. Analogous to denoising for diffusion models, we train a generative model to reverse edge collapse. Concretely, VertexRegen first generates a coarse mesh, $\\mathcal{M}_0$ , from scratch, before increasing the level of detail through the generation of vertex split records.", + "bbox": [ + 511, + 477, + 905, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Progressive Mesh Parameterization", + "text_level": 1, + "bbox": [ + 511, + 575, + 805, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We frame generation as an autoregressive sequential modeling task, with the full mesh sequence taking the form:", + "bbox": [ + 511, + 595, + 903, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c c} \\text {M : [ < b o s > , [ M _ {-} 0 s e q u e n c e ] , < s e p > ,} & \\# \\mathcal {M} _ {0} \\\\ \\text {[ v s p l i t \\_ 0 ] , \\ldots , [ v s p l i t \\_ n - 1 ] , < e o s > ]} & \\# \\text {v s p l i t s} \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 636, + 903, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we will define the formation of the subsequences [M_0 sequence] and [vsplit].", + "bbox": [ + 511, + 678, + 903, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{M}_0$ : Coarse Mesh Tokenization. For the initial coarse mesh, we follow the tokenization scheme defined in MeshXL [2]. In this formulation, embeddings are learned for discretized coordinates in an $N^3$ grid. A vertex is represented via the sequential look-up of $x$ -value, $y$ -value, and $z$ -value. A face is then constructed as the concatenation of its 3 vertices, totaling 9 tokens. The full mesh sequence is then defined as the concatenation of its constituent faces.", + "bbox": [ + 511, + 726, + 905, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv: \\left[ < x >, < y >, < z > \\right], \\quad \\# v e r t e x\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 845, + 859, + 861 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF: \\left[ \\begin{array}{l} [ v _ {- 1} ], [ v _ {- 2} ], [ v _ {- 3} ] \\end{array} \\right], \\quad \\# \\text {f a c e}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 864, + 844, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM: \\left[ [ F _ {- 1} ], [ F _ {- 2} ], \\dots , [ F _ {- N} ] \\right] \\quad \\# \\text {m e s h}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 881, + 851, + 896 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12572", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ef934b6144eba8d2af06d3526da99653091dac78c54247047ff6a896829d209.jpg", + "image_caption": [ + "Figure 3. Face count distribution of the coarsest mesh $\\mathcal{M}_0$ and original mesh $\\mathcal{M}$ . The coarsest meshes $(\\mathcal{M}_0)$ contain significantly fewer faces than original $(\\mathcal{M})$ , with an average of 18 and 457 faces, respectively." + ], + "image_footnote": [], + "bbox": [ + 135, + 90, + 441, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [37], we sort vertices in $z - y - x$ order (lower to higher). Within each face, vertices are cylindrically permuted to have the lowest-indexed vertex be first.", + "bbox": [ + 89, + 329, + 483, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although MeshXL [2] leverages this tokenization scheme to produce full detailed meshes, in VertexRegen only the coarsest level of the mesh is parameterized in this way. These coarse meshes consist of substantially fewer faces than the full detailed mesh, as demonstrated in Fig. 3, and the resulting coarse mesh only accounts for $5.68\\%$ of the total sequence length on average.", + "bbox": [ + 89, + 375, + 483, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vertex Split Generation. Consider the $k$ -th vertex split operation $\\mathrm{vsplit}_k$ , which converts a lower-detailed mesh $\\mathcal{M}_k$ into a higher-detailed mesh $\\mathcal{M}_{k+1}$ . We denote the neighbor vertices of vertex $v$ in $\\mathcal{M}_k$ as $N_k(v)$ . After determining the target vertex $v_s$ to split (in $\\mathcal{M}_k$ ) and the new vertex $v_t$ (in $\\mathcal{M}_{k+1}$ ), we need to obtain $N_{k+1}(v_s)$ and $N_{k+1}(v_t)$ from the vertex split record to connect $v_s$ and $v_t$ with correct neighbor vertices in $\\mathcal{M}_{k+1}$ , which are prohibitive to generate as the number of neighbor vertices $|N_{k+1}(v_s)|$ and $|N_{k+1}(v_t)|$ is often large.", + "bbox": [ + 89, + 500, + 483, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, as mesh $\\mathcal{M}_k$ is generated from $\\mathcal{M}_{k + 1}$ during edge collapse by merging $v_{t}$ into $v_{s}$ , we have", + "bbox": [ + 89, + 652, + 483, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN _ {k + 1} \\left(v _ {s}\\right) \\cap N _ {k + 1} \\left(v _ {t}\\right) = \\left\\{v _ {l}, v _ {r} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 695, + 482, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN _ {k + 1} \\left(v _ {s}\\right) \\cup N _ {k + 1} \\left(v _ {t}\\right) - \\left\\{v _ {s}, v _ {t} \\right\\} = N _ {k} \\left(v _ {s}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 724, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which indicates we will only need to record vertex $v_{l}$ and $v_{r}$ . They split the vertex ring surrounding $v_{s}$ in $\\mathcal{M}_k$ , where two halves of the vertices on the ring (top and bottom vertices in Fig. 2) belong to $v_{t}$ and $v_{s}$ respectively in $\\mathcal{M}_{k + 1}$ . However, an ambiguity arises in determining which half of the ring corresponds to $v_{t}$ 's neighbors in $\\mathcal{M}_{k + 1}$ . In the following, we demonstrate how the half-edge data structure [48] can be utilized to resolve this ambiguity.", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denote $\\mathcal{H}_{ij}^k$ as a half-edge pointing from $v_{i}$ to $v_{j}$ , with $v_{k}$ being the third vertex in the associated face. As shown in", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d020ef834e9f52cbf4765977de5065d6732f60ef43375c9c10b203dce4fe5895.jpg", + "image_caption": [ + "Figure 4. Illustration of half-edge data structure and the traversal process to determine the neighbors of $v_{s}$ and $v_{t}$ in $\\mathcal{M}_{k + 1}$ . Starting from half-edge $\\mathcal{H}_{ls}^{1}$ , we traverse the faces in clockwise direction until $\\mathcal{H}_{sr}^{2}$ , where we collect the vertices $v_{1}$ and $v_{2}$ into $N_{k + 1}(v_s)$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 92, + 901, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/819156d72a3c6e7f194cbad6db72f5450cd1b639e5afff7774fb590878e99a53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 320, + 663, + 406 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39b44188626f66fd98150f202dd5e908f6498201f422e5fca196bb4b331cd831.jpg", + "image_caption": [ + "Figure 5. Illustration of VertexRegen tokenization. The sequence begins with base mesh $\\mathcal{M}_0$ , followed by vertex split subsequences. A special token indicates either $v_{l}$ or $v_{r}$ does not exist." + ], + "image_footnote": [], + "bbox": [ + 666, + 320, + 903, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 4, we begin the traversal from $\\mathcal{H}_{ls}^{1}$ and proceed to the twin of the next half-edge, $\\mathcal{H}_{1s}^{2}$ . This operation is repeated until we reach $\\mathcal{H}_{,s}^{r}$ . Throughout this process, we traverse the faces above $(v_{l}, v_{s})$ and $(v_{s}, v_{r})$ clockwise, obtaining:", + "bbox": [ + 511, + 458, + 905, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{v _ {k} \\mid \\mathcal {H} _ {\\cdot s} ^ {k}, k \\neq r \\right\\} = N _ {k + 1} \\left(v _ {s}\\right) - \\left\\{v _ {l}, v _ {r}, v _ {t} \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 527, + 903, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_{s}^{k}$ is the half-edge encountered during traversal.", + "bbox": [ + 511, + 556, + 883, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In conjunction with Eqs. (1) and (2), $N_{k + 1}(v_t)$ can also be determined. When $(v_s, v_t)$ is a boundary in $\\mathcal{M}_{k + 1}$ , we follow either $\\mathcal{H}_{ls}$ clockwise or $\\mathcal{H}_{sr}$ counterclockwise until we reach a half-edge of the boundary.", + "bbox": [ + 511, + 571, + 905, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After identifying the neighbors of $v_{s}$ and $v_{t}$ in the mesh $\\mathcal{M}_{k + 1}$ , we add the new vertex $v_{t}$ to the mesh $\\mathcal{M}_k$ and reconnect the vertices in $N_{k + 1}(v_t)$ to $v_{t}$ . Finally, we restore two faces, $\\{v_s,v_l,v_t\\}$ and $\\{v_r,v_s,v_t\\}$ , which are associated with the half-edges $\\mathcal{H}_{sl}^{t}$ and $\\mathcal{H}_{rs}^{t}$ , respectively. Using the half-edge data structure, the orientation of the newly created faces remains consistent with the rest of the mesh. When $(v_{s},v_{t})$ lies on the boundary, only one face is restored.", + "bbox": [ + 511, + 632, + 905, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, with the above half-edge formulation, each vertex split operation can be uniquely defined by the selection of the target vertex $v_{s}$ and two neighbors $v_{l}$ and $v_{r}$ , as well as the position to place the new vertex $v_{t}$ .", + "bbox": [ + 511, + 753, + 905, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vertex Split Tokenization. Although $v_{s}, v_{l}$ and $v_{r}$ only require references to existing vertices in the mesh, in practice we implement this reference as a raw prediction of each vertex to avoid a vocabulary size proportional to the sequence length. Thus, in the majority of cases, each vertex", + "bbox": [ + 511, + 830, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a3ec9d00c549e222e74e59e34f64589d85b2bb278fa23b478001ff0b965514e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 154 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/838b5a3a02951f2826ef225c104dbe6f899341e017da11c477342322afb22898.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 157, + 903, + 238 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31f0bcb4672f8d3a55fb57b66296abf470de15be0171a09c39608e19cfa1081e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 241, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/db8404fafb7a72ff0cfd61a46683118a25a8bfb9e0f3a77fe6d6a38c6aca455a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 321, + 903, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1c90d14a5746db392b51a47595c6cb1863a98ac2404d890fba2f27cdc3b30b58.jpg", + "image_caption": [ + "Figure 6. Generation process for VertexRegen. $\\mathcal{M}_0$ represents the predicted initial coarsest mesh, followed by subsequent meshes generated through the predicted vertex split sequence." + ], + "image_footnote": [], + "bbox": [ + 94, + 406, + 903, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "split is represented by a subsequence of 12 tokens (or 10 tokens when $(v_{s}, v_{t})$ is a boundary edge):", + "bbox": [ + 89, + 640, + 482, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "```c\nvsplit: [, , , #v_s", + "bbox": [ + 119, + 680, + 452, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$< 1_{-}x>, < 1_{-}y>, < 1_{-}z> |< \\mathrm{nil}>$ , # $v_{l}$", + "bbox": [ + 194, + 698, + 450, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$< r_{-}x>, < r_{-}y>, < r_{-}z|< nil>$ ， #vr", + "bbox": [ + 194, + 715, + 450, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$< t\\_ x > , < t\\_ y > , < t\\_ z>$ #vt", + "bbox": [ + 194, + 734, + 450, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the special token $\\langle \\mathrm{nil} \\rangle$ signifies $v_{l}$ or $v_{r}$ does not exist, and only one of $v_{l}$ and $v_{r}$ is allowed to be $\\langle \\mathrm{nil} \\rangle$ for a valid vertex split subsequence. In Fig. 5 we illustrate the entire tokenization process.", + "bbox": [ + 89, + 758, + 482, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above progressive mesh serialization can then be trained with the standard next-token prediction target.", + "bbox": [ + 89, + 819, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vertex Split Decoding. The vertex split sequence must be consistent with the actual geometry to be valid. We", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "maintain a state machine and perform vertex split on the fly as new tokens are generated. After the coarsest mesh $\\mathcal{M}_0$ is generated, we initialize the state. During each generation step, we enforce $v_{s}$ to be a vertex in the current mesh $\\mathcal{M}_k$ , $(v_{s}, v_{l})$ , $(v_{s}, v_{r})$ are valid edges (if $v_{l}$ and $v_{r}$ are not $\\langle \\mathrm{nil} \\rangle$ ), and only one of $v_{l}$ and $v_{r}$ can be $\\langle \\mathrm{nil} \\rangle$ . Lastly, we decode $v_{t}$ , and perform vertex split with generated $v_{s}, v_{l}, v_{r}, v_{t}$ information.", + "bbox": [ + 511, + 640, + 906, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Conditional Generation", + "text_level": 1, + "bbox": [ + 511, + 771, + 732, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider generation conditioned on shapes. We adopt a pre-trained point-cloud encoder [58]. A LLaVa-style [21] projector is leveraged to project the condition features to the token embedding space. The projected feature tokens are prepended to the mesh sequence as the prefix. We supervise training with the next-token prediction objective while masking the loss for the prefix tokens.", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12574", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/887cfc1e08e1fd02eb1493072ea60203070ee823e86026e6e06a6970273a8e5c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTokenizationCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)
MeshXL [2]Flattened Coords.51.768.3050.843.81
MeshAnything V2 [3]AMT50.338.5052.254.84
EdgeRunner [43]EdgeBreaker [33]51.397.8149.443.22
VertexRegen (Ours)Progressive51.038.2950.222.89
", + "bbox": [ + 217, + 88, + 781, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons with state-of-the-art methods for unconditional mesh generation. Best results are bolded, second best are underlined. VertexRegen can generate meshes with comparable quality while enjoying the benefits of continuous level of detail.", + "bbox": [ + 89, + 196, + 906, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da9816eabdcbcd81189ff45c87dadff60fdf83e77fdf2d55f01b0477c27c1b85.jpg", + "image_caption": [ + "Figure 7. Unconditional generation under face count constraints. VertexRegen achieves significantly better COV, MMD, and 1-NNA in early stages of generation." + ], + "image_footnote": [], + "bbox": [ + 91, + 244, + 906, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 453, + 223, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 479, + 192, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We pre-train our model and the baselines presented using two primary sources of meshes: Objaverse-XL [9] and an additional set of licensed artist-created meshes. Initially, we select meshes containing fewer than 8,000 faces without applying decimation. We use the CGAL [10] library to implement edge collapse and vertex split operations, where we modify the vertex placement to either $v_{s}$ or $v_{t}$ . We filter out non-manifold meshes and those that cannot be processed, which results in a final dataset of approximately 1.5M meshes with an average of 457 faces. For unconditional generation evaluation, we construct a high-quality subset by further filtering with the alignment split of Objaverse-XL and with fewer than 800 faces. This yields approximately 18k meshes. We adopt a similar filtering process for shape-conditioned experiments.", + "bbox": [ + 88, + 505, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We apply robust data augmentations to input meshes, including random shift within range $[-0.1, 0.1]$ , random scaling between [0.9, 1.1], random rotation by $0^{\\circ}$ , $90^{\\circ}$ , $180^{\\circ}$ , or $270^{\\circ}$ . During pre-training, we handle sequences exceeding the context window by discarding vertex split subsequences, while for other baselines, we truncate the tokenized sequence directly. The set for quantitative evaluation is chosen to avoid truncation for any method evaluated.", + "bbox": [ + 88, + 734, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 453, + 730, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VertexRegen and all other baselines are built upon the pretrained OPT-350M [56], with newly initialized token embeddings and position embeddings. We train our model on 64 H100 GPUs for approximately four days, with an effective batch size of 256. We use the AdamW [25] optimizer with a weight decay of 0.1, and betas (0.9, 0.95). A cosine scheduler is employed, starting with an initial learning rate of $1 \\times 10^{4}$ and gradually decreasing to $5 \\times 10^{6}$ . We clip the gradient norm to 1.0. During inference, we adopt the top- $p$ sampling strategy with $p = 0.95$ by default.", + "bbox": [ + 509, + 478, + 906, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 511, + 643, + 607, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Unconditional Generation", + "text_level": 1, + "bbox": [ + 511, + 667, + 741, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow evaluation protocols from prior works [1, 2, 37], employing point-cloud-based metrics to assess unconditional generation. Specifically, we sample the same number of meshes as the evaluation dataset and randomly sample 2,048 points per mesh. Coverage (COV) measures the diversity of the generated samples, where higher values indicate greater diversity. Minimum Matching Distance (MMD) computes the average distance from each reference sample to its nearest neighbor in the generated set, serving as a measure of generation quality, with lower values being preferable. 1-Nearest Neighbor Accuracy (1-NNA) evaluates both diversity and quality, where an optimal value is achieved at $50\\%$ . Additionally, we compute Jensen-Shannon Divergence (JSD) to directly quantify the sim", + "bbox": [ + 509, + 688, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We neither used assets from Sketchfab nor obtained any from the Polycam website.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c4f79cf3f552eb86c3ded40706d7fa9a2e5f6189bc60a14e92d97d62ee9bfae.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison with state-of-the-art methods. VertexRegen is able to generate meshes with comparable quality to other baselines." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 897, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a80fd6086d5a71c985ca3b6d7d8b5efa4b9a0c2973eb713545b57244e96d0c34.jpg", + "image_caption": [ + "Figure 9. Qualitative results on shape-conditioned generation on meshes held out from training. The first column shows the point cloud used as the condition, followed by the generation sequence progressing from the coarsest mesh $\\mathcal{M}_0$ to the final output in the last column." + ], + "image_footnote": [], + "bbox": [ + 94, + 489, + 901, + 824 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12576", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ilarity between the generated and reference distributions. We present results in Tab. 1. Our method can generate meshes of comparable quality to other state-of-the-art methods, with the added advantage of a continuous level of detail. We also show qualitative examples and their generation process in Fig. 6, and more generation results comparing with state-of-the-art methods in Fig. 8.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generation with a face count constraint. We compare our method with the baselines under varying face count constraints. Our approach allows for generation to be paused at any point, accommodating the varying constraints. In contrast, to adapt to these constraints, baseline methods require either: (i) directly truncating the generation process, yielding incomplete intermediate meshes, or (ii) fine-tuning with an additional face count condition token.", + "bbox": [ + 88, + 218, + 482, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the direct truncation version, we evaluate unconditional generation metrics w.r.t. different face count constraints and present the results in Fig. 7. Additionally, we adopt a similar approach as proposed in EdgeRunner, where we split [1,800] range into 4 buckets and prepend the conditioning token to the mesh sequence. We compare VertexRegen and MeshXL fine-tuned with such conditioning scheme at face constraint 400, as shown in Tab. 2.", + "bbox": [ + 89, + 339, + 482, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to its ability to generate meshes with a continuous level of detail from coarse to fine, VertexRegen effectively captures the overall structure even with very limited face counts, yielding significantly better COV, MMD, and 1-NNA early in the process. As the face limit increases, all methods demonstrate improved results. Notably, for baselines such as EdgeRunner, while they enhance tokenization efficiency over MeshXL, they do not fundamentally address continuous level-of-detail generation and thus follow a similar trend to other baselines in the plot.", + "bbox": [ + 89, + 460, + 482, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Conditional Generation", + "text_level": 1, + "bbox": [ + 89, + 619, + 302, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show qualitative examples of shape-conditioned generation in Fig. 9. We condition on 4,096 sampled points with normals from dense meshes. VertexRegen can generate a coarse $\\mathcal{M}_0$ , starting as simply as a tetrahedron, and progressively refine it by generating a sequence of vertex splits.", + "bbox": [ + 89, + 640, + 482, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 727, + 254, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tokenization efficiency. We compute the tokenized sequence length of our proposed tokenization scheme (Sec. 3.2.1) on our dataset and report the average compression ratio relative to MeshXL (9 tokens per face) in Tab. 3. For VertexRegen, there are two primary sequence types: (i) a MeshXL-style sequence for the initial coarsest mesh $\\mathcal{M}_0$ and (ii) sequences encoding vertex splits. The latter requires 12 tokens for two non-boundary faces or 10 tokens per boundary face. As a result, the highest compression is achieved when $\\mathcal{M}_0$ is minimized and all vertex splits occur", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a22f183ec76f9591698f6db3b5857428809fe65d449d7b224ae3826165ccd9af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
@400 FacesCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
VertexRegen50.928.3151.032.88264147
MeshXL (w/ FCC)41.2010.0359.065.19308168
", + "bbox": [ + 519, + 89, + 903, + 152 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/da59ddc80e921a51a43d7518d086d416b398970160b2eb3a293e44e11018c6d9.jpg", + "table_caption": [ + "Table 2. Comparison between VertexRegen and MeshXL with face count condition (FCC) when face count constraint is 400." + ], + "table_footnote": [], + "table_body": "
MeshXLMeshAnything v2EdgeRunnerVertexRegen w/ HEw/o HE
Compression1.00.460.470.730.89
", + "bbox": [ + 519, + 202, + 901, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/90e337f1a55043eebcfec22e3fe0a474337de63382a52db87ff9971173642a4f.jpg", + "table_caption": [ + "Table 3. Compression ratio of different tokenization schemes, VertexRegen with and without leveraging a half-edge (HE) structure." + ], + "table_footnote": [], + "table_body": "
Guided DecodingCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
w/o51.128.3150.753.37211120
w/51.038.2950.222.89320176
", + "bbox": [ + 517, + 297, + 901, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Effects on geometry-guided decoding.", + "bbox": [ + 568, + 367, + 849, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on non-boundary faces. On average, VertexRegen achieves a compression ratio of 0.73, approaching the theoretical limit of 0.67. The discrepancy arises from the overhead introduced by $\\mathcal{M}_0$ (5.68% of all total tokenized length) and boundary vertex splits (3.64% of all vertex splits).", + "bbox": [ + 511, + 388, + 903, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Without the half-edge data structure, identifying which half of the ring surrounding $v_{s}$ is associated with $v_{s}$ or $v_{t}$ in $\\mathcal{M}_{k + 1}$ requires recording an additional vertex. This increases the token count to 15 per two non-boundary faces (or 13 per boundary face) and results in an average increase of $22\\%$ in tokenized sequence length.", + "bbox": [ + 511, + 465, + 905, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Guided decoding. In Tab. 4, we ablate the effects of geometry-constrained decoding. As one step of the vertex split operation is dependent on all the preceding operations, predicting an invalid vertex split may break the chain and end the generation prematurely. With guided decoding, the model is able to generate longer sequences with more faces.", + "bbox": [ + 511, + 583, + 905, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced VertexRegen, a novel mesh generation framework that enables continuous levels of detail through a generative process based on vertex splits. Unlike conventional auto-regressive approaches that synthesize meshes in a partial-to-complete manner, VertexRegen reinterprets mesh generation as the reversal of edge collapse, providing an effectively anytime solution to mesh generation. Our experimental results demonstrate that VertexRegen achieves competitive performance compared to state-of-the-art methods while offering the unique advantage of halting generation at any stage to obtain meshes at different levels of detail.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12577", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2, 6", + "[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Billzb Wang, Jingyi Yu, Gang Yu, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems, 37:97141-97166, 2024. 1, 2, 3, 4, 6", + "[3] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 6", + "[4] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2", + "[5] Yun-Chun Chen, Vladimir Kim, Noam Aigerman, and Alec Jacobson. Neural progressive meshes. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3", + "[6] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G. Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4456-4465, 2023. 2", + "[8] Angela Dai and Matthias Niessner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36:35799-35813, 2023. 6", + "[10] Andreas Fabri and Sylvain Pion. Cgal: the computational geometry algorithms library. In Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, page 538-539, New York, NY, USA, 2009. Association for Computing Machinery. 6", + "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1", + "[12] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2, 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2", + "[15] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 1, 2", + "[16] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[17] Hugues Hoppe. Progressive meshes. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 99–108, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3", + "[18] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2", + "[19] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2", + "[20] Peter Lindstrom and Greg Turk. Fast and memory efficient polygonal simplification. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 279-286. IEEE, 1998. 2", + "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5", + "[22] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. In Advances in Neural Information Processing Systems, pages 59314-59341. Curran Associates, Inc., 2024. 2", + "[23] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2", + "[24] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 2", + "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12578", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] David Luebke, Martin Reddy, Jonathan D. Cohen, Amitabh Varshney, Benjamin Watson, and Robert Huebner. Level of Detail for 3D Graphics. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 2002. 2", + "[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[28] Felix Mujkanovic, Ntumba Elie Nsampi, Christian Theobalt, Hans-Peter Seidel, and Thomas Leimkuhler. Neural gaussian scale-space fields. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 3", + "[29] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 2", + "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2", + "[31] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[32] Jovan Popovic and Hugues Hoppe. Progressive simplicial complexes. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 217-224, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2", + "[33] J. Rossignac. Edgebreaker: connectivity compression for triangle meshes. IEEE Transactions on Visualization and Computer Graphics, 5(1):47-61, 1999. 6", + "[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 1", + "[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2", + "[36] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2", + "[37] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2, 4, 6", + "[38] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "David Novotny. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. In Advances in Neural Information Processing Systems, pages 9532-9564. Curran Associates, Inc., 2024. 1", + "[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11367, 2021. 3", + "[40] Danhang Tang, Mingsong Dou, Peter Lincoln, Philip Davidson, Kaiwen Guo, Jonathan Taylor, Sean Fanello, Cem Keskin, Adarsh Kowdle, Sofien Bouaziz, et al. Real-time compression and streaming of 4d performances. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018.", + "[41] Danhang Tang, Saurabh Singh, Philip A. Chou, Christian Hane, Mingsong Dou, Sean Fanello, Jonathan Taylor, Philip Davidson, Onur G. Guleryuz, Yinda Zhang, Shahram Izadi, Andrea Tagliasacchi, Sofien Bouaziz, and Cem Keskin. Deep implicit volume compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[43] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgrunner: Auto-regressive auto-encoder for artistic mesh generation. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 6", + "[44] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2", + "[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2", + "[46] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2", + "[47] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 731-741, 2025. 2", + "[48] Kevin J Weiler. Topological structures for geometric modeling (Boundary representation, manifold, radial edge structure). Rensselaer Polytechnic Institute, 1986. 4", + "[49] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12579", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, Tong Zhang, Shenghua Gao, and C.L. Philip Chen. Scaling mesh generation via compressive tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11093-11103, 2025. 2", + "[51] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1", + "[52] Haiyang Xu, Yu Lei, Zeyuan Chen, Xiang Zhang, Yue Zhao, Yilin Wang, and Zhuowen Tu. Bayesian diffusion models for 3d shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10628-10638, 2024. 2", + "[53] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2", + "[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems, 2022. 1", + "[55] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2", + "[56] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6", + "[57] Qingcheng Zhao, Xiang Zhang, Haiyang Xu, Zeyuan Chen, Jianwen Xie, Yuan Gao, and Zhuowen Tu. Depr: Depth guided single-view scene reconstruction with instance-level diffusion. arXiv preprint arXiv:2507.22825, 2025. 2", + "[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems, pages 73969-73982. Curran Associates, Inc., 2023. 5", + "[59] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826–5835, 2021. 1, 2" + ], + "bbox": [ + 91, + 92, + 480, + 781 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12580", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_model.json b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2bbfcadf28840a768d41bc50aa33ff1259e9f33f --- /dev/null +++ b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "VertexRegen: Mesh Generation with Continuous Level of Detail" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.179, + 0.775, + 0.235 + ], + "angle": 0, + "content": "Xiang Zhang\\(^{1*}\\) Yawar Siddiqui\\(^{2}\\) Armen Avetisyan\\(^{2}\\) Chris Xie\\(^{2}\\) Jakob Engel\\(^{2}\\) Henry Howard-Jenkins\\(^{2}\\) \\(^{1}\\)UC San Diego \\(^{2}\\)Meta Reality Labs Research" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.25, + 0.891, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.494, + 0.908, + 0.539 + ], + "angle": 0, + "content": "Figure 1. Mesh generation process of VertexRegen (coarse-to-fine process) vs. previous work (partial-to-complete process). Prior work generates meshes face by face, with the step count corresponding to the face count in the figure. In contrast, VertexRegen produces meshes with a continuously increasing level of detail, where each step yields a valid mesh \\(\\mathcal{M}_k\\)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.552, + 0.329, + 0.568 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.486, + 0.766 + ], + "angle": 0, + "content": "We introduce VertexRegen, a novel mesh generation framework that enables generation at a continuous level of detail. Existing autoregressive methods generate meshes in a partial-to-complete manner and thus intermediate steps of generation represent incomplete structures. VertexRegen takes inspiration from progressive meshes and reformulates the process as the reversal of edge collapse, i.e. vertex split, learned through a generative model. Experimental results demonstrate that VertexRegen produces meshes of comparable quality to state-of-the-art methods while uniquely offering anytime generation with the flexibility to halt at any step to yield valid meshes with varying levels of detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.224, + 0.81 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.82, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Meshes are essential for 3D asset representation and are widely used in industries such as film, design, and gaming due to their compatibility with most 3D software and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.703 + ], + "angle": 0, + "content": "hardware. However, manually creating high-quality meshes is costly and time-consuming, prompting extensive research into automated 3D content creation. Largely, these approaches have used alternative representations such as neural fields [11, 27, 31, 38], voxels [34, 51] or point clouds [54, 59] which are later post-processed to meshes instead of direct mesh modeling. Unfortunately, these post-processed meshes often exhibit poor topology, overtessellation, and artifacts, lacking the quality of artist-crafted meshes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, there has been a surge of approaches that directly generate meshes, using autoregressive models to represent meshes as sequences of triangles [2-4, 15, 37, 43]. These methods capture the high-fidelity and aesthetic qualities of artist-created works without the need for post-generation conversion. Although significant progress has been made in improving tokenization schemes [2, 4, 37] and network architectures [15], these approaches do not fundamentally alter a key characteristic of generation: namely, to produce a valid mesh, the full sequence must be generated to completion. Consequently, these methods offer no control over the level of detail during generation; early stopping results in a mesh with missing faces. Simple extensions," + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.875, + 0.424, + 0.901 + ], + "angle": 0, + "content": "*Work conducted while the author was an intern at Meta. \nProject page: https://vertexregen.github.io" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12570" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "such as face-count conditioning tokens in EdgeRunner [43], have been proposed to introduce some coarse control over the detail by pre-specifying a target face count. However, each generation still must be completed in its entirety to yield a complete mesh, and thus a single generation still offers only a single level of detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.363 + ], + "angle": 0, + "content": "To allow for a continuous level of detail generation, we take inspiration from Hoppe's progressive mesh formulation [17]. Progressive meshes use two reversible operations to transition between levels of detail. The edge collapse operation simplifies a mesh into a coarser one by reducing one edge at a time. The vertex split operation reverses this simplification to add more detail by using information stored during edge collapse. By starting with a coarse mesh and keeping edge collapse records, this approach creates an efficient, lossless representation of the original mesh, allowing for continuous resolution adjustments by applying any number of vertex split operations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.604 + ], + "angle": 0, + "content": "In this paper, we build on this progressive mesh representation by learning the vertex split, i.e. reversing the edge collapse operation, as a generative problem. This allows the resulting mesh generation to inherit the properties of progressive meshes and provides a solution for anytime mesh generation, where the process can be stopped early to yield a coarser mesh rather than an incomplete one. By properly serializing the vertex split sequence, the entire generation process can be modeled using a Transformer trained with a next-token prediction objective, a widely adopted paradigm in mesh generation. We evaluate VertexRegen on the task of unconditional mesh generation, demonstrating superior results both qualitatively and quantitatively. We further provide examples of shape-conditioned generation. These experimental results illustrate the ability of our method to generate compelling meshes at a continuous resolution." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.605, + 0.435, + 0.619 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.622, + 0.483, + 0.667 + ], + "angle": 0, + "content": "- Inspired by Hoppe's progressive meshes [17], we reframe mesh generation as the reversal of edge collapse operations, i.e. generating vertex splits." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.668, + 0.483, + 0.697 + ], + "angle": 0, + "content": "- We formulate a token-efficient parameterization of a progressive mesh, through a half-edge data structure." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.698, + 0.483, + 0.758 + ], + "angle": 0, + "content": "- We propose VertexRegen for continuous level of detail mesh generation. VertexRegen creates meshes in a coarse-to-fine fashion, rather than partial-to-complete, uniquely providing a solution for anytime generation." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.622, + 0.483, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.233, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "3D Mesh Generation. Recent advancements in 3D shape generation have explored various representations, including point-clouds [30, 44, 52, 59], signed distance functions (SDFs) [7, 19, 22, 23, 36, 53, 57], neural radiance fields (NeRFs) [16, 18, 46, 55], and Gaussian splatting [14, 42]. These implicit representations require iso-surface extraction techniques [24, 35, 47] to output meshes, often resulting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "in over-tessellated and excessively smooth outputs, which pose challenges for downstream applications such as geometric processing and manipulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.544 + ], + "angle": 0, + "content": "In contrast, direct mesh generation inherently produces structured, well-defined geometry without the need for postprocessing or surface extraction, making it an increasingly prominent approach in recent years. Early methods tackle this task by generating meshes from surface patches [13], deforming ellipsoids [45], predicting mesh graphs [8], or employing binary space partitioning [6]. More recent techniques leverage generative models, particularly diffusion models and sequence-based approaches. PolyDiff [1] applies discrete diffusion, while PolyGen [29] autoregressively predicts vertices and faces using two separate networks. The sequence modeling paradigm has been further refined by representing the entire mesh as a single sequence. MeshGPT [37] introduces a tokenization scheme based on a vector quantized variational auto-encoder (VQVAE), while MeshXL [2] directly models discretized triangle soup sequences without compression. MeshAnything [4] extends MeshGPT by incorporating a point-cloud encoder for shape-conditioned generation. PivotMesh [49] introduces a hierarchical approach, generating pivot vertices before producing the full mesh. Further research has focused on optimizing tokenization and sequence modeling. Adjacency-aware compression techniques [3, 43, 50] improve tokenization efficiency and enable higher face counts within fixed context windows, while Meshtron [15] leverages Hourglass Transformers and sliding window attention to scale MeshXL sequences more effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.905, + 0.637 + ], + "angle": 0, + "content": "Despite these advancements, most existing approaches follow a partial-to-complete paradigm, where mesh regions are constructed sequentially at a uniform level of detail. Our work instead adopts a coarse-to-fine approach, continuously increasing the level of detail as new tokens are generated, achieving better control over geometric complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Level of Detail Representations. Level of Detail [26] (LOD) is a widely used technique in computer graphics to optimize rendering performance by reducing the complexity of 3D models based on their size, distance from the camera, or importance in a scene. Different LOD strategies depend on the underlying representation of 3D shapes. For meshes, progressive meshes [17] generate a mesh sequence starting from a coarse base model, gradually refining it through a series of transformations that incrementally add detail. Traditional mesh simplification methods [12, 20] can be adopted to construct such sequences by iteratively reducing polygon count while preserving geometric fidelity. Progressive simplicial complexes [32] extend progressive meshes to handle arbitrary meshes, including non-manifold and non-orientable surfaces." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recent studies have incorporated neural networks for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.48, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.205, + 0.483, + 0.261 + ], + "angle": 0, + "content": "Figure 2. Illustration of edge collapse and its inverse operation, i.e. vertex split. During edge collapse, vertex \\( v_{t} \\) is collapsed into \\( v_{s} \\), resulting in two degenerate triangles (shaded in yellow), which forms two new edges in the result mesh." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.484, + 0.477 + ], + "angle": 0, + "content": "LOD representations. Neural progressive meshes [5] propose a learned approach supporting LOD with a subdivision-based encoder-decoder. Similar representations have been explored for reconstruction approaches with signed distance functions (SDFs) [39-41] and neural fields [28]. However, these approaches are not generative; they typically encode and decode existing meshes [5] or optimize the LOD representation per shape [39]. In contrast, our approach is purely generative. VertexRegen learns to create meshes from scratch, progressively adding detail through autoregressive sequence generation. This allows us to generate new, high-fidelity meshes without relying on pre-existing structures, distinguishing our method from both traditional and neural LOD techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.182, + 0.506 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.355, + 0.532 + ], + "angle": 0, + "content": "3.1. Progressive Meshes Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.674 + ], + "angle": 0, + "content": "A progressive mesh [17] (PM) proposes an efficient and lossless continuous-resolution representation for arbitrary triangle meshes. The representation is built off two observations: (i) that a single mesh transformation, edge collapse, is sufficient for effective simplification of meshes; (ii) that edge collapse transformations are invertible via a vertex split operation. In the following, we provide details of the edge collapse operation and its inversion with vertex splits, each illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.765 + ], + "angle": 0, + "content": "The edge collapse operation, \\(\\operatorname{ecol}(v_s, v_t)\\), unifies two adjacent vertices \\(v_s\\) and \\(v_t\\) into a single vertex \\(v_s\\). The operation results in the vanishing of two faces, \\(\\{v_t, v_s, v_l\\}\\) and \\(\\{v_s, v_t, v_r\\}\\), as well as the vertex \\(v_t\\). In the general case, a new position for \\(v_s\\) is also specified; however, for the case of half-edge collapse, the original position for \\(v_s\\) is kept." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.796 + ], + "angle": 0, + "content": "An initial mesh, \\(\\mathcal{M}\\), can be simplified into a more coarse mesh, \\(\\mathcal{M}_n\\), by \\(n\\) successive edge collapse operations:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.807, + 0.434, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {0} \\xleftarrow {\\mathrm {e c o l} _ {0}} \\mathcal {M} _ {1} \\xleftarrow {\\mathrm {e c o l} _ {1}} \\dots \\xleftarrow {\\mathrm {e c o l} _ {n - 1}} \\mathcal {M}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The order in which the edge collapse operations are performed is determined such that each successive edge collapse results in the minimum increase in Quadratic Error Metrics (QEM) [12] with respect to the original mesh. This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "process can be repeated until there are no candidate edge collapses, e.g. without flipping face normals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.213 + ], + "angle": 0, + "content": "A vertex split, \\(\\mathrm{vsplit}(v_s,v_l,v_r,v_t)\\), defines the inverse of edge collapse. It restores \\(v_{t}\\) and the two faces, \\(\\{v_{t},v_{s},v_{l}\\}\\) and \\(\\{v_{s},v_{t},v_{r}\\}\\), which vanished during edge collapse, as well as \\(v_{s}\\) to its original position. In case \\((v_{s},v_{t})\\) is a boundary edge, either \\(v_{l}\\) or \\(v_{r}\\) will not exist, and there is only one face restored during the vertex split." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.304 + ], + "angle": 0, + "content": "The combination of edge collapse and its inverse vertex splits enables the progressive mesh representation, \\(\\mathrm{PM}(\\mathcal{M}) = (\\mathcal{M}_0,\\{\\mathrm{vsplit}_0,\\dots ,\\mathrm{vsplit}_{n - 1}\\})\\) to express an arbitrary triangle mesh, \\(\\mathcal{M}\\), as a combination of a coarse mesh, \\(\\mathcal{M}_0\\), obtained through edge collapse, and the sequence of \\(n\\) vertex split records required to reverse them:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.316, + 0.871, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {0} \\xrightarrow {\\mathrm {v s p l i t} _ {0}} \\mathcal {M} _ {1} \\xrightarrow {\\mathrm {v s p l i t} _ {1}} \\dots \\xrightarrow {\\mathrm {v s p l i t} _ {n - 1}} \\mathcal {M}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.35, + 0.649, + 0.365 + ], + "angle": 0, + "content": "3.2. VertexRegen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.906, + 0.476 + ], + "angle": 0, + "content": "Traditional progressive meshes require starting an initial detailed mesh to be simplified into a coarse mesh, \\(\\mathcal{M}_0\\), and recording each of these collapse steps as vertex split records to form the representation. We observe that the formation of the PM representation serves effectively as forward (edge collapse) and reverse (vertex split) processes that enable transition between detailed and coarse meshes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.569 + ], + "angle": 0, + "content": "VertexRegen frames the creation of a detailed mesh as the generation of a PM representation. Analogous to denoising for diffusion models, we train a generative model to reverse edge collapse. Concretely, VertexRegen first generates a coarse mesh, \\(\\mathcal{M}_0\\), from scratch, before increasing the level of detail through the generation of vertex split records." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.807, + 0.591 + ], + "angle": 0, + "content": "3.2.1. Progressive Mesh Parameterization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.905, + 0.627 + ], + "angle": 0, + "content": "We frame generation as an autoregressive sequential modeling task, with the full mesh sequence taking the form:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.637, + 0.904, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c c} \\text {M : [ < b o s > , [ M _ {-} 0 s e q u e n c e ] , < s e p > ,} & \\# \\mathcal {M} _ {0} \\\\ \\text {[ v s p l i t \\_ 0 ] , \\ldots , [ v s p l i t \\_ n - 1 ] , < e o s > ]} & \\# \\text {v s p l i t s} \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.679, + 0.905, + 0.71 + ], + "angle": 0, + "content": "In the following, we will define the formation of the subsequences [M_0 sequence] and [vsplit]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.906, + 0.838 + ], + "angle": 0, + "content": "\\(\\mathcal{M}_0\\): Coarse Mesh Tokenization. For the initial coarse mesh, we follow the tokenization scheme defined in MeshXL [2]. In this formulation, embeddings are learned for discretized coordinates in an \\(N^3\\) grid. A vertex is represented via the sequential look-up of \\(x\\)-value, \\(y\\)-value, and \\(z\\)-value. A face is then constructed as the concatenation of its 3 vertices, totaling 9 tokens. The full mesh sequence is then defined as the concatenation of its constituent faces." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.847, + 0.86, + 0.862 + ], + "angle": 0, + "content": "\\[\nv: \\left[ < x >, < y >, < z > \\right], \\quad \\# v e r t e x\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.865, + 0.845, + 0.88 + ], + "angle": 0, + "content": "\\[\nF: \\left[ \\begin{array}{l} [ v _ {- 1} ], [ v _ {- 2} ], [ v _ {- 3} ] \\end{array} \\right], \\quad \\# \\text {f a c e}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.882, + 0.852, + 0.897 + ], + "angle": 0, + "content": "\\[\nM: \\left[ [ F _ {- 1} ], [ F _ {- 2} ], \\dots , [ F _ {- N} ] \\right] \\quad \\# \\text {m e s h}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12572" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.092, + 0.442, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.485, + 0.326 + ], + "angle": 0, + "content": "Figure 3. Face count distribution of the coarsest mesh \\(\\mathcal{M}_0\\) and original mesh \\(\\mathcal{M}\\). The coarsest meshes \\((\\mathcal{M}_0)\\) contain significantly fewer faces than original \\((\\mathcal{M})\\), with an average of 18 and 457 faces, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.484, + 0.375 + ], + "angle": 0, + "content": "Following [37], we sort vertices in \\( z - y - x \\) order (lower to higher). Within each face, vertices are cylindrically permuted to have the lowest-indexed vertex be first." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.482 + ], + "angle": 0, + "content": "Although MeshXL [2] leverages this tokenization scheme to produce full detailed meshes, in VertexRegen only the coarsest level of the mesh is parameterized in this way. These coarse meshes consist of substantially fewer faces than the full detailed mesh, as demonstrated in Fig. 3, and the resulting coarse mesh only accounts for \\(5.68\\%\\) of the total sequence length on average." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.652 + ], + "angle": 0, + "content": "Vertex Split Generation. Consider the \\(k\\)-th vertex split operation \\(\\mathrm{vsplit}_k\\), which converts a lower-detailed mesh \\(\\mathcal{M}_k\\) into a higher-detailed mesh \\(\\mathcal{M}_{k+1}\\). We denote the neighbor vertices of vertex \\(v\\) in \\(\\mathcal{M}_k\\) as \\(N_k(v)\\). After determining the target vertex \\(v_s\\) to split (in \\(\\mathcal{M}_k\\)) and the new vertex \\(v_t\\) (in \\(\\mathcal{M}_{k+1}\\)), we need to obtain \\(N_{k+1}(v_s)\\) and \\(N_{k+1}(v_t)\\) from the vertex split record to connect \\(v_s\\) and \\(v_t\\) with correct neighbor vertices in \\(\\mathcal{M}_{k+1}\\), which are prohibitive to generate as the number of neighbor vertices \\(|N_{k+1}(v_s)|\\) and \\(|N_{k+1}(v_t)|\\) is often large." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.484, + 0.684 + ], + "angle": 0, + "content": "However, as mesh \\(\\mathcal{M}_k\\) is generated from \\(\\mathcal{M}_{k + 1}\\) during edge collapse by merging \\(v_{t}\\) into \\(v_{s}\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.696, + 0.483, + 0.713 + ], + "angle": 0, + "content": "\\[\nN _ {k + 1} \\left(v _ {s}\\right) \\cap N _ {k + 1} \\left(v _ {t}\\right) = \\left\\{v _ {l}, v _ {r} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.726, + 0.483, + 0.743 + ], + "angle": 0, + "content": "\\[\nN _ {k + 1} \\left(v _ {s}\\right) \\cup N _ {k + 1} \\left(v _ {t}\\right) - \\left\\{v _ {s}, v _ {t} \\right\\} = N _ {k} \\left(v _ {s}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.87 + ], + "angle": 0, + "content": "which indicates we will only need to record vertex \\( v_{l} \\) and \\( v_{r} \\). They split the vertex ring surrounding \\( v_{s} \\) in \\( \\mathcal{M}_k \\), where two halves of the vertices on the ring (top and bottom vertices in Fig. 2) belong to \\( v_{t} \\) and \\( v_{s} \\) respectively in \\( \\mathcal{M}_{k + 1} \\). However, an ambiguity arises in determining which half of the ring corresponds to \\( v_{t} \\)'s neighbors in \\( \\mathcal{M}_{k + 1} \\). In the following, we demonstrate how the half-edge data structure [48] can be utilized to resolve this ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Denote \\(\\mathcal{H}_{ij}^k\\) as a half-edge pointing from \\(v_{i}\\) to \\(v_{j}\\), with \\(v_{k}\\) being the third vertex in the associated face. As shown in" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.093, + 0.903, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.249, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Figure 4. Illustration of half-edge data structure and the traversal process to determine the neighbors of \\( v_{s} \\) and \\( v_{t} \\) in \\( \\mathcal{M}_{k + 1} \\). Starting from half-edge \\( \\mathcal{H}_{ls}^{1} \\), we traverse the faces in clockwise direction until \\( \\mathcal{H}_{sr}^{2} \\), where we collect the vertices \\( v_{1} \\) and \\( v_{2} \\) into \\( N_{k + 1}(v_s) \\)." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.321, + 0.665, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.321, + 0.904, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.414, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Figure 5. Illustration of VertexRegen tokenization. The sequence begins with base mesh \\(\\mathcal{M}_0\\), followed by vertex split subsequences. A special token indicates either \\(v_{l}\\) or \\(v_{r}\\) does not exist." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Fig. 4, we begin the traversal from \\(\\mathcal{H}_{ls}^{1}\\) and proceed to the twin of the next half-edge, \\(\\mathcal{H}_{1s}^{2}\\). This operation is repeated until we reach \\(\\mathcal{H}_{,s}^{r}\\). Throughout this process, we traverse the faces above \\((v_{l}, v_{s})\\) and \\((v_{s}, v_{r})\\) clockwise, obtaining:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.529, + 0.905, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\left\\{v _ {k} \\mid \\mathcal {H} _ {\\cdot s} ^ {k}, k \\neq r \\right\\} = N _ {k + 1} \\left(v _ {s}\\right) - \\left\\{v _ {l}, v _ {r}, v _ {t} \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.885, + 0.572 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_{s}^{k}\\) is the half-edge encountered during traversal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.906, + 0.633 + ], + "angle": 0, + "content": "In conjunction with Eqs. (1) and (2), \\(N_{k + 1}(v_t)\\) can also be determined. When \\((v_s, v_t)\\) is a boundary in \\(\\mathcal{M}_{k + 1}\\), we follow either \\(\\mathcal{H}_{ls}\\) clockwise or \\(\\mathcal{H}_{sr}\\) counterclockwise until we reach a half-edge of the boundary." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.633, + 0.906, + 0.753 + ], + "angle": 0, + "content": "After identifying the neighbors of \\( v_{s} \\) and \\( v_{t} \\) in the mesh \\( \\mathcal{M}_{k + 1} \\), we add the new vertex \\( v_{t} \\) to the mesh \\( \\mathcal{M}_k \\) and reconnect the vertices in \\( N_{k + 1}(v_t) \\) to \\( v_{t} \\). Finally, we restore two faces, \\( \\{v_s,v_l,v_t\\} \\) and \\( \\{v_r,v_s,v_t\\} \\), which are associated with the half-edges \\( \\mathcal{H}_{sl}^{t} \\) and \\( \\mathcal{H}_{rs}^{t} \\), respectively. Using the half-edge data structure, the orientation of the newly created faces remains consistent with the rest of the mesh. When \\( (v_{s},v_{t}) \\) lies on the boundary, only one face is restored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.754, + 0.906, + 0.814 + ], + "angle": 0, + "content": "Hence, with the above half-edge formulation, each vertex split operation can be uniquely defined by the selection of the target vertex \\( v_{s} \\) and two neighbors \\( v_{l} \\) and \\( v_{r} \\), as well as the position to place the new vertex \\( v_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Vertex Split Tokenization. Although \\( v_{s}, v_{l} \\) and \\( v_{r} \\) only require references to existing vertices in the mesh, in practice we implement this reference as a raw prediction of each vertex to avoid a vocabulary size proportional to the sequence length. Thus, in the majority of cases, each vertex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.158, + 0.904, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.242, + 0.904, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.322, + 0.904, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.407, + 0.904, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.584, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Figure 6. Generation process for VertexRegen. \\(\\mathcal{M}_0\\) represents the predicted initial coarsest mesh, followed by subsequent meshes generated through the predicted vertex split sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.671 + ], + "angle": 0, + "content": "split is represented by a subsequence of 12 tokens (or 10 tokens when \\((v_{s}, v_{t})\\) is a boundary edge):" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.681, + 0.454, + 0.696 + ], + "angle": 0, + "content": "```c\nvsplit: [, , , #v_s" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.699, + 0.451, + 0.714 + ], + "angle": 0, + "content": "\\(< 1_{-}x>, < 1_{-}y>, < 1_{-}z> |< \\mathrm{nil}>\\), # \\(v_{l}\\)" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.717, + 0.452, + 0.731 + ], + "angle": 0, + "content": "\\(< r_{-}x>, < r_{-}y>, < r_{-}z|< nil>\\) ， #vr" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.735, + 0.451, + 0.75 + ], + "angle": 0, + "content": "\\(< t\\_ x > , < t\\_ y > , < t\\_ z>\\) #vt" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.483, + 0.82 + ], + "angle": 0, + "content": "where the special token \\( \\langle \\mathrm{nil} \\rangle \\) signifies \\( v_{l} \\) or \\( v_{r} \\) does not exist, and only one of \\( v_{l} \\) and \\( v_{r} \\) is allowed to be \\( \\langle \\mathrm{nil} \\rangle \\) for a valid vertex split subsequence. In Fig. 5 we illustrate the entire tokenization process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.82, + 0.483, + 0.851 + ], + "angle": 0, + "content": "The above progressive mesh serialization can then be trained with the standard next-token prediction target." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Vertex Split Decoding. The vertex split sequence must be consistent with the actual geometry to be valid. We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.763 + ], + "angle": 0, + "content": "maintain a state machine and perform vertex split on the fly as new tokens are generated. After the coarsest mesh \\(\\mathcal{M}_0\\) is generated, we initialize the state. During each generation step, we enforce \\(v_{s}\\) to be a vertex in the current mesh \\(\\mathcal{M}_k\\), \\((v_{s}, v_{l})\\), \\((v_{s}, v_{r})\\) are valid edges (if \\(v_{l}\\) and \\(v_{r}\\) are not \\(\\langle \\mathrm{nil} \\rangle\\)), and only one of \\(v_{l}\\) and \\(v_{r}\\) can be \\(\\langle \\mathrm{nil} \\rangle\\). Lastly, we decode \\(v_{t}\\), and perform vertex split with generated \\(v_{s}, v_{l}, v_{r}, v_{t}\\) information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.733, + 0.788 + ], + "angle": 0, + "content": "3.3. Conditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We consider generation conditioned on shapes. We adopt a pre-trained point-cloud encoder [58]. A LLaVa-style [21] projector is leveraged to project the condition features to the token embedding space. The projected feature tokens are prepended to the mesh sequence as the prefix. We supervise training with the next-token prediction objective while masking the loss for the prefix tokens." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12574" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.218, + 0.089, + 0.782, + 0.196 + ], + "angle": 0, + "content": "
MethodTokenizationCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)
MeshXL [2]Flattened Coords.51.768.3050.843.81
MeshAnything V2 [3]AMT50.338.5052.254.84
EdgeRunner [43]EdgeBreaker [33]51.397.8149.443.22
VertexRegen (Ours)Progressive51.038.2950.222.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.197, + 0.907, + 0.225 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with state-of-the-art methods for unconditional mesh generation. Best results are bolded, second best are underlined. VertexRegen can generate meshes with comparable quality while enjoying the benefits of continuous level of detail." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.245, + 0.907, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.398, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Figure 7. Unconditional generation under face count constraints. VertexRegen achieves significantly better COV, MMD, and 1-NNA in early stages of generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.454, + 0.224, + 0.47 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.194, + 0.495 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.483, + 0.732 + ], + "angle": 0, + "content": "We pre-train our model and the baselines presented using two primary sources of meshes: Objaverse-XL [9] and an additional set of licensed artist-created meshes. Initially, we select meshes containing fewer than 8,000 faces without applying decimation. We use the CGAL [10] library to implement edge collapse and vertex split operations, where we modify the vertex placement to either \\( v_{s} \\) or \\( v_{t} \\). We filter out non-manifold meshes and those that cannot be processed, which results in a final dataset of approximately 1.5M meshes with an average of 457 faces. For unconditional generation evaluation, we construct a high-quality subset by further filtering with the alignment split of Objaverse-XL and with fewer than 800 faces. This yields approximately 18k meshes. We adopt a similar filtering process for shape-conditioned experiments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "We apply robust data augmentations to input meshes, including random shift within range \\([-0.1, 0.1]\\), random scaling between [0.9, 1.1], random rotation by \\(0^{\\circ}\\), \\(90^{\\circ}\\), \\(180^{\\circ}\\), or \\(270^{\\circ}\\). During pre-training, we handle sequences exceeding the context window by discarding vertex split subsequences, while for other baselines, we truncate the tokenized sequence directly. The set for quantitative evaluation is chosen to avoid truncation for any method evaluated." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.731, + 0.47 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.479, + 0.907, + 0.63 + ], + "angle": 0, + "content": "VertexRegen and all other baselines are built upon the pretrained OPT-350M [56], with newly initialized token embeddings and position embeddings. We train our model on 64 H100 GPUs for approximately four days, with an effective batch size of 256. We use the AdamW [25] optimizer with a weight decay of 0.1, and betas (0.9, 0.95). A cosine scheduler is employed, starting with an initial learning rate of \\(1 \\times 10^{4}\\) and gradually decreasing to \\(5 \\times 10^{6}\\). We clip the gradient norm to 1.0. During inference, we adopt the top-\\(p\\) sampling strategy with \\(p = 0.95\\) by default." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.608, + 0.659 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.742, + 0.683 + ], + "angle": 0, + "content": "4.3.1. Unconditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We follow evaluation protocols from prior works [1, 2, 37], employing point-cloud-based metrics to assess unconditional generation. Specifically, we sample the same number of meshes as the evaluation dataset and randomly sample 2,048 points per mesh. Coverage (COV) measures the diversity of the generated samples, where higher values indicate greater diversity. Minimum Matching Distance (MMD) computes the average distance from each reference sample to its nearest neighbor in the generated set, serving as a measure of generation quality, with lower values being preferable. 1-Nearest Neighbor Accuracy (1-NNA) evaluates both diversity and quality, where an optimal value is achieved at \\(50\\%\\). Additionally, we compute Jensen-Shannon Divergence (JSD) to directly quantify the sim" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "1We neither used assets from Sketchfab nor obtained any from the Polycam website." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12575" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.898, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.445, + 0.908, + 0.472 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison with state-of-the-art methods. VertexRegen is able to generate meshes with comparable quality to other baselines." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.491, + 0.903, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.826, + 0.908, + 0.855 + ], + "angle": 0, + "content": "Figure 9. Qualitative results on shape-conditioned generation on meshes held out from training. The first column shows the point cloud used as the condition, followed by the generation sequence progressing from the coarsest mesh \\(\\mathcal{M}_0\\) to the final output in the last column." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12576" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "ilarity between the generated and reference distributions. We present results in Tab. 1. Our method can generate meshes of comparable quality to other state-of-the-art methods, with the added advantage of a continuous level of detail. We also show qualitative examples and their generation process in Fig. 6, and more generation results comparing with state-of-the-art methods in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.219, + 0.483, + 0.339 + ], + "angle": 0, + "content": "Generation with a face count constraint. We compare our method with the baselines under varying face count constraints. Our approach allows for generation to be paused at any point, accommodating the varying constraints. In contrast, to adapt to these constraints, baseline methods require either: (i) directly truncating the generation process, yielding incomplete intermediate meshes, or (ii) fine-tuning with an additional face count condition token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.461 + ], + "angle": 0, + "content": "For the direct truncation version, we evaluate unconditional generation metrics w.r.t. different face count constraints and present the results in Fig. 7. Additionally, we adopt a similar approach as proposed in EdgeRunner, where we split [1,800] range into 4 buckets and prepend the conditioning token to the mesh sequence. We compare VertexRegen and MeshXL fine-tuned with such conditioning scheme at face constraint 400, as shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Due to its ability to generate meshes with a continuous level of detail from coarse to fine, VertexRegen effectively captures the overall structure even with very limited face counts, yielding significantly better COV, MMD, and 1-NNA early in the process. As the face limit increases, all methods demonstrate improved results. Notably, for baselines such as EdgeRunner, while they enhance tokenization efficiency over MeshXL, they do not fundamentally address continuous level-of-detail generation and thus follow a similar trend to other baselines in the plot." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.303, + 0.635 + ], + "angle": 0, + "content": "4.3.2. Conditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.718 + ], + "angle": 0, + "content": "We show qualitative examples of shape-conditioned generation in Fig. 9. We condition on 4,096 sampled points with normals from dense meshes. VertexRegen can generate a coarse \\(\\mathcal{M}_0\\), starting as simply as a tetrahedron, and progressively refine it by generating a sequence of vertex splits." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.255, + 0.743 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tokenization efficiency. We compute the tokenized sequence length of our proposed tokenization scheme (Sec. 3.2.1) on our dataset and report the average compression ratio relative to MeshXL (9 tokens per face) in Tab. 3. For VertexRegen, there are two primary sequence types: (i) a MeshXL-style sequence for the initial coarsest mesh \\(\\mathcal{M}_0\\) and (ii) sequences encoding vertex splits. The latter requires 12 tokens for two non-boundary faces or 10 tokens per boundary face. As a result, the highest compression is achieved when \\(\\mathcal{M}_0\\) is minimized and all vertex splits occur" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.09, + 0.905, + 0.154 + ], + "angle": 0, + "content": "
@400 FacesCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
VertexRegen50.928.3151.032.88264147
MeshXL (w/ FCC)41.2010.0359.065.19308168
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.155, + 0.905, + 0.184 + ], + "angle": 0, + "content": "Table 2. Comparison between VertexRegen and MeshXL with face count condition (FCC) when face count constraint is 400." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.203, + 0.903, + 0.25 + ], + "angle": 0, + "content": "
MeshXLMeshAnything v2EdgeRunnerVertexRegen w/ HEw/o HE
Compression1.00.460.470.730.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.252, + 0.905, + 0.28 + ], + "angle": 0, + "content": "Table 3. Compression ratio of different tokenization schemes, VertexRegen with and without leveraging a half-edge (HE) structure." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.299, + 0.903, + 0.367 + ], + "angle": 0, + "content": "
Guided DecodingCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
w/o51.128.3150.753.37211120
w/51.038.2950.222.89320176
" + }, + { + "type": "table_caption", + "bbox": [ + 0.569, + 0.368, + 0.85, + 0.382 + ], + "angle": 0, + "content": "Table 4. Effects on geometry-guided decoding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.905, + 0.465 + ], + "angle": 0, + "content": "on non-boundary faces. On average, VertexRegen achieves a compression ratio of 0.73, approaching the theoretical limit of 0.67. The discrepancy arises from the overhead introduced by \\(\\mathcal{M}_0\\) (5.68% of all total tokenized length) and boundary vertex splits (3.64% of all vertex splits)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.906, + 0.557 + ], + "angle": 0, + "content": "Without the half-edge data structure, identifying which half of the ring surrounding \\( v_{s} \\) is associated with \\( v_{s} \\) or \\( v_{t} \\) in \\( \\mathcal{M}_{k + 1} \\) requires recording an additional vertex. This increases the token count to 15 per two non-boundary faces (or 13 per boundary face) and results in an average increase of \\( 22\\% \\) in tokenized sequence length." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.676 + ], + "angle": 0, + "content": "Guided decoding. In Tab. 4, we ablate the effects of geometry-constrained decoding. As one step of the vertex split operation is dependent on all the preceding operations, predicting an invalid vertex split may break the chain and end the generation prematurely. With guided decoding, the model is able to generate longer sequences with more faces." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduced VertexRegen, a novel mesh generation framework that enables continuous levels of detail through a generative process based on vertex splits. Unlike conventional auto-regressive approaches that synthesize meshes in a partial-to-complete manner, VertexRegen reinterprets mesh generation as the reversal of edge collapse, providing an effectively anytime solution to mesh generation. Our experimental results demonstrate that VertexRegen achieves competitive performance compared to state-of-the-art methods while offering the unique advantage of halting generation at any stage to obtain meshes at different levels of detail." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12577" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Billzb Wang, Jingyi Yu, Gang Yu, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems, 37:97141-97166, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[4] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.429 + ], + "angle": 0, + "content": "[5] Yun-Chun Chen, Vladimir Kim, Noam Aigerman, and Alec Jacobson. Neural progressive meshes. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[6] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G. Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4456-4465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[8] Angela Dai and Matthias Niessner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 36:35799-35813, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[10] Andreas Fabri and Sylvain Pion. Cgal: the computational geometry algorithms library. In Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, page 538-539, New York, NY, USA, 2009. Association for Computing Machinery. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[13] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[15] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[16] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[17] Hugues Hoppe. Progressive meshes. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 99–108, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[18] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.907, + 0.497 + ], + "angle": 0, + "content": "[19] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[20] Peter Lindstrom and Greg Turk. Fast and memory efficient polygonal simplification. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 279-286. IEEE, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[22] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. In Advances in Neural Information Processing Systems, pages 59314-59341. Curran Associates, Inc., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[23] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.859 + ], + "angle": 0, + "content": "[24] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12578" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[26] David Luebke, Martin Reddy, Jonathan D. Cohen, Amitabh Varshney, Benjamin Watson, and Robert Huebner. Level of Detail for 3D Graphics. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[28] Felix Mujkanovic, Ntumba Elie Nsampi, Christian Theobalt, Hans-Peter Seidel, and Thomas Leimkuhler. Neural gaussian scale-space fields. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[29] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[31] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[32] Jovan Popovic and Hugues Hoppe. Progressive simplicial complexes. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 217-224, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[33] J. Rossignac. Edgebreaker: connectivity compression for triangle meshes. IEEE Transactions on Visualization and Computer Graphics, 5(1):47-61, 1999. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[36] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[37] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "David Novotny. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. In Advances in Neural Information Processing Systems, pages 9532-9564. Curran Associates, Inc., 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11367, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[40] Danhang Tang, Mingsong Dou, Peter Lincoln, Philip Davidson, Kaiwen Guo, Jonathan Taylor, Sean Fanello, Cem Keskin, Adarsh Kowdle, Sofien Bouaziz, et al. Real-time compression and streaming of 4d performances. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.416 + ], + "angle": 0, + "content": "[41] Danhang Tang, Saurabh Singh, Philip A. Chou, Christian Hane, Mingsong Dou, Sean Fanello, Jonathan Taylor, Philip Davidson, Onur G. Guleryuz, Yinda Zhang, Shahram Izadi, Andrea Tagliasacchi, Sofien Bouaziz, and Cem Keskin. Deep implicit volume compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[43] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgrunner: Auto-regressive auto-encoder for artistic mesh generation. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[44] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[46] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[47] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 731-741, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Kevin J Weiler. Topological structures for geometric modeling (Boundary representation, manifold, radial edge structure). Rensselaer Polytechnic Institute, 1986. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12579" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[50] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, Tong Zhang, Shenghua Gao, and C.L. Philip Chen. Scaling mesh generation via compressive tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11093-11103, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[51] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[52] Haiyang Xu, Yu Lei, Zeyuan Chen, Xiang Zhang, Yue Zhao, Yilin Wang, and Zhuowen Tu. Bayesian diffusion models for 3d shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10628-10638, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[53] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.402, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.459, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[55] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[56] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[57] Qingcheng Zhao, Xiang Zhang, Haiyang Xu, Zeyuan Chen, Jianwen Xie, Yuan Gao, and Zhuowen Tu. Depr: Depth guided single-view scene reconstruction with instance-level diffusion. arXiv preprint arXiv:2507.22825, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems, pages 73969-73982. Curran Associates, Inc., 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.727, + 0.482, + 0.782 + ], + "angle": 0, + "content": "[59] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826–5835, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12580" + } + ] +] \ No newline at end of file diff --git a/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_origin.pdf b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45d64bcb743702532a9c19e03b79481d2aea0a66 --- /dev/null +++ b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/342c3a2f-65c0-4b35-b717-21b0c8a1a0ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c1ad1cc7a5e5d50b7aeea9b50c51e88ec1895a937da3153560c5318911cf91b +size 7587864 diff --git a/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/full.md b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/full.md new file mode 100644 index 0000000000000000000000000000000000000000..439d85ab76b06f61230bbd5c9574fbf75fa4a1a6 --- /dev/null +++ b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/full.md @@ -0,0 +1,322 @@ +# VertexRegen: Mesh Generation with Continuous Level of Detail + +Xiang Zhang $^{1*}$ Yawar Siddiqui $^{2}$ Armen Avetisyan $^{2}$ Chris Xie $^{2}$ Jakob Engel $^{2}$ Henry Howard-Jenkins $^{2}$ $^{1}$ UC San Diego $^{2}$ Meta Reality Labs Research + +![](images/d0420cc4dd6766a5a06076fe8de70bc226cfe6c1a1d1e21f181e313f6c11f6e2.jpg) +Figure 1. Mesh generation process of VertexRegen (coarse-to-fine process) vs. previous work (partial-to-complete process). Prior work generates meshes face by face, with the step count corresponding to the face count in the figure. In contrast, VertexRegen produces meshes with a continuously increasing level of detail, where each step yields a valid mesh $\mathcal{M}_k$ . + +# Abstract + +We introduce VertexRegen, a novel mesh generation framework that enables generation at a continuous level of detail. Existing autoregressive methods generate meshes in a partial-to-complete manner and thus intermediate steps of generation represent incomplete structures. VertexRegen takes inspiration from progressive meshes and reformulates the process as the reversal of edge collapse, i.e. vertex split, learned through a generative model. Experimental results demonstrate that VertexRegen produces meshes of comparable quality to state-of-the-art methods while uniquely offering anytime generation with the flexibility to halt at any step to yield valid meshes with varying levels of detail. + +# 1. Introduction + +Meshes are essential for 3D asset representation and are widely used in industries such as film, design, and gaming due to their compatibility with most 3D software and + +hardware. However, manually creating high-quality meshes is costly and time-consuming, prompting extensive research into automated 3D content creation. Largely, these approaches have used alternative representations such as neural fields [11, 27, 31, 38], voxels [34, 51] or point clouds [54, 59] which are later post-processed to meshes instead of direct mesh modeling. Unfortunately, these post-processed meshes often exhibit poor topology, overtessellation, and artifacts, lacking the quality of artist-crafted meshes. + +Recently, there has been a surge of approaches that directly generate meshes, using autoregressive models to represent meshes as sequences of triangles [2-4, 15, 37, 43]. These methods capture the high-fidelity and aesthetic qualities of artist-created works without the need for post-generation conversion. Although significant progress has been made in improving tokenization schemes [2, 4, 37] and network architectures [15], these approaches do not fundamentally alter a key characteristic of generation: namely, to produce a valid mesh, the full sequence must be generated to completion. Consequently, these methods offer no control over the level of detail during generation; early stopping results in a mesh with missing faces. Simple extensions, + +such as face-count conditioning tokens in EdgeRunner [43], have been proposed to introduce some coarse control over the detail by pre-specifying a target face count. However, each generation still must be completed in its entirety to yield a complete mesh, and thus a single generation still offers only a single level of detail. + +To allow for a continuous level of detail generation, we take inspiration from Hoppe's progressive mesh formulation [17]. Progressive meshes use two reversible operations to transition between levels of detail. The edge collapse operation simplifies a mesh into a coarser one by reducing one edge at a time. The vertex split operation reverses this simplification to add more detail by using information stored during edge collapse. By starting with a coarse mesh and keeping edge collapse records, this approach creates an efficient, lossless representation of the original mesh, allowing for continuous resolution adjustments by applying any number of vertex split operations. + +In this paper, we build on this progressive mesh representation by learning the vertex split, i.e. reversing the edge collapse operation, as a generative problem. This allows the resulting mesh generation to inherit the properties of progressive meshes and provides a solution for anytime mesh generation, where the process can be stopped early to yield a coarser mesh rather than an incomplete one. By properly serializing the vertex split sequence, the entire generation process can be modeled using a Transformer trained with a next-token prediction objective, a widely adopted paradigm in mesh generation. We evaluate VertexRegen on the task of unconditional mesh generation, demonstrating superior results both qualitatively and quantitatively. We further provide examples of shape-conditioned generation. These experimental results illustrate the ability of our method to generate compelling meshes at a continuous resolution. + +Our contributions can be summarized as follows: + +- Inspired by Hoppe's progressive meshes [17], we reframe mesh generation as the reversal of edge collapse operations, i.e. generating vertex splits. +- We formulate a token-efficient parameterization of a progressive mesh, through a half-edge data structure. +- We propose VertexRegen for continuous level of detail mesh generation. VertexRegen creates meshes in a coarse-to-fine fashion, rather than partial-to-complete, uniquely providing a solution for anytime generation. + +# 2. Related Work + +3D Mesh Generation. Recent advancements in 3D shape generation have explored various representations, including point-clouds [30, 44, 52, 59], signed distance functions (SDFs) [7, 19, 22, 23, 36, 53, 57], neural radiance fields (NeRFs) [16, 18, 46, 55], and Gaussian splatting [14, 42]. These implicit representations require iso-surface extraction techniques [24, 35, 47] to output meshes, often resulting + +in over-tessellated and excessively smooth outputs, which pose challenges for downstream applications such as geometric processing and manipulation. + +In contrast, direct mesh generation inherently produces structured, well-defined geometry without the need for postprocessing or surface extraction, making it an increasingly prominent approach in recent years. Early methods tackle this task by generating meshes from surface patches [13], deforming ellipsoids [45], predicting mesh graphs [8], or employing binary space partitioning [6]. More recent techniques leverage generative models, particularly diffusion models and sequence-based approaches. PolyDiff [1] applies discrete diffusion, while PolyGen [29] autoregressively predicts vertices and faces using two separate networks. The sequence modeling paradigm has been further refined by representing the entire mesh as a single sequence. MeshGPT [37] introduces a tokenization scheme based on a vector quantized variational auto-encoder (VQVAE), while MeshXL [2] directly models discretized triangle soup sequences without compression. MeshAnything [4] extends MeshGPT by incorporating a point-cloud encoder for shape-conditioned generation. PivotMesh [49] introduces a hierarchical approach, generating pivot vertices before producing the full mesh. Further research has focused on optimizing tokenization and sequence modeling. Adjacency-aware compression techniques [3, 43, 50] improve tokenization efficiency and enable higher face counts within fixed context windows, while Meshtron [15] leverages Hourglass Transformers and sliding window attention to scale MeshXL sequences more effectively. + +Despite these advancements, most existing approaches follow a partial-to-complete paradigm, where mesh regions are constructed sequentially at a uniform level of detail. Our work instead adopts a coarse-to-fine approach, continuously increasing the level of detail as new tokens are generated, achieving better control over geometric complexity. + +Level of Detail Representations. Level of Detail [26] (LOD) is a widely used technique in computer graphics to optimize rendering performance by reducing the complexity of 3D models based on their size, distance from the camera, or importance in a scene. Different LOD strategies depend on the underlying representation of 3D shapes. For meshes, progressive meshes [17] generate a mesh sequence starting from a coarse base model, gradually refining it through a series of transformations that incrementally add detail. Traditional mesh simplification methods [12, 20] can be adopted to construct such sequences by iteratively reducing polygon count while preserving geometric fidelity. Progressive simplicial complexes [32] extend progressive meshes to handle arbitrary meshes, including non-manifold and non-orientable surfaces. + +Recent studies have incorporated neural networks for + +![](images/6d5a953d4abc6248f36699478fa034c38125043fcbb81d4ce61d00df28cee9f2.jpg) +Figure 2. Illustration of edge collapse and its inverse operation, i.e. vertex split. During edge collapse, vertex $v_{t}$ is collapsed into $v_{s}$ , resulting in two degenerate triangles (shaded in yellow), which forms two new edges in the result mesh. + +LOD representations. Neural progressive meshes [5] propose a learned approach supporting LOD with a subdivision-based encoder-decoder. Similar representations have been explored for reconstruction approaches with signed distance functions (SDFs) [39-41] and neural fields [28]. However, these approaches are not generative; they typically encode and decode existing meshes [5] or optimize the LOD representation per shape [39]. In contrast, our approach is purely generative. VertexRegen learns to create meshes from scratch, progressively adding detail through autoregressive sequence generation. This allows us to generate new, high-fidelity meshes without relying on pre-existing structures, distinguishing our method from both traditional and neural LOD techniques. + +# 3. Method + +# 3.1. Progressive Meshes Overview + +A progressive mesh [17] (PM) proposes an efficient and lossless continuous-resolution representation for arbitrary triangle meshes. The representation is built off two observations: (i) that a single mesh transformation, edge collapse, is sufficient for effective simplification of meshes; (ii) that edge collapse transformations are invertible via a vertex split operation. In the following, we provide details of the edge collapse operation and its inversion with vertex splits, each illustrated in Fig. 2. + +The edge collapse operation, $\operatorname{ecol}(v_s, v_t)$ , unifies two adjacent vertices $v_s$ and $v_t$ into a single vertex $v_s$ . The operation results in the vanishing of two faces, $\{v_t, v_s, v_l\}$ and $\{v_s, v_t, v_r\}$ , as well as the vertex $v_t$ . In the general case, a new position for $v_s$ is also specified; however, for the case of half-edge collapse, the original position for $v_s$ is kept. + +An initial mesh, $\mathcal{M}$ , can be simplified into a more coarse mesh, $\mathcal{M}_n$ , by $n$ successive edge collapse operations: + +$$ +\mathcal {M} _ {0} \xleftarrow {\mathrm {e c o l} _ {0}} \mathcal {M} _ {1} \xleftarrow {\mathrm {e c o l} _ {1}} \dots \xleftarrow {\mathrm {e c o l} _ {n - 1}} \mathcal {M} +$$ + +The order in which the edge collapse operations are performed is determined such that each successive edge collapse results in the minimum increase in Quadratic Error Metrics (QEM) [12] with respect to the original mesh. This + +process can be repeated until there are no candidate edge collapses, e.g. without flipping face normals. + +A vertex split, $\mathrm{vsplit}(v_s,v_l,v_r,v_t)$ , defines the inverse of edge collapse. It restores $v_{t}$ and the two faces, $\{v_{t},v_{s},v_{l}\}$ and $\{v_{s},v_{t},v_{r}\}$ , which vanished during edge collapse, as well as $v_{s}$ to its original position. In case $(v_{s},v_{t})$ is a boundary edge, either $v_{l}$ or $v_{r}$ will not exist, and there is only one face restored during the vertex split. + +The combination of edge collapse and its inverse vertex splits enables the progressive mesh representation, $\mathrm{PM}(\mathcal{M}) = (\mathcal{M}_0,\{\mathrm{vsplit}_0,\dots ,\mathrm{vsplit}_{n - 1}\})$ to express an arbitrary triangle mesh, $\mathcal{M}$ , as a combination of a coarse mesh, $\mathcal{M}_0$ , obtained through edge collapse, and the sequence of $n$ vertex split records required to reverse them: + +$$ +\mathcal {M} _ {0} \xrightarrow {\mathrm {v s p l i t} _ {0}} \mathcal {M} _ {1} \xrightarrow {\mathrm {v s p l i t} _ {1}} \dots \xrightarrow {\mathrm {v s p l i t} _ {n - 1}} \mathcal {M} +$$ + +# 3.2. VertexRegen + +Traditional progressive meshes require starting an initial detailed mesh to be simplified into a coarse mesh, $\mathcal{M}_0$ , and recording each of these collapse steps as vertex split records to form the representation. We observe that the formation of the PM representation serves effectively as forward (edge collapse) and reverse (vertex split) processes that enable transition between detailed and coarse meshes. + +VertexRegen frames the creation of a detailed mesh as the generation of a PM representation. Analogous to denoising for diffusion models, we train a generative model to reverse edge collapse. Concretely, VertexRegen first generates a coarse mesh, $\mathcal{M}_0$ , from scratch, before increasing the level of detail through the generation of vertex split records. + +# 3.2.1. Progressive Mesh Parameterization + +We frame generation as an autoregressive sequential modeling task, with the full mesh sequence taking the form: + +$$ +\begin{array}{c c} \text {M : [ < b o s > , [ M _ {-} 0 s e q u e n c e ] , < s e p > ,} & \# \mathcal {M} _ {0} \\ \text {[ v s p l i t \_ 0 ] , \ldots , [ v s p l i t \_ n - 1 ] , < e o s > ]} & \# \text {v s p l i t s} \end{array} +$$ + +In the following, we will define the formation of the subsequences [M_0 sequence] and [vsplit]. + +$\mathcal{M}_0$ : Coarse Mesh Tokenization. For the initial coarse mesh, we follow the tokenization scheme defined in MeshXL [2]. In this formulation, embeddings are learned for discretized coordinates in an $N^3$ grid. A vertex is represented via the sequential look-up of $x$ -value, $y$ -value, and $z$ -value. A face is then constructed as the concatenation of its 3 vertices, totaling 9 tokens. The full mesh sequence is then defined as the concatenation of its constituent faces. + +$$ +v: \left[ < x >, < y >, < z > \right], \quad \# v e r t e x +$$ + +$$ +F: \left[ \begin{array}{l} [ v _ {- 1} ], [ v _ {- 2} ], [ v _ {- 3} ] \end{array} \right], \quad \# \text {f a c e} +$$ + +$$ +M: \left[ [ F _ {- 1} ], [ F _ {- 2} ], \dots , [ F _ {- N} ] \right] \quad \# \text {m e s h} +$$ + +![](images/2ef934b6144eba8d2af06d3526da99653091dac78c54247047ff6a896829d209.jpg) +Figure 3. Face count distribution of the coarsest mesh $\mathcal{M}_0$ and original mesh $\mathcal{M}$ . The coarsest meshes $(\mathcal{M}_0)$ contain significantly fewer faces than original $(\mathcal{M})$ , with an average of 18 and 457 faces, respectively. + +Following [37], we sort vertices in $z - y - x$ order (lower to higher). Within each face, vertices are cylindrically permuted to have the lowest-indexed vertex be first. + +Although MeshXL [2] leverages this tokenization scheme to produce full detailed meshes, in VertexRegen only the coarsest level of the mesh is parameterized in this way. These coarse meshes consist of substantially fewer faces than the full detailed mesh, as demonstrated in Fig. 3, and the resulting coarse mesh only accounts for $5.68\%$ of the total sequence length on average. + +Vertex Split Generation. Consider the $k$ -th vertex split operation $\mathrm{vsplit}_k$ , which converts a lower-detailed mesh $\mathcal{M}_k$ into a higher-detailed mesh $\mathcal{M}_{k+1}$ . We denote the neighbor vertices of vertex $v$ in $\mathcal{M}_k$ as $N_k(v)$ . After determining the target vertex $v_s$ to split (in $\mathcal{M}_k$ ) and the new vertex $v_t$ (in $\mathcal{M}_{k+1}$ ), we need to obtain $N_{k+1}(v_s)$ and $N_{k+1}(v_t)$ from the vertex split record to connect $v_s$ and $v_t$ with correct neighbor vertices in $\mathcal{M}_{k+1}$ , which are prohibitive to generate as the number of neighbor vertices $|N_{k+1}(v_s)|$ and $|N_{k+1}(v_t)|$ is often large. + +However, as mesh $\mathcal{M}_k$ is generated from $\mathcal{M}_{k + 1}$ during edge collapse by merging $v_{t}$ into $v_{s}$ , we have + +$$ +N _ {k + 1} \left(v _ {s}\right) \cap N _ {k + 1} \left(v _ {t}\right) = \left\{v _ {l}, v _ {r} \right\} \tag {1} +$$ + +$$ +N _ {k + 1} \left(v _ {s}\right) \cup N _ {k + 1} \left(v _ {t}\right) - \left\{v _ {s}, v _ {t} \right\} = N _ {k} \left(v _ {s}\right) \tag {2} +$$ + +which indicates we will only need to record vertex $v_{l}$ and $v_{r}$ . They split the vertex ring surrounding $v_{s}$ in $\mathcal{M}_k$ , where two halves of the vertices on the ring (top and bottom vertices in Fig. 2) belong to $v_{t}$ and $v_{s}$ respectively in $\mathcal{M}_{k + 1}$ . However, an ambiguity arises in determining which half of the ring corresponds to $v_{t}$ 's neighbors in $\mathcal{M}_{k + 1}$ . In the following, we demonstrate how the half-edge data structure [48] can be utilized to resolve this ambiguity. + +Denote $\mathcal{H}_{ij}^k$ as a half-edge pointing from $v_{i}$ to $v_{j}$ , with $v_{k}$ being the third vertex in the associated face. As shown in + +![](images/d020ef834e9f52cbf4765977de5065d6732f60ef43375c9c10b203dce4fe5895.jpg) +Figure 4. Illustration of half-edge data structure and the traversal process to determine the neighbors of $v_{s}$ and $v_{t}$ in $\mathcal{M}_{k + 1}$ . Starting from half-edge $\mathcal{H}_{ls}^{1}$ , we traverse the faces in clockwise direction until $\mathcal{H}_{sr}^{2}$ , where we collect the vertices $v_{1}$ and $v_{2}$ into $N_{k + 1}(v_s)$ . + +![](images/819156d72a3c6e7f194cbad6db72f5450cd1b639e5afff7774fb590878e99a53.jpg) + +![](images/39b44188626f66fd98150f202dd5e908f6498201f422e5fca196bb4b331cd831.jpg) +Figure 5. Illustration of VertexRegen tokenization. The sequence begins with base mesh $\mathcal{M}_0$ , followed by vertex split subsequences. A special token indicates either $v_{l}$ or $v_{r}$ does not exist. + +Fig. 4, we begin the traversal from $\mathcal{H}_{ls}^{1}$ and proceed to the twin of the next half-edge, $\mathcal{H}_{1s}^{2}$ . This operation is repeated until we reach $\mathcal{H}_{,s}^{r}$ . Throughout this process, we traverse the faces above $(v_{l}, v_{s})$ and $(v_{s}, v_{r})$ clockwise, obtaining: + +$$ +\left\{v _ {k} \mid \mathcal {H} _ {\cdot s} ^ {k}, k \neq r \right\} = N _ {k + 1} \left(v _ {s}\right) - \left\{v _ {l}, v _ {r}, v _ {t} \right\} \tag {3} +$$ + +where $\mathcal{H}_{s}^{k}$ is the half-edge encountered during traversal. + +In conjunction with Eqs. (1) and (2), $N_{k + 1}(v_t)$ can also be determined. When $(v_s, v_t)$ is a boundary in $\mathcal{M}_{k + 1}$ , we follow either $\mathcal{H}_{ls}$ clockwise or $\mathcal{H}_{sr}$ counterclockwise until we reach a half-edge of the boundary. + +After identifying the neighbors of $v_{s}$ and $v_{t}$ in the mesh $\mathcal{M}_{k + 1}$ , we add the new vertex $v_{t}$ to the mesh $\mathcal{M}_k$ and reconnect the vertices in $N_{k + 1}(v_t)$ to $v_{t}$ . Finally, we restore two faces, $\{v_s,v_l,v_t\}$ and $\{v_r,v_s,v_t\}$ , which are associated with the half-edges $\mathcal{H}_{sl}^{t}$ and $\mathcal{H}_{rs}^{t}$ , respectively. Using the half-edge data structure, the orientation of the newly created faces remains consistent with the rest of the mesh. When $(v_{s},v_{t})$ lies on the boundary, only one face is restored. + +Hence, with the above half-edge formulation, each vertex split operation can be uniquely defined by the selection of the target vertex $v_{s}$ and two neighbors $v_{l}$ and $v_{r}$ , as well as the position to place the new vertex $v_{t}$ . + +Vertex Split Tokenization. Although $v_{s}, v_{l}$ and $v_{r}$ only require references to existing vertices in the mesh, in practice we implement this reference as a raw prediction of each vertex to avoid a vocabulary size proportional to the sequence length. Thus, in the majority of cases, each vertex + +![](images/a3ec9d00c549e222e74e59e34f64589d85b2bb278fa23b478001ff0b965514e4.jpg) + +![](images/838b5a3a02951f2826ef225c104dbe6f899341e017da11c477342322afb22898.jpg) + +![](images/31f0bcb4672f8d3a55fb57b66296abf470de15be0171a09c39608e19cfa1081e.jpg) + +![](images/db8404fafb7a72ff0cfd61a46683118a25a8bfb9e0f3a77fe6d6a38c6aca455a.jpg) + +![](images/1c90d14a5746db392b51a47595c6cb1863a98ac2404d890fba2f27cdc3b30b58.jpg) +Figure 6. Generation process for VertexRegen. $\mathcal{M}_0$ represents the predicted initial coarsest mesh, followed by subsequent meshes generated through the predicted vertex split sequence. + +split is represented by a subsequence of 12 tokens (or 10 tokens when $(v_{s}, v_{t})$ is a boundary edge): + +```c +vsplit: [, , , #v_s + +$< 1_{-}x>, < 1_{-}y>, < 1_{-}z> |< \mathrm{nil}>$ , # $v_{l}$ + +$< r_{-}x>, < r_{-}y>, < r_{-}z|< nil>$ ， #vr + +$< t\_ x > , < t\_ y > , < t\_ z>$ #vt + +where the special token $\langle \mathrm{nil} \rangle$ signifies $v_{l}$ or $v_{r}$ does not exist, and only one of $v_{l}$ and $v_{r}$ is allowed to be $\langle \mathrm{nil} \rangle$ for a valid vertex split subsequence. In Fig. 5 we illustrate the entire tokenization process. + +The above progressive mesh serialization can then be trained with the standard next-token prediction target. + +Vertex Split Decoding. The vertex split sequence must be consistent with the actual geometry to be valid. We + +maintain a state machine and perform vertex split on the fly as new tokens are generated. After the coarsest mesh $\mathcal{M}_0$ is generated, we initialize the state. During each generation step, we enforce $v_{s}$ to be a vertex in the current mesh $\mathcal{M}_k$ , $(v_{s}, v_{l})$ , $(v_{s}, v_{r})$ are valid edges (if $v_{l}$ and $v_{r}$ are not $\langle \mathrm{nil} \rangle$ ), and only one of $v_{l}$ and $v_{r}$ can be $\langle \mathrm{nil} \rangle$ . Lastly, we decode $v_{t}$ , and perform vertex split with generated $v_{s}, v_{l}, v_{r}, v_{t}$ information. + +# 3.3. Conditional Generation + +We consider generation conditioned on shapes. We adopt a pre-trained point-cloud encoder [58]. A LLaVa-style [21] projector is leveraged to project the condition features to the token embedding space. The projected feature tokens are prepended to the mesh sequence as the prefix. We supervise training with the next-token prediction objective while masking the loss for the prefix tokens. + +
MethodTokenizationCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)
MeshXL [2]Flattened Coords.51.768.3050.843.81
MeshAnything V2 [3]AMT50.338.5052.254.84
EdgeRunner [43]EdgeBreaker [33]51.397.8149.443.22
VertexRegen (Ours)Progressive51.038.2950.222.89
+ +Table 1. Quantitative comparisons with state-of-the-art methods for unconditional mesh generation. Best results are bolded, second best are underlined. VertexRegen can generate meshes with comparable quality while enjoying the benefits of continuous level of detail. + +![](images/da9816eabdcbcd81189ff45c87dadff60fdf83e77fdf2d55f01b0477c27c1b85.jpg) +Figure 7. Unconditional generation under face count constraints. VertexRegen achieves significantly better COV, MMD, and 1-NNA in early stages of generation. + +# 4. Experiments + +# 4.1. Datasets + +We pre-train our model and the baselines presented using two primary sources of meshes: Objaverse-XL [9] and an additional set of licensed artist-created meshes. Initially, we select meshes containing fewer than 8,000 faces without applying decimation. We use the CGAL [10] library to implement edge collapse and vertex split operations, where we modify the vertex placement to either $v_{s}$ or $v_{t}$ . We filter out non-manifold meshes and those that cannot be processed, which results in a final dataset of approximately 1.5M meshes with an average of 457 faces. For unconditional generation evaluation, we construct a high-quality subset by further filtering with the alignment split of Objaverse-XL and with fewer than 800 faces. This yields approximately 18k meshes. We adopt a similar filtering process for shape-conditioned experiments. + +We apply robust data augmentations to input meshes, including random shift within range $[-0.1, 0.1]$ , random scaling between [0.9, 1.1], random rotation by $0^{\circ}$ , $90^{\circ}$ , $180^{\circ}$ , or $270^{\circ}$ . During pre-training, we handle sequences exceeding the context window by discarding vertex split subsequences, while for other baselines, we truncate the tokenized sequence directly. The set for quantitative evaluation is chosen to avoid truncation for any method evaluated. + +# 4.2. Implementation Details + +VertexRegen and all other baselines are built upon the pretrained OPT-350M [56], with newly initialized token embeddings and position embeddings. We train our model on 64 H100 GPUs for approximately four days, with an effective batch size of 256. We use the AdamW [25] optimizer with a weight decay of 0.1, and betas (0.9, 0.95). A cosine scheduler is employed, starting with an initial learning rate of $1 \times 10^{4}$ and gradually decreasing to $5 \times 10^{6}$ . We clip the gradient norm to 1.0. During inference, we adopt the top- $p$ sampling strategy with $p = 0.95$ by default. + +# 4.3. Results + +# 4.3.1. Unconditional Generation + +We follow evaluation protocols from prior works [1, 2, 37], employing point-cloud-based metrics to assess unconditional generation. Specifically, we sample the same number of meshes as the evaluation dataset and randomly sample 2,048 points per mesh. Coverage (COV) measures the diversity of the generated samples, where higher values indicate greater diversity. Minimum Matching Distance (MMD) computes the average distance from each reference sample to its nearest neighbor in the generated set, serving as a measure of generation quality, with lower values being preferable. 1-Nearest Neighbor Accuracy (1-NNA) evaluates both diversity and quality, where an optimal value is achieved at $50\%$ . Additionally, we compute Jensen-Shannon Divergence (JSD) to directly quantify the sim + +![](images/7c4f79cf3f552eb86c3ded40706d7fa9a2e5f6189bc60a14e92d97d62ee9bfae.jpg) +Figure 8. Qualitative comparison with state-of-the-art methods. VertexRegen is able to generate meshes with comparable quality to other baselines. + +![](images/a80fd6086d5a71c985ca3b6d7d8b5efa4b9a0c2973eb713545b57244e96d0c34.jpg) +Figure 9. Qualitative results on shape-conditioned generation on meshes held out from training. The first column shows the point cloud used as the condition, followed by the generation sequence progressing from the coarsest mesh $\mathcal{M}_0$ to the final output in the last column. + +ilarity between the generated and reference distributions. We present results in Tab. 1. Our method can generate meshes of comparable quality to other state-of-the-art methods, with the added advantage of a continuous level of detail. We also show qualitative examples and their generation process in Fig. 6, and more generation results comparing with state-of-the-art methods in Fig. 8. + +Generation with a face count constraint. We compare our method with the baselines under varying face count constraints. Our approach allows for generation to be paused at any point, accommodating the varying constraints. In contrast, to adapt to these constraints, baseline methods require either: (i) directly truncating the generation process, yielding incomplete intermediate meshes, or (ii) fine-tuning with an additional face count condition token. + +For the direct truncation version, we evaluate unconditional generation metrics w.r.t. different face count constraints and present the results in Fig. 7. Additionally, we adopt a similar approach as proposed in EdgeRunner, where we split [1,800] range into 4 buckets and prepend the conditioning token to the mesh sequence. We compare VertexRegen and MeshXL fine-tuned with such conditioning scheme at face constraint 400, as shown in Tab. 2. + +Due to its ability to generate meshes with a continuous level of detail from coarse to fine, VertexRegen effectively captures the overall structure even with very limited face counts, yielding significantly better COV, MMD, and 1-NNA early in the process. As the face limit increases, all methods demonstrate improved results. Notably, for baselines such as EdgeRunner, while they enhance tokenization efficiency over MeshXL, they do not fundamentally address continuous level-of-detail generation and thus follow a similar trend to other baselines in the plot. + +# 4.3.2. Conditional Generation + +We show qualitative examples of shape-conditioned generation in Fig. 9. We condition on 4,096 sampled points with normals from dense meshes. VertexRegen can generate a coarse $\mathcal{M}_0$ , starting as simply as a tetrahedron, and progressively refine it by generating a sequence of vertex splits. + +# 4.4. Ablation Studies + +Tokenization efficiency. We compute the tokenized sequence length of our proposed tokenization scheme (Sec. 3.2.1) on our dataset and report the average compression ratio relative to MeshXL (9 tokens per face) in Tab. 3. For VertexRegen, there are two primary sequence types: (i) a MeshXL-style sequence for the initial coarsest mesh $\mathcal{M}_0$ and (ii) sequences encoding vertex splits. The latter requires 12 tokens for two non-boundary faces or 10 tokens per boundary face. As a result, the highest compression is achieved when $\mathcal{M}_0$ is minimized and all vertex splits occur + +
@400 FacesCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
VertexRegen50.928.3151.032.88264147
MeshXL (w/ FCC)41.2010.0359.065.19308168
+ +Table 2. Comparison between VertexRegen and MeshXL with face count condition (FCC) when face count constraint is 400. + +
MeshXLMeshAnything v2EdgeRunnerVertexRegen w/ HEw/o HE
Compression1.00.460.470.730.89
+ +Table 3. Compression ratio of different tokenization schemes, VertexRegen with and without leveraging a half-edge (HE) structure. + +
Guided DecodingCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
w/o51.128.3150.753.37211120
w/51.038.2950.222.89320176
+ +Table 4. Effects on geometry-guided decoding. + +on non-boundary faces. On average, VertexRegen achieves a compression ratio of 0.73, approaching the theoretical limit of 0.67. The discrepancy arises from the overhead introduced by $\mathcal{M}_0$ (5.68% of all total tokenized length) and boundary vertex splits (3.64% of all vertex splits). + +Without the half-edge data structure, identifying which half of the ring surrounding $v_{s}$ is associated with $v_{s}$ or $v_{t}$ in $\mathcal{M}_{k + 1}$ requires recording an additional vertex. This increases the token count to 15 per two non-boundary faces (or 13 per boundary face) and results in an average increase of $22\%$ in tokenized sequence length. + +Guided decoding. In Tab. 4, we ablate the effects of geometry-constrained decoding. As one step of the vertex split operation is dependent on all the preceding operations, predicting an invalid vertex split may break the chain and end the generation prematurely. With guided decoding, the model is able to generate longer sequences with more faces. + +# 5. Conclusion + +In this work, we introduced VertexRegen, a novel mesh generation framework that enables continuous levels of detail through a generative process based on vertex splits. Unlike conventional auto-regressive approaches that synthesize meshes in a partial-to-complete manner, VertexRegen reinterprets mesh generation as the reversal of edge collapse, providing an effectively anytime solution to mesh generation. Our experimental results demonstrate that VertexRegen achieves competitive performance compared to state-of-the-art methods while offering the unique advantage of halting generation at any stage to obtain meshes at different levels of detail. + +# References + +[1] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2, 6 +[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Billzb Wang, Jingyi Yu, Gang Yu, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems, 37:97141-97166, 2024. 1, 2, 3, 4, 6 +[3] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 6 +[4] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2 +[5] Yun-Chun Chen, Vladimir Kim, Noam Aigerman, and Alec Jacobson. Neural progressive meshes. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3 +[6] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G. Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4456-4465, 2023. 2 +[8] Angela Dai and Matthias Niessner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36:35799-35813, 2023. 6 +[10] Andreas Fabri and Sylvain Pion. Cgal: the computational geometry algorithms library. In Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, page 538-539, New York, NY, USA, 2009. Association for Computing Machinery. 6 +[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1 +[12] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2, 3 + +[13] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2 +[15] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 1, 2 +[16] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, 2024. 2 +[17] Hugues Hoppe. Progressive meshes. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 99–108, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3 +[18] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2 +[19] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2 +[20] Peter Lindstrom and Greg Turk. Fast and memory efficient polygonal simplification. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 279-286. IEEE, 1998. 2 +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5 +[22] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. In Advances in Neural Information Processing Systems, pages 59314-59341. Curran Associates, Inc., 2024. 2 +[23] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2 +[24] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 2 +[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 + +[26] David Luebke, Martin Reddy, Jonathan D. Cohen, Amitabh Varshney, Benjamin Watson, and Robert Huebner. Level of Detail for 3D Graphics. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 2002. 2 +[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[28] Felix Mujkanovic, Ntumba Elie Nsampi, Christian Theobalt, Hans-Peter Seidel, and Thomas Leimkuhler. Neural gaussian scale-space fields. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 3 +[29] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 2 +[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2 +[31] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[32] Jovan Popovic and Hugues Hoppe. Progressive simplicial complexes. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 217-224, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2 +[33] J. Rossignac. Edgebreaker: connectivity compression for triangle meshes. IEEE Transactions on Visualization and Computer Graphics, 5(1):47-61, 1999. 6 +[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 1 +[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2 +[36] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2 +[37] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2, 4, 6 +[38] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and + +David Novotny. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. In Advances in Neural Information Processing Systems, pages 9532-9564. Curran Associates, Inc., 2024. 1 +[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11367, 2021. 3 +[40] Danhang Tang, Mingsong Dou, Peter Lincoln, Philip Davidson, Kaiwen Guo, Jonathan Taylor, Sean Fanello, Cem Keskin, Adarsh Kowdle, Sofien Bouaziz, et al. Real-time compression and streaming of 4d performances. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. +[41] Danhang Tang, Saurabh Singh, Philip A. Chou, Christian Hane, Mingsong Dou, Sean Fanello, Jonathan Taylor, Philip Davidson, Onur G. Guleryuz, Yinda Zhang, Shahram Izadi, Andrea Tagliasacchi, Sofien Bouaziz, and Cem Keskin. Deep implicit volume compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2 +[43] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgrunner: Auto-regressive auto-encoder for artistic mesh generation. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 6 +[44] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2 +[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2 +[46] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2 +[47] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 731-741, 2025. 2 +[48] Kevin J Weiler. Topological structures for geometric modeling (Boundary representation, manifold, radial edge structure). Rensselaer Polytechnic Institute, 1986. 4 +[49] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 2 + +[50] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, Tong Zhang, Shenghua Gao, and C.L. Philip Chen. Scaling mesh generation via compressive tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11093-11103, 2025. 2 +[51] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1 +[52] Haiyang Xu, Yu Lei, Zeyuan Chen, Xiang Zhang, Yue Zhao, Yilin Wang, and Zhuowen Tu. Bayesian diffusion models for 3d shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10628-10638, 2024. 2 +[53] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2 +[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems, 2022. 1 +[55] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2 +[56] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6 +[57] Qingcheng Zhao, Xiang Zhang, Haiyang Xu, Zeyuan Chen, Jianwen Xie, Yuan Gao, and Zhuowen Tu. Depr: Depth guided single-view scene reconstruction with instance-level diffusion. arXiv preprint arXiv:2507.22825, 2025. 2 +[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems, pages 73969-73982. Curran Associates, Inc., 2023. 5 +[59] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826–5835, 2021. 1, 2 \ No newline at end of file diff --git a/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/images.zip b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4a623e0d78b4abad03833270802cef1ca8a8a3bf --- /dev/null +++ b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66fa7a93de783b041c57d7913b3a52e2046923980cc8b3c2e59354372430fee3 +size 720208 diff --git a/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/layout.json b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe1b8c87b0d398f0451df386af91d9fc1cb1b435 --- /dev/null +++ b/2025/VertexRegen_ Mesh Generation with Continuous Level of Detail/layout.json @@ -0,0 +1,10400 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "text", + "content": "VertexRegen: Mesh Generation with Continuous Level of Detail" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "spans": [ + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": "Xiang Zhang" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": " Yawar Siddiqui" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": " Armen Avetisyan" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": " Chris Xie" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": " Jakob Engel" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": " Henry Howard-Jenkins" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": "UC San Diego " + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 474, + 186 + ], + "type": "text", + "content": "Meta Reality Labs Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 198, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 61, + 198, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 61, + 198, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 545, + 383 + ], + "type": "image", + "image_path": "d0420cc4dd6766a5a06076fe8de70bc226cfe6c1a1d1e21f181e313f6c11f6e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 391, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 54, + 391, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 391, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Mesh generation process of VertexRegen (coarse-to-fine process) vs. previous work (partial-to-complete process). Prior work generates meshes face by face, with the step count corresponding to the face count in the figure. In contrast, VertexRegen produces meshes with a continuously increasing level of detail, where each step yields a valid mesh " + }, + { + "bbox": [ + 54, + 391, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 54, + 391, + 555, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 437, + 201, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 437, + 201, + 449 + ], + "spans": [ + { + "bbox": [ + 151, + 437, + 201, + 449 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 462, + 297, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 297, + 606 + ], + "type": "text", + "content": "We introduce VertexRegen, a novel mesh generation framework that enables generation at a continuous level of detail. Existing autoregressive methods generate meshes in a partial-to-complete manner and thus intermediate steps of generation represent incomplete structures. VertexRegen takes inspiration from progressive meshes and reformulates the process as the reversal of edge collapse, i.e. vertex split, learned through a generative model. Experimental results demonstrate that VertexRegen produces meshes of comparable quality to state-of-the-art methods while uniquely offering anytime generation with the flexibility to halt at any step to yield valid meshes with varying levels of detail." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 649, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 296, + 685 + ], + "type": "text", + "content": "Meshes are essential for 3D asset representation and are widely used in industries such as film, design, and gaming due to their compatibility with most 3D software and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 556 + ], + "type": "text", + "content": "hardware. However, manually creating high-quality meshes is costly and time-consuming, prompting extensive research into automated 3D content creation. Largely, these approaches have used alternative representations such as neural fields [11, 27, 31, 38], voxels [34, 51] or point clouds [54, 59] which are later post-processed to meshes instead of direct mesh modeling. Unfortunately, these post-processed meshes often exhibit poor topology, overtessellation, and artifacts, lacking the quality of artist-crafted meshes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Recently, there has been a surge of approaches that directly generate meshes, using autoregressive models to represent meshes as sequences of triangles [2-4, 15, 37, 43]. These methods capture the high-fidelity and aesthetic qualities of artist-created works without the need for post-generation conversion. Although significant progress has been made in improving tokenization schemes [2, 4, 37] and network architectures [15], these approaches do not fundamentally alter a key characteristic of generation: namely, to produce a valid mesh, the full sequence must be generated to completion. Consequently, these methods offer no control over the level of detail during generation; early stopping results in a mesh with missing faces. Simple extensions," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 693, + 259, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 693, + 259, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 693, + 259, + 713 + ], + "type": "text", + "content": "*Work conducted while the author was an intern at Meta. \nProject page: https://vertexregen.github.io" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12570" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "such as face-count conditioning tokens in EdgeRunner [43], have been proposed to introduce some coarse control over the detail by pre-specifying a target face count. However, each generation still must be completed in its entirety to yield a complete mesh, and thus a single generation still offers only a single level of detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 287 + ], + "type": "text", + "content": "To allow for a continuous level of detail generation, we take inspiration from Hoppe's progressive mesh formulation [17]. Progressive meshes use two reversible operations to transition between levels of detail. The edge collapse operation simplifies a mesh into a coarser one by reducing one edge at a time. The vertex split operation reverses this simplification to add more detail by using information stored during edge collapse. By starting with a coarse mesh and keeping edge collapse records, this approach creates an efficient, lossless representation of the original mesh, allowing for continuous resolution adjustments by applying any number of vertex split operations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 478 + ], + "type": "text", + "content": "In this paper, we build on this progressive mesh representation by learning the vertex split, i.e. reversing the edge collapse operation, as a generative problem. This allows the resulting mesh generation to inherit the properties of progressive meshes and provides a solution for anytime mesh generation, where the process can be stopped early to yield a coarser mesh rather than an incomplete one. By properly serializing the vertex split sequence, the entire generation process can be modeled using a Transformer trained with a next-token prediction objective, a widely adopted paradigm in mesh generation. We evaluate VertexRegen on the task of unconditional mesh generation, demonstrating superior results both qualitatively and quantitatively. We further provide examples of shape-conditioned generation. These experimental results illustrate the ability of our method to generate compelling meshes at a continuous resolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 479, + 266, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 479, + 266, + 490 + ], + "spans": [ + { + "bbox": [ + 67, + 479, + 266, + 490 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 492, + 295, + 600 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 492, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 295, + 528 + ], + "type": "text", + "content": "- Inspired by Hoppe's progressive meshes [17], we reframe mesh generation as the reversal of edge collapse operations, i.e. generating vertex splits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 529, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 295, + 552 + ], + "type": "text", + "content": "- We formulate a token-efficient parameterization of a progressive mesh, through a half-edge data structure." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 552, + 295, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 552, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 552, + 295, + 600 + ], + "type": "text", + "content": "- We propose VertexRegen for continuous level of detail mesh generation. VertexRegen creates meshes in a coarse-to-fine fashion, rather than partial-to-complete, uniquely providing a solution for anytime generation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "3D Mesh Generation. Recent advancements in 3D shape generation have explored various representations, including point-clouds [30, 44, 52, 59], signed distance functions (SDFs) [7, 19, 22, 23, 36, 53, 57], neural radiance fields (NeRFs) [16, 18, 46, 55], and Gaussian splatting [14, 42]. These implicit representations require iso-surface extraction techniques [24, 35, 47] to output meshes, often resulting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "in over-tessellated and excessively smooth outputs, which pose challenges for downstream applications such as geometric processing and manipulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 109, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 430 + ], + "type": "text", + "content": "In contrast, direct mesh generation inherently produces structured, well-defined geometry without the need for postprocessing or surface extraction, making it an increasingly prominent approach in recent years. Early methods tackle this task by generating meshes from surface patches [13], deforming ellipsoids [45], predicting mesh graphs [8], or employing binary space partitioning [6]. More recent techniques leverage generative models, particularly diffusion models and sequence-based approaches. PolyDiff [1] applies discrete diffusion, while PolyGen [29] autoregressively predicts vertices and faces using two separate networks. The sequence modeling paradigm has been further refined by representing the entire mesh as a single sequence. MeshGPT [37] introduces a tokenization scheme based on a vector quantized variational auto-encoder (VQVAE), while MeshXL [2] directly models discretized triangle soup sequences without compression. MeshAnything [4] extends MeshGPT by incorporating a point-cloud encoder for shape-conditioned generation. PivotMesh [49] introduces a hierarchical approach, generating pivot vertices before producing the full mesh. Further research has focused on optimizing tokenization and sequence modeling. Adjacency-aware compression techniques [3, 43, 50] improve tokenization efficiency and enable higher face counts within fixed context windows, while Meshtron [15] leverages Hourglass Transformers and sliding window attention to scale MeshXL sequences more effectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 553, + 504 + ], + "type": "text", + "content": "Despite these advancements, most existing approaches follow a partial-to-complete paradigm, where mesh regions are constructed sequentially at a uniform level of detail. Our work instead adopts a coarse-to-fine approach, continuously increasing the level of detail as new tokens are generated, achieving better control over geometric complexity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 700 + ], + "type": "text", + "content": "Level of Detail Representations. Level of Detail [26] (LOD) is a widely used technique in computer graphics to optimize rendering performance by reducing the complexity of 3D models based on their size, distance from the camera, or importance in a scene. Different LOD strategies depend on the underlying representation of 3D shapes. For meshes, progressive meshes [17] generate a mesh sequence starting from a coarse base model, gradually refining it through a series of transformations that incrementally add detail. Traditional mesh simplification methods [12, 20] can be adopted to construct such sequences by iteratively reducing polygon count while preserving geometric fidelity. Progressive simplicial complexes [32] extend progressive meshes to handle arbitrary meshes, including non-manifold and non-orientable surfaces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Recent studies have incorporated neural networks for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12571" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 293, + 159 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 293, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 293, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 293, + 159 + ], + "type": "image", + "image_path": "6d5a953d4abc6248f36699478fa034c38125043fcbb81d4ce61d00df28cee9f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 162, + 295, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 295, + 206 + ], + "type": "text", + "content": "Figure 2. Illustration of edge collapse and its inverse operation, i.e. vertex split. During edge collapse, vertex " + }, + { + "bbox": [ + 55, + 162, + 295, + 206 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 55, + 162, + 295, + 206 + ], + "type": "text", + "content": " is collapsed into " + }, + { + "bbox": [ + 55, + 162, + 295, + 206 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 55, + 162, + 295, + 206 + ], + "type": "text", + "content": ", resulting in two degenerate triangles (shaded in yellow), which forms two new edges in the result mesh." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 377 + ], + "type": "text", + "content": "LOD representations. Neural progressive meshes [5] propose a learned approach supporting LOD with a subdivision-based encoder-decoder. Similar representations have been explored for reconstruction approaches with signed distance functions (SDFs) [39-41] and neural fields [28]. However, these approaches are not generative; they typically encode and decode existing meshes [5] or optimize the LOD representation per shape [39]. In contrast, our approach is purely generative. VertexRegen learns to create meshes from scratch, progressively adding detail through autoregressive sequence generation. This allows us to generate new, high-fidelity meshes without relying on pre-existing structures, distinguishing our method from both traditional and neural LOD techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 111, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 111, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 111, + 400 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 217, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 217, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 217, + 421 + ], + "type": "text", + "content": "3.1. Progressive Meshes Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "content": "A progressive mesh [17] (PM) proposes an efficient and lossless continuous-resolution representation for arbitrary triangle meshes. The representation is built off two observations: (i) that a single mesh transformation, edge collapse, is sufficient for effective simplification of meshes; (ii) that edge collapse transformations are invertible via a vertex split operation. In the following, we provide details of the edge collapse operation and its inversion with vertex splits, each illustrated in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": "The edge collapse operation, " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\operatorname{ecol}(v_s, v_t)" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": ", unifies two adjacent vertices " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": " into a single vertex " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": ". The operation results in the vanishing of two faces, " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\{v_t, v_s, v_l\\}" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\{v_s, v_t, v_r\\}" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": ", as well as the vertex " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": ". In the general case, a new position for " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": " is also specified; however, for the case of half-edge collapse, the original position for " + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 55, + 533, + 296, + 605 + ], + "type": "text", + "content": " is kept." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": "An initial mesh, " + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": ", can be simplified into a more coarse mesh, " + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_n" + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": ", by " + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": " successive edge collapse operations:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 639, + 265, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 639, + 265, + 656 + ], + "spans": [ + { + "bbox": [ + 86, + 639, + 265, + 656 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {0} \\xleftarrow {\\mathrm {e c o l} _ {0}} \\mathcal {M} _ {1} \\xleftarrow {\\mathrm {e c o l} _ {1}} \\dots \\xleftarrow {\\mathrm {e c o l} _ {n - 1}} \\mathcal {M}", + "image_path": "a575b5e107a2ea09d9d5cdf07c4856bf2d7fc825a2db2f4b37f7f8d7e075d976.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The order in which the edge collapse operations are performed is determined such that each successive edge collapse results in the minimum increase in Quadratic Error Metrics (QEM) [12] with respect to the original mesh. This" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "process can be repeated until there are no candidate edge collapses, e.g. without flipping face normals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": "A vertex split, " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{vsplit}(v_s,v_l,v_r,v_t)" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": ", defines the inverse of edge collapse. It restores " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": " and the two faces, " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\{v_{t},v_{s},v_{l}\\}" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\{v_{s},v_{t},v_{r}\\}" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": ", which vanished during edge collapse, as well as " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": " to its original position. In case " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "(v_{s},v_{t})" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": " is a boundary edge, either " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "text", + "content": " will not exist, and there is only one face restored during the vertex split." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "content": "The combination of edge collapse and its inverse vertex splits enables the progressive mesh representation, " + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{PM}(\\mathcal{M}) = (\\mathcal{M}_0,\\{\\mathrm{vsplit}_0,\\dots ,\\mathrm{vsplit}_{n - 1}\\})" + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "content": " to express an arbitrary triangle mesh, " + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "content": ", as a combination of a coarse mesh, " + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "content": ", obtained through edge collapse, and the sequence of " + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "content": " vertex split records required to reverse them:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 336, + 250, + 533, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 250, + 533, + 266 + ], + "spans": [ + { + "bbox": [ + 336, + 250, + 533, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {0} \\xrightarrow {\\mathrm {v s p l i t} _ {0}} \\mathcal {M} _ {1} \\xrightarrow {\\mathrm {v s p l i t} _ {1}} \\dots \\xrightarrow {\\mathrm {v s p l i t} _ {n - 1}} \\mathcal {M}", + "image_path": "bae6b3cce47da0442a695cefbed31b0f539355eea74f866296f0278e7f2c40ee.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 397, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 397, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 397, + 289 + ], + "type": "text", + "content": "3.2. VertexRegen" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "content": "Traditional progressive meshes require starting an initial detailed mesh to be simplified into a coarse mesh, " + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "content": ", and recording each of these collapse steps as vertex split records to form the representation. We observe that the formation of the PM representation serves effectively as forward (edge collapse) and reverse (vertex split) processes that enable transition between detailed and coarse meshes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "type": "text", + "content": "VertexRegen frames the creation of a detailed mesh as the generation of a PM representation. Analogous to denoising for diffusion models, we train a generative model to reverse edge collapse. Concretely, VertexRegen first generates a coarse mesh, " + }, + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 378, + 554, + 450 + ], + "type": "text", + "content": ", from scratch, before increasing the level of detail through the generation of vertex split records." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 456, + 493, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 493, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 493, + 468 + ], + "type": "text", + "content": "3.2.1. Progressive Mesh Parameterization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 472, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 496 + ], + "type": "text", + "content": "We frame generation as an autoregressive sequential modeling task, with the full mesh sequence taking the form:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 504, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 504, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 504, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{c c} \\text {M : [ < b o s > , [ M _ {-} 0 s e q u e n c e ] , < s e p > ,} & \\# \\mathcal {M} _ {0} \\\\ \\text {[ v s p l i t \\_ 0 ] , \\ldots , [ v s p l i t \\_ n - 1 ] , < e o s > ]} & \\# \\text {v s p l i t s} \\end{array}", + "image_path": "d2c59fc551c193b781c5d56c5934268a8d0b6235235d05c6417568ca4ebdd7a5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 537, + 553, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 553, + 562 + ], + "type": "text", + "content": "In the following, we will define the formation of the subsequences [M_0 sequence] and [vsplit]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "text", + "content": ": Coarse Mesh Tokenization. For the initial coarse mesh, we follow the tokenization scheme defined in MeshXL [2]. In this formulation, embeddings are learned for discretized coordinates in an " + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "inline_equation", + "content": "N^3" + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "text", + "content": " grid. A vertex is represented via the sequential look-up of " + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "text", + "content": "-value, " + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "text", + "content": "-value, and " + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 575, + 554, + 663 + ], + "type": "text", + "content": "-value. A face is then constructed as the concatenation of its 3 vertices, totaling 9 tokens. The full mesh sequence is then defined as the concatenation of its constituent faces." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 341, + 670, + 526, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 670, + 526, + 682 + ], + "spans": [ + { + "bbox": [ + 341, + 670, + 526, + 682 + ], + "type": "interline_equation", + "content": "v: \\left[ < x >, < y >, < z > \\right], \\quad \\# v e r t e x", + "image_path": "ea85399a34d61499740ee1c57f76c50b1328b9817eee3f7d431b9e20786cbd83.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 343, + 685, + 517, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 685, + 517, + 696 + ], + "spans": [ + { + "bbox": [ + 343, + 685, + 517, + 696 + ], + "type": "interline_equation", + "content": "F: \\left[ \\begin{array}{l} [ v _ {- 1} ], [ v _ {- 2} ], [ v _ {- 3} ] \\end{array} \\right], \\quad \\# \\text {f a c e}", + "image_path": "e45812f7ad70c054d4b092983e2b4a978df2a9e3529e4203dd6d54bff81cc2e3.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 343, + 698, + 521, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 698, + 521, + 710 + ], + "spans": [ + { + "bbox": [ + 343, + 698, + 521, + 710 + ], + "type": "interline_equation", + "content": "M: \\left[ [ F _ {- 1} ], [ F _ {- 2} ], \\dots , [ F _ {- N} ] \\right] \\quad \\# \\text {m e s h}", + "image_path": "249b04a0b267c1abc6c80f809f22cff22bd82168746f17ef1c25c9f1d7e8cba3.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12572" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 72, + 270, + 210 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 270, + 210 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 270, + 210 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 270, + 210 + ], + "type": "image", + "image_path": "2ef934b6144eba8d2af06d3526da99653091dac78c54247047ff6a896829d209.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "text", + "content": "Figure 3. Face count distribution of the coarsest mesh " + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "text", + "content": " and original mesh " + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "text", + "content": ". The coarsest meshes " + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_0)" + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "text", + "content": " contain significantly fewer faces than original " + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 55, + 213, + 296, + 258 + ], + "type": "text", + "content": ", with an average of 18 and 457 faces, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 261, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 297 + ], + "type": "text", + "content": "Following [37], we sort vertices in " + }, + { + "bbox": [ + 55, + 261, + 296, + 297 + ], + "type": "inline_equation", + "content": "z - y - x" + }, + { + "bbox": [ + 55, + 261, + 296, + 297 + ], + "type": "text", + "content": " order (lower to higher). Within each face, vertices are cylindrically permuted to have the lowest-indexed vertex be first." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 381 + ], + "type": "text", + "content": "Although MeshXL [2] leverages this tokenization scheme to produce full detailed meshes, in VertexRegen only the coarsest level of the mesh is parameterized in this way. These coarse meshes consist of substantially fewer faces than the full detailed mesh, as demonstrated in Fig. 3, and the resulting coarse mesh only accounts for " + }, + { + "bbox": [ + 55, + 297, + 296, + 381 + ], + "type": "inline_equation", + "content": "5.68\\%" + }, + { + "bbox": [ + 55, + 297, + 296, + 381 + ], + "type": "text", + "content": " of the total sequence length on average." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": "Vertex Split Generation. Consider the " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": "-th vertex split operation " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\mathrm{vsplit}_k" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": ", which converts a lower-detailed mesh " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " into a higher-detailed mesh " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k+1}" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": ". We denote the neighbor vertices of vertex " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "N_k(v)" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": ". After determining the target vertex " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " to split (in " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": ") and the new vertex " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " (in " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k+1}" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": "), we need to obtain " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "N_{k+1}(v_s)" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "N_{k+1}(v_t)" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " from the vertex split record to connect " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "v_s" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " with correct neighbor vertices in " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k+1}" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": ", which are prohibitive to generate as the number of neighbor vertices " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "|N_{k+1}(v_s)|" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "inline_equation", + "content": "|N_{k+1}(v_t)|" + }, + { + "bbox": [ + 55, + 396, + 296, + 516 + ], + "type": "text", + "content": " is often large." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "content": "However, as mesh " + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "content": " is generated from " + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k + 1}" + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "content": " during edge collapse by merging " + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 551, + 295, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 551, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 295, + 564 + ], + "type": "interline_equation", + "content": "N _ {k + 1} \\left(v _ {s}\\right) \\cap N _ {k + 1} \\left(v _ {t}\\right) = \\left\\{v _ {l}, v _ {r} \\right\\} \\tag {1}", + "image_path": "6aee1c97d5850280847dd4995102b63602ccbd56d902d987103fe99ae2693f22.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 574, + 295, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 574, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 85, + 574, + 295, + 588 + ], + "type": "interline_equation", + "content": "N _ {k + 1} \\left(v _ {s}\\right) \\cup N _ {k + 1} \\left(v _ {t}\\right) - \\left\\{v _ {s}, v _ {t} \\right\\} = N _ {k} \\left(v _ {s}\\right) \\tag {2}", + "image_path": "d63bb8b5854ab7403f9f517c64d9afacc229baedcdfa006eb85e03921398a599.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "which indicates we will only need to record vertex " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ". They split the vertex ring surrounding " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ", where two halves of the vertices on the ring (top and bottom vertices in Fig. 2) belong to " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " respectively in " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k + 1}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ". However, an ambiguity arises in determining which half of the ring corresponds to " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "'s neighbors in " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k + 1}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ". In the following, we demonstrate how the half-edge data structure [48] can be utilized to resolve this ambiguity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Denote " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{ij}^k" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " as a half-edge pointing from " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_{k}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " being the third vertex in the associated face. As shown in" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 73, + 552, + 194 + ], + "blocks": [ + { + "bbox": [ + 318, + 73, + 552, + 194 + ], + "lines": [ + { + "bbox": [ + 318, + 73, + 552, + 194 + ], + "spans": [ + { + "bbox": [ + 318, + 73, + 552, + 194 + ], + "type": "image", + "image_path": "d020ef834e9f52cbf4765977de5065d6732f60ef43375c9c10b203dce4fe5895.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": "Figure 4. Illustration of half-edge data structure and the traversal process to determine the neighbors of " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k + 1}" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": ". Starting from half-edge " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{ls}^{1}" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": ", we traverse the faces in clockwise direction until " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{sr}^{2}" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": ", where we collect the vertices " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "v_{2}" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "inline_equation", + "content": "N_{k + 1}(v_s)" + }, + { + "bbox": [ + 313, + 197, + 555, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 254, + 406, + 322 + ], + "blocks": [ + { + "bbox": [ + 317, + 254, + 406, + 322 + ], + "lines": [ + { + "bbox": [ + 317, + 254, + 406, + 322 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 406, + 322 + ], + "type": "image", + "image_path": "819156d72a3c6e7f194cbad6db72f5450cd1b639e5afff7774fb590878e99a53.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 408, + 254, + 553, + 326 + ], + "blocks": [ + { + "bbox": [ + 408, + 254, + 553, + 326 + ], + "lines": [ + { + "bbox": [ + 408, + 254, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 408, + 254, + 553, + 326 + ], + "type": "image", + "image_path": "39b44188626f66fd98150f202dd5e908f6498201f422e5fca196bb4b331cd831.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "type": "text", + "content": "Figure 5. Illustration of VertexRegen tokenization. The sequence begins with base mesh " + }, + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "type": "text", + "content": ", followed by vertex split subsequences. A special token indicates either " + }, + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 313, + 327, + 554, + 361 + ], + "type": "text", + "content": " does not exist." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "text", + "content": "Fig. 4, we begin the traversal from " + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{ls}^{1}" + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "text", + "content": " and proceed to the twin of the next half-edge, " + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{1s}^{2}" + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "text", + "content": ". This operation is repeated until we reach " + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{,s}^{r}" + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "text", + "content": ". Throughout this process, we traverse the faces above " + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "inline_equation", + "content": "(v_{l}, v_{s})" + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "inline_equation", + "content": "(v_{s}, v_{r})" + }, + { + "bbox": [ + 313, + 363, + 554, + 412 + ], + "type": "text", + "content": " clockwise, obtaining:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 418, + 553, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 418, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 342, + 418, + 553, + 433 + ], + "type": "interline_equation", + "content": "\\left\\{v _ {k} \\mid \\mathcal {H} _ {\\cdot s} ^ {k}, k \\neq r \\right\\} = N _ {k + 1} \\left(v _ {s}\\right) - \\left\\{v _ {l}, v _ {r}, v _ {t} \\right\\} \\tag {3}", + "image_path": "c33d163b8035c7cd3b8d5f32bc2274b0c8200e1c3dba0fe9ab5030ce4d0ed894.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{s}^{k}" + }, + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "type": "text", + "content": " is the half-edge encountered during traversal." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "text", + "content": "In conjunction with Eqs. (1) and (2), " + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "inline_equation", + "content": "N_{k + 1}(v_t)" + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "text", + "content": " can also be determined. When " + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "inline_equation", + "content": "(v_s, v_t)" + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "text", + "content": " is a boundary in " + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k + 1}" + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "text", + "content": ", we follow either " + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{ls}" + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "text", + "content": " clockwise or " + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{sr}" + }, + { + "bbox": [ + 313, + 453, + 554, + 501 + ], + "type": "text", + "content": " counterclockwise until we reach a half-edge of the boundary." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": "After identifying the neighbors of " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " in the mesh " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k + 1}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": ", we add the new vertex " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " to the mesh " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " and reconnect the vertices in " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "N_{k + 1}(v_t)" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": ". Finally, we restore two faces, " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\{v_s,v_l,v_t\\}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\{v_r,v_s,v_t\\}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": ", which are associated with the half-edges " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{sl}^{t}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{rs}^{t}" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": ", respectively. Using the half-edge data structure, the orientation of the newly created faces remains consistent with the rest of the mesh. When " + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "inline_equation", + "content": "(v_{s},v_{t})" + }, + { + "bbox": [ + 313, + 501, + 554, + 596 + ], + "type": "text", + "content": " lies on the boundary, only one face is restored." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "text", + "content": "Hence, with the above half-edge formulation, each vertex split operation can be uniquely defined by the selection of the target vertex " + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "text", + "content": " and two neighbors " + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "text", + "content": ", as well as the position to place the new vertex " + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 597, + 554, + 644 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 658, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 554, + 714 + ], + "type": "text", + "content": "Vertex Split Tokenization. Although " + }, + { + "bbox": [ + 313, + 658, + 554, + 714 + ], + "type": "inline_equation", + "content": "v_{s}, v_{l}" + }, + { + "bbox": [ + 313, + 658, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 658, + 554, + 714 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 313, + 658, + 554, + 714 + ], + "type": "text", + "content": " only require references to existing vertices in the mesh, in practice we implement this reference as a raw prediction of each vertex to avoid a vocabulary size proportional to the sequence length. Thus, in the majority of cases, each vertex" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12573" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 122 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 122 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 122 + ], + "type": "image", + "image_path": "a3ec9d00c549e222e74e59e34f64589d85b2bb278fa23b478001ff0b965514e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 125, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 58, + 125, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 125, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 125, + 553, + 189 + ], + "type": "image", + "image_path": "838b5a3a02951f2826ef225c104dbe6f899341e017da11c477342322afb22898.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 191, + 553, + 253 + ], + "blocks": [ + { + "bbox": [ + 58, + 191, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 58, + 191, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 58, + 191, + 553, + 253 + ], + "type": "image", + "image_path": "31f0bcb4672f8d3a55fb57b66296abf470de15be0171a09c39608e19cfa1081e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 255, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 58, + 255, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 255, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 255, + 553, + 319 + ], + "type": "image", + "image_path": "db8404fafb7a72ff0cfd61a46683118a25a8bfb9e0f3a77fe6d6a38c6aca455a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 322, + 553, + 459 + ], + "blocks": [ + { + "bbox": [ + 58, + 322, + 553, + 459 + ], + "lines": [ + { + "bbox": [ + 58, + 322, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 58, + 322, + 553, + 459 + ], + "type": "image", + "image_path": "1c90d14a5746db392b51a47595c6cb1863a98ac2404d890fba2f27cdc3b30b58.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "type": "text", + "content": "Figure 6. Generation process for VertexRegen. " + }, + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 462, + 555, + 485 + ], + "type": "text", + "content": " represents the predicted initial coarsest mesh, followed by subsequent meshes generated through the predicted vertex split sequence." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 531 + ], + "type": "text", + "content": "split is represented by a subsequence of 12 tokens (or 10 tokens when " + }, + { + "bbox": [ + 55, + 507, + 295, + 531 + ], + "type": "inline_equation", + "content": "(v_{s}, v_{t})" + }, + { + "bbox": [ + 55, + 507, + 295, + 531 + ], + "type": "text", + "content": " is a boundary edge):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 539, + 277, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 539, + 277, + 551 + ], + "spans": [ + { + "bbox": [ + 73, + 539, + 277, + 551 + ], + "type": "text", + "content": "```c\nvsplit: [, , , #v_s" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 553, + 276, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 553, + 276, + 565 + ], + "spans": [ + { + "bbox": [ + 119, + 553, + 276, + 565 + ], + "type": "inline_equation", + "content": "< 1_{-}x>, < 1_{-}y>, < 1_{-}z> |< \\mathrm{nil}>" + }, + { + "bbox": [ + 119, + 553, + 276, + 565 + ], + "type": "text", + "content": ", # " + }, + { + "bbox": [ + 119, + 553, + 276, + 565 + ], + "type": "inline_equation", + "content": "v_{l}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 119, + 567, + 276, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 567, + 276, + 578 + ], + "spans": [ + { + "bbox": [ + 119, + 567, + 276, + 578 + ], + "type": "inline_equation", + "content": "< r_{-}x>, < r_{-}y>, < r_{-}z|< nil>" + }, + { + "bbox": [ + 119, + 567, + 276, + 578 + ], + "type": "text", + "content": " ， #vr" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 119, + 582, + 276, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 582, + 276, + 594 + ], + "spans": [ + { + "bbox": [ + 119, + 582, + 276, + 594 + ], + "type": "inline_equation", + "content": "< t\\_ x > , < t\\_ y > , < t\\_ z>" + }, + { + "bbox": [ + 119, + 582, + 276, + 594 + ], + "type": "text", + "content": " #vt" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": "where the special token " + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{nil} \\rangle" + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": " signifies " + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": " does not exist, and only one of " + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": " is allowed to be " + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{nil} \\rangle" + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": " for a valid vertex split subsequence. In Fig. 5 we illustrate the entire tokenization process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 649, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 295, + 673 + ], + "type": "text", + "content": "The above progressive mesh serialization can then be trained with the standard next-token prediction target." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Vertex Split Decoding. The vertex split sequence must be consistent with the actual geometry to be valid. We" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": "maintain a state machine and perform vertex split on the fly as new tokens are generated. After the coarsest mesh " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " is generated, we initialize the state. During each generation step, we enforce " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " to be a vertex in the current mesh " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "(v_{s}, v_{l})" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "(v_{s}, v_{r})" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " are valid edges (if " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " are not " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{nil} \\rangle" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": "), and only one of " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "v_{r}" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " can be " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{nil} \\rangle" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": ". Lastly, we decode " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": ", and perform vertex split with generated " + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "inline_equation", + "content": "v_{s}, v_{l}, v_{r}, v_{t}" + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": " information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 611, + 448, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 448, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 448, + 624 + ], + "type": "text", + "content": "3.3. Conditional Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "We consider generation conditioned on shapes. We adopt a pre-trained point-cloud encoder [58]. A LLaVa-style [21] projector is leveraged to project the condition features to the token embedding space. The projected feature tokens are prepended to the mesh sequence as the prefix. We supervise training with the next-token prediction objective while masking the loss for the prefix tokens." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12574" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 70, + 478, + 155 + ], + "blocks": [ + { + "bbox": [ + 133, + 70, + 478, + 155 + ], + "lines": [ + { + "bbox": [ + 133, + 70, + 478, + 155 + ], + "spans": [ + { + "bbox": [ + 133, + 70, + 478, + 155 + ], + "type": "table", + "html": "
MethodTokenizationCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)
MeshXL [2]Flattened Coords.51.768.3050.843.81
MeshAnything V2 [3]AMT50.338.5052.254.84
EdgeRunner [43]EdgeBreaker [33]51.397.8149.443.22
VertexRegen (Ours)Progressive51.038.2950.222.89
", + "image_path": "887cfc1e08e1fd02eb1493072ea60203070ee823e86026e6e06a6970273a8e5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 555, + 178 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 555, + 178 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with state-of-the-art methods for unconditional mesh generation. Best results are bolded, second best are underlined. VertexRegen can generate meshes with comparable quality while enjoying the benefits of continuous level of detail." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 194, + 555, + 312 + ], + "blocks": [ + { + "bbox": [ + 56, + 194, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 56, + 194, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 555, + 312 + ], + "type": "image", + "image_path": "da9816eabdcbcd81189ff45c87dadff60fdf83e77fdf2d55f01b0477c27c1b85.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "type": "text", + "content": "Figure 7. Unconditional generation under face count constraints. VertexRegen achieves significantly better COV, MMD, and 1-NNA in early stages of generation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 137, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 137, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 137, + 372 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 118, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 118, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 118, + 392 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 400, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 295, + 579 + ], + "type": "text", + "content": "We pre-train our model and the baselines presented using two primary sources of meshes: Objaverse-XL [9] and an additional set of licensed artist-created meshes. Initially, we select meshes containing fewer than 8,000 faces without applying decimation. We use the CGAL [10] library to implement edge collapse and vertex split operations, where we modify the vertex placement to either " + }, + { + "bbox": [ + 54, + 400, + 295, + 579 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 54, + 400, + 295, + 579 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 400, + 295, + 579 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 54, + 400, + 295, + 579 + ], + "type": "text", + "content": ". We filter out non-manifold meshes and those that cannot be processed, which results in a final dataset of approximately 1.5M meshes with an average of 457 faces. For unconditional generation evaluation, we construct a high-quality subset by further filtering with the alignment split of Objaverse-XL and with fewer than 800 faces. This yields approximately 18k meshes. We adopt a similar filtering process for shape-conditioned experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "text", + "content": "We apply robust data augmentations to input meshes, including random shift within range " + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "[-0.1, 0.1]" + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "text", + "content": ", random scaling between [0.9, 1.1], random rotation by " + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 54, + 582, + 295, + 677 + ], + "type": "text", + "content": ". During pre-training, we handle sequences exceeding the context window by discarding vertex split subsequences, while for other baselines, we truncate the tokenized sequence directly. The set for quantitative evaluation is chosen to avoid truncation for any method evaluated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 359, + 447, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 447, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 447, + 372 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "text", + "content": "VertexRegen and all other baselines are built upon the pretrained OPT-350M [56], with newly initialized token embeddings and position embeddings. We train our model on 64 H100 GPUs for approximately four days, with an effective batch size of 256. We use the AdamW [25] optimizer with a weight decay of 0.1, and betas (0.9, 0.95). A cosine scheduler is employed, starting with an initial learning rate of " + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "inline_equation", + "content": "1 \\times 10^{4}" + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "text", + "content": " and gradually decreasing to " + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "inline_equation", + "content": "5 \\times 10^{6}" + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "text", + "content": ". We clip the gradient norm to 1.0. During inference, we adopt the top-" + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "text", + "content": " sampling strategy with " + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "inline_equation", + "content": "p = 0.95" + }, + { + "bbox": [ + 312, + 379, + 555, + 498 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 372, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 372, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 372, + 521 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 529, + 454, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 454, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 454, + 540 + ], + "type": "text", + "content": "4.3.1. Unconditional Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": "We follow evaluation protocols from prior works [1, 2, 37], employing point-cloud-based metrics to assess unconditional generation. Specifically, we sample the same number of meshes as the evaluation dataset and randomly sample 2,048 points per mesh. Coverage (COV) measures the diversity of the generated samples, where higher values indicate greater diversity. Minimum Matching Distance (MMD) computes the average distance from each reference sample to its nearest neighbor in the generated set, serving as a measure of generation quality, with lower values being preferable. 1-Nearest Neighbor Accuracy (1-NNA) evaluates both diversity and quality, where an optimal value is achieved at " + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": ". Additionally, we compute Jensen-Shannon Divergence (JSD) to directly quantify the sim" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "1We neither used assets from Sketchfab nor obtained any from the Polycam website." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12575" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 549, + 345 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 549, + 345 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 549, + 345 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 549, + 345 + ], + "type": "image", + "image_path": "7c4f79cf3f552eb86c3ded40706d7fa9a2e5f6189bc60a14e92d97d62ee9bfae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 352, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 352, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 555, + 373 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison with state-of-the-art methods. VertexRegen is able to generate meshes with comparable quality to other baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 388, + 552, + 653 + ], + "blocks": [ + { + "bbox": [ + 58, + 388, + 552, + 653 + ], + "lines": [ + { + "bbox": [ + 58, + 388, + 552, + 653 + ], + "spans": [ + { + "bbox": [ + 58, + 388, + 552, + 653 + ], + "type": "image", + "image_path": "a80fd6086d5a71c985ca3b6d7d8b5efa4b9a0c2973eb713545b57244e96d0c34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 654, + 555, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 654, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 555, + 677 + ], + "type": "text", + "content": "Figure 9. Qualitative results on shape-conditioned generation on meshes held out from training. The first column shows the point cloud used as the condition, followed by the generation sequence progressing from the coarsest mesh " + }, + { + "bbox": [ + 55, + 654, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 654, + 555, + 677 + ], + "type": "text", + "content": " to the final output in the last column." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12576" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "ilarity between the generated and reference distributions. We present results in Tab. 1. Our method can generate meshes of comparable quality to other state-of-the-art methods, with the added advantage of a continuous level of detail. We also show qualitative examples and their generation process in Fig. 6, and more generation results comparing with state-of-the-art methods in Fig. 8." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 173, + 295, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 173, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 54, + 173, + 295, + 268 + ], + "type": "text", + "content": "Generation with a face count constraint. We compare our method with the baselines under varying face count constraints. Our approach allows for generation to be paused at any point, accommodating the varying constraints. In contrast, to adapt to these constraints, baseline methods require either: (i) directly truncating the generation process, yielding incomplete intermediate meshes, or (ii) fine-tuning with an additional face count condition token." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 365 + ], + "type": "text", + "content": "For the direct truncation version, we evaluate unconditional generation metrics w.r.t. different face count constraints and present the results in Fig. 7. Additionally, we adopt a similar approach as proposed in EdgeRunner, where we split [1,800] range into 4 buckets and prepend the conditioning token to the mesh sequence. We compare VertexRegen and MeshXL fine-tuned with such conditioning scheme at face constraint 400, as shown in Tab. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 485 + ], + "type": "text", + "content": "Due to its ability to generate meshes with a continuous level of detail from coarse to fine, VertexRegen effectively captures the overall structure even with very limited face counts, yielding significantly better COV, MMD, and 1-NNA early in the process. As the face limit increases, all methods demonstrate improved results. Notably, for baselines such as EdgeRunner, while they enhance tokenization efficiency over MeshXL, they do not fundamentally address continuous level-of-detail generation and thus follow a similar trend to other baselines in the plot." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 185, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 185, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 185, + 502 + ], + "type": "text", + "content": "4.3.2. Conditional Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 568 + ], + "type": "text", + "content": "We show qualitative examples of shape-conditioned generation in Fig. 9. We condition on 4,096 sampled points with normals from dense meshes. VertexRegen can generate a coarse " + }, + { + "bbox": [ + 55, + 507, + 295, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 507, + 295, + 568 + ], + "type": "text", + "content": ", starting as simply as a tetrahedron, and progressively refine it by generating a sequence of vertex splits." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 156, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 156, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 156, + 588 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Tokenization efficiency. We compute the tokenized sequence length of our proposed tokenization scheme (Sec. 3.2.1) on our dataset and report the average compression ratio relative to MeshXL (9 tokens per face) in Tab. 3. For VertexRegen, there are two primary sequence types: (i) a MeshXL-style sequence for the initial coarsest mesh " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " and (ii) sequences encoding vertex splits. The latter requires 12 tokens for two non-boundary faces or 10 tokens per boundary face. As a result, the highest compression is achieved when " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " is minimized and all vertex splits occur" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 553, + 121 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 553, + 121 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 121 + ], + "type": "table", + "html": "
@400 FacesCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
VertexRegen50.928.3151.032.88264147
MeshXL (w/ FCC)41.2010.0359.065.19308168
", + "image_path": "a22f183ec76f9591698f6db3b5857428809fe65d449d7b224ae3826165ccd9af.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 160, + 552, + 198 + ], + "blocks": [ + { + "bbox": [ + 314, + 122, + 553, + 145 + ], + "lines": [ + { + "bbox": [ + 314, + 122, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 122, + 553, + 145 + ], + "type": "text", + "content": "Table 2. Comparison between VertexRegen and MeshXL with face count condition (FCC) when face count constraint is 400." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 160, + 552, + 198 + ], + "lines": [ + { + "bbox": [ + 318, + 160, + 552, + 198 + ], + "spans": [ + { + "bbox": [ + 318, + 160, + 552, + 198 + ], + "type": "table", + "html": "
MeshXLMeshAnything v2EdgeRunnerVertexRegen w/ HEw/o HE
Compression1.00.460.470.730.89
", + "image_path": "da59ddc80e921a51a43d7518d086d416b398970160b2eb3a293e44e11018c6d9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 236, + 552, + 290 + ], + "blocks": [ + { + "bbox": [ + 314, + 199, + 553, + 221 + ], + "lines": [ + { + "bbox": [ + 314, + 199, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 314, + 199, + 553, + 221 + ], + "type": "text", + "content": "Table 3. Compression ratio of different tokenization schemes, VertexRegen with and without leveraging a half-edge (HE) structure." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 236, + 552, + 290 + ], + "lines": [ + { + "bbox": [ + 317, + 236, + 552, + 290 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 552, + 290 + ], + "type": "table", + "html": "
Guided DecodingCOV(%,↑)MMD(×103,↓)1-NNA(%)JSD(↓)|F||V|
w/o51.128.3150.753.37211120
w/51.038.2950.222.89320176
", + "image_path": "90e337f1a55043eebcfec22e3fe0a474337de63382a52db87ff9971173642a4f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 291, + 520, + 302 + ], + "lines": [ + { + "bbox": [ + 348, + 291, + 520, + 302 + ], + "spans": [ + { + "bbox": [ + 348, + 291, + 520, + 302 + ], + "type": "text", + "content": "Table 4. Effects on geometry-guided decoding." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 308, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 368 + ], + "type": "text", + "content": "on non-boundary faces. On average, VertexRegen achieves a compression ratio of 0.73, approaching the theoretical limit of 0.67. The discrepancy arises from the overhead introduced by " + }, + { + "bbox": [ + 313, + 308, + 553, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_0" + }, + { + "bbox": [ + 313, + 308, + 553, + 368 + ], + "type": "text", + "content": " (5.68% of all total tokenized length) and boundary vertex splits (3.64% of all vertex splits)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "text", + "content": "Without the half-edge data structure, identifying which half of the ring surrounding " + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "text", + "content": " is associated with " + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "inline_equation", + "content": "v_{s}" + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{k + 1}" + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "text", + "content": " requires recording an additional vertex. This increases the token count to 15 per two non-boundary faces (or 13 per boundary face) and results in an average increase of " + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 313, + 369, + 554, + 441 + ], + "type": "text", + "content": " in tokenized sequence length." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 462, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 535 + ], + "type": "text", + "content": "Guided decoding. In Tab. 4, we ablate the effects of geometry-constrained decoding. As one step of the vertex split operation is dependent on all the preceding operations, predicting an invalid vertex split may break the chain and end the generation prematurely. With guided decoding, the model is able to generate longer sequences with more faces." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduced VertexRegen, a novel mesh generation framework that enables continuous levels of detail through a generative process based on vertex splits. Unlike conventional auto-regressive approaches that synthesize meshes in a partial-to-complete manner, VertexRegen reinterprets mesh generation as the reversal of edge collapse, providing an effectively anytime solution to mesh generation. Our experimental results demonstrate that VertexRegen achieves competitive performance compared to state-of-the-art methods while offering the unique advantage of halting generation at any stage to obtain meshes at different levels of detail." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12577" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Antonio Alliegro, Yawar Siddiqui, Tatiana Tommasi, and Matthias Nießner. Polydiff: Generating 3d polygonal meshes with diffusion models. arXiv preprint arXiv:2312.11417, 2023. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Billzb Wang, Jingyi Yu, Gang Yu, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems, 37:97141-97166, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[3] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 304 + ], + "type": "text", + "content": "[4] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Zhongang Cai, Lei Yang, Gang Yu, Guosheng Lin, and Chi Zhang. Meshanything: Artist-created mesh generation with autoregressive transformers. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 339 + ], + "type": "text", + "content": "[5] Yun-Chun Chen, Vladimir Kim, Noam Aigerman, and Alec Jacobson. Neural progressive meshes. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 385 + ], + "type": "text", + "content": "[6] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 295, + 442 + ], + "type": "text", + "content": "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G. Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4456-4465, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 487 + ], + "type": "text", + "content": "[8] Angela Dai and Matthias Niessner. Scan2mesh: From unstructured range scans to 3d meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 489, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 295, + 544 + ], + "type": "text", + "content": "[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 62, + 489, + 295, + 544 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 62, + 489, + 295, + 544 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 36:35799-35813, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 545, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 601 + ], + "type": "text", + "content": "[10] Andreas Fabri and Sylvain Pion. Cgal: the computational geometry algorithms library. In Proceedings of the 17th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, page 538-539, New York, NY, USA, 2009. Association for Computing Machinery. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "text", + "content": "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[12] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 209-216, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[13] Thibault Groueix, Matthew Fisher, Vladimir G. Kim, Bryan C. Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "text", + "content": "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 217 + ], + "type": "text", + "content": "[15] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 217, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 554, + 271 + ], + "type": "text", + "content": "[16] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3d. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 273, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 554, + 316 + ], + "type": "text", + "content": "[17] Hugues Hoppe. Progressive meshes. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, page 99–108, New York, NY, USA, 1996. Association for Computing Machinery. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 317, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 554, + 348 + ], + "type": "text", + "content": "[18] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 350, + 555, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 555, + 393 + ], + "type": "text", + "content": "[19] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12642-12651, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 394, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 554, + 426 + ], + "type": "text", + "content": "[20] Peter Lindstrom and Greg Turk. Fast and memory efficient polygonal simplification. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 279-286. IEEE, 1998. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 427, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 554, + 470 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "type": "text", + "content": "[22] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, and Hao Su. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. In Advances in Neural Information Processing Systems, pages 59314-59341. Curran Associates, Inc., 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 612 + ], + "type": "text", + "content": "[23] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 554, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 680 + ], + "type": "text", + "content": "[24] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In Proceedings of the 14th Annual Conference on Computer Graphics and Interactive Techniques, page 163-169, New York, NY, USA, 1987. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12578" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[26] David Luebke, Martin Reddy, Jonathan D. Cohen, Amitabh Varshney, Benjamin Watson, and Robert Huebner. Level of Detail for 3D Graphics. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 2002. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[27] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "type": "text", + "content": "[28] Felix Mujkanovic, Ntumba Elie Nsampi, Christian Theobalt, Hans-Peter Seidel, and Thomas Leimkuhler. Neural gaussian scale-space fields. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "text", + "content": "[29] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 364 + ], + "type": "text", + "content": "[31] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "type": "text", + "content": "[32] Jovan Popovic and Hugues Hoppe. Progressive simplicial complexes. In Proceedings of the 24th Annual Conference on Computer Graphics and Interactive Techniques, page 217-224, USA, 1997. ACM Press/Addison-Wesley Publishing Co. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 454 + ], + "type": "text", + "content": "[33] J. Rossignac. Edgebreaker: connectivity compression for triangle meshes. IEEE Transactions on Visualization and Computer Graphics, 5(1):47-61, 1999. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "text", + "content": "[34] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 555 + ], + "type": "text", + "content": "[35] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "text", + "content": "[36] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20887-20897, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 678 + ], + "type": "text", + "content": "[37] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[38] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "type": "text", + "content": "David Novotny. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. In Advances in Neural Information Processing Systems, pages 9532-9564. Curran Associates, Inc., 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 194 + ], + "type": "text", + "content": "[39] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11367, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[40] Danhang Tang, Mingsong Dou, Peter Lincoln, Philip Davidson, Kaiwen Guo, Jonathan Taylor, Sean Fanello, Cem Keskin, Adarsh Kowdle, Sofien Bouaziz, et al. Real-time compression and streaming of 4d performances. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 555, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 329 + ], + "type": "text", + "content": "[41] Danhang Tang, Saurabh Singh, Philip A. Chou, Christian Hane, Mingsong Dou, Sean Fanello, Jonathan Taylor, Philip Davidson, Onur G. Guleryuz, Yinda Zhang, Shahram Izadi, Andrea Tagliasacchi, Sofien Bouaziz, and Cem Keskin. Deep implicit volume compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "text", + "content": "[43] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgrunner: Auto-regressive auto-encoder for artistic mesh generation. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "text", + "content": "[44] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 478, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 530 + ], + "type": "text", + "content": "[45] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 578 + ], + "type": "text", + "content": "[46] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "text", + "content": "[47] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 731-741, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "text", + "content": "[48] Kevin J Weiler. Topological structures for geometric modeling (Boundary representation, manifold, radial edge structure). Rensselaer Polytechnic Institute, 1986. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12579" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 619 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "text", + "content": "[50] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, Tong Zhang, Shenghua Gao, and C.L. Philip Chen. Scaling mesh generation via compressive tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11093-11103, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[51] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 206, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 261 + ], + "type": "text", + "content": "[52] Haiyang Xu, Yu Lei, Zeyuan Chen, Xiang Zhang, Yue Zhao, Yilin Wang, and Zhuowen Tu. Bayesian diffusion models for 3d shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10628-10638, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 262, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 316 + ], + "type": "text", + "content": "[53] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 318, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 294, + 361 + ], + "type": "text", + "content": "[54] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 294, + 407 + ], + "type": "text", + "content": "[55] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 294, + 460 + ], + "type": "text", + "content": "[56] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 463, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 507 + ], + "type": "text", + "content": "[57] Qingcheng Zhao, Xiang Zhang, Haiyang Xu, Zeyuan Chen, Jianwen Xie, Yuan Gao, and Zhuowen Tu. Depr: Depth guided single-view scene reconstruction with instance-level diffusion. arXiv preprint arXiv:2507.22825, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 509, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 573 + ], + "type": "text", + "content": "[58] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, BIN FU, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. In Advances in Neural Information Processing Systems, pages 73969-73982. Curran Associates, Inc., 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "text", + "content": "[59] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826–5835, 2021. 1, 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12580" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_content_list.json b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d706c97c2554eff45d6cba1383db29f66a54a2bd --- /dev/null +++ b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_content_list.json @@ -0,0 +1,1714 @@ +[ + { + "type": "text", + "text": "ViCTr: Vital Consistency Transfer for Pathology Aware Image Synthesis", + "text_level": 1, + "bbox": [ + 133, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Onkar Susladkar $^{1,3}$ , Gayatri Deshmukh $^{1}$ , Yalcin Tur $^{2}$ , Gorkem Durak $^{1}$ , Ulas Bagci $^{1}$ $^{1}$ Northwestern University $^{2}$ Stanford University $^{3}$ University of Illinois Ubrana-Champaign", + "bbox": [ + 125, + 179, + 870, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "onkarsus13@gmail.com, dgayatri9850@gmail.com, yalcintr@stanford.edu,", + "bbox": [ + 196, + 219, + 795, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gorkem.durak@northwestern.edu, ulas.bagci@northwestern.edu", + "bbox": [ + 238, + 237, + 751, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing medical images remains challenging due to limited annotated pathological data, modality domain gaps, and the complexity of representing diffuse pathologies such as liver cirrhosis. Existing methods often struggle to maintain anatomical fidelity while accurately modeling pathological features, frequently relying on priors derived from natural images or inefficient multi-step sampling. In this work, we introduce ViCTR (Vital Consistency Transfer), a novel two-stage framework that combines a rectified flow trajectory with a Tweedie-corrected diffusion process to achieve high-fidelity, pathology-aware image synthesis. First, we pretrain ViCTR on the ATLAS-8k dataset using Elastic Weight Consolidation (EWC) to preserve critical anatomical structures. We then fine-tune the model adversarially with Low-Rank Adaptation (LoRA) modules for precise control over pathology severity. By reformulating Tweedie's formula within a linear trajectory framework, ViCTR supports one-step sampling—reducing inference from 50 steps to just 4—without sacrificing anatomical realism. We evaluate ViCTR on BTCV (CT), AMOS (MRI), and CirrMRI600+ (cirrhosis) datasets. Results demonstrate state-of-the-art performance, achieving a Medical Fréchet Inception Distance (MFID) of 17.01 for cirrhosis synthesis—28% lower than existing approaches—and improving nnUNet segmentation by +3.8% mDSC when used for data augmentation. Radiologist reviews indicate that ViCTR-generated liver cirrhosis MRIs are clinically indistinguishable from real scans. To our knowledge, ViCTR is the first method to provide fine-grained, pathology-aware MRI synthesis with graded severity control, closing a critical gap in AI-driven medical imaging research.", + "bbox": [ + 88, + 316, + 485, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 814, + 220, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The exponential growth in computer vision capabilities has been driven by significant advances in artificial intelligence models [7, 32]. However, medical imaging faces a fundamental tension between model complexity and data avail", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ability that limits the application of state-of-the-art techniques. While recent breakthroughs in generative AI have demonstrated remarkable capabilities in synthetic data creation, these advances demand training datasets of unprecedented scale—a requirement that poses unique challenges in the medical domain [2, 33].", + "bbox": [ + 511, + 287, + 906, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike general computer vision applications, medical imaging faces several critical constraints: privacy regulations necessitating complex deidentification, inherent data fragmentation across healthcare institutions, and fundamental interoperability constraints. These barriers have created a growing disparity between the rapid advancement of general computer vision and the relatively slower progress in medical imaging applications. Current approaches to bridging this gap face two major limitations: insufficient feature preservation across anatomical structures and inadequate handling of pathological variations.", + "bbox": [ + 511, + 380, + 908, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This challenge is especially pronounced in abdominal imaging, where pathologies such as diffuse cirrhosis or multi-tissue disease processes manifest across multiple organ systems. Unlike the well-defined boundaries typical of tumors or simpler structures like bones in X-ray imaging, abdominal pathologies often involve subtle and heterogeneous changes in tissue characteristics, requiring more nuanced feature extraction and synthesis. The inherent complexity of MRI signals—spanning multiple sequences and high spatial resolution—presents additional computational hurdles, further distancing it from the more standardized nature of CT or simpler 2D radiographs.", + "bbox": [ + 511, + 550, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these gaps, our work introduces ViCTr (Vital Consistency Transfer), a two-stage framework that facilitates pathology-aware medical image synthesis with strong anatomical fidelity. We specifically target abdominal CT and MRI data, aiming to generate clinically relevant synthetic datasets that capture both normal anatomy and intricate pathological details. By providing robust augmentation material, our approach holds the potential to improve downstream tasks like segmentation and diagnosis, while alleviating issues of data scarcity and privacy constraints in medical imaging. Our contributions are:", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22772", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ccde6e94d4fd756b4cdc766a3dfee62e3782a1ac71e865a15b961163b194dacf.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed ViCTr methodology." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 229 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Novel Two-Stage Framework. We propose ViCTR, a method that fuses anatomical consistency and pathological realism for CT and MRI synthesis with high diversity.", + "- Tweedie's Formula in Rectified Flow. We introduce a rectified flow trajectory reformulation of Tweedie's formula, ensuring accurate initialization and reducing sampling bias.", + "- Wide Applicability. ViCTR integrates seamlessly with various diffusion models, supporting a broad range of medical imaging tasks.", + "- One-Step Sampling & Tweedie's Corrections. We reduce computational overhead by cutting the number of diffusion steps from dozens to a handful, expediting inference without sacrificing quality.", + "- Quantitative Improvements in Segmentation. We demonstrate that adding ViCTR-synthesized CT/MRI images to real-world training sets significantly boosts segmentation performance, underscoring the practical benefits of pathology-focused data augmentation.", + "- Enhanced Image Fidelity. ViCTR achieves consistently lower FID (Fréchet Inception Distance) scores across multiple datasets, indicating superior structural and textural coherence.", + "- First Abdominal MRI Pathology Synthesis. To our knowledge, we are the first to generate abdominal MRI pathologies with progressive severity, offering new possibilities for research and clinical applications in medical imaging." + ], + "bbox": [ + 89, + 279, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By emphasizing both anatomical fidelity and pathological variety, ViCTr presents a powerful step toward bridging the gap between limited medical imaging data and the demands of high-performing AI models.", + "bbox": [ + 89, + 704, + 482, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 777, + 240, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Generative Models for Medical Data Augmentation", + "text_level": 1, + "bbox": [ + 89, + 801, + 482, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic images have long been explored for data augmentation in medical imaging, with Generative Adversarial Networks (GANs) [25, 36, 50] initially offering promising results in tasks such as lesion synthesis and modality trans", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lation. However, GAN-based methods often struggle with mode collapse, requiring extensive architectural and training refinements to ensure sufficient diversity and realism in the synthesized images [12, 35].", + "bbox": [ + 511, + 279, + 905, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion Models in Medical Imaging", + "text_level": 1, + "bbox": [ + 511, + 352, + 834, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based approaches have emerged as a powerful alternative to GANs for high-quality synthetic image generation [4, 5, 13, 49]. Their inherent noise-to-image paradigm provides a more stable training regime and can produce richer data variations. In medical imaging, diffusion models [11] have been leveraged to improve segmentation and classification performance by generating diverse training samples [1, 10, 37, 46, 47]. Recent specialized frameworks, including segmentation-guided diffusion [26] and text-driven generation built upon RadImageNet [34, 54], highlight the adaptability of diffusion models to specific clinical scenarios. Efforts such as MixUp-enhanced augmentation [6, 28] and domain adaptation methods [55] demonstrate that diffusion techniques can mitigate biases and improve generalization in tasks like classification and risk prediction.", + "bbox": [ + 511, + 375, + 906, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. General-Purpose Diffusion for Data Augmentation", + "text_level": 1, + "bbox": [ + 511, + 614, + 903, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various general-purpose diffusion strategies have been applied to further enrich training datasets. DiffuseMix [21] employs conditional prompts to blend real and synthetic data while preserving labels, whereas DreamDA [18] combines diffusion-based perturbations with pseudo-labeling for semantic consistency. DetDiffusion [51] incorporates object-detection attributes into the diffusion process, while Effective Data Augmentation with Diffusion Models [48] explores diffusion-based techniques to boost performance in few-shot settings. These methods generally aim to enhance data diversity and improve the training of downstream models.", + "bbox": [ + 511, + 652, + 906, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.4. Rectified Flow Models", + "text_level": 1, + "bbox": [ + 511, + 847, + 720, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While traditional diffusion approaches rely on discretized noise schedules, Rectified Flow (ReFlow) models learn", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22773", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/29b383118786bec909fe7685f9ea896cf7d17c363f68fde4ed195210878dcebe.jpg", + "image_caption": [ + "Figure 2. Comparison of synthetic data distributions from various methods against the ground-truth (GT) on the AMOS dataset. The GT distribution (blue) reflects real data, while Reflow (red), Flow Straight and Fast (green), and Distribution Matching Distillation (purple) show varying alignment. Our proposed method, ViCTr (brown) achieves the closest match to GT, demonstrating superior distribution alignment and fidelity." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "smooth, near-linear trajectories to map between data distributions. This can result in more efficient sampling and reduced computational overhead. Recent works have leveraged rectified flows for improved convergence rates and lower numbers of function evaluations (NFE) [29, 30, 53], although most focus on general distribution transport rather than the strict alignment and pathology-aware realism needed in medical imaging. As shown in Figure 2, different ReFlow-based methods exhibit varying degrees of distribution matching, with our approach (ViCTr) offering closer alignment to the ground-truth distribution.", + "bbox": [ + 88, + 380, + 482, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.5. Domain-Specific Adaptation", + "text_level": 1, + "bbox": [ + 89, + 559, + 344, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain-focused adaptations continue to evolve. DiNO-Diffusion [24] introduces a self-supervised latent diffusion model to cope with limited annotated medical data, and Diverse Data Augmentation with Diffusions [17] enriches domain generalization using Stable Diffusion combined with cosine similarity filtering. Collectively, these methods illustrate the growing interest in customized diffusion pipelines for specialized domains, especially when high-quality annotated data are scarce.", + "bbox": [ + 89, + 582, + 482, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements have further showcased the potential of diffusion models for domain-specific adaptation. DiNO-Diffusion [24] addresses the challenge of limited annotated data in medical imaging by leveraging a self-supervised latent diffusion framework. Similarly, Diverse Data Augmentation with Diffusions [17] enhances domain generalization by integrating Stable Diffusion with cosine similarity-based filtering, enabling the generation of semantically diverse and high-quality data. However, these methods prioritize general-domain efficiency over the anatomical-pathological alignment critical for medical imaging, often failing to preserve fine-grained structures", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "like vascular networks or diffuse fibrosis patterns.", + "bbox": [ + 511, + 90, + 841, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we address the critical need for efficient, high-fidelity pathology-aware synthesis in abdominal imaging by merging rectified flow concepts with a Tweedie-corrected diffusion process. Our approach (ViCTr) stands out from prior methods by introducing a linearized sampling framework that enables one-step generation while maintaining anatomical fidelity—a crucial requirement for robust data augmentation in clinical applications.", + "bbox": [ + 511, + 107, + 906, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 511, + 246, + 612, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework, ViCTr, uses Rectified Flow [31] and Tweedie's Formula [43] for high-fidelity medical image synthesis with efficient one-step sampling. By rectifying the trajectory from a prior distribution $p_0$ to the target medical image distribution $p_{\\mathrm{target}}$ , we reduce sampling bias and retain key anatomical and pathological details.", + "bbox": [ + 511, + 272, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rectified Flow Trajectory", + "text_level": 1, + "bbox": [ + 511, + 376, + 746, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classical diffusion approaches gradually corrupt data samples $x \\in \\mathbb{R}^d$ into a Gaussian prior $p_0$ . While these methods have shown success, they often suffer from suboptimal sampling paths, cascading prediction errors, and diminished diversity in generated samples. We address these issues through Rectified Flow optimization [16, 31], which learns a continuous velocity field guiding the forward and reverse diffusion processes more directly.", + "bbox": [ + 511, + 398, + 905, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $(x_0, x_1)$ be a sample pair with $x_0 \\sim p_0$ and $x_1 \\sim p_{\\mathrm{target}}$ . We define an interpolated point $x_t = (1 - t)x_0 + tx_1$ where $t \\in [0, 1]$ . A velocity model $v_\\theta : \\mathbb{R}^d \\times [0, 1] \\to \\mathbb{R}^d$ then predicts how to transition from $x_t$ toward $x_1$ . We train $v_\\theta$ by minimizing:", + "bbox": [ + 511, + 522, + 905, + 598 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\theta} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t \\sim \\operatorname {U n i f o r m} (0, 1)} \\left[ \\| \\left(x _ {1} - x _ {0}\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\| ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 625, + 903, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ensuring the predicted flow $v_{\\theta}(x_t, t)$ aligns with the true path $(x_1 - x_o)$ . Once trained, the rectified flow is realized by solving the ODE:", + "bbox": [ + 511, + 670, + 906, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd x _ {t} = v _ {\\hat {\\theta}} \\left(x _ {t}, t\\right) d t, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 731, + 903, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "leading $x_0 \\sim p_0$ toward $x_1 \\sim p_{\\mathrm{target}}$ . To enable one-step sampling, we distill this multi-step process into a neural network $\\hat{\\mathcal{N}}: \\mathbb{R}^d \\to \\mathbb{R}^d$ , such that $\\tilde{\\mathcal{T}}(x_0) = x_0 + v(x_0, 0)$ , is trained to directly predict $x_1$ from $x_0$ . The loss function,", + "bbox": [ + 511, + 761, + 906, + 823 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} \\left[ \\left\\| \\left(x _ {1} - x _ {0}\\right) - v \\left(x _ {0}, 0\\right) \\right\\| ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 834, + 903, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "drives $\\mathcal{T}(\\hat{x}_0)$ to approximate the final, rectified state $x_{1}$", + "bbox": [ + 511, + 883, + 883, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22774", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Rectified Flow with Tweedie's Formula", + "text_level": 1, + "bbox": [ + 89, + 90, + 426, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Though rectified flow refines sampling, medical data distributions often need extra bias correction. Tweedie's Formula [43] addresses this by adjusting noisy observations to better approximate the posterior mean. For Gaussian variables $z \\sim \\mathcal{N}(\\mu_z, \\Sigma_z)$ , Tweedie's formula indicates:", + "bbox": [ + 89, + 112, + 483, + 189 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\mu_ {z} \\mid z ] = z + \\Sigma_ {z} \\nabla_ {z} \\log p (z)\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 198, + 392, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": ", where $\\nabla_z\\log p(z)$ is the gradient of the log probability. In typical diffusion, Tweedie's formula estimates $x_0$ from noisy $x_{t}$ , guiding predictions toward the data manifold.", + "bbox": [ + 89, + 224, + 482, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ViCTR incorporates this correction into the rectified flow ODE:", + "bbox": [ + 89, + 270, + 482, + 297 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd x _ {t} = v _ {\\hat {\\theta}} (x _ {t}, t) d t + (1 - \\bar {\\alpha} _ {t}) \\nabla_ {x _ {t}} \\log p (x _ {t}) d t,\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 310, + 439, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and updates the training objective to:", + "bbox": [ + 89, + 335, + 336, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\theta} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t \\sim \\operatorname {U n i f o r m} (0, 1)} \\left[ \\| \\left(x _ {1} - x _ {0}\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\right. \\\\ \\left. \\left. - \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {x _ {t}} \\log p (x _ {t}) \\right| ^ {2} \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 359, + 480, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\bar{\\alpha}_t$ denotes the cumulative product of variance decay factors in the diffusion schedule. The additional Tweedie term $(1 - \\bar{\\alpha}_t)\\nabla_{x_t}\\log p(x_t)$ corrects for sampling bias, driving $x_{t}$ more accurately toward the target distribution.", + "bbox": [ + 89, + 436, + 483, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We similarly extend the one-step distillation to integrate Tweedie's correction:", + "bbox": [ + 89, + 497, + 482, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {T}} (x _ {0}) = x _ {0} + v (x _ {0}, 0) + (1 - \\bar {\\alpha} _ {0}) \\nabla_ {x _ {0}} \\log p (x _ {0}),\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 536, + 452, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and optimize", + "bbox": [ + 89, + 563, + 181, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} \\left[ \\left\\| x _ {1} - \\hat {\\mathcal {T}} (x _ {0}) \\right\\| ^ {2} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 585, + 375, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This single-step approach balances computational efficiency with high-quality, pathology-aware generation—critical in medical applications where datasets are limited and synthetic realism is paramount.", + "bbox": [ + 89, + 628, + 482, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, ViCTr unifies rectified flow and Tweedie's correction to deliver anatomically consistent, pathology-aware sampling in a single forward pass. The resulting framework exhibits reduced inference costs, minimized sampling bias, and improved fidelity for generating high-resolution medical images.", + "bbox": [ + 89, + 688, + 483, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pathology Aware Image Synthesis", + "text_level": 1, + "bbox": [ + 89, + 787, + 385, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed ViCTr method adopts a two-stage training paradigm: Stage 1 establishes a foundational diffusion model tailored to medical imaging data, and Stage 2 fine-tunes this pre-trained model for downstream tasks, including semantic-guided multi-modal generation and counterfactual pathology synthesis.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training on ATLAS-8k Dataset (Stage 1)", + "text_level": 1, + "bbox": [ + 513, + 90, + 826, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the absence of large-scale, domain-specific pre-trained diffusion models, we begin by training on the ATLAS8k [40] dataset, which comprises abdominal CT scans and their segmentation annotations. We leverage these annotations as conditional guidance—enabling precise control over anatomical structures—and integrate textual prompts (e.g., \"create image having $\\langle \\langle \\text{organs} \\rangle \\rangle$ \" to further refine semantic consistency.", + "bbox": [ + 511, + 109, + 906, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Representations. Following Figure 1, we feed raw CT images $(X_I)$ and segmentation masks $(X_S)$ into a frozen VAE encoder, producing latent embeddings $(Z_o$ and $Z_s)$ . Simultaneously, textual prompts $(X_p)$ are processed by a pre-trained text encoder (details in Table 1) to yield prompt embeddings $(Z_p)$ . These latent representations guide the generative diffusion backbone $\\phi_{\\theta}$ during both forward and reverse diffusion.", + "bbox": [ + 511, + 231, + 906, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Forward Diffusion. We progressively inject noise into $Z_{o}$ forming a noisy representation $Z_{t}$ as", + "bbox": [ + 513, + 352, + 905, + 382 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (Z _ {t} | Z _ {o}) = (1 - t) \\cdot Z _ {o} + t \\cdot \\epsilon_ {\\mathrm {t r u e}},\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 392, + 826, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{true}$ is the true noise at step $t$ . We concatenate $Z_{t}$ with $Z_{s}$ and feed them into the diffusion model $\\phi_{\\theta}$ . Crucial layers are selectively unfrozen based on elastic weight consolidation, maintaining model stability while adapting to medical domain specifics.", + "bbox": [ + 511, + 419, + 905, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reverse Diffusion. In the reverse process,", + "bbox": [ + 532, + 494, + 815, + 510 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(Z _ {t - 1} \\mid Z _ {t}, Z _ {s}, Z _ {p}, t\\right) = Z _ {t} + \\delta T \\times \\phi_ {\\theta} \\left(Z _ {t}, Z _ {s}, Z _ {p}, t\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 521, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "iteratively removes noise to reconstruct $Z_{t-1}$ conditioned on the segmentation map $(Z_s)$ and text prompt $(Z_s)$ . The diffusion loss $L_{diff}$ between $Z_{t-1}$ and $\\epsilon_{true}$ is:", + "bbox": [ + 511, + 547, + 905, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i f f}} = - \\left| \\phi_ {\\theta} \\left(Z _ {t}, Z _ {s}, Z _ {p}, t\\right) - \\left(\\epsilon_ {\\text {t r u e}} - Z _ {o}\\right) \\right| ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 604, + 905, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After refining $Z_{t-1}$ via Tweedie's formula, the denoised latent $Z_o$ is fed into the frozen VAE decoder (matching $\\phi_\\theta$ ; see Table 1) to obtain a reconstructed CT image.", + "bbox": [ + 511, + 633, + 905, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Composite Loss. To ensure both local accuracy and global perceptual realism, we combine: Diffusion Loss $(L_{\\mathrm{diff}})$ , pixel-level reconstruction loss $(L_2)$ evaluating the gap between generated $X_{ig}$ and ground-truth $X_I$ , and structural similarity (SSIM) loss $(L_{SSIM})$ emphasizing higher-order textural correspondence. Optimizing this composite loss enables the diffusion backbone $\\phi_{\\theta}$ to capture essential anatomical features, establishing a versatile medical foundation model ready for fine-tuning.", + "bbox": [ + 511, + 679, + 906, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-tuning on downstream tasks (Stage 2)", + "text_level": 1, + "bbox": [ + 511, + 821, + 815, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on the Stage1 pre-trained model, Stage2 targets specialized tasks like semantic-guided CT/MRI generation and counterfactual pathology synthesis. As shown in Figure 1, a dual-network setup is adopted.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22775", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $\\phi_{\\text{base}}$ remains frozen, retaining the robust anatomical knowledge acquired in Stage 1.", + "2. $\\phi_{\\text {adapt }}$ is augmented with LoRA [19] modules within the previously fine-tuned layers. These modules are the only trainable parameters, ensuring targeted and stable adaptation." + ], + "bbox": [ + 89, + 90, + 480, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Input Consistency Loss. Similar to Stage1, each training sample includes prompts, segmentation maps, and either CT or MRI images (depending on the task). A frozen VAE encoder extracts latent representations $Z_{s}, Z_{o}$ , and $Z_{p}$ . Following forward diffusion of $Z_{o}$ into noisy $Z_{t}$ , both $\\phi_{\\text{base}}$ and $\\phi_{\\text{adapt}}$ process $[Z_{t}, Z_{s}]$ with text embedding $Z_{p}$ . We introduce a consistency loss $L_{\\text{consistency}}$ to align intermediate outputs of $\\phi_{\\text{base}}$ and $\\phi_{\\text{adapt}}$ , ensuring stable adaptation without drifting from core representations.", + "bbox": [ + 89, + 186, + 482, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Components. Diffusion loss $(L_{diff}^{\\mathrm{base}}$ , $L_{diff}^{\\mathrm{adapt}})$ encourages accurate noise prediction. Temporal consistency loss $L_{temporal}$ preserves smooth transitions across time steps in the reverse diffusion. Spatial consistency loss $L_{spatial}$ ensures alignment in the output reconstructions $(Z_{t-1}^{base}, (Z_{t-1}^{adapt})$ . Pixel-level losses $(L_2$ and $L_{SSIM})$ compare generated results to ground-truth scans, enforcing both local fidelity and structural coherence.", + "bbox": [ + 89, + 325, + 482, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CT and MRI Generation. We adapt to BTCV [15] (CT) and AMOS [23] (MRI) datasets, using organ segmentation masks for prompts. A standard 80/10/10 partition ensures robust evaluation.", + "bbox": [ + 89, + 454, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pathology Generation. For liver cirrhosis, we employ CurrMRI600+ [22] dataset (T1/T2 MRIs) with segmentation masks marking liver regions. Prompts specify \"low,\" \"mild,\" or \"severe\" cirrhosis intensity. As in the CT/MRI setting, the training/validation/testing split is 80/10/10. The inference details and inference diagram are in the supplementary material.", + "bbox": [ + 89, + 518, + 482, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This selective LoRA-based fine-tuning allows ViCTR to handle complex tasks—ranging from normal organ generation to pathology synthesis—while preserving the anatomical and text-driven guidance learned in Stage 1. The experiments (Section 4) demonstrate that this two-stage framework consistently yields high-quality, clinically realistic images in both CT and MRI settings, including nuanced liver cirrhosis modeling.", + "bbox": [ + 89, + 628, + 482, + 750 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2e5e90ccb33f0594f2e211f514deffece000b187c93d01c18f053ef3c05b10cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Diffusion MethodsDenoiserText Encoder
Stable Diffusion [44]UNet [45]Clip-B/16 [41]
Pixart-alpha [9]Transformer (DiT) [38]T5-XXXXL [42]
Stable Diffusion XL [39]Dual Unet [39]Clip-L/14
Flux [27]MultiModal Transformer [14]T5-XXXXL + Clip-L/15
Stable Diffusion-3 [14]MultiModal TransformerT5-XXXXL + Clip-B/16 + Clip-L/14
", + "bbox": [ + 91, + 782, + 513, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Selection of denoisers and text encoders for different diffusion methods ${\\phi }_{\\theta }$", + "bbox": [ + 89, + 868, + 482, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 513, + 90, + 746, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The implementation details for training and sampling are provided in the supplementary material.", + "bbox": [ + 511, + 114, + 903, + 145 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 513, + 152, + 705, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic Data Generation Results. Table 2 summarizes our comparative evaluation of synthetic medical image generation under vanilla fine-tuning versus the proposed ViCTR pipeline, spanning multiple datasets (BTCV, AMOS, Cir-rMRI600+) and diffusion backbones (Stable Diffusion, SD-XL, SD-3, Pixart-alpha, Flux). While vanilla fine-tuning applies a direct end-to-end approach after ATLAS-8k pretraining, it generally yields higher Fréchet Inception Distance (FID) and Medical FID (MFID) values across all datasets. Conversely, ViCTR integrates domain-specific loss functions and specialized architectural settings, consistently achieving lower FID/MFID scores and producing more realistic medical images.", + "bbox": [ + 511, + 175, + 903, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the well-known limitation of FID (which uses ImageNet features), we adopt M3D-CLIP [3] to compute a more domain-specific MFID. As shown in the table, ViCTR provides notable gains: for example, Stable Diffusion under ViCTR achieves FID/MFID of 21.98/19.02 (BTCV), 20.37/19.11 (AMOS), and 25.57/21.46 (Cir-rMRI600+). These results represent substantial improvements over vanilla fine-tuning and demonstrate consistent performance boosts across all tested architectures. Such gains highlight the robustness of ViCTR's two-stage framework in generating clinically meaningful synthetic images for both CT and MRI modalities.", + "bbox": [ + 511, + 372, + 903, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segmentation Results. To assess the quality and utility of ViCTR-generated synthetic data, we performed segmentation experiments on three datasets—BTCV, AMOS, and CurrMRI600+—tracking mean Dice Similarity Coefficient (mDSC) and mean Hausdorff Distance 95 (mHD95). Higher mDSC and lower mHD95 respectively indicate improved overlap accuracy and spatial precision. We examined four training configurations: (1) baseline using original data only, (2) standard augmentation (Random Crop, Rotate, Blur, Affine, Geometric Distortion at 0.4 probability), (3) $30\\%$ synthetic data via vanilla fine-tuning, and (4) $30\\%$ synthetic data from ViCTR. Table 3 highlights consistent performance gains from ViCTR across all datasets and backbone architectures. Notably, CurrMRI600+—a liver cirrhosis dataset—exhibited the most significant improvements, reflecting ViCTR's capacity to capture complex pathological features. These results confirm that ViCTR-generated images significantly boost segmentation metrics (mDSC and mHD95), highlighting the practical value of our two-stage framework for medical data augmentation.", + "bbox": [ + 511, + 553, + 903, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficiency of ViCTR. Table 2 summarizes the number of diffusion steps and average inference times (in seconds) for vanilla fine-tuning versus the ViCTR framework. Across all", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22776", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a89e077bf36c8f0e36f41f6e4c2714390fff6922da25f7f2eb632d3e71c9732d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselinesBTCV Dataset (CT Generation)AMOS Dataset (MRI Generation)CirrMRI600+ (Pathology Generation)Diffusion Steps | Inference Time
Fine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTR
Stable Diffusion [44]25.44 / 19.6721.98 / 19.0225.43 / 21.7620.37 / 19.1128.34 / 23.4325.57 / 21.4640 | 13.764 | 3.12
Stable Diffusion XL [39]23.47 / 18.2120.33 / 17.4424.11 / 20.2319.44 / 18.4527.34 / 22.1124.02 / 20.7630 | 14.554 | 3.45
Stable Diffusion-3 [14]19.07 / 16.2217.37 / 16.0222.32 / 19.7618.02 / 19.0824.49 / 21.7821.28 / 19.3450 | 18.983 | 2.78
Pixart-alpha [9]21.32 / 17.0919.22 / 16.9623.78 / 20.0418.76 / 18.5626.06 / 20.0723.04 / 18.9225 | 10.673 | 1.74
Flux [27]15.52 / 15.0113.28 / 14.0819.02 / 18.2815.55 / 16.5822.46 / 18.8819.96 / 17.0130 | 15.663 | 2.87
", + "bbox": [ + 91, + 87, + 978, + 164 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/896a7a5a02b0846a4023b5694836fe5a196965db1ec2c2a85690fbcec07d828f.jpg", + "table_caption": [ + "Table 2. Quantitative results on CT, MRI, and Cirrhosis generation. All baselines were pre-trained on ATLAS-8k and fine-tuned on target datasets. Metrics reported are Fréchet Inception Distance (FID) and Medical FID (MFID), shown as FID/MFID, with inference time in seconds. Lower values indicate better performance." + ], + "table_footnote": [], + "table_body": "
BaselinesOriginal Dataset (Org.)AugmentationOrg. + Synth. by FLUX 30% (vanilla Fine-tuning)Org. + Synth. by FLUX 30% (ViCtR)
mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)
BTCV dataset Segmentation results
UNet [45]76.7234.4278.4533.1779.3232.3781.2230.17
TransUnet [8]85.5232.3387.0131.4387.5430.1189.7829.12
nnUnet [20]80.4830.1982.5431.0283.3729.0185.1927.77
nnFormer [56]83.4730.0183.9830.8885.8828.7887.7226.22
MedSegDiff [52]87.9127.6788.6526.7289.7825.5291.9223.31
AMOS dataset Segmentation results
UNet [45]68.9234.5770.5532.3271.0231.7873.3429.11
TransUnet [8]71.3333.1272.5631.0973.4329.1977.5427.56
nnUnet [20]73.3332.3274.1130.1975.6828.7578.2926.32
nnFormer [56]75.7831.1976.4429.7677.7727.5781.3224.41
MedSegDiff [52]76.8329.9278.3826.9979.0325.4584.0222.18
CirrMRI600+ dataset Segmentation results
UNet [45]68.7436.7369.3835.4370.1235.1173.3932.11
TransUnet [8]70.7735.4270.9834.0171.7833.9274.5631.09
nnUnet [20]71.0234.3572.4932.7873.5631.2778.8930.25
nnFormer [56]74.8833.7875.2331.8876.4531.0979.4429.78
MedSegDiff [52]76.9230.7977.1130.3478.0329.8981.3727.34
", + "bbox": [ + 91, + 229, + 942, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Segmentation results on BTCV, AMOS, and CurrMRI600+ using various training settings and baselines. Metrics reported are mDSC (%) and mHD95 (mm). Top three results per setting are highlighted: blue (1st), green (2nd), and orange (3rd).", + "bbox": [ + 89, + 508, + 906, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluated models, ViCTR consistently reduces the required steps, thereby shortening inference duration without compromising image fidelity. This efficiency stems from two key innovations: (1) Tweedy's formula to approximate $Z_{0}$ and streamlines the reconstruction process, and (2) Two-Stage Training, enabling the model to learn both anatomical consistency and pathology-specific variations in a targeted manner. By minimizing extraneous steps, ViCTR can generate high-quality synthetic images more rapidly, making it a viable solution for resource-constrained clinical or research environments where computational overhead is a critical consideration.", + "bbox": [ + 88, + 564, + 482, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 756, + 272, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted extensive qualitative evaluations of ViCTr across multiple modalities (CT and MRI) and tasks, covering both anatomical and pathological image synthesis.", + "bbox": [ + 89, + 777, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Anatomical and Pathology-Driven Image Generation. Unlike prior methods that primarily generate normal anatomical scans or isolated tumors, ViCTr extends synthesis capabilities to complex pathologies, such as liver cirrhosis. Figure 3 shows pairs of non-cirrhotic MRIs alongside", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "their segmentation masks and the corresponding ViCTR-generated cirrhotic images. Notably, the synthesized images retain essential anatomical structures while introducing realistic cirrhotic texturing, an advancement that enables more diverse data augmentation and supports research into disease progression. These high-fidelity synthetic samples offer potential to enrich clinical training datasets, advance diagnostic algorithms, and support detailed analyses of pathology progression.", + "bbox": [ + 511, + 564, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Segmentation Outcomes. To further showcase ViCTR's impact on downstream tasks, we visually assessed segmentation quality on BTCV and AMOS (Figure 6). Focusing on regions like the left kidney and pancreas, models trained with ViCTR-augmented data show sharper boundaries and more precise delineation than standard fine-tuning, indicating improved spatial learning.", + "bbox": [ + 511, + 702, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparative Performance on BTCV and AMOS. Figures 4 (BTCV) and 5 (AMOS) compare diffusion-based image generation results using Stable Diffusion v3 (SD-3) and Flux backbones. On BTCV—a dataset with limited samples and larger spatial dimensions—vanilla finetuning with SD-3 often yields poor mask adherence, sug", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22777", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acbff25d698bee988afcb9bd7ad92bcea0f82dd02abecb03a68b130a14859499.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 478, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b4919450eea532967b234637106888ca96f72a6a0afc07acf66fdedeb80ccae.jpg", + "image_caption": [ + "Figure 3. Examples of synthetic cirrhotic images with different severity levels generated from non-cirrhotic scans using ViCTr. For T1-MRI, we show the segmentation mask, the original non-cirrhotic image, and the corresponding synthetic cirrhotic image." + ], + "image_footnote": [], + "bbox": [ + 96, + 340, + 480, + 540 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5590f8ff02fb0f7fd96cf80c038fe2b42c2fbfee3326fd783f47ea34103b500d.jpg", + "image_caption": [ + "Figure 4. Diffusion-Based Image Generation on BTCV MRI Dataset. From left to right: segmentation mask, SD-3 with standard fine-tuning, FLUX with standard fine-tuning, SD-3 with ViCTR, FLUX with ViCTR.", + "Figure 5. Diffusion-Based Image Generation on AMOS Dataset. From left to right: segmentation mask, SD-3 with Vanilla fine-tuning, SD-3 with our method, FLUX with Vanilla fine-tuning, FLUX with our method." + ], + "image_footnote": [], + "bbox": [ + 96, + 607, + 480, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gesting difficulties in learning accurate spatial constraints. In contrast, the same model shows markedly improved mask alignment on the AMOS dataset, attributed to its larger training set and more manageable image dimensions.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regardless of these dataset-specific variations, ViCTr achieves consistently high-quality outputs with robust mask adherence and anatomical accuracy across both BTCV and AMOS, underscoring its enhanced spatial learning and resilience. By maintaining performance even under data-sparse conditions, ViCTr demonstrates a clear advantage over conventional fine-tuning approaches, which can exhibit greater sensitivity to dataset size and complexity.", + "bbox": [ + 511, + 152, + 903, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, we evaluate the impact of ViCTR on segmentation tasks, illustrated in Figure 6. Both BTCV and AMOS datasets encompass various abdominal structures (e.g., left kidney, pancreas), and our results highlight how ViCTR-based training yields notably more precise organ boundaries compared to alternative methods. This improvement is particularly evident in challenging regions where standard segmentation often struggle, highlighting ViCTR's strength in enhancing spatial delineation and anatomical accuracy.", + "bbox": [ + 511, + 272, + 906, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 415, + 676, + 429 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb287f5517509101310520e15534411eaf9ab1f7efdaebcb58e4a63e24616151.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Ablation Based on Pre-training of Model
ViCTR (Without Stage-1)17.33
Ablation Based on Rectified Flow Algorithms
ViCTR (Without Proposed Tweedies Formula)18.78
ViCTR (With Reflow)20.19
ViCTR (With Flow Straight and Fast)21.37
ViCTR (With Distribution Matching Distillation)22.33
Ablation Study Based on Loss FunctionsAblation Based on LoRA Rank
Ldiff18.77r = 818.46
Ldiff + Lspatial18.21r = 1617.52
Ldiff + Lconsistency17.02r = 3216.66
Ldiff + Lspatial + Lconsistency15.55r = 6415.55
", + "bbox": [ + 514, + 445, + 901, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study for the effect of model pretraining, loss functions, and LoRA rank settings are shown.", + "bbox": [ + 511, + 583, + 903, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To rigorously validate our ViCTr framework, we conducted a series of ablation experiments on the BTCV dataset (Table 4), focusing on four key aspects: pretraining, rectified flow algorithms, loss function components, and LoRA rank selection. We used FID (Frechet Inception Distance) as our primary metric to gauge how each architectural choice affects both anatomical fidelity and pathological realism.", + "bbox": [ + 511, + 628, + 903, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Pretraining. We first examined the necessity of Stage 1 pretraining by comparing our full, two-stage pipeline to a variant that proceeds directly to downstream fine-tuning. Omitting Stage 1 degrades the model's FID score from 15.55 to 17.33 (Table 4), underscoring the vital role of an anatomical prior. These findings affirm that establishing robust structural representations in Stage 1 is crucial for achieving high-quality medical image synthesis.", + "bbox": [ + 511, + 750, + 905, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rectified Flow Algorithms. We next assessed our rectified flow formulation, augmented with Tweedie's formula,", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22778", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f77422e60a89252daac2dd04e67c939d78ee91f5c636912898c7b360e493bc01.jpg", + "image_caption": [ + "Figure 6. Segmentation results on BTCV and AMOS datasets' left kidney and pancreas. We compare the ground truth (blue), models trained on the default dataset, augmented dataset, default dataset plus $30\\%$ synthetic data generated by FLUX with standard fine-tuning, and default dataset plus $30\\%$ synthetic data generated by FLUX with ViCTr (red)." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 906, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "against alternative rectified flow-based models, including ReFlow, Flow Straight and Fast, and Distribution Matching Distillation. Removing Tweedie's correction alone increased the FID to 18.78, highlighting its significance for trajectory alignment. Across all baselines, we observed FID deteriorations, ranging from 13.28 to as high as 22.33, thereby demonstrating that our proposed method provides tighter distribution matching and superior synthesis fidelity.", + "bbox": [ + 88, + 419, + 480, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study Based on Loss Functions. We then investigated each ViCTr loss component—diffusion loss, spatial consistency, and consistency loss—to determine their relative contributions: the baseline model, utilizing only diffusion loss, produced an FID of 18.77 (Table 4). Incorporating spatial consistency loss yielded a moderate improvement to 18.21, highlighting its role in maintaining structural integrity. The addition of consistency loss further enhanced performance, reducing the FID to 17.02 by promoting image coherence across generations. When combining all three terms, we achieved an FID of 15.55, indicating that these components collectively offer significant improvements in anatomical accuracy and visual realism.", + "bbox": [ + 88, + 544, + 482, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Based on LoRA Rank. Finally, we explored how varying the LoRA rank $(r)$ influences generative performance. Starting at $r = 8$ and incrementing upwards, we found $r = 64$ to yield the best FID at 15.55. This result suggests that higher-dimensional adaptation spaces enable more nuanced parameter updates during fine-tuning, thereby improving visual quality and fidelity.", + "bbox": [ + 88, + 744, + 482, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Turing Tests. To further validate clinical plausibility, three radiologists participated in Visual Turing Tests using 15 randomly generated MRI scans depicting vary", + "bbox": [ + 88, + 854, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing levels of liver cirrhosis (mild, moderate, severe). All scans were uniformly judged to be clinically realistic, with identical outcomes observed when using additional random samples. Notably, the synthetic cirrhotic images correctly exhibited surface nodularity and textural irregularities, consistent with radiologic findings in mild-to-severe cirrhosis (Figure 3). These results confirm ViCTr's ability to synthesize pathology-specific features, enabling applications in training, algorithm development, and clinical research.", + "bbox": [ + 511, + 419, + 906, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 574, + 633, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented ViCTR, a novel two-stage framework designed to generate high-fidelity medical images by integrating robust anatomical pre-training with precise pathology-specific fine-tuning. By aligning Tweedie's formula with linear projection methods used in flow matching, ViCTR maintains accurate initial distribution estimates even amid diffusion processes. This setup is further enhanced by LoRA adapters, which preserve essential anatomical information while flexibly adapting to diverse pathologies. Experimental results demonstrate that ViCTR consistently surpasses conventional fine-tuning strategies, reducing FID scores and producing clinically realistic outputs. Beyond data augmentation for segmentation and classification, ViCTR can be extended to tasks such as modality translation, contrast synthesis, and professional training, offering a powerful and versatile tool for advancing AI-driven medical imaging.", + "bbox": [ + 511, + 601, + 906, + 859 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code: https://github.com/Onkarsus13/ViCTR-2D", + "bbox": [ + 535, + 875, + 875, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Weights: https://huggingface.co/onkarsus13/ViCTR-2D", + "bbox": [ + 514, + 887, + 897, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22779", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by the following NIH grants: R01-HL171376 and U01-CA268808.", + "bbox": [ + 89, + 114, + 482, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hazrat Ali, Shafaq Murad, and Zubair Shah. Spot the fake lungs: Generating synthetic medical images using neural diffusion models. In Artificial Intelligence and Cognitive Science, pages 32-39, Cham, 2023. Springer Nature Switzerland. 2", + "[2] Alhanoof Althnian, Duaa Alsaeed, Heyam Al-Baity, Amani Samha, Alanoud Bin Dris, Najla Alzakari, Afnan Abou Elwafa, and Heba Kurdi. Impact of dataset size on classification performance: an empirical evaluation in the medical domain. Applied Sciences, 11(2):796, 2021. 1", + "[3] Fan Bai, Yuxin Du, Tiejun Huang, Max Q. H. Meng, and Bo Zhao. M3d: Advancing 3d medical image analysis with multi-modal large language models, 2024. 5", + "[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 2", + "[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 2", + "[6] Luigi Carratino, Moustapha Cisse, Rodolphe Jenatton, and Jean-Philippe Vert. On mixup regularization. Journal of Machine Learning Research, 23(325):1-31, 2022. 2", + "[7] Junyi Chai, Hao Zeng, Anming Li, and Eric W.T. Ngai. Deep learning in computer vision: A critical review of emerging techniques and application scenarios. Machine Learning with Applications, 6:100134, 2021. 1", + "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6", + "[9] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation, 2024. 5, 6", + "[10] Qi Chen, Xiaoxi Chen, Haorui Song, Zhiwei Xiong, Alan Yuille, Chen Wei, and Zongwei Zhou. Towards generalizable tumor synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11147-11158, 2024. 2", + "[11] Gayatri Deshmukh, Onkar Kishor Susladkar, Debesh Jha, Elif Keles, Halil Ertugrul Aktas, Daniela P Ladner, Amir A Borhani, Gorkem Durak, Ulas Bagci, et al. Meddelinea: Scalable and efficient medical image segmentation via controllable diffusion transformers. In Medical Imaging with Deep Learning, 2025. 2" + ], + "bbox": [ + 93, + 186, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[13] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E. Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 5, 6", + "[15] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619-3629, 2020. 5", + "[16] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Flux that plays music, 2024. 3", + "[17] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 3", + "[18] Yunxiang Fu, Chaoqi Chen, Yu Qiao, and Yizhou Yu. Dreamda: Generative data augmentation with diffusion models. arXiv preprint arXiv:2403.12803, 2024. 2", + "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5", + "[20] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 6", + "[21] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27621-27630, 2024. 2", + "[22] Debesh Jha, Onkar Kishor Susladkar, Vandan Gorade, Elif Keles, Matthew Antalek, Deniz Seyithanoglu, Timurhan Cebeci, Halil Ertugrul Aktas, Gulbiz Dagoglu Kartal, Sabahattin Kaymakoglu, et al. Cirrmri600+: Large scale mri collection and segmentation of cirrhotic liver. arXiv preprint arXiv:2410.16296, 2024. 5", + "[23] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems, 35:36722-36732, 2022. 5", + "[24] Guillermo Jimenez-Perez, Pedro Osorio, Josef Cersovsky, Javier Montalt-Tordera, Jens Hooge, Steffen Vogler, and Sadegh Mohammadi. Dino-diffusion. scaling medical diffusion via self-supervised pre-training. arXiv preprint arXiv:2407.11594, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22780", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Sunho Kim, Byungjai Kim, and HyunWook Park. Synthesis of brain tumor multicontrast mr images for improved data augmentation. Medical Physics, 48(5):2185-2198, 2021. 2", + "[26] Nicholas Konz, Yuwen Chen, Haoyu Dong, and Maciej A Mazurowski. Anatomically-controllable medical image generation with segmentation-guided diffusion models. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 88–98. Springer, 2024. 2", + "[27] Black Forest Labs. black-forest-labs/FLUX.1-dev $\\cdot$ Hugging Face. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. [Online; accessed 05-October-2024]. 5, 6", + "[28] Hansang Lee, Haeil Lee, and Helen Hong. Genmix: Combining generative and mixture data augmentation for medical image classification. arXiv preprint arXiv:2405.20650, 2024. 2", + "[29] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. Advances in Neural Information Processing Systems, 37:63082-63109, 2025. 3", + "[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3", + "[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 3", + "[32] Supriya V. Mahadevkar, Bharti Khemani, Shruti Patil, Ketan Kotecha, Deepali R. Vora, Ajith Abraham, and Lubna Abdelkareim Gabralla. A review on machine learning styles in computer vision—techniques and future directions. IEEE Access, 10:107293–107329, 2022. 1", + "[33] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M. Alvarez, Zhiding Yu, Sanja Fidler, and Marc T. Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 275-284, 2022. 1", + "[34] Xueyan Mei, Zelong Liu, Philip M Robson, Brett Marinelli, Mingqian Huang, Amish Doshi, Adam Jacobi, Chendi Cao, Katherine E Link, Thomas Yang, Ying Wang, Hayit Greenspan, Timothy Deyer, Zahi A Fayad, and Yang Yang. RadImageNet: An open radiologic deep learning research dataset for effective transfer learning. Radiol Artif Intell, 4 (5):e210315, 2022. 2", + "[35] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, Jakob Nikolas Kather, and Daniel Truhn. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13(1):12098, 2023. 2", + "[36] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part III 20, pages 417-425. Springer, 2017. 2" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Kaliprasad Pani and Indu Chawla. Synthetic mri in action: A novel framework in data augmentation strategies for robust multi-modal brain tumor segmentation. Computers in Biology and Medicine, 183:109273, 2024. 2", + "[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5", + "[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5, 6", + "[40] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024. 4", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5", + "[43] Herbert E. Robbins. An Empirical Bayes Approach to Statistics, pages 388-394. Springer New York, New York, NY, 1992. 3, 4", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 5, 6", + "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5, 6", + "[46] Daniel G. Saragih, Atsuhiro Hibi, and Pascal N. Tyrrell. Using diffusion models to generate synthetic labeled data for medical image segmentation. International Journal of Computer Assisted Radiology and Surgery, 19(8):1615-1625, 2024. 2", + "[47] Vanshali Sharma, Abhishek Kumar, Debesh Jha, M.K. Bhuyan, Pradip K. Das, and Ulas Bagci. Controlpolynet: Towards controlled colon polyp synthesis for improved polyp segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2325-2334, 2024. 2", + "[48] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2", + "[49] Brandon Trabucco, Kyle Doherty, Max A Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffu" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion models. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[50] Toan Tran, Trung Pham, Gustavo Carneiro, Lyle Palmer, and Ian Reid. A bayesian data augmentation approach for learning deep models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2", + "[51] Yibo Wang, Ruiyuan Gao, Kai Chen, Kaiqiang Zhou, Yingjie Cai, Lanqing Hong, Zhenguo Li, Lihui Jiang, DitYan Yeung, Qiang Xu, et al. Detdiffusion: Synergizing generative and perceptive models for enhanced data generation and perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7246-7255, 2024. 2", + "[52] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In Medical Imaging with Deep Learning, pages 1623-1639. PMLR, 2024. 6", + "[53] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 3", + "[54] Zheyuan Zhang, Lanhong Yao, Bin Wang, Debesh Jha, Elif Keles, Alpay Medetalibeyoglu, and Ulas Bagci. Emit-diff: Enhancing medical image segmentation via text-guided diffusion model. arXiv preprint arXiv:2310.12868, 2023. 2", + "[55] Yuan Zhong, Suhan Cui, Jiaqi Wang, Xiaochen Wang, Ziyi Yin, Yaqing Wang, Houping Xiao, Mengdi Huai, Ting Wang, and Fenglong Ma. MedDiffusion: Boosting Health Risk Prediction via Diffusion-based Data Augmentation, pages 499-507. 2", + "[56] Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Lequan Yu, Liansheng Wang, and Yizhou Yu. nnformer: Interleaved transformer for volumetric segmentation. arXiv preprint arXiv:2109.03201, 2021. 6" + ], + "bbox": [ + 91, + 90, + 482, + 598 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22782", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_model.json b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0912c4c3ad8e0e4d43a4cba0e113114c8dda6f1a --- /dev/null +++ b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "ViCTr: Vital Consistency Transfer for Pathology Aware Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.872, + 0.219 + ], + "angle": 0, + "content": "Onkar Susladkar\\(^{1,3}\\), Gayatri Deshmukh\\(^{1}\\), Yalcin Tur\\(^{2}\\), Gorkem Durak\\(^{1}\\), Ulas Bagci\\(^{1}\\) \n\\(^{1}\\)Northwestern University \\(^{2}\\)Stanford University \\(^{3}\\)University of Illinois Ubrana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.22, + 0.797, + 0.234 + ], + "angle": 0, + "content": "onkarsus13@gmail.com, dgayatri9850@gmail.com, yalcintr@stanford.edu," + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.238, + 0.753, + 0.251 + ], + "angle": 0, + "content": "gorkem.durak@northwestern.edu, ulas.bagci@northwestern.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.789 + ], + "angle": 0, + "content": "Synthesizing medical images remains challenging due to limited annotated pathological data, modality domain gaps, and the complexity of representing diffuse pathologies such as liver cirrhosis. Existing methods often struggle to maintain anatomical fidelity while accurately modeling pathological features, frequently relying on priors derived from natural images or inefficient multi-step sampling. In this work, we introduce ViCTR (Vital Consistency Transfer), a novel two-stage framework that combines a rectified flow trajectory with a Tweedie-corrected diffusion process to achieve high-fidelity, pathology-aware image synthesis. First, we pretrain ViCTR on the ATLAS-8k dataset using Elastic Weight Consolidation (EWC) to preserve critical anatomical structures. We then fine-tune the model adversarially with Low-Rank Adaptation (LoRA) modules for precise control over pathology severity. By reformulating Tweedie's formula within a linear trajectory framework, ViCTR supports one-step sampling—reducing inference from 50 steps to just 4—without sacrificing anatomical realism. We evaluate ViCTR on BTCV (CT), AMOS (MRI), and CirrMRI600+ (cirrhosis) datasets. Results demonstrate state-of-the-art performance, achieving a Medical Fréchet Inception Distance (MFID) of 17.01 for cirrhosis synthesis—28% lower than existing approaches—and improving nnUNet segmentation by +3.8% mDSC when used for data augmentation. Radiologist reviews indicate that ViCTR-generated liver cirrhosis MRIs are clinically indistinguishable from real scans. To our knowledge, ViCTR is the first method to provide fine-grained, pathology-aware MRI synthesis with graded severity control, closing a critical gap in AI-driven medical imaging research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.222, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The exponential growth in computer vision capabilities has been driven by significant advances in artificial intelligence models [7, 32]. However, medical imaging faces a fundamental tension between model complexity and data avail" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.379 + ], + "angle": 0, + "content": "ability that limits the application of state-of-the-art techniques. While recent breakthroughs in generative AI have demonstrated remarkable capabilities in synthetic data creation, these advances demand training datasets of unprecedented scale—a requirement that poses unique challenges in the medical domain [2, 33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.909, + 0.548 + ], + "angle": 0, + "content": "Unlike general computer vision applications, medical imaging faces several critical constraints: privacy regulations necessitating complex deidentification, inherent data fragmentation across healthcare institutions, and fundamental interoperability constraints. These barriers have created a growing disparity between the rapid advancement of general computer vision and the relatively slower progress in medical imaging applications. Current approaches to bridging this gap face two major limitations: insufficient feature preservation across anatomical structures and inadequate handling of pathological variations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.734 + ], + "angle": 0, + "content": "This challenge is especially pronounced in abdominal imaging, where pathologies such as diffuse cirrhosis or multi-tissue disease processes manifest across multiple organ systems. Unlike the well-defined boundaries typical of tumors or simpler structures like bones in X-ray imaging, abdominal pathologies often involve subtle and heterogeneous changes in tissue characteristics, requiring more nuanced feature extraction and synthesis. The inherent complexity of MRI signals—spanning multiple sequences and high spatial resolution—presents additional computational hurdles, further distancing it from the more standardized nature of CT or simpler 2D radiographs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address these gaps, our work introduces ViCTr (Vital Consistency Transfer), a two-stage framework that facilitates pathology-aware medical image synthesis with strong anatomical fidelity. We specifically target abdominal CT and MRI data, aiming to generate clinically relevant synthetic datasets that capture both normal anatomy and intricate pathological details. By providing robust augmentation material, our approach holds the potential to improve downstream tasks like segmentation and diagnosis, while alleviating issues of data scarcity and privacy constraints in medical imaging. Our contributions are:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22772" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.238, + 0.67, + 0.253 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed ViCTr methodology." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.28, + 0.482, + 0.325 + ], + "angle": 0, + "content": "- Novel Two-Stage Framework. We propose ViCTR, a method that fuses anatomical consistency and pathological realism for CT and MRI synthesis with high diversity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.325, + 0.483, + 0.385 + ], + "angle": 0, + "content": "- Tweedie's Formula in Rectified Flow. We introduce a rectified flow trajectory reformulation of Tweedie's formula, ensuring accurate initialization and reducing sampling bias." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.385, + 0.483, + 0.431 + ], + "angle": 0, + "content": "- Wide Applicability. ViCTR integrates seamlessly with various diffusion models, supporting a broad range of medical imaging tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.431, + 0.483, + 0.491 + ], + "angle": 0, + "content": "- One-Step Sampling & Tweedie's Corrections. We reduce computational overhead by cutting the number of diffusion steps from dozens to a handful, expediting inference without sacrificing quality." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.491, + 0.483, + 0.567 + ], + "angle": 0, + "content": "- Quantitative Improvements in Segmentation. We demonstrate that adding ViCTR-synthesized CT/MRI images to real-world training sets significantly boosts segmentation performance, underscoring the practical benefits of pathology-focused data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.567, + 0.483, + 0.626 + ], + "angle": 0, + "content": "- Enhanced Image Fidelity. ViCTR achieves consistently lower FID (Fréchet Inception Distance) scores across multiple datasets, indicating superior structural and textural coherence." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.627, + 0.483, + 0.702 + ], + "angle": 0, + "content": "- First Abdominal MRI Pathology Synthesis. To our knowledge, we are the first to generate abdominal MRI pathologies with progressive severity, offering new possibilities for research and clinical applications in medical imaging." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.28, + 0.483, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.766 + ], + "angle": 0, + "content": "By emphasizing both anatomical fidelity and pathological variety, ViCTr presents a powerful step toward bridging the gap between limited medical imaging data and the demands of high-performing AI models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.779, + 0.241, + 0.794 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.483, + 0.833 + ], + "angle": 0, + "content": "2.1. Generative Models for Medical Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Synthetic images have long been explored for data augmentation in medical imaging, with Generative Adversarial Networks (GANs) [25, 36, 50] initially offering promising results in tasks such as lesion synthesis and modality trans" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.906, + 0.341 + ], + "angle": 0, + "content": "lation. However, GAN-based methods often struggle with mode collapse, requiring extensive architectural and training refinements to ensure sufficient diversity and realism in the synthesized images [12, 35]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.353, + 0.836, + 0.37 + ], + "angle": 0, + "content": "2.2. Diffusion Models in Medical Imaging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Diffusion-based approaches have emerged as a powerful alternative to GANs for high-quality synthetic image generation [4, 5, 13, 49]. Their inherent noise-to-image paradigm provides a more stable training regime and can produce richer data variations. In medical imaging, diffusion models [11] have been leveraged to improve segmentation and classification performance by generating diverse training samples [1, 10, 37, 46, 47]. Recent specialized frameworks, including segmentation-guided diffusion [26] and text-driven generation built upon RadImageNet [34, 54], highlight the adaptability of diffusion models to specific clinical scenarios. Efforts such as MixUp-enhanced augmentation [6, 28] and domain adaptation methods [55] demonstrate that diffusion techniques can mitigate biases and improve generalization in tasks like classification and risk prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.905, + 0.646 + ], + "angle": 0, + "content": "2.3. General-Purpose Diffusion for Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.835 + ], + "angle": 0, + "content": "Various general-purpose diffusion strategies have been applied to further enrich training datasets. DiffuseMix [21] employs conditional prompts to blend real and synthetic data while preserving labels, whereas DreamDA [18] combines diffusion-based perturbations with pseudo-labeling for semantic consistency. DetDiffusion [51] incorporates object-detection attributes into the diffusion process, while Effective Data Augmentation with Diffusion Models [48] explores diffusion-based techniques to boost performance in few-shot settings. These methods generally aim to enhance data diversity and improve the training of downstream models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.722, + 0.862 + ], + "angle": 0, + "content": "2.4. Rectified Flow Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "While traditional diffusion approaches rely on discretized noise schedules, Rectified Flow (ReFlow) models learn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22773" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.485, + 0.354 + ], + "angle": 0, + "content": "Figure 2. Comparison of synthetic data distributions from various methods against the ground-truth (GT) on the AMOS dataset. The GT distribution (blue) reflects real data, while Reflow (red), Flow Straight and Fast (green), and Distribution Matching Distillation (purple) show varying alignment. Our proposed method, ViCTr (brown) achieves the closest match to GT, demonstrating superior distribution alignment and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.381, + 0.483, + 0.549 + ], + "angle": 0, + "content": "smooth, near-linear trajectories to map between data distributions. This can result in more efficient sampling and reduced computational overhead. Recent works have leveraged rectified flows for improved convergence rates and lower numbers of function evaluations (NFE) [29, 30, 53], although most focus on general distribution transport rather than the strict alignment and pathology-aware realism needed in medical imaging. As shown in Figure 2, different ReFlow-based methods exhibit varying degrees of distribution matching, with our approach (ViCTr) offering closer alignment to the ground-truth distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.345, + 0.577 + ], + "angle": 0, + "content": "2.5. Domain-Specific Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Domain-focused adaptations continue to evolve. DiNO-Diffusion [24] introduces a self-supervised latent diffusion model to cope with limited annotated medical data, and Diverse Data Augmentation with Diffusions [17] enriches domain generalization using Stable Diffusion combined with cosine similarity filtering. Collectively, these methods illustrate the growing interest in customized diffusion pipelines for specialized domains, especially when high-quality annotated data are scarce." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Recent advancements have further showcased the potential of diffusion models for domain-specific adaptation. DiNO-Diffusion [24] addresses the challenge of limited annotated data in medical imaging by leveraging a self-supervised latent diffusion framework. Similarly, Diverse Data Augmentation with Diffusions [17] enhances domain generalization by integrating Stable Diffusion with cosine similarity-based filtering, enabling the generation of semantically diverse and high-quality data. However, these methods prioritize general-domain efficiency over the anatomical-pathological alignment critical for medical imaging, often failing to preserve fine-grained structures" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.843, + 0.107 + ], + "angle": 0, + "content": "like vascular networks or diffuse fibrosis patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.23 + ], + "angle": 0, + "content": "In this work, we address the critical need for efficient, high-fidelity pathology-aware synthesis in abdominal imaging by merging rectified flow concepts with a Tweedie-corrected diffusion process. Our approach (ViCTr) stands out from prior methods by introducing a linearized sampling framework that enables one-step generation while maintaining anatomical fidelity—a crucial requirement for robust data augmentation in clinical applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.247, + 0.613, + 0.263 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Our framework, ViCTr, uses Rectified Flow [31] and Tweedie's Formula [43] for high-fidelity medical image synthesis with efficient one-step sampling. By rectifying the trajectory from a prior distribution \\( p_0 \\) to the target medical image distribution \\( p_{\\mathrm{target}} \\), we reduce sampling bias and retain key anatomical and pathological details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.377, + 0.747, + 0.394 + ], + "angle": 0, + "content": "3.1. Rectified Flow Trajectory" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.906, + 0.522 + ], + "angle": 0, + "content": "Classical diffusion approaches gradually corrupt data samples \\( x \\in \\mathbb{R}^d \\) into a Gaussian prior \\( p_0 \\). While these methods have shown success, they often suffer from suboptimal sampling paths, cascading prediction errors, and diminished diversity in generated samples. We address these issues through Rectified Flow optimization [16, 31], which learns a continuous velocity field guiding the forward and reverse diffusion processes more directly." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.906, + 0.599 + ], + "angle": 0, + "content": "Let \\((x_0, x_1)\\) be a sample pair with \\(x_0 \\sim p_0\\) and \\(x_1 \\sim p_{\\mathrm{target}}\\). We define an interpolated point \\(x_t = (1 - t)x_0 + tx_1\\) where \\(t \\in [0, 1]\\). A velocity model \\(v_\\theta : \\mathbb{R}^d \\times [0, 1] \\to \\mathbb{R}^d\\) then predicts how to transition from \\(x_t\\) toward \\(x_1\\). We train \\(v_\\theta\\) by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.625, + 0.905, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\hat {\\theta} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t \\sim \\operatorname {U n i f o r m} (0, 1)} \\left[ \\| \\left(x _ {1} - x _ {0}\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\| ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.717 + ], + "angle": 0, + "content": "ensuring the predicted flow \\( v_{\\theta}(x_t, t) \\) aligns with the true path \\( (x_1 - x_o) \\). Once trained, the rectified flow is realized by solving the ODE:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.732, + 0.905, + 0.75 + ], + "angle": 0, + "content": "\\[\nd x _ {t} = v _ {\\hat {\\theta}} \\left(x _ {t}, t\\right) d t, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.762, + 0.907, + 0.824 + ], + "angle": 0, + "content": "leading \\(x_0 \\sim p_0\\) toward \\(x_1 \\sim p_{\\mathrm{target}}\\). To enable one-step sampling, we distill this multi-step process into a neural network \\(\\hat{\\mathcal{N}}: \\mathbb{R}^d \\to \\mathbb{R}^d\\), such that \\(\\tilde{\\mathcal{T}}(x_0) = x_0 + v(x_0, 0)\\), is trained to directly predict \\(x_1\\) from \\(x_0\\). The loss function," + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.835, + 0.905, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} \\left[ \\left\\| \\left(x _ {1} - x _ {0}\\right) - v \\left(x _ {0}, 0\\right) \\right\\| ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.884, + 0.884, + 0.902 + ], + "angle": 0, + "content": "drives \\(\\mathcal{T}(\\hat{x}_0)\\) to approximate the final, rectified state \\(x_{1}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22774" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.427, + 0.106 + ], + "angle": 0, + "content": "3.2. Rectified Flow with Tweedie's Formula" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.19 + ], + "angle": 0, + "content": "Though rectified flow refines sampling, medical data distributions often need extra bias correction. Tweedie's Formula [43] addresses this by adjusting noisy observations to better approximate the posterior mean. For Gaussian variables \\(z \\sim \\mathcal{N}(\\mu_z, \\Sigma_z)\\), Tweedie's formula indicates:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.199, + 0.393, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\mu_ {z} \\mid z ] = z + \\Sigma_ {z} \\nabla_ {z} \\log p (z)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.225, + 0.483, + 0.27 + ], + "angle": 0, + "content": ", where \\(\\nabla_z\\log p(z)\\) is the gradient of the log probability. In typical diffusion, Tweedie's formula estimates \\(x_0\\) from noisy \\(x_{t}\\), guiding predictions toward the data manifold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.483, + 0.299 + ], + "angle": 0, + "content": "ViCTR incorporates this correction into the rectified flow ODE:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.311, + 0.441, + 0.328 + ], + "angle": 0, + "content": "\\[\nd x _ {t} = v _ {\\hat {\\theta}} (x _ {t}, t) d t + (1 - \\bar {\\alpha} _ {t}) \\nabla_ {x _ {t}} \\log p (x _ {t}) d t,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.337, + 0.338, + 0.352 + ], + "angle": 0, + "content": "and updates the training objective to:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.36, + 0.482, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\theta} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t \\sim \\operatorname {U n i f o r m} (0, 1)} \\left[ \\| \\left(x _ {1} - x _ {0}\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\right. \\\\ \\left. \\left. - \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {x _ {t}} \\log p (x _ {t}) \\right| ^ {2} \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Here, \\(\\bar{\\alpha}_t\\) denotes the cumulative product of variance decay factors in the diffusion schedule. The additional Tweedie term \\((1 - \\bar{\\alpha}_t)\\nabla_{x_t}\\log p(x_t)\\) corrects for sampling bias, driving \\(x_{t}\\) more accurately toward the target distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.527 + ], + "angle": 0, + "content": "We similarly extend the one-step distillation to integrate Tweedie's correction:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.537, + 0.454, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {T}} (x _ {0}) = x _ {0} + v (x _ {0}, 0) + (1 - \\bar {\\alpha} _ {0}) \\nabla_ {x _ {0}} \\log p (x _ {0}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.564, + 0.182, + 0.579 + ], + "angle": 0, + "content": "and optimize" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.587, + 0.376, + 0.622 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} \\left[ \\left\\| x _ {1} - \\hat {\\mathcal {T}} (x _ {0}) \\right\\| ^ {2} \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.689 + ], + "angle": 0, + "content": "This single-step approach balances computational efficiency with high-quality, pathology-aware generation—critical in medical applications where datasets are limited and synthetic realism is paramount." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.78 + ], + "angle": 0, + "content": "In summary, ViCTr unifies rectified flow and Tweedie's correction to deliver anatomically consistent, pathology-aware sampling in a single forward pass. The resulting framework exhibits reduced inference costs, minimized sampling bias, and improved fidelity for generating high-resolution medical images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.387, + 0.805 + ], + "angle": 0, + "content": "3.3. Pathology Aware Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our proposed ViCTr method adopts a two-stage training paradigm: Stage 1 establishes a foundational diffusion model tailored to medical imaging data, and Stage 2 fine-tunes this pre-trained model for downstream tasks, including semantic-guided multi-modal generation and counterfactual pathology synthesis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.828, + 0.107 + ], + "angle": 0, + "content": "Pre-training on ATLAS-8k Dataset (Stage 1)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.11, + 0.907, + 0.231 + ], + "angle": 0, + "content": "In the absence of large-scale, domain-specific pre-trained diffusion models, we begin by training on the ATLAS8k [40] dataset, which comprises abdominal CT scans and their segmentation annotations. We leverage these annotations as conditional guidance—enabling precise control over anatomical structures—and integrate textual prompts (e.g., \"create image having \\(\\langle \\langle \\text{organs} \\rangle \\rangle\\)\" to further refine semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Latent Representations. Following Figure 1, we feed raw CT images \\((X_I)\\) and segmentation masks \\((X_S)\\) into a frozen VAE encoder, producing latent embeddings \\((Z_o\\) and \\(Z_s)\\). Simultaneously, textual prompts \\((X_p)\\) are processed by a pre-trained text encoder (details in Table 1) to yield prompt embeddings \\((Z_p)\\). These latent representations guide the generative diffusion backbone \\(\\phi_{\\theta}\\) during both forward and reverse diffusion." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.353, + 0.906, + 0.383 + ], + "angle": 0, + "content": "Forward Diffusion. We progressively inject noise into \\( Z_{o} \\) forming a noisy representation \\( Z_{t} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.393, + 0.828, + 0.41 + ], + "angle": 0, + "content": "\\[\nP (Z _ {t} | Z _ {o}) = (1 - t) \\cdot Z _ {o} + t \\cdot \\epsilon_ {\\mathrm {t r u e}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.906, + 0.495 + ], + "angle": 0, + "content": "where \\(\\epsilon_{true}\\) is the true noise at step \\(t\\). We concatenate \\(Z_{t}\\) with \\(Z_{s}\\) and feed them into the diffusion model \\(\\phi_{\\theta}\\). Crucial layers are selectively unfrozen based on elastic weight consolidation, maintaining model stability while adapting to medical domain specifics." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.496, + 0.816, + 0.511 + ], + "angle": 0, + "content": "Reverse Diffusion. In the reverse process," + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.522, + 0.907, + 0.539 + ], + "angle": 0, + "content": "\\[\nP \\left(Z _ {t - 1} \\mid Z _ {t}, Z _ {s}, Z _ {p}, t\\right) = Z _ {t} + \\delta T \\times \\phi_ {\\theta} \\left(Z _ {t}, Z _ {s}, Z _ {p}, t\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.595 + ], + "angle": 0, + "content": "iteratively removes noise to reconstruct \\( Z_{t-1} \\) conditioned on the segmentation map \\( (Z_s) \\) and text prompt \\( (Z_s) \\). The diffusion loss \\( L_{diff} \\) between \\( Z_{t-1} \\) and \\( \\epsilon_{true} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.605, + 0.906, + 0.624 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i f f}} = - \\left| \\phi_ {\\theta} \\left(Z _ {t}, Z _ {s}, Z _ {p}, t\\right) - \\left(\\epsilon_ {\\text {t r u e}} - Z _ {o}\\right) \\right| ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.679 + ], + "angle": 0, + "content": "After refining \\( Z_{t-1} \\) via Tweedie's formula, the denoised latent \\( Z_o \\) is fed into the frozen VAE decoder (matching \\( \\phi_\\theta \\); see Table 1) to obtain a reconstructed CT image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.907, + 0.815 + ], + "angle": 0, + "content": "Composite Loss. To ensure both local accuracy and global perceptual realism, we combine: Diffusion Loss \\((L_{\\mathrm{diff}})\\), pixel-level reconstruction loss \\((L_2)\\) evaluating the gap between generated \\(X_{ig}\\) and ground-truth \\(X_I\\), and structural similarity (SSIM) loss \\((L_{SSIM})\\) emphasizing higher-order textural correspondence. Optimizing this composite loss enables the diffusion backbone \\(\\phi_{\\theta}\\) to capture essential anatomical features, establishing a versatile medical foundation model ready for fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.816, + 0.837 + ], + "angle": 0, + "content": "Fine-tuning on downstream tasks (Stage 2)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Building on the Stage1 pre-trained model, Stage2 targets specialized tasks like semantic-guided CT/MRI generation and counterfactual pathology synthesis. As shown in Figure 1, a dual-network setup is adopted." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22775" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.48, + 0.121 + ], + "angle": 0, + "content": "1. \\(\\phi_{\\text{base}}\\) remains frozen, retaining the robust anatomical knowledge acquired in Stage 1." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.122, + 0.482, + 0.18 + ], + "angle": 0, + "content": "2. \\(\\phi_{\\text {adapt }}\\) is augmented with LoRA [19] modules within the previously fine-tuned layers. These modules are the only trainable parameters, ensuring targeted and stable adaptation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.187, + 0.483, + 0.323 + ], + "angle": 0, + "content": "Training Input Consistency Loss. Similar to Stage1, each training sample includes prompts, segmentation maps, and either CT or MRI images (depending on the task). A frozen VAE encoder extracts latent representations \\( Z_{s}, Z_{o} \\), and \\( Z_{p} \\). Following forward diffusion of \\( Z_{o} \\) into noisy \\( Z_{t} \\), both \\( \\phi_{\\text{base}} \\) and \\( \\phi_{\\text{adapt}} \\) process \\( [Z_{t}, Z_{s}] \\) with text embedding \\( Z_{p} \\). We introduce a consistency loss \\( L_{\\text{consistency}} \\) to align intermediate outputs of \\( \\phi_{\\text{base}} \\) and \\( \\phi_{\\text{adapt}} \\), ensuring stable adaptation without drifting from core representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.45 + ], + "angle": 0, + "content": "Loss Components. Diffusion loss \\((L_{diff}^{\\mathrm{base}}\\), \\(L_{diff}^{\\mathrm{adapt}})\\) encourages accurate noise prediction. Temporal consistency loss \\(L_{temporal}\\) preserves smooth transitions across time steps in the reverse diffusion. Spatial consistency loss \\(L_{spatial}\\) ensures alignment in the output reconstructions \\((Z_{t-1}^{base}, (Z_{t-1}^{adapt})\\). Pixel-level losses \\((L_2\\) and \\(L_{SSIM})\\) compare generated results to ground-truth scans, enforcing both local fidelity and structural coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.515 + ], + "angle": 0, + "content": "CT and MRI Generation. We adapt to BTCV [15] (CT) and AMOS [23] (MRI) datasets, using organ segmentation masks for prompts. A standard 80/10/10 partition ensures robust evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Pathology Generation. For liver cirrhosis, we employ CurrMRI600+ [22] dataset (T1/T2 MRIs) with segmentation masks marking liver regions. Prompts specify \"low,\" \"mild,\" or \"severe\" cirrhosis intensity. As in the CT/MRI setting, the training/validation/testing split is 80/10/10. The inference details and inference diagram are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.751 + ], + "angle": 0, + "content": "This selective LoRA-based fine-tuning allows ViCTR to handle complex tasks—ranging from normal organ generation to pathology synthesis—while preserving the anatomical and text-driven guidance learned in Stage 1. The experiments (Section 4) demonstrate that this two-stage framework consistently yields high-quality, clinically realistic images in both CT and MRI settings, including nuanced liver cirrhosis modeling." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.784, + 0.514, + 0.859 + ], + "angle": 0, + "content": "
Diffusion MethodsDenoiserText Encoder
Stable Diffusion [44]UNet [45]Clip-B/16 [41]
Pixart-alpha [9]Transformer (DiT) [38]T5-XXXXL [42]
Stable Diffusion XL [39]Dual Unet [39]Clip-L/14
Flux [27]MultiModal Transformer [14]T5-XXXXL + Clip-L/15
Stable Diffusion-3 [14]MultiModal TransformerT5-XXXXL + Clip-B/16 + Clip-L/14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.869, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Table 1. Selection of denoisers and text encoders for different diffusion methods \\( {\\phi }_{\\theta } \\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.747, + 0.107 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.146 + ], + "angle": 0, + "content": "The implementation details for training and sampling are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.154, + 0.707, + 0.17 + ], + "angle": 0, + "content": "4.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.176, + 0.905, + 0.373 + ], + "angle": 0, + "content": "Synthetic Data Generation Results. Table 2 summarizes our comparative evaluation of synthetic medical image generation under vanilla fine-tuning versus the proposed ViCTR pipeline, spanning multiple datasets (BTCV, AMOS, Cir-rMRI600+) and diffusion backbones (Stable Diffusion, SD-XL, SD-3, Pixart-alpha, Flux). While vanilla fine-tuning applies a direct end-to-end approach after ATLAS-8k pretraining, it generally yields higher Fréchet Inception Distance (FID) and Medical FID (MFID) values across all datasets. Conversely, ViCTR integrates domain-specific loss functions and specialized architectural settings, consistently achieving lower FID/MFID scores and producing more realistic medical images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.905, + 0.552 + ], + "angle": 0, + "content": "To address the well-known limitation of FID (which uses ImageNet features), we adopt M3D-CLIP [3] to compute a more domain-specific MFID. As shown in the table, ViCTR provides notable gains: for example, Stable Diffusion under ViCTR achieves FID/MFID of 21.98/19.02 (BTCV), 20.37/19.11 (AMOS), and 25.57/21.46 (Cir-rMRI600+). These results represent substantial improvements over vanilla fine-tuning and demonstrate consistent performance boosts across all tested architectures. Such gains highlight the robustness of ViCTR's two-stage framework in generating clinically meaningful synthetic images for both CT and MRI modalities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.905, + 0.855 + ], + "angle": 0, + "content": "Segmentation Results. To assess the quality and utility of ViCTR-generated synthetic data, we performed segmentation experiments on three datasets—BTCV, AMOS, and CurrMRI600+—tracking mean Dice Similarity Coefficient (mDSC) and mean Hausdorff Distance 95 (mHD95). Higher mDSC and lower mHD95 respectively indicate improved overlap accuracy and spatial precision. We examined four training configurations: (1) baseline using original data only, (2) standard augmentation (Random Crop, Rotate, Blur, Affine, Geometric Distortion at 0.4 probability), (3) \\(30\\%\\) synthetic data via vanilla fine-tuning, and (4) \\(30\\%\\) synthetic data from ViCTR. Table 3 highlights consistent performance gains from ViCTR across all datasets and backbone architectures. Notably, CurrMRI600+—a liver cirrhosis dataset—exhibited the most significant improvements, reflecting ViCTR's capacity to capture complex pathological features. These results confirm that ViCTR-generated images significantly boost segmentation metrics (mDSC and mHD95), highlighting the practical value of our two-stage framework for medical data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Efficiency of ViCTR. Table 2 summarizes the number of diffusion steps and average inference times (in seconds) for vanilla fine-tuning versus the ViCTR framework. Across all" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22776" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.088, + 0.979, + 0.165 + ], + "angle": 0, + "content": "
BaselinesBTCV Dataset (CT Generation)AMOS Dataset (MRI Generation)CirrMRI600+ (Pathology Generation)Diffusion Steps | Inference Time
Fine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTR
Stable Diffusion [44]25.44 / 19.6721.98 / 19.0225.43 / 21.7620.37 / 19.1128.34 / 23.4325.57 / 21.4640 | 13.764 | 3.12
Stable Diffusion XL [39]23.47 / 18.2120.33 / 17.4424.11 / 20.2319.44 / 18.4527.34 / 22.1124.02 / 20.7630 | 14.554 | 3.45
Stable Diffusion-3 [14]19.07 / 16.2217.37 / 16.0222.32 / 19.7618.02 / 19.0824.49 / 21.7821.28 / 19.3450 | 18.983 | 2.78
Pixart-alpha [9]21.32 / 17.0919.22 / 16.9623.78 / 20.0418.76 / 18.5626.06 / 20.0723.04 / 18.9225 | 10.673 | 1.74
Flux [27]15.52 / 15.0113.28 / 14.0819.02 / 18.2815.55 / 16.5822.46 / 18.8819.96 / 17.0130 | 15.663 | 2.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.174, + 0.908, + 0.217 + ], + "angle": 0, + "content": "Table 2. Quantitative results on CT, MRI, and Cirrhosis generation. All baselines were pre-trained on ATLAS-8k and fine-tuned on target datasets. Metrics reported are Fréchet Inception Distance (FID) and Medical FID (MFID), shown as FID/MFID, with inference time in seconds. Lower values indicate better performance." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.23, + 0.943, + 0.501 + ], + "angle": 0, + "content": "
BaselinesOriginal Dataset (Org.)AugmentationOrg. + Synth. by FLUX 30% (vanilla Fine-tuning)Org. + Synth. by FLUX 30% (ViCtR)
mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)
BTCV dataset Segmentation results
UNet [45]76.7234.4278.4533.1779.3232.3781.2230.17
TransUnet [8]85.5232.3387.0131.4387.5430.1189.7829.12
nnUnet [20]80.4830.1982.5431.0283.3729.0185.1927.77
nnFormer [56]83.4730.0183.9830.8885.8828.7887.7226.22
MedSegDiff [52]87.9127.6788.6526.7289.7825.5291.9223.31
AMOS dataset Segmentation results
UNet [45]68.9234.5770.5532.3271.0231.7873.3429.11
TransUnet [8]71.3333.1272.5631.0973.4329.1977.5427.56
nnUnet [20]73.3332.3274.1130.1975.6828.7578.2926.32
nnFormer [56]75.7831.1976.4429.7677.7727.5781.3224.41
MedSegDiff [52]76.8329.9278.3826.9979.0325.4584.0222.18
CirrMRI600+ dataset Segmentation results
UNet [45]68.7436.7369.3835.4370.1235.1173.3932.11
TransUnet [8]70.7735.4270.9834.0171.7833.9274.5631.09
nnUnet [20]71.0234.3572.4932.7873.5631.2778.8930.25
nnFormer [56]74.8833.7875.2331.8876.4531.0979.4429.78
MedSegDiff [52]76.9230.7977.1130.3478.0329.8981.3727.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.51, + 0.908, + 0.54 + ], + "angle": 0, + "content": "Table 3. Segmentation results on BTCV, AMOS, and CurrMRI600+ using various training settings and baselines. Metrics reported are mDSC (%) and mHD95 (mm). Top three results per setting are highlighted: blue (1st), green (2nd), and orange (3rd)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.565, + 0.483, + 0.745 + ], + "angle": 0, + "content": "evaluated models, ViCTR consistently reduces the required steps, thereby shortening inference duration without compromising image fidelity. This efficiency stems from two key innovations: (1) Tweedy's formula to approximate \\( Z_{0} \\) and streamlines the reconstruction process, and (2) Two-Stage Training, enabling the model to learn both anatomical consistency and pathology-specific variations in a targeted manner. By minimizing extraneous steps, ViCTR can generate high-quality synthetic images more rapidly, making it a viable solution for resource-constrained clinical or research environments where computational overhead is a critical consideration." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.757, + 0.274, + 0.772 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We conducted extensive qualitative evaluations of ViCTr across multiple modalities (CT and MRI) and tasks, covering both anatomical and pathological image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Anatomical and Pathology-Driven Image Generation. Unlike prior methods that primarily generate normal anatomical scans or isolated tumors, ViCTr extends synthesis capabilities to complex pathologies, such as liver cirrhosis. Figure 3 shows pairs of non-cirrhotic MRIs alongside" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.701 + ], + "angle": 0, + "content": "their segmentation masks and the corresponding ViCTR-generated cirrhotic images. Notably, the synthesized images retain essential anatomical structures while introducing realistic cirrhotic texturing, an advancement that enables more diverse data augmentation and supports research into disease progression. These high-fidelity synthetic samples offer potential to enrich clinical training datasets, advance diagnostic algorithms, and support detailed analyses of pathology progression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Qualitative Segmentation Outcomes. To further showcase ViCTR's impact on downstream tasks, we visually assessed segmentation quality on BTCV and AMOS (Figure 6). Focusing on regions like the left kidney and pancreas, models trained with ViCTR-augmented data show sharper boundaries and more precise delineation than standard fine-tuning, indicating improved spatial learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Comparative Performance on BTCV and AMOS. Figures 4 (BTCV) and 5 (AMOS) compare diffusion-based image generation results using Stable Diffusion v3 (SD-3) and Flux backbones. On BTCV—a dataset with limited samples and larger spatial dimensions—vanilla finetuning with SD-3 often yields poor mask adherence, sug" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.48, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.284, + 0.483, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Examples of synthetic cirrhotic images with different severity levels generated from non-cirrhotic scans using ViCTr. For T1-MRI, we show the segmentation mask, the original non-cirrhotic image, and the corresponding synthetic cirrhotic image." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.342, + 0.481, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Figure 4. Diffusion-Based Image Generation on BTCV MRI Dataset. From left to right: segmentation mask, SD-3 with standard fine-tuning, FLUX with standard fine-tuning, SD-3 with ViCTR, FLUX with ViCTR." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.608, + 0.481, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.895 + ], + "angle": 0, + "content": "Figure 5. Diffusion-Based Image Generation on AMOS Dataset. From left to right: segmentation mask, SD-3 with Vanilla fine-tuning, SD-3 with our method, FLUX with Vanilla fine-tuning, FLUX with our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "gesting difficulties in learning accurate spatial constraints. In contrast, the same model shows markedly improved mask alignment on the AMOS dataset, attributed to its larger training set and more manageable image dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Regardless of these dataset-specific variations, ViCTr achieves consistently high-quality outputs with robust mask adherence and anatomical accuracy across both BTCV and AMOS, underscoring its enhanced spatial learning and resilience. By maintaining performance even under data-sparse conditions, ViCTr demonstrates a clear advantage over conventional fine-tuning approaches, which can exhibit greater sensitivity to dataset size and complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Second, we evaluate the impact of ViCTR on segmentation tasks, illustrated in Figure 6. Both BTCV and AMOS datasets encompass various abdominal structures (e.g., left kidney, pancreas), and our results highlight how ViCTR-based training yields notably more precise organ boundaries compared to alternative methods. This improvement is particularly evident in challenging regions where standard segmentation often struggle, highlighting ViCTR's strength in enhancing spatial delineation and anatomical accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.416, + 0.678, + 0.43 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.446, + 0.903, + 0.574 + ], + "angle": 0, + "content": "
Ablation Based on Pre-training of Model
ViCTR (Without Stage-1)17.33
Ablation Based on Rectified Flow Algorithms
ViCTR (Without Proposed Tweedies Formula)18.78
ViCTR (With Reflow)20.19
ViCTR (With Flow Straight and Fast)21.37
ViCTR (With Distribution Matching Distillation)22.33
Ablation Study Based on Loss FunctionsAblation Based on LoRA Rank
Ldiff18.77r = 818.46
Ldiff + Lspatial18.21r = 1617.52
Ldiff + Lconsistency17.02r = 3216.66
Ldiff + Lspatial + Lconsistency15.55r = 6415.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.584, + 0.905, + 0.613 + ], + "angle": 0, + "content": "Table 4. Ablation study for the effect of model pretraining, loss functions, and LoRA rank settings are shown." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.748 + ], + "angle": 0, + "content": "To rigorously validate our ViCTr framework, we conducted a series of ablation experiments on the BTCV dataset (Table 4), focusing on four key aspects: pretraining, rectified flow algorithms, loss function components, and LoRA rank selection. We used FID (Frechet Inception Distance) as our primary metric to gauge how each architectural choice affects both anatomical fidelity and pathological realism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Impact of Pretraining. We first examined the necessity of Stage 1 pretraining by comparing our full, two-stage pipeline to a variant that proceeds directly to downstream fine-tuning. Omitting Stage 1 degrades the model's FID score from 15.55 to 17.33 (Table 4), underscoring the vital role of an anatomical prior. These findings affirm that establishing robust structural representations in Stage 1 is crucial for achieving high-quality medical image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Rectified Flow Algorithms. We next assessed our rectified flow formulation, augmented with Tweedie's formula," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.356, + 0.907, + 0.399 + ], + "angle": 0, + "content": "Figure 6. Segmentation results on BTCV and AMOS datasets' left kidney and pancreas. We compare the ground truth (blue), models trained on the default dataset, augmented dataset, default dataset plus \\(30\\%\\) synthetic data generated by FLUX with standard fine-tuning, and default dataset plus \\(30\\%\\) synthetic data generated by FLUX with ViCTr (red)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.42, + 0.482, + 0.542 + ], + "angle": 0, + "content": "against alternative rectified flow-based models, including ReFlow, Flow Straight and Fast, and Distribution Matching Distillation. Removing Tweedie's correction alone increased the FID to 18.78, highlighting its significance for trajectory alignment. Across all baselines, we observed FID deteriorations, ranging from 13.28 to as high as 22.33, thereby demonstrating that our proposed method provides tighter distribution matching and superior synthesis fidelity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.545, + 0.483, + 0.742 + ], + "angle": 0, + "content": "Ablation Study Based on Loss Functions. We then investigated each ViCTr loss component—diffusion loss, spatial consistency, and consistency loss—to determine their relative contributions: the baseline model, utilizing only diffusion loss, produced an FID of 18.77 (Table 4). Incorporating spatial consistency loss yielded a moderate improvement to 18.21, highlighting its role in maintaining structural integrity. The addition of consistency loss further enhanced performance, reducing the FID to 17.02 by promoting image coherence across generations. When combining all three terms, we achieved an FID of 15.55, indicating that these components collectively offer significant improvements in anatomical accuracy and visual realism." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.746, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Ablation Based on LoRA Rank. Finally, we explored how varying the LoRA rank \\((r)\\) influences generative performance. Starting at \\(r = 8\\) and incrementing upwards, we found \\(r = 64\\) to yield the best FID at 15.55. This result suggests that higher-dimensional adaptation spaces enable more nuanced parameter updates during fine-tuning, thereby improving visual quality and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Visual Turing Tests. To further validate clinical plausibility, three radiologists participated in Visual Turing Tests using 15 randomly generated MRI scans depicting vary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.557 + ], + "angle": 0, + "content": "ing levels of liver cirrhosis (mild, moderate, severe). All scans were uniformly judged to be clinically realistic, with identical outcomes observed when using additional random samples. Notably, the synthetic cirrhotic images correctly exhibited surface nodularity and textural irregularities, consistent with radiologic findings in mild-to-severe cirrhosis (Figure 3). These results confirm ViCTr's ability to synthesize pathology-specific features, enabling applications in training, algorithm development, and clinical research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.634, + 0.59 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.86 + ], + "angle": 0, + "content": "We presented ViCTR, a novel two-stage framework designed to generate high-fidelity medical images by integrating robust anatomical pre-training with precise pathology-specific fine-tuning. By aligning Tweedie's formula with linear projection methods used in flow matching, ViCTR maintains accurate initial distribution estimates even amid diffusion processes. This setup is further enhanced by LoRA adapters, which preserve essential anatomical information while flexibly adapting to diverse pathologies. Experimental results demonstrate that ViCTR consistently surpasses conventional fine-tuning strategies, reducing FID scores and producing clinically realistic outputs. Beyond data augmentation for segmentation and classification, ViCTR can be extended to tasks such as modality translation, contrast synthesis, and professional training, offering a powerful and versatile tool for advancing AI-driven medical imaging." + }, + { + "type": "page_footnote", + "bbox": [ + 0.536, + 0.875, + 0.877, + 0.888 + ], + "angle": 0, + "content": "\\(^{1}\\)Code: https://github.com/Onkarsus13/ViCTR-2D" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.888, + 0.898, + 0.901 + ], + "angle": 0, + "content": "Weights: https://huggingface.co/onkarsus13/ViCTR-2D" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.875, + 0.898, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22779" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.146 + ], + "angle": 0, + "content": "This research is supported by the following NIH grants: R01-HL171376 and U01-CA268808." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.256 + ], + "angle": 0, + "content": "[1] Hazrat Ali, Shafaq Murad, and Zubair Shah. Spot the fake lungs: Generating synthetic medical images using neural diffusion models. In Artificial Intelligence and Cognitive Science, pages 32-39, Cham, 2023. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.259, + 0.484, + 0.329 + ], + "angle": 0, + "content": "[2] Alhanoof Althnian, Duaa Alsaeed, Heyam Al-Baity, Amani Samha, Alanoud Bin Dris, Najla Alzakari, Afnan Abou Elwafa, and Heba Kurdi. Impact of dataset size on classification performance: an empirical evaluation in the medical domain. Applied Sciences, 11(2):796, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.331, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[3] Fan Bai, Yuxin Du, Tiejun Huang, Max Q. H. Meng, and Bo Zhao. M3d: Advancing 3d medical image analysis with multi-modal large language models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.374, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.432, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[6] Luigi Carratino, Moustapha Cisse, Rodolphe Jenatton, and Jean-Philippe Vert. On mixup regularization. Journal of Machine Learning Research, 23(325):1-31, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[7] Junyi Chai, Hao Zeng, Anming Li, and Eric W.T. Ngai. Deep learning in computer vision: A critical review of emerging techniques and application scenarios. Machine Learning with Applications, 6:100134, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[9] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[10] Qi Chen, Xiaoxi Chen, Haorui Song, Zhiwei Xiong, Alan Yuille, Chen Wei, and Zongwei Zhou. Towards generalizable tumor synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11147-11158, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Gayatri Deshmukh, Onkar Kishor Susladkar, Debesh Jha, Elif Keles, Halil Ertugrul Aktas, Daniela P Ladner, Amir A Borhani, Gorkem Durak, Ulas Bagci, et al. Meddelinea: Scalable and efficient medical image segmentation via controllable diffusion transformers. In Medical Imaging with Deep Learning, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.187, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[13] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E. Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[15] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619-3629, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.904, + 0.371 + ], + "angle": 0, + "content": "[16] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Flux that plays music, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[17] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[18] Yunxiang Fu, Chaoqi Chen, Yu Qiao, and Yizhou Yu. Dreamda: Generative data augmentation with diffusion models. arXiv preprint arXiv:2403.12803, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[20] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[21] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27621-27630, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[22] Debesh Jha, Onkar Kishor Susladkar, Vandan Gorade, Elif Keles, Matthew Antalek, Deniz Seyithanoglu, Timurhan Cebeci, Halil Ertugrul Aktas, Gulbiz Dagoglu Kartal, Sabahattin Kaymakoglu, et al. Cirrmri600+: Large scale mri collection and segmentation of cirrhotic liver. arXiv preprint arXiv:2410.16296, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[23] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems, 35:36722-36732, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Guillermo Jimenez-Perez, Pedro Osorio, Josef Cersovsky, Javier Montalt-Tordera, Jens Hooge, Steffen Vogler, and Sadegh Mohammadi. Dino-diffusion. scaling medical diffusion via self-supervised pre-training. arXiv preprint arXiv:2407.11594, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22780" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Sunho Kim, Byungjai Kim, and HyunWook Park. Synthesis of brain tumor multicontrast mr images for improved data augmentation. Medical Physics, 48(5):2185-2198, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[26] Nicholas Konz, Yuwen Chen, Haoyu Dong, and Maciej A Mazurowski. Anatomically-controllable medical image generation with segmentation-guided diffusion models. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 88–98. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[27] Black Forest Labs. black-forest-labs/FLUX.1-dev \\(\\cdot\\) Hugging Face. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. [Online; accessed 05-October-2024]. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[28] Hansang Lee, Haeil Lee, and Helen Hong. Genmix: Combining generative and mixture data augmentation for medical image classification. arXiv preprint arXiv:2405.20650, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[29] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. Advances in Neural Information Processing Systems, 37:63082-63109, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[32] Supriya V. Mahadevkar, Bharti Khemani, Shruti Patil, Ketan Kotecha, Deepali R. Vora, Ajith Abraham, and Lubna Abdelkareim Gabralla. A review on machine learning styles in computer vision—techniques and future directions. IEEE Access, 10:107293–107329, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[33] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M. Alvarez, Zhiding Yu, Sanja Fidler, and Marc T. Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 275-284, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[34] Xueyan Mei, Zelong Liu, Philip M Robson, Brett Marinelli, Mingqian Huang, Amish Doshi, Adam Jacobi, Chendi Cao, Katherine E Link, Thomas Yang, Ying Wang, Hayit Greenspan, Timothy Deyer, Zahi A Fayad, and Yang Yang. RadImageNet: An open radiologic deep learning research dataset for effective transfer learning. Radiol Artif Intell, 4 (5):e210315, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[35] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, Jakob Nikolas Kather, and Daniel Truhn. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13(1):12098, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part III 20, pages 417-425. Springer, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[37] Kaliprasad Pani and Indu Chawla. Synthetic mri in action: A novel framework in data augmentation strategies for robust multi-modal brain tumor segmentation. Computers in Biology and Medicine, 183:109273, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[40] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[43] Herbert E. Robbins. An Empirical Bayes Approach to Statistics, pages 388-394. Springer New York, New York, NY, 1992. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.594 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[46] Daniel G. Saragih, Atsuhiro Hibi, and Pascal N. Tyrrell. Using diffusion models to generate synthetic labeled data for medical image segmentation. International Journal of Computer Assisted Radiology and Surgery, 19(8):1615-1625, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[47] Vanshali Sharma, Abhishek Kumar, Debesh Jha, M.K. Bhuyan, Pradip K. Das, and Ulas Bagci. Controlpolynet: Towards controlled colon polyp synthesis for improved polyp segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2325-2334, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[48] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Brandon Trabucco, Kyle Doherty, Max A Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffu" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22781" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "sion models. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[50] Toan Tran, Trung Pham, Gustavo Carneiro, Lyle Palmer, and Ian Reid. A bayesian data augmentation approach for learning deep models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[51] Yibo Wang, Ruiyuan Gao, Kai Chen, Kaiqiang Zhou, Yingjie Cai, Lanqing Hong, Zhenguo Li, Lihui Jiang, DitYan Yeung, Qiang Xu, et al. Detdiffusion: Synergizing generative and perceptive models for enhanced data generation and perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7246-7255, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[52] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In Medical Imaging with Deep Learning, pages 1623-1639. PMLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[53] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[54] Zheyuan Zhang, Lanhong Yao, Bin Wang, Debesh Jha, Elif Keles, Alpay Medetalibeyoglu, and Ulas Bagci. Emit-diff: Enhancing medical image segmentation via text-guided diffusion model. arXiv preprint arXiv:2310.12868, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.541 + ], + "angle": 0, + "content": "[55] Yuan Zhong, Suhan Cui, Jiaqi Wang, Xiaochen Wang, Ziyi Yin, Yaqing Wang, Houping Xiao, Mengdi Huai, Ting Wang, and Fenglong Ma. MedDiffusion: Boosting Health Risk Prediction via Diffusion-based Data Augmentation, pages 499-507. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[56] Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Lequan Yu, Liansheng Wang, and Yizhou Yu. nnformer: Interleaved transformer for volumetric segmentation. arXiv preprint arXiv:2109.03201, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "22782" + } + ] +] \ No newline at end of file diff --git a/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_origin.pdf b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..905e43b8ad7f7cc2b0e7ca9bc35dac494b69259a --- /dev/null +++ b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/08391b0d-1c48-48a7-aee4-a1aed1fb2b8c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a555e0d5b398642542baab324f49741fb8d4dbed9ca4ec591e1d93e94e2c8752 +size 5192131 diff --git a/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/full.md b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9ff94cac50fd5d2bf642224bddc21c4edaad1d55 --- /dev/null +++ b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/full.md @@ -0,0 +1,342 @@ +# ViCTr: Vital Consistency Transfer for Pathology Aware Image Synthesis + +Onkar Susladkar $^{1,3}$ , Gayatri Deshmukh $^{1}$ , Yalcin Tur $^{2}$ , Gorkem Durak $^{1}$ , Ulas Bagci $^{1}$ $^{1}$ Northwestern University $^{2}$ Stanford University $^{3}$ University of Illinois Ubrana-Champaign + +onkarsus13@gmail.com, dgayatri9850@gmail.com, yalcintr@stanford.edu, + +gorkem.durak@northwestern.edu, ulas.bagci@northwestern.edu + +# Abstract + +Synthesizing medical images remains challenging due to limited annotated pathological data, modality domain gaps, and the complexity of representing diffuse pathologies such as liver cirrhosis. Existing methods often struggle to maintain anatomical fidelity while accurately modeling pathological features, frequently relying on priors derived from natural images or inefficient multi-step sampling. In this work, we introduce ViCTR (Vital Consistency Transfer), a novel two-stage framework that combines a rectified flow trajectory with a Tweedie-corrected diffusion process to achieve high-fidelity, pathology-aware image synthesis. First, we pretrain ViCTR on the ATLAS-8k dataset using Elastic Weight Consolidation (EWC) to preserve critical anatomical structures. We then fine-tune the model adversarially with Low-Rank Adaptation (LoRA) modules for precise control over pathology severity. By reformulating Tweedie's formula within a linear trajectory framework, ViCTR supports one-step sampling—reducing inference from 50 steps to just 4—without sacrificing anatomical realism. We evaluate ViCTR on BTCV (CT), AMOS (MRI), and CirrMRI600+ (cirrhosis) datasets. Results demonstrate state-of-the-art performance, achieving a Medical Fréchet Inception Distance (MFID) of 17.01 for cirrhosis synthesis—28% lower than existing approaches—and improving nnUNet segmentation by +3.8% mDSC when used for data augmentation. Radiologist reviews indicate that ViCTR-generated liver cirrhosis MRIs are clinically indistinguishable from real scans. To our knowledge, ViCTR is the first method to provide fine-grained, pathology-aware MRI synthesis with graded severity control, closing a critical gap in AI-driven medical imaging research. + +# 1. Introduction + +The exponential growth in computer vision capabilities has been driven by significant advances in artificial intelligence models [7, 32]. However, medical imaging faces a fundamental tension between model complexity and data avail + +ability that limits the application of state-of-the-art techniques. While recent breakthroughs in generative AI have demonstrated remarkable capabilities in synthetic data creation, these advances demand training datasets of unprecedented scale—a requirement that poses unique challenges in the medical domain [2, 33]. + +Unlike general computer vision applications, medical imaging faces several critical constraints: privacy regulations necessitating complex deidentification, inherent data fragmentation across healthcare institutions, and fundamental interoperability constraints. These barriers have created a growing disparity between the rapid advancement of general computer vision and the relatively slower progress in medical imaging applications. Current approaches to bridging this gap face two major limitations: insufficient feature preservation across anatomical structures and inadequate handling of pathological variations. + +This challenge is especially pronounced in abdominal imaging, where pathologies such as diffuse cirrhosis or multi-tissue disease processes manifest across multiple organ systems. Unlike the well-defined boundaries typical of tumors or simpler structures like bones in X-ray imaging, abdominal pathologies often involve subtle and heterogeneous changes in tissue characteristics, requiring more nuanced feature extraction and synthesis. The inherent complexity of MRI signals—spanning multiple sequences and high spatial resolution—presents additional computational hurdles, further distancing it from the more standardized nature of CT or simpler 2D radiographs. + +To address these gaps, our work introduces ViCTr (Vital Consistency Transfer), a two-stage framework that facilitates pathology-aware medical image synthesis with strong anatomical fidelity. We specifically target abdominal CT and MRI data, aiming to generate clinically relevant synthetic datasets that capture both normal anatomy and intricate pathological details. By providing robust augmentation material, our approach holds the potential to improve downstream tasks like segmentation and diagnosis, while alleviating issues of data scarcity and privacy constraints in medical imaging. Our contributions are: + +![](images/ccde6e94d4fd756b4cdc766a3dfee62e3782a1ac71e865a15b961163b194dacf.jpg) +Figure 1. Overview of the proposed ViCTr methodology. + +- Novel Two-Stage Framework. We propose ViCTR, a method that fuses anatomical consistency and pathological realism for CT and MRI synthesis with high diversity. +- Tweedie's Formula in Rectified Flow. We introduce a rectified flow trajectory reformulation of Tweedie's formula, ensuring accurate initialization and reducing sampling bias. +- Wide Applicability. ViCTR integrates seamlessly with various diffusion models, supporting a broad range of medical imaging tasks. +- One-Step Sampling & Tweedie's Corrections. We reduce computational overhead by cutting the number of diffusion steps from dozens to a handful, expediting inference without sacrificing quality. +- Quantitative Improvements in Segmentation. We demonstrate that adding ViCTR-synthesized CT/MRI images to real-world training sets significantly boosts segmentation performance, underscoring the practical benefits of pathology-focused data augmentation. +- Enhanced Image Fidelity. ViCTR achieves consistently lower FID (Fréchet Inception Distance) scores across multiple datasets, indicating superior structural and textural coherence. +- First Abdominal MRI Pathology Synthesis. To our knowledge, we are the first to generate abdominal MRI pathologies with progressive severity, offering new possibilities for research and clinical applications in medical imaging. + +By emphasizing both anatomical fidelity and pathological variety, ViCTr presents a powerful step toward bridging the gap between limited medical imaging data and the demands of high-performing AI models. + +# 2. Related Works + +# 2.1. Generative Models for Medical Data Augmentation + +Synthetic images have long been explored for data augmentation in medical imaging, with Generative Adversarial Networks (GANs) [25, 36, 50] initially offering promising results in tasks such as lesion synthesis and modality trans + +lation. However, GAN-based methods often struggle with mode collapse, requiring extensive architectural and training refinements to ensure sufficient diversity and realism in the synthesized images [12, 35]. + +# 2.2. Diffusion Models in Medical Imaging + +Diffusion-based approaches have emerged as a powerful alternative to GANs for high-quality synthetic image generation [4, 5, 13, 49]. Their inherent noise-to-image paradigm provides a more stable training regime and can produce richer data variations. In medical imaging, diffusion models [11] have been leveraged to improve segmentation and classification performance by generating diverse training samples [1, 10, 37, 46, 47]. Recent specialized frameworks, including segmentation-guided diffusion [26] and text-driven generation built upon RadImageNet [34, 54], highlight the adaptability of diffusion models to specific clinical scenarios. Efforts such as MixUp-enhanced augmentation [6, 28] and domain adaptation methods [55] demonstrate that diffusion techniques can mitigate biases and improve generalization in tasks like classification and risk prediction. + +# 2.3. General-Purpose Diffusion for Data Augmentation + +Various general-purpose diffusion strategies have been applied to further enrich training datasets. DiffuseMix [21] employs conditional prompts to blend real and synthetic data while preserving labels, whereas DreamDA [18] combines diffusion-based perturbations with pseudo-labeling for semantic consistency. DetDiffusion [51] incorporates object-detection attributes into the diffusion process, while Effective Data Augmentation with Diffusion Models [48] explores diffusion-based techniques to boost performance in few-shot settings. These methods generally aim to enhance data diversity and improve the training of downstream models. + +# 2.4. Rectified Flow Models + +While traditional diffusion approaches rely on discretized noise schedules, Rectified Flow (ReFlow) models learn + +![](images/29b383118786bec909fe7685f9ea896cf7d17c363f68fde4ed195210878dcebe.jpg) +Figure 2. Comparison of synthetic data distributions from various methods against the ground-truth (GT) on the AMOS dataset. The GT distribution (blue) reflects real data, while Reflow (red), Flow Straight and Fast (green), and Distribution Matching Distillation (purple) show varying alignment. Our proposed method, ViCTr (brown) achieves the closest match to GT, demonstrating superior distribution alignment and fidelity. + +smooth, near-linear trajectories to map between data distributions. This can result in more efficient sampling and reduced computational overhead. Recent works have leveraged rectified flows for improved convergence rates and lower numbers of function evaluations (NFE) [29, 30, 53], although most focus on general distribution transport rather than the strict alignment and pathology-aware realism needed in medical imaging. As shown in Figure 2, different ReFlow-based methods exhibit varying degrees of distribution matching, with our approach (ViCTr) offering closer alignment to the ground-truth distribution. + +# 2.5. Domain-Specific Adaptation + +Domain-focused adaptations continue to evolve. DiNO-Diffusion [24] introduces a self-supervised latent diffusion model to cope with limited annotated medical data, and Diverse Data Augmentation with Diffusions [17] enriches domain generalization using Stable Diffusion combined with cosine similarity filtering. Collectively, these methods illustrate the growing interest in customized diffusion pipelines for specialized domains, especially when high-quality annotated data are scarce. + +Recent advancements have further showcased the potential of diffusion models for domain-specific adaptation. DiNO-Diffusion [24] addresses the challenge of limited annotated data in medical imaging by leveraging a self-supervised latent diffusion framework. Similarly, Diverse Data Augmentation with Diffusions [17] enhances domain generalization by integrating Stable Diffusion with cosine similarity-based filtering, enabling the generation of semantically diverse and high-quality data. However, these methods prioritize general-domain efficiency over the anatomical-pathological alignment critical for medical imaging, often failing to preserve fine-grained structures + +like vascular networks or diffuse fibrosis patterns. + +In this work, we address the critical need for efficient, high-fidelity pathology-aware synthesis in abdominal imaging by merging rectified flow concepts with a Tweedie-corrected diffusion process. Our approach (ViCTr) stands out from prior methods by introducing a linearized sampling framework that enables one-step generation while maintaining anatomical fidelity—a crucial requirement for robust data augmentation in clinical applications. + +# 3. Methods + +Our framework, ViCTr, uses Rectified Flow [31] and Tweedie's Formula [43] for high-fidelity medical image synthesis with efficient one-step sampling. By rectifying the trajectory from a prior distribution $p_0$ to the target medical image distribution $p_{\mathrm{target}}$ , we reduce sampling bias and retain key anatomical and pathological details. + +# 3.1. Rectified Flow Trajectory + +Classical diffusion approaches gradually corrupt data samples $x \in \mathbb{R}^d$ into a Gaussian prior $p_0$ . While these methods have shown success, they often suffer from suboptimal sampling paths, cascading prediction errors, and diminished diversity in generated samples. We address these issues through Rectified Flow optimization [16, 31], which learns a continuous velocity field guiding the forward and reverse diffusion processes more directly. + +Let $(x_0, x_1)$ be a sample pair with $x_0 \sim p_0$ and $x_1 \sim p_{\mathrm{target}}$ . We define an interpolated point $x_t = (1 - t)x_0 + tx_1$ where $t \in [0, 1]$ . A velocity model $v_\theta : \mathbb{R}^d \times [0, 1] \to \mathbb{R}^d$ then predicts how to transition from $x_t$ toward $x_1$ . We train $v_\theta$ by minimizing: + +$$ +\hat {\theta} = \arg \min _ {\theta} \mathbb {E} _ {t \sim \operatorname {U n i f o r m} (0, 1)} \left[ \| \left(x _ {1} - x _ {0}\right) - v _ {\theta} \left(x _ {t}, t\right) \| ^ {2} \right], \tag {1} +$$ + +ensuring the predicted flow $v_{\theta}(x_t, t)$ aligns with the true path $(x_1 - x_o)$ . Once trained, the rectified flow is realized by solving the ODE: + +$$ +d x _ {t} = v _ {\hat {\theta}} \left(x _ {t}, t\right) d t, \tag {2} +$$ + +leading $x_0 \sim p_0$ toward $x_1 \sim p_{\mathrm{target}}$ . To enable one-step sampling, we distill this multi-step process into a neural network $\hat{\mathcal{N}}: \mathbb{R}^d \to \mathbb{R}^d$ , such that $\tilde{\mathcal{T}}(x_0) = x_0 + v(x_0, 0)$ , is trained to directly predict $x_1$ from $x_0$ . The loss function, + +$$ +\mathcal {L} = \mathbb {E} \left[ \left\| \left(x _ {1} - x _ {0}\right) - v \left(x _ {0}, 0\right) \right\| ^ {2} \right], \tag {3} +$$ + +drives $\mathcal{T}(\hat{x}_0)$ to approximate the final, rectified state $x_{1}$ + +# 3.2. Rectified Flow with Tweedie's Formula + +Though rectified flow refines sampling, medical data distributions often need extra bias correction. Tweedie's Formula [43] addresses this by adjusting noisy observations to better approximate the posterior mean. For Gaussian variables $z \sim \mathcal{N}(\mu_z, \Sigma_z)$ , Tweedie's formula indicates: + +$$ +\mathbb {E} [ \mu_ {z} \mid z ] = z + \Sigma_ {z} \nabla_ {z} \log p (z) +$$ + +, where $\nabla_z\log p(z)$ is the gradient of the log probability. In typical diffusion, Tweedie's formula estimates $x_0$ from noisy $x_{t}$ , guiding predictions toward the data manifold. + +ViCTR incorporates this correction into the rectified flow ODE: + +$$ +d x _ {t} = v _ {\hat {\theta}} (x _ {t}, t) d t + (1 - \bar {\alpha} _ {t}) \nabla_ {x _ {t}} \log p (x _ {t}) d t, +$$ + +and updates the training objective to: + +$$ +\begin{array}{l} \hat {\theta} = \arg \min _ {\theta} \mathbb {E} _ {t \sim \operatorname {U n i f o r m} (0, 1)} \left[ \| \left(x _ {1} - x _ {0}\right) - v _ {\theta} \left(x _ {t}, t\right) \right. \\ \left. \left. - \left(1 - \bar {\alpha} _ {t}\right) \nabla_ {x _ {t}} \log p (x _ {t}) \right| ^ {2} \right]. \\ \end{array} +$$ + +Here, $\bar{\alpha}_t$ denotes the cumulative product of variance decay factors in the diffusion schedule. The additional Tweedie term $(1 - \bar{\alpha}_t)\nabla_{x_t}\log p(x_t)$ corrects for sampling bias, driving $x_{t}$ more accurately toward the target distribution. + +We similarly extend the one-step distillation to integrate Tweedie's correction: + +$$ +\hat {\mathcal {T}} (x _ {0}) = x _ {0} + v (x _ {0}, 0) + (1 - \bar {\alpha} _ {0}) \nabla_ {x _ {0}} \log p (x _ {0}), +$$ + +and optimize + +$$ +\mathcal {L} = \mathbb {E} \left[ \left\| x _ {1} - \hat {\mathcal {T}} (x _ {0}) \right\| ^ {2} \right]. +$$ + +This single-step approach balances computational efficiency with high-quality, pathology-aware generation—critical in medical applications where datasets are limited and synthetic realism is paramount. + +In summary, ViCTr unifies rectified flow and Tweedie's correction to deliver anatomically consistent, pathology-aware sampling in a single forward pass. The resulting framework exhibits reduced inference costs, minimized sampling bias, and improved fidelity for generating high-resolution medical images. + +# 3.3. Pathology Aware Image Synthesis + +Our proposed ViCTr method adopts a two-stage training paradigm: Stage 1 establishes a foundational diffusion model tailored to medical imaging data, and Stage 2 fine-tunes this pre-trained model for downstream tasks, including semantic-guided multi-modal generation and counterfactual pathology synthesis. + +# Pre-training on ATLAS-8k Dataset (Stage 1) + +In the absence of large-scale, domain-specific pre-trained diffusion models, we begin by training on the ATLAS8k [40] dataset, which comprises abdominal CT scans and their segmentation annotations. We leverage these annotations as conditional guidance—enabling precise control over anatomical structures—and integrate textual prompts (e.g., "create image having $\langle \langle \text{organs} \rangle \rangle$ " to further refine semantic consistency. + +Latent Representations. Following Figure 1, we feed raw CT images $(X_I)$ and segmentation masks $(X_S)$ into a frozen VAE encoder, producing latent embeddings $(Z_o$ and $Z_s)$ . Simultaneously, textual prompts $(X_p)$ are processed by a pre-trained text encoder (details in Table 1) to yield prompt embeddings $(Z_p)$ . These latent representations guide the generative diffusion backbone $\phi_{\theta}$ during both forward and reverse diffusion. + +Forward Diffusion. We progressively inject noise into $Z_{o}$ forming a noisy representation $Z_{t}$ as + +$$ +P (Z _ {t} | Z _ {o}) = (1 - t) \cdot Z _ {o} + t \cdot \epsilon_ {\mathrm {t r u e}}, +$$ + +where $\epsilon_{true}$ is the true noise at step $t$ . We concatenate $Z_{t}$ with $Z_{s}$ and feed them into the diffusion model $\phi_{\theta}$ . Crucial layers are selectively unfrozen based on elastic weight consolidation, maintaining model stability while adapting to medical domain specifics. + +Reverse Diffusion. In the reverse process, + +$$ +P \left(Z _ {t - 1} \mid Z _ {t}, Z _ {s}, Z _ {p}, t\right) = Z _ {t} + \delta T \times \phi_ {\theta} \left(Z _ {t}, Z _ {s}, Z _ {p}, t\right), \tag {4} +$$ + +iteratively removes noise to reconstruct $Z_{t-1}$ conditioned on the segmentation map $(Z_s)$ and text prompt $(Z_s)$ . The diffusion loss $L_{diff}$ between $Z_{t-1}$ and $\epsilon_{true}$ is: + +$$ +L _ {\text {d i f f}} = - \left| \phi_ {\theta} \left(Z _ {t}, Z _ {s}, Z _ {p}, t\right) - \left(\epsilon_ {\text {t r u e}} - Z _ {o}\right) \right| ^ {2}. \tag {5} +$$ + +After refining $Z_{t-1}$ via Tweedie's formula, the denoised latent $Z_o$ is fed into the frozen VAE decoder (matching $\phi_\theta$ ; see Table 1) to obtain a reconstructed CT image. + +Composite Loss. To ensure both local accuracy and global perceptual realism, we combine: Diffusion Loss $(L_{\mathrm{diff}})$ , pixel-level reconstruction loss $(L_2)$ evaluating the gap between generated $X_{ig}$ and ground-truth $X_I$ , and structural similarity (SSIM) loss $(L_{SSIM})$ emphasizing higher-order textural correspondence. Optimizing this composite loss enables the diffusion backbone $\phi_{\theta}$ to capture essential anatomical features, establishing a versatile medical foundation model ready for fine-tuning. + +# Fine-tuning on downstream tasks (Stage 2) + +Building on the Stage1 pre-trained model, Stage2 targets specialized tasks like semantic-guided CT/MRI generation and counterfactual pathology synthesis. As shown in Figure 1, a dual-network setup is adopted. + +1. $\phi_{\text{base}}$ remains frozen, retaining the robust anatomical knowledge acquired in Stage 1. +2. $\phi_{\text {adapt }}$ is augmented with LoRA [19] modules within the previously fine-tuned layers. These modules are the only trainable parameters, ensuring targeted and stable adaptation. + +Training Input Consistency Loss. Similar to Stage1, each training sample includes prompts, segmentation maps, and either CT or MRI images (depending on the task). A frozen VAE encoder extracts latent representations $Z_{s}, Z_{o}$ , and $Z_{p}$ . Following forward diffusion of $Z_{o}$ into noisy $Z_{t}$ , both $\phi_{\text{base}}$ and $\phi_{\text{adapt}}$ process $[Z_{t}, Z_{s}]$ with text embedding $Z_{p}$ . We introduce a consistency loss $L_{\text{consistency}}$ to align intermediate outputs of $\phi_{\text{base}}$ and $\phi_{\text{adapt}}$ , ensuring stable adaptation without drifting from core representations. + +Loss Components. Diffusion loss $(L_{diff}^{\mathrm{base}}$ , $L_{diff}^{\mathrm{adapt}})$ encourages accurate noise prediction. Temporal consistency loss $L_{temporal}$ preserves smooth transitions across time steps in the reverse diffusion. Spatial consistency loss $L_{spatial}$ ensures alignment in the output reconstructions $(Z_{t-1}^{base}, (Z_{t-1}^{adapt})$ . Pixel-level losses $(L_2$ and $L_{SSIM})$ compare generated results to ground-truth scans, enforcing both local fidelity and structural coherence. + +CT and MRI Generation. We adapt to BTCV [15] (CT) and AMOS [23] (MRI) datasets, using organ segmentation masks for prompts. A standard 80/10/10 partition ensures robust evaluation. + +Pathology Generation. For liver cirrhosis, we employ CurrMRI600+ [22] dataset (T1/T2 MRIs) with segmentation masks marking liver regions. Prompts specify "low," "mild," or "severe" cirrhosis intensity. As in the CT/MRI setting, the training/validation/testing split is 80/10/10. The inference details and inference diagram are in the supplementary material. + +This selective LoRA-based fine-tuning allows ViCTR to handle complex tasks—ranging from normal organ generation to pathology synthesis—while preserving the anatomical and text-driven guidance learned in Stage 1. The experiments (Section 4) demonstrate that this two-stage framework consistently yields high-quality, clinically realistic images in both CT and MRI settings, including nuanced liver cirrhosis modeling. + +
Diffusion MethodsDenoiserText Encoder
Stable Diffusion [44]UNet [45]Clip-B/16 [41]
Pixart-alpha [9]Transformer (DiT) [38]T5-XXXXL [42]
Stable Diffusion XL [39]Dual Unet [39]Clip-L/14
Flux [27]MultiModal Transformer [14]T5-XXXXL + Clip-L/15
Stable Diffusion-3 [14]MultiModal TransformerT5-XXXXL + Clip-B/16 + Clip-L/14
+ +Table 1. Selection of denoisers and text encoders for different diffusion methods ${\phi }_{\theta }$ + +# 4. Experiments and Results + +The implementation details for training and sampling are provided in the supplementary material. + +# 4.1. Quantitative Results + +Synthetic Data Generation Results. Table 2 summarizes our comparative evaluation of synthetic medical image generation under vanilla fine-tuning versus the proposed ViCTR pipeline, spanning multiple datasets (BTCV, AMOS, Cir-rMRI600+) and diffusion backbones (Stable Diffusion, SD-XL, SD-3, Pixart-alpha, Flux). While vanilla fine-tuning applies a direct end-to-end approach after ATLAS-8k pretraining, it generally yields higher Fréchet Inception Distance (FID) and Medical FID (MFID) values across all datasets. Conversely, ViCTR integrates domain-specific loss functions and specialized architectural settings, consistently achieving lower FID/MFID scores and producing more realistic medical images. + +To address the well-known limitation of FID (which uses ImageNet features), we adopt M3D-CLIP [3] to compute a more domain-specific MFID. As shown in the table, ViCTR provides notable gains: for example, Stable Diffusion under ViCTR achieves FID/MFID of 21.98/19.02 (BTCV), 20.37/19.11 (AMOS), and 25.57/21.46 (Cir-rMRI600+). These results represent substantial improvements over vanilla fine-tuning and demonstrate consistent performance boosts across all tested architectures. Such gains highlight the robustness of ViCTR's two-stage framework in generating clinically meaningful synthetic images for both CT and MRI modalities. + +Segmentation Results. To assess the quality and utility of ViCTR-generated synthetic data, we performed segmentation experiments on three datasets—BTCV, AMOS, and CurrMRI600+—tracking mean Dice Similarity Coefficient (mDSC) and mean Hausdorff Distance 95 (mHD95). Higher mDSC and lower mHD95 respectively indicate improved overlap accuracy and spatial precision. We examined four training configurations: (1) baseline using original data only, (2) standard augmentation (Random Crop, Rotate, Blur, Affine, Geometric Distortion at 0.4 probability), (3) $30\%$ synthetic data via vanilla fine-tuning, and (4) $30\%$ synthetic data from ViCTR. Table 3 highlights consistent performance gains from ViCTR across all datasets and backbone architectures. Notably, CurrMRI600+—a liver cirrhosis dataset—exhibited the most significant improvements, reflecting ViCTR's capacity to capture complex pathological features. These results confirm that ViCTR-generated images significantly boost segmentation metrics (mDSC and mHD95), highlighting the practical value of our two-stage framework for medical data augmentation. + +Efficiency of ViCTR. Table 2 summarizes the number of diffusion steps and average inference times (in seconds) for vanilla fine-tuning versus the ViCTR framework. Across all + +
BaselinesBTCV Dataset (CT Generation)AMOS Dataset (MRI Generation)CirrMRI600+ (Pathology Generation)Diffusion Steps | Inference Time
Fine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTR
Stable Diffusion [44]25.44 / 19.6721.98 / 19.0225.43 / 21.7620.37 / 19.1128.34 / 23.4325.57 / 21.4640 | 13.764 | 3.12
Stable Diffusion XL [39]23.47 / 18.2120.33 / 17.4424.11 / 20.2319.44 / 18.4527.34 / 22.1124.02 / 20.7630 | 14.554 | 3.45
Stable Diffusion-3 [14]19.07 / 16.2217.37 / 16.0222.32 / 19.7618.02 / 19.0824.49 / 21.7821.28 / 19.3450 | 18.983 | 2.78
Pixart-alpha [9]21.32 / 17.0919.22 / 16.9623.78 / 20.0418.76 / 18.5626.06 / 20.0723.04 / 18.9225 | 10.673 | 1.74
Flux [27]15.52 / 15.0113.28 / 14.0819.02 / 18.2815.55 / 16.5822.46 / 18.8819.96 / 17.0130 | 15.663 | 2.87
+ +Table 2. Quantitative results on CT, MRI, and Cirrhosis generation. All baselines were pre-trained on ATLAS-8k and fine-tuned on target datasets. Metrics reported are Fréchet Inception Distance (FID) and Medical FID (MFID), shown as FID/MFID, with inference time in seconds. Lower values indicate better performance. + +
BaselinesOriginal Dataset (Org.)AugmentationOrg. + Synth. by FLUX 30% (vanilla Fine-tuning)Org. + Synth. by FLUX 30% (ViCtR)
mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)
BTCV dataset Segmentation results
UNet [45]76.7234.4278.4533.1779.3232.3781.2230.17
TransUnet [8]85.5232.3387.0131.4387.5430.1189.7829.12
nnUnet [20]80.4830.1982.5431.0283.3729.0185.1927.77
nnFormer [56]83.4730.0183.9830.8885.8828.7887.7226.22
MedSegDiff [52]87.9127.6788.6526.7289.7825.5291.9223.31
AMOS dataset Segmentation results
UNet [45]68.9234.5770.5532.3271.0231.7873.3429.11
TransUnet [8]71.3333.1272.5631.0973.4329.1977.5427.56
nnUnet [20]73.3332.3274.1130.1975.6828.7578.2926.32
nnFormer [56]75.7831.1976.4429.7677.7727.5781.3224.41
MedSegDiff [52]76.8329.9278.3826.9979.0325.4584.0222.18
CirrMRI600+ dataset Segmentation results
UNet [45]68.7436.7369.3835.4370.1235.1173.3932.11
TransUnet [8]70.7735.4270.9834.0171.7833.9274.5631.09
nnUnet [20]71.0234.3572.4932.7873.5631.2778.8930.25
nnFormer [56]74.8833.7875.2331.8876.4531.0979.4429.78
MedSegDiff [52]76.9230.7977.1130.3478.0329.8981.3727.34
+ +Table 3. Segmentation results on BTCV, AMOS, and CurrMRI600+ using various training settings and baselines. Metrics reported are mDSC (%) and mHD95 (mm). Top three results per setting are highlighted: blue (1st), green (2nd), and orange (3rd). + +evaluated models, ViCTR consistently reduces the required steps, thereby shortening inference duration without compromising image fidelity. This efficiency stems from two key innovations: (1) Tweedy's formula to approximate $Z_{0}$ and streamlines the reconstruction process, and (2) Two-Stage Training, enabling the model to learn both anatomical consistency and pathology-specific variations in a targeted manner. By minimizing extraneous steps, ViCTR can generate high-quality synthetic images more rapidly, making it a viable solution for resource-constrained clinical or research environments where computational overhead is a critical consideration. + +# 4.2. Qualitative Results + +We conducted extensive qualitative evaluations of ViCTr across multiple modalities (CT and MRI) and tasks, covering both anatomical and pathological image synthesis. + +Anatomical and Pathology-Driven Image Generation. Unlike prior methods that primarily generate normal anatomical scans or isolated tumors, ViCTr extends synthesis capabilities to complex pathologies, such as liver cirrhosis. Figure 3 shows pairs of non-cirrhotic MRIs alongside + +their segmentation masks and the corresponding ViCTR-generated cirrhotic images. Notably, the synthesized images retain essential anatomical structures while introducing realistic cirrhotic texturing, an advancement that enables more diverse data augmentation and supports research into disease progression. These high-fidelity synthetic samples offer potential to enrich clinical training datasets, advance diagnostic algorithms, and support detailed analyses of pathology progression. + +Qualitative Segmentation Outcomes. To further showcase ViCTR's impact on downstream tasks, we visually assessed segmentation quality on BTCV and AMOS (Figure 6). Focusing on regions like the left kidney and pancreas, models trained with ViCTR-augmented data show sharper boundaries and more precise delineation than standard fine-tuning, indicating improved spatial learning. + +Comparative Performance on BTCV and AMOS. Figures 4 (BTCV) and 5 (AMOS) compare diffusion-based image generation results using Stable Diffusion v3 (SD-3) and Flux backbones. On BTCV—a dataset with limited samples and larger spatial dimensions—vanilla finetuning with SD-3 often yields poor mask adherence, sug + +![](images/acbff25d698bee988afcb9bd7ad92bcea0f82dd02abecb03a68b130a14859499.jpg) + +![](images/2b4919450eea532967b234637106888ca96f72a6a0afc07acf66fdedeb80ccae.jpg) +Figure 3. Examples of synthetic cirrhotic images with different severity levels generated from non-cirrhotic scans using ViCTr. For T1-MRI, we show the segmentation mask, the original non-cirrhotic image, and the corresponding synthetic cirrhotic image. + +![](images/5590f8ff02fb0f7fd96cf80c038fe2b42c2fbfee3326fd783f47ea34103b500d.jpg) +Figure 4. Diffusion-Based Image Generation on BTCV MRI Dataset. From left to right: segmentation mask, SD-3 with standard fine-tuning, FLUX with standard fine-tuning, SD-3 with ViCTR, FLUX with ViCTR. +Figure 5. Diffusion-Based Image Generation on AMOS Dataset. From left to right: segmentation mask, SD-3 with Vanilla fine-tuning, SD-3 with our method, FLUX with Vanilla fine-tuning, FLUX with our method. + +gesting difficulties in learning accurate spatial constraints. In contrast, the same model shows markedly improved mask alignment on the AMOS dataset, attributed to its larger training set and more manageable image dimensions. + +Regardless of these dataset-specific variations, ViCTr achieves consistently high-quality outputs with robust mask adherence and anatomical accuracy across both BTCV and AMOS, underscoring its enhanced spatial learning and resilience. By maintaining performance even under data-sparse conditions, ViCTr demonstrates a clear advantage over conventional fine-tuning approaches, which can exhibit greater sensitivity to dataset size and complexity. + +Second, we evaluate the impact of ViCTR on segmentation tasks, illustrated in Figure 6. Both BTCV and AMOS datasets encompass various abdominal structures (e.g., left kidney, pancreas), and our results highlight how ViCTR-based training yields notably more precise organ boundaries compared to alternative methods. This improvement is particularly evident in challenging regions where standard segmentation often struggle, highlighting ViCTR's strength in enhancing spatial delineation and anatomical accuracy. + +# 4.3. Ablation Studies + +
Ablation Based on Pre-training of Model
ViCTR (Without Stage-1)17.33
Ablation Based on Rectified Flow Algorithms
ViCTR (Without Proposed Tweedies Formula)18.78
ViCTR (With Reflow)20.19
ViCTR (With Flow Straight and Fast)21.37
ViCTR (With Distribution Matching Distillation)22.33
Ablation Study Based on Loss FunctionsAblation Based on LoRA Rank
Ldiff18.77r = 818.46
Ldiff + Lspatial18.21r = 1617.52
Ldiff + Lconsistency17.02r = 3216.66
Ldiff + Lspatial + Lconsistency15.55r = 6415.55
+ +Table 4. Ablation study for the effect of model pretraining, loss functions, and LoRA rank settings are shown. + +To rigorously validate our ViCTr framework, we conducted a series of ablation experiments on the BTCV dataset (Table 4), focusing on four key aspects: pretraining, rectified flow algorithms, loss function components, and LoRA rank selection. We used FID (Frechet Inception Distance) as our primary metric to gauge how each architectural choice affects both anatomical fidelity and pathological realism. + +Impact of Pretraining. We first examined the necessity of Stage 1 pretraining by comparing our full, two-stage pipeline to a variant that proceeds directly to downstream fine-tuning. Omitting Stage 1 degrades the model's FID score from 15.55 to 17.33 (Table 4), underscoring the vital role of an anatomical prior. These findings affirm that establishing robust structural representations in Stage 1 is crucial for achieving high-quality medical image synthesis. + +Rectified Flow Algorithms. We next assessed our rectified flow formulation, augmented with Tweedie's formula, + +![](images/f77422e60a89252daac2dd04e67c939d78ee91f5c636912898c7b360e493bc01.jpg) +Figure 6. Segmentation results on BTCV and AMOS datasets' left kidney and pancreas. We compare the ground truth (blue), models trained on the default dataset, augmented dataset, default dataset plus $30\%$ synthetic data generated by FLUX with standard fine-tuning, and default dataset plus $30\%$ synthetic data generated by FLUX with ViCTr (red). + +against alternative rectified flow-based models, including ReFlow, Flow Straight and Fast, and Distribution Matching Distillation. Removing Tweedie's correction alone increased the FID to 18.78, highlighting its significance for trajectory alignment. Across all baselines, we observed FID deteriorations, ranging from 13.28 to as high as 22.33, thereby demonstrating that our proposed method provides tighter distribution matching and superior synthesis fidelity. + +Ablation Study Based on Loss Functions. We then investigated each ViCTr loss component—diffusion loss, spatial consistency, and consistency loss—to determine their relative contributions: the baseline model, utilizing only diffusion loss, produced an FID of 18.77 (Table 4). Incorporating spatial consistency loss yielded a moderate improvement to 18.21, highlighting its role in maintaining structural integrity. The addition of consistency loss further enhanced performance, reducing the FID to 17.02 by promoting image coherence across generations. When combining all three terms, we achieved an FID of 15.55, indicating that these components collectively offer significant improvements in anatomical accuracy and visual realism. + +Ablation Based on LoRA Rank. Finally, we explored how varying the LoRA rank $(r)$ influences generative performance. Starting at $r = 8$ and incrementing upwards, we found $r = 64$ to yield the best FID at 15.55. This result suggests that higher-dimensional adaptation spaces enable more nuanced parameter updates during fine-tuning, thereby improving visual quality and fidelity. + +Visual Turing Tests. To further validate clinical plausibility, three radiologists participated in Visual Turing Tests using 15 randomly generated MRI scans depicting vary + +ing levels of liver cirrhosis (mild, moderate, severe). All scans were uniformly judged to be clinically realistic, with identical outcomes observed when using additional random samples. Notably, the synthetic cirrhotic images correctly exhibited surface nodularity and textural irregularities, consistent with radiologic findings in mild-to-severe cirrhosis (Figure 3). These results confirm ViCTr's ability to synthesize pathology-specific features, enabling applications in training, algorithm development, and clinical research. + +# 5. Conclusion + +We presented ViCTR, a novel two-stage framework designed to generate high-fidelity medical images by integrating robust anatomical pre-training with precise pathology-specific fine-tuning. By aligning Tweedie's formula with linear projection methods used in flow matching, ViCTR maintains accurate initial distribution estimates even amid diffusion processes. This setup is further enhanced by LoRA adapters, which preserve essential anatomical information while flexibly adapting to diverse pathologies. Experimental results demonstrate that ViCTR consistently surpasses conventional fine-tuning strategies, reducing FID scores and producing clinically realistic outputs. Beyond data augmentation for segmentation and classification, ViCTR can be extended to tasks such as modality translation, contrast synthesis, and professional training, offering a powerful and versatile tool for advancing AI-driven medical imaging. + +# Acknowledgments + +This research is supported by the following NIH grants: R01-HL171376 and U01-CA268808. + +# References + +[1] Hazrat Ali, Shafaq Murad, and Zubair Shah. Spot the fake lungs: Generating synthetic medical images using neural diffusion models. In Artificial Intelligence and Cognitive Science, pages 32-39, Cham, 2023. Springer Nature Switzerland. 2 +[2] Alhanoof Althnian, Duaa Alsaeed, Heyam Al-Baity, Amani Samha, Alanoud Bin Dris, Najla Alzakari, Afnan Abou Elwafa, and Heba Kurdi. Impact of dataset size on classification performance: an empirical evaluation in the medical domain. Applied Sciences, 11(2):796, 2021. 1 +[3] Fan Bai, Yuxin Du, Tiejun Huang, Max Q. H. Meng, and Bo Zhao. M3d: Advancing 3d medical image analysis with multi-modal large language models, 2024. 5 +[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 2 +[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 2 +[6] Luigi Carratino, Moustapha Cisse, Rodolphe Jenatton, and Jean-Philippe Vert. On mixup regularization. Journal of Machine Learning Research, 23(325):1-31, 2022. 2 +[7] Junyi Chai, Hao Zeng, Anming Li, and Eric W.T. Ngai. Deep learning in computer vision: A critical review of emerging techniques and application scenarios. Machine Learning with Applications, 6:100134, 2021. 1 +[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6 +[9] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation, 2024. 5, 6 +[10] Qi Chen, Xiaoxi Chen, Haorui Song, Zhiwei Xiong, Alan Yuille, Chen Wei, and Zongwei Zhou. Towards generalizable tumor synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11147-11158, 2024. 2 +[11] Gayatri Deshmukh, Onkar Kishor Susladkar, Debesh Jha, Elif Keles, Halil Ertugrul Aktas, Daniela P Ladner, Amir A Borhani, Gorkem Durak, Ulas Bagci, et al. Meddelinea: Scalable and efficient medical image segmentation via controllable diffusion transformers. In Medical Imaging with Deep Learning, 2025. 2 + +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[13] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E. Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 5, 6 +[15] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619-3629, 2020. 5 +[16] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Flux that plays music, 2024. 3 +[17] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 3 +[18] Yunxiang Fu, Chaoqi Chen, Yu Qiao, and Yizhou Yu. Dreamda: Generative data augmentation with diffusion models. arXiv preprint arXiv:2403.12803, 2024. 2 +[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5 +[20] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 6 +[21] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27621-27630, 2024. 2 +[22] Debesh Jha, Onkar Kishor Susladkar, Vandan Gorade, Elif Keles, Matthew Antalek, Deniz Seyithanoglu, Timurhan Cebeci, Halil Ertugrul Aktas, Gulbiz Dagoglu Kartal, Sabahattin Kaymakoglu, et al. Cirrmri600+: Large scale mri collection and segmentation of cirrhotic liver. arXiv preprint arXiv:2410.16296, 2024. 5 +[23] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems, 35:36722-36732, 2022. 5 +[24] Guillermo Jimenez-Perez, Pedro Osorio, Josef Cersovsky, Javier Montalt-Tordera, Jens Hooge, Steffen Vogler, and Sadegh Mohammadi. Dino-diffusion. scaling medical diffusion via self-supervised pre-training. arXiv preprint arXiv:2407.11594, 2024. 3 + +[25] Sunho Kim, Byungjai Kim, and HyunWook Park. Synthesis of brain tumor multicontrast mr images for improved data augmentation. Medical Physics, 48(5):2185-2198, 2021. 2 +[26] Nicholas Konz, Yuwen Chen, Haoyu Dong, and Maciej A Mazurowski. Anatomically-controllable medical image generation with segmentation-guided diffusion models. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 88–98. Springer, 2024. 2 +[27] Black Forest Labs. black-forest-labs/FLUX.1-dev $\cdot$ Hugging Face. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. [Online; accessed 05-October-2024]. 5, 6 +[28] Hansang Lee, Haeil Lee, and Helen Hong. Genmix: Combining generative and mixture data augmentation for medical image classification. arXiv preprint arXiv:2405.20650, 2024. 2 +[29] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. Advances in Neural Information Processing Systems, 37:63082-63109, 2025. 3 +[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3 +[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 3 +[32] Supriya V. Mahadevkar, Bharti Khemani, Shruti Patil, Ketan Kotecha, Deepali R. Vora, Ajith Abraham, and Lubna Abdelkareim Gabralla. A review on machine learning styles in computer vision—techniques and future directions. IEEE Access, 10:107293–107329, 2022. 1 +[33] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M. Alvarez, Zhiding Yu, Sanja Fidler, and Marc T. Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 275-284, 2022. 1 +[34] Xueyan Mei, Zelong Liu, Philip M Robson, Brett Marinelli, Mingqian Huang, Amish Doshi, Adam Jacobi, Chendi Cao, Katherine E Link, Thomas Yang, Ying Wang, Hayit Greenspan, Timothy Deyer, Zahi A Fayad, and Yang Yang. RadImageNet: An open radiologic deep learning research dataset for effective transfer learning. Radiol Artif Intell, 4 (5):e210315, 2022. 2 +[35] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, Jakob Nikolas Kather, and Daniel Truhn. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13(1):12098, 2023. 2 +[36] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part III 20, pages 417-425. Springer, 2017. 2 + +[37] Kaliprasad Pani and Indu Chawla. Synthetic mri in action: A novel framework in data augmentation strategies for robust multi-modal brain tumor segmentation. Computers in Biology and Medicine, 183:109273, 2024. 2 +[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5 +[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5, 6 +[40] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024. 4 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5 +[43] Herbert E. Robbins. An Empirical Bayes Approach to Statistics, pages 388-394. Springer New York, New York, NY, 1992. 3, 4 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 5, 6 +[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5, 6 +[46] Daniel G. Saragih, Atsuhiro Hibi, and Pascal N. Tyrrell. Using diffusion models to generate synthetic labeled data for medical image segmentation. International Journal of Computer Assisted Radiology and Surgery, 19(8):1615-1625, 2024. 2 +[47] Vanshali Sharma, Abhishek Kumar, Debesh Jha, M.K. Bhuyan, Pradip K. Das, and Ulas Bagci. Controlpolynet: Towards controlled colon polyp synthesis for improved polyp segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2325-2334, 2024. 2 +[48] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2 +[49] Brandon Trabucco, Kyle Doherty, Max A Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffu + +sion models. In The Twelfth International Conference on Learning Representations, 2024. 2 +[50] Toan Tran, Trung Pham, Gustavo Carneiro, Lyle Palmer, and Ian Reid. A bayesian data augmentation approach for learning deep models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2 +[51] Yibo Wang, Ruiyuan Gao, Kai Chen, Kaiqiang Zhou, Yingjie Cai, Lanqing Hong, Zhenguo Li, Lihui Jiang, DitYan Yeung, Qiang Xu, et al. Detdiffusion: Synergizing generative and perceptive models for enhanced data generation and perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7246-7255, 2024. 2 +[52] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In Medical Imaging with Deep Learning, pages 1623-1639. PMLR, 2024. 6 +[53] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 3 +[54] Zheyuan Zhang, Lanhong Yao, Bin Wang, Debesh Jha, Elif Keles, Alpay Medetalibeyoglu, and Ulas Bagci. Emit-diff: Enhancing medical image segmentation via text-guided diffusion model. arXiv preprint arXiv:2310.12868, 2023. 2 +[55] Yuan Zhong, Suhan Cui, Jiaqi Wang, Xiaochen Wang, Ziyi Yin, Yaqing Wang, Houping Xiao, Mengdi Huai, Ting Wang, and Fenglong Ma. MedDiffusion: Boosting Health Risk Prediction via Diffusion-based Data Augmentation, pages 499-507. 2 +[56] Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Lequan Yu, Liansheng Wang, and Yizhou Yu. nnformer: Interleaved transformer for volumetric segmentation. arXiv preprint arXiv:2109.03201, 2021. 6 \ No newline at end of file diff --git a/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/images.zip b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0c2e4c74896138c9972ca52489dbb32fd294a8e --- /dev/null +++ b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91939c7132f18d374813fd5725a91d0853e0bd1e24b3c98c2d80e1c495b96fd0 +size 677096 diff --git a/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/layout.json b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2bffb1e6f6e945361176730701e816b6695fd023 --- /dev/null +++ b/2025/ViCTr_ Vital Consistency Transfer for Pathology Aware Image Synthesis/layout.json @@ -0,0 +1,9164 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "type": "text", + "content": "ViCTr: Vital Consistency Transfer for Pathology Aware Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": "Onkar Susladkar" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": ", Gayatri Deshmukh" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": ", Yalcin Tur" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": ", Gorkem Durak" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": ", Ulas Bagci" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": "Northwestern University " + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": "Stanford University " + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 142, + 533, + 173 + ], + "type": "text", + "content": "University of Illinois Ubrana-Champaign" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 174, + 487, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 174, + 487, + 185 + ], + "spans": [ + { + "bbox": [ + 120, + 174, + 487, + 185 + ], + "type": "text", + "content": "onkarsus13@gmail.com, dgayatri9850@gmail.com, yalcintr@stanford.edu," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 188, + 460, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 188, + 460, + 198 + ], + "spans": [ + { + "bbox": [ + 146, + 188, + 460, + 198 + ], + "type": "text", + "content": "gorkem.durak@northwestern.edu, ulas.bagci@northwestern.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 251, + 297, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 624 + ], + "type": "text", + "content": "Synthesizing medical images remains challenging due to limited annotated pathological data, modality domain gaps, and the complexity of representing diffuse pathologies such as liver cirrhosis. Existing methods often struggle to maintain anatomical fidelity while accurately modeling pathological features, frequently relying on priors derived from natural images or inefficient multi-step sampling. In this work, we introduce ViCTR (Vital Consistency Transfer), a novel two-stage framework that combines a rectified flow trajectory with a Tweedie-corrected diffusion process to achieve high-fidelity, pathology-aware image synthesis. First, we pretrain ViCTR on the ATLAS-8k dataset using Elastic Weight Consolidation (EWC) to preserve critical anatomical structures. We then fine-tune the model adversarially with Low-Rank Adaptation (LoRA) modules for precise control over pathology severity. By reformulating Tweedie's formula within a linear trajectory framework, ViCTR supports one-step sampling—reducing inference from 50 steps to just 4—without sacrificing anatomical realism. We evaluate ViCTR on BTCV (CT), AMOS (MRI), and CirrMRI600+ (cirrhosis) datasets. Results demonstrate state-of-the-art performance, achieving a Medical Fréchet Inception Distance (MFID) of 17.01 for cirrhosis synthesis—28% lower than existing approaches—and improving nnUNet segmentation by +3.8% mDSC when used for data augmentation. Radiologist reviews indicate that ViCTR-generated liver cirrhosis MRIs are clinically indistinguishable from real scans. To our knowledge, ViCTR is the first method to provide fine-grained, pathology-aware MRI synthesis with graded severity control, closing a critical gap in AI-driven medical imaging research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 135, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "The exponential growth in computer vision capabilities has been driven by significant advances in artificial intelligence models [7, 32]. However, medical imaging faces a fundamental tension between model complexity and data avail" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": "ability that limits the application of state-of-the-art techniques. While recent breakthroughs in generative AI have demonstrated remarkable capabilities in synthetic data creation, these advances demand training datasets of unprecedented scale—a requirement that poses unique challenges in the medical domain [2, 33]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 301, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 556, + 434 + ], + "type": "text", + "content": "Unlike general computer vision applications, medical imaging faces several critical constraints: privacy regulations necessitating complex deidentification, inherent data fragmentation across healthcare institutions, and fundamental interoperability constraints. These barriers have created a growing disparity between the rapid advancement of general computer vision and the relatively slower progress in medical imaging applications. Current approaches to bridging this gap face two major limitations: insufficient feature preservation across anatomical structures and inadequate handling of pathological variations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 436, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 581 + ], + "type": "text", + "content": "This challenge is especially pronounced in abdominal imaging, where pathologies such as diffuse cirrhosis or multi-tissue disease processes manifest across multiple organ systems. Unlike the well-defined boundaries typical of tumors or simpler structures like bones in X-ray imaging, abdominal pathologies often involve subtle and heterogeneous changes in tissue characteristics, requiring more nuanced feature extraction and synthesis. The inherent complexity of MRI signals—spanning multiple sequences and high spatial resolution—presents additional computational hurdles, further distancing it from the more standardized nature of CT or simpler 2D radiographs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "To address these gaps, our work introduces ViCTr (Vital Consistency Transfer), a two-stage framework that facilitates pathology-aware medical image synthesis with strong anatomical fidelity. We specifically target abdominal CT and MRI data, aiming to generate clinically relevant synthetic datasets that capture both normal anatomy and intricate pathological details. By providing robust augmentation material, our approach holds the potential to improve downstream tasks like segmentation and diagnosis, while alleviating issues of data scarcity and privacy constraints in medical imaging. Our contributions are:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22772" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 182 + ], + "type": "image", + "image_path": "ccde6e94d4fd756b4cdc766a3dfee62e3782a1ac71e865a15b961163b194dacf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 188, + 410, + 200 + ], + "lines": [ + { + "bbox": [ + 200, + 188, + 410, + 200 + ], + "spans": [ + { + "bbox": [ + 200, + 188, + 410, + 200 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed ViCTr methodology." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 221, + 295, + 555 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 221, + 294, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 294, + 257 + ], + "type": "text", + "content": "- Novel Two-Stage Framework. We propose ViCTR, a method that fuses anatomical consistency and pathological realism for CT and MRI synthesis with high diversity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 304 + ], + "type": "text", + "content": "- Tweedie's Formula in Rectified Flow. We introduce a rectified flow trajectory reformulation of Tweedie's formula, ensuring accurate initialization and reducing sampling bias." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 341 + ], + "type": "text", + "content": "- Wide Applicability. ViCTR integrates seamlessly with various diffusion models, supporting a broad range of medical imaging tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 295, + 388 + ], + "type": "text", + "content": "- One-Step Sampling & Tweedie's Corrections. We reduce computational overhead by cutting the number of diffusion steps from dozens to a handful, expediting inference without sacrificing quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 449 + ], + "type": "text", + "content": "- Quantitative Improvements in Segmentation. We demonstrate that adding ViCTR-synthesized CT/MRI images to real-world training sets significantly boosts segmentation performance, underscoring the practical benefits of pathology-focused data augmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 495 + ], + "type": "text", + "content": "- Enhanced Image Fidelity. ViCTR achieves consistently lower FID (Fréchet Inception Distance) scores across multiple datasets, indicating superior structural and textural coherence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 496, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 496, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 496, + 295, + 555 + ], + "type": "text", + "content": "- First Abdominal MRI Pathology Synthesis. To our knowledge, we are the first to generate abdominal MRI pathologies with progressive severity, offering new possibilities for research and clinical applications in medical imaging." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 558, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 606 + ], + "type": "text", + "content": "By emphasizing both anatomical fidelity and pathological variety, ViCTr presents a powerful step toward bridging the gap between limited medical imaging data and the demands of high-performing AI models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 616, + 147, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 147, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 147, + 628 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 659 + ], + "type": "text", + "content": "2.1. Generative Models for Medical Data Augmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Synthetic images have long been explored for data augmentation in medical imaging, with Generative Adversarial Networks (GANs) [25, 36, 50] initially offering promising results in tasks such as lesion synthesis and modality trans" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 221, + 554, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 554, + 270 + ], + "type": "text", + "content": "lation. However, GAN-based methods often struggle with mode collapse, requiring extensive architectural and training refinements to ensure sufficient diversity and realism in the synthesized images [12, 35]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 279, + 511, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 511, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 511, + 293 + ], + "type": "text", + "content": "2.2. Diffusion Models in Medical Imaging" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 477 + ], + "type": "text", + "content": "Diffusion-based approaches have emerged as a powerful alternative to GANs for high-quality synthetic image generation [4, 5, 13, 49]. Their inherent noise-to-image paradigm provides a more stable training regime and can produce richer data variations. In medical imaging, diffusion models [11] have been leveraged to improve segmentation and classification performance by generating diverse training samples [1, 10, 37, 46, 47]. Recent specialized frameworks, including segmentation-guided diffusion [26] and text-driven generation built upon RadImageNet [34, 54], highlight the adaptability of diffusion models to specific clinical scenarios. Efforts such as MixUp-enhanced augmentation [6, 28] and domain adaptation methods [55] demonstrate that diffusion techniques can mitigate biases and improve generalization in tasks like classification and risk prediction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 487, + 553, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 511 + ], + "type": "text", + "content": "2.3. General-Purpose Diffusion for Data Augmentation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": "Various general-purpose diffusion strategies have been applied to further enrich training datasets. DiffuseMix [21] employs conditional prompts to blend real and synthetic data while preserving labels, whereas DreamDA [18] combines diffusion-based perturbations with pseudo-labeling for semantic consistency. DetDiffusion [51] incorporates object-detection attributes into the diffusion process, while Effective Data Augmentation with Diffusion Models [48] explores diffusion-based techniques to boost performance in few-shot settings. These methods generally aim to enhance data diversity and improve the training of downstream models." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 671, + 441, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 441, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 441, + 682 + ], + "type": "text", + "content": "2.4. Rectified Flow Models" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "While traditional diffusion approaches rely on discretized noise schedules, Rectified Flow (ReFlow) models learn" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22773" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 194 + ], + "type": "image", + "image_path": "29b383118786bec909fe7685f9ea896cf7d17c363f68fde4ed195210878dcebe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 202, + 296, + 280 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 280 + ], + "type": "text", + "content": "Figure 2. Comparison of synthetic data distributions from various methods against the ground-truth (GT) on the AMOS dataset. The GT distribution (blue) reflects real data, while Reflow (red), Flow Straight and Fast (green), and Distribution Matching Distillation (purple) show varying alignment. Our proposed method, ViCTr (brown) achieves the closest match to GT, demonstrating superior distribution alignment and fidelity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 301, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 301, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 295, + 434 + ], + "type": "text", + "content": "smooth, near-linear trajectories to map between data distributions. This can result in more efficient sampling and reduced computational overhead. Recent works have leveraged rectified flows for improved convergence rates and lower numbers of function evaluations (NFE) [29, 30, 53], although most focus on general distribution transport rather than the strict alignment and pathology-aware realism needed in medical imaging. As shown in Figure 2, different ReFlow-based methods exhibit varying degrees of distribution matching, with our approach (ViCTr) offering closer alignment to the ground-truth distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 211, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 211, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 211, + 456 + ], + "type": "text", + "content": "2.5. Domain-Specific Adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 568 + ], + "type": "text", + "content": "Domain-focused adaptations continue to evolve. DiNO-Diffusion [24] introduces a self-supervised latent diffusion model to cope with limited annotated medical data, and Diverse Data Augmentation with Diffusions [17] enriches domain generalization using Stable Diffusion combined with cosine similarity filtering. Collectively, these methods illustrate the growing interest in customized diffusion pipelines for specialized domains, especially when high-quality annotated data are scarce." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Recent advancements have further showcased the potential of diffusion models for domain-specific adaptation. DiNO-Diffusion [24] addresses the challenge of limited annotated data in medical imaging by leveraging a self-supervised latent diffusion framework. Similarly, Diverse Data Augmentation with Diffusions [17] enhances domain generalization by integrating Stable Diffusion with cosine similarity-based filtering, enabling the generation of semantically diverse and high-quality data. However, these methods prioritize general-domain efficiency over the anatomical-pathological alignment critical for medical imaging, often failing to preserve fine-grained structures" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 515, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 515, + 84 + ], + "type": "text", + "content": "like vascular networks or diffuse fibrosis patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 85, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 182 + ], + "type": "text", + "content": "In this work, we address the critical need for efficient, high-fidelity pathology-aware synthesis in abdominal imaging by merging rectified flow concepts with a Tweedie-corrected diffusion process. Our approach (ViCTr) stands out from prior methods by introducing a linearized sampling framework that enables one-step generation while maintaining anatomical fidelity—a crucial requirement for robust data augmentation in clinical applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 195, + 375, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 375, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 375, + 208 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 289 + ], + "type": "text", + "content": "Our framework, ViCTr, uses Rectified Flow [31] and Tweedie's Formula [43] for high-fidelity medical image synthesis with efficient one-step sampling. By rectifying the trajectory from a prior distribution " + }, + { + "bbox": [ + 313, + 216, + 555, + 289 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 216, + 555, + 289 + ], + "type": "text", + "content": " to the target medical image distribution " + }, + { + "bbox": [ + 313, + 216, + 555, + 289 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 216, + 555, + 289 + ], + "type": "text", + "content": ", we reduce sampling bias and retain key anatomical and pathological details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 298, + 457, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 457, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 457, + 312 + ], + "type": "text", + "content": "3.1. Rectified Flow Trajectory" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 413 + ], + "type": "text", + "content": "Classical diffusion approaches gradually corrupt data samples " + }, + { + "bbox": [ + 313, + 316, + 554, + 413 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 316, + 554, + 413 + ], + "type": "text", + "content": " into a Gaussian prior " + }, + { + "bbox": [ + 313, + 316, + 554, + 413 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 316, + 554, + 413 + ], + "type": "text", + "content": ". While these methods have shown success, they often suffer from suboptimal sampling paths, cascading prediction errors, and diminished diversity in generated samples. We address these issues through Rectified Flow optimization [16, 31], which learns a continuous velocity field guiding the forward and reverse diffusion processes more directly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "(x_0, x_1)" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": " be a sample pair with " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "x_0 \\sim p_0" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "x_1 \\sim p_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": ". We define an interpolated point " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "x_t = (1 - t)x_0 + tx_1" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "t \\in [0, 1]" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": ". A velocity model " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "v_\\theta : \\mathbb{R}^d \\times [0, 1] \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": " then predicts how to transition from " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": " toward " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": ". We train " + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "inline_equation", + "content": "v_\\theta" + }, + { + "bbox": [ + 313, + 414, + 554, + 474 + ], + "type": "text", + "content": " by minimizing:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 495, + 553, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 495, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 318, + 495, + 553, + 530 + ], + "type": "interline_equation", + "content": "\\hat {\\theta} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t \\sim \\operatorname {U n i f o r m} (0, 1)} \\left[ \\| \\left(x _ {1} - x _ {0}\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\| ^ {2} \\right], \\tag {1}", + "image_path": "17bef26d0240dd3b1554f5c4106b51acc91dd1a71fab24bb1a12cb877645c9f8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 531, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 567 + ], + "type": "text", + "content": "ensuring the predicted flow " + }, + { + "bbox": [ + 313, + 531, + 555, + 567 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_t, t)" + }, + { + "bbox": [ + 313, + 531, + 555, + 567 + ], + "type": "text", + "content": " aligns with the true path " + }, + { + "bbox": [ + 313, + 531, + 555, + 567 + ], + "type": "inline_equation", + "content": "(x_1 - x_o)" + }, + { + "bbox": [ + 313, + 531, + 555, + 567 + ], + "type": "text", + "content": ". Once trained, the rectified flow is realized by solving the ODE:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 395, + 579, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 579, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 395, + 579, + 553, + 594 + ], + "type": "interline_equation", + "content": "d x _ {t} = v _ {\\hat {\\theta}} \\left(x _ {t}, t\\right) d t, \\tag {2}", + "image_path": "ba5d08efaf109541244a5f251bba3335b3db53709a625a3dfbd3444f5b3770c9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "content": "leading " + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "inline_equation", + "content": "x_0 \\sim p_0" + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "content": " toward " + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "inline_equation", + "content": "x_1 \\sim p_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "content": ". To enable one-step sampling, we distill this multi-step process into a neural network " + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{N}}: \\mathbb{R}^d \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{T}}(x_0) = x_0 + v(x_0, 0)" + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "content": ", is trained to directly predict " + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 603, + 555, + 652 + ], + "type": "text", + "content": ". The loss function," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 661, + 553, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 661, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 361, + 661, + 553, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} \\left[ \\left\\| \\left(x _ {1} - x _ {0}\\right) - v \\left(x _ {0}, 0\\right) \\right\\| ^ {2} \\right], \\tag {3}", + "image_path": "d5909985351b5916d1bfdecbc2bab05ffc9a8f14ce1e104550a014a55a4d9c46.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 700, + 541, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 700, + 541, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 700, + 541, + 714 + ], + "type": "text", + "content": "drives " + }, + { + "bbox": [ + 313, + 700, + 541, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\hat{x}_0)" + }, + { + "bbox": [ + 313, + 700, + 541, + 714 + ], + "type": "text", + "content": " to approximate the final, rectified state " + }, + { + "bbox": [ + 313, + 700, + 541, + 714 + ], + "type": "inline_equation", + "content": "x_{1}" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22774" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 261, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 261, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 261, + 83 + ], + "type": "text", + "content": "3.2. Rectified Flow with Tweedie's Formula" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": "Though rectified flow refines sampling, medical data distributions often need extra bias correction. Tweedie's Formula [43] addresses this by adjusting noisy observations to better approximate the posterior mean. For Gaussian variables " + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "inline_equation", + "content": "z \\sim \\mathcal{N}(\\mu_z, \\Sigma_z)" + }, + { + "bbox": [ + 55, + 89, + 296, + 150 + ], + "type": "text", + "content": ", Tweedie's formula indicates:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 157, + 240, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 157, + 240, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 157, + 240, + 171 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\mu_ {z} \\mid z ] = z + \\Sigma_ {z} \\nabla_ {z} \\log p (z)", + "image_path": "ca52dde54c3c98ddc13d47aded086f1c24ece4e6c50e0309b4bcfb13079ec4ee.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "inline_equation", + "content": "\\nabla_z\\log p(z)" + }, + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "text", + "content": " is the gradient of the log probability. In typical diffusion, Tweedie's formula estimates " + }, + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "text", + "content": " from noisy " + }, + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 178, + 295, + 213 + ], + "type": "text", + "content": ", guiding predictions toward the data manifold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 214, + 295, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 236 + ], + "type": "text", + "content": "ViCTR incorporates this correction into the rectified flow ODE:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 246, + 269, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 246, + 269, + 259 + ], + "spans": [ + { + "bbox": [ + 81, + 246, + 269, + 259 + ], + "type": "interline_equation", + "content": "d x _ {t} = v _ {\\hat {\\theta}} (x _ {t}, t) d t + (1 - \\bar {\\alpha} _ {t}) \\nabla_ {x _ {t}} \\log p (x _ {t}) d t,", + "image_path": "68bde1eb2e63281c23a4c3e68d5c6adf02154d912ae817111f6e9cfeee1fde8b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 266, + 206, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 206, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 206, + 278 + ], + "type": "text", + "content": "and updates the training objective to:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 285, + 294, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 294, + 340 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\theta} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t \\sim \\operatorname {U n i f o r m} (0, 1)} \\left[ \\| \\left(x _ {1} - x _ {0}\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\right. \\\\ \\left. \\left. - \\left(1 - \\bar {\\alpha} _ {t}\\right) \\nabla_ {x _ {t}} \\log p (x _ {t}) \\right| ^ {2} \\right]. \\\\ \\end{array}", + "image_path": "9d3bdc084f2db3c981a200287e0adc26bcfcdf1f2671f8d4531ba52dc9e86727.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "text", + "content": " denotes the cumulative product of variance decay factors in the diffusion schedule. The additional Tweedie term " + }, + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "inline_equation", + "content": "(1 - \\bar{\\alpha}_t)\\nabla_{x_t}\\log p(x_t)" + }, + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "text", + "content": " corrects for sampling bias, driving " + }, + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 346, + 296, + 393 + ], + "type": "text", + "content": " more accurately toward the target distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "type": "text", + "content": "We similarly extend the one-step distillation to integrate Tweedie's correction:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 73, + 425, + 277, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 425, + 277, + 439 + ], + "spans": [ + { + "bbox": [ + 73, + 425, + 277, + 439 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {T}} (x _ {0}) = x _ {0} + v (x _ {0}, 0) + (1 - \\bar {\\alpha} _ {0}) \\nabla_ {x _ {0}} \\log p (x _ {0}),", + "image_path": "8537d7aa578164478bd42a6f884f7950699512eb81b939f5ed00ae535f353805.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 446, + 111, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 111, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 111, + 458 + ], + "type": "text", + "content": "and optimize" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 122, + 464, + 230, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 464, + 230, + 492 + ], + "spans": [ + { + "bbox": [ + 122, + 464, + 230, + 492 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} \\left[ \\left\\| x _ {1} - \\hat {\\mathcal {T}} (x _ {0}) \\right\\| ^ {2} \\right].", + "image_path": "816ad087e01a54e79649b4f44f45fdb3e69922f2849f9e0945829ad9357f788e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "text", + "content": "This single-step approach balances computational efficiency with high-quality, pathology-aware generation—critical in medical applications where datasets are limited and synthetic realism is paramount." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 617 + ], + "type": "text", + "content": "In summary, ViCTr unifies rectified flow and Tweedie's correction to deliver anatomically consistent, pathology-aware sampling in a single forward pass. The resulting framework exhibits reduced inference costs, minimized sampling bias, and improved fidelity for generating high-resolution medical images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 624, + 236, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 236, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 236, + 637 + ], + "type": "text", + "content": "3.3. Pathology Aware Image Synthesis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Our proposed ViCTr method adopts a two-stage training paradigm: Stage 1 establishes a foundational diffusion model tailored to medical imaging data, and Stage 2 fine-tunes this pre-trained model for downstream tasks, including semantic-guided multi-modal generation and counterfactual pathology synthesis." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 72, + 506, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 506, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 506, + 84 + ], + "type": "text", + "content": "Pre-training on ATLAS-8k Dataset (Stage 1)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "type": "text", + "content": "In the absence of large-scale, domain-specific pre-trained diffusion models, we begin by training on the ATLAS8k [40] dataset, which comprises abdominal CT scans and their segmentation annotations. We leverage these annotations as conditional guidance—enabling precise control over anatomical structures—and integrate textual prompts (e.g., \"create image having " + }, + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\langle \\langle \\text{organs} \\rangle \\rangle" + }, + { + "bbox": [ + 313, + 87, + 555, + 182 + ], + "type": "text", + "content": "\" to further refine semantic consistency." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": "Latent Representations. Following Figure 1, we feed raw CT images " + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "inline_equation", + "content": "(X_I)" + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": " and segmentation masks " + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "inline_equation", + "content": "(X_S)" + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": " into a frozen VAE encoder, producing latent embeddings " + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "inline_equation", + "content": "(Z_o" + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "inline_equation", + "content": "Z_s)" + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": ". Simultaneously, textual prompts " + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "inline_equation", + "content": "(X_p)" + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": " are processed by a pre-trained text encoder (details in Table 1) to yield prompt embeddings " + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "inline_equation", + "content": "(Z_p)" + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": ". These latent representations guide the generative diffusion backbone " + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 313, + 183, + 555, + 277 + ], + "type": "text", + "content": " during both forward and reverse diffusion." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 279, + 554, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 279, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 314, + 279, + 554, + 303 + ], + "type": "text", + "content": "Forward Diffusion. We progressively inject noise into " + }, + { + "bbox": [ + 314, + 279, + 554, + 303 + ], + "type": "inline_equation", + "content": "Z_{o}" + }, + { + "bbox": [ + 314, + 279, + 554, + 303 + ], + "type": "text", + "content": " forming a noisy representation " + }, + { + "bbox": [ + 314, + 279, + 554, + 303 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 314, + 279, + 554, + 303 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 361, + 311, + 506, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 311, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 361, + 311, + 506, + 324 + ], + "type": "interline_equation", + "content": "P (Z _ {t} | Z _ {o}) = (1 - t) \\cdot Z _ {o} + t \\cdot \\epsilon_ {\\mathrm {t r u e}},", + "image_path": "b6dd0d9f6592c4d73c4fa37e9c86d711f856d5029f3045d73bf9aa2ae6dca7cb.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\epsilon_{true}" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": " is the true noise at step " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": ". We concatenate " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "Z_{s}" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": " and feed them into the diffusion model " + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 313, + 332, + 554, + 392 + ], + "type": "text", + "content": ". Crucial layers are selectively unfrozen based on elastic weight consolidation, maintaining model stability while adapting to medical domain specifics." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 326, + 392, + 499, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 392, + 499, + 404 + ], + "spans": [ + { + "bbox": [ + 326, + 392, + 499, + 404 + ], + "type": "text", + "content": "Reverse Diffusion. In the reverse process," + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 319, + 413, + 555, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 413, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 319, + 413, + 555, + 426 + ], + "type": "interline_equation", + "content": "P \\left(Z _ {t - 1} \\mid Z _ {t}, Z _ {s}, Z _ {p}, t\\right) = Z _ {t} + \\delta T \\times \\phi_ {\\theta} \\left(Z _ {t}, Z _ {s}, Z _ {p}, t\\right), \\tag {4}", + "image_path": "8db47576078b7efe28e24e7b16d3999c688a0727f4cf57cf51eac5ab0fb874f2.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": "iteratively removes noise to reconstruct " + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "inline_equation", + "content": "Z_{t-1}" + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": " conditioned on the segmentation map " + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "inline_equation", + "content": "(Z_s)" + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": " and text prompt " + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "inline_equation", + "content": "(Z_s)" + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": ". The diffusion loss " + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "inline_equation", + "content": "L_{diff}" + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "inline_equation", + "content": "Z_{t-1}" + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\epsilon_{true}" + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 342, + 479, + 554, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 479, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 342, + 479, + 554, + 494 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i f f}} = - \\left| \\phi_ {\\theta} \\left(Z _ {t}, Z _ {s}, Z _ {p}, t\\right) - \\left(\\epsilon_ {\\text {t r u e}} - Z _ {o}\\right) \\right| ^ {2}. \\tag {5}", + "image_path": "999fcb2b008a244ec422a0802f98a092673ba8ba321e72ba2e2c3db1cde3174b.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "text", + "content": "After refining " + }, + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "inline_equation", + "content": "Z_{t-1}" + }, + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "text", + "content": " via Tweedie's formula, the denoised latent " + }, + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "inline_equation", + "content": "Z_o" + }, + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "text", + "content": " is fed into the frozen VAE decoder (matching " + }, + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "inline_equation", + "content": "\\phi_\\theta" + }, + { + "bbox": [ + 313, + 502, + 554, + 537 + ], + "type": "text", + "content": "; see Table 1) to obtain a reconstructed CT image." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "content": "Composite Loss. To ensure both local accuracy and global perceptual realism, we combine: Diffusion Loss " + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "inline_equation", + "content": "(L_{\\mathrm{diff}})" + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "content": ", pixel-level reconstruction loss " + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "inline_equation", + "content": "(L_2)" + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "content": " evaluating the gap between generated " + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "inline_equation", + "content": "X_{ig}" + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "content": " and ground-truth " + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "inline_equation", + "content": "X_I" + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "content": ", and structural similarity (SSIM) loss " + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "inline_equation", + "content": "(L_{SSIM})" + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "content": " emphasizing higher-order textural correspondence. Optimizing this composite loss enables the diffusion backbone " + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 313, + 538, + 555, + 645 + ], + "type": "text", + "content": " to capture essential anatomical features, establishing a versatile medical foundation model ready for fine-tuning." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 651, + 499, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 499, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 499, + 662 + ], + "type": "text", + "content": "Fine-tuning on downstream tasks (Stage 2)" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Building on the Stage1 pre-trained model, Stage2 targets specialized tasks like semantic-guided CT/MRI generation and counterfactual pathology synthesis. As shown in Figure 1, a dual-network setup is adopted." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22775" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 142 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 293, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 293, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 293, + 95 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 55, + 72, + 293, + 95 + ], + "type": "inline_equation", + "content": "\\phi_{\\text{base}}" + }, + { + "bbox": [ + 55, + 72, + 293, + 95 + ], + "type": "text", + "content": " remains frozen, retaining the robust anatomical knowledge acquired in Stage 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 294, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 294, + 142 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 294, + 142 + ], + "type": "text", + "content": "2. " + }, + { + "bbox": [ + 55, + 96, + 294, + 142 + ], + "type": "inline_equation", + "content": "\\phi_{\\text {adapt }}" + }, + { + "bbox": [ + 55, + 96, + 294, + 142 + ], + "type": "text", + "content": " is augmented with LoRA [19] modules within the previously fine-tuned layers. These modules are the only trainable parameters, ensuring targeted and stable adaptation." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": "Training Input Consistency Loss. Similar to Stage1, each training sample includes prompts, segmentation maps, and either CT or MRI images (depending on the task). A frozen VAE encoder extracts latent representations " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "Z_{s}, Z_{o}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "Z_{p}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": ". Following forward diffusion of " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "Z_{o}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": " into noisy " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": ", both " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "\\phi_{\\text{base}}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "\\phi_{\\text{adapt}}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": " process " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "[Z_{t}, Z_{s}]" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": " with text embedding " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "Z_{p}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": ". We introduce a consistency loss " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "L_{\\text{consistency}}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": " to align intermediate outputs of " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "\\phi_{\\text{base}}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "inline_equation", + "content": "\\phi_{\\text{adapt}}" + }, + { + "bbox": [ + 55, + 148, + 295, + 255 + ], + "type": "text", + "content": ", ensuring stable adaptation without drifting from core representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": "Loss Components. Diffusion loss " + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "inline_equation", + "content": "(L_{diff}^{\\mathrm{base}}" + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "inline_equation", + "content": "L_{diff}^{\\mathrm{adapt}})" + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": " encourages accurate noise prediction. Temporal consistency loss " + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "inline_equation", + "content": "L_{temporal}" + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": " preserves smooth transitions across time steps in the reverse diffusion. Spatial consistency loss " + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "inline_equation", + "content": "L_{spatial}" + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": " ensures alignment in the output reconstructions " + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "inline_equation", + "content": "(Z_{t-1}^{base}, (Z_{t-1}^{adapt})" + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": ". Pixel-level losses " + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "inline_equation", + "content": "(L_2" + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "inline_equation", + "content": "L_{SSIM})" + }, + { + "bbox": [ + 55, + 258, + 295, + 356 + ], + "type": "text", + "content": " compare generated results to ground-truth scans, enforcing both local fidelity and structural coherence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "text", + "content": "CT and MRI Generation. We adapt to BTCV [15] (CT) and AMOS [23] (MRI) datasets, using organ segmentation masks for prompts. A standard 80/10/10 partition ensures robust evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "content": "Pathology Generation. For liver cirrhosis, we employ CurrMRI600+ [22] dataset (T1/T2 MRIs) with segmentation masks marking liver regions. Prompts specify \"low,\" \"mild,\" or \"severe\" cirrhosis intensity. As in the CT/MRI setting, the training/validation/testing split is 80/10/10. The inference details and inference diagram are in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "type": "text", + "content": "This selective LoRA-based fine-tuning allows ViCTR to handle complex tasks—ranging from normal organ generation to pathology synthesis—while preserving the anatomical and text-driven guidance learned in Stage 1. The experiments (Section 4) demonstrate that this two-stage framework consistently yields high-quality, clinically realistic images in both CT and MRI settings, including nuanced liver cirrhosis modeling." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 56, + 620, + 314, + 680 + ], + "blocks": [ + { + "bbox": [ + 56, + 620, + 314, + 680 + ], + "lines": [ + { + "bbox": [ + 56, + 620, + 314, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 314, + 680 + ], + "type": "table", + "html": "
Diffusion MethodsDenoiserText Encoder
Stable Diffusion [44]UNet [45]Clip-B/16 [41]
Pixart-alpha [9]Transformer (DiT) [38]T5-XXXXL [42]
Stable Diffusion XL [39]Dual Unet [39]Clip-L/14
Flux [27]MultiModal Transformer [14]T5-XXXXL + Clip-L/15
Stable Diffusion-3 [14]MultiModal TransformerT5-XXXXL + Clip-B/16 + Clip-L/14
", + "image_path": "2e5e90ccb33f0594f2e211f514deffece000b187c93d01c18f053ef3c05b10cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "type": "text", + "content": "Table 1. Selection of denoisers and text encoders for different diffusion methods " + }, + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "type": "inline_equation", + "content": "{\\phi }_{\\theta }" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 72, + 457, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 457, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 457, + 84 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 553, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 115 + ], + "type": "text", + "content": "The implementation details for training and sampling are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 121, + 432, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 432, + 134 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 432, + 134 + ], + "type": "text", + "content": "4.1. Quantitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 139, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 553, + 295 + ], + "type": "text", + "content": "Synthetic Data Generation Results. Table 2 summarizes our comparative evaluation of synthetic medical image generation under vanilla fine-tuning versus the proposed ViCTR pipeline, spanning multiple datasets (BTCV, AMOS, Cir-rMRI600+) and diffusion backbones (Stable Diffusion, SD-XL, SD-3, Pixart-alpha, Flux). While vanilla fine-tuning applies a direct end-to-end approach after ATLAS-8k pretraining, it generally yields higher Fréchet Inception Distance (FID) and Medical FID (MFID) values across all datasets. Conversely, ViCTR integrates domain-specific loss functions and specialized architectural settings, consistently achieving lower FID/MFID scores and producing more realistic medical images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 295, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 437 + ], + "type": "text", + "content": "To address the well-known limitation of FID (which uses ImageNet features), we adopt M3D-CLIP [3] to compute a more domain-specific MFID. As shown in the table, ViCTR provides notable gains: for example, Stable Diffusion under ViCTR achieves FID/MFID of 21.98/19.02 (BTCV), 20.37/19.11 (AMOS), and 25.57/21.46 (Cir-rMRI600+). These results represent substantial improvements over vanilla fine-tuning and demonstrate consistent performance boosts across all tested architectures. Such gains highlight the robustness of ViCTR's two-stage framework in generating clinically meaningful synthetic images for both CT and MRI modalities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 438, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 553, + 677 + ], + "type": "text", + "content": "Segmentation Results. To assess the quality and utility of ViCTR-generated synthetic data, we performed segmentation experiments on three datasets—BTCV, AMOS, and CurrMRI600+—tracking mean Dice Similarity Coefficient (mDSC) and mean Hausdorff Distance 95 (mHD95). Higher mDSC and lower mHD95 respectively indicate improved overlap accuracy and spatial precision. We examined four training configurations: (1) baseline using original data only, (2) standard augmentation (Random Crop, Rotate, Blur, Affine, Geometric Distortion at 0.4 probability), (3) " + }, + { + "bbox": [ + 313, + 438, + 553, + 677 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 438, + 553, + 677 + ], + "type": "text", + "content": " synthetic data via vanilla fine-tuning, and (4) " + }, + { + "bbox": [ + 313, + 438, + 553, + 677 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 438, + 553, + 677 + ], + "type": "text", + "content": " synthetic data from ViCTR. Table 3 highlights consistent performance gains from ViCTR across all datasets and backbone architectures. Notably, CurrMRI600+—a liver cirrhosis dataset—exhibited the most significant improvements, reflecting ViCTR's capacity to capture complex pathological features. These results confirm that ViCTR-generated images significantly boost segmentation metrics (mDSC and mHD95), highlighting the practical value of our two-stage framework for medical data augmentation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Efficiency of ViCTR. Table 2 summarizes the number of diffusion steps and average inference times (in seconds) for vanilla fine-tuning versus the ViCTR framework. Across all" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22776" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 599, + 130 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 599, + 130 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 599, + 130 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 599, + 130 + ], + "type": "table", + "html": "
BaselinesBTCV Dataset (CT Generation)AMOS Dataset (MRI Generation)CirrMRI600+ (Pathology Generation)Diffusion Steps | Inference Time
Fine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTRFine-tuned VanillaViCTR
Stable Diffusion [44]25.44 / 19.6721.98 / 19.0225.43 / 21.7620.37 / 19.1128.34 / 23.4325.57 / 21.4640 | 13.764 | 3.12
Stable Diffusion XL [39]23.47 / 18.2120.33 / 17.4424.11 / 20.2319.44 / 18.4527.34 / 22.1124.02 / 20.7630 | 14.554 | 3.45
Stable Diffusion-3 [14]19.07 / 16.2217.37 / 16.0222.32 / 19.7618.02 / 19.0824.49 / 21.7821.28 / 19.3450 | 18.983 | 2.78
Pixart-alpha [9]21.32 / 17.0919.22 / 16.9623.78 / 20.0418.76 / 18.5626.06 / 20.0723.04 / 18.9225 | 10.673 | 1.74
Flux [27]15.52 / 15.0113.28 / 14.0819.02 / 18.2815.55 / 16.5822.46 / 18.8819.96 / 17.0130 | 15.663 | 2.87
", + "image_path": "a89e077bf36c8f0e36f41f6e4c2714390fff6922da25f7f2eb632d3e71c9732d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 182, + 577, + 396 + ], + "blocks": [ + { + "bbox": [ + 55, + 137, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 555, + 171 + ], + "type": "text", + "content": "Table 2. Quantitative results on CT, MRI, and Cirrhosis generation. All baselines were pre-trained on ATLAS-8k and fine-tuned on target datasets. Metrics reported are Fréchet Inception Distance (FID) and Medical FID (MFID), shown as FID/MFID, with inference time in seconds. Lower values indicate better performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 182, + 577, + 396 + ], + "lines": [ + { + "bbox": [ + 56, + 182, + 577, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 577, + 396 + ], + "type": "table", + "html": "
BaselinesOriginal Dataset (Org.)AugmentationOrg. + Synth. by FLUX 30% (vanilla Fine-tuning)Org. + Synth. by FLUX 30% (ViCtR)
mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)mDSC (↑)mHD95 (↓)
BTCV dataset Segmentation results
UNet [45]76.7234.4278.4533.1779.3232.3781.2230.17
TransUnet [8]85.5232.3387.0131.4387.5430.1189.7829.12
nnUnet [20]80.4830.1982.5431.0283.3729.0185.1927.77
nnFormer [56]83.4730.0183.9830.8885.8828.7887.7226.22
MedSegDiff [52]87.9127.6788.6526.7289.7825.5291.9223.31
AMOS dataset Segmentation results
UNet [45]68.9234.5770.5532.3271.0231.7873.3429.11
TransUnet [8]71.3333.1272.5631.0973.4329.1977.5427.56
nnUnet [20]73.3332.3274.1130.1975.6828.7578.2926.32
nnFormer [56]75.7831.1976.4429.7677.7727.5781.3224.41
MedSegDiff [52]76.8329.9278.3826.9979.0325.4584.0222.18
CirrMRI600+ dataset Segmentation results
UNet [45]68.7436.7369.3835.4370.1235.1173.3932.11
TransUnet [8]70.7735.4270.9834.0171.7833.9274.5631.09
nnUnet [20]71.0234.3572.4932.7873.5631.2778.8930.25
nnFormer [56]74.8833.7875.2331.8876.4531.0979.4429.78
MedSegDiff [52]76.9230.7977.1130.3478.0329.8981.3727.34
", + "image_path": "896a7a5a02b0846a4023b5694836fe5a196965db1ec2c2a85690fbcec07d828f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 427 + ], + "type": "text", + "content": "Table 3. Segmentation results on BTCV, AMOS, and CurrMRI600+ using various training settings and baselines. Metrics reported are mDSC (%) and mHD95 (mm). Top three results per setting are highlighted: blue (1st), green (2nd), and orange (3rd)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 447, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 447, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 295, + 590 + ], + "type": "text", + "content": "evaluated models, ViCTR consistently reduces the required steps, thereby shortening inference duration without compromising image fidelity. This efficiency stems from two key innovations: (1) Tweedy's formula to approximate " + }, + { + "bbox": [ + 54, + 447, + 295, + 590 + ], + "type": "inline_equation", + "content": "Z_{0}" + }, + { + "bbox": [ + 54, + 447, + 295, + 590 + ], + "type": "text", + "content": " and streamlines the reconstruction process, and (2) Two-Stage Training, enabling the model to learn both anatomical consistency and pathology-specific variations in a targeted manner. By minimizing extraneous steps, ViCTR can generate high-quality synthetic images more rapidly, making it a viable solution for resource-constrained clinical or research environments where computational overhead is a critical consideration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 167, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 167, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 167, + 611 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": "We conducted extensive qualitative evaluations of ViCTr across multiple modalities (CT and MRI) and tasks, covering both anatomical and pathological image synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Anatomical and Pathology-Driven Image Generation. Unlike prior methods that primarily generate normal anatomical scans or isolated tumors, ViCTr extends synthesis capabilities to complex pathologies, such as liver cirrhosis. Figure 3 shows pairs of non-cirrhotic MRIs alongside" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 447, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 555 + ], + "type": "text", + "content": "their segmentation masks and the corresponding ViCTR-generated cirrhotic images. Notably, the synthesized images retain essential anatomical structures while introducing realistic cirrhotic texturing, an advancement that enables more diverse data augmentation and supports research into disease progression. These high-fidelity synthetic samples offer potential to enrich clinical training datasets, advance diagnostic algorithms, and support detailed analyses of pathology progression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 640 + ], + "type": "text", + "content": "Qualitative Segmentation Outcomes. To further showcase ViCTR's impact on downstream tasks, we visually assessed segmentation quality on BTCV and AMOS (Figure 6). Focusing on regions like the left kidney and pancreas, models trained with ViCTR-augmented data show sharper boundaries and more precise delineation than standard fine-tuning, indicating improved spatial learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Comparative Performance on BTCV and AMOS. Figures 4 (BTCV) and 5 (AMOS) compare diffusion-based image generation results using Stable Diffusion v3 (SD-3) and Flux backbones. On BTCV—a dataset with limited samples and larger spatial dimensions—vanilla finetuning with SD-3 often yields poor mask adherence, sug" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22777" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 293, + 216 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 293, + 216 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 293, + 216 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 293, + 216 + ], + "type": "image", + "image_path": "acbff25d698bee988afcb9bd7ad92bcea0f82dd02abecb03a68b130a14859499.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 270, + 294, + 428 + ], + "blocks": [ + { + "bbox": [ + 55, + 224, + 295, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 269 + ], + "type": "text", + "content": "Figure 3. Examples of synthetic cirrhotic images with different severity levels generated from non-cirrhotic scans using ViCTr. For T1-MRI, we show the segmentation mask, the original non-cirrhotic image, and the corresponding synthetic cirrhotic image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 270, + 294, + 428 + ], + "lines": [ + { + "bbox": [ + 59, + 270, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 294, + 428 + ], + "type": "image", + "image_path": "2b4919450eea532967b234637106888ca96f72a6a0afc07acf66fdedeb80ccae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 481, + 294, + 654 + ], + "blocks": [ + { + "bbox": [ + 55, + 436, + 295, + 479 + ], + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 479 + ], + "type": "text", + "content": "Figure 4. Diffusion-Based Image Generation on BTCV MRI Dataset. From left to right: segmentation mask, SD-3 with standard fine-tuning, FLUX with standard fine-tuning, SD-3 with ViCTR, FLUX with ViCTR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 481, + 294, + 654 + ], + "lines": [ + { + "bbox": [ + 59, + 481, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 59, + 481, + 294, + 654 + ], + "type": "image", + "image_path": "5590f8ff02fb0f7fd96cf80c038fe2b42c2fbfee3326fd783f47ea34103b500d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 665, + 295, + 708 + ], + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 708 + ], + "type": "text", + "content": "Figure 5. Diffusion-Based Image Generation on AMOS Dataset. From left to right: segmentation mask, SD-3 with Vanilla fine-tuning, SD-3 with our method, FLUX with Vanilla fine-tuning, FLUX with our method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "gesting difficulties in learning accurate spatial constraints. In contrast, the same model shows markedly improved mask alignment on the AMOS dataset, attributed to its larger training set and more manageable image dimensions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "type": "text", + "content": "Regardless of these dataset-specific variations, ViCTr achieves consistently high-quality outputs with robust mask adherence and anatomical accuracy across both BTCV and AMOS, underscoring its enhanced spatial learning and resilience. By maintaining performance even under data-sparse conditions, ViCTr demonstrates a clear advantage over conventional fine-tuning approaches, which can exhibit greater sensitivity to dataset size and complexity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": "Second, we evaluate the impact of ViCTR on segmentation tasks, illustrated in Figure 6. Both BTCV and AMOS datasets encompass various abdominal structures (e.g., left kidney, pancreas), and our results highlight how ViCTR-based training yields notably more precise organ boundaries compared to alternative methods. This improvement is particularly evident in challenging regions where standard segmentation often struggle, highlighting ViCTR's strength in enhancing spatial delineation and anatomical accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 329, + 414, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 329, + 414, + 340 + ], + "spans": [ + { + "bbox": [ + 314, + 329, + 414, + 340 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 353, + 552, + 454 + ], + "blocks": [ + { + "bbox": [ + 315, + 353, + 552, + 454 + ], + "lines": [ + { + "bbox": [ + 315, + 353, + 552, + 454 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 552, + 454 + ], + "type": "table", + "html": "
Ablation Based on Pre-training of Model
ViCTR (Without Stage-1)17.33
Ablation Based on Rectified Flow Algorithms
ViCTR (Without Proposed Tweedies Formula)18.78
ViCTR (With Reflow)20.19
ViCTR (With Flow Straight and Fast)21.37
ViCTR (With Distribution Matching Distillation)22.33
Ablation Study Based on Loss FunctionsAblation Based on LoRA Rank
Ldiff18.77r = 818.46
Ldiff + Lspatial18.21r = 1617.52
Ldiff + Lconsistency17.02r = 3216.66
Ldiff + Lspatial + Lconsistency15.55r = 6415.55
", + "image_path": "fb287f5517509101310520e15534411eaf9ab1f7efdaebcb58e4a63e24616151.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 553, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 485 + ], + "type": "text", + "content": "Table 4. Ablation study for the effect of model pretraining, loss functions, and LoRA rank settings are shown." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 498, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 592 + ], + "type": "text", + "content": "To rigorously validate our ViCTr framework, we conducted a series of ablation experiments on the BTCV dataset (Table 4), focusing on four key aspects: pretraining, rectified flow algorithms, loss function components, and LoRA rank selection. We used FID (Frechet Inception Distance) as our primary metric to gauge how each architectural choice affects both anatomical fidelity and pathological realism." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": "Impact of Pretraining. We first examined the necessity of Stage 1 pretraining by comparing our full, two-stage pipeline to a variant that proceeds directly to downstream fine-tuning. Omitting Stage 1 degrades the model's FID score from 15.55 to 17.33 (Table 4), underscoring the vital role of an anatomical prior. These findings affirm that establishing robust structural representations in Stage 1 is crucial for achieving high-quality medical image synthesis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Rectified Flow Algorithms. We next assessed our rectified flow formulation, augmented with Tweedie's formula," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22778" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 555, + 274 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 274 + ], + "type": "image", + "image_path": "f77422e60a89252daac2dd04e67c939d78ee91f5c636912898c7b360e493bc01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "type": "text", + "content": "Figure 6. Segmentation results on BTCV and AMOS datasets' left kidney and pancreas. We compare the ground truth (blue), models trained on the default dataset, augmented dataset, default dataset plus " + }, + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "type": "text", + "content": " synthetic data generated by FLUX with standard fine-tuning, and default dataset plus " + }, + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 281, + 555, + 316 + ], + "type": "text", + "content": " synthetic data generated by FLUX with ViCTr (red)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 332, + 294, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 332, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 332, + 294, + 429 + ], + "type": "text", + "content": "against alternative rectified flow-based models, including ReFlow, Flow Straight and Fast, and Distribution Matching Distillation. Removing Tweedie's correction alone increased the FID to 18.78, highlighting its significance for trajectory alignment. Across all baselines, we observed FID deteriorations, ranging from 13.28 to as high as 22.33, thereby demonstrating that our proposed method provides tighter distribution matching and superior synthesis fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 431, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 431, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 54, + 431, + 295, + 587 + ], + "type": "text", + "content": "Ablation Study Based on Loss Functions. We then investigated each ViCTr loss component—diffusion loss, spatial consistency, and consistency loss—to determine their relative contributions: the baseline model, utilizing only diffusion loss, produced an FID of 18.77 (Table 4). Incorporating spatial consistency loss yielded a moderate improvement to 18.21, highlighting its role in maintaining structural integrity. The addition of consistency loss further enhanced performance, reducing the FID to 17.02 by promoting image coherence across generations. When combining all three terms, we achieved an FID of 15.55, indicating that these components collectively offer significant improvements in anatomical accuracy and visual realism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "text", + "content": "Ablation Based on LoRA Rank. Finally, we explored how varying the LoRA rank " + }, + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "text", + "content": " influences generative performance. Starting at " + }, + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "inline_equation", + "content": "r = 8" + }, + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "text", + "content": " and incrementing upwards, we found " + }, + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "inline_equation", + "content": "r = 64" + }, + { + "bbox": [ + 54, + 590, + 295, + 675 + ], + "type": "text", + "content": " to yield the best FID at 15.55. This result suggests that higher-dimensional adaptation spaces enable more nuanced parameter updates during fine-tuning, thereby improving visual quality and fidelity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 295, + 715 + ], + "type": "text", + "content": "Visual Turing Tests. To further validate clinical plausibility, three radiologists participated in Visual Turing Tests using 15 randomly generated MRI scans depicting vary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "content": "ing levels of liver cirrhosis (mild, moderate, severe). All scans were uniformly judged to be clinically realistic, with identical outcomes observed when using additional random samples. Notably, the synthetic cirrhotic images correctly exhibited surface nodularity and textural irregularities, consistent with radiologic findings in mild-to-severe cirrhosis (Figure 3). These results confirm ViCTr's ability to synthesize pathology-specific features, enabling applications in training, algorithm development, and clinical research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 455, + 388, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 388, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 388, + 467 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 476, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 681 + ], + "type": "text", + "content": "We presented ViCTR, a novel two-stage framework designed to generate high-fidelity medical images by integrating robust anatomical pre-training with precise pathology-specific fine-tuning. By aligning Tweedie's formula with linear projection methods used in flow matching, ViCTR maintains accurate initial distribution estimates even amid diffusion processes. This setup is further enhanced by LoRA adapters, which preserve essential anatomical information while flexibly adapting to diverse pathologies. Experimental results demonstrate that ViCTR consistently surpasses conventional fine-tuning strategies, reducing FID scores and producing clinically realistic outputs. Beyond data augmentation for segmentation and classification, ViCTR can be extended to tasks such as modality translation, contrast synthesis, and professional training, offering a powerful and versatile tool for advancing AI-driven medical imaging." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 693, + 536, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 693, + 536, + 703 + ], + "spans": [ + { + "bbox": [ + 328, + 693, + 536, + 703 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 328, + 693, + 536, + 703 + ], + "type": "text", + "content": "Code: https://github.com/Onkarsus13/ViCTR-2D" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 703, + 549, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 703, + 549, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 703, + 549, + 713 + ], + "type": "text", + "content": "Weights: https://huggingface.co/onkarsus13/ViCTR-2D" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22779" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 115 + ], + "type": "text", + "content": "This research is supported by the following NIH grants: R01-HL171376 and U01-CA268808." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 202 + ], + "type": "text", + "content": "[1] Hazrat Ali, Shafaq Murad, and Zubair Shah. Spot the fake lungs: Generating synthetic medical images using neural diffusion models. In Artificial Intelligence and Cognitive Science, pages 32-39, Cham, 2023. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 205, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 260 + ], + "type": "text", + "content": "[2] Alhanoof Althnian, Duaa Alsaeed, Heyam Al-Baity, Amani Samha, Alanoud Bin Dris, Najla Alzakari, Afnan Abou Elwafa, and Heba Kurdi. Impact of dataset size on classification performance: an empirical evaluation in the medical domain. Applied Sciences, 11(2):796, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 262, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 295, + 294 + ], + "type": "text", + "content": "[3] Fan Bai, Yuxin Du, Tiejun Huang, Max Q. H. Meng, and Bo Zhao. M3d: Advancing 3d medical image analysis with multi-modal large language models, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 296, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 296, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 61, + 296, + 296, + 339 + ], + "type": "text", + "content": "[4] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. In ICLR 2023 Workshop on Trustworthy and Reliable Large-Scale Machine Learning Models, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 342, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 342, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 342, + 296, + 396 + ], + "type": "text", + "content": "[5] Fadi Boutros, Jonas Henry Grebe, Arjan Kuijper, and Naser Damer. Idiff-face: Synthetic-based face recognition through fizzy identity-conditioned diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19650-19661, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 430 + ], + "type": "text", + "content": "[6] Luigi Carratino, Moustapha Cisse, Rodolphe Jenatton, and Jean-Philippe Vert. On mixup regularization. Journal of Machine Learning Research, 23(325):1-31, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 476 + ], + "type": "text", + "content": "[7] Junyi Chai, Hao Zeng, Anming Li, and Eric W.T. Ngai. Deep learning in computer vision: A critical review of emerging techniques and application scenarios. Machine Learning with Applications, 6:100134, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 478, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 531 + ], + "type": "text", + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 534, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 296, + 589 + ], + "type": "text", + "content": "[9] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation, 2024. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 646 + ], + "type": "text", + "content": "[10] Qi Chen, Xiaoxi Chen, Haorui Song, Zhiwei Xiong, Alan Yuille, Chen Wei, and Zongwei Zhou. Towards generalizable tumor synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11147-11158, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[11] Gayatri Deshmukh, Onkar Kishor Susladkar, Debesh Jha, Elif Keles, Halil Ertugrul Aktas, Daniela P Ladner, Amir A Borhani, Gorkem Durak, Ulas Bagci, et al. Meddelinea: Scalable and efficient medical image segmentation via controllable diffusion transformers. In Medical Imaging with Deep Learning, 2025. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "text", + "content": "[13] Lisa Dunlap, Alyssa Umino, Han Zhang, Jiezhi Yang, Joseph E. Gonzalez, and Trevor Darrell. Diversify your vision datasets with automatic diffusion-based augmentation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "text", + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[15] Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619-3629, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "text", + "content": "[16] Zhengcong Fei, Mingyuan Fan, Changqian Yu, and Junshi Huang. Flux that plays music, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[17] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[18] Yunxiang Fu, Chaoqi Chen, Yu Qiao, and Yizhou Yu. Dreamda: Generative data augmentation with diffusion models. arXiv preprint arXiv:2403.12803, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[20] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 525 + ], + "type": "text", + "content": "[21] Khawar Islam, Muhammad Zaigham Zaheer, Arif Mahmood, and Karthik Nandakumar. Diffusemix: Label-preserving data augmentation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27621-27630, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 526, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 590 + ], + "type": "text", + "content": "[22] Debesh Jha, Onkar Kishor Susladkar, Vandan Gorade, Elif Keles, Matthew Antalek, Deniz Seyithanoglu, Timurhan Cebeci, Halil Ertugrul Aktas, Gulbiz Dagoglu Kartal, Sabahattin Kaymakoglu, et al. Cirrmri600+: Large scale mri collection and segmentation of cirrhotic liver. arXiv preprint arXiv:2410.16296, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 593, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 657 + ], + "type": "text", + "content": "[23] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems, 35:36722-36732, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[24] Guillermo Jimenez-Perez, Pedro Osorio, Josef Cersovsky, Javier Montalt-Tordera, Jens Hooge, Steffen Vogler, and Sadegh Mohammadi. Dino-diffusion. scaling medical diffusion via self-supervised pre-training. arXiv preprint arXiv:2407.11594, 2024. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22780" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[25] Sunho Kim, Byungjai Kim, and HyunWook Park. Synthesis of brain tumor multicontrast mr images for improved data augmentation. Medical Physics, 48(5):2185-2198, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 171 + ], + "type": "text", + "content": "[26] Nicholas Konz, Yuwen Chen, Haoyu Dong, and Maciej A Mazurowski. Anatomically-controllable medical image generation with segmentation-guided diffusion models. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 88–98. Springer, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[27] Black Forest Labs. black-forest-labs/FLUX.1-dev " + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": " Hugging Face. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024. [Online; accessed 05-October-2024]. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 258 + ], + "type": "text", + "content": "[28] Hansang Lee, Haeil Lee, and Helen Hong. Genmix: Combining generative and mixture data augmentation for medical image classification. arXiv preprint arXiv:2405.20650, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[29] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. Advances in Neural Information Processing Systems, 37:63082-63109, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 327 + ], + "type": "text", + "content": "[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 360 + ], + "type": "text", + "content": "[31] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "text", + "content": "[32] Supriya V. Mahadevkar, Bharti Khemani, Shruti Patil, Ketan Kotecha, Deepali R. Vora, Ajith Abraham, and Lubna Abdelkareim Gabralla. A review on machine learning styles in computer vision—techniques and future directions. IEEE Access, 10:107293–107329, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 482 + ], + "type": "text", + "content": "[33] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M. Alvarez, Zhiding Yu, Sanja Fidler, and Marc T. Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 275-284, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 483, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 294, + 558 + ], + "type": "text", + "content": "[34] Xueyan Mei, Zelong Liu, Philip M Robson, Brett Marinelli, Mingqian Huang, Amish Doshi, Adam Jacobi, Chendi Cao, Katherine E Link, Thomas Yang, Ying Wang, Hayit Greenspan, Timothy Deyer, Zahi A Fayad, and Yang Yang. RadImageNet: An open radiologic deep learning research dataset for effective transfer learning. Radiol Artif Intell, 4 (5):e210315, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "type": "text", + "content": "[35] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, Jakob Nikolas Kather, and Daniel Truhn. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13(1):12098, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 637, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 713 + ], + "type": "text", + "content": "[36] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part III 20, pages 417-425. Springer, 2017. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[37] Kaliprasad Pani and Indu Chawla. Synthetic mri in action: A novel framework in data augmentation strategies for robust multi-modal brain tumor segmentation. Computers in Biology and Medicine, 183:109273, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[38] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "type": "text", + "content": "[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[40] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 326 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[42] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 414 + ], + "type": "text", + "content": "[43] Herbert E. Robbins. An Empirical Bayes Approach to Statistics, pages 388-394. Springer New York, New York, NY, 1992. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 470 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 471, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 536 + ], + "type": "text", + "content": "[45] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "text", + "content": "[46] Daniel G. Saragih, Atsuhiro Hibi, and Pascal N. Tyrrell. Using diffusion models to generate synthetic labeled data for medical image segmentation. International Journal of Computer Assisted Radiology and Surgery, 19(8):1615-1625, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "type": "text", + "content": "[47] Vanshali Sharma, Abhishek Kumar, Debesh Jha, M.K. Bhuyan, Pradip K. Das, and Ulas Bagci. Controlpolynet: Towards controlled colon polyp synthesis for improved polyp segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2325-2334, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[48] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Brandon Trabucco, Kyle Doherty, Max A Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffu" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22781" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 474 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "sion models. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[50] Toan Tran, Trung Pham, Gustavo Carneiro, Lyle Palmer, and Ian Reid. A bayesian data augmentation approach for learning deep models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 217 + ], + "type": "text", + "content": "[51] Yibo Wang, Ruiyuan Gao, Kai Chen, Kaiqiang Zhou, Yingjie Cai, Lanqing Hong, Zhenguo Li, Lihui Jiang, DitYan Yeung, Qiang Xu, et al. Detdiffusion: Synergizing generative and perceptive models for enhanced data generation and perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7246-7255, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[52] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In Medical Imaging with Deep Learning, pages 1623-1639. PMLR, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 329 + ], + "type": "text", + "content": "[53] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[54] Zheyuan Zhang, Lanhong Yao, Bin Wang, Debesh Jha, Elif Keles, Alpay Medetalibeyoglu, and Ulas Bagci. Emit-diff: Enhancing medical image segmentation via text-guided diffusion model. arXiv preprint arXiv:2310.12868, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 428 + ], + "type": "text", + "content": "[55] Yuan Zhong, Suhan Cui, Jiaqi Wang, Xiaochen Wang, Ziyi Yin, Yaqing Wang, Houping Xiao, Mengdi Huai, Ting Wang, and Fenglong Ma. MedDiffusion: Boosting Health Risk Prediction via Diffusion-based Data Augmentation, pages 499-507. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "text", + "content": "[56] Hong-Yu Zhou, Jiansen Guo, Yinghao Zhang, Lequan Yu, Liansheng Wang, and Yizhou Yu. nnformer: Interleaved transformer for volumetric segmentation. arXiv preprint arXiv:2109.03201, 2021. 6" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22782" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ViLLa_ Video Reasoning Segmentation with Large Language Model/9374d1a1-c3e0-498b-83e7-5023233fa5ab_content_list.json b/2025/ViLLa_ Video Reasoning Segmentation with Large Language Model/9374d1a1-c3e0-498b-83e7-5023233fa5ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..561145342689b16d19c544cec759f26e619e3820 --- /dev/null +++ b/2025/ViLLa_ Video Reasoning Segmentation with Large Language Model/9374d1a1-c3e0-498b-83e7-5023233fa5ab_content_list.json @@ -0,0 +1,2002 @@ +[ + { + "type": "text", + "text": "ViLLa: Video Reasoning Segmentation with Large Language Model", + "text_level": 1, + "bbox": [ + 155, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rongkun Zheng $^{1}$ Lu Qi $^{2}$ Xi Chen $^{1}$ Yi Wang $^{3,4}$", + "bbox": [ + 282, + 179, + 712, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kun Wang $^{5}$ Hengshuang Zhao $^{1*}$", + "bbox": [ + 372, + 198, + 648, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2Insta360 3Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 161, + 215, + 834, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Shanghai Innovation Institute $^{5}$ SenseTime Research", + "bbox": [ + 290, + 233, + 707, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zrk22@connect, hszhao@cs}.hku.hk", + "bbox": [ + 303, + 253, + 691, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent efforts in video reasoning segmentation (VRS) integrate large language models (LLMs) with perception models to localize and track objects via textual instructions, achieving barely satisfactory results in simple scenarios. However, they struggled to discriminate and deduce the objects from user queries in more real-world scenes featured by long durations, multiple objects, rapid motion, and heavy occlusions. In this work, we analyze the underlying causes of these limitations, and present ViLLa: Video reasoning segmentation with Large Language Model. Remarkably, our ViLLa manages to tackle these challenges through multiple core innovations: (1) a context synthesizer that dynamically encodes the user intent with video contexts for accurate reasoning, resolving ambiguities in complex queries, and (2) a hierarchical temporal synchronizer that disentangles multi-object interactions across complex temporal scenarios by modelling multi-object interactions at local and global temporal scales. To enable efficient processing of long videos, ViLLa incorporates (3) a key segment sampler that adaptively partitions long videos into shorter but semantically dense segments for less redundancy. What's more, to promote research in this unexplored area, we construct a VRS benchmark, VideoReasonSeg, featuring different complex scenarios. Our model also exhibits impressive state-of-the-art results on VideoReasonSeg, Ref-YouTube-VOS, Ref-DAVIS17, MeViS, and ReVOS. Both quantitative and qualitative experiments demonstrate that our method effectively enhances video reasoning segmentation capabilities for multimodal LLMs.", + "bbox": [ + 88, + 335, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 789, + 220, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Capitalizing on the achievements of Large Language Models (LLMs) [10, 38, 39, 75], the development of large multimodal models (LMMs) [3, 25, 32, 34, 58, 80] has notably enhanced visual perception capabilities and user-interaction", + "bbox": [ + 89, + 815, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "experiences to new heights. However, in the absence of explicit instructions specifying target objects or categories, the majority of models [1, 24, 25, 31, 32, 71, 80], including both perception models and language ones, struggle to generate detailed and granular outputs such as instance masks. Instead, these models typically generate only general textual descriptions for images or videos. This limitation hinders the practical application of multimodal systems in various scenarios, such as autonomous driving [57], image and video editing [30], robotics [48], and augmented reality [46].", + "bbox": [ + 511, + 305, + 906, + 457 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies [23, 45, 70] have investigated the application of LLMs to generate object masks in a novel reasoning segmentation task that requires complex reasoning to interpret sophisticated instructions. These approaches introduce efficient language models capable of producing multiple open-set targets while accommodating diverse reasoning complexities. However, despite their successes, these methods remain limited to image-level reasoning and do not extend fine-grained reasoning capabilities to the temporal dimension. To address this issue, researchers [4, 67] managed to equip the image reasoning segmentation model with temporal understanding ability, and propose VRS. VISA [67] employs a video understanding model [28] to select the key frame for LISA [23] to segment and subsequently utilizes an object tracker [8] to maintain temporal consistency. However, the quality of segmentation is highly dependent on the accuracy of key frame selection, which is susceptible to error accumulation through propagation as the number of input frames increases. On the other hand, VideoLISA [4] adopts a sparse dense sampling strategy to compress visual features and extends the single segmentation token to track an instance across frames, especially when the movement is rapid. Nevertheless, the sampling strategy fails to adaptively select the key video segments, making it challenging to capture the essential information from the video. Also, because a single segmentation token is unable to segment and track multiple instances across frames, the segmentation and tracking quality decrease significantly when the video scene becomes intricate (e.g., involving multiple objects with occlusion),", + "bbox": [ + 511, + 463, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23667", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as proven in various previous Video Instance Segmentation (VIS) approaches [20, 78], which utilize multiple-level of tokens to segment and track instances.", + "bbox": [ + 89, + 90, + 485, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the aforementioned challenges, we propose ViLLa, an effective and efficient LMM for video reasoning segmentation. ViLLa proficiently perceives and reasons videos with multiple objects, rapid motion, and diverse reasoning complexities (as shown in Fig. 1). Specifically, our ViLLa has three key components: Key Segment Extractor, Context Synthesizer, and Hierarchical Temporal Synchronizer. These designs put effort into solving the complex video reasoning segmentation from different perspectives. The Key Segment Extractor selects the most relevant video segments, which reduces redundant visual tokens and effectively turns long videos into shorter segments. To make the responses more pertinent to users' intentions, Context Synthesizer synthesizes frame context information from visual features to text embeddings to generate contextually enriched text and highly input-responsive visual embeddings. To build up stronger temporal relations to multi-target segmentation tokens in complex scenes, the Hierarchical Temporal Synchronizer communicates and consolidates the multi-scale segmentation embeddings across temporal dimension via attention-based fusion strategy. Tested on multiple referring video object segmentation benchmarks, our model could surpass previous state-of-the-art LLM-based methods by a large margin. In addition, to validate its effectiveness, we establish a benchmark for video reasoning segmentation evaluation, called VideoReasonSeg. Utilizing a GPT-4V [39]-aided data curation pipeline, we curate 3k high quality video samples for video reasoning segmentation spanning multiple-choice QAs and video-instruction-mask samples, and the video scenarios contain diverse reasoning complexities (as presented in the appendix), which offers convincing metrics.", + "bbox": [ + 91, + 138, + 486, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 609, + 410, + 625 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We analyze the limitations of current VRS approaches and propose a framework named ViLLa that segments and tracks the desired objects in more complex scenarios.", + "- We develop three key modules that resolves complex scenario reasoning from different perspectives, namely key segment extractor, context synthesizer, and hierarchical temporal synchronizer. Our extractor selects the most relevant video segments from the complex input video, while the synthesizer combines and infuses user intention and context information, and the synchronizer communicates multi-scale segmentation tokens spatio-temporally.", + "- We develop VideoReasonSeg, a comprehensive benchmark for VRS that spans different complex reasoning scenarios, comprising 3k videos and 15k object descriptions for instruction tuning and evaluation. This benchmark plays a crucial role in evaluating and fostering the exploration of reasoning capabilities in video-based models.", + "- We conduct extensive experiments on challenging refer" + ], + "bbox": [ + 91, + 628, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ring segmentation benchmarks, ReVOS, and our proposed VideoReasonSeg. The state-of-the-art results demonstrate the effectiveness of our method.", + "bbox": [ + 526, + 90, + 906, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 150, + 663, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Multimodal Models. Large multimodal models (LMMs) can be broadly categorized into two groups based on their utilization of large language models (LLMs). The first category comprises models [35, 56, 74] that are either trained from scratch or leverage smaller language models like BERT for text processing. These models typically employ a combination of contrastive and generative objectives to tackle a range of multimodal tasks (e.g., Coca [74]). However, their limited language understanding capacity often restricts their performance in tasks that demand massive commonsense reasoning abilities.", + "bbox": [ + 511, + 175, + 906, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The emergence of LLMs in recent years has paved the way for a new paradigm in the development of LMMs, in which LLMs are enhanced with multimodal comprehension abilities [40, 50, 51, 76]. This approach typically includes the use of adapters to align visual and textual representations within LLMs, as exemplified by Flamingo [1], BLIP-2 [25], MiniGPT-4 [80], LLaMA-Adapter [16], LLaVA [32], InstructBLIP [11], InternGPT [34], QwenVL [3], InternVideo2 [59]. Some video-related MLLM (VideoLLM) VideoChatGPT [37], and Valley [36] utilize ChatGPT [38] to generate video instruction-tuning data, aiming to enhance instruction-following capabilities for real-world video comprehension. Even though these models have demonstrated improved performance in vision-language tasks, their primary limitation lies in generating only textual outputs.", + "bbox": [ + 511, + 342, + 908, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Segmentation. Video Instance Segmentation (VIS) [69] aims to simultaneously detect, segment, and track object instances in videos based on predefined object categories. Numerous studies [26, 29, 60, 63, 79] have proposed diverse designs (such as tracking by detection [2], masked attention [7], object token association [20, 27, 65], contrastive memory [19, 72], and more) to effectively model temporal relationships. For example, IDOL [66], which is built upon Deformable-DETR [81], incorporates a contrastive learning head for associations [17]. DVIS [78] proposed a decoupling strategy by dividing VIS into image segmentation, tracking, and refinement. The Referring Video Object Segmentation (RVOS) task was introduced to segment the target instance referred to in a given text. Early works [21, 47] used the spatial-temporal memory mechanism to strengthen temporal relations. The researchers then started building offline reference models [53, 54, 61, 68] that take the whole clip as input. Currently, query-based Transformer models [5, 64] begin to dominate. MTTR [5] utilized an instance-level segmentation pipeline to determine the most suitable sequence for the referred object. ReferFormer [64] transformed text expressions into queries to attend to relevant regions.", + "bbox": [ + 511, + 569, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23668", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "USER: